@slexisvn/mlfw 0.1.0 → 0.1.2

package/dist/index.node.js

@@ -0,0 +1,428 @@
+var mN=Object.defineProperty;var l=(n,e)=>mN(n,"name",{value:e,configurable:!0});var M=(n,e)=>()=>(n&&(e=n(n=0)),e);var at=(n,e)=>{for(var t in e)mN(n,t,{get:e[t],enumerable:!0})};function Ah(n,e){let t=e||"mlc",r=n.indexOf("->"),s=r>=0?n.substring(0,r).trim():n.trim(),o=r>=0?n.substring(r+2).trim():"",i=s.indexOf("("),a=s.lastIndexOf(")"),u,p="",c=s.substring(0,i).trim(),f=c.indexOf(".");f>=0?(u=c.substring(0,f),p=c.substring(f+1)):u=c;let d=s.substring(i+1,a).trim(),h=d.length>0?tP(d):[],m=o.length>0?nP(o):[{kind:"Tensor"}];return new t0(t,u,p,h,m)}function tP(n){return gN(n,",").map(t=>{let r=t.trim(),s=r.indexOf("="),o=r,i;s>=0&&(o=r.substring(0,s).trim(),i=r.substring(s+1).trim());let a=o.endsWith("!");a&&(o=o.substring(0,o.length-1).trim());let u=o.lastIndexOf(" "),p,c;u>=0?(p=o.substring(0,u).trim(),c=o.substring(u+1).trim()):(p=o,c="");let f=n0.get(p)||"Scalar";return new e0(c,f,i,a)})}function nP(n){let e=n.trim();if(e.startsWith("(")&&e.endsWith(")")){let r=e.substring(1,e.length-1);return gN(r,",").map(o=>{let i=o.trim();return{kind:n0.get(i)||"Tensor"}})}return[{kind:n0.get(e)||"Tensor"}]}function gN(n,e){let t=[],r=0,s=0;for(let o=0;o<n.length;o++){let i=n[o];i==="("||i==="["?r++:i===")"||i==="]"?r--:i===e&&r===0&&(t.push(n.substring(s,o)),s=o+1)}return t.push(n.substring(s)),t}var eP,e0,t0,n0,r0=M(()=>{"use strict";eP=new Set(["Tensor","Tensor[]"]),e0=class{static{l(this,"SchemaArg")}name;kind;defaultValue;isOut;constructor(e,t,r,s){this.name=e,this.kind=t,this.defaultValue=r??null,this.isOut=s??!1}get isTensor(){return eP.has(this.kind)}},t0=class{static{l(this,"OperatorSchema")}namespace;name;overload;args;returns;_key;_tensorArgIndices;constructor(e,t,r,s,o){this.namespace=e,this.name=t,this.overload=r||"",this.args=Object.freeze([...s]),this.returns=Object.freeze([...o]),this._key=null,this._tensorArgIndices=null}qualifiedName(){return`${this.namespace}::${this.name}`}key(){return this._key||(this._key=this.overload?`${this.namespace}::${this.name}.${this.overload}`:`${this.namespace}::${this.name}`),this._key}get tensorArgIndices(){if(!this._tensorArgIndices){let e=[];for(let t=0;t<this.args.length;t++)this.args[t].isTensor&&e.push(t);this._tensorArgIndices=Object.freeze(e)}return this._tensorArgIndices}get numTensorArgs(){return this.tensorArgIndices.length}},n0=new Map([["Tensor","Tensor"],["Scalar","Scalar"],["int","int"],["float","float"],["bool","bool"],["int[]","int[]"],["Tensor[]","Tensor[]"],["Dtype","Dtype"],["Device","Device"],["str","str"],["MemoryFormat","MemoryFormat"]]);l(Ah,"parseSchema");l(tP,"_parseArgs");l(nP,"_parseReturns");l(gN,"_splitTopLevel")});function yN(n){return typeof n=="object"&&n!==null&&"_impl"in n}function _N(n){return n instanceof Yr?n:yN(n)?Yr.tensor(n):typeof n=="number"?Number.isInteger(n)?Yr.int(n):Yr.float(n):typeof n=="boolean"?Yr.bool(n):typeof n=="string"?Yr.string(n):Array.isArray(n)?n.length>0&&yN(n[0])?Yr.tensorList(n):Yr.intList(n):(n==null,Yr.none())}var Yr,Mr,sp=M(()=>{"use strict";Yr=class n{static{l(this,"IValue")}tag;value;constructor(e,t){this.tag=e,this.value=t}static tensor(e){return new n(0,e)}static int(e){return new n(1,e)}static float(e){return new n(2,e)}static bool(e){return new n(3,e)}static intList(e){return new n(4,e)}static tensorList(e){return new n(5,e)}static string(e){return new n(6,e)}static none(){return new n(7,null)}static device(e){return new n(8,e)}static dtype(e){return new n(9,e)}isTensor(){return this.tag===0}isInt(){return this.tag===1}isFloat(){return this.tag===2}isBool(){return this.tag===3}isIntList(){return this.tag===4}isTensorList(){return this.tag===5}isString(){return this.tag===6}isNone(){return this.tag===7}toTensor(){return this.value}toInt(){return this.value}toFloat(){return this.value}toBool(){return this.value}toIntList(){return this.value}toTensorList(){return this.value}toString(){return this.value}toDevice(){return this.value}toDtype(){return this.value}},Mr=class n{static{l(this,"KernelFunction")}_boxed;_unboxed;constructor(e,t){this._boxed=e||null,this._unboxed=t||null}static fromBoxed(e){return new n(e)}static fromUnboxed(e){return new n(void 0,e)}static fromBoth(e,t){return new n(e,t)}get isBoxed(){return this._boxed!==null}get isUnboxed(){return this._unboxed!==null}callUnboxed(e,...t){return this._unboxed?this._unboxed(e,...t):this._callBoxedAsUnboxed(e,t)}callBoxed(e,t){return this._boxed?this._boxed(e,t):this._callUnboxedAsBoxed(e,t)}_callBoxedAsUnboxed(e,t){let r=t.map(i=>_N(i)),s=this._boxed;if(!s)return;let o=s(e,r);return Array.isArray(o)?o.length===1?o[0].value:o.map(i=>i.value):o instanceof Yr?o.value:o}_callUnboxedAsBoxed(e,t){let r=t.map(i=>i.value),s=this._unboxed;if(!s)return[];let o=s(e,...r);return o==null?[]:[_N(o)]}};l(yN,"hasImpl");l(_N,"_toIValue")});function bN(n){Jo=n}var Jo,Vt,ys=M(()=>{"use strict";r0();sp();Jo=null;l(bN,"_setDispatcher");Vt=class{static{l(this,"Library")}_namespace;_kind;_registrations;constructor(e,t){this._namespace=e,this._kind=t,this._registrations=[]}def(e){let t=Ah(e,this._namespace);return Jo&&Jo.registerOp(t),this._registrations.push({type:"def",schema:t}),this}impl(e,t,r){let s=r instanceof Mr?r:Mr.fromUnboxed(r);return Jo&&Jo.registerKernel(`${this._namespace}::${e}`,t,s),this._registrations.push({type:"impl",name:e,key:t,kernelFn:s}),this}implBoxed(e,t,r){let s=Mr.fromBoxed(r);return Jo&&Jo.registerKernel(`${this._namespace}::${e}`,t,s),this._registrations.push({type:"impl",name:e,key:t,kernelFn:s}),this}fallback(e,t){let r=t instanceof Mr?t:Mr.fromBoxed(t);return Jo&&Jo.registerFallback(e,r),this._registrations.push({type:"fallback",key:e,kernelFn:r}),this}replay(e){for(let t of this._registrations)t.type==="def"?e.registerOp(t.schema):t.type==="impl"?e.registerKernel(`${this._namespace}::${t.name}`,t.key,t.kernelFn):t.type==="fallback"&&e.registerFallback(t.key,t.kernelFn)}}});function Uu(n){let e=rP[n];if(e===void 0)throw new Error(`No backend key for device: ${n}`);return e}function xN(n){return sP[n]??40}function wN(n){return n=n-(n>>1&1431655765),n=(n&858993459)+(n>>2&858993459),(n+(n>>4)&252645135)*16843009>>24}var Xn,TN,rP,sP,Jr,op,qq,ip,Wq,An=M(()=>{"use strict";Xn=(w=>(w[w.CPU=0]="CPU",w[w.GPU=1]="GPU",w[w.WASM=2]="WASM",w[w.META=3]="META",w[w.LAZY=4]="LAZY",w[w.CUSTOM_0=5]="CUSTOM_0",w[w.CUSTOM_1=6]="CUSTOM_1",w[w.CUSTOM_2=7]="CUSTOM_2",w[w.CUSTOM_3=8]="CUSTOM_3",w[w.BATCHED=20]="BATCHED",w[w.VMAP=24]="VMAP",w[w.FUNCTIONALIZE=28]="FUNCTIONALIZE",w[w.AUTOCAST=32]="AUTOCAST",w[w.AUTOGRAD=40]="AUTOGRAD",w[w.AUTOGRAD_CPU=41]="AUTOGRAD_CPU",w[w.AUTOGRAD_GPU=42]="AUTOGRAD_GPU",w[w.AUTOGRAD_WASM=43]="AUTOGRAD_WASM",w[w.TRACING=48]="TRACING",w[w.NUM_KEYS=49]="NUM_KEYS",w))(Xn||{}),TN=new Array(49).fill(null);for(let[n,e]of Object.entries(Xn))n!=="NUM_KEYS"&&typeof e=="number"&&e<49&&(TN[e]=n);rP=Object.freeze({cpu:0,gpu:1,wasm:2,webgpu:5,meta:3,lazy:4}),sP=Object.freeze({0:41,1:42,2:43});l(Uu,"backendKeyForDevice");l(xN,"autogradKeyForBackend");Jr=class n{static{l(this,"DispatchKeySet")}_lo;_hi;constructor(e,t){this._lo=e|0,this._hi=t|0}static fromKey(e){return e<32?new n(1<<e,0):new n(0,1<<e-32)}static fromKeys(...e){let t=0,r=0;for(let s=0;s<e.length;s++){let o=e[s];o<32?t|=1<<o:r|=1<<o-32}return new n(t,r)}add(e){return e<32?new n(this._lo|1<<e,this._hi):new n(this._lo,this._hi|1<<e-32)}remove(e){return e<32?new n(this._lo&~(1<<e),this._hi):new n(this._lo,this._hi&~(1<<e-32))}has(e){return e<32?(this._lo&1<<e)!==0:(this._hi&1<<e-32)!==0}without(e){return this.remove(e)}union(e){return new n(this._lo|e._lo,this._hi|e._hi)}intersect(e){return new n(this._lo&e._lo,this._hi&e._hi)}subtract(e){return new n(this._lo&~e._lo,this._hi&~e._hi)}isEmpty(){return this._lo===0&&this._hi===0}equals(e){return this._lo===e._lo&&this._hi===e._hi}highestPriority(){return this._hi!==0?63-Math.clz32(this._hi):this._lo!==0?31-Math.clz32(this._lo):-1}lowestPriority(){return this._lo!==0?31-Math.clz32(this._lo&-this._lo):this._hi!==0?63-Math.clz32(this._hi&-this._hi):-1}count(){return wN(this._lo)+wN(this._hi)}*[Symbol.iterator](){let e=this._hi;for(;e!==0;){let r=31-Math.clz32(e);yield r+32,e&=~(1<<r)}let t=this._lo;for(;t!==0;){let r=31-Math.clz32(t);yield r,t&=~(1<<r)}}toString(){let e=[];for(let t of this)e.push(TN[t]||String(t));return`DispatchKeySet(${e.join(", ")})`}};l(wN,"_popcount32");op=new Jr(0,0),qq=Jr.fromKeys(0,1,2,3,4,5,6,7,8),ip=Jr.fromKeys(40,41,42,43),Wq=Jr.fromKeys(20,24,28,32,40,41,42,43,48)});var Oh,SN=M(()=>{"use strict";An();Oh=class{static{l(this,"OperatorEntry")}_schema;_kernels;_catchAll;constructor(e){this._schema=e,this._kernels=new Array(49).fill(null),this._catchAll=null}get schema(){return this._schema}registerKernel(e,t){this._kernels[e]=t}removeKernel(e){this._kernels[e]=null}lookupKernel(e){return this._kernels[e]}hasKernel(e){return this._kernels[e]!==null}get catchAll(){return this._catchAll}setCatchAll(e){this._catchAll=e}bestKernel(e){for(let t of e){let r=this._kernels[t];if(r)return{key:t,kernel:r}}return this._catchAll?{key:-1,kernel:this._catchAll}:null}registeredKeys(){let e=[];for(let t=0;t<this._kernels.length;t++)this._kernels[t]&&e.push(t);return e}}});var Bh,vN=M(()=>{"use strict";Bh=class{static{l(this,"OperatorHandle")}_entry;_schema;constructor(e,t){this._entry=e,this._schema=t}get entry(){return this._entry}get schema(){return this._schema}get name(){return this._schema.name}get qualifiedName(){return this._schema.qualifiedName()}get key(){return this._schema.key()}get tensorArgIndices(){return this._schema.tensorArgIndices}lookupKernel(e){return this._entry.lookupKernel(e)}bestKernel(e){return this._entry.bestKernel(e)}}});var Rh,kN=M(()=>{"use strict";An();Rh=class{static{l(this,"FallbackTable")}_kernels;constructor(){this._kernels=new Array(49).fill(null)}register(e,t){this._kernels[e]=t}remove(e){this._kernels[e]=null}lookup(e){return this._kernels[e]}has(e){return this._kernels[e]!==null}registeredKeys(){let e=[];for(let t=0;t<this._kernels.length;t++)this._kernels[t]&&e.push(t);return e}}});function oP(n){return typeof n=="object"&&n!==null&&"then"in n&&typeof n.then=="function"}function NN(n,e){Ra.push(void 0,n);let t;try{t=e()}catch(r){throw Ra.pop(),r}return oP(t)?t.then(r=>(Ra.pop(),r),r=>{throw Ra.pop(),r}):(Ra.pop(),t)}var s0,Ra,o0=M(()=>{"use strict";An();s0=class{static{l(this,"GuardStack")}_frames;constructor(){this._frames=[]}push(e,t){this._frames.push({exclude:e||op,include:t||op})}pop(){return this._frames.pop()}apply(e){let t=e;for(let r=this._frames.length-1;r>=0;r--){let s=this._frames[r];t=t.subtract(s.exclude),t=t.union(s.include)}return t}get depth(){return this._frames.length}clear(){this._frames.length=0}},Ra=new s0;l(oP,"isThenable");l(NN,"withIncludedKeys")});function iP(n){return typeof n=="object"&&n!==null&&"dispatchKeySet"in n}function a0(n,e){if(!e)return n;if(iP(e)&&e.dispatchKeySet)return n.union(e.dispatchKeySet);if(Array.isArray(e))for(let t=0;t<e.length;t++)n=a0(n,e[t]);return n}function u0(n,e){let t=op;if(e){let r=e.tensorArgIndices;for(let s=0;s<r.length;s++){let o=r[s];o<n.length&&(t=a0(t,n[o]))}}else for(let r=0;r<n.length;r++)t=a0(t,n[r]);return t}var i0,Lt,Ma=M(()=>{"use strict";An();SN();vN();kN();o0();ys();r0();l(iP,"hasDispatchKeySet");i0=class{static{l(this,"Dispatcher")}_entries;_handles;_fallbacks;constructor(){this._entries=new Map,this._handles=new Map,this._fallbacks=new Rh}registerOp(e){let t=e.key(),r=this._handles.get(t);if(r)return r;let s=new Oh(e);this._entries.set(t,s);let o=new Bh(s,e);return this._handles.set(t,o),o}findOp(e){let t=e.includes("::")?e:`mlc::${e}`;return this._handles.get(t)||null}findOrRegisterOp(e){let t=this.findOp(e);if(!t){let r=Ah(`${e}() -> Tensor`,"mlc");t=this.registerOp(r)}return t}registerKernel(e,t,r){let s=e.includes("::")?e:`mlc::${e}`,o=this._entries.get(s);if(!o)throw new Error(`Op '${s}' not registered`);o.registerKernel(t,r)}registerFallback(e,t){this._fallbacks.register(e,t)}dispatch(e,t,...r){let s=Ra.apply(t);return this._dispatchInternal(e,s,r)}redispatch(e,t,...r){return this._dispatchInternal(e,t,r)}_dispatchInternal(e,t,r){let s=t.highestPriority();if(s<0)throw new Error(`No dispatch key found for op '${e.name}'`);let o=e.lookupKernel(s);if(o||(o=this._fallbacks.lookup(s)),!o){let a=e.entry.catchAll;a&&(o=a)}if(!o)throw new Error(`No kernel registered for op '${e.name}' with dispatch key ${s}`);let i=t.without(s);return o.callUnboxed(i,...r)}callOp(e,...t){let r=this.findOp(e);if(!r)throw new Error(`Op '${e}' not found`);let s=u0(t,r.schema);return this.dispatch(r,s,...t)}listOps(){return[...this._handles.keys()]}hasOp(e){let t=e.includes("::")?e:`mlc::${e}`;return this._handles.has(t)}get fallbacks(){return this._fallbacks}};l(a0,"_unionArg");l(u0,"computeKeySet");Lt=new i0;bN(Lt)});function qu(n){return uP+n}var aP,uP,Pe,po=M(()=>{"use strict";aP={ceildiv:"ceilDiv"},uP="_sym_";l(qu,"symVarName");Pe=class n{static{l(this,"SymInt")}constructor(e,t=null,r=[]){this.type=e,this.name=t,this.args=r}static var(e){return new n("var",e)}static const(e){return e}static add(e,t){return typeof e=="number"&&typeof t=="number"?e+t:e===0?t:t===0?e:n.equals(e,t)?n.mul(2,e):new n("add",null,[e,t])}static sub(e,t){return typeof e=="number"&&typeof t=="number"?e-t:t===0?e:n.equals(e,t)?0:new n("sub",null,[e,t])}static neg(e){return typeof e=="number"?-e:e instanceof n&&e.type==="neg"?e.args[0]:new n("neg",null,[e])}static mul(e,t){return typeof e=="number"&&typeof t=="number"?e*t:e===0||t===0?0:e===1?t:t===1?e:new n("mul",null,[e,t])}static div(e,t){if(t===0)throw new Error("SymInt.div: division by zero");return typeof e=="number"&&typeof t=="number"?Math.floor(e/t):e===0?0:t===1?e:n.equals(e,t)?1:new n("div",null,[e,t])}static mod(e,t){if(t===0)throw new Error("SymInt.mod: modulo by zero");return typeof e=="number"&&typeof t=="number"?(e%t+t)%t:e===0||t===1||n.equals(e,t)?0:new n("mod",null,[e,t])}static max(e,t){return typeof e=="number"&&typeof t=="number"?Math.max(e,t):e===t||n.equals(e,t)?e:new n("max",null,[e,t])}static min(e,t){return typeof e=="number"&&typeof t=="number"?Math.min(e,t):e===t||n.equals(e,t)?e:new n("min",null,[e,t])}static ceilDiv(e,t){return typeof e=="number"&&typeof t=="number"?Math.ceil(e/t):new n("ceildiv",null,[e,t])}static equals(e,t){if(e===t)return!0;if(typeof e=="number"||typeof t=="number"||!(e instanceof n)||!(t instanceof n)||e.type!==t.type)return!1;if(e.type==="var")return e.name===t.name;if(e.args.length!==t.args.length)return!1;if(e.type==="add"||e.type==="mul"||e.type==="max"||e.type==="min"){let r=n.equals(e.args[0],t.args[0])&&n.equals(e.args[1],t.args[1]),s=n.equals(e.args[0],t.args[1])&&n.equals(e.args[1],t.args[0]);return r||s}for(let r=0;r<e.args.length;r++)if(!n.equals(e.args[r],t.args[r]))return!1;return!0}static substitute(e,t,r){if(typeof e=="number"||!(e instanceof n))return e;if(e.type==="var")return e.name===t?r:e;let s=e.args.map(i=>n.substitute(i,t,r)),o=n[aP[e.type]||e.type];return o&&s.length===2?o(s[0],s[1]):o&&s.length===1?o(s[0]):new n(e.type,e.name,s)}static evaluate(e,t){if(typeof e=="number"||!(e instanceof n))return e;if(e.type==="var"){if(t.has(e.name))return t.get(e.name);throw new Error(`Unbound symbolic variable: ${e.name}`)}let r=e.args.map(s=>n.evaluate(s,t));switch(e.type){case"add":return r[0]+r[1];case"sub":return r[0]-r[1];case"mul":return r[0]*r[1];case"div":return Math.floor(r[0]/r[1]);case"mod":return(r[0]%r[1]+r[1])%r[1];case"max":return Math.max(r[0],r[1]);case"min":return Math.min(r[0],r[1]);case"neg":return-r[0];case"ceildiv":return Math.ceil(r[0]/r[1]);default:throw new Error(`Unknown SymInt op: ${e.type}`)}}static freeVars(e,t=new Set){if(typeof e=="number"||!(e instanceof n))return t;if(e.type==="var")return t.add(e.name),t;for(let r of e.args)n.freeVars(r,t);return t}static isConst(e){return typeof e=="number"}static toConst(e){return typeof e=="number"?e:null}toString(){return this.type==="var"?this.name:this.type==="add"?`(${this.args[0]} + ${this.args[1]})`:this.type==="sub"?`(${this.args[0]} - ${this.args[1]})`:this.type==="mul"?`(${this.args[0]} * ${this.args[1]})`:this.type==="div"?`(${this.args[0]} / ${this.args[1]})`:this.type==="mod"?`(${this.args[0]} % ${this.args[1]})`:this.type==="max"?`max(${this.args[0]}, ${this.args[1]})`:this.type==="min"?`min(${this.args[0]}, ${this.args[1]})`:this.type==="neg"?`(-${this.args[0]})`:this.type==="ceildiv"?`ceildiv(${this.args[0]}, ${this.args[1]})`:"unknown"}}});function Mh(n,e){return n===e?!0:n instanceof Pe&&e instanceof Pe?Pe.equals(n,e):!1}function ap(n,e){let t=1;for(let r=0;r<n.length;r++){let s=n[r];if(typeof s!="number"||s<0)return e;t*=s}return t}function f0(n){let e=1;for(let t=0;t<n.length;t++){let r=n[t];if(typeof r=="number"){if(r<0)return X;e=Pe.mul(e,r)}else if(r instanceof Pe)e=Pe.mul(e,r);else return X}return e}function fo(n){let e=AN[n];if(e===void 0)throw new Error(`Unknown dtype: ${n}`);return e}function jt(n){return lP.has(n)}function d0(n){return cP.has(n)}function pP(n,e){return Mh(n,e)?n:n===1?e:e===1?n:n===X?e===X?X:e:e===X?n:n instanceof Pe||e instanceof Pe?X:null}function p0(n,e){return n===e?!0:n&&typeof n.equals=="function"?n.equals(e):!1}function Gn(n){if(n instanceof I)return`tensor<${n.shape.map(t=>t===X?"?":String(t)).join("x")}x${n.dtype}>`;if(n instanceof Ia)return`tuple<${n.elements.map(Gn).join(", ")}>`;if(n instanceof l0)return"token";if(n instanceof c0){let e=n.inputs.map(Gn).join(", "),t=n.outputs.map(Gn).join(", ");return`(${e}) -> (${t})`}return"unknown"}var er,AN,lP,cP,X,Gt,I,Ia,l0,c0,Re=M(()=>{"use strict";po();l(Mh,"dimEquals");er=(f=>(f.F16="f16",f.BF16="bf16",f.F32="f32",f.F64="f64",f.I8="i8",f.I16="i16",f.I32="i32",f.I64="i64",f.UI8="ui8",f.BOOL="bool",f.INDEX="index",f))(er||{}),AN=Object.freeze({f16:2,bf16:2,f32:4,f64:8,i8:1,i16:2,i32:4,i64:8,ui8:1,bool:1,index:4}),lP=new Set(["f16","bf16","f32","f64"]),cP=new Set(["i8","i16","i32","i64","ui8"]),X=-1;l(ap,"shapeProduct");l(f0,"symbolicShapeProduct");l(fo,"scalarBytes");l(jt,"isFloatType");l(d0,"isIntType");Gt=class n{static{l(this,"Layout")}order;_hash;constructor(e){this.order=Object.freeze([...e]),this._hash=null}static rowMajor(e){let t=new Array(e);for(let r=0;r<e;r++)t[r]=r;return new n(t)}static columnMajor(e){let t=new Array(e);for(let r=0;r<e;r++)t[r]=e-1-r;return new n(t)}get rank(){return this.order.length}isIdentity(){for(let e=0;e<this.order.length;e++)if(this.order[e]!==e)return!1;return!0}inverse(){let e=new Array(this.order.length);for(let t=0;t<this.order.length;t++)e[this.order[t]]=t;return new n(e)}compose(e){if(this.order.length!==e.order.length)throw new Error("Cannot compose layouts of different ranks");let t=new Array(this.order.length);for(let r=0;r<this.order.length;r++)t[r]=this.order[e.order[r]];return new n(t)}computeStrides(e){let t=e.length,r=new Array(t),s=1;for(let o=t-1;o>=0;o--){let i=this.order[o];r[i]=s,e[i]===X||e[i]instanceof Pe?s=X:s!==X&&(s*=e[i])}return r}equals(e){if(this===e)return!0;if(!(e instanceof n)||this.order.length!==e.order.length)return!1;for(let t=0;t<this.order.length;t++)if(this.order[t]!==e.order[t])return!1;return!0}hash(){if(this._hash!==null)return this._hash;let e=2166136261;for(let t=0;t<this.order.length;t++)e=(e^this.order[t])*16777619&2147483647;return this._hash=e,e}};l(pP,"broadcastDim");I=class n{static{l(this,"TensorType")}shape;dtype;layout;_hash;constructor(e,t,r=null){this.shape=Object.freeze([...e]),this.dtype=t,this.layout=r||Gt.rowMajor(e.length),this._hash=null}get rank(){return this.shape.length}get isScalar(){return this.shape.length===0}get isFullyStatic(){return this.shape.every(e=>typeof e=="number"&&e>=0)}get hasDynamic(){return this.shape.some(e=>e===X||e instanceof Pe)}numel(){return ap(this.shape,X)}symbolicNumel(){return f0(this.shape)}sizeInBytes(){let e=this.numel();return e===X?X:e*AN[this.dtype]}strides(){return this.layout.computeStrides(this.shape)}withShape(e){return new n(e,this.dtype,this.layout)}withDtype(e){return new n(this.shape,e,this.layout)}withLayout(e){return new n(this.shape,this.dtype,e)}equals(e){if(this===e)return!0;if(!(e instanceof n)||this.dtype!==e.dtype||this.shape.length!==e.shape.length)return!1;for(let t=0;t<this.shape.length;t++)if(!Mh(this.shape[t],e.shape[t]))return!1;return this.layout.equals(e.layout)}shapeEquals(e){if(!(e instanceof n)||this.dtype!==e.dtype||this.shape.length!==e.shape.length)return!1;for(let t=0;t<this.shape.length;t++)if(!Mh(this.shape[t],e.shape[t]))return!1;return!0}shapeCompatible(e){if(this.shape.length!==e.shape.length)return!1;for(let t=0;t<this.shape.length;t++){let r=this.shape[t],s=e.shape[t];if(!(r===X||s===X)&&!Mh(r,s)&&typeof r=="number"&&typeof s=="number")return!1}return!0}hash(){if(this._hash!==null)return this._hash;let e=2166136261;for(let t=0;t<this.shape.length;t++){let r=this.shape[t],s=typeof r=="number"?r&65535:32767;e=(e^s)*16777619&2147483647}return e=(e^this.dtype.charCodeAt(0))*16777619&2147483647,this._hash=e,e}static broadcastShape(...e){let t=0;for(let s=0;s<e.length;s++)e[s].length>t&&(t=e[s].length);let r=new Array(t);for(let s=0;s<t;s++){let o=1;for(let i=0;i<e.length;i++){let a=e[i],u=s-(t-a.length);if(u<0)continue;let p=pP(o,a[u]);if(p===null)return null;o=p}r[s]=o}return r}static broadcastCompatible(...e){return n.broadcastShape(...e)!==null}},Ia=class n{static{l(this,"TupleType")}elements;constructor(e){this.elements=Object.freeze([...e])}equals(e){if(this===e)return!0;if(!(e instanceof n)||this.elements.length!==e.elements.length)return!1;for(let t=0;t<this.elements.length;t++)if(!p0(this.elements[t],e.elements[t]))return!1;return!0}},l0=class n{static{l(this,"TokenType")}equals(e){return e instanceof n}},c0=class n{static{l(this,"FunctionType")}inputs;outputs;constructor(e,t){this.inputs=Object.freeze([...e]),this.outputs=Object.freeze([...t])}equals(e){if(this===e)return!0;if(!(e instanceof n)||this.inputs.length!==e.inputs.length||this.outputs.length!==e.outputs.length)return!1;for(let t=0;t<this.inputs.length;t++)if(!p0(this.inputs[t],e.inputs[t]))return!1;for(let t=0;t<this.outputs.length;t++)if(!p0(this.outputs[t],e.outputs[t]))return!1;return!0}};l(p0,"typeEquals");l(Gn,"typeToString")});var fP,up,Li,lp,Ih=M(()=>{"use strict";fP=0,up=class{static{l(this,"UseLink")}constructor(e,t){this.user=e,this.operandIndex=t,this.prev=null,this.next=null}},Li=class{static{l(this,"Value")}constructor(e,t=null,r=0){this.type=e,this.definingOp=t,this.resultIndex=r,this.id=fP++,this._useHead=null,this._useTail=null,this._useCount=0}get hasUses(){return this._useHead!==null}get useCount(){return this._useCount}addUse(e){e.prev=this._useTail,e.next=null,this._useTail?this._useTail.next=e:this._useHead=e,this._useTail=e,this._useCount++}removeUse(e){e.prev?e.prev.next=e.next:this._useHead=e.next,e.next?e.next.prev=e.prev:this._useTail=e.prev,e.prev=null,e.next=null,this._useCount--}*uses(){let e=this._useHead;for(;e;){let t=e.next;yield e,e=t}}getUsers(){let e=[],t=this._useHead;for(;t;)e.push(t.user),t=t.next;return e}replaceAllUsesWith(e){if(this===e)return;let t=this._useHead!==null,r=this._useHead;for(;r;)r.user.operands[r.operandIndex]=e,r=r.next;if(this._useHead&&(e._useTail?(e._useTail.next=this._useHead,this._useHead.prev=e._useTail):e._useHead=this._useHead,e._useTail=this._useTail,e._useCount+=this._useCount),this._useHead=null,this._useTail=null,this._useCount=0,t){let s=this._owningFunction();s&&s.bumpVersion()}}_owningFunction(){return this.definingOp?this.definingOp.getParentFunction():null}isBlockArgument(){return!1}},lp=class extends Li{static{l(this,"BlockArgument")}constructor(e,t,r){super(e,null,0),this.ownerBlock=t,this.argIndex=r}_owningFunction(){return this.ownerBlock?this.ownerBlock._owningFunction():null}isBlockArgument(){return!0}}});var On,Ut,Ea=M(()=>{"use strict";Ih();On=class{static{l(this,"Block")}constructor(e=[]){this.parentRegion=null,this._head=null,this._tail=null,this._size=0,this.arguments=e.map((t,r)=>new lp(t,this,r))}get size(){return this._size}get isEmpty(){return this._head===null}get firstOp(){return this._head}get lastOp(){return this._tail}get parentOp(){return this.parentRegion?this.parentRegion.parentOp:null}_owningFunction(){let e=this;for(;e;){if(e._parentFunction)return e._parentFunction;let t=e.parentOp;e=t?t.parentBlock:null}return null}_notifyMutation(){let e=this._owningFunction();e&&e.bumpVersion()}getArgument(e){return this.arguments[e]}addArgument(e){let t=new lp(e,this,this.arguments.length);return this.arguments.push(t),this._notifyMutation(),t}removeArguments(e){let t=e instanceof Set?e:new Set(e);if(t.size===0)return this;for(let s=0;s<this.arguments.length;s++)if(t.has(s)&&this.arguments[s].hasUses)throw new Error(`removeArguments: block argument ${s} still has uses`);let r=[];for(let s=0;s<this.arguments.length;s++)t.has(s)||r.push(this.arguments[s]);this.arguments=r;for(let s=0;s<r.length;s++)r[s].argIndex=s;return this._notifyMutation(),this}pushOp(e){if(e.parentBlock!==null)throw new Error(`pushOp: operation '${e.opName}' is already attached to a block`);e.parentBlock=this,e._prev=this._tail,e._next=null,this._tail?this._tail._next=e:this._head=e,this._tail=e,this._size++,this._notifyMutation()}insertBefore(e,t){if(e.parentBlock!==null)throw new Error(`insertBefore: operation '${e.opName}' is already attached to a block`);if(!t||t.parentBlock!==this)throw new Error("insertBefore: reference operation is not in this block");e.parentBlock=this,e._prev=t._prev,e._next=t,t._prev?t._prev._next=e:this._head=e,t._prev=e,this._size++,this._notifyMutation()}insertAfter(e,t){if(e.parentBlock!==null)throw new Error(`insertAfter: operation '${e.opName}' is already attached to a block`);if(!t||t.parentBlock!==this)throw new Error("insertAfter: reference operation is not in this block");e.parentBlock=this,e._prev=t,e._next=t._next,t._next?t._next._prev=e:this._tail=e,t._next=e,this._size++,this._notifyMutation()}removeOp(e){e.parentBlock===this&&(e._prev?e._prev._next=e._next:this._head=e._next,e._next?e._next._prev=e._prev:this._tail=e._prev,e._prev=null,e._next=null,e.parentBlock=null,this._size--,this._notifyMutation())}*ops(){let e=this._head;for(;e;){let t=e._next;yield e,e=t}}*opsReverse(){let e=this._tail;for(;e;){let t=e._prev;yield e,e=t}}opsArray(){let e=[],t=this._head;for(;t;)e.push(t),t=t._next;return e}[Symbol.iterator](){return this.ops()}},Ut=class{static{l(this,"Region")}constructor(e=[]){this.parentOp=null,this.blocks=[];for(let t=0;t<e.length;t++)this.addBlock(e[t])}get entryBlock(){return this.blocks[0]||null}get isEmpty(){return this.blocks.length===0}addBlock(e){if(e.parentRegion!==null)throw new Error("addBlock: block already belongs to a region");return e.parentRegion=this,this.blocks.push(e),e}insertBlock(e,t){if(t.parentRegion!==null)throw new Error("insertBlock: block already belongs to a region");t.parentRegion=this,this.blocks.splice(e,0,t)}removeBlock(e){let t=this.blocks.indexOf(e);t!==-1&&(this.blocks.splice(t,1),e.parentRegion=null)}*[Symbol.iterator](){yield*this.blocks}}});var gn,Y,Z,Eh,Bn=M(()=>{"use strict";gn=Object.freeze({NONE:0,READ:1,WRITE:2,ALLOCATE:4,CONTROL:8}),Y=Object.freeze({COMMUTATIVE:"commutative",ASSOCIATIVE:"associative",IDEMPOTENT:"idempotent",ELEMENTWISE:"elementwise",SAME_OPERAND_AND_RESULT_TYPE:"same_type",SAME_OPERAND_AND_RESULT_SHAPE:"same_shape",TERMINATOR:"terminator",CONSTANT:"constant",BROADCAST:"broadcast",REDUCTION:"reduction",VIEW:"view",INJECTIVE:"injective",OUT_EWISE_FUSABLE:"out_ewise_fusable",OPAQUE:"opaque"}),Z=class{static{l(this,"OpDef")}constructor(e){this.name=e.name,this.numOperands=e.numOperands,this.numResults=e.numResults!==void 0?e.numResults:1,this.attrs=Object.freeze(e.attrs||[]),this.sideEffects=e.sideEffects||gn.NONE,this.traits=new Set(e.traits||[]),this.inferResultTypes=e.inferResultTypes||null,this.propagateSymbolicShapes=e.propagateSymbolicShapes||null,this.verify=e.verify||null,this.getMemoryEffects=e.getMemoryEffects||null,this.fold=e.fold||null,this.getCanonicalizationPatterns=e.getCanonicalizationPatterns||null,this.getFlops=e.getFlops||null,this.hasRegions=e.hasRegions||!1,this.numRegions=e.numRegions||0,this.regionSpecs=e.regions||null,this.genericAttrs=new Map(Object.entries(e.opAttrs||{}))}setAttr(e,t){return this.genericAttrs.set(e,t),this}getAttr(e){return this.genericAttrs.has(e)?this.genericAttrs.get(e):null}hasAttr(e){return this.genericAttrs.has(e)}hasTrait(e){return this.traits.has(e)}addTrait(e){return this.traits.add(e),this}get isCommutative(){return this.traits.has(Y.COMMUTATIVE)}get isAssociative(){return this.traits.has(Y.ASSOCIATIVE)}get isElementwise(){return this.traits.has(Y.ELEMENTWISE)}get isTerminator(){return this.traits.has(Y.TERMINATOR)}get isConstant(){return this.traits.has(Y.CONSTANT)}get isReduction(){return this.traits.has(Y.REDUCTION)}get isBroadcast(){return this.traits.has(Y.BROADCAST)}get isInjective(){return this.traits.has(Y.INJECTIVE)}get isOutEWiseFusable(){return this.traits.has(Y.OUT_EWISE_FUSABLE)}get isOpaque(){return this.traits.has(Y.OPAQUE)}get hasSideEffects(){return this.sideEffects!==gn.NONE}},Eh=class{static{l(this,"OpRegistry")}constructor(){this._defs=new Map}register(e){if(this._defs.has(e.name))throw new Error(`Op '${e.name}' already registered`);this._defs.set(e.name,e)}registerOpAttr(e,t,r){let s=this._defs.get(e);if(!s)throw new Error(`registerOpAttr: op '${e}' not registered`);return s.setAttr(t,r),s}registerTrait(e,t){let r=this._defs.get(e);if(!r)throw new Error(`registerTrait: op '${e}' not registered`);return r.addTrait(t),r}get(e){return this._defs.get(e)||null}has(e){return this._defs.has(e)}allOps(){return[...this._defs.values()]}names(){return[...this._defs.keys()]}}});var dt,Wu,cp=M(()=>{"use strict";dt=class{static{l(this,"Pattern")}constructor(e,t=1){this.name=e,this.benefit=t,this.rootOpName=null}match(e){return!1}rewrite(e,t){return!1}},Wu=class{static{l(this,"PatternSet")}constructor(){this.patterns=[],this._byOp=new Map,this._generic=[],this._sorted=!1}add(e){if(this.patterns.push(e),this._sorted=!1,e.rootOpName){let t=this._byOp.get(e.rootOpName);t||(t=[],this._byOp.set(e.rootOpName,t)),t.push(e)}else this._generic.push(e)}_ensureSorted(){if(this._sorted)return;let e=l((t,r)=>r.benefit-t.benefit,"cmp");for(let[,t]of this._byOp)t.sort(e);this._generic.sort(e),this._sorted=!0}get(){return[...this.patterns].sort((e,t)=>t.benefit-e.benefit)}getForOp(e){this._ensureSorted();let t=this._byOp.get(e);if(!t)return this._generic;if(this._generic.length===0)return t;let r=new Array(t.length+this._generic.length),s=0,o=0,i=0;for(;s<t.length&&o<this._generic.length;)t[s].benefit>=this._generic[o].benefit?r[i++]=t[s++]:r[i++]=this._generic[o++];for(;s<t.length;)r[i++]=t[s++];for(;o<this._generic.length;)r[i++]=this._generic[o++];return r}hasPatterns(){return this.patterns.length>0}}});function Fs(n){return(pp[n]||Ch).wasm}function Dh(n){return(pp[n]||Ch).load}function RN(n){return(pp[n]||Ch).store}function fp(n){return(pp[n]||Ch).bytes}function h0(n){return BN[n]||null}function zs(n,e){let t=BN[n];return t&&t[e]||null}function Ci(n){return ON[n]||dP}function Ph(n){return Ci(n).js}function Zr(n){return Ci(n).c}function m0(n){return Ci(n).cPtr}function g0(n){return Ci(n).suffix}function hP(n){return Ci(n).mathSuffix}function Rn(n){return Ci(n).bytes}function gr(n){return Ci(n).isFloat}function mr(n){return Ci(n).isInt}function MN(n,e){return n==="sum"||n==="mean"?0:n==="prod"?1:n==="max"?mr(e)?(Lh[e]||Lh.i32)[0]:-1/0:n==="min"?mr(e)?(Lh[e]||Lh.i32)[1]:1/0:0}function $h(n,e){let t=mP[n];return t?t+hP(e):n}function Ku(n){let e=gP[n];if(!e)throw new Error(`unsupported compare direction '${n}'`);return e}function IN(n){let e=yP[n];if(!e)throw new Error(`unsupported compare direction '${n}'`);return e}function EN(n){return _P.has(n)}function jn(n){return(y0[n]||LN).wgsl}function Vs(n){return(y0[n]||LN).bytes}function DN(n){return CN[n]||n}function PN(n){return n in CN}var ON,pp,Ch,BN,dP,Lh,mP,gP,yP,_P,y0,LN,CN,tr=M(()=>{"use strict";ON={f16:{js:"Uint16Array",c:"__half",cPtr:"__half*",bytes:2,suffix:"h",mathSuffix:"h",isFloat:!0,isInt:!1},bf16:{js:"Uint16Array",c:"__nv_bfloat16",cPtr:"__nv_bfloat16*",bytes:2,suffix:"",mathSuffix:"",isFloat:!0,isInt:!1},f32:{js:"Float32Array",c:"float",cPtr:"float*",bytes:4,suffix:"f",mathSuffix:"f",isFloat:!0,isInt:!1},f64:{js:"Float64Array",c:"double",cPtr:"double*",bytes:8,suffix:"",mathSuffix:"",isFloat:!0,isInt:!1},i8:{js:"Int8Array",c:"int8_t",cPtr:"int8_t*",bytes:1,suffix:"",mathSuffix:"",isFloat:!1,isInt:!0},i16:{js:"Int16Array",c:"int16_t",cPtr:"int16_t*",bytes:2,suffix:"",mathSuffix:"",isFloat:!1,isInt:!0},i32:{js:"Int32Array",c:"int",cPtr:"int*",bytes:4,suffix:"",mathSuffix:"",isFloat:!1,isInt:!0},i64:{js:"BigInt64Array",c:"int64_t",cPtr:"int64_t*",bytes:8,suffix:"LL",mathSuffix:"",isFloat:!1,isInt:!0},ui8:{js:"Uint8Array",c:"uint8_t",cPtr:"uint8_t*",bytes:1,suffix:"",mathSuffix:"",isFloat:!1,isInt:!0},bool:{js:"Uint8Array",c:"bool",cPtr:"bool*",bytes:1,suffix:"",mathSuffix:"",isFloat:!1,isInt:!1},index:{js:"Int32Array",c:"int",cPtr:"int*",bytes:4,suffix:"",mathSuffix:"",isFloat:!1,isInt:!0}},pp={f16:{wasm:"f32",load:"i32.load16_u",store:"i32.store16",bytes:2},bf16:{wasm:"f32",load:"i32.load16_u",store:"i32.store16",bytes:2},f32:{wasm:"f32",load:"f32.load",store:"f32.store",bytes:4},f64:{wasm:"f64",load:"f64.load",store:"f64.store",bytes:8},i8:{wasm:"i32",load:"i32.load8_s",store:"i32.store8",bytes:1},i16:{wasm:"i32",load:"i32.load16_s",store:"i32.store16",bytes:2},i32:{wasm:"i32",load:"i32.load",store:"i32.store",bytes:4},i64:{wasm:"i64",load:"i64.load",store:"i64.store",bytes:8},ui8:{wasm:"i32",load:"i32.load8_u",store:"i32.store8",bytes:1},bool:{wasm:"i32",load:"i32.load8_u",store:"i32.store8",bytes:1},index:{wasm:"i32",load:"i32.load",store:"i32.store",bytes:4}},Ch=pp.f32,BN={f32:{laneType:"f32x4",lanes:4,laneBytes:4,vecLoad:"v128.load",vecStore:"v128.store",splat:"f32x4.splat",extractLane:"f32x4.extract_lane",replaceLane:"f32x4.replace_lane",add:"f32x4.add",sub:"f32x4.sub",mul:"f32x4.mul",div:"f32x4.div",neg:"f32x4.neg",abs:"f32x4.abs",sqrt:"f32x4.sqrt",min:"f32x4.min",max:"f32x4.max",ceil:"f32x4.ceil",floor:"f32x4.floor",eq:"f32x4.eq",ne:"f32x4.ne",lt:"f32x4.lt",le:"f32x4.le",gt:"f32x4.gt",ge:"f32x4.ge",bitselect:"v128.bitselect"},i32:{laneType:"i32x4",lanes:4,laneBytes:4,vecLoad:"v128.load",vecStore:"v128.store",splat:"i32x4.splat",extractLane:"i32x4.extract_lane",replaceLane:"i32x4.replace_lane",add:"i32x4.add",sub:"i32x4.sub",mul:"i32x4.mul",div:null,neg:null,abs:"i32x4.abs",sqrt:null,min:"i32x4.min_s",max:"i32x4.max_s",ceil:null,floor:null,eq:"i32x4.eq",ne:"i32x4.ne",lt:"i32x4.lt_s",le:"i32x4.le_s",gt:"i32x4.gt_s",ge:"i32x4.ge_s",bitselect:"v128.bitselect"}};l(Fs,"wasmType");l(Dh,"wasmLoad");l(RN,"wasmStore");l(fp,"wasmBytes");l(h0,"wasmSimdEntry");l(zs,"wasmVecOp");dP=ON.f32;l(Ci,"dtypeInfo");l(Ph,"jsTypedArray");l(Zr,"cType");l(m0,"cPtrType");l(g0,"cLiteralSuffix");l(hP,"cMathFuncSuffix");l(Rn,"dtypeBytes");l(gr,"isDtypeFloat");l(mr,"isDtypeInt");Lh={i8:[-128,127],i16:[-32768,32767],i32:[-2147483648,2147483647],i64:[Number.MIN_SAFE_INTEGER,Number.MAX_SAFE_INTEGER],ui8:[0,255],index:[-2147483648,2147483647]};l(MN,"reduceInitValue");mP={exp:"exp",log:"log",sqrt:"sqrt",tanh:"tanh",abs:"fabs",sin:"sin",cos:"cos",ceil:"ceil",floor:"floor",max:"fmax",min:"fmin",pow:"pow",round:"round",fmod:"fmod",rsqrt:"rsqrt",erf:"erf",erfc:"erfc",lgamma:"lgamma",gamma:"tgamma",log2:"log2",log10:"log10",exp2:"exp2"};l($h,"cMathFunc");gP={eq:"==",ne:"!=",lt:"<",le:"<=",gt:">",ge:">="},yP={eq:"===",ne:"!==",lt:"<",le:"<=",gt:">",ge:">="};l(Ku,"cCompareOp");l(IN,"jsCompareOp");_P=new Set(["exp","log","sqrt","tanh","abs","ceil","floor","sin","cos","max","min","pow","round","sign","log2","log10"]);l(EN,"isJSMathFunc");y0={f16:{wgsl:"f16",bytes:2},bf16:{wgsl:"f32",bytes:4},f32:{wgsl:"f32",bytes:4},f64:{wgsl:"f32",bytes:4},i8:{wgsl:"i32",bytes:4},i16:{wgsl:"i32",bytes:4},i32:{wgsl:"i32",bytes:4},i64:{wgsl:"i32",bytes:4},ui8:{wgsl:"u32",bytes:4},bool:{wgsl:"u32",bytes:4},index:{wgsl:"u32",bytes:4}},LN=y0.f32;l(jn,"wgslType");l(Vs,"wgslBytes");CN={exp:"exp",log:"log",sqrt:"sqrt",tanh:"tanh",abs:"abs",sin:"sin",cos:"cos",ceil:"ceil",floor:"floor",max:"max",min:"min",pow:"pow",round:"round",sign:"sign",rsqrt:"inverseSqrt",fabs:"abs",log2:"log2",exp2:"exp2"};l(DN,"wgslMathFunc");l(PN,"hasWgslMathFunc")});function Yu(n,e){let t={};return n.match(e,t)?t:null}var Fh,_0,b0,Hu,Gs,$N=M(()=>{"use strict";Fh=class{static{l(this,"DFPattern")}match(e,t){return!1}},_0=class extends Fh{static{l(this,"AnyPattern")}match(){return!0}},b0=class extends Fh{static{l(this,"OpPattern")}constructor(e,t){super(),this.name=e,this.operandPatterns=t}match(e,t){if(!e||e.opName!==this.name)return!1;if(this.operandPatterns.length===0)return!0;if(e.numOperands<this.operandPatterns.length)return!1;for(let r=0;r<this.operandPatterns.length;r++){let s=e.getOperand(r).definingOp;if(!this.operandPatterns[r].match(s,t))return!1}return!0}},Hu=l(()=>new _0,"wildcard"),Gs=l((n,...e)=>new b0(n,e),"isOp");l(Yu,"matchPattern")});function ho(n,e){return n&&n.opName==="constant"&&n.getAttr("value")===e}function T0(n){return new w0(n)}var bP,wP,TP,xP,SP,zh,Ju,Vh,Gh,jh,Uh,w0,Zu,Qu,Xu,el,tl,nl,rl,qh,Wh,Kh,Hh,Yh,Jh,Zh,Qh,Xh,em,sl=M(()=>{"use strict";cp();Re();tr();$N();bP=Gs("transpose",Gs("transpose",Hu())),wP=Gs("reshape",Gs("reshape",Hu())),TP=Gs("neg",Gs("neg",Hu())),xP=Gs("exp",Gs("log",Hu())),SP=Gs("log",Gs("exp",Hu()));l(ho,"isConstantVal");zh=class extends dt{static{l(this,"FoldTrivialReshape")}constructor(){super("fold_trivial_reshape",10),this.rootOpName="reshape"}match(e){let t=e.getOperand(0).type,r=e.getResult(0).type;return t instanceof I&&r instanceof I&&t.shapeEquals(r)}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},Ju=class extends dt{static{l(this,"ReshapeReshape")}constructor(){super("reshape_reshape",10),this.rootOpName="reshape"}match(e){return Yu(wP,e)!==null}rewrite(e,t){let r=e.getOperand(0).definingOp.getOperand(0),s=e.getAttr("new_shape"),o=t.reshape(r,s);return e.replaceAllResultsWith([o.getResult(0)]),e.erase(),!0}},Vh=class extends dt{static{l(this,"FoldTrivialTranspose")}constructor(){super("fold_trivial_transpose",10),this.rootOpName="transpose"}match(e){let t=e.getAttr("permutation");if(!t)return!1;for(let r=0;r<t.length;r++)if(t[r]!==r)return!1;return!0}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},Gh=class extends dt{static{l(this,"TransposeTranspose")}constructor(){super("transpose_transpose",10),this.rootOpName="transpose"}match(e){return Yu(bP,e)!==null}rewrite(e,t){let r=e.getOperand(0).definingOp.getOperand(0),s=e.getOperand(0).definingOp.getAttr("permutation"),o=e.getAttr("permutation"),i=new Array(o.length);for(let u=0;u<o.length;u++)i[u]=s[o[u]];let a=t.transpose(r,i);return e.replaceAllResultsWith([a.getResult(0)]),e.erase(),!0}},jh=class extends dt{static{l(this,"FoldTrivialPad")}constructor(){super("fold_trivial_pad",10),this.rootOpName="pad"}match(e){let t=e.getAttr("low"),r=e.getAttr("high"),s=e.getAttr("interior")||[];return!(t&&t.some(o=>o!==0)||r&&r.some(o=>o!==0)||s&&s.some(o=>o!==0))}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},Uh=class extends dt{static{l(this,"FoldTrivialSlice")}constructor(){super("fold_trivial_slice",10),this.rootOpName="slice"}match(e){let t=e.getOperand(0).type,r=e.getResult(0).type;if(!(t instanceof I)||!(r instanceof I))return!1;let s=e.getAttr("starts"),o=e.getAttr("strides")||s.map(()=>1);return s.some(i=>i!==0)||o.some(i=>i!==1)?!1:t.shapeEquals(r)}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},w0=class extends dt{static{l(this,"CommutativeConstantRight")}constructor(e=null){super(`commutative_constant_right${e?"_"+e:""}`,5),this.rootOpName=e}match(e){if(e.numOperands!==2)return!1;let t=e.getOperand(0).definingOp,r=e.getOperand(1).definingOp;return t&&t.opName==="constant"&&!(r&&r.opName==="constant")}rewrite(e,t){let r=e.getOperand(0),s=e.getOperand(1);return e.replaceOperand(0,s),e.replaceOperand(1,r),!0}};l(T0,"commutativeConstantRightFor");Zu=class extends dt{static{l(this,"AddZero")}constructor(){super("add_zero",5),this.rootOpName="add"}match(e){return ho(e.getOperand(1).definingOp,0)||ho(e.getOperand(0).definingOp,0)}rewrite(e,t){let r=ho(e.getOperand(1).definingOp,0)?e.getOperand(0):e.getOperand(1);return r.type.equals(e.getResult(0).type)?(e.replaceAllResultsWith([r]),e.erase(),!0):!1}},Qu=class extends dt{static{l(this,"SubZero")}constructor(){super("sub_zero",5),this.rootOpName="sub"}match(e){return ho(e.getOperand(1).definingOp,0)}rewrite(e,t){let r=e.getOperand(0);return r.type.equals(e.getResult(0).type)?(e.replaceAllResultsWith([r]),e.erase(),!0):!1}},Xu=class extends dt{static{l(this,"SubSelf")}constructor(e=!1){super("sub_self",5),this.rootOpName="sub",this.fastMath=e}match(e){return e.getOperand(0)!==e.getOperand(1)?!1:mr(e.getResult(0).type.dtype)||this.fastMath}rewrite(e,t){let s=t.scalarConstant(0,e.getResult(0).type.dtype).getResult(0),o=e.getResult(0).type.shape;return o.length>0&&(s=t.broadcast(s,o,[]).getResult(0)),e.replaceAllResultsWith([s]),e.erase(),!0}},el=class extends dt{static{l(this,"MulOne")}constructor(){super("mul_one",5),this.rootOpName="mul"}match(e){return ho(e.getOperand(1).definingOp,1)||ho(e.getOperand(0).definingOp,1)}rewrite(e,t){let r=ho(e.getOperand(1).definingOp,1)?e.getOperand(0):e.getOperand(1);return r.type.equals(e.getResult(0).type)?(e.replaceAllResultsWith([r]),e.erase(),!0):!1}},tl=class extends dt{static{l(this,"MulZero")}constructor(e=!1){super("mul_zero",5),this.rootOpName="mul",this.fastMath=e}match(e){return!mr(e.getResult(0).type.dtype)&&!this.fastMath?!1:ho(e.getOperand(1).definingOp,0)||ho(e.getOperand(0).definingOp,0)}rewrite(e,t){let s=t.scalarConstant(0,e.getResult(0).type.dtype).getResult(0),o=e.getResult(0).type.shape;return o.length>0&&(s=t.broadcast(s,o,[]).getResult(0)),e.replaceAllResultsWith([s]),e.erase(),!0}},nl=class extends dt{static{l(this,"DivOne")}constructor(){super("div_one",5),this.rootOpName="div"}match(e){return ho(e.getOperand(1).definingOp,1)}rewrite(e,t){let r=e.getOperand(0);return r.type.equals(e.getResult(0).type)?(e.replaceAllResultsWith([r]),e.erase(),!0):!1}},rl=class extends dt{static{l(this,"DoubleNeg")}constructor(){super("double_neg",5),this.rootOpName="neg"}match(e){return Yu(TP,e)!==null}rewrite(e,t){let r=e.getOperand(0).definingOp.getOperand(0);return e.replaceAllResultsWith([r]),e.erase(),!0}},qh=class extends dt{static{l(this,"ExpLog")}constructor(e=!1){super("exp_log",5),this.rootOpName="exp",this.fastMath=e}match(e){return this.fastMath&&Yu(xP,e)!==null}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0).definingOp.getOperand(0)]),e.erase(),!0}},Wh=class extends dt{static{l(this,"LogExp")}constructor(e=!1){super("log_exp",5),this.rootOpName="log",this.fastMath=e}match(e){return this.fastMath&&Yu(SP,e)!==null}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0).definingOp.getOperand(0)]),e.erase(),!0}},Kh=class extends dt{static{l(this,"DivSelf")}constructor(e=!1){super("div_self",5),this.rootOpName="div",this.fastMath=e}match(e){return this.fastMath?e.getOperand(0)===e.getOperand(1):!1}rewrite(e,t){let s=t.scalarConstant(1,e.getResult(0).type.dtype).getResult(0),o=e.getResult(0).type.shape;return o.length>0&&(s=t.broadcast(s,o,[]).getResult(0)),e.replaceAllResultsWith([s]),e.erase(),!0}},Hh=class extends dt{static{l(this,"MulNegNeg")}constructor(){super("mul_neg_neg",4),this.rootOpName="mul"}match(e){let t=e.getOperand(0).definingOp,r=e.getOperand(1).definingOp;return t&&t.opName==="neg"&&r&&r.opName==="neg"}rewrite(e,t){let r=e.getOperand(0).definingOp.getOperand(0),s=e.getOperand(1).definingOp.getOperand(0),o=t.mul(r,s);return e.replaceAllResultsWith([o.getResult(0)]),e.erase(),!0}},Yh=class extends dt{static{l(this,"AddNegToSub")}constructor(){super("add_neg_to_sub",4),this.rootOpName="add"}match(e){let t=e.getOperand(1).definingOp;return t&&t.opName==="neg"}rewrite(e,t){let r=e.getOperand(0),s=e.getOperand(1).definingOp.getOperand(0),o=t.sub(r,s);return e.replaceAllResultsWith([o.getResult(0)]),e.erase(),!0}},Jh=class extends dt{static{l(this,"SubNegToAdd")}constructor(){super("sub_neg_to_add",4),this.rootOpName="sub"}match(e){let t=e.getOperand(1).definingOp;return t&&t.opName==="neg"}rewrite(e,t){let r=e.getOperand(0),s=e.getOperand(1).definingOp.getOperand(0),o=t.add(r,s);return e.replaceAllResultsWith([o.getResult(0)]),e.erase(),!0}},Zh=class extends dt{static{l(this,"DoubleConvert")}constructor(){super("double_convert",6),this.rootOpName="convert"}match(e){let t=e.getOperand(0).definingOp;if(!t||t.opName!=="convert")return!1;let r=t.getOperand(0).type.dtype,s=e.getAttr("target_dtype");return r===s}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0).definingOp.getOperand(0)]),e.erase(),!0}},Qh=class extends dt{static{l(this,"LayoutTransformIdentity")}constructor(){super("layout_transform_identity",10),this.rootOpName="layout_transform"}match(e){let t=e.getAttr("src_layout"),r=e.getAttr("dst_layout");if(!t||!r||t.length!==r.length)return!1;for(let s=0;s<t.length;s++)if(t[s]!==r[s])return!1;return!0}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},Xh=class extends dt{static{l(this,"FoldTransposeIntoDot")}constructor(){super("fold_transpose_into_dot",10),this.rootOpName="dot"}match(e){for(let t=0;t<2;t++){let r=e.getOperand(t).definingOp;if(!r||r.opName!=="transpose")continue;let s=r.getAttr("permutation");if(!(!s||s.length!==2)&&!(s[0]!==1||s[1]!==0))return!0}return!1}rewrite(e,t){let r=[e.getOperand(0),e.getOperand(1)],s=[...e.getAttr("lhs_contracting")],o=[...e.getAttr("rhs_contracting")],i=[...e.getAttr("lhs_batch")||[]],a=[...e.getAttr("rhs_batch")||[]];for(let p=0;p<2;p++){let c=r[p].definingOp;if(!c||c.opName!=="transpose")continue;let f=c.getAttr("permutation");if(!f||f.length!==2||f[0]!==1||f[1]!==0)continue;r[p]=c.getOperand(0);let d=p===0?s:o,h=p===0?i:a;for(let m=0;m<d.length;m++)d[m]=d[m]===0?1:0;for(let m=0;m<h.length;m++)h[m]=h[m]===0?1:0}let u=t.dot(r[0],r[1],s,o);return i.length>0&&u.setAttr("lhs_batch",i),a.length>0&&u.setAttr("rhs_batch",a),e.replaceAllResultsWith([u.getResult(0)]),e.erase(),!0}},em=class extends dt{static{l(this,"LayoutTransformCompose")}constructor(){super("layout_transform_compose",10),this.rootOpName="layout_transform"}match(e){let t=e.getOperand(0).definingOp;return t&&t.opName==="layout_transform"}rewrite(e,t){let r=e.getOperand(0).definingOp,s=r.getAttr("src_layout"),o=r.getAttr("dst_layout"),i=e.getAttr("dst_layout"),a=new Array(s.length);for(let c=0;c<i.length;c++)a[c]=s[o.indexOf(i[c])];let u=r.getOperand(0),p=t._inferAndBuild("layout_transform",[u],{src_layout:s,dst_layout:a});return e.replaceAllResultsWith([p.getResult(0)]),e.erase(),!0}}});function La(n){if(n.length!==2)return null;let e=n[0],t=n[1];if(!(e instanceof I)||!(t instanceof I)||e.dtype!==t.dtype)return null;let r=I.broadcastShape(e.shape,t.shape);return r?[new I(r,e.dtype)]:null}function tm(n){if(n.length!==1)return null;let e=n[0];return e instanceof I?[new I(e.shape,e.dtype)]:null}function FN(n){if(n.length!==2)return null;let e=n[0],t=n[1];if(!(e instanceof I)||!(t instanceof I))return null;let r=I.broadcastShape(e.shape,t.shape);return r?[new I(r,"bool")]:null}function x0(n){if(n.length!==1)return null;let e=n[0];return!(e instanceof I)||!jt(e.dtype)?null:[new I(e.shape,e.dtype)]}function Ca(n){let e=[];if(n.numOperands!==2)return e.push(`${n.opName} expects 2 operands, got ${n.numOperands}`),e;let t=n.getOperand(0).type,r=n.getOperand(1).type;return t instanceof I||e.push(`${n.opName} operand 0 is not tensor`),r instanceof I||e.push(`${n.opName} operand 1 is not tensor`),t instanceof I&&r instanceof I&&t.dtype!==r.dtype&&e.push(`${n.opName} dtype mismatch: ${t.dtype} vs ${r.dtype}`),e}function hp(n){let e=[];return n.numOperands!==1?(e.push(`${n.opName} expects 1 operand, got ${n.numOperands}`),e):(n.getOperand(0).type instanceof I||e.push(`${n.opName} operand is not tensor`),e)}function S0(n){let e=hp(n);if(e.length===0){let t=n.getOperand(0).type.dtype;jt(t)||e.push(`${n.opName} requires float type, got ${t}`)}return e}var ol,v0,nm=M(()=>{"use strict";Re();Bn();l(La,"inferBinaryElementwise");l(tm,"inferUnaryElementwise");l(FN,"inferCompare");l(x0,"inferUnaryFloat");l(Ca,"verifyBinaryElementwise");l(hp,"verifyUnaryElementwise");l(S0,"verifyUnaryFloat");ol=[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],v0=[...ol,Y.COMMUTATIVE,Y.ASSOCIATIVE]});function rm(n){return e=>{if(!(typeof e[0]!="number"||typeof e[1]!="number"))return n(e[0],e[1])}}function vP(n){return e=>typeof e[0]=="number"?n(e[0]):void 0}function zN(n){n.register(new Z({name:"add",numOperands:2,numResults:1,traits:v0,inferResultTypes:La,verify:Ca,getCanonicalizationPatterns(){return[T0("add"),new Zu]},fold:rm((e,t)=>e+t)})),n.register(new Z({name:"mul",numOperands:2,numResults:1,traits:v0,inferResultTypes:La,verify:Ca,getCanonicalizationPatterns(){return[T0("mul"),new el,new tl]},fold:rm((e,t)=>e*t)})),n.register(new Z({name:"sub",numOperands:2,numResults:1,traits:ol,inferResultTypes:La,verify:Ca,getCanonicalizationPatterns(){return[new Qu,new Xu]},fold:rm((e,t)=>e-t)})),n.register(new Z({name:"div",numOperands:2,numResults:1,traits:ol,inferResultTypes:La,verify:Ca,getCanonicalizationPatterns(){return[new nl]},fold:rm((e,t)=>e/t)}));for(let e of["rem","pow"])n.register(new Z({name:e,numOperands:2,numResults:1,traits:ol,inferResultTypes:La,verify:Ca}));n.register(new Z({name:"neg",numOperands:1,numResults:1,traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:tm,verify:hp,getCanonicalizationPatterns(){return[new rl]},fold:vP(e=>-e)}));for(let e of["maximum","minimum"])n.register(new Z({name:e,numOperands:2,numResults:1,traits:[...ol,Y.COMMUTATIVE],inferResultTypes:La,verify:Ca}))}var VN=M(()=>{"use strict";Bn();sl();nm();l(rm,"scalarBinaryFold");l(vP,"scalarUnaryFold");l(zN,"register")});function GN(n){for(let e of["abs","floor","ceil","round","sign","square","reciprocal"])n.register(new Z({name:e,numOperands:1,numResults:1,traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:tm,verify:hp}));n.register(new Z({name:"exp",numOperands:1,numResults:1,traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:x0,fold(e){return typeof e[0]=="number"?Math.exp(e[0]):void 0},verify:S0}));for(let e of["log","sqrt","rsqrt","tanh","sin","cos","erf","erfc","lgamma","gamma","log2","log10","exp2"])n.register(new Z({name:e,numOperands:1,numResults:1,traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:x0,verify:S0}))}var jN=M(()=>{"use strict";Bn();nm();l(GN,"register")});function kP(n){if(n.length!==1)return null;let e=n[0];return!(e instanceof I)||e.dtype!=="bool"?null:[new I(e.shape,"bool")]}function NP(n){if(n.length!==2)return null;let[e,t]=n;if(!(e instanceof I)||e.dtype!=="bool"||!(t instanceof I)||t.dtype!=="bool")return null;let r=I.broadcastShape(e.shape,t.shape);return r?[new I(r,"bool")]:null}function UN(n){n.register(new Z({name:"compare",numOperands:2,numResults:1,attrs:[{name:"direction",type:"string",required:!0}],traits:[Y.ELEMENTWISE],inferResultTypes:FN,verify(e){let t=[];if(e.numOperands!==2)return t.push("compare expects 2 operands"),t;if(!e.hasAttr("direction"))t.push("compare missing direction attr");else{let r=e.getAttr("direction");AP.has(r)||t.push(`compare invalid direction: ${r}`)}return t}})),n.register(new Z({name:"select",numOperands:3,numResults:1,traits:[Y.ELEMENTWISE],inferResultTypes(e){if(e.length!==3)return null;let[t,r,s]=e;if(!(t instanceof I)||t.dtype!=="bool"||!(r instanceof I)||!(s instanceof I)||r.dtype!==s.dtype)return null;let o=I.broadcastShape(t.shape,r.shape,s.shape);return o?[new I(o,r.dtype)]:null}})),n.register(new Z({name:"where",numOperands:3,numResults:1,traits:[Y.ELEMENTWISE],inferResultTypes(e){if(e.length!==3)return null;let[t,r,s]=e;if(!(t instanceof I)||!(r instanceof I)||!(s instanceof I)||r.dtype!==s.dtype)return null;let o=I.broadcastShape(t.shape,r.shape,s.shape);return o?[new I(o,r.dtype)]:null}})),n.register(new Z({name:"logical_not",numOperands:1,numResults:1,traits:[Y.ELEMENTWISE],inferResultTypes:kP}));for(let e of["logical_and","logical_or"])n.register(new Z({name:e,numOperands:2,numResults:1,traits:[Y.ELEMENTWISE],inferResultTypes:NP}));n.register(new Z({name:"clamp",numOperands:3,numResults:1,traits:[Y.ELEMENTWISE],inferResultTypes(e){if(e.length!==3)return null;let t=e[1];return t instanceof I?[new I(t.shape,t.dtype)]:null}}))}var AP,qN=M(()=>{"use strict";Bn();Re();nm();l(kP,"inferUnaryBool");l(NP,"inferBinaryBool");AP=new Set(["eq","ne","lt","le","gt","ge"]);l(UN,"register")});function WN(n){n.register(new Z({name:"stop_gradient",numOperands:1,numResults:1,traits:[Y.VIEW],inferResultTypes(e){return e.length!==1?null:[e[0]]},propagateSymbolicShapes(e,t){let r=t.get(e.getOperand(0));return r?[r]:null}})),n.register(new Z({name:"reverse",numOperands:1,numResults:1,traits:[Y.INJECTIVE],attrs:[{name:"dimensions",type:"array",required:!0}],inferResultTypes(e){return e.length!==1?null:[e[0]]},propagateSymbolicShapes(e,t){let r=t.get(e.getOperand(0));return r?[r]:null}})),n.register(new Z({name:"broadcast_in_dim",numOperands:1,numResults:1,traits:[Y.BROADCAST],attrs:[{name:"broadcast_dimensions",type:"array",required:!0},{name:"result_shape",type:"array",required:!0}],inferResultTypes(e,t){if(e.length!==1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("result_shape"):t.result_shape;return s?[new I(s,r.dtype)]:null},propagateSymbolicShapes(e,t){let r=t.get(e.getOperand(0));if(!r)return null;let s=e.getAttr("broadcast_dimensions"),o=e.getAttr("result_shape");if(!s||!o)return null;let i=o.map(a=>a===X?null:a);for(let a=0;a<s.length;a++)(typeof r[a]!="number"||i[s[a]]===null)&&(i[s[a]]=r[a]);return[i]},fold(e){if(e.length===1&&typeof e[0]=="number")return e[0]},verify(e){let t=[];if(e.hasAttr("broadcast_dimensions")||t.push("broadcast_in_dim missing broadcast_dimensions"),e.hasAttr("result_shape")||t.push("broadcast_in_dim missing result_shape"),e.numOperands!==1&&t.push("broadcast_in_dim expects 1 operand"),t.length===0){let r=e.getAttr("broadcast_dimensions"),s=e.getAttr("result_shape"),o=e.getOperand(0).type;r.length!==o.rank&&t.push(`broadcast_dimensions length ${r.length} != input rank ${o.rank}`);for(let a=0;a<r.length;a++)r[a]<0||r[a]>=s.length?t.push(`broadcast_dimensions[${a}]=${r[a]} out of range for result rank ${s.length}`):o instanceof I&&o.shape[a]!==X&&o.shape[a]!==1&&s[r[a]]!==X&&o.shape[a]!==s[r[a]]&&t.push(`broadcast_in_dim: input dim ${a} size ${o.shape[a]} incompatible with result dim ${r[a]} size ${s[r[a]]}`);let i=new Set;for(let a of r)i.has(a)&&t.push(`broadcast_dimensions has duplicate: ${a}`),i.add(a)}return t}})),n.register(new Z({name:"reshape",numOperands:1,numResults:1,traits:[Y.VIEW],attrs:[{name:"new_shape",type:"array",required:!0}],inferResultTypes(e,t){if(e.length!==1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("new_shape"):t.new_shape;return s?[new I(s,r.dtype)]:null},verify(e){let t=[];if(e.hasAttr("new_shape")||t.push("reshape missing new_shape"),e.numOperands!==1&&t.push("reshape expects 1 operand"),t.length===0){let r=e.getOperand(0).type,s=e.getAttr("new_shape");if(r instanceof I&&r.isFullyStatic){let o=s.filter(i=>i===X).length;if(o>1&&t.push("reshape can have at most one dynamic dimension"),o===0){let i=r.numel(),a=s.reduce((u,p)=>u*p,1);i!==a&&t.push(`reshape numel mismatch: input ${i} vs output ${a}`)}}}return t},propagateSymbolicShapes(e,t){let r=t.get(e.getOperand(0));if(!r)return null;let s=e.getAttr("new_shape"),o=[];for(let i of s)if(i===-1){let a=r.find(u=>typeof u!="number");o.push(a||-1)}else o.push(i);return[o]},getCanonicalizationPatterns(){return[new zh,new Ju]},fold(e){return e[0]}})),n.register(new Z({name:"transpose",numOperands:1,numResults:1,traits:[Y.VIEW],attrs:[{name:"permutation",type:"array",required:!0}],inferResultTypes(e,t){if(e.length!==1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("permutation"):t.permutation;if(!s)return null;let o=s.map(i=>r.shape[i]);return[new I(o,r.dtype)]},propagateSymbolicShapes(e,t){let r=t.get(e.getOperand(0));if(!r)return null;let s=e.getAttr("permutation");return s?[s.map(o=>r[o])]:null},getCanonicalizationPatterns(){return[new Vh]},fold(e,t,r){let s=e[0];if(s==null||typeof s=="number"||typeof s.length!="number")return;let o=t.get?t.get("permutation"):t.permutation;if(!o)return;let i=r&&r[0]?r[0].getResult(0).type:null,a=i instanceof I?i.shape:null;if(!a||a.length!==o.length)return;let u=new Array(a.length),p=1;for(let m=a.length-1;m>=0;m--)u[m]=p,p*=a[m];let c=o.map(m=>a[m]),f=s.length,d=new Array(f),h=new Array(o.length).fill(0);for(let m=0;m<f;m++){let g=0;for(let y=0;y<o.length;y++)g+=h[y]*u[o[y]];d[m]=s[g];for(let y=o.length-1;y>=0&&!(++h[y]<c[y]);y--)h[y]=0}return d},verify(e){let t=[];if(!e.hasAttr("permutation"))return t.push("transpose missing permutation"),t;if(e.numOperands!==1)return t.push("transpose expects 1 operand"),t;let r=e.getAttr("permutation"),s=e.getOperand(0).type;s instanceof I&&r.length!==s.rank&&t.push(`transpose permutation length ${r.length} != input rank ${s.rank}`);let o=new Set;for(let i of r)o.has(i)&&t.push(`transpose duplicate in permutation: ${i}`),o.add(i);return t}})),n.register(new Z({name:"slice",numOperands:1,numResults:1,traits:[Y.VIEW],attrs:[{name:"starts",type:"array",required:!0},{name:"limits",type:"array",required:!0},{name:"strides",type:"array",required:!1}],inferResultTypes(e,t){if(e.length!==1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("starts"):t.starts,o=t.get?t.get("limits"):t.limits,i=(t.get?t.get("strides"):t.strides)||s.map(()=>1),a=[];for(let u=0;u<s.length;u++)a.push(Math.ceil((o[u]-s[u])/i[u]));return[new I(a,r.dtype)]},getCanonicalizationPatterns(){return[new Uh]}})),n.register(new Z({name:"concat",numOperands:-1,numResults:1,attrs:[{name:"dimension",type:"number",required:!0}],inferResultTypes(e,t){if(e.length<1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("dimension"):t.dimension;if(s===void 0)return null;let o=[...r.shape];for(let i=1;i<e.length;i++){let a=e[i];if(!(a instanceof I)||a.dtype!==r.dtype||a.rank!==r.rank)return null;o[s]===X||a.shape[s]===X?o[s]=X:o[s]+=a.shape[s]}return[new I(o,r.dtype)]}})),n.register(new Z({name:"pad",numOperands:2,numResults:1,traits:[Y.INJECTIVE],attrs:[{name:"low",type:"array",required:!0},{name:"high",type:"array",required:!0},{name:"interior",type:"array",required:!1}],inferResultTypes(e,t){if(e.length<1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("low"):t.low,o=t.get?t.get("high"):t.high,i=(t.get?t.get("interior"):t.interior)||s.map(()=>0),a=[];for(let u=0;u<r.rank;u++)r.shape[u]===X?a.push(X):a.push(s[u]+r.shape[u]+(r.shape[u]-1)*i[u]+o[u]);return[new I(a,r.dtype)]},getCanonicalizationPatterns(){return[new jh]}})),n.register(new Z({name:"gather",numOperands:2,numResults:1,traits:[Y.INJECTIVE],attrs:[{name:"offset_dims",type:"array",required:!0},{name:"collapsed_slice_dims",type:"array",required:!0},{name:"start_index_map",type:"array",required:!0},{name:"slice_sizes",type:"array",required:!0},{name:"index_vector_dim",type:"number",required:!0}],inferResultTypes(e,t){if(e.length!==2)return null;let r=e[0],s=e[1];if(!(r instanceof I)||!(s instanceof I))return null;let o=t.get?t.get("offset_dims"):t.offset_dims,i=new Set(t.get?t.get("collapsed_slice_dims"):t.collapsed_slice_dims),a=t.get?t.get("slice_sizes"):t.slice_sizes,u=t.get?t.get("index_vector_dim"):t.index_vector_dim,p=[];for(let y=0;y<s.rank;y++)y!==u&&p.push(s.shape[y]);let c=[];for(let y=0;y<a.length;y++)i.has(y)||c.push(a[y]);let f=new Set(o),d=[],h=0,m=0,g=p.length+c.length;for(let y=0;y<g;y++)f.has(y)?d.push(c[m++]):d.push(p[h++]);return[new I(d,r.dtype)]}})),n.register(new Z({name:"scatter",numOperands:3,numResults:1,traits:[Y.INJECTIVE],attrs:[{name:"update_window_dims",type:"array",required:!0},{name:"inserted_window_dims",type:"array",required:!0},{name:"scatter_dims_to_operand_dims",type:"array",required:!0},{name:"index_vector_dim",type:"number",required:!0}],hasRegions:!0,numRegions:1,sideEffects:2,inferResultTypes(e){if(e.length<1)return null;let t=e[0];return t instanceof I?[new I(t.shape,t.dtype)]:null}})),n.register(new Z({name:"split",numOperands:1,numResults:-1,attrs:[{name:"dimension",type:"number",required:!0},{name:"split_sizes",type:"array",required:!0}],inferResultTypes(e,t){if(e.length<1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("dimension"):t.dimension,o=t.get?t.get("split_sizes"):t.split_sizes;return s===void 0||!o?null:o.map(i=>{let a=[...r.shape];return a[s]=i,new I(a,r.dtype)})}}))}var KN=M(()=>{"use strict";Bn();Re();sl();l(WN,"register")});function HN(n){n.register(new Z({name:"reduce",numOperands:2,numResults:1,attrs:[{name:"dimensions",type:"array",required:!0},{name:"reduce_type",type:"string",required:!0}],traits:[Y.REDUCTION],hasRegions:!0,numRegions:1,inferResultTypes(t,r){if(t.length<1)return null;let s=t[0];if(!(s instanceof I))return null;let o=r.get?r.get("dimensions"):r.dimensions;if(!o)return null;let i=new Set(o),a=[];for(let u=0;u<s.rank;u++)i.has(u)||a.push(s.shape[u]);return[new I(a,s.dtype)]},propagateSymbolicShapes(t,r){let s=r.get(t.getOperand(0));if(!s)return null;let o=t.getAttr("dimensions");if(!o)return null;let i=new Set(o),a=[];for(let u=0;u<s.length;u++)i.has(u)||a.push(s[u]);return[a]},verify(t){let r=[];if(t.hasAttr("dimensions")||r.push("reduce missing dimensions"),!t.hasAttr("reduce_type"))r.push("reduce missing reduce_type");else{let s=t.getAttr("reduce_type");OP.has(s)||r.push(`reduce invalid reduce_type: ${s}`)}return r}}));function e(t,r){if(t.length<1)return null;let s=t[0];if(!(s instanceof I))return null;let o=r.get?r.get("axis"):r.axis;if(o===void 0)return null;let i=(r.get?r.get("keep_dims"):r.keep_dims)||!1,a=[];for(let u=0;u<s.rank;u++)u===o?i&&a.push(1):a.push(s.shape[u]);return[new I(a,"i32")]}l(e,"inferArgReduceTypes"),n.register(new Z({name:"argmax",numOperands:1,numResults:1,attrs:[{name:"axis",type:"number",required:!0},{name:"keep_dims",type:"boolean",required:!1}],traits:[Y.REDUCTION],inferResultTypes:e})),n.register(new Z({name:"argmin",numOperands:1,numResults:1,attrs:[{name:"axis",type:"number",required:!0},{name:"keep_dims",type:"boolean",required:!1}],traits:[Y.REDUCTION],inferResultTypes:e}))}var OP,YN=M(()=>{"use strict";Bn();Re();OP=new Set(["sum","max","min","prod","mean","and","or"]);l(HN,"register")});function BP(n,e){let t=n.getAttr("scale"),r=e.getAttr("scale"),s=n.getAttr("zero_point"),o=e.getAttr("zero_point"),i=n.getAttr("scheme"),a=e.getAttr("scheme");if(i!==a)return!1;if(typeof t=="number"&&typeof r=="number"){if(t!==r)return!1}else return!1;if(typeof s=="number"&&typeof o=="number"){if(s!==o)return!1}else return!1;return!0}var sm,om,im,k0=M(()=>{"use strict";cp();Re();l(BP,"quantParamsMatch");sm=class extends dt{static{l(this,"QuantizeDequantizeIdentity")}constructor(){super("quantize_dequantize_identity",20),this.rootOpName="quantize"}match(e){let t=e.getOperand(0).definingOp;return!t||t.opName!=="dequantize"?!1:BP(e,t)}rewrite(e,t){let r=e.getOperand(0).definingOp.getOperand(0);if(!r.type||!e.getResult(0).type)return!1;let s=r.type,o=e.getResult(0).type;return s instanceof I&&o instanceof I&&s.dtype===o.dtype?(e.replaceAllResultsWith([r]),e.erase(),!0):!1}},om=class extends dt{static{l(this,"ConstantQuantize")}constructor(){super("constant_quantize",15),this.rootOpName="quantize"}match(e){let t=e.getOperand(0).definingOp;return!t||t.opName!=="constant"?!1:typeof t.getAttr("value")=="number"}rewrite(e,t){let s=e.getOperand(0).definingOp.getAttr("value"),o=e.getAttr("scale"),i=e.getAttr("zero_point");if(typeof o!="number"||typeof i!="number")return!1;let a=e.getAttr("target_dtype"),u=fo(a)*8,p=a==="ui8",c=p?0:-(1<<u-1),f=p?(1<<u)-1:(1<<u-1)-1,d=Math.max(c,Math.min(f,Math.round(s/o+i))),h=e.getResult(0).type,m=t.constant(d,h);return e.replaceAllResultsWith([m.getResult(0)]),e.erase(),!0}},im=class extends dt{static{l(this,"DequantizeFoldIntoDot")}constructor(){super("dequantize_fold_into_dot",15),this.rootOpName="dot"}match(e){if(e.numOperands!==2)return!1;let t=e.getOperand(0).definingOp,r=e.getOperand(1).definingOp;return t&&t.opName==="dequantize"&&r&&r.opName==="dequantize"}rewrite(e,t){let r=e.getOperand(0).definingOp,s=e.getOperand(1).definingOp,o=r.getOperand(0),i=s.getOperand(0),a=r.getAttr("scale"),u=s.getAttr("scale"),p=r.getAttr("zero_point"),c=s.getAttr("zero_point");if(typeof a!="number"||typeof u!="number"||typeof p!="number"||typeof c!="number"||p!==0||c!==0)return!1;let f=a*u,d=e.getAttr("lhs_contracting"),h=e.getAttr("rhs_contracting"),m=e.getAttr("lhs_batch")||[],g=e.getAttr("rhs_batch")||[],y=new I(e.getResult(0).type.shape,"i32"),_=t._buildOp("quantized_dot",[o,i],[y],{lhs_contracting:d,rhs_contracting:h,lhs_batch:m,rhs_batch:g,lhs_scale:a,lhs_zero_point:p,rhs_scale:u,rhs_zero_point:c,output_scale:f,output_zero_point:0}),b=t._inferAndBuild("dequantize",[_.getResult(0)],{scale:f,zero_point:0,scheme:r.getAttr("scheme"),target_dtype:e.getResult(0).type.dtype});return e.replaceAllResultsWith([b.getResult(0)]),e.erase(),!0}}});function ZN(n){n.register(new Z({name:"dot",numOperands:2,numResults:1,opAttrs:{gpuCapable:!0},traits:[Y.OPAQUE,Y.OUT_EWISE_FUSABLE],attrs:[{name:"lhs_contracting",type:"array",required:!0},{name:"rhs_contracting",type:"array",required:!0},{name:"lhs_batch",type:"array",required:!1},{name:"rhs_batch",type:"array",required:!1}],getFlops(e){let t=e.getOperand(0).type,r=e.getOperand(1).type;if(!(t instanceof I)||!(r instanceof I))return 0;let s=e.getAttr("lhs_contracting")||[],o=1;for(let u of s)t.shape[u]!==X&&(o*=t.shape[u]);let i=e.getResult(0).type;if(!(i instanceof I))return 0;let a=i.numel();return a===X?0:2*a*o},inferResultTypes:am,getCanonicalizationPatterns(){return[new Xh,new im]},verify(e){let t=[];if(e.numOperands!==2)return t.push("dot expects 2 operands"),t;e.hasAttr("lhs_contracting")||t.push("dot missing lhs_contracting"),e.hasAttr("rhs_contracting")||t.push("dot missing rhs_contracting");let r=e.getOperand(0).type,s=e.getOperand(1).type;if(r instanceof I&&s instanceof I){r.dtype!==s.dtype&&t.push(`dot dtype mismatch: ${r.dtype} vs ${s.dtype}`);let o=e.getAttr("lhs_contracting")||[],i=e.getAttr("rhs_contracting")||[];if(o.length!==i.length)t.push(`dot contracting dimensions count mismatch: lhs ${o.length} vs rhs ${i.length}`);else for(let a=0;a<o.length;a++){let u=o[a],p=i[a];u>=r.rank&&t.push(`dot lhs_contracting[${a}]=${u} out of range (rank ${r.rank})`),p>=s.rank&&t.push(`dot rhs_contracting[${a}]=${p} out of range (rank ${s.rank})`);let c=r.shape[u],f=s.shape[p];u<r.rank&&p<s.rank&&typeof c=="number"&&c!==X&&typeof f=="number"&&f!==X&&c!==f&&t.push(`dot contracting dim size mismatch at [${a}]: lhs dim ${u} size ${c} vs rhs dim ${p} size ${f}`)}}return t}})),n.register(new Z({name:"cublas_gemm",numOperands:2,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"lhs_contracting",type:"array",required:!0},{name:"rhs_contracting",type:"array",required:!0},{name:"lhs_batch",type:"array",required:!1},{name:"rhs_batch",type:"array",required:!1}],inferResultTypes:am,getFlops(e){let t=e.getOperand(0).type,r=e.getResult(0).type;if(!(t instanceof I)||!(r instanceof I))return 0;let s=1;for(let i of e.getAttr("lhs_contracting")||[])t.shape[i]!==X&&(s*=t.shape[i]);let o=r.numel();return o===X?0:2*o*s},verify(e){let t=[];return e.numOperands!==2?(t.push("cublas_gemm expects 2 operands"),t):(e.hasAttr("lhs_contracting")||t.push("cublas_gemm missing lhs_contracting"),e.hasAttr("rhs_contracting")||t.push("cublas_gemm missing rhs_contracting"),t)}})),n.register(new Z({name:"conv",numOperands:2,numResults:1,opAttrs:{gpuCapable:!0},traits:[Y.OPAQUE],attrs:[{name:"strides",type:"array",required:!0},{name:"padding",type:"array",required:!0},{name:"dilation",type:"array",required:!1},{name:"groups",type:"number",required:!1},{name:"input_layout",type:"string",required:!0},{name:"kernel_layout",type:"string",required:!0}],getFlops(e){let t=e.getResult(0).type,r=e.getOperand(1).type;if(!(t instanceof I)||!(r instanceof I))return 0;let s=t.numel(),o=r.numel();return s===X||o===X?0:2*s*o/(r.shape[0]||1)},inferResultTypes:N0,verify(e){let t=[];return e.numOperands!==2?(t.push("conv expects 2 operands"),t):(e.hasAttr("strides")||t.push("conv missing strides"),e.hasAttr("padding")||t.push("conv missing padding"),e.hasAttr("input_layout")||t.push("conv missing input_layout"),e.hasAttr("kernel_layout")||t.push("conv missing kernel_layout"),t)}}))}function am(n,e,t){if(n.length!==2)return null;let r=n[0],s=n[1];if(!(r instanceof I)||!(s instanceof I))return null;let o=t&&!Array.isArray(t)?t:{};if(!o.allowMixedDtype&&r.dtype!==s.dtype)return null;let i=new Set(e.get?e.get("lhs_contracting"):e.lhs_contracting),a=new Set(e.get?e.get("rhs_contracting"):e.rhs_contracting),u=new Set((e.get?e.get("lhs_batch"):e.lhs_batch)||[]),p=new Set((e.get?e.get("rhs_batch"):e.rhs_batch)||[]),c=[];for(let f=0;f<r.rank;f++)u.has(f)&&c.push(r.shape[f]);for(let f=0;f<r.rank;f++)!i.has(f)&&!u.has(f)&&c.push(r.shape[f]);for(let f=0;f<s.rank;f++)!a.has(f)&&!p.has(f)&&c.push(s.shape[f]);return[new I(c,o.outputDtype||r.dtype)]}function N0(n,e,t){if(n.length!==2)return null;let r=n[0],s=n[1];if(!(r instanceof I)||!(s instanceof I))return null;let o=t&&!Array.isArray(t)?t:{};if(!o.allowMixedDtype&&r.dtype!==s.dtype)return null;let i=e.get?e.get("strides"):e.strides,a=e.get?e.get("padding"):e.padding,u=(e.get?e.get("dilation"):e.dilation)||i.map(()=>1),p=i.length,c=r.shape[0],f=s.shape[0],d=[];for(let h=0;h<p;h++){let m=r.shape[h+2],g=s.shape[h+2],y=a[h][0]+a[h][1];if(m===X||g===X)d.push(X);else{let _=(g-1)*u[h]+1;d.push(Math.floor((m+y-_)/i[h])+1)}}return[new I([c,f,...d],o.outputDtype||r.dtype)]}var A0=M(()=>{"use strict";Bn();Re();sl();k0();l(ZN,"register");l(am,"inferDotResultTypes");l(N0,"inferConvResultTypes")});function QN(n){let e={numOperands:0,numResults:1,attrs:[{name:"value",type:"any",required:!0},{name:"tensor_type",type:"object",required:!0}],traits:[Y.CONSTANT],inferResultTypes(t,r){let s=r.get?r.get("tensor_type"):r.tensor_type;return s?[s]:null}};for(let t of["constant","scalar_constant"])n.register(new Z({name:t,...e}));n.register(new Z({name:"iota",numOperands:0,numResults:1,attrs:[{name:"iota_dimension",type:"number",required:!0},{name:"tensor_type",type:"object",required:!0}],inferResultTypes(t,r){let s=r.get?r.get("tensor_type"):r.tensor_type;return s?[s]:null}})),n.register(new Z({name:"tuple",numOperands:-1,numResults:1,inferResultTypes(t){return[new Ia(t)]}})),n.register(new Z({name:"get_tuple_element",numOperands:1,numResults:1,attrs:[{name:"index",type:"number",required:!0}],inferResultTypes(t,r){if(t.length!==1)return null;let s=t[0];if(!s||!s.elements)return null;let o=r.get?r.get("index"):r.index;return o===void 0||o<0||o>=s.elements.length?null:[s.elements[o]]}})),n.register(new Z({name:"convert",numOperands:1,numResults:1,attrs:[{name:"target_dtype",type:"string",required:!0}],traits:[Y.ELEMENTWISE],inferResultTypes(t,r){if(t.length!==1)return null;let s=t[0];if(!(s instanceof I))return null;let o=r.get?r.get("target_dtype"):r.target_dtype;return[new I(s.shape,o)]}})),n.register(new Z({name:"one_hot",numOperands:1,numResults:1,attrs:[{name:"depth",type:"number",required:!0},{name:"axis",type:"number",required:!1},{name:"on_value",type:"number",required:!1},{name:"off_value",type:"number",required:!1},{name:"dtype",type:"string",required:!1}],inferResultTypes(t,r){if(t.length<1)return null;let s=t[0];if(!(s instanceof I))return null;let o=r.get?r.get("depth"):r.depth,i=(r.get?r.get("axis"):r.axis)??-1,a=(r.get?r.get("dtype"):r.dtype)||"f32",u=[...s.shape],p=i<0?u.length+1+i:i;return u.splice(p,0,o),[new I(u,a)]}}))}var XN=M(()=>{"use strict";Bn();Re();l(QN,"register")});function eA(n){n.register(new Z({name:"return",numOperands:-1,numResults:0,traits:[Y.TERMINATOR]})),n.register(new Z({name:"yield",numOperands:-1,numResults:0,traits:[Y.TERMINATOR]})),n.register(new Z({name:"if",numOperands:1,numResults:-1,hasRegions:!0,numRegions:2,sideEffects:gn.CONTROL,inferResultTypes(e,t,r){return r||null}})),n.register(new Z({name:"while",numOperands:-1,numResults:-1,hasRegions:!0,numRegions:2,sideEffects:gn.CONTROL,inferResultTypes(e){return[...e]}})),n.register(new Z({name:"scan",numOperands:-1,numResults:-1,hasRegions:!0,numRegions:1,sideEffects:gn.CONTROL,attrs:[{name:"num_carry",type:"number",required:!0},{name:"num_xs",type:"number",required:!0}],inferResultTypes(e,t,r){return r||null}})),n.register(new Z({name:"custom_call",numOperands:-1,numResults:-1,attrs:[{name:"call_target_name",type:"string",required:!0},{name:"backend_config",type:"any",required:!1}],sideEffects:gn.WRITE,inferResultTypes(e,t,r){return r||null}})),n.register(new Z({name:"fused_dot_epilogue",numOperands:-1,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"lhs_contracting",type:"array",required:!0},{name:"rhs_contracting",type:"array",required:!0},{name:"lhs_batch",type:"array",required:!1},{name:"rhs_batch",type:"array",required:!1},{name:"epilogue_ops",type:"array",required:!0},{name:"epilogue_tags",type:"array",required:!0},{name:"num_dot_operands",type:"number",required:!0},{name:"num_extra_inputs",type:"number",required:!0}],inferResultTypes(e,t,r){return r||null}})),n.register(new Z({name:"fusion",numOperands:-1,numResults:-1,hasRegions:!0,numRegions:1,attrs:[{name:"fusion_kind",type:"string",required:!1}],inferResultTypes(e,t,r){return r||null}}))}var tA=M(()=>{"use strict";Bn();l(eA,"register")});function nA(n){n.register(new Z({name:"layout_transform",numOperands:1,numResults:1,traits:[Y.VIEW],attrs:[{name:"src_layout",type:"array",required:!0},{name:"dst_layout",type:"array",required:!0}],inferResultTypes(e,t){if(e.length!==1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("dst_layout"):t.dst_layout;return s?[new I(r.shape,r.dtype,new Gt(s))]:null},verify(e){let t=[];if(e.numOperands!==1)return t.push("layout_transform expects 1 operand"),t;if(e.hasAttr("src_layout")||t.push("layout_transform missing src_layout"),e.hasAttr("dst_layout")||t.push("layout_transform missing dst_layout"),t.length===0){let r=e.getOperand(0).type,s=e.getAttr("src_layout"),o=e.getAttr("dst_layout");r instanceof I&&(s.length!==r.rank&&t.push(`src_layout length ${s.length} != input rank ${r.rank}`),o.length!==r.rank&&t.push(`dst_layout length ${o.length} != input rank ${r.rank}`))}return t},getCanonicalizationPatterns(){return[new Qh,new em]}}))}var rA=M(()=>{"use strict";Bn();Re();sl();l(nA,"register")});function sA(n){n.register(new Z({name:"quantize",numOperands:1,numResults:1,traits:[Y.ELEMENTWISE],attrs:[{name:"scale",type:"any",required:!0},{name:"zero_point",type:"any",required:!0},{name:"scheme",type:"string",required:!0},{name:"target_dtype",type:"string",required:!0},{name:"axis",type:"number",required:!1}],inferResultTypes(e,t){if(e.length!==1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("target_dtype"):t.target_dtype;return s?[new I(r.shape,s)]:null},verify(e){let t=[];if(e.numOperands!==1)return t.push("quantize expects 1 operand"),t;let r=e.getOperand(0).type;r instanceof I&&!jt(r.dtype)&&t.push(`quantize input must be float, got ${r.dtype}`);let s=e.getAttr("target_dtype");s&&!O0.has(s)&&t.push(`quantize target_dtype must be i8 or ui8, got ${s}`);let o=e.getAttr("scale");return typeof o=="number"&&o<=0&&t.push("quantize scale must be positive"),t},fold(e,t){if(typeof e[0]!="number")return;let r=t.get?t.get("scale"):t.scale,s=t.get?t.get("zero_point"):t.zero_point;if(!(typeof r!="number"||typeof s!="number"))return Math.round(e[0]/r+s)},getCanonicalizationPatterns(){return[new sm,new om]}})),n.register(new Z({name:"dequantize",numOperands:1,numResults:1,traits:[Y.ELEMENTWISE],attrs:[{name:"scale",type:"any",required:!0},{name:"zero_point",type:"any",required:!0},{name:"scheme",type:"string",required:!0},{name:"target_dtype",type:"string",required:!0},{name:"axis",type:"number",required:!1}],inferResultTypes(e,t){if(e.length!==1)return null;let r=e[0];if(!(r instanceof I))return null;let s=t.get?t.get("target_dtype"):t.target_dtype;return s?[new I(r.shape,s)]:null},verify(e){let t=[];if(e.numOperands!==1)return t.push("dequantize expects 1 operand"),t;let r=e.getOperand(0).type;r instanceof I&&!RP.has(r.dtype)&&t.push("dequantize input must be i8, ui8, or i32, got "+r.dtype);let s=e.getAttr("target_dtype");return s&&!jt(s)&&t.push(`dequantize target_dtype must be float, got ${s}`),t},fold(e,t){if(typeof e[0]!="number")return;let r=t.get?t.get("scale"):t.scale,s=t.get?t.get("zero_point"):t.zero_point;if(!(typeof r!="number"||typeof s!="number"))return(e[0]-s)*r}})),n.register(new Z({name:"quantized_dot",numOperands:2,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"lhs_contracting",type:"array",required:!0},{name:"rhs_contracting",type:"array",required:!0},{name:"lhs_batch",type:"array",required:!1},{name:"rhs_batch",type:"array",required:!1},{name:"lhs_scale",type:"number",required:!0},{name:"lhs_zero_point",type:"number",required:!0},{name:"rhs_scale",type:"number",required:!0},{name:"rhs_zero_point",type:"number",required:!0},{name:"output_scale",type:"number",required:!0},{name:"output_zero_point",type:"number",required:!0}],getFlops(e){let t=e.getOperand(0).type,r=e.getOperand(1).type;if(!(t instanceof I)||!(r instanceof I))return 0;let s=e.getAttr("lhs_contracting")||[],o=1;for(let u of s)t.shape[u]!==X&&(o*=t.shape[u]);let i=e.getResult(0).type;if(!(i instanceof I))return 0;let a=i.numel();return a===X?0:2*a*o},inferResultTypes(e,t){return am(e,t,{outputDtype:"i32",allowMixedDtype:!0})},verify(e){let t=[];if(e.numOperands!==2)return t.push("quantized_dot expects 2 operands"),t;e.hasAttr("lhs_contracting")||t.push("quantized_dot missing lhs_contracting"),e.hasAttr("rhs_contracting")||t.push("quantized_dot missing rhs_contracting"),e.hasAttr("lhs_scale")||t.push("quantized_dot missing lhs_scale"),e.hasAttr("rhs_scale")||t.push("quantized_dot missing rhs_scale");let r=e.getOperand(0).type,s=e.getOperand(1).type;return r instanceof I&&!O0.has(r.dtype)&&t.push(`quantized_dot lhs must be i8/ui8, got ${r.dtype}`),s instanceof I&&!O0.has(s.dtype)&&t.push(`quantized_dot rhs must be i8/ui8, got ${s.dtype}`),t}})),n.register(new Z({name:"quantized_conv",numOperands:2,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"strides",type:"array",required:!0},{name:"padding",type:"array",required:!0},{name:"dilation",type:"array",required:!1},{name:"groups",type:"number",required:!1},{name:"input_layout",type:"string",required:!0},{name:"kernel_layout",type:"string",required:!0},{name:"input_scale",type:"number",required:!0},{name:"input_zero_point",type:"number",required:!0},{name:"kernel_scale",type:"number",required:!0},{name:"kernel_zero_point",type:"number",required:!0},{name:"output_scale",type:"number",required:!0},{name:"output_zero_point",type:"number",required:!0}],inferResultTypes(e,t){return N0(e,t,{outputDtype:"i32",allowMixedDtype:!0})},verify(e){let t=[];return e.numOperands!==2?(t.push("quantized_conv expects 2 operands"),t):(e.hasAttr("strides")||t.push("quantized_conv missing strides"),e.hasAttr("padding")||t.push("quantized_conv missing padding"),e.hasAttr("input_layout")||t.push("quantized_conv missing input_layout"),e.hasAttr("kernel_layout")||t.push("quantized_conv missing kernel_layout"),e.hasAttr("input_scale")||t.push("quantized_conv missing input_scale"),e.hasAttr("kernel_scale")||t.push("quantized_conv missing kernel_scale"),t)}}))}var O0,RP,oA=M(()=>{"use strict";Bn();Re();k0();A0();O0=new Set(["i8","ui8"]),RP=new Set(["i8","ui8","i32"]);l(sA,"register")});function MP(n){if(n.length<1)return null;let e=n[0];return e instanceof I?[new I(e.shape,e.dtype)]:null}function js(n){if(n.length<1)return null;let e=n[0];return!(e instanceof I)||!jt(e.dtype)?null:[new I(e.shape,e.dtype)]}function Zo(n){let e=[];if(n.numOperands<1)return e.push(`${n.opName} expects at least 1 operand`),e;let t=n.getOperand(0).type;return t instanceof I&&!jt(t.dtype)&&e.push(`${n.opName} requires float input, got ${t.dtype}`),e}function iA(n){n.register(new Z({name:"all_reduce",numOperands:1,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"reduce_op",type:"string",required:!1},{name:"mesh_axis",type:"number",required:!1}],inferResultTypes:MP})),n.register(new Z({name:"all_gather",numOperands:1,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"mesh_axis",type:"number",required:!1},{name:"gather_dim",type:"number",required:!1}],inferResultTypes(e,t){let r=e[0];if(!(r instanceof I))return null;let s=t&&t.has("mesh_axis")?t.get("mesh_axis"):0,o=t&&t.has("gather_dim")?t.get("gather_dim"):1,i=[...r.shape];return i[o]=i[o]*i[s],[new I(i,r.dtype)]}})),n.register(new Z({name:"scaled_dot_product_attention",numOperands:3,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"scale",type:"number",required:!0},{name:"causal",type:"boolean",required:!1}],inferResultTypes(e){let t=e[0],r=e[2];return!(t instanceof I)||!(r instanceof I)?null:[new I([...t.shape.slice(0,t.rank-1),r.shape[r.rank-1]],t.dtype)]}})),n.register(new Z({name:"softmax",numOperands:1,numResults:1,opAttrs:{gpuCapable:!0},attrs:[{name:"axis",type:"number",required:!0}],inferResultTypes:js,verify:Zo})),n.register(new Z({name:"log_softmax",numOperands:1,numResults:1,attrs:[{name:"axis",type:"number",required:!0}],inferResultTypes:js,verify:Zo})),n.register(new Z({name:"gelu",numOperands:1,numResults:1,traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:js,verify:Zo})),n.register(new Z({name:"sigmoid",numOperands:1,numResults:1,traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:js,verify:Zo})),n.register(new Z({name:"silu",numOperands:1,numResults:1,traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:js,verify:Zo})),n.register(new Z({name:"layer_norm",numOperands:3,numResults:1,opAttrs:{gpuCapable:!0},attrs:[{name:"axis",type:"number",required:!0},{name:"epsilon",type:"number",required:!0}],inferResultTypes:js,verify(e){let t=[];if(e.numOperands!==3)return t.push("layer_norm expects 3 operands (input, gamma, beta)"),t;let r=e.getOperand(0).type;return r instanceof I&&!jt(r.dtype)&&t.push(`layer_norm requires float input, got ${r.dtype}`),t}})),n.register(new Z({name:"batch_norm",numOperands:5,numResults:1,opAttrs:{gpuCapable:!0},attrs:[{name:"axis",type:"number",required:!0},{name:"epsilon",type:"number",required:!0}],inferResultTypes:js,verify(e){let t=[];if(e.numOperands!==5)return t.push("batch_norm expects 5 operands (input, gamma, beta, mean, var)"),t;let r=e.getOperand(0).type;return r instanceof I&&!jt(r.dtype)&&t.push(`batch_norm requires float input, got ${r.dtype}`),t}}));for(let e of["selu","mish","hardswish","hardsigmoid"])n.register(new Z({name:e,numOperands:1,numResults:1,traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:js,verify:Zo}));n.register(new Z({name:"elu",numOperands:1,numResults:1,attrs:[{name:"alpha",type:"number"}],traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:js,verify:Zo})),n.register(new Z({name:"leaky_relu",numOperands:1,numResults:1,attrs:[{name:"negative_slope",type:"number"}],traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:js,verify:Zo})),n.register(new Z({name:"celu",numOperands:1,numResults:1,attrs:[{name:"alpha",type:"number"}],traits:[Y.ELEMENTWISE,Y.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:js,verify:Zo})),n.register(new Z({name:"embedding",numOperands:2,numResults:1,inferResultTypes(e){if(e.length<2)return null;let t=e[0],r=e[1];if(!(t instanceof I)||!(r instanceof I))return null;let s=[...r.shape,t.shape[t.rank-1]];return[new I(s,t.dtype)]}}))}var aA=M(()=>{"use strict";Bn();Re();l(MP,"inferSameAsInput");l(js,"inferSameAsInputFloat");l(Zo,"verifyUnaryFloat");l(iA,"register")});function uA(n){n.register(new Z({name:"copy_to_device",numOperands:1,numResults:1,attrs:[{name:"src_device",type:"string",required:!0},{name:"dst_device",type:"string",required:!0}],sideEffects:gn.READ|gn.WRITE,traits:[Y.INJECTIVE],inferResultTypes(e){return e.length!==1?null:[e[0]]},getFlops(){return 0}}))}var lA=M(()=>{"use strict";Bn();l(uA,"register")});function cA(n,e,t,r,s,o){let i=n+r+s-e;return(o?Math.ceil(i/t):Math.floor(i/t))+1}function pA(n){n.register(new Z({name:"pool2d",numOperands:1,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"pool_type",type:"string",required:!0},{name:"kernel_size",type:"array",required:!0},{name:"strides",type:"array",required:!0},{name:"padding",type:"array",required:!0},{name:"ceil_mode",type:"boolean",required:!1},{name:"count_include_pad",type:"boolean",required:!1},{name:"layout",type:"string",required:!1}],inferResultTypes(e,t){if(e.length<1)return null;let r=e[0];if(!(r instanceof I)||r.rank!==4)return null;let s=t.get?t.get("kernel_size"):t.kernel_size,o=t.get?t.get("strides"):t.strides,i=t.get?t.get("padding"):t.padding,a=(t.get?t.get("ceil_mode"):t.ceil_mode)||!1,u=cA(r.shape[2],s[0],o[0],i[0][0],i[0][1],a),p=cA(r.shape[3],s[1],o[1],i[1][0],i[1][1],a);return[new I([r.shape[0],r.shape[1],u,p],r.dtype)]}}))}var fA=M(()=>{"use strict";Bn();Re();l(cA,"computePoolOutputDim");l(pA,"register")});function dA(n){n.register(new Z({name:"resize",numOperands:1,numResults:1,traits:[Y.OPAQUE],attrs:[{name:"output_size",type:"array",required:!0},{name:"method",type:"string",required:!0},{name:"coordinate_mode",type:"string",required:!1},{name:"layout",type:"string",required:!1}],inferResultTypes(e,t){if(e.length<1)return null;let r=e[0];if(!(r instanceof I)||r.rank!==4)return null;let s=t.get?t.get("output_size"):t.output_size;return[new I([r.shape[0],r.shape[1],s[0],s[1]],r.dtype)]}}))}var hA=M(()=>{"use strict";Bn();Re();l(dA,"register")});function EP(n=IP){let e=new Eh;for(let t of n)t(e);return e}var IP,ne,Ct=M(()=>{"use strict";Bn();VN();jN();qN();KN();YN();A0();XN();tA();rA();oA();aA();lA();fA();hA();IP=[zN,GN,UN,WN,HN,ZN,QN,eA,nA,sA,iA,uA,pA,dA];l(EP,"buildRegistry");ne=EP()});function Qo(n,e,t="ignore"){let r=[],s=new Map;for(let o of n){if(s.get(o)!==void 0)continue;s.set(o,1);let i=[{op:o,i:0}];for(;i.length>0;){let a=i[i.length-1],u=a.op;if(a.i<u.numOperands){let p=u.getOperand(a.i).definingOp;if(a.i++,p&&e(p)){let c=s.get(p);if(c===void 0)s.set(p,1),i.push({op:p,i:0});else if(c===1){if(t==="throw")throw new Error("topo sort: cycle detected");if(t==="null")return null}}continue}s.set(u,2),r.push(u),i.pop()}}return r}function um(n,e="throw"){let t=Array.isArray(n)?n:[...n],r=new Set(t);return Qo(t,s=>r.has(s),e)}var Da=M(()=>{"use strict";l(Qo,"topoSortByOperands");l(um,"topoSortOpSet")});function mA(n){return Array.isArray(n)?n.map(mA):n}function gA(n){if(typeof n=="number")return n*2654435761&2147483647;if(typeof n=="string"){let e=0;for(let t=0;t<n.length;t++)e=(e<<5)-e+n.charCodeAt(t)&2147483647;return e}if(typeof n=="boolean")return n?1:0;if(Array.isArray(n)){let e=2654435769;for(let t=0;t<n.length;t++)e=(e^gA(n[t]))*16777619&2147483647;return e}if(ArrayBuffer.isView(n)&&n.buffer instanceof ArrayBuffer){let e=new Uint8Array(n.buffer,n.byteOffset,n.byteLength),t=2654435769;t=(t^e.length)*16777619&2147483647;let r=e.length>256?Math.ceil(e.length/256):1;for(let s=0;s<e.length;s+=r)t=(t^e[s])*16777619&2147483647;return t}return typeof n=="object"&&n!==null&&typeof n.hash=="function"?n.hash():0}function yA(n,e){if(n===e)return!0;if(typeof n!=typeof e)return!1;if(Array.isArray(n)){if(!Array.isArray(e)||n.length!==e.length)return!1;for(let t=0;t<n.length;t++)if(!yA(n[t],e[t]))return!1;return!0}return typeof n=="object"&&n!==null&&typeof n.equals=="function"?n.equals(e):!1}function CP(n,e=new Map){let t=new Ut;for(let r of n.blocks){let s=r.arguments.map(p=>p.type),o=new r.constructor(s);for(let p=0;p<r.arguments.length;p++)e.set(r.arguments[p],o.arguments[p]);let i=r.opsArray(),a=new Set(i),u=new Map;for(let p of Qo(i,c=>a.has(c),"ignore"))u.set(p,p.clone(e));for(let p of i)o.pushOp(u.get(p));t.addBlock(o)}return t}var LP,Me,yr=M(()=>{"use strict";Ih();Ea();Ct();Da();LP=0,Me=class n{static{l(this,"Operation")}constructor(e,t,r,s=null,o=null){if(this.opName=e,this.id=LP++,this.parentBlock=null,this._prev=null,this._next=null,this.attributes=new Map,s)if(s instanceof Map)for(let[i,a]of s)this.attributes.set(i,a);else for(let i of Object.keys(s))this.attributes.set(i,s[i]);this.operands=new Array(t.length),this._operandLinks=new Array(t.length);for(let i=0;i<t.length;i++){this.operands[i]=t[i];let a=new up(this,i);t[i].addUse(a),this._operandLinks[i]=a}this.results=new Array(r.length);for(let i=0;i<r.length;i++)this.results[i]=new Li(r[i],this,i);if(this.regions=[],o)for(let i=0;i<o.length;i++){let a=o[i]instanceof Ut?o[i]:new Ut;a.parentOp=this,this.regions.push(a)}}get numOperands(){return this.operands.length}get numResults(){return this.results.length}get numRegions(){return this.regions.length}getOperand(e){return this.operands[e]}getResult(e){return this.results[e]}getRegion(e){return this.regions[e]}getAttr(e){return this.attributes.get(e)}hasAttr(e){return this.attributes.has(e)}setAttr(e,t){this.attributes.set(e,t)}removeAttr(e){return this.attributes.delete(e)}replaceOperand(e,t){if(e<0||e>=this.operands.length)throw new Error(`replaceOperand: index ${e} out of range for '${this.opName}' (${this.operands.length} operands)`);if(!(t instanceof Li))throw new Error(`replaceOperand: new operand for '${this.opName}' is not a Value`);let r=this.operands[e];if(r===t)return;r.removeUse(this._operandLinks[e]),this.operands[e]=t;let s=new up(this,e);t.addUse(s),this._operandLinks[e]=s,this.parentBlock&&this.parentBlock._notifyMutation()}dropAllOperands(){let e=this.operands.length>0;for(let t=0;t<this.operands.length;t++)this.operands[t].removeUse(this._operandLinks[t]);this.operands=[],this._operandLinks=[],e&&this.parentBlock&&this.parentBlock._notifyMutation()}erase(){for(let e=0;e<this.results.length;e++)if(this.results[e].hasUses)throw new Error(`Cannot erase ${this.opName}: result ${e} still has uses`);this.dropAllOperands(),this.parentBlock&&this.parentBlock.removeOp(this)}replaceAllResultsWith(e){if(!e||e.length!==this.results.length)throw new Error(`replaceAllResultsWith: '${this.opName}' has ${this.results.length} results, got ${e?e.length:0}`);for(let t=0;t<this.results.length;t++)this.results[t].replaceAllUsesWith(e[t])}isTerminator(){let e=ne.get(this.opName);return!!(e&&e.isTerminator)}hasSideEffects(){let e=ne.get(this.opName);return!!(e&&e.hasSideEffects)}getParentFunction(){return this.parentBlock?this.parentBlock._owningFunction():null}clone(e=new Map){let t=this.operands.map(i=>e.get(i)||i),r=this.regions.map(i=>CP(i,e)),s=new Map;for(let[i,a]of this.attributes)s.set(i,mA(a));let o=new n(this.opName,t,this.results.map(i=>i.type),s,r);for(let i=0;i<this.results.length;i++)e.set(this.results[i],o.results[i]);return o}structuralHash(){let e=2166136261;for(let t=0;t<this.opName.length;t++)e=(e^this.opName.charCodeAt(t))*16777619&2147483647;for(let t=0;t<this.operands.length;t++)e=(e^this.operands[t].id)*16777619&2147483647;for(let[t,r]of this.attributes){for(let s=0;s<t.length;s++)e=(e^t.charCodeAt(s))*16777619&2147483647;e=(e^gA(r))*16777619&2147483647}return this.regions.length>0&&(e=(e^2654435769+this.regions.length)*16777619&2147483647),e}structuralEquals(e){if(this.opName!==e.opName||this.operands.length!==e.operands.length||this.results.length!==e.results.length||this.attributes.size!==e.attributes.size||this.regions.length>0||e.regions.length>0)return!1;for(let t=0;t<this.operands.length;t++)if(this.operands[t]!==e.operands[t])return!1;for(let[t,r]of this.attributes)if(!e.attributes.has(t)||!yA(r,e.attributes.get(t)))return!1;for(let t=0;t<this.results.length;t++)if(!this.results[t].type.equals(e.results[t].type))return!1;return!0}};l(mA,"cloneAttrValue");l(gA,"hashAttrValue");l(yA,"attrValueEquals");l(CP,"cloneRegion")});function*_A(n){if(!(!n.regions||n.regions.length===0))for(let e of n.regions)for(let t of e.blocks)for(let r of t.ops())yield r,yield*_A(r)}function*bA(n){if(!(!n.regions||n.regions.length===0))for(let e of n.regions)for(let t of e.blocks){yield t;for(let r of t.ops())yield*bA(r)}}function DP(n){let e=n.opsArray(),t=new Set(e);return Qo(e,r=>t.has(r),"ignore")}function lm(n){let e=new yn(n.name,n.inputTypes,n.outputTypes),t=new Map,r=n.entryBlock,s=e.entryBlock;for(let i=0;i<r.arguments.length;i++)t.set(r.arguments[i],s.arguments[i]);let o=new Map;for(let i of DP(r))o.set(i,i.clone(t));for(let i of r)s.pushOp(o.get(i));return e._version=n._version,e}var yn,Us=M(()=>{"use strict";Ea();Da();l(_A,"opsInRegions");l(bA,"blocksInRegions");yn=class{static{l(this,"GraphFunction")}constructor(e,t,r){this.name=e,this.inputTypes=Object.freeze([...t]),this.outputTypes=Object.freeze([...r]),this.body=new Ut;let s=new On(t);s._parentFunction=this,this.body.addBlock(s),this._version=0}get entryBlock(){return this.body.entryBlock}get args(){return this.entryBlock.arguments}bumpVersion(){this._version++}get version(){return this._version}*ops(){for(let e of this.body)yield*e}*opsRecursive(){for(let e of this.ops())yield e,yield*_A(e)}*blocksRecursive(){for(let e of this.body){yield e;for(let t of e)yield*bA(t)}}opsArray(){let e=[];for(let t of this.body)for(let r of t)e.push(r);return e}numOps(){let e=0;for(let t of this.body)e+=t.size;return e}findOp(e){for(let t of this.ops())if(e(t))return t;return null}findOps(e){let t=[];for(let r of this.ops())e(r)&&t.push(r);return t}getReturnOp(){let e=this.entryBlock.lastOp;return e&&e.opName==="return"?e:null}getReturnValues(){let e=this.getReturnOp();return e?[...e.operands]:[]}verify(){let e=[];if(!this.entryBlock)return e.push("Function has no entry block"),e;this.entryBlock.arguments.length!==this.inputTypes.length&&e.push(`Entry block has ${this.entryBlock.arguments.length} args but function expects ${this.inputTypes.length}`);let t=this.getReturnOp();return t?t.numOperands!==this.outputTypes.length&&e.push(`Return has ${t.numOperands} operands but function declares ${this.outputTypes.length} outputs`):e.push("Function body has no return op"),e}};l(DP,"topoOrderTopLevel");l(lm,"cloneGraphFunction")});function wA(n){let e=new Mn(n.name);for(let t of n)e.addFunction(lm(t));return e._version=n._version,e}var Mn,Di=M(()=>{"use strict";Us();Mn=class{static{l(this,"GraphModule")}constructor(e="module"){this.name=e,this._functions=new Map,this._version=0}get version(){return this._version}addFunction(e){return this._functions.set(e.name,e),e._module=this,this._version++,e}getFunction(e){return this._functions.get(e)||null}hasFunction(e){return this._functions.has(e)}removeFunction(e){let t=this._functions.delete(e);return t&&this._version++,t}get functionCount(){return this._functions.size}functionNames(){return[...this._functions.keys()]}*functions(){yield*this._functions.values()}*[Symbol.iterator](){yield*this._functions.values()}verify(){let e=[];this._functions.size===0&&e.push("Module has no functions");for(let t of this._functions.values()){let r=t.verify();for(let s=0;s<r.length;s++)e.push(`${t.name}: ${r[s]}`)}return e}};l(wA,"cloneGraphModule")});function TA(n){return n?n instanceof Ia?`tuple(${n.types.map(TA).join(", ")})`:n.shape!==void 0&&n.dtype!==void 0?`[${n.shape.join(",")}]:${n.dtype}`:String(n):"?"}function xA(n,e,t){let r=n.rank,s=e<0?r+e:e,o=n.shape.map((a,u)=>u===s?1:a),i=[];for(let a=0;a<r;a++)a!==s&&i.push(a<s?a:a-1+t);return{offsetDims:i,collapsedSliceDims:[s],startIndexMap:[s],indexVectorDim:t,sliceSizes:o}}function PP(n,e){let t=Math.max(n.length,e.length),r=new Array(t);for(let s=0;s<t;s++){let o=s<n.length?n[n.length-1-s]:1,i=s<e.length?e[e.length-1-s]:1;r[t-1-s]=o===1?i:o}return r}function Xo(n,e){let t=[];for(let r=0;r<n;r++)r!==e&&t.push(r);return t}function $P(n){if(!n)return new Map;if(n instanceof Map)return n;let e=new Map;for(let t of Object.keys(n))e.set(t,n[t]);return e}function mp(n,e,t,r){let s=new yn(n,e,t),o=new $n(s);return r(o,s.args),s}var $n,_s=M(()=>{"use strict";Re();yr();Ea();Us();Di();Ct();l(TA,"describeType");l(xA,"indexSelectGatherOpts");l(PP,"bcastBatchDims");$n=class n{static{l(this,"IRBuilder")}constructor(e){this.func=e,this.block=e.entryBlock,this._insertionPoint=null}setInsertionPoint(e){this._insertionPoint=e}setInsertionPointToEnd(){this._insertionPoint=null}_insert(e){return this._insertionPoint?this.block.insertBefore(e,this._insertionPoint):this.block.pushOp(e),e}_buildOp(e,t,r,s=null,o=null){let i=new Me(e,t,r,s,o);return this._insert(i)}_inferAndBuild(e,t,r=null,s=null,o=null){let i=ne.get(e),a=o;if(!a&&i&&i.inferResultTypes){let u=t.map(c=>c.type),p=r instanceof Map?r:$P(r);a=i.inferResultTypes(u,p,o)}if(!a){let u=t.map(c=>TA(c.type)).join(", "),p=i?i.inferResultTypes?"inferResultTypes returned no types":"op has no inferResultTypes and none were given":"op is not registered";throw new Error(`Cannot infer result types for op '${e}' (${p}); operands: [${u}]`)}return this._buildOp(e,t,a,r,s)}constant(e,t){return this._buildOp("constant",[],[t],{value:e,tensor_type:t})}scalarConstant(e,t="f32"){let r=new I([],t);return this.constant(e,r)}tensorConstant(e,t,r="f32"){let s=new I(t,r);return this.constant(e,s)}iota(e,t){return this._buildOp("iota",[],[t],{iota_dimension:e,tensor_type:t})}add(e,t){return this._inferAndBuild("add",[e,t])}sub(e,t){return this._inferAndBuild("sub",[e,t])}mul(e,t){return this._inferAndBuild("mul",[e,t])}div(e,t){return this._inferAndBuild("div",[e,t])}rem(e,t){return this._inferAndBuild("rem",[e,t])}pow(e,t){return this._inferAndBuild("pow",[e,t])}maximum(e,t){return this._inferAndBuild("maximum",[e,t])}minimum(e,t){return this._inferAndBuild("minimum",[e,t])}neg(e){return this._inferAndBuild("neg",[e])}stopGradient(e){return this._inferAndBuild("stop_gradient",[e])}reverse(e,t){return this._inferAndBuild("reverse",[e],{dimensions:t})}scaledDotProductAttention(e,t,r,s,o=!1){return this._inferAndBuild("scaled_dot_product_attention",[e,t,r],{scale:s,causal:o})}allReduce(e,t={}){return this._inferAndBuild("all_reduce",[e],{reduce_op:t.reduceOp||"sum",mesh_axis:t.meshAxis??0})}allGather(e,t={}){return this._inferAndBuild("all_gather",[e],{mesh_axis:t.meshAxis??0,gather_dim:t.gatherDim??1})}abs(e){return this._inferAndBuild("abs",[e])}exp(e){return this._inferAndBuild("exp",[e])}log(e){return this._inferAndBuild("log",[e])}sqrt(e){return this._inferAndBuild("sqrt",[e])}rsqrt(e){return this._inferAndBuild("rsqrt",[e])}tanh(e){return this._inferAndBuild("tanh",[e])}sin(e){return this._inferAndBuild("sin",[e])}cos(e){return this._inferAndBuild("cos",[e])}floor(e){return this._inferAndBuild("floor",[e])}ceil(e){return this._inferAndBuild("ceil",[e])}sign(e){return this._inferAndBuild("sign",[e])}erf(e){return this._inferAndBuild("erf",[e])}log2(e){return this._inferAndBuild("log2",[e])}log10(e){return this._inferAndBuild("log10",[e])}exp2(e){return this._inferAndBuild("exp2",[e])}square(e){return this._inferAndBuild("square",[e])}reciprocal(e){return this._inferAndBuild("reciprocal",[e])}logicalNot(e){return this._inferAndBuild("logical_not",[e])}logicalAnd(e,t){return this._inferAndBuild("logical_and",[e,t])}logicalOr(e,t){return this._inferAndBuild("logical_or",[e,t])}compare(e,t,r){return this._inferAndBuild("compare",[e,t],{direction:r})}select(e,t,r){return this._inferAndBuild("select",[e,t,r])}clamp(e,t,r){return this._inferAndBuild("clamp",[e,t,r])}broadcast(e,t,r){return this._inferAndBuild("broadcast_in_dim",[e],{result_shape:t,broadcast_dimensions:r})}reshape(e,t){return this._inferAndBuild("reshape",[e],{new_shape:t})}transpose(e,t){return this._inferAndBuild("transpose",[e],{permutation:t})}slice(e,t,r,s=null){let o={starts:t,limits:r};return s&&(o.strides=s),this._inferAndBuild("slice",[e],o)}concat(e,t){return this._inferAndBuild("concat",e,{dimension:t})}pad(e,t,r,s,o=null){let i={low:r,high:s};return o&&(i.interior=o),this._inferAndBuild("pad",[e,t],i)}reduce(e,t,r,s){let o=new I([],e.type.dtype),i=new Ut,a=new On([o,o]);return i.addBlock(a),this._inferAndBuild("reduce",[e,t],{dimensions:r,reduce_type:s},[i])}dot(e,t,r,s,o=[],i=[]){return this._inferAndBuild("dot",[e,t],{lhs_contracting:r,rhs_contracting:s,lhs_batch:o,rhs_batch:i})}matmul(e,t){let r=e.type.rank,s=t.type.rank;if(r===1&&s===1)return this.dot(e,t,[0],[0]);let o=e,i=t,a=!1,u=!1;r===1&&(o=this.reshape(e,[1,e.type.shape[0]]).getResult(0),a=!0),s===1&&(i=this.reshape(t,[t.type.shape[0],1]).getResult(0),u=!0);let p=o.type.rank,c=i.type.rank,f=i.type.shape.slice(0,c-2),d;if(f.length===0)d=this.dot(o,i,[p-1],[0]);else{let g=o.type.shape.slice(0,p-2),y=PP(g,f),_=y.length,b=this._broadcastBatch(o,g,y),w=this._broadcastBatch(i,f,y),x=Array.from({length:_},(S,T)=>T);d=this.dot(b,w,[_+1],[_],x,x)}if(!a&&!u)return d;let h=d.getResult(0).type.shape,m=new Set;return u&&m.add(h.length-1),a&&m.add(h.length-2),this.reshape(d.getResult(0),h.filter((g,y)=>!m.has(y)))}_broadcastBatch(e,t,r){let s=e.type.shape.slice(e.type.rank-2),o=[...r,...s];if(t.length===r.length&&t.every((u,p)=>u===r[p]))return e;let i=r.length-t.length,a=[];for(let u=0;u<t.length;u++)a.push(i+u);return a.push(o.length-2,o.length-1),this.broadcast(e,o,a).getResult(0)}conv(e,t,r,s,o={}){return this._inferAndBuild("conv",[e,t],{strides:r,padding:s,dilation:o.dilation||r.map(()=>1),groups:o.groups||1,input_layout:o.inputLayout||"NCHW",kernel_layout:o.kernelLayout||"OIHW"})}convert(e,t){return this._inferAndBuild("convert",[e],{target_dtype:t})}customCall(e,t,r,s=null){let o={call_target_name:e};return s&&(o.backend_config=s),this._buildOp("custom_call",t,r,o)}fusion(e,t,r,s){let o=new Ut,i=new On(e.map(u=>u.type));o.addBlock(i);let a=this._buildOp("fusion",e,t,{fusion_kind:r},[o]);if(s){let u=new n(this.func);u.block=i,s(u,i.arguments)}return a}ifOp(e,t,r,s){let o=new Ut,i=new On([]);o.addBlock(i);let a=new Ut,u=new On([]);a.addBlock(u);let p=this._buildOp("if",[e],t,null,[o,a]);if(r){let c=new n(this.func);c.block=i,r(c)}if(s){let c=new n(this.func);c.block=u,s(c)}return p}whileOp(e,t,r){let s=e.map(c=>c.type),o=new Ut,i=new On(s);o.addBlock(i);let a=new Ut,u=new On(s);a.addBlock(u);let p=this._buildOp("while",e,s,null,[o,a]);if(t){let c=new n(this.func);c.block=i,t(c,i.arguments)}if(r){let c=new n(this.func);c.block=u,r(c,u.arguments)}return p}scanOp(e,t,r){let s=e.map(g=>g.type.withShape(g.type.shape.slice(1))),o=t.map(g=>g.type),i=new Ut,a=new On([...s,...o]);i.addBlock(a);let u=new n(this.func);u.block=a;let p=a.arguments.slice(0,s.length),c=a.arguments.slice(s.length),[f,d]=r(u,p,c);if(u.yieldOp([...f,...d]),e.length===0)throw new Error("scanOp requires at least one xs input");let h=e[0].type.shape[0];if(typeof h!="number"||h<0)throw new Error(`scanOp requires a static, non-negative leading dim on xs, got ${h}`);for(let g=1;g<e.length;g++)if(e[g].type.shape[0]!==h)throw new Error("scanOp requires all xs inputs to share the same leading length");let m=d.map(g=>g.type.withShape([h,...g.type.shape]));return this._buildOp("scan",[...e,...t],[...o,...m],{num_carry:t.length,num_xs:e.length},[i])}returnOp(e){return this._buildOp("return",e,[])}yieldOp(e){return this._buildOp("yield",e,[])}relu(e){let t=this.scalarConstant(0,e.type.dtype),r=this.broadcast(t.getResult(0),e.type.shape,[]);return this.maximum(e,r.getResult(0))}softmax(e,t=-1){let r=t<0?e.type.rank+t:t;return this._inferAndBuild("softmax",[e],{axis:r})}logSoftmax(e,t=-1){let r=t<0?e.type.rank+t:t;return this._inferAndBuild("log_softmax",[e],{axis:r})}sigmoid(e){return this._inferAndBuild("sigmoid",[e])}gelu(e){return this._inferAndBuild("gelu",[e])}silu(e){return this._inferAndBuild("silu",[e])}elu(e,t=1){return this._inferAndBuild("elu",[e],{alpha:t})}leakyRelu(e,t=.01){return this._inferAndBuild("leaky_relu",[e],{negative_slope:t})}celu(e,t=1){return this._inferAndBuild("celu",[e],{alpha:t})}selu(e){return this._inferAndBuild("selu",[e])}mish(e){return this._inferAndBuild("mish",[e])}hardswish(e){return this._inferAndBuild("hardswish",[e])}hardsigmoid(e){return this._inferAndBuild("hardsigmoid",[e])}layernorm(e,t,r,s=-1,o=1e-5){let i=s<0?e.type.rank+s:s;return this._inferAndBuild("layer_norm",[e,t,r],{axis:i,epsilon:o})}batchnorm(e,t,r,s,o,i=1,a=1e-5){return this._inferAndBuild("batch_norm",[e,t,r,s,o],{axis:i,epsilon:a})}where(e,t,r){return this._inferAndBuild("where",[e,t,r])}split(e,t,r){return this._inferAndBuild("split",[e],{dimension:t,split_sizes:r})}oneHot(e,t,r={}){return this._inferAndBuild("one_hot",[e],{depth:t,axis:r.axis??-1,on_value:r.onValue??1,off_value:r.offValue??0,dtype:r.dtype||e.type.dtype})}embedding(e,t){return this._inferAndBuild("embedding",[e,t])}gather(e,t,r){return this._inferAndBuild("gather",[e,t],{offset_dims:r.offsetDims,collapsed_slice_dims:r.collapsedSliceDims,start_index_map:r.startIndexMap,index_vector_dim:r.indexVectorDim,slice_sizes:r.sliceSizes})}scatter(e,t,r,s){let o=new I([],e.type.dtype),i=new Ut;return i.addBlock(new On([o,o])),this._inferAndBuild("scatter",[e,t,r],{update_window_dims:s.updateWindowDims,inserted_window_dims:s.insertedWindowDims,scatter_dims_to_operand_dims:s.scatterDimsToOperandDims,index_vector_dim:s.indexVectorDim},[i])}scatterAdd(e,t,r,s){return this.scatter(e,t,r,s)}_dimCoordIndices(e,t,r){let s=r.length,o=e.type.dtype==="i32"?e:this.convert(e,"i32").getResult(0),i=[...r,1],a=[];for(let u=0;u<s;u++){let p=u===t?o:this.iota(u,new I(r,"i32")).getResult(0);a.push(this.reshape(p,i).getResult(0))}return this.concat(a,s).getResult(0)}gatherDim(e,t,r){let s=e.type.rank,o=r<0?s+r:r,i=[],a=[];for(let p=0;p<s;p++)i.push(p),a.push(1);let u=this._dimCoordIndices(t,o,t.type.shape);return this.gather(e,u,{offsetDims:[],collapsedSliceDims:i,startIndexMap:i,indexVectorDim:s,sliceSizes:a})}scatterAddDim(e,t,r,s){let o=e.type.rank,i=s<0?o+s:s,a=[];for(let p=0;p<o;p++)a.push(p);let u=this._dimCoordIndices(t,i,t.type.shape);return this.scatterAdd(e,u,r,{updateWindowDims:[],insertedWindowDims:a,scatterDimsToOperandDims:a,indexVectorDim:o})}argmax(e,t,r=!1){let s=t<0?e.type.rank+t:t;return this._inferAndBuild("argmax",[e],{axis:s,keep_dims:r})}argmin(e,t,r=!1){let s=t<0?e.type.rank+t:t;return this._inferAndBuild("argmin",[e],{axis:s,keep_dims:r})}pool2d(e,t,r,s,o,i={}){return this._inferAndBuild("pool2d",[e],{pool_type:t,kernel_size:r,strides:s,padding:o,ceil_mode:i.ceilMode||!1,count_include_pad:i.countIncludePad||!1,layout:i.layout||"NCHW"})}resize(e,t,r,s={}){return this._inferAndBuild("resize",[e],{output_size:t,method:r,coordinate_mode:s.coordinateMode||"asymmetric",layout:s.layout||"NCHW"})}};l(Xo,"broadcastDimsExcluding");l($P,"toMap");l(mp,"buildFunction")});function Un(n,e,t){if(t==null)return new P(n,e,t);let r=e instanceof L,s=t instanceof L;if(r&&s){let o=e.value,i=t.value;switch(n){case"+":return new L(o+i);case"-":return new L(o-i);case"*":return new L(o*i);case"//":if(i!==0)return new L(Math.trunc(o/i));break;case"%":if(i!==0)return new L((o%i+i)%i);break}}if(s){let o=t.value;if((n==="+"||n==="-")&&o===0||n==="*"&&o===1)return e;if(n==="*"&&o===0)return new L(0);if(n==="//"&&o===1)return e;if(n==="%"&&o===1)return new L(0)}if(r){let o=e.value;if(n==="+"&&o===0||n==="*"&&o===1)return t;if(n==="*"&&o===0)return new L(0)}return new P(n,e,t)}var z,Bt,Pi,he,fe,B0,Pa,J,q,Se,et,Ue,te,$i,il,Fn,al,Ne,P,Ke,ut,Oe,L,de,Fe=M(()=>{"use strict";z=Object.freeze({SERIAL:"serial",PARALLEL:"parallel",VECTORIZED:"vectorized",UNROLLED:"unrolled",THREAD_BINDING:"thread_binding",RECURRENCE:"recurrence"}),Bt=class n{static{l(this,"TensorNode")}constructor(){this.type=this.constructor.name.replace(/^_+/,""),this._parent=null,this._parentKey=null,this._parentIdx=-1}_setChild(e,t,r=-1){t instanceof n&&(t._parent=this,t._parentKey=e,t._parentIdx=r)}_setChildren(e,t){if(t)for(let r=0;r<t.length;r++)t[r]instanceof n&&(t[r]._parent=this,t[r]._parentKey=e,t[r]._parentIdx=r)}replaceWith(e){let t=this._parent;return t?(this._parentIdx>=0?t[this._parentKey][this._parentIdx]=e:t[this._parentKey]=e,e instanceof n&&(e._parent=t,e._parentKey=this._parentKey,e._parentIdx=this._parentIdx),this._parent=null,this._parentKey=null,this._parentIdx=-1,!0):!1}},Pi=class extends Bt{static{l(this,"PrimFunc")}constructor(e,t,r,s=new Map,o=[],i=null){super(),this.name=e,this.params=t,this.body=r,this.bufferMap=s,this.shapeParams=o,this.shapeParamMap=i||new Map,this._setChild("body",r)}},he=class extends Bt{static{l(this,"ForNode")}constructor(e,t,r,s,o,i=null){super(),this.loopVar=e,this.min=t,this.extent=r,this.kind=s,this.body=o,this.threadTag=i,this._setChild("body",o)}},fe=class extends Bt{static{l(this,"BlockNode")}constructor(e,t,r,s,o,i=null){super(),this.name=e,this.iterVars=t,this.reads=r,this.writes=s,this.body=o,this.initBody=i,this._setChild("body",o),this._setChild("initBody",i)}},B0=Object.freeze({DATA_PAR:"DataPar",COMM_REDUCE:"CommReduce"}),Pa=class extends Bt{static{l(this,"BlockRealizeNode")}constructor(e,t,r=B0.DATA_PAR){super(),this.iterVar=e,this.binding=t,this.kind=r}},J=class extends Bt{static{l(this,"BufferStoreNode")}constructor(e,t,r){super(),this.buffer=e,this.indices=t,this.value=r}},q=class extends Bt{static{l(this,"BufferLoadNode")}constructor(e,t){super(),this.buffer=e,this.indices=t}},Se=class extends Bt{static{l(this,"IfThenElseNode")}constructor(e,t,r=null){super(),this.condition=e,this.thenBody=t,this.elseBody=r,this._setChild("thenBody",t),this._setChild("elseBody",r)}},et=class extends Bt{static{l(this,"LetStmtNode")}constructor(e,t,r){super(),this.variable=e,this.value=t,this.body=r,this._setChild("body",r)}},Ue=class extends Bt{static{l(this,"AllocateNode")}constructor(e,t,r){super(),this.buffer=e,this.scope=t,this.body=r,this._setChild("body",r)}},te=class extends Bt{static{l(this,"SeqNode")}constructor(e){super(),this.stmts=e,this._setChildren("stmts",e)}},$i=class extends Bt{static{l(this,"WhileNode")}constructor(e,t,r){super(),this.condVar=e,this.condBody=t,this.loopBody=r,this._setChild("condBody",t),this._setChild("loopBody",r)}},il=class extends Bt{static{l(this,"EvaluateNode")}constructor(e){super(),this.value=e}},Fn=class extends Bt{static{l(this,"SyncThreadsNode")}constructor(){super()}},al=class extends Bt{static{l(this,"VecCopyNode")}constructor(e,t,r,s,o){super(),this.dstBuffer=e,this.dstIndex=t,this.srcBuffer=r,this.srcIndex=s,this.width=o}},Ne=class extends Bt{static{l(this,"CallExternNode")}constructor(e,t,r){super(),this.externName=e,this.args=t,this.dtype=r}},P=class extends Bt{static{l(this,"MathOpNode")}constructor(e,t,r=null){super(),this.op=e,this.a=t,this.b=r}},Ke=class extends Bt{static{l(this,"CompareNode")}constructor(e,t,r){super(),this.direction=e,this.a=t,this.b=r}},ut=class extends Bt{static{l(this,"CastNode")}constructor(e,t,r){super(),this.expr=e,this.fromDtype=t,this.toDtype=r}},Oe=class extends Bt{static{l(this,"VariableNode")}constructor(e,t){super(),this.name=e,this.dtype=t}},L=class extends Bt{static{l(this,"IntImmNode")}constructor(e){super(),this.value=e}},de=class extends Bt{static{l(this,"FloatImmNode")}constructor(e){super(),this.value=e}};l(Un,"mathOp")});var ul,R0=M(()=>{"use strict";ul=Object.freeze({GLOBAL:"global",SHARED:"shared",LOCAL:"local",REGISTER:"register"})});var me,ei=M(()=>{"use strict";tr();Re();me=class{static{l(this,"Buffer")}constructor(e,t,r,s,o=null,i=0,a=64){if(this.name=e,this.shape=t,this.dtype=r,this.scope=s,this.offset=i,this.alignment=a,this.broadcastDims=null,o)this.strides=o;else{this.strides=new Array(t.length);let u=1;for(let p=t.length-1;p>=0;p--)this.strides[p]=u,u!==X&&(typeof t[p]=="number"?u*=t[p]:u=X)}}get rank(){return this.shape.length}get isScalar(){return this.shape.length===0}numel(){return ap(this.shape,-1)}symbolicNumel(){return f0(this.shape)}sizeInBytes(){let e=this.numel();return e<0?-1:e*Rn(this.dtype)}}});function FP(n,e,t){switch(n){case"add":return new P("+",e,t);case"sub":return new P("-",e,t);case"mul":return new P("*",e,t);case"div":return new P("//",e,t);case"mod":return new P("%",e,t);case"neg":return new P("-",e);case"ceildiv":return new P("//",new P("-",new P("+",e,t),new L(1)),t);default:throw new Error(`symIntToNode: unsupported op '${n}' in extent/index context`)}}function ll(n,e){if(typeof n=="number")return new L(n);if(!(n instanceof Pe))return new L(n);if(n.type==="var")return e(n.name);let t=ll(n.args[0],e),r=n.args.length>1?ll(n.args[1],e):null;return FP(n.type,t,r)}var M0=M(()=>{"use strict";po();Fe();l(FP,"symOpToNode");l(ll,"symIntToNode")});function _n(n){let e=ne.get(n);return e!==null&&e.isConstant}function Xt(n){let e=ne.get(n);return e!==null&&e.isTerminator}function cl(n){let e=ne.get(n);return e!==null&&e.isBroadcast}var Ir=M(()=>{"use strict";Ct();l(_n,"isConstantOp");l(Xt,"isTerminatorOp");l(cl,"isBroadcastOp")});function fm(n,{name:e,compute:t,plevel:r=10,targetKind:s=null}){let o=I0.get(n);return o||(o=new gp(n),I0.set(n,o)),o.addImplementation(e,t,r,s),cm.clear(),o}function dm(n,e=null){let t=I0.get(n);if(!t)return null;let r=e?e.kind:null,s=`${n}|${r===null?"":r}`;if(cm.has(s))return cm.get(s);let o=new gp(n);for(let a of t.implementations)(a.targetKind===null||a.targetKind===r)&&o.implementations.push(a);let i=o.implementations.length>0?o:null;return cm.set(s,i),i}function E0(n,e=null){let t=dm(n,e);return t?t.best():null}var pm,gp,I0,cm,L0=M(()=>{"use strict";pm=class{static{l(this,"OpImplementation")}constructor(e,t,r=10,s=null){this.name=e,this.compute=t,this.plevel=r,this.targetKind=s}},gp=class{static{l(this,"OpStrategy")}constructor(e){this.opName=e,this.implementations=[]}addImplementation(e,t,r=10,s=null){let o=new pm(e,t,r,s),i=this.implementations.findIndex(a=>a.name===e);return i>=0?this.implementations[i]=o:this.implementations.push(o),this}get candidates(){return this.implementations}best(){let e=null;for(let t of this.implementations)(e===null||t.plevel>e.plevel)&&(e=t);return e}},I0=new Map,cm=new Map;l(fm,"registerOpStrategy");l(dm,"getOpStrategy");l(E0,"selectImplementation")});function D0(n,e){return dm(n,e)?!0:_n(n)}function Ie(n,e,t=zP){fm(n,{name:`${n}.generic`,compute:e,plevel:t,targetKind:null})}function vA(n,e,t,r=VP){fm(n,{name:`${n}.${e}`,compute:t,plevel:r,targetKind:e})}function Fi(n,e,t=null){if(t){let s=t.getLoweringRule(n);if(s)return s}let r=E0(n,e);return r?r.compute:void 0}function Rt(n,e,t){let r=e.length,s=n.allocVarArray("i",r),o=n.allocBindArray("v",s),i=new Array(r);for(let u=0;u<r;u++)i[u]=o[u].iterVar;let a=t?n.extentNodes(e,t):null;return{loopVars:s,loopBinds:o,indices:i,extentNodes:a}}function _p(n,e,t){let r=n;for(let s=e.length-1;s>=0;s--)r=new he(e[s],new L(0),t[s],z.SERIAL,r);return r}function GP(n,e,t){let r=n;for(let s=e.length-1;s>=0;s--)r=new he(e[s],new L(0),new L(t[s]),z.SERIAL,r);return r}function Mt(n,e,t,r){return r?_p(n,e,r):GP(n,e,t)}function bp(n,e,t,r,s,{prefix:o,initBlockName:i,accBlockName:a,initVal:u,accLeaf:p}){let c=jP(n,e,t,r),f=_r(n,o,Array.from({length:s.shape.length},(A,N)=>N),s.shape,s),d=new J(s,f.indices,u()),h=new fe(n.blockName(i),f.ivs,[],[{buffer:s}],d),m=f.wrap(h),g=new q(t,c.lhsIdx),y=new q(r,c.rhsIdx),_=e.getAttr("lhs_prologue_cast"),b=e.getAttr("rhs_prologue_cast");_&&(g=new ut(g,t.dtype,_)),b&&(y=new ut(y,r.dtype,b));let w=p(g,y),x=new P("+",new q(s,c.outIdx),w),S=new J(s,c.outIdx,x),T=new fe(n.blockName(a),c.allIvs,[{buffer:t},{buffer:r}],[{buffer:s}],S),v=c.wrapAccBody(T);return{geo:c,initBody:m,accBody:v}}function gm(n,e,t,r,s,{prefix:o,blockPrefix:i,initVal:a,guardFill:u,leafBuilder:p}){let c=e.getAttr("strides"),f=e.getAttr("padding"),d=e.getAttr("dilation")||c.map(()=>1),h=e.getAttr("groups")||1,m=mm(e.getAttr("input_layout")),g=mm(e.getAttr("kernel_layout")),y=c.length,_=t.shape[m.N],b=r.shape[g.O],w=r.shape[g.I],x=s.shape,S=_r(n,o+"i",Array.from({length:x.length},(ye,Xe)=>Xe),x,s),T=new J(s,S.indices,a()),v=new fe(n.blockName(i+"_init"),S.ivs,[],[{buffer:s}],T),A=S.wrap(v),N=n.allocVar(o+"n"),k=n.allocVar(o+"oc"),O=n.allocVar(o+"ic"),R=n.allocVarArray(o+"o",y),B=n.allocVarArray(o+"k",y),E=[N,k,...R,O,...B],F=n.allocBindArray(o+"v",E),C=F[0].iterVar,H=F[1].iterVar,G=F.slice(2,2+y),U=F[2+y].iterVar,D=F.slice(3+y),V=new Array(x.length);V[m.N]=C,V[m.C]=H;let ue=Object.keys(m).filter(ye=>ye!=="N"&&ye!=="C").sort();for(let ye=0;ye<y;ye++)V[m[ue[ye]]]=G[ye].iterVar;let ce=new Array(t.shape.length);ce[m.N]=C;let Le=Math.floor(b/h);h>1?ce[m.C]=new P("+",new P("*",new P("//",H,new L(Le)),new L(w)),U):ce[m.C]=U;let j=new Array(r.shape.length);j[g.O]=H,j[g.I]=U;let W=null;for(let ye=0;ye<y;ye++){let Xe=ue[ye],Nn=Xe.toUpperCase(),Qt=Un("+",Un("*",G[ye].iterVar,new L(c[ye])),Un("+",Un("*",D[ye].iterVar,new L(d[ye])),new L(-f[ye][0])));if(ce[m[Xe]]=Qt,j[g[Nn]]=D[ye].iterVar,f[ye][0]!==0||f[ye][1]!==0){let an=new Ke("ge",Qt,new L(0)),un=new Ke("lt",Qt,new L(t.shape[m[Xe]])),mn=new P("*",an,un);W=W?new P("*",W,mn):mn}}let ie=p(ce,j),je=W?new Se(W,ie,u()):ie,ke=new q(s,V),rt=new P("+",ke,je),zt=new J(s,V,rt),on=new fe(n.blockName(i+"_acc"),F,[{buffer:t},{buffer:r}],[{buffer:s}],zt),Zt=new Array(y);for(let ye=0;ye<y;ye++){let Xe=ue[ye].toUpperCase();Zt[ye]=r.shape[g[Xe]]}let it=on;for(let ye=y-1;ye>=0;ye--){let Xe=ue[ye].toUpperCase();it=new he(B[ye],new L(0),n.extentNode(Zt[ye],r,g[Xe]),z.SERIAL,it)}it=new he(O,new L(0),n.extentNode(w,r,g.I),z.SERIAL,it);for(let ye=y-1;ye>=0;ye--){let Xe=m[ue[ye]];it=new he(R[ye],new L(0),n.extentNode(x[Xe],s,Xe),z.SERIAL,it)}return it=new he(k,new L(0),n.extentNode(b,r,g.O),z.SERIAL,it),it=new he(N,new L(0),n.extentNode(_,t,m.N),z.SERIAL,it),new te([A,it])}function _r(n,e,t,r,s){let o=t.length,i=new Array(o),a=new Array(o),u=new Array(o),p=new Array(o);for(let c=0;c<o;c++)i[c]=n.allocVar(`${e}${t[c]}`),a[c]=new Pa(n.allocVar(`${e}v${t[c]}`),i[c]),u[c]=a[c].iterVar,p[c]=n.extentNode(r[t[c]],s,t[c]);return{vars:i,ivs:a,indices:u,extentNodes:p,wrap(c){return _p(c,i,p)}}}function ti(n,e,t){let r=n.shape.length;if(n.broadcastDims){let a=n.broadcastDims,u=new Array(r);for(let p=0;p<r;p++)u[p]=n.shape[p]===1?new L(0):t[a[p]];return u}let o=e.shape.length-r,i=new Array(r);for(let a=0;a<r;a++)i[a]=n.shape[a]===1?new L(0):t[o+a];return i}function pl(n){let e=new Array(n.length);for(let t=0;t<n.length;t++)e[t]={buffer:n[t]};return e}function yp(){let n=0;for(let r=0;r<arguments.length;r++)n+=arguments[r].length;let e=new Array(n),t=0;for(let r=0;r<arguments.length;r++){let s=arguments[r];for(let o=0;o<s.length;o++)e[t++]=s[o]}return e}function C0(n){let e=new Array(n.length);for(let t=0;t<n.length;t++)e[t]=n[t].iterVar;return e}function mo(n,e,t,r,s){let o=r[0],{loopVars:i,loopBinds:a,indices:u,extentNodes:p}=Rt(n,o.shape,o),c=new Array(t.length);for(let m=0;m<t.length;m++){let g=ti(t[m],o,u);c[m]=new q(t[m],g)}let f=s(e,c,o.dtype),d=new J(o,u,f),h=new fe(n.blockName(`${e.opName}_block`),a,pl(t),[{buffer:o}],d);return Mt(h,i,o.shape,p)}function fl(n,e){let t=e.getResult(0),r=e.getAttr("value"),s=t.type&&t.type.shape||[];if(typeof r=="number"&&s.length>0&&!n.bufferMap.has(t)){let g=t.type&&t.type.dtype||"f32",y=new Array(s.length).fill(1),_=new me(`buf_${n.varCounter++}`,y,g,ul.GLOBAL);_.broadcastDims=Array.from({length:s.length},(w,x)=>x),n.bufferMap.set(t,_);let b=mr(g)?new L(r):new de(r);return new J(_,y.map(()=>new L(0)),b)}let o=n.getOrAllocBuffer(t),i=mr(o.dtype),a=l(g=>i?new L(g):new de(g),"imm");if(r&&typeof r!="number"&&typeof r.length=="number"){if(o.shape.length===0)return new J(o,[],a(r[0]));let g=new Array(o.shape.length),y=1;for(let b=o.shape.length-1;b>=0;b--)g[b]=y,y*=o.shape[b];let _=[];for(let b=0;b<r.length;b++){let w=new Array(o.shape.length);for(let x=0;x<o.shape.length;x++)w[x]=new L(Math.floor(b/g[x])%o.shape[x]);_.push(new J(o,w,a(r[b])))}return new te(_)}let u=a(typeof r=="number"?r:0);if(o.shape.length===0)return new J(o,[],u);let{loopVars:p,loopBinds:c,indices:f,extentNodes:d}=Rt(n,o.shape,o),h=new J(o,f,u),m=new fe(n.blockName(`${e.opName}_block`),c,[],[{buffer:o}],h);return Mt(m,p,o.shape,d)}function mm(n){let e={};for(let t=0;t<n.length;t++)e[n[t]]=t;return e}function SA(n,e){if(!n.broadcastDims)return e;let t=n.broadcastDims,r=new Array(n.shape.length);for(let s=0;s<n.shape.length;s++)r[s]=n.shape[s]===1?new L(0):e[t[s]];return r}function jP(n,e,t,r){let s=e.getAttr("lhs_contracting")||[],o=e.getAttr("rhs_contracting")||[],i=e.getAttr("lhs_batch")||[],a=e.getAttr("rhs_batch")||[],u=e.getOperand(0).type.shape,p=e.getOperand(1).type.shape,c=new Set(s),f=new Set(i),d=new Set(o),h=new Set(a),m=[];for(let C=0;C<u.length;C++)!c.has(C)&&!f.has(C)&&m.push(C);let g=[];for(let C=0;C<p.length;C++)!d.has(C)&&!h.has(C)&&g.push(C);let y=n.allocVarArray("b",i.length),_=n.allocVarArray("ls",m.length),b=n.allocVarArray("rs",g.length),w=n.allocVarArray("c",s.length),x=n.allocBindArray("vb",y),S=n.allocBindArray("vls",_),T=n.allocBindArray("vrs",b),v=n.allocBindArray("vc",w),A=yp(C0(x),C0(S),C0(T)),N=new Array(u.length);for(let C=0;C<i.length;C++)N[i[C]]=x[C].iterVar;for(let C=0;C<m.length;C++)N[m[C]]=S[C].iterVar;for(let C=0;C<s.length;C++)N[s[C]]=v[C].iterVar;let k=new Array(p.length);for(let C=0;C<a.length;C++)k[a[C]]=x[C].iterVar;for(let C=0;C<g.length;C++)k[g[C]]=T[C].iterVar;for(let C=0;C<o.length;C++)k[o[C]]=v[C].iterVar;let O=SA(t,N),R=SA(r,k),B=yp(x,S,T,v),E=[{vars:y,dims:i,shape:u,buf:t},{vars:_,dims:m,shape:u,buf:t},{vars:b,dims:g,shape:p,buf:r},{vars:w,dims:s,shape:u,buf:t}];function F(C){let H=C;for(let G=E.length-1;G>=0;G--){let{vars:U,dims:D,shape:V,buf:ue}=E[G];for(let ce=U.length-1;ce>=0;ce--){let Le=ue.broadcastDims?new L(V[D[ce]]):n.extentNode(V[D[ce]],ue,D[ce]);H=new he(U[ce],new L(0),Le,z.SERIAL,H)}}return H}return l(F,"wrapAccBody"),{outIdx:A,lhsIdx:O,rhsIdx:R,allIvs:B,wrapAccBody:F}}var zP,VP,hm,Qr=M(()=>{"use strict";Re();po();R0();ei();tr();Fe();M0();Ir();L0();L0();zP=10,VP=20;l(D0,"hasLoweringRule");l(Ie,"registerLoweringRule");l(vA,"registerTargetLoweringRule");l(Fi,"getLoweringRule");hm=class{static{l(this,"LoweringContext")}constructor(){this.bufferMap=new Map,this.varCounter=0,this.shapeParams=new Map,this.symbolToVar=new Map,this.symVars=new Map,this._blockCounter=0}blockName(e){return`${e}_${this._blockCounter++}`}allocVar(e,t="int32"){return new Oe(`${e}_${this.varCounter++}`,t)}getOrAllocBuffer(e){let t=this.bufferMap.get(e);if(t)return t;let r=e.type,s=r.shape||[],o=r.dtype||"f32",i=r.layout?r.layout.computeStrides(s):null;return t=new me(`buf_${this.varCounter++}`,s,o,ul.GLOBAL,i),e.symbolicShape&&(t.symbolicShape=e.symbolicShape),this.bufferMap.set(e,t),this._registerDynamicDims(t),t}allocFreshBuffer(e){let t=e.type,r=t.shape||[],s=t.dtype||"f32",o=t.layout?t.layout.computeStrides(r):null,i=new me(`buf_${this.varCounter++}`,r,s,ul.GLOBAL,o);return e.symbolicShape&&(i.symbolicShape=e.symbolicShape),this._registerDynamicDims(i),i}_registerDynamicDims(e){for(let t=0;t<e.shape.length;t++){let r=e.shape[t];r===X?this.extentNode(X,e,t):r instanceof Pe&&this._registerSymIntDim(e,t,r)}}_symVarNode(e){let t=this.symVars.get(e);return t||(t=new Oe(qu(e),"int32"),this.symVars.set(e,t)),t}_registerSymIntDim(e,t,r){for(let s of Pe.freeVars(r))this._symVarNode(s);if(r.type==="var"){let s=`${e.name}:${t}`;this.shapeParams.has(s)||this.shapeParams.set(s,this._symVarNode(r.name))}}symIntToExtentNode(e){return ll(e,t=>this._symVarNode(t))}_shapeParamVar(e,t){let r=t>=0?`${e.name}:${t}`:`${e.name}:dyn`,s=this.shapeParams.get(r);if(s)return s;let o=e.symbolicShape&&t>=0&&typeof e.symbolicShape[t]!="number"?e.symbolicShape[t]:null;return o!==null&&this.symbolToVar.has(o)?s=this.symbolToVar.get(o):(s=this.allocVar("_ds"),o!==null&&this.symbolToVar.set(o,s)),this.shapeParams.set(r,s),s}extentNode(e,t,r=-1){return e instanceof Pe?this.symIntToExtentNode(e):e!==X?new L(e):this._shapeParamVar(t,r)}extentNodes(e,t){let r=new Array(e.length);for(let s=0;s<e.length;s++)r[s]=this.extentNode(e[s],t,s);return r}allocVarArray(e,t){let r=new Array(t);for(let s=0;s<t;s++)r[s]=this.allocVar(`${e}${s}`);return r}allocBindArray(e,t){let r=new Array(t.length);for(let s=0;s<t.length;s++)r[s]=new Pa(this.allocVar(`${e}${s}`),t[s]);return r}};l(Rt,"makeLoopNest");l(_p,"wrapLoopsWithNodes");l(GP,"wrapLoops");l(Mt,"wrapInLoops");l(bp,"emitMatmulInitAcc");l(gm,"buildConvNest");l(_r,"buildSpatialNest");l(ti,"computeBroadcastIndices");l(pl,"bufRefs");l(yp,"concatIterVars");l(C0,"extractIterVars");l(mo,"lowerPointwise");l(fl,"lowerConstant");l(mm,"parseLayout");l(SA,"physicalDotIndices");l(jP,"buildDotGeometry")});function HP(n){let e=ne.get(n);return e?e.getAttr(P0):null}function wp(){let n=[];for(let e of ne.allOps())e.hasAttr(P0)&&n.push(e.name);return n}function $0(n,e,t){let r=HP(n);return r?kA[n]?kA[n](e):e.length===2&&UP.has(r)?new P(r,e[0],e[1]):e.length===2&&qP.has(r)?new P(r,e[0],e[1]):e.length===1&&WP.has(r)?new P(r,e[0]):e.length===1&&r==="-"?new P("-",e[0]):new Ne(r,e,t):null}function NA(){for(let n of wp())Ie(n,(e,t,r,s)=>mo(e,t,r,s,(o,i,a)=>$0(o.opName,i,a)));Ie("compare",(n,e,t,r)=>mo(n,e,t,r,(s,o)=>new Ke(s.getAttr("direction")||"eq",o[0],o[1]))),Ie("select",(n,e,t,r)=>mo(n,e,t,r,(s,o)=>new Se(o[0],o[1],o[2]))),Ie("clamp",(n,e,t,r)=>mo(n,e,t,r,(s,o,i)=>new Ne("min",[new Ne("max",[o[1],o[0]],i),o[2]],i))),Ie("convert",(n,e,t,r)=>mo(n,e,t,r,(s,o)=>new ut(o[0],t[0].dtype,r[0].dtype))),Ie("copy_to_device",(n,e,t,r)=>mo(n,e,t,r,(s,o)=>o[0]))}var UP,qP,WP,kA,P0,KP,ym=M(()=>{"use strict";Fe();Ct();Qr();UP=new Set(["+","-","*","/"]),qP=new Set(["&&","||"]),WP=new Set(["!"]),kA={square:l(n=>new P("*",n[0],n[0]),"square"),reciprocal:l(n=>new P("/",new de(1),n[0]),"reciprocal")},P0="elementwiseScalarOp",KP={add:"+",sub:"-",mul:"*",div:"/",max:"max",min:"min",exp:"exp",log:"log",sqrt:"sqrt",rsqrt:"rsqrt",tanh:"tanh",abs:"abs",ceil:"ceil",floor:"floor",neg:"-",maximum:"max",minimum:"min",sin:"sin",cos:"cos",round:"round",sign:"sign",pow:"pow",rem:"fmod",erf:"erf",erfc:"erfc",lgamma:"lgamma",gamma:"gamma",log2:"log2",log10:"log10",exp2:"exp2",square:"square",reciprocal:"reciprocal",logical_not:"!",logical_and:"&&",logical_or:"||"};for(let[n,e]of Object.entries(KP))ne.has(n)&&ne.registerOpAttr(n,P0,e);l(HP,"elementwiseScalarOp");l(wp,"elementwiseOpNames");l($0,"buildElementwiseExpr");l(NA,"register")});function OA(){Ie("broadcast_in_dim",AA),Ie("broadcast",AA),Ie("transpose",(n,e,t,r)=>{let s=e.getAttr("permutation"),o=t[0],i=r[0],{loopVars:a,loopBinds:u,indices:p,extentNodes:c}=Rt(n,i.shape,i),f=new Array(o.shape.length);for(let m=0;m<s.length;m++)f[s[m]]=p[m];let d=new J(i,p,new q(o,f)),h=new fe(n.blockName("transpose_block"),u,[{buffer:o}],[{buffer:i}],d);return Mt(h,a,i.shape,c)}),Ie("reverse",(n,e,t,r)=>{let s=new Set(e.getAttr("dimensions")),o=t[0],i=r[0],{loopVars:a,loopBinds:u,indices:p,extentNodes:c}=Rt(n,i.shape,i),f=new Array(o.shape.length);for(let m=0;m<o.shape.length;m++)f[m]=s.has(m)?new P("-",new P("-",c[m],new L(1)),p[m]):p[m];let d=new J(i,p,new q(o,f)),h=new fe(n.blockName("reverse_block"),u,[{buffer:o}],[{buffer:i}],d);return Mt(h,a,i.shape,c)}),Ie("reshape",(n,e,t,r)=>{let s=t[0],o=r[0],{loopVars:i,loopBinds:a,indices:u,extentNodes:p}=Rt(n,o.shape,o),c;if(s.shape.length===o.shape.length&&s.shape.every((m,g)=>m===o.shape[g]))c=u;else{for(let _=1;_<s.shape.length;_++)if(typeof s.shape[_]!="number"||s.shape[_]<0)throw new Error("reshape lowering requires static non-leading input dims");for(let _=1;_<o.shape.length;_++)if(typeof o.shape[_]!="number"||o.shape[_]<0)throw new Error("reshape lowering requires static non-leading output dims");let m=u[o.shape.length-1],g=1;for(let _=o.shape.length-2;_>=0;_--)g*=o.shape[_+1],m=Un("+",m,Un("*",u[_],new L(g)));c=new Array(s.shape.length);let y=m;for(let _=s.shape.length-1;_>=0;_--)_===0?c[_]=y:(c[_]=Un("%",y,new L(s.shape[_])),y=Un("//",y,new L(s.shape[_])))}let d=new J(o,u,new q(s,c)),h=new fe(n.blockName("reshape_block"),a,[{buffer:s}],[{buffer:o}],d);return Mt(h,i,o.shape,p)}),Ie("slice",(n,e,t,r)=>{let s=t[0],o=r[0],i=e.getAttr("starts"),a=e.getAttr("strides")||i.map(()=>1),{loopVars:u,loopBinds:p,indices:c,extentNodes:f}=Rt(n,o.shape,o),d=new Array(s.shape.length);for(let g=0;g<s.shape.length;g++){let y=new L(i[g]);a[g]===1?d[g]=new P("+",y,c[g]):d[g]=new P("+",y,new P("*",c[g],new L(a[g])))}let h=new J(o,c,new q(s,d)),m=new fe(n.blockName("slice_block"),p,[{buffer:s}],[{buffer:o}],h);return Mt(m,u,o.shape,f)}),Ie("pad",(n,e,t,r)=>{let s=t[0],o=t[1],i=r[0],a=e.getAttr("low"),u=e.getAttr("interior")||a.map(()=>0),{loopVars:p,loopBinds:c,indices:f,extentNodes:d}=Rt(n,i.shape,i),h=new Array(s.shape.length),m=new L(1);for(let x=0;x<s.shape.length;x++){let S=new P("+",f[x],new L(-a[x]));if(u[x]>0){let A=u[x]+1,N=new P("%",S,new L(A)),k=new Ke("eq",N,new L(0));m=new P("*",m,k),h[x]=new P("//",S,new L(A))}else h[x]=S;let T=new Ke("ge",h[x],new L(0)),v=new Ke("lt",h[x],new L(s.shape[x]));m=new P("*",m,new P("*",T,v))}let g=new q(s,h),y=new q(o,[]),_=new Se(m,g,y),b=new J(i,f,_),w=new fe(n.blockName("pad_block"),c,[{buffer:s},{buffer:o}],[{buffer:i}],b);return Mt(w,p,i.shape,d)}),Ie("concat",(n,e,t,r)=>{let s=r[0],o=e.getAttr("dimension"),i=[],a=0;for(let u=0;u<t.length;u++){let p=t[u],{loopVars:c,loopBinds:f,indices:d,extentNodes:h}=Rt(n,p.shape,p),m=new Array(p.shape.length);for(let _=0;_<p.shape.length;_++)m[_]=_===o&&a>0?new P("+",d[_],new L(a)):d[_];let g=new J(s,m,new q(p,d)),y=new fe(n.blockName("concat"),f,[{buffer:p}],[{buffer:s}],g);i.push(Mt(y,c,p.shape,h)),a+=p.shape[o]}return i.length===1?i[0]:new te(i)}),Ie("iota",(n,e,t,r)=>{let s=r[0],o=e.getAttr("iota_dimension"),{loopVars:i,loopBinds:a,indices:u,extentNodes:p}=Rt(n,s.shape,s),c=new ut(u[o],"index",s.dtype),f=new J(s,u,c),d=new fe(n.blockName("iota_block"),a,[],[{buffer:s}],f);return Mt(d,i,s.shape,p)})}function AA(n,e,t,r){let s=t[0],o=r[0],i=e.getAttr("broadcast_dimensions")||[],{loopVars:a,loopBinds:u,indices:p,extentNodes:c}=Rt(n,o.shape,o),f=new Array(s.shape.length);for(let m=0;m<s.shape.length;m++){let g=i.length>0?i[m]:m+(o.shape.length-s.shape.length);f[m]=s.shape[m]===1?new L(0):p[g]}let d=new J(o,p,new q(s,f)),h=new fe(n.blockName("broadcast_block"),u,[{buffer:s}],[{buffer:o}],d);return Mt(h,a,o.shape,c)}var BA=M(()=>{"use strict";Fe();Qr();l(OA,"register");l(AA,"lowerBroadcast")});function ZP(n,e){if(mr(n)){let t=e?YP[n]??-2147483648:JP[n]??2147483647;return new L(t)}return new de(e?-1/0:1/0)}function RA(){Ie("reduce",(e,t,r,s)=>{let o=r[0],i=r[1],a=s[0],u=t.getAttr("dimensions")||[],p=t.getAttr("reduce_type")||"sum",c=new Set(u),f=[],d=[];for(let B=0;B<o.shape.length;B++)(c.has(B)?d:f).push(B);let h=_r(e,"si",f,o.shape,o),m=new J(a,h.indices,new q(i,[])),g=new fe(e.blockName("reduce_init"),h.ivs,[{buffer:i}],[{buffer:a}],m),y=f.length>0?h.wrap(g):g,_=_r(e,"sa",f,o.shape,o),b=e.allocVarArray("r",d.length),w=e.allocBindArray("rv",b);for(let B of w)B.kind=B0.COMM_REDUCE;let x=new Array(o.shape.length);for(let B=0;B<f.length;B++)x[f[B]]=_.ivs[B].iterVar;for(let B=0;B<d.length;B++)x[d[B]]=w[B].iterVar;let S=new q(a,_.indices),T=new q(o,x),v=QP[p];if(!v)throw new Error(`reduction lowering: unsupported reduce_type '${p}'`);let A=new J(a,_.indices,v(S,T,a.dtype)),N=new Array(d.length);for(let B=0;B<d.length;B++)N[B]=e.extentNode(o.shape[d[B]],o,d[B]);let k=new fe(e.blockName("reduce_acc"),yp(_.ivs,w),[{buffer:o}],[{buffer:a}],A),O=_p(k,b,N);O=_.wrap(O);let R=[y,O];if(p==="mean"){let B=1,E=[];for(let D=0;D<d.length;D++){let V=o.shape[d[D]];V===X?E.push(e.extentNode(X,o,d[D])):B*=V}let F=_r(e,"sm",f,o.shape,o),C=new q(a,F.indices),H;if(E.length===0)H=new P("*",C,new de(1/B));else{let D=new L(B);for(let V of E)D=new P("*",D,V);H=new P("/",C,D)}let G=new J(a,F.indices,H),U=new fe(e.blockName("mean_div"),F.ivs,[{buffer:a}],[{buffer:a}],G);R.push(f.length>0?F.wrap(U):U)}return new te(R)});function n(e,t){Ie(e,(r,s,o,i)=>{let a=o[0],u=i[0],p=s.getAttr("axis"),c=s.getAttr("keep_dims")||!1,f=new Set([p]),d=[],h=p;for(let V=0;V<a.shape.length;V++)f.has(V)||d.push(V);let m=l(V=>{if(!c)return V.indices;let ue=new Array(a.shape.length);for(let ce=0;ce<d.length;ce++)ue[d[ce]]=V.indices[ce];return ue[h]=new L(0),ue},"outIndicesFor"),g=new me("_argval_"+r.varCounter,d.map(V=>a.shape[V]),a.dtype,"global");r.varCounter++;let y=_r(r,"ai",d,a.shape,a),_=new J(g,y.indices,ZP(a.dtype,t==="gt")),b=new J(u,m(y),new L(0)),w=new fe(r.blockName("arg_init"),y.ivs,[],[{buffer:g},{buffer:u}],new te([_,b])),x=d.length>0?y.wrap(w):w,S=_r(r,"as",d,a.shape,a),T=r.allocVar("ar"),v=r.allocBindArray("arv",[T]),A=new Array(a.shape.length);for(let V=0;V<d.length;V++)A[d[V]]=S.ivs[V].iterVar;A[h]=v[0].iterVar;let N=new q(a,A),k=new q(g,S.indices),O=new Ke(t,N,k),R=new Se(O,N,k),B=m(S),E=new q(u,B),F=new Se(O,v[0].iterVar,E),C=new J(u,B,F),H=new J(g,S.indices,R),G=new fe(r.blockName("arg_acc"),yp(S.ivs,v),[{buffer:a},{buffer:g}],[{buffer:g},{buffer:u}],new te([C,H])),U=r.extentNode(a.shape[h],a,h),D=_p(G,[T],[U]);return D=S.wrap(D),new te([x,D])})}l(n,"registerArgReduce"),n("argmax","gt"),n("argmin","lt")}var YP,JP,QP,MA=M(()=>{"use strict";Fe();ei();Re();Qr();tr();YP={i8:-128,i16:-32768,i32:-2147483648,i64:-2147483648,ui8:0,ui16:0,ui32:0,bool:0},JP={i8:127,i16:32767,i32:2147483647,i64:2147483647,ui8:255,ui16:65535,ui32:4294967295,bool:1};l(ZP,"argReduceSentinel");QP={sum:l((n,e)=>new P("+",n,e),"sum"),mean:l((n,e)=>new P("+",n,e),"mean"),prod:l((n,e)=>new P("*",n,e),"prod"),max:l((n,e,t)=>new Ne("max",[n,e],t),"max"),min:l((n,e,t)=>new Ne("min",[n,e],t),"min"),and:l((n,e)=>new P("*",n,e),"and"),or:l((n,e,t)=>new Ne("max",[n,e],t),"or")};l(RA,"register")});function IA(n,e){return jt(e)?new ut(n,e,"i32"):n}function EA(){Ie("dot",(n,e,t,r)=>{let{initBody:s,accBody:o}=bp(n,e,t[0],t[1],r[0],{prefix:"di",initBlockName:"matmul_init",accBlockName:"matmul",initVal:l(()=>new de(0),"initVal"),accLeaf:l((i,a)=>new P("*",i,a),"accLeaf")});return new te([s,o])}),Ie("conv",(n,e,t,r)=>{let s=t[0],o=t[1];return gm(n,e,s,o,r[0],{prefix:"c",blockPrefix:"conv",initVal:l(()=>new de(0),"initVal"),guardFill:l(()=>new de(0),"guardFill"),leafBuilder:l((i,a)=>new P("*",new q(s,i),new q(o,a)),"leafBuilder")})}),Ie("gather",(n,e,t,r)=>{let s=t[0],o=t[1],i=r[0],a=new Set(e.getAttr("offset_dims")),u=new Set(e.getAttr("collapsed_slice_dims")),p=e.getAttr("start_index_map"),c=e.getAttr("index_vector_dim"),{loopVars:f,loopBinds:d,indices:h,extentNodes:m}=Rt(n,i.shape,i),g=[],y=[];for(let S=0;S<i.shape.length;S++)a.has(S)?y.push(h[S]):g.push(h[S]);let _=new Array(s.shape.length),b=0;for(let S=0;S<s.shape.length;S++)u.has(S)?_[S]=new L(0):_[S]=y[b++];for(let S=0;S<p.length;S++){let T=new Array(o.shape.length),v=0;for(let k=0;k<o.shape.length;k++)k===c?T[k]=new L(S):T[k]=g[v++];let A=IA(new q(o,T),o.dtype),N=p[S];_[N]=new P("+",_[N],A)}let w=new J(i,h,new q(s,_)),x=new fe(n.blockName("gather_block"),d,[{buffer:s},{buffer:o}],[{buffer:i}],w);return Mt(x,f,i.shape,m)}),Ie("scatter",(n,e,t,r)=>{let s=t[0],o=t[1],i=t[2],a=r[0],u=new Set(e.getAttr("inserted_window_dims")),p=e.getAttr("scatter_dims_to_operand_dims"),c=e.getAttr("index_vector_dim"),f=new Set(e.getAttr("update_window_dims")),d=Rt(n,s.shape,s),h=new J(a,d.indices,new q(s,d.indices)),m=new fe(n.blockName("scatter_copy"),d.loopBinds,[{buffer:s}],[{buffer:a}],h),g=Mt(m,d.loopVars,s.shape,d.extentNodes),{loopVars:y,loopBinds:_,indices:b,extentNodes:w}=Rt(n,i.shape,i),x=[],S=[];for(let E=0;E<i.shape.length;E++)f.has(E)?S.push(b[E]):x.push(b[E]);let T=new Array(s.shape.length),v=0;for(let E=0;E<s.shape.length;E++)u.has(E)?T[E]=new L(0):T[E]=S[v++];for(let E=0;E<p.length;E++){let F=new Array(o.shape.length),C=0;for(let U=0;U<o.shape.length;U++)U===c?F[U]=new L(E):F[U]=x[C++];let H=IA(new q(o,F),o.dtype),G=p[E];T[G]=new P("+",T[G],H)}let A=new q(i,b),N=new q(a,T),k=new P("+",N,A),O=new J(a,T,k),R=new fe(n.blockName("scatter_update"),_,[{buffer:i},{buffer:o}],[{buffer:a}],O),B=Mt(R,y,i.shape,w);return new te([g,B])}),Ie("fused_dot_epilogue",(n,e,t,r)=>{let s=e.getAttr("num_dot_operands")||2,o=t[0],i=t[1],a=t.slice(s),u=r[0],p=e.getAttr("epilogue_tags")||[],{initBody:c,accBody:f}=bp(n,e,o,i,u,{prefix:"ei",initBlockName:"matmul_init",accBlockName:"matmul_acc",initVal:l(()=>new de(0),"initVal"),accLeaf:l((x,S)=>new P("*",x,S),"accLeaf")});if(p.length===0)return new te([c,f]);let d=_r(n,"ep",Array.from({length:u.shape.length},(x,S)=>S),u.shape,u),h=d.indices,m=new q(u,h),g={v:0};for(let x of p){let S=Xr.get(x);S&&(m=S(m,a,g,u,h))}let y=pl([u,...a]),_=new J(u,h,m),b=new fe(n.blockName("epilogue"),d.ivs,y,[{buffer:u}],_),w=d.wrap(b);return new te([c,f,w])}),Ie("cublas_gemm",Fi("dot"))}var Xr,LA=M(()=>{"use strict";Fe();Re();Qr();l(IA,"asIndexValue");Xr=new Map;Xr.set("bias",(n,e,t,r,s)=>{let o=e[t.v++];return new P("+",n,new q(o,ti(o,r,s)))});Xr.set("residual_add",(n,e,t,r,s)=>{let o=e[t.v++];return new P("+",n,new q(o,ti(o,r,s)))});Xr.set("scale",(n,e,t,r,s)=>{let o=e[t.v++];return new P("*",n,new q(o,ti(o,r,s)))});Xr.set("relu",(n,e,t,r)=>new Ne("max",[n,new de(0)],r.dtype));Xr.set("clamp",(n,e,t,r,s)=>{let o=e[t.v++],i=e[t.v++];return new Ne("min",[new Ne("max",[n,new q(o,ti(o,r,s))],r.dtype),new q(i,ti(i,r,s))],r.dtype)});Xr.set("neg",n=>new P("-",n));Xr.set("exp",(n,e,t,r)=>new Ne("exp",[n],r.dtype));Xr.set("tanh",(n,e,t,r)=>new Ne("tanh",[n],r.dtype));Xr.set("sqrt",(n,e,t,r)=>new Ne("sqrt",[n],r.dtype));Xr.set("abs",(n,e,t,r)=>new Ne("abs",[n],r.dtype));Xr.set("log",(n,e,t,r)=>new Ne("log",[n],r.dtype));Xr.set("activation",n=>n);l(EA,"register")});function dl(n,e,t){let{loopVars:r,loopBinds:s,indices:o,extentNodes:i}=Rt(n,t.shape,t),a=new J(t,o,new q(e,o)),u=new fe(n.blockName("cf_copy"),s,[{buffer:e}],[{buffer:t}],a);return Mt(u,r,t.shape,i)}function XP(n,e,t,r){let{loopVars:s,loopBinds:o,indices:i,extentNodes:a}=Rt(n,t.shape,t),u=new J(t,i,new q(e,[r,...i])),p=new fe(n.blockName("scan_in"),o,[{buffer:e}],[{buffer:t}],u);return Mt(p,s,t.shape,a)}function e$(n,e,t,r){let{loopVars:s,loopBinds:o,indices:i,extentNodes:a}=Rt(n,e.shape,e),u=new J(t,[r,...i],new q(e,i)),p=new fe(n.blockName("scan_out"),o,[{buffer:e}],[{buffer:t}],u);return Mt(p,s,e.shape,a)}function Tp(n,e,t){let r=e.entryBlock,s=new Map;for(let i=0;i<r.arguments.length;i++)s.set(r.arguments[i],t[i]),t[i]&&n.bufferMap.set(r.arguments[i],t[i]);let o=[];for(let i of r.ops()){if(i.opName==="yield"){let d=new Array(i.numOperands);for(let h=0;h<i.numOperands;h++)d[h]=s.get(i.getOperand(h))||n.getOrAllocBuffer(i.getOperand(h));return{stmts:o,yieldBuffers:d}}if(_n(i.opName)){o.push(fl(n,i));continue}let a=new Array(i.numOperands);for(let d=0;d<i.numOperands;d++)a[d]=s.get(i.getOperand(d))||i.getOperand(d);let u=new Array(a.length);for(let d=0;d<a.length;d++)u[d]=a[d]instanceof me?a[d]:n.getOrAllocBuffer(a[d]);let p=new Array(i.numResults);for(let d=0;d<i.numResults;d++){let h={type:i.getResult(d).type};p[d]=n.getOrAllocBuffer(h),s.set(i.getResult(d),p[d]),n.bufferMap.set(i.getResult(d),p[d])}let c=Fi(i.opName);if(!c)throw new Error(`No lowering rule for op '${i.opName}' inside region`);let f=c(n,i,u,p);f&&o.push(f)}return{stmts:o,yieldBuffers:[]}}function CA(){Ie("if",(n,e,t,r)=>{let s=t[0],o=new q(s,[]),i=e.regions[0],a=e.regions[1],u=new Array(e.numResults);for(let h=0;h<e.numResults;h++)u[h]=r[h]||n.getOrAllocBuffer(e.getResult(h)),n.bufferMap.set(e.getResult(h),u[h]);let p=Tp(n,i,[]),c=p.stmts.slice();for(let h=0;h<u.length&&h<p.yieldBuffers.length;h++){let m=p.yieldBuffers[h];m&&m!==u[h]&&c.push(dl(n,m,u[h]))}let f=c.length===1?c[0]:new te(c),d=null;if(a&&a.entryBlock){let h=Tp(n,a,[]),m=h.stmts.slice();for(let g=0;g<u.length&&g<h.yieldBuffers.length;g++){let y=h.yieldBuffers[g];y&&y!==u[g]&&m.push(dl(n,y,u[g]))}m.length>0&&(d=m.length===1?m[0]:new te(m))}return new Se(o,f,d)}),Ie("while",(n,e,t,r)=>{let s=e.regions[0],o=e.regions[1],i=new Array(t.length),a=[];for(let y=0;y<t.length;y++){let _=r[y]||n.getOrAllocBuffer({type:e.getResult(y).type});i[y]=_,a.push(dl(n,t[y],_))}for(let y=0;y<e.numResults;y++)n.bufferMap.set(e.getResult(y),i[y]);let u=new me(`_wcond_${n.varCounter++}`,[],"bool",ul.GLOBAL),p=Tp(n,s,i),c=p.stmts.slice(),f=p.yieldBuffers[0];c.push(new J(u,[],new q(f,[])));let d=c.length===1?c[0]:new te(c),h=Tp(n,o,i),m=h.stmts.slice();for(let y=0;y<i.length&&y<h.yieldBuffers.length;y++){let _=h.yieldBuffers[y];_&&_!==i[y]&&m.push(dl(n,_,i[y]))}let g=m.length===1?m[0]:new te(m);return new te([...a,new $i(u,d,g)])}),Ie("scan",(n,e,t,r)=>{let s=e.getAttr("num_xs"),o=e.getAttr("num_carry"),i=t.slice(0,s),a=t.slice(s),u=new Array(o),p=[];for(let b=0;b<o;b++){let w=r[b]||n.getOrAllocBuffer({type:e.getResult(b).type});u[b]=w,p.push(dl(n,a[b],w)),n.bufferMap.set(e.getResult(b),w)}let c=r.length-o,f=new Array(c);for(let b=0;b<c;b++){let w=r[o+b]||n.getOrAllocBuffer({type:e.getResult(o+b).type});f[b]=w,n.bufferMap.set(e.getResult(o+b),w)}let d=n.allocVar("t"),h=n.extentNode(i[0].shape[0],i[0],0),m=[],g=new Array(s);for(let b=0;b<s;b++){let w=n.getOrAllocBuffer({type:{shape:i[b].shape.slice(1),dtype:i[b].dtype}});g[b]=w,m.push(XP(n,i[b],w,d))}let y=Tp(n,e.regions[0],[...g,...u]);for(let b of y.stmts)m.push(b);for(let b=0;b<o;b++){let w=y.yieldBuffers[b];w&&w!==u[b]&&m.push(dl(n,w,u[b]))}for(let b=0;b<c;b++)m.push(e$(n,y.yieldBuffers[o+b],f[b],d));m.push(new Fn);let _=m.length===1?m[0]:new te(m);return new te([...p,new he(d,new L(0),h,z.RECURRENCE,_)])})}var DA=M(()=>{"use strict";Fe();ei();R0();Qr();l(dl,"copyBuffer");l(XP,"sliceCopyIn");l(e$,"sliceCopyOut");l(Tp,"lowerRegionBody");l(CA,"register")});function PA(){Ie("layout_transform",(n,e,t,r)=>{let s=t[0],o=r[0],i=e.getAttr("src_layout"),a=e.getAttr("dst_layout"),{loopVars:u,loopBinds:p,indices:c,extentNodes:f}=Rt(n,o.shape,o),d=new J(o,c,new q(s,c)),h=new fe(n.blockName("layout_transform_block"),p,[{buffer:s}],[{buffer:o}],d);return Mt(h,u,o.shape,f)})}var $A=M(()=>{"use strict";Fe();Qr();l(PA,"register")});function t$(n){let e=n==="ui8",t=8;return{cMin:e?0:-(1<<t-1),cMax:e?(1<<t)-1:(1<<t-1)-1}}function _m(n,{scale:e,zeroPoint:t,targetDtype:r}){let s=r||"i8",{cMin:o,cMax:i}=t$(s),a=new P("/",n,new de(e)),u=t!==0?new P("+",a,new de(t)):a,p=new Ne("round",[u],"f32"),c=new Ne("min",[new Ne("max",[p,new de(o)],"f32"),new de(i)],"f32");return new ut(c,"f32",s)}function bm(n,{scale:e,zeroPoint:t,srcDtype:r,targetDtype:s}){let o=s||"f32",i=new ut(n,r||"i8",o),a=t!==0?new P("-",i,new de(t)):i;return new P("*",a,new de(e))}var F0=M(()=>{"use strict";Fe();l(t$,"clampRange");l(_m,"buildQuantizeExpr");l(bm,"buildDequantizeExpr")});function FA(){Ie("quantize",(n,e,t,r)=>{let s=e.getAttr("scale"),o=e.getAttr("zero_point"),i=e.getAttr("target_dtype")||"i8";return mo(n,e,t,r,(a,u)=>_m(u[0],{scale:s,zeroPoint:o,targetDtype:i}))}),Ie("dequantize",(n,e,t,r)=>{let s=e.getAttr("scale"),o=e.getAttr("zero_point"),i=t[0].dtype||"i8",a=e.getAttr("target_dtype")||"f32";return mo(n,e,t,r,(u,p)=>bm(p[0],{scale:s,zeroPoint:o,srcDtype:i,targetDtype:a}))}),Ie("quantized_dot",(n,e,t,r)=>{let s=t[0],o=t[1],i=e.getAttr("lhs_zero_point")||0,a=e.getAttr("rhs_zero_point")||0,{initBody:u,accBody:p}=bp(n,e,s,o,r[0],{prefix:"qdi",initBlockName:"qmatmul_init",accBlockName:"qmatmul_acc",initVal:l(()=>new L(0),"initVal"),accLeaf:l((c,f)=>{let d=new ut(c,s.dtype,"i32"),h=new ut(f,o.dtype,"i32"),m=i!==0?new P("-",d,new L(i)):d,g=a!==0?new P("-",h,new L(a)):h;return new P("*",m,g)},"accLeaf")});return new te([u,p])}),Ie("quantized_conv",(n,e,t,r)=>{let s=t[0],o=t[1],i=e.getAttr("input_zero_point")||0,a=e.getAttr("kernel_zero_point")||0;return gm(n,e,s,o,r[0],{prefix:"qc",blockPrefix:"qconv",initVal:l(()=>new L(0),"initVal"),guardFill:l(()=>new L(0),"guardFill"),leafBuilder:l((u,p)=>{let c=new ut(new q(s,u),s.dtype,"i32"),f=new ut(new q(o,p),o.dtype,"i32"),d=i!==0?new P("-",c,new L(i)):c,h=a!==0?new P("-",f,new L(a)):f;return new P("*",d,h)},"leafBuilder")})})}var zA=M(()=>{"use strict";Fe();Qr();F0();l(FA,"register")});function ni(n){return es.has(n)}function n$(){for(let n of wp())es.set(n,(e,t,r)=>$0(e.opName,t,r));es.set("compare",(n,e)=>new Ke(n.getAttr("direction")||"eq",e[0],e[1])),es.set("select",(n,e)=>new Se(e[0],e[1],e[2])),es.set("clamp",(n,e,t)=>new Ne("min",[new Ne("max",[e[1],e[0]],t),e[2]],t)),es.set("convert",(n,e)=>new ut(e[0],n.getOperand(0).type.dtype,n.getAttr("target_dtype")||n.getResult(0).type.dtype)),es.set("broadcast_in_dim",(n,e)=>e[0]),es.set("broadcast",(n,e)=>e[0]),es.set("iota",()=>{throw new Error("iota fusion must be handled by the index-aware path in lowerFusion")}),es.set("quantize",(n,e)=>_m(e[0],{scale:n.getAttr("scale"),zeroPoint:n.getAttr("zero_point"),targetDtype:n.getAttr("target_dtype")||"i8"})),es.set("dequantize",(n,e)=>bm(e[0],{scale:n.getAttr("scale"),zeroPoint:n.getAttr("zero_point"),srcDtype:n.getOperand(0).type?.dtype||"i8",targetDtype:n.getAttr("target_dtype")||"f32"}))}function VA(n,e){let t=e.numOperands,r=e.numResults,s=new Array(t);for(let N=0;N<t;N++)s[N]=n.getOrAllocBuffer(e.getOperand(N));let o=new Array(r);for(let N=0;N<r;N++)o[N]=n.getOrAllocBuffer(e.getResult(N));let i=o[0],{loopVars:a,loopBinds:u,indices:p,extentNodes:c}=Rt(n,i.shape,i),f=new Map,d=e.regions[0].entryBlock,h=new Map,m=[...d.ops()];for(let N=m.length-1;N>=0;N--){let k=m[N];if(cl(k.opName)){let O=k.getAttr("broadcast_dimensions");O&&O.length>0&&h.set(k.getOperand(0),O);continue}if(!(k.opName==="yield"||_n(k.opName)))for(let O=0;O<k.numResults;O++){let R=h.get(k.getResult(O));if(R)for(let B=0;B<k.numOperands;B++)h.has(k.getOperand(B))||h.set(k.getOperand(B),R)}}let g=new Map,y=d.arguments;for(let N=0;N<y.length;N++){let k=h.get(y[N]);k&&g.set(N,k)}for(let N=0;N<d.arguments.length;N++){let k=g.get(N),O;if(k){let R=s[N];O=new Array(R.shape.length);for(let B=0;B<R.shape.length;B++)O[B]=R.shape[B]===1?new L(0):p[k[B]]}else O=ti(s[N],i,p);f.set(d.arguments[N],new q(s[N],O))}let _=new Map;for(let N of d.ops())for(let k=0;k<N.numOperands;k++){let O=N.getOperand(k);_.set(O,(_.get(O)||0)+1)}let b=new Map,w=0,x=[];function S(N){let k=f.get(N);if(k===void 0)throw new Error(`Fusion lowering: unmapped operand from '${N.definingOp?N.definingOp.opName:"unknown"}'`);if((_.get(N)||0)>1&&!r$.has(k.type)){if(!b.has(N)){let O=k.type==="CompareNode"?"i32":k.type==="CastNode"?k.toDtype:N.type&&N.type.dtype?N.type.dtype:i.dtype,R=n.allocVar(`cse${w++}`,O);b.set(N,R),x.push({variable:R,value:k}),f.set(N,R)}return b.get(N)}return k}l(S,"getExpr");let T=[];for(let N of d.ops()){if(N.opName==="yield"){for(let B=0;B<N.numOperands;B++)T.push(new J(o[B],p,S(N.getOperand(B))));break}if(_n(N.opName)){let B=N.getAttr("value");f.set(N.getResult(0),new de(typeof B=="number"?B:0));continue}if(N.opName==="iota"){let B=N.getAttr("iota_dimension")??N.getAttr("dimension")??0;f.set(N.getResult(0),p[B]);continue}let k=es.get(N.opName);if(!k)throw new Error(`Fusion lowering: unsupported op '${N.opName}' inside fusion body`);let O=new Array(N.numOperands);for(let B=0;B<N.numOperands;B++)O[B]=S(N.getOperand(B));let R=N.getResult(0).type.dtype;f.set(N.getResult(0),k(N,O,R))}let v=T.length===1?T[0]:new te(T);for(let N=x.length-1;N>=0;N--)v=new et(x[N].variable,x[N].value,v);let A=new fe(n.blockName("fusion_block"),u,pl(s),pl(o),v);return Mt(A,a,i.shape,c)}function s$(n,e){if(n.length!==e.length)return!1;for(let t=0;t<n.length;t++)if(n[t]!==e[t])return!1;return!0}function GA(n){let e=n.regions[0];if(!e)return!1;for(let t of e.entryBlock.ops())if(t.opName!=="yield"){if(_n(t.opName)){if(typeof t.getAttr("value")=="number")continue;return!1}if(!es.has(t.opName))return!1}if(n.numResults>1){let t=n.getResult(0).type.shape;for(let r=1;r<n.numResults;r++)if(!s$(n.getResult(r).type.shape,t))return!1}return!0}function jA(n,e,t){let r=e.regions[0].entryBlock,s=new Map;for(let i=0;i<r.arguments.length;i++)s.set(r.arguments[i],e.getOperand(i));let o=new Map;for(let i of r.ops())if(i.opName==="yield"){for(let a=0;a<i.numOperands;a++)o.set(i.getOperand(a),e.getResult(a));break}for(let i of r.ops()){if(i.opName==="yield")continue;let a=new Array(i.numOperands);for(let d=0;d<i.numOperands;d++)a[d]=s.get(i.getOperand(d))||i.getOperand(d);let u=new Array(a.length);for(let d=0;d<a.length;d++)u[d]=n.getOrAllocBuffer(a[d]);let p=new Array(i.numResults);for(let d=0;d<i.numResults;d++){let h=i.getResult(d),m=o.get(h);if(m){let g=n.getOrAllocBuffer(m);p[d]=g,s.set(h,m)}else{let g={type:h.type};p[d]=n.getOrAllocBuffer(g),s.set(h,g)}}if(_n(i.opName)){t.push(fl(n,i));continue}let c=Fi(i.opName);if(!c)throw new Error(`Fusion lowering: no lowering rule for op '${i.opName}' inside fusion body`);let f=c(n,i,u,p);f&&t.push(f)}}function UA(){n$()}var es,r$,qA=M(()=>{"use strict";Fe();Qr();ym();F0();Ir();es=new Map;l(ni,"canInlineFuse");l(n$,"initBuiltinFusionBuilders");r$=new Set(["BufferLoadNode","VariableNode","IntImmNode","FloatImmNode"]);l(VA,"lowerFusion");l(s$,"shapesEqual");l(GA,"canLowerAsElementwiseFusion");l(jA,"lowerFusionAsIndividualOps");l(UA,"register")});function WA(n,e,t,r,s){let o=new P("-",new P("*",n,new L(e)),new L(r)),i=new P("+",o,new L(t)),a=new Ne("max",[o,new L(0)],"index"),u=new Ne("min",[i,new L(s)],"index"),p=new P("-",u,a);return new Ne("max",[p,new L(0)],"index")}function o$(n,e,t,r,s,o,i,a,u,p,c){let f=WA(n,t,s,i,u),d=WA(e,r,o,a,p),h=new P("*",f,d);return new ut(h,"index",c)}function KA(){Ie("pool2d",(n,e,t,r)=>{let s=t[0],o=r[0],i=e.getAttr("pool_type"),a=e.getAttr("kernel_size"),u=e.getAttr("strides"),p=e.getAttr("padding"),c=e.getAttr("count_include_pad")||!1,f=mm(e.getAttr("layout")||"NCHW"),d=f.N,h=f.C,m=f.H,g=f.W,y=s.shape[d],_=s.shape[h],b=s.shape[m],w=s.shape[g],x=o.shape[m],S=o.shape[g],T=a[0],v=a[1],A=u[0],N=u[1],k=p[0][0],O=p[1][0],R=i==="max",B=R?new de(-1/0):new de(0),E=_r(n,"pi",[0,1,2,3],o.shape,o),F=new J(o,E.indices,B),C=new fe(n.blockName("pool_init"),E.ivs,[],[{buffer:o}],F),H=E.wrap(C),G=n.allocVar("pn"),U=n.allocVar("pc"),D=n.allocVar("poh"),V=n.allocVar("pow"),ue=n.allocVar("pkh"),ce=n.allocVar("pkw"),Le=[G,U,D,V,ue,ce],j=n.allocBindArray("pv",Le),W=j[0].iterVar,ie=j[1].iterVar,je=j[2].iterVar,ke=j[3].iterVar,rt=j[4].iterVar,zt=j[5].iterVar,on=new P("+",new P("*",je,new L(A)),new P("-",rt,new L(k))),Zt=new P("+",new P("*",ke,new L(N)),new P("-",zt,new L(O))),it=new Ke("ge",on,new L(0)),ye=new Ke("lt",on,new L(b)),Xe=new Ke("ge",Zt,new L(0)),Nn=new Ke("lt",Zt,new L(w)),Qt=new P("*",new P("*",it,ye),new P("*",Xe,Nn)),an=new Array(4);an[d]=W,an[h]=ie,an[m]=je,an[g]=ke;let un=new Array(4);un[d]=W,un[h]=ie,un[m]=on,un[g]=Zt;let mn=new q(s,un),cr=new q(o,an),Vn;if(R){let xt=new Se(Qt,mn,new de(-1/0));Vn=new Ne("max",[cr,xt],o.dtype)}else{let xt=new Se(Qt,mn,new de(0));Vn=new P("+",cr,xt)}let Or=new J(o,an,Vn),yt=new fe(n.blockName("pool_acc"),j,[{buffer:s}],[{buffer:o}],Or);yt=new he(ce,new L(0),new L(v),z.SERIAL,yt),yt=new he(ue,new L(0),new L(T),z.SERIAL,yt),yt=new he(V,new L(0),new L(S),z.SERIAL,yt),yt=new he(D,new L(0),new L(x),z.SERIAL,yt),yt=new he(U,new L(0),new L(_),z.SERIAL,yt),yt=new he(G,new L(0),new L(y),z.SERIAL,yt);let fr=[H,yt];if(!R){let xt=_r(n,"pd",[0,1,2,3],o.shape,o),Vr=xt.indices[m],Ds=xt.indices[g],uo=c?new de(T*v):o$(Vr,Ds,A,N,T,v,k,O,b,w,o.dtype),Gr=new q(o,xt.indices),dr=c?new P("*",Gr,new de(1/(T*v))):new P("/",Gr,uo),Br=new J(o,xt.indices,dr),qo=new fe(n.blockName("pool_div"),xt.ivs,[{buffer:o}],[{buffer:o}],Br);fr.push(xt.wrap(qo))}return new te(fr)})}var HA=M(()=>{"use strict";Fe();Qr();l(WA,"spatialCount");l(o$,"avgPoolDivisorExpr");l(KA,"register")});function YA(){Ie("resize",(n,e,t,r)=>{let s=t[0],o=r[0],i=e.getAttr("method"),a=s.shape[2],u=s.shape[3],p=o.shape[2],c=o.shape[3],f=_r(n,"rz",[0,1,2,3],o.shape,o),[d,h,m,g]=f.indices,y=a/p,_=u/c,b;if(i==="nearest"){let S=new Ne("floor",[new P("*",m,new de(y))],"f32"),T=new Ne("floor",[new P("*",g,new de(_))],"f32"),v=new ut(new Ne("min",[new Ne("max",[S,new de(0)],"f32"),new de(a-1)],"f32"),"f32","i32"),A=new ut(new Ne("min",[new Ne("max",[T,new de(0)],"f32"),new de(u-1)],"f32"),"f32","i32");b=new q(s,[d,h,v,A])}else{let S=new P("*",m,new de(y)),T=new P("*",g,new de(_)),v=new Ne("floor",[S],"f32"),A=new Ne("floor",[T],"f32"),N=new Ne("min",[new P("+",v,new de(1)),new de(a-1)],"f32"),k=new Ne("min",[new P("+",A,new de(1)),new de(u-1)],"f32"),O=new P("-",S,v),R=new P("-",T,A),B=new Ne("max",[v,new de(0)],"f32"),E=new Ne("max",[A,new de(0)],"f32"),F=new ut(B,"f32","i32"),C=new ut(E,"f32","i32"),H=new ut(N,"f32","i32"),G=new ut(k,"f32","i32"),U=new q(s,[d,h,F,C]),D=new q(s,[d,h,F,G]),V=new q(s,[d,h,H,C]),ue=new q(s,[d,h,H,G]),ce=new P("-",new de(1),R),Le=new P("-",new de(1),O),j=new P("+",new P("*",U,ce),new P("*",D,R)),W=new P("+",new P("*",V,ce),new P("*",ue,R));b=new P("+",new P("*",j,Le),new P("*",W,O))}let w=new J(o,f.indices,b),x=new fe(n.blockName("resize_block"),f.ivs,[{buffer:s}],[{buffer:o}],w);return f.wrap(x)})}var JA=M(()=>{"use strict";Fe();Qr();l(YA,"register")});function z0(n,e){let[t,r,s]=e;return{Q:t,K:r,V:s,O:null,scale:n.getAttr("scale"),causal:!!n.getAttr("causal"),dtype:t.dtype,B:t.shape[0],H:t.shape[1],Lq:t.shape[2],Dk:t.shape[3],Lk:r.shape[2],Dv:s.shape[3]}}function ZA(n,e,t,r){let{Q:s,K:o,V:i,scale:a,causal:u,dtype:p,B:c,H:f,Lq:d,Dk:h,Lk:m,Dv:g}=z0(e,t),y=r[0],_=new me(n.blockName("fa_m"),[1],p,"local"),b=new me(n.blockName("fa_l"),[1],p,"local"),w=new me(n.blockName("fa_s"),[1],p,"local"),x=new me(n.blockName("fa_mn"),[1],p,"local"),S=new me(n.blockName("fa_p"),[1],p,"local"),T=new me(n.blockName("fa_cr"),[1],p,"local"),v=new me(n.blockName("fa_o"),[g],p,"local"),A=new Oe(n.blockName("fa_b"),"int32"),N=new Oe(n.blockName("fa_h"),"int32"),k=new Oe(n.blockName("fa_i"),"i32"),O=new Oe(n.blockName("fa_j"),"int32"),R=new Oe(n.blockName("fa_d"),"int32"),B=new Oe(n.blockName("fa_di"),"int32"),E=new Oe(n.blockName("fa_do"),"int32"),F=new Oe(n.blockName("fa_dw"),"int32"),C=new fe(n.blockName("fa_initm"),[],[],[{buffer:_}],mt(_,[_e],new de(-1/0))),H=new fe(n.blockName("fa_initl"),[],[],[{buffer:b}],mt(b,[_e],new de(0))),G=Er(B,g,new fe(n.blockName("fa_inito"),[],[],[{buffer:v}],mt(v,[B],new de(0)))),U=new fe(n.blockName("fa_sinit"),[],[],[{buffer:w}],mt(w,[_e],new de(0))),D=Er(R,h,new fe(n.blockName("fa_sred"),[],[{buffer:s},{buffer:o},{buffer:w}],[{buffer:w}],mt(w,[_e],lt("+",Ce(w,[_e]),lt("*",Ce(s,[A,N,k,R]),Ce(o,[A,N,O,R])))))),V=new te([U,D]),ue=new fe(n.blockName("fa_supd"),[],[{buffer:w},{buffer:_},{buffer:b}],[{buffer:w},{buffer:x},{buffer:S},{buffer:T},{buffer:b}],new te([mt(w,[_e],lt("*",Ce(w,[_e]),new de(a))),mt(x,[_e],hl("max",Ce(_,[_e]),Ce(w,[_e]))),mt(S,[_e],hl("exp",lt("-",Ce(w,[_e]),Ce(x,[_e])))),mt(T,[_e],hl("exp",lt("-",Ce(_,[_e]),Ce(x,[_e])))),mt(b,[_e],lt("+",lt("*",Ce(b,[_e]),Ce(T,[_e])),Ce(S,[_e])))])),ce=Er(E,g,new fe(n.blockName("fa_oupd"),[],[{buffer:v},{buffer:T},{buffer:S},{buffer:i}],[{buffer:v}],mt(v,[E],lt("+",lt("*",Ce(v,[E]),Ce(T,[_e])),lt("*",Ce(S,[_e]),Ce(i,[A,N,O,E])))))),Le=new fe(n.blockName("fa_mupd"),[],[{buffer:x}],[{buffer:_}],mt(_,[_e],Ce(x,[_e]))),j=new te([V,ue,ce,Le]),W=m-d,ie=u?new Se(new Ke("le",O,lt("+",k,zi(W))),j):j,je=Er(O,m,ie),ke=Er(F,g,new fe(n.blockName("fa_finw"),[],[{buffer:v},{buffer:b}],[{buffer:y}],mt(y,[A,N,k,F],lt("/",Ce(v,[F]),Ce(b,[_e]))))),rt=new te([C,H,G,je,ke]);for(let ye of[v,T,S,x,w,b,_])rt=new Ue(ye,ye.scope,rt);let zt=Math.min(d,256),on=Math.ceil(d/zt),Zt;if(on===1)Zt=new he(k,_e,new L(d),z.THREAD_BINDING,rt,"threadIdx.x");else{let ye=new Oe(n.blockName("fa_ii"),"i32"),Xe=new Oe(n.blockName("fa_io"),"i32"),Nn=lt("+",lt("*",Xe,new L(zt)),ye),Qt=on*zt===d?rt:new Se(new Ke("lt",k,new L(d)),rt),an=new he(ye,_e,new L(zt),z.THREAD_BINDING,new et(k,Nn,Qt),"threadIdx.x");Zt=new he(Xe,_e,new L(on),z.THREAD_BINDING,an,"blockIdx.z")}let it=new he(N,_e,new L(f),z.THREAD_BINDING,Zt,"blockIdx.y");return new he(A,_e,new L(c),z.THREAD_BINDING,it,"blockIdx.x")}function i$(n,e,t,r,s){let{Q:o,K:i,V:a,scale:u,causal:p,dtype:c,B:f,H:d,Lq:h,Dk:m,Lk:g,Dv:y}=z0(e,t),_=r[0],b=Math.ceil(g/s),w=Math.ceil(h/s),x=g-h,S=new me(n.blockName("fa_Ks"),[s,m],c,"shared"),T=new me(n.blockName("fa_Vs"),[s,y],c,"shared"),v=new me(n.blockName("fa_q"),[m],c,"local"),A=new me(n.blockName("fa_m"),[1],c,"local"),N=new me(n.blockName("fa_l"),[1],c,"local"),k=new me(n.blockName("fa_s"),[1],c,"local"),O=new me(n.blockName("fa_mn"),[1],c,"local"),R=new me(n.blockName("fa_p"),[1],c,"local"),B=new me(n.blockName("fa_cr"),[1],c,"local"),E=new me(n.blockName("fa_o"),[y],c,"local"),F=new Oe(n.blockName("fa_b"),"int32"),C=new Oe(n.blockName("fa_h"),"int32"),H=new Oe(n.blockName("fa_qb"),"int32"),G=new Oe(n.blockName("fa_t"),"int32"),U=new Oe(n.blockName("fa_i"),"i32"),D=new Oe(n.blockName("fa_kt"),"int32"),V=new Oe(n.blockName("fa_jj"),"int32"),ue=new Oe(n.blockName("fa_dq"),"int32"),ce=new Oe(n.blockName("fa_dk"),"int32"),Le=new Oe(n.blockName("fa_dv"),"int32"),j=new Oe(n.blockName("fa_ds"),"int32"),W=new Oe(n.blockName("fa_doo"),"int32"),ie=new Oe(n.blockName("fa_dw"),"int32"),je=new Oe(n.blockName("fa_din"),"int32"),ke=lt("+",lt("*",H,zi(s)),G),rt=lt("+",lt("*",D,zi(s)),G),zt=lt("+",lt("*",D,zi(s)),V),on=new Ke("lt",U,zi(h)),Zt=new Ke("lt",rt,zi(g)),it=Er(ue,m,mt(v,[ue],Ce(o,[F,C,U,ue]))),ye=mt(A,[_e],new de(-1/0)),Xe=mt(N,[_e],new de(0)),Nn=Er(je,y,mt(E,[je],new de(0))),Qt=new Se(on,new te([it,ye,Xe,Nn])),an=new Se(Zt,Er(ce,m,mt(S,[G,ce],Ce(i,[F,C,rt,ce])))),un=new Se(Zt,Er(Le,y,mt(T,[G,Le],Ce(a,[F,C,rt,Le])))),mn=mt(k,[_e],new de(0)),cr=Er(j,m,mt(k,[_e],lt("+",Ce(k,[_e]),lt("*",Ce(v,[j]),Ce(S,[V,j]))))),Vn=new te([mt(k,[_e],lt("*",Ce(k,[_e]),new de(u))),mt(O,[_e],hl("max",Ce(A,[_e]),Ce(k,[_e]))),mt(R,[_e],hl("exp",lt("-",Ce(k,[_e]),Ce(O,[_e])))),mt(B,[_e],hl("exp",lt("-",Ce(A,[_e]),Ce(O,[_e])))),mt(N,[_e],lt("+",lt("*",Ce(N,[_e]),Ce(B,[_e])),Ce(R,[_e])))]),Or=Er(W,y,mt(E,[W],lt("+",lt("*",Ce(E,[W]),Ce(B,[_e])),lt("*",Ce(R,[_e]),Ce(T,[V,W]))))),pr=mt(A,[_e],Ce(O,[_e])),yt=new te([mn,cr,Vn,Or,pr]),fr=new Ke("lt",zt,zi(g)),xt=p?new Se(new Ke("le",zt,lt("+",U,zi(x))),yt):yt,Vr=Er(V,s,new Se(fr,xt)),Ds=new Se(on,Vr),uo=new te([an,un,new Fn,Ds,new Fn]),Gr=Er(D,b,uo),dr=new Se(on,Er(ie,y,mt(_,[F,C,U,ie],lt("/",Ce(E,[ie]),Ce(N,[_e]))))),Br=new et(U,ke,new te([Qt,Gr,dr]));for(let ms of[E,B,R,O,k,N,A,v])Br=new Ue(ms,ms.scope,Br);let qo=wm(G,"threadIdx.x",s,Br),jr=new Ue(S,"shared",new Ue(T,"shared",qo)),Ur=wm(H,"blockIdx.z",w,jr),Ps=wm(C,"blockIdx.y",d,Ur);return wm(F,"blockIdx.x",f,Ps)}function QA(){Ie("scaled_dot_product_attention",ZA),vA("scaled_dot_product_attention","cuda",(n,e,t,r)=>{let{Lq:s,Dk:o,Dv:i,dtype:a}=z0(e,t),u=n.target,p=u&&u.sharedMemoryBytes||0,c=u&&u.maxThreadsPerBlock||1024,f=Math.min(c,s,Math.floor(p/((o+i)*Rn(a)*2)));return f<8?ZA(n,e,t,r):i$(n,e,t,r,f)})}var _e,zi,Ce,mt,lt,hl,Er,wm,XA=M(()=>{"use strict";Fe();ei();Qr();tr();_e=new L(0),zi=l(n=>new L(n),"I"),Ce=l((n,e)=>new q(n,e),"ld"),mt=l((n,e,t)=>new J(n,e,t),"st"),lt=l((n,e,t)=>new P(n,e,t),"mop"),hl=l((n,...e)=>new Ne(n,e),"ext"),Er=l((n,e,t,r=z.SERIAL)=>new he(n,_e,new L(e),r,t),"forL"),wm=l((n,e,t,r)=>new he(n,_e,new L(t),z.THREAD_BINDING,r,e),"forT");l(z0,"dims");l(ZA,"buildNaive");l(i$,"buildTiled");l(QA,"register")});function e1(n,e,t=new Set){let r=ne.get(e.opName);if(!r||!r.getAttr("broadcastViewSafe"))return!1;if(e.opName!=="fusion")return!0;let s=e.regions[0];if(!s)return!1;let o=s.entryBlock;for(let i=0;i<e.numOperands;i++){if(e.getOperand(i)!==n)continue;let a=o.arguments[i];for(let u of a.getUsers()){let p=`${a.id}:${u.id}`;if(!t.has(p)&&(t.add(p),!e1(a,u,t)))return!1}}return!0}function u$(n){return um([...n.ops()],"ignore")}function Tm(n,e=null,t=null){let r=new hm;r.target=e;let s=[],o=new Map;for(let g of n.args){let y=r.allocVar("arg");s.push(y),o.set(y,r.getOrAllocBuffer(g))}let i=n.getReturnOp(),a=new Set;for(let g of n.args)a.add(r.getOrAllocBuffer(g));let u=[],p=new Set;for(let g=0;g<i.numOperands;g++){let y=r.allocVar("ret");s.push(y);let _=r.getOrAllocBuffer(i.getOperand(g));if(a.has(_)||p.has(_)){let b=r.allocFreshBuffer(i.getOperand(g));o.set(y,b),u.push({src:_,dst:b})}else o.set(y,_),p.add(_)}let c=new Set;for(let g=0;g<i.numOperands;g++)c.add(i.getOperand(g));let f=[];for(let g of n.ops())_n(g.opName)&&f.push(fl(r,g));for(let g of u$(n)){if(Xt(g.opName)||_n(g.opName))continue;if(g.opName==="fusion"){GA(g)?f.push(VA(r,g)):jA(r,g,f);continue}if(cl(g.opName)&&!c.has(g.getResult(0))&&g.getOperand(0).getUsers().length===1&&g.getResult(0).getUsers().every(x=>e1(g.getResult(0),x))){let x=r.getOrAllocBuffer(g.getOperand(0)),S=g.getResult(0).type.shape,T=g.getAttr("broadcast_dimensions"),v;if(T&&T.length>0)v=T;else{let A=S.length-x.shape.length;v=Array.from({length:x.shape.length},(N,k)=>k+A)}x.broadcastDims=v,r.bufferMap.set(g.getResult(0),x);continue}let y=Fi(g.opName,e,t);if(!y)throw new Error(`No lowering rule defined for op: ${g.opName}`);let _=new Array(g.numOperands);for(let x=0;x<g.numOperands;x++)_[x]=r.getOrAllocBuffer(g.getOperand(x));let b=new Array(g.numResults);for(let x=0;x<g.numResults;x++)b[x]=r.getOrAllocBuffer(g.getResult(x));let w=y(r,g,_,b);w&&f.push(w)}for(let{src:g,dst:y}of u){let{loopVars:_,loopBinds:b,indices:w,extentNodes:x}=Rt(r,g.shape,g),S=new q(g,w),T=new J(y,w,S),v=new fe(r.blockName("copy_block"),b,[{buffer:g}],[{buffer:y}],T);f.push(Mt(v,_,g.shape,x))}let d=[],h=new Set;for(let g of r.shapeParams.values())h.has(g.name)||(h.add(g.name),d.push(g));for(let[g,y]of r.symVars)if(!h.has(y.name))throw new Error(`Symbolic dimension '${g}' has no input dimension to bind it to at runtime`);for(let g of d)s.push(g);let m=new Pi(n.name,s,f.length===1?f[0]:new te(f),o,d,new Map(r.shapeParams));return n._partitionTarget&&(m._partitionTarget=n._partitionTarget),m}var a$,$a=M(()=>{"use strict";Fe();Da();Ct();Qr();Ir();ym();BA();MA();LA();DA();$A();zA();qA();HA();JA();XA();ym();a$=["compare","select","clamp","convert","copy_to_device","dot","fusion"];for(let n of[...wp(),...a$])ne.has(n)&&ne.registerOpAttr(n,"broadcastViewSafe",!0);l(e1,"broadcastViewSafeForUser");NA();OA();RA();EA();CA();PA();FA();UA();KA();YA();QA();l(u$,"topologicalOps");l(Tm,"lowerGraphToPrimFunc")});var t1={};at(t1,{CPUTarget:()=>go,CUDATarget:()=>xm,TargetFeatures:()=>Fa,TargetKind:()=>Dt,WasmTarget:()=>Sm,WebGPUTarget:()=>vm});var Dt,Fa,go,xm,Sm,vm,yo=M(()=>{"use strict";Dt=Object.freeze({CPU:"cpu",CUDA:"cuda",WEBGPU:"webgpu",WASM:"wasm",ACCELERATOR:"accelerator"}),Fa=class{static{l(this,"TargetFeatures")}constructor(e){this.kind=e.kind,this.name=e.name,this.vectorWidth=e.vectorWidth||1,this.numCores=e.numCores||1,this.maxThreadsPerBlock=e.maxThreadsPerBlock||1,this.maxBlockDimX=e.maxBlockDimX||1,this.maxBlockDimY=e.maxBlockDimY||1,this.maxBlockDimZ=e.maxBlockDimZ||1,this.maxGridDimX=e.maxGridDimX||1,this.maxGridDimY=e.maxGridDimY||1,this.maxGridDimZ=e.maxGridDimZ||1,this.sharedMemoryBytes=e.sharedMemoryBytes||0,this.memoryBudgetBytes=e.memoryBudgetBytes||0,this.registersPerThread=e.registersPerThread||0,this.warpSize=e.warpSize||1,this.memoryBandwidthGBs=e.memoryBandwidthGBs||0,this.computeTFLOPs=e.computeTFLOPs||0,this.cacheLineSizeBytes=e.cacheLineSizeBytes||64,this.l1CacheBytes=e.l1CacheBytes||0,this.l2CacheBytes=e.l2CacheBytes||0,this.supportsFloat16=e.supportsFloat16??!1,this.supportsTensorCore=e.supportsTensorCore??!1,this.libraryOps=e.libraryOps||new Set,this.enableEpilogueFusion=e.enableEpilogueFusion??!1,this.preferredConvLayout=e.preferredConvLayout||null,this.layoutAwareOps=e.layoutAwareOps instanceof Set?e.layoutAwareOps:new Set(e.layoutAwareOps||[]),this.preferredBlockFactor=e.preferredBlockFactor||0,this.supportsBlockedLayout=e.supportsBlockedLayout??!1,this.supportsInt8=e.supportsInt8??!1,this.simd=e.simd??!1,this.host=e.host||null,this.attrs=new Map(Object.entries(e.attrs||{}))}getAttr(e,t=null){return this.attrs.has(e)?this.attrs.get(e):t}hasAttr(e){return this.attrs.has(e)}withAttr(e,t){return this.attrs.set(e,t),this}isGPU(){return this.kind===Dt.CUDA||this.kind===Dt.WEBGPU}isWebGPU(){return this.kind===Dt.WEBGPU}isCPU(){return this.kind===Dt.CPU}isWasm(){return this.kind===Dt.WASM}supportsThreadBinding(){return this.isGPU()||this.isWebGPU()}supportsVectorization(){return this.vectorWidth>1}maxParallelism(){return this.isGPU()?this.maxThreadsPerBlock*this.maxGridDimX:this.numCores}supportsSimd(){return this.simd&&this.vectorWidth>1}hasLibraryOp(e){return this.libraryOps.has(e)}},go=l((n={})=>new Fa({kind:Dt.CPU,name:"cpu_generic",vectorWidth:8,numCores:8,cacheLineSizeBytes:64,l1CacheBytes:32*1024,l2CacheBytes:256*1024,memoryBandwidthGBs:50,computeTFLOPs:.5,supportsBlockedLayout:!0,preferredBlockFactor:8,supportsInt8:!0,...n}),"CPUTarget"),xm=l((n={})=>new Fa({kind:Dt.CUDA,name:"cuda_generic",vectorWidth:1,numCores:80,maxThreadsPerBlock:1024,maxBlockDimX:1024,maxBlockDimY:1024,maxBlockDimZ:64,maxGridDimX:2147483647,maxGridDimY:65535,maxGridDimZ:65535,sharedMemoryBytes:48*1024,registersPerThread:255,warpSize:32,memoryBandwidthGBs:900,computeTFLOPs:15,supportsFloat16:!0,supportsTensorCore:!1,libraryOps:new Set(["dot","conv"]),enableEpilogueFusion:!0,supportsInt8:!0,...n}),"CUDATarget"),Sm=l((n={})=>new Fa({kind:Dt.WASM,name:"wasm_generic",vectorWidth:4,numCores:1,cacheLineSizeBytes:64,memoryBandwidthGBs:10,computeTFLOPs:.1,supportsInt8:!0,simd:!0,...n}),"WasmTarget"),vm=l((n={})=>new Fa({kind:Dt.WEBGPU,name:"webgpu_generic",vectorWidth:1,numCores:32,maxThreadsPerBlock:256,maxBlockDimX:256,maxBlockDimY:256,maxBlockDimZ:64,maxGridDimX:65535,maxGridDimY:65535,maxGridDimZ:65535,sharedMemoryBytes:16384,warpSize:32,memoryBandwidthGBs:400,computeTFLOPs:8,supportsFloat16:!0,...n}),"WebGPUTarget")});function ml(n,e,t,r,s=!1){if(e.length===0)return"0";if(e.length===1)return t(e[0]);let o=[];for(let i=0;i<e.length;i++){let a=t(e[i]);if(s&&a==="0")continue;let u=n.strides[i];u===1?o.push(a):typeof u=="number"&&u>=0?o.push(`${a} * ${u}`):o.push(`${a} * ${r(n,i)}`)}return o.length===0?"0":o.join(" + ")}var km=M(()=>{"use strict";l(ml,"flattenRowMajorIndex")});function Vi(n){return c$[n]||"f32"}function br(n){if(!n)return"f32";if(n._dtype)return n._dtype;switch(n.type){case"IntImmNode":return"i32";case"FloatImmNode":return"f32";case"LIRFlatLoadNode":return Vi(n.dtype);case"BufferLoadNode":return Vi(n.buffer.dtype);case"CastNode":return Vi(n.toDtype);case"CallExternNode":return Vi(n.dtype||"f32");case"CompareNode":return"i32";case"VariableNode":return Vi(n.dtype||"i32");case"MathOpNode":{let e=br(n.a);if(gr(e))return e;if(n.b){let t=br(n.b);if(gr(t))return t}return e}case"IfThenElseNode":return br(n.thenBody);default:return"f32"}}function si(n){!n||typeof n!="object"||!n.type||(n._dtype=br(n))}function n1(n){return l$.has(n)}var ri,Nm,xp,Sp,Am,gl,Om,l$,c$,vp=M(()=>{"use strict";tr();ri=class n{static{l(this,"LIRNode")}constructor(){this.type=this.constructor.name.replace(/^_+/,""),this._parent=null,this._parentKey=null,this._parentIdx=-1}_setChild(e,t,r=-1){(t instanceof n||t&&t._parent!==void 0)&&(t._parent=this,t._parentKey=e,t._parentIdx=r)}_setChildren(e,t){if(t)for(let r=0;r<t.length;r++){let s=t[r];(s instanceof n||s&&s._parent!==void 0)&&(s._parent=this,s._parentKey=e,s._parentIdx=r)}}replaceWith(e){let t=this._parent;return t?(this._parentIdx>=0?t[this._parentKey][this._parentIdx]=e:t[this._parentKey]=e,e&&(e instanceof n||e._parent!==void 0)&&(e._parent=t,e._parentKey=this._parentKey,e._parentIdx=this._parentIdx),this._parent=null,this._parentKey=null,this._parentIdx=-1,!0):!1}},Nm=class extends ri{static{l(this,"LIRFunc")}constructor(e,t,r,s,o,i,a){super(),this.name=e,this.params=t,this.body=r,this.bufferMap=s,this.shapeParams=o,this.shapeParamMap=i,this.metadata=a,this._setChild("body",r)}},xp=class extends ri{static{l(this,"LIRFlatLoadNode")}constructor(e,t,r){super(),this.buffer=e,this.offsetExpr=t,this.dtype=r,this._setChild("offsetExpr",t)}},Sp=class extends ri{static{l(this,"LIRFlatStoreNode")}constructor(e,t,r,s){super(),this.buffer=e,this.offsetExpr=t,this.value=r,this.dtype=s,this._setChild("offsetExpr",t),this._setChild("value",r)}},Am=class extends ri{static{l(this,"LIRAccumulatorNode")}constructor(e){super(),this.localName=e.localName,this.dtype=e.dtype,this.op=e.op||"+",this.initLoad=e.initLoad,this.loopVar=e.loopVar,this.extent=e.extent,this.loopKind=e.loopKind,this.body=e.body,this.flushStore=e.flushStore,this.initBody=e.initBody||null,this._setChild("initLoad",e.initLoad),this._setChild("body",e.body),this._setChild("flushStore",e.flushStore),this._setChild("initBody",e.initBody||null)}},gl=class extends ri{static{l(this,"LIRBindingsNode")}constructor(e,t){super(),this.bindings=e,this.body=t,this._setChild("body",t)}},Om=class{static{l(this,"LIRMetadata")}constructor(){this.locals=new Map,this.externCalls=new Map,this.memoryLayout={bufferOffsets:new Map,totalBytes:0,alignment:16},this.threadBindings=new Map,this.sharedBuffers=[],this.zeroBuffers=new Set,this.constantBuffers=new Map,this.usedBuffers=new Map,this.allocatedBuffers=new Set,this.paramBuffers=new Set}},l$=new Set(["sqrt","abs","ceil","floor","min","max"]),c$={int32:"i32",index:"i32",i32:"i32",float32:"f32",f32:"f32",f64:"f64",f16:"f32",bf16:"f32",i8:"i32",i16:"i32",i64:"i64",ui8:"i32",bool:"i32"};l(Vi,"normalizeDtype");l(br,"inferDtype");l(si,"annotateDtype");l(n1,"isWasmNativeOp")});function nr(n){return n instanceof Bt||n instanceof ri}function Ae(n,e,t={}){return{key:n,kind:e,array:!!t.array,poly:!!t.poly,bind:!!t.bind,param:!!t.param,region:!!t.region,iterVarBinding:!!t.iterVarBinding,bindingsExpr:!!t.bindingsExpr}}function G0(n){return{kinds:n.kinds||"both",descendParams:n.descendParams===!0,bindVars:n.bindVars!==!1}}function p$(n,e){return e.iterVarBinding?{read(){return(n.iterVars||[]).map(t=>t&&t.binding).filter(nr)},write(t){let r=0;for(let s of n.iterVars||[])s&&nr(s.binding)&&(s.binding=t[r++])}}:e.bindingsExpr?{read(){return(n.bindings||[]).map(t=>t&&t.expr).filter(nr)},write(t){let r=0;for(let s of n.bindings||[])s&&nr(s.expr)&&(s.expr=t[r++])}}:e.poly?{read(){let t=n[e.key];return Array.isArray(t)?t.filter(nr):nr(t)?[t]:[]},write(t){let r=n[e.key];if(Array.isArray(r)){let s=0;n[e.key]=r.map(o=>nr(o)?t[s++]:o)}else t.length&&(n[e.key]=t[0])}}:e.array?{read(){let t=n[e.key];return Array.isArray(t)?t.filter(nr):[]},write(t){let r=n[e.key];if(!Array.isArray(r))return;let s=0;n[e.key]=r.map(o=>nr(o)?t[s++]:o)}}:{read(){return nr(n[e.key])?[n[e.key]]:[]},write(t){t.length&&nr(n[e.key])&&(n[e.key]=t[0])}}}function j0(n,e={}){let t=G0(e),r=V0[n.type];if(r===void 0)throw new Error(`ir_visitor: no child schema for node type '${n.type}'`);let s=[];for(let o of r)o.region||o.param&&!t.descendParams||o.bind&&!t.bindVars||t.kinds!=="both"&&o.kind!==t.kinds||s.push(p$(n,o));return s}function ln(n,e={}){if(!nr(n)||V0[n.type]===void 0)return[];let t=[];for(let r of j0(n,e)){let s=r.read();for(let o=0;o<s.length;o++)t.push(s[o])}return t}function r1(n,e,t,r){n._parent=e,n._parentKey=t,n._parentIdx=r}function f$(n){let e=V0[n.type];if(e)for(let t of e){if(t.iterVarBinding||t.bindingsExpr||t.region)continue;let r=n[t.key];if(Array.isArray(r))for(let s=0;s<r.length;s++)nr(r[s])&&r1(r[s],n,t.key,s);else nr(r)&&r1(r,n,t.key,-1)}}function s1(n,e,t,r,s,o){if(o.stop||!nr(n))return;let i={parent:r,depth:s},a=!0;if(e.pre){let u=e.pre(n,i);if(u===Bm){o.stop=!0;return}u===!1&&(a=!1)}if(a)for(let u of j0(n,t)){let p=u.read();for(let c=0;c<p.length;c++)if(s1(p[c],e,t,n,s+1,o),o.stop)return}e.post&&e.post(n,i)}function qs(n,e,t={}){s1(n,typeof e=="function"?{pre:e}:e||{},G0(t),null,0,{stop:!1})}function yl(n,e,t={}){let r=[];return qs(n,s=>{e(s)&&r.push(s)},t),r}function kp(n,e,t={}){let r=!1;return qs(n,s=>{if(e(s))return r=!0,Bm},t),r}function Gi(n,e,t={}){let r=G0(t);return o1(n,e,r)}function o1(n,e,t){if(!nr(n))return n;let r=!1;for(let o of j0(n,t)){let i=o.read();if(i.length===0)continue;let a=!1,u=new Array(i.length);for(let p=0;p<i.length;p++){let c=o1(i[p],e,t);c!==i[p]&&(a=!0),u[p]=c}a&&(o.write(u),r=!0)}r&&f$(n);let s=e(n);return s??n}var Bm,V0,ts=M(()=>{"use strict";Fe();vp();Bm=Symbol("ir-visitor-stop");l(nr,"isIRNode");l(Ae,"field");V0={VariableNode:[],IntImmNode:[],FloatImmNode:[],SyncThreadsNode:[],MathOpNode:[Ae("a","expr"),Ae("b","expr")],CompareNode:[Ae("a","expr"),Ae("b","expr")],CastNode:[Ae("expr","expr")],CallExternNode:[Ae("args","expr",{array:!0})],BufferLoadNode:[Ae("indices","expr",{array:!0})],IfThenElseNode:[Ae("condition","expr"),Ae("thenBody","stmt"),Ae("elseBody","stmt")],BlockRealizeNode:[Ae("binding","expr",{poly:!0}),Ae("iterVar","expr",{bind:!0})],LIRFlatLoadNode:[Ae("offsetExpr","expr")],PrimFunc:[Ae("params","expr",{array:!0,param:!0}),Ae("body","stmt")],ForNode:[Ae("min","expr"),Ae("extent","expr"),Ae("loopVar","expr",{bind:!0}),Ae("body","stmt")],BlockNode:[Ae("iterVars","expr",{iterVarBinding:!0}),Ae("reads","expr",{region:!0,array:!0}),Ae("writes","expr",{region:!0,array:!0}),Ae("initBody","stmt"),Ae("body","stmt")],SeqNode:[Ae("stmts","stmt",{array:!0})],VecCopyNode:[Ae("dstIndex","expr"),Ae("srcIndex","expr")],LetStmtNode:[Ae("value","expr"),Ae("variable","expr",{bind:!0}),Ae("body","stmt")],AllocateNode:[Ae("body","stmt")],WhileNode:[Ae("condBody","stmt"),Ae("condVar","expr",{bind:!0}),Ae("loopBody","stmt")],EvaluateNode:[Ae("value","expr")],BufferStoreNode:[Ae("indices","expr",{array:!0}),Ae("value","expr")],LIRFunc:[Ae("body","stmt")],LIRFlatStoreNode:[Ae("offsetExpr","expr"),Ae("value","expr")],LIRAccumulatorNode:[Ae("loopVar","expr",{bind:!0}),Ae("extent","expr"),Ae("initLoad","stmt"),Ae("initBody","stmt"),Ae("body","expr"),Ae("flushStore","stmt")],LIRBindingsNode:[Ae("bindings","expr",{bindingsExpr:!0}),Ae("body","stmt")]};l(G0,"normOpts");l(p$,"buildAccessor");l(j0,"childAccessors");l(ln,"irChildNodes");l(r1,"setParent");l(f$,"relink");l(s1,"walkInner");l(qs,"walk");l(yl,"collect");l(kp,"some");l(Gi,"transform");l(o1,"transformInner")});function d$(n,e,t,r){switch(n){case"add":return`(${e} + ${t})`;case"sub":return`(${e} - ${t})`;case"mul":return`(${e} * ${t})`;case"neg":return`(-${e})`;case"div":return r==="js"?`((${e} / ${t}) | 0)`:`(${e} / ${t})`;case"mod":return r==="js"?`((${e} % ${t} + ${t}) % ${t})`:`(${e} % ${t})`;case"ceildiv":return r==="js"?`(((${e} + ${t} - 1) / ${t}) | 0)`:`((${e} + ${t} - 1) / ${t})`;case"max":return r==="js"?`Math.max(${e}, ${t})`:`max(${e}, ${t})`;case"min":return r==="js"?`Math.min(${e}, ${t})`:`min(${e}, ${t})`;default:throw new Error(`emitSymInt: unsupported op '${n}'`)}}function U0(n,e,t="c"){if(typeof n=="number"||!(n instanceof Pe))return String(n);if(n.type==="var")return e({name:qu(n.name)});if(t==="wat")throw new Error("emitSymInt: compound symbolic expressions are not supported on the WASM backend");let r=U0(n.args[0],e,t),s=n.args.length>1?U0(n.args[1],e,t):null;return d$(n.type,r,s,t)}function za(n){let e=n.indexOf(".");if(e<0)return null;let t=n.charCodeAt(e+1)-120;if(t<0||t>2)return null;let r=n.substring(0,e);return r==="threadIdx"?{space:"thread",axis:t}:r==="blockIdx"?{space:"block",axis:t}:null}function Rm(n,e){let t=[e];for(;t.length>0;){let r=t.pop();if(r)switch(r.type){case"SeqNode":for(let s=r.stmts.length-1;s>=0;s--)t.push(r.stmts[s]);continue;case"AllocateNode":n._visitAllocateNode(r),t.push(r.body);continue;case"ForNode":n._visitForNode(r);continue;case"BlockNode":n._visitBlockNode(r);continue;case"IfThenElseNode":n._visitIfStmt(r);continue;case"LetStmtNode":n._visitLetStmtNode(r);continue;case"BufferStoreNode":n._visitBufferStoreNode(r);continue;case"LIRFlatStoreNode":n._visitLIRFlatStore(r);continue;case"LIRBindingsNode":n._visitLIRBindings(r);continue;case"LIRAccumulatorNode":n._visitLIRAccumulator(r);continue;case"WhileNode":n._visitWhileNode(r);continue;case"SyncThreadsNode":n._emitSync();continue;case"VecCopyNode":n._visitVecCopyNode(r);continue;case"EvaluateNode":continue;default:throw new Error(`${n.constructor.name}: unhandled statement node '${r.type}'`)}}}function _l(n,e){let t=[n];for(;t.length>0;){let r=t.pop();if(!(!r||typeof r!="object")&&e(r)!==!1)for(let s of h$){let o=r[s];if(o!=null)if(Array.isArray(o))for(let i=o.length-1;i>=0;i--)t.push(o[i]);else t.push(o)}}}function Mm(n){let e=n;for(;e;){if(e.type==="ForNode"||e.type==="BlockNode"){e=e.body;continue}if(e.type==="BufferStoreNode"||e.type==="LIRFlatStoreNode"){let t=e.value;return t.type==="FloatImmNode"&&t.value===0||t.type==="IntImmNode"&&t.value===0}return!1}return!1}function ji(n,e,t,r,s,o="c"){let i=e.shape[t];if(i instanceof Pe)return U0(i,r,o);if(n&&n.shapeParamMap){let a=n.shapeParamMap.get(`${e.name}:${t}`);if(a)return r(a)}throw new Error(`${s} codegen: missing shape param for ${e.name}:${t}`)}function Im(n){let e=1;for(let t of n.shape)typeof t=="number"&&t>0&&(e*=t);return e}function Ui(n,e,t){let r=[];for(let s=e;s<n.shape.length;s++){let o=n.shape[s];typeof o=="number"&&o>=0?r.push(String(o)):r.push(t(n,s))}return r.length===0?"1":r.join(" * ")}function Em(n,e){let t=n.get(e);if(!t)return 0;let r=0;for(let s of t)s.extent>r&&(r=s.extent);return r}var h$,Np=M(()=>{"use strict";po();l(d$,"symOpToString");l(U0,"emitSymInt");l(za,"parseThreadAxis");l(Rm,"visitStatements");h$=["body","stmts","thenBody","elseBody","loopBody","condBody","initBody"];l(_l,"walkStmtTree");l(Mm,"isZeroFillBody");l(ji,"resolveShapeParam");l(Im,"estimateBufferSize");l(Ui,"dynamicDimProduct");l(Em,"maxBindingExtent")});function Wi(n){let e=n>=0?1:-1,t=Math.abs(n),r=1/(1+Va*t),s=0;for(let o=qi.length-1;o>=0;o--)s=s*r+qi[o];return e*(1-s*r*Math.exp(-t*t))}function Ki(n){return 1-Wi(n)}function cn(n){if(n<.5)return Math.log(Math.PI/Math.abs(Math.sin(Math.PI*n)))-cn(1-n);let e=n-1,t=oi[0];for(let s=1;s<oi.length;s++)t+=oi[s]/(e+s);let r=e+7+.5;return .5*Math.log(2*Math.PI)+(e+.5)*Math.log(r)-r+Math.log(t)}function Hi(n){return n<.5?Math.PI/(Math.sin(Math.PI*n)*Hi(1-n)):Math.exp(cn(n))}function ja(n){let e=0,t=n;for(;t<ii;)e-=1/t,t+=1;let r=1/t,s=r*r;e+=Math.log(t)-.5*r;let o=s;for(let i of Ga)e+=i*o,o*=s;return e}var oi,qi,Va,ii,Ga,Yi=M(()=>{"use strict";oi=[.9999999999998099,676.5203681218851,-1259.1392167224028,771.3234287776531,-176.6150291621406,12.507343278686905,-.13857109526572012,9984369578019572e-21,15056327351493116e-23],qi=[.254829592,-.284496736,1.421413741,-1.453152027,1.061405429],Va=.3275911,ii=6,Ga=[-1/12,1/120,-1/252,1/240,-1/132];l(Wi,"erfScalar");l(Ki,"erfcScalar");l(cn,"lgammaScalar");l(Hi,"gammaScalar");l(ja,"digammaScalar")});function u1(n){return ai[0]=(n&32767)<<13,Ji[0]*=a1,Ji[0]>=65536&&(ai[0]|=2139095040),ai[0]|=(n&32768)<<16,Ji[0]}function l1(n){Ji[0]=n;let e=ai[0]&2147483648,t=(ai[0]^e)>>>0,r;if(t>=1199570944)r=t>2139095040?32256:31744;else if(t<947912704)ai[0]=t,Ji[0]+=.5,r=ai[0]-1056964608&65535;else{let s=t>>13&1;t=t+i1>>>0,t=t+s>>>0,r=t>>13&65535}return(r|e>>>16)&65535}function Lm(n){return ai[0]=n<<16>>>0,Ji[0]}function Cm(n){Ji[0]=n;let e=ai[0];if((e>>>23&255)===255)return e>>>16&65535;let t=32767+(e>>>16&1)>>>0;return e+t>>>0>>>16&65535}function Dm(n,e){return n==="f16"?l1(Number(e)):n==="bf16"?Cm(Number(e)):n==="i64"?typeof e=="bigint"?e:BigInt(Math.trunc(e)):e}function Pm(n,e){return n==="f16"?u1(Number(e)):n==="bf16"?Lm(Number(e)):n==="i64"&&typeof e=="bigint"?Number(e):e}var Ji,ai,i1,a1,q0,$m,bl=M(()=>{"use strict";Ji=new Float32Array(1),ai=new Uint32Array(Ji.buffer),i1=(-112<<23>>>0)+4095>>>0,a1=5192296858534828e18;l(u1,"f16ToF32");l(l1,"f32ToF16");l(Lm,"bf16ToF32");l(Cm,"f32ToBf16");l(Dm,"coerceForStorage");l(Pm,"readFromStorage");q0={F16_ADD_BIAS:i1|0,F16_MAGIC_MUL:a1},$m=globalThis;$m.__mlfw_f16_to_f32=u1;$m.__mlfw_f32_to_f16=l1;$m.__mlfw_bf16_to_f32=Lm;$m.__mlfw_f32_to_bf16=Cm});function p1(n){return`((x_erf => { ${g$} })(${n}))`}function _$(n){return`((x_lg => { const lg = ${f1}; return x_lg < 0.5 ? Math.log(Math.PI / Math.abs(Math.sin(Math.PI * x_lg))) - lg(1 - x_lg) : lg(x_lg); })(${n}))`}function b$(n){return`((x_g => { const lg = ${f1}; return x_g < 0.5 ? Math.PI / (Math.sin(Math.PI * x_g) * Math.exp(lg(1 - x_g))) : Math.exp(lg(x_g)); })(${n}))`}var W0,c1,m$,g$,y$,f1,Fm,d1=M(()=>{"use strict";Fe();tr();km();ts();Np();Yi();bl();W0={f16:"__mlfw_f16_to_f32",bf16:"__mlfw_bf16_to_f32"},c1={f16:"__mlfw_f32_to_f16",bf16:"__mlfw_f32_to_bf16"},m$=qi.slice().reverse().reduce((n,e)=>`(${e} + t * ${n})`,"0"),g$=`const t = 1.0 / (1.0 + ${Va} * Math.abs(x_erf)); const p = t * ${m$}; return (x_erf >= 0 ? 1 : -1) * (1.0 - p * Math.exp(-x_erf * x_erf));`;l(p1,"_erfExpr");y$=oi.map((n,e)=>e===0?`${n}`:`${n} / (zz + ${e})`).join(" + "),f1=`(zg => { const zz = zg - 1; const s = ${y$}; const t = zz + ${7} + 0.5; return ${.5*Math.log(2*Math.PI)} + (zz + 0.5) * Math.log(t) - t + Math.log(s); })`;l(_$,"_lgammaExpr");l(b$,"_gammaExpr");Fm=class{static{l(this,"CPUCodegen")}constructor(e){this.target=e,this._indent=0,this._lines=[],this._loopStack=[]}generate(e){this._indent=0,this._lines=[],this._aliases=new Map,this._accTarget=null,this._accVar=null,this._accCounter=0;let t=e.type==="LIRFunc",r=new Set,s=[];for(let[,c]of e.bufferMap)s.push(c.name),r.add(c.name);for(let c of e.shapeParams)s.push(c.name);this._paramBuffers=r,this._readBuffers=new Set;let o,i,a,u;t?(o=e.metadata.usedBuffers,i=e.metadata.allocatedBuffers,a=e.metadata.zeroBuffers,u=e.metadata.constantBuffers):(o=new Map,i=new Set,this._scanTree(e.body,o,i,this._readBuffers),a=this._findZeroOnlyBuffers(e.body,r),u=this._constantBuffers),this._zeroBuffers=a,this._constantBuffers=u,this._localBuffers=new Set,this._primFunc=e,this._emit(`function ${e.name}(${s.join(", ")}) {`),this._indent++;let p=0;for(let[,c]of o)if(c.poolByteOffset!==void 0&&c.poolByteOffset!==null){let f=c.numel();f>0&&(p=Math.max(p,c.poolByteOffset+f*Rn(c.dtype)))}p>0&&this._emit(`const _mem_pool = new ArrayBuffer(${p});`);for(let[c,f]of o)if(!a.has(c)&&!u.has(c)&&!r.has(c)&&!i.has(c)){let d=f.numel();if(d>0)this._emit(`const ${c} = ${this._allocRhs(f,d)};`);else if(d<0){let h=[];for(let m=0;m<f.shape.length;m++)h.push(typeof f.shape[m]=="number"&&f.shape[m]>=0?String(f.shape[m]):this._resolveShapeParam(f,m));this._emit(`const ${c} = new ${Ph(f.dtype)}(${h.join(" * ")});`)}}for(let[,c]of o)r.has(c.name)||this._localBuffers.add(c.name);return this._visitNode(e.body),this._indent--,this._emit("}"),this._cleanupSource(this._lines.join(`
+`))}_emit(e){this._lines.push("  ".repeat(this._indent)+e)}_allocRhs(e,t){let r=Ph(e.dtype);return e.poolByteOffset!==void 0&&e.poolByteOffset!==null&&t>0?`new ${r}(_mem_pool, ${e.poolByteOffset}, ${t})`:`new ${r}(${t})`}_wrapLoad(e,t){let r=W0[e];return r?`${r}(${t})`:t}_wrapStoreVal(e,t){let r=c1[e];return r?`${r}(${t})`:e==="i64"?`BigInt(${t})`:t}_zeroLit(e){return e==="i64"?"0n":"0"}_visitNode(e){let t=e;for(;t;)switch(t.type){case"SeqNode":for(let r=0;r<t.stmts.length-1;r++)this._visitNode(t.stmts[r]);t=t.stmts[t.stmts.length-1];continue;case"AllocateNode":{let r=t.buffer,s=r.numel();s>0?this._emit(`const ${r.name} = ${this._allocRhs(r,s)};`):s<0&&this._emit(`const ${r.name} = new ${Ph(r.dtype)}(${this._dynamicNumel(r)});`),t=t.body;continue}case"LetStmtNode":this._emit(`const ${t.variable.name} = ${this._exprToJS(t.value)};`),t=t.body;continue;case"ForNode":this._visitForNode(t);return;case"BlockNode":this._visitBlockNode(t);return;case"IfThenElseNode":this._visitIfThenElseStmt(t);return;case"BufferStoreNode":this._visitBufferStoreNode(t);return;case"LIRFlatStoreNode":this._visitLIRFlatStore(t);return;case"LIRBindingsNode":this._visitLIRBindings(t);return;case"LIRAccumulatorNode":this._visitLIRAccumulator(t);return;case"WhileNode":this._visitWhileNode(t);return;case"EvaluateNode":return;case"SyncThreadsNode":return;default:throw new Error(`CPU codegen: unhandled statement node '${t.type}'`)}}_visitForNode(e){if(this._isRedundantZeroFill(e))return;let t=e.loopVar.name;if(e.extent.type==="IntImmNode"&&e.extent.value===1){this._aliases.set(t,"0"),this._visitNode(e.body);return}let r=this._exprToJS(e.extent);if(e.kind===z.UNROLLED){let o=e.extent.type==="IntImmNode"?e.extent.value:null;if(o&&o<=32&&!this._isZeroFillBody(e.body)){for(let i=0;i<o;i++)this._emit("{ const "+t+" = "+i+";"),this._indent++,this._loopStack.push(t),this._visitNode(e.body),this._loopStack.pop(),this._indent--,this._emit("}");return}}let s=this._detectReductionAcc(e);if(s){let o="_acc_"+(this._accCounter=(this._accCounter||0)+1),i=this._accTarget,a=this._accVar;this._accTarget=s,this._accVar=o,this._emit("let "+o+" = "+this._wrapLoad(s.dtype,s.bufName+"["+s.idxExpr+"]")+";"),this._emit("for (let "+t+" = 0; "+t+" < "+r+"; "+t+"++) {"),this._indent++,this._loopStack.push(t),this._visitNode(e.body),this._loopStack.pop(),this._indent--,this._emit("}"),this._emit(s.bufName+"["+s.idxExpr+"] = "+this._wrapStoreVal(s.dtype,o)+";"),this._accTarget=i,this._accVar=a;return}this._emit("for (let "+t+" = 0; "+t+" < "+r+"; "+t+"++) {"),this._indent++,this._loopStack.push(t),this._visitNode(e.body),this._loopStack.pop(),this._indent--,this._emit("}")}_visitLIRFlatStore(e){if(!(this._zeroBuffers&&this._zeroBuffers.has(e.buffer.name))&&!(this._constantBuffers&&this._constantBuffers.has(e.buffer.name))){if(this._accTarget&&e.buffer.name===this._accTarget.bufName&&this._exprToJS(e.offsetExpr)===this._accTarget.idxExpr){this._emit(this._accVar+" = "+this._exprToJS(e.value)+";");return}this._emit(e.buffer.name+"["+this._exprToJS(e.offsetExpr)+"] = "+this._wrapStoreVal(e.dtype||e.buffer.dtype,this._exprToJS(e.value))+";")}}_visitLIRBindings(e){for(let t of e.bindings){let r=this._exprToJS(t.expr);this._aliases.set(t.name,r)}this._visitNode(e.body)}_visitLIRAccumulator(e){let t=e.localName,r=this._exprToJS(e.initLoad);this._emit("let "+t+" = "+r+";");let s=this._accTarget,o=this._accVar;this._accTarget={bufName:e.flushStore.buffer.name,idxExpr:this._exprToJS(e.flushStore.offsetExpr)},this._accVar=t;let i=e.loopVar.name,a=this._exprToJS(e.extent);this._emit("for (let "+i+" = 0; "+i+" < "+a+"; "+i+"++) {"),this._indent++,this._loopStack.push(i);let u=this._exprToJS(e.body),p=e.op||"+",c;p==="max"?c="Math.max("+t+", "+u+")":p==="min"?c="Math.min("+t+", "+u+")":c="("+t+" "+p+" "+u+")",this._emit(t+" = "+c+";"),this._loopStack.pop(),this._indent--,this._emit("}"),this._emit(e.flushStore.buffer.name+"["+this._accTarget.idxExpr+"] = "+this._wrapStoreVal(e.flushStore.dtype||e.flushStore.buffer.dtype,t)+";"),this._accTarget=s,this._accVar=o}_detectReductionAcc(e){let t=e.body;if(!t||t.type!=="BlockNode")return null;let r=t.body;if(!r||r.type!=="BufferStoreNode")return null;let s=r,o=s.value;if(!o||o.type!=="MathOpNode"||o.op!=="+")return null;let i=null;if(o.a&&o.a.type==="BufferLoadNode"&&o.a.buffer.name===s.buffer.name?i=o.a:o.b&&o.b.type==="BufferLoadNode"&&o.b.buffer.name===s.buffer.name&&(i=o.b),!i)return null;for(let c of t.iterVars)c.iterVar&&c.binding&&this._aliases.set(c.iterVar.name,this._exprToJS(c.binding));let a=this._flatIndex(s.buffer,s.indices),u=this._flatIndex(i.buffer,i.indices);if(a!==u)return null;let p=e.loopVar.name;return a.includes(p)?null:{bufName:s.buffer.name,idxExpr:a,dtype:s.buffer.dtype}}_visitBlockNode(e){for(let t of e.iterVars)if(t.iterVar&&t.binding){let r=this._exprToJS(t.binding);this._aliases.set(t.iterVar.name,r)}if(e.initBody){let t=this._loopStack.length>0?this._loopStack[this._loopStack.length-1]:null;t?(this._emit("if ("+t+" === 0) {"),this._indent++,this._visitNode(e.initBody),this._indent--,this._emit("}")):this._visitNode(e.initBody)}this._visitNode(e.body)}_visitIfThenElseStmt(e){this._emit(`if (${this._exprToJS(e.condition)}) {`),this._indent++,this._visitNode(e.thenBody),this._indent--,e.elseBody&&(this._emit("} else {"),this._indent++,this._visitNode(e.elseBody),this._indent--),this._emit("}")}_visitWhileNode(e){this._visitNode(e.condBody),this._emit(`while (${e.condVar.name}[0]) {`),this._indent++,this._visitNode(e.loopBody),this._visitNode(e.condBody),this._indent--,this._emit("}")}_visitBufferStoreNode(e){if(!(this._zeroBuffers&&this._zeroBuffers.has(e.buffer.name))&&!(this._constantBuffers&&this._constantBuffers.has(e.buffer.name))){if(this._accTarget&&e.buffer.name===this._accTarget.bufName&&this._flatIndex(e.buffer,e.indices)===this._accTarget.idxExpr){this._emit(this._accVar+" = "+this._exprToJS(e.value)+";");return}this._emit(e.buffer.name+"["+this._flatIndex(e.buffer,e.indices)+"] = "+this._wrapStoreVal(e.buffer.dtype,this._exprToJS(e.value))+";")}}_exprToJS(e){if(!e)return"0";let t=[],r=[{node:e,phase:0}];for(;r.length>0;){let s=r[r.length-1],o=s.node;if(!o){r.pop(),t.push("0");continue}switch(o.type){case"IntImmNode":r.pop(),t.push(String(o.value));continue;case"FloatImmNode":r.pop(),t.push(String(o.value));continue;case"VariableNode":r.pop(),t.push(this._aliases.get(o.name)||o.name);continue;case"BufferLoadNode":{if(r.pop(),this._zeroBuffers&&this._zeroBuffers.has(o.buffer.name))t.push(this._zeroLit(o.buffer.dtype));else if(this._constantBuffers&&this._constantBuffers.has(o.buffer.name)){let i=this._constantBuffers.get(o.buffer.name);t.push(o.buffer.dtype==="i64"?`BigInt(${i})`:i)}else this._accTarget&&o.buffer.name===this._accTarget.bufName&&this._flatIndex(o.buffer,o.indices)===this._accTarget.idxExpr?t.push(this._accVar):t.push(this._wrapLoad(o.buffer.dtype,o.buffer.name+"["+this._flatIndex(o.buffer,o.indices)+"]"));continue}case"LIRFlatLoadNode":{if(r.pop(),this._zeroBuffers&&this._zeroBuffers.has(o.buffer.name))t.push(this._zeroLit(o.dtype||o.buffer.dtype));else if(this._constantBuffers&&this._constantBuffers.has(o.buffer.name)){let i=String(this._constantBuffers.get(o.buffer.name));t.push((o.dtype||o.buffer.dtype)==="i64"?`BigInt(${i})`:i)}else this._accTarget&&o.buffer.name===this._accTarget.bufName&&this._exprToJS(o.offsetExpr)===this._accTarget.idxExpr?t.push(this._accVar):t.push(this._wrapLoad(o.dtype||o.buffer.dtype,o.buffer.name+"["+this._exprToJS(o.offsetExpr)+"]"));continue}case"MathOpNode":if(s.phase===0)s.phase=1,r.push({node:o.a,phase:0});else if(s.phase===1&&o.b)s.phase=2,r.push({node:o.b,phase:0});else if(r.pop(),!o.b)t.push(`(${o.op}${t.pop()})`);else{let i=t.pop(),a=t.pop();(o.op==="+"||o.op==="-")&&i==="0"?t.push(a):o.op==="+"&&a==="0"?t.push(i):o.op==="*"&&(a==="0"||i==="0")?t.push("0"):o.op==="*"&&i==="1"?t.push(a):o.op==="*"&&a==="1"?t.push(i):o.op==="%"?t.push(`((${a} % ${i} + ${i}) % ${i})`):o.op==="//"?t.push(`((${a} / ${i}) | 0)`):t.push(`(${a} ${o.op} ${i})`)}continue;case"CompareNode":if(s.phase===0)s.phase=1,r.push({node:o.a,phase:0});else if(s.phase===1)s.phase=2,r.push({node:o.b,phase:0});else{r.pop();let i=t.pop(),a=t.pop();t.push(`(${a} ${IN(o.direction)} ${i})`)}continue;case"IfThenElseNode":if(s.phase===0)s.phase=1,r.push({node:o.condition,phase:0});else if(s.phase===1)s.phase=2,r.push({node:o.thenBody,phase:0});else if(s.phase===2)s.phase=3,r.push({node:o.elseBody,phase:0});else{r.pop();let i=t.pop(),a=t.pop(),u=t.pop();t.push(`(${u} ? ${a} : ${i})`)}continue;case"CastNode":if(s.phase===0)s.phase=1,r.push({node:o.expr,phase:0});else{r.pop();let i=t.pop();o.toDtype==="bool"?t.push(`(${i} ? 1 : 0)`):mr(o.toDtype)?t.push(`(${i} | 0)`):W0[o.toDtype]?t.push(`${W0[o.toDtype]}(${c1[o.toDtype]}(${i}))`):t.push(`(+${i})`)}continue;case"CallExternNode":if(s.phase<o.args.length){let i=s.phase;s.phase++,r.push({node:o.args[i],phase:0})}else{r.pop();let i=[];for(let u=0;u<o.args.length;u++)i.unshift(t.pop());let a=i.join(", ");if(EN(o.externName))t.push(`Math.${o.externName}(${a})`);else if(o.externName==="rsqrt")t.push(`(1.0 / Math.sqrt(${a}))`);else if(o.externName==="fmod")t.push(`((${i[0]} % ${i[1]} + ${i[1]}) % ${i[1]})`);else if(o.externName==="exp2")t.push(`Math.pow(2, ${a})`);else if(o.externName==="erf")t.push(p1(a));else if(o.externName==="erfc")t.push(`(1.0 - ${p1(a)})`);else if(o.externName==="lgamma")t.push(_$(a));else if(o.externName==="gamma")t.push(b$(a));else if(o.externName==="log10")t.push(`(Math.log(${a}) * ${1/Math.LN10})`);else throw new Error(`CPU codegen: unsupported extern function "${o.externName}"`)}continue;default:throw new Error(`CPU codegen: unhandled expr node '${o.type}'`)}}return t.length>0?t[0]:"0"}_dynamicNumel(e){return Ui(e,0,(t,r)=>this._resolveShapeParam(t,r))}_flatIndex(e,t){return ml(e,t,r=>this._exprToJS(r),(r,s)=>this._computeDynamicStride(r,s),!0)}_computeDynamicStride(e,t){return Ui(e,t+1,(r,s)=>this._resolveShapeParam(r,s))}_resolveShapeParam(e,t){return ji(this._primFunc,e,t,r=>r.name,"CPU","js")}_cleanupSource(e){let t=e.split(`
+`),r=new Array(t.length).fill(null),s=new Map;for(let i=0;i<t.length;i++){let a=t[i].match(/^\s*const (\w+) = new \w+Array\(\d+\);\s*$/);a&&(r[i]=a[1]);let u=t[i].match(/[A-Za-z_]\w*/g);if(u)for(let p of u)s.set(p,(s.get(p)||0)+1)}let o=[];for(let i=0;i<t.length;i++){if(r[i]!==null&&s.get(r[i])===1)continue;let a=t[i];if(/^\s*\}\s*$/.test(a)&&o.length>0&&/^\s*for\s*\(.*\{\s*$/.test(o[o.length-1])){o.pop();continue}o.push(a)}return o.join(`
+`)}_isRedundantZeroFill(e){let t=e.body;for(;t;){if(t.type==="ForNode"){if(t.extent.type==="IntImmNode"&&t.extent.value===1){t=t.body;continue}return this._isRedundantZeroFill(t)}if(t.type==="BlockNode"){t=t.body;continue}if(t.type==="BufferStoreNode"||t.type==="LIRFlatStoreNode"){let r=t.value;if(!(r.type==="FloatImmNode"&&r.value===0||r.type==="IntImmNode"&&r.value===0)||!t.buffer)return!1;let o=t.buffer.name,i=this._zeroBuffers&&this._zeroBuffers.has(o),a=this._readBuffers&&this._readBuffers.has(o);return i||!a}return!1}return!1}_isZeroFillBody(e){return Mm(e)}_scanTree(e,t,r,s){let o=[e];for(;o.length>0;){let i=o.pop();if(!(!i||typeof i!="object")){switch(i.type){case"BufferStoreNode":case"BufferLoadNode":i.buffer&&t.set(i.buffer.name,i.buffer),i.type==="BufferLoadNode"&&i.buffer&&s&&s.add(i.buffer.name);break;case"AllocateNode":i.buffer&&r.add(i.buffer.name);break}if(i.reads)for(let a of i.reads)a.buffer&&t.set(a.buffer.name,a.buffer);if(i.writes)for(let a of i.writes)a.buffer&&t.set(a.buffer.name,a.buffer);for(let a of ln(i))o.push(a)}}}_findZeroOnlyBuffers(e,t){let r=new Map,s=[e];for(;s.length>0;){let i=s.pop();if(!(!i||typeof i!="object")){if(i.type==="BufferStoreNode"){let a=i.buffer.name;t.has(a)||(r.has(a)||r.set(a,[]),r.get(a).push(i.value))}for(let a of ln(i))s.push(a)}}let o=new Set;this._constantBuffers=new Map;for(let[i,a]of r){if(a.length===0)continue;if(a.every(c=>c.type==="FloatImmNode"&&c.value===0||c.type==="IntImmNode"&&c.value===0)){o.add(i);continue}let u=a[0];(u.type==="FloatImmNode"||u.type==="IntImmNode")&&a.every(c=>c.type===u.type&&c.value===u.value)&&this._constantBuffers.set(i,String(u.value))}return o}}});function m1(n,e){h1.set(n,e)}function g1(n){return h1.get(n)||null}var h1,y1=M(()=>{"use strict";h1=new Map;l(m1,"registerCudaIntrin");l(g1,"getCudaIntrin");m1("wmma_16x16x16_f16f16f32",(n,e)=>{let{M:t,N:r,K:s,a:o,b:i,c:a}=e;n._blockDim=[32,1,1],n._gridDim=[Math.ceil(t/16),Math.ceil(r/16),1],n._emit("const int warpM = blockIdx.x;"),n._emit("const int warpN = blockIdx.y;"),n._emit("fragment<accumulator, 16, 16, 16, float> cf;"),n._emit("fill_fragment(cf, 0.0f);"),n._emit(`for (int kk = 0; kk < ${s}; kk += 16) {`),n._indent++,n._emit("fragment<matrix_a, 16, 16, 16, half, row_major> af;"),n._emit("fragment<matrix_b, 16, 16, 16, half, row_major> bf;"),n._emit(`load_matrix_sync(af, ${o} + warpM * 16 * ${s} + kk, ${s});`),n._emit(`load_matrix_sync(bf, ${i} + kk * ${r} + warpN * 16, ${r});`),n._emit("mma_sync(cf, af, bf, cf);"),n._indent--,n._emit("}"),n._emit(`store_matrix_sync(${a} + warpM * 16 * ${r} + warpN * 16, cf, ${r}, mem_row_major);`)});m1("gemm_pipelined_f32",(n,e)=>{let{M:t,N:r,K:s,a:o,b:i,c:a,tile:u=16}=e;n._blockDim=[u,u,1],n._gridDim=[Math.ceil(r/u),Math.ceil(t/u),1];let p=u;n._emit(`__shared__ float As[2][${p}][${p}];`),n._emit(`__shared__ float Bs[2][${p}][${p}];`),n._emit(`const int row = blockIdx.y * ${p} + threadIdx.y;`),n._emit(`const int col = blockIdx.x * ${p} + threadIdx.x;`),n._emit("float acc = 0.0f;"),n._emit(`const int nTiles = ${s} / ${p};`),n._emit(`__pipeline_memcpy_async(&As[0][threadIdx.y][threadIdx.x], &${o}[row * ${s} + threadIdx.x], sizeof(float));`),n._emit(`__pipeline_memcpy_async(&Bs[0][threadIdx.y][threadIdx.x], &${i}[threadIdx.y * ${r} + col], sizeof(float));`),n._emit("__pipeline_commit();"),n._emit("for (int t = 0; t < nTiles; t++) {"),n._indent++,n._emit("const int cur = t & 1, nxt = (t + 1) & 1;"),n._emit("if (t + 1 < nTiles) {"),n._indent++,n._emit(`__pipeline_memcpy_async(&As[nxt][threadIdx.y][threadIdx.x], &${o}[row * ${s} + (t + 1) * ${p} + threadIdx.x], sizeof(float));`),n._emit(`__pipeline_memcpy_async(&Bs[nxt][threadIdx.y][threadIdx.x], &${i}[((t + 1) * ${p} + threadIdx.y) * ${r} + col], sizeof(float));`),n._emit("__pipeline_commit();"),n._indent--,n._emit("}"),n._emit("__pipeline_wait_prior(t + 1 < nTiles ? 1 : 0);"),n._emit("__syncthreads();"),n._emit(`for (int kk = 0; kk < ${p}; kk++) acc += As[cur][threadIdx.y][kk] * Bs[cur][kk][threadIdx.x];`),n._emit("__syncthreads();"),n._indent--,n._emit("}"),n._emit(`${a}[row * ${r} + col] = acc;`)})});var _1,K0,zm,b1=M(()=>{"use strict";Fe();tr();km();ts();Np();y1();_1=256*1024,K0=class{static{l(this,"CUDAKernel")}constructor(e,t,r,s,o,i,a,u){this.name=e,this.source=t,this.blockDim=r,this.gridDim=s,this.sharedMemBytes=o,this.params=i,this.outputIndices=a,this.scratch=u||[]}},zm=class{static{l(this,"CUDACodegen")}constructor(e){this.target=e,this._indent=0,this._lines=[],this._threadBindings=new Map,this._sharedBuffers=[],this._blockDim=[1,1,1],this._gridDim=[1,1,1],this._defaultDtype="f32",this._storeBuffers=new Set,this._promotedBuffers=new Set,this._promotedBufferDecls=[],this._declaredLocals=new Set,this._needsBarriers=!1,this._globalScratch=[],this._scratchNames=new Set,this._serializeThreads=!1}generate(e){if(this._indent=0,this._lines=[],this._threadBindings.clear(),this._sharedBuffers=[],this._blockDim=[1,1,1],this._gridDim=[1,1,1],this._didParallelReduce=!1,this._primFunc=e,this._storeBuffers=new Set,this._promotedBuffers=new Set,this._promotedBufferDecls=[],this._declaredLocals=new Set,this._needsBarriers=!1,this._globalScratch=[],this._scratchNames=new Set,this._serializeThreads=!1,e.type==="LIRFunc"){for(let[m,g]of e.metadata.threadBindings){this._threadBindings.set(m,g);for(let y of g)y.isDynamic||this._applyBindingDim(m,y.extent)}this._sharedBuffers=e.metadata.sharedBuffers}else this._scanBindings(e.body);this._scanStoreTargets(e.body),this._analyzeSharing(e),this._collectGlobalScratch(e);let r=[],s=[],o=[],i=0;for(let[,m]of e.bufferMap)s.push(m.name),r.push(`${m0(m.dtype)} ${m.name}`),this._defaultDtype=m.dtype,this._storeBuffers.has(m.name)&&o.push(i),i++;for(let m of this._globalScratch)s.push(m.name),r.push(`${m0(m.dtype)} ${m.name}`);for(let m of e.shapeParams)s.push(m.name),r.push(`int ${m.name}`);this._emit(`__global__ void ${e.name}(${r.join(", ")}) {`),this._indent++;let a=new Set;for(let m of this._sharedBuffers){if(a.has(m.name))continue;a.add(m.name);let g=m.numel(),y=m.align16?"__align__(16) ":"";this._emit(`__shared__ ${y}${Zr(m.dtype)} ${m.name}[${g>0?g:1}];`)}for(let m of this._promotedBufferDecls)this._emit(`__shared__ ${Zr(m.dtype)} ${m.name}[${m.size}];`);let u=new Set;if(!this._serializeThreads)for(let[m,g]of this._threadBindings)for(let y of g)u.has(y.varName)||(this._emit(`const int ${y.varName} = ${m};`),u.add(y.varName));if(this._emitMissingLocalDecls(e),e._tensorIntrin){let m=g1(e._tensorIntrin.name);if(!m)throw new Error(`CUDA codegen: unknown tensor intrinsic '${e._tensorIntrin.name}'`);m(this,e._tensorIntrin.info)}else this._visitNode(e.body);this._indent--,this._emit("}");let p=this.target,c=this._serializeThreads?[1,1,1]:[Math.min(this._blockDim[0],p.maxBlockDimX),Math.min(this._blockDim[1],p.maxBlockDimY),Math.min(this._blockDim[2],p.maxBlockDimZ)],f=this._serializeThreads?[1,1,1]:[Math.min(this._gridDim[0],p.maxGridDimX),Math.min(this._gridDim[1],p.maxGridDimY),Math.min(this._gridDim[2],p.maxGridDimZ)],d=c[0]*c[1]*c[2];if(d>p.maxThreadsPerBlock)throw new Error(`[codegen] kernel '${e.name}' block ${c.join("x")} = ${d} threads exceeds maxThreadsPerBlock ${p.maxThreadsPerBlock}`);let h=this._sharedBuffers.reduce((m,g)=>m+Math.max(g.sizeInBytes(),0),0)+this._promotedBufferDecls.reduce((m,g)=>m+Math.max(g.size,0)*Rn(g.dtype),0);if(h>p.sharedMemoryBytes)throw new Error(`[codegen] kernel '${e.name}' shared memory ${h} bytes exceeds device limit ${p.sharedMemoryBytes}`);return new K0(e.name,this._lines.join(`
+`),c,f,h,s,o,this._globalScratch)}_scanBindings(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="ForNode"&&r.kind===z.THREAD_BINDING&&r.threadTag){let s=r.extent.type==="IntImmNode"?r.extent.value:0,o=r.extent.type!=="IntImmNode",i={varName:r.loopVar.name,extent:s,isDynamic:o,extentNode:r.extent};this._threadBindings.has(r.threadTag)?this._threadBindings.get(r.threadTag).push(i):this._threadBindings.set(r.threadTag,[i]),o||this._applyBindingDim(r.threadTag,s)}if(r.type==="AllocateNode"&&r.scope==="shared"&&this._sharedBuffers.push(r.buffer),r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody)}}}_applyBindingDim(e,t){let r=za(e);r&&(r.space==="thread"?this._blockDim[r.axis]=Math.max(this._blockDim[r.axis],t):this._gridDim[r.axis]=Math.max(this._gridDim[r.axis],t))}_emit(e){this._lines.push("  ".repeat(this._indent)+e)}_visitNode(e){Rm(this,e)}_emitSync(){this._emit("__syncthreads();")}_matchFullReduction(e){let t=[],r=e;for(;r&&r.type==="ForNode";){if(r.kind!==z.SERIAL)return null;let d=r.extent.type==="IntImmNode"?r.extent.value:0;if(d<=0)return null;t.push({extC:this._exprToC(r.extent),extVal:d,varName:r.loopVar.name}),r=r.body}if(!r||r.type!=="BlockNode")return null;let s=r,o=s.body;if(!o||o.type!=="BufferStoreNode"||s.iterVars&&s.iterVars.length!==t.length||typeof o.buffer.numel!="function"||o.buffer.numel()!==1)return null;let i=o.value;if(!i||i.type!=="MathOpNode"||i.op!=="+")return null;let a=o.indices.map(d=>this._exprToC(d)).join(","),u=l(d=>d&&d.type==="BufferLoadNode"&&d.buffer===o.buffer&&d.indices.map(h=>this._exprToC(h)).join(",")===a,"isOutLoad"),p=null;if(u(i.a))p=i.b;else if(u(i.b))p=i.a;else return null;let c=new Set(t.map(d=>d.varName));if(a.split(/[^A-Za-z0-9_]/).some(d=>c.has(d)))return null;let f=t.reduce((d,h)=>d*h.extVal,1);return f<2048?null:{loops:t,block:s,store:o,valExpr:p,total:f}}_emitParallelReduction(e,t){this._blockDim=[256,1,1],this._didParallelReduce=!0;let s=Zr(t.store.buffer.dtype),o=this._flatIndex(t.store.buffer,t.store.indices);this._emit(`__shared__ ${s} _redsh[256];`),this._emit(`${s} _racc = 0;`),this._emit(`for (int _rf = threadIdx.x; _rf < ${t.total}; _rf += 256) {`),this._indent++,this._emit("int _rem = _rf;");for(let i=t.loops.length-1;i>=0;i--)this._emit(`const int ${t.loops[i].varName} = _rem % ${t.loops[i].extC}; _rem /= ${t.loops[i].extC};`);for(let i of t.block.iterVars)i.iterVar&&i.binding&&this._emit(`const int ${i.iterVar.name} = ${this._exprToC(i.binding)};`);this._emit(`_racc = _racc + ${this._exprToC(t.valExpr)};`),this._indent--,this._emit("}"),this._emit("_redsh[threadIdx.x] = _racc;"),this._emit("__syncthreads();"),this._emit("for (int _rs = 256 / 2; _rs > 0; _rs >>= 1) {"),this._indent++,this._emit("if (threadIdx.x < _rs) _redsh[threadIdx.x] = _redsh[threadIdx.x] + _redsh[threadIdx.x + _rs];"),this._emit("__syncthreads();"),this._indent--,this._emit("}"),this._emit(`if (threadIdx.x == 0) ${t.store.buffer.name}[${o}] = _redsh[0];`)}_visitForNode(e){if(e.kind===z.THREAD_BINDING&&this._serializeThreads){let s=e.loopVar.name,o=this._exprToC(e.extent);this._emit(`for (int ${s} = 0; ${s} < ${o}; ${s}++) {`),this._indent++,this._visitNode(e.body),this._indent--,this._emit("}");return}if(e.kind===z.THREAD_BINDING){let s=e.extent.type==="IntImmNode"?e.extent.value:0,o=e.threadTag,i=this._getMaxBindingExtent(o);s>0&&i>0&&s<i?(this._emit(`if (${o} < ${s}) {`),this._indent++,this._visitNode(e.body),this._indent--,this._emit("}")):this._visitNode(e.body),this._needsBarriers&&this._emit("__syncthreads();");return}if(this._threadBindings.size===0&&!this._didParallelReduce){let s=this._matchFullReduction(e);if(s){this._emitParallelReduction(e,s);return}}let t=e.loopVar.name,r=this._exprToC(e.extent);(e.kind===z.UNROLLED||e.kind===z.VECTORIZED)&&this._emit("#pragma unroll"),this._emit(`for (int ${t} = 0; ${t} < ${r}; ${t}++) {`),this._indent++,this._visitNode(e.body),this._indent--,this._emit("}")}_visitBlockNode(e){for(let t of e.iterVars)t.iterVar&&t.binding&&this._emit(`const int ${t.iterVar.name} = ${this._exprToC(t.binding)};`);e.initBody&&this._visitNode(e.initBody),this._visitNode(e.body)}_visitAllocateNode(e){if(e.scope!=="shared"){if(this._promotedBuffers.has(e.buffer.name)||this._scratchNames.has(e.buffer.name)||this._declaredLocals.has(e.buffer.name))return;this._declaredLocals.add(e.buffer.name);let t=e.buffer.numel();if(t>0){let r=e.buffer.align16?"__align__(16) ":"";this._emit(`${r}${Zr(e.buffer.dtype)} ${e.buffer.name}[${t}];`)}else this._emit(`${Zr(e.buffer.dtype)}* ${e.buffer.name} = (${Zr(e.buffer.dtype)}*)alloca(${this._dynamicNumel(e.buffer)} * sizeof(${Zr(e.buffer.dtype)}));`)}}_visitIfStmt(e){this._emit(`if (${this._exprToC(e.condition)}) {`),this._indent++,this._visitNode(e.thenBody),this._indent--,e.elseBody&&(this._emit("} else {"),this._indent++,this._visitNode(e.elseBody),this._indent--),this._emit("}")}_visitLetStmtNode(e){let t=e.variable.dtype||this._defaultDtype;this._emit(`${Zr(t)} ${e.variable.name} = ${this._exprToC(e.value)};`),this._visitNode(e.body)}_visitWhileNode(e){this._visitNode(e.condBody);let t=Array.isArray(e.condVar.shape)?`${e.condVar.name}[0]`:e.condVar.name;this._emit(`while (${t}) {`),this._indent++,this._visitNode(e.loopBody),this._visitNode(e.condBody),this._indent--,this._emit("}")}_visitBufferStoreNode(e){this._emit(`${e.buffer.name}[${this._flatIndex(e.buffer,e.indices)}] = ${this._exprToC(e.value)};`)}_visitVecCopyNode(e){let t=`${Zr(e.dstBuffer.dtype)}${e.width}`,r=`${e.dstBuffer.name}[${this._flatIndex(e.dstBuffer,[e.dstIndex])}]`,s=`${e.srcBuffer.name}[${this._flatIndex(e.srcBuffer,[e.srcIndex])}]`;this._emit(`*reinterpret_cast<${t}*>(&${r}) = *reinterpret_cast<const ${t}*>(&${s});`)}_visitLIRFlatStore(e){this._emit(`${e.buffer.name}[${this._exprToC(e.offsetExpr)}] = ${this._exprToC(e.value)};`)}_visitLIRBindings(e){for(let t of e.bindings)this._emit(`const int ${t.name} = ${this._exprToC(t.expr)};`);this._visitNode(e.body)}_visitLIRAccumulator(e){let t=e.localName,r=e.dtype||this._defaultDtype;this._emit(`${Zr(r)} ${t} = ${this._exprToC(e.initLoad)};`);let s=e.loopVar.name,o=this._exprToC(e.extent);this._emit(`for (int ${s} = 0; ${s} < ${o}; ${s}++) {`),this._indent++;let i=this._exprToC(e.body),a=e.op||"+",u;a==="max"||a==="min"?mr(e.dtype)?u=`((${t}) ${a==="max"?">":"<"} (${i}) ? (${t}) : (${i}))`:u=`${a==="max"?"fmaxf":"fminf"}(${t}, ${i})`:u=`(${t} ${a} ${i})`,this._emit(`${t} = ${u};`),this._indent--,this._emit("}"),this._emit(`${e.flushStore.buffer.name}[${this._exprToC(e.flushStore.offsetExpr)}] = ${t};`)}_exprToC(e){if(!e)return"0";switch(e.type){case"IntImmNode":return String(e.value);case"FloatImmNode":return this._emitFloatLiteral(e.value);case"VariableNode":return e.name;case"BufferLoadNode":return`${e.buffer.name}[${this._flatIndex(e.buffer,e.indices)}]`;case"LIRFlatLoadNode":return`${e.buffer.name}[${this._exprToC(e.offsetExpr)}]`;case"MathOpNode":{let t=this._exprToC(e.a);if(!e.b)return`(${e.op}${t})`;let r=this._exprToC(e.b);return e.op==="//"?`(${t} / ${r})`:`(${t} ${e.op} ${r})`}case"CompareNode":return`(${this._exprToC(e.a)} ${Ku(e.direction)} ${this._exprToC(e.b)})`;case"IfThenElseNode":return`(${this._exprToC(e.condition)} ? ${this._exprToC(e.thenBody)} : ${this._exprToC(e.elseBody)})`;case"CastNode":return`((${Zr(e.toDtype)})(${this._exprToC(e.expr)}))`;case"CallExternNode":return this._emitExternCall(e);default:throw new Error(`CUDA codegen: unhandled expr node '${e.type}'`)}}_emitFloatLiteral(e){if(e===1/0)return"INFINITY";if(e===-1/0)return"(-INFINITY)";let t=g0(this._defaultDtype),r=String(e);return`${/[.eEnN]/.test(r)?r:r+".0"}${t}`}_emitExternCall(e){let t=e.args.length,r=new Array(t);for(let a=0;a<t;a++)r[a]=this._exprToC(e.args[a]);let s=r.join(", "),o=e.dtype||this._defaultDtype;if(e.externName==="rsqrt")return`${$h("rsqrt",o)}(${s})`;if(e.externName==="sign"){let a=r[0],u=`0.0${g0(o)}`;return`((${a} > ${u}) - (${a} < ${u}))`}if(e.externName==="min"||e.externName==="max"){if(mr(o)){let a=e.externName==="min"?"<":">";return`((${r[0]}) ${a} (${r[1]}) ? (${r[0]}) : (${r[1]}))`}return`${$h(e.externName,o)}(${s})`}return`${$h(e.externName,o)}(${s})`}_flatIndex(e,t){return ml(e,t,r=>this._exprToC(r),(r,s)=>this._computeDynamicStride(r,s),!1)}_computeDynamicStride(e,t){return Ui(e,t+1,(r,s)=>this._resolveShapeParam(r,s))}_dynamicNumel(e){return Ui(e,0,(t,r)=>this._resolveShapeParam(t,r))}_getMaxBindingExtent(e){return Em(this._threadBindings,e)}_scanStoreTargets(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r){(r.type==="BufferStoreNode"||r.type==="LIRFlatStoreNode")&&this._storeBuffers.add(r.buffer.name),r.type==="VecCopyNode"&&this._storeBuffers.add(r.dstBuffer.name),r.type==="LIRAccumulatorNode"&&r.flushStore&&this._storeBuffers.add(r.flushStore.buffer.name);for(let s of ln(r))t.push(s)}}}_hasCrossBlockGlobalRAW(e){let t=new Set;for(let[,a]of e.bufferMap)t.add(a.name);if(t.size===0)return!1;let r=new Map,s=new Map,o=l((a,u,p)=>{let c=a.get(u);c||(c=new Set,a.set(u,c)),c.add(p)},"record"),i=l((a,u)=>{if(!a||typeof a!="object")return;let p=u;if(a.type==="ForNode"&&a.kind===z.THREAD_BINDING&&a.threadTag){let c=a.extent&&a.extent.type==="IntImmNode"?a.extent.value:0;p=u.concat(`${a.threadTag}:${c}`)}(a.type==="BufferStoreNode"||a.type==="LIRFlatStoreNode")&&a.buffer&&t.has(a.buffer.name)&&o(r,a.buffer.name,[...p].sort().join(",")),(a.type==="BufferLoadNode"||a.type==="LIRFlatLoadNode")&&a.buffer&&t.has(a.buffer.name)&&o(s,a.buffer.name,[...p].sort().join(","));for(let c of ln(a))i(c,p)},"walk");i(e.body,[]);for(let[a,u]of r){let p=s.get(a);if(!p)continue;let c=new Set(u);for(let f of p)c.add(f);if(c.size>1)return!0}return!1}_findThreadPrivateAllocs(e,t){let r=l((s,o)=>{if(!s||typeof s!="object")return;let i=o;if(s.type==="ForNode"&&s.kind===z.THREAD_BINDING&&s.threadTag){let a=za(s.threadTag);a&&a.space==="thread"&&(i=!0)}i&&s.type==="AllocateNode"&&s.buffer&&t.add(s.buffer.name);for(let a of ln(s))r(a,i)},"walk");r(e,!1)}_findCrossThreadBuffers(e){let t=new Set;for(let[,h]of e.bufferMap)t.add(h.name);let r=new Set(this._sharedBuffers.map(h=>h.name)),s=new Set;this._findThreadPrivateAllocs(e.body,s);let o=l(h=>h&&!t.has(h.name)&&!r.has(h.name)&&!s.has(h.name),"isIntermediate"),i=l(h=>typeof h.numel!="function"||h.numel()!==1,"isMulti"),a=new Set,u=new Set,p=new Set,c=new Set,f=l((h,m)=>{if(!h||typeof h!="object")return;let g=m;if(h.type==="ForNode"&&h.kind===z.THREAD_BINDING&&h.threadTag){let y=h.extent&&h.extent.type==="IntImmNode"?h.extent.value:0,_=this._getMaxBindingExtent(h.threadTag);y>0&&_>0&&y<_&&(g=!0)}(h.type==="BufferStoreNode"||h.type==="LIRFlatStoreNode")&&o(h.buffer)&&(i(h.buffer)?a.add(h.buffer.name):g&&p.add(h.buffer.name)),(h.type==="BufferLoadNode"||h.type==="LIRFlatLoadNode")&&o(h.buffer)&&(i(h.buffer)?u.add(h.buffer.name):c.add(h.buffer.name));for(let y of ln(h))f(y,g)},"walk");f(e.body,!1);let d=new Set;for(let h of a)u.has(h)&&d.add(h);for(let h of p)c.has(h)&&d.add(h);return d}_crossThreadBuffersAreBlockLocal(e,t){let r=!0,s=l((o,i)=>{if(!o||typeof o!="object")return;let a=i;if(o.type==="ForNode"&&o.kind===z.THREAD_BINDING&&o.threadTag){let u=za(o.threadTag);u&&u.space==="block"&&(a=!0)}a&&(o.type==="BufferStoreNode"||o.type==="LIRFlatStoreNode")&&o.buffer&&t.has(o.buffer.name)&&(r=!1);for(let u of ln(o))s(u,a)},"walk");return s(e.body,!1),r}_promoteCrossThreadToShared(e,t){let r=new Set;for(let[,p]of e.bufferMap)r.add(p.name);let s=new Map;this._scanBufferRefs(e.body,s);let o=this.target.sharedMemoryBytes||49152,i=new Set(this._sharedBuffers.map(p=>p.name)),a=this._sharedBuffers.reduce((p,c)=>p+Math.max(c.sizeInBytes(),0),0),u=[];for(let p of t){if(r.has(p)||this._promotedBuffers.has(p)||i.has(p))continue;let c=s.get(p);if(!c)return!1;let f=c.numel(),d=f>0?f:this._estimateBufferSize(c);if(d<=0||(a+=d*Rn(c.dtype),a>o))return!1;u.push({name:p,dtype:c.dtype,size:d})}for(let p of u)this._promotedBuffers.add(p.name),this._promotedBufferDecls.push(p);return!0}_analyzeSharing(e){if(e._tensorIntrin){this._needsBarriers=!1;return}if(e.gpuRegisterBlocked){this._needsBarriers=!1;return}let t=this._gridDim[0]*this._gridDim[1]*this._gridDim[2];if(t>1&&this._hasCrossBlockGlobalRAW(e)){this._serializeThreads=!0,this._needsBarriers=!1;return}if(this._threadBindings.size>0){let u=this._blockDim[0]*this._blockDim[1]*this._blockDim[2],p=this._findCrossThreadBuffers(e);if(u*t>1&&p.size>0){if(this._crossThreadBuffersAreBlockLocal(e,p)&&this._promoteCrossThreadToShared(e,p)){this._needsBarriers=!0;return}this._serializeThreads=!0,this._needsBarriers=!1;return}}let r=!1;for(let[u,p]of this._threadBindings){let c=new Set;for(let d of p)d.extent>0&&c.add(d.extent);if(c.size<=1)continue;let f=za(u);if(f&&f.space==="block"){this._serializeThreads=!0,this._needsBarriers=!1;return}r=!0}if(!r)return;this._needsBarriers=!0;let s=new Set;for(let[,u]of e.bufferMap)s.add(u.name);let o=[e.body];for(;o.length>0;){let u=o.pop();if(u){if(u.type==="AllocateNode"&&u.scope!=="shared"&&!s.has(u.buffer.name)){let p=u.buffer.numel(),c=p>0?p:this._estimateBufferSize(u.buffer);c>0&&(this._promotedBuffers.add(u.buffer.name),this._promotedBufferDecls.push({name:u.buffer.name,dtype:u.buffer.dtype,size:c}))}for(let p of ln(u))o.push(p)}}let i=new Map;this._scanBufferRefs(e.body,i);let a=new Set;this._scanAllocateNodes(e.body,a);for(let[u,p]of i){if(s.has(u)||a.has(u)||this._promotedBuffers.has(u))continue;let c=p.numel(),f=c>0?c:this._estimateBufferSize(p);f>0&&(this._promotedBuffers.add(u),this._promotedBufferDecls.push({name:u,dtype:p.dtype,size:f}))}}_emitMissingLocalDecls(e){let t=new Set;for(let[,o]of e.bufferMap)t.add(o.name);let r=new Set;this._scanAllocateNodes(e.body,r);let s=new Map;this._scanBufferRefs(e.body,s);for(let[o,i]of s){if(t.has(o)||r.has(o)||this._promotedBuffers.has(o)||this._scratchNames.has(o)||this._declaredLocals.has(o))continue;let a=i.numel(),u=a>0?a:this._estimateBufferSize(i);u>0&&(this._declaredLocals.add(o),this._emit(`${Zr(i.dtype)} ${o}[${u}];`))}}_collectGlobalScratch(e){let t=this._serializeThreads;if(this._threadBindings.size>0&&!t)return;let r=32768,s=new Set;for(let[,d]of e.bufferMap)s.add(d.name);let o=[],i=new Set,a=l((d,h)=>{if(!h||s.has(d)||this._promotedBuffers.has(d)||this._scratchNames.has(d)||i.has(d))return;let m=typeof h.numel=="function"?h.numel():0,g=m>0?m:this._estimateBufferSize(h);g<=0||(i.add(d),o.push({name:d,dtype:h.dtype,size:g,bytes:g*Rn(h.dtype)}))},"consider"),u=[e.body];for(;u.length>0;){let d=u.pop();if(d){d.type==="AllocateNode"&&d.scope!=="shared"&&a(d.buffer.name,d.buffer);for(let h of ln(d))u.push(h)}}let p=new Map;this._scanBufferRefs(e.body,p);for(let[d,h]of p)a(d,h);let c=l(d=>{this._scratchNames.add(d.name),this._globalScratch.push({name:d.name,dtype:d.dtype,size:d.size})},"offload");if(!t){for(let d of o)d.size>r&&c(d);return}let f=o.reduce((d,h)=>d+h.bytes,0);if(!(f<=_1))for(let d of[...o].sort((h,m)=>m.bytes-h.bytes)){if(f<=_1)break;c(d),f-=d.bytes}}_scanAllocateNodes(e,t){let r=[e];for(;r.length>0;){let s=r.pop();if(s){s.type==="AllocateNode"&&t.add(s.buffer.name);for(let o of ln(s))r.push(o)}}}_scanBufferRefs(e,t){let r=[e];for(;r.length>0;){let s=r.pop();if(s){(s.type==="BufferLoadNode"||s.type==="BufferStoreNode"||s.type==="LIRFlatLoadNode"||s.type==="LIRFlatStoreNode")&&s.buffer&&t.set(s.buffer.name,s.buffer),s.type==="VecCopyNode"&&(s.dstBuffer&&t.set(s.dstBuffer.name,s.dstBuffer),s.srcBuffer&&t.set(s.srcBuffer.name,s.srcBuffer)),s.type==="LIRAccumulatorNode"&&(s.flushStore&&s.flushStore.buffer&&t.set(s.flushStore.buffer.name,s.flushStore.buffer),s.initLoad&&s.initLoad.buffer&&t.set(s.initLoad.buffer.name,s.initLoad.buffer));for(let o of ln(s))r.push(o)}}}_estimateBufferSize(e){return Im(e)}_resolveShapeParam(e,t){return ji(this._primFunc,e,t,r=>r.name,"CUDA","c")}}});var wl,Vm,w1=M(()=>{"use strict";Fe();tr();vp();bl();ts();Np();wl=new Set(["f16","bf16"]),Vm=class{static{l(this,"WasmCodegen")}constructor(e){this.target=e,this._lines=[],this._indent=0,this._locals=new Map,this._localCounter=0,this._imports=new Map,this._bufferOffsets=new Map,this._totalMemBytes=0,this._defaultDtype="f32",this._vectorMode=null,this._vecTmpCounter=0,this._loopVarStack=[]}generate(e){this._lines=[],this._indent=0,this._locals.clear(),this._localCounter=0,this._imports.clear(),this._bufferOffsets.clear(),this._totalMemBytes=0,this._primFunc=e,this._wasmAcc=null,this._waccCounter=0,this._hasParallel=!1,this._parallelExtent=0,this._vectorMode=null,this._vecTmpCounter=0;let t=e.type==="LIRFunc";if(t){for(let[c,f]of e.metadata.memoryLayout.bufferOffsets)this._bufferOffsets.set(c,f);this._totalMemBytes=e.metadata.memoryLayout.totalBytes;for(let[c,f]of e.metadata.externCalls){let d=this._mathImportSig(c,f.argCount);this._imports.set(c,d)}for(let[c,f]of e.metadata.locals)this._ensureLocal(c,Fs(f))}else this._layoutBuffers(e),this._scanMathImports(e.body);this._scanParallel(e.body);let r=[];for(let[,c]of e.bufferMap)r.push(c.name);for(let c of e.shapeParams)r.push(c.name);this._emit("(module"),this._indent++;let s=Math.max(1,Math.ceil(this._totalMemBytes/65536)),o=Math.max(256,s);this._emit(`(memory (export "memory") ${s} ${o})`);for(let[c,f]of this._imports)this._emit(`(import "math" "${c}" (func $math_${c} ${f}))`);let i=[];for(let[,c]of e.bufferMap)i.push("(param i32)");let a=[];for(let c of e.shapeParams)i.push("(param i32)"),this._ensureLocal(c.name,"i32"),a.push(c.name);this._hasParallel&&(i.push("(param i32)"),i.push("(param i32)"),this._ensureLocal("_par_start","i32"),this._ensureLocal("_par_end","i32"),r.push("_par_start"),r.push("_par_end")),this._emit('(func (export "'+e.name+'") '+i.join(" ")),this._indent++,t||(this._prescanLocals(e.body),this._vecTmpCounter=0),this.target.supportsSimd()&&(this._prescanVecLocalsAll(e.body),this._vecTmpCounter=0),this._fixLetStmtLocals(e.body),this._intMinMaxDepth=0,this._prescanIntMinMax(e.body),this._ensureHalfScratch(e);let u=[];for(let[c,f]of this._locals)u.push("(local $"+c+" "+f+")");u.length>0&&this._emit(u.join(" "));let p=e.bufferMap.size;for(let c=0;c<a.length;c++)this._emit("(local.get "+(p+c)+")"),this._emit("local.set $"+a[c]);if(this._hasParallel){let c=p+a.length;this._emit("(local.get "+c+")"),this._emit("local.set $_par_start"),this._emit("(local.get "+(c+1)+")"),this._emit("local.set $_par_end")}return this._visitNode(e.body),this._indent--,this._emit(")"),this._indent--,this._emit(")"),{name:e.name,wat:this._lines.join(`
+`),memoryPages:s,bufferOffsets:new Map(this._bufferOffsets),imports:this._imports,params:r,parallel:this._hasParallel?{extent:this._parallelExtent,outputIndices:this._findOutputIndices(e),poolSafe:this._isParallelSafe(e)}:null}}_isParallelSafe(e){let t=[],r=[e.body];for(;r.length>0;){let p=r.pop();if(p){if(p.type==="ForNode"&&p.kind===z.PARALLEL&&t.push(p),p.body&&r.push(p.body),p.stmts)for(let c of p.stmts)r.push(c);p.thenBody&&r.push(p.thenBody),p.elseBody&&r.push(p.elseBody),p.loopBody&&r.push(p.loopBody)}}if(t.length!==1)return!1;let s=t[0];if(!(e.body&&e.body.stmts?e.body.stmts:[e.body]).includes(s))return!1;let i=l((p,c)=>{let f=[p];for(;f.length>0;){let d=f.pop();if(d){if((d.type==="BufferStoreNode"||d.type==="LIRFlatStoreNode")&&c.add(d),d.type==="LIRAccumulatorNode"&&d.flushStore&&c.add(d.flushStore),d.body&&f.push(d.body),d.stmts)for(let h of d.stmts)f.push(h);d.thenBody&&f.push(d.thenBody),d.elseBody&&f.push(d.elseBody),d.loopBody&&f.push(d.loopBody)}}},"collectStores"),a=new Set,u=new Set;i(e.body,a),i(s.body,u);for(let p of a)if(!u.has(p))return!1;return!0}_ensureLocal(e,t){this._locals.has(e)||this._locals.set(e,t)}_emit(e){this._lines.push("  ".repeat(this._indent)+e)}_emitLoadOp(e){if(wl.has(e)){this._emitHalfDecode(e);return}this._emit(Dh(e))}_emitStoreOp(e){if(wl.has(e)){this._emitHalfEncode(e);return}this._emit(RN(e))}_emitHalfDecode(e){if(this._emit("i32.load16_u"),this._emit("local.set $_half_i"),e==="bf16"){this._emit("(local.get $_half_i)"),this._emit("(i32.const 16)"),this._emit("i32.shl"),this._emit("f32.reinterpret_i32");return}this._emit("(local.get $_half_i)"),this._emit("(i32.const 32767)"),this._emit("i32.and"),this._emit("(i32.const 13)"),this._emit("i32.shl"),this._emit("f32.reinterpret_i32"),this._emit("(f32.const "+q0.F16_MAGIC_MUL+")"),this._emit("f32.mul"),this._emit("local.set $_half_f"),this._emit("(local.get $_half_f)"),this._emit("i32.reinterpret_f32"),this._emit("local.set $_half_i2"),this._emit("(local.get $_half_i2)"),this._emit("(i32.const 2139095040)"),this._emit("i32.or"),this._emit("(local.get $_half_i2)"),this._emit("(local.get $_half_f)"),this._emit("(f32.const 65536)"),this._emit("f32.ge"),this._emit("select"),this._emit("(local.get $_half_i)"),this._emit("(i32.const 32768)"),this._emit("i32.and"),this._emit("(i32.const 16)"),this._emit("i32.shl"),this._emit("i32.or"),this._emit("f32.reinterpret_i32")}_emitHalfEncode(e){if(this._emit("local.set $_half_f"),this._emit("(local.get $_half_f)"),this._emit("i32.reinterpret_f32"),this._emit("local.set $_half_i"),e==="bf16"){this._emit("(local.get $_half_i)"),this._emit("(i32.const 16)"),this._emit("i32.shr_u"),this._emit("(i32.const 1)"),this._emit("i32.and"),this._emit("(i32.const 32767)"),this._emit("i32.add"),this._emit("(local.get $_half_i)"),this._emit("i32.add"),this._emit("(i32.const 16)"),this._emit("i32.shr_u"),this._emit("i32.store16");return}this._emit("(local.get $_half_i)"),this._emit("(i32.const 2147483647)"),this._emit("i32.and"),this._emit("local.set $_half_i2"),this._emit("(i32.const 32256)"),this._emit("(i32.const 31744)"),this._emit("(local.get $_half_i2)"),this._emit("(i32.const 2139095040)"),this._emit("i32.gt_s"),this._emit("select"),this._emit("(local.get $_half_i2)"),this._emit("f32.reinterpret_i32"),this._emit("(f32.const 0.5)"),this._emit("f32.add"),this._emit("i32.reinterpret_f32"),this._emit("(i32.const 1056964608)"),this._emit("i32.sub"),this._emit("(local.get $_half_i2)"),this._emit("(i32.const "+q0.F16_ADD_BIAS+")"),this._emit("i32.add"),this._emit("(local.get $_half_i2)"),this._emit("(i32.const 13)"),this._emit("i32.shr_u"),this._emit("(i32.const 1)"),this._emit("i32.and"),this._emit("i32.add"),this._emit("(i32.const 13)"),this._emit("i32.shr_u"),this._emit("(local.get $_half_i2)"),this._emit("(i32.const 947912704)"),this._emit("i32.lt_s"),this._emit("select"),this._emit("(local.get $_half_i2)"),this._emit("(i32.const 1199570944)"),this._emit("i32.ge_s"),this._emit("select"),this._emit("(local.get $_half_i)"),this._emit("(i32.const 16)"),this._emit("i32.shr_u"),this._emit("(i32.const 32768)"),this._emit("i32.and"),this._emit("i32.or"),this._emit("i32.store16")}_ensureHalfScratch(e){let t=!1;for(let[,r]of e.bufferMap)if(wl.has(r.dtype)){t=!0;break}if(!t&&e.metadata&&e.metadata.locals){for(let[,r]of e.metadata.locals)if(wl.has(r)){t=!0;break}}t||(t=this._treeHasHalf(e.body)),t&&(this._ensureLocal("_half_f","f32"),this._ensureLocal("_half_i","i32"),this._ensureLocal("_half_i2","i32"))}_treeHasHalf(e){let t=[e];for(;t.length>0;){let r=t.pop();if(!(!r||typeof r!="object")){if((r.type==="BufferLoadNode"||r.type==="BufferStoreNode")&&r.buffer&&wl.has(r.buffer.dtype)||(r.type==="LIRFlatLoadNode"||r.type==="LIRFlatStoreNode")&&wl.has(r.dtype))return!0;if(r.body&&t.push(r.body),r.value&&typeof r.value=="object"&&t.push(r.value),r.stmts)for(let s of r.stmts)t.push(s);if(r.a&&t.push(r.a),r.b&&t.push(r.b),r.expr&&t.push(r.expr),r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),r.initBody&&t.push(r.initBody),r.condition&&t.push(r.condition),r.offsetExpr&&t.push(r.offsetExpr),r.args)for(let s of r.args)t.push(s);if(r.indices)for(let s of r.indices)t.push(s)}}return!1}_visitNode(e){let t=e;for(;t;)switch(t.type){case"SeqNode":for(let r=0;r<t.stmts.length-1;r++)this._visitNode(t.stmts[r]);t=t.stmts[t.stmts.length-1];continue;case"AllocateNode":t=t.body;continue;case"LetStmtNode":{let r=this._vectorMode;if(r&&r.simd&&this._dependsOnVecVar(t.value)){let o=t.variable.name+"_vlet";this._ensureLocal(o,"v128"),r.vecLets||(r.vecLets=new Set),r.vecLets.add(t.variable.name),this._emitVecExpr(t.value),this._emit("local.set $"+o),t=t.body;continue}let s=br(t.value)||t.variable.dtype||this._defaultDtype;this._locals.set(t.variable.name,Fs(s)),this._emitCoercedTo(t.value,this._numPrefix(s)),this._emit(`local.set $${t.variable.name}`),t=t.body;continue}case"ForNode":this._visitFor(t);return;case"BlockNode":this._visitBlock(t);return;case"IfThenElseNode":this._visitIf(t);return;case"BufferStoreNode":if(this._vectorMode){this._emitVecStore(t);return}this._visitStore(t);return;case"LIRFlatStoreNode":if(this._vectorMode){this._emitVecFlatStore(t);return}this._visitLIRFlatStore(t);return;case"LIRBindingsNode":this._visitLIRBindings(t);return;case"LIRAccumulatorNode":this._visitLIRAccumulator(t);return;case"WhileNode":this._visitWhile(t);return;case"EvaluateNode":return;case"SyncThreadsNode":return;default:throw new Error(`WASM codegen: unhandled statement node '${t.type}'`)}}_findOutputIndices(e){let t=new Set,r=[e.body];for(;r.length>0;){let i=r.pop();if(i){if((i.type==="BufferStoreNode"||i.type==="LIRFlatStoreNode")&&i.buffer&&t.add(i.buffer.name),i.type==="LIRAccumulatorNode"&&i.flushStore&&i.flushStore.buffer&&t.add(i.flushStore.buffer.name),i.body&&r.push(i.body),i.stmts)for(let a of i.stmts)r.push(a);i.thenBody&&r.push(i.thenBody),i.elseBody&&r.push(i.elseBody),i.loopBody&&r.push(i.loopBody)}}let s=[],o=0;for(let[,i]of e.bufferMap)t.has(i.name)&&s.push(o),o++;return s}_scanParallel(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="ForNode"&&r.kind===z.PARALLEL){this._hasParallel=!0,this._parallelExtent=r.extent&&r.extent.type==="IntImmNode"?r.extent.value:0;return}if(r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody)}}}_visitFor(e){let t=e.loopVar.name,r=this._constExtent(e.extent);if(e.kind===z.PARALLEL){if(r!==null&&this._parallelExtent&&r!==this._parallelExtent){this._emitForLoop(t,e.extent,e.body);return}this._emit("(local.get $_par_start)"),this._emit("local.set $"+t),this._emit("(block $break_"+t),this._indent++,this._emit("(loop $loop_"+t),this._indent++,this._emit("(local.get $"+t+")"),this._emit("(local.get $_par_end)"),this._emit("i32.ge_s"),this._emit("br_if $break_"+t),this._visitNode(e.body),this._emit("(local.get $"+t+")"),this._emit("(i32.const 1)"),this._emit("i32.add"),this._emit("local.set $"+t),this._emit("br $loop_"+t),this._indent--,this._emit(")"),this._indent--,this._emit(")");return}if(e.kind===z.VECTORIZED&&this.target.supportsSimd()&&r!==null){this._visitVectorizedFor(e);return}if((e.kind===z.UNROLLED||e.kind===z.VECTORIZED)&&r!==null&&r<=32&&!this._isZeroFillBody(e.body)){for(let o=0;o<r;o++)this._emit("(i32.const "+o+")"),this._emit("local.set $"+t),this._visitNode(e.body);return}let s=this._detectWasmAcc(e);if(s){let o="_wacc_"+(this._waccCounter=(this._waccCounter||0)+1);this._ensureLocal(o,Fs(s.buf.dtype)),this._emitAddr(s.buf,s.outerIndices),this._emitLoadOp(s.buf.dtype),this._emit("local.set $"+o),this._wasmAcc={local:o,bufName:s.buf.name,indices:s.indices},this._emitForLoop(t,e.extent,e.body),this._emitAddr(s.buf,s.outerIndices),this._emit("(local.get $"+o+")"),this._emitStoreOp(s.buf.dtype),this._wasmAcc=null;return}this._emitForLoop(t,e.extent,e.body)}_visitLIRFlatStore(e){if(this._wasmAcc&&e.buffer.name===this._wasmAcc.bufName){this._emitCoercedTo(e.value,this._numPrefix(e.dtype)),this._emit("local.set $"+this._wasmAcc.local);return}this._emitFlatAddr(e.buffer,e.offsetExpr),this._emitCoercedTo(e.value,this._numPrefix(e.dtype)),this._emitStoreOp(e.dtype)}_visitLIRBindings(e){for(let t of e.bindings)this._emitExpr(t.expr),this._emit(`local.set $${t.name}`);this._visitNode(e.body)}_vecAccumOperandsUnitStride(e){let t=e.loopVar&&e.loopVar.name;if(!t)return!1;let r=l(i=>{let a=[i];for(;a.length>0;){let u=a.pop();if(!(!u||typeof u!="object")){if(u.type==="VariableNode"&&u.name===t)return!0;if(u.a&&a.push(u.a),u.b&&a.push(u.b),u.expr&&a.push(u.expr),u.args)for(let p of u.args)a.push(p);if(u.indices)for(let p of u.indices)a.push(p);u.offsetExpr&&a.push(u.offsetExpr)}}return!1},"usesVar"),s=l(i=>{let a=[i];for(;a.length>0;){let u=a.pop();if(!(!u||typeof u!="object")){if(u.type==="MathOpNode"&&u.op==="*"&&(r(u.a)||r(u.b)))return!0;if(u.a&&a.push(u.a),u.b&&a.push(u.b),u.expr&&a.push(u.expr),u.args)for(let p of u.args)a.push(p)}}return!1},"stridedMul"),o=[e.body];for(;o.length>0;){let i=o.pop();if(!(!i||typeof i!="object")){if(i.type==="BufferLoadNode"&&Array.isArray(i.indices)){for(let a=0;a<i.indices.length-1;a++)if(r(i.indices[a]))return!1}if(i.type==="LIRFlatLoadNode"&&i.offsetExpr&&s(i.offsetExpr))return!1;if(i.a&&o.push(i.a),i.b&&o.push(i.b),i.expr&&o.push(i.expr),i.args)for(let a of i.args)o.push(a);i.body&&o.push(i.body)}}return!0}_accumInstr(e,t){if(e==="*")return"mul";if(e==="max"||e==="min"){if(gr(t))return e;throw new Error(`wasm accumulator: integer ${e} reduction not supported (dtype ${t})`)}return"add"}_visitLIRAccumulator(e){let t=e.localName,r=e.dtype;this._ensureLocal(t,Fs(r));let s=this._constExtent(e.extent),o=e.op||"+",i=o==="+"&&s!==null&&e.loopKind===z.VECTORIZED&&this.target.supportsSimd()?h0(r):null,a=i?this.target.vectorWidth:0;if(i&&s>=a&&this._vecAccumOperandsUnitStride(e)){this._visitVecAccumulator(e,i,a,s);return}this._emitExpr(e.initLoad),this._emit("local.set $"+t);let u=this._wasmAcc;this._wasmAcc={local:t,bufName:e.flushStore.buffer.name};let p=e.loopVar.name;this._emit("(i32.const 0)"),this._emit("local.set $"+p),this._emit("(block $break_"+p),this._indent++,this._emit("(loop $loop_"+p),this._indent++,this._emit("(local.get $"+p+")"),this._emitExpr(e.extent),this._emit("i32.ge_s"),this._emit("br_if $break_"+p),this._emit("(local.get $"+t+")"),this._emitCoercedTo(e.body,this._numPrefix(r)),this._emit(this._numPrefix(r)+"."+this._accumInstr(o,r)),this._emit("local.set $"+t),this._emit("(local.get $"+p+")"),this._emit("(i32.const 1)"),this._emit("i32.add"),this._emit("local.set $"+p),this._emit("br $loop_"+p),this._indent--,this._emit(")"),this._indent--,this._emit(")"),this._emitFlatAddr(e.flushStore.buffer,e.flushStore.offsetExpr),this._emit("(local.get $"+t+")"),this._emitStoreOp(e.flushStore.dtype),this._wasmAcc=u}_visitVecAccumulator(e,t,r,s){let o=e.localName,i=e.dtype,a=e.loopVar.name,u=gr(i),p=u?"f32.add":"i32.add",c=zs(i,"add"),f=Math.floor(s/r)*r,d=f,h=o+"_vec";this._ensureLocal(h,"v128"),this._emitExpr(e.initLoad),this._emit("local.set $"+o),this._emit(u?"(f32.const 0)":"(i32.const 0)"),this._emit(t.splat),this._emit("local.set $"+h);let m=this._wasmAcc;this._wasmAcc={local:o,bufName:e.flushStore.buffer.name},this._vectorMode={dtype:i,lanes:r,loopVar:a,simd:t,laneVars:this._computeLaneVars(e)},this._emit("(i32.const 0)"),this._emit("local.set $"+a),this._emit("(block $vbreak_"+a),this._indent++,this._emit("(loop $vloop_"+a),this._indent++,this._emit("(local.get $"+a+")"),this._emit("(i32.const "+f+")"),this._emit("i32.ge_s"),this._emit("br_if $vbreak_"+a),this._emit("(local.get $"+h+")"),this._emitVecExpr(e.body),this._emit(c),this._emit("local.set $"+h),this._emit("(local.get $"+a+")"),this._emit("(i32.const "+r+")"),this._emit("i32.add"),this._emit("local.set $"+a),this._emit("br $vloop_"+a),this._indent--,this._emit(")"),this._indent--,this._emit(")"),this._vectorMode=null,this._emit("(local.get $"+o+")");for(let g=0;g<r;g++)this._emit("(local.get $"+h+")"),this._emit(t.extractLane+" "+g),this._emit(p);this._emit("local.set $"+o),d<s&&(this._emit("(i32.const "+d+")"),this._emit("local.set $"+a),this._emit("(block $tbreak_"+a),this._indent++,this._emit("(loop $tloop_"+a),this._indent++,this._emit("(local.get $"+a+")"),this._emit("(i32.const "+s+")"),this._emit("i32.ge_s"),this._emit("br_if $tbreak_"+a),this._emit("(local.get $"+o+")"),this._emitCoerced(e.body,u),this._emit(p),this._emit("local.set $"+o),this._emit("(local.get $"+a+")"),this._emit("(i32.const 1)"),this._emit("i32.add"),this._emit("local.set $"+a),this._emit("br $tloop_"+a),this._indent--,this._emit(")"),this._indent--,this._emit(")")),this._emitFlatAddr(e.flushStore.buffer,e.flushStore.offsetExpr),this._emit("(local.get $"+o+")"),this._emitStoreOp(e.flushStore.dtype),this._wasmAcc=m}_emitFlatAddr(e,t){let r=this._bufferOffsets.get(e.name)||0,s=fp(e.dtype||"f32");if(!t||t.type==="IntImmNode"&&t.value===0){this._emit(`(i32.const ${r})`);return}let o=this._vectorMode,i=s+"#"+this._exprKey(t);if(o&&o.addrLocal&&this._exprKey(t)!=="?"){(!o._addrEmitted||o._addrKey!==i)&&(this._emitExpr(t),this._emit(`(i32.const ${s})`),this._emit("i32.mul"),this._emit("local.set $"+o.addrLocal),o._addrEmitted=!0,o._addrKey=i),this._emit("(local.get $"+o.addrLocal+")"),r>0&&(this._emit(`(i32.const ${r})`),this._emit("i32.add"));return}this._emitExpr(t),this._emit(`(i32.const ${s})`),this._emit("i32.mul"),r>0&&(this._emit(`(i32.const ${r})`),this._emit("i32.add"))}_emitForLoop(e,t,r){this._emit("(i32.const 0)"),this._emit("local.set $"+e),this._emit("(block $break_"+e),this._indent++,this._emit("(loop $loop_"+e),this._indent++,this._emit("(local.get $"+e+")"),this._emitExpr(t),this._emit("i32.ge_s"),this._emit("br_if $break_"+e),this._loopVarStack.push(e),this._visitNode(r),this._loopVarStack.pop(),this._emit("(local.get $"+e+")"),this._emit("(i32.const 1)"),this._emit("i32.add"),this._emit("local.set $"+e),this._emit("br $loop_"+e),this._indent--,this._emit(")"),this._indent--,this._emit(")")}_visitBlock(e){for(let t of e.iterVars)if(t.iterVar&&t.binding){this._emitExpr(t.binding),this._emit(`local.set $${t.iterVar.name}`);let r=this._vectorMode;r&&r.laneVars&&this._dependsOnVecVar(t.binding)&&r.laneVars.add(t.iterVar.name)}if(e.initBody){let t=this._loopVarStack.length>0?this._loopVarStack[this._loopVarStack.length-1]:null;t?(this._emit("(local.get $"+t+")"),this._emit("i32.eqz"),this._emit("(if"),this._indent++,this._emit("(then"),this._indent++,this._visitNode(e.initBody),this._indent--,this._emit(")"),this._indent--,this._emit(")")):this._visitNode(e.initBody)}this._visitNode(e.body)}_visitStore(e){if(this._wasmAcc&&this._isAccTarget(e.buffer,e.indices)){this._emitCoercedTo(e.value,this._numPrefix(e.buffer.dtype)),this._emit("local.set $"+this._wasmAcc.local);return}this._emitAddr(e.buffer,e.indices),this._emitCoercedTo(e.value,this._numPrefix(e.buffer.dtype)),this._emitStoreOp(e.buffer.dtype)}_visitIf(e){this._emitExpr(e.condition),this._emit("(if"),this._indent++,this._emit("(then"),this._indent++,this._visitNode(e.thenBody),this._indent--,this._emit(")"),e.elseBody&&(this._emit("(else"),this._indent++,this._visitNode(e.elseBody),this._indent--,this._emit(")")),this._indent--,this._emit(")")}_visitWhile(e){this._visitNode(e.condBody),this._emit("(block $wbreak"),this._indent++,this._emit("(loop $wloop"),this._indent++,this._emitAddr(e.condVar,[]),this._emitLoadOp(e.condVar.dtype),this._emit("i32.eqz"),this._emit("br_if $wbreak"),this._visitNode(e.loopBody),this._visitNode(e.condBody),this._emit("br $wloop"),this._indent--,this._emit(")"),this._indent--,this._emit(")")}_emitAddr(e,t){let r=this._bufferOffsets.get(e.name)||0,s=fp(e.dtype);if(t.length===0){this._emit(`(i32.const ${r})`);return}let o=this._vectorMode;if(o&&o.addrLocal&&t.length>0){let i=s+"#"+t.map(a=>this._exprKey(a)).join(",");(!o._addrEmitted||o._addrKey!==i)&&(this._emitFlatIndex(e,t),this._emit(`(i32.const ${s})`),this._emit("i32.mul"),this._emit("local.set $"+o.addrLocal),o._addrEmitted=!0,o._addrKey=i),this._emit("(local.get $"+o.addrLocal+")"),r>0&&(this._emit(`(i32.const ${r})`),this._emit("i32.add"));return}this._emitFlatIndex(e,t),this._emit(`(i32.const ${s})`),this._emit("i32.mul"),r>0&&(this._emit(`(i32.const ${r})`),this._emit("i32.add"))}_emitFlatIndex(e,t){if(t.length===1){this._emitExpr(t[0]);return}let r=!0;for(let s=0;s<t.length;s++){this._emitExpr(t[s]);let o=e.strides[s];typeof o=="number"&&o>=0?o!==1&&(this._emit(`(i32.const ${o})`),this._emit("i32.mul")):(this._emitDynamicStride(e,s),this._emit("i32.mul")),r||this._emit("i32.add"),r=!1}}_emitDynamicStride(e,t){let r=0;for(let s=t+1;s<e.shape.length;s++){let o=e.shape[s];if(typeof o=="number"&&o>=0)this._emit(`(i32.const ${o})`);else{let i=this._resolveShapeParam(e,s);this._emit(`(local.get $${i})`)}r>0&&this._emit("i32.mul"),r++}r===0&&this._emit("(i32.const 1)")}_resolveShapeParam(e,t){return ji(this._primFunc,e,t,r=>r.name,"WASM","wat")}_emitExpr(e){if(!e){this._emit("(i32.const 0)");return}switch(e.type){case"IntImmNode":this._emit(`(i32.const ${e.value})`);break;case"FloatImmNode":this._emit(`(f32.const ${e.value})`);break;case"VariableNode":this._emit(`(local.get $${e.name})`);break;case"BufferLoadNode":this._wasmAcc&&this._isAccTarget(e.buffer,e.indices)?this._emit("(local.get $"+this._wasmAcc.local+")"):(this._emitAddr(e.buffer,e.indices),this._emitLoadOp(e.buffer.dtype));break;case"LIRFlatLoadNode":this._wasmAcc&&e.buffer.name===this._wasmAcc.bufName?this._emit("(local.get $"+this._wasmAcc.local+")"):(this._emitFlatAddr(e.buffer,e.offsetExpr),this._emitLoadOp(e.dtype));break;case"MathOpNode":this._emitMathOp(e);break;case"CompareNode":this._emitCompare(e);break;case"CastNode":this._emitCast(e);break;case"CallExternNode":this._emitCallExtern(e);break;case"IfThenElseNode":{let t=e._dtype||br(e.thenBody),r=this._numPrefix(t);this._emitExpr(e.condition),gr(this._wasmExprDtype(e.condition))&&(this._emit("(f32.const 0)"),this._emit("f32.ne")),this._emit("(if (result "+r+")"),this._indent++,this._emit("(then"),this._indent++,this._emitCoercedTo(e.thenBody,r),this._indent--,this._emit(")"),this._emit("(else"),this._indent++,this._emitCoercedTo(e.elseBody,r),this._indent--,this._emit(")"),this._indent--,this._emit(")");break}default:this._emit("(i32.const 0)");break}}_numPrefix(e){return e==="f64"?"f64":e==="i64"?"i64":gr(e)?"f32":"i32"}_exprPrefix(e){return this._numPrefix(e&&e._dtype||br(e))}_convertTo(e,t){e!==t&&(t==="f64"?e==="f32"?this._emit("f64.promote_f32"):e==="i32"?this._emit("f64.convert_i32_s"):e==="i64"&&this._emit("f64.convert_i64_s"):t==="f32"?e==="f64"?this._emit("f32.demote_f64"):e==="i32"?this._emit("f32.convert_i32_s"):e==="i64"&&this._emit("f32.convert_i64_s"):t==="i32"?e==="f64"?this._emit("i32.trunc_f64_s"):e==="f32"?this._emit("i32.trunc_f32_s"):e==="i64"&&this._emit("i32.wrap_i64"):t==="i64"&&(e==="i32"?this._emit("i64.extend_i32_s"):e==="f32"?this._emit("i64.trunc_f32_s"):e==="f64"&&this._emit("i64.trunc_f64_s")))}_emitCoercedTo(e,t){this._emitExpr(e),this._convertTo(this._exprPrefix(e),t)}_emitCoerced(e,t){this._emitCoercedTo(e,t?"f32":"i32")}_emitMathOp(e){let t=this._exprPrefix(e.a),r=e.b?this._exprPrefix(e.b):t,s=this._joinPrefix(t,r);if(!e.b){e.op==="-"?s==="f32"||s==="f64"?(this._emitCoercedTo(e.a,s),this._emit(s+".neg")):(this._emit("("+s+".const 0)"),this._emitCoercedTo(e.a,s),this._emit(s+".sub")):e.op==="!"&&(this._emitExpr(e.a),this._emit("i32.eqz"));return}if(e.op==="&&"){this._emitExpr(e.a),this._emitExpr(e.b),this._emit("i32.and");return}if(e.op==="||"){this._emitExpr(e.a),this._emitExpr(e.b),this._emit("i32.or");return}let o=s==="f32"||s==="f64";if(!o){if(e.op==="/"||e.op==="//"){this._emitIntDiv(e,s);return}if(e.op==="%"){this._emitIntRem(e,s);return}}switch(this._emitCoercedTo(e.a,s),this._emitCoercedTo(e.b,s),e.op){case"+":this._emit(`${s}.add`);break;case"-":this._emit(`${s}.sub`);break;case"*":this._emit(`${s}.mul`);break;case"/":this._emit(o?`${s}.div`:`${s}.div_s`);break;case"%":this._emit(`${s}.rem_s`);break;case"//":this._emit(`${s}.div_s`);break;case"<":this._emit(o?`${s}.lt`:`${s}.lt_s`);break;case">":this._emit(o?`${s}.gt`:`${s}.gt_s`);break;case"<=":this._emit(o?`${s}.le`:`${s}.le_s`);break;case">=":this._emit(o?`${s}.ge`:`${s}.ge_s`);break;default:this._emit(`${s}.add`);break}}_joinPrefix(e,t){return e==="f64"||t==="f64"?"f64":e==="f32"||t==="f32"?"f32":e==="i64"||t==="i64"?"i64":"i32"}_emitCompare(e){let t=this._exprPrefix(e.a),r=this._exprPrefix(e.b),s=this._joinPrefix(t,r),o=s==="f32"||s==="f64";this._emitCoercedTo(e.a,s),this._emitCoercedTo(e.b,s);let a={eq:"eq",ne:"ne",lt:o?"lt":"lt_s",le:o?"le":"le_s",gt:o?"gt":"gt_s",ge:o?"ge":"ge_s"}[e.direction];if(!a)throw new Error(`WASM codegen: unhandled compare direction '${e.direction}'`);this._emit(s+"."+a)}_emitCast(e){this._emitExpr(e.expr),this._convertTo(this._numPrefix(e.fromDtype),this._numPrefix(e.toDtype))}_emitCallExtern(e){if((e.externName==="min"||e.externName==="max")&&!gr(e.dtype)){this._emitIntMinMax(e);return}if(e.externName==="abs"&&!gr(e.dtype)){this._emitIntAbs(e);return}let t=new Set(["sqrt","abs","ceil","floor","min","max","rsqrt"]),r=e.externName==="abs"||t.has(e.externName)?this._numPrefix(e.dtype)==="i32"?"f32":this._numPrefix(e.dtype):"f32";if(e.externName==="rsqrt"){this._emit("("+r+".const 1)");for(let s of e.args)this._emitCoercedTo(s,r);this._emit(r+".sqrt"),this._emit(r+".div");return}for(let s of e.args)this._emitCoercedTo(s,r);switch(e.externName){case"sqrt":this._emit(r+".sqrt");break;case"abs":this._emit(r+".abs");break;case"ceil":this._emit(r+".ceil");break;case"floor":this._emit(r+".floor");break;case"min":this._emit(r+".min");break;case"max":this._emit(r+".max");break;default:this._imports.has(e.externName)&&(this._emit(`call $math_${e.externName}`),this._numPrefix(e.dtype)==="f64"&&this._emit("f64.promote_f32"));break}}_emitIntAbs(e){let t=this._intMinMaxEmitDepth||0,r="_iabs"+t;this._intMinMaxEmitDepth=t+1,this._emitExpr(e.args[0]),this._emit("local.set $"+r),this._intMinMaxEmitDepth=t,this._emit("(i32.const 0)"),this._emit("(local.get $"+r+")"),this._emit("i32.sub"),this._emit("(local.get $"+r+")"),this._emit("(local.get $"+r+")"),this._emit("(i32.const 0)"),this._emit("i32.lt_s"),this._emit("select")}_emitIntMinMax(e){let t=this._intMinMaxEmitDepth||0,r="_immm_a"+t,s="_immm_b"+t;this._intMinMaxEmitDepth=t+1,this._emitExpr(e.args[0]),this._emit("local.set $"+r),this._emitExpr(e.args[1]),this._emit("local.set $"+s),this._intMinMaxEmitDepth=t,this._emit("(local.get $"+r+")"),this._emit("(local.get $"+s+")"),this._emit("(local.get $"+r+")"),this._emit("(local.get $"+s+")"),this._emit(e.externName==="min"?"i32.lt_s":"i32.gt_s"),this._emit("select")}_isIntDivNode(e){if(e.type!=="MathOpNode"||!e.b||e.op!=="/"&&e.op!=="//"&&e.op!=="%")return!1;let t=this._joinPrefix(this._exprPrefix(e.a),this._exprPrefix(e.b));return t==="i32"||t==="i64"}_emitIntDiv(e,t){let r=this._intDivEmitDepth||0,s="_idiv_a"+r,o="_idiv_b"+r;this._intDivEmitDepth=r+1,this._emitCoercedTo(e.a,t),this._emit("local.set $"+s),this._emitCoercedTo(e.b,t),this._emit("local.set $"+o),this._intDivEmitDepth=r;let i=t==="i64"?"-9223372036854775808":"-2147483648";this._emit("(local.get $"+s+")"),this._emit("("+t+".const 1)"),this._emit("(local.get $"+o+")"),this._emit("(local.get $"+o+")"),this._emit(t+".eqz"),this._emit("(local.get $"+s+")"),this._emit("("+t+".const "+i+")"),this._emit(t+".eq"),this._emit("(local.get $"+o+")"),this._emit("("+t+".const -1)"),this._emit(t+".eq"),this._emit("i32.and"),this._emit("i32.or"),this._emit("select"),this._emit(t+".div_s"),this._emit("local.set $"+s),this._emit("("+t+".const 0)"),this._emit("(local.get $"+s+")"),this._emit("(local.get $"+o+")"),this._emit(t+".eqz"),this._emit("select")}_emitIntRem(e,t){let r=this._intDivEmitDepth||0,s="_idiv_a"+r,o="_idiv_b"+r;this._intDivEmitDepth=r+1,this._emitCoercedTo(e.a,t),this._emit("local.set $"+s),this._emitCoercedTo(e.b,t),this._emit("local.set $"+o),this._intDivEmitDepth=r,this._emit("(local.get $"+s+")"),this._emit("("+t+".const 1)"),this._emit("(local.get $"+o+")"),this._emit("(local.get $"+o+")"),this._emit(t+".eqz"),this._emit("select"),this._emit(t+".rem_s"),this._emit("local.set $"+s),this._emit("("+t+".const 0)"),this._emit("(local.get $"+s+")"),this._emit("(local.get $"+o+")"),this._emit(t+".eqz"),this._emit("select")}_prescanIntMinMax(e){let t=l((r,s,o)=>{if(!r||typeof r!="object")return;let i=s,a=o;if(r.type==="CallExternNode"&&(r.externName==="min"||r.externName==="max")&&!gr(r.dtype))this._ensureLocal("_immm_a"+s,"i32"),this._ensureLocal("_immm_b"+s,"i32"),s+1>this._intMinMaxDepth&&(this._intMinMaxDepth=s+1),i=s+1;else if(r.type==="CallExternNode"&&r.externName==="abs"&&!gr(r.dtype))this._ensureLocal("_iabs"+s,"i32"),s+1>this._intMinMaxDepth&&(this._intMinMaxDepth=s+1),i=s+1;else if(this._isIntDivNode(r)){let u=Fs(this._joinPrefix(this._exprPrefix(r.a),this._exprPrefix(r.b)));this._ensureLocal("_idiv_a"+o,u),this._ensureLocal("_idiv_b"+o,u),a=o+1}if(r.body&&t(r.body,i,a),r.value&&typeof r.value=="object"&&t(r.value,i,a),r.a&&t(r.a,i,a),r.b&&t(r.b,i,a),r.expr&&t(r.expr,i,a),r.condition&&t(r.condition,i,a),r.offsetExpr&&t(r.offsetExpr,i,a),r.thenBody&&t(r.thenBody,i,a),r.elseBody&&t(r.elseBody,i,a),r.initBody&&t(r.initBody,i,a),r.stmts)for(let u of r.stmts)t(u,i,a);if(r.args)for(let u of r.args)t(u,i,a);if(r.indices)for(let u of r.indices)t(u,i,a);if(r.bindings)for(let u of r.bindings)t(u.expr,i,a);if(r.iterVars)for(let u of r.iterVars)u.binding&&t(u.binding,i,a)},"visit");t(e,0,0)}_mathImportSig(e,t){return`${Array(t).fill("(param f32)").join(" ")} (result f32)`}_constExtent(e){return e.type==="IntImmNode"?e.value:null}_isZeroFillBody(e){return Mm(e)}_collectBindings(e,t){let r=[e];for(;r.length>0;){let s=r.pop();if(!(!s||typeof s!="object")){if(s.type==="BlockNode"&&s.iterVars)for(let o of s.iterVars)o.iterVar&&o.binding&&t.push({name:o.iterVar.name,expr:o.binding});if(s.type==="LIRBindingsNode"&&s.bindings)for(let o of s.bindings)t.push({name:o.name,expr:o.expr});if(s.body&&r.push(s.body),s.stmts)for(let o of s.stmts)r.push(o);s.thenBody&&r.push(s.thenBody),s.elseBody&&r.push(s.elseBody),s.initBody&&r.push(s.initBody),s.loopBody&&r.push(s.loopBody)}}}_computeLaneVars(e){let t=new Set([e.loopVar.name]),r=[];this._collectBindings(e.body,r);let s=l(i=>{let a=[],u=[i];for(;u.length>0;){let p=u.pop();if(!(!p||typeof p!="object")&&(p.type==="VariableNode"&&a.push(p.name),p.a&&u.push(p.a),p.b&&u.push(p.b),p.expr&&u.push(p.expr),p.args))for(let c of p.args)u.push(c)}return a},"varsIn"),o=!0;for(;o;){o=!1;for(let i of r)t.has(i.name)||s(i.expr).some(a=>t.has(a))&&(t.add(i.name),o=!0)}return t}_vecLoadsContiguous(e,t){let r=l(i=>{let a=[i];for(;a.length>0;){let u=a.pop();if(!(!u||typeof u!="object")){if(u.type==="VariableNode"&&t.has(u.name))return!0;if(u.a&&a.push(u.a),u.b&&a.push(u.b),u.expr&&a.push(u.expr),u.args)for(let p of u.args)a.push(p);if(u.indices)for(let p of u.indices)a.push(p);u.offsetExpr&&a.push(u.offsetExpr)}}return!1},"usesLane"),s=l(i=>{let a=[i];for(;a.length>0;){let u=a.pop();if(!(!u||typeof u!="object")){if(u.type==="MathOpNode"&&u.op==="*"&&(r(u.a)||r(u.b)))return!0;if(u.a&&a.push(u.a),u.b&&a.push(u.b),u.expr&&a.push(u.expr),u.args)for(let p of u.args)a.push(p)}}return!1},"stridedMul"),o=[e];for(;o.length>0;){let i=o.pop();if(!(!i||typeof i!="object")){if(i.type==="BufferLoadNode"&&Array.isArray(i.indices)){for(let a=0;a<i.indices.length-1;a++)if(r(i.indices[a]))return!1}if(i.type==="LIRFlatLoadNode"&&i.offsetExpr&&s(i.offsetExpr))return!1;if(i.a&&o.push(i.a),i.b&&o.push(i.b),i.expr&&o.push(i.expr),i.args)for(let a of i.args)o.push(a);if(i.value&&typeof i.value=="object"&&o.push(i.value),i.body&&o.push(i.body),i.stmts)for(let a of i.stmts)o.push(a);i.thenBody&&o.push(i.thenBody),i.elseBody&&o.push(i.elseBody),i.loopBody&&o.push(i.loopBody)}}return!0}_vecStoresLaneIndexed(e,t){let r=[e];for(;r.length>0;){let s=r.pop();if(!s)continue;let o=null;if(s.type==="BufferStoreNode"?o=s.indices:s.type==="LIRFlatStoreNode"&&(o=s.offsetExpr),o!=null){let i=[],a=Array.isArray(o)?[...o]:[o];for(;a.length>0;){let u=a.pop();if(!(!u||typeof u!="object")){if(u.type==="VariableNode"&&i.push(u.name),u.a&&a.push(u.a),u.b&&a.push(u.b),u.expr&&a.push(u.expr),u.args)for(let p of u.args)a.push(p);if(u.indices)for(let p of u.indices)a.push(p);u.offsetExpr&&a.push(u.offsetExpr)}}if(!i.some(u=>t.has(u)))return!1}if(s.body&&r.push(s.body),s.stmts)for(let i of s.stmts)r.push(i);s.thenBody&&r.push(s.thenBody),s.elseBody&&r.push(s.elseBody),s.loopBody&&r.push(s.loopBody)}return!0}_visitVectorizedFor(e){let t=e.loopVar.name,r=this._constExtent(e.extent),s=this._inferBodyDtype(e.body)||this._defaultDtype,o=this.target.vectorWidth,i=h0(s);if(!i||r<o||this._treeHasHalf(e.body)){this._emitForLoop(t,e.extent,e.body);return}let a=this._computeLaneVars(e);if(!this._vecStoresLaneIndexed(e.body,a)||!this._vecLoadsContiguous(e.body,a)){this._emitForLoop(t,e.extent,e.body);return}let u=Math.floor(r/o)*o,p=r-u,c=fp(s),h=this._countBufAccesses(e.body)>=2?"_vaddr_"+t:null;if(u>0&&(this._vectorMode={dtype:s,lanes:o,loopVar:t,simd:i,addrLocal:h,laneVars:a},u===o?(this._emit("(i32.const 0)"),this._emit("local.set $"+t),this._emitVecAddrReset(),this._visitNode(e.body)):(this._emit("(i32.const 0)"),this._emit("local.set $"+t),this._emit("(block $vbreak_"+t),this._indent++,this._emit("(loop $vloop_"+t),this._indent++,this._emit("(local.get $"+t+")"),this._emit("(i32.const "+u+")"),this._emit("i32.ge_s"),this._emit("br_if $vbreak_"+t),this._emitVecAddrReset(),this._visitNode(e.body),this._emit("(local.get $"+t+")"),this._emit("(i32.const "+o+")"),this._emit("i32.add"),this._emit("local.set $"+t),this._emit("br $vloop_"+t),this._indent--,this._emit(")"),this._indent--,this._emit(")")),this._vectorMode=null),p>0)for(let m=u;m<r;m++)this._emit("(i32.const "+m+")"),this._emit("local.set $"+t),this._visitNode(e.body)}_emitVecAddrReset(){this._vectorMode&&this._vectorMode.addrLocal&&(this._vectorMode._addrEmitted=!1,this._vectorMode._addrKey=null)}_countBufAccesses(e){let t=0,r=[e];for(;r.length>0;){let s=r.pop();if(!(!s||typeof s!="object")){if((s.type==="BufferLoadNode"||s.type==="BufferStoreNode"||s.type==="LIRFlatLoadNode"||s.type==="LIRFlatStoreNode")&&t++,s.body&&r.push(s.body),s.value&&typeof s.value=="object"&&r.push(s.value),s.a&&r.push(s.a),s.b&&r.push(s.b),s.stmts)for(let o of s.stmts)r.push(o);if(s.args)for(let o of s.args)r.push(o);if(s.indices)for(let o of s.indices)r.push(o);s.thenBody&&r.push(s.thenBody),s.elseBody&&r.push(s.elseBody),s.expr&&r.push(s.expr),s.condition&&r.push(s.condition),s.offsetExpr&&r.push(s.offsetExpr)}}return t}_inferBodyDtype(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="BufferStoreNode"&&r.buffer)return r.buffer.dtype;if(r.type==="LIRFlatStoreNode")return r.dtype||this._defaultDtype;if(r.type==="BufferLoadNode"&&r.buffer)return r.buffer.dtype;if(r.type==="LIRFlatLoadNode")return r.dtype||this._defaultDtype;if(r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),r.value&&typeof r.value=="object"&&t.push(r.value)}}return null}_emitVecStore(e){let t=this._vectorMode;this._emitAddr(e.buffer,e.indices),this._emitVecExpr(e.value),this._emit(t.simd.vecStore)}_emitVecFlatStore(e){let t=this._vectorMode;this._emitFlatAddr(e.buffer,e.offsetExpr),this._emitVecExpr(e.value),this._emit(t.simd.vecStore)}_dependsOnVecVar(e){let t=this._vectorMode;if(!t)return!0;let r=t.laneVars,s=Array.isArray(e)?[...e]:[e];for(;s.length>0;){let o=s.pop();if(!(!o||typeof o!="object")){if(o.type==="VariableNode"&&(o.name===t.loopVar||r&&r.has(o.name)))return!0;if(o.a&&s.push(o.a),o.b&&s.push(o.b),o.expr&&s.push(o.expr),o.args)for(let i of o.args)s.push(i);if(o.indices)for(let i of o.indices)s.push(i);o.offsetExpr&&s.push(o.offsetExpr)}}return!1}_emitVecExpr(e){if(!e){this._emit("(i32.const 0)");return}let t=this._vectorMode,r=t.dtype;switch(e.type){case"BufferLoadNode":this._dependsOnVecVar(e.indices)?(this._emitAddr(e.buffer,e.indices),this._emit(t.simd.vecLoad)):(this._emitAddr(e.buffer,e.indices),this._emit(Dh(e.buffer.dtype)),this._emit(t.simd.splat));break;case"LIRFlatLoadNode":this._dependsOnVecVar(e.offsetExpr)?(this._emitFlatAddr(e.buffer,e.offsetExpr),this._emit(t.simd.vecLoad)):(this._emitFlatAddr(e.buffer,e.offsetExpr),this._emit(Dh(e.dtype)),this._emit(t.simd.splat));break;case"FloatImmNode":this._emit("(f32.const "+e.value+")"),this._emit(t.simd.splat);break;case"IntImmNode":gr(r)?(this._emit("(f32.const "+e.value+")"),this._emit(t.simd.splat)):(this._emit("(i32.const "+e.value+")"),this._emit(t.simd.splat));break;case"VariableNode":t.vecLets&&t.vecLets.has(e.name)?this._emit("(local.get $"+e.name+"_vlet)"):e.name===t.loopVar?this._emit("(local.get $"+e.name+")"):this._locals.get(e.name)==="f32"?(this._emit("(local.get $"+e.name+")"),this._emit(t.simd.splat)):(this._emit("(local.get $"+e.name+")"),gr(r)&&this._emit("f32.convert_i32_s"),this._emit(t.simd.splat));break;case"MathOpNode":this._emitVecMathOp(e);break;case"CompareNode":this._emitVecCompare(e);break;case"CallExternNode":this._emitVecCallExtern(e);break;case"CastNode":this._emitVecExpr(e.expr),gr(e.toDtype)&&this._isVecMaskExpr(e.expr)&&(this._emit("(f32.const 1)"),this._emit("f32x4.splat"),this._emit("v128.and"));break;case"IfThenElseNode":this._emitVecSelect(e);break;default:this._emitExpr(e);break}}_isVecMaskExpr(e){return e?e.type==="CompareNode"||e.type==="MathOpNode"&&(e.op==="&&"||e.op==="||"||e.op==="!"):!1}_emitVecMathOp(e){let t=this._vectorMode,r=t.dtype;if(!e.b){if(e.op==="-"){let u=zs(r,"neg");u?(this._emitVecExpr(e.a),this._emit(u)):(this._emit("(i32.const 0)"),this._emit(t.simd.splat),this._emitVecExpr(e.a),this._emit(zs(r,"sub")))}else e.op==="!"&&(this._emitVecExpr(e.a),this._emit("v128.not"));return}if(e.op==="&&"){this._emitVecExpr(e.a),this._emitVecExpr(e.b),this._emit("v128.and");return}if(e.op==="||"){this._emitVecExpr(e.a),this._emitVecExpr(e.b),this._emit("v128.or");return}let o={"+":"add","-":"sub","*":"mul","/":"div"}[e.op];if(o){let u=zs(r,o);if(u){this._emitVecExpr(e.a),this._emitVecExpr(e.b),this._emit(u);return}}let a={"<":"lt",">":"gt","<=":"le",">=":"ge"}[e.op];if(a){let u=zs(r,a);if(u){this._emitVecExpr(e.a),this._emitVecExpr(e.b),this._emit(u);return}}this._emitExpr(e),this._emit(t.simd.splat)}_emitVecCompare(e){let t=this._vectorMode,r=t.dtype,s=zs(r,e.direction);s?(this._emitVecExpr(e.a),this._emitVecExpr(e.b),this._emit(s)):(this._emitExpr(e),this._emit(t.simd.splat))}_emitVecCallExtern(e){let t=this._vectorMode,r=t.dtype,s=zs(r,e.externName);if(s){e.externName==="min"||e.externName==="max"?(this._emitVecExpr(e.args[0]),this._emitVecExpr(e.args[1])):this._emitVecExpr(e.args[0]),this._emit(s);return}if(e.externName==="rsqrt"){let o=zs(r,"sqrt");if(o){this._emit("(f32.const 1)"),this._emit(t.simd.splat),this._emitVecExpr(e.args[0]),this._emit(o),this._emit(zs(r,"div"));return}}this._emitScalarizeFallback(e)}_emitVecSelect(e){let t=this._vectorMode;this._emitVecExpr(e.thenBody),this._emitVecExpr(e.elseBody),this._emitVecExpr(e.condition),this._emit(t.simd.bitselect)}_emitScalarizeFallback(e){let t=this._vectorMode,r=t.lanes,s=t.simd.extractLane,o=t.simd.replaceLane,i=t.simd.splat,a="_vtmp_"+this._vecTmpCounter++;this._ensureLocal(a,"v128");let u=[];for(let p=0;p<r;p++){let c="_vl_"+a+"_"+p;this._ensureLocal(c,Fs(t.dtype)),u.push(c)}this._emitVecExpr(e.args[0]),this._emit("local.set $"+a);for(let p=0;p<r;p++){if(this._emit("(local.get $"+a+")"),this._emit(s+" "+p),e.args.length>1){let c="_vtmp2_"+a;if(p===0){this._ensureLocal(c,"v128");let f=this._vectorMode;this._vectorMode=t,this._emitVecExpr(e.args[1]),this._vectorMode=f,this._emit("local.set $"+c)}this._emit("(local.get $"+c+")"),this._emit(s+" "+p)}this._imports.has(e.externName)&&this._emit("call $math_"+e.externName),this._emit("local.set $"+u[p])}this._emit("(local.get $"+u[r-1]+")"),this._emit(i);for(let p=r-2;p>=0;p--)this._emit("(local.get $"+u[p]+")"),this._emit(o+" "+p)}_prescanVecLocalsAll(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="ForNode"&&r.kind===z.VECTORIZED&&(this._prescanVecLocals(r.body),this._countBufAccesses(r.body)>=2&&this._ensureLocal("_vaddr_"+r.loopVar.name,"i32"),this._prescanVecLets(r)),r.type==="LIRAccumulatorNode"&&r.loopKind===z.VECTORIZED&&(this._ensureLocal(r.localName+"_vec","v128"),this._prescanVecLocals(r.body)),r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),r.initBody&&t.push(r.initBody),r.loopBody&&t.push(r.loopBody)}}}_prescanVecLets(e){let t=this._computeLaneVars(e),r=l(o=>{let i=[o];for(;i.length>0;){let a=i.pop();if(!(!a||typeof a!="object")){if(a.type==="VariableNode"&&t.has(a.name))return!0;if(a.a&&i.push(a.a),a.b&&i.push(a.b),a.expr&&i.push(a.expr),a.args)for(let u of a.args)i.push(u);if(a.indices)for(let u of a.indices)i.push(u);a.offsetExpr&&i.push(a.offsetExpr)}}return!1},"dependsOn"),s=[e.body];for(;s.length>0;){let o=s.pop();if(!(!o||typeof o!="object")){if(o.type==="LetStmtNode"&&o.variable&&r(o.value)&&this._ensureLocal(o.variable.name+"_vlet","v128"),o.body&&s.push(o.body),o.stmts)for(let i of o.stmts)s.push(i);o.thenBody&&s.push(o.thenBody),o.elseBody&&s.push(o.elseBody),o.initBody&&s.push(o.initBody)}}}_prescanVecLocals(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="CallExternNode"&&r.externName&&!zs(this._defaultDtype,r.externName)&&r.externName!=="rsqrt"){let o="_vtmp_"+this._vecTmpCounter;this._ensureLocal(o,"v128");let i=this.target.vectorWidth;for(let a=0;a<i;a++)this._ensureLocal("_vl_"+o+"_"+a,Fs(this._defaultDtype));r.args.length>1&&this._ensureLocal("_vtmp2_"+o,"v128"),this._vecTmpCounter++}if(r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);if(r.value&&typeof r.value=="object"&&r.value.type&&t.push(r.value),r.a&&typeof r.a=="object"&&t.push(r.a),r.b&&typeof r.b=="object"&&t.push(r.b),r.args)for(let s of r.args)typeof s=="object"&&t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody)}}}_layoutBuffers(e){let t=0,r=16;this._dynamicBuffers=new Set;let s=l(i=>{t=Math.ceil(t/r)*r,this._bufferOffsets.set(i.name,t);let a=i.numel();!i.shape.some(p=>typeof p!="number"||p<0)&&a>0?t+=a*fp(i.dtype):(this._dynamicBuffers.add(i.name),t+=65536)},"place");for(let[,i]of e.bufferMap)s(i);let o=new Map;this._collectBuffers(e.body,o);for(let[i,a]of o)this._bufferOffsets.has(i)||s(a);this._totalMemBytes=t}_collectBuffers(e,t){let r=[e];for(;r.length>0;){let s=r.pop();if(s){if((s.type==="BufferStoreNode"||s.type==="BufferLoadNode")&&s.buffer&&t.set(s.buffer.name,s.buffer),s.reads)for(let o of s.reads)o.buffer&&t.set(o.buffer.name,o.buffer);if(s.writes)for(let o of s.writes)o.buffer&&t.set(o.buffer.name,o.buffer);for(let o of ln(s))r.push(o)}}}_scanMathImports(e){let t=[e];for(;t.length>0;){let r=t.pop();if(!(!r||typeof r!="object")){if(r.type==="CallExternNode"&&r.externName){let s=r.externName;if(s!=="sqrt"&&s!=="min"&&s!=="max"){let o=this._mathImportSig(s,r.args.length);this._imports.set(s,o)}}if(r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);if(r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),r.initBody&&t.push(r.initBody),r.value&&typeof r.value=="object"&&r.value.type&&t.push(r.value),r.a&&typeof r.a=="object"&&t.push(r.a),r.b&&typeof r.b=="object"&&t.push(r.b),r.expr&&typeof r.expr=="object"&&t.push(r.expr),r.args)for(let s of r.args)typeof s=="object"&&t.push(s);if(r.indices)for(let s of r.indices)typeof s=="object"&&t.push(s);r.condition&&typeof r.condition=="object"&&t.push(r.condition)}}}_prescanLocals(e){this._waccCounter=0;let t=[e];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="ForNode"){this._ensureLocal(r.loopVar.name,"i32");let s=this._accPatternDtype(r);s&&this._ensureLocal("_wacc_"+ ++this._waccCounter,Fs(s)),r.kind===z.VECTORIZED&&this.target.supportsSimd()&&this._prescanVecLocals(r.body)}if(r.type==="BlockNode")for(let s of r.iterVars)s.iterVar&&this._ensureLocal(s.iterVar.name,"i32");if(r.type==="LetStmtNode"&&r.variable){let s=br(r.value)||r.variable.dtype||this._defaultDtype;this._ensureLocal(r.variable.name,Fs(s))}if(r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),r.initBody&&t.push(r.initBody)}}this._waccCounter=0}_wasmExprDtype(e){if(!e||e.type==="CompareNode"||e.type==="MathOpNode"&&(e.op==="!"||e.op==="&&"||e.op==="||"||e.op==="<"||e.op===">"||e.op==="<="||e.op===">="))return"i32";if(e.type==="VariableNode"){let t=this._locals.get(e.name);if(t)return t==="f32"?"f32":"i32"}return br(e)}_fixLetStmtLocals(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="LetStmtNode"&&r.variable&&r.value){let s=br(r.value);s&&this._locals.set(r.variable.name,Fs(s))}if(r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),r.initBody&&t.push(r.initBody)}}}_accPatternDtype(e){let t=e.body;if(!t||t.type!=="BlockNode")return null;let r=t.body;if(!r||r.type!=="BufferStoreNode")return null;let s=r.value;return!s||s.type!=="MathOpNode"||s.op!=="+"?null:s.a&&s.a.type==="BufferLoadNode"&&s.a.buffer.name===r.buffer.name||s.b&&s.b.type==="BufferLoadNode"&&s.b.buffer.name===r.buffer.name?r.buffer.dtype:null}_inferDtype(e){return e&&e._dtype||br(e)}_detectWasmAcc(e){let t=e.body;if(!t||t.type!=="BlockNode")return null;let r=t.body;if(!r||r.type!=="BufferStoreNode")return null;let s=r,o=s.value;if(!o||o.type!=="MathOpNode"||o.op!=="+")return null;let i=null;if(o.a&&o.a.type==="BufferLoadNode"&&o.a.buffer.name===s.buffer.name?i=o.a:o.b&&o.b.type==="BufferLoadNode"&&o.b.buffer.name===s.buffer.name&&(i=o.b),!i)return null;let a=this._indicesKey(s.buffer,s.indices),u=this._indicesKey(i.buffer,i.indices);if(a!==u)return null;let p=s.indices.map(c=>{if(c.type!=="VariableNode")return c;for(let f of t.iterVars)if(f.iterVar&&f.iterVar.name===c.name&&f.binding)return f.binding;return c});return{buf:s.buffer,indices:s.indices,outerIndices:p}}_isAccTarget(e,t){return!this._wasmAcc||e.name!==this._wasmAcc.bufName?!1:this._indicesKey(e,t)===this._indicesKey({name:this._wasmAcc.bufName,shape:e.shape,strides:e.strides},this._wasmAcc.indices)}_indicesKey(e,t){let r=[];for(let s=0;s<t.length;s++)r.push(this._exprKey(t[s]));return e.name+":"+r.join(",")}_exprKey(e){return e?e.type==="VariableNode"?"$"+e.name:e.type==="IntImmNode"?String(e.value):e.type==="MathOpNode"?"("+this._exprKey(e.a)+e.op+(e.b?this._exprKey(e.b):"")+")":"?":"?"}}});var Tl,H0=M(()=>{"use strict";Tl=class{static{l(this,"MinHeap")}_items;_compare;constructor(e){this._items=[],this._compare=e}get size(){return this._items.length}peek(){return this._items.length>0?this._items[0]:null}push(e){let t=this._items;t.push(e);let r=t.length-1;for(;r>0;){let s=r-1>>1;if(this._compare(t[r],t[s])>=0)break;[t[r],t[s]]=[t[s],t[r]],r=s}}pop(){let e=this._items;if(e.length===0)return null;let t=e[0],r=e.pop();if(e.length>0){e[0]=r;let s=0,o=e.length;for(;;){let i=2*s+1,a=2*s+2,u=s;if(i<o&&this._compare(e[i],e[u])<0&&(u=i),a<o&&this._compare(e[a],e[u])<0&&(u=a),u===s)break;[e[s],e[u]]=[e[u],e[s]],s=u}}return t}}});function Zi(n){let e=String(n);return/[.e]/.test(e)?e:`${e}.0`}function T1(n){let e=`(1.0 / (1.0 + ${Zi(Va)} * abs(${n})))`,t=qi.slice().reverse().reduce((r,s)=>`(${Zi(s)} + ${e} * ${r})`,"0.0");return`((select(-1.0, 1.0, ${n} >= 0.0)) * (1.0 - ${e} * ${t} * exp(-${n} * ${n})))`}function Gm(n){let e=`(${n} - 1.0)`,t=oi.map((s,o)=>o===0?Zi(s):`${Zi(s)} / (${e} + ${Zi(o)})`).join(" + "),r=`(${e} + ${Zi(7+.5)})`;return`(${Zi(.5*Math.log(2*Math.PI))} + (${e} + 0.5) * log(${r}) - ${r} + log(${t}))`}function x$(n){let e=`(log(${jm} / abs(sin(${jm} * ${n}))) - ${Gm(`(1.0 - ${n})`)})`;return`(select(${Gm(n)}, ${e}, ${n} < 0.5))`}function S$(n){let e=`(${jm} / (sin(${jm} * ${n}) * exp(${Gm(`(1.0 - ${n})`)})))`;return`(select(exp(${Gm(n)}), ${e}, ${n} < 0.5))`}function x1(n,e){for(let t of v$)n[t]&&e(n[t]);if(n.stmts)for(let t of n.stmts)e(t);if(n.indices)for(let t of n.indices)e(t);if(n.args)for(let t of n.args)e(t);n.initLoad&&e(n.initLoad),n.flushStore&&e(n.flushStore)}var w$,T$,jm,v$,k$,Y0,Um,S1=M(()=>{"use strict";Fe();tr();km();H0();ts();Np();Yi();w$=new Set(["!","&&","||"]),T$=new Set(["<",">","<=",">=","==","!="]);l(Zi,"_wgslFloat");l(T1,"_wgslErf");l(Gm,"_wgslLanczosCore");jm=Zi(Math.PI);l(x$,"_wgslLgamma");l(S$,"_wgslGamma");v$=["body","loopBody","condBody","initBody","thenBody","elseBody","value","a","b","condition","expr","offsetExpr"];l(x1,"walkFullChildren");k$={"threadIdx.x":"local_invocation_id.x","threadIdx.y":"local_invocation_id.y","threadIdx.z":"local_invocation_id.z","blockIdx.x":"workgroup_id.x","blockIdx.y":"workgroup_id.y","blockIdx.z":"workgroup_id.z"},Y0=class{static{l(this,"WebGPUKernel")}constructor(e,t,r,s,o,i,a){this.name=e,this.source=t,this.workgroupSize=r,this.dispatchSize=s,this.sharedMemBytes=o,this.params=i,this.bindings=a}},Um=class{static{l(this,"WebGPUCodegen")}constructor(e){this.target=e,this._indent=0,this._lines=[],this._threadBindings=new Map,this._sharedBuffers=[],this._workgroupSize=[1,1,1],this._dispatchSize=[1,1,1],this._defaultDtype="f32",this._storeBuffers=new Set}generate(e){if(this._indent=0,this._lines=[],this._threadBindings.clear(),this._sharedBuffers=[],this._workgroupSize=[1,1,1],this._dispatchSize=[1,1,1],this._primFunc=e,this._storeBuffers=new Set,this._promotedBuffers=new Set,this._promotedBufferDecls=[],this._wgPoolOffsets=null,this._wgPoolDecls=[],this._needsBarriers=!1,this._serializeThreads=!1,this._localSlots=null,this._slotDecls=[],this._scalarSlotNames=new Set,this._crossThread=null,this._crossExtent=null,e.type==="LIRFunc"){for(let[T,v]of e.metadata.threadBindings){this._threadBindings.set(T,v);for(let A of v)A.isDynamic||this._applyBindingDim(T,A.extent)}this._sharedBuffers=e.metadata.sharedBuffers}else this._scanBindings(e.body);this._scanStoreTargets(e.body),this._analyzeSharing(e);let r=[],s=[],o=0,i=this._checkF16Usage(e);i&&this._emit("enable f16;"),i&&this._emit("");let a=e.bufferMap.size+(e.shapeParams.length>0?1:0),u=[...e.bufferMap.values()],p=a>6&&u.every(T=>T.numel()>0);this._packedMode=p,this._packedOffsets=null;let c=new Map;{let T=0;for(let[,v]of e.bufferMap)c.set(v.name,T++)}{let T=null;for(let[,v]of e.bufferMap)if(this._storeBuffers.has(v.name)){T=v.dtype;break}this._defaultDtype=T||(u.length?u[0].dtype:"f32")}let f={f32:"f32",i32:"i32",u32:"u32",f16:"f16"};if(p){this._packedOffsets=new Map;let T=l(N=>Math.ceil(N/4)*4,"align4"),v=new Map;for(let[,N]of e.bufferMap){r.push(N.name);let k=this._storeBuffers.has(N.name),O=jn(N.dtype),R=(k?"w:":"r:")+O,B=v.get(R);B||(B={isWrite:k,wt:O,dtype:f[O]||N.dtype,bufs:[],off:0},v.set(R,B));let E=N.numel(),F=(k?"_pw_":"_pr_")+O;this._packedOffsets.set(N.name,{storage:F,offset:B.off}),B.bufs.push({name:N.name,offset:B.off,size:E,dtype:N.dtype,argIndex:c.get(N.name)}),B.off+=T(E)}let A=[...v.values()].sort((N,k)=>N.isWrite===k.isWrite?N.wt.localeCompare(k.wt):N.isWrite?1:-1);for(let N of A){let k=(N.isWrite?"_pw_":"_pr_")+N.wt,O=N.isWrite?"storage, read_write":"storage, read";s.push({index:o,name:k,mode:N.isWrite?"read_write":"read",packed:N.bufs,packedSize:N.off,dtype:N.dtype}),this._emit(`@group(0) @binding(${o}) var<${O}> ${k}: array<${N.wt}>;`),o++}}else for(let[,T]of e.bufferMap){r.push(T.name);let v=this._storeBuffers.has(T.name)?"read_write":"read",A=v==="read_write"?"storage, read_write":"storage, read";s.push({index:o,name:T.name,mode:v,dtype:T.dtype,argIndex:c.get(T.name)}),this._emit(`@group(0) @binding(${o}) var<${A}> ${T.name}: array<${jn(T.dtype)}>;`),o++}if(e.shapeParams.length>0){this._emit(""),this._emit("struct ShapeParams {"),this._indent++;for(let T of e.shapeParams)r.push(T.name),this._emit(`${T.name}: u32,`);this._indent--,this._emit("}"),s.push({index:o,name:"_shapes",mode:"read"}),this._emit(`@group(0) @binding(${o}) var<uniform> _shapes: ShapeParams;`),o++}this._emit("");for(let T of this._sharedBuffers){let v=T.numel();this._emit(`var<workgroup> ${T.name}: array<${jn(T.dtype)}, ${v>0?v:1}>;`)}for(let T of this._promotedBufferDecls)this._emit(`var<workgroup> ${T.name}: array<${jn(T.dtype)}, ${T.size}>;`);for(let T of this._wgPoolDecls)this._emit(`var<workgroup> ${T.pool}: array<${jn(T.dtype)}, ${T.size>0?T.size:1}>;`);(this._sharedBuffers.length>0||this._promotedBufferDecls.length>0||this._wgPoolDecls.length>0)&&this._emit("");let d=[],h=!this._serializeThreads&&this._hasBindingPrefix("threadIdx"),m=!this._serializeThreads&&this._hasBindingPrefix("blockIdx");h&&d.push("@builtin(local_invocation_id) _lid: vec3u"),m&&d.push("@builtin(workgroup_id) _wid: vec3u"),!h&&!m&&d.push("@builtin(global_invocation_id) _gid: vec3u");let g=this.target.maxThreadsPerBlock||256;for(;this._workgroupSize[0]*this._workgroupSize[1]*this._workgroupSize[2]>g;){let T=0;this._workgroupSize[1]>this._workgroupSize[T]&&(T=1),this._workgroupSize[2]>this._workgroupSize[T]&&(T=2),this._workgroupSize[T]=Math.max(1,this._workgroupSize[T]>>1)}let y=this._workgroupSize[0],_=this._workgroupSize[1],b=this._workgroupSize[2];if(this._emit(`@compute @workgroup_size(${y}, ${_}, ${b})`),this._emit(`fn ${e.name}(${d.join(", ")}) {`),this._indent++,!this._serializeThreads){let T=new Set;for(let[v,A]of this._threadBindings)if(k$[v]){for(let k of A)if(!T.has(k.varName)){let O=this._wgslBuiltinAccess(v);this._emit(`let ${k.varName}: i32 = i32(${O});`),T.add(k.varName)}}}this._assignLocalSlots(e),this._emitMissingLocalDecls(e),this._visitNode(e.body),this._indent--,this._emit("}");let w=this.target,x=[Math.min(this._workgroupSize[0],w.maxBlockDimX),Math.min(this._workgroupSize[1],w.maxBlockDimY),Math.min(this._workgroupSize[2],w.maxBlockDimZ)],S=[Math.min(this._dispatchSize[0],w.maxGridDimX),Math.min(this._dispatchSize[1],w.maxGridDimY),Math.min(this._dispatchSize[2],w.maxGridDimZ)];return new Y0(e.name,this._lines.join(`
+`),x,S,this._sharedBuffers.reduce((T,v)=>T+Math.max(v.sizeInBytes(),0),0),r,s)}_checkF16Usage(e){for(let[,t]of e.bufferMap)if(t.dtype==="f16")return!0;return!1}_hasBindingPrefix(e){for(let t of this._threadBindings.keys())if(t.startsWith(e))return!0;return!1}_wgslBuiltinAccess(e){let t=e.indexOf(".");if(t<0)return"_gid.x";let r=e.substring(0,t),s=e.substring(t+1);return r==="threadIdx"?`_lid.${s}`:r==="blockIdx"?`_wid.${s}`:`_gid.${s}`}_scanStoreTargets(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r){(r.type==="BufferStoreNode"||r.type==="LIRFlatStoreNode")&&this._storeBuffers.add(r.buffer.name),r.type==="LIRAccumulatorNode"&&r.flushStore&&this._storeBuffers.add(r.flushStore.buffer.name);for(let s of ln(r))t.push(s)}}}_scanBindings(e){_l(e,t=>{if(t.type==="ForNode"&&t.kind===z.THREAD_BINDING&&t.threadTag){let r=t.extent.type==="IntImmNode"?t.extent.value:0,s=t.extent.type!=="IntImmNode",o={varName:t.loopVar.name,extent:r,isDynamic:s,extentNode:t.extent};this._threadBindings.has(t.threadTag)?this._threadBindings.get(t.threadTag).push(o):this._threadBindings.set(t.threadTag,[o]),s||this._applyBindingDim(t.threadTag,r)}t.type==="AllocateNode"&&t.scope==="shared"&&this._sharedBuffers.push(t.buffer)})}_applyBindingDim(e,t){let r=za(e);r&&(r.space==="thread"?this._workgroupSize[r.axis]=Math.max(this._workgroupSize[r.axis],t):this._dispatchSize[r.axis]=Math.max(this._dispatchSize[r.axis],t))}_getMaxBindingExtent(e){return Em(this._threadBindings,e)}_hasRecurrence(e){let t=!1;return _l(e.body,r=>{if(r.type==="SyncThreadsNode")return t=!0,!1}),t}_analyzeSharing(e){let t=new Set;for(let[,f]of this._threadBindings)for(let d of f)d.extent>0&&t.add(d.extent);let r=this._threadBindings.size>0?this._findCrossThreadBuffers(e):new Set,s=this._threadBindings.size>0?this._findCrossExtentBuffers(e):new Set;this._crossThread=r,this._crossExtent=s;let o=this._hasRecurrence(e);if(!o&&t.size<=1&&r.size===0&&s.size===0)return;let i=this.target.sharedMemoryBytes||16384,a=this._sharedBuffers.reduce((f,d)=>f+Math.max(d.sizeInBytes(),0),0),u=new Set;for(let[,f]of e.bufferMap)u.add(f.name);let p=this._collectPromotionCandidates(e,u);p.sort((f,d)=>(r.has(d.name)?1:0)-(r.has(f.name)?1:0));let c=this._dispatchSize[0]*this._dispatchSize[1]*this._dispatchSize[2];if(o){let f=this.target.maxThreadsPerBlock||256,d=this._workgroupSize[0]*this._workgroupSize[1]*this._workgroupSize[2],h=this._packWorkgroupPool(e,p),m=a+h.bytes<=i;if(c===1&&d<=f&&m){this._needsBarriers=!0,this._wgPoolOffsets=h.offsets,this._wgPoolDecls=h.decls;for(let g of p)this._promotedBuffers.add(g.name);return}this._serializeThreads=!0,this._workgroupSize=[1,1,1],this._dispatchSize=[1,1,1],this._needsBarriers=!1}else this._serializeThreads=(r.size>0||s.size>0)&&c>1,this._serializeThreads&&(this._workgroupSize=[1,1,1],this._dispatchSize=[1,1,1]),this._needsBarriers=!this._serializeThreads;for(let f of p){let d=f.size*(Vs(f.dtype)||4);a+d<=i&&(this._promotedBuffers.add(f.name),this._promotedBufferDecls.push(f),a+=d)}}_collectPromotionCandidates(e,t){let r=[];_l(e.body,a=>{if(a.type==="AllocateNode"&&a.scope!=="shared"&&!t.has(a.buffer.name)){let u=a.buffer.numel(),p=u>0?u:this._estimateBufferSize(a.buffer);p>0&&r.push({name:a.buffer.name,dtype:a.buffer.dtype,size:p})}});let s=new Map;this._scanBufferRefs(e.body,s);let o=new Set;this._scanAllocateNodes(e.body,o);let i=new Set(r.map(a=>a.name));for(let[a,u]of s){if(t.has(a)||o.has(a)||i.has(a))continue;let p=u.numel(),c=p>0?p:this._estimateBufferSize(u);c>0&&(r.push({name:a,dtype:u.dtype,size:c}),i.add(a))}return r}_findRecurrenceBody(e){let t=null;return _l(e,r=>{if(!t&&r.type==="ForNode"&&r.kind===z.RECURRENCE)return t=r.body,!1}),t}_namesTouchedOutside(e,t,r){let s=new Set,o=l(i=>{if(!(!i||i===t)){(i.type==="BufferLoadNode"||i.type==="BufferStoreNode"||i.type==="LIRFlatLoadNode"||i.type==="LIRFlatStoreNode")&&i.buffer&&r.has(i.buffer.name)&&s.add(i.buffer.name),i.type==="LIRAccumulatorNode"&&(i.flushStore&&i.flushStore.buffer&&r.has(i.flushStore.buffer.name)&&s.add(i.flushStore.buffer.name),i.initLoad&&i.initLoad.buffer&&r.has(i.initLoad.buffer.name)&&s.add(i.initLoad.buffer.name));for(let a of["stmts","body","initBody","condBody","loopBody","thenBody","elseBody","value","a","b","condition","expr","offsetExpr","extent","indices","args"]){let u=i[a];if(u)if(Array.isArray(u))for(let p of u)o(p);else typeof u=="object"&&o(u)}}},"walk");return o(e),s}_packWorkgroupPool(e,t){let r=new Set(t.map(d=>d.name)),s=this._findRecurrenceBody(e.body),{minPos:o,maxPos:i}=s?this._livenessWalk(s,r):{minPos:new Map,maxPos:new Map},a=s?this._namesTouchedOutside(e.body,s,r):r,u=new Map;for(let d of t){u.has(d.dtype)||u.set(d.dtype,[]);let h=a.has(d.name)||!o.has(d.name),m=h?0:o.get(d.name),g=h?Number.MAX_SAFE_INTEGER:i.get(d.name);u.get(d.dtype).push({name:d.name,size:d.size,first:m,last:g})}let p=new Map,c=[],f=0;for(let[d,h]of u){h.sort((_,b)=>_.first-b.first||b.size-_.size);let m=[],g=`_wg_${jn(d)}`,y=0;for(let _ of h){let b=m.filter(x=>x.first<=_.last&&_.first<=x.last).map(x=>[x.offset,x.offset+x.size]).sort((x,S)=>x[0]-S[0]),w=0;for(let[x,S]of b){if(w+_.size<=x)break;w<S&&(w=S)}m.push({offset:w,size:_.size,first:_.first,last:_.last}),p.set(_.name,{pool:g,offset:w}),w+_.size>y&&(y=w+_.size)}c.push({pool:g,dtype:d,size:y}),f+=y*(Vs(d)||4)}return{offsets:p,decls:c,bytes:f}}_findCrossThreadBuffers(e){let t=new Set;for(let[,a]of e.bufferMap)t.add(a.name);let r=new Set,s=l((a,u)=>{if(u.length===0)return!1;let p=new Set;for(let c of a)this._collectVarNames(c,p);return u.some(c=>p.has(c))},"indexUsesLoopVar"),o=l((a,u)=>{let p=new Set;this._collectVarNames(a,p);for(let c of p)if(u.includes(c))return!0;return!1},"aliasOf"),i=l((a,u)=>{if(!a)return;let p=u;if((a.type==="ForNode"&&a.kind!==z.THREAD_BINDING||a.type==="WhileNode"||a.type==="LIRAccumulatorNode")&&a.loopVar&&(p=[...u,a.loopVar.name]),a.type==="LIRBindingsNode"&&a.bindings)for(let f of a.bindings)!p.includes(f.name)&&o(f.expr,p)&&(p=[...p,f.name]);if(a.type==="BlockNode"&&a.iterVars)for(let f of a.iterVars)f.iterVar&&f.binding&&!p.includes(f.iterVar.name)&&o(f.binding,p)&&(p=[...p,f.iterVar.name]);a.type==="LetStmtNode"&&a.variable&&!p.includes(a.variable.name)&&o(a.value,p)&&(p=[...p,a.variable.name]),a.type==="BufferLoadNode"&&a.buffer&&!t.has(a.buffer.name)&&s(a.indices||[],p)&&r.add(a.buffer.name),a.type==="LIRFlatLoadNode"&&a.buffer&&!t.has(a.buffer.name)&&a.offsetExpr&&s([a.offsetExpr],p)&&r.add(a.buffer.name),x1(a,f=>i(f,p))},"walk");return i(e.body,[]),r}_findCrossExtentBuffers(e){let t=new Set;for(let[,c]of e.bufferMap)t.add(c.name);let r=new Map,s=new Map,o=new Set,i=l(c=>c&&(c.type==="FloatImmNode"||c.type==="IntImmNode"),"isConst"),a=l((c,f,d)=>{if(t.has(f))return;let h=c.get(f);h||(h=new Set,c.set(f,h)),h.add(d)},"record"),u=l((c,f)=>{if(!c)return;let d=f;if(c.type==="ForNode"&&c.kind===z.THREAD_BINDING){let h=c.extent&&c.extent.type==="IntImmNode"?c.extent.value:0;h>0&&(d=f*h)}c.type==="BufferStoreNode"&&c.buffer&&(a(r,c.buffer.name,d),i(c.value)||o.add(c.buffer.name)),c.type==="LIRFlatStoreNode"&&c.buffer&&(a(r,c.buffer.name,d),i(c.value)||o.add(c.buffer.name)),c.type==="LIRAccumulatorNode"&&c.flushStore&&c.flushStore.buffer&&(a(r,c.flushStore.buffer.name,d),o.add(c.flushStore.buffer.name)),c.type==="BufferLoadNode"&&c.buffer&&a(s,c.buffer.name,d),c.type==="LIRFlatLoadNode"&&c.buffer&&a(s,c.buffer.name,d),x1(c,h=>u(h,d))},"walk");u(e.body,1);let p=new Set;for(let[c,f]of s){if(!o.has(c))continue;let d=r.get(c);if(d){for(let h of f)if(!d.has(h)){p.add(c);break}}}return p}_collectVarNames(e,t){let r=[e];for(;r.length>0;){let s=r.pop();if(!(!s||typeof s!="object")){if(s.type==="VariableNode"){t.add(s.name);continue}for(let o of["a","b","condition","thenBody","elseBody","expr","offsetExpr"])s[o]&&r.push(s[o]);if(s.indices)for(let o of s.indices)r.push(o);if(s.args)for(let o of s.args)r.push(o)}}}_emit(e){this._lines.push("  ".repeat(this._indent)+e)}_visitNode(e){Rm(this,e)}_emitSync(){this._needsBarriers&&(this._emit("storageBarrier();"),this._emit("workgroupBarrier();"))}_visitForNode(e){if(e.kind===z.THREAD_BINDING){if(this._serializeThreads){let a=e.loopVar.name,u=this._exprToWGSL(e.extent);this._emit(`for (var ${a}: i32 = 0; ${a} < ${u}; ${a} = ${a} + 1) {`),this._indent++,this._visitNode(e.body),this._indent--,this._emit("}");return}let s=e.extent.type==="IntImmNode"?e.extent.value:0,o=e.threadTag,i=this._getMaxBindingExtent(o);if(s>0&&i>0&&s<i){let a=this._wgslBuiltinAccess(o);this._emit(`if (i32(${a}) < ${s}) {`),this._indent++,this._visitNode(e.body),this._indent--,this._emit("}")}else this._visitNode(e.body);this._needsBarriers&&this._emit("workgroupBarrier();");return}let t=e.loopVar.name,r=this._exprToWGSL(e.extent);this._emit(`for (var ${t}: i32 = 0; ${t} < ${r}; ${t} = ${t} + 1) {`),this._indent++,this._visitNode(e.body),this._indent--,this._emit("}")}_visitBlockNode(e){for(let t of e.iterVars)t.iterVar&&t.binding&&this._emit(`let ${t.iterVar.name}: i32 = ${this._exprToWGSL(t.binding)};`);e.initBody&&this._visitNode(e.initBody),this._visitNode(e.body)}_emitMissingLocalDecls(){for(let e of this._slotDecls)e.scalar?this._emit(`var ${e.name}: ${jn(e.dtype)};`):this._emit(`var ${e.name}: array<${jn(e.dtype)}, ${e.size}>;`)}_collectLocalBuffers(e){let t=new Set;for(let[,a]of e.bufferMap)t.add(a.name);let r=new Set;for(let a of this._sharedBuffers)r.add(a.name);let s=new Map,o=l(a=>!t.has(a)&&!r.has(a)&&!this._promotedBuffers.has(a),"isLocal"),i=new Map;this._scanBufferRefs(e.body,i);for(let[a,u]of i)o(a)&&s.set(a,u);return _l(e.body,a=>{a.type==="AllocateNode"&&a.scope!=="shared"&&a.buffer&&o(a.buffer.name)&&s.set(a.buffer.name,a.buffer)}),s}_assignLocalSlots(e){let t=this._collectLocalBuffers(e);if(t.size===0)return;let{minPos:r,maxPos:s}=this._computeBufferLiveness(e,t),o=[...t.keys()].sort((h,m)=>r.get(h)-r.get(m)||s.get(h)-s.get(m));this._localSlots=new Map;let i=new Map,a=0,u=0,p=this._workgroupSize[0]*this._workgroupSize[1]*this._workgroupSize[2]*(this._dispatchSize[0]*this._dispatchSize[1]*this._dispatchSize[2]),c=this._crossThread||new Set,f=this._crossExtent||new Set,d=l((h,m)=>!this._serializeThreads&&this._threadBindings.size>0&&!c.has(h)&&!f.has(h)&&m.numel()>1&&m.numel()<=p,"scalarEligible");for(let h of o){let m=t.get(h),g=m.numel()>0?m.numel():this._estimateBufferSize(m),y=Math.max(g,1),_=r.get(h),b=s.get(h);if(d(h,m)){let T=`_s${u++}`;this._slotDecls.push({name:T,dtype:m.dtype,size:1,scalar:!0}),this._localSlots.set(h,T),this._scalarSlotNames.add(T);continue}let w=i.get(m.dtype);w||(w=new Tl((T,v)=>T.freeAt-v.freeAt),i.set(m.dtype,w));let x=null,S=w.peek();if(S&&S.freeAt<_&&(x=w.pop()),x)y>x.decl.size&&(x.decl.size=y);else{let T={name:`_lt${a++}`,dtype:m.dtype,size:y};this._slotDecls.push(T),x={decl:T}}x.freeAt=b,this._localSlots.set(h,x.decl.name),w.push(x)}}_computeBufferLiveness(e,t){return this._livenessWalk(e.body,t)}_livenessWalk(e,t){let r=new Map,s=new Map,o=new Set(["ForNode","WhileNode","LIRAccumulatorNode"]),i=["stmts","body","initBody","condBody","loopBody","thenBody","elseBody","value","a","b","condition","expr","offsetExpr","extent","indices","args"],a=0,u=[],p=null,c=0,f=l(h=>{t.has(h)&&(r.has(h)||r.set(h,a),s.set(h,a),u.length>0&&p.add(h))},"touch"),d=l(h=>{if(!h)return;a++;let m=o.has(h.type),g=!1;m&&(u.length===0&&(p=new Set,c=a,g=!0),u.push(h)),(h.type==="BufferLoadNode"||h.type==="BufferStoreNode"||h.type==="LIRFlatLoadNode"||h.type==="LIRFlatStoreNode")&&h.buffer&&f(h.buffer.name),h.type==="LIRAccumulatorNode"&&(h.flushStore&&h.flushStore.buffer&&f(h.flushStore.buffer.name),h.initLoad&&h.initLoad.buffer&&f(h.initLoad.buffer.name));for(let y of i){let _=h[y];if(_)if(Array.isArray(_))for(let b of _)d(b);else typeof _=="object"&&d(_)}if(m&&(u.pop(),g)){let y=a;for(let _ of p)r.set(_,Math.min(r.get(_),c)),s.set(_,Math.max(s.get(_),y));p=null}},"walk");return d(e),{minPos:r,maxPos:s}}_scanAllocateNodes(e,t){let r=[e];for(;r.length>0;){let s=r.pop();if(s){s.type==="AllocateNode"&&t.add(s.buffer.name);for(let o of ln(s))r.push(o)}}}_scanBufferRefs(e,t){let r=[e];for(;r.length>0;){let s=r.pop();if(s){(s.type==="BufferLoadNode"||s.type==="BufferStoreNode"||s.type==="LIRFlatLoadNode"||s.type==="LIRFlatStoreNode")&&s.buffer&&t.set(s.buffer.name,s.buffer),s.type==="LIRAccumulatorNode"&&(s.flushStore&&s.flushStore.buffer&&t.set(s.flushStore.buffer.name,s.flushStore.buffer),s.initLoad&&s.initLoad.buffer&&t.set(s.initLoad.buffer.name,s.initLoad.buffer));for(let o of ln(s))r.push(o)}}}_estimateBufferSize(e){return Im(e)}_visitAllocateNode(e){if(e.scope!=="shared"){if(this._promotedBuffers.has(e.buffer.name)||this._localSlots&&this._localSlots.has(e.buffer.name))return;let t=e.buffer.numel(),r=t>0?t:this._estimateBufferSize(e.buffer);r>0&&this._emit(`var ${e.buffer.name}: array<${jn(e.buffer.dtype)}, ${r}>;`)}}_visitIfStmt(e){this._emit(`if (${this._boolExpr(e.condition)}) {`),this._indent++,this._visitNode(e.thenBody),this._indent--,e.elseBody&&(this._emit("} else {"),this._indent++,this._visitNode(e.elseBody),this._indent--),this._emit("}")}_visitLetStmtNode(e){let t=e.variable.dtype||this._defaultDtype,r=jn(t),s=this._numericExpr(e.value,t);this._emit(`var ${e.variable.name}: ${r} = ${s};`),this._visitNode(e.body)}_visitWhileNode(e){this._visitNode(e.condBody);let t=Array.isArray(e.condVar.shape)?`${this._packedBufAccess(e.condVar.name,"0")} != 0`:e.condVar.name;this._emit(`while (${t}) {`),this._indent++,this._visitNode(e.loopBody),this._visitNode(e.condBody),this._indent--,this._emit("}")}_visitBufferStoreNode(e){let t=this._flatIndex(e.buffer,e.indices),r=this._packedBufAccess(e.buffer.name,t),s=this._numericExpr(e.value,e.buffer.dtype);this._emit(`${r} = ${s};`)}_visitLIRFlatStore(e){let t=this._exprToWGSL(e.offsetExpr),r=this._packedBufAccess(e.buffer.name,t),s=this._numericExpr(e.value,e.buffer.dtype);this._emit(`${r} = ${s};`)}_visitLIRBindings(e){for(let t of e.bindings)this._emit(`let ${t.name}: i32 = ${this._numericExpr(t.expr,"i32")};`);this._visitNode(e.body)}_visitLIRAccumulator(e){let t=e.localName,r=jn(e.dtype||this._defaultDtype);this._emit(`var ${t}: ${r} = ${this._exprToWGSL(e.initLoad)};`);let s=e.loopVar.name,o=this._exprToWGSL(e.extent);this._emit(`for (var ${s}: i32 = 0; ${s} < ${o}; ${s} = ${s} + 1) {`),this._indent++;let i=e.op||"+",a=this._exprToWGSL(e.body);this._emit(`${t} = ${i==="max"||i==="min"?`${i}(${t}, ${a})`:`(${t} ${i} ${a})`};`),this._indent--,this._emit("}");let u=this._exprToWGSL(e.flushStore.offsetExpr),p=this._packedBufAccess(e.flushStore.buffer.name,u);this._emit(`${p} = ${t};`)}_isBoolExpr(e){return e?!!(e.type==="CompareNode"||e.type==="MathOpNode"&&(w$.has(e.op)||T$.has(e.op))):!1}_numericExpr(e,t){if(this._isBoolExpr(e)){let r=t?jn(t):"i32";return r==="f32"?`select(0.0, 1.0, ${this._exprToWGSL(e)})`:`select(${r}(0), ${r}(1), ${this._exprToWGSL(e)})`}return this._exprToWGSL(e)}_boolExpr(e){return this._isBoolExpr(e)?this._exprToWGSL(e):`(${this._exprToWGSL(e)} != 0)`}_exprToWGSL(e){if(!e)return"0";switch(e.type){case"IntImmNode":return String(e.value);case"FloatImmNode":return this._emitFloatLiteral(e.value);case"VariableNode":return this._resolveVariable(e.name);case"BufferLoadNode":return this._packedBufAccess(e.buffer.name,this._flatIndex(e.buffer,e.indices));case"LIRFlatLoadNode":return this._packedBufAccess(e.buffer.name,this._exprToWGSL(e.offsetExpr));case"MathOpNode":{if(!e.b)return e.op==="!"?`(!${this._boolExpr(e.a)})`:`(${e.op}${this._numericExpr(e.a)})`;if(e.op==="&&")return`(${this._boolExpr(e.a)} && ${this._boolExpr(e.b)})`;if(e.op==="||")return`(${this._boolExpr(e.a)} || ${this._boolExpr(e.b)})`;let t=this._numericExpr(e.a),r=this._numericExpr(e.b);return e.op==="//"?`(${t} / ${r})`:e.op==="%"?`(${t} % ${r})`:`(${t} ${e.op} ${r})`}case"CompareNode":return`(${this._numericExpr(e.a)} ${Ku(e.direction)} ${this._numericExpr(e.b)})`;case"IfThenElseNode":return`select(${this._exprToWGSL(e.elseBody)}, ${this._exprToWGSL(e.thenBody)}, ${this._boolExpr(e.condition)})`;case"CastNode":return`${jn(e.toDtype)}(${this._exprToWGSL(e.expr)})`;case"CallExternNode":return this._emitExternCall(e);default:throw new Error(`WebGPU codegen: unhandled expr node '${e.type}'`)}}_resolveVariable(e){if(this._primFunc&&this._primFunc.shapeParams){for(let t of this._primFunc.shapeParams)if(t.name===e)return`i32(_shapes.${e})`}return e}_emitFloatLiteral(e){return e===1/0?"f32(0x1.fffffep+127)":e===-1/0?"f32(-0x1.fffffep+127)":Number.isInteger(e)?`${e}.0`:String(e)}_emitExternCall(e){let t=e.args.length,r=new Array(t);for(let i=0;i<t;i++)r[i]=this._exprToWGSL(e.args[i]);let s=r.join(", ");if(e.externName==="fmod")return`(${r[0]} % ${r[1]})`;if(e.externName==="erf")return T1(r[0]);if(e.externName==="erfc")return`(1.0 - ${T1(r[0])})`;if(e.externName==="lgamma")return x$(r[0]);if(e.externName==="gamma")return S$(r[0]);if(e.externName==="log10")return`(log(${r[0]}) * ${1/Math.LN10})`;let o=DN(e.externName);if(o===e.externName&&!PN(e.externName))throw new Error(`WebGPU codegen: unsupported extern function "${e.externName}"`);return`${o}(${s})`}_packedBufAccess(e,t){if(this._wgPoolOffsets&&this._wgPoolOffsets.has(e)){let r=this._wgPoolOffsets.get(e);return r.offset===0?`${r.pool}[${t}]`:`${r.pool}[${r.offset}u + u32(${t})]`}if(this._packedMode&&this._packedOffsets&&this._packedOffsets.has(e)){let r=this._packedOffsets.get(e);return r.offset===0?`${r.storage}[${t}]`:`${r.storage}[${r.offset}u + u32(${t})]`}if(this._localSlots&&this._localSlots.has(e)){let r=this._localSlots.get(e);return this._scalarSlotNames.has(r)?r:`${r}[${t}]`}return`${e}[${t}]`}_flatIndex(e,t){return ml(e,t,r=>this._exprToWGSL(r),(r,s)=>this._computeDynamicStride(r,s),!1)}_computeDynamicStride(e,t){return Ui(e,t+1,(r,s)=>this._resolveShapeParam(r,s))}_resolveShapeParam(e,t){return ji(this._primFunc,e,t,r=>`i32(_shapes.${r.name})`,"WebGPU","wgsl")}}});function v1(n){let{source:e,name:t,metadata:r}=n,s=[];for(let i=0;i<r.paramCount;i++)s.push(`buf_${i}`);let o=[];return o.push(e),o.push(""),o.push(`// ${t}(${s.map(i=>`/* ${i}: Float32Array */`).join(", ")});`),o.join(`
+`)}var k1=M(()=>{"use strict";l(v1,"buildSnippet")});function re(n){let e=[];do{let t=n&127;n>>>=7,n&&(t|=128),e.push(t)}while(n);return e}function C$(n){let e=[],t=!0;for(;t;){let r=n&127;n>>=7,n===0&&!(r&64)||n===-1&&r&64?t=!1:r|=128,e.push(r)}return e}function D$(n){n=BigInt(n);let e=[],t=!0;for(;t;){let r=Number(n&0x7fn);n>>=7n,n===0n&&!(r&64)||n===-1n&&r&64?t=!1:r|=128,e.push(r)}return e}function qm(n){let e=new TextEncoder().encode(n);return[...re(e.length),...e]}function P$(n){let e=new ArrayBuffer(4);return new Float32Array(e)[0]=n,[...new Uint8Array(e)]}function $$(n){let e=new ArrayBuffer(8);return new Float64Array(e)[0]=n,[...new Uint8Array(e)]}function _o(n,e){for(let t=0;t<e.length;t++)n.push(e[t]);return n}function xl(n,e){let t=[n];return _o(t,re(e.length)),_o(t,e),t}function Ua(n){let e=[];_o(e,re(n.length));for(let t of n)_o(e,t);return e}function F$(n){let e=[],t=0;for(;t<n.length;){let r=n[t];if(r<=" "){t++;continue}if(r===";"&&n[t+1]===";"){t=n.indexOf(`
+`,t),t<0&&(t=n.length);continue}if(r==="("&&n[t+1]===";"){t=n.indexOf(";)",t+2),t=t<0?n.length:t+2;continue}if(r==="("||r===")"){e.push(r),t++;continue}if(r==='"'){let o=t+1;for(;o<n.length&&n[o]!=='"';)o++;e.push(n.substring(t,o+1)),t=o+1;continue}let s=t;for(;s<n.length&&n[s]>" "&&n[s]!=="("&&n[s]!==")";)s++;e.push(n.substring(t,s)),t=s}return e}function z$(n){let e=0,t=l(()=>n[e++],"eat"),r=l(()=>n[e],"peek"),s=l(y=>{if(t()!==y)throw new Error("expected "+y+" at "+(e-1))},"expect");s("("),s("module");let o=[],i=1,a=256,u="",p=[],c=[],f=[],d=[],h=-1,m=-1;for(;r()!==")";){if(r()!=="("){e++;continue}e++;let y=t();if(y==="memory"){for(;r()==="(";)g();r()!==")"&&(i=parseInt(t())||1),r()!==")"&&(a=parseInt(t())||256),s(")")}else if(y==="import"){let _=t().replace(/"/g,""),b=t().replace(/"/g,"");s("("),s("func"),r().startsWith("$")&&e++;let w=[],x=[];for(;r()==="(";){e++;let S=t();if(S==="param")for(;r()!==")";){let T=t();T==="f32"?w.push(Rp):T==="f64"?w.push(Bp):T==="i32"?w.push(Op):T==="v128"&&w.push(Mp)}else if(S==="result")for(;r()!==")";){let T=t();T==="f32"?x.push(Rp):T==="f64"?x.push(Bp):T==="i32"?x.push(Op):T==="v128"&&x.push(Mp)}s(")")}s(")"),s(")"),o.push({module:_,name:b,params:w,results:x})}else if(y==="func"){for(;r()==="(";){let b=e;e++;let w=t();if(w==="export")u=t().replace(/"/g,""),s(")");else if(w==="param"){let x="";for(;r()!==")";){let S=t();if(S.startsWith("$")){x=S.replace("$","");continue}let T=null;S==="i32"?T=Op:S==="i64"?T=Q0:S==="f32"?T=Rp:S==="f64"?T=Bp:S==="v128"&&(T=Mp),T!==null&&(p.push(T),c.push(x),x="")}s(")")}else if(w==="result"){for(;r()!==")";)t();s(")")}else if(w==="local"){for(;r()!==")";){let x=t();x.startsWith("$")?d.push(x.replace("$","")):x==="i32"?f.push(Op):x==="i64"?f.push(Q0):x==="f32"?f.push(Rp):x==="f64"?f.push(Bp):x==="v128"&&f.push(Mp)}s(")")}else{e=b;break}}h=e;let _=1;for(;_>0;){let b=t();b==="("?_++:b===")"&&_--}m=e-1}else g()}function g(){if(r()!=="("){e++;return}let y=0;do{let _=t();_==="("?y++:_===")"&&y--}while(y>0)}return l(g,"skipSExpr"),{imports:o,memMin:i,memMax:a,funcExportName:u,funcParams:p,funcParamNames:c,funcLocals:f,funcLocalNames:d,bodyTokens:n.slice(h,m)}}function V$(n,e,t){let r=[],s=0,o=l(()=>n[s],"peek"),i=l(()=>n[s++],"eat"),a=[],u=-1;for(let y of e.values())y>u&&(u=y);function p(y){let _=y.replace("$","");if(e.has(_))return e.get(_);let b=parseInt(_,10);if(!isNaN(b))return b>u&&(u=b),b;let w=u+1;return u=w,e.set(_,w),w}l(p,"localIdx");function c(){return o()&&o().startsWith("$")?i().replace("$",""):""}l(c,"eatLabel");function f(y,_){let b=y.indexOf(_);return b>=0?y.substring(b+_.length):""}l(f,"extractName");function d(){for(;s<n.length&&o()!==")";){let y=o();if(y==="("){s++;let _=i();if(_==="i32.const")r.push(65),r.push(...C$(parseInt(i()))),m(")");else if(_==="i64.const")r.push(66),r.push(...D$(i())),m(")");else if(_==="f32.const")r.push(67),r.push(...P$(parseFloat(i()))),m(")");else if(_==="f64.const")r.push(68),r.push(...$$(parseFloat(i()))),m(")");else if(_==="local.get")r.push(32),r.push(...re(p(i()))),m(")");else if(_==="local.set")r.push(33),r.push(...re(p(i()))),m(")");else if(_==="block"){let b=i().replace("$","");a.push(b),r.push(2,J0),d(),r.push(11),a.pop(),m(")")}else if(_==="loop"){let b=i().replace("$","");a.push(b),r.push(3,J0),d(),r.push(11),a.pop(),m(")")}else if(_==="if"){let b=J0;if(o()==="("){let w=s;if(s++,o()==="result"){i();let x=i();b=x==="f32"?Rp:x==="f64"?Bp:x==="v128"?Mp:x==="i64"?Q0:Op,m(")")}else s=w}for(r.push(4,b),a.push("_if");s<n.length&&o()!==")"&&o()==="(";){let w=s;s++;let x=i();if(x==="then")d(),m(")");else if(x==="else")r.push(5),d(),m(")");else{s=w;break}}r.push(11),a.pop(),m(")")}else s-=2,g();continue}if(s++,y==="br_if")r.push(13),r.push(...re(h(c())));else if(y==="br")r.push(12),r.push(...re(h(c())));else if(y.startsWith("call")){r.push(16);let _=f(y,"math_")||(o()&&o().startsWith("$")?f(i(),"math_"):""),b=t.get(_);if(b===void 0)throw new Error(`WAT encoder: call to unresolved import '${_}'`);r.push(...re(b))}else y==="local.get"?(r.push(32),r.push(...re(p(c())))):y==="local.set"?(r.push(33),r.push(...re(p(c())))):y.endsWith(".extract_lane")||y.endsWith(".replace_lane")?(r.push(...Z0.get(y)),r.push(parseInt(i(),10))):Z0.has(y)&&r.push(...Z0.get(y))}}l(d,"emitBlock");function h(y){for(let _=a.length-1;_>=0;_--)if(a[_]===y)return a.length-1-_;return 0}l(h,"resolveBr");function m(y){if(i()!==y)throw new Error("expect "+y)}l(m,"expect");function g(){if(o()!=="("){s++;return}let y=0;do{let _=i();_==="("?y++:_===")"&&y--}while(y>0)}return l(g,"skipSExpr"),d(),r}function Sl(n){let e=F$(n),t=z$(e),r=new Map,s=t.funcParams.length;for(let T=0;T<t.funcParamNames.length;T++)t.funcParamNames[T]&&r.set(t.funcParamNames[T],T);for(let T=0;T<t.funcLocalNames.length;T++)r.set(t.funcLocalNames[T],s+T);let o=new Map;for(let T=0;T<t.imports.length;T++)o.set(t.imports[T].name,T);let i=[],a=new Map;function u(T,v){let A=T.join(",")+">"+v.join(",");if(a.has(A))return a.get(A);let N=i.length;return i.push({params:T,results:v}),a.set(A,N),N}l(u,"getType");for(let T of t.imports)T.typeIdx=u(T.params,T.results);let p=u(t.funcParams,[]),c=xl(O$,Ua(i.map(T=>[L$,...re(T.params.length),...T.params,...re(T.results.length),...T.results]))),f=[];t.imports.length>0&&(f=xl(B$,Ua(t.imports.map(T=>[...qm(T.module),...qm(T.name),0,...re(T.typeIdx)]))));let d=xl(R$,Ua([[...re(p)]])),h=xl(M$,Ua([[1,...re(t.memMin),...re(t.memMax)]])),m=t.imports.length,g=xl(I$,Ua([[...qm("memory"),2,...re(0)],[...qm(t.funcExportName),0,...re(m)]])),y=[];if(t.funcLocals.length>0){let T=[],v=t.funcLocals[0],A=1;for(let N=1;N<t.funcLocals.length;N++)t.funcLocals[N]===v?A++:(T.push([...re(A),v]),v=t.funcLocals[N],A=1);T.push([...re(A),v]),_o(y,Ua(T))}else y.push(0);let _=V$(t.bodyTokens,r,o),b=[];_o(b,y),_o(b,_),b.push(11);let w=[];_o(w,re(b.length)),_o(w,b);let x=xl(E$,Ua([w])),S=[];for(let T of[N$,A$,c,f,d,h,g,x])_o(S,T);return new Uint8Array(S)}var N$,A$,O$,B$,R$,M$,I$,E$,Op,Q0,Bp,Rp,Mp,L$,J0,Z0,Wm=M(()=>{"use strict";N$=[0,97,115,109],A$=[1,0,0,0],O$=1,B$=2,R$=3,M$=5,I$=7,E$=10,Op=127,Q0=126,Bp=124,Rp=125,Mp=123,L$=96,J0=64;l(re,"uleb");l(C$,"sleb");l(D$,"slebBig");l(qm,"encStr");l(P$,"encF32");l($$,"encF64");l(_o,"pushAll");l(xl,"section");l(Ua,"vec");Z0=new Map([["i32.const",[65]],["f32.const",[67]],["local.get",[32]],["local.set",[33]],["i32.add",[106]],["i32.sub",[107]],["i32.mul",[108]],["i32.div_s",[109]],["i32.rem_s",[111]],["i32.eq",[70]],["i32.ne",[71]],["i32.lt_s",[72]],["i32.gt_s",[74]],["i32.le_s",[76]],["i32.ge_s",[78]],["i32.eqz",[69]],["i32.lt_u",[73]],["i32.gt_u",[75]],["i32.le_u",[77]],["i32.ge_u",[79]],["i32.and",[113]],["i32.or",[114]],["i32.xor",[115]],["i32.shl",[116]],["i32.shr_s",[117]],["i32.shr_u",[118]],["i32.reinterpret_f32",[188]],["f32.reinterpret_i32",[190]],["i32.trunc_f32_s",[168]],["f32.add",[146]],["f32.sub",[147]],["f32.mul",[148]],["f32.div",[149]],["f32.neg",[140]],["f32.abs",[139]],["f32.ceil",[141]],["f32.floor",[142]],["f32.sqrt",[145]],["f32.min",[150]],["f32.max",[151]],["f32.eq",[91]],["f32.ne",[92]],["f32.lt",[93]],["f32.gt",[94]],["f32.le",[95]],["f32.ge",[96]],["f32.convert_i32_s",[178]],["f64.add",[160]],["f64.sub",[161]],["f64.mul",[162]],["f64.div",[163]],["f64.min",[164]],["f64.max",[165]],["f64.neg",[154]],["f64.abs",[153]],["f64.ceil",[155]],["f64.floor",[156]],["f64.sqrt",[159]],["f64.eq",[97]],["f64.ne",[98]],["f64.lt",[99]],["f64.gt",[100]],["f64.le",[101]],["f64.ge",[102]],["f64.convert_i32_s",[183]],["f64.promote_f32",[187]],["f32.demote_f64",[182]],["i32.trunc_f64_s",[170]],["i64.load",[41,3,0]],["i64.store",[55,3,0]],["i64.add",[124]],["i64.sub",[125]],["i64.mul",[126]],["i64.div_s",[127]],["i64.rem_s",[129]],["i64.and",[131]],["i64.or",[132]],["i64.xor",[133]],["i64.shl",[134]],["i64.shr_s",[135]],["i64.shr_u",[136]],["i64.eqz",[80]],["i64.eq",[81]],["i64.ne",[82]],["i64.lt_s",[83]],["i64.gt_s",[85]],["i64.le_s",[87]],["i64.ge_s",[89]],["i64.extend_i32_s",[172]],["i32.wrap_i64",[167]],["f64.convert_i64_s",[185]],["i64.trunc_f64_s",[176]],["f32.convert_i64_s",[180]],["i64.trunc_f32_s",[174]],["select",[27]],["f32.load",[42,2,0]],["f32.store",[56,2,0]],["f64.load",[43,3,0]],["f64.store",[57,3,0]],["i32.load",[40,2,0]],["i32.store",[54,2,0]],["i32.load8_s",[44,0,0]],["i32.load8_u",[45,0,0]],["i32.store8",[58,0,0]],["i32.load16_s",[46,1,0]],["i32.load16_u",[47,1,0]],["i32.store16",[59,1,0]],["v128.load",[253,...re(0),4,0]],["v128.store",[253,...re(11),4,0]],["v128.bitselect",[253,...re(82)]],["v128.and",[253,...re(78)]],["v128.or",[253,...re(80)]],["v128.not",[253,...re(77)]],["f32x4.splat",[253,...re(19)]],["i32x4.splat",[253,...re(17)]],["f32x4.add",[253,...re(228)]],["f32x4.sub",[253,...re(229)]],["f32x4.mul",[253,...re(230)]],["f32x4.div",[253,...re(231)]],["f32x4.neg",[253,...re(225)]],["f32x4.abs",[253,...re(224)]],["f32x4.sqrt",[253,...re(227)]],["f32x4.ceil",[253,...re(103)]],["f32x4.floor",[253,...re(104)]],["f32x4.min",[253,...re(232)]],["f32x4.max",[253,...re(233)]],["f32x4.eq",[253,...re(65)]],["f32x4.ne",[253,...re(66)]],["f32x4.lt",[253,...re(67)]],["f32x4.gt",[253,...re(68)]],["f32x4.le",[253,...re(69)]],["f32x4.ge",[253,...re(70)]],["f32x4.extract_lane",[253,...re(31)]],["f32x4.replace_lane",[253,...re(32)]],["i32x4.add",[253,...re(174)]],["i32x4.sub",[253,...re(177)]],["i32x4.mul",[253,...re(181)]],["i32x4.abs",[253,...re(160)]],["i32x4.min_s",[253,...re(182)]],["i32x4.max_s",[253,...re(184)]],["i32x4.eq",[253,...re(55)]],["i32x4.ne",[253,...re(56)]],["i32x4.lt_s",[253,...re(57)]],["i32x4.gt_s",[253,...re(58)]],["i32x4.le_s",[253,...re(59)]],["i32x4.ge_s",[253,...re(60)]],["i32x4.extract_lane",[253,...re(27)]],["i32x4.replace_lane",[253,...re(28)]],["f64x2.splat",[253,...re(20)]],["f64x2.add",[253,...re(240)]],["f64x2.sub",[253,...re(241)]],["f64x2.mul",[253,...re(242)]],["f64x2.div",[253,...re(243)]],["f64x2.neg",[253,...re(237)]],["f64x2.abs",[253,...re(236)]],["f64x2.sqrt",[253,...re(239)]],["f64x2.min",[253,...re(244)]],["f64x2.max",[253,...re(245)]],["f64x2.eq",[253,...re(71)]],["f64x2.ne",[253,...re(72)]],["f64x2.lt",[253,...re(73)]],["f64x2.gt",[253,...re(74)]],["f64x2.le",[253,...re(75)]],["f64x2.ge",[253,...re(76)]],["f64x2.extract_lane",[253,...re(33)]],["f64x2.replace_lane",[253,...re(34)]]]);l(F$,"tokenize");l(z$,"parseModule");l(V$,"encodeBody");l(Sl,"encodeWat")});function N1(n){let{source:e,name:t,metadata:r}=n,{bufferOffsets:s,imports:o}=r,i=[...s.entries()],a=Sl(e),u=[];if(u.push("(async () => {"),u.push(""),u.push("const binary = new Uint8Array(["+a.join(",")+"]);"),u.push(""),o&&o.size>0){u.push("const mathImports = {");for(let[c]of o)c==="fmod"?u.push("  fmod: (a, b) => a % b,"):c==="rsqrt"?u.push("  rsqrt: x => 1 / Math.sqrt(x),"):u.push(`  ${c}: Math.${c},`);u.push("};"),u.push("const { instance } = await WebAssembly.instantiate(binary, { math: mathImports });")}else u.push("const { instance } = await WebAssembly.instantiate(binary);");u.push("const memory = instance.exports.memory;"),u.push("");for(let[c,f]of i)u.push(`// ${c}: new Float32Array(memory.buffer, ${f}, N).set(data);`);u.push(""),u.push(`instance.exports.${t}(${i.map(([,c])=>c).join(", ")});`),u.push("");let p=i[i.length-1];return p&&(u.push(`const result = new Float32Array(memory.buffer, ${p[1]}, N);`),u.push("console.log(result);")),u.push("})();"),u.join(`
+`)}var A1=M(()=>{"use strict";Wm();l(N1,"buildSnippet")});function O1(n){let{source:e,name:t,metadata:r}=n,{bindings:s,dispatchSize:o}=r,i=[];i.push("(async () => {"),i.push("const adapter = await navigator.gpu.requestAdapter();"),i.push("const device = await adapter.requestDevice();"),i.push(""),i.push("const wgsl = "+JSON.stringify(e)+";"),i.push(""),i.push("const pipeline = device.createComputePipeline({"),i.push('  layout: "auto",'),i.push("  compute: { module: device.createShaderModule({ code: wgsl }), entryPoint: "+JSON.stringify(t)+" }"),i.push("});"),i.push(""),i.push("const entries = [];"),i.push("const gpuBuffers = [];"),i.push("");for(let a of s)a.name==="_shapes"?G$(i,a):a.packed?j$(i,a):a.mode==="read_write"?U$(i,a):q$(i,a),i.push("");return i.push("const bindGroup = device.createBindGroup({ layout: pipeline.getBindGroupLayout(0), entries });"),i.push("const encoder = device.createCommandEncoder();"),i.push("const pass = encoder.beginComputePass();"),i.push("pass.setPipeline(pipeline);"),i.push("pass.setBindGroup(0, bindGroup);"),i.push(`pass.dispatchWorkgroups(${o[0]}, ${o[1]}, ${o[2]});`),i.push("pass.end();"),i.push(""),W$(i,s),i.push(""),i.push("gpuBuffers.forEach(b => b.destroy());"),i.push("device.destroy();"),i.push("})();"),i.join(`
+`)}function G$(n,e){n.push(`// binding ${e.index}: uniform (_shapes)`),n.push("{"),n.push("  const shapeData = new Uint32Array([/* shape values */]);"),n.push("  const size = Math.max(Math.ceil(shapeData.byteLength / 16) * 16, 16);"),n.push("  const buf = device.createBuffer({ size, usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST });"),n.push("  device.queue.writeBuffer(buf, 0, shapeData);"),n.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),n.push("  gpuBuffers.push(buf);"),n.push("}")}function j$(n,e){let t=e.mode==="read_write",r=t?"GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST":"GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST";if(n.push(`// binding ${e.index}: packed ${e.name} (${e.packed.length} buffers, ${e.packedSize} elems)`),n.push("{"),n.push(`  const buf = device.createBuffer({ size: ${e.packedSize*4}, usage: ${r}, mappedAtCreation: true });`),n.push("  const mapped = new Float32Array(buf.getMappedRange());"),t)n.push("  mapped.fill(0);");else for(let s of e.packed)n.push(`  // mapped.set(${s.name}_data, ${s.offset}); // size=${s.size}`);n.push("  buf.unmap();"),n.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),n.push("  gpuBuffers.push(buf);"),n.push("}")}function U$(n,e){n.push(`// binding ${e.index}: output ${e.name}`),n.push("{"),n.push("  const size = N * 4; // N = output element count"),n.push("  const buf = device.createBuffer({ size, usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST, mappedAtCreation: true });"),n.push("  new Float32Array(buf.getMappedRange()).fill(0);"),n.push("  buf.unmap();"),n.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),n.push("  gpuBuffers.push(buf);"),n.push("}")}function q$(n,e){n.push(`// binding ${e.index}: input ${e.name}`),n.push("{"),n.push(`  const data = ${e.name}_data; // Float32Array`),n.push("  const buf = device.createBuffer({ size: Math.max(data.byteLength, 4), usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST, mappedAtCreation: true });"),n.push("  new Float32Array(buf.getMappedRange()).set(data);"),n.push("  buf.unmap();"),n.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),n.push("  gpuBuffers.push(buf);"),n.push("}")}function W$(n,e){if(e.filter(r=>r.mode==="read_write").length===0){n.push("device.queue.submit([encoder.finish()]);");return}for(let r=0;r<e.length;r++){let s=e[r];if(s.mode==="read_write")if(s.packed)for(let o of s.packed)n.push("{"),n.push(`  const rb = device.createBuffer({ size: ${o.size*4}, usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST });`),n.push(`  encoder.copyBufferToBuffer(gpuBuffers[${r}], ${o.offset*4}, rb, 0, ${o.size*4});`),n.push("  device.queue.submit([encoder.finish()]);"),n.push("  await rb.mapAsync(GPUMapMode.READ);"),n.push(`  console.log("${o.name}:", new Float32Array(rb.getMappedRange()));`),n.push("  rb.unmap(); rb.destroy();"),n.push("}");else n.push("{"),n.push(`  const src = gpuBuffers[${r}];`),n.push("  const rb = device.createBuffer({ size: src.size, usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST });"),n.push("  encoder.copyBufferToBuffer(src, 0, rb, 0, src.size);"),n.push("  device.queue.submit([encoder.finish()]);"),n.push("  await rb.mapAsync(GPUMapMode.READ);"),n.push(`  console.log("${s.name}:", new Float32Array(rb.getMappedRange()));`),n.push("  rb.unmap(); rb.destroy();"),n.push("}")}}var B1=M(()=>{"use strict";l(O1,"buildSnippet");l(G$,"emitShapeBinding");l(j$,"emitPackedBinding");l(U$,"emitOutputBinding");l(q$,"emitInputBinding");l(W$,"emitReadback")});function R1(n){let{source:e,name:t,metadata:r}=n,{blockDim:s,gridDim:o,sharedMemBytes:i,params:a}=r,u=[];return u.push(e),u.push(""),u.push("// Launch:"),u.push(`//   dim3 block(${s[0]}, ${s[1]}, ${s[2]});`),u.push(`//   dim3 grid(${o[0]}, ${o[1]}, ${o[2]});`),i>0?u.push(`//   ${t}<<<grid, block, ${i}>>>(${a.join(", ")});`):u.push(`//   ${t}<<<grid, block>>>(${a.join(", ")});`),u.join(`
+`)}var M1=M(()=>{"use strict";l(R1,"buildSnippet")});function Km(n,e){I1.set(n,e),e.runtimeKind&&e.snippet&&E1.set(e.runtimeKind,e.snippet)}function L1(n){return I1.get(n)||null}function C1(n){return E1.get(n)||null}var I1,E1,D1=M(()=>{"use strict";yo();d1();b1();w1();S1();k1();A1();B1();M1();I1=new Map,E1=new Map;l(Km,"registerCodegen");l(L1,"getCodegenEntry");l(C1,"getSnippetBuilder");Km(Dt.CPU,{runtimeKind:"js",snippet:v1,compile(n,e){return{source:new Fm(e).generate(n),metadata:{kind:"js",paramCount:n.params.length}}}});Km(Dt.WASM,{runtimeKind:"wasm",snippet:N1,compile(n,e){let t=new Vm(e).generate(n),r={kind:"wasm",memoryPages:t.memoryPages,bufferOffsets:t.bufferOffsets,imports:t.imports,params:t.params,bufferMap:n.bufferMap};return t.parallel&&(r.parallel=t.parallel),{source:t.wat,metadata:r}}});Km(Dt.WEBGPU,{runtimeKind:"webgpu",snippet:O1,compile(n,e){let t=new Um(e).generate(n);return{source:t.source,metadata:{kind:"webgpu",workgroupSize:t.workgroupSize,dispatchSize:t.dispatchSize,sharedMemBytes:t.sharedMemBytes,params:t.params,bindings:t.bindings}}}});Km(Dt.CUDA,{runtimeKind:"cuda",snippet:R1,compile(n,e,t){if(t&&t.matmulBackend==="cublas"&&n.cublasInfo)return{source:"",metadata:{kind:"cuda",cublas:n.cublasInfo,outputIndices:[n.cublasInfo.cIdx]}};let r=new zm(e).generate(n);return{source:r.source,metadata:{kind:"cuda",blockDim:r.blockDim,gridDim:r.gridDim,sharedMemBytes:r.sharedMemBytes,params:r.params,outputIndices:r.outputIndices,scratch:r.scratch}}}})});function P1(n){let e=[],t=[n.body];for(;t.length>0;){let m=t.pop();if(m){if(m.type==="BlockNode"){e.push(m),t.push(m.body);continue}if(m.body&&t.push(m.body),m.stmts)for(let g of m.stmts)t.push(g);m.thenBody&&t.push(m.thenBody),m.elseBody&&t.push(m.elseBody)}}let r=null;for(let m of e)if(m.name.includes("matmul"))m.reads.length>=2&&m.writes.length>=1&&(r=m);else return null;if(!r)return null;let s=r.reads[0].buffer,o=r.reads[1].buffer,i=r.writes[0].buffer;if(s.dtype!=="f32"||o.dtype!=="f32"||i.dtype!=="f32"||s.shape.length!==2||o.shape.length!==2||i.shape.length!==2)return null;let a=i.shape[0],u=i.shape[1],p=s.shape[1];if(![a,u,p].every(m=>typeof m=="number"))return null;let c=[];for(let[,m]of n.bufferMap)c.push(m.name);let f=c.indexOf(s.name),d=c.indexOf(o.name),h=c.indexOf(i.name);return f<0||d<0||h<0?null:{M:a,N:u,K:p,aIdx:f,bIdx:d,cIdx:h}}var Ip,bo,Ep=M(()=>{"use strict";D1();l(P1,"detectPureMatmul");Ip=class{static{l(this,"CompiledKernel")}constructor(e,t,r,s={}){this.name=e,this.source=t,this.target=r,this.metadata=s}snippet(){let e=C1(this.metadata.kind);if(!e)throw new Error(`No snippet for kind: ${this.metadata.kind}`);return e(this)}},bo=class{static{l(this,"BackendPipeline")}constructor(e,t={}){this.target=e,this.matmulBackend=t.matmulBackend||"native",this.context=t.context||null}compile(e){let t=this.context&&this.context.getCodegenEntry(this.target.kind)||L1(this.target.kind);if(!t)throw new Error(`Unsupported target kind: ${this.target.kind}`);let{source:r,metadata:s}=t.compile(e,this.target,this);return new Ip(e.name,r,this.target,s)}compileAll(e){return e.map(t=>this.compile(t))}}});function Hm(n,e){$1.set(n,e)}function F1(n){return $1.get(n)||null}var $1,Ym=M(()=>{"use strict";$1=new Map;l(Hm,"registerMeasurer");l(F1,"getMeasurer")});var z1={};at(z1,{runWasmParallel:()=>Z$});import{Worker as K$}from"node:worker_threads";import H$ from"node:os";function J$(){if(!X0){let n=Math.max(1,Math.min(H$.cpus().length-1,16));X0=new eT(n)}return X0}async function Z$(n,e,t,r,s,o){let i=J$();i._ensureWorkers();let{binary:a,bufferOffsets:u}=n,p=[...u.values()],c=Math.min(p.length,t.length),{extent:f,outputIndices:d}=s,h=new Set(d||[]),m=Math.min(i.numWorkers,Math.max(1,f)),g=Math.ceil(f/m),y=[];for(let T=0;T<c;T++){let v=t[T],N=(v instanceof Float32Array?v.length:0)/f;y.push(N>=1&&Number.isInteger(N)?N:0)}let _=p.slice(0,c);if(r)for(let T of r)_.push(T);let b=new SharedArrayBuffer(4*m),w=a.buffer.slice(a.byteOffset,a.byteOffset+a.byteLength),x=[];for(let T=0;T<m;T++){let v=T*g,A=Math.min(v+g,f);if(v>=f)break;let N=[];for(let O=0;O<c;O++){let R=t[O];R instanceof Float32Array&&N.push({offset:p[O],length:R.length,data:R.buffer.slice(0),fullLength:R.length,elemStart:0})}let k=[..._,v,A];x.push(i.dispatch(T,{wasmBinary:w,mathImportNames:o,bufferEntries:N,callArgs:k,controlBuffer:b,workerIdx:T,outputIndices:d||[]}))}let S=await Promise.all(x);for(let T=0;T<c;T++){if(!h.has(T)&&h.size>0)continue;let v=t[T];if(!(v instanceof Float32Array))continue;let A=y[T];if(A>0)for(let N of S){let k=N.outputs[T];if(!k)continue;let R=N.workerIdx*g,B=Math.min(R+g,f),E=new Float32Array(k),F=R*A,C=B*A;v.set(E.subarray(F,C),F)}else{let N=S.find(k=>k.workerIdx===0);N&&N.outputs[T]&&v.set(new Float32Array(N.outputs[T]))}}}var Y$,X0,eT,V1=M(()=>{"use strict";Yi();Y$=`
+const { parentPort } = require('node:worker_threads');
+
+const LANCZOS_G = ${7};
+const LANCZOS_COEFFS = ${JSON.stringify(oi)};
+const ERF_A = ${JSON.stringify(qi)};
+const ERF_P = ${Va};
+const DIGAMMA_SHIFT = ${ii};
+const DIGAMMA_SERIES = ${JSON.stringify(Ga)};
+const erfScalar = ${Wi.toString()};
+const erfcScalar = ${Ki.toString()};
+const lgammaScalar = ${cn.toString()};
+const gammaScalar = ${Hi.toString()};
+const digammaScalar = ${ja.toString()};
+
+const MATH = {
+  exp: Math.exp, log: Math.log, sin: Math.sin, cos: Math.cos,
+  tan: Math.tan, tanh: Math.tanh, pow: Math.pow,
+  fmod: (a, b) => a % b,
+  rsqrt: x => 1 / Math.sqrt(x),
+  sign: Math.sign, round: Math.round,
+  erf: erfScalar, erfc: erfcScalar, lgamma: lgammaScalar, gamma: gammaScalar, digamma: digammaScalar,
+};
+
+parentPort.on('message', (msg) => {
+  const { wasmBinary, mathImportNames, bufferEntries,
+          callArgs, controlBuffer, workerIdx, outputIndices } = msg;
+
+  const compiled = new WebAssembly.Module(new Uint8Array(wasmBinary));
+
+  const mathImports = {};
+  for (const name of mathImportNames) {
+    mathImports[name] = MATH[name] || Math[name] || (x => x);
+  }
+
+  const instance = new WebAssembly.Instance(compiled, { math: mathImports });
+  const wasmMem = instance.exports.memory;
+  const fn = Object.values(instance.exports).find(v => typeof v === 'function');
+
+  for (const entry of bufferEntries) {
+    new Float32Array(wasmMem.buffer, entry.offset, entry.length).set(new Float32Array(entry.data));
+  }
+
+  fn(...callArgs);
+
+  const outSet = new Set(outputIndices);
+  const outputs = [];
+  const transferable = [];
+  for (let i = 0; i < bufferEntries.length; i++) {
+    if (outSet.has(i)) {
+      const entry = bufferEntries[i];
+      const out = new Float32Array(entry.length);
+      out.set(new Float32Array(wasmMem.buffer, entry.offset, entry.length));
+      outputs.push(out.buffer);
+      transferable.push(out.buffer);
+    } else {
+      outputs.push(null);
+    }
+  }
+  const ctrl = new Int32Array(controlBuffer);
+  parentPort.postMessage({ workerIdx, outputs }, transferable);
+  Atomics.store(ctrl, workerIdx, 1);
+  Atomics.notify(ctrl, workerIdx);
+});
+`,X0=null;l(J$,"getPool");eT=class{static{l(this,"WasmThreadPool")}constructor(e){this.numWorkers=e,this._workers=[],this._ready=[]}_ensureWorkers(){if(!(this._workers.length>0))for(let e=0;e<this.numWorkers;e++){let t=new K$(Y$,{eval:!0});t.unref(),this._workers.push(t)}}dispatch(e,t){return this._ensureWorkers(),new Promise(r=>{let s=l(o=>{o.workerIdx===e&&(this._workers[e].removeListener("message",s),r(o))},"handler");this._workers[e].on("message",s),this._workers[e].postMessage(t)})}terminate(){for(let e of this._workers)e.terminate();this._workers=[]}};l(Z$,"runWasmParallel")});import{readdirSync as Lp,existsSync as qa}from"fs";import{join as ui,delimiter as G1}from"path";function q1(){let n=[];return process.env.CUDA_PATH&&n.push(process.env.CUDA_PATH),!Wa&&process.env.CUDA_HOME&&n.push(process.env.CUDA_HOME),n}function W1(){let n=[];if(Wa){let e="C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA";if(qa(e))for(let t of Lp(e))n.push(ui(e,t))}else{let e="/usr/local";if(qa(e))for(let t of Lp(e))(t==="cuda"||t.startsWith("cuda-"))&&n.push(ui(e,t))}return n}function Q$(n){return Wa?[ui(n,"bin")]:[ui(n,"lib64"),ui(n,"targets/x86_64-linux/lib")]}function X$(){return Wa?[]:["/usr/lib/x86_64-linux-gnu","/lib/x86_64-linux-gnu"]}function U1(n){let e=n.match(/\d+/g);return e?e.map(Number):[]}function eF(n,e){let t=U1(n),r=U1(e),s=Math.max(t.length,r.length);for(let o=0;o<s;o++){let i=(t[o]||0)-(r[o]||0);if(i!==0)return i}return 0}function tF(n){let e=process.env[j1]||"";e.split(G1).includes(n)||(process.env[j1]=e?n+G1+e:n)}function bs(n){let{pattern:e,fallback:t,extraDirs:r}=Wa?n.win:n.linux,s=[];for(let o of[...q1(),...W1()])for(let i of Q$(o))s.push(i);for(let o of X$())s.push(o);if(r)for(let o of r())s.push(o);for(let o of s){if(!qa(o))continue;let i=Lp(o).filter(a=>e.test(a)&&!a.includes(".alt"));if(i.length>0)return tF(o),i.sort(eF),ui(o,i.pop())}return t}function nF(){for(let n of[...q1(),...W1()]){let e=ui(n,"include");if(qa(e))return e}return Wa?null:qa("/usr/local/cuda/include")?"/usr/local/cuda/include":null}function rF(){let n=[],e="C:/Program Files/NVIDIA/CUDNN";if(qa(e))for(let t of Lp(e).sort().reverse()){let r=ui(e,t,"bin");if(qa(r))for(let s of Lp(r).filter(o=>o.startsWith("12.")).sort().reverse())n.push(ui(r,s,"x64"))}return n}var Wa,j1,Jm,K1,H1,Zm,Y1,J1,Z1,vl=M(()=>{"use strict";Wa=process.platform==="win32",j1=Wa?"PATH":"LD_LIBRARY_PATH";l(q1,"envRoots");l(W1,"toolkitRoots");l(Q$,"libDirs");l(X$,"systemDirs");l(U1,"versionKey");l(eF,"compareVersion");l(tF,"prependSearchPath");l(bs,"loadCudaLib");l(nF,"resolveIncludeDir");Jm=nF(),K1={win:{pattern:/^nvcuda\.dll$/,fallback:"nvcuda.dll"},linux:{pattern:/^libcuda\.so(\.\d+)*$/,fallback:"libcuda.so.1"}},H1={win:{pattern:/^nvrtc64_\d+_\d+\.dll$/,fallback:"nvrtc64_120_0.dll"},linux:{pattern:/^libnvrtc\.so(\.\d+)*$/,fallback:"libnvrtc.so.12"}},Zm={win:{pattern:/^cudart64_\d+\.dll$/,fallback:"cudart64_12.dll"},linux:{pattern:/^libcudart\.so(\.\d+)*$/,fallback:"libcudart.so.12"}},Y1={win:{pattern:/^cublas64_\d+\.dll$/,fallback:"cublas64_12.dll"},linux:{pattern:/^libcublas\.so(\.\d+)*$/,fallback:"libcublas.so.12"}},J1={win:{pattern:/^cusolver64_\d+\.dll$/,fallback:"cusolver64_11.dll"},linux:{pattern:/^libcusolver\.so(\.\d+)*$/,fallback:"libcusolver.so.11"}};l(rF,"winCudnnDirs");Z1={win:{pattern:/^cudnn64_9\.dll$/,fallback:"cudnn64_9.dll",extraDirs:rF},linux:{pattern:/^libcudnn\.so(\.\d+)*$/,fallback:"libcudnn.so.9"}}});import Q1 from"koffi";function $e(n,e){if(e!==0)throw new Error("CUDA driver error "+e+" in "+n)}function tO(n){let e=[0n];li.getProgramLogSize(n,e);let t=Number(e[0]);if(t<=1)return"";let r=new Uint8Array(t);li.getProgramLog(n,r);let s="";for(let o of r){if(o===0)break;s+=String.fromCharCode(o)}return s}var gt,Ka,we,li,X1,eO,ws=M(()=>{"use strict";vl();gt=Q1.load(bs(K1)),Ka=Q1.load(bs(H1)),we={init:gt.func("int cuInit(uint)"),deviceGet:gt.func("int cuDeviceGet(_Out_ int *dev, int ordinal)"),deviceGetAttribute:gt.func("int cuDeviceGetAttribute(_Out_ int *pi, int attrib, int dev)"),primaryCtxRetain:gt.func("int cuDevicePrimaryCtxRetain(_Out_ void **pctx, int dev)"),primaryCtxRelease:gt.func("int cuDevicePrimaryCtxRelease_v2(int dev)"),ctxSynchronize:gt.func("int cuCtxSynchronize()"),moduleLoadData:gt.func("int cuModuleLoadData(_Out_ void **mod, void *image)"),moduleGetFunction:gt.func("int cuModuleGetFunction(_Out_ void **func, void *mod, str name)"),memGetInfo:gt.func("int cuMemGetInfo_v2(_Out_ uint64 *free, _Out_ uint64 *total)"),memAlloc:gt.func("int cuMemAlloc_v2(_Out_ uint64 *dptr, size_t bytes)"),memFree:gt.func("int cuMemFree_v2(uint64 dptr)"),memcpyHtoD:gt.func("int cuMemcpyHtoD_v2(uint64 dst, void *src, size_t n)"),memcpyDtoH:gt.func("int cuMemcpyDtoH_v2(void *dst, uint64 src, size_t n)"),memcpyDtoD:gt.func("int cuMemcpyDtoD_v2(uint64 dst, uint64 src, size_t n)"),memcpyHtoDAsync:gt.func("int cuMemcpyHtoDAsync_v2(uint64 dst, void *src, size_t n, void *stream)"),memcpyDtoHAsync:gt.func("int cuMemcpyDtoHAsync_v2(void *dst, uint64 src, size_t n, void *stream)"),memcpyDtoDAsync:gt.func("int cuMemcpyDtoDAsync_v2(uint64 dst, uint64 src, size_t n, void *stream)"),memsetD8:gt.func("int cuMemsetD8_v2(uint64 dst, uint8 uc, size_t n)"),memsetD8Async:gt.func("int cuMemsetD8Async(uint64 dst, uint8 uc, size_t n, void *stream)"),launchKernel:gt.func("int cuLaunchKernel(void *f, uint gx, uint gy, uint gz, uint bx, uint by, uint bz, uint shmem, void *stream, void **params, void **extra)"),streamCreate:gt.func("int cuStreamCreate(_Out_ void **stream, uint flags)"),streamSynchronize:gt.func("int cuStreamSynchronize(void *stream)"),ctxSetCurrent:gt.func("int cuCtxSetCurrent(void *ctx)"),streamBeginCapture:gt.func("int cuStreamBeginCapture_v2(void *stream, int mode)"),streamEndCapture:gt.func("int cuStreamEndCapture(void *stream, _Out_ void **graph)"),graphInstantiate:gt.func("int cuGraphInstantiateWithFlags(_Out_ void **exec, void *graph, uint64 flags)"),graphLaunch:gt.func("int cuGraphLaunch(void *exec, void *stream)"),graphExecDestroy:gt.func("int cuGraphExecDestroy(void *exec)"),graphDestroy:gt.func("int cuGraphDestroy(void *graph)")},li={createProgram:Ka.func("int nvrtcCreateProgram(_Out_ void **prog, str src, str name, int n, void *h, void *inc)"),compileProgram:Ka.func("int nvrtcCompileProgram(void *prog, int n, str *opts)"),destroyProgram:Ka.func("int nvrtcDestroyProgram(void **prog)"),getPTXSize:Ka.func("int nvrtcGetPTXSize(void *prog, _Out_ size_t *sz)"),getPTX:Ka.func("int nvrtcGetPTX(void *prog, _Out_ uint8_t *ptx)"),getProgramLogSize:Ka.func("int nvrtcGetProgramLogSize(void *prog, _Out_ size_t *sz)"),getProgramLog:Ka.func("int nvrtcGetProgramLog(void *prog, _Out_ uint8_t *log)")},X1=75,eO=76;l($e,"checkCU");l(tO,"readProgramLog")});function He(){if(kl)return $e("cuCtxSetCurrent",we.ctxSetCurrent(kl.ctx)),kl;$e("cuInit",we.init(0));let n=[0];$e("cuDeviceGet",we.deviceGet(n,0));let e=[0],t=[0];$e("cuDeviceGetAttribute",we.deviceGetAttribute(e,X1,n[0])),$e("cuDeviceGetAttribute",we.deviceGetAttribute(t,eO,n[0]));let r=[null];$e("cuDevicePrimaryCtxRetain",we.primaryCtxRetain(r,n[0])),$e("cuCtxSetCurrent",we.ctxSetCurrent(r[0]));let s=[null];$e("cuStreamCreate",we.streamCreate(s,0));let o=[0n],i=[0n];return $e("cuMemGetInfo",we.memGetInfo(o,i)),kl={dev:n[0],ctx:r[0],stream:s[0],arch:"sm_"+e[0]+t[0],totalMem:Number(i[0])},process.on("exit",()=>{try{we.primaryCtxRelease(kl.dev)}catch{}}),kl}var kl,Ts=M(()=>{"use strict";ws();kl=null;l(He,"getDevice")});function Qm(n){let e=2166136261;for(let t=0;t<n.length;t++)e^=n.charCodeAt(t),e=Math.imul(e,16777619);return(e>>>0).toString(16)}function Xm(n,e){let t=Qm(n),r=rO.get(t);if(r)return r;let{arch:s}=He(),o="";n.includes("__half")&&(o+=`#include <cuda_fp16.h>
+`),n.includes("__nv_bfloat16")&&(o+=`#include <cuda_bf16.h>
+`),/u?int(8|16|64)_t/.test(n)&&(o+=oF),/mma_sync|wmma::|fragment</.test(n)&&(o+=`#include <mma.h>
+using namespace nvcuda::wmma;
+`),n.includes("__pipeline_memcpy_async")&&(o+=`#include <cuda_pipeline.h>
+`);let i=o+sF+`extern "C" {
+`+n+`
+}
+`,a=[null];$e("nvrtcCreateProgram",li.createProgram(a,i,e+".cu",0,null,null));let u=["--gpu-architecture="+s];if(Jm&&u.push("--include-path="+Jm),li.compileProgram(a[0],u.length,u)!==0){let d=tO(a[0]);throw li.destroyProgram(a),new Error("NVRTC compile failed for kernel "+e+`:
+`+d+`
+--- source ---
+`+n)}let c=[0n];$e("nvrtcGetPTXSize",li.getPTXSize(a[0],c));let f=new Uint8Array(Number(c[0]));return $e("nvrtcGetPTX",li.getPTX(a[0],f)),li.destroyProgram(a),rO.set(t,f),f}function tT(n){let e=sO.get(n);return e||(e=wo(n.source,n.name),sO.set(n,e),e)}function wo(n,e){let t=Qm(n)+":"+e,r=nO.get(t);if(r)return r;let s=Xm(n,e),o=[null];$e("cuModuleLoadData",we.moduleLoadData(o,s));let i=[null];$e("cuModuleGetFunction",we.moduleGetFunction(i,o[0],e));let a={func:i[0],module:o[0]};return nO.set(t,a),a}var nO,sF,oF,rO,sO,nT=M(()=>{"use strict";ws();Ts();nO=new Map,sF=`#ifndef INFINITY
+#define INFINITY __int_as_float(0x7f800000)
+#endif
+#ifndef NAN
+#define NAN __int_as_float(0x7fffffff)
+#endif
+`,oF=`typedef signed char int8_t;
+typedef short int16_t;
+typedef int int32_t;
+typedef long long int64_t;
+typedef unsigned char uint8_t;
+typedef unsigned short uint16_t;
+typedef unsigned int uint32_t;
+typedef unsigned long long uint64_t;
+`;l(Qm,"hashSource");rO=new Map;l(Xm,"compileToPTX");sO=new WeakMap;l(tT,"getProgramFor");l(wo,"getProgram")});var iO={};at(iO,{acquire:()=>ht,alloc:()=>oO,copyDeviceToHost:()=>qt,copyDeviceToHostAsync:()=>rT,copyHostToDevice:()=>xs,copyHostToDeviceAsync:()=>eg,drainPool:()=>tg,free:()=>ci,release:()=>_t});function oO(n){let e=[0n];return $e("cuMemAlloc",we.memAlloc(e,n)),e[0]}function xs(n,e){$e("cuMemcpyHtoD",we.memcpyHtoD(n,e,e.byteLength))}function qt(n,e){$e("cuMemcpyDtoH",we.memcpyDtoH(n,e,n.byteLength))}function eg(n,e){$e("cuMemcpyHtoDAsync",we.memcpyHtoDAsync(n,e,e.byteLength,He().stream))}function rT(n,e){$e("cuMemcpyDtoHAsync",we.memcpyDtoHAsync(n,e,n.byteLength,He().stream))}function ci(n){we.memFree(n)}function ht(n){let e=Cp.get(n);return e&&e.length>0?e.pop():oO(n)}function _t(n,e){let t=Cp.get(e);t||(t=[],Cp.set(e,t)),t.push(n)}function tg(){let n=0;for(let e of Cp.values())for(let t of e)ci(t),n++;return Cp.clear(),n}var Cp,pi=M(()=>{"use strict";ws();Ts();l(oO,"alloc");l(xs,"copyHostToDevice");l(qt,"copyDeviceToHost");l(eg,"copyHostToDeviceAsync");l(rT,"copyDeviceToHostAsync");l(ci,"free");Cp=new Map;l(ht,"acquire");l(_t,"release");l(tg,"drainPool")});function oT(n){aO=n}function In(){return aO}function ng(n){uO=n}function zn(){return uO}function Nl(n){lO=n}function cO(){return lO}function pO(n){sT=n}function iT(){sT&&sT()}var aO,sT,uO,lO,fi=M(()=>{"use strict";aO=!1,sT=null,uO=!1,lO=!1;l(oT,"setEagerDeferred");l(In,"isEagerDeferred");l(ng,"setEagerCapturing");l(zn,"isEagerCapturing");l(Nl,"setCudaGraphArmed");l(cO,"isCudaGraphArmed");l(pO,"setEagerFlushHook");l(iT,"eagerFlush")});function iF(n){let e=Buffer.alloc(8);return e.writeBigUInt64LE(BigInt(n)),e}function rr(n){return new rg("f32",n)}function aT(n){let e=Buffer.alloc(4);return n instanceof rg?(n.kind==="f32"?e.writeFloatLE(n.value):e.writeInt32LE(n.value|0),e):(Number.isInteger(n)?e.writeInt32LE(n|0):e.writeFloatLE(n),e)}function ns(n,e,t,r,s,o,i=!0){let{stream:a}=He(),u=[];for(let p of s)u.push(iF(p));for(let p of o)u.push(aT(p));$e("cuLaunchKernel",we.launchKernel(n,e[0],e[1],e[2],t[0],t[1],t[2],r,a,u,null)),i&&!zn()&&$e("cuStreamSynchronize",we.streamSynchronize(a))}var rg,uT=M(()=>{"use strict";ws();Ts();fi();l(iF,"devicePtrParam");rg=class{static{l(this,"TypedScalar")}constructor(e,t){this.kind=e,this.value=t}};l(rr,"f32");l(aT,"scalarParam");l(ns,"launch")});import dO from"koffi";function Qi(){aF(0)}function cT(){let n=lF();if(n!==0)throw new Error("cudaDeviceSynchronize failed: "+n)}function fF(n){let e=[null],t=uF(e,n);if(t!==0)throw new Error("cudaMalloc failed: "+t);return e[0]}function Pp(n,e){let t=hO(n,e,e.byteLength,cF);if(t!==0)throw new Error("cudaMemcpy H2D failed: "+t);fO&&fO(e.byteLength)}function sg(n,e){let t=hO(n,e,n.byteLength,pF);if(t!==0)throw new Error("cudaMemcpy D2H failed: "+t)}function pT(n){return typeof n=="bigint"?n:typeof n=="number"?BigInt(n):dO.address(n)}function og(n){let e=lT.get(n);return e&&e.length>0?e.pop():fF(n)}function fT(n,e){let t=lT.get(e);t||(t=[],lT.set(e,t)),t.push(n)}var Dp,aF,uF,hO,s9,lF,cF,pF,fO,lT,ig=M(()=>{"use strict";vl();Dp=dO.load(bs(Zm)),aF=Dp.func("int cudaSetDevice(int d)"),uF=Dp.func("int cudaMalloc(_Out_ void **p, size_t s)"),hO=Dp.func("int cudaMemcpy(void *dst, void *src, size_t n, int kind)"),s9=Dp.func("int cudaFree(void *p)"),lF=Dp.func("int cudaDeviceSynchronize()"),cF=1,pF=2,fO=null;l(Qi,"setDevice");l(cT,"devSync");l(fF,"devAlloc");l(Pp,"devH2D");l(sg,"devD2H");l(pT,"devAddr");lT=new Map;l(og,"acquireDevice");l(fT,"releaseDevice")});var gO={};at(gO,{CU_STREAM_CAPTURE_MODE_GLOBAL:()=>dF,CU_STREAM_CAPTURE_MODE_RELAXED:()=>hF,CU_STREAM_CAPTURE_MODE_THREAD_LOCAL:()=>mO,beginEagerCapture:()=>dT,destroyEagerGraph:()=>gT,endEagerCapture:()=>ag,isGraphCapturing:()=>mF,replay:()=>hT,syncStream:()=>mT});function mF(){return zn()}function dT(n=mO){let{stream:e}=He();ng(!0);try{$e("cuStreamBeginCapture",we.streamBeginCapture(e,n))}catch(t){throw ng(!1),t}}function ag(){let{stream:n}=He(),e=[null],t=we.streamEndCapture(n,e);ng(!1),$e("cuStreamEndCapture",t);let r=[null];return $e("cuGraphInstantiateWithFlags",we.graphInstantiate(r,e[0],0n)),{graph:e[0],exec:r[0]}}function hT(n){let{stream:e}=He();$e("cuGraphLaunch",we.graphLaunch(n,e))}function mT(){let{stream:n}=He();$e("cuStreamSynchronize",we.streamSynchronize(n))}function gT(n){n&&(n.exec&&we.graphExecDestroy(n.exec),n.graph&&we.graphDestroy(n.graph))}var dF,mO,hF,ug=M(()=>{"use strict";ws();Ts();fi();dF=0,mO=1,hF=2;l(mF,"isGraphCapturing");l(dT,"beginEagerCapture");l(ag,"endEagerCapture");l(hT,"replay");l(mT,"syncStream");l(gT,"destroyEagerGraph")});var wO={};at(wO,{cublasGemmBatchedDevice:()=>TT,cublasGemmDevice:()=>zp,cublasMatmul:()=>vF,cublasMatmulDevice:()=>mg});import _O from"koffi";function hg(){if(!$p){let n=[null],e=_F(n);if(e!==0)throw new Error("cublasCreate failed: "+e);$p=n[0],bF($p,He().stream),wF($p,3)}return $p}function _T(n){let e=wT.get(n);if(e&&e.length>0)return e.pop();let t=[null];if(gF(t,n))throw new Error("cudaMalloc failed");return t[0]}function bT(n,e){let t=wT.get(e);t||(t=[],wT.set(e,t)),t.push(n)}function mg(n,e,t,r,s,o,i=!1){pg(0);let a=hg(),c=bO(a,i?1:0,0,e,n,t,fg,s,i?t:e,r,t,dg,o,e);if(c!==0)throw new Error("cublasSgemm failed: "+c)}function zp(n,e,t,r,s,o,i,a){pg(0);let u=hg(),h=TF(u,i?cg:lg,s?cg:lg,e,n,t,fg,o,i?t:e,r,s?n:t,dg,a,e);if(h!==0)throw new Error("cublasSgemm failed: "+h)}function TT(n,e,t,r,s,o,i,a,u,p,c,f){pg(0);let d=hg(),h=i?cg:lg,m=p?cg:lg,g=i?e:r,_=xF(d,m,h,t,e,r,fg,a,p?r:t,BigInt(u),s,g,BigInt(o),dg,c,t,BigInt(f),n);if(_!==0)throw new Error("cublasSgemmStridedBatched failed: "+_)}function vF(n,e,t,r,s,o,i=!1){pg(0);let a=hg(),u=_T(r.byteLength),p=_T(s.byteLength),c=_T(o.byteLength);try{yT(u,r,r.byteLength,yO),yT(p,s,s.byteLength,yO);let h=bO(a,i?1:0,0,e,n,t,fg,p,i?t:e,u,t,dg,c,e);if(h!==0)throw new Error("cublasSgemm failed: "+h);yF(),yT(o,c,o.byteLength,SF)}finally{bT(u,r.byteLength),bT(p,s.byteLength),bT(c,o.byteLength)}}var Fp,Al,pg,gF,yT,h9,yF,_F,bF,wF,bO,TF,xF,yO,SF,lg,cg,fg,dg,$p,wT,Vp=M(()=>{"use strict";Ts();vl();Fp=_O.load(bs(Zm)),Al=_O.load(bs(Y1)),pg=Fp.func("int cudaSetDevice(int d)"),gF=Fp.func("int cudaMalloc(_Out_ void **p, size_t s)"),yT=Fp.func("int cudaMemcpy(void *dst, void *src, size_t n, int kind)"),h9=Fp.func("int cudaFree(void *p)"),yF=Fp.func("int cudaDeviceSynchronize()"),_F=Al.func("int cublasCreate_v2(_Out_ void **h)"),bF=Al.func("int cublasSetStream_v2(void *h, void *streamId)"),wF=Al.func("int cublasSetMathMode(void *h, int mode)"),bO=Al.func("int cublasSgemm_v2(void *h, int ta, int tb, int m, int n, int k, float *alpha, void *A, int lda, void *B, int ldb, float *beta, void *C, int ldc)"),TF=Al.func("int cublasSgemm_v2(void *h, int ta, int tb, int m, int n, int k, float *alpha, uint64 A, int lda, uint64 B, int ldb, float *beta, uint64 C, int ldc)"),xF=Al.func("int cublasSgemmStridedBatched(void *h, int ta, int tb, int m, int n, int k, float *alpha, uint64 A, int lda, int64 sa, uint64 B, int ldb, int64 sb, float *beta, uint64 C, int ldc, int64 sc, int batch)"),yO=1,SF=2,lg=0,cg=1,fg=new Float32Array([1]),dg=new Float32Array([0]),$p=null;l(hg,"handle");wT=new Map;l(_T,"acquire");l(bT,"release");l(mg,"cublasMatmulDevice");l(zp,"cublasGemmDevice");l(TT,"cublasGemmBatchedDevice");l(vF,"cublasMatmul")});function vO(n){xT=n}function kO(){return xT}function NO(n,e){let t=Qm(n)+":"+e,r=TO.get(t);if(r)return r;let s=Xm(n,e);Qi();let o=[null];$e("cuModuleLoadData",we.moduleLoadData(o,s));let i=[null];return $e("cuModuleGetFunction",we.moduleGetFunction(i,o[0],e)),TO.set(t,i[0]),i[0]}function kF(n){let e=Buffer.alloc(8);return e.writeBigUInt64LE(n),e}function AO(n,e,t,r,s,o,i=null){let a=[];for(let u of s)a.push(kF(u));for(let u of o)a.push(aT(u));$e("cuLaunchKernel",we.launchKernel(n,e[0],e[1],e[2],t[0],t[1],t[2],r,i,a,null))}function NF(n,e,t){let r=n.inputSlots.concat(n.outputSlots),s=n.kernel.metadata;if(s.cublas){let{M:o,N:i,K:a,aIdx:u,bIdx:p,cIdx:c,transB:f}=s.cublas;mg(o,i,a,e[r[u]],e[r[p]],e[r[c]],f)}else{let o=r.map(i=>pT(e[i]));AO(AF(n),s.gridDim,s.blockDim,0,o,n.shapeValues||[],t)}}function AF(n){return OO.get(n.name)}function OF(n,e,t){Qi();let r=He().stream,s=new Map;for(let u of t)u.kernel.metadata.cublas||s.set(u.name,NO(u.kernel.source,u.kernel.name));OO=s;let o=new Set;for(let u of t)for(let p of u.outputSlots)o.add(p);let i=new Set(n.argSlots),a=xO.get(n);if(a)for(let u=0;u<n.numSlots;u++)e[u]&&!o.has(u)&&Pp(a.dptr[u],e[u].data);else{let u=new Array(n.numSlots).fill(null);for(let p=0;p<n.numSlots;p++){let c=e[p];c&&(u[p]=og(Math.max(c.data.byteLength,1)),o.has(p)||Pp(u[p],c.data))}dT();try{for(let c of t)NF(c,u,r);let p=ag();a={dptr:u,graph:p.graph,exec:p.exec}}catch(p){try{ag()}catch{}throw p}xO.set(n,a)}hT(a.exec),mT();for(let u=0;u<n.numSlots;u++)a.dptr[u]!==null&&o.has(u)&&i.has(u)&&sg(e[u].data,a.dptr[u])}function BO(n,e,t){for(let r of n){let s=r.inputSlots.concat(r.outputSlots),o=r.kernel.metadata;if(o.cublas){let{M:i,N:a,K:u,aIdx:p,bIdx:c,cIdx:f,transB:d}=o.cublas;mg(i,a,u,e[s[p]],e[s[c]],e[s[f]],d)}else{let i=s.map(a=>pT(e[a]));AO(t.get(r.name),o.gridDim,o.blockDim,0,i,r.shapeValues||[])}}}function BF(n,e,t,r,s){let o=SO.get(n);o||(o={dptr:new Array(n.numSlots).fill(null),sizes:new Array(n.numSlots).fill(0),pin:new Map},SO.set(n,o));let{dptr:i,sizes:a,pin:u}=o,p=new Set(n.argSlots);for(let c=0;c<n.numSlots;c++){let f=e[c];if(!f)continue;let d=Math.max(f.data.byteLength,1);if((i[c]===null||a[c]!==d)&&(i[c]!==null&&fT(i[c],a[c]),i[c]=og(d),a[c]=d,u.delete(c)),s.has(c))continue;let h=f.resident;if(h){let m=u.get(c);if(m&&m.key===h.key&&m.version===h.version)continue;u.set(c,{key:h.key,version:h.version})}Pp(i[c],f.data)}BO(t,i,r),cT();for(let c=0;c<n.numSlots;c++)i[c]!==null&&s.has(c)&&p.has(c)&&sg(e[c].data,i[c])}async function RO(n,e,t,r){if(xT)return OF(n,e,t);for(let p of t)p.kernel.metadata.cublas||Xm(p.kernel.source,p.kernel.name);Qi();let s=new Map;for(let p of t)p.kernel.metadata.cublas||s.set(p.name,NO(p.kernel.source,p.kernel.name));let o=new Set;for(let p of t)for(let c of p.outputSlots)o.add(c);if(r&&r.resident){BF(n,e,t,s,o);return}let i=new Array(n.numSlots).fill(null),a=new Array(n.numSlots).fill(0);for(let p=0;p<n.numSlots;p++){let c=e[p];if(!c)continue;let f=Math.max(c.data.byteLength,1);a[p]=f,i[p]=og(f),o.has(p)||Pp(i[p],c.data)}BO(t,i,s),cT();let u=new Set(n.argSlots);for(let p=0;p<n.numSlots;p++)i[p]!==null&&o.has(p)&&u.has(p)&&sg(e[p].data,i[p]);for(let p=0;p<n.numSlots;p++)i[p]!==null&&fT(i[p],a[p])}var TO,xO,SO,xT,OO,MO=M(()=>{"use strict";ws();nT();ig();Ts();ug();Vp();uT();TO=new Map,xO=new WeakMap,SO=new WeakMap,xT=!1;l(vO,"setCudaGraphEnabled");l(kO,"isCudaGraphEnabled");l(NO,"loadFunctionOnPrimary");l(kF,"bufferParam");l(AO,"launchOnPrimary");l(NF,"_runStepGraphed");OO=null;l(AF,"funcsFor");l(OF,"runCudaPlanGraphed");l(BO,"_launchSteps");l(BF,"_runCudaPlanResident");l(RO,"runCudaPlan")});var EO={};at(EO,{clearCapturePins:()=>LF,deviceBufferDptr:()=>CF,deviceBufferForInplace:()=>gg,deviceBufferForInput:()=>To,deviceBufferForOutput:()=>Xi,downloadAndValidate:()=>vT,downloadStaticAsync:()=>PF,flushDeferred:()=>yg,hostReadHook:()=>BT,isEagerDeferred:()=>In,pinResident:()=>kT,releaseAllResident:()=>OT,setCudaGraphArmed:()=>Nl,setEagerDeferred:()=>oT,unpinResident:()=>EF,uploadIfStale:()=>Ha,uploadStaticAsync:()=>DF});function MF(n){let e=Ss.get(n);if(e)return Ss.delete(n),Ss.set(n,e),e;if(Ss.size>=RF){let t=Ss.keys().next().value,r=Ss.get(t);Ss.delete(t),_t(r.dptr,r.bytes)}return e={dptr:ht(n.byteLength),bytes:n.byteLength,valid:!1},Ss.set(n,e),e}function Ha(n){let e=MF(n);return e.valid||(xs(e.dptr,n),e.valid=!0),e.dptr}function vT(n,e){qt(n,e);let t=Ss.get(n);t&&(t.valid=!0)}function kT(n){jp.add(n)}function EF(n){jp.delete(n)}function LF(){Up.clear()}function CF(n){let e=vs.get(n);return e?e.dptr:null}function DF(n,e){eg(n,e)}function PF(n,e){rT(n,e)}function $F(){return ST||(ST=Math.floor(He().totalMem*IF)),ST}function NT(n,e){if(zn())throw new Error("illegal device sync during CUDA graph capture");$e("cuStreamSynchronize",we.streamSynchronize(He().stream)),qt(n,e)}function FF(n){let e=$F();for(let[t,r]of vs){if(Gp+n<=e)break;jp.has(t)||Up.has(t)||(r.hostStale&&NT(t,r.dptr),_t(r.dptr,r.bytes),Gp-=r.bytes,vs.delete(t))}}function AT(n){let e=vs.get(n);return e?(vs.delete(n),vs.set(n,e)):(zn()||FF(n.byteLength),e={dptr:ht(n.byteLength),bytes:n.byteLength,deviceFresh:!1,hostStale:!1},vs.set(n,e),Gp+=e.bytes),zn()&&Up.add(n),e}function IO(n,e){zn()?eg(n,e):xs(n,e)}function To(n){let e=AT(n);return e.deviceFresh||(IO(e.dptr,n),e.deviceFresh=!0),e.dptr}function Xi(n){let e=AT(n);return e.deviceFresh=!0,e.hostStale=!0,e.dptr}function gg(n){let e=AT(n);return e.deviceFresh||(IO(e.dptr,n),e.deviceFresh=!0),e.hostStale=!0,e.dptr}function yg(){for(let[n,e]of vs)jp.has(n)||Up.has(n)||(_t(e.dptr,e.bytes),Gp-=e.bytes,vs.delete(n))}function OT(){for(let[n,e]of vs)e.hostStale&&NT(n,e.dptr),_t(e.dptr,e.bytes);vs.clear(),Gp=0,jp.clear(),Up.clear();for(let n of Ss.values())_t(n.dptr,n.bytes);Ss.clear()}function BT(n){let e=vs.get(n);if(zn()){if(e&&e.hostStale)throw new Error("illegal host read of device-resident tensor during CUDA graph capture");return}let t=Ss.get(n);t&&(t.valid=!1),e&&(e.hostStale&&(NT(n,e.dptr),e.hostStale=!1),e.deviceFresh=!1)}var RF,Ss,IF,vs,jp,Up,Gp,ST,qp=M(()=>{"use strict";pi();Ts();ws();fi();RF=1024,Ss=new Map;l(MF,"_safeEntry");l(Ha,"uploadIfStale");l(vT,"downloadAndValidate");IF=.5,vs=new Map,jp=new Set,Up=new Set,Gp=0,ST=0;l(kT,"pinResident");l(EF,"unpinResident");l(LF,"clearCapturePins");l(CF,"deviceBufferDptr");l(DF,"uploadStaticAsync");l(PF,"downloadStaticAsync");l($F,"_cap");l(NT,"_syncDownload");l(FF,"_evict");l(AT,"_defEntry");l(IO,"_upload");l(To,"deviceBufferForInput");l(Xi,"deviceBufferForOutput");l(gg,"deviceBufferForInplace");l(yg,"flushDeferred");l(OT,"releaseAllResident");l(BT,"hostReadHook");pO(yg);oT(!0)});function St(n){return VF[n]||Float32Array}function _g(n,e){if(n===e)return n;let t=Ol.get(n)??0,r=Ol.get(e)??0;return t>=r?n:e}function st(n,e){let t=DO.get(n);t||(t=new Set,DO.set(n,t)),t.add(e)}function Ya(n){return zF(n)}var K,zF,VF,LO,CO,Ol,DO,rs=M(()=>{"use strict";Re();K=er,zF=fo,VF=Object.freeze({[K.F16]:Uint16Array,[K.BF16]:Uint16Array,[K.F32]:Float32Array,[K.F64]:Float64Array,[K.I8]:Int8Array,[K.I16]:Int16Array,[K.I32]:Int32Array,[K.I64]:BigInt64Array,[K.UI8]:Uint8Array,[K.BOOL]:Uint8Array,[K.INDEX]:Int32Array});l(St,"typedArrayCtor");LO=[K.F16,K.BF16,K.F32,K.F64],CO=[K.UI8,K.I8,K.I16,K.I32,K.I64],Ol=new Map;for(let n=0;n<LO.length;n++)Ol.set(LO[n],100+n);for(let n=0;n<CO.length;n++)Ol.set(CO[n],n);Ol.set(K.BOOL,-1);Ol.set(K.INDEX,50);l(_g,"resultDtype");DO=new Map;l(st,"_addCastPair");st(K.BOOL,K.I8);st(K.BOOL,K.I16);st(K.BOOL,K.I32);st(K.BOOL,K.I64);st(K.BOOL,K.F16);st(K.BOOL,K.F32);st(K.BOOL,K.F64);st(K.UI8,K.I16);st(K.UI8,K.I32);st(K.UI8,K.I64);st(K.UI8,K.F16);st(K.UI8,K.F32);st(K.UI8,K.F64);st(K.I8,K.I16);st(K.I8,K.I32);st(K.I8,K.I64);st(K.I8,K.F16);st(K.I8,K.F32);st(K.I8,K.F64);st(K.I16,K.I32);st(K.I16,K.I64);st(K.I16,K.F32);st(K.I16,K.F64);st(K.I32,K.I64);st(K.I32,K.F64);st(K.F16,K.F32);st(K.F16,K.F64);st(K.BF16,K.F32);st(K.BF16,K.F64);st(K.BOOL,K.BF16);st(K.UI8,K.BF16);st(K.I8,K.BF16);st(K.F32,K.F64);l(Ya,"dtypeSize")});function ea(){return PO}function GF(n){PO=n}var Ja,Ws,bg,RT,xo,Bl,PO,qn=M(()=>{"use strict";An();Ja=class{static{l(this,"Device")}type;index;constructor(e,t=0){this.type=e,this.index=t}dispatchKey(){return Uu(this.type)}equals(e){return this.type===e.type&&this.index===e.index}hash(){let e=2166136261;for(let t=0;t<this.type.length;t++)e=(e^this.type.charCodeAt(t))*16777619&2147483647;return e=(e^this.index)*16777619&2147483647,e}toString(){return this.index===0?this.type:`${this.type}:${this.index}`}},Ws=new Ja("cpu"),bg=new Ja("gpu"),RT=new Ja("wasm"),xo=new Ja("webgpu"),Bl=new Ja("meta"),PO=Ws;l(ea,"getDefaultDevice");l(GF,"setDefaultDevice")});function MT(n){return $O.get(n)||$O.get("cpu")}var wg,Tg,$O,Za,IT=M(()=>{"use strict";rs();qn();wg=class{static{l(this,"CPUAllocator")}allocate(e,t){let r=St(t);return new r(e)}free(e){}},Tg=class{static{l(this,"MetaAllocator")}allocate(){return null}free(e){}},$O=new Map([["cpu",new wg],["wasm",new wg],["meta",new Tg],["lazy",new Tg]]);l(MT,"getAllocator");Za=class n{static{l(this,"StorageImpl")}static#e=null;_data;_nbytes;_device;_allocator;_refCount;static setHostReadHook(e){n.#e=e}constructor(e,t,r,s){this._data=e,this._nbytes=t,this._device=r,this._allocator=s,this._refCount=1}static allocate(e,t,r){let s=MT(r.type),o=Ya(t),i=Math.max(Math.ceil(e/o),1),a=s.allocate(i,t);return new n(a,e,r,s)}static fromData(e,t){let r=e?e.byteLength:0,s=MT(t.type);return new n(e,r,t,s)}retain(){return this._refCount++,this}release(){this._refCount--,this._refCount===0&&(this._allocator&&this._data&&this._allocator.free(this._data),this._data=null)}get refCount(){return this._refCount}get data(){return n.#e&&this._data&&n.#e(this._data),this._data}get rawData(){return this._data}get nbytes(){return this._nbytes}get device(){return this._device}get isValid(){return this._refCount>0&&this._data!==null}get isMeta(){return this._data===null}resize(e,t){if(e<=this._nbytes&&this._data)return;let r=Ya(t),s=Math.max(Math.ceil(e/r),1),o=this._allocator.allocate(s,t);if(this._data&&o){let i=Math.min(this._data.length,o.length);for(let a=0;a<i;a++)o[a]=this._data[a]}this._allocator&&this._data&&this._allocator.free(this._data),this._data=o,this._nbytes=e}clone(){let e=MT(this._device.type),t=null;if(this._data){n.#e&&n.#e(this._data);let r=this._data.constructor;t=new r(this._data.length),t.set(this._data)}return new n(t,this._nbytes,this._device,e)}}});function zO(n){let e=FO.get(n)??null;return e||(e=Lt.findOp(n),e&&FO.set(n,e)),e||null}var FO,VO=M(()=>{"use strict";Ma();FO=new Map;l(zO,"getHandle")});function Ks(n,e){return n<0?e+n:n}function En(n){let e=n.length,t=new Array(e),r=1;for(let s=e-1;s>=0;s--)t[s]=r,r*=n[s];return t}function bn(n){let e=1;for(let t=0;t<n.length;t++)e*=n[t];return e}function ET(n,e){let t=n.length;if(t===0)return!0;let r=1;for(let s=t-1;s>=0;s--){let o=n[s];if(o===0)return!0;if(o!==1&&e[s]!==r)return!1;r*=o}return!0}function Rl(n,e){let t=n.length,r=e.length,s=Math.max(t,r),o=new Array(s);for(let i=0;i<s;i++){let a=i<t?n[t-1-i]:1,u=i<r?e[r-1-i]:1;if(a===u)o[s-1-i]=a;else if(a===1)o[s-1-i]=u;else if(u===1)o[s-1-i]=a;else return null}return o}function xg(n,e){let t=n.length,r=e.length;if(t===1&&r===1)return[];let s=t===1?[1,n[0]]:[...n],o=r===1?[e[0],1]:[...e],i=s.length,a=o.length,u=Rl(s.slice(0,i-2),o.slice(0,a-2));if(u===null)return null;let p=[...u,s[i-2],o[a-1]],c=new Set;return r===1&&c.add(p.length-1),t===1&&c.add(p.length-2),p.filter((f,d)=>!c.has(d))}function GO(n,e,t){let r=bn(n),s=-1,o=1,i=new Array(t.length);for(let a=0;a<t.length;a++){let u=t[a];if(u===-1){if(s!==-1)return null;s=a,i[a]=-1}else i[a]=u,o*=u}if(s!==-1){if(o===0||(i[s]=r/o|0,i[s]*o!==r))return null}else if(bn(i)!==r)return null;return ET(n,e)?{sizes:i,strides:En(i),needsCopy:!1}:{sizes:i,strides:En(i),needsCopy:!0}}var ss=M(()=>{"use strict";l(Ks,"normalizeAxis");l(En,"computeStrides");l(bn,"computeNumel");l(ET,"isContiguous");l(Rl,"broadcastShapes");l(xg,"matmulOutputShape");l(GO,"inferReshape")});function jO(n,e){if(n.length<4)return!1;let t=n.length,r=1,s=t===4?jF:UF;if(!s)return!1;for(let o=t-1;o>=0;o--){let i=s[o];if(i===void 0||n[i]===0||e[i]!==r)return!1;r*=n[i]}return!0}var jF,UF,LT=M(()=>{"use strict";jF=Object.freeze([0,2,3,1]),UF=Object.freeze([0,2,3,4,1]);l(jO,"isChannelsLast")});var Pt,di=M(()=>{"use strict";An();ss();LT();LT();Pt=class{static{l(this,"TensorImpl")}_storage;_storageOffset;_sizes;_strides;_dtype;_device;_numel;_keySet;_autogradMeta;_version;_contiguousCache;constructor(e,t,r,s,o,i){this._storage=e,this._storageOffset=t,this._sizes=Object.freeze([...r]),this._strides=Object.freeze(s?[...s]:En(r)),this._dtype=o,this._device=i,this._numel=bn(r),this._keySet=Jr.fromKey(Uu(i.type)),this._autogradMeta=null,this._version=0,this._contiguousCache=null}get storage(){return this._storage}get storageOffset(){return this._storageOffset}size(e){let t=e<0?this._sizes.length+e:e;return this._sizes[t]}stride(e){let t=e<0?this._strides.length+e:e;return this._strides[t]}sizes(){return this._sizes}strides(){return this._strides}dim(){return this._sizes.length}numel(){return this._numel}get dtype(){return this._dtype}get device(){return this._device}isContiguous(e){return!e||e==="contiguous"?(this._contiguousCache===null&&(this._contiguousCache=ET(this._sizes,this._strides)),this._contiguousCache):e==="channels_last"?jO(this._sizes,this._strides):!1}setSizesAndStrides(e,t){this._sizes=Object.freeze([...e]),this._strides=Object.freeze(t?[...t]:En(e)),this._numel=bn(e),this._contiguousCache=null}setStorageOffset(e){this._storageOffset=e}bumpVersion(){this._version++}get version(){return this._version}get autogradMeta(){return this._autogradMeta}setAutogradMeta(e){this._autogradMeta=e,this._updateKeySet()}keySet(){return this._keySet}addKeyToSet(e){this._keySet=this._keySet.add(e)}removeKeyFromSet(e){this._keySet=this._keySet.remove(e)}_updateKeySet(){let e=Jr.fromKey(Uu(this._device.type));if(this._autogradMeta&&this._autogradMeta.requiresGrad){let t=Uu(this._device.type);e=e.add(xN(t))}this._keySet=e}get isMeta(){return this._storage&&this._storage.isMeta}shallowCopyFrom(e){this._storage=e._storage,this._storageOffset=e._storageOffset,this._sizes=e._sizes,this._strides=e._strides,this._dtype=e._dtype,this._device=e._device,this._numel=e._numel,this._keySet=e._keySet,this._contiguousCache=null}}});var So,Wp=M(()=>{"use strict";So=class{static{l(this,"AutogradMeta")}_grad;_gradFn;_outputNr;_requiresGrad;_retainGrad;_gradAccumulator;_versionAtCreation;constructor(){this._grad=null,this._gradFn=null,this._outputNr=0,this._requiresGrad=!1,this._retainGrad=!1,this._gradAccumulator=null,this._versionAtCreation=0}get grad(){return this._grad}set grad(e){this._grad=e}get gradFn(){return this._gradFn}setGradFn(e,t){this._gradFn=e,this._outputNr=t??0}get outputNr(){return this._outputNr}get requiresGrad(){return this._requiresGrad}set requiresGrad(e){this._requiresGrad=e}get retainGrad(){return this._retainGrad}set retainGrad(e){this._retainGrad=e}get isLeaf(){return this._gradFn===null}get versionAtCreation(){return this._versionAtCreation}set versionAtCreation(e){this._versionAtCreation=e}getGradAccumulator(){if(this._gradAccumulator){let e=this._gradAccumulator instanceof WeakRef?this._gradAccumulator.deref():this._gradAccumulator;if(e)return e}return null}setGradAccumulator(e){this._gradAccumulator=typeof WeakRef<"u"?new WeakRef(e):e}clearGrad(){this._grad=null}clearGradFn(){this._gradFn=null,this._outputNr=0}}});function qO(n,e,t,r,s,o){let i=e[s];if(s===e.length-1){let u=new Array(i);if(o)for(let p=0;p<i;p++)u[p]=Pm(o,n[r+p*t[s]]);else for(let p=0;p<i;p++)u[p]=n[r+p*t[s]];return u}let a=new Array(i);for(let u=0;u<i;u++)a[u]=qO(n,e,t,r+u*t[s],s+1,o);return a}function qF(){if(!CT)throw new Error("Autograd engine not initialized. Import autograd/engine.js first.");return CT}function WO(n){CT=n}var UO,Ze,CT,sr=M(()=>{"use strict";di();Wp();bl();UO=new Set(["f16","bf16","i64"]),Ze=class n{static{l(this,"Tensor")}_impl;constructor(e){this._impl=e}get impl(){return this._impl}get shape(){return this._impl.sizes()}get strides(){return this._impl.strides()}get dtype(){return this._impl.dtype}get device(){return this._impl.device}get ndim(){return this._impl.dim()}get rank(){return this._impl.dim()}get numel(){return this._impl.numel()}get length(){return this._impl.numel()}get isContiguous(){return this._impl.isContiguous()}get dispatchKeySet(){return this._impl.keySet()}get storage(){return this._impl.storage}get storageOffset(){return this._impl.storageOffset}get data(){let e=this._impl.storage;if(!e||e.isMeta)return null;let t=e.data,r=this._impl.storageOffset;return t&&r===0&&this._impl.isContiguous()&&t.length===this.numel,t}get requiresGrad(){let e=this._impl.autogradMeta;return e?e.requiresGrad:!1}get gradFn(){let e=this._impl.autogradMeta;return e?e.gradFn:null}get grad(){let e=this._impl.autogradMeta;return e?e.grad:null}set grad(e){this._ensureAutogradMeta().grad=e}get isLeaf(){let e=this._impl.autogradMeta;return e?e.isLeaf:!0}get version(){return this._impl.version}requiresGrad_(e=!0){let t=this._ensureAutogradMeta();return t.requiresGrad=e,this._impl._updateKeySet(),this}retainGrad(){return this._ensureAutogradMeta().retainGrad=!0,this}detach(){let e=new Pt(this._impl.storage,this._impl.storageOffset,this._impl.sizes(),this._impl.strides(),this._impl.dtype,this._impl.device);return new n(e)}item(){if(this.numel!==1)throw new Error(`item() requires tensor with exactly 1 element, got ${this.numel}`);let t=this._impl.storage.data[this._impl.storageOffset];return UO.has(this._impl.dtype)?Pm(this._impl.dtype,t):t}toArray(){let e=this.shape,t=this.strides,r=this._impl.storage.data,s=this._impl.storageOffset,o=UO.has(this._impl.dtype)?this._impl.dtype:null;return e.length===0?o?Pm(o,r[s]):r[s]:qO(r,e,t,s,0,o)}toString(){return`Tensor(shape=[${this.shape.join(", ")}], dtype=${this.dtype}, device=${this.device})`}*[Symbol.iterator](){let e=this.shape[0];if(e===void 0)throw new Error("Cannot iterate over a 0-d tensor");for(let t=0;t<e;t++)yield this._select(0,t)}_select(e,t){let r=this._impl.sizes(),s=this._impl.strides(),o=[],i=[];for(let p=0;p<r.length;p++)p!==e&&(o.push(r[p]),i.push(s[p]));let a=this._impl.storageOffset+t*s[e],u=new Pt(this._impl.storage,a,o,i,this._impl.dtype,this._impl.device);return new n(u)}_ensureAutogradMeta(){if(!this._impl.autogradMeta){let e=new So;e.versionAtCreation=this._impl.version,this._impl.setAutogradMeta(e)}return this._impl.autogradMeta}backward(e){let{backward:t}=qF();t(this,e)}};l(qO,"_toNestedArray");CT=null;l(qF,"require_autograd_engine");l(WO,"setAutogradEngine")});var Wn,Xa=M(()=>{"use strict";IT();Wn=class n{static{l(this,"Storage")}_impl;constructor(e){this._impl=e}static allocate(e,t,r){return new n(Za.allocate(e,t,r))}static fromData(e,t){return new n(Za.fromData(e,t))}get impl(){return this._impl}get data(){return this._impl.data}get rawData(){return this._impl.rawData}get nbytes(){return this._impl.nbytes}get device(){return this._impl.device}get isValid(){return this._impl.isValid}get isMeta(){return this._impl.isMeta}retain(){return this._impl.retain(),this}release(){this._impl.release()}clone(){return new n(this._impl.clone())}resize(e,t){this._impl.resize(e,t)}}});function Wt(n,e){let t=e?.dtype??K.F32,r=e?.device??ea(),s=e?.requiresGrad??!1;if(ArrayBuffer.isView(n)&&"length"in n)return WF(n,e?.shape,t,r,s);if(Array.isArray(n)){let{flat:o,shape:i}=HF(n),a=e?.shape??i;return KF(o,a,t,r,s)}if(typeof n=="number")return KO(n,t,r,s);throw new Error("Unsupported data type for tensor()")}function eu(n,e,t,r){let s=r?.device??ea(),o=r?.requiresGrad??!1,i=En(e),a=Wn.fromData(n,s),u=r?.offset??0,p=new Pt(a,u,e,i,t,s),c=new Ze(p);return o&&c.requiresGrad_(!0),c}function Kp(n,e){let t=e?.dtype??K.F32,r=e?.device??ea();return KO(n,t,r,e?.requiresGrad??!1)}function KO(n,e,t,r){let s=St(e),o=new s(1),i=o;i[0]=DT.has(e)?Dm(e,n):n;let a=Wn.fromData(o,t),u=new Pt(a,0,[],[],e,t),p=new Ze(u);return r&&p.requiresGrad_(!0),p}function WF(n,e,t,r,s){let o=St(t),i=new o(n.length),a=i;if(DT.has(t))for(let h=0;h<n.length;h++)a[h]=Dm(t,n[h]);else for(let h=0;h<n.length;h++)a[h]=n[h];let u=e??[n.length],p=En(u),c=Wn.fromData(i,r),f=new Pt(c,0,u,p,t,r),d=new Ze(f);return s&&d.requiresGrad_(!0),d}function KF(n,e,t,r,s){let o=St(t),i=new o(n.length),a=i;if(DT.has(t))for(let d=0;d<n.length;d++)a[d]=Dm(t,n[d]);else for(let d=0;d<n.length;d++)a[d]=n[d];let u=En(e),p=Wn.fromData(i,r),c=new Pt(p,0,e,u,t,r),f=new Ze(c);return s&&f.requiresGrad_(!0),f}function HF(n){let e=[],t=n;for(;Array.isArray(t);)e.push(t.length),t=t[0];let r=[];return HO(n,r,e,0),{flat:r,shape:e}}function HO(n,e,t,r){if(r===t.length-1){for(let s=0;s<n.length;s++)e.push(n[s]);return}for(let s=0;s<n.length;s++)HO(n[s],e,t,r+1)}var DT,os=M(()=>{"use strict";sr();di();Xa();rs();qn();ss();bl();l(Wt,"tensor");l(eu,"fromBuffer");l(Kp,"scalar");DT=new Set(["f16","bf16","i64"]);l(KO,"_fromScalar");l(WF,"_fromTypedArray");l(KF,"_fromFlatArray");l(HF,"_flattenNested");l(HO,"_recurFlatten")});var Ht={};at(Ht,{_dispatch:()=>ct,abs:()=>Hp,add:()=>ae,argmax:()=>Cg,argmin:()=>iB,argsort:()=>gB,batch_norm:()=>$T,broadcast_in_dim:()=>cB,cat:()=>ru,ceil:()=>nB,chunk:()=>pB,clamp:()=>Mg,clone:()=>lB,contiguous:()=>ls,conv2d:()=>FT,cos:()=>ZO,cumsum:()=>hB,div:()=>Kt,dot:()=>uB,embedding:()=>QF,eq:()=>Yp,erf:()=>QO,erfc:()=>XO,exp:()=>as,expand:()=>Ll,fill:()=>ZF,flip:()=>dB,floor:()=>tB,gamma:()=>kg,gather:()=>Eg,ge:()=>Rg,gelu:()=>Ng,gt:()=>Il,index_select:()=>ra,layer_norm:()=>PT,le:()=>Bg,lgamma:()=>eB,log:()=>ko,log_softmax:()=>$g,lt:()=>sB,matmul:()=>tt,max:()=>No,maximum:()=>ta,mean:()=>Tr,min:()=>oB,minimum:()=>vg,mul:()=>ee,narrow:()=>ir,ne:()=>rB,neg:()=>is,one_hot:()=>Ig,pad:()=>tu,permute:()=>El,pool2d:()=>Fg,pow:()=>vo,prod:()=>aB,relu:()=>Ml,remainder:()=>JF,repeat:()=>Dl,reshape:()=>Hn,roll:()=>fB,rsqrt:()=>YO,scatter:()=>Lg,scatter_add:()=>nu,select:()=>en,sigmoid:()=>wr,sign:()=>Og,silu:()=>Ag,sin:()=>JO,slice:()=>Dg,softmax:()=>mi,sort:()=>mB,split:()=>hi,sqrt:()=>na,squeeze:()=>Cl,stack:()=>us,sub:()=>ze,sum:()=>wt,tanh:()=>Hs,tile:()=>Pg,topk:()=>yB,transpose:()=>vt,unsqueeze:()=>Lr,where:()=>or});function YF(n){return typeof n=="object"&&n!==null&&"_impl"in n}function Kn(n,e){return YF(n)?n:Kp(n,{dtype:e.dtype,device:e.device})}function ct(n,...e){let t=zO(n);if(!t)return Lt.callOp(n,...e);let r=u0(e,t.schema);return Lt.dispatch(t,r,...e)}function se(n,...e){return ct(n,...e)}function Sg(n,...e){return ct(n,...e)}function ae(n,e){return se("add",n,Kn(e,n))}function ze(n,e){return se("sub",n,Kn(e,n))}function ee(n,e){return se("mul",n,Kn(e,n))}function Kt(n,e){return se("div",n,Kn(e,n))}function is(n){return se("neg",n)}function vo(n,e){return se("pow",n,Kn(e,n))}function JF(n,e){return se("rem",n,Kn(e,n))}function ta(n,e){return se("maximum",n,Kn(e,n))}function vg(n,e){return se("minimum",n,Kn(e,n))}function as(n){return se("exp",n)}function ko(n){return se("log",n)}function na(n){return se("sqrt",n)}function YO(n){return se("rsqrt",n)}function Hp(n){return se("abs",n)}function JO(n){return se("sin",n)}function ZO(n){return se("cos",n)}function Hs(n){return se("tanh",n)}function QO(n){return se("erf",n)}function XO(n){return se("erfc",n)}function eB(n){return se("lgamma",n)}function kg(n){return se("gamma",n)}function wr(n){return se("sigmoid",n)}function Ml(n){return se("relu",n)}function Ng(n){return se("gelu",n)}function Ag(n){return se("silu",n)}function Og(n){return se("sign",n)}function tB(n){return se("floor",n)}function nB(n){return se("ceil",n)}function Yp(n,e){return se("eq",n,Kn(e,n))}function rB(n,e){return se("ne",n,Kn(e,n))}function sB(n,e){return se("lt",n,Kn(e,n))}function Bg(n,e){return se("le",n,Kn(e,n))}function Il(n,e){return se("gt",n,Kn(e,n))}function Rg(n,e){return se("ge",n,Kn(e,n))}function or(n,e,t){return se("where",n,e,t)}function Mg(n,e,t){return se("clamp",n,Kn(e,n),Kn(t,n))}function tu(n,e,t,r=0){return se("pad",n,Kn(r,n),e,t)}function Ig(n,e){return se("one_hot",n,e)}function ra(n,e,t){return se("index_select",n,t,e)}function Eg(n,e,t){return se("gather",n,t,e)}function nu(n,e,t,r){return se("scatter_add",n,t,r,e)}function Lg(n,e,t,r){return se("scatter",n,e,t,r)}function wt(n,e,t){return se("sum",n,e,t)}function Tr(n,e,t){return se("mean",n,e,t)}function No(n,e,t){return se("max",n,e,t)}function oB(n,e,t){return se("min",n,e,t)}function Cg(n,e,t){return se("argmax",n,e,t)}function iB(n,e,t){return se("argmin",n,e,t)}function aB(n,e,t){return se("prod",n,e,t)}function tt(n,e){let t=zT();if(t){let r=t(n,e);if(r!==null)return r}return se("matmul",n,e)}function uB(n,e){return se("dot",n,e)}function ru(n,e){return se("cat",n,e)}function us(n,e){return se("stack",n,e)}function lB(n){return se("clone",n)}function ZF(n,e){return se("fill",n,e)}function Hn(n,e){return se("reshape",n,e)}function vt(n,e,t){return se("transpose",n,e,t)}function El(n,e){return se("permute",n,e)}function cB(n,e,t){return se("broadcast_in_dim",n,e,t)}function Ll(n,e){return se("expand",n,e)}function Dg(n,e,t,r=null,s=1){return se("slice",n,e,t,r,s)}function Lr(n,e){return se("unsqueeze",n,e)}function Cl(n,e=null){return se("squeeze",n,e)}function ir(n,e,t,r){return se("narrow",n,e,t,r)}function en(n,e,t){return se("select",n,e,t)}function ls(n){return se("contiguous",n)}function Dl(n,e){return se("repeat",n,e)}function Pg(n,e){return se("tile",n,e)}function hi(n,e,t=0){if(typeof e!="number")return Sg("split",n,e,t);let r=n.shape.length,s=t<0?r+t:t,o=n.shape[s],i=[];for(let a=0;a<o;a+=e)i.push(Math.min(e,o-a));return Sg("split",n,i,t)}function pB(n,e,t=0){return Sg("chunk",n,e,t)}function fB(n,e,t=0){return se("roll",n,e,t)}function dB(n,e){return se("flip",n,Array.isArray(e)?e:[e])}function hB(n,e=0){return se("cumsum",n,e)}function mB(n,e=-1,t=!1){return se("sort",n,e,t)}function gB(n,e=-1,t=!1){return se("argsort",n,e,t)}function yB(n,e,t=-1,r=!0){return Sg("topk",n,e,t,r)}function mi(n,e){return se("softmax",n,e)}function $g(n,e){return se("log_softmax",n,e)}function PT(n,e,t,r,s){return se("layer_norm",n,e,t,r,s)}function $T(n,e,t,r,s,o,i){return se("batch_norm",n,e,t,r,s,o,i)}function FT(n,e,t,r,s,o){return se("conv2d",n,e,t,r,s,o)}function Fg(n,e,t,r,s){return se("pool2d",n,e,t,r,s)}function QF(n,e){return se("embedding",n,e)}var be=M(()=>{"use strict";Ma();VO();os();wn();l(YF,"hasTensorImpl");l(Kn,"_asTensor");l(ct,"_dispatch");l(se,"_dispatchTensor");l(Sg,"_dispatchTensorArray");l(ae,"add");l(ze,"sub");l(ee,"mul");l(Kt,"div");l(is,"neg");l(vo,"pow");l(JF,"remainder");l(ta,"maximum");l(vg,"minimum");l(as,"exp");l(ko,"log");l(na,"sqrt");l(YO,"rsqrt");l(Hp,"abs");l(JO,"sin");l(ZO,"cos");l(Hs,"tanh");l(QO,"erf");l(XO,"erfc");l(eB,"lgamma");l(kg,"gamma");l(wr,"sigmoid");l(Ml,"relu");l(Ng,"gelu");l(Ag,"silu");l(Og,"sign");l(tB,"floor");l(nB,"ceil");l(Yp,"eq");l(rB,"ne");l(sB,"lt");l(Bg,"le");l(Il,"gt");l(Rg,"ge");l(or,"where");l(Mg,"clamp");l(tu,"pad");l(Ig,"one_hot");l(ra,"index_select");l(Eg,"gather");l(nu,"scatter_add");l(Lg,"scatter");l(wt,"sum");l(Tr,"mean");l(No,"max");l(oB,"min");l(Cg,"argmax");l(iB,"argmin");l(aB,"prod");l(tt,"matmul");l(uB,"dot");l(ru,"cat");l(us,"stack");l(lB,"clone");l(ZF,"fill");l(Hn,"reshape");l(vt,"transpose");l(El,"permute");l(cB,"broadcast_in_dim");l(Ll,"expand");l(Dg,"slice");l(Lr,"unsqueeze");l(Cl,"squeeze");l(ir,"narrow");l(en,"select");l(ls,"contiguous");l(Dl,"repeat");l(Pg,"tile");l(hi,"split");l(pB,"chunk");l(fB,"roll");l(dB,"flip");l(hB,"cumsum");l(mB,"sort");l(gB,"argsort");l(yB,"topk");l(mi,"softmax");l($g,"log_softmax");l(PT,"layer_norm");l($T,"batch_norm");l(FT,"conv2d");l(Fg,"pool2d");l(QF,"embedding")});function _B(n,e,t,r){let s=n.length,o=t<0?s+t:t,i=r<0?s+r:r,a=[...n],u=[...e];return a[o]=n[i],a[i]=n[o],u[o]=e[i],u[i]=e[o],{sizes:a,strides:u}}function bB(n,e,t){let r=n.length,s=new Array(r),o=new Array(r);for(let i=0;i<r;i++){let a=t[i],u=a<0?r+a:a;s[i]=n[u],o[i]=e[u]}return{sizes:s,strides:o}}function wB(n,e,t){let r=n.length,s=t.length,o=s-r,i=new Array(s),a=new Array(s);for(let u=0;u<s;u++){let p=u-o;if(p<0)a[u]=t[u],i[u]=0;else if(n[p]===1&&t[u]!==1)a[u]=t[u],i[u]=0;else if(n[p]===t[u]||t[u]===-1)a[u]=n[p],i[u]=e[p];else throw new Error(`Cannot expand size ${n[p]} to ${t[u]} at dim ${u}`)}return{sizes:a,strides:i}}function TB(n,e,t,r,s,o){let i=n.length,a=t<0?i+t:t,u=n[a],p=r??0,c=s??u,f=o??1;p<0&&(p+=u),c<0&&(c+=u),p=Math.max(0,Math.min(p,u)),c=Math.max(0,Math.min(c,u));let d=Math.max(0,Math.ceil((c-p)/f)),h=p*e[a],m=[...n],g=[...e];return m[a]=d,g[a]=e[a]*f,{sizes:m,strides:g,offsetDelta:h}}function xB(n,e,t){let r=n.length,s=t<0?r+1+t:t,o=[...n],i=[...e],a=s<r?n[s]*e[s]:1;return o.splice(s,0,1),i.splice(s,0,a),{sizes:o,strides:i}}function SB(n,e,t){if(t!=null){let o=n.length,i=t<0?o+t:t;if(n[i]!==1)return{sizes:[...n],strides:[...e]};let a=[...n],u=[...e];return a.splice(i,1),u.splice(i,1),{sizes:a,strides:u}}let r=[],s=[];for(let o=0;o<n.length;o++)n[o]!==1&&(r.push(n[o]),s.push(e[o]));return{sizes:r,strides:s}}function vB(n,e,t,r,s){let o=n.length,i=t<0?o+t:t,a=[...n];a[i]=s;let u=r*e[i];return{sizes:a,strides:[...e],offsetDelta:u}}function kB(n,e,t,r){let s=n.length,o=t<0?s+t:t,i=r<0?n[o]+r:r,a=[],u=[];for(let c=0;c<s;c++)c!==o&&(a.push(n[c]),u.push(e[c]));let p=i*e[o];return{sizes:a,strides:u,offsetDelta:p}}var NB=M(()=>{"use strict";l(_B,"computeTranspose");l(bB,"computePermute");l(wB,"computeExpand");l(TB,"computeSlice");l(xB,"computeUnsqueeze");l(SB,"computeSqueeze");l(vB,"computeNarrow");l(kB,"computeSelect")});function Ys(n,e,t,r){let s=new Pt(n._impl.storage,n._impl.storageOffset+(r||0),e,t,n._impl.dtype,n._impl.device),o=n._impl.autogradMeta;return o&&s.setAutogradMeta(o),new Ze(s)}function XF(n,e){let t=GO(n.shape,n.strides,e);if(!t)throw new Error(`Cannot reshape tensor of shape [${n.shape}] to [${e}]`);if(!t.needsCopy)return Ys(n,t.sizes,t.strides,0);let r=n.isContiguous?n:OB(n);return Ys(r,t.sizes,t.strides,0)}function ez(n,e,t){let{sizes:r,strides:s}=_B(n.shape,n.strides,e,t);return Ys(n,r,s,0)}function tz(n,e){let{sizes:t,strides:r}=bB(n.shape,n.strides,e);return Ys(n,t,r,0)}function nz(n,e){let{sizes:t,strides:r}=wB(n.shape,n.strides,e);return Ys(n,t,r,0)}function rz(n,e,t){let r=n.shape,s=n.strides,o=[...e],i=new Array(e.length).fill(0);for(let a=0;a<t.length;a++){let u=t[a];i[u]=r[a]===1&&e[u]!==1?0:s[a]}return Ys(n,o,i,0)}function sz(n,e,t,r,s){let{sizes:o,strides:i,offsetDelta:a}=TB(n.shape,n.strides,e,t,r,s);return Ys(n,o,i,a)}function oz(n,e){let{sizes:t,strides:r}=xB(n.shape,n.strides,e);return Ys(n,t,r,0)}function iz(n,e){let{sizes:t,strides:r}=SB(n.shape,n.strides,e);return Ys(n,t,r,0)}function az(n,e,t,r){let{sizes:s,strides:o,offsetDelta:i}=vB(n.shape,n.strides,e,t,r);return Ys(n,s,o,i)}function uz(n,e,t){let{sizes:r,strides:s,offsetDelta:o}=kB(n.shape,n.strides,e,t);return Ys(n,r,s,o)}function AB(n){VT=n}function OB(n){let e=n._impl;return n.isContiguous&&e.storageOffset===0&&e.storage.rawData.length===n.numel?n:lz(n)}function lz(n){if(VT){let g=VT(n);if(g)return g}let e=n.shape,t=n.strides,r=n._impl.storage.data,s=n._impl.storageOffset,o=n.numel,i=En(e),a=n._impl.dtype,u=Wn.allocate(o*Ya(a),a,n._impl.device),p=u.data,c=e.length,f=new Int32Array(c),d=s;for(let g=0;g<o;g++){p[g]=r[d];for(let y=c-1;y>=0&&(f[y]++,d+=t[y],!(f[y]<e[y]));y--)d-=f[y]*t[y],f[y]=0}let h=new Pt(u,0,e,i,n._impl.dtype,n._impl.device),m=n._impl.autogradMeta;return m&&h.setAutogradMeta(m),new Ze(h)}var VT,BB,GT=M(()=>{"use strict";sr();di();Xa();rs();ss();NB();l(Ys,"makeView");l(XF,"reshapeKernel");l(ez,"transposeKernel");l(tz,"permuteKernel");l(nz,"expandKernel");l(rz,"broadcastInDimKernel");l(sz,"sliceKernel");l(oz,"unsqueezeKernel");l(iz,"squeezeKernel");l(az,"narrowKernel");l(uz,"selectKernel");VT=null;l(AB,"setGpuContiguousHook");l(OB,"contiguousKernel");BB=Object.freeze({reshape:l((n,e,t)=>XF(e,t),"reshape"),transpose:l((n,e,t,r)=>ez(e,t,r),"transpose"),permute:l((n,e,t)=>tz(e,t),"permute"),broadcast_in_dim:l((n,e,t,r)=>rz(e,t,r),"broadcast_in_dim"),expand:l((n,e,t)=>nz(e,t),"expand"),slice:l((n,e,t,r,s,o)=>sz(e,t,r,s,o),"slice"),unsqueeze:l((n,e,t)=>oz(e,t),"unsqueeze"),squeeze:l((n,e,t)=>iz(e,t),"squeeze"),narrow:l((n,e,t,r,s)=>az(e,t,r,s),"narrow"),select:l((n,e,t,r)=>uz(e,t,r),"select"),contiguous:l((n,e)=>OB(e),"contiguous")});l(lz,"copyContiguous")});var cz,zg,ve,Ao=M(()=>{"use strict";cz=0,zg=class{static{l(this,"Edge")}node;inputNr;constructor(e,t){this.node=e,this.inputNr=t}},ve=class{static{l(this,"AutogradNode")}_id;_numInputs;_nextEdges;_savedTensors;_inputMetadata;_opArgs;constructor(e){this._id=cz++,this._numInputs=e||0,this._nextEdges=[],this._savedTensors=[],this._inputMetadata=[],this._opArgs=null}setOpArgs(e){this._opArgs=e}opArgs(){return this._opArgs}get id(){return this._id}get numInputs(){return this._numInputs}get nextEdges(){return this._nextEdges}addNextEdge(e,t){this._nextEdges.push(new zg(e,t))}setNextEdge(e,t,r){for(;this._nextEdges.length<=e;)this._nextEdges.push(null);this._nextEdges[e]=new zg(t,r)}saveTensor(e){this._savedTensors.push(e)}savedTensors(){return this._savedTensors}saveInputMetadata(e,t,r){this._inputMetadata[e]={shape:t,dtype:r}}inputMetadata(e){return this._inputMetadata[e]||null}apply(e){throw new Error(`${this.name()}.apply() not implemented`)}name(){return this.constructor.name}releaseVariables(){this._savedTensors=[],this._inputMetadata=[]}}});var EB={};at(EB,{arange:()=>RB,empty:()=>Tn,eye:()=>jg,full:()=>ot,linspace:()=>IB,ones:()=>tn,randn:()=>Gg,randperm:()=>MB,zeros:()=>Ve});function Vg(n){return{dtype:n?.dtype??K.F32,device:n?.device??ea(),requiresGrad:n?.requiresGrad??!1}}function Jp(n,e,t,r){let s=En(n),i=bn(n)*Ya(e),a=Wn.allocate(i,e,t),u=new Pt(a,0,n,s,e,t),p=new Ze(u);return r&&p.requiresGrad_(!0),p}function Tn(n,e){let{dtype:t,device:r,requiresGrad:s}=Vg(e);return Jp(n,t,r,s)}function jT(n,e,t){let r=Tn(n,e),s=r.data;if(s)return s.fill(t),r;if(r.device&&r.device.type==="meta"){let o=Tn(n,{dtype:e?.dtype});return o.data.fill(t),o}return r}function Ve(n,e){return jT(n,e,0)}function tn(n,e){return jT(n,e,1)}function ot(n,e,t){return jT(n,t,e)}function Gg(n,e){let t=Tn(n,e),r=t.data;if(r){let s=r.length;for(let o=0;o<s;o+=2){let i=Math.random()||1e-10,a=Math.random(),u=Math.sqrt(-2*Math.log(i)),p=6.283185307179586*a;r[o]=u*Math.cos(p),o+1<s&&(r[o+1]=u*Math.sin(p))}}return t}function RB(n,e,t,r){let s,o,i;e===void 0&&t===void 0?(s=0,o=n,i=1):t===void 0?(s=n,o=e,i=1):(s=n,o=e,i=t);let a=Math.max(0,Math.ceil((o-s)/i)),{dtype:u,device:p,requiresGrad:c}=Vg(r),f=Jp([a],u,p,c),d=f.data;if(d)for(let h=0;h<a;h++)d[h]=s+h*i;return f}function jg(n,e,t){let r=e??n,{dtype:s,device:o,requiresGrad:i}=Vg(t),a=Jp([n,r],s,o,i),u=a.data;if(u){u.fill(0);let p=Math.min(n,r);for(let c=0;c<p;c++)u[c*r+c]=1}return a}function MB(n,e){let t=e?.dtype??K.I32,r=e?.device??ea(),s=Jp([n],t,r,e?.requiresGrad??!1),o=s.data;for(let i=0;i<n;i++)o[i]=i;for(let i=n-1;i>0;i--){let a=Math.floor(Math.random()*(i+1)),u=o[i];o[i]=o[a],o[a]=u}return s}function IB(n,e,t,r){let{dtype:s,device:o,requiresGrad:i}=Vg(r),a=Jp([t],s,o,i),u=a.data;if(u&&t>0)if(t===1)u[0]=n;else{let p=(e-n)/(t-1);for(let c=0;c<t;c++)u[c]=n+c*p}return a}var kt=M(()=>{"use strict";sr();di();Xa();rs();qn();ss();l(Vg,"_defaultOpts");l(Jp,"_makeTensor");l(Tn,"empty");l(jT,"_filled");l(Ve,"zeros");l(tn,"ones");l(ot,"full");l(Gg,"randn");l(RB,"arange");l(jg,"eye");l(MB,"randperm");l(IB,"linspace")});function pz(n,e,t){n[e]+=t}var Ug,qg,Wg,Kg,Pl,Hg,UT=M(()=>{"use strict";Ao();be();kt();be();ss();l(pz,"_addAt");Ug=class extends ve{static{l(this,"CatBackward")}constructor(){super(0)}apply(e){let t=e[0],r=this.opArgs(),s=t.shape.length,o=Ks(r&&r.length>1?r[1]??0:0,s),i=[],a=0,u=0;for(;this.inputMetadata(u);){let c=this.inputMetadata(u).shape[o];i.push(ls(ir(t,o,a,c))),a+=c,u++}return i}},qg=class extends ve{static{l(this,"StackBackward")}constructor(){super(0)}apply(e){let t=e[0],r=this.opArgs(),s=t.shape.length,o=Ks(r&&r.length>1?r[1]??0:0,s),i=[],a=0;for(;this.inputMetadata(a);)i.push(ls(en(t,o,a))),a++;return i}},Wg=class extends ve{static{l(this,"ClampBackward")}constructor(){super(3)}apply(e){let t=e[0],[r,s,o]=this.savedTensors(),i=Ve(t.shape,{dtype:t.dtype,device:t.device}),a=Rg(r.detach(),s.detach()),u=or(a,t,i),p=Bg(r.detach(),o.detach());return[or(p,u,i),null,null]}},Kg=class extends ve{static{l(this,"PadBackward")}constructor(){super(2)}apply(e){let t=e[0],s=this.opArgs()[2],i=this.inputMetadata(0).shape,a=t;for(let u=0;u<i.length;u++){let p=s[u]||0;a=ir(a,u,p,i[u])}return[ls(a),null]}},Pl=class n extends ve{static{l(this,"IndexSelectBackward")}static#e=null;static setGpuBackward(e){n.#e=e}constructor(){super(2)}apply(e){let t=e[0],[,r]=this.savedTensors(),o=this.inputMetadata(0).shape,i=o.length,a=this.opArgs(),u=Ks(a&&a.length>2?a[2]??0:0,i);if(n.#e){let v=n.#e(t,r,o,u);if(v)return[v]}let p=Ve(o,{dtype:t.dtype,device:t.device}),c=p._impl.storage.data,f=p.strides,d=ls(t),h=d._impl.storage.data,m=d._impl.storageOffset,g=d.shape,y=d.strides,_=ls(r),b=_._impl.storage.data,w=_._impl.storageOffset,x=g.length,S=new Int32Array(x),T=m;for(let v=0;v<d.numel;v++){let A=0;for(let N=0;N<x;N++){let k=N===u?Number(b[w+S[N]]):S[N];A+=k*f[N]}pz(c,A,h[T]);for(let N=x-1;N>=0;N--){if(S[N]++,S[N]<g[N]){T+=y[N];break}T-=(g[N]-1)*y[N],S[N]=0}}return[p,null]}},Hg=class extends ve{static{l(this,"WhereBackward")}constructor(){super(3)}apply(e){let t=e[0],[r]=this.savedTensors(),s=Ve(t.shape,{dtype:t.dtype,device:t.device}),o=or(r.detach(),t,s),i=or(r.detach(),s,t);return[null,o,i]}}});import fz from"koffi";function Yg(n,e,t){zn()?we.memcpyDtoDAsync(n,e,t,He().stream):we.memcpyDtoD(n,e,t)}function dz(n,e){zn()?we.memsetD8Async(n,0,e,He().stream):we.memsetD8(n,0,e)}function DB(){if(Zp!==null)return Zp;try{let n=fz.load(bs(Z1));pn={create:n.func("int cudnnCreate(_Out_ void **h)"),setStream:n.func("int cudnnSetStream(void *h, void *stream)"),getErrorString:n.func("str cudnnGetErrorString(int status)"),createRNNDesc:n.func("int cudnnCreateRNNDescriptor(_Out_ void **d)"),setRNNDesc:n.func("int cudnnSetRNNDescriptor_v8(void *d, int algo, int cellMode, int biasMode, int dirMode, int inputMode, int dataType, int mathPrec, int mathType, int inputSize, int hiddenSize, int projSize, int numLayers, void *dropoutDesc, uint auxFlags)"),createRNNDataDesc:n.func("int cudnnCreateRNNDataDescriptor(_Out_ void **d)"),setRNNDataDesc:n.func("int cudnnSetRNNDataDescriptor(void *d, int dataType, int layout, int maxSeqLength, int batchSize, int vectorSize, void *seqLengthArray, void *paddingFill)"),createTensorDesc:n.func("int cudnnCreateTensorDescriptor(_Out_ void **d)"),setTensorNdDesc:n.func("int cudnnSetTensorNdDescriptor(void *d, int dataType, int nbDims, void *dimA, void *strideA)"),createDropoutDesc:n.func("int cudnnCreateDropoutDescriptor(_Out_ void **d)"),setDropoutDesc:n.func("int cudnnSetDropoutDescriptor(void *d, void *h, float dropout, uint64 states, size_t stateSize, uint64 seed)"),dropoutGetStatesSize:n.func("int cudnnDropoutGetStatesSize(void *h, _Out_ size_t *sz)"),getRNNWeightSpaceSize:n.func("int cudnnGetRNNWeightSpaceSize(void *h, void *d, _Out_ size_t *sz)"),getRNNTempSpaceSizes:n.func("int cudnnGetRNNTempSpaceSizes(void *h, void *d, int fwdMode, void *xDesc, _Out_ size_t *work, _Out_ size_t *reserve)"),getRNNWeightParams:n.func("int cudnnGetRNNWeightParams(void *h, void *d, int pseudoLayer, size_t wss, uint64 weightSpace, int linLayerID, void *mDesc, _Out_ uint64 *mAddr, void *bDesc, _Out_ uint64 *bAddr)"),rnnForward:n.func("int cudnnRNNForward(void *h, void *d, int fwdMode, uint64 devSeqLengths, void *xDesc, uint64 x, void *yDesc, uint64 y, void *hDesc, uint64 hx, uint64 hy, void *cDesc, uint64 cx, uint64 cy, size_t wss, uint64 weightSpace, size_t workSize, uint64 workSpace, size_t reserveSize, uint64 reserveSpace)"),rnnBackwardData:n.func("int cudnnRNNBackwardData_v8(void *h, void *d, uint64 devSeqLengths, void *yDesc, uint64 y, uint64 dy, void *xDesc, uint64 dx, void *hDesc, uint64 hx, uint64 dhy, uint64 dhx, void *cDesc, uint64 cx, uint64 dcy, uint64 dcx, size_t wss, uint64 weightSpace, size_t workSize, uint64 workSpace, size_t reserveSize, uint64 reserveSpace)"),rnnBackwardWeights:n.func("int cudnnRNNBackwardWeights_v8(void *h, void *d, int addGrad, uint64 devSeqLengths, void *xDesc, uint64 x, void *hDesc, uint64 hx, void *yDesc, uint64 y, size_t wss, uint64 dweightSpace, size_t workSize, uint64 workSpace, size_t reserveSize, uint64 reserveSpace)")},Zp=!0}catch{Zp=!1}return Zp}function PB(){return DB()}function Ln(n,e){if(e!==0)throw new Error("cuDNN "+n+" failed: "+e+" ("+pn.getErrorString(e)+")")}function Qp(){if(!DB())throw new Error("cuDNN not available");let n=He();if(Qi(),!Jg){let e=su();Ln("create",pn.create(e)),Jg=e[0],Ln("setStream",pn.setStream(Jg,n.stream))}return Jg}function Sz(){if(!qT){let n=Qp(),e=su();Ln("createDropout",pn.createDropoutDesc(e));let t=[0n];Ln("dropStatesSize",pn.dropoutGetStatesSize(n,t));let r=ht(Math.max(Number(t[0]),1));Ln("setDropout",pn.setDropoutDesc(e[0],n,0,r,t[0],0n)),qT=e[0]}return qT}function vz(n,e){let t=su();return Ln("createTensor",pn.createTensorDesc(t)),Ln("setTensorNd",pn.setTensorNdDesc(t[0],Zg,n.length,new Int32Array(n),new Int32Array(e))),t[0]}function LB(n,e,t,r){let s=su();return Ln("createRNNData",pn.createRNNDataDesc(s)),Ln("setRNNData",pn.setRNNDataDesc(s[0],Zg,wz,n,e,t,r,null)),s[0]}function kz({inputSize:n,hiddenSize:e,seqLen:t,batch:r,numLayers:s,cellMode:o=$B,gates:i=4}){let a=`${o}_${n}_${e}_${t}_${r}_${s}`,u=CB.get(a);if(u)return u;let p=Qp(),c=su();Ln("createRNN",pn.createRNNDesc(c)),Ln("setRNN",pn.setRNNDesc(c[0],hz,o,gz,yz,_z,Zg,Zg,bz,n,e,e,s,Sz(),0));let f=c[0],d=new Int32Array(r).fill(t),h=LB(t,r,n,d),m=LB(t,r,e,d),g=vz([s,r,e],[r*e,e,1]),y=[0n];Ln("wss",pn.getRNNWeightSpaceSize(p,f,y));let _=[0n],b=[0n];Ln("temp",pn.getRNNTempSpaceSizes(p,f,VB,h,_,b));let w=ht(r*4);return we.memcpyHtoD(w,d,r*4),u={rd:f,xDesc:h,yDesc:m,hDesc:g,devSeq:w,wss:y[0],wssN:Number(y[0]),workSize:_[0],workN:Math.max(Number(_[0]),1),reserveSize:b[0],reserveN:Math.max(Number(b[0]),1),inputSize:n,hiddenSize:e,seqLen:t,batch:r,numLayers:s,gates:i},CB.set(a,u),u}function GB(n,e,t,r){let s=Qp(),o=su(),i=su();Ln("cwm",pn.createTensorDesc(o)),Ln("cwb",pn.createTensorDesc(i));let{numLayers:a,hiddenSize:u,inputSize:p,wss:c,gates:f}=n;for(let d=0;d<a;d++){let h=d===0?p:u,m=t[d];for(let g=0;g<2*f;g++){let y=g<f,_=g%f,b=[0n],w=[0n];Ln("wparam",pn.getRNNWeightParams(s,n.rd,d,c,e,g,o[0],b,i[0],w));let S=(y?u*h:u*u)*4,T=u*4,v=BigInt(y?m.x2hW:m.h2hW)+BigInt(_*S),A=BigInt(y?m.x2hB:m.h2hB)+BigInt(_*T);r?(Yg(v,b[0],S),Yg(A,w[0],T)):(Yg(b[0],v,S),Yg(w[0],A,T))}}}function jB(n,e,t,r,s,o,i,a,u=!0){let p=Qp(),c=kz({...t,numLayers:e.length}),f=ht(c.wssN);GB(c,f,e,!1);let d=ht(c.workN),h=u?ht(c.reserveN):0n;return Ln("forward",pn.rnnForward(p,c.rd,u?VB:Tz,c.devSeq,c.xDesc,n,c.yDesc,o,c.hDesc,r||0n,i,c.hDesc,s||0n,a,c.wss,f,c.workSize,d,u?c.reserveSize:0n,h)),{p:c,weightSpace:f,workSpace:d,reserveSpace:h,training:u}}function UB(n){n&&(_t(n.weightSpace,n.p.wssN),_t(n.workSpace,n.p.workN),n.reserveSpace&&n.reserveSpace!==0n&&_t(n.reserveSpace,n.p.reserveN))}function qB(n,e,t,r,s,o,i,a,u,p,c,f){let{p:d,weightSpace:h,workSpace:m,reserveSpace:g}=n,y=Qp(),_=ht(d.wssN);dz(_,d.wssN),Ln("backwardData",pn.rnnBackwardData(y,d.rd,d.devSeq,d.yDesc,t,o,d.xDesc,u,d.hDesc,r||0n,i||0n,p||0n,d.hDesc,s||0n,a||0n,c||0n,d.wss,h,d.workSize,m,d.reserveSize,g)),Ln("backwardWeights",pn.rnnBackwardWeights(y,d.rd,xz,d.devSeq,d.xDesc,e,d.hDesc,r||0n,d.yDesc,t,d.wss,_,d.workSize,m,d.reserveSize,g)),GB(d,_,f,!0),_t(h,d.wssN),_t(m,d.workN),_t(g,d.reserveN),_t(_,d.wssN)}var pn,Zp,Zg,hz,$B,mz,gz,yz,_z,bz,FB,zB,wz,VB,Tz,xz,su,Jg,qT,CB,WT=M(()=>{"use strict";ws();Ts();ig();pi();vl();fi();l(Yg,"d2d");l(dz,"memZero");Zp=null;l(DB,"ensure");l(PB,"cudnnAvailable");Zg=0,hz=0,$B=2,mz=3,gz=2,yz=0,_z=0,bz=2,FB=$B,zB=mz,wz=0,VB=1,Tz=0,xz=0;l(Ln,"ck");su=l(()=>[null],"o2"),Jg=null,qT=null;l(Qp,"handle");l(Sz,"dropoutDesc");l(vz,"tensorDesc");l(LB,"dataDesc");CB=new Map;l(kz,"plan");l(GB,"packWeights");l(jB,"cudnnRNNForward");l(UB,"releaseRNNForward");l(qB,"cudnnRNNBackward")});var HB={};at(HB,{GradMode:()=>Oo,enableGrad:()=>KB,noGrad:()=>WB});function WB(n){let e=sa;sa=!1;try{return n()}finally{sa=e}}function KB(n){let e=sa;sa=!0;try{return n()}finally{sa=e}}var sa,Oo,ou=M(()=>{"use strict";sa=!0,Oo={isEnabled(){return sa},setEnabled(n){sa=n}};l(WB,"noGrad");l(KB,"enableGrad")});function Qg(n,e){for(let t=0;t<e.length;t++){n.saveInputMetadata(t,[...e[t].shape],e[t].dtype);let r=e[t]._impl.autogradMeta;if(r&&r.requiresGrad)if(r.gradFn)n.setNextEdge(t,r.gradFn,r.outputNr||0);else{let s=r.getGradAccumulator();s||(s=new Xp(e[t]),r.setGradAccumulator(s)),n.setNextEdge(t,s,0)}else n.setNextEdge(t,null,0)}}var Xp,Xg=M(()=>{"use strict";Ao();be();Xp=class extends ve{static{l(this,"GradAccumulator")}_variable;constructor(e){super(0),this._variable=typeof WeakRef<"u"?new WeakRef(e):{deref:l(()=>e,"deref")}}apply(e){let t=e[0],r=this._variable.deref();if(!r)return[];let s=r._impl.autogradMeta;return s?(s.grad===null?s.grad=t:s.grad=ae(s.grad,t),[]):[]}name(){return"GradAccumulator"}};l(Qg,"wireInputEdges")});function Cr(n){return In()?To(n):Ha(n)}function Ro(n,e){if(In())return Xi(n);let t=ht(n.byteLength);return e.push([n,t]),t}function YB(n){for(let[e,t]of n)qt(e,t),_t(t,e.byteLength)}function KT(n,e,t){let r=new So;r.setGradFn(n,t),r.requiresGrad=!0,e._impl.setAutogradMeta(r),e._impl._updateKeySet()}function JB(n,e,t,r,s=null,o=null){let i=Nz[n],{inputSize:a,hiddenSize:u,seqLen:p,batch:c}=r,f=t.length,d=[],h={...r,numLayers:f,cellMode:i.cellMode,gates:i.gates},m=Bo(e),g=Cr(m),y=t.map(D=>({x2hW:Cr(Bo(D.x2h.weight)),x2hB:Cr(Bo(D.x2h.bias)),h2hW:Cr(Bo(D.h2h.weight)),h2hB:Cr(Bo(D.h2h.bias))})),_=s?Bo(s):null,b=i.hasCell&&o?Bo(o):null,w=s?Cr(_):0n,x=b?Cr(b):0n,S=f*c*u,T=[f,c,u],v=new Float32Array(p*c*u),A=new Float32Array(S),N=i.hasCell?new Float32Array(S):null,k=Ro(v,d),O=Ro(A,d),R=N?Ro(N,d):0n,B=[e];for(let D of t)B.push(D.x2h.weight,D.x2h.bias,D.h2h.weight,D.h2h.bias);s&&(B.push(s),i.hasCell&&B.push(o));let E=!1;for(let D of B)if(D._impl.autogradMeta&&D.requiresGrad){E=!0;break}let F=Oo.isEnabled()&&E,C=jB(g,y,h,w,x,k,O,R,F);YB(d);let H=Ee(v,[p,c,u],e.dtype,e.device),G=Ee(A,T,e.dtype,e.device),U=N?Ee(N,T,e.dtype,e.device):null;if(F){let D={xArr:m,yArr:v,hxArr:_,cxArr:b,opts:h,numLayers:f,hasInit:!!s,hasCell:i.hasCell,dtype:e.dtype,device:e.device,inputShape:[...e.shape],stateShape:T,weightShapes:t.map(ue=>({x2hW:[...ue.x2h.weight.shape],x2hB:[...ue.x2h.bias.shape],h2hW:[...ue.h2h.weight.shape],h2hB:[...ue.h2h.bias.shape]}))},V=new HT(C,D,B.length);Qg(V,B),KT(V,H,0),KT(V,G,1),U&&KT(V,U,2)}else UB(C);return i.hasCell?[H,G,U]:[H,G]}function ZB(n,e,t,r=null,s=null){return JB("lstm",n,e,t,r,s)}function QB(n,e,t,r=null){return JB("gru",n,e,t,r,null)}var Bo,ey,Nz,HT,XB=M(()=>{"use strict";Ao();Wp();ou();Xg();wn();qp();pi();WT();Bo=l(n=>YT(n),"carr"),ey=l(n=>n.reduce((e,t)=>e*t,1),"prod");l(Cr,"devIn");l(Ro,"devOut");l(YB,"flushOut");Nz={lstm:{cellMode:FB,gates:4,hasCell:!0},gru:{cellMode:zB,gates:3,hasCell:!1}},HT=class extends ve{static{l(this,"CudnnRNNBackward")}constructor(e,t,r){super(r),this.fwd=e,this.info=t}apply(e){let[t,r,s]=e,o=this.info,{seqLen:i,batch:a,inputSize:u,hiddenSize:p}=o.opts,c=o.numLayers,f=c*a*p,d=[],h=Cr(o.xArr),m=Cr(o.yArr),g=o.hasInit?Cr(o.hxArr):0n,y=o.hasCell&&o.hasInit?Cr(o.cxArr):0n,_=Cr(Bo(t)),b=r?Cr(Bo(r)):0n,w=o.hasCell&&s?Cr(Bo(s)):0n,x=new Float32Array(i*a*u),S=Ro(x,d),T=null,v=null,A=0n,N=0n;o.hasInit&&(T=new Float32Array(f),A=Ro(T,d),o.hasCell&&(v=new Float32Array(f),N=Ro(v,d)));let k=o.weightShapes.map(E=>({x2hW:new Float32Array(ey(E.x2hW)),x2hB:new Float32Array(ey(E.x2hB)),h2hW:new Float32Array(ey(E.h2hW)),h2hB:new Float32Array(ey(E.h2hB))})),O=k.map(E=>({x2hW:Ro(E.x2hW,d),x2hB:Ro(E.x2hB,d),h2hW:Ro(E.h2hW,d),h2hB:Ro(E.h2hB,d)}));qB(this.fwd,h,m,g,y,_,b,w,S,A,N,O),YB(d);let R=l((E,F)=>Ee(E,F,o.dtype,o.device),"w"),B=[R(x,o.inputShape)];for(let E=0;E<k.length;E++){let F=k[E],C=o.weightShapes[E];B.push(R(F.x2hW,C.x2hW),R(F.x2hB,C.x2hB),R(F.h2hW,C.h2hW),R(F.h2hB,C.h2hB))}return o.hasInit&&(B.push(R(T,o.stateShape)),o.hasCell&&B.push(R(v,o.stateShape))),B}};l(KT,"attach");l(JB,"cudnnRNNOp");l(ZB,"cudnnLSTMOp");l(QB,"cudnnGRUOp")});function eR(n){return In()?To(n):Ha(n)}function Az(n,e){if(In())return Xi(n);let t=ht(n.byteLength);return e.push([n,t]),t}function Oz(n){for(let[e,t]of n)qt(e,t),_t(t,e.byteLength)}function tR(n){let e=nR(n);if(e)return{data:n._impl.storage.rawData,lay:e};let t=ZT();if(In()&&t){let s=t(n._impl.storage.rawData,n.shape,n.strides,n._impl.storageOffset,n.dtype),o=n.shape,i=o.length,a=o[i-2],u=o[i-1],p=1;for(let c=0;c<i-2;c++)p*=o[c];return{data:s,lay:{trans:!1,batchStride:a*u,batch:p,rows:a,cols:u}}}let r=ls(n);return{data:r._impl.storage.rawData,lay:nR(r)}}function nR(n){if(n.storageOffset!==0)return null;let e=n.shape,t=n.strides,r=e.length;if(r<2)return null;let s=e[r-2],o=e[r-1],i=t[r-2],a=t[r-1],u;if(i===o&&a===1)u=!1;else if(i===1&&a===s)u=!0;else return null;let p=s*o;for(let f=r-3;f>=0;f--){if(t[f]!==p)return null;p*=e[f]}let c=1;for(let f=0;f<r-2;f++)c*=e[f];return{trans:u,batchStride:s*o,batch:c,rows:s,cols:o}}function Bz(n,e){let t=new So;t.setGradFn(n,0),t.requiresGrad=!0,e._impl.setAutogradMeta(t),e._impl._updateKeySet()}function rR(n,e){if(!n.device||n.device.type!=="gpu"||e.device.type!=="gpu"||n.dtype!=="f32"||e.dtype!=="f32")return null;let t=n.shape.length,r=e.shape.length;if(t<2||r<2||t!==r)return null;let s=n.shape[t-2],o=n.shape[t-1],i=e.shape[r-1];if(e.shape[r-2]!==o)return null;let a=tR(n),u=tR(e),p=a.lay,c=u.lay;if(!p||!c||p.batch!==c.batch)return null;let f=p.batch,d=[],h=eR(a.data),m=eR(u.data),g=new Float32Array(f*s*i),y=Az(g,d);f>1?TT(f,s,i,o,h,p.batchStride,p.trans,m,c.batchStride,c.trans,y,s*i):zp(s,i,o,h,p.trans,m,c.trans,y),Oz(d);let _=[...n.shape.slice(0,t-2),s,i],b=Ee(g,_,n.dtype,n.device);if(Oo.isEnabled()&&(n._impl.autogradMeta&&n.requiresGrad||e._impl.autogradMeta&&e.requiresGrad)){let w=new JT(n,e);Qg(w,[n,e]),Bz(w,b)}return b}var JT,sR=M(()=>{"use strict";Ao();Wp();ou();Xg();wn();be();qn();qp();pi();Vp();be();l(eR,"devIn");l(Az,"devOut");l(Oz,"flushOut");l(tR,"operandData");l(nR,"operandLayout");JT=class extends ve{static{l(this,"MatmulBackward")}constructor(e,t){super(2),this.A=e,this.B=t}apply(e){let t=e[0],r=this.A,s=this.B,o=r.shape.length,i=s.shape.length,a=tt(t,s.transpose(i-2,i-1)),u=tt(r.transpose(o-2,o-1),t);return[a,u]}};l(Bz,"attach");l(rR,"gpuMatmul")});function $l(){if(oR)return;oR=!0;let n=new Vt("mlc","DEF");for(let e of Rz)n.def(e)}var Rz,oR,ty=M(()=>{"use strict";ys();Rz=["svd(Tensor input) -> (Tensor, Tensor, Tensor)","eigh(Tensor input) -> (Tensor, Tensor)","cholesky(Tensor input) -> Tensor","inv(Tensor input) -> Tensor","pinv(Tensor input) -> Tensor","det(Tensor input) -> Tensor","cov(Tensor input) -> Tensor","solve(Tensor a, Tensor b) -> Tensor","lstsq(Tensor a, Tensor b) -> Tensor"],oR=!1;l($l,"ensureLinalgSchemas")});var ef=M(()=>{"use strict"});import Mz from"koffi";function Iz(){if(tf!==null)return tf;try{let n=Mz.load(bs(J1));Dr={create:n.func("int cusolverDnCreate(_Out_ void **h)"),destroy:n.func("int cusolverDnDestroy(void *h)"),setStream:n.func("int cusolverDnSetStream(void *h, void *stream)"),gesvdBufferSize:n.func("int cusolverDnSgesvd_bufferSize(void *h, int m, int n, _Out_ int *lwork)"),gesvd:n.func("int cusolverDnSgesvd(void *h, int8 jobu, int8 jobvt, int m, int n, uint64 A, int lda, uint64 S, uint64 U, int ldu, uint64 VT, int ldvt, uint64 work, int lwork, uint64 rwork, uint64 info)"),syevdBufferSize:n.func("int cusolverDnSsyevd_bufferSize(void *h, int jobz, int uplo, int n, uint64 A, int lda, uint64 W, _Out_ int *lwork)"),syevd:n.func("int cusolverDnSsyevd(void *h, int jobz, int uplo, int n, uint64 A, int lda, uint64 W, uint64 work, int lwork, uint64 info)"),potrfBufferSize:n.func("int cusolverDnSpotrf_bufferSize(void *h, int uplo, int n, uint64 A, int lda, _Out_ int *lwork)"),potrf:n.func("int cusolverDnSpotrf(void *h, int uplo, int n, uint64 A, int lda, uint64 work, int lwork, uint64 info)"),getrfBufferSize:n.func("int cusolverDnSgetrf_bufferSize(void *h, int m, int n, uint64 A, int lda, _Out_ int *lwork)"),getrf:n.func("int cusolverDnSgetrf(void *h, int m, int n, uint64 A, int lda, uint64 work, uint64 ipiv, uint64 info)"),getrs:n.func("int cusolverDnSgetrs(void *h, int trans, int n, int nrhs, uint64 A, int lda, uint64 ipiv, uint64 B, int ldb, uint64 info)")},tf=!0}catch{tf=!1}return tf}function rf(){if(!Iz())throw new Error("cuSOLVER library not found; install the CUDA Toolkit (cusolver) to use GPU linalg");let n=He();if(Qi(),!ny){let e=[null];cs("create",Dr.create(e)),ny=e[0],cs("setStream",Dr.setStream(ny,n.stream))}return ny}function cs(n,e){if(e!==0)throw new Error("cuSOLVER "+n+" failed: "+e)}function Fl(){$e("cuStreamSynchronize",we.streamSynchronize(He().stream))}function sf(){let n=[];return{up(e){let t=ht(e.byteLength);return xs(t,e),n.push([t,e.byteLength]),t},out(e){let t=Math.max(e,1),r=ht(t);return n.push([r,t]),r},free(){for(let[e,t]of n)_t(e,t)}}}function nf(n){let e=new Int32Array(1);return qt(e,n),e[0]}function QT(n,e,t){let r=rf(),s=sf();try{let o=s.up(n),i=[0];cs("gesvd_bufferSize",Dr.gesvdBufferSize(r,e,t,i));let a=s.out(t*4),u=s.out(e*t*4),p=s.out(t*t*4),c=s.out(i[0]*4),f=s.out(4);cs("gesvd",Dr.gesvd(r,iR,iR,e,t,o,e,a,u,e,p,t,c,i[0],0n,f)),Fl();let d=nf(f);if(d<0)throw new Error("cuSOLVER gesvd: invalid argument "+-d);if(d>0)throw new Error("cuSOLVER gesvd: did not converge");let h=new Float32Array(t),m=new Float32Array(e*t),g=new Float32Array(t*t);return qt(h,a),qt(m,u),qt(g,p),{S:h,U:m,VT:g}}finally{s.free()}}function uR(n,e){let t=rf(),r=sf();try{let s=r.up(n),o=r.out(e*4),i=[0];cs("syevd_bufferSize",Dr.syevdBufferSize(t,aR,ry,e,s,e,o,i));let a=r.out(i[0]*4),u=r.out(4);cs("syevd",Dr.syevd(t,aR,ry,e,s,e,o,a,i[0],u)),Fl();let p=nf(u);if(p!==0)throw new Error("cuSOLVER syevd: failed to converge (info "+p+")");let c=new Float32Array(e),f=new Float32Array(e*e);return qt(c,o),qt(f,s),{W:c,V:f}}finally{r.free()}}function lR(n,e){let t=rf(),r=sf();try{let s=r.up(n),o=[0];cs("potrf_bufferSize",Dr.potrfBufferSize(t,ry,e,s,e,o));let i=r.out(o[0]*4),a=r.out(4);cs("potrf",Dr.potrf(t,ry,e,s,e,i,o[0],a)),Fl();let u=nf(a);if(u>0)throw new Error("linalg.cholesky: matrix is not positive definite");if(u<0)throw new Error("cuSOLVER potrf: invalid argument "+-u);let p=new Float32Array(e*e);return qt(p,s),p}finally{r.free()}}function cR(n,e){let t=rf(),r=sf();try{let s=r.up(n),o=[0];cs("getrf_bufferSize",Dr.getrfBufferSize(t,e,e,s,e,o));let i=r.out(o[0]*4),a=r.out(e*4),u=r.out(4);cs("getrf",Dr.getrf(t,e,e,s,e,i,a,u)),Fl();let p=new Float32Array(e*e),c=new Int32Array(e);return qt(p,s),qt(c,a),{LU:p,ipiv:c}}finally{r.free()}}function XT(n,e,t,r,s){let o=rf(),i=sf();try{let a=i.up(n),u=[0];cs("getrf_bufferSize",Dr.getrfBufferSize(o,e,e,a,e,u));let p=i.out(u[0]*4),c=i.out(e*4),f=i.out(4);cs("getrf",Dr.getrf(o,e,e,a,e,p,c,f)),Fl();let d=nf(f);if(d>0)throw new Error("linalg: matrix is singular");let h=i.up(t);if(cs("getrs",Dr.getrs(o,s,e,r,a,e,c,h,e,f)),Fl(),d=nf(f),d<0)throw new Error("cuSOLVER getrs: invalid argument "+-d);let m=new Float32Array(e*r);return qt(m,h),m}finally{i.free()}}var Dr,tf,iR,aR,ry,ny,pR=M(()=>{"use strict";Ts();ig();ws();pi();vl();tf=null;l(Iz,"ensure");iR=83,aR=1,ry=0,ny=null;l(rf,"handle");l(cs,"ck");l(Fl,"sync");l(sf,"scope");l(nf,"readInt");l(QT,"csGesvd");l(uR,"csSyevd");l(lR,"csPotrf");l(cR,"csGetrf");l(XT,"csSolveLU")});function Js(n){return Float32Array.from(Qe(n))}function gi(n,e){if(n.ndim!==2)throw new Error(`linalg.${e}: expected a 2-D matrix, got ${n.ndim}-D`)}function Zs(n,e){if(n.dtype!=="f32")throw new Error(`linalg.${e}: GPU backend supports f32 tensors only, got '${n.dtype}'`)}function af(n,e){if(n.shape[0]!==n.shape[1])throw new Error(`linalg.${e}: matrix must be square`)}function Ez(){$e("cuStreamSynchronize",we.streamSynchronize(He().stream))}function ex(n,e,t){let r=Math.min(e,t),s=new Float32Array(e*r),o=new Float32Array(r),i=new Float32Array(t*r);if(e>=t){let a=new Float32Array(e*t);for(let p=0;p<e;p++)for(let c=0;c<t;c++)a[p+c*e]=n[p*t+c];let u=QT(a,e,t);for(let p=0;p<r;p++){o[p]=u.S[p];for(let c=0;c<e;c++)s[c*r+p]=u.U[c+p*e];for(let c=0;c<t;c++)i[c*r+p]=u.VT[p+c*t]}}else{let a=QT(n,t,e);for(let u=0;u<r;u++){o[u]=a.S[u];for(let p=0;p<e;p++)s[p*r+u]=a.VT[u+p*e];for(let p=0;p<t;p++)i[p*r+u]=a.U[p+u*t]}}return{U:s,S:o,V:i,k:r}}function dR(n,e){gi(e,"svd"),Zs(e,"svd");let[t,r]=e.shape,{U:s,S:o,V:i,k:a}=ex(Js(e),t,r);return[Ee(s,[t,a],e.dtype,e.device),Ee(o,[a],e.dtype,e.device),Ee(i,[r,a],e.dtype,e.device)]}function hR(n,e){gi(e,"eigh"),Zs(e,"eigh"),af(e,"eigh");let t=e.shape[0],{W:r,V:s}=uR(Js(e),t),o=new Float32Array(t*t);for(let i=0;i<t;i++)for(let a=0;a<t;a++)o[i*t+a]=s[i+a*t];return[Ee(r,[t],e.dtype,e.device),Ee(o,[t,t],e.dtype,e.device)]}function mR(n,e){gi(e,"cholesky"),Zs(e,"cholesky"),af(e,"cholesky");let t=e.shape[0],r=lR(Js(e),t),s=new Float32Array(t*t);for(let o=0;o<t;o++)for(let i=0;i<=o;i++)s[o*t+i]=r[o+i*t];return Ee(s,[t,t],e.dtype,e.device)}function gR(n,e,t){gi(e,"solve"),Zs(e,"solve"),af(e,"solve"),Zs(t,"solve");let r=e.shape[0],s=t.ndim===1,o=s?1:t.shape[1];if(t.shape[0]!==r)throw new Error("linalg.solve: right-hand side rows must match matrix");let i=Js(t),a=new Float32Array(r*o);for(let c=0;c<r;c++)for(let f=0;f<o;f++)a[c+f*r]=i[c*o+f];let u=XT(Js(e),r,a,o,fR),p=new Float32Array(r*o);for(let c=0;c<r;c++)for(let f=0;f<o;f++)p[c*o+f]=u[c+f*r];return Ee(p,s?[r]:[r,o],e.dtype,e.device)}function yR(n,e){gi(e,"inv"),Zs(e,"inv"),af(e,"inv");let t=e.shape[0],r=new Float32Array(t*t);for(let i=0;i<t;i++)r[i+i*t]=1;let s=XT(Js(e),t,r,t,fR),o=new Float32Array(t*t);for(let i=0;i<t;i++)for(let a=0;a<t;a++)o[i*t+a]=s[i+a*t];return Ee(o,[t,t],e.dtype,e.device)}function _R(n,e){gi(e,"det"),Zs(e,"det"),af(e,"det");let t=e.shape[0],{LU:r,ipiv:s}=cR(Js(e),t),o=1,i=1;for(let a=0;a<t;a++)i*=r[a+a*t],s[a]!==a+1&&(o=-o);return Ee(new Float32Array([o*i]),[],e.dtype,e.device)}function bR(n,e){gi(e,"pinv"),Zs(e,"pinv");let[t,r]=e.shape,{U:s,S:o,V:i,k:a}=ex(Js(e),t,r),u=1e-12*(a?o[0]:0),p=new Float32Array(r*t);for(let c=0;c<r;c++)for(let f=0;f<t;f++){let d=0;for(let h=0;h<a;h++){let m=o[h];m>u&&(d+=i[c*a+h]*(1/m)*s[f*a+h])}p[c*t+f]=d}return Ee(p,[r,t],e.dtype,e.device)}function wR(n,e,t){gi(e,"lstsq"),Zs(e,"lstsq"),Zs(t,"lstsq");let[r,s]=e.shape,o=t.ndim===1,i=o?1:t.shape[1];if(t.shape[0]!==r)throw new Error("linalg.lstsq: right-hand side rows must match matrix");let a=Js(t),{U:u,S:p,V:c,k:f}=ex(Js(e),r,s),d=1e-12*(f?p[0]:0),h=new Float32Array(f*i);for(let g=0;g<f;g++){let y=p[g];for(let _=0;_<i;_++){let b=0;for(let w=0;w<r;w++)b+=u[w*f+g]*a[w*i+_];h[g*i+_]=y>d?b/y:0}}let m=new Float32Array(s*i);for(let g=0;g<s;g++)for(let y=0;y<i;y++){let _=0;for(let b=0;b<f;b++)_+=c[g*f+b]*h[b*i+y];m[g*i+y]=_}return Ee(m,o?[s]:[s,i],e.dtype,e.device)}function TR(n,e){gi(e,"cov"),Zs(e,"cov");let[t,r]=e.shape,s=Js(e),o=new Float32Array(r);for(let f=0;f<r;f++){let d=0;for(let h=0;h<t;h++)d+=s[h*r+f];o[f]=d/t}let i=new Float32Array(t*r);for(let f=0;f<t;f++)for(let d=0;d<r;d++)i[f*r+d]=s[f*r+d]-o[d];let a=t>1?t-1:1;He();let u=ht(i.byteLength),p=new Float32Array(r*r),c=ht(p.byteLength);try{xs(u,i),zp(r,r,t,u,!0,u,!1,c),Ez(),qt(p,c)}finally{_t(u,i.byteLength),_t(c,p.byteLength)}for(let f=0;f<r*r;f++)p[f]/=a;return Ee(p,[r,r],e.dtype,e.device)}var fR,xR=M(()=>{"use strict";wn();ef();pR();Vp();pi();ws();Ts();fR=1;l(Js,"hostData");l(gi,"require2D");l(Zs,"requireF32");l(af,"requireSquare");l(Ez,"syncStream");l(ex,"_svdRowMajor");l(dR,"gpuSvd");l(hR,"gpuEigh");l(mR,"gpuCholesky");l(gR,"gpuSolve");l(yR,"gpuInv");l(_R,"gpuDet");l(bR,"gpuPinv");l(wR,"gpuLstsq");l(TR,"gpuCov")});function vR(){if(SR)return;SR=!0,$l();let n=new Vt("mlc","IMPL");for(let[e,t]of Object.entries(Lz))n.impl(e,1,t)}var Lz,SR,kR=M(()=>{"use strict";ys();An();ty();xR();Lz={svd:dR,eigh:hR,cholesky:mR,inv:yR,pinv:bR,det:_R,cov:TR,solve:gR,lstsq:wR},SR=!1;l(vR,"registerCudaLinalg")});var oy={};at(oy,{adamGraphKernelSource:()=>CR,deviceAdam:()=>DR,deviceClipGradNorm:()=>qz,deviceConcat:()=>MR,deviceContiguous:()=>rx,freeOptimizerDeviceState:()=>IR,isCudaGraphEnabled:()=>kO,measureCudaKernel:()=>NR,preloadCublas:()=>AR,releaseCudaMemory:()=>Wz,runCudaKernel:()=>Cz,runCudaKernelResident:()=>nx,runCudaKernelSync:()=>OR,runCudaPlan:()=>RO,setCudaGraphEnabled:()=>vO,stepIncKernelSource:()=>LR,teardownAfterFit:()=>Kz});function NR(n,e,t=[],r={}){He();let{func:s}=wo(n.source,n.name),o=n.metadata,i=r.warmup??5,a=r.repeat??30,u=r.minWarmupMs??25,p=r.maxWarmup??1e5,c=o.gridDim,f=o.blockDim,d=0,h=e.map(g=>Math.max(g,1)),m=h.map(g=>ht(g));try{let g=performance.now(),y=0;for(;(y<i||performance.now()-g<u)&&(ns(s,c,f,d,m,t),!(++y>=p)););let _=[];for(let b=0;b<a;b++){let w=performance.now();ns(s,c,f,d,m,t),_.push(performance.now()-w)}return _}finally{for(let g=0;g<m.length;g++)_t(m[g],h[g])}}async function AR(){return uf||(uf=await Promise.resolve().then(()=>(Vp(),wO))),uf}function OR(n,e,t){let r=n.metadata;if(r.cublas){if(!uf)throw new Error("cuBLAS module not preloaded; call preloadCublas() before sync execution");let{M:f,N:d,K:h,aIdx:m,bIdx:g,cIdx:y,transB:_}=r.cublas;uf.cublasMatmul(f,d,h,e[m],e[g],e[y],_);return}He();let{func:s}=tT(n),o=[],i=[];for(let f of e)ArrayBuffer.isView(f)?o.push(f):i.push(f);if(t)for(let f of t)i.push(f);let a=r.outputIndices||o.map((f,d)=>d),u=[];for(let f of o){let d=ht(f.byteLength);xs(d,f),u.push(d)}let p=BR(r.scratch);ns(s,r.gridDim,r.blockDim,0,[...u,...p.ptrs],i);let c=new Set(a);for(let f=0;f<o.length;f++)c.has(f)&&qt(o[f],u[f]);for(let f=0;f<u.length;f++)_t(u[f],o[f].byteLength);tx(p)}function nx(n,e,t){let r=n.metadata;if(r.cublas)throw new Error("cuBLAS kernels are not supported on the eager device-resident path");He();let{func:s}=tT(n),o=[],i=[];for(let c of e)ArrayBuffer.isView(c)?o.push(c):i.push(c);if(t)for(let c of t)i.push(c);let a=r._outputSet||(r._outputSet=new Set(r.outputIndices||o.map((c,f)=>f))),u=new Array(o.length),p=BR(r.scratch);if(In()){for(let c=0;c<o.length;c++)u[c]=a.has(c)?Xi(o[c]):To(o[c]);ns(s,r.gridDim,r.blockDim,0,[...u,...p.ptrs],i,!1),tx(p);return}for(let c=0;c<o.length;c++)u[c]=Ha(o[c]);ns(s,r.gridDim,r.blockDim,0,[...u,...p.ptrs],i,!1);for(let c=0;c<o.length;c++)a.has(c)&&vT(o[c],u[c]);tx(p)}function BR(n){if(!n||n.length===0)return{ptrs:[],bufs:[]};let e=[],t=[];for(let r of n){let s=r.size*St(r.dtype).BYTES_PER_ELEMENT,o=ht(s);e.push(o),t.push([o,s])}return{ptrs:e,bufs:t}}function tx(n){for(let[e,t]of n.bufs)_t(e,t)}async function Cz(n,e,t){n.metadata.cublas&&await AR(),OR(n,e,t)}function sy(n){return Dz[n]||"float"}function RR(n,e){return{gridDim:[Math.max(Math.ceil(n/256),1),1,1],blockDim:[256,1,1],sharedMemBytes:0,outputIndices:e}}function Pz(n,e){return`extern "C" __global__ void ${e}(const ${n}* in, ${n}* out, int n, int rank,
+  int s0,int s1,int s2,int s3,int s4,int s5,int s6,int s7,
+  int t0,int t1,int t2,int t3,int t4,int t5,int t6,int t7, int off) {
+  int i = blockIdx.x*blockDim.x + threadIdx.x; if (i >= n) return;
+  int shp[8] = {s0,s1,s2,s3,s4,s5,s6,s7};
+  int strd[8] = {t0,t1,t2,t3,t4,t5,t6,t7};
+  long long src = off; int rem = i;
+  for (int d = rank-1; d >= 0; d--) { int x = rem % shp[d]; rem /= shp[d]; src += (long long)x*strd[d]; }
+  out[i] = in[src];
+}`}function rx(n,e,t,r,s){let o=sy(s),i=`gather_${s}`,a=Mo.get(i);a||(a=Pz(o,i),Mo.set(i,a));let u=e.length,p=1;for(let h=0;h<u;h++)p*=e[h];let c=new(St(s))(Math.max(p,1)),f=new Array(8).fill(1),d=new Array(8).fill(0);for(let h=0;h<u;h++)f[h]=e[h],d[h]=t[h];return nx({source:a,name:i,metadata:RR(p,[1])},[n,c,p,u,...f,...d,r|0],null),c}function $z(n,e){return`extern "C" __global__ void ${e}(const ${n}* in, ${n}* out, int pre, int dk, int tail, int total, int offset) {
+  int i = blockIdx.x*blockDim.x + threadIdx.x; int n = pre*dk*tail; if (i >= n) return;
+  int t = i % tail; int r = i / tail; int j = r % dk; int p = r / dk;
+  out[(long long)p*total*tail + (long long)(offset+j)*tail + t] = in[i];
+}`}function MR(n,e,t,r,s,o,i){let a=sy(i),u=`catcopy_${i}`,p=Mo.get(u);p||(p=$z(a,u),Mo.set(u,p));let c=n==="stack",f=s.length,d=r<0?f+r:r,h=1;for(let _=0;_<d;_++)h*=s[_];let m=1;for(let _=d+1;_<f;_++)m*=s[_];let g=s[d],y=0;for(let _=0;_<e.length;_++){let b=c?1:t[_][d],w=h*b*m;nx({source:p,name:u,metadata:RR(w,[1])},[e[_],o,h,b,m,g,y],null),y+=b}}function Fz(n,e){return`extern "C" __global__ void ${e}(${n}* w, const ${n}* g, ${n}* m, ${n}* v, int n, float b1, float b2, float ob1, float ob2, float eps, float ss, float bc2s, float wd) {
+  int i = blockIdx.x*blockDim.x + threadIdx.x; if (i >= n) return;
+  float gi = (float)g[i] + wd * (float)w[i];
+  float mi = b1*(float)m[i] + ob1*gi;
+  float vi = b2*(float)v[i] + ob2*gi*gi;
+  m[i] = (${n})mi; v[i] = (${n})vi;
+  w[i] = (${n})((float)w[i] - ss * mi / (sqrtf(vi)/bc2s + eps));
+}`}function zz(n){return`extern "C" __global__ void ${n}(int* t) {
+  if (threadIdx.x == 0 && blockIdx.x == 0) t[0] = t[0] + 1;
+}`}function Vz(n,e){return`extern "C" __global__ void ${e}(${n}* w, const ${n}* g, ${n}* m, ${n}* v, const int* t, int n, float b1, float b2, float ob1, float ob2, float eps, float lr, float wd) {
+  int i = blockIdx.x*blockDim.x + threadIdx.x; if (i >= n) return;
+  int step = t[0];
+  double bc1 = 1.0 - pow((double)b1, (double)step);
+  double bc2 = 1.0 - pow((double)b2, (double)step);
+  float ss = (float)((double)lr / bc1);
+  float bc2s = (float)sqrt(bc2);
+  float gi = (float)g[i] + wd * (float)w[i];
+  float mi = b1*(float)m[i] + ob1*gi;
+  float vi = b2*(float)v[i] + ob2*gi*gi;
+  m[i] = (${n})mi; v[i] = (${n})vi;
+  w[i] = (${n})((float)w[i] - ss * mi / (sqrtf(vi)/bc2s + eps));
+}`}function Gz(n){return`extern "C" __global__ void ${n}(const float* g, double* acc, int n) {
+  extern __shared__ double sdata[];
+  int tid = threadIdx.x;
+  int i = blockIdx.x*blockDim.x + threadIdx.x;
+  double v = (i < n) ? (double)g[i]*(double)g[i] : 0.0;
+  sdata[tid] = v;
+  __syncthreads();
+  for (int s = blockDim.x/2; s > 0; s >>= 1) { if (tid < s) sdata[tid] += sdata[tid+s]; __syncthreads(); }
+  if (tid == 0) atomicAdd(acc, sdata[0]);
+}`}function jz(n){return`extern "C" __global__ void ${n}(const double* acc, const float* mp, float* coef) {
+  if (threadIdx.x == 0 && blockIdx.x == 0) {
+    double norm = sqrt(acc[0]);
+    double c = (double)mp[0] / (norm + (double)mp[1]);
+    coef[0] = c < 1.0 ? (float)c : 1.0f;
+  }
+}`}function Uz(n){return`extern "C" __global__ void ${n}(float* g, const float* coef, int n) {
+  int i = blockIdx.x*blockDim.x + threadIdx.x; if (i < n) g[i] = g[i] * coef[0];
+}`}function qz(n,e,t=1e-6){He(),yi===null&&(yi=ht(8),lf=ht(4),cf=ht(8));let r=He().stream;zn()||xs(cf,new Float32Array([e,t])),zn()?we.memsetD8Async(yi,0,8,r):we.memsetD8(yi,0,8);let s=wo(Gz("clip_accum"),"clip_accum").func,o=wo(jz("clip_coef"),"clip_coef").func,i=wo(Uz("clip_scale"),"clip_scale").func,a=[];for(let u of n){if(!u.grad)continue;let p=u.grad._impl.storage.rawData;if(!p)continue;let c=p.length,f=gg(p);a.push([f,c]),ns(s,[Math.ceil(c/256),1,1],[256,1,1],256*8,[f,yi],[c],!1)}if(a.length!==0){ns(o,[1,1,1],[1,1,1],0,[yi,cf,lf],[],!1);for(let[u,p]of a)ns(i,[Math.ceil(p/256),1,1],[256,1,1],0,[u,lf],[p],!1)}}function IR(n){let e=n&&n._state;if(!(!e||typeof e.values!="function"))for(let t of e.values())t&&(t._mDev!==void 0&&(ci(t._mDev),delete t._mDev),t._vDev!==void 0&&(ci(t._vDev),delete t._vDev),t._tDev!==void 0&&(ci(t._tDev),delete t._tDev))}function ER(){yi!==null&&(ci(yi),ci(lf),ci(cf),yi=null,lf=null,cf=null)}function Wz(){return OT(),ER(),tg()}function Kz(n,e){let t=n&&n.__eagerGraphRunner;if(t){try{gT(t.captured)}catch{}n._cudaGraphPhase=t.phase,delete n.__eagerGraphRunner}if(e)for(let r of e)IR(r);return Nl(!1),yg(),ER(),tg()}function LR(){let n="step_inc",e=Mo.get(n);return e||(e=zz(n),Mo.set(n,e)),{source:e,name:n}}function CR(n="f32"){let e=`adam_graph_${n}`,t=Mo.get(e);return t||(t=Vz(sy(n),e),Mo.set(e,t)),{source:t,name:e}}function Hz(n,e){xs(n,new Int32Array([e|0]))}function DR(n,e,t){if(!In())return!1;let r=n.dtype||"f32";if(r!=="f32")return!1;let s=n._impl.storage.rawData,o=n.grad._impl.storage.rawData;if(!s||!o)return!1;let i=s.length;He();let a=cO();e._mDev||(e._mDev=ht(i*4),we.memsetD8(e._mDev,0,i*4),e._vDev=ht(i*4),we.memsetD8(e._vDev,0,i*4),a&&(e._tDev=ht(4),we.memsetD8(e._tDev,0,4)),kT(s));let u=gg(s),p=To(o);if(zn()){if(e._tDev===void 0)throw new Error("CUDA graph Adam requires a device step counter; warmup must run before capture");let h=LR(),m=CR(r),g=wo(h.source,h.name).func,y=wo(m.source,m.name).func;return ns(g,[1,1,1],[1,1,1],0,[e._tDev],[],!1),ns(y,[Math.ceil(i/256),1,1],[256,1,1],0,[u,p,e._mDev,e._vDev,e._tDev],[i,rr(t.beta1),rr(t.beta2),rr(t.omb1),rr(t.omb2),rr(t.eps),rr(t.lr),rr(t.wd)],!1),!0}let c=`adam_${r}`,f=Mo.get(c);f||(f=Fz(sy(r),c),Mo.set(c,f));let{func:d}=wo(f,c);return ns(d,[Math.ceil(i/256),1,1],[256,1,1],0,[u,p,e._mDev,e._vDev],[i,rr(t.beta1),rr(t.beta2),rr(t.omb1),rr(t.omb2),rr(t.eps),rr(t.stepSize),rr(t.bc2sqrt),rr(t.wd)],!1),a&&e._tDev!==void 0&&Hz(e._tDev,e.step),!0}function Yz(n){if(!In()||!n.device||n.device.type!=="gpu")return null;let e=rx(n._impl.storage.rawData,n.shape,n.strides,n._impl.storageOffset,n.dtype);return Ee(e,[...n.shape],n.dtype,n.device)}function Zz(n,e,t,r){if(!In()||r!==0||t.length!==2||!n.device||n.device.type!=="gpu"||e.dtype!=="i32")return null;let s=t[0],o=t[1],i=ls(n),a=ls(e),u=a.numel;if(i.numel!==u*o)return null;He();let p=new Float32Array(s*o),{func:c}=wo(Jz,"index_add0"),f=Xi(p),d=He().stream;zn()?we.memsetD8Async(f,0,s*o*4,d):we.memsetD8(f,0,s*o*4);let h=To(i._impl.storage.rawData),m=To(a._impl.storage.rawData);return ns(c,[Math.ceil(u*o/256),1,1],[256,1,1],0,[h,m,f],[u,o],!1),Ee(p,[s,o],n.dtype,n.device)}var uf,Dz,Mo,yi,lf,cf,Jz,iy=M(()=>{"use strict";Ts();nT();pi();uT();MO();qp();fi();ug();ws();IT();Ym();wn();be();GT();UT();qn();rs();WT();XB();sR();kR();Za.setHostReadHook(BT);l(NR,"measureCudaKernel");uf=null;l(AR,"preloadCublas");l(OR,"runCudaKernelSync");l(nx,"runCudaKernelResident");l(BR,"_acquireScratch");l(tx,"_releaseScratch");l(Cz,"runCudaKernel");Dz={f32:"float",f64:"double",i64:"long long",i32:"int",i16:"short",i8:"signed char",u8:"unsigned char",bool:"unsigned char"};l(sy,"_ctype");Mo=new Map;l(RR,"_meta");l(Pz,"_gatherKernel");l(rx,"deviceContiguous");l($z,"_catKernel");l(MR,"deviceConcat");l(Fz,"_adamKernel");l(zz,"_stepIncKernel");l(Vz,"_adamGraphKernel");l(Gz,"_clipAccumKernel");l(jz,"_clipCoefKernel");l(Uz,"_clipScaleKernel");yi=null,lf=null,cf=null;l(qz,"deviceClipGradNorm");l(IR,"freeOptimizerDeviceState");l(ER,"freeClipScratch");l(Wz,"releaseCudaMemory");l(Kz,"teardownAfterFit");l(LR,"stepIncKernelSource");l(CR,"adamGraphKernelSource");l(Hz,"_setDeviceInt");l(DR,"deviceAdam");l(Yz,"gpuContiguousTensor");AB(Yz);Jz=`extern "C" __global__ void index_add0(const float* g, const int* idx, float* out, int K, int E) {
+  int t = blockIdx.x*blockDim.x + threadIdx.x; if (t >= K*E) return;
+  int i = t / E, j = t % E;
+  atomicAdd(&out[idx[i]*E + j], g[i*E + j]);
+}`;l(Zz,"gpuIndexSelectBackward");Pl.setGpuBackward(Zz);sx(rx);ox(MR);ux(DR);lx(rR);PB()&&(ix(ZB),ax(QB));vR();Hm("cuda",NR)});var JR={};at(JR,{ensureWebGPUEager:()=>s3,flushWebGPUEager:()=>HR,instantiateWebGPU:()=>e3,prewarmPipelines:()=>WR,recordWebGPUEager:()=>hf,resetDevice:()=>Qz,runWebGPUKernel:()=>t3,runWebGPUPlan:()=>n3,webgpuEagerOp:()=>a3,webgpuEagerReady:()=>o3,webgpuRNN:()=>p3});function Vl(n){switch(jn(n)){case"i32":return Int32Array;case"u32":return Uint32Array;case"f16":return Uint16Array;default:return Float32Array}}function ly(n,e,t,r){if(t==="bf16")for(let s=0;s<e.length;s++)n[r+s]=Lm(e[s]);else if(t==="i64")for(let s=0;s<e.length;s++)n[r+s]=Number(BigInt.asIntN(32,e[s]));else n.set(e,r)}function hx(n,e,t,r,s){if(t==="bf16")for(let o=0;o<s;o++)n[o]=Cm(e[r+o]);else if(t==="i64")for(let o=0;o<s;o++)n[o]=BigInt(e[r+o]);else n.set(e.subarray(r,r+s))}function Eo(n){return Math.ceil(n/4)*4}async function mx(){return It||pf||(pf=(async()=>{let n=typeof navigator<"u"&&navigator.gpu;if(!n)try{let i=await import("webgpu");VR=i,n=i.create([]),i.globals&&(ff=i.globals.GPUBufferUsage,ay=i.globals.GPUMapMode,uy=i.globals.GPUShaderStage)}catch{throw new Error('WebGPU not available: install the "webgpu" npm package or run in a browser with WebGPU support')}!ff&&typeof GPUBufferUsage<"u"&&(ff=GPUBufferUsage,ay=GPUMapMode,uy=GPUShaderStage);let e=await n.requestAdapter();if(!e)throw new Error("WebGPU: no adapter found");let t=e.limits||{},r={},s=["maxStorageBuffersPerShaderStage","maxStorageBufferBindingSize","maxBufferSize","maxBindingsPerBindGroup"];for(let i of s)t[i]!==void 0&&(r[i]=t[i]);let o=[];return e.features&&e.features.has("shader-f16")&&o.push("shader-f16"),It=await e.requestDevice({requiredLimits:r,requiredFeatures:o}),!PR&&typeof process<"u"&&process.on&&(PR=!0,process.on("exit",()=>{It&&(It.destroy(),It=null)})),It})(),pf)}function Gl(){return ff}function gx(){return ay.READ}function Qz(){It&&(It.destroy(),It=null),pf=null,ff=null,ay=null,uy=null,VR=null}function Xz(n){return n.name==="_shapes"?"uniform":n.mode==="read_write"?"storage":"read-only-storage"}function GR(n,e){let t=n.createShaderModule({code:e.source}),r=[];for(let i of e.metadata.bindings)r.push({binding:i.index,visibility:uy.COMPUTE,buffer:{type:Xz(i)}});let s=n.createBindGroupLayout({entries:r}),o=n.createPipelineLayout({bindGroupLayouts:[s]});return{shaderModule:t,bindGroupLayout:s,pipelineLayout:o}}function jR(n,e){let{shaderModule:t,bindGroupLayout:r,pipelineLayout:s}=GR(n,e);return{pipeline:n.createComputePipeline({layout:s,compute:{module:t,entryPoint:e.name}}),bindGroupLayout:r}}function UR(n,e){let t=cy.get(e);return t||(t=jR(n,e),cy.set(e,t)),t}async function e3(n){let e=await mx(),{pipeline:t,bindGroupLayout:r}=jR(e,n);return{device:e,pipeline:t,bindGroupLayout:r,kernel:n,workgroupSize:n.metadata.workgroupSize,dispatchSize:n.metadata.dispatchSize,bindings:n.metadata.bindings}}function qR(n){let e=new Map,t=0;for(let r of n)if(r.name!=="_shapes")if(r.packed)for(let s of r.packed)e.set(s.name,s.argIndex!==void 0?s.argIndex:t++);else e.set(r.name,r.argIndex!==void 0?r.argIndex:t++);return e}async function t3(n,e,t){let{device:r,pipeline:s,bindGroupLayout:o,bindings:i,dispatchSize:a}=n,u=Gl(),p=[],c=[],f=qR(i);for(let y=0;y<i.length;y++){let _=i[y];if(_.name==="_shapes"){let k=new Uint32Array(t||[]),O=Math.max(Math.ceil(k.byteLength/16)*16,16),R=r.createBuffer({size:O,usage:u.UNIFORM|u.COPY_DST});r.queue.writeBuffer(R,0,k),p.push(R),c.push({binding:_.index,resource:{buffer:R}});continue}let b=Vs(_.dtype),w=Vl(_.dtype);if(_.packed){let k=Eo(_.packedSize*b),R=_.mode==="read_write"?u.STORAGE|u.COPY_SRC|u.COPY_DST:u.STORAGE|u.COPY_DST,B=r.createBuffer({size:Math.max(k,4),usage:R,mappedAtCreation:!0}),E=new w(B.getMappedRange());for(let F of _.packed){let C=f.get(F.name),H=e[C];H&&ly(E,H,F.dtype,F.offset)}B.unmap(),p.push(B),c.push({binding:_.index,resource:{buffer:B}});continue}let x=f.get(_.name),S=e[x],T=Eo(S.length*b),A=_.mode==="read_write"?u.STORAGE|u.COPY_SRC|u.COPY_DST:u.STORAGE|u.COPY_DST,N=r.createBuffer({size:Math.max(T,4),usage:A,mappedAtCreation:!0});ly(new w(N.getMappedRange()),S,_.dtype,0),N.unmap(),p.push(N),c.push({binding:_.index,resource:{buffer:N}})}let d=r.createBindGroup({layout:o,entries:c}),h=r.createCommandEncoder(),m=h.beginComputePass();m.setPipeline(s),m.setBindGroup(0,d),m.dispatchWorkgroups(a[0],a[1],a[2]),m.end();let g=[];for(let y=0;y<i.length;y++){let _=i[y];if(_.name==="_shapes"||_.mode!=="read_write")continue;let b=p[y],w=Vs(_.dtype),x=Vl(_.dtype);if(_.packed)for(let S of _.packed){let T=S.offset*w,v=Eo(S.size*w),A=r.createBuffer({size:v,usage:u.MAP_READ|u.COPY_DST});h.copyBufferToBuffer(b,T,A,0,v),g.push({readBuf:A,tensorIdx:f.get(S.name),size:S.size,dtype:S.dtype,ViewCtor:x})}else{let S=b.size,T=r.createBuffer({size:S,usage:u.MAP_READ|u.COPY_DST});h.copyBufferToBuffer(b,0,T,0,S),g.push({readBuf:T,tensorIdx:f.get(_.name),size:e[f.get(_.name)].length,dtype:_.dtype,ViewCtor:x})}}r.queue.submit([h.finish()]);for(let y of g){await y.readBuf.mapAsync(gx());let _=new y.ViewCtor(y.readBuf.getMappedRange());hx(e[y.tensorIdx],_,y.dtype,0,y.size),y.readBuf.unmap(),y.readBuf.destroy()}for(let y of p)y.destroy()}async function WR(n,e){if(typeof n.createComputePipelineAsync!="function")return;let t=[],r=new Set;for(let s of e){if(!s||r.has(s)||cy.has(s))continue;r.add(s);let{shaderModule:o,bindGroupLayout:i,pipelineLayout:a}=GR(n,s);t.push(n.createComputePipelineAsync({layout:a,compute:{module:o,entryPoint:s.name}}).then(u=>cy.set(s,{pipeline:u,bindGroupLayout:i}),()=>{}))}t.length>0&&await Promise.all(t)}async function n3(n,e,t){let r=await mx(),s=Gl();await WR(r,t.map(b=>b.kernel));let o=new Set;for(let b of t)for(let w of b.outputSlots)o.add(w);let i=new Array(n.numSlots).fill(null),a=new Array(n.numSlots).fill("f32");for(let b=0;b<n.numSlots;b++){let w=e[b];if(!w)continue;let x=w.dtype||"f32";a[b]=x;let S=Math.max(Eo(w.data.length*Vs(x)),4),T=!o.has(b),v=r.createBuffer({size:S,usage:s.STORAGE|s.COPY_DST|s.COPY_SRC,mappedAtCreation:T});T&&(ly(new(Vl(x))(v.getMappedRange()),w.data,x,0),v.unmap()),i[b]=v}let u=32,p=[],c=[],f={encoder:r.createCommandEncoder(),pending:0},d=l(()=>{++f.pending>=u&&(r.queue.submit([f.encoder.finish()]),f.encoder=r.createCommandEncoder(),f.pending=0)},"maybeFlush"),h=l(b=>{let w=f.encoder,{pipeline:x,bindGroupLayout:S}=UR(r,b.kernel),T=b.inputSlots.concat(b.outputSlots),v=[],A=[],N=0;for(let B of b.kernel.metadata.bindings)if(B.name==="_shapes"){let E=new Uint32Array(b.shapeValues||[]),F=Math.max(Math.ceil(E.byteLength/16)*16,16),C=r.createBuffer({size:F,usage:s.UNIFORM|s.COPY_DST});r.queue.writeBuffer(C,0,E),p.push(C),v.push({binding:B.index,resource:{buffer:C}})}else if(B.packed){let E=Vs(B.dtype),F=r.createBuffer({size:Math.max(Eo(B.packedSize*E),4),usage:s.STORAGE|s.COPY_DST|s.COPY_SRC});c.push(F);let C=B.mode==="read_write";for(let H of B.packed){let G=T[H.argIndex],U=Eo(H.size*E);C?A.push({slot:G,src:F,srcOff:H.offset*E,bytes:U}):w.copyBufferToBuffer(i[G],0,F,H.offset*E,U)}v.push({binding:B.index,resource:{buffer:F}})}else v.push({binding:B.index,resource:{buffer:i[T[N++]]}});let k=r.createBindGroup({layout:S,entries:v}),O=w.beginComputePass();O.setPipeline(x),O.setBindGroup(0,k);let R=b.kernel.metadata.dispatchSize;O.dispatchWorkgroups(R[0],R[1],R[2]),O.end();for(let B of A)w.copyBufferToBuffer(B.src,B.srcOff,i[B.slot],0,B.bytes);d()},"encodeStep"),m=n.scanLoops?[...n.scanLoops].sort((b,w)=>b.loopStart-w.loopStart):n.scanLoop?[n.scanLoop]:null;if(m&&m.length){let b=0;for(let w of m){for(;b<w.loopStart;b++)h(t[b]);for(let x of w.carry)f.encoder.copyBufferToBuffer(i[x.initSlot],0,i[x.a],0,x.bytes);for(let x=0;x<w.T;x++){for(let S of w.xs)f.encoder.copyBufferToBuffer(i[S.xsSlot],x*S.stepBytes,i[S.xtSlot],0,S.stepBytes);for(let S=w.loopStart;S<w.loopEnd;S++)h(t[S]);for(let S of w.ys)f.encoder.copyBufferToBuffer(i[S.ytSlot],0,i[S.ysSlot],x*S.stepBytes,S.stepBytes);for(let S of w.carry){let T=i[S.a];i[S.a]=i[S.b],i[S.b]=T}}for(let x of w.carry)f.encoder.copyBufferToBuffer(i[x.a],0,i[x.finalSlot],0,x.bytes);b=w.loopEnd}for(;b<t.length;b++)h(t[b])}else for(let b of t)h(b);let g=f.encoder,y=new Set(n.argSlots),_=[];for(let b=0;b<n.numSlots;b++){if(!i[b]||!o.has(b)||!y.has(b))continue;let w=e[b],x=Math.max(Eo(w.data.length*Vs(a[b])),4),S=r.createBuffer({size:x,usage:s.MAP_READ|s.COPY_DST});g.copyBufferToBuffer(i[b],0,S,0,x),_.push({rb:S,dtype:a[b],size:w.data.length,dst:w.data})}r.queue.submit([g.finish()]);for(let b of _)await b.rb.mapAsync(gx()),hx(b.dst,new(Vl(b.dtype))(b.rb.getMappedRange()),b.dtype,0,b.size),b.rb.unmap(),b.rb.destroy();for(let b of i)b&&b.destroy();for(let b of p)b.destroy();for(let b of c)b.destroy()}function KR(n){let e=dx.get(n);if(e&&e.length)return e.pop();let t=Gl();return It.createBuffer({size:n,usage:t.STORAGE|t.COPY_SRC|t.COPY_DST})}function $R(n,e){let t=dx.get(e);t||(t=[],dx.set(e,t)),t.push(n)}async function s3(){return await mx(),Qs||(Qs=It.createCommandEncoder()),It}function o3(){return!!It&&!!Qs}function FR(n,e,t){let r=df.get(n);if(r)return r.buf;let s=Gl(),o=Math.max(Eo(n.length*Vs(e)),4);if(t){let a=It.createBuffer({size:o,usage:s.STORAGE|s.COPY_SRC|s.COPY_DST,mappedAtCreation:!0});return ly(new(Vl(e))(a.getMappedRange()),n,e,0),a.unmap(),df.set(n,{buf:a,bytes:o,dtype:e,persistent:t}),a}let i=KR(o);return df.set(n,{buf:i,bytes:o,dtype:e,persistent:!1}),i}function hf(n,e,t){let{pipeline:r,bindGroupLayout:s}=UR(It,n),{bindings:o,dispatchSize:i}=n.metadata,a=qR(o),u=Gl(),p=[],c=[];for(let h of o){if(h.name==="_shapes"){let y=new Uint32Array(t||[]),_=Math.max(Math.ceil(y.byteLength/16)*16,16),b=It.createBuffer({size:_,usage:u.UNIFORM|u.COPY_DST});It.queue.writeBuffer(b,0,y),px.push(b),p.push({binding:h.index,resource:{buffer:b}});continue}if(h.packed){let y=Vs(h.dtype),_=Math.max(Eo(h.packedSize*y),4),b=KR(_);fx.push({buf:b,bytes:_});let w=h.mode==="read_write";for(let x of h.packed){let S=FR(e[a.get(x.name)],x.dtype,!w),T=Eo(x.size*y);w?c.push({pbuf:b,off:x.offset*y,sub:S,bytes:T}):Qs.copyBufferToBuffer(S,0,b,x.offset*y,T)}p.push({binding:h.index,resource:{buffer:b}});continue}let m=e[a.get(h.name)],g=FR(m,h.dtype,h.mode!=="read_write");p.push({binding:h.index,resource:{buffer:g}})}let f=It.createBindGroup({layout:s,entries:p}),d=Qs.beginComputePass();d.setPipeline(r),d.setBindGroup(0,f),d.dispatchWorkgroups(i[0],i[1],i[2]),d.end();for(let h of c)Qs.copyBufferToBuffer(h.pbuf,h.off,h.sub,0,h.bytes);++cx>=r3&&(It.queue.submit([Qs.finish()]),Qs=It.createCommandEncoder(),cx=0)}async function HR(){if(!It)return;let n=Gl(),e=[];for(let[t,r]of df){if(r.persistent)continue;let s=It.createBuffer({size:r.bytes,usage:n.MAP_READ|n.COPY_DST});Qs.copyBufferToBuffer(r.buf,0,s,0,r.bytes),e.push({rb:s,ha:t,buf:r.buf,bytes:r.bytes,dtype:r.dtype})}It.queue.submit([Qs.finish()]),Qs=It.createCommandEncoder(),cx=0;for(let t of px)t.destroy();px.length=0,await Promise.all(e.map(t=>t.rb.mapAsync(gx())));for(let t of e)hx(t.ha,new(Vl(t.dtype))(t.rb.getMappedRange()),t.dtype,0,t.ha.length),t.rb.unmap(),t.rb.destroy();for(let t of fx)$R(t.buf,t.bytes);fx.length=0;for(let t of e)$R(t.buf,t.bytes),df.delete(t.ha)}function i3(n,e,t,r,s){let o=`${n.join(",")}|${e.join(",")}|${t}|${r}`,i=zR.get(o);if(i)return i;let a=jn(r),u="";for(let c=n.length-1;c>=0;c--)u+=`  let idx${c} = rem % ${n[c]}u; rem = rem / ${n[c]}u; src = src + idx${c} * ${e[c]}u;
+`;return i={name:"contig",source:`@group(0) @binding(0) var<storage, read> inp : array<${a}>;
+@group(0) @binding(1) var<storage, read_write> outp : array<${a}>;
+@compute @workgroup_size(64)
+fn contig(@builtin(global_invocation_id) gid : vec3<u32>) {
+  let i = gid.x;
+  if (i >= ${s}u) { return; }
+  var rem = i;
+  var src = ${t}u;
+`+u+`  outp[i] = inp[src];
+}`,metadata:{bindings:[{index:0,name:"inp",dtype:r,mode:"read"},{index:1,name:"outp",dtype:r,mode:"read_write"}],dispatchSize:[Math.ceil(s/64),1,1]}},zR.set(o,i),i}function zl(n){let e=n._impl.storage.rawData;if(n.isContiguous&&n._impl.storageOffset===0&&e.length===n.numel)return e;let t=new e.constructor(n.numel);return hf(i3(n.shape,n.strides,n._impl.storageOffset,n.dtype,n.numel),[e,t],void 0),t}function a3(n,e,t){let r=e.map(zl);r.push(t),hf(n,r,void 0)}function u3(n,e,t){let r=`lstm|${n}|${e}|${t}`,s=py.get(r);if(s)return s;let o=4*e,i=o*n,a=o*n+o*e,u=a+o;return s={name:"rnn",source:`fn sig(x: f32) -> f32 { return 1.0 / (1.0 + exp(-x)); }
+@group(0) @binding(0) var<storage, read> xin : array<f32>;
+@group(0) @binding(1) var<storage, read> w : array<f32>;
+@group(0) @binding(2) var<storage, read> h0 : array<f32>;
+@group(0) @binding(3) var<storage, read> c0 : array<f32>;
+@group(0) @binding(4) var<storage, read_write> ys : array<f32>;
+@group(0) @binding(5) var<storage, read_write> hn : array<f32>;
+@group(0) @binding(6) var<storage, read_write> cn : array<f32>;
+var<workgroup> wh : array<f32, ${e}>;
+var<workgroup> wc : array<f32, ${e}>;
+var<workgroup> wg : array<f32, ${o}>;
+@compute @workgroup_size(${Io})
+fn rnn(@builtin(local_invocation_id) lid : vec3<u32>) {
+  let tid = lid.x;
+  for (var p = tid; p < ${e}u; p = p + ${Io}u) { wh[p] = h0[p]; wc[p] = c0[p]; }
+  workgroupBarrier();
+  for (var t = 0u; t < ${t}u; t = t + 1u) {
+    for (var j = tid; j < ${o}u; j = j + ${Io}u) {
+      var acc = w[${a}u + j] + w[${u}u + j];
+      let xb = t * ${n}u;
+      let wxb = j * ${n}u;
+      for (var e = 0u; e < ${n}u; e = e + 1u) { acc = acc + xin[xb + e] * w[wxb + e]; }
+      let whb = ${i}u + j * ${e}u;
+      for (var q = 0u; q < ${e}u; q = q + 1u) { acc = acc + wh[q] * w[whb + q]; }
+      wg[j] = acc;
+    }
+    workgroupBarrier();
+    for (var j = tid; j < ${e}u; j = j + ${Io}u) {
+      let ii = sig(wg[j]);
+      let ff = sig(wg[${e}u + j]);
+      let gg = tanh(wg[${2*e}u + j]);
+      let oo = sig(wg[${3*e}u + j]);
+      let cv = ff * wc[j] + ii * gg;
+      wc[j] = cv;
+      let nh = oo * tanh(cv);
+      wh[j] = nh;
+      ys[t * ${e}u + j] = nh;
+    }
+    workgroupBarrier();
+  }
+  for (var p = tid; p < ${e}u; p = p + ${Io}u) { hn[p] = wh[p]; cn[p] = wc[p]; }
+}`,metadata:{bindings:[{index:0,name:"xin",dtype:"f32",mode:"read"},{index:1,name:"w",dtype:"f32",mode:"read"},{index:2,name:"h0",dtype:"f32",mode:"read"},{index:3,name:"c0",dtype:"f32",mode:"read"},{index:4,name:"ys",dtype:"f32",mode:"read_write"},{index:5,name:"hn",dtype:"f32",mode:"read_write"},{index:6,name:"cn",dtype:"f32",mode:"read_write"}],dispatchSize:[1,1,1]}},py.set(r,s),s}function YR(n,e,t,r){let s=`${e}|${t}|${r}`;if(n._webgpuPacked&&n._webgpuPackedTag===s)return n._webgpuPacked;let o=l(f=>f.contiguous()._impl.storage.rawData,"raw"),i=o(n.x2h.weight),a=o(n.h2h.weight),u=n.x2h.bias?o(n.x2h.bias):new Float32Array(r),p=n.h2h.bias?o(n.h2h.bias):new Float32Array(r),c=new Float32Array(r*e+r*t+r+r);return c.set(i,0),c.set(a,r*e),c.set(u,r*e+r*t),c.set(p,r*e+r*t+r),n._webgpuPacked=c,n._webgpuPackedTag=s,c}function l3(n,e,t){return YR(n,e,t,4*t)}function c3(n,e,t){let r=`gru|${n}|${e}|${t}`,s=py.get(r);if(s)return s;let o=3*e,i=o*n,a=o*n+o*e,u=a+o;return s={name:"rnn",source:`fn sig(x: f32) -> f32 { return 1.0 / (1.0 + exp(-x)); }
+@group(0) @binding(0) var<storage, read> xin : array<f32>;
+@group(0) @binding(1) var<storage, read> w : array<f32>;
+@group(0) @binding(2) var<storage, read> h0 : array<f32>;
+@group(0) @binding(3) var<storage, read_write> ys : array<f32>;
+@group(0) @binding(4) var<storage, read_write> hn : array<f32>;
+var<workgroup> wh : array<f32, ${e}>;
+var<workgroup> wgx : array<f32, ${o}>;
+var<workgroup> wgh : array<f32, ${o}>;
+@compute @workgroup_size(${Io})
+fn rnn(@builtin(local_invocation_id) lid : vec3<u32>) {
+  let tid = lid.x;
+  for (var p = tid; p < ${e}u; p = p + ${Io}u) { wh[p] = h0[p]; }
+  workgroupBarrier();
+  for (var t = 0u; t < ${t}u; t = t + 1u) {
+    for (var j = tid; j < ${o}u; j = j + ${Io}u) {
+      var ax = w[${a}u + j];
+      var ah = w[${u}u + j];
+      let xb = t * ${n}u;
+      let wxb = j * ${n}u;
+      for (var e = 0u; e < ${n}u; e = e + 1u) { ax = ax + xin[xb + e] * w[wxb + e]; }
+      let whb = ${i}u + j * ${e}u;
+      for (var q = 0u; q < ${e}u; q = q + 1u) { ah = ah + wh[q] * w[whb + q]; }
+      wgx[j] = ax;
+      wgh[j] = ah;
+    }
+    workgroupBarrier();
+    for (var j = tid; j < ${e}u; j = j + ${Io}u) {
+      let rr = sig(wgx[j] + wgh[j]);
+      let zz = sig(wgx[${e}u + j] + wgh[${e}u + j]);
+      let nn = tanh(wgx[${2*e}u + j] + rr * wgh[${2*e}u + j]);
+      let nh = (1.0 - zz) * nn + zz * wh[j];
+      wh[j] = nh;
+      ys[t * ${e}u + j] = nh;
+    }
+    workgroupBarrier();
+  }
+  for (var p = tid; p < ${e}u; p = p + ${Io}u) { hn[p] = wh[p]; }
+}`,metadata:{bindings:[{index:0,name:"xin",dtype:"f32",mode:"read"},{index:1,name:"w",dtype:"f32",mode:"read"},{index:2,name:"h0",dtype:"f32",mode:"read"},{index:3,name:"ys",dtype:"f32",mode:"read_write"},{index:4,name:"hn",dtype:"f32",mode:"read_write"}],dispatchSize:[1,1,1]}},py.set(r,s),s}function p3(n,e,t,r,s){if(t.kind==="gru")return f3(n,e,t,r);if(t.batch!==1)return null;let o=t.hiddenSize,i=t.seqLen,a=e.length,u=zl(n),p=[],c=[];for(let h=0;h<a;h++){let m=h===0?t.inputSize:o,g=l3(e[h],m,o),y,_;r!=null&&s!=null?(y=zl(en(r,0,h)),_=zl(en(s,0,h))):(y=new Float32Array(o),_=new Float32Array(o));let b=new Float32Array(i*o),w=new Float32Array(o),x=new Float32Array(o);hf(u3(m,o,i),[u,g,y,_,b,w,x],void 0),u=b,p.push(w),c.push(x)}let f=Ee(u,[i,1,o],"f32",xo),d=l(h=>a===1?Ee(h[0],[1,1,o],"f32",xo):us(h.map(m=>Ee(m,[1,o],"f32",xo)),0),"wrapState");return[f,d(p),d(c)]}function f3(n,e,t,r){if(t.batch!==1)return null;let s=t.hiddenSize,o=t.seqLen,i=e.length,a=zl(n),u=[];for(let f=0;f<i;f++){let d=f===0?t.inputSize:s,h=YR(e[f],d,s,3*s),m=r!=null?zl(en(r,0,f)):new Float32Array(s),g=new Float32Array(o*s),y=new Float32Array(s);hf(c3(d,s,o),[a,h,m,g,y],void 0),a=g,u.push(y)}let p=Ee(a,[o,1,s],"f32",xo),c=i===1?Ee(u[0],[1,1,s],"f32",xo):us(u.map(f=>Ee(f,[1,s],"f32",xo)),0);return[p,c]}var It,pf,ff,ay,uy,VR,PR,cy,Qs,cx,r3,df,px,fx,dx,zR,Io,py,yx=M(()=>{"use strict";tr();bl();wn();qn();be();be();It=null,pf=null,ff=null,ay=null,uy=null,VR=null,PR=!1;l(Vl,"wgslViewCtor");l(ly,"packTensorInto");l(hx,"unpackTensorFrom");l(Eo,"align4");l(mx,"ensureDevice");l(Gl,"bufUsage");l(gx,"mapModeRead");l(Qz,"resetDevice");l(Xz,"bindingBufferType");l(GR,"pipelineParts");l(jR,"createPipeline");cy=new WeakMap;l(UR,"pipelineFor");l(e3,"instantiateWebGPU");l(qR,"buildParamIndex");l(t3,"runWebGPUKernel");l(WR,"prewarmPipelines");l(n3,"runWebGPUPlan");Qs=null,cx=0,r3=64,df=new Map,px=[],fx=[],dx=new Map;l(KR,"acquireStorage");l($R,"releaseStorage");l(s3,"ensureWebGPUEager");l(o3,"webgpuEagerReady");l(FR,"eagerBufferFor");l(hf,"recordWebGPUEager");l(HR,"flushWebGPUEager");zR=new Map;l(i3,"contigKernel");l(zl,"webgpuEagerInput");l(a3,"webgpuEagerOp");Io=256,py=new Map;l(u3,"lstmKernelWGSL");l(YR,"rnnPackedWeights");l(l3,"lstmPackedWeights");l(c3,"gruKernelWGSL");l(p3,"webgpuRNN");l(f3,"webgpuGRU")});var vx={};at(vx,{getBackend:()=>Sx,hasBackend:()=>d3,preloadCudaRuntime:()=>ZR,preloadWebGPU:()=>gf,registerBackend:()=>yf});function yf(n,e){xx.set(n,e)}function Sx(n){return xx.get(n)||null}function d3(n){return xx.has(n)}function h3(){return _x||(_x=Promise.resolve().then(()=>(V1(),z1))),_x}function Tx(){return bx||(bx=Promise.resolve().then(()=>(iy(),oy))),bx}async function ZR(){if(mf)return mf;let n=await Tx();return n.preloadCublas&&await n.preloadCublas(),mf=n,n}async function gf(){if(wx)return wx;let n=await Promise.resolve().then(()=>(yx(),JR)),{setWebGPUEagerFn:e,setWebgpuRNN:t}=await Promise.resolve().then(()=>(wn(),XR));return await n.ensureWebGPUEager(),e(n.webgpuEagerOp),t(n.webgpuRNN),wx=n,n}function QR(n){let e;try{e=Sl(n.source)}catch(o){throw new Error("encodeWat: "+o.message+`
+`+n.source)}let t;try{t=new WebAssembly.Module(e)}catch(o){throw new Error("WASM: "+o.message+`
+`+n.source)}let r={};if(n.metadata.imports)for(let[o]of n.metadata.imports)r[o]=m3[o]||Math[o]||(i=>i);let s=new WebAssembly.Instance(t,{math:r});return{exports:s.exports,memory:s.exports.memory,bufferOffsets:n.metadata.bufferOffsets,funcName:n.name,binary:e,parallel:n.metadata.parallel||null,mathNames:n.metadata.imports?[...n.metadata.imports.keys()]:[]}}function fy(n,e,t,r,s){let{exports:o,memory:i,bufferOffsets:a,funcName:u}=n,p=o[u],c=[...a.values()],f=Math.min(c.length,e.length);for(let h=0;h<f;h++){let m=e[h];ArrayBuffer.isView(m)&&new m.constructor(i.buffer,c[h],m.length).set(m)}let d=c.slice(0,f);if(t)for(let h of t)d.push(h);r!==void 0&&s!==void 0&&d.push(r,s),p(...d);for(let h=0;h<f;h++){let m=e[h];ArrayBuffer.isView(m)&&m.set(new m.constructor(i.buffer,c[h],m.length))}}function g3(n,e,t=[],r={}){let s=r.warmup??5,o=r.repeat??30,i=QR(n),a=e.map(f=>new Float32Array(Math.max(1,Math.ceil(f/4)))),u=t||[],p=l(()=>{i.parallel?fy(i,a,u,0,i.parallel.extent):fy(i,a,u)},"once");for(let f=0;f<s;f++)p();let c=[];for(let f=0;f<o;f++){let d=performance.now();p(),c.push(performance.now()-d)}return c}var xx,_x,bx,mf,wx,m3,_f=M(()=>{"use strict";Wm();Yi();Ym();xx=new Map;l(yf,"registerBackend");l(Sx,"getBackend");l(d3,"hasBackend");_x=null;l(h3,"getWasmPool");bx=null;l(Tx,"getCudaRuntime");mf=null;l(ZR,"preloadCudaRuntime");wx=null;l(gf,"preloadWebGPU");m3={exp:Math.exp,log:Math.log,sin:Math.sin,cos:Math.cos,tan:Math.tan,tanh:Math.tanh,pow:Math.pow,fmod:l((n,e)=>n%e,"fmod"),rsqrt:l(n=>1/Math.sqrt(n),"rsqrt"),sign:Math.sign,round:Math.round,erf:Wi,erfc:Ki,lgamma:cn,gamma:Hi,digamma:ja};l(QR,"instantiateWasm");l(fy,"runWasmKernel");yf("js",{instantiate(n){return new Function("return "+n.source)()},runSync(n,e,t){let r=t?[...e,...t]:e;return n(...r)},runAsync(n,e,t){let r=t?[...e,...t]:e;return n(...r)},isAsync(){return!1}});yf("wasm",{instantiate(n){return QR(n)},runSync(n,e,t){n.parallel?fy(n,e,t,0,n.parallel.extent):fy(n,e,t)},async runAsync(n,e,t){if(n.parallel&&n.parallel.poolSafe){let{runWasmParallel:r}=await h3();await r(n,n.funcName,e,t,n.parallel,n.mathNames);return}this.runSync(n,e,t)},isAsync(n){return!!(n&&n.parallel&&n.parallel.poolSafe)}});l(g3,"measureWasm");Hm("wasm",g3);yf("webgpu",{instantiate(n){return gf().then(e=>e.instantiateWebGPU(n))},runSync(){throw new Error("WebGPU kernel requires async execution \u2014 use runAsync()")},async runAsync(n,e,t){let{runWebGPUKernel:r}=await gf();await r(n,e,t)},async runPlan(n,e,t){let{runWebGPUPlan:r}=await gf();await r(n,e,t)},isAsync(){return!0}});yf("cuda",{instantiate(n){return{kernel:n}},runSync(n,e,t){if(!mf)throw new Error("CUDA sync runtime not preloaded \u2014 call preloadCudaRuntime() before synchronous execution");mf.runCudaKernelResident(n.kernel,e,t)},async runAsync(n,e,t){let{runCudaKernel:r}=await Tx();await r(n.kernel,e,t)},async runPlan(n,e,t,r){let{runCudaPlan:s}=await Tx();await s(n,e,t,r)},isAsync(){return!0}})});function kx(n){return y3[n]||Float32Array}function _3(n){return n instanceof Float64Array?"f64":n instanceof Int32Array?"i32":n instanceof Int16Array?"i16":n instanceof Int8Array?"i8":n instanceof Uint16Array?"ui16":n instanceof Uint8Array?"ui8":n instanceof BigInt64Array?"i64":"f32"}function eM(n,e){if(!(!n.returnFixups||n.returnFixups.length===0))for(let t of n.returnFixups){let r=e[n.argSlots[t.pos]];if(!(!r||!r.data))if(t.kind==="copy"){let s=e[t.srcSlot];s&&s.data&&r.data.set(s.data.subarray(0,r.data.length))}else t.kind==="const"&&r.data.fill(t.value)}}var y3,Lo,Nx,jl,dy=M(()=>{"use strict";_f();Ep();y3={f16:Uint16Array,bf16:Uint16Array,f32:Float32Array,f64:Float64Array,i8:Int8Array,i16:Int16Array,i32:Int32Array,i64:BigInt64Array,ui8:Uint8Array,bool:Uint8Array,index:Int32Array};l(kx,"typedArrayCtor");l(_3,"dtypeOfTypedArray");l(eM,"_applyReturnFixups");Lo=class n{static{l(this,"RuntimeTensor")}data;shape;dtype;strides;resident;constructor(e,t,r,s=null){this.data=e,this.shape=[...t],this.dtype=r,this.strides=s?[...s]:n.defaultStrides(t)}static defaultStrides(e){let t=new Array(e.length),r=1;for(let s=e.length-1;s>=0;s--)t[s]=r,r*=e[s];return t}get numel(){let e=1;for(let t=0;t<this.shape.length;t++)e*=this.shape[t];return e}get rank(){return this.shape.length}static zeros(e,t="f32"){let r=1;for(let s=0;s<e.length;s++)r*=e[s];return r=Math.max(r,1),new n(new(kx(t))(r),e,t)}static fromArray(e,t,r="f32"){let s=kx(r);return new n(new s(e),t,r)}get(e){let t=0;for(let r=0;r<e.length;r++)t+=e[r]*this.strides[r];return this.data[t]}set(e,t){let r=0;for(let s=0;s<e.length;s++)r+=e[s]*this.strides[s];this.data[r]=t}},Nx=class{static{l(this,"KernelRegistry")}_kernels;constructor(){this._kernels=new Map}register(e,t){this._kernels.set(e,t)}get(e){return this._kernels.get(e)||null}has(e){return this._kernels.has(e)}names(){return[...this._kernels.keys()]}},jl=class n{static{l(this,"RuntimeModule")}name;kernels;_instances;_shapeParamMaps;_bufferMaps;constructor(e){this.name=e,this.kernels=new Nx,this._instances=new Map}addCompiledKernel(e){this.kernels.register(e.name,e);let t=Sx(e.metadata.kind);if(!t)throw new Error("No runtime backend registered for kind: "+e.metadata.kind);this._instances.set(e.name,{backend:t,instance:t.instantiate(e)})}setShapeParamMap(e,t,r){this._shapeParamMaps||(this._shapeParamMaps=new Map),this._shapeParamMaps.set(e,t),r&&(this._bufferMaps||(this._bufferMaps=new Map),this._bufferMaps.set(e,r))}_prepareArgs(e,t){let r=[],s=new Map;for(let a=0;a<t.length;a++){let u=t[a];u instanceof Lo?(r.push(u.data),s.set(a,u.shape)):r.push(u)}let o=this._shapeParamMaps&&this._shapeParamMaps.get(e),i=null;if(o&&o.size>0){let a=this._bufferMaps&&this._bufferMaps.get(e);i=n._extractShapeParams(o,s,t,a)}return{tensorArgs:r,shapeValues:i}}run(e,...t){let r=this._instances.get(e);if(!r)throw new Error("Kernel '"+e+"' not found or not executable");if(r.instance instanceof Promise)throw new Error("Kernel '"+e+"' requires async execution \u2014 use runAsync()");let{tensorArgs:s,shapeValues:o}=this._prepareArgs(e,t);return r.backend.runSync(r.instance,s,o)}async runAsync(e,...t){let r=this._instances.get(e);if(!r)throw new Error("Kernel '"+e+"' not found or not executable");let{tensorArgs:s,shapeValues:o}=this._prepareArgs(e,t),i=await r.instance;return r.backend.runAsync(i,s,o)}isAsync(e){let t=this._instances.get(e);if(!t)return!1;let r=t.instance instanceof Promise?null:t.instance;return t.backend.isAsync(r)}async runPlanAsync(e,t,r){let s=new Array(e.numSlots).fill(null);for(let a=0;a<t.length;a++){let u=t[a];s[e.argSlots[a]]=u instanceof Lo?u:new Lo(u,[u.length],_3(u))}for(let a of e.intermediates){let u=1;for(let p of a.shape)u*=p;s[a.slot]=new Lo(new(kx(a.dtype))(Math.max(u,1)),a.shape,a.dtype)}for(let a of e.steps){let u=this._instances.get(a.name);u&&u.instance instanceof Promise&&(u.instance=await u.instance)}let o=this._uniformPlanBackend(e),i=e.steps.some(a=>{let u=this.kernels.get(a.name);return u&&u.metadata&&u.metadata.scratch&&u.metadata.scratch.length>0});if(o&&o.runPlan&&!i){let a=e.steps.map(u=>{let p=[];for(let f of u.inputSlots)p.push(s[f]);for(let f of u.outputSlots)p.push(s[f]);let{shapeValues:c}=this._prepareArgs(u.name,p);return{name:u.name,inputSlots:u.inputSlots,outputSlots:u.outputSlots,kernel:this.kernels.get(u.name),shapeValues:c}});await o.runPlan(e,s,a,r),eM(e,s);return}for(let a of e.steps){let u=[];for(let p of a.inputSlots)u.push(s[p]);for(let p of a.outputSlots)u.push(s[p]);await this.runAsync(a.name,...u)}eM(e,s)}_uniformPlanBackend(e){let t=null;for(let r of e.steps){let s=this._instances.get(r.name);if(!s||s.instance instanceof Promise)return null;if(t===null)t=s.backend;else if(s.backend!==t)return null}return t}static _extractShapeParams(e,t,r,s){let o=new Map;if(s){let u=0;for(let[p,c]of s){let f=typeof p=="string"?p:c&&c.name;typeof f=="string"&&o.set(f,u),u++}}let i=new Map,a=[];for(let[u,p]of e){if(i.has(p.name))continue;i.set(p.name,!0);let c=u.lastIndexOf(":"),f=u.substring(0,c),d=parseInt(u.substring(c+1),10),h=null;if(o.has(f)){let m=t.get(o.get(f));m&&d<m.length&&m[d]>0&&(h=m[d])}if(h===null){for(let[,m]of t)if(d<m.length&&m[d]>0){h=m[d];break}}a.push(h!==null?h:1)}return a}getKernelSource(e){let t=this.kernels.get(e);return t?t.source:null}getKernelSnippet(e){let t=this.kernels.get(e);return t?t.snippet():null}listKernels(){return this.kernels.names()}serialize(){let e=[];for(let t of this.kernels.names()){let r=this.kernels.get(t);e.push({name:r.name,source:r.source,target:r.target.name,metadata:r.metadata})}return{name:this.name,kernels:e}}static deserialize(e){let t=new n(e.name);for(let r of e.kernels)t.addCompiledKernel(new Ip(r.name,r.source,{name:r.target},r.metadata));return t}}});var hy,tM=M(()=>{"use strict";hy=class{static{l(this,"AnalysisManager")}constructor(){this._cache=new WeakMap}getAnalysis(e,t){let r=this._cache.get(t);r||(r=new Map,this._cache.set(t,r));let s=r.get(e);if(!s||s.version!==t.version){let o=this._resolveDeps(e,t,r);s={data:e.compute(t,o),version:t.version},r.set(e,s)}return s.data}_resolveDeps(e,t,r){let s=e.dependencies;if(!s||s.length===0)return{};let o={};for(let i of s){let a=i.depKey||i.name;o[a]=this.getAnalysis(i,t)}return o}invalidate(e,t=null){let r=this._cache.get(e);if(!r)return;if(!t){this._cache.delete(e);return}let s=l(u=>t.has(u)||t.has(u.name),"isPreserved"),o=new Map,i=l(u=>{let p=o.get(u);if(p!==void 0)return p;let c=u.dependencies,f=!1;if(c){for(let d of c)if(!s(d)||i(d)){f=!0;break}}return o.set(u,f),f},"isStale"),a=[];for(let u of r.keys())if(!s(u)||i(u))a.push(u);else{let p=r.get(u);p&&(p.version=e.version)}for(let u of a)r.delete(u)}invalidateFunctions(e,t=null){for(let r of e)this.invalidate(r,t)}invalidateAll(){this._cache=new WeakMap}}});var oe,my,gy,qe,yy,fn=M(()=>{"use strict";oe=Object.freeze({UNCHANGED:0,CHANGED:1,FAILED:2}),my=class{static{l(this,"Pass")}constructor(e){this.name=e,this.preservedAnalyses=new Set,this.invalidatedAnalyses=new Set,this.requiredAnalyses=[],this.optLevel=0,this.trace=null}run(e,t){throw new Error("Not implemented")}},gy=class{static{l(this,"PassContext")}constructor({optLevel:e=1/0,disabledPasses:t=[],requiredPasses:r=[],config:s={}}={}){this.optLevel=e,this.disabledPasses=t instanceof Set?t:new Set(t),this.requiredPasses=r instanceof Set?r:new Set(r),this.config=s instanceof Map?s:new Map(Object.entries(s))}shouldRun(e){return this.disabledPasses.has(e.name)?!1:this.requiredPasses.has(e.name)?!0:!((e.optLevel||0)>this.optLevel)}},qe=class extends my{static{l(this,"FunctionPass")}run(e,t){throw new Error("Not implemented")}},yy=class extends my{static{l(this,"ModulePass")}run(e,t){throw new Error("Not implemented")}}});var le,Yn,b3,_y,nn=M(()=>{"use strict";fn();le=Object.freeze({SILENT:0,INFO:1,VERBOSE:2,DEBUG:3}),Yn=class{static{l(this,"CompilationError")}constructor(e,t,r,s){this.phase=e,this.funcName=t,this.message=r,this.passName=s||null}toString(){let e="["+this.phase+"]";return this.funcName&&(e+=" "+this.funcName),this.passName&&(e+=" ("+this.passName+")"),e+=": "+this.message,e}},b3=l(()=>{},"NOOP"),_y=class{static{l(this,"TraceLog")}constructor(e={}){this.level=e.level??le.SILENT,this.sink=typeof e.sink=="function"?e.sink:b3,this.irSnapshot={afterGraphPasses:!1,afterLowering:!1,afterScheduling:!1,...e.irSnapshot||{}},this._compileStart=0}emit(e){e.level>this.level||(e.timestamp=performance.now(),this.sink(e))}phaseStart(e){this.emit({type:"phase",action:"start",phase:e,level:le.INFO})}phaseEnd(e,t){this.emit({type:"phase",action:"end",phase:e,durationMs:t,level:le.INFO})}passRun(e,t,r,s,o){this.emit({type:"pass",passName:e,changed:t===oe.CHANGED,durationMs:r,opCountBefore:s,opCountAfter:o,level:le.VERBOSE})}functionEvent(e,t,r){this.emit({type:"function",phase:e,funcName:t,...r,level:le.INFO})}irDump(e,t){this.emit({type:"ir_snapshot",label:e,text:t,level:le.DEBUG})}memoryStats(e,t){this.emit({type:"memory",funcName:e,...t,level:le.VERBOSE})}autotuneStats(e,t){this.emit({type:"autotune",funcName:e,...t,level:le.VERBOSE})}codegenStats(e,t){this.emit({type:"codegen",funcName:e,...t,level:le.VERBOSE})}errorEvent(e,t,r,s){this.emit({type:"error",phase:e,funcName:t,message:r,passName:s||null,level:le.INFO})}warn(e,t,r,s){this.emit({type:"warning",phase:e,funcName:t,message:r,detail:s||null,level:le.INFO})}explain(e,t,r,s,o){this.emit({type:"explain",category:e,subject:t,decision:r,reason:s,...o||{},level:le.DEBUG})}get explainsEnabled(){return this.level>=le.DEBUG}shouldSnapshot(e){return this.level>=le.DEBUG&&!!this.irSnapshot[e]}}});function bf(n){if(typeof n.numOps=="function")return n.numOps();if(typeof n[Symbol.iterator]=="function"){let e=0;for(let t of n)typeof t.numOps=="function"&&(e+=t.numOps());return e}return-1}var Ul,iu,by=M(()=>{"use strict";tM();fn();nn();nn();l(bf,"countOps");Ul=class{static{l(this,"FixedPointGroup")}constructor(e,t,r=8){this.name=e,this.passes=t,this.maxIterations=r}},iu=class{static{l(this,"PassManager")}constructor(){this.passes=[],this.analysisManager=new hy,this.trace=null,this.verifyHook=null,this.instruments=[]}addPass(e){this.passes.push(e)}addInstrument(e){this.instruments.push(e)}_notify(e,t,r,s){for(let o of this.instruments)typeof o[e]=="function"&&o[e](t,r,s)}setTrace(e){this.trace=e}setVerifyHook(e){this.verifyHook=e}_verifyAfter(e,t,r){if(!this.verifyHook)return null;let s=this.verifyHook(t,r);if(!s||s.length===0)return null;let o=r?t.name||"<module>":t.name;return new Yn("verification",o,`pass '${e.name}' produced invalid IR: ${s.join("; ")}`,e.name)}_applyPass(e,t,r,s){this.trace&&(e.trace=this.trace),this.instruments.length&&this._notify("runBeforePass",e,t,null);let o=r.verbose,i=r.resilient,a=!1,u=!1;if(e instanceof yy){let p=o?bf(t):-1,c=o?performance.now():0,f;try{f=e.run(t,this.analysisManager)}catch(d){if(!i)throw d;return this.analysisManager.invalidateAll(),s.push(oe.FAILED),r.errors.push(new Yn("graphPasses",t.name||"<module>",d.message,e.name)),{changed:a,fatal:!1}}if(s.push(f),o&&this.trace.passRun(e.name,f,performance.now()-c,p,bf(t)),f===oe.CHANGED){a=!0,r.anyChanged=!0,this.analysisManager.invalidateFunctions(t,e.preservedAnalyses);let d=this._verifyAfter(e,t,!0);d&&(r.errors.push(d),i||(u=!0))}else f===oe.FAILED&&(this.analysisManager.invalidateAll(),r.errors.push(new Yn("graphPasses",t.name||"<module>",`pass '${e.name}' failed`,e.name)),i||(u=!0))}else if(e instanceof qe){let p=!1;for(let c of t){if(r.failedFunctions.has(c.name))continue;let f=o?bf(c):-1,d=o?performance.now():0;if(i)try{for(let m of e.requiredAnalyses)this.analysisManager.getAnalysis(m,c);let h=e.run(c,this.analysisManager);if(o&&this.trace.passRun(e.name,h,performance.now()-d,f,bf(c)),h===oe.CHANGED){p=!0,r.anyChanged=!0,c.bumpVersion(),this.analysisManager.invalidate(c,e.preservedAnalyses);let m=this._verifyAfter(e,c,!1);m&&(r.errors.push(m),r.failedFunctions.add(c.name))}else h===oe.FAILED&&(this.analysisManager.invalidate(c),r.errors.push(new Yn("graphPasses",c.name,`pass '${e.name}' failed`,e.name)),r.failedFunctions.add(c.name))}catch(h){r.errors.push(new Yn("graphPasses",c.name,h.message,e.name)),r.failedFunctions.add(c.name)}else{for(let m of e.requiredAnalyses)this.analysisManager.getAnalysis(m,c);let h=e.run(c,this.analysisManager);if(o&&this.trace.passRun(e.name,h,performance.now()-d,f,bf(c)),h===oe.CHANGED){p=!0,r.anyChanged=!0,c.bumpVersion(),this.analysisManager.invalidate(c,e.preservedAnalyses);let m=this._verifyAfter(e,c,!1);if(m){r.errors.push(m),r.failedFunctions.add(c.name),u=!0;break}}else if(h===oe.FAILED){this.analysisManager.invalidate(c),r.errors.push(new Yn("graphPasses",c.name,`pass '${e.name}' failed`,e.name)),r.failedFunctions.add(c.name),u=!0;break}}}s.push(p?oe.CHANGED:oe.UNCHANGED),a=p}return this.instruments.length&&this._notify("runAfterPass",e,t,a?oe.CHANGED:oe.UNCHANGED),e.trace=null,{changed:a,fatal:u}}_runGroup(e,t,r,s){let o=e.maxIterations>0?e.maxIterations:1;for(let i=0;i<o;i++){let a=!1;for(let u of e.passes){if(!r.passContext.shouldRun(u))continue;let{changed:p,fatal:c}=this._applyPass(u,t,r,s);if(c)return!0;p&&(a=!0)}if(!a)return!1}return this.trace&&this.trace.passRun(`${e.name}:max-iter`,oe.UNCHANGED,0,-1,-1),!1}run(e,t={}){let r={verbose:this.trace&&this.trace.level>=le.VERBOSE,resilient:t.errorMode==="resilient",errors:[],failedFunctions:new Set,anyChanged:!1,passContext:t.passContext||new gy},s=[];for(let o of this.passes){if(!(o instanceof Ul)&&!r.passContext.shouldRun(o))continue;if(o instanceof Ul?this._runGroup(o,e,r,s):this._applyPass(o,e,r,s).fatal)return{changed:r.anyChanged,results:s,errors:r.errors,failedFunctions:r.failedFunctions.size>0?r.failedFunctions:null}}return{changed:r.anyChanged,results:s,errors:r.errors.length>0?r.errors:null,failedFunctions:r.failedFunctions.size>0?r.failedFunctions:null}}}});function dn(n,e){Ax.set(n,e)}function Ox(n,e,t,r){let s=n.neg(e).getResult(0),o=n.exp(s).getResult(0),i=n.broadcast(n.scalarConstant(1,t).getResult(0),r,[]).getResult(0),a=n.add(i,o).getResult(0);return n.div(i,a).getResult(0)}function xn(n,e,t,r){return n.broadcast(n.scalarConstant(e,t).getResult(0),r,[]).getResult(0)}var Ax,ql,Bx=M(()=>{"use strict";fn();_s();Re();nn();Ax=new Map;l(dn,"registerDecomposition");ql=class extends qe{static{l(this,"DecompositionPass")}constructor(e=null){super("DecompositionPass"),this.target=e}_shouldDecompose(e){if(!this.target)return!0;let t=this.target.getAttr?this.target.getAttr("nativeOps"):null;return!(t&&t.has(e.opName))}run(e){let t=[];for(let o of e.opsRecursive())Ax.has(o.opName)&&this._shouldDecompose(o)&&t.push(o);if(t.length===0)return oe.UNCHANGED;let r=new $n(e),s=[];for(let o of t){if(!o.parentBlock)continue;let i=Ax.get(o.opName);r.block=o.parentBlock,r.setInsertionPoint(o),s.push(o.opName),i(o,r)}if(this.trace&&this.trace.level>=le.DEBUG){let o={};for(let i of s)o[i]=(o[i]||0)+1;this.trace.emit({type:"pass_detail",passName:this.name,decomposed:o,totalDecomposed:s.length,level:le.DEBUG})}return oe.CHANGED}};dn("stop_gradient",n=>{n.replaceAllResultsWith([n.getOperand(0)]),n.erase()});dn("all_reduce",(n,e)=>{let t=n.getOperand(0),r=n.getAttr("mesh_axis")??0,s=n.getAttr("reduce_op")||"sum",o=t.type.shape,i=t.type.dtype,a=s==="max"?-1/0:s==="min"?1/0:s==="prod"||s==="and"?1:0,u=e.reduce(t,e.scalarConstant(a,i).getResult(0),[r],s).getResult(0),p=[];for(let f=0;f<o.length;f++)f!==r&&p.push(f);let c=e.broadcast(u,o,p).getResult(0);n.replaceAllResultsWith([c]),n.erase()});dn("all_gather",(n,e)=>{let t=n.getOperand(0),r=t.type.shape,s=n.getAttr("mesh_axis")??0,o=n.getAttr("gather_dim")??1,i=r[s],a=r.filter((m,g)=>g!==s),u=o<s?o:o-1,p=[];for(let m=0;m<i;m++){let g=r.map((b,w)=>w===s?m:0),y=r.map((b,w)=>w===s?m+1:b),_=e.slice(t,g,y).getResult(0);p.push(e.reshape(_,a).getResult(0))}let c=p.length===1?p[0]:e.concat(p,u).getResult(0),f=[...r];f[o]=i*r[o];let d=[];for(let m=0;m<f.length;m++)m!==s&&d.push(m);let h=e.broadcast(c,f,d).getResult(0);n.replaceAllResultsWith([h]),n.erase()});dn("softmax",(n,e)=>{let t=n.getOperand(0),r=n.getAttr("axis"),s=t.type.rank,o=t.type.dtype,i=t.type.shape,a=Xo(s,r),u=e.reduce(t,e.scalarConstant(-1/0,o).getResult(0),[r],"max"),p=e.broadcast(u.getResult(0),i,a),c=e.sub(t,p.getResult(0)),f=e.exp(c.getResult(0)),d=e.reduce(f.getResult(0),e.scalarConstant(0,o).getResult(0),[r],"sum"),h=e.broadcast(d.getResult(0),i,a),m=e.div(f.getResult(0),h.getResult(0));n.replaceAllResultsWith([m.getResult(0)]),n.erase()});dn("log_softmax",(n,e)=>{let t=n.getOperand(0),r=n.getAttr("axis"),s=t.type.rank,o=t.type.dtype,i=t.type.shape,a=Xo(s,r),u=e.reduce(t,e.scalarConstant(-1/0,o).getResult(0),[r],"max"),p=e.broadcast(u.getResult(0),i,a),c=e.sub(t,p.getResult(0)),f=e.exp(c.getResult(0)),d=e.reduce(f.getResult(0),e.scalarConstant(0,o).getResult(0),[r],"sum"),h=e.log(d.getResult(0)),m=e.broadcast(h.getResult(0),i,a),g=e.sub(c.getResult(0),m.getResult(0));n.replaceAllResultsWith([g.getResult(0)]),n.erase()});l(Ox,"emitSigmoid");dn("sigmoid",(n,e)=>{let t=n.getOperand(0),r=Ox(e,t,t.type.dtype,t.type.shape);n.replaceAllResultsWith([r]),n.erase()});dn("gelu",(n,e)=>{let t=n.getOperand(0),r=t.type.dtype,s=t.type.shape,o=e.broadcast(e.scalarConstant(1.702,r).getResult(0),s,[]).getResult(0),i=e.mul(o,t).getResult(0),a=Ox(e,i,r,s),u=e.mul(t,a).getResult(0);n.replaceAllResultsWith([u]),n.erase()});dn("silu",(n,e)=>{let t=n.getOperand(0),r=Ox(e,t,t.type.dtype,t.type.shape),s=e.mul(t,r).getResult(0);n.replaceAllResultsWith([s]),n.erase()});dn("layer_norm",(n,e)=>{let t=n.getOperand(0),r=n.getOperand(1),s=n.getOperand(2),o=n.getAttr("axis"),i=n.getAttr("epsilon"),a=t.type.rank,u=t.type.dtype,p=t.type.shape,c=Xo(a,o),f=e.reduce(t,e.scalarConstant(0,u).getResult(0),[o],"mean"),d=e.broadcast(f.getResult(0),p,c),h=e.sub(t,d.getResult(0)),m=e.mul(h.getResult(0),h.getResult(0)),g=e.reduce(m.getResult(0),e.scalarConstant(0,u).getResult(0),[o],"mean"),y=e.broadcast(e.scalarConstant(i,u).getResult(0),g.getResult(0).type.shape,[]),_=e.add(g.getResult(0),y.getResult(0)),b=e.rsqrt(_.getResult(0)),w=e.broadcast(b.getResult(0),p,c),x=e.mul(h.getResult(0),w.getResult(0)),S=e.broadcast(r,p,[o]),T=e.mul(x.getResult(0),S.getResult(0)),v=e.broadcast(s,p,[o]),A=e.add(T.getResult(0),v.getResult(0));n.replaceAllResultsWith([A.getResult(0)]),n.erase()});dn("batch_norm",(n,e)=>{let t=n.getOperand(0),r=n.getOperand(1),s=n.getOperand(2),o=n.getOperand(3),i=n.getOperand(4),a=n.getAttr("axis"),u=n.getAttr("epsilon"),p=t.type.rank,c=t.type.dtype,f=t.type.shape,d=e.broadcast(e.scalarConstant(u,c).getResult(0),i.type.shape,[]),h=e.add(i,d.getResult(0)),m=e.rsqrt(h.getResult(0)),g=e.broadcast(o,f,[a]),y=e.sub(t,g.getResult(0)),_=e.broadcast(m.getResult(0),f,[a]),b=e.mul(y.getResult(0),_.getResult(0)),w=e.broadcast(r,f,[a]),x=e.mul(b.getResult(0),w.getResult(0)),S=e.broadcast(s,f,[a]),T=e.add(x.getResult(0),S.getResult(0));n.replaceAllResultsWith([T.getResult(0)]),n.erase()});dn("where",(n,e)=>{let t=n.getOperand(0);if(t.type.dtype!=="bool"){let s=e.broadcast(e.scalarConstant(0,t.type.dtype).getResult(0),t.type.shape,[]);t=e.compare(t,s.getResult(0),"ne").getResult(0)}let r=e.select(t,n.getOperand(1),n.getOperand(2));n.replaceAllResultsWith([r.getResult(0)]),n.erase()});dn("split",(n,e)=>{let t=n.getOperand(0),r=n.getAttr("dimension"),s=n.getAttr("split_sizes"),o=t.type.shape,i=[],a=0;for(let u of s){let p=o.map((f,d)=>d===r?a:0),c=o.map((f,d)=>d===r?a+u:f);i.push(e.slice(t,p,c).getResult(0)),a+=u}n.replaceAllResultsWith(i),n.erase()});dn("one_hot",(n,e)=>{let t=n.getOperand(0),r=n.getAttr("depth"),s=n.getAttr("axis")??-1,o=n.getAttr("on_value")??1,i=n.getAttr("off_value")??0,a=n.getResult(0).type,u=a.dtype,p=a.shape,c=s<0?t.type.rank+1+s:s,f=new I(p,"i32"),d=e._inferAndBuild("iota",[],{iota_dimension:c,tensor_type:f}),h=t.type.rank,m=[];for(let S=0;S<h;S++)m.push(S<c?S:S+1);let g=e.broadcast(t,p,m),y=e.convert(g.getResult(0),"i32"),_=e.compare(y.getResult(0),d.getResult(0),"eq"),b=e.broadcast(e.scalarConstant(o,u).getResult(0),p,[]),w=e.broadcast(e.scalarConstant(i,u).getResult(0),p,[]),x=e.select(_.getResult(0),b.getResult(0),w.getResult(0));n.replaceAllResultsWith([x.getResult(0)]),n.erase()});l(xn,"bcast");dn("elu",(n,e)=>{let t=n.getOperand(0),r=n.getAttr("alpha")??1,s=t.type.dtype,o=t.type.shape,i=xn(e,0,s,o),a=e.compare(t,i,"gt").getResult(0),u=xn(e,1,s,o),p=e.exp(t).getResult(0),c=e.sub(p,u).getResult(0),f=xn(e,r,s,o),d=e.mul(f,c).getResult(0),h=e.select(a,t,d);n.replaceAllResultsWith([h.getResult(0)]),n.erase()});dn("leaky_relu",(n,e)=>{let t=n.getOperand(0),r=n.getAttr("negative_slope")??.01,s=t.type.dtype,o=t.type.shape,i=xn(e,0,s,o),a=e.compare(t,i,"gt").getResult(0),u=xn(e,r,s,o),p=e.mul(u,t).getResult(0),c=e.select(a,t,p);n.replaceAllResultsWith([c.getResult(0)]),n.erase()});dn("celu",(n,e)=>{let t=n.getOperand(0),r=n.getAttr("alpha")??1,s=t.type.dtype,o=t.type.shape,i=xn(e,0,s,o),a=e.maximum(t,i).getResult(0),u=xn(e,r,s,o),p=e.div(t,u).getResult(0),c=e.exp(p).getResult(0),f=xn(e,1,s,o),d=e.sub(c,f).getResult(0),h=e.mul(u,d).getResult(0),m=e.minimum(i,h).getResult(0),g=e.add(a,m);n.replaceAllResultsWith([g.getResult(0)]),n.erase()});dn("selu",(n,e)=>{let t=n.getOperand(0),r=t.type.dtype,s=t.type.shape,o=1.0507009873554805,i=1.6732632423543772,a=xn(e,0,r,s),u=e.compare(t,a,"gt").getResult(0),p=xn(e,1,r,s),c=e.exp(t).getResult(0),f=e.sub(c,p).getResult(0),d=xn(e,i,r,s),h=e.mul(d,f).getResult(0),m=e.select(u,t,h).getResult(0),g=xn(e,o,r,s),y=e.mul(g,m);n.replaceAllResultsWith([y.getResult(0)]),n.erase()});dn("mish",(n,e)=>{let t=n.getOperand(0),r=t.type.dtype,s=t.type.shape,o=xn(e,1,r,s),i=e.exp(t).getResult(0),a=e.add(o,i).getResult(0),u=e.log(a).getResult(0),p=e.tanh(u).getResult(0),c=e.mul(t,p);n.replaceAllResultsWith([c.getResult(0)]),n.erase()});dn("hardswish",(n,e)=>{let t=n.getOperand(0),r=t.type.dtype,s=t.type.shape,o=xn(e,3,r,s),i=xn(e,0,r,s),a=xn(e,6,r,s),u=e.add(t,o).getResult(0),p=e.minimum(e.maximum(u,i).getResult(0),a).getResult(0),c=e.div(p,a).getResult(0),f=e.mul(t,c);n.replaceAllResultsWith([f.getResult(0)]),n.erase()});dn("hardsigmoid",(n,e)=>{let t=n.getOperand(0),r=t.type.dtype,s=t.type.shape,o=xn(e,6,r,s),i=xn(e,.5,r,s),a=xn(e,0,r,s),u=xn(e,1,r,s),p=e.div(t,o).getResult(0),c=e.add(p,i).getResult(0),f=e.minimum(e.maximum(c,a).getResult(0),u);n.replaceAllResultsWith([f.getResult(0)]),n.erase()});dn("embedding",(n,e)=>{let t=n.getOperand(0),r=n.getOperand(1),s=t.type.shape[t.type.rank-1],o=r.type.rank,i=e._inferAndBuild("gather",[t,r],{offset_dims:Array.from({length:1},(a,u)=>o+u),collapsed_slice_dims:[0],start_index_map:[0],slice_sizes:[1,s],index_vector_dim:o});n.replaceAllResultsWith([i.getResult(0)]),n.erase()})});var Wl,Rx=M(()=>{"use strict";fn();_s();nn();Wl=class{static{l(this,"PatternApplicator")}constructor(e){this.patternSet=e}applyPatterns(e,t=10,r=null){let s=new $n(e),o=0,i=[...e.opsRecursive()],a=0,u=new Set(i),p=l(h=>{!h||!h.parentBlock||u.has(h)||(u.add(h),i.push(h))},"enqueue"),c=Math.max(t,1)*Math.max(i.length,1)*4+1e3,f=0,d=!1;for(;a<i.length;){if(++f>c){d=!0;break}let h=i[a++];if(u.delete(h),!h.parentBlock)continue;let m=this.patternSet.getForOp(h.opName);for(let g of m){if(!g.match(h))continue;let y=h.parentBlock,_=h._prev,b=h._next,w=[];for(let T=0;T<h.numResults;T++)for(let v of h.getResult(T).getUsers())w.push(v);for(let T=0;T<h.numOperands;T++){let v=h.getOperand(T).definingOp;v&&w.push(v)}if(s.block=y,s.setInsertionPoint(h),!g.rewrite(h,s))continue;o++;for(let T of w)p(T);let x=_?_._next:y._head,S=y._size+2;for(;x&&x!==b&&S-- >0;)p(x),x=x._next;p(h);break}}return r&&(d&&r.level>=le.INFO&&r.emit({type:"pass_detail",passName:"PatternApplicator",message:`pattern rewriting hit safety budget (${c}) without converging`,totalRewrites:o,level:le.INFO}),r.level>=le.DEBUG&&o>0&&r.emit({type:"pass_detail",passName:"PatternApplicator",totalRewrites:o,patternCount:this.patternSet.patterns.length,level:le.DEBUG})),o>0?oe.CHANGED:oe.UNCHANGED}}});function w3(){if(wf)return wf;wf=new Wu;for(let n of ne.allOps())if(n.getCanonicalizationPatterns){let e=n.getCanonicalizationPatterns();if(e)for(let t of e)wf.add(t)}return wf}var wf,au,Mx=M(()=>{"use strict";fn();cp();Rx();Ct();wf=null;l(w3,"getCanonicalizationPatterns");au=class extends qe{static{l(this,"CanonicalizePass")}constructor(){super("canonicalize")}run(e,t){let r=w3();return new Wl(r).applyPatterns(e,10,this.trace)}}});var wy,Ix,Kl,nM=M(()=>{"use strict";Ct();Bn();wy=class{static{l(this,"MemoryEffect")}constructor(e,t){this.kind=e,this.value=t}},Ix=class{static{l(this,"MemoryEffectResult")}constructor(e,t){this.opEffects=e,this._valueEffects=t}hasSideEffect(e){let t=this.opEffects.get(e);return t&&t.length>0}getEffects(e){return this.opEffects.get(e)||[]}getEffectsOn(e){return this._valueEffects.get(e)||[]}getReadersOf(e){let t=this._valueEffects.get(e);if(!t)return[];let r=[];for(let s=0;s<t.length;s++)t[s].effect.kind===gn.READ&&r.push(t[s].op);return r}getWritersOf(e){let t=this._valueEffects.get(e);if(!t)return[];let r=[];for(let s=0;s<t.length;s++)t[s].effect.kind===gn.WRITE&&r.push(t[s].op);return r}},Kl=class{static get name(){return"memory_effect"}static get depKey(){return"memoryEffect"}static get dependencies(){return[]}static compute(e){let t=new Map,r=new Map,s=l((i,a,u)=>{let p=r.get(i);p||(p=[],r.set(i,p)),p.push({op:a,effect:u})},"addValueEffect"),o=typeof e.opsRecursive=="function"?e.opsRecursive():e.ops();for(let i of o){let a=ne.get(i.opName),u=[],p=gn.NONE;if(a&&a.sideEffects?p=a.sideEffects:i.hasSideEffects&&i.hasSideEffects()&&(p=gn.WRITE),p!==gn.NONE)if(a&&a.getMemoryEffects){let c=a.getMemoryEffects(i);for(let f of c)u.push(f),s(f.value,i,f)}else{if(p&gn.READ)for(let c=0;c<i.numOperands;c++){let f=new wy(gn.READ,i.getOperand(c));u.push(f),s(i.getOperand(c),i,f)}if(p&gn.WRITE)for(let c=0;c<i.numResults;c++){let f=new wy(gn.WRITE,i.getResult(c));u.push(f),s(i.getResult(c),i,f)}}t.set(i,u)}return new Ix(t,r)}}});var Co,Ex=M(()=>{"use strict";fn();nM();Ir();nn();Co=class extends qe{static{l(this,"DCEPass")}constructor(){super("dce"),this.preservedAnalyses=new Set,this.requiredAnalyses=[Kl]}run(e,t){let r=!1,s=t?t.getAnalysis(Kl,e):Kl.compute(e),o=[];for(let a of e.opsRecursive())this._isDead(a,s)&&o.push(a);let i=0;for(;o.length>0;){let a=o.pop();if(!a.parentBlock||!this._isDead(a,s))continue;let u=[];for(let p=0;p<a.numOperands;p++){let c=a.getOperand(p).definingOp;c&&c.parentBlock&&u.push(c)}a.erase(),r=!0,i++;for(let p of u)p.parentBlock&&this._isDead(p,s)&&o.push(p)}return this.trace&&this.trace.level>=le.DEBUG&&i>0&&this.trace.emit({type:"pass_detail",passName:this.name,erasedCount:i,level:le.DEBUG}),r?oe.CHANGED:oe.UNCHANGED}_isDead(e,t){if(Xt(e.opName)||e.regions&&e.regions.length>0)return!1;for(let r=0;r<e.numResults;r++)if(e.getResult(r).hasUses)return!1;return!t.hasSideEffect(e)}}});var Ty,xy,rM=M(()=>{"use strict";Ty=class n{static{l(this,"ScheduleStep")}constructor(e,t){this.primitive=e,this.args=t}serialize(){return{primitive:this.primitive,args:this.args}}static deserialize(e){return new n(e.primitive,e.args)}},xy=class n{static{l(this,"ScheduleTrace")}constructor(){this.steps=[]}record(e,t){let r=new Ty(e,t);return this.steps.push(r),r}serialize(){return this.steps.map(e=>e.serialize())}static deserialize(e){let t=new n;for(let r of e)t.steps.push(Ty.deserialize(r));return t}replay(e){for(let t of this.steps){let r=e[t.primitive];if(typeof r!="function")throw new Error(`Unknown schedule primitive: ${t.primitive}`);e._replaying=!0;try{r.call(e,...t.args)}finally{e._replaying=!1}}}get length(){return this.steps.length}clear(){this.steps.length=0}}});function uu(n){return Number.isFinite(n)}function sM(n){return Number.isNaN(n)?iM:n}function oM(n){return Number.isNaN(n)?aM:n}var iM,aM,ps,Hl,Xs,Yl=M(()=>{"use strict";po();iM=-1/0,aM=1/0;l(uu,"isFinite2");l(sM,"boundLo");l(oM,"boundHi");ps=class{static{l(this,"IntBound")}constructor(e,t){this.min=e,this.max=t}isConst(){return this.min===this.max&&uu(this.min)}},Hl=new ps(iM,aM),Xs=class{static{l(this,"Analyzer")}constructor(){this._varBounds=new Map}bind(e,t,r){return this._varBounds.set(e,new ps(t,r)),this}getVarBound(e){return this._varBounds.get(e)||null}setVarBound(e,t){return t?this._varBounds.set(e,t):this._varBounds.delete(e),this}bindShape(e){for(let[t,r]of e)typeof r=="number"&&r>0&&this.bind(t,0,r-1);return this}constIntBound(e){if(typeof e=="number")return new ps(e,e);if(!(e instanceof Pe))return Hl;if(e.type==="var")return this._varBounds.get(e.name)||Hl;let t=e.args.length>0?this.constIntBound(e.args[0]):null,r=e.args.length>1?this.constIntBound(e.args[1]):null;switch(e.type){case"add":return new ps(sM(t.min+r.min),oM(t.max+r.max));case"sub":return new ps(sM(t.min-r.max),oM(t.max-r.min));case"neg":return new ps(-t.max,-t.min);case"mul":return this._mulBound(t,r);case"max":return new ps(Math.max(t.min,r.min),Math.max(t.max,r.max));case"min":return new ps(Math.min(t.min,r.min),Math.min(t.max,r.max));case"div":return this._divBound(t,r,Math.floor);case"ceildiv":return this._divBound(t,r,Math.ceil);case"mod":return this._modBound(t,r);default:return Hl}}_mulBound(e,t){if(!uu(e.min)||!uu(e.max)||!uu(t.min)||!uu(t.max))return Hl;let r=[e.min*t.min,e.min*t.max,e.max*t.min,e.max*t.max];return new ps(Math.min(...r),Math.max(...r))}_divBound(e,t,r){if(t.isConst()&&t.min>0){let s=uu(e.min)?r(e.min/t.min):e.min,o=uu(e.max)?r(e.max/t.min):e.max;return new ps(s,o)}return Hl}_modBound(e,t){return t.isConst()&&t.min>0?new ps(0,t.min-1):Hl}canProveGreaterEqual(e,t){return this.constIntBound(Pe.sub(e,t)).min>=0}canProveLess(e,t){return this.constIntBound(Pe.sub(e,t)).max<0}canProveNonNegative(e){return this.canProveGreaterEqual(e,0)}canProveEqual(e,t){if(Pe.equals(e,t))return!0;let r=this.constIntBound(Pe.sub(e,t));return r.min===0&&r.max===0}}});function oa(n){if(n==null)return null;if(typeof n=="number")return n;switch(n.type){case"IntImmNode":return n.value;case"VariableNode":return Pe.var(n.name);case"MathOpNode":{if(n.b===null||n.b===void 0){if(n.op==="-"){let s=oa(n.a);return s===null?null:Pe.neg(s)}return null}let e=T3[n.op];if(!e)return null;let t=oa(n.a);if(t===null)return null;let r=oa(n.b);return r===null||(n.op==="//"||n.op==="%")&&(typeof r!="number"||r<=0)?null:Pe[e](t,r)}case"CallExternNode":{if(n.args.length!==2||n.externName!=="max"&&n.externName!=="min")return null;let e=oa(n.args[0]);if(e===null)return null;let t=oa(n.args[1]);return t===null?null:n.externName==="max"?Pe.max(e,t):Pe.min(e,t)}default:return null}}function cu(n,e){let t=oa(e);return t===null?null:n.constIntBound(t)}function x3(n,e,t){let r=oa(e);if(r===null)return null;let s=oa(t);return s===null?null:n.constIntBound(Pe.sub(r,s))}function Sy(n,e,t,r,s){let o=s[e];if(!o)return!1;let i=x3(n,t,r);return i===null?!1:o(i)}function _i(n,e){if(e==null)return!1;if(e.type==="IntImmNode")return e.value!==0;if(e.type==="CompareNode")return Sy(n,e.direction,e.a,e.b,uM);if(e.type==="MathOpNode"){if(e.op==="*"&&e.b)return _i(n,e.a)&&_i(n,e.b);if(Lx.has(e.op))return Sy(n,fM[e.op],e.a,e.b,uM)}return!1}function lu(n,e){if(e==null)return!1;if(e.type==="IntImmNode")return e.value===0;if(e.type==="CompareNode")return Sy(n,e.direction,e.a,e.b,lM);if(e.type==="MathOpNode"){if(e.op==="*"&&e.b)return lu(n,e.a)||lu(n,e.b);if(Lx.has(e.op))return Sy(n,fM[e.op],e.a,e.b,lM)}return!1}function cM(n){return!n||n.type!=="MathOpNode"||n.op!=="*"?null:n.b&&n.b.type==="IntImmNode"&&n.b.value>0?{factor:n.a,c:n.b.value}:n.a&&n.a.type==="IntImmNode"&&n.a.value>0?{factor:n.b,c:n.a.value}:null}function pM(n,e,t,r){let s=cu(n,e);return s===null?!1:s.min>=t&&s.max<=r}var T3,Lx,fM,uM,lM,vy,Tf=M(()=>{"use strict";Yl();po();Fe();T3={"+":"add","-":"sub","*":"mul","//":"div","%":"mod"},Lx=new Set(["<","<=",">",">=","==","!="]),fM={"<":"lt","<=":"le",">":"gt",">=":"ge","==":"eq","!=":"ne"};l(oa,"irToSymInt");l(cu,"irBound");l(x3,"diffBound");uM={lt:l(n=>n.max<0,"lt"),le:l(n=>n.max<=0,"le"),gt:l(n=>n.min>0,"gt"),ge:l(n=>n.min>=0,"ge"),eq:l(n=>n.min===0&&n.max===0,"eq"),ne:l(n=>n.min>0||n.max<0,"ne")},lM={lt:l(n=>n.min>=0,"lt"),le:l(n=>n.min>0,"le"),gt:l(n=>n.max<=0,"gt"),ge:l(n=>n.max<0,"ge"),eq:l(n=>n.min>0||n.max<0,"eq"),ne:l(n=>n.min===0&&n.max===0,"ne")};l(Sy,"proveCompare");l(_i,"proveTrue");l(lu,"proveFalse");l(cM,"asScaledVar");l(pM,"boundWithin");vy=class{static{l(this,"RewriteSimplify")}constructor(e=new Xs){this.analyzer=e}simplify(e){if(e==null||typeof e!="object")return e;switch(e.type){case"IntImmNode":case"VariableNode":return e;case"MathOpNode":return this._simplifyMathOp(e);case"CompareNode":return this._simplifyCompare(e);default:return e}}_simplifyCompare(e){let t=this.simplify(e.a),r=this.simplify(e.b);return _i(this.analyzer,new Ke(e.direction,t,r))?new L(1):lu(this.analyzer,new Ke(e.direction,t,r))?new L(0):t===e.a&&r===e.b?e:new Ke(e.direction,t,r)}_simplifyMathOp(e){if(e.b===null||e.b===void 0){let o=this.simplify(e.a);return o===e.a?e:new P(e.op,o)}let t=this.simplify(e.a),r=this.simplify(e.b),s=Un(e.op,t,r);if(!s||s.type!=="MathOpNode")return s;if(s.op==="//"&&s.b.type==="IntImmNode"&&s.b.value>0){let o=s.b.value,i=cM(s.a);if(i&&i.c===o)return i.factor;if(pM(this.analyzer,s.a,0,o-1))return new L(0)}if(s.op==="%"&&s.b.type==="IntImmNode"&&s.b.value>0){let o=s.b.value,i=cM(s.a);if(i&&i.c===o)return new L(0);if(pM(this.analyzer,s.a,0,o-1))return s.a}if(Lx.has(s.op)){if(_i(this.analyzer,s))return new L(1);if(lu(this.analyzer,s))return new L(0)}return s}}});function dM(n,e,t){if(typeof t!="number"||t<0)return"unknown";let r=cu(n,e);return r===null?"unknown":r.min>=0&&r.max<=t-1?"in":r.min>t-1||r.max<0?"oob":"unknown"}function Nt(n,e){if(!(!n||typeof n!="object"))switch(n.type){case"VariableNode":n.name&&e.add(n.name);return;case"BufferLoadNode":if(n.indices)for(let t of n.indices)Nt(t,e);return;case"BufferStoreNode":if(n.indices)for(let t of n.indices)Nt(t,e);Nt(n.value,e);return;case"MathOpNode":Nt(n.a,e),n.b&&Nt(n.b,e);return;case"CompareNode":Nt(n.a,e),Nt(n.b,e);return;case"CastNode":Nt(n.expr,e);return;case"CallExternNode":for(let t of n.args)Nt(t,e);return;case"IfThenElseNode":Nt(n.condition,e),Nt(n.thenBody,e),n.elseBody&&Nt(n.elseBody,e);return;case"SeqNode":for(let t of n.stmts)Nt(t,e);return;case"LetStmtNode":Nt(n.value,e),Nt(n.body,e);return;default:return}}function ky(n,e){let t=new Set((n.writes||[]).map(s=>s.buffer&&s.buffer.name)),r=[n.body,n.initBody];for(;r.length>0;){let s=r.pop();if(!(!s||typeof s!="object")){if(s.type==="BufferStoreNode"&&s.buffer&&t.has(s.buffer.name)&&s.indices)for(let o of s.indices)Nt(o,e);if(s.type==="BufferStoreNode"){s.value&&r.push(s.value);continue}if(s.type==="SeqNode"){for(let o of s.stmts)r.push(o);continue}if(s.type==="IfThenElseNode"){r.push(s.thenBody),s.elseBody&&r.push(s.elseBody);continue}if(s.type==="LetStmtNode"){r.push(s.body);continue}}}}function S3(n,e){let t=new Set;if(Nt(e.body,t),e.initBody&&Nt(e.initBody,t),!t.has(n))return!1;let r=new Set;return ky(e,r),!r.has(n)}function pu(n,e){if(!(!n||typeof n!="object"))switch(n.type){case"BlockNode":e.push(n),pu(n.body,e),n.initBody&&pu(n.initBody,e);return;case"ForNode":case"AllocateNode":case"LetStmtNode":pu(n.body,e);return;case"SeqNode":for(let t of n.stmts)pu(t,e);return;case"IfThenElseNode":pu(n.thenBody,e),n.elseBody&&pu(n.elseBody,e);return;default:return}}function Cx(n){let e=n.loopVar.name,t=[];pu(n.body,t);for(let r of t)if(S3(e,r))return r.name;return null}var Dx=M(()=>{"use strict";Yl();po();Tf();l(dM,"classifyBufferIndex");l(Nt,"collectVarsUsed");l(ky,"collectWriteIndexVars");l(S3,"loopVarIsReductionOf");l(pu,"collectBlocksUnder");l(Cx,"loopCarriesReduction")});var Jl,Px=M(()=>{"use strict";Fe();Dx();Yl();Tf();Jl=class n{static{l(this,"ScheduleValidator")}static validate(e){let t=[],r={boundVars:new Set,threadBindings:new Map,parallelExtents:new Map,parLoops:[],innermostLoopVar:null,loopStack:[],analyzer:new Xs,condDepth:0,errors:t};for(let s of e.params)r.boundVars.add(s.name);return n._visitNode(e.body,r),n._checkPartitionConsistency(r),t}static _checkPartitionConsistency(e){if(e.parallelExtents.size<=1)return;let t=[...e.parallelExtents.entries()].map(([r,s])=>`'${s}'(extent ${r})`);e.errors.push(`Ambiguous parallel partition: ${e.parallelExtents.size} distinct parallel extents [${t.join(", ")}] \u2014 runtime partitions a single axis, mismatched-extent parallel loops corrupt buffers`)}static _visitNode(e,t){if(e)switch(e.type){case"ForNode":n._visitFor(e,t);break;case"BlockNode":n._visitBlock(e,t);break;case"SeqNode":for(let r of e.stmts)n._visitNode(r,t);break;case"IfThenElseNode":n._visitExpr(e.condition,t),t.condDepth++,n._visitNode(e.thenBody,t),e.elseBody&&n._visitNode(e.elseBody,t),t.condDepth--;break;case"AllocateNode":n._visitNode(e.body,t);break;case"LetStmtNode":t.boundVars.add(e.variable.name),n._visitNode(e.body,t),t.boundVars.delete(e.variable.name);break;case"BufferStoreNode":n._validateBufferAccess(e,t);break;case"EvaluateNode":break;default:break}}static _visitFor(e,t){let r=e.loopVar.name;if(t.boundVars.has(r)&&t.errors.push(`Duplicate loop variable: ${r}`),e.kind===z.THREAD_BINDING&&(e.threadTag?(t.threadBindings.has(e.threadTag)&&t.errors.push(`Duplicate thread binding '${e.threadTag}': already bound to '${t.threadBindings.get(e.threadTag)}'`),t.threadBindings.set(e.threadTag,r)):t.errors.push(`Thread-bound loop '${r}' missing threadTag`)),e.kind===z.VECTORIZED){let a=e.extent;a.type==="IntImmNode"&&a.value<=0&&t.errors.push(`Vectorized loop '${r}' has non-positive extent ${a.value}`)}if(e.kind===z.PARALLEL){let a=e.extent;a&&a.type==="IntImmNode"&&!t.parallelExtents.has(a.value)&&t.parallelExtents.set(a.value,r)}(e.kind===z.PARALLEL||e.kind===z.THREAD_BINDING)&&n._checkNoNestedParallel(e.body,r,t);let s=e.kind===z.PARALLEL||e.kind===z.VECTORIZED;s&&t.parLoops.push({varName:r,kind:e.kind}),t.boundVars.add(r);let o=t.analyzer.getVarBound(r);e.extent&&e.extent.type==="IntImmNode"&&e.extent.value>0?t.analyzer.bind(r,0,e.extent.value-1):t.analyzer.setVarBound(r,null);let i=t.innermostLoopVar;t.innermostLoopVar=r,t.loopStack.push(r),n._visitNode(e.body,t),t.loopStack.pop(),t.innermostLoopVar=i,t.boundVars.delete(r),t.analyzer.setVarBound(r,o),s&&t.parLoops.pop(),e.kind===z.THREAD_BINDING&&e.threadTag&&t.threadBindings.delete(e.threadTag)}static _checkNoNestedParallel(e,t,r){e&&e.type==="ForNode"&&e.kind===z.PARALLEL&&r.errors.push(`Parallel loop '${e.loopVar.name}' nested inside parallel/thread-bound loop '${t}'`)}static _visitBlock(e,t){let r=[];for(let s of e.iterVars)s.iterVar&&(t.boundVars.add(s.iterVar.name),r.push([s.iterVar.name,t.analyzer.getVarBound(s.iterVar.name)]),t.analyzer.setVarBound(s.iterVar.name,s.binding?cu(t.analyzer,s.binding):null));if(e.initBody&&t.innermostLoopVar){let s=new Set;ky(e,s);let o=new Set;Nt(e.body,o),Nt(e.initBody,o);let i=0;for(let a of t.loopStack)o.has(a)&&!s.has(a)&&i++;(s.has(t.innermostLoopVar)||i>1)&&t.errors.push(`Reduction block '${e.name}' violates the init contract: codegen zeroes the accumulator at the single innermost loop, so an init-bearing block needs exactly one enclosing reduction loop and it must be innermost (innermost '${t.innermostLoopVar}', ${i} enclosing reduction loops)`)}if(t.parLoops.length>0){let s=new Set;Nt(e.body,s),e.initBody&&Nt(e.initBody,s);let o=new Set;ky(e,o);for(let i of t.parLoops)if(s.has(i.varName)&&!o.has(i.varName)){let a=i.kind===z.VECTORIZED?"Vectorized":"Parallel";t.errors.push(`${a} loop '${i.varName}' carries a reduction in block '${e.name}': the loop variable is read but never written, so parallel iterations race on the accumulator`)}}n._visitNode(e.body,t),e.initBody&&n._visitNode(e.initBody,t);for(let s of e.iterVars)s.iterVar&&t.boundVars.delete(s.iterVar.name);for(let[s,o]of r)t.analyzer.setVarBound(s,o)}static _validateBufferAccess(e,t){if(e.buffer?(n._checkRank(e,t),n._checkBounds(e,t)):t.errors.push("BufferStore with null buffer"),e.indices)for(let r of e.indices)n._visitExpr(r,t);e.value&&n._visitExpr(e.value,t)}static _checkRank(e,t){e.indices&&e.buffer.shape&&e.indices.length!==e.buffer.shape.length&&t.errors.push(`Buffer '${e.buffer.name}' rank mismatch: ${e.indices.length} indices for rank-${e.buffer.shape.length} buffer`)}static _checkBounds(e,t){if(!(t.condDepth>0)&&!(!e.buffer||!e.buffer.shape||!e.indices)&&e.indices.length===e.buffer.shape.length)for(let r=0;r<e.indices.length;r++){let s=e.buffer.shape[r];dM(t.analyzer,e.indices[r],s)==="oob"&&t.errors.push(`Buffer '${e.buffer.name}' access is out of bounds on axis ${r}: index is provably outside [0, ${s-1}]`)}}static _visitExpr(e,t){if(!(!e||typeof e!="object"))switch(e.type){case"BufferLoadNode":if(e.buffer?(n._checkRank(e,t),n._checkBounds(e,t)):t.errors.push("BufferLoad with null buffer"),e.indices)for(let r of e.indices)n._visitExpr(r,t);break;case"BufferStoreNode":n._validateBufferAccess(e,t);break;case"MathOpNode":case"CompareNode":n._visitExpr(e.a,t),n._visitExpr(e.b,t);break;case"CastNode":n._visitExpr(e.expr,t);break;case"CallExternNode":for(let r of e.args)n._visitExpr(r,t);break;case"IfThenElseNode":n._visitExpr(e.condition,t),t.condDepth++,n._visitNode(e.thenBody,t),e.elseBody&&n._visitNode(e.elseBody,t),t.condDepth--;break}}}});var $x,Zl,Fx=M(()=>{"use strict";Fe();$x=class{static{l(this,"SRef")}constructor(e,t=null){this.node=e,this.parent=t,this.children=[]}get type(){return this.node.type}get isLoop(){return this.node.type==="ForNode"}get isBlock(){return this.node.type==="BlockNode"}get isRoot(){return this.parent===null}ancestors(){let e=[],t=this.parent;for(;t;)e.push(t),t=t.parent;return e}loopAncestors(){let e=[],t=this.parent;for(;t;)t.isLoop&&e.push(t),t=t.parent;return e}childBlocks(){let e=[],t=[...this.children];for(;t.length>0;){let r=t.pop();if(r.isBlock)e.push(r);else for(let s of r.children)t.push(s)}return e}childLoops(){return this.children.filter(e=>e.isLoop)}},Zl=class{static{l(this,"SRefTree")}constructor(e){this._nodeToSRef=new Map,this._blockNameToSRef=new Map,this._loopSRefs=new Set,this._blockSRefs=new Set,this.root=this._build(e.body,null)}_register(e){this._nodeToSRef.set(e.node,e),e.isLoop?this._loopSRefs.add(e):e.isBlock&&(this._blockSRefs.add(e),this._blockNameToSRef.set(e.node.name,e))}_unregisterSubtree(e){let t=[e];for(;t.length>0;){let r=t.pop();this._nodeToSRef.delete(r.node),r.isLoop?this._loopSRefs.delete(r):r.isBlock&&(this._blockSRefs.delete(r),this._blockNameToSRef.get(r.node.name)===r&&this._blockNameToSRef.delete(r.node.name));for(let s of r.children)t.push(s)}}_buildSubtree(e,t){let r=[],s=[{node:e,parentSRef:t,isTop:!0}];for(;s.length>0;){let{node:o,parentSRef:i,isTop:a}=s.pop();if(o)switch(o.type){case"ForNode":case"BlockNode":{let u=new $x(o,i);this._register(u),a?r.push(u):i.children.push(u),s.push({node:o.body,parentSRef:u,isTop:!1}),o.type==="BlockNode"&&o.initBody&&s.push({node:o.initBody,parentSRef:u,isTop:!1});break}case"SeqNode":for(let u=o.stmts.length-1;u>=0;u--)s.push({node:o.stmts[u],parentSRef:i,isTop:a});break;case"IfThenElseNode":o.elseBody&&s.push({node:o.elseBody,parentSRef:i,isTop:a}),s.push({node:o.thenBody,parentSRef:i,isTop:a});break;case"AllocateNode":case"LetStmtNode":s.push({node:o.body,parentSRef:i,isTop:a});break}}return r}_build(e,t){let r=this._buildSubtree(e,t);if(t)for(let s of r)t.children.push(s);return this._nodeToSRef.get(e)||null}getSRef(e){return this._nodeToSRef.get(e)||null}getBlockSRef(e){return this._blockNameToSRef.get(e)||null}allBlocks(){return[...this._blockSRefs]}allLoops(){return[...this._loopSRefs]}loopsOf(e){let t=this._blockNameToSRef.get(e);return t?t.loopAncestors().filter(r=>r.node.kind!==z.RECURRENCE).reverse():[]}rebuildFrom(e){this._nodeToSRef.clear(),this._blockNameToSRef.clear(),this._loopSRefs.clear(),this._blockSRefs.clear(),this.root=this._build(e,null)}replaceNode(e,t){let r=this._nodeToSRef.get(e);if(!r)return!1;let s=r.parent,o=this.root===r;this._unregisterSubtree(r);let i=this._buildSubtree(t,s);if(s){let a=s.children.indexOf(r);if(a>=0)s.children.splice(a,1,...i);else for(let u of i)s.children.push(u)}else o&&(this.root=this._nodeToSRef.get(t)||null);return!0}removeNode(e){let t=this._nodeToSRef.get(e);if(!t)return!1;let r=t.parent,s=this.root===t;if(this._unregisterSubtree(t),r){let o=r.children.indexOf(t);o>=0&&r.children.splice(o,1)}else s&&(this.root=null);return!0}}});var zx,Vx,Ny,hM=M(()=>{"use strict";Fx();Fe();zx=class{static{l(this,"LoopBinding")}constructor(e,t,r,s,o){this.loopSRef=e,this.iterVar=t,this.extent=r,this.kind=s,this.threadTag=o}},Vx=class{static{l(this,"BlockBinding")}constructor(e,t,r,s){this.blockSRef=e,this.iterVars=t,this.readBuffers=r,this.writeBuffers=s}},Ny=class{static{l(this,"ScheduleState")}constructor(e){this.primFunc=e,this._dirty=!0,this._tree=null,this._loopBindings=null,this._blockBindings=null}_ensureBuilt(){if(this._dirty){this._tree=new Zl(this.primFunc),this._loopBindings=new Map,this._blockBindings=new Map;for(let e of this._tree.allLoops()){let t=e.node,r=t.extent&&t.extent.type==="IntImmNode"?t.extent.value:null;this._loopBindings.set(t.loopVar.name,new zx(e,t.loopVar.name,r,t.kind,t.threadTag))}for(let e of this._tree.allBlocks()){let t=e.node,r=[];for(let i of t.iterVars)i.iterVar&&r.push(i.iterVar.name);let s=[];for(let i of t.reads)s.push(i.buffer.name);let o=[];for(let i of t.writes)o.push(i.buffer.name);this._blockBindings.set(t.name,new Vx(e,r,s,o))}this._dirty=!1}}get tree(){return this._ensureBuilt(),this._tree}invalidate(){this._dirty=!0,this._tree=null,this._loopBindings=null,this._blockBindings=null}getLoopBinding(e){return this._ensureBuilt(),this._loopBindings.get(e)||null}getBlockBinding(e){return this._ensureBuilt(),this._blockBindings.get(e)||null}getBlock(e){let t=this.tree.getBlockSRef(e);if(!t)throw new Error(`Block '${e}' not found in schedule state`);return t}getLoopsOf(e){return this.tree.loopsOf(e)}threadBindingSummary(){this._ensureBuilt();let e={};for(let[,t]of this._loopBindings)t.kind===z.THREAD_BINDING&&t.threadTag&&(e[t.threadTag]={varName:t.iterVar,extent:t.extent});return e}blockDim(){let e=this.threadBindingSummary();return[e["threadIdx.x"]?.extent||1,e["threadIdx.y"]?.extent||1,e["threadIdx.z"]?.extent||1]}gridDim(){let e=this.threadBindingSummary();return[e["blockIdx.x"]?.extent||1,e["blockIdx.y"]?.extent||1,e["blockIdx.z"]?.extent||1]}allBlockNames(){return this._ensureBuilt(),[...this._blockBindings.keys()]}allLoopVarNames(){return this._ensureBuilt(),[...this._loopBindings.keys()]}summary(){return{blocks:this.allBlockNames(),loops:this.allLoopVarNames(),threadBindings:this.threadBindingSummary(),blockDim:this.blockDim(),gridDim:this.gridDim()}}}});var Ay,mM=M(()=>{"use strict";ts();Ay=class{static{l(this,"ScheduleMutator")}constructor(e){this.func=e}replaceNode(e,t){if(e._parent){e.replaceWith(t);return}(this.func.body===e||this.func.body===void 0)&&(this.func.body=t,this.func._setChild&&this.func._setChild("body",t))}removeNode(e){let t=e._parent;if(t&&t.type==="SeqNode"&&Array.isArray(t.stmts)){let r=t.stmts.indexOf(e);if(r>=0){t.stmts.splice(r,1),t._setChildren&&t._setChildren("stmts",t.stmts);return}}throw new Error("removeNode: node parent is not a SeqNode; cannot remove without duplicating it")}redirectReads(e,t,r){e&&qs(e,s=>{s.type==="BufferLoadNode"&&s.buffer===t&&(s.buffer=r)})}redirectBuffer(e,t,r){e&&qs(e,s=>{(s.type==="BufferLoadNode"||s.type==="BufferStoreNode")&&s.buffer===t&&(s.buffer=r)})}}});function Oy(n,e,t){if(!n||typeof n!="object")return n;if(Array.isArray(n))return n.map(e);let r=Object.create(Object.getPrototypeOf(n));switch(r.type=n.type,r._parent=null,r._parentKey=null,r._parentIdx=-1,n.type){case"ForNode":return r.loopVar=n.loopVar,r.min=e(n.min),r.extent=e(n.extent),r.kind=n.kind,r.body=e(n.body),r.threadTag=n.threadTag,r._setChild("body",r.body),r;case"BlockNode":return r.name=n.name,r.iterVars=n.iterVars.map(e),r.reads=n.reads,r.writes=n.writes,r.body=e(n.body),r.initBody=n.initBody?e(n.initBody):null,r._setChild("body",r.body),r._setChild("initBody",r.initBody),r;case"SeqNode":return r.stmts=n.stmts.map(e),r._setChildren("stmts",r.stmts),r;case"IfThenElseNode":return r.condition=e(n.condition),r.thenBody=e(n.thenBody),r.elseBody=n.elseBody?e(n.elseBody):null,r._setChild("thenBody",r.thenBody),r._setChild("elseBody",r.elseBody),r;case"BufferStoreNode":return r.buffer=n.buffer,r.indices=n.indices.map(e),r.value=e(n.value),r;case"BufferLoadNode":return r.buffer=n.buffer,r.indices=n.indices.map(e),r;case"MathOpNode":return r.op=n.op,r.a=e(n.a),r.b=e(n.b),r;case"CompareNode":return r.direction=n.direction,r.a=e(n.a),r.b=e(n.b),r;case"CastNode":return r.expr=e(n.expr),r.fromDtype=n.fromDtype,r.toDtype=n.toDtype,r;case"CallExternNode":return r.externName=n.externName,r.args=n.args.map(e),r.dtype=n.dtype,r;default:return t(n,r,e)}}var Gx=M(()=>{"use strict";l(Oy,"cloneIRShared")});function ia(n,e,t){return Gi(n,r=>r.type==="VariableNode"&&r.name===e?t():r,{bindVars:!1})}function k3(n,e,t,r){return Gi(n,s=>s.type==="BufferLoadNode"&&s.buffer&&s.buffer.name===e?(r.n++,t(s)):s,{bindVars:!1})}function jx(n,e){return kp(n,t=>t.type==="BufferLoadNode"&&t.buffer&&t.buffer.name===e)}function Xl(n,e){if(!n||typeof n!="object")return!1;if((n.type==="BufferLoadNode"||n.type==="BufferStoreNode")&&n.indices){for(let t of n.indices)if(jx(t,e))return!0}for(let t of["a","b","expr","value","condition","thenBody","elseBody","body","initBody"])if(n[t]&&Xl(n[t],e))return!0;if(n.args){for(let t of n.args)if(Xl(t,e))return!0}if(n.indices){for(let t of n.indices)if(Xl(t,e))return!0}if(n.stmts){for(let t of n.stmts)if(Xl(t,e))return!0}return!1}function rn(n){return Oy(n,rn,(e,t,r)=>{if(e.type==="BlockRealizeNode")return t.iterVar=e.iterVar,t.binding=r(e.binding),t;for(let s of Object.keys(e))s==="_parent"||s==="_parentKey"||s==="_parentIdx"||(t[s]=e[s]);return t})}function By(n,e){if(n){if(n.type==="BufferStoreNode"){e.push(n);return}if(n.type==="SeqNode"){for(let t of n.stmts)By(t,e);return}(n.type==="ForNode"||n.type==="BlockNode")&&By(n.body,e)}}function gM(n){return n.indices.map(e=>e&&e.type==="VariableNode"?e.name:null)}function yM(n,e,t){let r=e.indices.map(i=>i.name),s=r.map((i,a)=>`__inl_${t}_${a}`),o={n:0};return k3(n,e.buffer.name,i=>{let a=rn(e.value);for(let u=0;u<r.length;u++)a=ia(a,r[u],()=>new Oe(s[u],"int32"));for(let u=0;u<r.length;u++)a=ia(a,s[u],()=>rn(i.indices[u]));return a},o),o.n}function _M(n,e){let t=[n];for(;t.length>0;){let r=t.pop();if(!(!r||typeof r!="object")){if(r.type==="BlockNode"&&r.reads){let s=r.reads.filter(o=>!(o.buffer&&o.buffer.name===e));s.length!==r.reads.length&&(r.reads=s)}if(r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),r.initBody&&t.push(r.initBody)}}}function fu(n){return n.type==="IntImmNode"?n.value:null}function Ql(n,e="int32"){return new Oe(`${n}_${N3++}`,e)}function xf(n,e,t){if(!(!n||typeof n!="object")){n.type==="BufferLoadNode"&&n.buffer&&e.set(n.buffer.name,n.buffer),n.type==="BufferStoreNode"&&n.buffer&&t.set(n.buffer.name,n.buffer);for(let r of["a","b","value","expr","condition","thenBody","elseBody","body","initBody"])n[r]&&xf(n[r],e,t);if(n.indices)for(let r of n.indices)xf(r,e,t);if(n.args)for(let r of n.args)xf(r,e,t);if(n.stmts)for(let r of n.stmts)xf(r,e,t)}}var v3,N3,Cn,du=M(()=>{"use strict";Fe();ei();rM();Px();hM();mM();Fx();Dx();Gx();ts();v3=new Set(["+","*","min","max"]);l(ia,"substituteVar");l(k3,"replaceBufferLoads");l(jx,"loadsBuffer");l(Xl,"usesBufferInIndex");l(rn,"cloneExprTree");l(By,"collectStores");l(gM,"storeIndexNames");l(yM,"inlineStoreValue");l(_M,"dropBufferReads");l(fu,"getConstExtent");N3=0;l(Ql,"freshVar");Cn=class{static{l(this,"Schedule")}constructor(e){this.func=e,this.trace=new xy,this.state=new Ny(e),this._replaying=!1,this._srefTree=new Zl(e),this.mutator=new Ay(e)}_replaceInTree(e,t){this._srefTree.replaceNode(e,t)||this._srefTree.rebuildFrom(this.func.body),this.state.invalidate()}_removeFromTree(e){this._srefTree.removeNode(e)||this._srefTree.rebuildFrom(this.func.body),this.state.invalidate()}getBlock(e){let t=this._srefTree.getBlockSRef(e);if(!t)throw new Error(`Block '${e}' not found`);return t.node}getLoops(e){return this._srefTree.loopsOf(e).map(t=>t.node)}_resolveLoop(e){if(typeof e!="string")return e;let t=null,r=l(s=>{if(!(!s||typeof s!="object"||t)){if(s.type==="ForNode"&&s.loopVar&&s.loopVar.name===e){t=s;return}if(s.body&&r(s.body),s.initBody&&r(s.initBody),s.stmts)for(let o of s.stmts)r(o);s.thenBody&&r(s.thenBody),s.elseBody&&r(s.elseBody)}},"walk");return r(this.func.body),t||e}split(e,t){e=this._resolveLoop(e);let r=fu(e.extent);if(r===null)throw new Error(`Cannot split loop '${e.loopVar.name}' with non-constant extent`);if(t<=0||!Number.isInteger(t))throw new Error(`Split factor must be a positive integer, got ${t}`);let s=Math.ceil(r/t),o=Ql(`${e.loopVar.name}_o`),i=Ql(`${e.loopVar.name}_i`),a=e.loopVar.name,u=rn(e.body),p=new he(i,new L(0),new L(t),e.kind,u,e.threadTag);if(r%t!==0){let d=new P("+",new P("*",o,new L(t)),i),h=new P("<",d,new L(r)),m=new Se(h,p.body);p.body=m,p._setChild("body",m)}let f=new he(o,new L(0),new L(s),e.kind,p,e.threadTag);return ia(p.body,a,()=>new P("+",new P("*",o,new L(t)),i)),this.mutator.replaceNode(e,f),this._replaceInTree(e,f),this._replaying||this.trace.record("split",[e.loopVar.name,t]),[f,p]}reorder(...e){if(e.length===1&&Array.isArray(e[0])&&(e=e[0]),e=e.map(c=>this._resolveLoop(c)),e.length<2)return;for(let c of e)if(c.type!=="ForNode")throw new Error("reorder expects ForNode arguments");let t=new Set(e);if(t.size!==e.length)throw new Error("reorder: duplicate loop in requested order");let r=null,s=1/0,o=new Map,i=l((c,f)=>{if(c){if(c.type==="ForNode"){t.has(c)&&(o.set(c,f),f<s&&(s=f,r=c)),i(c.body,f+1);return}if(c.type==="SeqNode")for(let d of c.stmts)i(d,f);else c.type==="IfThenElseNode"?i(c.thenBody,f):(c.type==="BlockNode"||c.type==="AllocateNode"||c.type==="LetStmtNode")&&i(c.body,f)}},"findDepths");if(i(this.func.body,0),o.size!==e.length)throw new Error("reorder: not all requested loops were found in the function nest");let{wrappers:a,innermostBody:u}=this._collectReorderNest(t,r);this.mutator.replaceNode(r,e[0]);for(let c of e)c!==e[0]&&(c._parent=null,c._parentKey=null,c._parentIdx=-1);let p=u;for(let c=a.length-1;c>=0;c--)this._setWrapperChild(a[c],p),p=a[c];for(let c=0;c<e.length;c++){let f=c<e.length-1?e[c+1]:p;e[c].body=f,e[c]._setChild("body",f)}this._replaceInTree(r,e[0]),this._replaying||this.trace.record("reorder",[e.map(c=>c.loopVar.name)])}_setWrapperChild(e,t){e.type==="IfThenElseNode"?(e._parent=null,e._parentKey=null,e._parentIdx=-1,e.thenBody=t,e._setChild("thenBody",t)):(e._parent=null,e._parentKey=null,e._parentIdx=-1,e.body=t,e._setChild("body",t))}_collectReorderNest(e,t){let r=new Set(e),s=[],o=t,i=!1,a=null;for(;o;)if(o.type==="ForNode"){if(e.has(o)){if(r.delete(o),i=!0,r.size===0){a=o.body;break}o=o.body;continue}if(i)throw new Error(`reorder: loops are not a perfect nest \u2014 non-reordered loop '${o.loopVar.name}' is interleaved between reordered loops`);o=o.body}else if(o.type==="IfThenElseNode"){if(o.elseBody)throw new Error("reorder: cannot reorder across a conditional with an else-branch");i&&s.push(o),o=o.thenBody}else if(o.type==="AllocateNode"||o.type==="LetStmtNode")i&&s.push(o),o=o.body;else if(o.type==="BlockNode"){if(i)throw new Error("reorder: a compute block separates the reordered loops");o=o.body}else if(o.type==="SeqNode"){if(o.stmts.length!==1)throw new Error("reorder: loops are not a perfect nest \u2014 multiple statements separate the reordered loops");o=o.stmts[0]}else break;if(r.size>0)throw new Error("reorder: loops do not form a single perfect nest");return{wrappers:s,innermostBody:a}}fuseLoops(e,t){if(e=this._resolveLoop(e),t=this._resolveLoop(t),e.type!=="ForNode"||t.type!=="ForNode")throw new Error("fuseLoops expects two ForNode arguments");if(e.body!==t)throw new Error("fuseLoops requires inner loop to be direct child of outer loop");let r=fu(e.extent),s=fu(t.extent);if(r===null||s===null)throw new Error("Cannot fuse loops with non-constant extents");let o=r*s,i=Ql(`${e.loopVar.name}_${t.loopVar.name}_fused`),a=e.loopVar.name,u=t.loopVar.name,p=new he(i,new L(0),new L(o),e.kind,t.body);return ia(p.body,a,()=>new P("//",i,new L(s))),ia(p.body,u,()=>new P("%",i,new L(s))),this.mutator.replaceNode(e,p),this._replaceInTree(e,p),this._replaying||this.trace.record("fuseLoops",[a,u]),p}tile(e,t,r){if(t.length!==r.length)throw new Error("tile: loopIndices and tileSizes must have same length");let s=this.getLoops(e),o=t.map(f=>{if(f>=s.length)throw new Error(`tile: loop index ${f} out of range`);return s[f]}),i=[],a=[];for(let f=0;f<o.length;f++){let h=this.getLoops(e).find(y=>y.loopVar.name===o[f].loopVar.name||y===o[f]);if(!h)throw new Error(`tile: lost track of loop at index ${f}`);let[m,g]=this.split(h,r[f]);i.push(m),a.push(g)}let u=this.getLoops(e),p=[],c=[];for(let f of u)i.some(d=>d.loopVar.name===f.loopVar.name)?p.push(f):a.some(d=>d.loopVar.name===f.loopVar.name)&&c.push(f);return p.length>0&&c.length>0&&this.reorder(...p,...c),{outerLoops:p,innerLoops:c}}vectorize(e){if(e=this._resolveLoop(e),e.type!=="ForNode")throw new Error("vectorize expects ForNode");if(fu(e.extent)===null)throw new Error("Cannot vectorize loop with non-constant extent");let r=Cx(e);if(r!==null)throw new Error(`Cannot vectorize reduction loop '${e.loopVar.name}' (loop-carried dependency in block '${r}')`);e.kind=z.VECTORIZED,this.state.invalidate(),this._replaying||this.trace.record("vectorize",[e.loopVar.name])}unroll(e){if(e=this._resolveLoop(e),e.type!=="ForNode")throw new Error("unroll expects ForNode");e.kind=z.UNROLLED,this.state.invalidate(),this._replaying||this.trace.record("unroll",[e.loopVar.name])}parallelize(e){if(e=this._resolveLoop(e),e.type!=="ForNode")throw new Error("parallelize expects ForNode");let t=Cx(e);if(t!==null)throw new Error(`Cannot parallelize reduction loop '${e.loopVar.name}' (loop-carried dependency in block '${t}')`);e.kind=z.PARALLEL,this.state.invalidate(),this._replaying||this.trace.record("parallelize",[e.loopVar.name])}bindThread(e,t){if(e=this._resolveLoop(e),e.type!=="ForNode")throw new Error("bindThread expects ForNode");let r=["blockIdx.x","blockIdx.y","blockIdx.z","threadIdx.x","threadIdx.y","threadIdx.z"];if(!r.includes(t))throw new Error(`Invalid thread tag: ${t}. Must be one of: ${r.join(", ")}`);e.kind=z.THREAD_BINDING,e.threadTag=t,this.state.invalidate(),this._replaying||this.trace.record("bindThread",[e.loopVar.name,t])}rfactor(e,t,r){let s=this.getBlock(e),o=this.getLoops(e),i=o.find(C=>C.loopVar.name===t);if(!i)throw new Error(`rfactor: reduction loop '${t}' not found for block '${e}'`);let a=fu(i.extent);if(a===null)throw new Error(`rfactor: reduction loop '${t}' has non-constant extent`);if(!Number.isInteger(r)||r<=1||r>=a||a%r!==0)throw new Error(`rfactor: factor ${r} must divide reduction extent ${a} with 1 < factor < ${a}`);let u=s.body;if(!u||u.type!=="BufferStoreNode"||!u.value||u.value.type!=="MathOpNode")throw new Error(`rfactor: block '${e}' body is not a single accumulating store`);let p=u.buffer,c=u.indices,f=u.value.op,d=l(C=>C&&C.type==="BufferLoadNode"&&C.buffer===p,"isAccLoad"),h;if(d(u.value.a))h=u.value.b;else if(d(u.value.b))h=u.value.a;else throw new Error(`rfactor: accumulator load not found in block '${e}' body`);if(!v3.has(f))throw new Error(`rfactor: op '${f}' is not associative+commutative; cannot factor reduction`);let m=s.initBody&&s.initBody.type==="BufferStoreNode"&&s.initBody.value?s.initBody.value:new L(0),g=o.filter(C=>C.loopVar.name!==t),y=a/r,_=new me(`${p.name}_rf`,[r,...p.shape],p.dtype,p.scope),b=Ql(`${t}_rfi`),w=Ql(`${t}_rfo`),x=Ql(`${t}_rfp`),S=l(C=>[C,...c.map(rn)],"cfIdx"),T=ia(rn(h),t,()=>new P("+",new P("*",w,new L(r)),b)),v=new J(_,S(b),new P(f,new q(_,S(b)),T)),A=new J(_,S(b),rn(m)),N=new fe(`${e}_rf_p`,[],s.reads.map(C=>({buffer:C.buffer})),[{buffer:_}],v,A),k=new he(w,new L(0),new L(y),z.SERIAL,N);k=new he(b,new L(0),new L(r),z.SERIAL,k);for(let C=g.length-1;C>=0;C--)k=new he(g[C].loopVar,new L(0),rn(g[C].extent),z.SERIAL,k);let O=new J(p,c.map(rn),new P(f,new q(p,c.map(rn)),new q(_,S(x)))),R=new J(p,c.map(rn),rn(m)),B=new fe(`${e}_rf_c`,[],[{buffer:_}],[{buffer:p}],O,R),E=new he(x,new L(0),new L(r),z.SERIAL,B);for(let C=g.length-1;C>=0;C--)E=new he(g[C].loopVar,new L(0),rn(g[C].extent),z.SERIAL,E);let F=new te([k,E]);return this.mutator.replaceNode(o[0],F),this._replaceInTree(o[0],F),this._replaying||this.trace.record("rfactor",[e,t,r]),_}decomposeReduction(e){let t=this.getBlock(e);if(!t.initBody)throw new Error(`decomposeReduction: block '${e}' has no initBody`);let r=this.getLoops(e),s=t.body;if(!s||s.type!=="BufferStoreNode")throw new Error(`decomposeReduction: block '${e}' body is not a store`);let o=s.buffer,i=new Set;for(let g of s.indices)Nt(g,i);let a=r.filter(g=>i.has(g.loopVar.name));if(r.filter(g=>!i.has(g.loopVar.name)).length===0)throw new Error(`decomposeReduction: block '${e}' has no reduction loop`);let p=new J(o,s.indices.map(rn),rn(t.initBody.value)),f=new fe(`${e}_init`,[],[],[{buffer:o}],p);for(let g=a.length-1;g>=0;g--)f=new he(a[g].loopVar,new L(0),rn(a[g].extent),z.SERIAL,f);let h=new fe(`${e}_upd`,[],t.reads.map(g=>({buffer:g.buffer})),[{buffer:o}],rn(s));for(let g=r.length-1;g>=0;g--)h=new he(r[g].loopVar,new L(0),rn(r[g].extent),z.SERIAL,h);let m=new te([f,h]);this.mutator.replaceNode(r[0],m),this._replaceInTree(r[0],m),this._replaying||this.trace.record("decomposeReduction",[e])}cacheWrite(e,t,r="local"){let s=this.getBlock(e),o=this.getLoops(e);if(o.length===0)throw new Error("cacheWrite: block has no enclosing loops");let i=(s.writes||[]).find(y=>y.buffer&&y.buffer.name===t);if(!i)throw new Error(`cacheWrite: block '${e}' does not write '${t}'`);let a=i.buffer,u=new me(`${t}_${e}_cachew`,[...a.shape],a.dtype,r);this.mutator.redirectBuffer(s.body,a,u),s.initBody&&this.mutator.redirectBuffer(s.initBody,a,u),i.buffer=u;let p=a.shape.map((y,_)=>new Oe(`${u.name}_o${_}`,"int32")),c=new J(a,p,new q(u,p)),d=new fe(`${u.name}_flush`,p.map(y=>({iterVar:y,binding:y})),[{buffer:u}],[{buffer:a}],c);for(let y=a.shape.length-1;y>=0;y--)d=new he(p[y],new L(0),new L(a.shape[y]),z.SERIAL,d);let h=o[0],m=new te([]),g=new Ue(u,r,m);this.mutator.replaceNode(h,g),m.stmts.push(h,d),this._replaceInTree(h,g),this._replaying||this.trace.record("cacheWrite",[e,t,r])}setScope(e,t,r){let o=(this.getBlock(e).writes||[]).find(i=>i.buffer&&i.buffer.name===t);if(!o)throw new Error(`setScope: block '${e}' does not write '${t}'`);o.buffer.scope=r,this.state.invalidate(),this._replaying||this.trace.record("setScope",[e,t,r])}storageAlign(e,t,r,s,o){let i=this.getBlock(e),a=[...i.writes||[],...i.reads||[]].find(u=>u.buffer&&u.buffer.name===t);if(!a)throw new Error(`storageAlign: block '${e}' does not access '${t}'`);if(!Number.isInteger(s)||s<=0)throw new Error("storageAlign: factor must be a positive integer");a.buffer.storageAlign={axis:r,factor:s,offset:o||0},this.state.invalidate(),this._replaying||this.trace.record("storageAlign",[e,t,r,s,o||0])}_removeBlockNest(e,t){let r=this.getLoops(e),s=r.length>0?r[0]:t,o=new te([]);this.mutator.replaceNode(s,o),this._replaceInTree(s,o)}computeInline(e){let t=this.getBlock(e);if(!t)throw new Error(`computeInline: block '${e}' not found`);if(t.initBody)throw new Error("computeInline: cannot inline a reduction block (has init)");let r=[];By(t.body,r);let s=r[0];if(!s)throw new Error("computeInline: producer has no single store to inline");let o=s.buffer;if(gM(s).some(i=>i===null))throw new Error("computeInline: producer indices must be simple loop variables");if(jx(s.value,o.name))throw new Error("computeInline: producer is self-referential (recurrence), cannot inline");if(Xl(this.func.body,o.name))throw new Error(`computeInline: buffer '${o.name}' is used inside an index expression (indirect), cannot safely inline`);if(yM(this.func.body,s,e)===0)throw new Error(`computeInline: buffer '${o.name}' has no consumers to inline into`);_M(this.func.body,o.name),this._removeBlockNest(e,t),this._replaying||this.trace.record("computeInline",[e])}computeInlineBlock(e){let t=this.getBlock(e);if(!t)throw new Error(`computeInlineBlock: block '${e}' not found`);if(t.initBody)throw new Error("computeInlineBlock: cannot inline a reduction block (has init)");let r=[];if(By(t.body,r),r.length===0)throw new Error("computeInlineBlock: producer has no stores to inline");let s=new Set(r.map(i=>i.buffer.name));if(s.size!==r.length)throw new Error("computeInlineBlock: buffer written more than once in block");for(let i of r){if(gM(i).some(a=>a===null))throw new Error("computeInlineBlock: producer indices must be simple loop variables");for(let a of s)if(jx(i.value,a))throw new Error("computeInlineBlock: producer store depends on a co-produced buffer");if(Xl(this.func.body,i.buffer.name))throw new Error(`computeInlineBlock: buffer '${i.buffer.name}' is used inside an index expression (indirect), cannot safely inline`)}let o=0;for(let i of r)o+=yM(this.func.body,i,`${e}_${i.buffer.name}`);if(o===0)throw new Error(`computeInlineBlock: block '${e}' has no consumers to inline into`);for(let i of s)_M(this.func.body,i);this._removeBlockNest(e,t),this._replaying||this.trace.record("computeInlineBlock",[e])}_relocateBlockToLoop(e,t,r){if(!this.getBlock(e))throw new Error(`computeAt: block '${e}' not found`);let o=this._resolveLoop(t);if(!o||o.type!=="ForNode")throw new Error("computeAt: target must be a loop");let i=this.getLoops(e);if(i.length!==1)throw new Error("computeAt: aligned case requires exactly one enclosing loop on the moved block");let a=i[0];if(a===o)throw new Error("computeAt: block already at target loop");let u=fu(a.extent),p=fu(o.extent);if(u===null||p===null||u!==p)throw new Error("computeAt: aligned case requires equal static extent on the block and target loops");let c=rn(a.body);ia(c,a.loopVar.name,()=>o.loopVar);let f=new te([]);this.mutator.replaceNode(a,f);let d=o.body;if(d&&d.type==="SeqNode")r?d.stmts.unshift(c):d.stmts.push(c);else{let h=r?new te([c,d]):new te([d,c]);o.body=h,o._setChild("body",h)}return this._replaceInTree(a,f),this._replaceInTree(o,o),o.loopVar.name}computeAt(e,t){let r=this._relocateBlockToLoop(e,t,!0);this._replaying||this.trace.record("computeAt",[e,r])}reverseComputeAt(e,t){let r=this._relocateBlockToLoop(e,t,!1);this._replaying||this.trace.record("reverseComputeAt",[e,r])}cacheRead(e,t,r="local"){let s=this.getBlock(e),o=this.getLoops(e);if(o.length===0)throw new Error("cacheRead: block has no enclosing loops");let i=(s.reads||[]).find(y=>y.buffer&&y.buffer.name===t);if(!i)throw new Error(`cacheRead: block '${e}' does not read '${t}'`);let a=i.buffer,u=new me(`${t}_${e}_cache`,[...a.shape],a.dtype,r),p=a.shape.map((y,_)=>new Oe(`${u.name}_i${_}`,"int32")),c=new J(u,p,new q(a,p)),d=new fe(`${u.name}_fill`,p.map(y=>({iterVar:y,binding:y})),[{buffer:a}],[{buffer:u}],c);for(let y=a.shape.length-1;y>=0;y--)d=new he(p[y],new L(0),new L(a.shape[y]),z.SERIAL,d);this.mutator.redirectReads(s.body,a,u),s.initBody&&this.mutator.redirectReads(s.initBody,a,u),i.buffer=u;let h=o[0],m=new te([d]),g=new Ue(u,r,m);this.mutator.replaceNode(h,g),m.stmts.push(h),this._replaceInTree(h,g),this._replaying||this.trace.record("cacheRead",[e,t,r])}fuseConsumer(e,t){let r=this.getBlock(e),s=this.getBlock(t),o=this.getLoops(e),i=this.getLoops(t);if(!r.body||r.body.type!=="BufferStoreNode")throw new Error(`fuseConsumer: producer '${e}' body is not a store`);let a=new Set;for(let g of r.body.indices)Nt(g,a);let u=o.filter(g=>a.has(g.loopVar.name));if(u.length===0||i.length!==u.length)throw new Error(`fuseConsumer: producer/consumer spatial rank mismatch (${u.length} vs ${i.length})`);if(!i[0]._parent||i[0]._parent.type!=="SeqNode")throw new Error("fuseConsumer: consumer loop nest is not a direct SeqNode sibling; cannot fuse without duplicating it");let p=u[u.length-1],c=rn(p.body),f=rn(s.body);for(let g=0;g<i.length;g++){let y=u[g].loopVar.name;f=ia(f,i[g].loopVar.name,()=>new Oe(y,"int32"))}let d=new fe(`${t}_fused`,[],s.reads.map(g=>({buffer:g.buffer})),s.writes.map(g=>({buffer:g.buffer})),f),h=new te([c,d]);for(let g=u.length-1;g>=0;g--){let y=u[g];h=new he(y.loopVar,new L(0),rn(y.extent),y.kind,h,y.threadTag)}let m=i[0];this.mutator.replaceNode(o[0],h),this.mutator.removeNode(m),this._replaceInTree(o[0],h),this._removeFromTree(m),this._replaying||this.trace.record("fuseConsumer",[e,t])}annotate(e,t,r){if(e=this._resolveLoop(e),e.type!=="ForNode")throw new Error("annotate expects ForNode");e.annotations||(e.annotations={}),e.annotations[t]=r,this.state.invalidate(),this._replaying||this.trace.record("annotate",[e.loopVar.name,t,r])}tensorize(e,t){if(typeof e!="string")throw new Error("tensorize expects an intrinsic name");if(!t||typeof t.M!="number"||typeof t.N!="number"||typeof t.K!="number")throw new Error("tensorize expects info { M, N, K, a, b, c }");this.func._tensorIntrin={name:e,info:t},this.state.invalidate()}blockize(e,t=null){let r=this._resolveLoop(e);if(!r||r.type!=="ForNode")throw new Error("blockize expects a loop");let s=new Map,o=new Map;xf(r,s,o);let i=t||`blockized_${r.loopVar.name}`,a=new fe(i,[],[...s.values()].map(u=>({buffer:u})),[...o.values()].map(u=>({buffer:u})),new te([]));return this.mutator.replaceNode(r,a),a.body=r,a._setChild("body",r),this._replaceInTree(r,a),this._replaying||this.trace.record("blockize",[r.loopVar.name]),a}getTrace(){return this.trace}verify(){return Jl.validate(this.func)}};l(xf,"collectBufferAccess")});function Tt(n,e){let t=Ry.get(n);return t&&t.has(e)?t.get(e):(t||(t=new Map,Ry.set(n,t)),O3(n.body,t,[]),t.has(e)?t.get(e):null)}function Xx(n){n&&Ry.delete(n)}function A3(n,e){let t=Ry.get(n);t&&t.delete(e)}function bM(n,e){let t=[n];for(;t.length>0;){let r=t.pop();if(!(!r||typeof r!="object")){if(r.type==="VariableNode"){e.add(r.name);continue}for(let s of["a","b","condition","thenBody","elseBody","expr","value","offsetExpr","extent"])r[s]&&t.push(r[s]);if(r.indices)for(let s of r.indices)t.push(s);if(r.args)for(let s of r.args)t.push(s)}}}function wM(n){let e=new Set((n.writes||[]).map(o=>o.buffer&&o.buffer.name)),t=new Set,r=[n.body,n.initBody];for(;r.length>0;){let o=r.pop();if(o){if(o.type==="BufferStoreNode"&&o.buffer&&e.has(o.buffer.name))for(let i of o.indices)bM(i,t);if(o.body&&r.push(o.body),o.stmts)for(let i of o.stmts)r.push(i);o.thenBody&&r.push(o.thenBody),o.elseBody&&r.push(o.elseBody),o.value&&r.push(o.value)}}let s=new Set;for(let o of n.iterVars||[]){let i=new Set;if(bM(o.binding,i),!(o.iterVar&&t.has(o.iterVar.name)||[...i].some(u=>t.has(u))))for(let u of i)s.add(u)}return s}function O3(n,e,t){let r=[{node:n,loops:[...t]}];for(;r.length>0;){let{node:s,loops:o}=r.pop();if(s)if(s.type==="ForNode")r.push({node:s.body,loops:[...o,s]});else if(s.type==="BlockNode"){if(!e.has(s.name)){let i=wM(s);e.set(s.name,{loopCount:o.length,hasReduction:s.initBody!==null||i.size>0,reductionLoopVars:i,readBuffers:s.reads.map(a=>a.buffer.name),writeBuffers:s.writes.map(a=>a.buffer.name),loops:[...o]})}r.push({node:s.body,loops:o})}else if(s.type==="SeqNode")for(let i=s.stmts.length-1;i>=0;i--)r.push({node:s.stmts[i],loops:o});else s.type==="IfThenElseNode"?(s.elseBody&&r.push({node:s.elseBody,loops:o}),r.push({node:s.thenBody,loops:o})):s.type==="AllocateNode"?r.push({node:s.body,loops:o}):s.type==="LetStmtNode"&&r.push({node:s.body,loops:o})}}function TM(n){return n.hasReduction&&n.readBuffers.length===2&&n.writeBuffers.length===1&&n.loopCount>=3}function B3(n,e){let t=Tt(n,e);return t?t.loops.some(r=>r.extent&&r.extent.type!=="IntImmNode"):!1}function xM(n){let e=0,t=[n.body];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="BlockNode"&&(e++,e>1))return!0;if(r.type==="ForNode"&&t.push(r.body),r.type==="SeqNode")for(let s of r.stmts)t.push(s)}}return!1}function eS(n){return kp(n.body,e=>e.type==="BlockNode"&&(e.initBody!==null||wM(e).size>0),{kinds:"stmt"})}function tS(n){return kp(n.body,e=>e.type==="ForNode"&&e.kind===z.RECURRENCE,{kinds:"stmt"})}function nS(n,e,t){let r=Math.min(t.maxThreadsPerBlock,256),s=e.extent;if(s.type==="IntImmNode"&&s.value>r){let[o,i]=n.split(e,r);n.bindThread(i,"threadIdx.x"),tS(n.func)||n.bindThread(o,"blockIdx.x")}else n.bindThread(e,"threadIdx.x")}function ec(n,e){return e.reductionLoopVars?e.reductionLoopVars.has(n.loopVar.name):!1}function rS(n,e){return n.body===e}function R3(n){return yl(n,e=>e.type==="BlockNode",{kinds:"stmt"}).map(e=>e.name)}var eo,Ry,Ux,qx,Wx,Kx,Hx,Yx,Jx,Zx,Qx,to,Do=M(()=>{"use strict";yo();Fe();ts();eo=class{static{l(this,"ScheduleRule")}constructor(e){this.name=e}matches(e,t,r){throw new Error("ScheduleRule.matches must be implemented")}apply(e,t,r){throw new Error("ScheduleRule.apply must be implemented")}},Ry=new WeakMap;l(Tt,"classifyBlock");l(Xx,"invalidateClassifyCache");l(A3,"invalidateClassifyBlock");l(bM,"collectVarNames");l(wM,"computeReductionLoopVars");l(O3,"collectBlockInfo");l(TM,"isMatmulShape");l(B3,"blockHasNonConstExtent");l(xM,"hasMultipleBlocks");Ux=class extends eo{static{l(this,"ElementwiseCPURule")}constructor(){super("elementwise_cpu")}matches(e,t,r){if(r.kind!==Dt.CPU)return!1;let s=Tt(e,t);if(!s||s.hasReduction||s.loopCount<1||s.loops.length>0&&xM(s.loops[0]))return!1;let o=1;for(let i of s.loops){let a=i.extent&&i.extent.type==="IntImmNode"?i.extent.value:1;o*=a}return o>=r.numCores*r.vectorWidth}apply(e,t,r){let s=e.getLoops(t);if(s.length===0)return;if(s.length===1){let a=s[0].extent;if(a.type==="IntImmNode"&&a.value>=r.vectorWidth*2){let[u,p]=e.split(s[0],r.vectorWidth);e.parallelize(u),e.vectorize(p);return}e.parallelize(s[0]);return}e.parallelize(s[0]);let o=s[s.length-1],i=o.extent;if(i.type==="IntImmNode"&&i.value>=r.vectorWidth&&i.value%r.vectorWidth===0){let[a,u]=e.split(o,r.vectorWidth);e.vectorize(u)}}};l(eS,"primFuncHasReduction");l(tS,"primFuncHasRecurrence");l(nS,"bindFusedSpatialGPU");qx=class extends eo{static{l(this,"ElementwiseGPURule")}constructor(){super("elementwise_gpu")}matches(e,t,r){if(!r.isGPU())return!1;let s=Tt(e,t);return s?!s.hasReduction&&s.loopCount>=1:!1}apply(e,t,r){let s=e.getLoops(t);if(s.length===0)return;let o=s[0];for(let p=1;p<s.length;p++){let f=e.getLoops(t).find(d=>d.loopVar.name===s[p].loopVar.name);f&&rS(o,f)&&(o=e.fuseLoops(o,f))}let i=o.extent;if(i.type!=="IntImmNode"){e.bindThread(o,"threadIdx.x");return}let a=i.value,u=Math.min(r.maxThreadsPerBlock,1024);if(eS(e.func)&&a<=u){e.bindThread(o,"threadIdx.x");return}nS(e,o,r)}},Wx=class extends eo{static{l(this,"ReductionCPURule")}constructor(){super("reduction_cpu")}matches(e,t,r){if(r.kind!==Dt.CPU)return!1;let s=Tt(e,t);return s?s.hasReduction:!1}apply(e,t,r){let s=e.getLoops(t);if(s.length===0)return;let o=[],i=[],a=Tt(e.func,t);for(let u of s)a&&ec(u,a)?i.push(u):o.push(u);o.length>0&&e.parallelize(o[0])}},Kx=class extends eo{static{l(this,"ReductionGPURule")}constructor(){super("reduction_gpu")}matches(e,t,r){if(!r.isGPU())return!1;let s=Tt(e,t);return s?s.hasReduction:!1}apply(e,t,r){let s=e.getLoops(t);if(s.length===0)return;let o=[],i=Tt(e.func,t);for(let u of s)(!i||!ec(u,i))&&o.push(u);if(o.length===0)return;let a=o[0];for(let u=1;u<o.length;u++){let c=e.getLoops(t).find(f=>f.loopVar.name===o[u].loopVar.name);c&&rS(a,c)&&(a=e.fuseLoops(a,c))}nS(e,a,r)}},Hx=class extends eo{static{l(this,"MatmulTiledCPURule")}constructor(){super("matmul_tiled_cpu")}matches(e,t,r){if(r.kind!==Dt.CPU)return!1;let s=Tt(e,t);if(!s||!TM(s))return!1;let o=r.l1CacheBytes||32768,i=Math.max(8,Math.min(64,Math.floor(Math.sqrt(o/4))));return s.loops.reduce((u,p)=>{let c=p.extent&&p.extent.type==="IntImmNode"?p.extent.value:0;return c>u?c:u},0)>=i}apply(e,t,r){let s=e.getLoops(t);if(s.length<3)return;let o=r.l1CacheBytes||32768,i=Math.max(8,Math.min(64,Math.floor(Math.sqrt(o/4)))),a=[],u=[];for(let c=0;c<Math.min(2,s.length);c++){let f=s[c].extent.type==="IntImmNode"?s[c].extent.value:null;f&&f>=i&&(a.push(c),u.push(i))}if(a.length===0)return;let{outerLoops:p}=e.tile(t,a,u);p.length>0&&e.parallelize(p[0])}},Yx=class extends eo{static{l(this,"MatmulTiledGPURule")}constructor(){super("matmul_tiled_gpu")}matches(e,t,r){if(!r.isGPU())return!1;let s=Tt(e,t);if(!s||!TM(s))return!1;let o=r.sharedMemoryBytes||49152,a=Math.max(16,Math.min(128,Math.floor(Math.sqrt(o/8))));return s.loops.reduce((p,c)=>{let f=c.extent&&c.extent.type==="IntImmNode"?c.extent.value:0;return f>p?f:p},0)>=a}apply(e,t,r){let s=e.getLoops(t);if(s.length<3)return;let o=Tt(e.func,t),i=s.filter(u=>!o||!ec(u,o));if(i.length===0)return;let a=i[0];for(let u=1;u<i.length;u++){let c=e.getLoops(t).find(f=>f.loopVar.name===i[u].loopVar.name);c&&rS(a,c)&&(a=e.fuseLoops(a,c))}nS(e,a,r)}},Jx=class extends eo{static{l(this,"ElementwiseWasmRule")}constructor(){super("elementwise_wasm")}matches(e,t,r){if(r.kind!==Dt.WASM)return!1;let s=Tt(e,t);return!(!s||s.hasReduction||s.loopCount<1||s.loops.length>0&&xM(s.loops[0]))}apply(e,t,r){let s=e.getLoops(t);if(s.length===0)return;let o=r.numCores||1,i=r.vectorWidth||4;if(o>1&&s.length>=1){let p=s[0].extent,c=p&&p.type==="IntImmNode"?p.value:0;if(c>=o*4){if(s.length===1&&r.supportsSimd&&r.supportsSimd()&&c>=i*2&&c%i===0){let[f,d]=e.split(s[0],i);e.parallelize(f),e.vectorize(d);return}if(e.parallelize(s[0]),s.length>1){let f=s[s.length-1],d=f.extent;if(d.type==="IntImmNode"&&d.value>=i&&d.value%i===0){let[,h]=e.split(f,i);e.vectorize(h)}}return}}let a=s[s.length-1],u=a.extent;if(u.type==="IntImmNode"&&u.value>=i*2){let[p,c]=e.split(a,i);e.vectorize(c);return}u.type==="IntImmNode"&&u.value>=i&&e.vectorize(a)}},Zx=class extends eo{static{l(this,"ReductionWasmRule")}constructor(){super("reduction_wasm")}matches(e,t,r){if(r.kind!==Dt.WASM)return!1;let s=r.numCores>1,o=r.supportsSimd&&r.supportsSimd();if(!s&&!o)return!1;let i=Tt(e,t);return i?i.hasReduction&&i.loopCount>=2:!1}apply(e,t,r){let s=e.getLoops(t);if(s.length<2)return;let o=Tt(e.func,t),i=[],a=[];for(let u of s)!o||!ec(u,o)?i.push(u):a.push(u);if(i.length>0){let u=i[0].extent;(u&&u.type==="IntImmNode"?u.value:0)>=(r.numCores||1)*4&&e.parallelize(i[0])}if(r.supportsSimd&&r.supportsSimd()&&a.length>0){let u=a[a.length-1];(u.extent&&u.extent.type==="IntImmNode"?u.extent.value:0)>=r.vectorWidth*2&&e.vectorize(u)}}},Qx=class extends eo{static{l(this,"FallbackRule")}constructor(){super("fallback")}matches(){return!0}apply(e,t,r){let s=e.getLoops(t);s.length!==0&&r.isCPU()&&s.length>=1&&e.parallelize(s[0])}};l(ec,"isReductionLoop");l(rS,"findDirectChild");to=class n{static{l(this,"SchedulePolicy")}constructor(e,t=null,r=null){this.target=e,this.rules=t||n.defaultRules(),this.trace=r}static defaultRules(){return[new Hx,new Yx,new Wx,new Kx,new Zx,new Ux,new qx,new Jx,new Qx]}selectRule(e,t){for(let r of this.rules)if(r.matches(e,t,this.target))return r;return null}applyToBlock(e,t){if(this.target.isGPU()&&B3(e.func,t))return this._explain(t,"none","block has dynamic loop extents; runs sequentially (no dynamic grid)"),null;let r=this.selectRule(e.func,t);return r?(r.apply(e,t,this.target),A3(e.func,t),this._explain(t,r.name,`matched rule '${r.name}' for ${this.target.name}`),r.name):(this._explain(t,"none","no schedule rule matched; runs sequentially"),null)}_explain(e,t,r){this.trace&&this.trace.explainsEnabled&&this.trace.explain("schedule",e,t,r,{target:this.target.name})}applyToAllBlocks(e){Xx(e.func);let t=R3(e.func.body),r=new Set,s=new Map;for(let o of t){if(r.has(o))continue;r.add(o);let i=this.applyToBlock(e,o);i&&s.set(o,i)}return s}};l(R3,"collectAllBlockNames")});function My(n,e,t,r){if(M3[e])return I3(n,e,t,r);let s=L3[e];return s?s(n,t,r||{}):typeof n[e]=="function"?E3(n,e,t,r):n._inferAndBuild(e,t,r||null)}function I3(n,e,t,r){let s=t[0].type.rank,o=r?.dim,i=o!=null?(Array.isArray(o)?o:[o]).map(f=>f<0?s+f:f):Array.from({length:s},(f,d)=>d),a=n.scalarConstant(MN(e,t[0].type.dtype),t[0].type.dtype),u=n.reduce(t[0],a.getResult(0),i,e);if(!r?.keepdim)return u;let p=new Set(i),c=t[0].type.shape.map((f,d)=>p.has(d)?1:f);return n.reshape(u.getResult(0),c)}function E3(n,e,t,r){return t.length===1?De(n,e,t[0]):t.length===2?De(n,e,t[0],t[1]):t.length===3?De(n,e,t[0],t[1],t[2]):n._inferAndBuild(e,t,r||null)}function De(n,e,...t){return n[e].call(n,...t)}function bi(n,e){return n<0?e+n:n}function C3(n,e){let t=1,r=-1,s=1;for(let i of e)s*=i;let o=[...n];for(let i=0;i<o.length;i++)o[i]===-1?r=i:t*=o[i];return r>=0&&(o[r]=t===0?0:s/t),o}function SM(n,e){return n<0?e+1+n:n}function D3(n,e,t,r,s){let o=n.length,i=bi(e,o),a=n[i],u=t??0,p=r??a,c=s??1;u<0&&(u+=a),p<0&&(p+=a),u=Math.max(0,Math.min(u,a)),p=Math.max(0,Math.min(p,a));let f=new Array(o).fill(0),d=[...n],h=new Array(o).fill(1);return f[i]=u,d[i]=p,h[i]=c,{starts:f,limits:d,strides:h}}var M3,L3,sS=M(()=>{"use strict";_s();Re();tr();M3=Object.freeze({sum:"sum",mean:"mean",max:"max",min:"min",prod:"prod"});l(My,"buildMappedOp");l(I3,"buildReduce");l(E3,"callBuilder");l(De,"callMethod");L3=Object.freeze({matmul:l((n,e)=>De(n,"matmul",e[0],e[1]),"matmul"),dot:l((n,e)=>De(n,"dot",e[0],e[1],[e[0].type.rank-1],[0]),"dot"),clone:l((n,e)=>n._inferAndBuild("add",[e[0],n.scalarConstant(0,e[0].type.dtype).getResult(0)],null),"clone"),relu:l((n,e)=>De(n,"relu",e[0]),"relu"),sigmoid:l((n,e)=>De(n,"sigmoid",e[0]),"sigmoid"),gelu:l((n,e)=>De(n,"gelu",e[0]),"gelu"),silu:l((n,e)=>De(n,"silu",e[0]),"silu"),softmax:l((n,e,t)=>De(n,"softmax",e[0],t?.dim??-1),"softmax"),log_softmax:l((n,e,t)=>De(n,"logSoftmax",e[0],t?.dim??-1),"log_softmax"),layer_norm:l((n,e,t)=>De(n,"layernorm",e[0],e[1],e[2],t?.axis??-1,t?.eps??1e-5),"layer_norm"),batch_norm:l((n,e,t)=>De(n,"batchnorm",e[0],e[1],e[2],e[3],e[4],t?.axis??1,t?.eps??1e-5),"batch_norm"),conv2d:l((n,e,t)=>De(n,"conv",e[0],e[1],t?.strides??[1,1],t?.padding??[[0,0],[0,0]],{dilation:t?.dilation??[1,1],groups:t?.groups??1}),"conv2d"),pool2d:l((n,e,t)=>De(n,"pool2d",e[0],t?.pool_type??"max",t?.kernel_size??[2,2],t?.strides??[2,2],t?.padding??[[0,0],[0,0]]),"pool2d"),embedding:l((n,e)=>De(n,"embedding",e[0],e[1]),"embedding"),argmax:l((n,e,t)=>De(n,"argmax",e[0],t?.dim??0,t?.keepdim??!1),"argmax"),argmin:l((n,e,t)=>De(n,"argmin",e[0],t?.dim??0,t?.keepdim??!1),"argmin"),maximum:l((n,e)=>De(n,"maximum",e[0],e[1]),"maximum"),minimum:l((n,e)=>De(n,"minimum",e[0],e[1]),"minimum"),eq:l((n,e)=>De(n,"compare",e[0],e[1],"eq"),"eq"),ne:l((n,e)=>De(n,"compare",e[0],e[1],"ne"),"ne"),lt:l((n,e)=>De(n,"compare",e[0],e[1],"lt"),"lt"),le:l((n,e)=>De(n,"compare",e[0],e[1],"le"),"le"),gt:l((n,e)=>De(n,"compare",e[0],e[1],"gt"),"gt"),ge:l((n,e)=>De(n,"compare",e[0],e[1],"ge"),"ge"),clamp:l((n,e)=>De(n,"clamp",e[1],e[0],e[2]),"clamp"),pad:l((n,e,t)=>De(n,"pad",e[0],e[1],t.low,t.high),"pad"),one_hot:l((n,e,t)=>De(n,"oneHot",e[0],t.depth,{dtype:"f32"}),"one_hot"),index_select:l((n,e,t)=>De(n,"gather",e[0],e[1],xA(e[0].type,t?.dim??0,e[1].type.rank)),"index_select"),gather:l((n,e,t)=>De(n,"gatherDim",e[0],e[1],t?.dim??0),"gather"),scatter_add:l((n,e,t)=>De(n,"scatterAddDim",e[0],e[1],e[2],t?.dim??0),"scatter_add"),cat:l((n,e,t)=>De(n,"concat",e,bi(t?.dim??0,e[0].type.rank)),"cat"),stack:l((n,e,t)=>{let r=SM(t?.dim??0,e[0].type.rank),s=e.map(o=>{let i=[...o.type.shape];return i.splice(r,0,1),n.reshape(o,i).getResult(0)});return De(n,"concat",s,r)},"stack"),reshape:l((n,e,t)=>n.reshape(e[0],C3(t.shape??t.new_shape,e[0].type.shape)),"reshape"),transpose:l((n,e,t)=>{let r=e[0].type.rank,s=bi(t?.dim0??0,r),o=bi(t?.dim1??1,r),i=Array.from({length:r},(a,u)=>u);return i[s]=o,i[o]=s,n.transpose(e[0],i)},"transpose"),permute:l((n,e,t)=>De(n,"transpose",e[0],t.dims),"permute"),broadcast_in_dim:l((n,e,t)=>De(n,"broadcast",e[0],t.result_shape,t.broadcast_dimensions),"broadcast_in_dim"),expand:l((n,e,t)=>{let r=e[0].type.rank,s=t.shape,o=s.length-r,i=s.map((u,p)=>u===-1?e[0].type.shape[p-o]:u),a=Array.from({length:r},(u,p)=>p+o);return De(n,"broadcast",e[0],i,a)},"expand"),slice:l((n,e,t)=>{let r=D3(e[0].type.shape,t.dim,t.start,t.end,t.step);return De(n,"slice",e[0],r.starts,r.limits,r.strides)},"slice"),unsqueeze:l((n,e,t)=>{let r=[...e[0].type.shape];return r.splice(SM(t.dim,r.length),0,1),n.reshape(e[0],r)},"unsqueeze"),squeeze:l((n,e,t)=>{let r=[...e[0].type.shape];if(t.dim===void 0||t.dim===null)return n.reshape(e[0],r.filter(o=>o!==1));let s=bi(t.dim,r.length);return r[s]===1&&r.splice(s,1),n.reshape(e[0],r)},"squeeze"),narrow:l((n,e,t)=>{let r=bi(t.dim,e[0].type.rank),s=new Array(e[0].type.rank).fill(0),o=[...e[0].type.shape],i=new Array(e[0].type.rank).fill(1);return s[r]=t.start,o[r]=t.start+t.length,De(n,"slice",e[0],s,o,i)},"narrow"),select:l((n,e,t)=>{let r=e[0].type.rank,s=bi(t.dim,r),o=t.index,i=o<0?e[0].type.shape[s]+o:o,a=new Array(r).fill(0),u=[...e[0].type.shape];a[s]=i,u[s]=i+1;let p=De(n,"slice",e[0],a,u,new Array(r).fill(1)).getResult(0);return n.reshape(p,e[0].type.shape.filter((c,f)=>f!==s))},"select"),split:l((n,e,t)=>De(n,"split",e[0],bi(t.dim??0,e[0].type.rank),t.sizes),"split"),chunk:l((n,e,t)=>{let r=bi(t.dim??0,e[0].type.rank),s=e[0].type.shape[r],o=Math.ceil(s/t.chunks),i=[];for(let a=0;a<s;a+=o)i.push(Math.min(o,s-a));return De(n,"split",e[0],r,i)},"chunk")});l(bi,"normalizeDim");l(C3,"resolveShape");l(SM,"normalizeInsertDim");l(D3,"sliceAttrs")});function $3(n,e,t,r){let s=n;for(let o=0;o<e.length;o++)s+="|"+e[o].shape.join(",")+":"+e[o].dtype;if(t)for(let[o,i]of Object.entries(t))s+="|"+o+"="+JSON.stringify(i);return s+="|"+r.name,s}function F3(n){let e=kM.get(n);return e||(e=new jl("jit_"+n),kM.set(n,e)),e}function z3(n){let e=1;for(let t of n.shape)e*=typeof t=="number"&&t>0?t:1;return Math.max(e,1)}function V3(n,e,t){let r=[];for(let[,s]of t.bufferMap)r.push(new(St(s.dtype))(z3(s)));n.run(e.name,...r)}function G3(n,e,t,r){try{let s=NM(n,e);if(s.shapeParams&&s.shapeParams.length>0)return null;new to(e).applyToAllBlocks(new Cn(s));let o=t.compile(s);return r.addCompiledKernel(o),e.isWebGPU()||V3(r,o,s),o}catch{return null}}function U3(n,e,t){let r=e.map(a=>new I(a.shape,a.dtype)),s=n+"_jit_"+j3++,o=mp(s,r,[],(a,u)=>{let p;p=My(a,n,u,t),a.returnOp([p.getResult(0)])}),i=o.getReturnOp();return i&&i.operands.length>0&&(o.outputTypes=Object.freeze([i.operands[0].type])),o}function AM(n,e,t=null,r){let s=$3(n,e,t,r),o=vM.get(s);if(o)return o;let i=U3(n,e,t),a=new Mn(n+"_jit_mod");a.addFunction(i);let u=new iu;u.addPass(new ql),u.addPass(new au),u.addPass(new Co),u.run(a);let p=F3(r.name),c=new P3(r),d=typeof r.isGPU=="function"&&r.isGPU()?G3(i,r,c,p):null;d||(d=c.compile(NM(i,r)),p.addCompiledKernel(d));let h=i.getReturnOp(),m=h&&h.operands.length>0?h.operands[0].type.dtype:null;return o={funcName:d.name,runtime:p,numInputs:e.length,outDtype:m,compiled:d},vM.set(s,o),o}var vM,kM,NM,P3,j3,OM=M(()=>{"use strict";_s();Di();Re();$a();Ep();dy();by();Bx();Mx();Ex();du();Do();rs();sS();vM=new Map,kM=new Map,NM=Tm,P3=bo;l($3,"_cacheKey");l(F3,"_getRuntime");l(z3,"_bufferNumel");l(V3,"_trialLaunch");l(G3,"_compileScheduledGPU");j3=0;l(U3,"_buildGraphFunc");l(AM,"jitCompile")});function Iy(n){return n.slice(0,n.indexOf("(")).split(".")[0].trim()}function BM(){return Ey.map(n=>n.schema)}function Ly(n){return W3.get(n)||null}function K3(n){let e=n.indexOf("(")+1,t=n.lastIndexOf(")"),r=n.slice(e,t).trim();if(!r)return[];let s=[];for(let o of RM(r)){let i=o.trim(),a=i.lastIndexOf(" ");if(a<0)continue;let u=i.slice(0,a).trim();if(u==="Tensor"||u==="Tensor[]")continue;let p=i.slice(a+1).trim().replace(/=.*/,"").trim();s.push(p)}return s}function RM(n){let e=[],t=0,r=0;for(let s=0;s<n.length;s++){let o=n[s];o==="("||o==="["?t++:o===")"||o==="]"?t--:o===","&&t===0&&(e.push(n.slice(r,s)),r=s+1)}return e.push(n.slice(r)),e}function H3(n){let t=n.lastIndexOf("->"),r=t>=0?n.slice(t+2).trim():"Tensor";return r==="Tensor"?{kind:"tensor",arity:1}:r==="Tensor[]"?{kind:"tensor_list"}:r.startsWith("(")&&r.endsWith(")")?{kind:"tuple",arity:RM(r.slice(1,-1)).length}:{kind:"value",arity:1}}var Ey,q3,W3,Yee,Cy=M(()=>{"use strict";Ey=Object.freeze([{schema:"add(Tensor self, Tensor other) -> Tensor"},{schema:"sub(Tensor self, Tensor other) -> Tensor"},{schema:"mul(Tensor self, Tensor other) -> Tensor"},{schema:"div(Tensor self, Tensor other) -> Tensor"},{schema:"neg(Tensor self) -> Tensor"},{schema:"pow(Tensor self, Tensor exponent) -> Tensor"},{schema:"rem(Tensor self, Tensor other) -> Tensor"},{schema:"maximum(Tensor self, Tensor other) -> Tensor"},{schema:"minimum(Tensor self, Tensor other) -> Tensor"},{schema:"exp(Tensor self) -> Tensor"},{schema:"log(Tensor self) -> Tensor"},{schema:"sqrt(Tensor self) -> Tensor"},{schema:"rsqrt(Tensor self) -> Tensor"},{schema:"abs(Tensor self) -> Tensor"},{schema:"sin(Tensor self) -> Tensor"},{schema:"cos(Tensor self) -> Tensor"},{schema:"tanh(Tensor self) -> Tensor"},{schema:"erf(Tensor self) -> Tensor"},{schema:"erfc(Tensor self) -> Tensor"},{schema:"lgamma(Tensor self) -> Tensor"},{schema:"gamma(Tensor self) -> Tensor"},{schema:"sigmoid(Tensor self) -> Tensor"},{schema:"relu(Tensor self) -> Tensor"},{schema:"gelu(Tensor self) -> Tensor"},{schema:"silu(Tensor self) -> Tensor"},{schema:"sign(Tensor self) -> Tensor"},{schema:"floor(Tensor self) -> Tensor"},{schema:"ceil(Tensor self) -> Tensor"},{schema:"eq(Tensor self, Tensor other) -> Tensor"},{schema:"ne(Tensor self, Tensor other) -> Tensor"},{schema:"lt(Tensor self, Tensor other) -> Tensor"},{schema:"le(Tensor self, Tensor other) -> Tensor"},{schema:"gt(Tensor self, Tensor other) -> Tensor"},{schema:"ge(Tensor self, Tensor other) -> Tensor"},{schema:"where(Tensor condition, Tensor self, Tensor other) -> Tensor"},{schema:"clamp(Tensor self, Tensor min, Tensor max) -> Tensor"},{schema:"pad(Tensor self, Tensor value, int[] low, int[] high) -> Tensor"},{schema:"one_hot(Tensor indices, int depth) -> Tensor"},{schema:"index_select(Tensor self, Tensor index, int dim) -> Tensor"},{schema:"gather(Tensor self, Tensor index, int dim) -> Tensor"},{schema:"scatter_add(Tensor self, Tensor index, Tensor src, int dim) -> Tensor"},{schema:"scatter(Tensor self, int dim, Tensor index, Tensor src) -> Tensor"},{schema:"sum(Tensor self, int[] dim, bool keepdim) -> Tensor"},{schema:"mean(Tensor self, int[] dim, bool keepdim) -> Tensor"},{schema:"max(Tensor self, int[] dim, bool keepdim) -> Tensor"},{schema:"min(Tensor self, int[] dim, bool keepdim) -> Tensor"},{schema:"prod(Tensor self, int[] dim, bool keepdim) -> Tensor"},{schema:"argmax(Tensor self, int dim, bool keepdim) -> Tensor"},{schema:"argmin(Tensor self, int dim, bool keepdim) -> Tensor"},{schema:"matmul(Tensor self, Tensor other) -> Tensor"},{schema:"dot(Tensor self, Tensor other) -> Tensor"},{schema:"cat(Tensor[] tensors, int dim) -> Tensor"},{schema:"stack(Tensor[] tensors, int dim) -> Tensor"},{schema:"clone(Tensor self) -> Tensor"},{schema:"fill(Tensor self, Scalar value) -> Tensor"},{schema:"reshape(Tensor self, int[] shape) -> Tensor",ir:"reshape",scalarArgs:["shape"]},{schema:"transpose(Tensor self, int dim0, int dim1) -> Tensor",ir:"transpose"},{schema:"permute(Tensor self, int[] dims) -> Tensor",ir:"permute"},{schema:"broadcast_in_dim(Tensor self, int[] result_shape, int[] broadcast_dimensions) -> Tensor",ir:"broadcast_in_dim"},{schema:"expand(Tensor self, int[] shape) -> Tensor",ir:"expand",scalarArgs:["shape"]},{schema:"slice(Tensor self, int dim, int start, int end, int step) -> Tensor"},{schema:"unsqueeze(Tensor self, int dim) -> Tensor"},{schema:"squeeze(Tensor self, int dim) -> Tensor"},{schema:"narrow(Tensor self, int dim, int start, int length) -> Tensor"},{schema:"select(Tensor self, int dim, int index) -> Tensor"},{schema:"contiguous(Tensor self) -> Tensor"},{schema:"repeat(Tensor self, int[] reps) -> Tensor"},{schema:"tile(Tensor self, int[] reps) -> Tensor"},{schema:"split(Tensor self, int[] sizes, int dim) -> Tensor[]",returns:{kind:"tensor_list"}},{schema:"chunk(Tensor self, int chunks, int dim) -> Tensor[]",returns:{kind:"tensor_list"}},{schema:"roll(Tensor self, int shift, int dim) -> Tensor"},{schema:"flip(Tensor self, int[] dims) -> Tensor"},{schema:"cumsum(Tensor self, int dim) -> Tensor"},{schema:"sort(Tensor self, int dim, bool descending) -> Tensor"},{schema:"argsort(Tensor self, int dim, bool descending) -> Tensor"},{schema:"topk(Tensor self, int k, int dim, bool largest) -> (Tensor, Tensor)",returns:{kind:"tuple",arity:2}},{schema:"softmax(Tensor self, int dim) -> Tensor"},{schema:"log_softmax(Tensor self, int dim) -> Tensor"},{schema:"layer_norm(Tensor input, Tensor weight, Tensor bias, int axis, float eps) -> Tensor"},{schema:"batch_norm(Tensor input, Tensor weight, Tensor bias, Tensor mean, Tensor var, int axis, float eps) -> Tensor"},{schema:"conv2d(Tensor input, Tensor weight, int[] strides, int[] padding, int[] dilation, int groups) -> Tensor"},{schema:"pool2d(Tensor input, str pool_type, int[] kernel_size, int[] strides, int[] padding) -> Tensor"},{schema:"embedding(Tensor weight, Tensor indices) -> Tensor"}]),q3=new Map(Ey.filter(n=>n.scalarArgs).map(n=>[Iy(n.schema),n.scalarArgs])),W3=new Map(Ey.map(n=>[Iy(n.schema),q3.get(Iy(n.schema))||K3(n.schema)])),Yee=new Map(Ey.map(n=>{let e=Iy(n.schema);return[e,Object.freeze({...n,name:e,returns:n.returns||H3(n.schema)})]}));l(Iy,"opNameFromSchema");l(BM,"tensorOpSchemas");l(Ly,"scalarArgNames");l(K3,"scalarArgNamesFromSchema");l(RM,"splitTopLevel");l(H3,"returnsFromSchema")});var XR={};at(XR,{getCudnnGRU:()=>uS,getCudnnLSTM:()=>aS,getGpuAdamFn:()=>lS,getGpuContiguousFn:()=>ZT,getGpuMatmul:()=>zT,getWebgpuRNN:()=>Sf,gpuContiguousArray:()=>YT,registerJITKernels:()=>pS,setCudnnGRU:()=>ax,setCudnnLSTM:()=>ix,setGpuAdamFn:()=>ux,setGpuConcatFn:()=>ox,setGpuContiguousFn:()=>sx,setGpuMatmul:()=>lx,setWebGPUEagerFn:()=>eV,setWebgpuRNN:()=>oV,tensorToContiguous:()=>Qe,wrapResult:()=>Ee});function eV(n){IM=n}function oS(n){return typeof n=="object"&&n!==null&&"_impl"in n}function tV(n,e){let t=[],r={},s=Ly(n),o=0;for(let i=0;i<e.length;i++){let a=e[i];if(oS(a))t.push(a);else if(Array.isArray(a)&&a.length>0&&oS(a[0]))for(let u of a)oS(u)&&t.push(u);else a!=null&&(s&&o<s.length&&(r[s[o]]=a),o++)}return{tensors:t,scalars:r}}function hu(n,e){return typeof n=="number"?n:e}function tc(n,e){return Array.isArray(n)?n:e}function MM(n){return Array.isArray(n)?n:[[0,0],[0,0]]}function EM(n,e,t){if(e.length===0)return[];if(new Set(["sum","mean","max","min","prod","argmax","argmin"]).has(n)){let o=e[0].shape,i=t.dim,a=t.keepdim;if(i==null)return a?o.map(()=>1):[];let u=new Set((Array.isArray(i)?i:[i]).map(c=>typeof c=="number"&&c<0?o.length+c:Number(c))),p=[];for(let c=0;c<o.length;c++)u.has(c)?a&&p.push(1):p.push(o[c]);return p}if(n==="matmul")return xg(e[0].shape,e[1].shape)||[];if(n==="dot")return[];if(n==="transpose"){let o=[...e[0].shape],i=hu(t.dim0,0),a=hu(t.dim1,1),u=o[i];return o[i]=o[a],o[a]=u,o}if(n==="conv2d"){let o=e[0].shape,i=e[1].shape,a=tc(t.strides,[1,1]),u=MM(t.padding),p=tc(t.dilation,[1,1]),c=[];for(let f=0;f<2;f++){let d=u[f][0]+u[f][1],h=(i[f+2]-1)*p[f]+1;c.push(Math.floor((o[f+2]+d-h)/a[f])+1)}return[o[0],i[0],...c]}if(n==="pool2d"){let o=e[0].shape,i=tc(t.kernel_size,[2,2]),a=tc(t.strides,i),u=MM(t.padding),p=[];for(let c=0;c<2;c++){let f=u[c][0]+u[c][1];p.push(Math.floor((o[c+2]+f-i[c])/a[c])+1)}return[o[0],o[1],...p]}if(n==="clamp"){let o=[...e[0].shape];for(let i=1;i<e.length;i++)o=Rl(o,e[i].shape)||o;return o}if(n==="pad"){let o=e[0].shape,i=tc(t.low,[]),a=tc(t.high,[]);return o.map((u,p)=>u+(i[p]||0)+(a[p]||0))}if(n==="one_hot")return[...e[0].shape,hu(t.depth,0)];if(n==="cat"){let o=e[0].shape.length,i=hu(t.dim,0),a=i<0?o+i:i,u=[...e[0].shape];return u[a]=e.reduce((p,c)=>p+c.shape[a],0),u}if(n==="stack"){let o=e[0].shape.length,i=hu(t.dim,0),a=i<0?o+1+i:i,u=[...e[0].shape];return u.splice(a,0,e.length),u}if(n==="index_select"){let o=[...e[0].shape],i=o.length,a=hu(t.dim,0),u=a<0?i+a:a;return o[u]=e[1].shape.reduce((p,c)=>p*c,1),o}if(n==="gather")return[...e[1].shape];if(n==="scatter_add")return[...e[0].shape];if(n==="softmax"||n==="log_softmax")return[...e[0].shape];if(n==="layer_norm"||n==="batch_norm")return[...e[0].shape];if(n==="embedding"){let o=e[0].shape;return[...e[1].shape,o[1]]}return e.length===1?[...e[0].shape]:Rl(e[0].shape,e[1].shape)||[...e[0].shape]}function Qe(n){let e=n._impl.storage.data,t=n._impl.storageOffset,r=n.numel;if(n.isContiguous&&t===0&&e.length===r)return n.data||e;let s=n.shape,o=n.strides,i=e.constructor,a=new i(r),u=a,p=s.length,c=new Int32Array(p),f=t;for(let d=0;d<r;d++){u[d]=e[f];for(let h=p-1;h>=0;h--){if(c[h]++,c[h]<s[h]){f+=o[h];break}f-=(s[h]-1)*o[h],c[h]=0}}return a}function Ee(n,e,t,r){let s=En(e),o=Wn.fromData(n,r),i=new Pt(o,0,e,s,t,r);return new Ze(i)}function Dy(n,e,t){let r=1;for(let s=e;s<t;s++)r*=n[s];return r}function nV(n,e,t,r){let s=e.length,o=t<0?s+1+t:t,i=Dy(e,0,o),a=Dy(e,o,s),u=n.length;for(let p=0;p<i;p++)for(let c=0;c<u;c++)r.set(n[c].subarray(p*a,(p+1)*a),(p*u+c)*a)}function rV(n,e,t,r){let s=e[0].length,o=t<0?s+t:t,i=Dy(e[0],0,o),a=Dy(e[0],o+1,s),u=0;for(let p of e)u+=p[o];for(let p=0;p<i;p++){let c=p*u*a;for(let f=0;f<n.length;f++){let d=e[f][o]*a;r.set(n[f].subarray(p*d,(p+1)*d),c),c+=d}}}function sV(n,e,t){let r=EM(n,e,t),s=e[0].dtype,o=St(s),i=new o(Math.max(bn(r),1)),a=hu(t.dim,0);if(In()&&iS){let p=e.map(c=>cS(c));return iS(n,p,e.map(c=>[...c.shape]),a,r,i,s),Ee(i,r,s,e[0].device)}let u=e.map(p=>Qe(p));return n==="stack"?nV(u,e[0].shape,a,i):rV(u,e.map(p=>p.shape),a,i),Ee(i,r,s,e[0].device)}function sx(n){Py=n}function ZT(){return Py}function ox(n){iS=n}function ix(n){LM=n}function aS(){return LM}function ax(n){CM=n}function uS(){return CM}function oV(n){DM=n}function Sf(){return DM}function ux(n){PM=n}function lS(){return PM}function lx(n){$M=n}function zT(){return $M}function cS(n){if(n.isContiguous&&n._impl.storageOffset===0){let e=n._impl.storage.rawData;if(e&&e.length===n.numel)return e}return In()&&Py?Py(n._impl.storage.rawData,n.shape,n.strides,n._impl.storageOffset,n.dtype):Qe(n)}function YT(n){return cS(n)}function iV(n,e){let t=X3[e];if(!t)return null;let r=e===1,s=e===5,o=r&&(n==="stack"||n==="cat");return(i,...a)=>{let{tensors:u,scalars:p}=tV(n,a);if(u.length===0)throw new Error(`JIT dispatch: no tensor args for op '${n}'`);if(o)return sV(n,u,p);let c=t(),f=AM(n,u,p,c),d=EM(n,u,p),h=f.outDtype||_g(u[0].dtype,u.length>1?u[1].dtype:u[0].dtype),m=bn(d),g=St(h),y=new g(Math.max(m,1));if(s)return IM(f.compiled,u,y),Ee(y,d,h,u[0].device);let _=u.map(b=>r?cS(b):Qe(b));return _.push(y),f.runtime.run(f.funcName,..._),Ee(y,d,h,u[0].device)}}function pS(){let n=Lt.listOps(),e=[0,1,2,5];for(let t of n){let r=Lt.findOp(t);if(!r)continue;let s=r.name;for(let o of e){if(r.entry.hasKernel(o))continue;let i=iV(s,o);i&&r.entry.registerKernel(o,Mr.fromUnboxed(i))}}}var Y3,J3,Z3,Q3,X3,IM,Py,iS,LM,CM,DM,PM,$M,wn=M(()=>{"use strict";An();sp();Ma();OM();fi();yo();sr();di();Xa();ss();rs();Cy();X3={0:()=>Y3??=go(),1:()=>J3??=xm(),2:()=>Z3??=Sm(),5:()=>Q3??=vm()},IM=null;l(eV,"setWebGPUEagerFn");l(oS,"hasTensorImpl");l(tV,"_extractTensorsAndScalars");l(hu,"numberScalar");l(tc,"numberArrayScalar");l(MM,"paddingScalar");l(EM,"_inferOutputShape");l(Qe,"tensorToContiguous");l(Ee,"wrapResult");l(Dy,"_prod");l(nV,"_hostStack");l(rV,"_hostCat");l(sV,"_runHostConcatLike");Py=null;l(sx,"setGpuContiguousFn");l(ZT,"getGpuContiguousFn");iS=null;l(ox,"setGpuConcatFn");LM=null;l(ix,"setCudnnLSTM");l(aS,"getCudnnLSTM");CM=null;l(ax,"setCudnnGRU");l(uS,"getCudnnGRU");DM=null;l(oV,"setWebgpuRNN");l(Sf,"getWebgpuRNN");PM=null;l(ux,"setGpuAdamFn");l(lS,"getGpuAdamFn");$M=null;l(lx,"setGpuMatmul");l(zT,"getGpuMatmul");l(cS,"_gpuInputArray");l(YT,"gpuContiguousArray");l(iV,"_wrapOpForJIT");l(pS,"registerJITKernels")});var CG,bt,yu=M(()=>{"use strict";sr();di();Xa();An();qn();ss();CG=Jr.fromKey(48),bt=class extends Ze{static{l(this,"SymbolicTensor")}_irValue;_tracer;_symbolicShape;constructor(e,t,r,s,o){let i=En(t),a=Wn.allocate(0,r,Bl),u=new Pt(a,0,t,i,r,Bl);super(u),this._irValue=e,this._tracer=s,this._symbolicShape=o}get irValue(){return this._irValue}get tracer(){return this._tracer}get symbolicShape(){return this._symbolicShape}get dispatchKeySet(){return super.dispatchKeySet.union(CG)}get isSymbolic(){return!0}}});function DG(n){let e=[...n.entries()];return e.sort((t,r)=>t[0]<r[0]?-1:t[0]>r[0]?1:0),e}function b_(n){return n==null?"null":typeof n=="number"?Number.isFinite(n)?String(n):n===1/0?"inf":n===-1/0?"-inf":"nan":typeof n=="boolean"?String(n):typeof n=="string"?`"${n}"`:n instanceof I?Gn(n):Array.isArray(n)?n.length===0?"[]":Array.isArray(n[0])?"["+n.map(e=>b_(e)).join(", ")+"]":"["+n.map(e=>b_(e)).join(", ")+"]":typeof n=="object"&&n.constructor===Object?"{"+Object.entries(n).sort((t,r)=>t[0].localeCompare(r[0])).map(([t,r])=>`${t}: ${b_(r)}`).join(", ")+"}":String(n)}function PG(n){return new Ef().printModule(n)}var Ef,RS=M(()=>{"use strict";Re();Ef=class{static{l(this,"IRPrinter")}constructor(e={}){this.indent=0,this.indentStr=e.indentStr||"  ",this.valueNames=new Map,this._nextValueId=0}printModule(e){let t=[];t.push(`module @${e.name} {`),this.indent++;let r=!0;for(let s of e)r||t.push(""),r=!1,this.printFunction(s,t);return this.indent--,t.push("}"),t.join(`
+`)}printFunction(e,t=null){let r=t===null;r&&(t=[]),this.valueNames.clear(),this._nextValueId=0;let s=e.args.map((i,a)=>`${this._nameValue(i)}: ${Gn(i.type)}`),o=e.outputTypes.map(i=>Gn(i)).join(", ");t.push(`${this._indentPrefix()}func @${e.name}(${s.join(", ")}) -> (${o}) {`),this.indent++;for(let i of e.body)this.printBlock(i,t,i===e.entryBlock);return this.indent--,t.push(`${this._indentPrefix()}}`),r?t.join(`
+`):void 0}printBlock(e,t,r=!1){if(!r&&e.arguments.length>0){let s=e.arguments.map(o=>`${this._nameValue(o)}: ${Gn(o.type)}`);t.push(`${this._indentPrefix()}^bb(${s.join(", ")}):`)}for(let s of e)this.printOperation(s,t)}printOperation(e,t=null){let r=t===null;r&&(t=[]);let s=this._indentPrefix(),o=s;if(e.numResults>0){let i=[];for(let a=0;a<e.numResults;a++)i.push(this._nameValue(e.getResult(a)));o+=i.join(", ")+" = "}if(o+=e.opName,e.numOperands>0){let i=[];for(let a=0;a<e.numOperands;a++)i.push(this._valueName(e.getOperand(a)));o+="("+i.join(", ")+")"}else o+="()";if(e.attributes.size>0){let i=[];for(let[a,u]of DG(e.attributes))i.push(`${a} = ${b_(u)}`);o+=" {"+i.join(", ")+"}"}if(e.numResults>0){let i=[];for(let a=0;a<e.numResults;a++)i.push(Gn(e.getResult(a).type));o+=" : "+i.join(", ")}t.push(o);for(let i=0;i<e.regions.length;i++){let a=e.regions[i];t.push(`${s}{`),this.indent++;for(let u of a){if(u.arguments.length>0){let p=u.arguments.map(c=>`${this._nameValue(c)}: ${Gn(c.type)}`);t.push(`${this._indentPrefix()}^bb(${p.join(", ")}):`)}for(let p of u)this.printOperation(p,t)}this.indent--,t.push(`${s}}`)}return r?t.join(`
+`):void 0}_nameValue(e){if(this.valueNames.has(e))return this.valueNames.get(e);let t=`%${this._nextValueId++}`;return this.valueNames.set(e,t),t}_valueName(e){return this.valueNames.has(e)?this.valueNames.get(e):"%?"}_indentPrefix(){let e="";for(let t=0;t<this.indent;t++)e+=this.indentStr;return e}};l(DG,"sortedEntries");l(b_,"formatAttrValue");l(PG,"printModule")});var n6,R_,cE=M(()=>{"use strict";Re();po();n6={eq:l((n,e)=>n===e,"eq"),ne:l((n,e)=>n!==e,"ne"),gt:l((n,e)=>n>e,"gt"),ge:l((n,e)=>n>=e,"ge"),lt:l((n,e)=>n<e,"lt"),le:l((n,e)=>n<=e,"le")},R_=class{static{l(this,"ShapeEnv")}_symbols;_guards;_bindings;_nextId;constructor(){this._symbols=new Map,this._guards=[],this._bindings=new Map,this._nextId=0}allocate(e,t,r){let s=`s${this._nextId++}`;return this._symbols.set(s,{hint:r,inputIdx:e,dimIdx:t}),s}produceShapeSpec(e,t,r){let s=new Array(t.length),o=new Array(t.length);for(let i=0;i<t.length;i++)if(r&&r.has(i)){let a=this.allocate(e,i,t[i]);s[i]=X,o[i]=a}else{let a=this.allocate(e,i,t[i]);this.guardRelation(a,"eq",t[i]),s[i]=t[i],o[i]=t[i]}return{irShape:s,symShape:o}}guardRelation(e,t,r){this._guards.push({lhs:e,op:t,rhs:r})}guardDivisible(e,t){this._guards.push({type:"divisible",sym:e,divisor:t})}bindInputShapes(e){this._bindings.clear();for(let[t,r]of this._symbols)this._bindings.set(t,e[r.inputIdx].shape[r.dimIdx])}evaluateGuards(){for(let e=0;e<this._guards.length;e++){let t=this._guards[e];if("type"in t){if(this._resolve(t.sym)%t.divisor!==0)return{passed:!1,failedGuard:t};continue}let r=this._resolve(t.lhs),s=this._resolve(t.rhs);if(!n6[t.op](r,s))return{passed:!1,failedGuard:t}}return{passed:!0,failedGuard:null}}resolveSymbolicShape(e){let t=new Array(e.length);for(let r=0;r<e.length;r++)t[r]=this._resolve(e[r]);return t}_resolve(e){return typeof e=="number"?e:typeof e=="string"?this._bindings.get(e):e instanceof Pe?Pe.evaluate(e,this._bindings):e}get symbols(){return this._symbols}get guards(){return this._guards}get bindings(){return this._bindings}}});function As(){return M_}var M_,I_,_u=M(()=>{"use strict";_s();Di();Us();Re();Ct();yu();cE();sS();M_=null;l(As,"getActiveTracer");I_=class{static{l(this,"Tracer")}_name;_shapeEnv;_inputTypes;_inputSymShapes;_outputTypes;_outputSymShapes;_inputs;_func;_builder;_module;_capturedParams;_capturedParamOrder;constructor(e){this._name=e||"traced",this._shapeEnv=new R_,this._inputTypes=[],this._inputSymShapes=[],this._outputTypes=[],this._outputSymShapes=[],this._inputs=[],this._func=null,this._builder=null,this._module=null,this._capturedParams=new Map,this._capturedParamOrder=[]}get shapeEnv(){return this._shapeEnv}createInput(e,t,r){let s=this._inputTypes.length,{irShape:o,symShape:i}=this._shapeEnv.produceShapeSpec(s,e,r);if(r)for(let u=0;u<i.length;u++)typeof i[u]=="string"&&this._shapeEnv.guardRelation(i[u],"gt",0);let a=new I(o,t);return this._inputTypes.push(a),this._inputSymShapes.push(i),{shape:o,dtype:t,tensorType:a}}_initGraph(){this._func=new yn(this._name,this._inputTypes,[]),this._func.inputTypes=[...this._func.inputTypes],this._builder=new $n(this._func),this._module=new Mn(this._name);let e=[],t=this._func.args;for(let r=0;r<t.length;r++){let s=t[r],o=this._inputTypes[r];s.symbolicShape=this._inputSymShapes[r];let i=new bt(s,o.shape,o.dtype,this,this._inputSymShapes[r]);e.push(i)}return this._inputs=e,e}recordOp(e,t,r){let s=[];for(let u of t)u instanceof bt&&s.push(u.irValue);let o=this._requireBuilder(),i=My(o,e,s,r),a=[];for(let u=0;u<i.numResults;u++){let p=i.getResult(u),c=p.type,f=this._propagateSymbolicShape(e,i,t,c,u);p.symbolicShape=f,a.push(new bt(p,c.shape,c.dtype,this,f))}return a.length===1?a[0]:a}_propagateSymbolicShape(e,t,r,s,o=0){let i=r.filter(c=>c instanceof bt),a=ne.get(t.opName||e);if(a&&a.propagateSymbolicShapes){let c=new Map;for(let d of i)c.set(d.irValue,d.symbolicShape);let f=a.propagateSymbolicShapes(t,c);if(f&&f[o])return f[o]}let u=s.shape,p=new Array(u.length);for(let c=0;c<u.length;c++){if(u[c]!==X){p[c]=u[c];continue}let f=null;for(let d of i){let h=d.symbolicShape;if(!h)continue;let m=u.length-h.length,g=c-m;if(g>=0&&g<h.length&&typeof h[g]=="string"){f=h[g];break}}p[c]=f!==null?f:X}return p}scan(e,t,r){let s=l(d=>d instanceof bt?d.irValue:this.captureConstant(d).irValue,"toIr"),o=e.map(s),i=t.map(s),u=this._requireBuilder().scanOp(o,i,(d,h,m)=>{let g=this._requireBuilder();this._builder=d;try{let y=l(w=>new bt(w,w.type.shape,w.type.dtype,this,[...w.type.shape]),"wrap"),[_,b]=r(m.map(y),h.map(y));return[_.map(w=>w.irValue),b.map(w=>w.irValue)]}finally{this._builder=g}}),p=t.length,c=[],f=[];for(let d=0;d<u.numResults;d++){let h=u.getResult(d),m=new bt(h,h.type.shape,h.type.dtype,this,[...h.type.shape]);d<p?c.push(m):f.push(m)}return[c,f]}captureConstant(e){let t=this._capturedParams.get(e);if(t)return t;if(e.shape.length===0&&e.data){let u=e.data[0],c=this._requireBuilder().scalarConstant(u,e.dtype).getResult(0),f=new bt(c,[],e.dtype,this,[]);return this._capturedParams.set(e,f),f}let r=new I(e.shape,e.dtype),s=this._requireFunc();s.inputTypes.push(r);let i=s.entryBlock.addArgument(r),a=new bt(i,e.shape,e.dtype,this,[...e.shape]);return this._capturedParams.set(e,a),this._capturedParamOrder.push(e),a}get capturedParams(){return this._capturedParamOrder}markOutput(e){e instanceof bt&&(this._requireBuilder().returnOp([e.irValue]),this._outputSymShapes=[e.symbolicShape]),this._outputTypes=[new I(e.shape,e.dtype)]}markOutputs(e){let t=e.map(r=>r.irValue);this._requireBuilder().returnOp(t),this._outputTypes=e.map(r=>new I(r.shape,r.dtype)),this._outputSymShapes=e.map(r=>r.symbolicShape)}get outputSymShapes(){return this._outputSymShapes}getGraphModule(){let e=this._requireFunc(),t=this._requireModule();return e.outputTypes=Object.freeze(this._outputTypes),Object.isFrozen(e.inputTypes)||(e.inputTypes=Object.freeze(e.inputTypes)),t.addFunction(e),t}activate(){M_=this}deactivate(){M_===this&&(M_=null)}_requireBuilder(){if(!this._builder)throw new Error("Tracer graph has not been initialized");return this._builder}_requireFunc(){if(!this._func)throw new Error("Tracer graph has not been initialized");return this._func}_requireModule(){if(!this._module)throw new Error("Tracer graph has not been initialized");return this._module}}});function NE(){return[...kE]}function z_(n,e,t,r=kE){return r.filter(s=>s.phase===n).sort((s,o)=>s.priority-o.priority).map(s=>s.factory(e,t)).filter(Boolean)}var kE,ev=M(()=>{"use strict";kE=[];l(NE,"snapshotGraphPasses");l(z_,"graphPassesForPhase")});function AE(n){return n instanceof Map?n:n&&typeof n=="object"?new Map(Object.entries(n)):new Map}var V_,OE=M(()=>{"use strict";ev();l(AE,"toMap");V_=class{static{l(this,"CompilerContext")}constructor({loweringRules:e=null,codegenEntries:t=null,graphPasses:r=null}={}){this.loweringRules=AE(e),this.codegenEntries=AE(t),this.graphPasses=r||NE()}get hasOverrides(){return this.loweringRules.size>0||this.codegenEntries.size>0}getLoweringRule(e){return this.loweringRules.get(e)||null}getCodegenEntry(e){return this.codegenEntries.get(e)||null}passesForPhase(e,t,r){return z_(e,t,r,this.graphPasses)}}});function G_(n){return new tv().print(n)}var tv,nv=M(()=>{"use strict";tr();tv=class{static{l(this,"TensorIRPrinter")}constructor(){this.indent=0,this.out=[]}print(e){return this.out=[],this.visit(e),this.out.join("")}push(e){this.out.push(e)}newline(){this.out.push(`
+`+"  ".repeat(this.indent))}visit(e){if(!e)return;let t="visit"+e.type;this[t]?this[t](e):this.push(`[UnknownNode: ${e.type}]`)}visitPrimFunc(e){this.push(`prim_func ${e.name}(${e.params.map(t=>t.name).join(", ")}) {`),this.indent++,this.newline();for(let[t,r]of e.bufferMap)this.push(`${r.name} = buffer_map(${t.name}, shape=[${r.shape.join(",")}], dtype=${r.dtype})`),this.newline();this.visit(e.body),this.indent--,this.newline(),this.push("}")}visitSeqNode(e){for(let t=0;t<e.stmts.length;t++)this.visit(e.stmts[t]),t<e.stmts.length-1&&this.newline()}visitForNode(e){let t=e.kind==="serial"?"":`@${e.kind} `,r=e.threadTag?`[${e.threadTag}] `:"";this.push(`for ${e.loopVar.name} in 0..`),this.visit(e.extent),this.push(` ${t}${r}{`),this.indent++,this.newline(),this.visit(e.body),this.indent--,this.newline(),this.push("}")}visitBlockNode(e){this.push(`block ${e.name} {`),this.indent++,this.newline();for(let t of e.iterVars)this.push(`bind ${t.iterVar.name} = `),this.visit(t.binding),this.newline();e.reads.length>0&&(this.push("reads(["),this.push(e.reads.map(t=>`${t.buffer.name}[...]`).join(", ")),this.push("])"),this.newline()),e.writes.length>0&&(this.push("writes(["),this.push(e.writes.map(t=>`${t.buffer.name}[...]`).join(", ")),this.push("])"),this.newline()),e.initBody&&(this.push("init {"),this.indent++,this.newline(),this.visit(e.initBody),this.indent--,this.newline(),this.push("}"),this.newline()),this.visit(e.body),this.indent--,this.newline(),this.push("}")}visitBufferStoreNode(e){this.push(`${e.buffer.name}[`);for(let t=0;t<e.indices.length;t++)this.visit(e.indices[t]),t<e.indices.length-1&&this.push(", ");this.push("] = "),this.visit(e.value)}visitBufferLoadNode(e){this.push(`${e.buffer.name}[`);for(let t=0;t<e.indices.length;t++)this.visit(e.indices[t]),t<e.indices.length-1&&this.push(", ");this.push("]")}visitIfThenElseNode(e){this.push("if ("),this.visit(e.condition),this.push(") {"),this.indent++,this.newline(),this.visit(e.thenBody),this.indent--,this.newline(),this.push("}"),e.elseBody&&(this.push(" else {"),this.indent++,this.newline(),this.visit(e.elseBody),this.indent--,this.newline(),this.push("}"))}visitLetStmtNode(e){this.push(`let ${e.variable.name} = `),this.visit(e.value),this.newline(),this.visit(e.body)}visitAllocateNode(e){this.push(`allocate ${e.buffer.name}[${e.buffer.shape.join(", ")}] (${e.scope}) {`),this.indent++,this.newline(),this.visit(e.body),this.indent--,this.newline(),this.push("}")}visitEvaluateNode(e){this.push("evaluate "),this.visit(e.value)}visitMathOpNode(e){this.push("("),this.visit(e.a),e.b&&(this.push(` ${e.op} `),this.visit(e.b)),this.push(")")}visitCompareNode(e){this.push("("),this.visit(e.a),this.push(` ${Ku(e.direction)} `),this.visit(e.b),this.push(")")}visitCallExternNode(e){this.push(`${e.externName}(`);for(let t=0;t<e.args.length;t++)this.visit(e.args[t]),t<e.args.length-1&&this.push(", ");this.push(")")}visitVariableNode(e){this.push(e.name)}visitIntImmNode(e){this.push(e.value.toString())}visitFloatImmNode(e){this.push(e.value.toString())}visitCastNode(e){this.push(`cast<${e.toDtype}>(`),this.visit(e.expr),this.push(")")}};l(G_,"printTensorIR")});var j_,BE=M(()=>{"use strict";nn();nv();j_=class{static{l(this,"TirPassManager")}constructor(){this.passes=[],this.trace=null,this.verifyHook=null}addPass(e){this.passes.push(e)}setTrace(e){this.trace=e}setVerifyHook(e){this.verifyHook=e}run(e,t){for(let r of this.passes)this._runPass(r,e,t);return e}_runPass(e,t,r){let s=r.trace;e.trace=s,s.phaseStart(e.phase);let o=performance.now();e.begin(r);for(let i=0;i<t.length;i++){let a=t[i];if(!r.failed.has(a.name))try{let u=e.run(a,r);u&&u!==a&&(t[i]=u)}catch(u){if(r.errors.push(new Yn(e.phase,a.name,u.message)),r.failed.add(a.name),s.errorEvent(e.phase,a.name,u.message),!r.resilient)break}}if(e.end(r),s.phaseEnd(e.phase,performance.now()-o),e.snapshotPoint&&s.shouldSnapshot(e.snapshotPoint))for(let i of t)r.failed.has(i.name)||s.irDump(e.snapshotPoint+":"+i.name,G_(i));this.verifyHook&&this._verifyFuncs(t,r),e.trace=null}_verifyFuncs(e,t){for(let r of e){if(t.failed.has(r.name))continue;let s=this.verifyHook(r);if(s&&s.length>0){let o=s.join("; ");if(t.resilient)t.errors.push(new Yn("verification",r.name,o)),t.failed.add(r.name);else throw new Error("TensorIR verification failed for "+r.name+": "+o)}}}}});function RE(n){let e=new Wu;return e.add(new Zu),e.add(new Qu),e.add(new Xu(n)),e.add(new el),e.add(new tl(n)),e.add(new nl),e.add(new rl),e.add(new Gh),e.add(new Ju),e.add(new Hh),e.add(new Yh),e.add(new Jh),e.add(new Zh),n&&(e.add(new Kh(n)),e.add(new qh(n)),e.add(new Wh(n))),e}var N6,A6,U_,ME=M(()=>{"use strict";fn();cp();Rx();sl();l(RE,"buildAlgebraicPatterns");N6=RE(!1),A6=RE(!0),U_=class extends qe{static{l(this,"AlgebraicSimplificationPass")}constructor(e={}){super("algebraic_simplify"),this.preservedAnalyses=new Set,this.patterns=e.fastMath?A6:N6}run(e,t){return new Wl(this.patterns).applyPatterns(e,10,this.trace)}}});function IE(n){let e=ne.get(n);return!!(e&&e.isConstant)}function O6(n,e){return!d0(e)||typeof n!="number"?!0:Number.isInteger(n)&&Number.isSafeInteger(n)}function EE(n,e,t){if(t.has(n))return t.get(n);let r=B6(n,e,t);return t.set(n,r),r}function B6(n,e,t){let r=n.definingOp;if(!r)return;if(IE(r.opName))return r.getAttr("value");if(e.has(r))return;e.add(r);let s=ne.get(r.opName);if(!s||!s.fold||s.hasSideEffects||r.regions.length>0)return;let o=new Array(r.numOperands),i=new Array(r.numOperands);for(let a=0;a<r.numOperands;a++){let u=EE(r.getOperand(a),e,t);if(u===void 0)return;o[a]=u,i[a]=r.getOperand(a).definingOp}try{return s.fold(o,r.attributes,i)}catch{return}}var q_,LE=M(()=>{"use strict";fn();_s();Ct();nn();Re();l(IE,"isConstantProducer");l(O6,"isFoldResultRepresentable");l(EE,"resolveConstantValue");l(B6,"computeConstantValue");q_=class extends qe{static{l(this,"ConstantFoldPass")}constructor(){super("constant_fold")}run(e,t){let r=!1,s=0,o=new $n(e),i=new Map;for(let a of[...e.opsRecursive()]){if(!a.parentBlock||IE(a.opName))continue;let u=ne.get(a.opName);if(!u||a.regions.length>0||u.hasSideEffects||u.getMemoryEffects&&u.getMemoryEffects(a).length>0||!u.fold||a.numOperands===0)continue;let p=new Array(a.numOperands),c=new Array(a.numOperands),f=!0;for(let d=0;d<a.numOperands;d++){let h=EE(a.getOperand(d),new Set,i);if(h===void 0){f=!1;break}p[d]=h,c[d]=a.getOperand(d).definingOp}if(f)try{let d=u.fold(p,a.attributes,c);if(d===void 0||!O6(d,a.getResult(0).type.dtype))continue;o.block=a.parentBlock,o.setInsertionPoint(a);let h=o.constant(d,a.getResult(0).type);a.replaceAllResultsWith([h.getResult(0)]),a.erase(),r=!0,s++}catch(d){this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,foldError:a.opName,message:d.message,level:le.DEBUG})}}return this.trace&&this.trace.level>=le.DEBUG&&s>0&&this.trace.emit({type:"pass_detail",passName:this.name,foldedCount:s,level:le.DEBUG}),r?oe.CHANGED:oe.UNCHANGED}}});var W_,CE=M(()=>{"use strict";fn();Ct();nn();W_=class extends qe{static{l(this,"CSEPass")}constructor(){super("cse"),this.preservedAnalyses=new Set}run(e,t){let r=!1,s=0,o=typeof e.blocksRecursive=="function"?[...e.blocksRecursive()]:[...e.body];for(let i of o){let a=new Map;for(let u of[...i.ops()]){if(!u.parentBlock||u.regions&&u.regions.length>0)continue;let p=ne.get(u.opName);if(p&&p.hasSideEffects||p&&p.getMemoryEffects&&p.getMemoryEffects(u).length>0)continue;let c=u.structuralHash();if(!a.has(c)){a.set(c,[u]);continue}let f=a.get(c),d=!1;for(let h of f)if(h.parentBlock&&h.structuralEquals(u)){let m=[];for(let g=0;g<h.numResults;g++)m.push(h.getResult(g));u.replaceAllResultsWith(m),u.erase(),r=!0,s++,d=!0;break}d||f.push(u)}}return this.trace&&this.trace.level>=le.DEBUG&&s>0&&this.trace.emit({type:"pass_detail",passName:this.name,eliminated:s,level:le.DEBUG}),r?oe.CHANGED:oe.UNCHANGED}}});function sv(n,e){let t=rv[n],r=rv[e];return t===void 0||r===void 0?!1:n===We.REDUCTION?e===We.ELEMENTWISE:e===We.REDUCTION?t<=rv[We.INJECTIVE]:!0}function Tc(n){let e=!1,t=!1,r=!1,s=!1;for(let o of n){let i=ne.get(o.opName);if(!i||i.isOpaque){r=!0;continue}i.isReduction?e=!0:i.isInjective?t=!0:i.isBroadcast?s=!0:i.isElementwise||(r=!0)}return r?We.OPAQUE:e?We.REDUCTION:t?We.INJECTIVE:s?We.BROADCAST:We.ELEMENTWISE}function Ou(n){let e=ne.get(n.opName);return!e||e.isOpaque?We.OPAQUE:e.isReduction?We.REDUCTION:e.isInjective?We.INJECTIVE:e.isBroadcast?We.BROADCAST:e.isElementwise?We.ELEMENTWISE:We.OPAQUE}function DE(n,e){if(n.length!==e.length)return!1;for(let t=0;t<n.length;t++)if(!(n[t]===X||e[t]===X)&&n[t]!==e[t])return!1;return!0}var We,rv,bc,wc,Bu=M(()=>{"use strict";Re();Ct();$a();Ir();We=Object.freeze({ELEMENTWISE:"kElementwise",BROADCAST:"kBroadcast",REDUCTION:"kReduction",INJECTIVE:"kInjective",OPAQUE:"kOpaque",HORIZONTAL:"kHorizontal"}),rv={[We.ELEMENTWISE]:0,[We.BROADCAST]:1,[We.INJECTIVE]:2,[We.REDUCTION]:3};l(sv,"canFusePatterns");l(Tc,"classifyFusionKind");l(Ou,"classifyOpPattern");bc=Object.freeze({legal:!0}),wc=class{static{l(this,"FusionLegality")}constructor(e={}){this.maxFusionSize=e.maxFusionSize||512,this.maxSharedMemory=e.maxSharedMemory||49152,this.libraryOps=e.libraryOps||new Set,this.allowReductionFusion=e.allowReductionFusion!==!1,this._lowerableCache=new Map}isOpLowerable(e){let t=this._lowerableCache.get(e);return t===void 0&&(t=Xt(e)||ni(e)||D0(e),this._lowerableCache.set(e,t)),t}canFuse(e,t){if(!e||!t)return{legal:!1,reason:"null op"};if(e===t)return{legal:!1,reason:"same op"};if(e.regions.length>0&&e.opName!=="fusion"&&e.opName!=="reduce")return{legal:!1,reason:"producer has control flow regions"};if(t.regions.length>0&&t.opName!=="fusion"&&t.opName!=="reduce")return{legal:!1,reason:"consumer has control flow regions"};let r=ne.get(e.opName),s=ne.get(t.opName);if(!r)return{legal:!1,reason:"unknown producer op"};if(!s)return{legal:!1,reason:"unknown consumer op"};if(!this.isOpLowerable(e.opName))return{legal:!1,reason:`producer op '${e.opName}' has no lowering rule`};if(!this.isOpLowerable(t.opName))return{legal:!1,reason:`consumer op '${t.opName}' has no lowering rule`};if(r.isOpaque)return{legal:!1,reason:"producer is opaque (use EpilogueFusionPass for dot/conv epilogues)"};if(s.isOpaque)return{legal:!1,reason:"consumer is opaque"};let o=Ou(e),i=Ou(t);return sv(o,i)?(o===We.REDUCTION||i===We.REDUCTION)&&!this.allowReductionFusion?{legal:!1,reason:"reduction fusion disabled by target"}:o===We.ELEMENTWISE&&i===We.ELEMENTWISE?this._checkElementwisePair(e,t):(o===We.BROADCAST||o===We.REDUCTION)&&i===We.ELEMENTWISE?bc:this._checkProducerConsumerShapes(e,t):{legal:!1,reason:`cannot fuse pattern ${o} -> ${i}`}}canMergeGroups(e,t){if(e.size+t.size>this.maxFusionSize)return{legal:!1,reason:"merged group exceeds max fusion size"};let r=0;for(let s of e.ops){let o=ne.get(s.opName);if(o&&o.isReduction&&r++,o&&o.isOpaque)return{legal:!1,reason:"opaque op in merge (use EpilogueFusionPass)"}}for(let s of t.ops){let o=ne.get(s.opName);if(o&&o.isReduction&&r++,o&&o.isOpaque)return{legal:!1,reason:"opaque op in merge (use EpilogueFusionPass)"}}return r>1?{legal:!1,reason:"merged group would contain multiple reductions"}:bc}_checkElementwisePair(e,t){let r=this._getOutputShape(e),s=this._getOutputShape(t);return r&&s&&!DE(r,s)?{legal:!1,reason:`elementwise shape mismatch: [${r}] vs [${s}]`}:bc}_checkProducerConsumerShapes(e,t){let r=this._getOutputShape(e);if(!r)return bc;for(let s=0;s<t.numOperands;s++)if(t.getOperand(s).definingOp===e){let o=t.getOperand(s).type;if(o instanceof I)return DE(r,o.shape)?bc:{legal:!1,reason:`shape mismatch on data edge: [${r}] vs [${o.shape}]`}}return bc}_getOutputShape(e){for(let t=0;t<e.numResults;t++){let r=e.getResult(t).type;if(r instanceof I)return r.shape}return null}};l(DE,"shapesCompatible")});function Ru(n){return um(n,"throw")}function Nd(n,{labelOf:e,sameLabel:t=l((i,a)=>i===a,"sameLabel"),canMerge:r=l(()=>!0,"canMerge"),onAttach:s=l(()=>{},"onAttach"),sort:o=Ru}){let i=o(n),a=new Map,u=new Map,p=[],c=0,f=l((m,g)=>{if(m===g)return!0;let y=[g],_=new Set;for(;y.length>0;){let b=y.pop();if(b===m)return!0;if(_.has(b))continue;_.add(b);let w=u.get(b);if(w)for(let x of w)y.push(x)}return!1},"isUpstreamOf"),d=l(m=>{let g=new Set;for(let y=0;y<m.numOperands;y++){let _=m.getOperand(y).definingOp;if(!_)continue;let b=a.get(_);b&&g.add(b)}return g},"operandParts"),h=l((m,g)=>{for(let y of d(m)){if(y===g)continue;let _=u.get(g);_||(_=new Set,u.set(g,_)),_.add(y)}},"recordEdges");for(let m of i){let g=e(m);if(g==null)continue;let y=!1;for(let _=0;_<m.numOperands;_++){let b=m.getOperand(_).definingOp;if(!b)continue;let w=a.get(b);if(!w||!t(w.label,g)||!r(w,m,g))continue;let x=!1;for(let S of d(m))if(S!==w&&f(w,S)){x=!0;break}if(!x){w.ops.push(m),w.opSet.add(m),s(w,m),a.set(m,w),h(m,w),y=!0;break}}if(!y){let _={id:c++,label:g,ops:[m],opSet:new Set([m])};s(_,m),p.push(_),a.set(m,_),h(m,_)}}return{partitions:p,opToPart:a,preds:u}}function pa(n,e,{pullConstants:t=!1,isConstant:r=l(()=>!1,"isConstant")}={}){let s=[],o=new Set,i=[],a=new Set,u=[],p=new Set;for(let c of e){for(let f=0;f<c.numOperands;f++){let d=c.getOperand(f),h=d.definingOp;if(!(h&&n.has(h))){if(t&&h&&r(h)){p.has(h)||(p.add(h),u.push(h));continue}o.has(d)||(o.add(d),s.push(d))}}for(let f=0;f<c.numResults;f++){let d=c.getResult(f);if(a.has(d))continue;let h=!1;for(let m of d.uses())if(!n.has(m.user)){h=!0;break}h&&(a.add(d),i.push(d))}}return{inputs:s,outputs:i,constDefs:u}}function ov(n,e){let t=new Map,r=new Map;for(let a of n)t.set(a,0),r.set(a,[]);for(let a of n){let u=e.get(a);if(u)for(let p of u)r.has(p)&&(r.get(p).push(a),t.set(a,t.get(a)+1))}let s=[];for(let a of n)t.get(a)===0&&s.push(a);let o=[],i=0;for(;i<s.length;){let a=s[i++];o.push(a);for(let u of r.get(a)){let p=t.get(u)-1;t.set(u,p),p===0&&s.push(u)}}return o.length===n.length?o:null}var Ad=M(()=>{"use strict";Da();l(Ru,"topoSortOps");l(Nd,"buildPartitions");l(pa,"computePartitionIO");l(ov,"topoSortPartitions")});function R6(n){for(let e=0;e<n.numResults;e++){let t=n.getResult(e).type;if(t instanceof I)return t.shape.join(",")}return null}var wi,K_,H_=M(()=>{"use strict";Re();Ct();Bu();$a();Ad();wi=class{static{l(this,"FusionGroup")}constructor(e){this.id=e,this.ops=[],this.opSet=new Set,this.kind=null,this._inputValues=null,this._outputValues=null,this.minTopoPos=1/0,this.maxTopoPos=-1/0}addOp(e,t){this.opSet.has(e)||(this.ops.push(e),this.opSet.add(e),this._inputValues=null,this._outputValues=null,t!==void 0&&(t<this.minTopoPos&&(this.minTopoPos=t),t>this.maxTopoPos&&(this.maxTopoPos=t)))}hasOp(e){return this.opSet.has(e)}merge(e){for(let t of e.ops)this.addOp(t);e.minTopoPos<this.minTopoPos&&(this.minTopoPos=e.minTopoPos),e.maxTopoPos>this.maxTopoPos&&(this.maxTopoPos=e.maxTopoPos)}computeIO(){if(this._inputValues&&this._outputValues)return;let{inputs:e,outputs:t}=pa(this.opSet,this.ops);this._inputValues=e,this._outputValues=t}getInputValues(){return this.computeIO(),this._inputValues}getOutputValues(){return this.computeIO(),this._outputValues}get size(){return this.ops.length}classifyKind(){return this.kind=Tc(this.ops),this.kind}allOpsInlineFusable(){for(let e of this.ops){let t=ne.get(e.opName);if(!t)return!1;if(!(t.isReduction||t.isConstant)&&!ni(e.opName))return!1}return!0}};l(R6,"outputShapeKey");K_=class{static{l(this,"FusionGroupBuilder")}constructor(e){this.legality=e,this._nextId=0,this._topoIndex=null}buildProducerConsumerGroups(e){this._topoIndex=new Map;let t=0;for(let i of e.ops())this._topoIndex.set(i,t++);let r=[],s=new Map;for(let i of e.ops()){let a=ne.get(i.opName);if(!a||a.isConstant||a.isTerminator||a.isReduction)continue;let u=this._topoIndex.get(i);for(let p=0;p<i.numOperands;p++){let c=i.getOperand(p).definingOp;if(!c)continue;let f=ne.get(c.opName);if(!f||f.isConstant||f.isReduction)continue;let d=s.get(i),h=s.get(c);if(d&&d===h||!this.legality.canFuse(c,i).legal)continue;let g=this._topoIndex.get(c);if(!d&&!h){let y=new wi(this._nextId++);y.addOp(c,g),y.addOp(i,u),s.set(c,y),s.set(i,y)}else if(d&&!h)d.size<this.legality.maxFusionSize&&!this._wouldCreateCycle(d,c)&&(d.addOp(c,g),s.set(c,d));else if(!d&&h)h.size<this.legality.maxFusionSize&&!this._wouldCreateCycle(h,i)&&(h.addOp(i,u),s.set(i,h));else if(this.legality.canMergeGroups(d,h).legal&&!this._mergeWouldCreateCycle(d,h)){d.merge(h);for(let _ of h.ops)s.set(_,d)}}}let o=new Set;for(let i of s.values())o.has(i)||i.size<2||(o.add(i),i.classifyKind(),r.push(i));return r}_wouldCreateCycle(e,t){for(let r=0;r<t.numOperands;r++){let s=t.getOperand(r).definingOp;if(!(!s||e.hasOp(s)||this._topoIndex.get(s)<e.minTopoPos)&&this._dependsOnGroup(s,e))return!0}return!1}_dependsOnOps(e,t,r){let s=new Set,o=[e];for(s.add(e);o.length>0;){let i=o.pop();for(let a=0;a<i.numOperands;a++){let u=i.getOperand(a).definingOp;if(!u||s.has(u))continue;if(t.has(u))return!0;this._topoIndex.get(u)<r||(s.add(u),o.push(u))}}return!1}_dependsOnGroup(e,t){return this._dependsOnOps(e,t.opSet,t.minTopoPos)}_mergeWouldCreateCycle(e,t){let r=Math.min(e.minTopoPos,t.minTopoPos),s=Math.max(e.maxTopoPos,t.maxTopoPos),o=new Set([...e.opSet,...t.opSet]);for(let i of o)for(let a=0;a<i.numOperands;a++){let u=i.getOperand(a).definingOp;if(!u||o.has(u))continue;let p=this._topoIndex.get(u);if(!(p<r||p>s)&&this._dependsOnOps(u,o,r))return!0}return!1}_bucketable(e,t){return t&&!t.isConstant&&!t.isTerminator&&!t.isOpaque&&e.numResults>0&&R6(e)!==null}buildHorizontalGroups(e){let t=[...e.ops()];this._topoIndex=new Map;for(let u=0;u<t.length;u++)this._topoIndex.set(t[u],u);let r=[],s=new Map,o=this.legality.maxFusionSize||t.length,i=new Map,a=0;for(let u=0;u<t.length;u++){let p=t[u];if(s.has(p))continue;let c=ne.get(p.opName);if(!this._bucketable(p,c))continue;let f=p.getResult(0).type;a++,i.set(p,a);let d=new wi(this._nextId++);d.addOp(p,u),s.set(p,d);let h=Math.min(t.length,u+1+o);for(let m=u+1;m<h;m++){let g=t[m],y=!1;for(let b=0;b<g.numOperands;b++){let w=g.getOperand(b).definingOp;if(w&&i.get(w)===a){y=!0;break}}if(y){i.set(g,a);continue}if(d.size>=o||s.has(g))continue;let _=ne.get(g.opName);this._bucketable(g,_)&&f.equals(g.getResult(0).type)&&(this._sharesInput(p,g)||c.isElementwise&&_.isElementwise)&&(d.addOp(g,m),s.set(g,d),i.set(g,a))}d.size>=2&&(d.kind=We.HORIZONTAL,r.push(d))}return r}buildAllGroups(e){let t=this.buildProducerConsumerGroups(e),r=this.buildHorizontalGroups(e),s=new Map;for(let c of t)for(let f of c.ops)s.set(f,c);let o=new Set(s.keys()),i=[];for(let c of r){let f=!1;for(let d of c.ops)if(o.has(d)){f=!0;break}f||i.push(c)}let a=[...e.ops()],u=[];for(let c of a)for(let f=0;f<c.numOperands;f++){let d=c.getOperand(f).definingOp;d&&u.push(c,d)}for(let c of i)for(let f of c.ops)s.set(f,c);if(!this._condensedHasCycle(a,u,s))return[...t,...i];for(let c of i)for(let f of c.ops)s.delete(f);let p=[...t];for(let c of i){for(let f of c.ops)s.set(f,c);if(this._condensedHasCycle(a,u,s)){for(let f of c.ops)s.delete(f);continue}p.push(c)}return p}_condensedHasCycle(e,t,r){let s=l(f=>r.get(f)||f,"repOf"),o=new Map,i=new Set;for(let f of e)i.add(s(f));for(let f=0;f<t.length;f+=2){let d=s(t[f]),h=s(t[f+1]);if(h===d)continue;i.add(h),i.add(d);let m=o.get(h);m||(m=new Set,o.set(h,m)),m.add(d)}let a=0,u=1,p=2,c=new Map;for(let f of i)c.set(f,a);for(let f of i){if(c.get(f)!==a)continue;let d=[f];for(;d.length>0;){let h=d[d.length-1],m=c.get(h);if(m===a){c.set(h,u);let g=o.get(h);if(g)for(let y of g){let _=c.get(y);if(_===u)return!0;_===a&&d.push(y)}}else m===u&&c.set(h,p),d.pop()}}return!1}_sharesInput(e,t){if(e.numOperands===0||t.numOperands===0)return!1;let r=new Set;for(let s=0;s<e.numOperands;s++)r.add(e.getOperand(s));for(let s=0;s<t.numOperands;s++)if(r.has(t.getOperand(s)))return!0;return!1}_hasDependency(e,t){for(let r=0;r<e.numResults;r++){let s=e.getResult(r);for(let o=0;o<t.numOperands;o++)if(t.getOperand(o)===s)return!0}return!1}}});var M6,fa,Y_=M(()=>{"use strict";Re();Ct();M6=Object.freeze({memory:1,launch:1e3}),fa=class{static{l(this,"FusionCostModel")}constructor(e={}){this.memoryBandwidthGBs=e.memoryBandwidthGBs||900,this.computeTFLOPs=e.computeTFLOPs||15,this.launchOverheadUs=e.launchOverheadUs||5,this.minBenefitRatio=e.minBenefitRatio||1.05,this.maxRegistersPerThread=e.maxRegistersPerThread||255,this.maxSharedMemory=e.maxSharedMemory||49152,this.maxCodeSizeOps=e.maxCodeSizeOps||256,this.libraryOps=e.libraryOps||new Set,this.registerBytesPerOp=e.registerBytesPerOp||8,this.policy=e.policy||null,this.benefitWeights={...M6,...e.benefitWeights||{}}}edgeBenefit(e){let t=this.benefitWeights;return t.launch*this.launchOverheadUs+t.memory*e}estimateOpCost(e){let t=this.estimateFLOPs(e),r=this.estimateBytes(e);return{flops:t,bytes:r,arithmeticIntensity:r>0?t/r:0}}estimateFLOPs(e){let t=ne.get(e.opName);if(t&&t.getFlops)return t.getFlops(e);let r=1;for(let s=0;s<e.numResults;s++){let o=e.getResult(s).type;if(o instanceof I){let i=o.numel();i!==X&&(r=i);break}}if(t&&t.isReduction&&e.numOperands>0){let s=e.getOperand(0).type;if(s instanceof I){let o=s.numel();if(o!==X)return o}}return r}estimateBytes(e){let t=0;for(let r=0;r<e.numOperands;r++){let s=e.getOperand(r).type;if(s instanceof I){let o=s.sizeInBytes();o!==X&&(t+=o)}}for(let r=0;r<e.numResults;r++){let s=e.getResult(r).type;if(s instanceof I){let o=s.sizeInBytes();o!==X&&(t+=o)}}return t}estimateGroupCost(e){let t=e.opSet,r=new Map,s=0,o=0,i=0,a=0,u=!1,p=!1,c=0,f=0,d=0,h=0,m=new Set;for(let x of e.ops){let S=this.estimateFLOPs(x);r.set(x,S),s+=S,o+=this.estimateBytes(x),this.libraryOps.has(x.opName)&&a++;let T=ne.get(x.opName);if(T){let v=this._outputElements(x);T.isReduction&&(u=!0,c+=v),T.isElementwise&&(p=!0,f+=v)}for(let v=0;v<x.numResults;v++){let A=x.getResult(v);m.add(A);let N=0,k=!1;for(let O of A.uses())t.has(O.user)?N++:k=!0;N>1&&(i+=(N-1)*S),(N>0||k)&&d++}for(let v=0;v<x.numOperands;v++){let A=x.getOperand(v);if(!m.has(A))continue;let N=!1;for(let k of A.uses())if(t.has(k.user)&&k.user!==x){N=!0;break}N||d--}d>h&&(h=d)}let g=e.getInputValues(),y=e.getOutputValues(),_=0;for(let x of g)if(x.type instanceof I){let S=x.type.sizeInBytes();S!==X&&(_+=S)}for(let x of y)if(x.type instanceof I){let S=x.type.sizeInBytes();S!==X&&(_+=S)}let b=0;for(let x of e.ops)for(let S=0;S<x.numResults;S++){let T=x.getResult(S),v=0;for(let A of T.uses())if(t.has(A.user)&&(v++,v>1))break;if(v>1){let A=T.type;if(A instanceof I){let N=A.sizeInBytes();N!==X&&(b+=N)}}}let w=0;return u&&p&&f>0&&c>0&&(w=Math.abs(f-c)),{unfusedFLOPs:s,unfusedBytes:o,fusedFLOPs:s+i,fusedBytes:_,recomputeCost:i,memorySaved:o-_,launchSaved:(e.size-1)*this.launchOverheadUs,registerPressure:h*this.registerBytesPerOp,sharedMemoryUsage:b,parallelismLoss:w,libraryCallLoss:a}}shouldFuse(e){if(e.size<2)return{fuse:!1,reason:"group too small"};if(this.policy&&typeof this.policy.shouldFuse=="function"){let r=this.policy.shouldFuse(e,this);if(r)return r}let t=this.estimateGroupCost(e);if(t.libraryCallLoss>0)return{fuse:!1,reason:"fusion would lose library call opportunity",cost:t};if(t.registerPressure>this.maxRegistersPerThread)return{fuse:!1,reason:`register pressure ${t.registerPressure} exceeds limit ${this.maxRegistersPerThread}`,cost:t};if(t.sharedMemoryUsage>this.maxSharedMemory)return{fuse:!1,reason:`shared memory ${t.sharedMemoryUsage} exceeds limit ${this.maxSharedMemory}`,cost:t};if(e.size>this.maxCodeSizeOps)return{fuse:!1,reason:`code size ${e.size} exceeds limit ${this.maxCodeSizeOps}`,cost:t};if(t.memorySaved<=0&&t.launchSaved<=0)return{fuse:!1,reason:"no memory or launch benefit",cost:t};if(t.parallelismLoss>0){let r=t.memorySaved+t.launchSaved*1e3;if(t.parallelismLoss>r*this.minBenefitRatio)return{fuse:!1,reason:"parallelism loss outweighs fusion benefit",cost:t}}return{fuse:!0,reason:`saves ${t.memorySaved} bytes, ${t.launchSaved}us launch`,cost:t}}_outputElements(e){for(let t=0;t<e.numResults;t++){let r=e.getResult(t).type;if(r instanceof I){let s=r.numel();if(s!==X)return s}}return 0}}});function xc(n){let e=null;for(let t of n.ops())e=t;return e&&e.opName==="yield"?e:null}function Sc(n){let e=0,t=n.regions[0]?.entryBlock;if(!t)return 0;for(let r of t.ops())r.opName!=="yield"&&e++;return e}function vc(n){let e=0,t=n.regions[0]?.entryBlock;if(!t)return 0;for(let r of t.ops()){let s=ne.get(r.opName);s&&s.isReduction&&e++}return e}function kc(n){let e=n.regions[0]?.entryBlock;if(!e)return!1;for(let t of e.ops())if(t.opName!=="yield"&&!ni(t.opName))return!1;return!0}function I6(n){let e=new Map,t=0;for(let r=n.firstOp;r;r=r._next)e.set(r,t++);return e}function iv(n){let e=I6(n);return(t,r)=>{let s=e.get(t),o=e.get(r);return s!==void 0&&o!==void 0&&s<o}}function Nc(n,e){let t=Qo(n.ops,g=>n.hasOp(g),"null");if(t===null||t.length===0)return null;n._inputValues=null,n._outputValues=null;let r=n.getInputValues(),s=n.getOutputValues(),o=r.map(g=>g.type),i=s.map(g=>g.type),a=new Ut,u=new On(o);a.addBlock(u);let p=new Map;for(let g=0;g<r.length;g++)p.set(r[g],u.arguments[g]);for(let g of t)u.pushOp(g.clone(p));let c=s.map(g=>{let y=p.get(g);if(y===void 0)throw new Error("Fusion materialization: output value not found in valueMap");return y});u.pushOp(new Me("yield",c,[]));let f=new Me("fusion",r,i,{fusion_kind:n.kind||e},[a]),d=t[0].parentBlock;if(!d)return null;let h=iv(d),m=null;for(let g of r){let y=g.definingOp;!y||n.hasOp(y)||(!m||!h(y,m))&&(m=y)}m&&m.parentBlock===d?d.insertAfter(f,m):d.insertBefore(f,t[0]);for(let g=0;g<s.length;g++)s[g].replaceAllUsesWith(f.getResult(g));for(let g of t)g.dropAllOperands(),g.parentBlock&&g.parentBlock.removeOp(g);return f}var Ac=M(()=>{"use strict";Ct();$a();yr();Ea();Da();l(xc,"getYieldOp");l(Sc,"countInnerOps");l(vc,"countReductions");l(kc,"allInnerOpsFusable");l(I6,"blockPositionIndex");l(iv,"makeComesBefore");l(Nc,"materializeFusionGroup")});var J_,PE=M(()=>{"use strict";fn();Bu();H_();Y_();Ac();nn();J_=class extends qe{static{l(this,"FusionPass")}constructor(e={}){super("FusionPass");let t=e.target||{};this.legality=new wc({maxFusionSize:t.maxFusionSize||e.maxFusionSize,maxSharedMemory:t.sharedMemoryBytes||e.maxSharedMemory,libraryOps:t.libraryOps||e.libraryOps,allowReductionFusion:e.allowReductionFusion}),this.costModel=new fa({memoryBandwidthGBs:t.memoryBandwidthGBs,computeTFLOPs:t.computeTFLOPs,maxRegistersPerThread:t.registersPerThread,maxSharedMemory:t.sharedMemoryBytes,libraryOps:t.libraryOps,policy:t.getAttr?t.getAttr("fusionPolicy"):null,...e.cost}),this.groupBuilder=new K_(this.legality)}run(e,t){let r=this.groupBuilder.buildAllGroups(e),s=[];for(let i of r)this._createsCycle(e,i)?this._traceDecision(i,!1,"fusing would create a dependency cycle"):s.push(i);let o=[];for(let i of s){if(!i.allOpsInlineFusable()){this._traceDecision(i,!1,"group contains ops without inline fusion support");continue}let a=this.costModel.shouldFuse(i);this._traceDecision(i,a.fuse,a.reason),a.fuse&&o.push(i)}if(o.length===0)return oe.UNCHANGED;for(let i of o)Nc(i,We.ELEMENTWISE);return oe.CHANGED}_traceDecision(e,t,r){if(!this.trace||this.trace.level<le.DEBUG)return;let s=e.ops.map(o=>o.opName);this.trace.emit({type:"fusion_decision",passName:this.name,groupSize:e.ops.length,ops:s,anchor:s[s.length-1]||null,fuse:t,reason:r||null,level:le.DEBUG}),this.trace.explain("fusion",s.join("+"),t?"fused":"not-fused",r||null,{groupSize:s.length})}_createsCycle(e,t){let r=t.opSet,s=t.getOutputValues(),o=new Set,i=[];for(let u of s)for(let p of u.uses())r.has(p.user)||(i.push(p.user),o.add(p.user));let a=0;for(;a<i.length;){let u=i[a++];if(r.has(u))return!0;for(let p=0;p<u.numResults;p++)for(let c of u.getResult(p).uses())o.has(c.user)||(o.add(c.user),i.push(c.user))}return!1}}});function Z_(n){return cl(n.opName)||_n(n.opName)}function $E(n){if(Z_(n))return!0;let e=ne.get(n.opName);return e!==null&&e.isElementwise}function L6(n,e){let t=n.getOperand(0).definingOp;return t&&e.has(t)&&!Z_(t)?n.getOperand(1).definingOp:t}function C6(n,e){if(!n.parentBlock||n.parentBlock!==e.parentBlock)return!1;let t=n.parentBlock.firstOp;for(;t;){if(t===n)return!0;if(t===e)return!1;t=t._next}return!1}function D6(n,e){for(let t of n)if(t!==e){for(let r=0;r<t.numResults;r++)for(let s of t.getResult(r).uses())if(!n.has(s.user))return!0}return!1}function P6(n,e){let t=ne.get(n.opName),r=t&&t.getAttr("epilogueTag");return r?r(n,e):"activation"}function $6(n){let e=[],t=new Set,r=new Set;function s(d){if(!(r.has(d)||d===n)){r.add(d);for(let h=0;h<d.numOperands;h++){let m=d.getOperand(h).definingOp;m&&m!==n&&!r.has(m)&&Z_(m)&&s(m)}e.push(d),t.add(d)}}l(s,"absorb");let o=[],i=n.getResult(0);for(let d of i.uses())$E(d.user)&&o.push(d.user);for(;o.length>0;){let d=o.pop();if(!r.has(d)){s(d);for(let h=0;h<d.numResults;h++)for(let m of d.getResult(h).uses())!r.has(m.user)&&$E(m.user)&&o.push(m.user)}}let a=[];for(let d of e)Z_(d)||a.push(P6(d,t));let u=e[e.length-1];for(let d=e.length-1;d>=0;d--){let h=e[d];for(let m=0;m<h.numResults;m++){for(let g of h.getResult(m).uses())if(!t.has(g.user)){u=h,d=-1;break}if(d<0)break}}let p=new Set;for(let d=0;d<n.numOperands;d++)p.add(n.getOperand(d));let c=[],f=new Set;for(let d of e)for(let h=0;h<d.numOperands;h++){let m=d.getOperand(h);if(f.has(m))continue;f.add(m);let g=m.definingOp;g!==n&&(g&&t.has(g)||p.has(m)||c.push(m))}return{chain:e,chainSet:t,tags:a,lastOp:u,extras:c}}function F6(n){let e=null,t=null,r=null,s=null,o=new Set;for(let i=0;i<2;i++){let a=n.getOperand(i),u=a.definingOp;if(!u||u.opName!=="convert")continue;let p=!1;for(let f of a.uses())if(f.user!==n){p=!0;break}if(p)continue;let c=u.getAttr("target_dtype")||u.getResult(0).type.dtype;i===0?(e=c,r=u.getOperand(0)):(t=c,s=u.getOperand(0)),o.add(u)}return{lhsCast:e,rhsCast:t,lhsInput:r,rhsInput:s,removed:o}}var E6,Q_,FE=M(()=>{"use strict";fn();yr();Ct();Ir();nn();l(Z_,"isPassthrough");l($E,"isEpilogueCandidate");E6=new Map([["add",(n,e)=>e.has(L6(n,e))?"residual_add":"bias"],["sub",()=>"bias"],["mul",()=>"scale"],["maximum",()=>"relu"],["clamp",()=>"clamp"],["neg",()=>"neg"],["exp",()=>"exp"],["tanh",()=>"tanh"],["sqrt",()=>"sqrt"],["abs",()=>"abs"],["log",()=>"log"]]);l(L6,"resolveOtherOperand");for(let[n,e]of E6)ne.has(n)&&ne.registerOpAttr(n,"epilogueTag",e);l(C6,"comesBefore");l(D6,"hasEscapingUse");l(P6,"classifyTag");l($6,"collectChainAndAnalyze");l(F6,"collectPrologue");Q_=class extends qe{static{l(this,"EpilogueFusionPass")}constructor(e={}){super("EpilogueFusionPass"),this.maxEpilogueOps=e.maxEpilogueOps||16,this.target=e.target||null}run(e){if(this.target&&!this.target.enableEpilogueFusion)return oe.UNCHANGED;let t=!1,r=[];for(let s of e.ops()){let o=ne.get(s.opName);o&&o.isOutEWiseFusable&&r.push(s)}for(let s of r){let o=$6(s);if(o.chain.length>this.maxEpilogueOps)continue;let i=F6(s),a=o.chain.length>0,u=i.lhsCast!==null||i.rhsCast!==null;if(!a&&!u)continue;let{chain:p,tags:c,extras:f}=o,d=0;for(let v of c)v==="bias"||v==="residual_add"||v==="scale"?d++:v==="clamp"&&(d+=2);if(d!==f.length)continue;let h=i.lhsInput||s.getOperand(0),m=i.rhsInput||s.getOperand(1),g=[h,m,...f],y=a?o.lastOp:s,_=y.getResult(0).type,b=new Map(s.attributes);b.set("epilogue_ops",p.map(v=>v.opName)),b.set("epilogue_tags",c),b.set("num_dot_operands",2),b.set("num_extra_inputs",f.length),i.lhsCast&&b.set("lhs_prologue_cast",i.lhsCast),i.rhsCast&&b.set("rhs_prologue_cast",i.rhsCast);let w=new Me("fused_dot_epilogue",g,[_],b),x=s.parentBlock;if(!x)continue;let S=new Set(p);S.add(s);for(let v of i.removed)S.add(v);if(D6(S,y))continue;let T=null;for(let v of g){let A=v.definingOp;!A||S.has(A)||(!T||!C6(A,T))&&(T=A)}T&&T.parentBlock===x?x.insertAfter(w,T):x.insertBefore(w,s),y.getResult(0).replaceAllUsesWith(w.getResult(0));for(let v=p.length-1;v>=0;v--)p[v].dropAllOperands(),p[v].parentBlock&&p[v].parentBlock.removeOp(p[v]);s.dropAllOperands(),s.parentBlock&&s.parentBlock.removeOp(s);for(let v of i.removed)v.dropAllOperands(),v.parentBlock&&v.parentBlock.removeOp(v);t=!0}return this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,dotsFound:r.length,changed:t,level:le.DEBUG}),t?oe.CHANGED:oe.UNCHANGED}}});var X_,zE=M(()=>{"use strict";fn();yr();Ea();Re();Ct();nn();Bu();Ac();X_=class extends qe{static{l(this,"FusionMergerPass")}constructor(e={}){super("FusionMergerPass"),this.maxFusionSize=e.maxFusionSize||512,this.maxReductions=e.maxReductions||1,this.launchOverheadUs=e.launchOverheadUs||5,this.minMemorySavings=e.minMemorySavings||0}run(e){let t=[];for(let u of e.ops())u.opName==="fusion"&&t.push(u);if(t.length<2)return oe.UNCHANGED;let r=new Set(t),s=this._buildProducerConsumerEdges(t,r);if(s.length===0)return oe.UNCHANGED;let o=!1,i=0,a=new Set;for(let{producer:u,consumer:p,sharedResults:c}of s)a.has(u)||a.has(p)||this._canMerge(u,p)&&this._shouldMerge(u,p,c)&&(this._mergeCreatesCycle(u,p,r)||(this._merge(u,p,c),a.add(u),a.add(p),o=!0,i++));return this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,fusionOps:t.length,edges:s.length,mergeCount:i,level:le.DEBUG}),o?oe.CHANGED:oe.UNCHANGED}_buildProducerConsumerEdges(e,t){let r=[],s=new Map;for(let o of e)for(let i=0;i<o.numResults;i++){let a=o.getResult(i);for(let u of a.uses()){if(!t.has(u.user))continue;let p=u.user;if(p===o)continue;let c=`${o.id}|${p.id}`,f=s.get(c);f||(f={producer:o,consumer:p,sharedResults:new Map},s.set(c,f),r.push(f)),f.sharedResults.set(i,u.operandIndex)}}return r}_canMerge(e,t){let r=Sc(e),s=Sc(t);if(r+s>this.maxFusionSize)return!1;let o=0;return o+=vc(e),o+=vc(t),!(o>this.maxReductions||!kc(e)||!kc(t))}_shouldMerge(e,t,r){let s=0;for(let[u]of r){let p=e.getResult(u);if(p.type instanceof I){let c=p.type.sizeInBytes();c!==X&&(s+=c)}}let o=!0;for(let u=0;u<e.numResults;u++){let p=e.getResult(u);for(let c of p.uses())if(c.user!==t){o=!1;break}if(!o)break}if(o)return!0;let i=this._estimateRecomputeCost(e);return s+this.launchOverheadUs*1e3>i}_mergeCreatesCycle(e,t,r){let s=new Set,o=[];for(let a=0;a<t.numOperands;a++){let p=t.getOperand(a).definingOp;p&&p!==e&&p.opName!=="constant"&&(s.has(p)||(s.add(p),o.push(p)))}let i=0;for(;i<o.length;){let a=o[i++];if(a===e)return!0;for(let u=0;u<a.numOperands;u++){let p=a.getOperand(u).definingOp;p&&!s.has(p)&&(s.add(p),o.push(p))}}return!1}_merge(e,t,r){let s=e.regions[0].entryBlock,o=t.regions[0].entryBlock,i=xc(s),a=xc(o);if(!i||!a)return;let u=new Map;for(let[N,k]of r)u.set(k,N);let p=[],c=new Map;for(let N=0;N<e.numOperands;N++)c.set(N,p.length),p.push(e.getOperand(N));let f=new Map;for(let N=0;N<t.numOperands;N++){if(u.has(N))continue;let k=t.getOperand(N),O=-1;for(let R=0;R<p.length;R++)if(p[R]===k){O=R;break}O>=0?f.set(N,O):(f.set(N,p.length),p.push(k))}let d=p.map(N=>N.type),h=new Ut,m=new On(d);h.addBlock(m);let g=new Map;for(let N=0;N<s.arguments.length;N++)g.set(s.arguments[N],m.arguments[c.get(N)]);for(let N of s.ops())N.opName!=="yield"&&m.pushOp(N.clone(g));for(let N=0;N<i.numOperands;N++){let k=i.getOperand(N),O=g.get(k)||k;for(let[R,B]of u)B===N&&g.set(o.arguments[R],O)}for(let N=0;N<o.arguments.length;N++){if(u.has(N))continue;let k=f.get(N);k!==void 0&&g.set(o.arguments[N],m.arguments[k])}for(let N of o.ops())N.opName!=="yield"&&m.pushOp(N.clone(g));let y=[],_=[],b=[];for(let N=0;N<e.numResults;N++){let k=!1;for(let O of e.getResult(N).uses())if(O.user!==t){k=!0;break}if(k){let O=i.getOperand(N),R=g.get(O)||O;b.push({resultIdx:N,value:R}),y.push(R),_.push(e.getResult(N).type)}}for(let N=0;N<t.numResults;N++){let k=a.getOperand(N),O=g.get(k)||k;y.push(O),_.push(t.getResult(N).type)}let w=new Me("yield",y,[]);m.pushOp(w);let x=[];for(let N of m.ops())N.opName!=="yield"&&x.push(N);let S=Tc(x),T=new Me("fusion",p,_,{fusion_kind:S},[h]),v=t;if(!v.parentBlock)return;v.parentBlock.insertBefore(T,v);let A=0;for(let{resultIdx:N}of b)e.getResult(N).replaceAllUsesWith(T.getResult(A)),A++;for(let N=0;N<t.numResults;N++)t.getResult(N).replaceAllUsesWith(T.getResult(A)),A++;e.dropAllOperands(),e.parentBlock&&e.parentBlock.removeOp(e),t.dropAllOperands(),t.parentBlock&&t.parentBlock.removeOp(t)}_estimateRecomputeCost(e){let t=0,r=e.regions[0]?.entryBlock;if(!r)return 0;for(let s of r.ops()){if(s.opName==="yield")continue;let o=ne.get(s.opName);if(o&&o.getFlops)t+=o.getFlops(s);else for(let i=0;i<s.numResults;i++){let a=s.getResult(i).type;if(a instanceof I){let u=a.numel();u!==X&&(t+=u);break}}}return t}}});function z6(n,e){let t=Math.min(n.id,e.id),r=Math.max(n.id,e.id);return`${t}|${r}`}var Od,VE=M(()=>{"use strict";fn();yr();Ea();Re();Bu();nn();Ac();Od=class extends qe{static{l(this,"MultiOutputFusionPass")}constructor(e={}){super("MultiOutputFusionPass"),this.maxFusionSize=e.maxFusionSize||512,this.maxReductions=e.maxReductions||1,this.minSharedInputBytes=e.minSharedInputBytes||0,this.maxOutputs=e.maxOutputs||8,this.maxConsumersPerInput=e.maxConsumersPerInput||64}run(e){let t=[];for(let a of e.ops())a.opName==="fusion"&&t.push(a);if(t.length<2)return oe.UNCHANGED;let r=this._findCandidates(t);if(r.length===0)return oe.UNCHANGED;let s=!1,o=0,i=new Set;for(let{left:a,right:u,sharedInputs:p,sharedBytes:c}of r)i.has(a)||i.has(u)||this._canMerge(a,u)&&(this._mergeCreatesCycle(a,u)||(this._mergeMultiOutput(a,u,p),i.add(a),i.add(u),s=!0,o++));return this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,fusionOps:t.length,candidates:r.length,mergeCount:o,level:le.DEBUG}),s?oe.CHANGED:oe.UNCHANGED}_findCandidates(e){let t=new Map;for(let o of e)for(let i=0;i<o.numOperands;i++){let u=o.getOperand(i).id,p=t.get(u);p||(p=[],t.set(u,p)),p.push(o)}let r=new Map;for(let[o,i]of t){if(i.length<2)continue;let a=Math.min(i.length,this.maxConsumersPerInput);for(let u=0;u<a;u++)for(let p=u+1;p<a;p++){let c=i[u],f=i[p];if(c===f)continue;let d=z6(c,f),h=r.get(d);h||(h={left:c,right:f,sharedInputs:new Set,sharedBytes:0},r.set(d,h)),h.sharedInputs.add(o)}}let s=[];for(let o of r.values()){let i=0;for(let a of o.sharedInputs){let u=t.get(a);if(!u||u.length===0)continue;let p=u[0];for(let c=0;c<p.numOperands;c++){let f=p.getOperand(c);if(f.id===a&&f.type instanceof I){let d=f.type.sizeInBytes();d!==X&&(i+=d);break}}}o.sharedBytes=i,i>=this.minSharedInputBytes&&s.push(o)}return s.sort((o,i)=>i.sharedBytes-o.sharedBytes),s}_canMerge(e,t){if(this._hasProducerConsumerEdge(e,t)||this._hasProducerConsumerEdge(t,e))return!1;let r=Sc(e),s=Sc(t);if(r+s>this.maxFusionSize||e.numResults+t.numResults>this.maxOutputs)return!1;let i=0;return i+=vc(e),i+=vc(t),!(i>this.maxReductions||!kc(e)||!kc(t))}_mergeCreatesCycle(e,t){let r=new Set,s=[];for(let i=0;i<e.numResults;i++){let a=e.getResult(i);for(let u of a.uses())u.user!==t&&!r.has(u.user)&&(r.add(u.user),s.push(u.user))}let o=0;for(;o<s.length;){let i=s[o++];if(i===t)return!0;for(let a=0;a<i.numResults;a++)for(let u of i.getResult(a).uses())r.has(u.user)||(r.add(u.user),s.push(u.user))}r.clear(),s.length=0;for(let i=0;i<t.numResults;i++){let a=t.getResult(i);for(let u of a.uses())u.user!==e&&!r.has(u.user)&&(r.add(u.user),s.push(u.user))}for(o=0;o<s.length;){let i=s[o++];if(i===e)return!0;for(let a=0;a<i.numResults;a++)for(let u of i.getResult(a).uses())r.has(u.user)||(r.add(u.user),s.push(u.user))}return!1}_mergeMultiOutput(e,t,r){let s=e.regions[0].entryBlock,o=t.regions[0].entryBlock,i=xc(s),a=xc(o);if(!i||!a)return;let u=[],p=new Map,c=new Map;for(let k=0;k<e.numOperands;k++){let O=e.getOperand(k),R=O.id,B=p.get(R);B===void 0&&(B=u.length,p.set(R,B),u.push(O)),c.set(k,B)}let f=new Map;for(let k=0;k<t.numOperands;k++){let O=t.getOperand(k),R=O.id,B=p.get(R);B===void 0&&(B=u.length,p.set(R,B),u.push(O)),f.set(k,B)}let d=u.map(k=>k.type),h=new Ut,m=new On(d);h.addBlock(m);let g=new Map;for(let k=0;k<s.arguments.length;k++)g.set(s.arguments[k],m.arguments[c.get(k)]);for(let k of s.ops())k.opName!=="yield"&&m.pushOp(k.clone(g));for(let k=0;k<o.arguments.length;k++)g.set(o.arguments[k],m.arguments[f.get(k)]);for(let k of o.ops())k.opName!=="yield"&&m.pushOp(k.clone(g));let y=[],_=[];for(let k=0;k<i.numOperands;k++){let O=i.getOperand(k);y.push(g.get(O)||O),_.push(e.getResult(k).type)}for(let k=0;k<a.numOperands;k++){let O=a.getOperand(k);y.push(g.get(O)||O),_.push(t.getResult(k).type)}let b=new Me("yield",y,[]);m.pushOp(b);let w=[];for(let k of m.ops())k.opName!=="yield"&&w.push(k);let x=Tc(w),S=new Me("fusion",u,_,{fusion_kind:x},[h]),T=e.parentBlock;if(!T)return;let v=iv(T),A=null;for(let k of u){let O=k.definingOp;!O||O===e||O===t||(!A||!v(O,A))&&(A=O)}A&&A.parentBlock===T?T.insertAfter(S,A):T.insertBefore(S,e);for(let k=0;k<e.numResults;k++)e.getResult(k).replaceAllUsesWith(S.getResult(k));let N=e.numResults;for(let k=0;k<t.numResults;k++)t.getResult(k).replaceAllUsesWith(S.getResult(N+k));e.dropAllOperands(),e.parentBlock&&e.parentBlock.removeOp(e),t.dropAllOperands(),t.parentBlock&&t.parentBlock.removeOp(t)}_hasProducerConsumerEdge(e,t){for(let r=0;r<e.numResults;r++)for(let s of e.getResult(r).uses())if(s.user===t)return!0;return!1}};l(z6,"pairKey")});var av,ft,ro=M(()=>{"use strict";av=class{static{l(this,"UseDefResult")}constructor(e,t,r,s,o){this.topologicalOrder=e,this.valueToOp=t,this.opUsers=r,this.depth=s,this.height=o}},ft=class{static get name(){return"use_def"}static get depKey(){return"useDef"}static get dependencies(){return[]}static compute(e){let t=[],r=new Map,s=new Map,o=new Set,i=new Set,a=l(c=>{if(o.has(c))return;i.add(c);let f=[{op:c,i:0}];for(;f.length>0;){let d=f[f.length-1],h=d.op;if(d.i<h.numOperands){let m=h.getOperand(d.i).definingOp;if(d.i++,m&&!o.has(m)){if(i.has(m))throw new Error("Cycle detected in UseDefAnalysis");i.add(m),f.push({op:m,i:0})}continue}i.delete(h),o.add(h),t.push(h),f.pop()}},"visit");for(let c of e.ops()){for(let f=0;f<c.numResults;f++)r.set(c.getResult(f),c);s.set(c,new Set)}let u=new Map,p=new Map;for(let c of e.ops())a(c);for(let c of t){let f=0;for(let d=0;d<c.numOperands;d++){let h=c.getOperand(d).definingOp;if(h){s.get(h).add(c);let m=u.get(h)||0;m+1>f&&(f=m+1)}}u.set(c,f)}for(let c=t.length-1;c>=0;c--){let f=t[c],d=0,h=s.get(f)||new Set;for(let m of h){let g=p.get(m)||0;g+1>d&&(d=g+1)}p.set(f,d)}return new av(t,r,s,u,p)}}});var uv,eb,GE=M(()=>{"use strict";ro();Ir();uv=class{static{l(this,"DominanceResult")}constructor(e){this.idom=e}postDominates(e,t){let r=t;for(;r;){if(r===e)return!0;r=this.idom.get(r)}return!1}immediatePDom(e){return this.idom.get(e)||null}pathToPDom(e){let t=[],r=this.idom.get(e);for(;r&&(t.push(r),r!==this.idom.get(r));)r=this.idom.get(r);return t}},eb=class{static get name(){return"post_dominance"}static get depKey(){return"postDominance"}static get dependencies(){return[ft]}static compute(e,t={}){let s=(t.useDef||ft.compute(e)).topologicalOrder,o=new Map;for(let m of s)o.set(m,[]);for(let m of s)for(let g=0;g<m.numResults;g++){let y=m.getResult(g);for(let _ of y.uses()){let b=_.user;o.has(b)&&o.get(m).push(b)}}let i=new Set;for(let m of s)Xt(m.opName)&&i.add(m);let a={opName:"__pdom_root__"},u=Math.max(1,Math.ceil(Math.log2(s.length+2))+1),p=new Map,c=new Map,f=new Map;c.set(a,0),f.set(a,new Array(u).fill(a));let d=l((m,g)=>{p.set(m,g),c.set(m,c.get(g)+1);let y=new Array(u);y[0]=g;for(let _=1;_<u;_++)y[_]=f.get(y[_-1])[_-1];f.set(m,y)},"link"),h=l((m,g)=>{if(c.get(m)<c.get(g)){let _=m;m=g,g=_}let y=c.get(m)-c.get(g);for(let _=0;_<u;_++)y>>_&1&&(m=f.get(m)[_]);if(m===g)return m;for(let _=u-1;_>=0;_--)f.get(m)[_]!==f.get(g)[_]&&(m=f.get(m)[_],g=f.get(g)[_]);return f.get(m)[0]},"lca");for(let m=s.length-1;m>=0;m--){let g=s[m];if(i.has(g)){d(g,a);continue}let y=null;for(let _ of o.get(g))p.has(_)&&(y=y===null?_:h(y,_));y!==null&&d(g,y)}for(let[m,g]of p)g===a&&p.delete(m);return new uv(p,s)}}});function jE(n){return Xt(n)||_n(n)}var tb,UE=M(()=>{"use strict";fn();Ct();Bu();H_();Y_();GE();nn();ro();Ac();Ir();l(jE,"isSkipOp");tb=class extends qe{static{l(this,"DominatorFusionPass")}constructor(e={}){super("DominatorFusionPass"),this.requiredAnalyses=[ft];let t=e.target||{};this.maxFusionSize=t.maxFusionSize||e.maxFusionSize||512,this.maxReductions=e.maxReductions||1,this.libraryOps=t.libraryOps||e.libraryOps||new Set,this.costModel=new fa({memoryBandwidthGBs:t.memoryBandwidthGBs,computeTFLOPs:t.computeTFLOPs,maxRegistersPerThread:t.registersPerThread,maxSharedMemory:t.sharedMemoryBytes,libraryOps:t.libraryOps,...e.cost})}run(e,t){let r=t?t.getAnalysis(ft,e):ft.compute(e),s=eb.compute(e,{useDef:r}),o=r.topologicalOrder,i=this._buildGroups(o,s);if(i.length===0)return oe.UNCHANGED;let a=[];for(let u of i){if(!u.allOpsInlineFusable()){this._explain(u,!1,"group contains ops without inline fusion support");continue}let p=this.costModel.shouldFuse(u);this._explain(u,p.fuse,p.reason),p.fuse&&a.push(u)}if(a.length===0)return oe.UNCHANGED;for(let u of a)Nc(u,We.ELEMENTWISE);return this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,groupsBuilt:i.length,groupsFused:a.length,level:le.DEBUG}),oe.CHANGED}_explain(e,t,r){if(!this.trace||!this.trace.explainsEnabled)return;let s=e.ops.map(o=>o.opName);this.trace.explain("fusion",s.join("+"),t?"fused":"not-fused",r||null,{groupSize:s.length,strategy:"dominator"})}_buildGroups(e,t){let r=new Map,s=[],o=0;for(let p=e.length-1;p>=0;p--){let c=e[p];if(jE(c.opName))continue;let f=Ou(c);if(f===We.OPAQUE||this.libraryOps.has(c.opName))continue;let d=t.immediatePDom(c);if(!d||jE(d.opName))continue;let h=Ou(d);if(h===We.OPAQUE||this.libraryOps.has(d.opName)||!this._canFusePatterns(f,h)||!this._pathAllFusable(c,d,e,t))continue;let m=r.get(d);if(m){if(m.size+1>this.maxFusionSize||!this._checkReductionLimit(m,c))continue;m.addOp(c),r.set(c,m),this._absorbIntermediates(c,d,m,r,e,t)}else{let g=r.get(c);if(g){if(g.size+1>this.maxFusionSize||!this._checkReductionLimit(g,d))continue;g.addOp(d),r.set(d,g),this._absorbIntermediates(c,d,g,r,e,t)}else{let y=new wi(o++);if(y.addOp(c),y.addOp(d),!this._checkGroupReductions(y))continue;r.set(c,y),r.set(d,y),s.push(y),this._absorbIntermediates(c,d,y,r,e,t)}}}let i=new Set,a=[];for(let p of s)i.has(p)||p.size<2||(i.add(p),p.classifyKind(),a.push(p));for(let p of r.values())i.has(p)||p.size<2||(i.add(p),p.classifyKind(),a.push(p));let u=[];for(let p of a){if(!this._checkGroupReductions(p)){this._explain(p,!1,`group exceeds the ${this.maxReductions}-reduction limit`);continue}if(this._createsCycle(p)){this._explain(p,!1,"fusing would create a dependency cycle");continue}u.push(p)}return u}_createsCycle(e){let t=new Set;for(let o of e.getInputValues())o.definingOp&&!e.hasOp(o.definingOp)&&t.add(o.definingOp);if(t.size===0)return!1;let r=new Set,s=[];for(let o of e.getOutputValues())for(let i of o.uses())e.hasOp(i.user)||s.push(i.user);for(;s.length>0;){let o=s.pop();if(!r.has(o)){if(r.add(o),t.has(o))return!0;for(let i=0;i<o.numResults;i++)for(let a of o.getResult(i).uses())e.hasOp(a.user)||s.push(a.user)}}return!1}_canFusePatterns(e,t){return sv(e,t)}_pathAllFusable(e,t,r,s){let o=new Set,i=[e];for(o.add(e),o.add(t);i.length>0;){let a=i.pop();for(let u=0;u<a.numResults;u++)for(let p of a.getResult(u).uses()){let c=p.user;if(c===t||o.has(c)||!s.postDominates(t,c))continue;if(o.add(c),Ou(c)===We.OPAQUE||this.libraryOps.has(c.opName))return!1;i.push(c)}}return!0}_absorbIntermediates(e,t,r,s,o,i){let a=new Set,u=[e];for(a.add(e),a.add(t);u.length>0;){let p=u.pop();for(let c=0;c<p.numResults;c++)for(let f of p.getResult(c).uses()){let d=f.user;d===t||a.has(d)||!i.postDominates(t,d)||(a.add(d),Ou(d)===We.OPAQUE||this.libraryOps.has(d.opName))||(r.size<this.maxFusionSize&&this._checkReductionLimit(r,d)&&(r.addOp(d),s.set(d,r)),u.push(d))}}}_checkGroupReductions(e){let t=0;for(let r of e.ops){let s=ne.get(r.opName);s&&s.isReduction&&t++}return t<=this.maxReductions}_checkReductionLimit(e,t){let r=ne.get(t.opName);if(!r||!r.isReduction)return!0;let s=0;for(let o of e.ops){let i=ne.get(o.opName);i&&i.isReduction&&s++}return s<this.maxReductions}}});var nb,qE=M(()=>{"use strict";nb=class{static{l(this,"GraphCycles")}constructor(e,t){this._n=e,this._parent=new Int32Array(e),this._rank=new Int32Array(e),this._nodeAtRank=new Int32Array(e),this._out=new Array(e),this._in=new Array(e);for(let r=0;r<e;r++)this._parent[r]=r,this._rank[r]=r,this._nodeAtRank[r]=r,this._out[r]=new Set,this._in[r]=new Set;if(t)for(let[r,s]of t)r!==s&&(this._out[r].add(s),this._in[s].add(r))}find(e){let t=this._parent,r=e;for(;t[r]!==r;)r=t[r];for(;t[e]!==r;){let s=t[e];t[e]=r,e=s}return r}wouldCreateCycle(e,t){let r=this.find(e),s=this.find(t);if(r===s)return!1;let o=this._rank[r]<this._rank[s]?r:s,i=o===r?s:r;return this._hasIntermediatePath(o,i)}_hasIntermediatePath(e,t){let r=this._rank[t],s=new Set([e]),o=[e];for(;o.length>0;){let i=o.pop();for(let a of this._out[i]){let u=this.find(a);if(!(u===i||u===e)){if(u===t){if(i!==e)return!0;continue}this._rank[u]>=r||s.has(u)||(s.add(u),o.push(u))}}}return!1}merge(e,t){let r=this.find(e),s=this.find(t);if(r===s)return r;let o=this._rank[r]<this._rank[s]?r:s,i=o===r?s:r,a=this._rank[o],u=this._rank[i],p=this._out[r].size+this._in[r].size,c=this._out[s].size+this._in[s].size,f=p>=c?r:s,d=f===r?s:r,h=this._out[f],m=this._in[f];for(let g of this._in[d]){let y=this.find(g);y===f||y===d||(this._out[y].delete(d),this._out[y].add(f),m.add(y))}for(let g of this._out[d]){let y=this.find(g);y===f||y===d||(this._in[y].delete(d),this._in[y].add(f),h.add(y))}return h.delete(d),m.delete(d),h.delete(f),m.delete(f),this._parent[d]=f,this._nodeAtRank[this._rank[d]]=-1,this._reorder(a,u),f}_reorder(e,t){let r=[],s=[];for(let c=e;c<=t;c++){let f=this._nodeAtRank[c];if(!(f<0)){if(this.find(f)!==f){this._nodeAtRank[c]=-1;continue}r.push(c),s.push(f)}}if(s.length<=1){for(let c=0;c<s.length;c++)this._rank[s[c]]=r[c],this._nodeAtRank[r[c]]=s[c];return}let o=new Set(s),i=new Map;for(let c of s)i.set(c,0);for(let c of s)for(let f of this._out[c]){let d=this.find(f);d!==c&&o.has(d)&&i.set(d,i.get(d)+1)}let a=[];for(let c of s)i.get(c)===0&&a.push(c);let u=[],p=0;for(;p<a.length;){let c=a[p++];u.push(c);for(let f of this._out[c]){let d=this.find(f);if(d!==c&&o.has(d)){let h=i.get(d)-1;i.set(d,h),h===0&&a.push(d)}}}for(let c=0;c<u.length;c++){let f=u[c],d=r[c];this._rank[f]=d,this._nodeAtRank[d]=f}}}});var rb,WE=M(()=>{"use strict";rb=class{static{l(this,"MaxHeap")}constructor(){this._items=[]}get size(){return this._items.length}isEmpty(){return this._items.length===0}push(e,t){let r=this._items;r.push({priority:e,value:t});let s=r.length-1;for(;s>0;){let o=s-1>>1;if(r[o].priority>=r[s].priority)break;let i=r[o];r[o]=r[s],r[s]=i,s=o}}pop(){let e=this._items,t=e.length;if(t===0)return;let r=e[0],s=e.pop();if(t>1){e[0]=s;let o=e.length,i=0;for(;;){let a=i*2+1,u=a+1,p=i;if(a<o&&e[a].priority>e[p].priority&&(p=a),u<o&&e[u].priority>e[p].priority&&(p=u),p===i)break;let c=e[p];e[p]=e[i],e[i]=c,i=p}}return r.value}}});function V6(n){let e=ne.get(n.opName);return!e||e.isConstant||e.isTerminator||e.isOpaque?!1:e.isReduction?!0:ni(n.opName)}var sb,KE=M(()=>{"use strict";fn();Ct();Re();Bu();H_();Y_();ro();qE();WE();Ac();$a();nn();l(V6,"isFusibleOp");sb=class extends qe{static{l(this,"PriorityFusionPass")}constructor(e={}){super("PriorityFusionPass"),this.requiredAnalyses=[ft];let t=e.target||{};this.maxReductions=e.maxReductions??1,this.legality=new wc({maxFusionSize:t.maxFusionSize||e.maxFusionSize,maxSharedMemory:t.sharedMemoryBytes||e.maxSharedMemory,libraryOps:t.libraryOps||e.libraryOps,allowReductionFusion:e.allowReductionFusion});let r=t.getAttr&&t.getAttr("fusionBenefitWeights")||e.benefitWeights;this.costModel=new fa({memoryBandwidthGBs:t.memoryBandwidthGBs,computeTFLOPs:t.computeTFLOPs,maxRegistersPerThread:t.registersPerThread,maxSharedMemory:t.sharedMemoryBytes,libraryOps:t.libraryOps,policy:t.getAttr?t.getAttr("fusionPolicy"):null,benefitWeights:r,...e.cost})}run(e,t){let s=(t?t.getAnalysis(ft,e):ft.compute(e)).topologicalOrder,o=s.length;if(o===0)return oe.UNCHANGED;let i=new Map;for(let T=0;T<o;T++)i.set(s[T],T);let a=[];for(let T=0;T<o;T++){let v=s[T];for(let A=0;A<v.numOperands;A++){let N=v.getOperand(A).definingOp;N&&i.has(N)&&a.push([i.get(N),T])}}let u=new nb(o,a),p=new Int32Array(o),c=new Map;for(let T=0;T<o;T++)if(V6(s[T])){let v=new wi(T);v.addOp(s[T]),c.set(T,v)}let f=new rb,d=l((T,v)=>{let A=new wi(-1);for(let N of T.ops)A.addOp(N);for(let N of v.ops)A.addOp(N);return A},"mergedGroup"),h=l(T=>{if(T.size<2||!T.allOpsInlineFusable())return!1;let v=0;for(let A of T.ops){let N=ne.get(A.opName);N&&N.isReduction&&v++}return v>this.maxReductions||(T.classifyKind(),T.kind===We.OPAQUE)?!1:this.costModel.shouldFuse(T).fuse===!0},"legalMerge"),m=l((T,v)=>{let A=0;for(let N=0;N<v.numOperands;N++){let k=v.getOperand(N);if(k.definingOp===T&&k.type instanceof I){let O=k.type.sizeInBytes();O!==X&&(A+=O)}}return A},"edgeBytes"),g=l((T,v)=>{let A=u.find(i.get(T)),N=u.find(i.get(v));if(A===N||!c.has(A)||!c.has(N)||!this.legality.canFuse(T,v).legal)return;let k=this.costModel.edgeBenefit(m(T,v));f.push(k,{a:A,b:N,va:p[A],vb:p[N],prodOp:T,consOp:v})},"pushCandidate"),y=l((T,v)=>{let A=new Set;for(let N of T.ops){for(let k=0;k<N.numResults;k++)for(let O of N.getResult(k).uses()){let R=i.get(O.user);if(R===void 0)continue;let B=u.find(R);if(B===v||!c.has(B))continue;let E=v*o+B;A.has(E)||(A.add(E),g(N,O.user))}for(let k=0;k<N.numOperands;k++){let O=N.getOperand(k).definingOp;if(!O)continue;let R=i.get(O);if(R===void 0)continue;let B=u.find(R);if(B===v||!c.has(B))continue;let E=B*o+v;A.has(E)||(A.add(E),g(O,N))}}},"reEval"),_=new Set;for(let T=0;T<o;T++){if(!c.has(T))continue;let v=s[T];for(let A=0;A<v.numOperands;A++){let N=v.getOperand(A).definingOp;if(!N)continue;let k=i.get(N);if(k===void 0||!c.has(k))continue;let O=k*o+T;_.has(O)||(_.add(O),g(N,v))}}let b=0;for(;!f.isEmpty();){let T=f.pop(),v=u.find(T.a),A=u.find(T.b);if(v===A||p[v]!==T.va||p[A]!==T.vb)continue;let N=c.get(v),k=c.get(A);if(!N||!k||!this.legality.canFuse(T.prodOp,T.consOp).legal)continue;let O=d(N,k);if(!h(O)||u.wouldCreateCycle(v,A))continue;let R=u.merge(v,A),B=N.size>=k.size?N:k,E=B===N?k:N;B.merge(E),c.delete(v),c.delete(A),c.set(R,B),p[R]=(p[v]>p[A]?p[v]:p[A])+1,b++,y(B,R)}if(b===0)return oe.UNCHANGED;let w=[],x=new Set;for(let T of c.values())x.has(T)||T.size<2||(x.add(T),w.push(T));let S=0;for(let T of w)T.classifyKind(),this._explain(T),Nc(T,We.ELEMENTWISE)&&S++;return this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,ops:o,merges:b,fusions:S,level:le.DEBUG}),S>0?oe.CHANGED:oe.UNCHANGED}_explain(e){if(!this.trace||!this.trace.explainsEnabled)return;let t=e.ops.map(r=>r.opName);this.trace.explain("fusion",t.join("+"),"fused",null,{groupSize:t.length,strategy:"priority"})}}});function G6(n,e){return n===e?!0:!n||!e?!1:typeof n.equals=="function"?n.equals(e):!1}var da,ob,HE=M(()=>{"use strict";Re();da=class{static{l(this,"LayoutPreference")}constructor(e,t,r=0){this.inputs=e,this.outputs=t,this.cost=r}},ob=class{static{l(this,"LayoutPolicy")}constructor(e){this.target=e,this._rules=new Map,this._initDefaultRules()}registerRule(e,t){this._rules.set(e,t)}getPreference(e){let t=this._rules.get(e.opName);return t?t(e,this.target):null}estimateConversionCost(e,t,r){if(!(r instanceof I)||G6(e,t))return 0;let s=r.numel();return s<0?1024:s*2}estimateBenefit(e,t,r){if(!(t instanceof I))return 0;let s=t.numel();if(s<0)return 0;let o=e.opName;if(o==="dot"||o==="conv"||o==="matmul")return s*4*r;if(o==="reduce")return s*2*r;let i=this.target.cacheLineBytes||64;return s*4<=i*4?0:Math.floor(s*.5)}_initDefaultRules(){this._rules.set("conv",(e,t)=>{let s=e.getOperand(0).type?.rank||4;if(t.preferredConvLayout)return new da([t.preferredConvLayout,null],[t.preferredConvLayout]);if(t.isGPU()&&s===4){let o=new Gt([0,2,3,1]);return new da([o,null],[o])}if(t.isCPU()&&s===4){let o=new Gt([0,2,3,1]);return new da([o,null],[o])}return null}),this._rules.set("dot",(e,t)=>{let r=e.getOperand(0).type,s=e.getOperand(1).type;if(!r||!s)return null;let o=Gt.rowMajor(r.rank);if(t.isCPU()&&s.rank===2){let a=Gt.columnMajor(s.rank);return new da([o,a],[o])}let i=Gt.rowMajor(s.rank);return new da([o,i],[o])}),this._rules.set("reduce",(e,t)=>{let r=e.getResult(0).type;return r?new da([null],[Gt.rowMajor(r.rank)]):null})}};l(G6,"layoutEquals")});function YE(n,e){let t=new Map,r=null,s=0;for(let i=0;i<n.numOperands;i++){let a=n.getOperand(i),u=e.get(a);if(!u)continue;let p=u.hash?u.hash():0,c=(t.get(p)||0)+1;t.set(p,c),c>s&&(s=c,r=u)}if(r)return r;let o=n.getResult(0);return o&&o.type instanceof I?Gt.rowMajor(o.type.rank):Gt.rowMajor(1)}function j6(n,e){return n===e?!0:!n||!e?!1:typeof n.equals=="function"?n.equals(e):!1}function JE(n){if(n instanceof Gt)return n;if(n&&typeof n.toLayout=="function")try{return n.toLayout()}catch{}return n&&n.baseOrder&&!n.isBlocked?.()?new Gt(n.baseOrder):n&&n.order?new Gt(n.order):null}var lv,ib,ZE=M(()=>{"use strict";ro();Re();Ct();Bn();Ir();lv=class{static{l(this,"LayoutAnalysisResult")}constructor(e,t,r){this.assignments=e,this.conversions=t,this.totalCost=r}},ib=class{static get name(){return"layout"}static get depKey(){return"layout"}static get dependencies(){return[ft]}static compute(e,t,r){let s=t.useDef,o=new Map,i=s.topologicalOrder;for(let p of e.args)p.type instanceof I&&o.set(p,p.type.layout||Gt.rowMajor(p.type.rank));for(let p=0;p<i.length;p++){let c=i[p];if(Xt(c.opName))continue;let f=r?r.getPreference(c):null;if(f&&f.outputs.length>0){for(let m=0;m<c.numResults;m++){let g=c.getResult(m);if(!(g.type instanceof I))continue;let y=f.outputs[m]||null;y?o.set(g,y):o.set(g,YE(c,o))}continue}let d=ne.get(c.opName),h=d&&d.hasTrait(Y.ELEMENTWISE);for(let m=0;m<c.numResults;m++){let g=c.getResult(m);g.type instanceof I&&(h?o.set(g,YE(c,o)):o.set(g,Gt.rowMajor(g.type.rank)))}}let a=[],u=0;for(let p=0;p<i.length;p++){let c=i[p],f=r?r.getPreference(c):null;if(f)for(let d=0;d<c.numOperands;d++){let h=c.getOperand(d);if(!(h.type instanceof I))continue;let m=o.get(h);if(!m)continue;let g=f.inputs[d];if(!g||j6(m,g))continue;let y=JE(m),_=JE(g);!y||!_||y.equals(_)||(a.push({value:h,consumer:c,operandIdx:d,from:y,to:_}),r&&(u+=r.estimateConversionCost(y,_,h.type)))}}return new lv(o,a,u)}};l(YE,"resolveFromInputs");l(j6,"layoutEquals");l(JE,"toPlainLayout")});function U6(n,e,t){let r=n.id,s=e.hash?e.hash():0,o=t.hash?t.hash():0;return`${r}:${s}:${o}`}var ab,QE=M(()=>{"use strict";fn();yr();Re();HE();ZE();ro();nn();ab=class extends qe{static{l(this,"LayoutTransformPass")}constructor(e={}){super("LayoutTransformPass"),this.target=e.target||null,this._policy=null}run(e,t){if(!this.target)return oe.UNCHANGED;this._policy||(this._policy=new ob(this.target));let r=t?t.getAnalysis(ft,e):ft.compute(e),s=ib.compute(e,{useDef:r},this._policy);if(s.conversions.length===0)return oe.UNCHANGED;let o=new Map;for(let p of s.conversions){let{value:c,consumer:f,operandIdx:d,from:h,to:m}=p,g=U6(c,h,m),y=o.get(g);y||(y={value:c,from:h,to:m,consumers:[],cost:this._policy.estimateConversionCost(h,m,c.type),benefit:0},o.set(g,y)),y.consumers.push({consumer:f,operandIdx:d});let _=this.target.layoutAwareOps&&this.target.layoutAwareOps.has(f.opName);y.benefit+=_?this._policy.estimateBenefit(f,c.type,1):0}let i=0,a=0,u=[];for(let p of o.values())p.benefit<p.cost||(u.push(p),i+=p.cost,a+=p.benefit);if(u.length===0||i>a)return oe.UNCHANGED;for(let p of u){let c=p.from instanceof Gt?p.from.order:Array.from({length:p.value.type.rank},(y,_)=>_),f=p.to instanceof Gt?p.to.order:Array.from({length:p.value.type.rank},(y,_)=>_),d=new I(p.value.type.shape,p.value.type.dtype,p.to),h=new Me("layout_transform",[p.value],[d],{src_layout:[...c],dst_layout:[...f]}),m=p.value.definingOp;m&&m.parentBlock?m.parentBlock.insertAfter(h,m):p.consumers[0].consumer.parentBlock&&p.consumers[0].consumer.parentBlock.insertBefore(h,p.consumers[0].consumer);let g=h.getResult(0);for(let y of p.consumers)y.consumer.replaceOperand(y.operandIdx,g)}return this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,conversions:s.conversions.length,uniqueTransforms:u.length,level:le.DEBUG}),oe.CHANGED}};l(U6,"valueLayoutKey")});function XE(n,e){if(n===e)return!0;if(typeof n=="number"&&typeof e=="number")return n===e;if(typeof n!=typeof e||n.length!==e.length)return!1;for(let t=0;t<n.length;t++)if(n[t]!==e[t])return!1;return!0}function eL(n){let e=0;for(let t=0;t<n.length;t++)e=(e<<5)-e+n.charCodeAt(t)&2147483647;return e}function K6(n){if(typeof n=="number"){let t=new Float64Array([n]),r=new Uint32Array(t.buffer);return(r[0]^r[1])&2147483647}let e=n.length;for(let t=0;t<Math.min(n.length,8);t++){let r=new Float64Array([n[t]]),s=new Uint32Array(r.buffer);e=(e^(s[0]^s[1]))*16777619&2147483647}return e}var Pr,q6,W6,vr,cv=M(()=>{"use strict";Re();Pr=Object.freeze({PER_TENSOR_SYMMETRIC:"per_tensor_symmetric",PER_TENSOR_ASYMMETRIC:"per_tensor_asymmetric",PER_CHANNEL:"per_channel",PER_GROUP:"per_group"}),q6=new Set(Object.values(Pr)),W6=new Set(["i8","ui8"]),vr=class n{static{l(this,"QuantizationParams")}constructor(e){this.scheme=e.scheme,this.scale=e.scale,this.zeroPoint=e.zeroPoint,this.axis=e.axis??null,this.groupSize=e.groupSize??null,this.dtype=e.dtype||"i8",this.numBits=e.numBits||fo(this.dtype)*8,this._hash=null}clampRange(){if(this.isSymmetric()){let r=2**(this.numBits-1)-1;return[-r,r]}if(this.dtype==="ui8")return[0,2**this.numBits-1];let e=-(2**(this.numBits-1)),t=2**(this.numBits-1)-1;return[e,t]}quantize(e){let[t,r]=this.clampRange(),s=Math.round(e/this.getScalarScale()+this.getScalarZeroPoint());return Math.max(t,Math.min(r,s))}dequantize(e){return(e-this.getScalarZeroPoint())*this.getScalarScale()}quantizeArray(e){let[t,r]=this.clampRange(),s=new Array(e.length);if(this.isPerChannel())throw new Error("Use quantizeArrayPerChannel(floatArr, shape) for per-channel quantization");let o=this.getScalarScale(),i=this.getScalarZeroPoint();for(let a=0;a<e.length;a++)s[a]=Math.max(t,Math.min(r,Math.round(e[a]/o+i)));return s}dequantizeArray(e){let t=new Array(e.length);if(this.isPerChannel())throw new Error("Use dequantizeArrayPerChannel(intArr, shape) for per-channel dequantization");let r=this.getScalarScale(),s=this.getScalarZeroPoint();for(let o=0;o<e.length;o++)t[o]=(e[o]-s)*r;return t}_channelStride(e){let t=1;for(let r=this.axis+1;r<e.length;r++)t*=e[r];return t}quantizeArrayPerChannel(e,t){let[r,s]=this.clampRange(),o=this._channelStride(t),i=t[this.axis],a=new Array(e.length);for(let u=0;u<e.length;u++){let p=Math.floor(u/o)%i,c=this.getScaleForChannel(p),f=this.getZeroPointForChannel(p);a[u]=Math.max(r,Math.min(s,Math.round(e[u]/c+f)))}return a}dequantizeArrayPerChannel(e,t){let r=this._channelStride(t),s=t[this.axis],o=new Array(e.length);for(let i=0;i<e.length;i++){let a=Math.floor(i/r)%s;o[i]=(e[i]-this.getZeroPointForChannel(a))*this.getScaleForChannel(a)}return o}quantizeArrayPerGroup(e){let[t,r]=this.clampRange(),s=new Array(e.length);for(let o=0;o<e.length;o++){let i=Math.floor(o/this.groupSize),a=this.scale[i],u=this.zeroPoint[i];s[o]=Math.max(t,Math.min(r,Math.round(e[o]/a+u)))}return s}dequantizeArrayPerGroup(e){let t=new Array(e.length);for(let r=0;r<e.length;r++){let s=Math.floor(r/this.groupSize);t[r]=(e[r]-this.zeroPoint[s])*this.scale[s]}return t}getScaleForGroup(e){return this.scale[e]}getZeroPointForGroup(e){return this.zeroPoint[e]}getScaleForChannel(e){return this.isPerChannel()?this.scale[e]:this.getScalarScale()}getZeroPointForChannel(e){return this.isPerChannel()?this.zeroPoint[e]:this.getScalarZeroPoint()}getScalarScale(){return typeof this.scale=="number"?this.scale:this.scale[0]}getScalarZeroPoint(){return typeof this.zeroPoint=="number"?this.zeroPoint:this.zeroPoint[0]}numChannels(){return this.isPerChannel()?typeof this.scale=="number"?1:this.scale.length:1}isPerChannel(){return this.scheme===Pr.PER_CHANNEL}isPerGroup(){return this.scheme===Pr.PER_GROUP}isSymmetric(){return this.scheme===Pr.PER_TENSOR_SYMMETRIC}equals(e){return this===e?!0:!(!(e instanceof n)||this.scheme!==e.scheme||this.dtype!==e.dtype||this.numBits!==e.numBits||this.axis!==e.axis||this.groupSize!==e.groupSize||!XE(this.scale,e.scale)||!XE(this.zeroPoint,e.zeroPoint))}hash(){if(this._hash!==null)return this._hash;let e=2166136261;return e=(e^eL(this.scheme))*16777619&2147483647,e=(e^eL(this.dtype))*16777619&2147483647,e=(e^this.numBits)*16777619&2147483647,e=(e^K6(this.scale))*16777619&2147483647,this._hash=e,e}serialize(){let e=this.isPerChannel()||this.isPerGroup();return{scheme:this.scheme,scale:e?[...this.scale]:this.scale,zeroPoint:e?[...this.zeroPoint]:this.zeroPoint,axis:this.axis,groupSize:this.groupSize,dtype:this.dtype,numBits:this.numBits}}static deserialize(e){return new n(e)}static fromRange(e,t,r,s="i8",o=8){if(r===Pr.PER_TENSOR_SYMMETRIC){let d=Math.max(Math.abs(e),Math.abs(t)),h=2**(o-1)-1,m=d/h||1e-10;return new n({scheme:r,scale:m,zeroPoint:0,dtype:s,numBits:o})}let[i,a]=s==="ui8"?[0,2**o-1]:[-(2**(o-1)),2**(o-1)-1],p=(t-e||1e-10)/(a-i),c=Math.round(i-e/p),f=Math.max(i,Math.min(a,c));return new n({scheme:r,scale:p,zeroPoint:f,dtype:s,numBits:o})}static fromRangePerChannel(e,t,r,s="i8",o=8){let i=e.length,a=new Float64Array(i),u=new Int32Array(i),p=2**(o-1)-1;for(let c=0;c<i;c++){let f=Math.max(Math.abs(e[c]),Math.abs(t[c]));a[c]=f/p||1e-10,u[c]=0}return new n({scheme:Pr.PER_CHANNEL,scale:a,zeroPoint:u,axis:r,dtype:s,numBits:o})}static defaultForActivation(e,t="i8",r=8){return n.fromRange(-6,6,e,t,r)}static fromConstantArray(e,t,r="i8",s=8){let o=1/0,i=-1/0;for(let a=0;a<e.length;a++)e[a]<o&&(o=e[a]),e[a]>i&&(i=e[a]);return isFinite(o)||(o=-1),isFinite(i)||(i=1),o===i&&(o-=.5,i+=.5),n.fromRange(o,i,t,r,s)}static fromConstantArrayPerChannel(e,t,r,s="i8",o=8){let i=t[r],a=1;for(let c=r+1;c<t.length;c++)a*=t[c];let u=new Array(i).fill(1/0),p=new Array(i).fill(-1/0);for(let c=0;c<e.length;c++){let f=Math.floor(c/a)%i,d=e[c];d<u[f]&&(u[f]=d),d>p[f]&&(p[f]=d)}for(let c=0;c<i;c++)isFinite(u[c])||(u[c]=-1),isFinite(p[c])||(p[c]=1),u[c]===p[c]&&(u[c]-=.5,p[c]+=.5);return n.fromRangePerChannel(u,p,r,s,o)}static fromConstantArrayPerGroup(e,t,r="i8",s=4){let o=Math.ceil(e.length/t),i=new Float64Array(o),a=new Int32Array(o),u=2**(s-1)-1;for(let p=0;p<o;p++){let c=p*t,f=Math.min(c+t,e.length),d=0;for(let h=c;h<f;h++){let m=Math.abs(e[h]);m>d&&(d=m)}i[p]=d/u||1e-10,a[p]=0}return new n({scheme:Pr.PER_GROUP,scale:i,zeroPoint:a,groupSize:t,dtype:r,numBits:s})}static isQuantizableDtype(e){return W6.has(e)}static isValidScheme(e){return q6.has(e)}};l(XE,"scaleEquals");l(eL,"hashStr");l(K6,"hashScaleValue")});function Z6(n){for(let e=0;e<n.numOperands;e++){let t=n.getOperand(e).definingOp;if(t&&t.opName==="constant")return!0}return!1}function Q6(n,e,t){for(let r=0;r<n.numOperands;r++){let s=n.getOperand(r);if(!e.has(s)&&(!(s.type instanceof I)||!jt(s.type.dtype)))return!1}return n.numOperands>0}var H6,Y6,J6,ub,lb,tL=M(()=>{"use strict";fn();yr();Re();Ct();ro();cv();nn();Ir();H6=new Set(["softmax","sqrt","div","rsqrt","log","exp","tanh"]),Y6=new Set(["dot","conv","add","mul","sub"]),J6=new Map([["dot","quantized_dot"],["conv","quantized_conv"]]);for(let[n,e]of J6)ne.has(n)&&ne.registerOpAttr(n,"quantizedVariant",e);ub=class{static{l(this,"QuantizationConfig")}constructor(e={}){this.scheme=e.scheme||Pr.PER_TENSOR_SYMMETRIC,this.calibration=e.calibration||null,this.targetDtype=e.targetDtype||"i8",this.excludeOps=e.excludeOps||H6,this.quantizableOps=e.quantizableOps||Y6,this.sensitivityThreshold=e.sensitivityThreshold||0,this.sensitivityResult=e.sensitivityResult||null,this.weightOnly=e.weightOnly||!1,this.target=e.target||null}},lb=class extends qe{static{l(this,"QuantizationPass")}constructor(e={}){super("QuantizationPass"),this.requiredAnalyses=[ft],this.config=e instanceof ub?e:new ub(e)}run(e,t){let s=(t?t.getAnalysis(ft,e):ft.compute(e)).topologicalOrder,o=new Set;this._paramsByValue=new Map;let i=this.config,a=!1;if(i.target&&!i.target.supportsInt8)return oe.UNCHANGED;for(let p=0;p<s.length;p++){let c=s[p];if(Xt(c.opName))continue;if(i.excludeOps.has(c.opName)||!i.quantizableOps.has(c.opName)){for(let h=0;h<c.numOperands;h++)o.has(c.getOperand(h))&&(a=this._insertDequantBefore(c,h,c.getOperand(h),i)||a);continue}if(i.sensitivityResult&&i.sensitivityThreshold>0&&i.sensitivityResult.isSensitive(c,i.sensitivityThreshold)||i.weightOnly&&!Z6(c))continue;let f=ne.get(c.opName),d=f?f.getAttr("quantizedVariant"):null;if(d&&Q6(c,o,i)){i.scheme===Pr.PER_CHANNEL&&this._canPerChannelDot(c,o)?a=this._replacePerChannelDot(c,i)||a:a=this._replaceWithNativeQuantized(c,d,o,i)||a;continue}a=this._insertDequantQuantBoundary(c,o,i)||a}let u=e.getReturnOp();if(u)for(let p=0;p<u.numOperands;p++){let c=u.getOperand(p);o.has(c)&&(a=this._insertDequantBefore(u,p,c,i)||a)}return this.trace&&this.trace.level>=le.DEBUG&&a&&this.trace.emit({type:"pass_detail",passName:this.name,opsProcessed:s.length,changed:a,level:le.DEBUG}),a?oe.CHANGED:oe.UNCHANGED}_getQuantParams(e,t){let r=fo(t.targetDtype)*8;if(t.calibration&&t.calibration.hasData(e))return t.calibration.getQuantParams(e,t.scheme,t.targetDtype);let s=e.definingOp;if(s&&s.opName==="constant"){let o=s.getAttr("value");if(typeof o=="number")return vr.fromRange(-Math.abs(o)||-1,Math.abs(o)||1,t.scheme,t.targetDtype,r);if(o&&typeof o.length=="number")return vr.fromConstantArray(o,t.scheme,t.targetDtype,r)}return e.type instanceof I&&jt(e.type.dtype)?vr.defaultForActivation(t.scheme,t.targetDtype,r):null}_resolveQuantParams(e,t){let r=this._paramsByValue.get(e);if(r)return r;let s=e.definingOp;return s&&s.opName==="quantize"?new vr({scheme:s.getAttr("scheme")||t.scheme,scale:s.getAttr("scale"),zeroPoint:s.getAttr("zero_point")||0,dtype:s.getAttr("target_dtype")||t.targetDtype}):this._getQuantParams(e,t)}_insertQuantizeAfter(e,t,r){let s=e.getResult(t),o=this._getQuantParams(s,r);if(!o)return null;let i=new I(s.type.shape,r.targetDtype),a=new Me("quantize",[s],[i],{scale:o.getScalarScale(),zero_point:o.getScalarZeroPoint(),scheme:r.scheme,target_dtype:r.targetDtype,...o.axis!==null?{axis:o.axis}:{}});e.parentBlock&&e.parentBlock.insertAfter(a,e);let u=a.getResult(0);return this._paramsByValue.set(u,o),u}_insertDequantBefore(e,t,r,s){let o=this._resolveQuantParams(r,s);if(!o)return!1;let i="f32",a=new I(r.type.shape,i),u=new Me("dequantize",[r],[a],{scale:o.getScalarScale(),zero_point:o.getScalarZeroPoint(),scheme:s.scheme,target_dtype:i});return e.parentBlock&&(e.parentBlock.insertBefore(u,e),e.replaceOperand(t,u.getResult(0))),!0}_insertDequantQuantBoundary(e,t,r){let s=!1;for(let o=0;o<e.numOperands;o++){let i=e.getOperand(o);t.has(i)&&(s=this._insertDequantBefore(e,o,i,r)||s)}for(let o=0;o<e.numResults;o++){let i=e.getResult(o);if(!(i.type instanceof I)||!jt(i.type.dtype))continue;let a=this._insertQuantizeAfter(e,o,r);if(a){let u=[...i.uses()];for(let p of u)p.user!==a.definingOp&&p.user.replaceOperand(p.operandIndex,a);t.add(a),s=!0}}return s}_replaceWithNativeQuantized(e,t,r,s){let o=[],i={};for(let[f,d]of e.attributes||[])i[f]=d;for(let f=0;f<e.numOperands;f++){let d=e.getOperand(f);if(r.has(d)){o.push(d);let _=this._resolveQuantParams(d,s);if(_){let b=f===0?"lhs":"rhs";if(t==="quantized_dot")i[`${b}_scale`]=_.getScalarScale(),i[`${b}_zero_point`]=_.getScalarZeroPoint();else{let w=f===0?"input":"kernel";i[`${w}_scale`]=_.getScalarScale(),i[`${w}_zero_point`]=_.getScalarZeroPoint()}}continue}let h=this._getQuantParams(d,s);if(!h)return!1;let m=new I(d.type.shape,s.targetDtype),g=new Me("quantize",[d],[m],{scale:h.getScalarScale(),zero_point:h.getScalarZeroPoint(),scheme:s.scheme,target_dtype:s.targetDtype});e.parentBlock&&e.parentBlock.insertBefore(g,e),o.push(g.getResult(0));let y=f===0?t==="quantized_dot"?"lhs":"input":t==="quantized_dot"?"rhs":"kernel";i[`${y}_scale`]=h.getScalarScale(),i[`${y}_zero_point`]=h.getScalarZeroPoint()}let a=i.lhs_scale||i.input_scale||1,u=i.rhs_scale||i.kernel_scale||1;i.output_scale=a*u,i.output_zero_point=0;let p=new I(e.getResult(0).type.shape,"i32"),c=new Me(t,o,[p],i);if(e.parentBlock){e.parentBlock.insertBefore(c,e);let f=new I(e.getResult(0).type.shape,e.getResult(0).type.dtype),d=i.output_scale||1,h=new Me("dequantize",[c.getResult(0)],[f],{scale:d,zero_point:i.output_zero_point||0,scheme:s.scheme,target_dtype:e.getResult(0).type.dtype});e.parentBlock.insertBefore(h,e),e.replaceAllResultsWith([h.getResult(0)]),e.erase()}return!0}_canPerChannelDot(e,t){if(e.opName!=="dot")return!1;let r=e.getOperand(0),s=e.getOperand(1);if(t.has(r)||t.has(s)||!(r.type instanceof I)||r.type.shape.length!==2||!(s.type instanceof I)||s.type.shape.length!==2)return!1;let o=s.definingOp;if(!o||o.opName!=="constant")return!1;let i=o.getAttr("value");if(!i||typeof i=="number"||typeof i.length!="number")return!1;let a=e.getAttr("rhs_contracting")||[],u=e.getAttr("lhs_contracting")||[];return!(a.length!==1||u.length!==1||(e.getAttr("rhs_batch")||[]).length!==0||(e.getAttr("lhs_batch")||[]).length!==0)}_activationParams(e,t){let r=fo(t.targetDtype)*8,s=Pr.PER_TENSOR_SYMMETRIC;return t.calibration&&t.calibration.hasData(e)?t.calibration.getQuantParams(e,s,t.targetDtype):vr.defaultForActivation(s,t.targetDtype,r)}_replacePerChannelDot(e,t){let r=e.getOperand(0),s=e.getOperand(1),o=s.type.shape,i=s.definingOp.getAttr("value"),a=fo(t.targetDtype)*8,p=e.getAttr("rhs_contracting")[0]===0?1:0,c=vr.fromConstantArrayPerChannel([...i],o,p,t.targetDtype,a),f=c.quantizeArrayPerChannel([...i],o),d=this._activationParams(r,t),h=d.getScalarScale(),m=d.getScalarZeroPoint(),g=e.parentBlock;if(!g)return!1;let y=new I(o,t.targetDtype),_=new Me("constant",[],[y],{value:f,tensor_type:y});g.insertBefore(_,e);let b=new I(r.type.shape,t.targetDtype),w=new Me("quantize",[r],[b],{scale:h,zero_point:m,scheme:Pr.PER_TENSOR_SYMMETRIC,target_dtype:t.targetDtype});g.insertBefore(w,e);let x=e.getResult(0).type.shape,S={};for(let[U,D]of e.attributes||[])S[U]=D;S.lhs_scale=h,S.lhs_zero_point=m,S.rhs_scale=1,S.rhs_zero_point=0,S.output_scale=1,S.output_zero_point=0;let T=new I(x,"i32"),v=new Me("quantized_dot",[w.getResult(0),_.getResult(0)],[T],S);g.insertBefore(v,e);let A=new I(x,"f32"),N=new Me("convert",[v.getResult(0)],[A],{target_dtype:"f32"});g.insertBefore(N,e);let k=o[p],O=new Array(k);for(let U=0;U<k;U++)O[U]=h*c.getScaleForChannel(U);let R=new Set(e.getAttr("lhs_contracting")||[]),B=0;for(let U=0;U<r.type.shape.length;U++)R.has(U)||B++;let E=B,F=new I([k],"f32"),C=new Me("constant",[],[F],{value:O,tensor_type:F});g.insertBefore(C,e);let H=new Me("broadcast_in_dim",[C.getResult(0)],[A],{broadcast_dimensions:[E],result_shape:x});g.insertBefore(H,e);let G=new Me("mul",[N.getResult(0),H.getResult(0)],[A],{});return g.insertBefore(G,e),e.replaceAllResultsWith([G.getResult(0)]),e.erase(),!0}};l(Z6,"hasConstantOperand");l(Q6,"allOperandsCanQuantize")});var pv,cb,nL=M(()=>{"use strict";ro();Re();pv=class{static{l(this,"LivenessResult")}constructor(e,t,r,s,o,i,a){this.liveIn=e,this.liveOut=t,this.intervals=r,this.opIndex=s,this.peakPressure=o,this.peakOp=i,this.pressureAtOp=a}interfere(e,t){if(e===t)return!0;let r=this.intervals.get(e),s=this.intervals.get(t);return r&&s?r.start<=s.end&&s.start<=r.end:!1}liveAtOp(e){return this.liveIn.get(e)||new Set}intervalOf(e){return this.intervals.get(e)||null}},cb=class n{static get name(){return"liveness"}static get depKey(){return"liveness"}static get dependencies(){return[ft]}static buildIntervals(e,t){let r=new Map;for(let o=0;o<t.length;o++)r.set(t[o],o);let s=new Map;for(let o of e.args)s.set(o,{start:-1,end:-1});for(let o=0;o<t.length;o++){let i=t[o];for(let a=0;a<i.numResults;a++)s.set(i.getResult(a),{start:o,end:o})}for(let o=0;o<t.length;o++){let i=t[o];for(let a=0;a<i.numOperands;a++){let u=i.getOperand(a),p=s.get(u);p&&p.end<o&&(p.end=o)}}return{intervals:s,opIndex:r}}static compute(e,t={}){let r=t.useDef||ft.compute(e),s=r.topologicalOrder,{intervals:o,opIndex:i}=n.buildIntervals(e,s),a=new Map,u=new Map;for(let d of s)a.set(d,new Set),u.set(d,new Set);for(let d=s.length-1;d>=0;d--){let h=s[d],m=u.get(h),g=r.opUsers.get(h);if(g)for(let _ of g){let b=a.get(_);if(b)for(let w of b)m.add(w)}let y=a.get(h);for(let _ of m)y.add(_);for(let _=0;_<h.numResults;_++)y.delete(h.getResult(_));for(let _=0;_<h.numOperands;_++)y.add(h.getOperand(_))}for(let[d,h]of i){let m=a.get(d);for(let g of m){let y=o.get(g);y&&y.end<h&&(y.end=h)}}let p=0,c=null,f=new Map;for(let d=0;d<s.length;d++){let h=s[d],m=a.get(h),g=0;for(let y of m)if(y.type instanceof I){let _=y.type.sizeInBytes();_!==X&&(g+=_)}f.set(h,g),g>p&&(p=g,c=h)}return new pv(a,u,o,i,p,c,f)}}});var pb,fb,rL=M(()=>{"use strict";fn();yr();Re();Ct();nn();ro();nL();Ir();pb=class{static{l(this,"RematerializationConfig")}constructor(e={}){this.memoryBudget=e.memoryBudget||1/0,this.maxIterations=e.maxIterations||100,this.maxRecomputeCost=e.maxRecomputeCost||1/0,this.excludeOps=e.excludeOps||new Set}},fb=class extends qe{static{l(this,"RematerializationPass")}constructor(e={}){super("RematerializationPass"),this.config=e instanceof pb?e:new pb(e)}run(e,t){if(this.config.memoryBudget===1/0)return oe.UNCHANGED;let r=!1,s=0,o=0;for(;s<this.config.maxIterations;){let i=ft.compute(e),{peakPressure:a,candidates:u,opIndex:p}=this._analyzeIntervalPressure(e,i);if(o=a,a<=this.config.memoryBudget||u.length===0)break;u.sort((c,f)=>f.score-c.score),this._rematerialize(e,u[0],p),r=!0,s++}return this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,iterations:s,peakPressure:o,budget:this.config.memoryBudget,changed:r,level:le.DEBUG}),r?oe.CHANGED:oe.UNCHANGED}_analyzeIntervalPressure(e,t){let r=t.topologicalOrder,{intervals:s,opIndex:o}=cb.buildIntervals(e,r),i=[];for(let[m,g]of s){if(!(m.type instanceof I))continue;let y=m.type.sizeInBytes();y===X||y<=0||(i.push({idx:g.start,delta:y,value:m}),i.push({idx:g.end+1,delta:-y,value:m}))}i.sort((m,g)=>m.idx-g.idx||m.delta-g.delta);let a=0,u=0,p=0,c=new Set,f=new Set,d=0;for(let m=-1;m<=r.length;m++){for(;d<i.length&&i[d].idx<=m;)a+=i[d].delta,i[d].delta>0&&i[d].value&&f.add(i[d].value),i[d].delta<0&&i[d].value&&f.delete(i[d].value),d++;if(a>u){u=a,p=m,c.clear();for(let g of f)c.add(g)}}let h=[];for(let m of c){if(!this._canRematerialize(m))continue;let g=m.definingOp,y=this._computeMemorySaved(m);if(y<=0)continue;let _=this._estimateRecomputeCost(g);_>=this.config.maxRecomputeCost||_!==0&&h.push({value:m,definingOp:g,memorySaved:y,recomputeCost:_,score:y/_})}return{peakPressure:u,peakIdx:p,candidates:h,opIndex:o}}_canRematerialize(e){if(e.isBlockArgument())return!1;let t=e.definingOp;if(!t||Xt(t.opName)||_n(t.opName)||this.config.excludeOps.has(t.opName)||t.regions.length>0||t.hasSideEffects()||e.useCount<=1)return!1;for(let r=0;r<t.numOperands;r++){let s=t.getOperand(r);if(s.definingOp&&s.definingOp.hasSideEffects())return!1}return!0}_computeMemorySaved(e){if(!(e.type instanceof I))return 0;let t=e.type.sizeInBytes();return t===X?0:t}_operandExtensionCost(e,t,r){let s=0;for(let o=0;o<e.numOperands;o++){let i=e.getOperand(o),a=0;for(let u of i.uses()){let p=r.get(u.user)||0;p>a&&(a=p)}if(t>a&&i.type instanceof I){let u=i.type.sizeInBytes();u!==X&&(s+=u)}}return s}_estimateRecomputeCost(e){let t=ne.get(e.opName);if(!t||t.isOpaque)return 1/0;if(t.getFlops){let r=t.getFlops(e);if(r>0)return r}if(t.isElementwise||t.isBroadcast||t.isView){let r=0;for(let s=0;s<e.numResults;s++){let o=e.getResult(s).type;if(o instanceof I){let i=o.numel();if(i!==X){r=i;break}}}return r||1}if(t.isReduction)for(let r=0;r<e.numOperands;r++){let s=e.getOperand(r).type;if(s instanceof I){let o=s.numel();if(o!==X)return o}}return 1/0}_rematerialize(e,t,r){let{value:s,definingOp:o}=t,i=[];for(let c of s.uses())i.push({user:c.user,operandIndex:c.operandIndex});if(i.length<=1)return;i.sort((c,f)=>(r.get(c.user)||0)-(r.get(f.user)||0));let a=r.get(i[i.length-1].user)||0,u=this._operandExtensionCost(o,a,r),p=this._computeMemorySaved(s);if(!(u>=p))for(let c=1;c<i.length;c++){let f=i[c],d=f.user;if(!d.parentBlock)continue;let h=new Me(o.opName,[...o.operands],o.results.map(m=>m.type),new Map(o.attributes));d.parentBlock.insertBefore(h,d),d.replaceOperand(f.operandIndex,h.getResult(s.resultIndex))}}}});function $o(n){return _n(n.opName)}function sL(n){let e=0;for(let t=0;t<n.numResults;t++){let r=n.getResult(t).type;if(!r||!r.isFullyStatic)continue;let s=r.sizeInBytes();s>e&&(e=s)}return e}function e4(n){let e=n.definingOp;if(e&&e.opName==="broadcast"){let t=e.getOperand(0);e=t&&t.definingOp}if(e&&$o(e)){let t=e.getAttr("value");if(typeof t=="number")return t}return 0}function fv(n){if(n.opName!=="dot"&&n.opName!=="cublas_gemm")return null;let e=n.getOperand(0).type,t=n.getOperand(1).type,r=n.getResult(0).type;if(!e||!t||!r||e.dtype!=="f32"||t.dtype!=="f32"||r.dtype!=="f32")return null;let s=n.getAttr("lhs_batch")||[],o=n.getAttr("rhs_batch")||[];if(s.length>0||o.length>0)return null;let i=n.getAttr("lhs_contracting")||[],a=n.getAttr("rhs_contracting")||[];if(i.length!==1||a.length!==1||t.rank!==2||i[0]!==e.rank-1||a[0]!==0&&a[0]!==1||!e.isFullyStatic||!t.isFullyStatic||!r.isFullyStatic)return null;let u=n.getOperand(0).definingOp,p=n.getOperand(1).definingOp;if(u&&$o(u)||p&&$o(p))return null;let c=a[0]===1,f=e.shape[e.rank-1];if((c?t.shape[1]:t.shape[0])!==f)return null;let h=1;for(let g=0;g<e.rank-1;g++)h*=e.shape[g];let m=c?t.shape[0]:t.shape[1];return h<=0||m<=0||f<=0?null:{M:h,N:m,K:f,transB:c}}function oL(n){return{labelOf:l(e=>n.get(e),"labelOf"),canMerge:l((e,t)=>Math.max(e.maxBuf||0,sL(t))<=X6,"canMerge"),onAttach:l((e,t)=>{e.maxBuf=Math.max(e.maxBuf||0,sL(t))},"onAttach")}}function db(n,e,t){let r=n.opSet,s=Ru(n.ops),{inputs:o,outputs:i,constDefs:a}=pa(r,s,{pullConstants:!0,isConstant:$o});for(let d of o)if(!d.type||!d.type.isFullyStatic)return null;for(let d of i)if(!d.type||!d.type.isFullyStatic)return null;let u=new yn(e,o.map(d=>d.type),i.map(d=>d.type)),p=new Map;for(let d=0;d<o.length;d++)p.set(o[d],u.args[d]);for(let d of a)u.entryBlock.pushOp(d.clone(p));for(let d of s)u.entryBlock.pushOp(d.clone(p));let c=i.map(d=>p.get(d));if(c.some(d=>d===void 0))return null;u.entryBlock.pushOp(new Me("return",c,[]));let f=n.ops.length===1&&t.has(n.ops[0])?n.ops[0]:null;return{part:n,subFunc:u,inputs:o,outputs:i,dotOp:f}}function Bd(n){if(t4.has(n.opName))return!0;if(n.regions)for(let e of n.regions){let t=e.entryBlock;if(t){for(let r of t.ops())if(Bd(r))return!0}}return!1}function n4(n){let e=0;for(let t=0;t<n.numResults;t++){let r=n.getResult(t).type;if(!r||!r.isFullyStatic)continue;let s=1;for(let o of r.shape)s*=o;s>e&&(e=s)}return e}function iL(n,e=256){if(n.functionCount!==1)return!1;let t=n.functions().next().value,r=new Map,s=l(o=>{if(!o)return 0;let i=r.get(o);if(i!==void 0)return i;r.set(o,0);let a=Bd(o)?n4(o):0;for(let u=0;u<o.numOperands;u++){let p=s(o.getOperand(u).definingOp);p>a&&(a=p)}return r.set(o,a),a},"maxBoundaryInSubtree");for(let o of t.ops())if(!(Xt(o.opName)||!Bd(o))){for(let i=0;i<o.numOperands;i++)if(s(o.getOperand(i).definingOp)>e)return!0}return!1}function aL(n,e,t){let r=new Map,s=0,o=l(h=>{let m=r.get(h);return m===void 0&&(m=s++,r.set(h,m)),m},"getSlot");for(let h of n.args)o(h);for(let h of t)for(let m of h.outputs)o(m);let i=[];for(let h of n.args)i.push(o(h));let a=[],u=new Set;for(let h=0;h<e.numOperands;h++){let m=e.getOperand(h),g=i.length,y=m.isBlockArgument&&m.isBlockArgument(),_=m.definingOp&&$o(m.definingOp);if(!y&&!_&&r.has(m)){let b=r.get(m);if(!u.has(b)){u.add(b),i.push(b);continue}i.push(s++),a.push({pos:g,kind:"copy",srcSlot:b});continue}if(y){i.push(s++),a.push({pos:g,kind:"copy",srcSlot:o(m)});continue}if(_){i.push(s++),a.push({pos:g,kind:"const",value:e4(m)});continue}return null}let p=[];for(let h of t){let m=[];for(let y of h.inputs){let _=r.get(y);if(_===void 0)return null;m.push(_)}let g=h.outputs.map(y=>r.get(y));p.push({name:h.subFunc.name,inputSlots:m,outputSlots:g})}let c=new Set(i),f=[],d=new Set;for(let[h,m]of r)if(!(c.has(m)||d.has(m))){if(d.add(m),!h.type||!h.type.isFullyStatic)return null;f.push({slot:m,shape:[...h.type.shape],dtype:h.type.dtype})}return{plan:{numSlots:s,argSlots:i,intermediates:f,steps:p,returnFixups:a}}}function ha(n,e=2){if(n.functionCount!==1)return null;let t=n.functions().next().value,r=t.getReturnOp();if(!r)return null;let s=[],o=new Map,i=0;for(let g of t.ops())Xt(g.opName)||$o(g)||(Bd(g)?o.set(g,"boundary#"+i++):o.set(g,"native"),s.push(g));if(i<e||s.length===0)return null;let{partitions:a,preds:u}=Nd(s,oL(o));if(a.length<2)return null;let p=ov(a,u);if(!p)return null;let c=t.name,f=[],d=new Map,h=0;for(let g of p){let y=db(g,c+"_p"+h++,d);if(!y)return null;f.push(y)}let m=aL(t,r,f);if(!m)return null;n.removeFunction(t.name);for(let g of f)n.addFunction(g.subFunc);return{plan:m.plan}}function uL(n){if(n.functionCount!==1)return null;let e=n.functions().next().value,t=e.getReturnOp();if(!t)return null;let r=[],s=new Map,o=new Map,i=0,a=0;for(let y of e.ops()){if(Xt(y.opName)||$o(y))continue;let _=fv(y);_?(s.set(y,"cublas#"+i),o.set(y,_),i++):Bd(y)?s.set(y,"boundary#"+a++):s.set(y,"native"),r.push(y)}if(i+a===0||r.length===0)return null;let{partitions:u,preds:p}=Nd(r,oL(s));if(u.length<2)return null;let c=ov(u,p);if(!c)return null;let f=e.name,d=[],h=0;for(let y of c){let _=db(y,f+"_p"+h++,o);if(!_)return null;d.push(_)}let m=aL(e,t,d);if(!m)return null;let g=new Map;for(let y of d){if(!y.dotOp)continue;let _=o.get(y.dotOp),b=y.inputs.indexOf(y.dotOp.getOperand(0)),w=y.inputs.indexOf(y.dotOp.getOperand(1)),x=y.outputs.indexOf(y.dotOp.getResult(0));b<0||w<0||x<0||g.set(y.subFunc.name,{M:_.M,N:_.N,K:_.K,transB:_.transB,aIdx:b,bIdx:w,cIdx:y.inputs.length+x})}if(g.size===0)return null;n.removeFunction(e.name);for(let y of d)n.addFunction(y.subFunc);return{plan:m.plan,cublasInfos:g}}var X6,t4,hb=M(()=>{"use strict";Us();yr();Ad();Ir();l($o,"isConstantOp");X6=32*1024;l(sL,"maxResultBytes");l(e4,"constScalarOf");l(fv,"cublasDotInfo");l(oL,"bufferLimitedConfig");l(db,"materializePartition");t4=new Set(["dot","cublas_gemm","reduce","conv","quantized_conv","scaled_dot_product_attention"]);l(Bd,"containsBoundaryOp");l(n4,"maxBoundaryResultBytes");l(iL,"hasDependentBoundaries");l(aL,"buildExecutionPlan");l(ha,"splitGraphForNative");l(uL,"splitGraphForCublas")});function cL(n){return lL.push(n),n}var lL,mb,pL=M(()=>{"use strict";fn();lL=[];l(cL,"registerBackendOpRewrite");mb=class extends qe{static{l(this,"BackendOpRewritePass")}constructor(e={}){super(e.name||"BackendOpRewritePass"),this.config=e,this.rewrites=e.rewrites||lL}run(e){let t=!1;for(let r of[...e.ops()]){let s=r.parentBlock;if(s){for(let o of this.rewrites)if(o.match(r,this.config)){o.build(r,s,this.config),t=!0;break}}}return t?oe.CHANGED:oe.UNCHANGED}}});var r4,gb,fL=M(()=>{"use strict";yr();hb();pL();r4=cL({name:"dot->cublas_gemm",match:l(n=>n.opName==="dot"&&n.numOperands===2&&fv(n)!==null,"match"),build:l((n,e)=>{let t=n.results.map(s=>s.type),r=new Me("cublas_gemm",[n.getOperand(0),n.getOperand(1)],t,new Map(n.attributes),null);e.insertBefore(r,n),n.replaceAllResultsWith(r.results),n.erase()},"build")}),gb=class extends mb{static{l(this,"CublasRewritePass")}constructor(e={}){super({...e,name:"CublasRewritePass",rewrites:[r4]})}}});function dL(n,e,{cudaMatmulChain:t=!1,context:r=null}={}){let s=r?a=>r.passesForPhase(a,n,e):a=>z_(a,n,e),o=[];for(let a of s("pre"))o.push(a);if(o.push(new ql(e)),o.push(new Ul("canonicalize",[new au,new U_({fastMath:n.optimization.fastMath}),new q_,new W_,new Co],n.optimization.maxSimplifyIterations)),n.optimization.layout&&e&&(o.push(new ab({target:e})),o.push(new Co)),n.quantization.enabled&&(o.push(new lb({...n.quantization,target:e})),o.push(new au),o.push(new Co)),n.matmulBackend!=="cublas"&&!t&&(n.fusion.epilogue!==void 0?n.fusion.epilogue:e&&e.enableEpilogueFusion)&&(o.push(new Q_({target:e})),o.push(new Co)),n.fusion.enabled){let a=n.fusion,u=a.launchOverheadUs??s4;a.strategy==="dominator"?o.push(new tb({target:e,...a})):a.strategy==="priority"?(o.push(new sb({target:e,cost:{launchOverheadUs:u},...a})),o.push(new Od({maxFusionSize:e?.maxFusionSize,...a}))):(o.push(new J_({target:e,cost:{launchOverheadUs:u},...a})),o.push(new X_({maxFusionSize:e?.maxFusionSize,...a})),o.push(new Od({maxFusionSize:e?.maxFusionSize,...a}))),o.push(new Co)}if(n.matmulBackend==="cublas"&&o.push(new gb),n.optimization.rematerialization){let a={...n.optimization.rematConfig};a.memoryBudget===void 0&&e&&e.memoryBudgetBytes>0&&(a.memoryBudget=e.memoryBudgetBytes),o.push(new fb(a))}for(let a of s("post"))o.push(a);return o}var s4,hL=M(()=>{"use strict";by();Mx();ME();LE();CE();Ex();PE();FE();zE();VE();UE();KE();QE();tL();Bx();rL();fL();ev();s4=5;l(dL,"buildGraphPipeline")});var ar,ma=M(()=>{"use strict";ar=class{static{l(this,"PrimFuncPass")}constructor(e,t=null){this.name=e,this.phase=t||e,this.snapshotPoint=null,this.trace=null}begin(e){}run(e,t){throw new Error("PrimFuncPass.run not implemented")}end(e){}}});var yb,dv,Mu,hv=M(()=>{"use strict";Fe();yb=["iterCount","depth","parallelLoops","vectorizedLoops","unrolledLoops","threadBoundLoops","serialLoops","threadBlockSize","gridSize","underReduction","numMathOps","numExternCalls","numReads","numWrites","stride1Accesses","stridedAccesses","reuseCount","touchedBytes","arithmeticIntensity","vectorized","parallelized","innermostExtent"],dv=class{static{l(this,"ScheduleFeatures")}constructor(e){this.numLoops=e.numLoops||0,this.numBlocks=e.numBlocks||0,this.totalIterations=e.totalIterations||0,this.maxLoopDepth=e.maxLoopDepth||0,this.numParallelLoops=e.numParallelLoops||0,this.numVectorizedLoops=e.numVectorizedLoops||0,this.numUnrolledLoops=e.numUnrolledLoops||0,this.numThreadBound=e.numThreadBound||0,this.numSerialLoops=e.numSerialLoops||0,this.totalBufferBytes=e.totalBufferBytes||0,this.numBufferReads=e.numBufferReads||0,this.numBufferWrites=e.numBufferWrites||0,this.numMathOps=e.numMathOps||0,this.numExternCalls=e.numExternCalls||0,this.arithmeticIntensity=e.arithmeticIntensity||0,this.innermostExtent=e.innermostExtent||0,this.outermostExtent=e.outermostExtent||0,this.hasReduction=e.hasReduction||!1,this.reductionDepth=e.reductionDepth||0,this.threadBlockSize=e.threadBlockSize||0,this.gridSize=e.gridSize||0,this.strideOneAccesses=e.strideOneAccesses||0,this.nonStrideOneAccesses=e.nonStrideOneAccesses||0}},Mu=class n{static{l(this,"FeatureExtractor")}static extract(e){let t={numLoops:0,numBlocks:0,totalIterations:1,maxLoopDepth:0,currentDepth:0,numParallelLoops:0,numVectorizedLoops:0,numUnrolledLoops:0,numThreadBound:0,numSerialLoops:0,totalBufferBytes:0,buffersSeen:new Set,numBufferReads:0,numBufferWrites:0,numMathOps:0,numExternCalls:0,hasReduction:!1,reductionDepth:0,threadBlockSize:1,gridSize:1,innermostExtent:0,outermostExtent:0,loopExtents:[],strideOneAccesses:0,nonStrideOneAccesses:0};n._visitIterative(e.body,t);let r=t.totalBufferBytes,s=t.numMathOps+t.numExternCalls;return t.arithmeticIntensity=r>0?s/r:0,new dv(t)}static extractStatements(e){let t=[],r=[],s=0,o=[{node:e.body,action:"enter"}];for(;o.length>0;){let{node:i,action:a}=o.pop();if(i){if(a==="leaveFor"){r.pop();continue}if(a==="leaveBlock"){s--;continue}switch(i.type){case"ForNode":r.push(i),o.push({node:null,action:"leaveFor"}),o.push({node:i.body,action:"enter"});break;case"BlockNode":i.initBody&&(s++,o.push({node:null,action:"leaveBlock"}),o.push({node:i.initBody,action:"enter"})),o.push({node:i.body,action:"enter"});break;case"SeqNode":for(let u=i.stmts.length-1;u>=0;u--)o.push({node:i.stmts[u],action:"enter"});break;case"AllocateNode":case"LetStmtNode":o.push({node:i.body,action:"enter"});break;case"IfThenElseNode":i.elseBody&&o.push({node:i.elseBody,action:"enter"}),o.push({node:i.thenBody,action:"enter"});break;case"BufferStoreNode":t.push(n._statementVector(i,r,s));break}}}return t}static _statementVector(e,t,r){let s=1,o=0,i=0,a=0,u=0,p=0,c=1,f=1;for(let T of t){let v=T.extent&&T.extent.type==="IntImmNode"?T.extent.value:1;switch(s*=v,T.kind){case z.PARALLEL:o++;break;case z.VECTORIZED:i++;break;case z.UNROLLED:a++;break;case z.THREAD_BINDING:u++,T.threadTag&&T.threadTag.startsWith("threadIdx")?c*=v:T.threadTag&&T.threadTag.startsWith("blockIdx")&&(f*=v);break;default:p++;break}}let d={math:0,extern:0};n._countExpr(e.value,d);let h=t.map(T=>T.loopVar.name),m=[{buffer:e.buffer,indices:e.indices}];n._collectLoads(e.value,m);let g=0,y=0,_=0,b=0;for(let T of m){let v=T.buffer&&T.buffer.sizeInBytes?T.buffer.sizeInBytes():0;v>0&&(b+=v);let A=T.indices&&T.indices.length>0?T.indices[T.indices.length-1]:null;A&&A.type==="VariableNode"?g++:y++;let N=new Set;if(T.indices)for(let k of T.indices)n._collectVars(k,N);for(let k of h)N.has(k)||_++}let w=t.length>0?t[t.length-1]:null,x=w&&w.extent&&w.extent.type==="IntImmNode"?w.extent.value:0,S={iterCount:s,depth:t.length,parallelLoops:o,vectorizedLoops:i,unrolledLoops:a,threadBoundLoops:u,serialLoops:p,threadBlockSize:c,gridSize:f,underReduction:r>0?1:0,numMathOps:d.math,numExternCalls:d.extern,numReads:m.length-1,numWrites:1,stride1Accesses:g,stridedAccesses:y,reuseCount:_,touchedBytes:b,arithmeticIntensity:b>0?(d.math+d.extern)/b:0,vectorized:i>0?1:0,parallelized:o+u>0?1:0,innermostExtent:x};return yb.map(T=>S[T]||0)}static _countExpr(e,t){if(!(!e||typeof e!="object"))switch(e.type){case"MathOpNode":t.math++,n._countExpr(e.a,t),e.b&&n._countExpr(e.b,t);break;case"CompareNode":t.math++,n._countExpr(e.a,t),n._countExpr(e.b,t);break;case"CallExternNode":t.extern++;for(let r of e.args)n._countExpr(r,t);break;default:break}}static _collectLoads(e,t){if(!(!e||typeof e!="object")){if(e.type==="BufferLoadNode"){t.push({buffer:e.buffer,indices:e.indices});return}if(e.a&&n._collectLoads(e.a,t),e.b&&n._collectLoads(e.b,t),e.args)for(let r of e.args)n._collectLoads(r,t)}}static _collectVars(e,t){if(!(!e||typeof e!="object")){if(e.type==="VariableNode"){t.add(e.name);return}if(e.a&&n._collectVars(e.a,t),e.b&&n._collectVars(e.b,t),e.args)for(let r of e.args)n._collectVars(r,t);if(e.indices)for(let r of e.indices)n._collectVars(r,t)}}static _visitIterative(e,t){let r=[{node:e,action:"enter"}];for(;r.length>0;){let{node:s,action:o}=r.pop();if(s){if(o==="leave_for"){t.currentDepth--;continue}switch(s.type){case"ForNode":{t.numLoops++,t.currentDepth++,t.currentDepth>t.maxLoopDepth&&(t.maxLoopDepth=t.currentDepth);let i=s.extent.type==="IntImmNode"?s.extent.value:1;switch(t.loopExtents.push(i),t.numLoops===1&&(t.outermostExtent=i),t.innermostExtent=i,s.kind){case z.PARALLEL:t.numParallelLoops++;break;case z.VECTORIZED:t.numVectorizedLoops++;break;case z.UNROLLED:t.numUnrolledLoops++;break;case z.THREAD_BINDING:t.numThreadBound++,s.threadTag&&s.threadTag.startsWith("threadIdx")?t.threadBlockSize*=i:s.threadTag&&s.threadTag.startsWith("blockIdx")&&(t.gridSize*=i);break;default:t.numSerialLoops++;break}t.totalIterations*=i,r.push({node:null,action:"leave_for"}),r.push({node:s.body,action:"enter"});break}case"BlockNode":t.numBlocks++,s.initBody&&(t.hasReduction=!0,t.reductionDepth=t.currentDepth);for(let i of s.reads)n._visitBuffer(i.buffer,t);for(let i of s.writes)n._visitBuffer(i.buffer,t);r.push({node:s.body,action:"enter"}),s.initBody&&r.push({node:s.initBody,action:"enter"});break;case"SeqNode":for(let i=s.stmts.length-1;i>=0;i--)r.push({node:s.stmts[i],action:"enter"});break;case"AllocateNode":n._visitBuffer(s.buffer,t),r.push({node:s.body,action:"enter"});break;case"IfThenElseNode":s.elseBody&&r.push({node:s.elseBody,action:"enter"}),r.push({node:s.thenBody,action:"enter"});break;case"LetStmtNode":r.push({node:s.body,action:"enter"});break;case"BufferStoreNode":t.numBufferWrites++,n._visitBuffer(s.buffer,t),n._checkStride(s.buffer,s.indices,t),n._visitExpr(s.value,t);break}}}}static _visitBuffer(e,t){if(!e||t.buffersSeen.has(e))return;t.buffersSeen.add(e);let r=e.sizeInBytes();r>0&&(t.totalBufferBytes+=r)}static _visitExpr(e,t){if(e)switch(e.type){case"MathOpNode":t.numMathOps++,n._visitExpr(e.a,t),e.b&&n._visitExpr(e.b,t);break;case"CallExternNode":t.numExternCalls++;for(let r of e.args)n._visitExpr(r,t);break;case"BufferLoadNode":t.numBufferReads++,n._checkStride(e.buffer,e.indices,t);break;case"CompareNode":t.numMathOps++,n._visitExpr(e.a,t),n._visitExpr(e.b,t);break;default:break}}static _checkStride(e,t,r){if(!e||!t||t.length===0)return;let s=t[t.length-1];for(;s&&s.type==="MathOpNode"&&s.b&&s.b.type==="IntImmNode";)if(s.op==="+"&&s.b.value===0)s=s.a;else if(s.op==="*"&&s.b.value===1)s=s.a;else break;s&&s.type==="VariableNode"?r.strideOneAccesses++:r.nonStrideOneAccesses++}}});function mL(n,e){if(n.length===0)return 0;let t=0;for(let r of n)t+=e[r];return t/n.length}function mv(n,e,t,r,s,o,i){let a=t.length;if(r>=s||a<o*2)return{leaf:mL(t,e)};let u=n[0].length,p=0,c=0,f=new Set(t);for(let g of t){let y=e[g];p+=y,c+=y*y}let d=null,h=null;for(let g=0;g<u;g++){let y=i[g].filter(w=>f.has(w)),_=0,b=0;for(let w=0;w<y.length-1;w++){let x=e[y[w]];_+=x,b+=x*x;let S=n[y[w]][g],T=n[y[w+1]][g];if(S===T)continue;let v=w+1,A=a-v;if(v<o||A<o)continue;let N=p-_,k=c-b,O=b-_*_/v+(k-N*N/A);(!d||O<d.cost)&&(d={f:g,thr:(S+T)/2,sIdx:w,cost:O},h=y)}}if(!d)return{leaf:mL(t,e)};let m=h;return{f:d.f,thr:d.thr,left:mv(n,e,m.slice(0,d.sIdx+1),r+1,s,o,i),right:mv(n,e,m.slice(d.sIdx+1),r+1,s,o,i)}}function o4(n){let e=n.length,t=e>0?n[0].length:0,r=new Array(t),s=[];for(let o=0;o<e;o++)s.push(o);for(let o=0;o<t;o++)r[o]=s.slice().sort((i,a)=>n[i][o]-n[a][o]);return r}function gL(n,e){for(;n&&n.leaf===void 0;)n=e[n.f]<=n.thr?n.left:n.right;return n?n.leaf:0}var Rd,yL=M(()=>{"use strict";l(mL,"leafValue");l(mv,"buildTree");l(o4,"buildFeatureOrder");l(gL,"predictTree");Rd=class n{static{l(this,"GradientBoostedTrees")}constructor(e={}){this.numTrees=e.numTrees??60,this.maxDepth=e.maxDepth??3,this.lr=e.lr??.1,this.minSamples=e.minSamples??1,this.trees=null,this.base=0}fit(e,t){let r=e.length;if(r===0)return;let s=0;for(let p of t)s+=p;s/=r;let o=new Array(r).fill(s),i=[];for(let p=0;p<r;p++)i.push(p);let a=o4(e),u=[];for(let p=0;p<this.numTrees;p++){let c=new Array(r);for(let d=0;d<r;d++)c[d]=t[d]-o[d];let f=mv(e,c,i,0,this.maxDepth,this.minSamples,a);for(let d=0;d<r;d++)o[d]+=this.lr*gL(f,e[d]);u.push(f)}this.trees=u,this.base=s}predict(e){if(!this.trees)return 0;let t=this.base;for(let r of this.trees)t+=this.lr*gL(r,e);return t}serialize(){return{trees:this.trees,base:this.base,lr:this.lr}}static deserialize(e){let t=new n({lr:e.lr});return t.trees=e.trees,t.base=e.base,t}}});function _L(n){let e=n[0].length,t=new Array(e+1).fill(0);for(let r of n)for(let s=0;s<e;s++){let o=r[s]||0;u4.has(s)?o>t[s]&&(t[s]=o):t[s]+=o}for(let r of l4)r<e&&(t[r]/=n.length);return t[e]=n.length,t}var i4,a4,u4,l4,gv,c4,_b,Oc,Bc,yv=M(()=>{"use strict";hv();yL();i4=new Set(["depth","threadBlockSize","gridSize","underReduction","vectorized","parallelized","innermostExtent"]),a4=new Set(["arithmeticIntensity"]),u4=new Set(yb.map((n,e)=>i4.has(n)?e:-1).filter(n=>n>=0)),l4=new Set(yb.map((n,e)=>a4.has(n)?e:-1).filter(n=>n>=0));l(_L,"aggregateStatements");gv=class{static{l(this,"CostEstimate")}constructor(e,t){this.score=e,this.breakdown=t}},c4={parallelism:2,vectorization:1.5,memoryCoalescing:2,occupancy:1,arithmeticIntensity:1,loopOverhead:-.5,codeSize:-.3},_b=class{static{l(this,"AnalyticalCostModel")}constructor(e,t={}){this.target=e,this._weights={...c4,...e&&e.costModelWeights?e.costModelWeights:{},...t.weights||{}}}estimate(e){let t=Mu.extract(e);return this.estimateFromFeatures(t)}score(e){return this.estimate(e).score}estimateFromFeatures(e){let t=this._scoreParallelism(e),r=this._scoreVectorization(e),s=this._scoreMemoryAccess(e),o=this._scoreOccupancy(e),i=this._scoreIntensity(e),a=this._scoreOverhead(e),u=this._scoreCodeSize(e),p=this._weights,c=p.parallelism*t+p.vectorization*r+p.memoryCoalescing*s+p.occupancy*o+p.arithmeticIntensity*i+p.loopOverhead*a+p.codeSize*u;return new gv(c,{parallelism:t,vectorization:r,memoryCoalescing:s,occupancy:o,arithmeticIntensity:i,loopOverhead:a,codeSize:u})}_scoreParallelism(e){if(this.target.isGPU()){let r=e.threadBlockSize*e.gridSize,s=this.target.maxParallelism();return Math.min(1,r/Math.max(s*.1,1))}return e.numParallelLoops/Math.max(e.numLoops,1)}_scoreVectorization(e){return e.numLoops===0?0:this.target.isGPU()?e.strideOneAccesses/Math.max(e.strideOneAccesses+e.nonStrideOneAccesses,1):e.numVectorizedLoops>0?Math.min(1,e.innermostExtent/this.target.vectorWidth):0}_scoreMemoryAccess(e){let t=e.strideOneAccesses+e.nonStrideOneAccesses;return t===0?1:e.strideOneAccesses/t}_scoreOccupancy(e){if(!this.target.isGPU())return 1;if(e.threadBlockSize===0)return 0;let t=this.target.warpSize,r=Math.ceil(e.threadBlockSize/t),s=Math.floor(this.target.maxThreadsPerBlock/t);return Math.min(1,r/s)}_scoreIntensity(e){let t=this.target.supportsFloat16?1.5:1;return Math.min(1,e.arithmeticIntensity*10*t)}_scoreOverhead(e){return e.numSerialLoops/Math.max(e.numLoops,1)}_scoreCodeSize(e){return Math.min(1,(e.numMathOps+e.numExternCalls)/256)}compare(e,t){return this.estimate(e).score-this.estimate(t).score}},Oc=class n{static{l(this,"LearnedCostModel")}constructor(e=null,t={}){this.opts={numTrees:t.numTrees??60,maxDepth:t.maxDepth??3,lr:t.lr??.1,minSamples:t.minSamples??1},this._gbt=e?Rd.deserialize(e):null,this._X=[],this._Y=[]}addSample(e,t){!e||e.length===0||Number.isFinite(t)&&(this._X.push(_L(e)),this._Y.push(t))}train(){if(this._X.length===0)return;let e=new Rd(this.opts);e.fit(this._X,this._Y),this._gbt=e}predict(e){return!this._gbt||!e||e.length===0?0:this._gbt.predict(_L(e))}get trained(){return this._gbt!==null}get sampleCount(){return this._X.length}serialize(){return{gbt:this._gbt?this._gbt.serialize():null,numSamples:this._X.length}}static deserialize(e){return new n(e&&e.gbt?e.gbt:null)}},Bc=class{static{l(this,"GuidedCostModel")}constructor(e,t,r={}){this.analytical=e,this.learned=t,this.confidenceSamples=r.confidenceSamples??8}_learnedConfident(){return this.learned&&this.learned.trained&&this.learned.sampleCount>=this.confidenceSamples}score(e){return this._learnedConfident()?this.learned.predict(Mu.extractStatements(e)):this.analytical.score(e)}}});function*bL(n){let e=[n];for(;e.length>0;){let t=e.pop();if(t){if(yield t,t.body&&e.push(t.body),t.stmts)for(let r of t.stmts)e.push(r);t.thenBody&&e.push(t.thenBody),t.elseBody&&e.push(t.elseBody),t.initBody&&e.push(t.initBody)}}}function Md(n,e){for(let t of bL(n))if(t.type==="BlockNode"&&t.name===e)return t;return null}function Fo(n){let e=[];for(let t of bL(n))t.type==="BlockNode"&&e.push(t.name);return e}function Id(n,e){let t=Tt(n,e);if(!t)return{spatial:0,reduction:0,reads:0,hasReduction:!1};let r=0,s=0;for(let o of t.loops)t.reductionLoopVars.has(o.loopVar.name)?s++:r++;return{spatial:r,reduction:s,reads:t.readBuffers.length,hasReduction:t.hasReduction}}var Rc=M(()=>{"use strict";Do();l(bL,"walkStmts");l(Md,"findBlock");l(Fo,"collectAllBlockNames");l(Id,"analyzeBlockStructure")});function TL(n){return n.isGPU()?f4:p4}function _v(n){let e=0,t=0;for(let[r,s]of n.order)r==="S"?e=Math.max(e,s+1):t=Math.max(t,s+1);return{spatialLevels:e,reductionLevels:t}}var p4,f4,wL,bv=M(()=>{"use strict";p4={name:"mlt_cpu",order:[["S",0],["S",1],["S",2],["S",3],["R",0]],roles:{S0:"parallelize",S3:"vectorize"}},f4={name:"mlt_gpu",order:[["S",0],["S",1],["S",2],["R",0]],roles:{S0:"blockIdx",S1:"threadIdx",S2:"unroll"}},wL={name:"ssrsrs_cpu",order:[["S",0],["S",1],["R",0],["S",2],["R",1],["S",3]],roles:{S0:"parallelize",S3:"vectorize",R1:"unroll"}};l(TL,"getTileStructure");l(_v,"levelCounts")});var Os,$r,bb=M(()=>{"use strict";Os=class{static{l(this,"SearchVariable")}constructor(e,t){this.name=e,this.candidates=t}sample(e){return this.candidates[e(this.candidates.length)]}},$r=class{static{l(this,"ScheduleSketch")}constructor(e,t,r){this.name=e,this.variables=t,this._apply=r}instantiate(e){return(t,r,s)=>{this._apply(t,r,s,e)}}sampleParams(e){let t={};for(let r of this.variables)t[r.name]=r.sample(e);return t}}});function d4(n){let e=[];for(let t=1;t*t<=n;t++){if(n%t!==0)continue;e.push(t);let r=n/t;r!==t&&e.push(r)}return e.sort((t,r)=>t-r),e}function h4(n,e){let t=Math.min(n.length,e.length);for(let r=0;r<t;r++)if(n[r]!==e[r])return n[r]-e[r];return n.length-e.length}function m4(n,e){let t=new Set,r=[];n.sort(h4);for(let i of n){let a=i.join(",");t.has(a)||(t.add(a),r.push(i))}if(r.length<=e)return r;let s=new Set,o=[];for(let i=0;i<e;i++){let a=Math.floor(i*(r.length-1)/(e-1));s.has(a)||(s.add(a),o.push(r[a]))}return o}function wv(n,e,t=48){if(!Number.isFinite(n)||n<1||e<=1)return[[n]];let r=[],s=t*8,o=l((i,a,u)=>{if(!(r.length>=s)){if(a===e-1){r.push([...u,i]);return}for(let p of d4(i))if(u.push(p),o(i/p,a+1,u),u.pop(),r.length>=s)return}},"rec");return o(n,0,[]),m4(r,t)}var xL=M(()=>{"use strict";l(d4,"divisorsOf");l(h4,"cmpTuple");l(m4,"selectDiverse");l(wv,"enumerateFactorizations")});function wb(n){return n.extent&&n.extent.type==="IntImmNode"?n.extent.value:null}function vL(n,e,t,r){let s=[],o=t;for(let i=0;i<r.length-1;i++){let a=n.getLoops(e).find(f=>f.loopVar.name===o);if(!a)return null;let u=1;for(let f=i+1;f<r.length;f++)u*=r[f];let[p,c]=n.split(a,u);s.push(p.loopVar.name),o=c.loopVar.name}return s.push(o),s}function g4(n,e,t,r,s){let o=l(i=>n.getLoops(e).find(a=>a.loopVar.name===i),"find");for(let[i,a]of t.order){let u=t.roles[i+a];if(!u)continue;let p=i==="S"?r:s;if(u==="parallelize"){let c=p[0]&&a<p[0].length?o(p[0][a]):null;c&&n.parallelize(c)}else if(u==="vectorize"){let c=p[p.length-1],f=c&&a<c.length?o(c[a]):null;f&&n.vectorize(f)}else if(u==="unroll")for(let c of p){let f=a<c.length?o(c[a]):null;f&&n.unroll(f)}else(u==="blockIdx"||u==="threadIdx")&&p.forEach((c,f)=>{if(f>=SL.length)return;let d=a<c.length?o(c[a]):null;d&&n.bindThread(d,`${u}.${SL[f]}`)})}}function kL(n,e,t,r,s,o){let i=[];for(let f=0;f<r.length;f++){let d=vL(n,e,r[f],o[`s${f}`]);if(!d)return;i.push(d)}let a=[];for(let f=0;f<s.length;f++){let d=vL(n,e,s[f],o[`r${f}`]);if(!d)return;a.push(d)}let u=[];for(let[f,d]of t.order){let h=f==="S"?i:a;for(let m of h)d<m.length&&u.push(m[d])}let p=new Map(n.getLoops(e).map(f=>[f.loopVar.name,f])),c=u.map(f=>p.get(f)).filter(Boolean);c.length>=2&&n.reorder(...c),g4(n,e,t,i,a)}function NL(n){let e=n.loops.filter(r=>!n.reductionLoopVars.has(r.loopVar.name)),t=n.loops.filter(r=>n.reductionLoopVars.has(r.loopVar.name));return{spatialLoops:e,reductionLoops:t}}function AL(n,e,t,r){let s=[];return n.forEach((o,i)=>s.push(new Os(`s${i}`,wv(wb(o),t)))),e.forEach((o,i)=>s.push(new Os(`r${i}`,wv(wb(o),r)))),s}function OL(n,e){let{spatialLevels:t,reductionLevels:r}=_v(e),{spatialLoops:s,reductionLoops:o}=NL(n);if(s.length===0)return null;for(let p of[...s,...o])if(wb(p)===null)return null;let i=AL(s,o,t,r),a=s.map(p=>p.loopVar.name),u=o.map(p=>p.loopVar.name);return new $r(e.name,i,(p,c,f,d)=>{kL(p,c,e,a,u,d)})}function BL(n,e){let{spatialLevels:t,reductionLevels:r}=_v(e),{spatialLoops:s,reductionLoops:o}=NL(n);if(s.length===0||o.length===0)return null;for(let p of[...s,...o])if(wb(p)===null)return null;let i=AL(s,o,t,r),a=s.map(p=>p.loopVar.name),u=o.map(p=>p.loopVar.name);return new $r(e.name,i,(p,c,f,d)=>{p.decomposeReduction(c),kL(p,`${c}_upd`,e,a,u,d)})}var SL,RL=M(()=>{"use strict";bb();xL();bv();SL=["x","y","z"];l(wb,"staticExtent");l(vL,"multiLevelSplit");l(g4,"applyRoles");l(kL,"tileBlock");l(NL,"splitLoops");l(AL,"tilingVariables");l(OL,"createMultiLevelTilingSketch");l(BL,"createSSRSRSTilingSketch")});function IL(n){return Math.min(n&&n.maxThreadsPerBlock||256,256)}function _4(n){let e=new Set;for(let t=2;t*t<=n;t++){if(n%t!==0)continue;e.add(t);let r=n/t;r>1&&r<n&&e.add(r)}return[...e].sort((t,r)=>t-r)}function EL(n){return new $r("fused",[],(e,t,r,s)=>{e.fuseConsumer(t,n);let o=e.getLoops(t);o.length>0&&e.parallelize(o[0])})}function LL(n){let e=n.loops.filter(i=>n.reductionLoopVars.has(i.loopVar.name));if(e.length!==1)return null;let t=e[0],r=t.extent&&t.extent.type==="IntImmNode"?t.extent.value:null;if(r===null)return null;let s=_4(r);if(s.length===0)return null;let o=t.loopVar.name;return new $r("rfactor",[new Os("rf_factor",s)],(i,a,u,p)=>{i.rfactor(a,o,p.rf_factor);let c=i.getLoops(`${a}_rf_p`);c.length>0&&i.parallelize(c[0])})}function CL(){return new $r("elementwise_cpu",[new Os("vector_width",y4)],(n,e,t,r)=>{let s=n.getLoops(e);if(s.length===0)return;if(s.length===1){let a=s[0].extent;if(a.type==="IntImmNode"&&a.value>=r.vector_width*2){let[u,p]=n.split(s[0],r.vector_width);n.parallelize(u),n.vectorize(p)}else n.parallelize(s[0]);return}n.parallelize(s[0]);let o=s[s.length-1],i=o.extent;if(i.type==="IntImmNode"&&i.value>=r.vector_width){let[,a]=n.split(o,r.vector_width);n.vectorize(a)}})}function DL(){return new $r("elementwise_gpu",[new Os("block_size",ML)],(n,e,t,r)=>{let s=n.getLoops(e);if(s.length===0)return;let o=s[0];for(let c=1;c<s.length;c++){let d=n.getLoops(e).find(h=>h.loopVar.name===s[c].loopVar.name);d&&o.body===d&&(o=n.fuseLoops(o,d))}let i=o.extent;if(i.type!=="IntImmNode"){n.bindThread(o,"threadIdx.x");return}let a=i.value,u=Math.min(t.maxThreadsPerBlock,1024);if(eS(n.func)&&a<=u){n.bindThread(o,"threadIdx.x");return}let p=Math.min(r.block_size,IL(t));if(a>p){let[c,f]=n.split(o,p);n.bindThread(c,"blockIdx.x"),n.bindThread(f,"threadIdx.x")}else n.bindThread(o,"threadIdx.x")})}function PL(){return new $r("reduction_cpu",[],(n,e,t,r)=>{let s=n.getLoops(e);s.length>0&&n.parallelize(s[0])})}function $L(){return new $r("reduction_gpu",[new Os("block_size",ML)],(n,e,t,r)=>{let s=n.getLoops(e);if(s.length===0)return;let o=Tt(n.func,e),i=s.filter(c=>!o||!ec(c,o));if(i.length===0)return;let a=i[0];for(let c=1;c<i.length;c++){let d=n.getLoops(e).find(h=>h.loopVar.name===i[c].loopVar.name);d&&a.body===d&&(a=n.fuseLoops(a,d))}let u=a.extent;if(u.type!=="IntImmNode"){n.bindThread(a,"threadIdx.x");return}let p=Math.min(r.block_size,IL(t));if(u.value>p){let[c,f]=n.split(a,p);n.bindThread(c,"blockIdx.x"),n.bindThread(f,"threadIdx.x")}else n.bindThread(a,"threadIdx.x")})}var ML,y4,FL=M(()=>{"use strict";Do();bb();ML=[32,64,128,256,512,1024],y4=[1,2,4,8,16];l(IL,"gpuThreadCap");l(_4,"reductionDivisors");l(EL,"createFusedTilingSketch");l(LL,"createRfactorSketch");l(CL,"createElementwiseCPUSketch");l(DL,"createElementwiseGPUSketch");l(PL,"createReductionCPUSketch");l($L,"createReductionGPUSketch")});function Tb(n){return Oy(n,Tb,(e,t,r)=>{switch(e.type){case"PrimFunc":return t.name=e.name,t.params=e.params,t.body=r(e.body),t.bufferMap=new Map(e.bufferMap),t.shapeParams=e.shapeParams,t.shapeParamMap=e.shapeParamMap instanceof Map?new Map(e.shapeParamMap):e.shapeParamMap,t._setChild("body",t.body),t;case"AllocateNode":return t.buffer=e.buffer,t.scope=e.scope,t.body=r(e.body),t._setChild("body",t.body),t;case"LetStmtNode":return t.variable=e.variable,t.value=r(e.value),t.body=r(e.body),t._setChild("body",t.body),t;case"WhileNode":return t.condVar=e.condVar,t.condBody=r(e.condBody),t.loopBody=r(e.loopBody),t._setChild("condBody",t.condBody),t._setChild("loopBody",t.loopBody),t;default:for(let s of Object.keys(e)){if(s==="_parent"||s==="_parentKey"||s==="_parentIdx")continue;let o=e[s];o instanceof Map?t[s]=new Map(o):Array.isArray(o)?t[s]=o.map(r):typeof o=="object"&&o!==null&&o.type?t[s]=r(o):t[s]=o}return t}})}function Ti(n){return Tb(n)}function Iu(n){return Tb(n)}function Tv(n){return Tb(n)}function zL(n,e,t){let r=t.get(e);if(!r)return null;let s=[],o=r._parent;for(;o&&o!==n;)o.type==="ForNode"&&s.push(o),o=o._parent;s.reverse();let i=Tv(r);for(let p=s.length-1;p>=0;p--){let c=s[p],f=new he(c.loopVar,Tv(c.min),Tv(c.extent),c.kind,i,c.threadTag);f._setChild("body",i),i=f}let a=new Map;for(let p of r.reads)a.set(p.buffer.name,p.buffer);for(let p of r.writes)a.set(p.buffer.name,p.buffer);let u=[];for(let p of n.params)a.has(p.name)&&u.push(p);return new Pi("__tune_"+e,u,i,a,[])}var Ed=M(()=>{"use strict";Fe();Gx();l(Tb,"deepCloneIR");l(Ti,"clonePrimFunc");l(Iu,"cloneTensorIR");l(Tv,"cloneBlockSubtree");l(zL,"extractBlockMini")});function kv(n,e){let t=[];for(let r=1;r<=e;r*=2)r>=n&&t.push(r);return t}function Dd(n){if(!n)return null;let e=[];for(let t of n){if(!t||t.type!=="VariableNode")return null;e.push(t.name)}return e}function b4(n,e){let t=[n];for(;t.length>0;){let r=t.pop();if(!(!r||typeof r!="object")){if(r.type==="BufferStoreNode"&&r.buffer&&r.buffer.name===e&&r.value&&r.value.type==="MathOpNode"&&r.value.op==="+")return r;if(r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody)}}return null}function Cd(n){if(!n||n.broadcastDims||n.offset&&n.offset!==0)return!1;let e=n.shape,t=n.strides;if(!t||t.length!==e.length)return!1;let r=1;for(let s=e.length-1;s>=0;s--){if(typeof e[s]!="number"||e[s]<=0||t[s]!==r)return!1;r*=e[s]}return!0}function Nv(n,e){if(!Tt(n,e))return null;let r=Md(n.body,e);if(!r||r.reads.length<2||r.writes.length<1)return null;let s=r.writes[0].buffer;if(!s||s.shape.length<2)return null;let o=s.shape.length,i=b4(r.body,s.name);if(!i)return null;let a=Dd(i.indices);if(!a||a.length!==o)return null;let u=i.value,p=l(T=>T&&T.type==="BufferLoadNode"&&T.buffer&&T.buffer.name===s.name,"isCLoad"),c=p(u.a)?u.b:p(u.b)?u.a:null;if(!c||c.type!=="MathOpNode"||c.op!=="*")return null;let f=[c.a,c.b];if(!f.every(T=>T&&T.type==="BufferLoadNode"&&T.buffer))return null;let d=a[o-1],h=a.slice(0,o-1),m=null,g=null;for(let T of f){let v=Dd(T.indices);if(!v||v.length!==o)continue;let A=!0;for(let N=0;N<o-1;N++)if(v[N]!==h[N]){A=!1;break}if(A){m=T,g=v[o-1];break}}if(!m||g==null)return null;let y=f[0]===m?f[1]:f[0],_=Dd(y.indices);if(!_)return null;let b=m.buffer,w=y.buffer;if(b.shape.length!==o||b.dtype!=="f32"||w.dtype!=="f32"||s.dtype!=="f32")return null;for(let T=0;T<o-1;T++)if(b.shape[T]!==s.shape[T])return null;let x=s.shape[o-1],S=b.shape[o-1];if(![x,S].every(T=>typeof T=="number"&&T>0))return null;if(_.length===2&&w.shape.length===2){let T;if(_[0]===g&&_[1]===d)T=!1;else if(_[0]===d&&_[1]===g)T=!0;else return null;if(T){if(w.shape[0]!==x||w.shape[1]!==S)return null}else if(w.shape[0]!==S||w.shape[1]!==x)return null;let v=1;for(let A=0;A<o-1;A++){let N=s.shape[A];if(typeof N!="number"||N<=0)return null;v*=N}return o===2?{A:b,B:w,C:s,M:v,N:x,K:S,transB:T,batch:1}:!Cd(b)||!Cd(s)?null:{A:new me(b.name,[v,S],b.dtype,b.scope),B:w,C:new me(s.name,[v,x],s.dtype,s.scope),M:v,N:x,K:S,transB:T,batch:1}}if(o>=3&&_.length===o&&w.shape.length===o){let T=a.slice(0,o-2);for(let B=0;B<o-2;B++)if(_[B]!==T[B])return null;let v;if(_[o-2]===g&&_[o-1]===d)v=!1;else if(_[o-2]===d&&_[o-1]===g)v=!0;else return null;for(let B=0;B<o-2;B++)if(w.shape[B]!==s.shape[B])return null;if(v){if(w.shape[o-2]!==x||w.shape[o-1]!==S)return null}else if(w.shape[o-2]!==S||w.shape[o-1]!==x)return null;let A=s.shape[o-2];if(typeof A!="number"||A<=0)return null;let N=1;for(let B=0;B<o-2;B++){let E=s.shape[B];if(typeof E!="number"||E<=0)return null;N*=E}if(!Cd(b)||!Cd(w)||!Cd(s))return null;let k=new me(b.name,[N,A,S],b.dtype,b.scope),O=new me(w.name,v?[N,x,S]:[N,S,x],w.dtype,w.scope),R=new me(s.name,[N,A,x],s.dtype,s.scope);return{A:k,B:O,C:R,M:A,N:x,K:S,transB:v,batch:N}}return null}function qL(n,e,t=32){let r=n.maxThreadsPerBlock||1024,s=n.warpSize||32,o=n.sharedMemoryBytes||49152,i=n.registersPerThread||255,a=4,{M:u,N:p,K:c}=e,f=kv(2,8),d=kv(2,s/2),h=kv(s/4,s/2),m=2*s,g=Math.min(r,8*s),y=4*s,_=new Set,b=[];for(let w of f)for(let x of f)if(!(w*x>y)&&!(w*x+w+x+s>i))for(let S of d)for(let T of d){let v=T*S;if(v<m||v>g||v%s!==0)continue;let A=w*S,N=x*T;if(!(A>u||N>p||A>y||N>y))for(let k of h){if(k>c||(A*k+k*N)*a>o)continue;let R=`${A}_${N}_${k}_${w}_${x}`;_.has(R)||(_.add(R),b.push({BM:A,BN:N,BK:k,TM:w,TN:x,threads:v}))}}return b.sort((w,x)=>jL(x,s)-jL(w,s)),b.slice(0,t)}function jL(n,e){let t=n.TM*n.TN,r=-Math.abs(n.TM-n.TN),s=-Math.abs(Math.log2(n.BM)-Math.log2(n.BN)),o=n.threads>=4*e&&n.threads<=8*e?1:0,i=n.BK===e/4?1:0;return o*100+t*4+r*6+s*4+i}function xb(n,e){if(!e)return null;let t=qL(n,e,256);return t.find(s=>s.BM===64&&s.BN===64&&s.BK===8&&s.TM===4&&s.TN===4)||(t.length>0?t[0]:null)}function w4(n,e,t,r){let s=Iu(n.storeValue);return Gi(s,o=>o.type==="BufferLoadNode"&&o.buffer?o.buffer.name===n.inputName?e:n.scalarConsts.has(o.buffer.name)?Iu(n.scalarConsts.get(o.buffer.name)):o:o.type==="VariableNode"?o.name===n.iv0?Iu(t):o.name===n.iv1?Iu(r):o:o)}function Av(n,e,t=null){let{A:r,B:s,C:o,M:i,N:a,K:u,transB:p}=n,c=n.batch||1,{BM:f,BN:d,BK:h,TM:m,TN:g}=e,y=d/g,_=f/m,b=y*_,w=Math.ceil(u/h),x=f*h,S=h*d,T=Math.ceil(x/b),v=Math.ceil(S/b),A=i%f!==0,N=a%d!==0,k=u%h!==0,O=x%b!==0,R=S%b!==0,B=new me("rb_As",[x],"f32","shared"),E=new me("rb_Bs",[S],"f32","shared"),F=new me("rb_acc",[m*g],"f32","local"),C=new me("rb_af",[m],"f32","local"),H=new me("rb_bf",[g],"f32","local"),G=Jt("rb_bx"),U=Jt("rb_by"),D=Jt("rb_tx"),V=Jt("rb_ty"),ue=Jt("rb_tid"),ce=Jt("rb_brow"),Le=Jt("rb_bcol"),j=Jt("rb_k0"),W=Jt("rb_bz"),ie=c>1,je=l((Ot,lo)=>ie?[W,Ot,lo]:[Ot,lo],"gIdx"),ke=l((Ot,lo)=>Jn(Fr(Ot,nt(g)),lo),"accIdx"),rt=Jt("rb_im"),zt=Jt("rb_in"),on=xi(rt,m,xi(zt,g,new J(F,[ke(rt,zt)],xv()))),Zt=Jt("rb_la"),it=Jt("rb_aidx"),ye=Jn(ce,VL(it,nt(h))),Xe=Jn(j,GL(it,nt(h))),Nn=new q(r,je(ye,Xe));if(A||k){let Ot=A?so(ye,nt(i)):null;k&&(Ot=Ot?Sv(Ot,so(Xe,nt(u))):so(Xe,nt(u))),Nn=new Se(Ot,Nn,xv())}let Qt=new J(B,[it],Nn);O&&(Qt=new Se(so(it,nt(x)),Qt));let an=vv(Zt,T,new et(it,Jn(ue,Fr(Zt,nt(b))),Qt)),un=Jt("rb_lb"),mn=Jt("rb_bidx"),cr=Jn(j,VL(mn,nt(d))),Vn=Jn(Le,GL(mn,nt(d))),Or=new q(s,p?je(Vn,cr):je(cr,Vn));if(k||N){let Ot=k?so(cr,nt(u)):null;N&&(Ot=Ot?Sv(Ot,so(Vn,nt(a))):so(Vn,nt(a))),Or=new Se(Ot,Or,xv())}let pr=new J(E,[mn],Or);R&&(pr=new Se(so(mn,nt(S)),pr));let yt=vv(un,v,new et(mn,Jn(ue,Fr(un,nt(b))),pr)),fr=Jt("rb_kk"),xt=Jt("rb_fi"),Vr=Jt("rb_fj"),Ds=xi(xt,m,new J(C,[xt],new q(B,[Jn(Fr(Jn(Fr(V,nt(m)),xt),nt(h)),fr)]))),uo=xi(Vr,g,new J(H,[Vr],new q(E,[Jn(Fr(fr,nt(d)),Jn(Fr(D,nt(g)),Vr))]))),Gr=Jt("rb_mi"),dr=Jt("rb_ni"),Br=xi(Gr,m,xi(dr,g,new J(F,[ke(Gr,dr)],Jn(new q(F,[ke(Gr,dr)]),Fr(new q(C,[Gr]),new q(H,[dr])))))),qo=xi(fr,h,new te([Ds,uo,Br])),jr=Jt("rb_kt"),Ur=new et(j,Fr(jr,nt(h)),new te([an,yt,new Fn,qo,new Fn])),Ps=vv(jr,w,Ur),ms=Jt("rb_wm"),va=Jt("rb_wn"),Oi=Jn(Jn(ce,Fr(V,nt(m))),ms),Bi=Jn(Jn(Le,Fr(D,nt(g))),va),ka=new q(F,[ke(ms,va)]),np=t?t.outBuffer:o,zu=t?w4(t,ka,Oi,Bi):ka,Na=new J(np,je(Oi,Bi),zu);if(A||N){let Ot=A?so(Oi,nt(i)):null;N&&(Ot=Ot?Sv(Ot,so(Bi,nt(a))):so(Bi,nt(a))),Na=new Se(Ot,Na)}let Aa=xi(ms,m,xi(va,g,Na)),Wo=new te([on,Ps,Aa]),Oa=new Ue(F,"local",new Ue(C,"local",new Ue(H,"local",Wo))),Ri=new et(ue,Jn(Fr(V,nt(y)),D),new et(ce,Fr(U,nt(f)),new et(Le,Fr(G,nt(d)),Oa))),Ko=Math.ceil(a/d),Mi=Math.ceil(i/f),Ba=Ld(U,"blockIdx.y",Mi,Ld(G,"blockIdx.x",Ko,Ld(V,"threadIdx.y",_,Ld(D,"threadIdx.x",y,Ri))));return ie&&(Ba=Ld(W,"blockIdx.z",c,Ba)),new Ue(B,"shared",new Ue(E,"shared",Ba))}function T4(n){let e=new Os("config_index",n.map((r,s)=>s)),t=new $r("matmul_register_block_gpu",[e],(r,s,o,i)=>{let a=Nv(r.func,s);if(!a)return;let u=n[i.config_index];if(!u)return;let p=Av(a,u);r.func.body=p,r.func._setChild&&r.func._setChild("body",p),r.func.gpuRegisterBlocked=!0});return t.configs=n,t.enumerate=()=>n.map((r,s)=>({config_index:s})),t}function x4(n){let e=Fo(n.body),t=null;for(let s of e){let o=Id(n,s);if(o.hasReduction&&o.spatial>=2&&o.reads>=2){if(t)return null;t=s}}if(!t)return null;let r=Nv(n,t);if(!r)return null;for(let s of e){if(s===t)continue;let o=Tt(n,s);if(!o||o.hasReduction||o.readBuffers.length>0)return null;for(let i of o.writeBuffers)if(i!==r.C.name)return null}return{reductionBlock:t,dims:r}}function S4(n){let e=new Map,t=new Map;qs(n,r=>{if(r.type==="BufferStoreNode"&&r.buffer){let s=r.buffer.name;t.set(s,(t.get(s)||0)+1),(!r.indices||r.indices.length===0)&&r.value&&(r.value.type==="FloatImmNode"||r.value.type==="IntImmNode")&&e.set(s,r.value)}});for(let r of[...e.keys()])(t.get(r)||0)!==1&&e.delete(r);return e}function v4(n,e){let t=null;return qs(n,r=>{if(r.type==="BufferStoreNode"&&r.buffer&&r.buffer.name===e)return t=r,Bm}),t}function k4(n,e){let t=null,r=0;return qs(n,s=>{s.type==="BufferLoadNode"&&s.buffer&&s.buffer.name===e&&(r++,t=s)}),r===1?t:null}function WL(n){let e=Fo(n.body),t=null;for(let y of e){let _=Id(n,y);if(_.hasReduction&&_.spatial>=2&&_.reads>=2){if(t)return null;t=y}}if(!t)return null;let r=Nv(n,t);if(!r)return null;let s=r.C.name,o=new Set;for(let y of e){let _=Tt(n,y);if(!_)return null;for(let b of _.writeBuffers)o.add(b)}let i=[];for(let y of e){if(y===t)continue;let _=Tt(n,y);if(_.hasReduction)return null;if(_.readBuffers.length===0){if(_.writeBuffers.every(b=>b===s))continue;return null}i.push({name:y,info:_})}if(i.length===0)return{reductionBlock:t,dims:r,epilogue:null};if(i.length!==1||(r.batch||1)!==1)return null;let a=i[0];if(a.info.writeBuffers.length!==1)return null;let u=a.info.writeBuffers[0],p=S4(n.body),c=0;for(let y of a.info.readBuffers){if(y===s){c++;continue}if(o.has(y)&&!p.has(y))return null}if(c!==1)return null;let f=Md(n.body,a.name);if(!f)return null;let d=v4(f.body,u);if(!d)return null;let h=Dd(d.indices);if(!h||h.length!==2)return null;let m=k4(d.value,s);if(!m)return null;let g=Dd(m.indices);return!g||g.length!==h.length||g.some((y,_)=>y!==h[_])?null:{reductionBlock:t,dims:r,epilogue:{outBuffer:d.buffer,storeValue:d.value,inputName:s,iv0:h[0],iv1:h[1],scalarConsts:p}}}function KL(n,e,t){let r=x4(n);if(!r)return null;let s=UL.get(n);if(s===void 0){let o=qL(t,r.dims);s=o.length>0?T4(o):null,UL.set(n,s)}return s?e===r.reductionBlock?[s]:[]:null}var nt,xv,Jt,Jn,Fr,VL,GL,so,Sv,vv,xi,Ld,UL,Sb=M(()=>{"use strict";Fe();ei();Do();bb();Rc();ts();Ed();nt=l(n=>new L(n),"I"),xv=l(()=>new de(0),"FZERO"),Jt=l(n=>new Oe(n,"i32"),"IV"),Jn=l((n,e)=>Un("+",n,e),"ADD"),Fr=l((n,e)=>Un("*",n,e),"MUL"),VL=l((n,e)=>Un("//",n,e),"DIV"),GL=l((n,e)=>Un("%",n,e),"MOD"),so=l((n,e)=>new Ke("lt",n,e),"LT"),Sv=l((n,e)=>new P("&&",n,e),"AND"),vv=l((n,e,t)=>new he(n,nt(0),nt(e),z.SERIAL,t),"forS"),xi=l((n,e,t)=>new he(n,nt(0),nt(e),z.UNROLLED,t),"forU"),Ld=l((n,e,t,r)=>new he(n,nt(0),nt(t),z.THREAD_BINDING,r,e),"forT");l(kv,"pow2Range");l(Dd,"plainVars");l(b4,"findAccStore");l(Cd,"isContiguousRowMajor");l(Nv,"matmulTileDims");l(qL,"enumerateRegisterBlockConfigs");l(jL,"goodness");l(xb,"pickFixedConfig");l(w4,"foldEpilogue");l(Av,"buildRegisterBlockedMatmul");l(T4,"createMatmulRegisterBlockGPUSketch");l(x4,"analyzePureMatmul");l(S4,"collectScalarConstBuffers");l(v4,"findEpilogueStore");l(k4,"singleBufferLoad");l(WL,"analyzeMatmulEpilogue");UL=new WeakMap;l(KL,"richMatmulSketches")});function N4(n,e){let t=[n];for(;t.length;){let r=t.pop();if(!(!r||typeof r!="object")&&(r.type==="BlockNode"&&e(r),r.body&&t.push(r.body),r.initBody&&t.push(r.initBody),r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),Array.isArray(r.stmts)))for(let s of r.stmts)t.push(s)}}function $d(n){let e=[];N4(n.body,r=>{e.push({name:r.name,node:r,reads:(r.reads||[]).map(s=>s.buffer&&s.buffer.name).filter(Boolean),writes:(r.writes||[]).map(s=>s.buffer&&s.buffer.name).filter(Boolean)})});let t=new Map;for(let r of e)for(let s of r.reads)t.has(s)||t.set(s,[]),t.get(s).push(r.name);return{blocks:e,consumersByBuf:t,byName:new Map(e.map(r=>[r.name,r]))}}function Ov(n){let e=[];for(let t of n){if(!t||t.type!=="VariableNode")return null;e.push(t.name)}return e}function Pd(n,e,t){if(!n||typeof n!="object")return t;n.type==="BufferLoadNode"&&n.buffer&&n.buffer.name===e&&t.push(n);for(let r of["a","b","expr","value","condition","thenBody","elseBody","body","initBody"])n[r]&&Pd(n[r],e,t);if(Array.isArray(n.args))for(let r of n.args)Pd(r,e,t);if(Array.isArray(n.indices))for(let r of n.indices)Pd(r,e,t);if(Array.isArray(n.stmts))for(let r of n.stmts)Pd(r,e,t);return t}function Mc(n,e){if(!n||typeof n!="object")return e;n.type==="BufferLoadNode"&&n.buffer&&n.buffer.name&&e.add(n.buffer.name);for(let t of["a","b","expr","value","condition","thenBody","elseBody","body","initBody"])n[t]&&Mc(n[t],e);if(Array.isArray(n.args))for(let t of n.args)Mc(t,e);if(Array.isArray(n.indices))for(let t of n.indices)Mc(t,e);if(Array.isArray(n.stmts))for(let t of n.stmts)Mc(t,e);return e}function vb(n,e,t,r){let s=e.byName.get(t);if(!s||s.writes.length!==1)return null;let o=s.writes[0],i=[...new Set((e.consumersByBuf.get(o)||[]).filter(w=>w!==t))];if(i.length!==1)return null;let a=i[0],u=r(n,t),p=r(n,a);if(!u||!p||p.hasReduction)return null;let c=e.byName.get(a);if(!c||c.writes.length!==1||c.writes[0]===o)return null;let f=new Set(s.reads);if(Mc(s.node.body,f),Mc(s.node.initBody,f),f.has(c.writes[0]))return null;let d=s.node.body&&s.node.body.type==="BufferStoreNode"?s.node.body:null,h=c.node.body&&c.node.body.type==="BufferStoreNode"?c.node.body:null;if(!d||!h)return null;let m=u.loops.filter(w=>!u.reductionLoopVars.has(w.loopVar.name)).map(w=>w.loopVar.name),g=Ov(d.indices);if(!g||g.join(",")!==m.join(","))return null;let y=p.loops.map(w=>w.loopVar.name);if(y.length!==m.length)return null;let _=Ov(h.indices);if(!_||_.join(",")!==y.join(","))return null;let b=Pd(h.value,o,[]);if(b.length===0)return null;for(let w of b){let x=Ov(w.indices);if(!x||x.join(",")!==_.join(","))return null}return a}var Bv=M(()=>{"use strict";l(N4,"walkBlocks");l($d,"buildBlockDAG");l(Ov,"indexVarNames");l(Pd,"collectLoads");l(Mc,"collectLoadedBuffers");l(vb,"findFusibleConsumer")});function Rv(n){return n.isGPU()?$L():PL()}function A4(n){return n.isGPU()?DL():CL()}function O4(n,e,t,r){let s=Tt(n,e);if(!s)return[Rv(t)];let o=[],i=OL(s,TL(t));if(i&&o.push(i),t.kind===Dt.CPU){let a=BL(s,wL);a&&o.push(a);let u=LL(s);u&&o.push(u);let p=r?vb(n,r,e,Tt):null;p&&o.push(EL(p))}return o.push(Rv(t)),o}function Mv(n,{priority:e=100}={}){kb.some(t=>t.derive===n.derive&&t.matches===n.matches)||(kb.push({matches:n.matches,derive:n.derive,priority:e}),kb.sort((t,r)=>t.priority-r.priority))}function HL(n,e,t,r={}){if(r.richGpu&&t.isGPU()){let i=KL(n,e,t);if(i!==null)return i}if(t.kind!==Dt.CPU&&!t.isGPU())return[];let s=Id(n,e),o=r.dag||$d(n);for(let i of kb)if(i.matches(s,t))return i.derive(n,e,t,o);return[]}var kb,YL=M(()=>{"use strict";yo();Do();Rc();bv();RL();FL();Sb();Bv();l(Rv,"reductionSketch");l(A4,"elementwiseSketch");l(O4,"deriveMultiLevel");kb=[];l(Mv,"registerSketchRule");Mv({matches:l(n=>n.hasReduction&&n.spatial>=1&&n.reads>=2,"matches"),derive:O4},{priority:10});Mv({matches:l(n=>n.hasReduction,"matches"),derive:l((n,e,t)=>[Rv(t)],"derive")},{priority:20});Mv({matches:l(()=>!0,"matches"),derive:l((n,e,t)=>[A4(t)],"derive")},{priority:30});l(HL,"deriveSketches")});function Iv(n,e,t,r,s={}){return HL(n,e,t,s)}var JL=M(()=>{"use strict";YL();l(Iv,"getSketchesForBlock")});function Nb(n){let e=new Map,t=[n];for(;t.length>0;){let r=t.pop();if(r){if(r.type==="BlockNode"&&e.set(r.name,r),r.body&&t.push(r.body),r.stmts)for(let s of r.stmts)t.push(s);r.thenBody&&t.push(r.thenBody),r.elseBody&&t.push(r.elseBody),r.initBody&&t.push(r.initBody)}}return e}function Ab(n,e,t,r=null){let s=r||Nb(n.body),o=s.get(e)||null,i=[];if(o){let a=[];for(let p of o.reads)a.push(`${p.buffer.shape.join("x")}:${p.buffer.dtype}`);for(let p of o.writes)a.push(`${p.buffer.shape.join("x")}:${p.buffer.dtype}`);i.push(a.join(","));let u=[];if(sn(o.body,u),o.initBody&&sn(o.initBody,u),i.push(u.join(";")),o.writes.length===1){let p=o.writes[0].buffer.name,c=[];for(let f of B4(s).get(p)||[]){if(f===o)continue;let d=[];sn(f.body,d),f.initBody&&sn(f.initBody,d),c.push(d.join(";"))}c.length>0&&(c.sort(),i.push(`consumers:${c.join("|")}`))}}return i.push(t.name),i.push(t.kind),R4(i.join("|"))}function B4(n){if(n.__readersByBuffer)return n.__readersByBuffer;let e=new Map;for(let t of n.values()){if(!t.reads)continue;let r=new Set;for(let s of t.reads)s.buffer&&r.add(s.buffer.name);for(let s of r){let o=e.get(s);o||(o=[],e.set(s,o)),o.push(t)}}return n.__readersByBuffer=e,e}function sn(n,e){if(!(!n||typeof n!="object"))switch(n.type){case"BufferStoreNode":if(e.push("store"),n.indices)for(let t of n.indices)sn(t,e);sn(n.value,e);return;case"BufferLoadNode":if(e.push(`load:${n.buffer.shape.join("x")}:${n.buffer.dtype}`),n.indices)for(let t of n.indices)sn(t,e);return;case"MathOpNode":e.push(`math:${n.op}`),sn(n.a,e),n.b&&sn(n.b,e);return;case"CallExternNode":e.push(`call:${n.externName}`);for(let t of n.args)sn(t,e);return;case"CompareNode":e.push(`cmp:${n.direction}`),sn(n.a,e),sn(n.b,e);return;case"CastNode":e.push(`cast:${n.fromDtype}->${n.toDtype}`),sn(n.expr,e);return;case"IfThenElseNode":sn(n.condition,e),sn(n.thenBody,e),n.elseBody&&sn(n.elseBody,e);return;case"SeqNode":for(let t of n.stmts)sn(t,e);return;case"ForNode":sn(n.body,e);return;case"BlockNode":n.initBody&&sn(n.initBody,e),sn(n.body,e);return;case"LetStmtNode":sn(n.value,e),sn(n.body,e);return;default:return}}function R4(n){let e=2166136261;for(let t=0;t<n.length;t++)e^=n.charCodeAt(t),e=e*16777619>>>0;return e.toString(16).padStart(8,"0")}var Ev=M(()=>{"use strict";l(Nb,"buildBlockMap");l(Ab,"computeWorkloadKey");l(B4,"readersByBuffer");l(sn,"collectBlockOps");l(R4,"fnv1a")});function M4(n,e){let t=n.medianMs!=null?1:0,r=e.medianMs!=null?1:0;return t!==r?r-t:t===1?n.medianMs-e.medianMs:e.score-n.score}var ZL,Fd,zd,QL=M(()=>{"use strict";Ev();ZL="mlfw-codegen-1",Fd=class{static{l(this,"TuningRecord")}constructor(e,t,r,s,o,i){this.workloadKey=e,this.sketchName=t,this.params=r,this.score=s,this.traceData=o,this.version=i,this.timestamp=Date.now(),this.medianMs=null,this.minMs=null}};l(M4,"rankRecords");zd=class n{static{l(this,"TuningDatabase")}constructor(e=1){this.version=e,this._records=new Map}computeWorkloadKey(e,t,r,s=null){return Ab(e,t,r,s)}store(e,t){let r=this._records.get(e);r||(r=[],this._records.set(e,r)),r.push(t),r.sort(M4),r.length>10&&(r.length=10)}lookup(e){let t=this._records.get(e);return!t||t.length===0?null:t[0]}lookupTopK(e,t=5){let r=this._records.get(e);return r?r.slice(0,t):[]}has(e){return this._records.has(e)&&this._records.get(e).length>0}get size(){let e=0;for(let[,t]of this._records)e+=t.length;return e}serialize(){let e=[];for(let[t,r]of this._records)for(let s of r)e.push({workloadKey:s.workloadKey,sketchName:s.sketchName,params:s.params,score:s.score,traceData:s.traceData,version:s.version,timestamp:s.timestamp,medianMs:s.medianMs,minMs:s.minMs});return{version:this.version,codegenVersion:ZL,entries:e}}static deserialize(e){let t=new n(e.version);if(e.codegenVersion!==void 0&&e.codegenVersion!==ZL)return t;for(let r of e.entries){let s=new Fd(r.workloadKey,r.sketchName,r.params,r.score,r.traceData,r.version);s.timestamp=r.timestamp,s.medianMs=r.medianMs??null,s.minMs=r.minMs??null,t.store(r.workloadKey,s)}return t}saveToFile(e,t){return t.writeFile(e,JSON.stringify(this.serialize())),e}static loadFromFile(e,t){return t.exists(e)?n.deserialize(JSON.parse(t.readFile(e))):new n}clear(){this._records.clear()}}});function ga(n,e){if(!n)return 0;if(n.type==="ForNode"){let t=n.extent&&n.extent.type==="IntImmNode"?n.extent.value:1,r=n.threadTag?e:e*t;return ga(n.body,r)}if(n.type==="BlockNode")return Math.max(e,ga(n.body,e),n.initBody?ga(n.initBody,e):0);if(n.type==="SeqNode"){let t=e;for(let r of n.stmts)t=Math.max(t,ga(r,e));return t}return n.type==="IfThenElseNode"?Math.max(ga(n.thenBody,e),n.elseBody?ga(n.elseBody,e):e):n.type==="AllocateNode"||n.type==="LetStmtNode"?ga(n.body,e):e}function XL(n,e=.1){let t=n.slice().sort((g,y)=>g-y),r=t.length;if(r===0)return{median:0,min:0,trimmedMean:0,cv:0};let s=t[r>>1],o=t[0],i=Math.floor(r*e),a=i,u=r-i>a?r-i:r,p=0,c=0;for(let g=a;g<u;g++)p+=t[g],c++;let f=c>0?p/c:s,d=0;for(let g=a;g<u;g++){let y=t[g]-f;d+=y*y}let h=c>1?Math.sqrt(d/(c-1)):0,m=f>0?h/f:0;return{median:s,min:o,trimmedMean:f,cv:m}}var I4,Ob,Bb,e2=M(()=>{"use strict";Ep();I4=1e6;l(ga,"maxSerialTripCount");Ob=class{static{l(this,"BenchmarkResult")}constructor(e,t,r,s,o=null,i=0){this.medianMs=e,this.minMs=t,this.trimmedMeanMs=o??e,this.cv=i,this.samples=r,this._totalBytes=s}get throughputGBs(){return this.minMs<=0||!this._totalBytes?0:this._totalBytes/(this.minMs*1e6)}};l(XL,"robustStats");Bb=class{static{l(this,"BenchmarkRunner")}constructor(e,t={}){this.target=e,this.warmup=t.warmup??3,this.repeat=t.repeat??10,this.minRepeatMs=t.minRepeatMs??0,this.maxCv=t.maxCv??0,this.maxReMeasures=t.maxReMeasures??1,this.measurer=t.measurer||null,this._warn=t.warn||null,this._bufferCache=new Map}_record(e,t){this._warn&&this._warn(e,null,t)}_getOrAllocBuffers(e){let t=0,r=[];for(let[,i]of e.bufferMap){let a=Math.max(i.numel(),1);r.push(a);let u=i.sizeInBytes();u>0&&(t+=u)}let s=r.join(","),o=this._bufferCache.get(s);o||(o=r.map(i=>new Float32Array(i)),this._bufferCache.set(s,o));for(let i of o)for(let a=0;a<i.length;a++)i[a]=Math.random()*2-1;return{buffers:o,totalBytes:t}}run(e){if(!e||!e.body)return null;if(!this.target.isCPU())return this.measurer?this._runMeasured(e):null;let t=new bo(this.target),r;try{r=t.compile(e)}catch(p){return this._record("benchmark-compile",p),null}if(r.metadata.kind!=="js")return null;let s;try{s=new Function("return "+r.source)()}catch(p){return this._record("benchmark-construct-fn",p),null}let{buffers:o,totalBytes:i}=this._getOrAllocBuffers(e);for(let p=0;p<this.warmup;p++)try{s(...o)}catch(c){return this._record("benchmark-warmup-run",c),null}let a=[],u=null;for(let p=0;p<=this.maxReMeasures&&(this._collect(s,o,a),u=XL(a),!(this.maxCv<=0||u.cv<=this.maxCv));p++);return new Ob(u.median,u.min,a,i,u.trimmedMean,u.cv)}_runMeasured(e){if(e.shapeParams&&e.shapeParams.length>0||ga(e.body,1)>I4)return null;let t;try{t=new bo(this.target).compile(e)}catch(a){return this._record("measured-compile",a),null}let r=[],s=0;for(let[,a]of e.bufferMap){let u=Math.max(a.sizeInBytes(),1);r.push(u),s+=u}let o;try{o=this.measurer(t,r,[],{warmup:this.warmup,repeat:this.repeat})}catch(a){return this._record("measurer",a),null}if(!o||o.length===0)return null;let i=XL(o);return new Ob(i.median,i.min,o,s,i.trimmedMean,i.cv)}_collect(e,t,r){let s=0,o=this.repeat*3;for(let i=0;i<o&&(i<this.repeat||s<this.minRepeatMs);i++){let a=performance.now();e(...t);let u=performance.now()-a;r.push(u),s+=u}}}});var Rb,t2=M(()=>{"use strict";Rb=class{static{l(this,"Deadline")}constructor(e=1/0,t=null){this._clock=t||(()=>performance.now()),this._endAt=e===1/0||e==null?1/0:this._clock()+e}get expired(){return this._clock()>=this._endAt}remainingMs(){return this._endAt===1/0?1/0:Math.max(0,this._endAt-this._clock())}}});function n2(n){let e=null;for(let w of n.ops())if(w.opName==="conv"){if(e)return null;e=w}else if(w.opName!=="return")return null;if(!e||(e.getAttr("groups")||1)!==1)return null;let t=e.getAttr("input_layout"),r=e.getAttr("kernel_layout");if(t&&t!=="NCHW"||r&&r!=="OIHW")return null;let s=e.getOperand(0).type,o=e.getOperand(1).type,i=e.getResult(0).type;if(!s||!o||!i||s.shape.length!==4||o.shape.length!==4||i.shape.length!==4||s.dtype!=="f32"||o.dtype!=="f32"||i.dtype!=="f32")return null;let a=e.getAttr("strides")||[1,1],u=e.getAttr("padding")||[[0,0],[0,0]],p=e.getAttr("dilation")||[1,1],c=s.shape[0],f=s.shape[1],d=s.shape[2],h=s.shape[3],m=o.shape[0],g=o.shape[2],y=o.shape[3],_=i.shape[2],b=i.shape[3];return![c,f,d,h,m,g,y,_,b].every(w=>typeof w=="number"&&w>0)||o.shape[1]!==f||i.shape[1]!==m||i.shape[0]!==c?null:{N:c,Cin:f,H:d,W:h,O:m,Kh:g,Kw:y,Oh:_,Ow:b,sH:a[0],sW:a[1],pH:u[0][0],pW:u[1][0],dH:p[0],dW:p[1]}}function E4(n,e,t){let{weight:r,input:s,output:o}=n,{N:i,Cin:a,H:u,W:p,O:c,Kh:f,Kw:d,Oh:h,Ow:m,sH:g,sW:y,pH:_,pW:b,dH:w,dW:x}=e,{BM:S,BN:T,BK:v,TM:A,TN:N}=t,k=c,O=i*h*m,R=a*f*d,B=f*d,E=h*m,F=T/N,C=S/A,H=F*C,G=Math.ceil(R/v),U=S*v,D=v*T,V=Math.ceil(U/H),ue=Math.ceil(D/H),ce=new me("ig_As",[U],"f32","shared"),Le=new me("ig_Bs",[D],"f32","shared"),j=new me("ig_acc",[A*N],"f32","local"),W=new me("ig_af",[A],"f32","local"),ie=new me("ig_bf",[N],"f32","local"),je=Be("ig_bx"),ke=Be("ig_by"),rt=Be("ig_tx"),zt=Be("ig_ty"),on=Be("ig_tid"),Zt=Be("ig_brow"),it=Be("ig_bcol"),ye=Be("ig_k0"),Xe=l((Hr,Jw)=>Te(xe(Hr,$(N)),Jw),"accIdx"),Nn=Be("ig_im"),Qt=Be("ig_in"),an=kn(Nn,A,kn(Qt,N,new J(j,[Xe(Nn,Qt)],Vd()))),un=k%S!==0,mn=R%v!==0,cr=O%T!==0,Vn=U%H!==0,Or=D%H!==0,pr=Be("ig_la"),yt=Be("ig_aidx"),fr=Te(Zt,Dn(yt,$(v))),xt=Te(ye,hn(yt,$(v))),Vr=Dn(xt,$(B)),Ds=hn(xt,$(B)),uo=Dn(Ds,$(d)),Gr=hn(Ds,$(d)),dr=new q(r,[fr,Vr,uo,Gr]);if(un||mn){let Hr=un?kr(fr,$(k)):null;mn&&(Hr=Hr?zo(Hr,kr(xt,$(R))):kr(xt,$(R))),dr=new Se(Hr,dr,Vd())}let Br=new J(ce,[yt],dr);Vn&&(Br=new Se(kr(yt,$(U)),Br));let qo=Mb(pr,V,new et(yt,Te(on,xe(pr,$(H))),Br)),jr=Be("ig_lb"),Ur=Be("ig_bidx"),Ps=Te(ye,Dn(Ur,$(T))),ms=Te(it,hn(Ur,$(T))),va=Dn(Ps,$(B)),Oi=hn(Ps,$(B)),Bi=Dn(Oi,$(d)),ka=hn(Oi,$(d)),np=Dn(ms,$(E)),zu=hn(ms,$(E)),Na=Dn(zu,$(m)),Aa=hn(zu,$(m)),Wo=Ib(Te(xe(Na,$(g)),xe(Bi,$(w))),$(_)),Oa=Ib(Te(xe(Aa,$(y)),xe(ka,$(x))),$(b)),Ri=new q(s,[np,va,Wo,Oa]),Ko=zo(zo(Eb(Wo,$(0)),kr(Wo,$(u))),zo(Eb(Oa,$(0)),kr(Oa,$(p))));mn&&(Ko=zo(Ko,kr(Ps,$(R)))),cr&&(Ko=zo(Ko,kr(ms,$(O)))),Ri=new Se(Ko,Ri,Vd());let Mi=new J(Le,[Ur],Ri);Or&&(Mi=new Se(kr(Ur,$(D)),Mi));let Ba=Mb(jr,ue,new et(Ur,Te(on,xe(jr,$(H))),Mi)),Ot=Be("ig_kk"),lo=Be("ig_fi"),rp=Be("ig_fj"),Gw=kn(lo,A,new J(W,[lo],new q(ce,[Te(xe(Te(xe(zt,$(A)),lo),$(v)),Ot)]))),jw=kn(rp,N,new J(ie,[rp],new q(Le,[Te(xe(Ot,$(T)),Te(xe(rt,$(N)),rp))]))),Vu=Be("ig_mi"),Gu=Be("ig_ni"),Uw=kn(Vu,A,kn(Gu,N,new J(j,[Xe(Vu,Gu)],Te(new q(j,[Xe(Vu,Gu)]),xe(new q(W,[Vu]),new q(ie,[Gu])))))),qw=kn(Ot,v,new te([Gw,jw,Uw])),$s=Be("ig_kt"),gs=new et(ye,xe($s,$(v)),new te([qo,Ba,new Fn,qw,new Fn])),Rr=Mb($s,G,gs),hr=Be("ig_wm"),qr=Be("ig_wn"),Ii=Te(Te(Zt,xe(zt,$(A))),hr),co=Te(Te(it,xe(rt,$(N))),qr),Wr=Dn(co,$(E)),Kr=hn(co,$(E)),Ho=Dn(Kr,$(m)),Qn=hn(Kr,$(m)),Ei=new J(o,[Wr,Ii,Ho,Qn],new q(j,[Xe(hr,qr)]));if(un||cr){let Hr=un?kr(Ii,$(k)):null;cr&&(Hr=Hr?zo(Hr,kr(co,$(O))):kr(co,$(O))),Ei=new Se(Hr,Ei)}let Yo=kn(hr,A,kn(qr,N,Ei)),Ww=new te([an,Rr,Yo]),kh=new Ue(j,"local",new Ue(W,"local",new Ue(ie,"local",Ww))),Kw=new et(on,Te(xe(zt,$(F)),rt),new et(Zt,xe(ke,$(S)),new et(it,xe(je,$(T)),kh))),Hw=Math.ceil(O/T),Yw=Math.ceil(k/S),Nh=ya(ke,"blockIdx.y",Yw,ya(je,"blockIdx.x",Hw,ya(zt,"threadIdx.y",C,ya(rt,"threadIdx.x",F,Kw))));return new Ue(ce,"shared",new Ue(Le,"shared",Nh))}function r2(n,e,t){if(!s2(n,{M:e.O,N:e.N*e.Oh*e.Ow,K:e.Cin*e.Kh*e.Kw},t))return!1;let{BM:r,BN:s,BK:o,TM:i,TN:a}=t,u=e.O,p=e.N*e.Oh*e.Ow,c=e.Cin*e.Kh*e.Kw;if(u%r!==0||p%s!==0||c%o!==0||[r,s,o,i,a].some(d=>d%Lv!==0))return!1;let f=r/i*(s/a);return!(r*o%(Lv*f)!==0||o*s%f!==0)}function L4(n,e,t){let{weight:r,input:s,output:o}=n,{N:i,Cin:a,H:u,W:p,O:c,Kh:f,Kw:d,Oh:h,Ow:m,sH:g,sW:y,pH:_,pW:b,dH:w,dW:x}=e,{BM:S,BN:T,BK:v,TM:A,TN:N}=t,k=Lv,O=c,R=i*h*m,B=a*f*d,E=f*d,F=h*m,C=u*p,H=a*u*p,G=T/N,U=S/A,D=G*U,V=B/v,ue=S*v,ce=v*T,Le=ue/D,j=ce/D,W=Le/k,ie=v/k,je=(h-1)*g+(f-1)*w-_,ke=(m-1)*y+(d-1)*x-b,rt=_>0||je>=u,zt=b>0||ke>=p,on=new me(r.name,[O*B>0?O*B:1],r.dtype,r.scope),Zt=new me(s.name,[i*H>0?i*H:1],s.dtype,s.scope),it=new me("iv_As",[2*ue],"f32","shared");it.align16=!0;let ye=new me("iv_Bs",[2*ce],"f32","shared");ye.align16=!0;let Xe=new me("iv_acc",[A*N],"f32","local"),Nn=new me("iv_af",[A],"f32","local");Nn.align16=!0;let Qt=new me("iv_bf",[N],"f32","local");Qt.align16=!0;let an=new me("iv_ra",[Le],"f32","local");an.align16=!0;let un=new me("iv_rb",[j],"f32","local"),mn=Be("iv_bx"),cr=Be("iv_by"),Vn=Be("iv_tx"),Or=Be("iv_ty"),pr=Be("iv_tid"),yt=Be("iv_brow"),fr=Be("iv_bcol"),xt=l(($s,gs)=>Te(xe($s,$(N)),gs),"accIdx"),Vr=0,Ds=Be("iv_im"),uo=Be("iv_in"),Gr=kn(Ds,A,kn(uo,N,new J(Xe,[xt(Ds,uo)],Vd()))),dr=l($s=>{let gs=Vr++,Rr=Be("iv_la"+gs),hr=Be("iv_c"+gs),qr=Dn(hr,$(ie)),Ii=xe(hn(hr,$(ie)),$(k)),co=kn(Rr,W,new et(hr,Te(pr,xe(Rr,$(D))),new al(an,xe(Rr,$(k)),on,Te(xe(Te(yt,qr),$(B)),Te($s,Ii)),k))),Wr=Be("iv_lb"+gs),Kr=Be("iv_e"+gs),Ho=Dn(Kr,$(T)),Qn=hn(Kr,$(T)),Ei=Te(fr,Qn),Yo=Te($s,Ho),Ww=Dn(Yo,$(E)),kh=hn(Yo,$(E)),Kw=Dn(kh,$(d)),Hw=hn(kh,$(d)),Yw=Dn(Ei,$(F)),Nh=hn(Ei,$(F)),Hr=Dn(Nh,$(m)),Jw=hn(Nh,$(m)),Zw=Ib(Te(xe(Hr,$(g)),xe(Kw,$(w))),$(_)),Qw=Ib(Te(xe(Jw,$(y)),xe(Hw,$(x))),$(b)),QD=Te(Te(Te(xe(Yw,$(H)),xe(Ww,$(C))),xe(Zw,$(p))),Qw),Xw=new q(Zt,[QD]),ju=null;if(rt&&(ju=zo(Eb(Zw,$(0)),kr(Zw,$(u)))),zt){let hN=zo(Eb(Qw,$(0)),kr(Qw,$(p)));ju=ju?zo(ju,hN):hN}ju&&(Xw=new Se(ju,Xw,Vd()));let XD=kn(Wr,j,new et(Kr,Te(pr,xe(Wr,$(D))),new J(un,[Wr],Xw)));return new te([co,XD])},"prefetch"),Br=l(($s,gs)=>{let Rr=Vr++,hr=Be("iv_la"+Rr),qr=Be("iv_c"+Rr),Ii=Dn(qr,$(ie)),co=xe(hn(qr,$(ie)),$(k)),Wr=[];for(let Yo=0;Yo<k;Yo++)Wr.push(new J(it,[Te($s,Te(xe(Te(co,$(Yo)),$(S)),Ii))],new q(an,[Te(xe(hr,$(k)),$(Yo))])));let Kr=kn(hr,W,new et(qr,Te(pr,xe(hr,$(D))),new te(Wr))),Ho=Be("iv_lb"+Rr),Qn=Be("iv_e"+Rr),Ei=kn(Ho,j,new et(Qn,Te(pr,xe(Ho,$(D))),new J(ye,[Te(gs,Qn)],new q(un,[Ho]))));return new te([Kr,Ei])},"commit"),qo=l(($s,gs)=>{let Rr=Vr++,hr=Be("iv_kk"+Rr),qr=[],Ii=Te($s,Te(xe(hr,$(S)),xe(Or,$(A))));for(let Qn=0;Qn<A/k;Qn++)qr.push(new al(Nn,$(Qn*k),it,Te(Ii,$(Qn*k)),k));let co=Te(gs,Te(xe(hr,$(T)),xe(Vn,$(N))));for(let Qn=0;Qn<N/k;Qn++)qr.push(new al(Qt,$(Qn*k),ye,Te(co,$(Qn*k)),k));let Wr=Be("iv_mi"+Rr),Kr=Be("iv_ni"+Rr),Ho=kn(Wr,A,kn(Kr,N,new J(Xe,[xt(Wr,Kr)],Te(new q(Xe,[xt(Wr,Kr)]),xe(new q(Nn,[Wr]),new q(Qt,[Kr]))))));return kn(hr,v,new te([...qr,Ho]))},"computeMMA"),jr=Be("iv_kt"),Ur=Be("iv_p"),Ps=Be("iv_pN"),ms=xe(Ur,$(ue)),va=xe(Ur,$(ce)),Oi=xe(Ps,$(ue)),Bi=xe(Ps,$(ce)),ka=kr(jr,$(V-1)),np=new te([dr($(0)),Br($(0),$(0)),new Fn]),zu=new et(Ur,hn(jr,$(2)),new et(Ps,hn(Te(jr,$(1)),$(2)),new te([new Se(ka,dr(xe(Te(jr,$(1)),$(v)))),qo(ms,va),new Se(ka,new te([Br(Oi,Bi),new Fn]))]))),Na=Mb(jr,V,zu),Aa=Be("iv_wm"),Wo=Be("iv_wn"),Oa=Te(Te(yt,xe(Or,$(A))),Aa),Ri=Te(Te(fr,xe(Vn,$(N))),Wo),Ko=Dn(Ri,$(F)),Mi=hn(Ri,$(F)),Ba=Dn(Mi,$(m)),Ot=hn(Mi,$(m)),lo=new J(o,[Ko,Oa,Ba,Ot],new q(Xe,[xt(Aa,Wo)])),rp=kn(Aa,A,kn(Wo,N,lo)),Gw=new te([Gr,np,Na,rp]),jw=new Ue(Xe,"local",new Ue(Nn,"local",new Ue(Qt,"local",new Ue(an,"local",new Ue(un,"local",Gw))))),Vu=new et(pr,Te(xe(Or,$(G)),Vn),new et(yt,xe(cr,$(S)),new et(fr,xe(mn,$(T)),jw))),Gu=R/T,Uw=O/S,qw=ya(cr,"blockIdx.y",Uw,ya(mn,"blockIdx.x",Gu,ya(Or,"threadIdx.y",U,ya(Vn,"threadIdx.x",G,Vu))));return new Ue(it,"shared",new Ue(ye,"shared",qw))}function s2(n,e,t){let{BM:r,BN:s,BK:o,TM:i,TN:a}=t;if(![r,s,o,i,a].every(m=>typeof m=="number"&&m>0)||r%i!==0||s%a!==0)return!1;let u=s/a,p=r/i,c=u*p,f=n.warpSize||32;return!(c%f!==0||c>(n.maxThreadsPerBlock||1024)||(r*o+o*s)*4*2>(n.sharedMemoryBytes||49152)||i*a+i+a+f>(n.registersPerThread||255))}function D4(n,e){for(let t of C4)if(r2(n,e,t))return t;return null}function o2(n,e,t){let r=n.func,s=r.convInfo;if(!s)return!1;let i=Fo(r.body).find(_=>/^conv_acc_/.test(_));if(!i)return!1;let a=Md(r.body,i);if(!a||a.reads.length<2||a.writes.length<1)return!1;let u=a.reads[0].buffer,p=a.reads[1].buffer,c=a.writes[0].buffer;if(!u||!p||!c)return!1;let f=s.O,d=s.N*s.Oh*s.Ow,h=s.Cin*s.Kh*s.Kw;if(h<128||d<64)return!1;let m=t&&t.convConfig,g=t&&t.convNoVec,y=null;if(!g){let _=m?r2(e,s,m)?m:null:D4(e,s);_&&(y=L4({weight:p,input:u,output:c},s,_))}if(!y){let _=m?s2(e,{M:f,N:d,K:h},m)?m:null:xb(e,{M:f,N:d,K:h});if(!_)return!1;y=E4({weight:p,input:u,output:c},s,_)}return n.func.body=y,n.func._setChild&&n.func._setChild("body",y),n.func.gpuRegisterBlocked=!0,!0}var $,Vd,Be,Te,Ib,xe,Dn,hn,kr,Eb,zo,Mb,kn,ya,Lv,C4,Cv=M(()=>{"use strict";Fe();ei();Rc();Sb();$=l(n=>new L(n),"I"),Vd=l(()=>new de(0),"FZERO"),Be=l(n=>new Oe(n,"i32"),"IV"),Te=l((n,e)=>new P("+",n,e),"ADD"),Ib=l((n,e)=>new P("-",n,e),"SUB"),xe=l((n,e)=>new P("*",n,e),"MUL"),Dn=l((n,e)=>new P("//",n,e),"DIV"),hn=l((n,e)=>new P("%",n,e),"MOD"),kr=l((n,e)=>new Ke("lt",n,e),"LT"),Eb=l((n,e)=>new Ke("ge",n,e),"GE"),zo=l((n,e)=>new P("&&",n,e),"AND"),Mb=l((n,e,t)=>new he(n,$(0),$(e),z.SERIAL,t),"forS"),kn=l((n,e,t)=>new he(n,$(0),$(e),z.UNROLLED,t),"forU"),ya=l((n,e,t,r)=>new he(n,$(0),$(t),z.THREAD_BINDING,r,e),"forT");l(n2,"detectPureConv");l(E4,"buildImplicitGemmConv");Lv=4;l(r2,"vectorizableConvConfig");l(L4,"buildVectorizedImplicitGemmConv");l(s2,"validConvConfig");C4=[{BM:128,BN:64,BK:8,TM:8,TN:8},{BM:64,BN:64,BK:8,TM:8,TN:8},{BM:64,BN:64,BK:8,TM:4,TN:8},{BM:64,BN:32,BK:8,TM:8,TN:8},{BM:32,BN:64,BK:8,TM:4,TN:8}];l(D4,"pickVectorizedConvConfig");l(o2,"applyImplicitGemmConv")});function i2(n,e,t,r,s,o,i,a){let u=new me(n,[t,r],e.dtype,"shared"),p=new J(u,[s,o],new q(e,[i,a]));return{tile:u,fill:p}}function F4(n,e,t){let{A:r,B:s,C:o,M:i,N:a,K:u,transB:p}=n,c=u/t,f=Ic("ts_by"),d=Ic("ts_bx"),h=Ic("ts_ty"),m=Ic("ts_tx"),g=Ic("ts_ko"),y=Ic("ts_ki"),_=new me("ts_acc",[1],"f32","local"),b=Ec(Lb(f,ds(e)),h),w=Ec(Lb(d,ds(e)),m),x=Lb(g,ds(t)),S=i2("ts_As",r,e,t,h,m,b,Ec(x,m)),T=i2("ts_Bs",s,t,e,h,m,p?w:Ec(x,h),p?Ec(x,h):w),v=$4(y,t,new J(_,[ds(0)],Ec(new q(_,[ds(0)]),Lb(new q(S.tile,[h,y]),new q(T.tile,[y,m]))))),A=new te([S.fill,T.fill,new Fn,v,new Fn]),N=P4(g,c,A),k=new J(_,[ds(0)],new de(0)),O=new J(o,[b,w],new q(_,[ds(0)])),R=new Ue(_,"local",new te([k,N,O]));return Cb(f,"blockIdx.y",i/e,Cb(d,"blockIdx.x",a/e,new Ue(S.tile,"shared",new Ue(T.tile,"shared",Cb(h,"threadIdx.y",e,Cb(m,"threadIdx.x",e,R))))))}function z4(n,e){let{M:t,N:r,K:s}=e,o=16,i=16;return o*o>(n.maxThreadsPerBlock||1024)||(o*i+i*o)*4>(n.sharedMemoryBytes||49152)||t%o!==0||r%o!==0||s%i!==0||t<o||r<o||s<i?null:{BS:o,BK:i}}function V4(n,e){let r=Fo(n.func.body).filter(o=>/^q?conv_(init|acc)_/.test(o));if(r.length===0)return!1;let s=new to(e);for(let o of r)s.applyToBlock(n,o);return!0}function G4(n,e,t={}){if(!e.isGPU())return!1;let r=WL(n.func);if(!r)return!1;let s=r.dims;if(t&&t.primitiveMatmul&&(s.batch||1)===1&&!r.epilogue){let a=z4(e,s);if(a){let u=F4(s,a.BS,a.BK);return n.func.body=u,n.func._setChild&&n.func._setChild("body",u),n.func.gpuRegisterBlocked=!0,!0}}let o=xb(e,s);if(!o)return!1;let i=Av(s,o,r.epilogue);return n.func.body=i,n.func._setChild&&n.func._setChild("body",i),n.func.gpuRegisterBlocked=!0,!0}function Db(n,e,t={}){if(!e.isGPU()||e.isWebGPU&&e.isWebGPU())return!1;let r=G4(n,e,t);return r||(r=o2(n,e,t)),r||(r=V4(n,e)),r}var ds,Ic,Ec,Lb,P4,$4,Cb,Dv=M(()=>{"use strict";Sb();Rc();Cv();Do();Fe();ei();ds=l(n=>new L(n),"I"),Ic=l(n=>new Oe(n,"i32"),"IV"),Ec=l((n,e)=>new P("+",n,e),"ADD"),Lb=l((n,e)=>new P("*",n,e),"MUL"),P4=l((n,e,t)=>new he(n,ds(0),ds(e),z.SERIAL,t),"forS"),$4=l((n,e,t)=>new he(n,ds(0),ds(e),z.UNROLLED,t),"forU"),Cb=l((n,e,t,r)=>new he(n,ds(0),ds(t),z.THREAD_BINDING,r,e),"forT");l(i2,"stageTileToShared");l(F4,"buildTiledSharedMatmul");l(z4,"pickTiledConfig");l(V4,"applyDeterministicGpuConv");l(G4,"applyDeterministicGpuMatmul");l(Db,"applyDeterministicGpuSchedule")});function Pv(n){return n*1664525+1013904223&2147483647}function U4(n,e,t){let r={...n};for(let s of e.variables)t.rngFloat()<t.mutationRate&&(r[s.name]=s.sample(o=>t.rng(o)));return r}function a2(n={}){return n.strategy==="random"?new $v({numTrials:n.numTrials,seed:n.seed,deadline:n.deadline}):new Fv({populationSize:n.populationSize,numGenerations:n.numGenerations,mutationRate:n.mutationRate,eliteRatio:n.eliteRatio,seed:n.seed,deadline:n.deadline})}var Pb,$v,Fv,j4,u2=M(()=>{"use strict";l(Pv,"nextLcg");Pb=class{static{l(this,"SearchCandidate")}constructor(e,t,r){this.sketchName=e,this.params=t,this.score=r}},$v=class{static{l(this,"RandomSearch")}constructor(e={}){this.numTrials=e.numTrials||64,this.seed=e.seed||42,this.deadline=e.deadline||null,this._rngState=this.seed}_rng(e){return this._rngState=Pv(this._rngState),this._rngState%e}_expired(){return this.deadline?this.deadline.expired:!1}search(e,t){let r=[];for(let s of e){if(this._expired())break;for(let o=0;o<this.numTrials&&!this._expired();o++){let i=s.sampleParams(u=>this._rng(u)),a=t(s,i);a&&r.push(new Pb(s.name,i,a.score))}}return r.sort((s,o)=>o.score-s.score),{candidates:r,population:null}}},Fv=class{static{l(this,"EvolutionarySearch")}constructor(e={}){this.populationSize=e.populationSize||32,this.numGenerations=e.numGenerations||10,this.mutationRate=e.mutationRate||.3,this.eliteRatio=e.eliteRatio||.2,this.seed=e.seed||42,this.deadline=e.deadline||null,this._rngState=this.seed}_rng(e){return this._rngState=Pv(this._rngState),this._rngState%e}_rngFloat(){return this._rngState=Pv(this._rngState),this._rngState/2147483647}_expired(){return this.deadline?this.deadline.expired:!1}search(e,t,r=null){let s=r&&r.length?r:this._initPopulation(e),o=new Map,i=l((u,p)=>{let c=u.name+"|"+JSON.stringify(p);if(o.has(c))return o.get(c);let f=t(u,p);return o.set(c,f),f},"evalMemo");for(let u=0;u<this.numGenerations&&!this._expired();u++){let p=[];for(let h of s){let m=i(h.sketch,h.params);m&&p.push({...h,score:m.score})}if(p.sort((h,m)=>m.score-h.score),p.length===0)break;let c=Math.max(1,Math.floor(p.length*this.eliteRatio)),f=p.slice(0,c),d=f.map(h=>({sketch:h.sketch,params:{...h.params}}));for(;d.length<this.populationSize;){let h=f[this._rng(f.length)],m=f[this._rng(f.length)];if(h.sketch!==m.sketch){let g=this._rngFloat()<.5?h:m;d.push(this._mutate(g.sketch,{...g.params}))}else d.push(this._mutate(h.sketch,this._crossover(h,m)))}s=d}let a=[];for(let u of s){let p=i(u.sketch,u.params);p&&a.push(new Pb(u.sketch.name,u.params,p.score))}return a.sort((u,p)=>p.score-u.score),{candidates:a,population:s}}_initPopulation(e){let t=[];for(let r=0;r<this.populationSize;r++){let s=e[this._rng(e.length)],o=s.sampleParams(i=>this._rng(i));t.push({sketch:s,params:o})}return t}_crossover(e,t){if(e.sketch!==t.sketch)return{...e.params};let r={};for(let s of e.sketch.variables)r[s.name]=this._rngFloat()<.5?e.params[s.name]:t.params[s.name];return r}_mutate(e,t){let r={rngFloat:l(()=>this._rngFloat(),"rngFloat"),rng:l(i=>this._rng(i),"rng"),mutationRate:this.mutationRate},s=[U4,...j4],o=t;for(let i of s)o=i(o,e,r);return{sketch:e,params:o}}},j4=[];l(U4,"defaultResampleMutator");l(a2,"createSearchStrategy")});function Fb(n){let e={"threadIdx.x":1,"threadIdx.y":1,"threadIdx.z":1},t=new Set,r=l(s=>{if(!(!s||typeof s!="object"||t.has(s))){if(t.add(s),s.type==="ForNode"&&e[s.threadTag]!==void 0){let o=s.extent&&s.extent.type==="IntImmNode"?s.extent.value:1;o>e[s.threadTag]&&(e[s.threadTag]=o)}for(let o in s){let i=s[o];Array.isArray(i)?i.forEach(r):i&&typeof i=="object"&&r(i)}}},"visit");return r(n.body),q4.reduce((s,o)=>s*e[o],1)}var q4,$b,l2=M(()=>{"use strict";du();Px();hv();Ed();u2();yv();q4=["threadIdx.x","threadIdx.y","threadIdx.z"];l(Fb,"gpuThreadBlockSize");$b=class{static{l(this,"BlockTuningSession")}constructor(e){this.target=e.target,this.primFunc=e.primFunc,this.blockName=e.blockName,this.sketches=e.sketches,this.benchmarkRunner=e.benchmarkRunner||null,this.config=e.config,this.deadline=e.deadline||null,this._warn=e.warn||(()=>{}),this._warnedEvalSketches=new Set;let t=this.sketches.some(o=>o.name==="fused");t?(this.learnedModel=new Oc,this.costModel=new Bc(e.costModel.analytical,this.learnedModel)):(this.costModel=e.costModel,this.learnedModel=e.learnedModel);let r=t?null:zL(e.primFunc,e.blockName,e.blockMap);this.evalFunc=r||e.primFunc,this.evalBlockName=e.blockName,this.sketchByName=new Map;for(let o of this.sketches)this.sketchByName.set(o.name,o);let s=this.sketches.length===1&&typeof this.sketches[0].enumerate=="function";this.enumSketch=s?this.sketches[0]:null,this.enumParams=s?this.sketches[0].enumerate():null,this.strategy=a2({...this.config,deadline:this.deadline}),this.population=null,this._best=null,this.plateaued=!1}runRound(){let e=this._best?this._best.measuredScore:-1/0,t=this._produceCandidates();if(t.length===0)return this.plateaued=!0,0;if(this.benchmarkRunner)this._measureAndLearn(t);else{let s=t[0];this._consider({sketchName:s.sketchName,params:s.params,score:s.score,measuredScore:s.score}),this.plateaued=!0}let r=this._best?this._best.measuredScore:-1/0;return Math.max(0,r-e)}best(){return this._best?{sketchName:this._best.sketchName,params:this._best.params,score:this._best.score,medianMs:this._best.medianMs??null,minMs:this._best.minMs??null}:null}bestTrace(){if(!this._best)return null;let e=this.sketchByName.get(this._best.sketchName);if(!e)return null;try{let t=new Cn(Ti(this.primFunc));return e.instantiate(this._best.params)(t,this.blockName,this.target),t.trace.serialize()}catch(t){return this._warn("best-trace",this.blockName,t),null}}_produceCandidates(){if(this.enumSketch){let s=[];for(let o of this.enumParams){let i=this._evaluate(this.enumSketch,o);i&&s.push({sketchName:this.enumSketch.name,params:o,score:i.score})}return s.sort((o,i)=>i.score-o.score),s}let e=l((s,o)=>this._evaluate(s,o),"evaluator"),{candidates:t,population:r}=this.strategy.search(this.sketches,e,this.population);return this.population=r,t}_evaluate(e,t){try{let r=Ti(this.evalFunc),s=new Cn(r);if(e.instantiate(t)(s,this.evalBlockName,this.target),Jl.validate(r).length>0)return null;let i=this.target.maxThreadsPerBlock;return this.target.isGPU&&this.target.isGPU()&&i&&Fb(r)>i?null:{score:this.costModel.score(r)}}catch(r){return this._warnedEvalSketches.has(e.name)||(this._warnedEvalSketches.add(e.name),this._warn("evaluate-candidate",this.blockName,r)),null}}_measureAndLearn(e){let t=e.slice(0,this.config.topKForBenchmark);for(let r of t){if(this.deadline&&this.deadline.expired)break;let s=this._measure(r);if(!s)continue;let o=-s.result.medianMs;this.learnedModel.addSample(s.features,o),this._consider({sketchName:r.sketchName,params:r.params,score:o,measuredScore:o,medianMs:s.result.medianMs,minMs:s.result.minMs})}this.learnedModel.train()}_measure(e){let t=this.sketchByName.get(e.sketchName);if(!t)return null;let r,s;try{r=Ti(this.primFunc),t.instantiate(e.params)(new Cn(r),this.blockName,this.target),s=Ti(this.evalFunc),t.instantiate(e.params)(new Cn(s),this.evalBlockName,this.target)}catch(i){return this._warn("measure-candidate",this.blockName,i),null}let o=this.benchmarkRunner.run(r);return o?{result:o,features:Mu.extractStatements(s)}:null}_consider(e){(!this._best||e.measuredScore>this._best.measuredScore)&&(this._best=e)}}});var zv,zb,c2=M(()=>{"use strict";zv=class{static{l(this,"GradientSchedulerPolicy")}pick(e){let t=e.filter(i=>!i.plateaued);if(t.length===0)return null;let r=t.filter(i=>i.rounds===0);if(r.length>0)return r[0];let s=null,o=-1/0;for(let i of t){let a=i.gainEwma!==void 0?i.gainEwma:i.lastGain,u=i.weight*a;u>o&&(o=u,s=i)}return s}},zb=class{static{l(this,"TaskScheduler")}constructor(e=null){this.policy=e||new zv}run(e,t,r={}){let s=r.maxRoundsPerTask??8,o=r.plateauPatience??2,i=r.gainEwmaAlpha??.5;for(let a of e)a.rounds=0,a.lastGain=0,a.gainEwma=0,a.stale=0,a.plateaued=!1;for(;!(t&&t.expired);){let a=this.policy.pick(e);if(!a)break;let u=a.session.runRound();a.rounds++,a.lastGain=u,a.gainEwma=i*u+(1-i)*a.gainEwma,u<=0?a.stale++:a.stale=0,(a.session.plateaued||a.stale>=o||a.rounds>=s)&&(a.plateaued=!0)}}}});function W4(n){if(n.isCPU())return null;let e=F1(n.kind);if(!e)throw new Error("hardwareMeasure requested for target '"+n.kind+"' but no measurer is registered for it; the corresponding runtime must be loaded (Node: import '#io/cuda_runtime') before compiling");return e}var Vb,Gb,p2=M(()=>{"use strict";du();yv();JL();QL();e2();t2();Ev();Rc();Bv();Do();Dv();l2();Ed();Fe();c2();Ym();l(W4,"resolveMeasurer");Vb=class{static{l(this,"AutotuneConfig")}constructor(e={}){this.strategy=e.strategy||"evolutionary",this.numTrials=e.numTrials||64,this.populationSize=e.populationSize||32,this.numGenerations=e.numGenerations||10,this.mutationRate=e.mutationRate,this.eliteRatio=e.eliteRatio,this.seed=e.seed||42,this.timeBudgetMs=e.timeBudgetMs||3e4,this.clock=e.clock||null,this.tuningDB=e.tuningDB||null,this.useTuningDB=e.useTuningDB!==!1,this.measurer=e.measurer||null,this.hardwareMeasure=e.hardwareMeasure??!1,this.enableBenchmark=e.enableBenchmark??(this.hardwareMeasure||!!e.measurer),this.benchmarkWarmup=e.benchmarkWarmup??3,this.benchmarkRepeat=e.benchmarkRepeat??10,this.benchmarkMaxCv=e.benchmarkMaxCv??0,this.topKForBenchmark=e.topKForBenchmark??5,this.maxRoundsPerTask=e.maxRoundsPerTask??8,this.plateauPatience=e.plateauPatience??2,this.schedulerPolicy=e.schedulerPolicy||null,this.onWarning=e.onWarning||null}},Gb=class{static{l(this,"Autotuner")}constructor(e,t={},r=null){this.target=e,this.config=t instanceof Vb?t:new Vb(t),this.trace=r,this._funcName=null,this.config.hardwareMeasure&&(this.config.measurer=W4(e)),this.analyticalModel=new _b(e),this.learnedModel=new Oc,this.costModel=new Bc(this.analyticalModel,this.learnedModel),this.db=this.config.tuningDB instanceof zd?this.config.tuningDB:new zd;let s=l((o,i,a)=>this._warn(o,i,a),"warn");this.benchmarkRunner=this.config.enableBenchmark?new Bb(e,{warmup:this.config.benchmarkWarmup,repeat:this.config.benchmarkRepeat,maxCv:this.config.benchmarkMaxCv,measurer:this.config.measurer,warn:s}):null,this.scheduler=new zb(this.config.schedulerPolicy)}_warn(e,t,r){let s=r&&r.message?r.message:String(r);if(this.config.onWarning)try{this.config.onWarning({stage:e,func:this._funcName,block:t||null,message:s,error:r})}catch(o){}this.trace&&this.trace.warn("autotune",this._funcName,`${e}${t?" ["+t+"]":""}: ${s}`)}tune(e,t=null){this._funcName=e.name;let r=t?[t]:Fo(e.body),s=Nb(e.body),o=$d(e),i=new Rb(this.config.timeBudgetMs,this.config.clock),a=new Map,u=new Map;for(let f of r){let d=Ab(e,f,this.target,s);u.set(f,d);let h=a.get(d);if(h){h.weight++;continue}if(this.config.useTuningDB&&this.db.has(d)){a.set(d,{key:d,kind:"cache",cached:this.db.lookup(d),weight:1});continue}let m=Iv(e,f,this.target,s,{richGpu:this.config.richGpu??!!this.config.measurer,dag:o});if(m.length===0){a.set(d,{key:d,kind:"empty",weight:1});continue}let g=new $b({target:this.target,primFunc:e,blockName:f,blockMap:s,sketches:m,costModel:this.costModel,learnedModel:this.learnedModel,benchmarkRunner:this.benchmarkRunner,config:this.config,deadline:i,warn:l((y,_,b)=>this._warn(y,_,b),"warn")});a.set(d,{key:d,kind:"session",session:g,weight:1})}let p=[...a.values()].filter(f=>f.kind==="session");p.length>0&&this.scheduler.run(p,i,this.config);let c=new Map;for(let f of r){let d=a.get(u.get(f));if(d.kind==="cache"){c.set(f,{sketchName:d.cached.sketchName,params:d.cached.params,score:d.cached.score,fromCache:!0});continue}if(d.kind==="empty")continue;let h=d.session.best();if(h){if(this.config.useTuningDB&&!d.stored){let m=new Fd(d.key,h.sketchName,h.params,h.score,d.session.bestTrace(),this.db.version);m.medianMs=h.medianMs||null,m.minMs=h.minMs||null,this.db.store(d.key,m),d.stored=!0}c.set(f,{sketchName:h.sketchName,params:h.params,score:h.score,fromCache:!1,medianMs:h.medianMs,minMs:h.minMs})}}return c}tuneAndApply(e,t=null){let r=this.tune(e,t);if(r.size>0){let s=this._applyBestSchedule(e,r);if(s)return{func:s.func,results:r,applied:!0}}return{func:e,results:r,applied:!1}}_applyBestSchedule(e,t){let r=this._buildDefaultSchedule(e),s=!!r&&this._scheduleIsValid(r),o=s&&this._isStrongBackendSchedule(r),i=this._buildTunedSchedule(e,t),a=!!i&&this._scheduleIsValid(i),u=a&&this._isStrongBackendSchedule(i),p=!o||u&&this.config.measurer!=null;return a&&p?(this._adoptSchedule(e,i),{func:e}):(i&&!a?this._warn("tuned-schedule-invalid",null,new Error("tuned schedule exceeds target thread-block limit; falling back to default")):a&&o&&this._warn("baseline-preferred",null,new Error("cost-model-only tuning cannot displace the deterministic GPU schedule without hardware measurement; keeping the deterministic kernel")),s?(this._adoptSchedule(e,r),{func:e}):a?(this._adoptSchedule(e,i),{func:e}):(this._warn("no-valid-schedule",null,new Error("neither tuned nor default schedule is valid; leaving function unscheduled")),null))}_isStrongBackendSchedule(e){return!!e&&e.gpuRegisterBlocked===!0}_buildTunedSchedule(e,t){try{let r=Ti(e),s=new Cn(r),o=Nb(r.body),i=$d(r),a=new Set,u=[];for(let c of t)if(c[1].sketchName==="fused"){let f=vb(r,i,c[0],Tt);f&&a.add(f),u.unshift(c)}else u.push(c);let p=new Set;for(let[c,f]of u)if(!(a.has(c)||p.has(f))&&(p.add(f),!(!f.sketchName||!f.params)))try{let h=Iv(r,c,this.target,o,{richGpu:this.config.richGpu??!!this.config.measurer,dag:i}).find(m=>m.name===f.sketchName);h&&this._fitsThreadBlock(r,c,h,f.params)&&h.instantiate(f.params)(s,c,this.target)}catch(d){this._warn("apply-tuned-block",c,d);continue}return this._scheduleResidualBlocks(s,a),r}catch(r){return this._warn("build-tuned-schedule",null,r),null}}_scheduleResidualBlocks(e,t){let r=null;for(let s of Fo(e.func.body))if(!(t.has(s)||this._blockIsParallelized(e,s))){r||(r=new to(this.target));try{r.applyToBlock(e,s)}catch(o){this._warn("residual-block",s,o)}}}_blockIsParallelized(e,t){let r;try{r=e.getLoops(t)}catch(s){return this._warn("block-loops",t,s),!0}for(let s of r)if(s.kind===z.THREAD_BINDING||s.kind===z.PARALLEL||s.kind===z.VECTORIZED)return!0;return!1}_scheduleIsValid(e){return!this.target.isGPU||!this.target.isGPU()||!this.target.maxThreadsPerBlock?!0:Fb(e)<=this.target.maxThreadsPerBlock}_buildDefaultSchedule(e){try{let t=Ti(e),r=new Cn(t);return Db(r,this.target,this.config)||new to(this.target).applyToAllBlocks(r),t}catch(t){return this._warn("build-default-schedule",null,t),null}}_adoptSchedule(e,t){Object.assign(e,t),e._setChild("body",e.body)}_fitsThreadBlock(e,t,r,s){if(!this.target.isGPU||!this.target.isGPU()||!this.target.maxThreadsPerBlock)return!0;try{let o=Ti(e);return r.instantiate(s)(new Cn(o),t,this.target),Fb(o)<=this.target.maxThreadsPerBlock}catch(o){return this._warn("fits-thread-block",t,o),!1}}}});var jb,f2=M(()=>{"use strict";ma();du();Do();p2();Dv();jb=class extends ar{static{l(this,"SchedulePass")}constructor(e){super("SchedulePass","scheduling"),this.config=e,this.target=e.target,this.snapshotPoint="afterScheduling",this._autotuner=null,this._policy=null}begin(e){let t=this.config.scheduling;t.autotune?this._autotuner=new Gb(this.target,t,e.trace):t.enabled&&(this._policy=new to(this.target,null,e.trace))}run(e,t){let r=this.config.scheduling,s=t.trace;if(r.autotune){if(e.cublasInfo||e._tensorIntrin)return;let o=performance.now(),i=this._autotuner.tuneAndApply(e),a=performance.now()-o,u=0,p=0;if(i&&i.results){p=i.results.size;for(let[c,f]of i.results)f.fromCache&&u++,s.explainsEnabled&&s.explain("schedule",c,f.sketchName,`autotuned: best of search${f.fromCache?" (cached)":""}, score ${f.score!=null?f.score.toFixed(3):"n/a"}`,{target:this.target.name,params:f.params})}s.autotuneStats(e.name,{durationMs:a,blockCount:p,applied:!!(i&&i.applied),cacheHits:u})}else if(r.enabled||r.gpuTiling){if(e.cublasInfo||e._tensorIntrin)return;let o=performance.now(),i=new Cn(e);!Db(i,this.target,r)&&r.enabled&&this._policy.applyToAllBlocks(i),s.functionEvent("scheduling",e.name,{durationMs:performance.now()-o})}}}});function d2(n){let e={analyzer:new Xs,simp:null};e.simp=new vy(e.analyzer);let t=hs(n.body,e);return n.body=t,n._setChild("body",t),n}function K4(n,e,t){let r=n.analyzer.getVarBound(e);return t&&t.type==="IntImmNode"&&t.value>0?n.analyzer.bind(e,0,t.value-1):n.analyzer.setVarBound(e,null),r}function hs(n,e){if(!n||typeof n!="object")return n;switch(n.type){case"ForNode":{let t=K4(e,n.loopVar.name,n.extent),r=hs(n.body,e);e.analyzer.setVarBound(n.loopVar.name,t);let s=new he(n.loopVar,n.min,n.extent,n.kind,r,n.threadTag);return n.annotations&&(s.annotations=n.annotations),s}case"BlockNode":{let t=[];for(let i of n.iterVars)i.iterVar&&(t.push([i.iterVar.name,e.analyzer.getVarBound(i.iterVar.name)]),e.analyzer.setVarBound(i.iterVar.name,i.binding?cu(e.analyzer,i.binding):null));let r=n.iterVars.map(H4(e)),s=hs(n.body,e),o=n.initBody?hs(n.initBody,e):null;for(let[i,a]of t)e.analyzer.setVarBound(i,a);return new fe(n.name,r,n.reads,n.writes,s,o)}case"SeqNode":return new te(n.stmts.map(t=>hs(t,e)));case"IfThenElseNode":{let t=ur(n.condition,e);return _i(e.analyzer,t)?hs(n.thenBody,e):lu(e.analyzer,t)?n.elseBody?hs(n.elseBody,e):new te([]):new Se(t,hs(n.thenBody,e),n.elseBody?hs(n.elseBody,e):null)}case"BufferStoreNode":return new J(n.buffer,n.indices.map(t=>ur(t,e)),ur(n.value,e));case"LetStmtNode":return new et(n.variable,ur(n.value,e),hs(n.body,e));case"AllocateNode":return new Ue(n.buffer,n.scope,hs(n.body,e));case"WhileNode":return new $i(n.condVar,hs(n.condBody,e),hs(n.loopBody,e));case"EvaluateNode":return new il(ur(n.value,e));default:return n}}function H4(n){return e=>{if(!e.iterVar||!e.binding)return e;let t=ur(e.binding,n);return new Pa(e.iterVar,t,e.kind)}}function ur(n,e){if(!n||typeof n!="object"||!n.type)return n;switch(n.type){case"IntImmNode":case"FloatImmNode":case"VariableNode":return n;case"BufferLoadNode":return new q(n.buffer,n.indices.map(t=>ur(t,e)));case"MathOpNode":{let t=ur(n.a,e),r=n.b?ur(n.b,e):null;return e.simp.simplify(new P(n.op,t,r))}case"CompareNode":{let t=ur(n.a,e),r=ur(n.b,e);return e.simp.simplify(new Ke(n.direction,t,r))}case"CastNode":return new ut(ur(n.expr,e),n.fromDtype,n.toDtype);case"CallExternNode":return new Ne(n.externName,n.args.map(t=>ur(t,e)),n.dtype);case"IfThenElseNode":{let t=ur(n.condition,e),r=ur(n.thenBody,e),s=n.elseBody?ur(n.elseBody,e):null;return _i(e.analyzer,t)?r:s!==null&&lu(e.analyzer,t)?s:new Se(t,r,s)}default:return n}}var h2=M(()=>{"use strict";Yl();Tf();Fe();l(d2,"simplifyPrimFunc");l(K4,"bindLoopVar");l(hs,"simplifyStmt");l(H4,"simplifyIterVar");l(ur,"simplifyExpr")});var Ub,m2=M(()=>{"use strict";ma();h2();Ub=class extends ar{static{l(this,"SimplifyPass")}constructor(){super("SimplifyPass","simplify")}run(e,t){let r=performance.now();return d2(e),t.trace.functionEvent("simplify",e.name,{durationMs:performance.now()-r}),e}}});function g2(n){return!!n&&typeof n=="object"&&typeof n.name=="string"&&n.dtype!==void 0&&n.shape!==void 0&&n.type===void 0}var Vv,Gv,Y4,qb,y2=M(()=>{"use strict";Vv=class{static{l(this,"BufferInterval")}constructor(e,t,r,s){this.buffer=e,this.firstUse=t,this.lastUse=r,this.scope=s}get size(){return this.buffer.sizeInBytes()}overlaps(e){return this.firstUse<=e.lastUse&&e.firstUse<=this.lastUse}},Gv=class{static{l(this,"BufferLivenessResult")}constructor(e,t,r){this.intervals=e,this.stmtOrder=t,this.paramBuffers=r}isParam(e){return this.paramBuffers.has(e)}getTemporaries(){let e=[];for(let[t,r]of this.intervals)this.paramBuffers.has(t)||e.push(r);return e}interfere(e,t){let r=this.intervals.get(e),s=this.intervals.get(t);return!r||!s?!1:r.overlaps(s)}},Y4=new Set(["_parent","_parentKey","_parentIdx"]);l(g2,"isBuffer");qb=class{static{l(this,"BufferLiveness")}static analyze(e){let t=new Map,r=[],s=0,o=new Set;for(let[,f]of e.bufferMap)o.add(f);let i=[],a=l(f=>{if(!f)return;let d=t.get(f);d?(s<d.firstUse&&(d.firstUse=s),s>d.lastUse&&(d.lastUse=s)):(d=new Vv(f,s,s,f.scope),t.set(f,d)),i.push(f)},"touch"),u=l((f,d)=>{if(!(!f||typeof f!="object"||d.has(f))){if(d.add(f),g2(f)){a(f);return}g2(f.buffer)&&a(f.buffer);for(let h of Object.keys(f)){if(Y4.has(h)||h==="buffer")continue;let m=f[h];if(!(!m||typeof m!="object"))if(Array.isArray(m))for(let g of m)u(g,d);else u(m,d)}}},"touchAll"),p=l((f,d)=>{for(let h=f;h<i.length;h++){let m=t.get(i[h]);m&&d>m.lastUse&&(m.lastUse=d)}},"extendRegion"),c=l(f=>{if(f)switch(f.type){case"SeqNode":for(let d of f.stmts)c(d);break;case"ForNode":{let d=s,h=i.length;u(f.min,new Set),u(f.extent,new Set),c(f.body);let m=s>d?s-1:d;p(h,m);break}case"WhileNode":{let d=s,h=i.length;u(f.condVar,new Set),c(f.condBody),c(f.loopBody);let m=s>d?s-1:d;p(h,m);break}case"BlockNode":r.push({idx:s,node:f});for(let d of f.reads)a(d.buffer);for(let d of f.writes)a(d.buffer);u(f.body,new Set),f.initBody&&u(f.initBody,new Set),s++;break;case"AllocateNode":a(f.buffer),c(f.body);break;case"IfThenElseNode":{let d=s,h=i.length;u(f.condition,new Set),c(f.thenBody),f.elseBody&&c(f.elseBody);let m=s>d?s-1:d;p(h,m);break}case"LetStmtNode":u(f.value,new Set),c(f.body);break;case"EvaluateNode":u(f.value,new Set);break;default:u(f,new Set);break}},"walk");return c(e.body),new Gv(t,r,o)}}});function J4(n,e){if(n.shape.length!==e.shape.length)return!1;for(let t=0;t<n.shape.length;t++)if(n.shape[t]!==e.shape[t])return!1;return!0}function Lc(n,e){if(n===e)return!0;if(!n||!e||typeof n!="object"||typeof e!="object"||n.type!==e.type)return!1;switch(n.type){case"VariableNode":return n.name===e.name;case"IntImmNode":return n.value===e.value;case"FloatImmNode":return n.value===e.value;case"MathOpNode":return n.op===e.op&&Lc(n.a,e.a)&&Lc(n.b,e.b);case"CompareNode":return n.direction===e.direction&&Lc(n.a,e.a)&&Lc(n.b,e.b);case"CastNode":return n.toDtype===e.toDtype&&Lc(n.expr,e.expr);case"BufferLoadNode":return n.buffer===e.buffer&&Uv(n.indices,e.indices);default:return!1}}function Uv(n,e){if(n.length!==e.length)return!1;for(let t=0;t<n.length;t++)if(!Lc(n[t],e[t]))return!1;return!0}function _2(n,e){n&&qs(n,e)}function Z4(n,e,t){let r=[],s=[];for(let p of[n.body,n.initBody])p&&_2(p,c=>{c.type==="BufferStoreNode"&&c.buffer===t?r.push(c):c.type==="BufferLoadNode"&&c.buffer===e&&s.push(c)});if(r.length!==1||s.length===0)return!1;let o=r[0],i=o.indices;for(let p of s)if(!Uv(p.indices,i))return!1;let a=new Set;for(let p of[o.value,...o.indices])p&&_2(p,c=>{c.type==="BufferLoadNode"&&c.buffer===e&&a.add(c)});let u=!0;for(let p of s)if(!a.has(p)){u=!1;break}if(!u){let p=o.value;if(!(p&&p.type==="BufferLoadNode"&&p.buffer===e&&Uv(p.indices,i)))return!1}return!0}function Q4(n,e){if(n)for(let t of yl(n,r=>r.type==="BlockNode",{kinds:"stmt"}))e.push(t)}var jv,Wb,b2=M(()=>{"use strict";ts();jv=class{static{l(this,"InplaceCandidate")}constructor(e,t,r){this.srcBuffer=e,this.dstBuffer=t,this.reason=r}},Wb=class{static{l(this,"InplaceAnalysis")}static analyze(e,t,r=new Set){let s=[],o=[];Q4(e.body,o);let i=new Map;for(let p of t.stmtOrder)i.set(p.node,p.idx);let a=new Map;for(let p of t.stmtOrder)for(let c of p.node.reads){let f=a.get(c.buffer);(f===void 0||p.idx>f)&&a.set(c.buffer,p.idx)}let u=new Set;for(let p of o){let c=i.get(p);if(c!==void 0)for(let f of p.writes){let d=f.buffer;if(!(t.isParam(d)&&!r.has(d)))for(let h of p.reads){let m=h.buffer;if(m===d||u.has(m)||t.isParam(m)&&!r.has(m)||!J4(m,d)||m.dtype!==d.dtype||m.scope!==d.scope)continue;let g=t.intervals.get(m),y=t.intervals.get(d);if(!(!g||!y)&&g.lastUse<=y.firstUse){let _=a.get(m);if(_===void 0||_<=c){if(!Z4(p,m,d))continue;s.push(new jv(m,d,`${m.name} last used at ${g.lastUse}, ${d.name} first used at ${y.firstUse}`)),u.add(m);break}}}}}return s}};l(J4,"shapesMatch");l(Lc,"exprEqual");l(Uv,"indexListEqual");l(_2,"walkNodes");l(Z4,"isInplaceComputeSafe");l(Q4,"collectBlocks")});function w2(n,e){let t=0,r=n.length;for(;t<r;){let s=t+r>>1;n[s].offset<=e.offset?t=s+1:r=s}n.splice(t,0,e)}function T2(n){return{best:null,consider(e,t,r){return t<r?null:n==="best-fit"?((this.best===null||t<this.best.gap)&&(this.best={offset:e,gap:t}),null):e},result(e){return this.best!==null?this.best.offset:e}}}var qv,Wv,Kb,x2=M(()=>{"use strict";l(w2,"insertByOffset");l(T2,"gapSelector");qv=class{static{l(this,"MemoryBlock")}constructor(e,t,r){this.offset=e,this.size=t,this.buffer=r}get end(){return this.offset+this.size}overlaps(e){return this.offset<e.end&&e.offset<this.end}},Wv=class{static{l(this,"MemoryPool")}constructor(e,t=64,r="best-fit"){this.scope=e,this.alignment=t,this.strategy=r,this.blocks=[],this.peakUsage=0}allocate(e,t){let r=this._align(e),s=this._findFreeOffset(r);return this.placeAt(s,r,t)}placeAt(e,t,r){let s=this._align(t),o=new qv(e,s,r);w2(this.blocks,o);let i=e+s;return i>this.peakUsage&&(this.peakUsage=i),o}_align(e){return Math.ceil(e/this.alignment)*this.alignment}_findFreeOffset(e){let t=this.blocks,r=0,s=T2(this.strategy);for(let o of t){let i=this._align(r),a=o.offset-i,u=s.consider(i,a,e);if(u!==null)return u;o.end>r&&(r=o.end)}return s.result(this._align(r))}fragmentation(){if(this.peakUsage===0)return 0;let e=this.blocks.reduce((t,r)=>t+r.size,0);return Math.max(0,1-e/this.peakUsage)}release(e){let t=this.blocks.indexOf(e);t>=0&&this.blocks.splice(t,1)}},Kb=class{static{l(this,"BufferAssignment")}constructor(){this.assignments=new Map,this.pools=new Map,this.inplaceMap=new Map,this.effLastUse=new Map}assign(e,t=[],r=64,s="best-fit"){for(let c of t)this.inplaceMap.set(c.dstBuffer,c.srcBuffer);let o=new Map;for(let c of e)o.set(c.buffer,c);let i=new Map;for(let c of e)i.set(c.buffer,c.lastUse);let a=new Map;for(let[c,f]of this.inplaceMap)!o.has(c)||!o.has(f)||(a.has(f)||a.set(f,[]),a.get(f).push(c));if(a.size>0){let c=new Map;for(let f of e){let d=f.buffer;if(c.get(d)===1)continue;let h=[d];for(;h.length;){let m=h[h.length-1],g=a.get(m);if(c.get(m)===void 0){if(c.set(m,0),g)for(let y of g)c.get(y)===void 0&&h.push(y);continue}if(h.pop(),c.get(m)!==1){if(g){let y=i.get(m);for(let _ of g){let b=i.get(_);b>y&&(y=b)}i.set(m,y)}c.set(m,1)}}}}this.effLastUse=i;let u=[...e].sort((c,f)=>{let d=c.size,h=f.size,m=d>0,g=h>0;if(m&&g){let y=h-d;if(y!==0)return y}else if(m!==g)return m?-1:1;return c.firstUse-f.firstUse}),p=new Map;for(let c of u){let f=c.buffer;if(this.inplaceMap.has(f)){let w=this.inplaceMap.get(f),x=this.assignments.get(w);if(x){this.assignments.set(f,{offset:x.offset,size:c.size,scope:c.scope,pool:x.pool,inplaceOf:w});continue}}let d=c.size;if(d===0)continue;if(d<0){this.assignments.set(f,{offset:0,size:0,scope:c.scope,pool:c.scope,inplaceOf:null,isDynamic:!0});continue}let h=c.scope,m=this.pools.get(h);m||(m=new Wv(h,r),this.pools.set(h,m)),p.has(h)||p.set(h,[]);let g=p.get(h),y=i.get(f)??c.lastUse,_=this._interferenceOffset(g,c.firstUse,y,m._align(d),r,s),b=m.placeAt(_,d,f);w2(g,{firstUse:c.firstUse,lastUseEff:y,offset:b.offset,size:b.size}),this.assignments.set(f,{offset:b.offset,size:b.size,scope:h,pool:h,inplaceOf:null})}return this}_interferenceOffset(e,t,r,s,o,i){let a=[];for(let c of e)c.firstUse<=r&&t<=c.lastUseEff&&a.push([c.offset,c.offset+c.size]);let u=0,p=T2(i);for(let[c,f]of a){let d=Math.ceil(u/o)*o,h=c-d,m=p.consider(d,h,s);if(m!==null)return m;f>u&&(u=f)}return p.result(Math.ceil(u/o)*o)}getOffset(e){let t=this.assignments.get(e);return t?t.offset:-1}getAssignment(e){return this.assignments.get(e)||null}peakMemory(e=null){if(e){let r=this.pools.get(e);return r?r.peakUsage:0}let t=0;for(let[,r]of this.pools)t+=r.peakUsage;return t}}});function Yb(n,e,t){if(!n||typeof n!="object"||t.has(n))return!1;if(t.add(n),n.type==="BufferLoadNode"&&(!e||n.buffer===e))return!0;for(let r of Object.keys(n)){if(r==="_parent"||r==="_parentKey"||r==="_parentIdx")continue;let s=n[r];if(!(!s||typeof s!="object")){if(Array.isArray(s)){for(let o of s)if(Yb(o,e,t))return!0}else if(Yb(s,e,t))return!0}}return!1}function S2(n){let e=new Set,t=new Map,r=new Set,s=[n.body];for(;s.length>0;){let o=s.pop();if(!(!o||typeof o!="object"||r.has(o))){if(r.add(o),o.type==="BufferStoreNode"&&o.buffer){let i=o.buffer;for(let u of o.indices)if(Yb(u,null,new Set)){e.add(i);break}Yb(o.value,i,new Set)&&e.add(i);let a=t.get(i);a||(a=[],t.set(i,a)),a.push(o.value)}for(let i of Object.keys(o)){if(i==="_parent"||i==="_parentKey"||i==="_parentIdx")continue;let a=o[i];if(Array.isArray(a))for(let u of a)u&&typeof u=="object"&&s.push(u);else a&&typeof a=="object"&&s.push(a)}}}for(let[o,i]of t){let a=l(u=>u&&(u.type==="IntImmNode"||u.type==="FloatImmNode")&&u.value===0,"isConstZero");i.every(a)&&e.add(o),i.length===1&&i[0]&&(i[0].type==="IntImmNode"||i[0].type==="FloatImmNode")&&e.add(o)}return e}function X4(n,e){let t=new Set,r=[n];for(;r.length>0;){let s=r.pop();if(!(!s||typeof s!="object"||t.has(s))){t.add(s);for(let o of Object.keys(s)){if(o==="_parent"||o==="_parentKey"||o==="_parentIdx")continue;let i=s[o];if(i&&typeof i=="object"&&e.has(i)){s[o]=e.get(i);continue}if(Array.isArray(i)){for(let a=0;a<i.length;a++){let u=i[a];u&&typeof u=="object"&&e.has(u)?i[a]=e.get(u):u&&typeof u=="object"&&r.push(u)}continue}i&&typeof i=="object"&&r.push(i)}}}}var Kv,Hb,v2=M(()=>{"use strict";y2();b2();x2();Fe();H0();Kv=class{static{l(this,"MemoryPlan")}constructor(e,t,r){this.assignment=e,this.liveness=t,this.inplaceCandidates=r,this.aliasMap=new Map}peakMemory(e=null){return this.assignment.peakMemory(e)}getReport(){let e=new Map;for(let[s,o]of this.assignment.pools)e.set(s,{peakUsage:o.peakUsage,numBuffers:0,numReused:0});for(let[s,o]of this.assignment.assignments){let i=e.get(o.scope);i&&(i.numBuffers++,o.inplaceOf&&i.numReused++)}let t=this.liveness.getTemporaries().length,r=this.inplaceCandidates.length;return{peakMemory:this.assignment.peakMemory(),scopeBreakdown:e,totalTemporaries:t,totalInplace:r,materializedReuse:this.aliasMap.size,assignments:this.assignment.assignments}}},Hb=class{static{l(this,"MemoryPlanner")}constructor(e={}){this.alignment=e.alignment||64,this.enableInplace=e.enableInplace!==!1,this.allocStrategy=e.allocStrategy||"best-fit",this.poolAllocation=e.poolAllocation||!1}plan(e){let t=qb.analyze(e),r=t.getTemporaries(),s=[];this.enableInplace&&(s=Wb.analyze(e,t));let o=new Kb;return o.assign(r,s,this.alignment,this.allocStrategy),new Kv(o,t,s)}planAndRewrite(e){let t=this.plan(e);return{func:this._insertAllocations(e,t),plan:t}}_insertAllocations(e,t){let r=t.liveness.getTemporaries();if(r.length===0)return e;let s=new Map;this.poolAllocation?this._assignPoolOffsets(e,t,r):(s=this._buildReuseAliases(r,t,e),s.size>0&&X4(e.body,s)),t.aliasMap=s;let o=[...r].sort((u,p)=>p.firstUse-u.firstUse),i=e.body,a=new Set;for(let u of o){let p=u.buffer;if(s.has(p))continue;let c=t.assignment.getAssignment(p);c&&(c.inplaceOf||a.has(p)||(a.add(p),i=new Ue(p,c.isDynamic?"dynamic":c.scope,i)))}return e.body=i,e._setChild("body",i),e}_assignPoolOffsets(e,t,r){let s=S2(e);for(let o of r){let i=o.buffer;if(s.has(i)||i.scope!=="global")continue;let a=t.assignment.getAssignment(i);!a||a.inplaceOf||a.isDynamic||a.size>0&&(i.poolByteOffset=a.offset)}}_buildReuseAliases(e,t,r){let s=S2(r),o=new Set(t.assignment.inplaceMap.values()),i=t.assignment.effLastUse,a=l(c=>i.get(c.buffer)??c.lastUse,"lastUseOf"),u=new Map;for(let c of e){let f=c.buffer,d=t.assignment.getAssignment(f);if(!d||d.inplaceOf||d.isDynamic||o.has(f)||f.numel()<=0||s.has(f))continue;let h=`${f.scope}|${f.dtype}|${f.shape.join(",")}|${f.strides.join(",")}`,m=u.get(h);m||(m=[],u.set(h,m)),m.push(c)}let p=new Map;for(let c of u.values()){if(c.length<2)continue;c.sort((d,h)=>d.firstUse-h.firstUse||a(d)-a(h));let f=new Tl((d,h)=>d.lastUse-h.lastUse);for(let d of c){let h=f.peek();h&&h.lastUse<d.firstUse?(f.pop(),h.lastUse=a(d),f.push(h),p.set(d.buffer,h.rep)):f.push({rep:d.buffer,lastUse:a(d)})}}return p}};l(Yb,"exprLoadsBuffer");l(S2,"collectFreshZeroDependent");l(X4,"rewriteBufferAliases")});var Jb,k2=M(()=>{"use strict";ma();v2();Jb=class extends ar{static{l(this,"MemoryPlanPass")}constructor(e){super("MemoryPlanPass","memoryPlanning"),this.config=e,this._planner=null}begin(e){let t=this.config.memory.alignment||this.config.target?.cacheLineSizeBytes||64;this._planner=new Hb({alignment:t,enableInplace:this.config.memory.inplaceReuse,allocStrategy:this.config.memory.allocStrategy,poolAllocation:this.config.memory.poolAllocation})}run(e,t){if(e.gpuRegisterBlocked)return;let r=performance.now(),{plan:s}=this._planner.planAndRewrite(e),o=s.getReport();t.trace.memoryStats(e.name,{durationMs:performance.now()-r,peakMemory:o.peakMemory,totalTemporaries:o.totalTemporaries,totalInplace:o.totalInplace})}}});function Cc(n){return n&&n.type==="IntImmNode"?n.value:null}function ej(n){if(!n||n.type!=="ForNode"||!N2.has(n.kind)||Cc(n.min)!==0)return null;let e=Cc(n.extent);if(e===null)return null;let t=n.body;if(!t||t.type!=="ForNode"||!N2.has(t.kind)||Cc(t.min)!==0)return null;let r=Cc(t.extent);if(r===null||r<=0)return null;let s=t.body;if(!s||s.type!=="IfThenElseNode"||s.elseBody)return null;let o=tj(s.condition,n.loopVar.name,t.loopVar.name,r);if(o===null||o<=0)return null;let i=Math.floor(o/r),a=o-i*r;if(a===0||i<1||e!==i+1)return null;let u=new Xs;if(u.bind(n.loopVar.name,0,i-1),u.bind(t.loopVar.name,0,r-1),!_i(u,s.condition))return null;let p=s.thenBody,c=new he(t.loopVar,new L(0),new L(r),t.kind,p,t.threadTag),f=new he(n.loopVar,new L(0),new L(i),n.kind,c,n.threadTag),d=Gi(Iu(p),g=>g.type==="VariableNode"&&g.name===n.loopVar.name?new L(i):g),h=new Oe(t.loopVar.name,t.loopVar.dtype),m=new he(h,new L(0),new L(a),t.kind,d,t.threadTag);return new te([f,m])}function tj(n,e,t,r){if(!n)return null;let s,o;if(n.type==="MathOpNode"&&n.op==="<")s=n.a,o=n.b;else if(n.type==="CompareNode"&&n.direction==="lt")s=n.a,o=n.b;else return null;let i=Cc(o);if(i===null||!s||s.type!=="MathOpNode"||s.op!=="+")return null;let a=s.a,u=s.b;return!u||u.type!=="VariableNode"||u.name!==t||!a||a.type!=="MathOpNode"||a.op!=="*"||!a.a||a.a.type!=="VariableNode"||a.a.name!==e||Cc(a.b)!==r?null:i}var N2,Zb,A2=M(()=>{"use strict";ma();Fe();ts();Ed();Yl();Tf();N2=new Set([z.SERIAL,z.PARALLEL,z.VECTORIZED,z.UNROLLED]),Zb=class extends ar{static{l(this,"LoopPartitionPass")}constructor(){super("LoopPartitionPass","loopPartition")}run(e,t){let r=Gi(e.body,s=>ej(s)||s);return r&&r!==e.body&&(e.body=r,e._setChild("body",r)),e}};l(Cc,"intImm");l(ej,"partitionLoop");l(tj,"matchFlatGuard")});function Qb(n){let e=n.body;if(!e||e.type!=="BlockNode")return null;let t=e.body;if(!t||t.type!=="BufferStoreNode")return null;let r=t,s=r.value;if(!s||s.type!=="MathOpNode"||!nj.has(s.op))return null;let o=null,i=null;if(s.a&&s.a.type==="BufferLoadNode"&&s.a.buffer===r.buffer?(o=s.a,i=s.b):s.b&&s.b.type==="BufferLoadNode"&&s.b.buffer===r.buffer&&(o=s.b,i=s.a),!o)return null;let a=Hv(r.indices),u=Hv(o.indices);if(a!==u||a.includes("?"))return null;let p=r.indices.map(d=>{if(d.type!=="VariableNode")return d;for(let h of e.iterVars)if(h.iterVar&&h.iterVar.name===d.name&&h.binding)return h.binding;return d}),c=n.loopVar.name,f=Hv(p);return f.includes("?")||f.includes("$"+c)?null:{store:r,loadSide:o,valueSide:i,outerIndices:p,block:e,op:r.value.op}}function Hv(n){return n.map(Yv).join(",")}function Yv(n){return n?n.type==="VariableNode"?"$"+n.name:n.type==="IntImmNode"?String(n.value):n.type==="MathOpNode"?"("+Yv(n.a)+n.op+(n.b?Yv(n.b):"")+")":"?":"?"}var nj,Jv=M(()=>{"use strict";nj=new Set(["+","*","max","min"]);l(Qb,"detectAccumulator");l(Hv,"indicesKey");l(Yv,"exprKey")});function Vo(n){if(!(!n||typeof n!="object"||!n.type))switch(n.type){case"ForNode":{let e=Qb(n);n.accumulator=e,e||Vo(n.body);return}case"BlockNode":Vo(n.body),n.initBody&&Vo(n.initBody);return;case"SeqNode":for(let e of n.stmts)Vo(e);return;case"LetStmtNode":case"AllocateNode":Vo(n.body);return;case"IfThenElseNode":Vo(n.thenBody),n.elseBody&&Vo(n.elseBody);return;case"WhileNode":Vo(n.condBody),Vo(n.loopBody);return;default:return}}var Xb,O2=M(()=>{"use strict";ma();Jv();Xb=class extends ar{static{l(this,"AccumulatorDetectionPass")}constructor(){super("AccumulatorDetectionPass","accumulatorDetect")}run(e,t){return Vo(e.body),e}};l(Vo,"annotateStmt")});function sj(n){let e=[],t=[n.body];for(;t.length>0;){let s=t.pop();if(s){if(s.type==="BlockNode"){e.push(s),t.push(s.body);continue}if(s.body&&t.push(s.body),s.stmts)for(let o of s.stmts)t.push(o);s.thenBody&&t.push(s.thenBody),s.elseBody&&t.push(s.elseBody)}}if(e.length===0)return null;let r=null;for(let s of e){if(!s.name.includes("matmul"))return null;s.reads.length>=2&&s.writes.length>=1&&(r=s)}return r}function oj(n){let e=sj(n);if(!e)return null;let t=e.reads[0].buffer,r=e.reads[1].buffer,s=e.writes[0].buffer;if(!B2.has(t.dtype)||!B2.has(r.dtype)||s.dtype!=="f32"||t.shape.length!==2||r.shape.length!==2||s.shape.length!==2)return null;let o=s.shape[0],i=s.shape[1],a=t.shape[1];if(![o,i,a].every(p=>typeof p=="number"&&p>0&&p%rj===0))return null;let u=new Set;for(let[,p]of n.bufferMap)u.add(p.name);return!u.has(t.name)||!u.has(r.name)||!u.has(s.name)?null:{M:o,N:i,K:a,a:t.name,b:r.name,c:s.name}}var rj,B2,ew,R2=M(()=>{"use strict";ma();du();rj=16,B2=new Set(["f16","bf16"]);l(sj,"findMatmulBlock");l(oj,"detectWmmaMatmul");ew=class extends ar{static{l(this,"AutoTensorizePass")}constructor(e){super("AutoTensorizePass","scheduling"),this.config=e,this.target=e.target}run(e,t){if(e.cublasInfo||e._tensorIntrin||!this.target||!this.target.isGPU())return;let r=oj(e);r&&(new Cn(e).tensorize("wmma_16x16x16_f16f16f32",r),t&&t.trace&&t.trace.explainsEnabled&&t.trace.explain("tensorize",e.name,"wmma_16x16x16_f16f16f32",`auto-tensorized ${r.M}x${r.N}x${r.K} f16 GEMM`,{target:this.target.name}))}}});function ij(n){let e=[],t=new Map,r=new Map,s=[{node:n,block:null}];for(;s.length>0;){let{node:o,block:i}=s.pop();if(o.type==="BufferLoadNode"&&o.buffer)t.set(o.buffer.name,(t.get(o.buffer.name)||0)+1),i&&i.reads.add(o.buffer.name);else if(o.type==="BufferStoreNode"&&o.buffer&&i){i.writes.add(o.buffer.name);let u=r.get(o.buffer.name);u||(u=new Set,r.set(o.buffer.name,u)),u.add(i.name)}let a=i;o.type==="BlockNode"&&(a={name:o.name,writes:new Set,reads:new Set,hasInit:o.initBody!=null},e.push(a));for(let u of ln(o))s.push({node:u,block:a})}return{blocks:e,loadCount:t,storeWriters:r}}var tw,M2=M(()=>{"use strict";ma();du();Do();ts();l(ij,"analyzeFunc");tw=class extends ar{static{l(this,"InlineReindexPass")}constructor(e){super("InlineReindexPass","scheduling"),this.config=e,this.target=e.target,this.snapshotPoint="afterInlineReindex"}run(e,t){if(!this.target.isGPU()||this.target.isWebGPU()||e.cublasInfo||e._tensorIntrin||tS(e))return;let r=this.config.scheduling;if(!(r.enabled||r.gpuTiling||r.autotune))return;let s=new Set;for(let[,c]of e.bufferMap)s.add(c.name);let{blocks:o,loadCount:i,storeWriters:a}=ij(e.body),u=new Cn(e),p=!1;for(let c of o){if(c.hasInit||c.writes.size===0)continue;let f=[...c.writes];if(f.every(d=>!s.has(d)&&a.get(d).size===1)&&f.some(d=>(i.get(d)||0)>0)&&[...c.reads].every(d=>(a.get(d)?a.get(d).size:0)<=1))try{u.computeInlineBlock(c.name),p=!0}catch{}}p&&Xx(e)}}});function Zv(n,e,t=aj){return t.filter(r=>r.phase===n).sort((r,s)=>r.priority-s.priority).map(r=>r.factory(e)).filter(Boolean)}var aj,I2=M(()=>{"use strict";aj=[];l(Zv,"tirPassesForPhase")});function E2(n){let e=[];for(let t of Zv("pre",n))e.push(t);e.push(new tw(n)),n.optimization.tensorize&&e.push(new ew(n)),e.push(new jb(n)),n.optimization.loopPartition&&e.push(new Zb),e.push(new Ub),e.push(new Jb(n)),n.optimization.detectAccumulators&&e.push(new Xb);for(let t of Zv("post",n))e.push(t);return e}var L2=M(()=>{"use strict";f2();m2();k2();A2();O2();R2();M2();I2();l(E2,"buildTirPipeline")});var Gd,C2=M(()=>{"use strict";Fe();Gd=class{static{l(this,"TensorVerifier")}constructor(){this.errors=[],this.boundVars=new Set}verify(e){if(this.errors=[],this.boundVars.clear(),!(e instanceof Pi))return this.errors.push("Expected PrimFunc at root"),this.errors;for(let t of e.params)this.boundVars.add(t.name);if(e.shapeParams)for(let t of e.shapeParams)this.boundVars.add(t.name);return this.visit(e.body),this.errors}visit(e){let t=[e];for(;t.length>0;){let r=t.pop();if(r)switch(r.type){case"SeqNode":for(let s=r.stmts.length-1;s>=0;s--)t.push(r.stmts[s]);break;case"ForNode":this.boundVars.has(r.loopVar.name)&&this.errors.push(`Loop variable ${r.loopVar.name} already bound`),this.boundVars.add(r.loopVar.name),t.push({type:"_unbind",name:r.loopVar.name}),t.push(r.body);break;case"BlockNode":for(let s of r.iterVars)s.iterVar&&(this.boundVars.has(s.iterVar.name)&&this.errors.push(`Block variable ${s.iterVar.name} already bound`),this.boundVars.add(s.iterVar.name));for(let s=r.iterVars.length-1;s>=0;s--)r.iterVars[s].iterVar&&t.push({type:"_unbind",name:r.iterVars[s].iterVar.name});t.push(r.body),r.initBody&&t.push(r.initBody);break;case"AllocateNode":r.buffer||this.errors.push("Allocate missing buffer"),t.push(r.body);break;case"LetStmtNode":this._visitExpr(r.value),this.boundVars.add(r.variable.name),t.push({type:"_unbind",name:r.variable.name}),t.push(r.body);break;case"IfThenElseNode":this._visitExpr(r.condition),r.elseBody&&t.push(r.elseBody),t.push(r.thenBody);break;case"WhileNode":r.condVar||this.errors.push("WhileNode missing condition variable"),t.push(r.loopBody),t.push(r.condBody);break;case"BufferStoreNode":if(r.buffer||this.errors.push("BufferStore missing buffer"),(!r.indices||r.indices.length!==r.buffer.shape.length)&&this.errors.push(`BufferStore rank mismatch for ${r.buffer?r.buffer.name:"unknown"}`),r.indices)for(let s of r.indices)this._visitExpr(s);this._visitExpr(r.value);break;case"EvaluateNode":this._visitExpr(r.value);break;case"_unbind":this.boundVars.delete(r.name);break;default:this._visitExpr(r);break}}}_visitExpr(e){if(e)switch(e.type){case"BufferLoadNode":if(e.buffer||this.errors.push("BufferLoad missing buffer"),(!e.indices||e.indices.length!==e.buffer.shape.length)&&this.errors.push(`BufferLoad rank mismatch for ${e.buffer?e.buffer.name:"unknown"}`),e.indices)for(let t of e.indices)this._visitExpr(t);break;case"MathOpNode":this._visitExpr(e.a),e.b&&this._visitExpr(e.b);break;case"CompareNode":this._visitExpr(e.a),this._visitExpr(e.b);break;case"CastNode":this._visitExpr(e.expr);break;case"CallExternNode":e.externName||this.errors.push("CallExtern missing function name");for(let t of e.args)this._visitExpr(t);break;case"IfThenElseNode":this._visitExpr(e.condition),this._visitExpr(e.thenBody),e.elseBody&&this._visitExpr(e.elseBody);break;case"VariableNode":this.boundVars.has(e.name)||this.errors.push(`Unbound variable used: ${e.name}`);break;case"BlockRealizeNode":e.binding&&this._visitExpr(e.binding);break;case"IntImmNode":case"FloatImmNode":break}}}});function Qv(n){let e=[];n.functionCount===0&&e.push(new Et("Module has no functions"));for(let t of n)jd(t,e);return e}function jd(n,e=[]){if(!n.entryBlock)return e.push(new Et("Function has no entry block",null,n)),e;n.entryBlock.arguments.length!==n.inputTypes.length&&e.push(new Et(`Entry block has ${n.entryBlock.arguments.length} args, expected ${n.inputTypes.length}`,null,n));for(let s=0;s<n.entryBlock.arguments.length;s++){let o=n.entryBlock.arguments[s],i=n.inputTypes[s];i&&!o.type.equals(i)&&e.push(new Et(`Block arg ${s} type ${Gn(o.type)} != expected ${Gn(i)}`,null,n))}let t=new Set;for(let s of n.entryBlock.arguments)t.add(s);for(let s of n.body)D2(s,t);for(let s of n.body)P2(s,n,t,e);let r=n.getReturnOp();if(!r)e.push(new Et("Missing return op",null,n));else{r.numOperands!==n.outputTypes.length&&e.push(new Et(`Return has ${r.numOperands} operands, function declares ${n.outputTypes.length} outputs`,r,n));for(let s=0;s<Math.min(r.numOperands,n.outputTypes.length);s++){let o=r.getOperand(s).type,i=n.outputTypes[s];i instanceof I&&o instanceof I&&(o.dtype!==i.dtype&&e.push(new Et(`Return operand ${s} dtype ${o.dtype} != expected ${i.dtype}`,r,n)),o.shapeCompatible(i)||e.push(new Et(`Return operand ${s} shape incompatible: ${Gn(o)} vs ${Gn(i)}`,r,n)))}}return e}function D2(n,e){for(let t of n.arguments)e.add(t);for(let t of n)for(let r=0;r<t.numResults;r++){let s=t.getResult(r);s&&e.add(s)}}function uj(n,e,t){let r=new Set;for(let u of n)r.add(u);let s=1,o=2,i=new Map,a=new Set;for(let u of n){if(i.get(u)!==void 0)continue;let p=[{op:u,i:0}];for(i.set(u,s);p.length>0;){let c=p[p.length-1],f=c.op;if(c.i<f.numOperands){let d=f.getOperand(c.i);c.i++;let h=d&&d.definingOp;if(!h||!r.has(h))continue;let m=i.get(h);m===s?a.has(h)||(a.add(h),t.push(new Et("participates in a value dependency cycle",h,e))):m===void 0&&(i.set(h,s),p.push({op:h,i:0}));continue}i.set(f,o),p.pop()}}}function P2(n,e,t,r){uj(n,e,r);for(let s of n)lj(s,e,t,r);if(n.size>0){let s=n.lastOp;if(n.parentRegion&&n.parentRegion.parentOp){let o=n.parentRegion.parentOp.opName,i=ne.get(o);i&&i.hasRegions&&(s.isTerminator()||r.push(new Et(`Block in ${o} region must end with terminator, got '${s.opName}'`,s,e)))}}}function lj(n,e,t,r){for(let o=0;o<n.numOperands;o++){let i=n.getOperand(o);if(!i){r.push(new Et(`Operand ${o} is null`,n,e));continue}if(!(i instanceof Li)){r.push(new Et(`Operand ${o} is not a Value`,n,e));continue}t.has(i)||r.push(new Et(`Operand ${o} used before definition`,n,e))}for(let o=0;o<n.numResults;o++){let i=n.getResult(o);i||r.push(new Et(`Result ${o} is null`,n,e)),i&&i.definingOp!==n&&r.push(new Et(`Result ${o} definingOp mismatch`,n,e))}let s=ne.get(n.opName);if(!s){!["return","yield"].includes(n.opName)&&!ne.has(n.opName)&&r.push(new Et(`Unknown op '${n.opName}'`,n,e));return}s.numOperands>=0&&n.numOperands!==s.numOperands&&r.push(new Et(`'${n.opName}' expects ${s.numOperands} operands, got ${n.numOperands}`,n,e)),s.numResults>=0&&n.numResults!==s.numResults&&r.push(new Et(`'${n.opName}' expects ${s.numResults} results, got ${n.numResults}`,n,e));for(let o of s.attrs)o.required&&!n.hasAttr(o.name)&&r.push(new Et(`'${n.opName}' missing required attribute '${o.name}'`,n,e));if(s.hasRegions&&s.numRegions>0&&n.numRegions!==s.numRegions&&r.push(new Et(`'${n.opName}' expects ${s.numRegions} regions, got ${n.numRegions}`,n,e)),s.verify){let o=s.verify(n);if(o)for(let i of o)r.push(new Et(i,n,e))}if(s.inferResultTypes&&n.numResults>0){let o=[];for(let a=0;a<n.numOperands;a++)o.push(n.getOperand(a).type);let i=s.inferResultTypes(o,n.attributes,n.results.map(a=>a.type));if(i)for(let a=0;a<Math.min(i.length,n.numResults);a++){let u=n.getResult(a).type,p=i[a];p instanceof I&&u instanceof I&&(u.dtype!==p.dtype&&r.push(new Et(`Result ${a} dtype ${u.dtype} != inferred ${p.dtype}`,n,e)),u.shapeCompatible(p)||r.push(new Et(`Result ${a} shape ${Gn(u)} incompatible with inferred ${Gn(p)}`,n,e)))}}for(let o of n.regions){let i=new Set(t);for(let a of o)D2(a,i);for(let a of o)P2(a,e,i,r)}}var Et,$2=M(()=>{"use strict";Re();Ih();Ct();Et=class{static{l(this,"VerificationError")}constructor(e,t=null,r=null){this.message=e,this.op=t,this.func=r}toString(){let e="";return this.func&&(e+=`[${this.func.name}] `),this.op&&(e+=`op '${this.op.opName}' (id=${this.op.id}): `),e+this.message}};l(Qv,"verifyModule");l(jd,"verifyFunction");l(D2,"collectScopeDefs");l(uj,"detectCycles");l(P2,"verifyBlock");l(lj,"verifyOperation")});var nw,Xv,Dc,ek,tk=M(()=>{"use strict";cv();Re();nw=class{static{l(this,"ValueObserver")}constructor(){this.min=1/0,this.max=-1/0,this.count=0,this.histogram=null}update(e){for(let t=0;t<e.length;t++){let r=e[t];r<this.min&&(this.min=r),r>this.max&&(this.max=r)}this.count+=e.length,this.histogram&&this.histogram.update(e)}enableHistogram(e=2048){this.histogram=new Xv(e)}},Xv=class{static{l(this,"HistogramCollector")}constructor(e=2048){this.numBins=e,this.bins=new Float64Array(e),this.rangeMin=0,this.rangeMax=0,this.initialized=!1,this.totalCount=0}update(e){this.initialized||this._initRange(e);let t=this.rangeMax-this.rangeMin;if(t<=0)return;for(let s=0;s<e.length;s++){let o=e[s];if(o<this.rangeMin||o>this.rangeMax){this._expandAndRebucket(e);return}}let r=this.numBins/t;for(let s=0;s<e.length;s++){let o=Math.min(this.numBins-1,Math.floor((e[s]-this.rangeMin)*r));this.bins[o]++}this.totalCount+=e.length}_initRange(e){let t=1/0,r=-1/0;for(let s=0;s<e.length;s++)e[s]<t&&(t=e[s]),e[s]>r&&(r=e[s]);t===r&&(r=t+1),this.rangeMin=t,this.rangeMax=r,this.initialized=!0}_expandAndRebucket(e){let t=this.rangeMin,r=this.rangeMax;for(let c=0;c<e.length;c++)e[c]<t&&(t=e[c]),e[c]>r&&(r=e[c]);let s=new Float64Array(this.bins),o=this.rangeMin,a=this.rangeMax-o;this.rangeMin=t,this.rangeMax=r;let u=r-t,p=this.numBins/u;if(this.bins.fill(0),a>0){let c=a/this.numBins;for(let f=0;f<this.numBins;f++){if(s[f]===0)continue;let d=o+(f+.5)*c,h=Math.min(this.numBins-1,Math.floor((d-t)*p));this.bins[h]+=s[f]}}for(let c=0;c<e.length;c++){let f=Math.min(this.numBins-1,Math.floor((e[c]-t)*p));this.bins[f]++}this.totalCount+=e.length}computePercentileThreshold(e){let t=this.totalCount*e,r=0,s=(this.rangeMax-this.rangeMin)/this.numBins;for(let o=0;o<this.numBins;o++)if(r+=this.bins[o],r>=t)return this.rangeMin+(o+1)*s;return this.rangeMax}computeEntropyThreshold(e){if(this.totalCount===0)return this.rangeMax;let t=new Float64Array(this.numBins);for(let i=0;i<this.numBins;i++)t[i]=this.bins[i]/this.totalCount;let r=this.numBins,s=1/0;for(let i=e;i<=this.numBins;i++){let a=i/e,u=new Float64Array(i);for(let c=0;c<e;c++){let f=Math.floor(c*a),d=Math.min(i,Math.floor((c+1)*a)),h=0;for(let g=f;g<d;g++)h+=t[g];let m=d-f;if(m>0){let g=h/m;for(let y=f;y<d;y++)u[y]=g}}let p=0;for(let c=0;c<i;c++)t[c]>0&&u[c]>0&&(p+=t[c]*Math.log(t[c]/u[c]));p<s&&(s=p,r=i)}let o=(this.rangeMax-this.rangeMin)/this.numBins;return this.rangeMin+r*o}},Dc=class{static{l(this,"CalibrationCollector")}constructor(e="minmax"){this.observers=new Map,this.mode=e}attach(e){for(let t of e.ops())for(let r=0;r<t.numResults;r++){let s=t.getResult(r);if(s.type instanceof I&&jt(s.type.dtype)){let o=new nw;(this.mode==="entropy"||this.mode==="percentile")&&o.enableHistogram(),this.observers.set(s,o)}}for(let t of e.args)if(t.type instanceof I&&jt(t.type.dtype)){let r=new nw;(this.mode==="entropy"||this.mode==="percentile")&&r.enableHistogram(),this.observers.set(t,r)}}observe(e,t){let r=this.observers.get(e);r&&r.update(t)}getResult(){return new ek(this.observers,this.mode)}},ek=class{static{l(this,"CalibrationResult")}constructor(e,t){this._observers=e,this._mode=t}getRange(e){let t=this._observers.get(e);return!t||t.count===0?null:{min:t.min,max:t.max}}getQuantParams(e,t,r){let s=this.getRange(e);if(!s)return null;let o=t===Pr.PER_TENSOR_ASYMMETRIC;if(this._mode==="percentile"&&this._observers.get(e).histogram){if(o)return vr.fromRange(s.min,s.max,t,r);let a=this._observers.get(e).histogram.computePercentileThreshold(.999);return vr.fromRange(-a,a,t,r)}if(this._mode==="entropy"&&this._observers.get(e).histogram){if(o)return vr.fromRange(s.min,s.max,t,r);let i=this._observers.get(e).histogram,a=r==="ui8"?256:255,u=i.computeEntropyThreshold(a);return vr.fromRange(-u,u,t,r)}return vr.fromRange(s.min,s.max,t,r)}hasData(e){let t=this._observers.get(e);return t&&t.count>0}values(){return this._observers.keys()}}});function pj(n,e){let t=[],r=new Set;for(let s of n.ops())if(e.has(s.opName))for(let o=0;o<s.numOperands;o++){let i=s.getOperand(o);if(r.has(i)||!(i.type instanceof I)||!jt(i.type.dtype))continue;let a=i.definingOp;a&&a.opName==="constant"||(r.add(i),t.push(i))}return t}function fj(n,e){let t=e.map(h=>h.type),r=new yn(n.name,n.inputTypes,[...n.outputTypes,...t]),s=new Map,o=n.entryBlock.arguments,i=r.entryBlock.arguments;for(let h=0;h<o.length;h++)s.set(o[h],i[h]);let a=n.entryBlock.opsArray(),u=new Set(a),p=Qo(a,h=>u.has(h),"ignore"),c=new Map;for(let h of p)h.opName!=="return"&&c.set(h,h.clone(s));for(let h of a)h.opName!=="return"&&r.entryBlock.pushOp(c.get(h));let f=n.getReturnOp(),d=[];if(f)for(let h of f.operands)d.push(s.get(h)||h);for(let h of e)d.push(s.get(h));return r.entryBlock.pushOp(new Me("return",d,[],{})),r}function F2(n,e,t,r={}){if(e.isGPU&&e.isGPU())throw new Error("collectCalibration: synchronous calibration is unavailable for async (GPU) targets; precompute calibration on a CPU/WASM target or supply quantization.calibration directly");if(!t||t.length===0)throw new Error("collectCalibration: at least one calibration batch is required");let s=r.quantizableOps||cj,o=r.mode||"minmax",i=pj(n,s),a=new Dc(o);a.attach(n);let u=a.getResult();if(i.length===0)return u;let p=fj(n,i),c=new Mn("__calibrate__");c.addFunction(p);let f=n.outputTypes.length,h=(r.compileFn||dj)(c,e);for(let m of t){let g=Array.isArray(m)?m:[m],y=p.outputTypes.map(_=>new Float32Array(Math.max(1,_.numel())));h.run(n.name,...g,...y);for(let _=0;_<i.length;_++)a.observe(i[_],y[f+_])}return u}function dj(n,e){throw new Error("collectCalibration: opts.compileFn is required (pass the compileModule function to avoid a circular import)")}var cj,z2=M(()=>{"use strict";Us();Di();yr();Re();Da();tk();cj=new Set(["dot","conv","add","mul","sub"]);l(pj,"activationTargets");l(fj,"buildCaptureFunction");l(F2,"collectCalibration");l(dj,"defaultCompileFn")});function rk(n){let e=0;for(let t=0;t<n.numResults;t++){let r=n.getResult(t).type;r instanceof I&&r.isFullyStatic&&(e+=r.sizeInBytes())}return e}var nk,Pc,sk,rw,V2=M(()=>{"use strict";Ct();Re();Ad();nk=class{static{l(this,"Partition")}constructor(e,t){this.id=e,this.target=t,this.ops=[],this.opSet=new Set,this._inputValues=null,this._outputValues=null,this._memoryBytes=0}addOp(e){this.opSet.has(e)||(this.ops.push(e),this.opSet.add(e),this._inputValues=null,this._outputValues=null,this._memoryBytes+=rk(e))}hasOp(e){return this.opSet.has(e)}merge(e){for(let t of e.ops)this.addOp(t)}computeIO(){if(this._inputValues&&this._outputValues)return;let{inputs:e,outputs:t}=pa(this.opSet,this.ops);this._inputValues=e,this._outputValues=t}getInputValues(){return this.computeIO(),this._inputValues}getOutputValues(){return this.computeIO(),this._outputValues}get size(){return this.ops.length}get memoryBytes(){return this._memoryBytes}};l(rk,"estimateOpMemory");Pc=class{static{l(this,"PartitionerConfig")}constructor(e={}){this.targets=e.targets||[],this.defaultTarget=e.defaultTarget||null,this.opTargetOverrides=e.opTargetOverrides||new Map,this.memoryLimits=e.memoryLimits||new Map,this.minPartitionSize=e.minPartitionSize||1,this.costWeights={transferCost:1,loadBalance:.5,...e.costWeights||{}}}},sk=class{static{l(this,"PartitionResult")}constructor(e,t,r){this.partitions=e,this.opToPartition=t,this.transferEdges=r}getPartition(e){return this.opToPartition.get(e)||null}getPartitionsForTarget(e){return this.partitions.filter(t=>t.target===e||t.target.name===e.name)}get numPartitions(){return this.partitions.length}},rw=class{static{l(this,"GraphPartitioner")}constructor(e){this.config=e instanceof Pc?e:new Pc(e),this._supportCache=new Map,this._buildSupportMap()}partition(e){let t=this._collectPartitionableOps(e),r=this._assignTargets(t),s=this._buildPartitions(t,r),o=this._mergeSmallPartitions(s,r),i=this._computeTransferEdges(o),a=new Map;for(let u of o)for(let p of u.ops)a.set(p,u);return new sk(o,a,i)}_buildSupportMap(){for(let e of this.config.targets){let t=new Set;for(let r of ne.names())this._targetSupportsOp(e,r)&&t.add(r);this._supportCache.set(e,t)}}_targetSupportsOp(e,t){if(e.hasLibraryOp(t))return!0;let r=ne.get(t);return r?r.isConstant||r.isTerminator?!0:e.isGPU()?r.isElementwise||r.isReduction||r.isBroadcast||r.isInjective||r.getAttr("gpuCapable")===!0:e.isCPU()?!0:e.isWasm()?!r.isOpaque&&t!=="custom_call":!1:!1}_collectPartitionableOps(e){let t=[];for(let r of e.ops()){let s=ne.get(r.opName);!s||s.isTerminator||t.push(r)}return t}_assignTargets(e){let t=new Map;for(let r of e){let s=r.getAttr("device");if(s){let a=this._resolveDeviceAttr(s);if(a){t.set(r,a);continue}}let o=this.config.opTargetOverrides.get(r.opName);if(o){t.set(r,o);continue}let i=this._selectBestTarget(r);t.set(r,i)}return t}_resolveDeviceAttr(e){return typeof e=="string"?this.config.targets.find(t=>t.name===e||t.kind===e)||null:e}_selectBestTarget(e){let t=this.config.defaultTarget||this.config.targets[0],r=-1/0;for(let s of this.config.targets){let o=this._supportCache.get(s);if(!o||!o.has(e.opName))continue;let i=this._scoreTargetForOp(s,e);i>r&&(r=i,t=s)}return t}_scoreTargetForOp(e,t){let r=ne.get(t.opName);if(!r)return 0;let s=0;if(e.hasLibraryOp(t.opName)&&(s+=100),r.isReduction||r.isElementwise){let o=0;for(let i=0;i<t.numOperands;i++){let a=t.getOperand(i).type;a instanceof I&&a.isFullyStatic&&(o+=a.numel())}e.isGPU()&&o>1024?s+=50:e.isCPU()&&o<=1024&&(s+=30)}return s+=e.computeTFLOPs*10,s}_buildPartitions(e,t){let{partitions:r}=Nd(e,{sort:l(i=>this._topologicalSort(i),"sort"),labelOf:l(i=>t.get(i),"labelOf"),sameLabel:l((i,a)=>i===a||i.name===a.name,"sameLabel"),canMerge:l((i,a,u)=>this._fitsMemoryLimit(i,a,u),"canMerge"),onAttach:l((i,a)=>{i.memoryBytes=(i.memoryBytes||0)+rk(a)},"onAttach")}),s=new Map;for(let i of r){let a=new nk(i.id,i.label);for(let u of i.ops)a.addOp(u);s.has(i.label.name)||s.set(i.label.name,[]),s.get(i.label.name).push(a)}let o=[];for(let i of s.values())for(let a of i)o.push(a);return o}_fitsMemoryLimit(e,t,r){let s=this.config.memoryLimits.get(r.name);if(!s)return!0;let o=rk(t);return e.memoryBytes+o<=s}_mergeSmallPartitions(e,t){if(e.length<=1)return e;let r=new Map;for(let h of e)for(let m of h.ops)r.set(m,h);let s=new Set,o=new Map,i=new Map;for(let h of e)s.has(h)||o.set(h,new Set);for(let h of e){if(s.has(h))continue;let m=o.get(h);for(let g of h.ops)for(let y=0;y<g.numResults;y++)for(let _ of g.getResult(y).uses()){let b=r.get(_.user);b&&b!==h&&m.add(b)}}let a=l((h,m)=>{let g=o.get(h);for(let y of o.get(m))y!==h&&g.add(y);g.delete(m);for(let[y,_]of o)_.has(m)&&(_.delete(m),y!==h&&_.add(h));o.delete(m),i=new Map},"mergeSucc"),u=l(h=>{let m=i.get(h);if(m)return m;m=new Set;let g=[...o.get(h)];for(;g.length>0;){let y=g.pop();if(m.has(y))continue;m.add(y);let _=o.get(y);if(_)for(let b of _)g.push(b)}return i.set(h,m),m},"reachOf"),p=l((h,m)=>{for(let g of o.get(h))if(g!==m&&u(g).has(m))return!0;return!1},"pathThroughIntermediate"),c=l((h,m)=>p(h,m)||p(m,h),"mergeCreatesCycle"),f=[];for(let h=0;h<e.length;h++){let m=e[h];if(s.has(m))continue;if(m.size>=this.config.minPartitionSize){f.push(m);continue}let g=-1,y=-1/0;for(let _=0;_<e.length;_++){if(h===_)continue;let b=e[_];if(s.has(b)||b.target.name!==m.target.name||c(m,b))continue;let w=this.config.costWeights.transferCost*this._mergeScore(m,b);w>y&&(y=w,g=_)}if(g>=0){let _=e[g];a(_,m),_.merge(m);for(let b of m.ops)r.set(b,_);s.add(m)}else f.push(m)}let d=new Set;for(let h of f)for(let m of h.ops)d.add(m);for(let h of e){let m=!1;for(let g of h.ops)if(!d.has(g)){m=!0;break}if(m){f.push(h);for(let g of h.ops)d.add(g)}}return f}_mergeScore(e,t){let r=0,s=new Set;for(let o of e.ops)for(let i=0;i<o.numResults;i++)s.add(o.getResult(i));for(let o of t.ops)for(let i=0;i<o.numOperands;i++)s.has(o.getOperand(i))&&r++;return r}_computeTransferEdges(e){let t=[],r=new Set,s=new Map;for(let o of e)for(let i of o.ops)s.set(i,o);for(let o of e)for(let i of o.ops)for(let a=0;a<i.numOperands;a++){let u=i.getOperand(a).definingOp;if(!u)continue;let p=s.get(u);if(!p||p===o)continue;let c=i.getOperand(a),f=`${p.id}|${o.id}|${c.id}`;r.has(f)||(r.add(f),t.push({src:p,dst:o,value:c,sizeBytes:c.type instanceof I&&c.type.isFullyStatic?c.type.sizeInBytes():0}))}return t}_topologicalSort(e){let t=new Set(e),r=new Map,s=new Map;for(let u of e)r.set(u,0),s.set(u,[]);for(let u of e)for(let p=0;p<u.numOperands;p++){let c=u.getOperand(p).definingOp;c&&t.has(c)&&(s.get(c).push(u),r.set(u,r.get(u)+1))}let o=[];for(let u of e)r.get(u)===0&&o.push(u);let i=[],a=0;for(;a<o.length;){let u=o[a++];i.push(u);for(let p of s.get(u)){let c=r.get(p)-1;r.set(p,c),c===0&&o.push(p)}}return i}}});var sw,ow,G2=M(()=>{"use strict";fn();yr();Us();V2();Ad();nn();sw=class extends qe{static{l(this,"GraphPartitionPass")}constructor(e={}){super("GraphPartitionPass"),this.partitionerConfig=new Pc(e),this.partitionResult=null}run(e,t){if(this.partitionerConfig.targets.length<2)return oe.UNCHANGED;let r=new rw(this.partitionerConfig);return this.partitionResult=r.partition(e),this.partitionResult.numPartitions<=1?oe.UNCHANGED:(this._annotateOps(),this._insertTransferOps(e),this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,numPartitions:this.partitionResult.numPartitions,transferEdges:this.partitionResult.transferEdges.length,level:le.DEBUG}),e.bumpVersion(),oe.CHANGED)}_annotateOps(){for(let e of this.partitionResult.partitions)for(let t of e.ops)t.setAttr("partition_id",e.id),t.setAttr("partition_target",e.target.name)}_insertTransferOps(e){let t=e.entryBlock,r=this._buildOrderIndex(t),{useMap:s,firstInPart:o}=this._buildInsertionIndex(t);for(let i of this.partitionResult.transferEdges){let a=i.value,u=i.src.target.name,p=i.dst.target.name;if(u===p)continue;let c=new Me("copy_to_device",[a],[a.type],{src_device:u,dst_device:p});c.setAttr("partition_id",i.dst.id),c.setAttr("partition_target",p);let f=s.get(i.dst),d=f&&f.get(a)||o.get(i.dst)||null,h=a.definingOp;if(d&&h&&r.has(h)&&r.has(d)&&r.get(h)>=r.get(d))t.insertAfter(c,h);else if(d)t.insertBefore(c,d);else if(h&&r.has(h))t.insertAfter(c,h);else{let g=e.getReturnOp();g?t.insertBefore(c,g):t.pushOp(c)}let m=c.getResult(0);for(let g of i.dst.ops)for(let y=0;y<g.numOperands;y++)g.getOperand(y)===a&&g.replaceOperand(y,m)}}_buildInsertionIndex(e){let t=[...new Set(this.partitionResult.transferEdges.map(o=>o.dst))],r=new Map,s=new Map;for(let o of t)r.set(o,new Map);for(let o of e.ops())for(let i of t){if(!i.hasOp(o))continue;s.has(i)||s.set(i,o);let a=r.get(i);for(let u=0;u<o.numOperands;u++){let p=o.getOperand(u);a.has(p)||a.set(p,o)}}return{useMap:r,firstInPart:s}}_buildOrderIndex(e){let t=new Map,r=0;for(let s of e.ops())t.set(s,r++);return t}},ow=class extends qe{static{l(this,"PartitionMaterializationPass")}constructor(e={}){super("PartitionMaterializationPass"),this.targets=e.targets||[]}run(e,t){let r=this._collectPartitions(e);if(r.size<=1)return oe.UNCHANGED;let s=this._getModule(e);if(!s)return oe.UNCHANGED;let o=this._materializePartitions(e,r);for(let i of o)s.addFunction(i);return this._rewriteOriginalFunction(e,o,r),this.trace&&this.trace.level>=le.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,partitions:r.size,subFunctions:o.length,level:le.DEBUG}),e.bumpVersion(),oe.CHANGED}_collectPartitions(e){let t=new Map;for(let r of e.ops()){let s=r.getAttr("partition_id");s!==void 0&&(t.has(s)||t.set(s,{id:s,target:r.getAttr("partition_target"),ops:[]}),t.get(s).ops.push(r))}return t}_getModule(e){return e._module||null}_materializePartitions(e,t){let r=[];for(let[s,o]of t){let i=new Set(o.ops),{inputs:a,outputs:u}=pa(i,o.ops),p=a.map(_=>_.type),c=u.map(_=>_.type),f=`${e.name}_partition_${s}`,d=new yn(f,p,c);d._partitionTarget=o.target;let h=new Map;for(let _=0;_<a.length;_++)h.set(a[_],d.args[_]);let m=Ru(o.ops);for(let _ of m)d.entryBlock.pushOp(_.clone(h));let g=u.map(_=>h.get(_)||_),y=new Me("return",g,[]);d.entryBlock.pushOp(y),r.push(d)}return r}_rewriteOriginalFunction(e,t,r){for(let s of t)e.setAttr?.(`sub_${s.name}`,s._partitionTarget)}}});function Ud(n){return ap(n,-1)}function hj(n,e,t){let r=t.sharedMemoryBytes||16384,s=0;for(let u of e.entryBlock.ops())if(u.opName!=="yield")for(let p=0;p<u.numResults;p++){let c=u.getResult(p).type;if(!c||!c.shape)continue;let f=Ud(c.shape);if(f<0)return!0;let d=f*Rn(c.dtype);d>s&&(s=d)}let o=n.getAttr("num_carry"),i=n.getAttr("num_xs"),a=0;for(let u=0;u<o;u++){let p=n.getOperand(i+u).type,c=p&&p.shape?Ud(p.shape):-1;c>0&&(a+=c*Rn(p.dtype))}return 3*s+2*a>r}function mj(n,e){let r=n.regions[0].entryBlock,s=n.getAttr("num_carry"),o=[],i=null;for(let b of r.ops()){if(b.opName==="yield"){i=b;continue}o.push(b)}if(!i)return null;let a=new Set(o),u=new Set(r.arguments),p=[],c=new Set,f=[],d=new Set;for(let b of o)for(let w=0;w<b.numOperands;w++){let x=b.getOperand(w),S=x.definingOp;if(!u.has(x)&&!(S&&a.has(S))){if(S&&$o(S)){d.has(S)||(d.add(S),f.push(S));continue}c.has(x)||(c.add(x),p.push(x))}}let h=[...r.arguments,...p];for(let b of h)if(!b.type||!b.type.isFullyStatic)return null;let m=[];for(let b=0;b<i.numOperands;b++)m.push(i.getOperand(b));let g=new yn(e,h.map(b=>b.type),m.map(b=>b.type)),y=new Map;for(let b=0;b<h.length;b++)y.set(h[b],g.args[b]);for(let b of f)g.entryBlock.pushOp(b.clone(y));for(let b of o)g.entryBlock.pushOp(b.clone(y));let _=m.map(b=>y.get(b));return _.some(b=>b===void 0)?null:(g.entryBlock.pushOp(new Me("return",_,[])),{bodyFunc:g,captured:p,numCarry:s,numYs:m.length-s})}function gj(n,e,t,r,s){let{getSlot:o,newSlot:i,steps:a,addedFuncs:u}=r,p=new Mn(n.name+"_mod");p.addFunction(n);let c=ha(p,s);if(c){let f=c.plan,d=e.length,h=new Map,m=!0;for(let g of f.returnFixups||[]){if(g.kind!=="copy"){m=!1;break}h.set(g.pos,g.srcSlot)}if(m){let g=l(x=>{let S=d+x;return h.has(S)?h.get(S):f.argSlots[S]},"retSlot"),y=new Map;for(let x=0;x<d;x++)y.set(f.argSlots[x],o(e[x]));for(let x of f.intermediates)y.set(x.slot,i(x.shape,x.dtype));for(let x=0;x<t.length;x++){let S=g(x);y.has(S)||y.set(S,o(t[x]))}let _=l(x=>{let S=y.get(x);return S===void 0?null:S},"mapSlot"),b=[],w=!0;for(let x of f.steps){let S=x.inputSlots.map(_),T=x.outputSlots.map(_);if(S.includes(null)||T.includes(null)){w=!1;break}b.push({name:x.name,inputSlots:S,outputSlots:T})}if(w){for(let x of b)a.push(x);for(let x of p.functions())u.push(x);return!0}}}return a.push({name:n.name,inputSlots:e.map(o),outputSlots:t.map(o)}),u.push(n),!0}function j2(n,e,t){let r=db({ops:n,opSet:new Set(n)},e,new Map);return r?gj(r.subFunc,r.inputs,r.outputs,t,1):!1}function yj(n,e,t){let{getSlot:r,newSlot:s,steps:o,scanLoops:i,addedFuncs:a}=t,u=n.regions[0];if(!u||!u.entryBlock)return!1;let p=mj(n,e);if(!p)return!1;let{bodyFunc:c,captured:f,numCarry:d,numYs:h}=p,m=n.getAttr("num_xs"),g=new Mn(e+"_mod");g.addFunction(c);let y=ha(g,2),_=y?y.plan:null,b=[],w=[];for(let D=0;D<d;D++){let V=n.getOperand(m+D).type;b.push(V.shape),w.push(V.dtype)}let x=[],S=[];for(let D=0;D<m;D++){let V=n.getOperand(D).type;x.push(V.shape.slice(1)),S.push(V.dtype)}let T=[],v=[];for(let D=0;D<h;D++){let V=n.getResult(d+D).type;T.push(V.shape.slice(1)),v.push(V.dtype)}let A=b.map((D,V)=>s(D,w[V])),N=b.map((D,V)=>s(D,w[V])),k=x.map((D,V)=>s(D,S[V])),O=[],R=[],B=[],E=[];for(let D=0;D<d;D++)R.push(r(n.getOperand(m+D)));for(let D=0;D<d;D++)B.push(r(n.getResult(D)));for(let D of f)E.push(r(D));let F=[],C=[];for(let D=0;D<m;D++)F.push(r(n.getOperand(D)));for(let D=0;D<h;D++)C.push(r(n.getResult(d+D)));let H=o.length;if(!_)O=T.map((D,V)=>s(D,v[V])),o.push({name:c.name,inputSlots:[...k,...A,...E],outputSlots:[...N,...O]}),a.push(c);else{let D=m+d+f.length,V=new Map;for(let j of _.returnFixups||[]){if(j.kind!=="copy")return!1;V.set(j.pos,j.srcSlot)}let ue=l(j=>{let W=D+j;return V.has(W)?V.get(W):_.argSlots[W]},"bodyReturnSlot"),ce=new Map;for(let j=0;j<m;j++)ce.set(_.argSlots[j],k[j]);for(let j=0;j<d;j++)ce.set(_.argSlots[m+j],A[j]);for(let j=0;j<f.length;j++)ce.set(_.argSlots[m+d+j],E[j]);for(let j of _.intermediates)ce.set(j.slot,s(j.shape,j.dtype));for(let j=0;j<d;j++){let W=ue(j);ce.has(W)||ce.set(W,N[j])}for(let j=0;j<h;j++){let W=ue(d+j),ie=ce.get(W);ie===void 0&&(ie=s(T[j],v[j]),ce.set(W,ie)),O.push(ie)}let Le=l(j=>{let W=ce.get(j);return W===void 0?null:W},"mapSlot");for(let j of _.steps){let W=j.inputSlots.map(Le),ie=j.outputSlots.map(Le);if(W.includes(null)||ie.includes(null))return!1;o.push({name:j.name,inputSlots:W,outputSlots:ie})}for(let j of g.functions())a.push(j)}let G=o.length,U=n.getOperand(0).type.shape[0];return typeof U!="number"||U<0?!1:(i.push({T:U,loopStart:H,loopEnd:G,carry:b.map((D,V)=>({a:A[V],b:N[V],initSlot:R[V],finalSlot:B[V],bytes:Ud(D)*Rn(w[V])})),xs:k.map((D,V)=>({xtSlot:D,xsSlot:F[V],stepBytes:Ud(x[V])*Rn(S[V])})),ys:O.map((D,V)=>({ytSlot:D,ysSlot:C[V],stepBytes:Ud(T[V])*Rn(v[V])}))}),!0)}function ok(n,e,t=!1){if(!e||typeof e.isWebGPU!="function"||!e.isWebGPU()||n.functionCount!==1)return null;let r=n.functions().next().value,s=r.getReturnOp();if(!s)return null;let o=[];for(let R of r.ops())R.opName==="scan"&&o.push(R);if(o.length===0)return null;let i=new Set,a=l(R=>{if(!(!R||i.has(R))){i.add(R);for(let B=0;B<R.numOperands;B++)a(R.getOperand(B).definingOp)}},"visit");for(let R=0;R<s.numOperands;R++)a(s.getOperand(R).definingOp);for(let R of o)if(!i.has(R))return null;if(!t){let R=!1;for(let B of o){let E=B.regions[0];if(E&&E.entryBlock&&hj(B,E,e)){R=!0;break}}if(!R)return null}let u=[];for(let R of i)$o(R)||Xt(R.opName)||u.push(R);let p=Ru(u),c=new Map,f=0,d=l(R=>{let B=c.get(R);return B===void 0&&(B=f++,c.set(R,B)),B},"getSlot");for(let R of r.args)d(R);let h=[],m=l((R,B)=>{let E=f++;return h.push({slot:E,shape:[...R],dtype:B}),E},"newSlot"),g=[],y=[],_=[],b={getSlot:d,newSlot:m,steps:g,scanLoops:y,addedFuncs:_},w=new Set(o),x=[],S=0,T=0;for(let R of p)if(w.has(R)){if(x.length&&!j2(x,r.name+"_seg"+S++,b)||(x=[],!yj(R,r.name+"_scan"+T++,b)))return null}else x.push(R);if(x.length&&!j2(x,r.name+"_seg"+S++,b))return null;let v=r.args.map(d),A=[],N=new Set;for(let R=0;R<s.numOperands;R++){let B=s.getOperand(R),E=v.length;if(c.has(B)&&!(B.isBlockArgument&&B.isBlockArgument())){let F=d(B);N.has(F)?(v.push(f++),A.push({pos:E,kind:"copy",srcSlot:F})):(N.add(F),v.push(F))}else if(B.isBlockArgument&&B.isBlockArgument())v.push(f++),A.push({pos:E,kind:"copy",srcSlot:d(B)});else return null}let k=new Set(v),O=new Set;for(let[R,B]of c)if(!(k.has(B)||O.has(B))){if(O.add(B),!R.type||!R.type.isFullyStatic)return null;h.some(E=>E.slot===B)||h.push({slot:B,shape:[...R.type.shape],dtype:R.type.dtype})}n.removeFunction(r.name);for(let R of _)n.addFunction(R);return{plan:{numSlots:f,argSlots:v,intermediates:h,steps:g,returnFixups:A,scanLoops:y}}}var U2=M(()=>{"use strict";yr();Us();Di();hb();Ir();tr();Re();l(Ud,"numel");l(hj,"isScanOversized");l(mj,"buildScanBodyFunction");l(gj,"inlineNativeSplit");l(j2,"emitSegment");l(yj,"emitScanLoop");l(ok,"splitGraphForScan")});function qd(n){ik.push(n),ik.sort((e,t)=>(e.priority??100)-(t.priority??100))}function _j(n){for(let e of ik)if(e.applies(n))return e;return null}function q2(n,e){let t=_j(e);return t?t.run(n,e):null}var ik,W2=M(()=>{"use strict";hb();U2();ik=[];l(qd,"registerGraphSplitStrategy");l(_j,"selectGraphSplitStrategy");l(q2,"splitGraph");qd({name:"cublas",priority:10,applies:l(n=>n.config.matmulBackend==="cublas","applies"),run:l(n=>uL(n),"run")});qd({name:"cuda-attention",priority:15,applies:l(n=>n.cudaAttention,"applies"),run:l(n=>ha(n,1),"run")});qd({name:"cuda-native-matmul-chain",priority:20,applies:l(n=>n.cudaMatmulChain,"applies"),run:l(n=>ha(n),"run")});qd({name:"cuda-native-conv-chain",priority:25,applies:l(n=>n.cudaConvChain,"applies"),run:l(n=>ha(n),"run")});qd({name:"webgpu",priority:30,applies:l(n=>n.isWebGPU,"applies"),run:l((n,e)=>{let t=ok(n,e.target);return!t&&iL(n,e.target.maxThreadsPerBlock||256)&&(t=ok(n,e.target,!0),t||(t=ha(n,2))),t},"run")})});function K2(n,e){let t=new Om;for(let[,r]of n.bufferMap)t.paramBuffers.add(r.name);return bj(n.body,t,e),Tj(n,t,e),xj(n.body,t),t}function bj(n,e,t){let r=[n];for(;r.length>0;){let s=r.pop();if(!(!s||typeof s!="object")){switch(s.type){case"ForNode":if(s.loopVar&&e.locals.set(s.loopVar.name,"i32"),s.kind===z.THREAD_BINDING&&s.threadTag){let o=s.extent&&s.extent.type==="IntImmNode"?s.extent.value:0,i=!s.extent||s.extent.type!=="IntImmNode",a={varName:s.loopVar.name,extent:o,isDynamic:i,extentNode:s.extent};e.threadBindings.has(s.threadTag)?e.threadBindings.get(s.threadTag).push(a):e.threadBindings.set(s.threadTag,[a])}break;case"LetStmtNode":s.variable&&e.locals.set(s.variable.name,Vi(s.variable.dtype));break;case"BufferStoreNode":case"BufferLoadNode":s.buffer&&e.usedBuffers.set(s.buffer.name,s.buffer);break;case"AllocateNode":s.buffer&&e.allocatedBuffers.add(s.buffer.name),s.scope==="shared"&&s.buffer&&e.sharedBuffers.push(s.buffer);break;case"CallExternNode":s.externName&&!n1(s.externName)&&e.externCalls.set(s.externName,{argCount:s.args?s.args.length:0,dtype:s.dtype||"f32"});break;case"BlockNode":if(s.reads)for(let o of s.reads)o.buffer&&e.usedBuffers.set(o.buffer.name,o.buffer);if(s.writes)for(let o of s.writes)o.buffer&&e.usedBuffers.set(o.buffer.name,o.buffer);if(s.iterVars)for(let o of s.iterVars)o.iterVar&&e.locals.set(o.iterVar.name,Vi(o.iterVar.dtype)),o.binding&&typeof o.binding=="object"&&o.binding.type&&r.push(o.binding);break}if(s.body&&r.push(s.body),s.value&&typeof s.value=="object"&&s.value.type&&r.push(s.value),s.stmts)for(let o of s.stmts)r.push(o);if(s.thenBody&&r.push(s.thenBody),s.elseBody&&r.push(s.elseBody),s.initBody&&r.push(s.initBody),s.condBody&&r.push(s.condBody),s.loopBody&&r.push(s.loopBody),s.condition&&typeof s.condition=="object"&&s.condition.type&&r.push(s.condition),s.a&&typeof s.a=="object"&&s.a.type&&r.push(s.a),s.b&&typeof s.b=="object"&&s.b.type&&r.push(s.b),s.expr&&typeof s.expr=="object"&&s.expr.type&&r.push(s.expr),s.args)for(let o of s.args)typeof o=="object"&&o!==null&&o.type&&r.push(o);if(s.indices)for(let o of s.indices)typeof o=="object"&&o!==null&&o.type&&r.push(o)}}}function Tj(n,e,t){let r=e.memoryLayout.alignment,s=0,o=l(i=>{let a=i.shape.some(c=>typeof c!="number"||c<0),u=i.numel();if(!a&&u>=0)return u*Rn(i.dtype);let p=1;for(let c of i.shape)p*=typeof c=="number"&&c>0?c:1;return Math.max(wj,p*Rn(i.dtype))},"bufBytes");for(let[,i]of n.bufferMap)s=Math.ceil(s/r)*r,e.memoryLayout.bufferOffsets.set(i.name,s),s+=o(i);for(let[i,a]of e.usedBuffers)e.memoryLayout.bufferOffsets.has(i)||(s=Math.ceil(s/r)*r,e.memoryLayout.bufferOffsets.set(i,s),s+=o(a));e.memoryLayout.totalBytes=s}function xj(n,e){let t=new Map;for(let r of yl(n,s=>s.type==="BufferStoreNode"&&s.buffer&&!e.paramBuffers.has(s.buffer.name))){let s=r.buffer.name;t.has(s)||t.set(s,[]),t.get(s).push(r.value)}for(let[r,s]of t)if(s.every(i=>i&&i.type==="IntImmNode"&&i.value===0||i&&i.type==="FloatImmNode"&&i.value===0)&&e.zeroBuffers.add(r),s.length===1&&s[0]){let i=s[0];(i.type==="IntImmNode"||i.type==="FloatImmNode")&&e.constantBuffers.set(r,i.value)}}var wj,H2=M(()=>{"use strict";vp();tr();Fe();ts();l(K2,"scanMetadata");l(bj,"walkTree");wj=65536;l(Tj,"computeMemoryLayout");l(xj,"detectZeroBuffers")});function Wd(n,e,t){let r=typeof n.offset=="number"?n.offset:0;if(e.length===0)return new L(r);if(e.length===1)return r===0?e[0]:new P("+",e[0],new L(r));let s=[];r!==0&&s.push(new L(r));for(let o=0;o<e.length;o++){let i=e[o];if(i.type==="IntImmNode"&&i.value===0)continue;let a=n.strides[o];if(a===1)s.push(i);else if(typeof a=="number"&&a>=0)s.push(new P("*",i,new L(a)));else{let u=Sj(n,o,t);s.push(new P("*",i,u))}}return s.length===0?new L(0):s.reduce((o,i)=>new P("+",o,i))}function Sj(n,e,t){let r=[];for(let s=e+1;s<n.shape.length;s++){let o=n.shape[s];typeof o=="number"&&o>=0?r.push(new L(o)):r.push(vj(n,s,t))}return r.length===0?new L(1):r.reduce((s,o)=>new P("*",s,o))}function vj(n,e,t){let r=n.shape[e];if(r instanceof Pe)return ll(r,s=>new Oe(qu(s),"index"));if(t){let s=`${n.name}:${e}`,o=t.get(s);if(o)return new Oe(o.name,"index")}return new L(1)}var Y2=M(()=>{"use strict";Fe();po();M0();l(Wd,"flattenIndex");l(Sj,"computeDynamicStride");l(vj,"resolveShapeParam")});function J2(n,e){let t=K2(n,e),r={shapeParamMap:n.shapeParamMap,accCounter:0,metadata:t},s=Bs(n.body,r);return new Nm(n.name,n.params,s,n.bufferMap,n.shapeParams,n.shapeParamMap,t)}function Bs(n,e){if(!n||typeof n!="object")return n;switch(n.type){case"ForNode":return kj(n,e);case"BlockNode":return Nj(n,e);case"SeqNode":return Aj(n,e);case"BufferStoreNode":return Oj(n,e);case"LetStmtNode":return Bj(n,e);case"AllocateNode":return Rj(n,e);case"IfThenElseNode":return Mj(n,e);case"WhileNode":return Ij(n,e);case"EvaluateNode":return new il(Nr(n.value,e));case"SyncThreadsNode":return n;default:return n}}function kj(n,e){let t=n.accumulator!==void 0?n.accumulator:Qb(n);if(t)return Ej(n,t,e);let r=Bs(n.body,e);return new he(n.loopVar,n.min,n.extent,n.kind,r,n.threadTag)}function Nj(n,e){let t=[];for(let o of n.iterVars)o.iterVar&&o.binding&&t.push({name:o.iterVar.name,dtype:o.iterVar.dtype,expr:Nr(o.binding,e)});let r=Bs(n.body,e),s=n.initBody?Bs(n.initBody,e):null;return t.length>0?s?(r=new te([new gl(t,s),new gl(t,r)]),s=null):r=new gl(t,r):s&&(r=new te([s,r])),r}function Aj(n,e){let t=[];for(let r of n.stmts)t.push(Bs(r,e));return new te(t)}function Oj(n,e){let t=Wd(n.buffer,n.indices,e.shapeParamMap),r=Nr(n.value,e),s=n.buffer.dtype||br(n.value);return new Sp(n.buffer,t,r,s)}function Bj(n,e){return new et(n.variable,Nr(n.value,e),Bs(n.body,e))}function Rj(n,e){return new Ue(n.buffer,n.scope,Bs(n.body,e))}function Mj(n,e){return new Se(Nr(n.condition,e),Bs(n.thenBody,e),n.elseBody?Bs(n.elseBody,e):null)}function Ij(n,e){return new $i(n.condVar,Bs(n.condBody,e),Bs(n.loopBody,e))}function Nr(n,e){if(!n||typeof n!="object"||!n.type)return n;switch(n.type){case"BufferLoadNode":{let t=Wd(n.buffer,n.indices,e.shapeParamMap),r=new xp(n.buffer,t,n.buffer.dtype);return si(r),r}case"MathOpNode":{let t=Nr(n.a,e),r=n.b?Nr(n.b,e):null,s={...n,a:t,b:r};return Object.setPrototypeOf(s,Object.getPrototypeOf(n)),si(s),s}case"CompareNode":{let t=Nr(n.a,e),r=Nr(n.b,e),s={...n,a:t,b:r};return Object.setPrototypeOf(s,Object.getPrototypeOf(n)),si(s),s}case"CastNode":{let t=Nr(n.expr,e),r={...n,expr:t};return Object.setPrototypeOf(r,Object.getPrototypeOf(n)),si(r),r}case"CallExternNode":{let t=n.args.map(s=>Nr(s,e)),r={...n,args:t};return Object.setPrototypeOf(r,Object.getPrototypeOf(n)),si(r),r}case"IfThenElseNode":{let t=Nr(n.condition,e),r=Nr(n.thenBody,e),s=n.elseBody?Nr(n.elseBody,e):null,o={...n,condition:t,thenBody:r,elseBody:s};return Object.setPrototypeOf(o,Object.getPrototypeOf(n)),si(o),o}default:return si(n),n}}function Ej(n,e,t){let r=`_acc_${t.accCounter++}`,s=br(e.loadSide);t.metadata.locals.set(r,s);let o=new Map;for(let h of e.block.iterVars)h.iterVar&&h.binding&&o.set(h.iterVar.name,h.binding);let i=o.size>0?zr(e.valueSide,o):e.valueSide,a=Wd(e.store.buffer,e.outerIndices,t.shapeParamMap),u=new xp(e.store.buffer,a,e.store.buffer.dtype);si(u);let p=Nr(i,t),c=Wd(e.store.buffer,e.outerIndices,t.shapeParamMap),f=new Sp(e.store.buffer,c,null,e.store.buffer.dtype),d=e.block.initBody?o.size>0?Kd(e.block.initBody,o):e.block.initBody:null;return new Am({localName:r,dtype:s,op:e.op,initLoad:u,loopVar:n.loopVar,extent:n.extent,loopKind:n.kind,body:p,flushStore:f,initBody:d?Bs(d,t):null})}function zr(n,e){if(!n||typeof n!="object"||!n.type)return n;if(n.type==="VariableNode"&&e.has(n.name))return e.get(n.name);if(n.type==="BufferLoadNode"){let t=n.indices.map(o=>zr(o,e));if(!t.some((o,i)=>o!==n.indices[i]))return n;let s={...n,indices:t};return Object.setPrototypeOf(s,Object.getPrototypeOf(n)),s}if(n.type==="MathOpNode"){let t=zr(n.a,e),r=n.b?zr(n.b,e):null;if(t===n.a&&r===n.b)return n;let s={...n,a:t,b:r};return Object.setPrototypeOf(s,Object.getPrototypeOf(n)),s}if(n.type==="CompareNode"){let t=zr(n.a,e),r=zr(n.b,e);if(t===n.a&&r===n.b)return n;let s={...n,a:t,b:r};return Object.setPrototypeOf(s,Object.getPrototypeOf(n)),s}if(n.type==="CastNode"){let t=zr(n.expr,e);if(t===n.expr)return n;let r={...n,expr:t};return Object.setPrototypeOf(r,Object.getPrototypeOf(n)),r}if(n.type==="CallExternNode"){let t=n.args.map(o=>zr(o,e));if(!t.some((o,i)=>o!==n.args[i]))return n;let s={...n,args:t};return Object.setPrototypeOf(s,Object.getPrototypeOf(n)),s}if(n.type==="IfThenElseNode"){let t=zr(n.condition,e),r=zr(n.thenBody,e),s=n.elseBody?zr(n.elseBody,e):null;if(t===n.condition&&r===n.thenBody&&s===n.elseBody)return n;let o={...n,condition:t,thenBody:r,elseBody:s};return Object.setPrototypeOf(o,Object.getPrototypeOf(n)),o}return n}function Kd(n,e){if(!n||typeof n!="object"||!n.type)return n;switch(n.type){case"BufferStoreNode":{let t=n.indices.map(o=>zr(o,e)),r=zr(n.value,e);if(t.every((o,i)=>o===n.indices[i])&&r===n.value)return n;let s={...n,indices:t,value:r};return Object.setPrototypeOf(s,Object.getPrototypeOf(n)),s}case"SeqNode":{let t=n.stmts.map(s=>Kd(s,e)),r={...n,stmts:t};return Object.setPrototypeOf(r,Object.getPrototypeOf(n)),r}case"ForNode":{let t=Kd(n.body,e);if(t===n.body)return n;let r={...n,body:t};return Object.setPrototypeOf(r,Object.getPrototypeOf(n)),r}case"IfThenElseNode":{let t=zr(n.condition,e),r=Kd(n.thenBody,e),s=n.elseBody?Kd(n.elseBody,e):null,o={...n,condition:t,thenBody:r,elseBody:s};return Object.setPrototypeOf(o,Object.getPrototypeOf(n)),o}default:return n}}var Z2=M(()=>{"use strict";vp();H2();Y2();Fe();Jv();l(J2,"lowerToLIR");l(Bs,"lowerStmt");l(kj,"lowerForNode");l(Nj,"lowerBlockNode");l(Aj,"lowerSeqNode");l(Oj,"lowerBufferStore");l(Bj,"lowerLetStmt");l(Rj,"lowerAllocate");l(Mj,"lowerIfThenElse");l(Ij,"lowerWhile");l(Nr,"lowerExpr");l(Ej,"lowerAccumulator");l(zr,"substituteVars");l(Kd,"substituteVarsStmt")});function Q2(n){let e=[],t={errors:e,boundVars:new Set,path:[],bufferNames:new Set};if(!n||n.type!=="LIRFunc")return e.push(new Go("root must be LIRFunc")),e;for(let[,r]of n.bufferMap)t.bufferNames.add(r.name);for(let r of n.shapeParams)t.boundVars.add(r.name);if(n.metadata)for(let[r]of n.metadata.locals)t.boundVars.add(r);return Rs(n.body,t),e}function Rs(n,e){if(!(!n||typeof n!="object")){switch(e.path.push(n.type),n.type){case"ForNode":Lj(n,e);break;case"SeqNode":for(let t of n.stmts)Rs(t,e);break;case"LIRFlatStoreNode":Cj(n,e);break;case"LIRAccumulatorNode":Dj(n,e);break;case"LIRBindingsNode":Pj(n,e);break;case"LetStmtNode":{Pn(n.value,e);let t=e.boundVars.has(n.variable.name);e.boundVars.add(n.variable.name),Rs(n.body,e),t||e.boundVars.delete(n.variable.name);break}case"AllocateNode":{let t=n.buffer?e.bufferNames.has(n.buffer.name):!0;n.buffer&&e.bufferNames.add(n.buffer.name),Rs(n.body,e),n.buffer&&!t&&e.bufferNames.delete(n.buffer.name);break}case"IfThenElseNode":Pn(n.condition,e),Rs(n.thenBody,e),n.elseBody&&Rs(n.elseBody,e);break;case"WhileNode":Rs(n.condBody,e),Rs(n.loopBody,e);break;case"EvaluateNode":Pn(n.value,e);break;default:break}e.path.pop()}}function Lj(n,e){n.extent?Pn(n.extent,e):e.errors.push(new Go("ForNode missing extent",[...e.path]));let t=!0;n.loopVar?(t=e.boundVars.has(n.loopVar.name),e.boundVars.add(n.loopVar.name)):e.errors.push(new Go("ForNode missing loopVar",[...e.path])),Rs(n.body,e),n.loopVar&&!t&&e.boundVars.delete(n.loopVar.name)}function Cj(n,e){n.buffer||e.errors.push(new Go("LIRFlatStoreNode missing buffer",[...e.path])),Pn(n.offsetExpr,e),Pn(n.value,e)}function Dj(n,e){n.localName||e.errors.push(new Go("LIRAccumulatorNode missing localName",[...e.path])),n.dtype||e.errors.push(new Go("LIRAccumulatorNode missing dtype",[...e.path]));let t=n.localName?e.boundVars.has(n.localName):!0;n.localName&&e.boundVars.add(n.localName),Pn(n.initLoad,e),Pn(n.body,e),Rs(n.flushStore,e),n.initBody&&Rs(n.initBody,e),n.localName&&!t&&e.boundVars.delete(n.localName)}function Pj(n,e){let t=[];for(let r of n.bindings)Pn(r.expr,e),e.boundVars.has(r.name)||t.push(r.name),e.boundVars.add(r.name);Rs(n.body,e);for(let r of t)e.boundVars.delete(r)}function Pn(n,e){if(!(!n||typeof n!="object"||!n.type))switch(n.type){case"LIRFlatLoadNode":n.buffer||e.errors.push(new Go("LIRFlatLoadNode missing buffer",[...e.path])),Pn(n.offsetExpr,e);break;case"MathOpNode":Pn(n.a,e),n.b&&Pn(n.b,e);break;case"CompareNode":Pn(n.a,e),Pn(n.b,e);break;case"CastNode":Pn(n.expr,e);break;case"CallExternNode":if(n.args)for(let t of n.args)Pn(t,e);break;case"IfThenElseNode":Pn(n.condition,e),Pn(n.thenBody,e),n.elseBody&&Pn(n.elseBody,e);break;case"VariableNode":n.name!==void 0&&!e.boundVars.has(n.name)&&e.errors.push(new Go(`unbound variable '${n.name}'`,[...e.path]));break;case"IntImmNode":case"FloatImmNode":break;default:break}}var Go,X2=M(()=>{"use strict";Go=class{static{l(this,"LIRVerificationError")}constructor(e,t){this.message=e,this.nodePath=t||[]}toString(){let e=this.nodePath.length>0?` at ${this.nodePath.join(" > ")}`:"";return`LIR verification: ${this.message}${e}`}};l(Q2,"verifyLIR");l(Rs,"verifyStmt");l(Lj,"verifyForNode");l(Cj,"verifyFlatStore");l(Dj,"verifyAccumulator");l(Pj,"verifyBindings");l(Pn,"verifyExpr")});function eC(n,e,t={}){return new Si({target:e,...t}).compileFunction(n)}var iw,ak,Si,aw=M(()=>{"use strict";Di();Us();OE();by();BE();hL();L2();$a();Ep();dy();C2();$2();tk();z2();G2();W2();Cv();nn();RS();nv();Z2();X2();nn();iw=class{static{l(this,"CompilerConfig")}constructor(e={}){this.target=e.target,this.verify=e.verify!==!1,this.verifyMode=e.verify==="full"?"full":"normal",this.errorMode=e.errorMode||"strict";let t=this.target&&typeof this.target.isWebGPU=="function"&&this.target.isWebGPU(),s=this.target&&typeof this.target.isGPU=="function"&&this.target.isGPU()&&!t;this.fusion={enabled:!0,strategy:"priority",epilogue:void 0,...e.fusion},this.scheduling={enabled:t,autotune:!1,gpuTiling:s,...e.scheduling},this.matmulBackend=e.matmulBackend||"native",this.quantization={enabled:!1,...e.quantization},this.optimization={layout:!1,rematerialization:!1,rematConfig:{},fastMath:!1,maxSimplifyIterations:8,loopPartition:!1,detectAccumulators:!1,tensorize:!1,...e.optimization},this.memory={alignment:64,inplaceReuse:!0,allocStrategy:"best-fit",poolAllocation:!1,...e.memory},this.partition={enabled:!1,targets:[],defaultTarget:null,opTargetOverrides:new Map,memoryLimits:new Map,minPartitionSize:1,costWeights:{},...e.partition},this.passContext=e.passContext||null,this.loweringRules=e.loweringRules||null,this.codegenEntries=e.codegenEntries||null;let o=e.trace||{};this.trace={level:o.level??le.SILENT,sink:o.sink??null,irSnapshot:{afterGraphPasses:!1,afterLowering:!1,afterScheduling:!1,...o.irSnapshot||{}}}}get usePartition(){return this.partition.enabled&&this.partition.targets.length>=2}},ak=class{static{l(this,"CompilationResult")}constructor(e,t,r){this.module=e,this.trace=t,this.errors=r||[]}get succeeded(){return this.errors.length===0}get failedFunctions(){let e=new Set;for(let t of this.errors)t.funcName&&e.add(t.funcName);return e}run(e,...t){return this.module.run(e,...t)}async runAsync(e,...t){return this.module.runAsync(e,...t)}isAsync(e){return this.module.isAsync(e)}getSource(e){return this.module.getKernelSource(e)}getSnippet(e){return this.module.getKernelSnippet(e)}listKernels(){return this.module.listKernels()}},Si=class n{static{l(this,"Compiler")}constructor(e){if(this.config=e instanceof iw?e:new iw(e),!this.config.target)throw new Error("Compiler requires a target");this.context=new V_({loweringRules:this.config.loweringRules,codegenEntries:this.config.codegenEntries})}compile(e){let t=new _y(this.config.trace),r=this.config.errorMode==="resilient",s=[],o=new Set,i=performance.now();t.phaseStart("compile");let a={compiler:this,trace:t,errors:s,failed:o,resilient:r,original:e,working:r?wA(e):e,cudaMatmulChain:!1,split:null,primFuncs:null,lirFuncs:null,runtimeModule:null};for(let u of this._compilePhases())u.when&&!u.when(a)||u.run(a);if(t.phaseEnd("compile",performance.now()-i),!r&&s.length>0)throw new Error(s[0].toString());return new ak(a.runtimeModule,t,s)}_compilePhases(){return[{name:"verify:pre",when:l(e=>e.compiler.config.verify,"when"),run:l(e=>e.compiler._verifyGraph(e.working,"before graph passes",e.trace,e.errors,e.failed,e.resilient),"run")},{name:"calibrate",when:l(e=>{let t=e.compiler.config.quantization;return t.enabled&&t.calibrationData&&!t.calibration},"when"),run:l(e=>e.compiler._runCalibration(e.working,e.trace),"run")},{name:"graphPasses",run:l(e=>{e.cudaMatmulChain=e.compiler._runGraphPasses(e.working,e.original,e.trace,e.errors,e.failed,e.resilient)},"run")},{name:"partition",when:l(e=>e.compiler.config.usePartition,"when"),run:l(e=>e.compiler._runPartitioning(e.working,e.trace),"run")},{name:"split",run:l(e=>{let t=e.compiler.config,r=typeof t.target.isWebGPU=="function"&&t.target.isWebGPU(),s=typeof t.target.isGPU=="function"&&t.target.isGPU()&&!r,o=0,i=0;for(let p of e.working)for(let c of p.ops())c.opName==="conv"||c.opName==="quantized_conv"?o++:c.opName==="scaled_dot_product_attention"&&i++;let a=s&&o>=2,u=s&&i>0;e.split=q2(e.working,{config:t,target:t.target,cudaMatmulChain:e.cudaMatmulChain,cudaConvChain:a,cudaAttention:u,isWebGPU:r})},"run")},{name:"verify:post",when:l(e=>e.compiler.config.verify,"when"),run:l(e=>e.compiler._verifyGraph(e.working,"after graph passes",e.trace,e.errors,e.failed,e.resilient),"run")},{name:"lowering",run:l(e=>{if(e.primFuncs=e.compiler._lowerAll(e.working,e.trace,e.errors,e.failed,e.resilient),e.compiler.config.matmulBackend==="cublas")for(let t of e.primFuncs)t.cublasInfo=e.split&&e.split.cublasInfos?e.split.cublasInfos.get(t.name)||null:P1(t)},"run")},{name:"tirPasses",run:l(e=>e.compiler._runTirPasses(e),"run")},{name:"verify:tensor",when:l(e=>e.compiler.config.verify,"when"),run:l(e=>e.compiler._verifyAll(e.primFuncs,e.errors,e.failed,e.resilient),"run")},{name:"lirLowering",run:l(e=>{e.lirFuncs=e.compiler._lowerToLIR(e.primFuncs,e.trace,e.errors,e.failed,e.resilient)},"run")},{name:"codegen",run:l(e=>{e.runtimeModule=e.compiler._codegen(e.lirFuncs,e.trace,e.errors,e.failed,e.resilient),e.split&&(e.runtimeModule.executionPlan=e.split.plan)},"run")}]}compileFunction(e){let t=new Mn("single");return t.addFunction(e),this.compile(t)}calibrate(e,t="minmax"){let r=new Dc(t);for(let s of e)r.attach(s);return r}_runCalibration(e,t){let r=this.config.quantization,s=e.functionNames()[0],o=e.getFunction(s);if(!o)return;t.phaseStart("calibrate");let i=performance.now(),a=this.config.target,u=l((c,f)=>new n({target:f,verify:this.config.verify}).compile(c),"compileFn"),p=F2(o,a,r.calibrationData,{mode:r.calibrationMode||"minmax",quantizableOps:r.quantizableOps,compileFn:u});this.config.quantization={...r,calibration:p},t.phaseEnd("calibrate",performance.now()-i)}_runGraphPasses(e,t,r,s,o,i){let a=new iu,u=0;for(let m of e)for(let g of m.ops())g.opName==="dot"&&u++;let p=this.config.target,c=(p.getAttr&&p.getAttr("matmulChainThreshold"))??(p.kind==="cuda"?2:1/0),f=u>=c;for(let m of dL(this.config,this.config.target,{cudaMatmulChain:f,context:this.context}))a.addPass(m);a.setTrace(r),this.config.verifyMode==="full"&&a.setVerifyHook((m,g)=>(g?Qv(m):jd(m)).map(_=>_.toString())),r.phaseStart("graphPasses");let d=performance.now(),h=a.run(e,{errorMode:i?"resilient":"strict",passContext:this.config.passContext});if(h.errors){for(let m of h.errors)s.push(m),r.errorEvent(m.phase,m.funcName,m.message,m.passName);if(h.failedFunctions){for(let m of h.failedFunctions)if(o.add(m),i&&t&&t!==e){let g=t.getFunction(m);g&&e.addFunction(lm(g))}}}if(r.phaseEnd("graphPasses",performance.now()-d),r.shouldSnapshot("afterGraphPasses")){let m=new Ef;r.irDump("afterGraphPasses",m.printModule(e))}return f}_runPartitioning(e,t){let r=new iu;r.addPass(new sw(this.config.partition)),r.addPass(new ow({targets:this.config.partition.targets})),r.setTrace(t),t.phaseStart("partition");let s=performance.now();r.run(e),t.phaseEnd("partition",performance.now()-s)}_eachFunc(e,t,r,s,o,i,a){for(let u of e)if(!o.has(u.name))try{a(u)}catch(p){if(s.push(new Yn(t,u.name,p.message)),o.add(u.name),r.errorEvent(t,u.name,p.message),!i)break}}_lowerAll(e,t,r,s,o){t.phaseStart("lowering");let i=performance.now(),a=[];return this._eachFunc(e,"lowering",t,r,s,o,u=>{let p=performance.now(),c=Tm(u,this.config.target,this.context);this.config.target.isGPU&&this.config.target.isGPU()&&!(this.config.target.isWebGPU&&this.config.target.isWebGPU())&&(c.convInfo=n2(u)),t.functionEvent("lowering",u.name,{durationMs:performance.now()-p}),a.push(c),t.shouldSnapshot("afterLowering")&&t.irDump("afterLowering:"+u.name,G_(c))}),t.phaseEnd("lowering",performance.now()-i),a}_runTirPasses(e){let t=new j_;for(let r of E2(this.config))t.addPass(r);if(t.setTrace(e.trace),this.config.verifyMode==="full"){let r=new Gd;t.setVerifyHook(s=>r.verify(s))}t.run(e.primFuncs,{trace:e.trace,errors:e.errors,failed:e.failed,resilient:e.resilient})}_verifyGraph(e,t,r,s,o,i){if(i){for(let u of e){if(o.has(u.name))continue;let p=jd?jd(u):[];if(p.length>0){let c=p.map(f=>f.toString()).join("; ");s.push(new Yn("verification",u.name,c)),o.add(u.name),r.errorEvent("verification",u.name,c)}}return}let a=Qv(e);if(a.length>0)throw new Error("Graph verification failed ("+t+"): "+a.map(u=>u.toString()).join("; "))}_verifyAll(e,t,r,s){let o=new Gd;for(let i of e){if(r.has(i.name))continue;let a=o.verify(i);if(a.length>0){let u=a.join("; ");if(s)t.push(new Yn("verification",i.name,u)),r.add(i.name);else throw new Error("TensorIR verification failed for "+i.name+": "+u)}}}_lowerToLIR(e,t,r,s,o){t.phaseStart("lirLowering");let i=performance.now(),a=[];return this._eachFunc(e,"lirLowering",t,r,s,o,u=>{let p=performance.now(),c=J2(u,this.config.target);if(u.cublasInfo&&(c.cublasInfo=u.cublasInfo),u.gpuRegisterBlocked&&(c.gpuRegisterBlocked=!0),this.config.verifyMode==="full"){let f=Q2(c);if(f.length>0)throw new Error("LIR verification failed: "+f.map(d=>d.toString()).join("; "))}t.functionEvent("lirLowering",u.name,{durationMs:performance.now()-p}),a.push(c)}),t.phaseEnd("lirLowering",performance.now()-i),a}_codegen(e,t,r,s,o){t.phaseStart("codegen");let i=performance.now(),a=new jl("compiled"),u=this.config.usePartition,p={matmulBackend:this.config.matmulBackend,context:this.context},c=new Map,f=l(h=>(c.has(h.name)||c.set(h.name,new bo(h,p)),c.get(h.name)),"getBackend"),d=u?null:new bo(this.config.target,p);return this._eachFunc(e,"codegen",t,r,s,o,h=>{let m=performance.now(),g;if(u){let _=h._partitionTarget,b=_?this.config.partition.targets.find(w=>w.name===_):this.config.target;g=f(b||this.config.target)}else g=d;let y=g.compile(h);a.addCompiledKernel(y),h.shapeParamMap&&h.shapeParamMap.size>0&&a.setShapeParamMap(h.name,h.shapeParamMap,h.bufferMap),t.codegenStats(h.name,{durationMs:performance.now()-m,sourceSize:y.source.length,targetName:y.target.name})}),t.phaseEnd("codegen",performance.now()-i),a}};l(eC,"compileGraph")});function Fj(n){return typeof n=="object"&&n!==null}function zj(n){return Fj(n)&&"_impl"in n}function Vj(n){return(e,...t)=>{let r=As();if(!r)throw new Error(`TRACING dispatch key active but no tracer is set for op '${n}'`);if($j.has(n)){let c=Lt.findOp(n);return Lt.redispatch(c,e,...t)}let s=[],o={},i=Ly(n),a=0,u=l(c=>c instanceof bt||zj(c),"isTensor"),p=l(c=>{c instanceof bt||c.isSymbolic?s.push(c):s.push(r.captureConstant(c))},"pushTensor");for(let c of t)if(Array.isArray(c)&&c.length>0&&u(c[0]))for(let f of c)p(f);else u(c)?p(c):i&&(c!=null&&a<i.length&&(o[i[a]]=c),a++);return r.recordOp(n,s,o)}}function nC(){let n=Lt.listOps();for(let e of n){let t=Lt.findOp(e);if(!t)continue;let r=t.name,s=Mr.fromUnboxed(Vj(r));t.entry.registerKernel(48,s)}}var $j,rC=M(()=>{"use strict";An();sp();Ma();_u();yu();Cy();$j=new Set(["scatter","repeat","tile","split","chunk","roll","flip","cumsum","sort","argsort","topk"]);l(Fj,"_isObject");l(zj,"_hasImpl");l(Vj,"_tracingKernel");l(nC,"registerTracingDispatch")});function jj(n){return n&&n.shape&&n.shape.length>=2&&Gj.has(n.dtype)}function sC(n,e,t=jj){let r=n.graph.functions().next().value;if(!r)return n;let s=r.entryBlock,o=n.numUserInputs,i=n.capturedParams,a=[];for(let f=0;f<i.length;f++){let d=o+f,h=s.getArgument(d);h&&t(i[f],f,h)&&a.push({j:f,argIndex:d,arg:h,param:i[f]})}if(a.length===0)return n;for(let f of a){let d=e(f.param),h=f.arg.type,m=Me,g=new m("constant",[],[h],{value:d,tensor_type:h}),y=s.firstOp;y?s.insertBefore(g,y):s.pushOp(g),f.arg.replaceAllUsesWith(g.getResult(0))}let u=new Set(a.map(f=>f.argIndex));s.removeArguments(u),r.inputTypes=Object.freeze(r.inputTypes.filter((f,d)=>!u.has(d)));let p=new Set(a.map(f=>f.j)),c=i.filter((f,d)=>!p.has(d));return{...n,capturedParams:c}}var Gj,oC=M(()=>{"use strict";yr();Gj=new Set(["f16","f32","f64"]);l(jj,"defaultWeightPredicate");l(sC,"foldWeightParams")});function Hd(n,e,t){let r=t.get(e.id);if(r)return r;let s=n.scalarConstant(0,e.type.dtype).getResult(0);return n.broadcast(s,e.type.shape,[]).getResult(0)}var jo,Yd=M(()=>{"use strict";l(Hd,"gradOrZero");jo=class{static{l(this,"GradAccumulator")}constructor(e){this._builder=e,this._pending=new Map,this._reduced=new Map}accumulate(e,t){if(!t)return;let r=this._pending.get(e);r||(r=[],this._pending.set(e,r)),r.push(t),this._reduced.delete(e)}_treeReduce(e){let t=e;for(;t.length>1;){let r=[];for(let s=0;s<t.length;s+=2)s+1<t.length?r.push(this._builder.add(t[s],t[s+1]).getResult(0)):r.push(t[s]);t=r}return t[0]}get(e){if(this._reduced.has(e))return this._reduced.get(e);let t=this._pending.get(e);if(!t||t.length===0)return null;let r=this._treeReduce(t);return this._reduced.set(e,r),r}has(e){let t=this._pending.get(e);return!!t&&t.length>0}}});function pe(n,e){uk.set(n,e)}function Eu(n){return uk.get(n)||null}function pw(n){let e=uk.get(n);if(e)return e;if(lk.has(n))return null;throw new Error(`autodiff: op '${n}' is on the gradient path but has no VJP rule and is not a registered gradient barrier. Register one with registerVJPRule('${n}', ...) or registerGradientBarrier('${n}').`)}function ck(n,e){iC.set(n,e)}function fw(n){return iC.get(n)||null}function pk(n){lk.add(n)}function dw(n){return lk.has(n)}var uk,lk,iC,Is=M(()=>{"use strict";uk=new Map,lk=new Set;l(pe,"registerVJPRule");l(Eu,"getVJPRule");l(pw,"requireVJPRuleOrBarrier");iC=new Map;l(ck,"registerRegionVJP");l(fw,"getRegionVJP");l(pk,"registerGradientBarrier");l(dw,"isGradientBarrier")});var vi,fk=M(()=>{"use strict";vi=new Set(["scan","if"])});function Jd(n){let e=new Set(n.arguments.map(i=>i.id)),t=l(i=>{for(let a of i.ops()){for(let u of a.results)e.add(u.id);for(let u of a.regions||[])for(let p of u.blocks){for(let c of p.arguments)e.add(c.id);t(p)}}},"addLocals");t(n);let r=new Set,s=[],o=l(i=>{for(let a of i.ops()){for(let u of a.operands)e.has(u.id)||r.has(u.id)||u.definingOp&&u.definingOp.opName==="constant"||(r.add(u.id),s.push(u));for(let u of a.regions||[])for(let p of u.blocks)o(p)}},"scan");return o(n),s}function hw(n,e){let t=n.scalarConstant(0,e.type.dtype).getResult(0);return n.broadcast(t,e.type.shape,[]).getResult(0)}function aC(n,e,t){let r=e.type.shape,s=r.map((a,u)=>u===0?t:0),o=r.map((a,u)=>u===0?t+1:a),i=n.slice(e,s,o).getResult(0);return n.reshape(i,r.slice(1)).getResult(0)}function qj(n,e,t){let r=t.slice(1),s=e.map(o=>n.reshape(o,[1,...r]).getResult(0));return s.length===1?n.reshape(s[0],t).getResult(0):n.concat(s,0).getResult(0)}function mw(n,e,t,r,s,o,i=new Map,a=[]){let u=new Map,p=new Map;for(let _=0;_<e.arguments.length;_++)u.set(e.arguments[_].id,t[_]),p.set(e.arguments[_],t[_]);for(let[_,b]of r)u.set(_,b);for(let _ of a)r.has(_.id)&&p.set(_,r.get(_.id));let c=[],f=null;for(let _ of e.ops())_.opName==="yield"?f=_:c.push(_);let d=l(_=>{if(u.has(_.id))return u.get(_.id);let b=_.definingOp;if(b&&b.opName==="constant"){let w=i.get(_.id);return w===void 0&&(w=n._buildOp("constant",[],[_.type],new Map(b.attributes),null).getResult(0),i.set(_.id,w)),u.set(_.id,w),p.set(_,w),w}return _},"matOperand");for(let _ of c){if(vi.has(_.opName)){for(let S of _.operands)d(S),p.set(S,u.get(S.id)??S);let x=_.clone(p);n.block.pushOp(x);for(let S=0;S<_.numResults;S++)u.set(_.getResult(S).id,x.getResult(S)),p.set(_.getResult(S),x.getResult(S));continue}let b=_.operands.map(d),w=n._buildOp(_.opName,b,_.results.map(x=>x.type),new Map(_.attributes),null);for(let x=0;x<_.numResults;x++)u.set(_.getResult(x).id,w.getResult(x)),p.set(_.getResult(x),w.getResult(x))}let h=f.operands.map(_=>u.get(_.id));if(o)return{forwardYields:h};let m=new jo(n);for(let _=0;_<f.operands.length;_++)s[_]&&m.accumulate(f.operands[_].id,s[_]);for(let _=c.length-1;_>=0;_--){let b=c[_];if(b.opName==="constant")continue;let w=b.results.map(v=>m.get(v.id));if(w.every(v=>v===null))continue;if(vi.has(b.opName)){let v=fw(b.opName);v&&v(b,{accumulator:m,builder:n,materialize:d,needsGrad:Uj,scanCheckpoint:null});continue}let x=pw(b.opName);if(!x)continue;let S={builder:n,op:b,operands:b.operands.map(d),results:b.results.map(v=>u.get(v.id)),gradOutputs:w,attrs:b.attributes,full:l((v,A)=>n.broadcast(n.scalarConstant(v,A.dtype).getResult(0),A.shape,[]).getResult(0),"full")},T=x(S);if(T)for(let v=0;v<b.numOperands;v++)T[v]&&m.accumulate(b.getOperand(v).id,dk(n,T[v],b.getOperand(v).type.shape))}let g=e.arguments.map(_=>m.get(_.id)),y=new Map;for(let _ of r.keys())y.set(_,m.get(_));return{forwardYields:h,gradArgs:g,gradFree:y}}function Wj(n,e,t,r,s){let o=n.regions[0].blocks[0],i=n.regions[1].blocks[0],a=r(n.getOperand(0)),u=[];for(let y=0;y<n.numResults;y++)u.push(e.get(n.getResult(y).id));let p=Jd(o),c=Jd(i),f=new Map(p.map(y=>[y.id,r(y)])),d=new Map(c.map(y=>[y.id,r(y)])),{gradFree:h}=mw(t,o,[],f,u,!1,new Map,p),{gradFree:m}=mw(t,i,[],d,u,!1,new Map,c),g=new Map;for(let y of p)g.set(y.id,y);for(let y of c)g.set(y.id,y);for(let[y,_]of g){if(!s.has(y))continue;let b=h.get(y),w=m.get(y);if(!b&&!w)continue;let x=hw(t,_),S=t.broadcast(a,_.type.shape,[]).getResult(0);e.accumulate(y,t.select(S,b??x,w??x).getResult(0))}}function Kj(n,e){if(!n||e<=1)return null;if(n==="sqrt"||n===!0)return Math.max(1,Math.ceil(Math.sqrt(e)));if(typeof n=="number"&&n>=1){let t=Math.floor(n);return t>=e?null:t}return null}function Hj(n,e,t,r,s,o=null){let i=n.regions[0].blocks[0],a=n.getAttr("num_carry"),u=n.getAttr("num_xs"),p=n.numResults-a,c=[];for(let k=0;k<u;k++)c.push(n.getOperand(k));let f=[];for(let k=0;k<a;k++)f.push(n.getOperand(u+k));let d=c[0].type.shape[0],h=Jd(i),m=c.map(r),g=f.map(r),y=new Map(h.map(k=>[k.id,r(k)])),_=new Map,b=l(k=>m.map(O=>aC(t,O,k)),"sliceX"),w=l((k,O)=>mw(t,i,[...k,...O],y,null,!0,_,h).forwardYields.slice(0,a),"stepForward"),x=[];for(let k=0;k<p;k++)x.push(e.get(n.getResult(a+k).id));let S=[];for(let k=0;k<a;k++){let O=e.get(n.getResult(k).id);S.push(O??hw(t,g[k]))}let T=new Map,v=m.map(()=>new Array(d)),A=l((k,O,R)=>{let B=[...O,...R],E=x.map(G=>G===null?null:aC(t,G,k)),F=[...S,...E],{gradArgs:C,gradFree:H}=mw(t,i,B,y,F,!1,_,h);for(let G=0;G<u;G++)v[G][k]=C[G]??hw(t,O[G]);S=[];for(let G=0;G<a;G++)S.push(C[u+G]??hw(t,R[G]));for(let[G,U]of H){if(!U)continue;let D=T.get(G);T.set(G,D?t.add(D,U).getResult(0):U)}},"backwardStep"),N=Kj(o,d);if(N){let k=Math.ceil(d/N),O=new Array(k),R=g;O[0]=R;for(let B=0;B<d;B++){R=w(b(B),R);let E=(B+1)/N;Number.isInteger(E)&&E<k&&(O[E]=R)}for(let B=k-1;B>=0;B--){let E=B*N,F=Math.min(E+N,d),C=[],H=[O[B]],G=O[B];for(let U=E;U<F;U++){let D=b(U);C.push(D),G=w(D,G),H.push(G)}for(let U=F-1;U>=E;U--)A(U,C[U-E],H[U-E])}}else{let k=g,O=[k],R=[];for(let B=0;B<d;B++){let E=b(B);R.push(E),k=w(E,k),O.push(k)}for(let B=d-1;B>=0;B--)A(B,R[B],O[B])}for(let k=0;k<u;k++)s.has(c[k].id)&&e.accumulate(c[k].id,qj(t,v[k],m[k].type.shape));for(let k=0;k<a;k++)s.has(f[k].id)&&e.accumulate(f[k].id,S[k]);for(let[k,O]of T)O&&s.has(k)&&e.accumulate(k,O)}var Uj,uC=M(()=>{"use strict";Yd();Is();Zd();fk();Uj={has:l(()=>!0,"has")};ck("scan",(n,e)=>Hj(n,e.accumulator,e.builder,e.materialize,e.needsGrad,e.scanCheckpoint));ck("if",(n,e)=>Wj(n,e.accumulator,e.builder,e.materialize,e.needsGrad));l(Jd,"regionFreeVars");l(hw,"zeroLike");l(aC,"sliceStep");l(qj,"stackSteps");l(mw,"diffBodyStep");l(Wj,"buildCondBackward");l(Kj,"resolveSegmentLength");l(Hj,"buildScanBackward")});function Qd(n,{accumulator:e,builder:t,needsGrad:r,resolveValue:s,handleRegionOp:o=null}){for(let i=n.length-1;i>=0;i--){let a=n[i];if(a.opName==="return"||a.opName==="constant"||!a.results.some(g=>r.has(g.id)))continue;let p=[];for(let g=0;g<a.numResults;g++)p.push(e.get(a.getResult(g).id));if(p.every(g=>g===null)||o&&o(a))continue;let c=pw(a.opName);if(!c)continue;let f=new Array(a.numOperands);for(let g=0;g<a.numOperands;g++)f[g]=s(a.getOperand(g));let d=new Array(a.numResults);for(let g=0;g<a.numResults;g++)d[g]=s(a.getResult(g));let h=l((g,y)=>t.broadcast(t.scalarConstant(g,y.dtype).getResult(0),y.shape,[]).getResult(0),"full"),m=c({builder:t,op:a,operands:f,results:d,gradOutputs:p,attrs:a.attributes,full:h});if(m)for(let g=0;g<a.numOperands;g++){if(g>=m.length||!m[g])continue;let y=a.getOperand(g);r.has(y.id)&&e.accumulate(y.id,dk(t,m[g],y.type.shape))}}}function lC(n){let e=[];for(let t of n.regions)t.blocks[0]&&e.push(...Jd(t.blocks[0]));return e}function dk(n,e,t){let r=e.type.shape;if(r.length===t.length&&r.every((u,p)=>u===t[p]))return e;let s=r.length-t.length,o=[];for(let u=0;u<s;u++)o.push(u);for(let u=0;u<t.length;u++)t[u]===1&&r[s+u]!==1&&o.push(s+u);let i=e;if(o.length>0){let u=n.scalarConstant(0,e.type.dtype).getResult(0);i=n.reduce(e,u,o,"sum").getResult(0)}let a=i.type.shape;return a.length===t.length&&a.every((u,p)=>u===t[p])||(i=n.reshape(i,t).getResult(0)),i}var Yj,Xd,Zd=M(()=>{"use strict";Us();_s();ro();Yd();Is();uC();fk();Yj=new Set(["neg","abs","sign","floor","ceil"]);l(Qd,"backpropOps");l(lC,"regionControlFlowFreeVars");l(dk,"reduceGradToOperandShape");Xd=class{static{l(this,"BackwardGraphBuilder")}constructor(e={}){this._rematPolicy=e.rematPolicy||null,this._checkpointPolicy=e.checkpointPolicy||null,this._scanCheckpoint=e.scanCheckpoint||null}build(e){if(this._checkpointPolicy)return this._buildCheckpointed(e);let r=ft.compute(e).topologicalOrder,s=e.getReturnOp();if(!s)throw new Error("Forward function has no return op");let o=s.operands,i=e.args,a=this._computeGradReachability(e,r),{savedValues:u,savedValueIndices:p}=this._identifySavedValues(r,a,i),c=o.map(T=>T.type),f=u.map(T=>T.type),d=[...c,...f],h=[];for(let T=0;T<i.length;T++)a.has(i[T].id)&&h.push(i[T].type);let m=new yn(`backward_${e.name}`,d,h),g=new $n(m),y=m.args,_=y.slice(0,c.length),b=y.slice(c.length),w=new Map;for(let T=0;T<u.length;T++)w.set(u[T].id,b[T]);for(let T=0;T<i.length;T++)p.has(i[T].id)&&w.set(i[T].id,b[p.get(i[T].id)]);let x=new jo(g);for(let T=0;T<o.length;T++){let v=o[T];x.accumulate(v.id,_[T])}Qd(r,{accumulator:x,builder:g,needsGrad:a,resolveValue:l(T=>this._materialize(T,w,g),"resolveValue"),handleRegionOp:l(T=>{let v=fw(T.opName);return v?(v(T,{accumulator:x,builder:g,materialize:l(A=>this._materialize(A,w,g),"materialize"),needsGrad:a,scanCheckpoint:this._scanCheckpoint}),!0):!1},"handleRegionOp")});let S=[];for(let T=0;T<i.length;T++)a.has(i[T].id)&&S.push(Hd(g,i[T],x));return g.returnOp(S),{backwardFunc:m,savedValues:u,gradInputIndices:this._getGradInputIndices(i,a)}}_materialize(e,t,r){if(t.has(e.id))return t.get(e.id);if(!e.definingOp)return e;let s=new Set([e.id]),o=[{val:e,i:0}];for(;o.length>0;){let i=o[o.length-1],a=i.val,u=a.definingOp;if(t.has(a.id)||!u){s.delete(a.id),o.pop();continue}if(i.i<u.numOperands){let d=u.getOperand(i.i);i.i++,d.definingOp&&!t.has(d.id)&&!s.has(d.id)&&(s.add(d.id),o.push({val:d,i:0}));continue}let p=new Array(u.numOperands);for(let d=0;d<u.numOperands;d++){let h=u.getOperand(d);p[d]=t.has(h.id)?t.get(h.id):h}let c=u.results.map(d=>d.type),f=r._buildOp(u.opName,p,c,new Map(u.attributes),null);for(let d=0;d<u.numResults;d++)t.set(u.getResult(d).id,f.getResult(d));s.delete(a.id),o.pop()}return t.has(e.id)?t.get(e.id):e}_computeGradReachability(e,t){let r=new Set,s=e.getReturnOp();for(let o of s.operands)r.add(o.id);for(let o=t.length-1;o>=0;o--){let i=t[o];if(!(i.opName==="return"||!i.results.some(u=>r.has(u.id)))){if(vi.has(i.opName)){for(let u=0;u<i.numOperands;u++)r.add(i.getOperand(u).id);for(let u of lC(i))r.add(u.id);continue}if(Eu(i.opName)&&!dw(i.opName))for(let u=0;u<i.numOperands;u++)r.add(i.getOperand(u).id)}}return r}_identifySavedValues(e,t,r){let s=[],o=new Map,i=new Set(r.map(f=>f.id));for(let f of e)if(!(f.opName==="return"||f.opName==="constant"||!Eu(f.opName)||!f.results.some(m=>t.has(m.id)))){for(let m=0;m<f.numOperands;m++){let g=f.getOperand(m);i.has(g.id)&&!o.has(g.id)&&(o.set(g.id,s.length),s.push(g))}for(let m=0;m<f.numResults;m++){let g=f.getResult(m);t.has(g.id)&&!o.has(g.id)&&this._shouldSaveResult(f)&&(o.set(g.id,s.length),s.push(g))}}let a=new Set(o.keys()),u=new Map(r.map(f=>[f.id,f])),p=new Set,c=l(f=>{if(a.has(f.id)||p.has(f.id))return;p.add(f.id);let d=f.definingOp;if(!d){i.has(f.id)&&!o.has(f.id)&&(o.set(f.id,s.length),s.push(u.get(f.id)));return}for(let h=0;h<d.numOperands;h++)c(d.getOperand(h))},"collect");for(let f of e)if(!(f.opName==="return"||f.opName==="constant")&&f.results.some(d=>t.has(d.id))){if(vi.has(f.opName)){for(let d of f.operands)c(d);for(let d of lC(f))c(d);continue}if(Eu(f.opName)){for(let d=0;d<f.numOperands;d++)c(f.getOperand(d));for(let d=0;d<f.numResults;d++)c(f.getResult(d))}}return{savedValues:s,savedValueIndices:o}}_shouldSaveResult(e){return this._rematPolicy?!this._rematPolicy.shouldRematerialize(e):!Yj.has(e.opName)}_getGradInputIndices(e,t){let r=[];for(let s=0;s<e.length;s++)t.has(e[s].id)&&r.push(s);return r}_buildCheckpointed(e){let r=ft.compute(e).topologicalOrder,s=e.getReturnOp();if(!s)throw new Error("Forward function has no return op");let o=s.operands,i=e.args,a=this._computeGradReachability(e,r),u=this._checkpointPolicy.segment(r,e);for(let O of u)for(let R of O.ops)if(vi.has(R.opName))throw new Error(`Checkpointed backward does not support region control-flow op '${R.opName}'; build the backward without a checkpointPolicy, which differentiates scan/if via buildScanBackward/buildCondBackward.`);let p=new Set,c=[],f=new Map,d=new Map;for(let O of r)for(let R=0;R<O.numResults;R++){let B=O.getResult(R);d.set(B.id,B)}let h=new Set(i.map(O=>O.id));for(let O of i)a.has(O.id)&&!p.has(O.id)&&(p.add(O.id),f.set(O.id,c.length),c.push(O));for(let O of u){for(let R of O.boundaryInputs)if(!h.has(R)&&!p.has(R)){p.add(R);let B=d.get(R)||null;B&&(f.set(R,c.length),c.push(B))}for(let R of O.boundaryOutputs)if(!p.has(R)){p.add(R);let B=d.get(R)||null;B&&(f.set(R,c.length),c.push(B))}}for(let O of o)p.has(O.id)||(p.add(O.id),f.set(O.id,c.length),c.push(O));let m=o.map(O=>O.type),g=c.map(O=>O.type),y=[...m,...g],_=[];for(let O=0;O<i.length;O++)a.has(i[O].id)&&_.push(i[O].type);let b=new yn(`backward_${e.name}`,y,_),w=new $n(b),x=b.args,S=x.slice(0,m.length),T=x.slice(m.length),v=new Map;for(let O=0;O<c.length;O++)v.set(c[O].id,T[O]);let A=new jo(w);for(let O=0;O<o.length;O++)A.accumulate(o[O].id,S[O]);let N=new Map;for(let O of r)if(O.opName==="constant"){let R=O.getResult(0).type,B=w._buildOp("constant",[],[R],new Map(O.attributes),null);N.set(O.getResult(0).id,B.getResult(0))}for(let O=u.length-1;O>=0;O--){let R=u[O],B=new Map;for(let E of R.ops){let F=new Array(E.numOperands);for(let G=0;G<E.numOperands;G++){let U=E.getOperand(G),D=B.get(U.id)||v.get(U.id)||N.get(U.id);F[G]=D||U}let C=E.results.map(G=>G.type),H=w._buildOp(E.opName,F,C,new Map(E.attributes),null);for(let G=0;G<E.numResults;G++)B.set(E.getResult(G).id,H.getResult(G))}Qd(R.ops,{accumulator:A,builder:w,needsGrad:a,resolveValue:l(E=>B.get(E.id)||v.get(E.id)||N.get(E.id)||E,"resolveValue")})}let k=[];for(let O=0;O<i.length;O++)a.has(i[O].id)&&k.push(Hd(w,i[O],A));return w.returnOp(k),{backwardFunc:b,savedValues:c,gradInputIndices:this._getGradInputIndices(i,a)}}}});var Lu,gw=M(()=>{"use strict";Lu=class{static{l(this,"RematPolicy")}constructor(e={}){this._maxRematDepth=e.maxRematDepth||1,this._sizeThreshold=e.sizeThreshold||1024*1024,this._alwaysRemat=new Set(e.alwaysRemat||["neg","abs","sign","floor","ceil","exp","log","sqrt","rsqrt","sin","cos","tanh"]),this._neverRemat=new Set(e.neverRemat||["matmul","dot","conv","reduce","custom_call","pool2d"])}shouldRematerialize(e){if(this._alwaysRemat.has(e.opName))return!0;if(this._neverRemat.has(e.opName))return!1;let t=e.numResults>0?e.getResult(0).type:null;return!t||!t.shape||t.numel()>this._sizeThreshold?!1:this._isElementwise(e)}_isElementwise(e){if(e.numOperands===0||e.numResults===0)return!1;let t=e.getResult(0).type.shape;for(let r=0;r<e.numOperands;r++){let s=e.getOperand(r).type;if(!s||!s.shape||s.shape.length!==t.length)return!1;for(let o=0;o<t.length;o++)if(s.shape[o]!==t[o])return!1}return!0}}});var eh,hk=M(()=>{"use strict";Us();_s();ro();Yd();Is();gw();Zd();eh=class{static{l(this,"JointGraphBuilder")}constructor(e={}){this._rematPolicy=e.rematPolicy||new Lu(e.remat||{}),this._checkpointPolicy=e.checkpointPolicy||null}build(e){if(this._checkpointPolicy)return this._buildCheckpointed(e);let t=this._buildScaffold(e);return Qd(t.topoOrder,{accumulator:t.accumulator,builder:t.builder,needsGrad:t.needsGrad,resolveValue:l(r=>t.valueMap.get(r.id)||r,"resolveValue")}),this._finish(t)}_buildCheckpointed(e){let t=this._buildScaffold(e),r=this._checkpointPolicy.segment(t.topoOrder,e);for(let s=r.length-1;s>=0;s--){let o=r[s],i=new Map;for(let a of o.ops){let u=new Array(a.numOperands);for(let f=0;f<a.numOperands;f++){let d=a.getOperand(f);u[f]=i.get(d.id)||t.valueMap.get(d.id)||d}let p=a.results.map(f=>f.type),c=t.builder._buildOp(a.opName,u,p,new Map(a.attributes),null);for(let f=0;f<a.numResults;f++)i.set(a.getResult(f).id,c.getResult(f))}Qd(o.ops,{accumulator:t.accumulator,builder:t.builder,needsGrad:t.needsGrad,resolveValue:l(a=>i.get(a.id)||t.valueMap.get(a.id)||a,"resolveValue")})}return this._finish(t)}_buildScaffold(e){let r=ft.compute(e).topologicalOrder;this._assertNoRegionControlFlow(r);let s=e.getReturnOp();if(!s)throw new Error("Forward function has no return op");let o=s.operands,i=e.args,a=o.map(w=>w.type),u=[...e.inputTypes,...a],p=[...e.outputTypes,...e.inputTypes],c=new yn(`joint_${e.name}`,u,p),f=new $n(c),d=c.args,h=d.slice(0,e.inputTypes.length),m=d.slice(e.inputTypes.length),g=new Map;for(let w=0;w<i.length;w++)g.set(i[w].id,h[w]);for(let w of r){if(w.opName==="return")continue;let x=new Array(w.numOperands);for(let v=0;v<w.numOperands;v++){let A=w.getOperand(v);x[v]=g.get(A.id)||A}let S=w.results.map(v=>v.type),T=f._buildOp(w.opName,x,S,new Map(w.attributes),null);for(let v=0;v<w.numResults;v++)g.set(w.getResult(v).id,T.getResult(v))}let y=o.map(w=>g.get(w.id)),_=this._computeGradReachability(e,r),b=new jo(f);for(let w=0;w<o.length;w++)b.accumulate(o[w].id,m[w]);return{topoOrder:r,forwardInputs:i,forwardOutputs:o,fwdOutputValues:y,valueMap:g,builder:f,needsGrad:_,accumulator:b,jointFunc:c}}_finish(e){let t=[];for(let r=0;r<e.forwardInputs.length;r++)t.push(Hd(e.builder,e.forwardInputs[r],e.accumulator));return e.builder.returnOp([...e.fwdOutputValues,...t]),{jointFunc:e.jointFunc,numForwardOutputs:e.forwardOutputs.length,numGradInputs:t.length}}_computeGradReachability(e,t){let r=new Set,s=e.getReturnOp();for(let o of s.operands)r.add(o.id);for(let o=t.length-1;o>=0;o--){let i=t[o];if(!(i.opName==="return"||!i.results.some(u=>r.has(u.id)))&&Eu(i.opName)&&!dw(i.opName))for(let u=0;u<i.numOperands;u++)r.add(i.getOperand(u).id)}return r}_assertNoRegionControlFlow(e){for(let t of e)if(vi.has(t.opName))throw new Error(`JointGraphBuilder does not support region control-flow op '${t.opName}'; use BackwardGraphBuilder (separate mode) without a checkpointPolicy, which differentiates scan/if.`)}}});var cC=M(()=>{"use strict"});function pC(n,e){let t=n.gradOutputs[0],[r,s]=n.operands,o=n.full(0,r.type),i=n.builder.compare(r,s,e).getResult(0),a=n.builder.select(i,t,o).getResult(0),u=n.builder.select(i,o,t).getResult(0);return[a,u]}function fC(n){let e=n.gradOutputs[0],[t]=n.operands,r=n.full(0,e.type),s=n.builder.where(t,e,r).getResult(0),o=n.builder.where(t,r,e).getResult(0);return[null,s,o]}var dC=M(()=>{"use strict";Is();pe("add",n=>{let e=n.gradOutputs[0];return[e,e]});pe("sub",n=>{let e=n.gradOutputs[0],t=n.builder.neg(e).getResult(0);return[e,t]});l(pC,"_minMaxVJP");pe("maximum",n=>pC(n,"ge"));pe("minimum",n=>pC(n,"le"));pe("clamp",n=>{let e=n.gradOutputs[0],[t,r,s]=n.operands,o=n.full(0,r.type),i=n.builder.compare(r,t,"ge").getResult(0),a=n.builder.where(i,e,o).getResult(0),u=n.builder.compare(r,s,"le").getResult(0);return[null,n.builder.where(u,a,o).getResult(0),null]});l(fC,"_whereVJP");pe("where",fC);pe("select",fC);pe("mul",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=n.builder.mul(e,r).getResult(0),o=n.builder.mul(e,t).getResult(0);return[s,o]});pe("div",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=n.builder.div(e,r).getResult(0),o=n.builder.mul(r,r).getResult(0),i=n.builder.neg(t).getResult(0),a=n.builder.mul(e,i).getResult(0),u=n.builder.div(a,o).getResult(0);return[s,u]});pe("neg",n=>{let e=n.gradOutputs[0];return[n.builder.neg(e).getResult(0)]});pe("pow",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=n.full(1,t.type),o=n.builder.sub(r,s).getResult(0),i=n.builder.pow(t,o).getResult(0),a=n.builder.mul(r,i).getResult(0),u=n.builder.mul(e,a).getResult(0),p=n.builder.pow(t,r).getResult(0),c=n.builder.log(t).getResult(0),f=n.builder.mul(p,c).getResult(0),d=n.builder.mul(e,f).getResult(0);return[u,d]})});function hC(n,e){let t=n.full(2/Math.sqrt(Math.PI),e.type),r=n.builder.mul(e,e).getResult(0),s=n.builder.neg(r).getResult(0),o=n.builder.exp(s).getResult(0);return n.builder.mul(t,o).getResult(0)}function mC(n,e){let t=n.builder,r=n.full(1,e.type),s=t.add(e,n.full(ii,e.type)).getResult(0),o=t.div(r,s).getResult(0),i=t.log(s).getResult(0),a=t.sub(i,t.mul(n.full(.5,e.type),o).getResult(0)).getResult(0),u=t.mul(o,o).getResult(0),p=u;for(let c of Ga)a=t.add(a,t.mul(n.full(c,e.type),p).getResult(0)).getResult(0),p=t.mul(p,u).getResult(0);for(let c=0;c<ii;c++){let f=t.add(e,n.full(c,e.type)).getResult(0);a=t.sub(a,t.div(r,f).getResult(0)).getResult(0)}return a}var gC=M(()=>{"use strict";Is();Yi();pe("exp",n=>{let e=n.gradOutputs[0],t=n.results[0];return[n.builder.mul(e,t).getResult(0)]});pe("log",n=>{let e=n.gradOutputs[0],[t]=n.operands;return[n.builder.div(e,t).getResult(0)]});pe("sqrt",n=>{let e=n.gradOutputs[0],t=n.results[0],r=n.full(2,t.type),s=n.builder.mul(r,t).getResult(0);return[n.builder.div(e,s).getResult(0)]});pe("tanh",n=>{let e=n.gradOutputs[0],t=n.results[0],r=n.builder.mul(t,t).getResult(0),s=n.builder.sub(n.full(1,t.type),r).getResult(0);return[n.builder.mul(e,s).getResult(0)]});pe("sigmoid",n=>{let e=n.gradOutputs[0],t=n.results[0],r=n.builder.sub(n.full(1,t.type),t).getResult(0),s=n.builder.mul(t,r).getResult(0);return[n.builder.mul(e,s).getResult(0)]});pe("relu",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.full(0,t.type),s=n.builder.compare(t,r,"gt").getResult(0);return[n.builder.select(s,e,r).getResult(0)]});pe("gelu",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.full(1.702,t.type),s=n.full(1,t.type),o=n.builder.mul(r,t).getResult(0),i=n.builder.sigmoid(o).getResult(0),a=n.builder.sub(s,i).getResult(0),u=n.builder.mul(o,a).getResult(0),p=n.builder.add(s,u).getResult(0),c=n.builder.mul(i,p).getResult(0);return[n.builder.mul(e,c).getResult(0)]});pe("silu",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.builder.sigmoid(t).getResult(0),s=n.builder.sub(n.full(1,t.type),r).getResult(0),o=n.builder.mul(t,s).getResult(0),i=n.builder.add(r,n.builder.mul(r,o).getResult(0)).getResult(0);return[n.builder.mul(e,i).getResult(0)]});pe("sin",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.builder.cos(t).getResult(0);return[n.builder.mul(e,r).getResult(0)]});pe("cos",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.builder.sin(t).getResult(0),s=n.builder.neg(r).getResult(0);return[n.builder.mul(e,s).getResult(0)]});pe("abs",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.builder.sign(t).getResult(0);return[n.builder.mul(e,r).getResult(0)]});l(hC,"_erfDerivIR");l(mC,"_digammaIR");pe("erf",n=>{let e=n.gradOutputs[0],[t]=n.operands;return[n.builder.mul(e,hC(n,t)).getResult(0)]});pe("erfc",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.builder.neg(hC(n,t)).getResult(0);return[n.builder.mul(e,r).getResult(0)]});pe("lgamma",n=>{let e=n.gradOutputs[0],[t]=n.operands;return[n.builder.mul(e,mC(n,t)).getResult(0)]});pe("gamma",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.results[0],s=n.builder.mul(r,mC(n,t)).getResult(0);return[n.builder.mul(e,s).getResult(0)]});pe("log2",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.full(Math.LN2,t.type),s=n.builder.mul(t,r).getResult(0);return[n.builder.div(e,s).getResult(0)]});pe("log10",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.full(Math.LN10,t.type),s=n.builder.mul(t,r).getResult(0);return[n.builder.div(e,s).getResult(0)]});pe("exp2",n=>{let e=n.gradOutputs[0],t=n.results[0],r=n.full(Math.LN2,t.type),s=n.builder.mul(t,r).getResult(0);return[n.builder.mul(e,s).getResult(0)]});pe("square",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.full(2,t.type),s=n.builder.mul(r,t).getResult(0);return[n.builder.mul(e,s).getResult(0)]});pe("reciprocal",n=>{let e=n.gradOutputs[0],t=n.results[0],r=n.builder.mul(t,t).getResult(0),s=n.builder.neg(r).getResult(0);return[n.builder.mul(e,s).getResult(0)]});pe("rsqrt",n=>{let e=n.gradOutputs[0],t=n.results[0],r=n.builder.mul(n.builder.mul(t,t).getResult(0),t).getResult(0),s=n.full(-.5,t.type),o=n.builder.mul(s,r).getResult(0);return[n.builder.mul(e,o).getResult(0)]})});var yC=M(()=>{"use strict";Is();pe("dot",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=n.op.getAttr("lhs_contracting"),o=n.op.getAttr("rhs_contracting"),i=n.op.getAttr("lhs_batch")||[],a=n.op.getAttr("rhs_batch")||[],u=t.type.rank,p=r.type.rank,c=e.type.rank,f=[];for(let _=0;_<u;_++)!s.includes(_)&&!i.includes(_)&&f.push(_);let d=[];for(let _=0;_<p;_++)!o.includes(_)&&!a.includes(_)&&d.push(_);let h=[];for(let _=i.length;_<i.length+d.length;_++)h.push(_+f.length);let m=[];for(let _=i.length;_<i.length+f.length;_++)m.push(_);let g=n.builder.dot(e,r,h,d,Array.from({length:i.length},(_,b)=>b),a).getResult(0),y=n.builder.dot(t,e,f,m,i,Array.from({length:i.length},(_,b)=>b)).getResult(0);return[g,y]});pe("matmul",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=l(p=>{let c=Array.from({length:p},(f,d)=>d);return c[p-2]=p-1,c[p-1]=p-2,c},"swapLastTwo"),o=n.builder.transpose(r,s(r.type.rank)).getResult(0),i=n.builder.matmul(e,o).getResult(0),a=n.builder.transpose(t,s(t.type.rank)).getResult(0),u=n.builder.matmul(a,e).getResult(0);return[i,u]});pe("conv",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=n.builder,o=n.op.getAttr("strides"),i=n.op.getAttr("padding"),a=n.op.getAttr("dilation")||o.map(()=>1),u=n.op.getAttr("groups")||1,p=n.op.getAttr("input_layout"),c=n.op.getAttr("kernel_layout");if(!(o.length===2&&o.every(A=>A===1)&&a.every(A=>A===1)&&u===1&&p==="NCHW"&&c==="OIHW"))throw new Error("conv VJP supports only 2D stride-1 dilation-1 groups-1 NCHW/OIHW conv");let d=r.type.shape,h=d[2],m=d[3],[g,y]=i,_=s.reverse(s.transpose(r,[1,0,2,3]).getResult(0),[2,3]).getResult(0),b=[[h-1-g[0],h-1-g[1]],[m-1-y[0],m-1-y[1]]],w=s.conv(e,_,[1,1],b).getResult(0),x=s.transpose(t,[1,0,2,3]).getResult(0),S=s.transpose(e,[1,0,2,3]).getResult(0),T=s.conv(x,S,[1,1],[g,y]).getResult(0),v=s.transpose(T,[1,0,2,3]).getResult(0);return[w,v]})});var _C=M(()=>{"use strict";Is();pe("reduce",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.op.getAttr("reduce_type"),s=n.op.getAttr("dimensions"),o=t.type.shape,i=t.type.dtype,a=new Set(s),u=o.map((f,d)=>a.has(d)?1:f),p=o.map((f,d)=>d),c=n.builder.reshape(e,u).getResult(0);if(r==="sum")return[n.builder.broadcast(c,o,p).getResult(0),null];if(r==="mean"){let f=1;for(let y=0;y<o.length;y++)a.has(y)&&(f*=o[y]);let d=n.builder.broadcast(c,o,p).getResult(0),h=n.builder.scalarConstant(f,i).getResult(0),m=n.builder.broadcast(h,o,[]).getResult(0);return[n.builder.div(d,m).getResult(0),null]}if(r==="max"||r==="min"){let f=n.builder.reshape(n.results[0],u).getResult(0),d=n.builder.broadcast(f,o,p).getResult(0),h=n.builder.compare(t,d,"eq").getResult(0),m=n.builder.convert(h,i).getResult(0),g=n.builder.broadcast(c,o,p).getResult(0);return[n.builder.mul(g,m).getResult(0),null]}throw r==="prod"?new Error("reduce VJP for reduce_type='prod' is not implemented; provide a gradient rule or avoid differentiating reduce_prod"):new Error(`reduce VJP: unsupported reduce_type '${r}' on the gradient path (would silently drop the gradient)`)})});var bC=M(()=>{"use strict";Is();pe("reshape",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=t.type.shape;return[n.builder.reshape(e,r).getResult(0)]});pe("reverse",n=>[n.builder.reverse(n.gradOutputs[0],n.op.getAttr("dimensions")).getResult(0)]);pe("transpose",n=>{let e=n.gradOutputs[0],t=n.op.getAttr("permutation"),r=new Array(t.length);for(let s=0;s<t.length;s++)r[t[s]]=s;return[n.builder.transpose(e,r).getResult(0)]});pe("broadcast_in_dim",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.op.getAttr("broadcast_dimensions"),s=t.type.shape,o=n.results[0].type.shape,i=t.type.dtype,a=[];for(let f=0;f<o.length;f++)if(!r.includes(f))a.push(f);else{let d=r.indexOf(f);s[d]===1&&o[f]!==1&&a.push(f)}if(a.length===0)return[e];let u=n.builder.scalarConstant(0,i).getResult(0),p=n.builder.reduce(e,u,a,"sum").getResult(0),c=p.type.shape;return c.length!==s.length||!c.every((f,d)=>f===s[d])?[n.builder.reshape(p,s).getResult(0)]:[p]});pe("slice",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.op.getAttr("starts"),s=t.type.shape,o=e.type.shape,i=n.op.getAttr("strides")||s.map(()=>1),a=t.type.dtype,u=[...r],p=new Array(s.length),c=new Array(s.length);for(let d=0;d<s.length;d++)c[d]=i[d]-1,p[d]=s[d]-r[d]-(o[d]-1)*i[d]-1;let f=n.builder.scalarConstant(0,a).getResult(0);return[n.builder.pad(e,f,u,p,c).getResult(0)]});pe("concat",n=>{let e=n.gradOutputs[0],t=n.op.getAttr("dimension"),r=n.operands,s=[],o=0;for(let i=0;i<r.length;i++){let a=r[i].type.shape,u=new Array(a.length).fill(0),p=[...e.type.shape];u[t]=o,p[t]=o+a[t],s.push(n.builder.slice(e,u,p).getResult(0)),o+=a[t]}return s});pe("gather",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=n.builder.scalarConstant(0,t.type.dtype).getResult(0),o=n.builder.broadcast(s,t.type.shape,[]).getResult(0);return[n.builder.scatterAdd(o,r,e,{updateWindowDims:n.op.getAttr("offset_dims"),insertedWindowDims:n.op.getAttr("collapsed_slice_dims"),scatterDimsToOperandDims:n.op.getAttr("start_index_map"),indexVectorDim:n.op.getAttr("index_vector_dim")}).getResult(0),null]});pe("scatter",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=n.op.getAttr("inserted_window_dims"),o=t.type.shape.map((a,u)=>s.includes(u)?1:a),i=n.builder.gather(e,r,{offsetDims:n.op.getAttr("update_window_dims"),collapsedSliceDims:s,startIndexMap:n.op.getAttr("scatter_dims_to_operand_dims"),indexVectorDim:n.op.getAttr("index_vector_dim"),sliceSizes:o}).getResult(0);return[e,null,i]});pe("pad",n=>{let e=n.gradOutputs[0],t=n.op.getAttr("low"),r=n.op.getAttr("high"),[s]=n.operands,o=s.type.shape,i=[...t],a=new Array(o.length);for(let u=0;u<o.length;u++)a[u]=t[u]+o[u];return[n.builder.slice(e,i,a).getResult(0),null]})});var wC=M(()=>{"use strict";Is();_s();pe("softmax",n=>{let e=n.gradOutputs[0],t=n.results[0],r=n.op.getAttr("axis"),s=t.type.dtype,o=t.type.shape,i=n.builder.mul(e,t).getResult(0),a=n.builder.scalarConstant(0,s).getResult(0),u=n.builder.reduce(i,a,[r],"sum").getResult(0),p=Xo(o.length,r),c=n.builder.broadcast(u,o,p).getResult(0),f=n.builder.sub(e,c).getResult(0);return[n.builder.mul(t,f).getResult(0)]});pe("log_softmax",n=>{let e=n.gradOutputs[0],t=n.results[0],[r]=n.operands,s=n.op.getAttr("axis"),o=t.type.dtype,i=t.type.shape,a=n.builder.exp(t).getResult(0),u=n.builder.scalarConstant(0,o).getResult(0),p=n.builder.reduce(e,u,[s],"sum").getResult(0),c=Xo(i.length,s),f=n.builder.broadcast(p,i,c).getResult(0),d=n.builder.mul(a,f).getResult(0);return[n.builder.sub(e,d).getResult(0)]});pe("layer_norm",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=n.op.getAttr("axis"),o=n.op.getAttr("epsilon"),i=t.type.dtype,a=t.type.shape,u=[s<0?a.length+s:s],p=a[u[0]],c=Xo(a.length,u[0]),f=n.builder.scalarConstant(0,i).getResult(0),d=n.builder.reduce(t,f,u,"mean").getResult(0),h=n.builder.broadcast(d,a,c).getResult(0),m=n.builder.sub(t,h).getResult(0),g=n.builder.mul(m,m).getResult(0),y=n.builder.reduce(g,f,u,"mean").getResult(0),_=n.full(o,y.type),b=n.builder.add(y,_).getResult(0),w=n.builder.rsqrt(b).getResult(0),x=n.builder.broadcast(w,a,c).getResult(0),S=[];for(let ue=0;ue<a.length;ue++)u.includes(ue)&&S.push(ue);let T=n.builder.broadcast(r,a,S).getResult(0),v=n.builder.mul(e,T).getResult(0),A=n.full(p,t.type),N=n.builder.mul(A,v).getResult(0),k=n.builder.reduce(v,f,u,"sum").getResult(0),O=n.builder.broadcast(k,a,c).getResult(0),R=n.builder.mul(m,x).getResult(0),B=n.builder.mul(v,R).getResult(0),E=n.builder.reduce(B,f,u,"sum").getResult(0),F=n.builder.broadcast(E,a,c).getResult(0),C=n.builder.mul(R,F).getResult(0),H=n.builder.sub(N,n.builder.add(O,C).getResult(0)).getResult(0),G=n.builder.div(x,A).getResult(0),U=n.builder.mul(G,H).getResult(0),D=n.builder.reduce(n.builder.mul(e,R).getResult(0),f,c,"sum").getResult(0),V=n.builder.reduce(e,f,c,"sum").getResult(0);return[U,D,V]});pe("scaled_dot_product_attention",n=>{let e=n.gradOutputs[0],[t,r,s]=n.operands,o=n.op.getAttr("scale");if(n.op.getAttr("causal"))throw new Error("causal scaled_dot_product_attention VJP not supported");let i=n.builder,a=t.type.dtype,u=t.type.rank,p=[];for(let N=0;N<u;N++)p.push(N);p[u-2]=u-1,p[u-1]=u-2;let c=l(N=>i.transpose(N,p).getResult(0),"lastT"),f=i.matmul(t,c(r)).getResult(0),d=i.mul(f,n.full(o,f.type)).getResult(0),h=i.softmax(d,u-1).getResult(0),m=i.matmul(c(h),e).getResult(0),g=i.matmul(e,c(s)).getResult(0),y=i.scalarConstant(0,a).getResult(0),_=i.mul(g,h).getResult(0),b=i.reduce(_,y,[u-1],"sum").getResult(0),w=[];for(let N=0;N<u-1;N++)w.push(N);let x=i.broadcast(b,h.type.shape,w).getResult(0),S=i.mul(h,i.sub(g,x).getResult(0)).getResult(0),T=i.mul(S,n.full(o,S.type)).getResult(0),v=i.matmul(T,r).getResult(0),A=i.matmul(c(T),t).getResult(0);return[v,A,m]});pe("pool2d",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.builder,s=n.op.getAttr("pool_type"),o=n.op.getAttr("kernel_size"),i=n.op.getAttr("strides"),a=n.op.getAttr("padding"),u=n.op.getAttr("layout")||"NCHW",p=a.every(v=>v[0]===0&&v[1]===0),c=i[0]===o[0]&&i[1]===o[1];if(u!=="NCHW"||!p||!c||s!=="avg"&&s!=="max")throw new Error("pool2d VJP supports only non-overlapping (stride=kernel) avg/max pooling without padding, NCHW");let[f,d,h,m]=e.type.shape,[g,y]=o,_=t.type.shape,b=l(v=>r.reshape(r.broadcast(v,[f,d,h,g,m,y],[0,1,2,4]).getResult(0),_).getResult(0),"upsample"),w=b(e);if(s==="avg"){let v=n.full(g*y,t.type);return[r.div(w,v).getResult(0)]}let x=b(n.results[0]),S=r.compare(t,x,"eq").getResult(0),T=n.full(0,t.type);return[r.select(S,w,T).getResult(0)]});pe("batch_norm",n=>{let e=n.gradOutputs[0],[t,r,,s,o]=n.operands,i=n.op.getAttr("axis"),a=n.op.getAttr("epsilon"),u=t.type.dtype,p=t.type.shape,c=n.builder,f=Xo(p.length,i),d=c.scalarConstant(0,u).getResult(0),h=n.full(a,o.type),m=c.rsqrt(c.add(o,h).getResult(0)).getResult(0),g=c.broadcast(m,p,[i]).getResult(0),y=c.broadcast(s,p,[i]).getResult(0),_=c.broadcast(r,p,[i]).getResult(0),b=c.sub(t,y).getResult(0),w=c.mul(b,g).getResult(0),x=c.mul(e,_).getResult(0),S=c.mul(x,g).getResult(0),T=c.reduce(c.mul(e,w).getResult(0),d,f,"sum").getResult(0),v=c.reduce(e,d,f,"sum").getResult(0),A=c.neg(c.reduce(S,d,f,"sum").getResult(0)).getResult(0),N=c.mul(c.mul(g,g).getResult(0),g).getResult(0),k=n.full(-.5,t.type),O=c.mul(c.mul(c.mul(x,b).getResult(0),N).getResult(0),k).getResult(0),R=c.reduce(O,d,f,"sum").getResult(0);return[S,T,v,A,R]});pe("elu",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.results[0],s=n.op.getAttr("alpha")??1,o=n.full(0,t.type),i=n.full(1,t.type),a=n.full(s,t.type),u=n.builder.compare(t,o,"gt").getResult(0),p=n.builder.add(r,a).getResult(0),c=n.builder.select(u,i,p).getResult(0);return[n.builder.mul(e,c).getResult(0)]});pe("leaky_relu",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.op.getAttr("negative_slope")??.01,s=n.full(0,t.type),o=n.full(1,t.type),i=n.full(r,t.type),a=n.builder.compare(t,s,"gt").getResult(0),u=n.builder.select(a,o,i).getResult(0);return[n.builder.mul(e,u).getResult(0)]});pe("celu",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.op.getAttr("alpha")??1,s=n.full(0,t.type),o=n.full(1,t.type),i=n.builder.compare(t,s,"gt").getResult(0),a=n.full(r,t.type),u=n.builder.div(t,a).getResult(0),p=n.builder.exp(u).getResult(0),c=n.builder.select(i,o,p).getResult(0);return[n.builder.mul(e,c).getResult(0)]});pe("selu",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=1.0507009873554805,s=1.6732632423543772,o=n.full(0,t.type),i=n.full(r,t.type),a=n.builder.compare(t,o,"gt").getResult(0),u=n.full(s,t.type),p=n.builder.exp(t).getResult(0),c=n.builder.mul(u,p).getResult(0),f=n.builder.select(a,i,n.builder.mul(i,c).getResult(0)).getResult(0);return[n.builder.mul(e,f).getResult(0)]});pe("hardswish",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.full(-3,t.type),s=n.full(3,t.type),o=n.full(0,t.type),i=n.full(1,t.type),a=n.full(2,t.type),u=n.full(6,t.type),p=n.builder.compare(t,r,"le").getResult(0),c=n.builder.compare(t,s,"ge").getResult(0),f=n.builder.add(n.builder.mul(a,t).getResult(0),s).getResult(0),d=n.builder.div(f,u).getResult(0),h=n.builder.select(p,o,n.builder.select(c,i,d).getResult(0)).getResult(0);return[n.builder.mul(e,h).getResult(0)]});pe("hardsigmoid",n=>{let e=n.gradOutputs[0],[t]=n.operands,r=n.full(-3,t.type),s=n.full(3,t.type),o=n.full(0,t.type),i=n.full(1/6,t.type),a=n.builder.compare(t,r,"le").getResult(0),u=n.builder.compare(t,s,"ge").getResult(0),p=n.builder.select(a,o,n.builder.select(u,o,i).getResult(0)).getResult(0);return[n.builder.mul(e,p).getResult(0)]});pe("embedding",n=>{let e=n.gradOutputs[0],[t,r]=n.operands,s=r.type.rank,o=n.full(0,t.type);return[n.builder.scatter(o,r,e,{updateWindowDims:[s],insertedWindowDims:[0],scatterDimsToOperandDims:[0],indexVectorDim:s}).getResult(0),null]})});var TC=M(()=>{"use strict";Is();pk("stop_gradient");pe("stop_gradient",()=>[null]);for(let n of["compare","logical_not","argmax","argmin","iota"])pk(n)});var xC=M(()=>{"use strict";Is();Yd();Zd();hk();gw();cC();dC();gC();yC();_C();bC();wC();TC()});var SC={};at(SC,{compileWithBackward:()=>_w});function yw(n){return typeof n.then=="function"}function _w(n,e,t={}){let r=t.target??go(),s=t.mode||"separate",o=t.rematPolicy||new Lu(t.remat||{}),i={target:r,verify:!1,...t,backward:void 0,mode:void 0,rematPolicy:void 0,remat:void 0},a=t.dynamic_shapes||null,u=[],p=null,c=null,f=null;function d(k){let O=ww((...B)=>n.forward(...B),k,{name:n.constructor.name||"compiled",dynamicShapes:a}),R=l(B=>{let E=B.graph.functions().next().value,F=s==="joint"?m(E,B,o):h(E,B,o);return F.shapeEnv=B.shapeEnv,F.outputSymShapes=B.outputSymShapes,F},"finish");return yw(O)?O.then(R):R(O)}l(d,"_compile");function h(k,O,R){let B=new Xd({rematPolicy:R}),{backwardFunc:E,savedValues:F,gradInputIndices:C}=B.build(k),H=k.getReturnOp(),G=[...H.operands],U=G.length,D=new Map(k.args.map((ke,rt)=>[ke.id,rt])),V=new Map(G.map((ke,rt)=>[ke.id,rt])),ue=[],ce=new Map;for(let ke of F)D.has(ke.id)||V.has(ke.id)||ce.has(ke.id)||(ce.set(ke.id,U+ue.length),ue.push(ke));ue.length>0&&(H.erase(),new $n(k).returnOp([...G,...ue]),k.outputTypes=Object.freeze([...G.map(ke=>ke.type),...ue.map(ke=>ke.type)]));let Le=F.map(ke=>D.has(ke.id)?{kind:"arg",index:D.get(ke.id)}:V.has(ke.id)?{kind:"output",index:V.get(ke.id)}:{kind:"output",index:ce.get(ke.id)}),j=new Mn("forward");j.addFunction(k);let W=new Si(i).compile(j),ie=new Mn("backward");ie.addFunction(E);let je=new Si(i).compile(ie);return{mode:"separate",fwdResult:W,bwdResult:je,forwardFunc:k,backwardFunc:E,savedValues:F,savedSources:Le,numRealOutputs:U,gradInputIndices:C,capturedParams:O.capturedParams,numUserInputs:O.numUserInputs,outputTypes:O.outputTypes,shapeEnv:O.shapeEnv,outputSymShapes:O.outputSymShapes}}l(h,"_compileSeparate");function m(k,O,R){let B=new eh({rematPolicy:R}),{jointFunc:E,numForwardOutputs:F,numGradInputs:C}=B.build(k),H=new Mn("joint");return H.addFunction(E),{mode:"joint",result:new Si(i).compile(H),jointFunc:E,numForwardOutputs:F,numGradInputs:C,capturedParams:O.capturedParams,numUserInputs:O.numUserInputs,outputTypes:O.outputTypes,inputTypes:k.inputTypes,shapeEnv:O.shapeEnv,outputSymShapes:O.outputSymShapes}}l(m,"_compileJoint");function g(k,O){return k.outputSymShapes&&k.shapeEnv?k.shapeEnv.resolveSymbolicShape(k.outputSymShapes[O]):[...k.outputTypes[O].shape]}l(g,"_resolveOutputShape");function y(k,O,R){let B=k.module||k;return B.executionPlan?B.runPlanAsync(B.executionPlan,R):k.isAsync(O)?k.runAsync(O,...R):(k.run(O,...R),null)}l(y,"_runK");function _(k,O){let B=k.fwdResult.listKernels()[0],E=O.length>0?O[0].device:"cpu",F=O.map(j=>Qe(j)),H=k.capturedParams.map(j=>Qe(j)),G=k.forwardFunc.outputTypes,U=k.numRealOutputs,D=new Array(G.length),V=new Array(G.length);for(let j=0;j<G.length;j++){let W=j<U?g(k,j):[...G[j].shape],ie=G[j].dtype,je=bn(W),ke=St(ie);D[j]=new ke(Math.max(je,1)),V[j]=W}let ue=[...F,...H,...D],ce=l(()=>({results:U===1?Ee(D[0],V[0],G[0].dtype,E):Array.from({length:U},(W,ie)=>Ee(D[ie],V[ie],G[ie].dtype,E)),inputArrays:F,paramArrays:H,outputArrays:D,device:E}),"build"),Le=y(k.fwdResult,B,ue);return Le?Le.then(ce):ce()}l(_,"_executeSeparateForward");function b(k,O,R){let E=k.bwdResult.listKernels()[0],F=O.map(ie=>Qe(ie)),C=k.savedValues,H=k.savedSources,G=[...R.inputArrays,...R.paramArrays],U=new Array(C.length);for(let ie=0;ie<C.length;ie++){let je=H[ie];U[ie]=je.kind==="arg"?G[je.index]:R.outputArrays[je.index]}let D=k.backwardFunc,V=D.outputTypes.length,ue=new Array(V),ce=new Array(V);for(let ie=0;ie<V;ie++){let je=[...D.outputTypes[ie].shape],ke=D.outputTypes[ie].dtype,rt=bn(je),zt=St(ke);ue[ie]=new zt(Math.max(rt,1)),ce[ie]=je}let Le=[...F,...U,...ue],j=l(()=>ue.map((ie,je)=>Ee(ie,ce[je],D.outputTypes[je].dtype,R.device)),"build"),W=y(k.bwdResult,E,Le);return W?W.then(j):j()}l(b,"_executeSeparateBackward");function w(k){for(let O=0;O<u.length;O++){let R=u[O];R.shapeEnv.bindInputShapes(k);let{passed:B}=R.shapeEnv.evaluateGuards();if(B)return R}return null}l(w,"_findCachedEntry");function x(k,O){if(c=k,k.mode==="joint")return v(k,O);let R=_(k,O);return yw(R)?R.then(B=>(p=B,B.results)):(p=R,R.results)}l(x,"_forwardWith");function S(k){let O=w(k);if(O)return x(O,k);let R=d(k);return yw(R)?R.then(B=>(u.push(B),B.shapeEnv.bindInputShapes(k),x(B,k))):(u.push(R),R.shapeEnv.bindInputShapes(k),x(R,k))}l(S,"_runForward");function T(...k){return f?f.then(()=>S(k)):S(k)}l(T,"compiledForward");function v(k,O){let B=k.result.listKernels()[0],E=O.length>0?O[0].device:"cpu",F=O.map(W=>Qe(W)),H=k.capturedParams.map(W=>Qe(W)),G=k.jointFunc,U=G.outputTypes.length,D=new Array(U),V=new Array(U);for(let W=0;W<U;W++){let ie=[...G.outputTypes[W].shape],je=G.outputTypes[W].dtype,ke=bn(ie),rt=St(je);D[W]=new rt(Math.max(ke,1)),V[W]=ie}let ue=new Array(k.numForwardOutputs);for(let W=0;W<k.numForwardOutputs;W++){let ie=k.outputTypes[W],je=bn(ie.shape),ke=St(ie.dtype);ue[W]=new ke(Math.max(je,1))}p={inputArrays:F,paramArrays:H,gradOutputArrays:ue,outputArrays:D,outputShapes:V,device:E,compiled:k};let ce=[...F,...H,...ue,...D],Le=l(()=>{let W=[];for(let ie=0;ie<k.numForwardOutputs;ie++)W.push(Ee(D[ie],V[ie],G.outputTypes[ie].dtype,E));return W.length===1?W[0]:W},"build"),j=y(k.result,B,ce);return j?j.then(Le):Le()}l(v,"_executeJointForward");let A=T;A.backward=function(...k){if(!c||!p)throw new Error("Must run forward before backward");return c.mode==="joint"?N(c,k,p):b(c,k,p)};function N(k,O,R){let{inputArrays:B,paramArrays:E,outputArrays:F,outputShapes:C,device:H}=R,G=O.map(W=>Qe(W));for(let W=0;W<G.length;W++)R.gradOutputArrays[W].set(G[W]);let U=k.jointFunc,D=U.outputTypes.length,V=new Array(D),ue=new Array(D);for(let W=0;W<D;W++){let ie=[...U.outputTypes[W].shape],je=U.outputTypes[W].dtype,ke=bn(ie),rt=St(je);V[W]=new rt(Math.max(ke,1)),ue[W]=ie}let ce=[...B,...E,...G,...V],Le=l(()=>{let W=[];for(let ie=k.numForwardOutputs;ie<D;ie++)W.push(Ee(V[ie],ue[ie],U.outputTypes[ie].dtype,H));return W},"build"),j=y(k.result,k.result.listKernels()[0],ce);return j?j.then(Le):Le()}if(l(N,"_executeJointBackward"),A.original=n,A.backwardGraph=()=>{if(u.length===0)return null;let k=u[0];return k.mode==="joint"?k.jointFunc:k.backwardFunc},A.forwardGraph=()=>{if(u.length===0)return null;let k=u[0];return k.mode==="joint"?k.jointFunc:k.forwardFunc},A.capturedParams=()=>u.length?u[0].capturedParams:[],e){let k=d(e);yw(k)?f=k.then(O=>(u.push(O),f=null,O)):u.push(k)}return A}var bw=M(()=>{"use strict";mk();aw();yo();Di();Zd();_s();hk();gw();xC();wn();rs();ss();l(yw,"_isThenable");l(_w,"compileWithBackward")});function Cu(n){return typeof n.then=="function"}function Jj(){vC||(nC(),vC=!0)}function Zj(n,e){if(!n)return new Array(e.length).fill(null);let t=new Array(e.length);for(let r=0;r<e.length;r++){let s=n[r];if(s===!0){let o=new Set;for(let i=0;i<e[r].shape.length;i++)o.add(i);t[r]=o}else s instanceof Set?t[r]=s:t[r]=null}return t}function ww(n,e,t){Jj();let r=t?.name||n.name||"traced",s=new I_(r),o=Zj(t?.dynamicShapes,e);for(let f=0;f<e.length;f++)s.createInput(e[f].shape,e[f].dtype,o[f]);let i=e.length,a=s._initGraph();function u(f){Array.isArray(f)?s.markOutputs(f):s.markOutput(f),s.deactivate();let d=s.getGraphModule(),h=d.functions().next().value;return{graph:d,capturedParams:[...s.capturedParams],numUserInputs:i,outputTypes:h.outputTypes,shapeEnv:s.shapeEnv,outputSymShapes:s.outputSymShapes}}l(u,"_finalize"),s.activate();let p=Jr.fromKey(48),c=NN(p,()=>n(...a));if(Cu(c))return c.then(f=>u(f),f=>{throw s.deactivate(),f});try{return u(c)}catch(f){throw s.deactivate(),f}}function kC(n,e,t){let r=ww(n,e,t);return Cu(r)?r.then(s=>s.graph):r.graph}function Qj(n,e,t){let r=n.result.listKernels();if(r.length===0)throw new Error("No kernels compiled");let s=r[0],o=e.length>0?e[0].device:"cpu",i=new Array(e.length);for(let g=0;g<e.length;g++)i[g]=Qe(e[g]);let a=n.capturedParams,u=new Array(a.length);for(let g=0;g<a.length;g++)u[g]=Qe(a[g]);let p=n.outputTypes,c=n.outputSymShapes,f=new Array(p.length),d=new Array(p.length);for(let g=0;g<p.length;g++){let y=c&&t?t.resolveSymbolicShape(c[g]):p[g].shape,_=p[g].dtype,b=bn(y),w=St(_);f[g]=new w(Math.max(b,1)),d[g]=[...y]}let h=new Array(i.length+u.length+f.length),m=0;for(let g=0;g<i.length;g++)h[m++]=new Lo(i[g],e[g].shape,e[g].dtype);for(let g=0;g<u.length;g++){let y=new Lo(u[g],a[g].shape,a[g].dtype),_=a[g]._impl;_&&(y.resident={key:_.storage.rawData,version:_.version}),h[m++]=y}for(let g=0;g<f.length;g++)h[m++]=new Lo(f[g],d[g],p[g].dtype);return{funcName:s,device:o,outputTypes:p,outputArrays:f,outputShapes:d,allArgs:h}}function gk(n,e,t,r){if(e.length===1)return Ee(t[0],r[0],e[0].dtype,n);let s=new Array(e.length);for(let o=0;o<e.length;o++)s[o]=Ee(t[o],r[o],e[o].dtype,n);return s}function Xj(n,e,t){let{funcName:r,device:s,outputTypes:o,outputArrays:i,outputShapes:a,allArgs:u}=Qj(n,e,t),p=n.result.module.executionPlan;return p?n.result.module.runPlanAsync(p,u,{resident:!0}).then(()=>gk(s,o,i,a)):n.result.isAsync(r)?n.result.runAsync(r,...u).then(()=>gk(s,o,i,a)):(n.result.run(r,...u),gk(s,o,i,a))}function eU(n,e,t={}){if(t?.backward)return _w(n,e,t);let r=t?.target??go(),s={target:r,verify:!1,...t},o=t?.dynamic_shapes||null,i=t?.shapeBuckets||null,a=t?.foldWeights??t?.quantization?.foldWeights??!1,u=[];function p(w,x,S){if(!w||typeof w!="object"||w.repro)return w;try{w.repro={name:n.constructor?.name||"compiled",phase:S,target:r?.name,inputs:(x||[]).map(T=>({shape:T.shape,dtype:T.dtype})),config:{fusion:s.fusion,scheduling:s.scheduling,optimization:s.optimization,quantization:s.quantization,dynamicShapes:!!o}}}catch{return w}return w}l(p,"_attachRepro");function c(w){let x=a?sC(w,Qe):w;return{result:new Si(s).compile(x.graph),graph:x.graph,capturedParams:x.capturedParams,numUserInputs:x.numUserInputs,outputTypes:x.outputTypes,shapeEnv:x.shapeEnv,outputSymShapes:x.outputSymShapes}}l(c,"_finalize");function f(w,x){try{let S=ww((...T)=>n.forward(...T),w,{name:n.constructor.name||"compiled",dynamicShapes:x});return Cu(S)?S.then(c,T=>{throw p(T,w,"compile")}):c(S)}catch(S){throw p(S,w,"compile")}}l(f,"_compileWith");function d(w){return f(w,o)}l(d,"_compile");function h(w){return w.map((x,S)=>({shape:x,dtype:e[S].dtype}))}l(h,"_bucketInputs");function m(w){for(let x=0;x<u.length;x++){let S=u[x];S.shapeEnv.bindInputShapes(w);let{passed:T}=S.shapeEnv.evaluateGuards();if(T)return S}return null}l(m,"_findCachedEntry");function g(w,x){try{let S=Xj(w,x,w.shapeEnv);return Cu(S)?S.then(void 0,T=>{throw p(T,x,"run")}):S}catch(S){throw p(S,x,"run")}}l(g,"_execute");function y(...w){let x=m(w);if(!x){let S=d(w);if(Cu(S))return S.then(T=>(u.push(T),T.shapeEnv.bindInputShapes(w),g(T,w)));x=S,u.push(x),x.shapeEnv.bindInputShapes(w)}return g(x,w)}l(y,"compiledForward");let _=null;if(e){let w=[];if(i)for(let x of i)w.push(f(h(x),null));if(w.push(d(e)),w.some(x=>Cu(x)))_=Promise.all(w).then(x=>{for(let S of x)u.push(S)});else for(let x of w)Cu(x)||u.push(x)}let b=y;return b.original=n,b.graph=w=>kC((...S)=>n.forward(...S),w||e,{name:n.constructor.name||"compiled",dynamicShapes:o}),b.source=()=>{if(u.length===0)return null;let w=u[0],x=w.result.listKernels();return x.length>0?w.result.getSource(x[0]):null},b.kernels=()=>u.length===0?[]:u[0].result.listKernels(),b.snippet=()=>{if(u.length===0)return null;let w=u[0],x=w.result.listKernels();return x.length>0?w.result.getSnippet(x[0]):null},b.result=()=>u.length>0?u[0].result:null,b._ready=_,b}var vC,mk=M(()=>{"use strict";_u();rC();An();o0();aw();yo();wn();dy();rs();ss();oC();bw();vC=!1;l(Cu,"_isThenable");l(Jj,"_ensureTracing");l(Zj,"_normalizeDynamicShapes");l(ww,"_traceCore");l(kC,"trace");l(Qj,"_prepareExecution");l(gk,"_wrapOutputs");l(Xj,"executeCompiled");l(eU,"compile")});ys();An();wn();Cy();sr();di();Xa();qn();ss();rs();function nc(n,e){let t=En(n),r=Wn.allocate(0,e,Bl),s=new Pt(r,0,n,t,e,Bl);return new Ze(s)}l(nc,"_metaTensor");function aV(n,e,t){let r=Rl(e.shape,t.shape);if(!r)throw new Error(`Incompatible shapes: [${e.shape}] vs [${t.shape}]`);let s=_g(e.dtype,t.dtype);return nc(r,s)}l(aV,"_metaBinary");function uV(n,e){return nc([...e.shape],e.dtype)}l(uV,"_metaUnary");function lV(n,e,t,r){let s=[...e.shape],o=t!=null?Array.isArray(t)?t:[t]:Array.from({length:s.length},(u,p)=>p);if(o.length===s.length||t===void 0)return nc(r?s.map(()=>1):[],e.dtype);let i=[],a=new Set(o.map(u=>u<0?s.length+u:u));for(let u=0;u<s.length;u++)a.has(u)?r&&i.push(1):i.push(s[u]);return nc(i,e.dtype)}l(lV,"_metaReduction");function cV(n,e,t){let r=xg(e.shape,t.shape);if(r===null)throw new Error("metaMatmul: unsupported shapes");return nc(r,e.dtype)}l(cV,"metaMatmul");function pV(n,e){return nc([...e.shape],e.dtype)}l(pV,"metaClone");function fS(n,e){return Object.fromEntries(n.map(t=>[t,e]))}l(fS,"mapKernel");var FM=Object.freeze({...fS(["add","sub","mul","div","pow","rem","maximum","minimum"],aV),...fS(["neg","exp","log","sqrt","rsqrt","abs","sin","cos","tanh","erf","erfc","lgamma","gamma","sigmoid","relu","gelu","silu","sign","floor","ceil"],uV),...fS(["sum","mean","max","min","prod"],lV),matmul:cV,clone:pV});GT();be();kt();os();ss();function fV(n,e,t,r){let s=Ve(n.shape,{dtype:n.dtype}),o=nu(s,e,t,r),i=nu(s,e,t,tn(r.shape,{dtype:n.dtype}));return or(Il(i,Ve(i.shape,{dtype:i.dtype})),o,n)}l(fV,"scatterKernel");function dV(n,e){let t=n.shape,r=t.length;if(e.length<r)throw new Error("repeat: reps length must be >= tensor rank");let s=e.length-r,o=s>0?[...Array(s).fill(1),...t]:t.slice(),i=[],a=[],u=[];for(let p=0;p<o.length;p++)i.push(1,o[p]),a.push(e[p],o[p]),u.push(e[p]*o[p]);return Hn(Ll(Hn(n,i),a),u)}l(dV,"repeatKernel");function hV(n,e){let t=n.shape.length,r=e.length<t?[...Array(t-e.length).fill(1),...e]:e;return Dl(n,r)}l(hV,"tileKernel");function mV(n,e,t){let r=[],s=n.shape.length,o=t<0?s+t:t,i=0;for(let a of e)r.push(ir(n,o,i,a)),i+=a;return r}l(mV,"splitKernel");function gV(n,e,t){let r=n.shape.length,s=t<0?r+t:t,o=n.shape[s],i=Math.ceil(o/e);return hi(n,i,s)}l(gV,"chunkKernel");function yV(n,e,t=0){let r=n.shape.length,s=Ks(t,r),o=n.shape[s],i=(e%o+o)%o;return i===0?n:ru([ir(n,s,o-i,i),ir(n,s,0,o-i)],s)}l(yV,"rollKernel");function _V(n,e){let t=Array.isArray(e)?e:[e],r=n.shape.length,s=n;for(let o of t){let i=Ks(o,r),a=s.shape[i],u=new Array(a);for(let p=0;p<a;p++)u[p]=a-1-p;s=ra(s,i,Wt(u,{dtype:"i32"}))}return s}l(_V,"flipKernel");function bV(n,e=0){let t=n.shape.length,r=Ks(e,t),s=n.shape[r],o=n;for(let i=1;i<s;i*=2){let a=[...o.shape];a[r]=i;let u=ru([Ve(a,{dtype:o.dtype}),ir(o,r,0,s-i)],r);o=ae(o,u)}return o}l(bV,"cumsumKernel");function wV(n){let e=1;for(;e<n;)e<<=1;return e}l(wV,"nextPow2");function $y(n,e,t){let r=n.shape.length,s=r-1,o=n.shape[s],i=wV(o),a=e?-1/0:1/0,u=n;if(i>o){let f=new Array(r).fill(0),d=new Array(r).fill(0);d[s]=i-o,u=tu(n,f,d,a)}let p=new Array(r).fill(1);p[s]=i;let c=null;if(t){let f=new Array(i);for(let d=0;d<i;d++)f[d]=d;c=Hn(Wt(f,{dtype:"i32"}),p)}for(let f=2;f<=i;f<<=1)for(let d=f>>1;d>=1;d>>=1){let h=new Array(i),m=new Array(i);for(let x=0;x<i;x++){let S=x^d;h[x]=S;let T=(x&f)===0,v=e?!T:T;S>x?m[x]=v?1:0:m[x]=v?0:1}let g=Wt(h,{dtype:"i32"}),y=ra(u,s,g),_=vg(u,y),b=ta(u,y),w=Hn(Wt(m,{dtype:"f32"}),p);if(t){let x=ra(c,s,g),S=or(w,Yp(_,u),Yp(b,u));c=or(S,c,x)}u=or(w,_,b)}return i>o&&(u=ir(u,s,0,o),t&&(c=ir(c,s,0,o))),t?{values:u,indices:c}:u}l($y,"bitonicLastDim");function TV(n,e=-1,t=!1){let r=n.shape.length,s=Ks(e,r);if(s===r-1)return $y(n,t,!1);let o=vt(n,s,r-1),i=$y(o,t,!1);return vt(i,s,r-1)}l(TV,"sortKernel");function zM(n,e,t){let r=n.shape.length,s=Ks(e,r);if(s===r-1)return $y(n,t,!0);let o=vt(n,s,r-1),i=$y(o,t,!0);return{values:vt(i.values,s,r-1),indices:vt(i.indices,s,r-1)}}l(zM,"sortWithIndices");function xV(n,e=-1,t=!1){return zM(n,e,t).indices}l(xV,"argsortKernel");function SV(n,e,t=-1,r=!0){let s=n.shape.length,o=Ks(t,s),{values:i,indices:a}=zM(n,o,r);return[ir(i,o,0,e),ir(a,o,0,e)]}l(SV,"topkKernel");var VM=Object.freeze({scatter:l((n,e,t,r,s)=>fV(e,t,r,s),"scatter"),repeat:l((n,e,t)=>dV(e,t),"repeat"),tile:l((n,e,t)=>hV(e,t),"tile"),split:l((n,e,t,r)=>mV(e,t,r),"split"),chunk:l((n,e,t,r)=>gV(e,t,r),"chunk"),roll:l((n,e,t,r)=>yV(e,t,r),"roll"),flip:l((n,e,t)=>_V(e,t),"flip"),cumsum:l((n,e,t)=>bV(e,t),"cumsum"),sort:l((n,e,t,r)=>TV(e,t,r),"sort"),argsort:l((n,e,t,r)=>xV(e,t,r),"argsort"),topk:l((n,e,t,r,s)=>SV(e,t,r,s),"topk")});var vV=3,GM=!1,kV=[0,1,2,3,5];function jM(){if(GM)return;GM=!0;let n=new Vt("mlc","DEF");for(let t of BM())n.def(t);let e=new Vt("mlc","IMPL");for(let[t,r]of Object.entries(FM))e.impl(t,vV,r);for(let[t,r]of Object.entries({...BB,...VM}))for(let s of kV)e.impl(t,s,r);pS()}l(jM,"registerNativeOps");ys();An();ty();os();wn();qn();var NV=new Set(["cpu","wasm"]);function dS(n){let e=new Float64Array(n.length);for(let t=0;t<n.length;t++)e[t]=Number(n[t]);return e}l(dS,"float64From");function UM(n){if(!NV.has(n.device.type))throw new Error(`linalg/ml: host-readable device required (cpu or wasm), got '${n.device.type}'`)}l(UM,"requireHost");function Ye(n){if(n.ndim!==2)throw new Error(`linalg/ml: expected a 2-D matrix, got ${n.ndim}-D`);UM(n);let[e,t]=n.shape;return{data:dS(Qe(n)),rows:e,cols:t}}l(Ye,"hostMatrix");function $t(n){if(UM(n),n.ndim===1)return{data:dS(Qe(n)),rows:n.shape[0],cols:1,wasVector:!0};if(n.ndim===2)return{data:dS(Qe(n)),rows:n.shape[0],cols:n.shape[1],wasVector:!1};throw new Error(`linalg/ml: expected a 1-D or 2-D right-hand side, got ${n.ndim}-D`)}l($t,"hostColumns");function Q(n,e,t,r=Ws){return Wt(n,{shape:e,dtype:t,device:r})}l(Q,"toHostTensor");os();function qM(n,e){let t=Float64Array.from(n),r=new Int32Array(e);for(let o=0;o<e;o++)r[o]=o;let s=1;for(let o=0;o<e;o++){let i=o,a=Math.abs(t[o*e+o]);for(let p=o+1;p<e;p++){let c=Math.abs(t[p*e+o]);c>a&&(a=c,i=p)}if(a===0)throw new Error("linalg: matrix is singular");if(i!==o){for(let c=0;c<e;c++){let f=t[o*e+c];t[o*e+c]=t[i*e+c],t[i*e+c]=f}let p=r[o];r[o]=r[i],r[i]=p,s=-s}let u=t[o*e+o];for(let p=o+1;p<e;p++){let c=t[p*e+o]/u;t[p*e+o]=c;for(let f=o+1;f<e;f++)t[p*e+f]-=c*t[o*e+f]}}return{lu:t,piv:r,sign:s}}l(qM,"luFactor");function AV(n,e,t,r,s){let o=new Float64Array(t*s);for(let i=0;i<t;i++)for(let a=0;a<s;a++)o[i*s+a]=r[e[i]*s+a];for(let i=0;i<t;i++)for(let a=0;a<s;a++){let u=o[i*s+a];for(let p=0;p<i;p++)u-=n[i*t+p]*o[p*s+a];o[i*s+a]=u}for(let i=t-1;i>=0;i--){let a=n[i*t+i];for(let u=0;u<s;u++){let p=o[i*s+u];for(let c=i+1;c<t;c++)p-=n[i*t+c]*o[c*s+u];o[i*s+u]=p/a}}return o}l(AV,"luSolve");function hS(n,e,t,r){let{lu:s,piv:o}=qM(n,e);return AV(s,o,e,t,r)}l(hS,"solveHost");function WM(n,e){let t;try{t=qM(n,e)}catch{return 0}let r=t.sign;for(let s=0;s<e;s++)r*=t.lu[s*e+s];return r}l(WM,"detHost");function KM(n,e){let t=new Float64Array(e*e);for(let r=0;r<e;r++)for(let s=0;s<=r;s++){let o=n[r*e+s];for(let i=0;i<s;i++)o-=t[r*e+i]*t[s*e+i];if(r===s){if(o<=0)throw new Error("linalg.cholesky: matrix is not positive definite");t[r*e+s]=Math.sqrt(o)}else t[r*e+s]=o/t[s*e+s]}return t}l(KM,"choleskyHost");ef();function vf(n,e,t){let r=t?.tol??1e-12,s=t?.maxSweeps??100,o=Float64Array.from(n),i=new Float64Array(e*e);for(let c=0;c<e;c++)i[c*e+c]=1;for(let c=0;c<s;c++){let f=0;for(let d=0;d<e;d++)for(let h=d+1;h<e;h++)f+=o[d*e+h]*o[d*e+h];if(Math.sqrt(f)<r)break;for(let d=0;d<e;d++)for(let h=d+1;h<e;h++){let m=o[d*e+h];if(Math.abs(m)<1e-300)continue;let g=(o[h*e+h]-o[d*e+d])/(2*m),y=Math.sign(g||1)/(Math.abs(g)+Math.sqrt(g*g+1)),_=1/Math.sqrt(y*y+1),b=y*_;for(let w=0;w<e;w++){let x=o[w*e+d],S=o[w*e+h];o[w*e+d]=_*x-b*S,o[w*e+h]=b*x+_*S}for(let w=0;w<e;w++){let x=o[d*e+w],S=o[h*e+w];o[d*e+w]=_*x-b*S,o[h*e+w]=b*x+_*S}for(let w=0;w<e;w++){let x=i[w*e+d],S=i[w*e+h];i[w*e+d]=_*x-b*S,i[w*e+h]=b*x+_*S}}}let a=Array.from({length:e},(c,f)=>f).sort((c,f)=>o[c*e+c]-o[f*e+f]),u=new Float64Array(e),p=new Float64Array(e*e);for(let c=0;c<e;c++){let f=a[c];u[c]=o[f*e+f];for(let d=0;d<e;d++)p[d*e+c]=i[d*e+f]}return{values:u,vectors:p}}l(vf,"eighHost");ef();function RV(n,e,t,r){if(r){let o=new Float64Array(t*t);for(let i=0;i<t;i++)for(let a=i;a<t;a++){let u=0;for(let p=0;p<e;p++)u+=n[p*t+i]*n[p*t+a];o[i*t+a]=u,o[a*t+i]=u}return o}let s=new Float64Array(e*e);for(let o=0;o<e;o++)for(let i=o;i<e;i++){let a=0;for(let u=0;u<t;u++)a+=n[o*t+u]*n[i*t+u];s[o*e+i]=a,s[i*e+o]=a}return s}l(RV,"gram");function MV(n,e,t,r,s){let o=new Float64Array(e*s);for(let i=0;i<e;i++)for(let a=0;a<s;a++){let u=0;for(let p=0;p<t;p++)u+=n[i*t+p]*r[a*t+p];o[i*s+a]=u}return o}l(MV,"matmulRows");function mu(n,e,t,r,s=RV,o=MV){let i=r?.tol??1e-12,a=Math.min(e,t),u=new Float64Array(e*a),p=new Float64Array(a),c=new Float64Array(t*a);if(t<=e){let{values:f,vectors:d}=vf(s(n,e,t,!0),t,r),h=new Float64Array(a*t);for(let g=0;g<a;g++){let y=t-1-g;p[g]=Math.sqrt(Math.max(f[y],0));for(let _=0;_<t;_++){let b=d[_*t+y];c[_*a+g]=b,h[g*t+_]=b}}let m=o(n,e,t,h,a);for(let g=0;g<a;g++)if(p[g]>i)for(let y=0;y<e;y++)u[y*a+g]=m[y*a+g]/p[g]}else{let{values:f,vectors:d}=vf(s(n,e,t,!1),e,r),h=new Float64Array(a*e);for(let y=0;y<a;y++){let _=e-1-y;p[y]=Math.sqrt(Math.max(f[_],0));for(let b=0;b<e;b++){let w=d[b*e+_];u[b*a+y]=w,h[y*e+b]=w}}let m=new Float64Array(t*e);for(let y=0;y<e;y++)for(let _=0;_<t;_++)m[_*e+y]=n[y*t+_];let g=o(m,t,e,h,a);for(let y=0;y<a;y++)if(p[y]>i)for(let _=0;_<t;_++)c[_*a+y]=g[_*a+y]/p[y]}return{U:u,S:p,V:c,k:a}}l(mu,"svdHost");ef();function HM(n,e,t,r,s,o){let i=o?.rcond??1e-12,{U:a,S:u,V:p,k:c}=mu(n,e,t,o),f=i*(u.length?u[0]:0),d=new Float64Array(c*s);for(let m=0;m<c;m++){let g=u[m];for(let y=0;y<s;y++){let _=0;for(let b=0;b<e;b++)_+=a[b*c+m]*r[b*s+y];d[m*s+y]=g>f?_/g:0}}let h=new Float64Array(t*s);for(let m=0;m<t;m++)for(let g=0;g<s;g++){let y=0;for(let _=0;_<c;_++)y+=p[m*c+_]*d[_*s+g];h[m*s+g]=y}return h}l(HM,"lstsqHost");function YM(n,e,t,r){let s=r?.rcond??1e-12,{U:o,S:i,V:a,k:u}=mu(n,e,t,r),p=s*(i.length?i[0]:0),c=new Float64Array(t*e);for(let f=0;f<t;f++)for(let d=0;d<e;d++){let h=0;for(let m=0;m<u;m++){let g=i[m];g>p&&(h+=a[f*u+m]*(1/g)*o[d*u+m])}c[f*e+d]=h}return c}l(YM,"pinvHost");function JM(n,e,t){let r=new Float64Array(t);for(let i=0;i<t;i++){let a=0;for(let u=0;u<e;u++)a+=n[u*t+i];r[i]=a/e}let s=e>1?e-1:1,o=new Float64Array(t*t);for(let i=0;i<t;i++)for(let a=i;a<t;a++){let u=0;for(let c=0;c<e;c++)u+=(n[c*t+i]-r[i])*(n[c*t+a]-r[a]);let p=u/s;o[i*t+a]=p,o[a*t+i]=p}return o}l(JM,"covHost");function kf(n,e){let t=Ye(n);if(t.rows!==t.cols)throw new Error(`linalg.${e}: matrix must be square`);return t}l(kf,"squareMatrix");function ZM(n,e){let{data:t,rows:r}=kf(e,"cholesky");return Q(KM(t,r),[r,r],e.dtype,e.device)}l(ZM,"cpuCholesky");function QM(n,e,t){let{data:r,rows:s}=kf(e,"solve"),o=$t(t);if(o.rows!==s)throw new Error("linalg.solve: right-hand side rows must match matrix");let i=hS(r,s,o.data,o.cols);return Q(i,o.wasVector?[s]:[s,o.cols],e.dtype,e.device)}l(QM,"cpuSolve");function XM(n,e,t){let{data:r,rows:s,cols:o}=Ye(e),i=$t(t);if(i.rows!==s)throw new Error("linalg.lstsq: right-hand side rows must match matrix");let a=HM(r,s,o,i.data,i.cols);return Q(a,i.wasVector?[o]:[o,i.cols],e.dtype,e.device)}l(XM,"cpuLstsq");function eI(n,e){let{data:t,rows:r}=kf(e,"inv"),s=new Float64Array(r*r);for(let i=0;i<r;i++)s[i*r+i]=1;let o=hS(t,r,s,r);return Q(o,[r,r],e.dtype,e.device)}l(eI,"cpuInv");function tI(n,e){let{data:t,rows:r,cols:s}=Ye(e);return Q(YM(t,r,s),[s,r],e.dtype,e.device)}l(tI,"cpuPinv");function nI(n,e){let{data:t,rows:r}=kf(e,"det");return Kp(WM(t,r),{dtype:e.dtype})}l(nI,"cpuDet");function rI(n,e){let{data:t,rows:r,cols:s}=Ye(e);return Q(JM(t,r,s),[s,s],e.dtype,e.device)}l(rI,"cpuCov");function sI(n,e){let{data:t,rows:r}=kf(e,"eigh"),{values:s,vectors:o}=vf(t,r);return[Q(s,[r],e.dtype,e.device),Q(o,[r,r],e.dtype,e.device)]}l(sI,"cpuEigh");function Fy(n,e){let{data:t,rows:r,cols:s}=Ye(e),{U:o,S:i,V:a,k:u}=mu(t,r,s);return[Q(o,[r,u],e.dtype,e.device),Q(i,[u],e.dtype,e.device),Q(a,[s,u],e.dtype,e.device)]}l(Fy,"cpuSvd");var IV={svd:Fy,eigh:sI,cholesky:ZM,inv:eI,pinv:tI,det:nI,cov:rI,solve:QM,lstsq:XM},EV=[0,2],oI=!1;function iI(){if(oI)return;oI=!0,$l();let n=new Vt("mlc","IMPL");for(let[e,t]of Object.entries(IV))for(let r of EV)n.impl(e,r,t)}l(iI,"registerCpuLinalg");ys();An();ys();var LV=["kmeans(Tensor x, int n_clusters, int max_iter, int n_init, int seed) -> (Tensor, Tensor, Tensor)","kmeans_predict(Tensor x, Tensor centers) -> Tensor","knn_predict(Tensor x_train, Tensor y_train, Tensor x_query, int n_neighbors, bool classify) -> Tensor","gaussian_nb_fit(Tensor x, Tensor y) -> (Tensor, Tensor, Tensor, Tensor)","gaussian_nb_predict(Tensor x, Tensor means, Tensor variances, Tensor priors, Tensor classes) -> Tensor","elastic_net(Tensor x, Tensor y, float alpha, float l1_ratio, int max_iter, float tol, bool fit_intercept) -> (Tensor, Tensor)","decision_tree_fit(Tensor x, Tensor y, int max_depth, int min_split, int min_leaf, int max_features, bool classify, int seed) -> (Tensor, Tensor, Tensor, Tensor, Tensor)","decision_tree_predict(Tensor x, Tensor feature, Tensor threshold, Tensor left, Tensor right, Tensor value) -> Tensor"],aI=!1;function zy(){if(aI)return;aI=!0;let n=new Vt("mlc","DEF");for(let e of LV)n.def(e)}l(zy,"ensureMlSchemas");function xr(n){let e=(n??2654435769)>>>0;return l(function(){e|=0,e=e+1831565813|0;let r=Math.imul(e^e>>>15,1|e);return r=r+Math.imul(r^r>>>7,61|r)^r,((r^r>>>14)>>>0)/4294967296},"next")}l(xr,"makeRng");function rc(n,e){return Math.floor(n()*e)}l(rc,"randInt");function gS(n,e){let t=Array.from({length:n},(r,s)=>s);for(let r=n-1;r>0;r--){let s=rc(e,r+1),o=t[r];t[r]=t[s],t[s]=o}return t}l(gS,"shuffledIndices");function yS(n,e,t,r,s){let o=0;for(let i=0;i<s;i++){let a=n[e*s+i]-t[r*s+i];o+=a*a}return o}l(yS,"sqDist");function _S(n,e,t,r,s){let o=new Float64Array(r*t),i=rc(s,e);for(let u=0;u<t;u++)o[u]=n[i*t+u];let a=new Float64Array(e).fill(1/0);for(let u=1;u<r;u++){let p=0;for(let d=0;d<e;d++){let h=yS(n,d,o,u-1,t);h<a[d]&&(a[d]=h),p+=a[d]}let c=s()*p,f=e-1;for(let d=0;d<e;d++)if(c-=a[d],c<=0){f=d;break}for(let d=0;d<t;d++)o[u*t+d]=n[f*t+d]}return o}l(_S,"kmeansPlusPlus");function uI(n,e,t,r,s,o){let i=0;for(let a=0;a<e;a++){let u=0,p=1/0;for(let c=0;c<s;c++){let f=yS(n,a,r,c,t);f<p&&(p=f,u=c)}o[a]=u,i+=p}return i}l(uI,"assign");function bS(n,e,t,r,s,o){let i=new Float64Array(r*t),a=new Int32Array(r);for(let p=0;p<e;p++){let c=s[p];a[c]++;for(let f=0;f<t;f++)i[c*t+f]+=n[p*t+f]}let u=!1;for(let p=0;p<r;p++)if(a[p]!==0)for(let c=0;c<t;c++){let f=i[p*t+c]/a[p];f!==o[p*t+c]&&(u=!0),o[p*t+c]=f}return u}l(bS,"updateCenters");function CV(n,e,t,r,s,o){let i=_S(n,e,t,r,o),a=new Int32Array(e),u=1/0;for(let p=0;p<s&&(u=uI(n,e,t,i,r,a),!!bS(n,e,t,r,a,i));p++);return u=uI(n,e,t,i,r,a),{centers:i,labels:a,inertia:u}}l(CV,"lloyd");function Vy(n,e,t,r,s,o){let{data:i,rows:a,cols:u}=Ye(e),p=null;for(let f=0;f<s;f++){let d=xr(o+f*2654435769),h=CV(i,a,u,t,r,d);(!p||h.inertia<p.inertia)&&(p=h)}let c=new Float64Array(a);for(let f=0;f<a;f++)c[f]=p.labels[f];return[Q(p.centers,[t,u],e.dtype,e.device),Q(c,[a],e.dtype,e.device),Q(new Float64Array([p.inertia]),[1],e.dtype,e.device)]}l(Vy,"cpuKmeans");function Gy(n,e,t){let{data:r,rows:s,cols:o}=Ye(e),i=Ye(t),a=new Float64Array(s);for(let u=0;u<s;u++){let p=0,c=1/0;for(let f=0;f<i.rows;f++){let d=yS(r,u,i.data,f,o);d<c&&(c=d,p=f)}a[u]=p}return Q(a,[s],e.dtype,e.device)}l(Gy,"cpuKmeansPredict");function wS(n,e,t,r){for(let s=0;s<t;s++)e[s]=s;for(let s=0;s<r;s++){let o=s;for(let a=s+1;a<t;a++)n[e[a]]<n[e[o]]&&(o=a);let i=e[s];e[s]=e[o],e[o]=i}}l(wS,"topKSelect");function TS(n,e,t,r){if(r){let o=new Map,i=n[e[0]],a=0;for(let u=0;u<t;u++){let p=n[e[u]],c=(o.get(p)||0)+1;o.set(p,c),c>a&&(a=c,i=p)}return i}let s=0;for(let o=0;o<t;o++)s+=n[e[o]];return s/t}l(TS,"knnVote");function jy(n,e,t,r,s,o){let i=Ye(e),a=Ye(r),u=$t(t),p=i.cols,c=Math.min(s,i.rows),f=new Float64Array(a.rows),d=new Float64Array(i.rows),h=new Int32Array(i.rows);for(let m=0;m<a.rows;m++){for(let g=0;g<i.rows;g++){let y=0;for(let _=0;_<p;_++){let b=a.data[m*p+_]-i.data[g*p+_];y+=b*b}d[g]=y}wS(d,h,i.rows,c),f[m]=TS(u.data,h,c,o)}return Q(f,[a.rows],r.dtype,r.device)}l(jy,"cpuKnnPredict");be();os();function ks(n,e){let t=Wt(Int32Array.from(e),{shape:[e.length],dtype:"i32",device:n.device});return ct("index_select",n,t,0)}l(ks,"takeRows");function sc(n){return Ye(n)}l(sc,"matrixOf");function Yt(n){let e=$t(n);if(!e.wasVector&&e.cols!==1)throw new Error("ml: expected a 1-D target or single-column matrix");return{data:e.data,n:e.rows}}l(Yt,"vectorOf");function Nf(n,e,t,r){return Q(n,[e,t],r??"f32")}l(Nf,"matrix");function xS(n,e,t){return Q(n,[e],t??"f32")}l(xS,"vector");function Ns(n,e){let t=[],r=new Map,s=new Int32Array(e);for(let o=0;o<e;o++){let i=n[o],a=r.get(i);a===void 0&&(a=t.length,r.set(i,a),t.push(i)),s[o]=a}return{y:s,classes:t}}l(Ns,"encodeLabels");var DV=1e-9;function lI(n,e,t){let{data:r,rows:s,cols:o}=Ye(e),i=$t(t),{y:a,classes:u}=Ns(i.data,i.rows),p=u.length,c=new Float64Array(p*o),f=new Float64Array(p*o),d=new Float64Array(p),h=new Int32Array(p);for(let _=0;_<s;_++){let b=a[_];h[b]++;for(let w=0;w<o;w++)c[b*o+w]+=r[_*o+w]}for(let _=0;_<p;_++){for(let b=0;b<o;b++)c[_*o+b]/=h[_];d[_]=h[_]/s}for(let _=0;_<s;_++){let b=a[_];for(let w=0;w<o;w++){let x=r[_*o+w]-c[b*o+w];f[b*o+w]+=x*x}}let m=0;for(let _=0;_<p;_++)for(let b=0;b<o;b++)f[_*o+b]/=h[_],f[_*o+b]>m&&(m=f[_*o+b]);let g=DV*m;for(let _=0;_<p*o;_++)f[_]+=g;let y=new Float64Array(p);for(let _=0;_<p;_++)y[_]=u[_];return[Q(c,[p,o],e.dtype,e.device),Q(f,[p,o],e.dtype,e.device),Q(d,[p],e.dtype,e.device),Q(y,[p],e.dtype,e.device)]}l(lI,"cpuGaussianNbFit");function cI(n,e,t,r,s,o){let{data:i,rows:a,cols:u}=Ye(e),p=Ye(t),c=Ye(r),f=$t(s),d=$t(o),h=p.rows,m=new Float64Array(a);for(let g=0;g<a;g++){let y=0,_=-1/0;for(let b=0;b<h;b++){let w=Math.log(f.data[b]);for(let x=0;x<u;x++){let S=c.data[b*u+x],T=i[g*u+x]-p.data[b*u+x];w+=-.5*(Math.log(2*Math.PI*S)+T*T/S)}w>_&&(_=w,y=b)}m[g]=d.data[y]}return Q(m,[a],e.dtype,e.device)}l(cI,"cpuGaussianNbPredict");function PV(n,e){return n>e?n-e:n<-e?n+e:0}l(PV,"softThreshold");function SS(n,e,t,r,s){let o=new Float64Array(t),i=0;if(s){for(let c=0;c<t;c++){let f=0;for(let d=0;d<e;d++)f+=n[d*t+c];o[c]=f/e}for(let c=0;c<e;c++)i+=r[c];i/=e}let a=new Float64Array(e*t);for(let c=0;c<e;c++)for(let f=0;f<t;f++)a[c*t+f]=n[c*t+f]-o[f];let u=new Float64Array(e);for(let c=0;c<e;c++)u[c]=r[c]-i;let p=new Float64Array(t);for(let c=0;c<t;c++){let f=0;for(let d=0;d<e;d++)f+=a[d*t+c]*a[d*t+c];p[c]=f/e}return{xc:a,r:u,z:p,meanX:o,meanY:i}}l(SS,"elasticNetPrep");function vS(n,e,t,r){let s=e;for(let o=0;o<r;o++)s-=n[o]*t[o];return s}l(vS,"elasticNetIntercept");function $V(n,e,t,r,s,o,i,a,u,p){for(let c=0;c<u;c++){let f=0;for(let d=0;d<t;d++){if(s[d]===0)continue;let h=0;for(let _=0;_<e;_++)h+=n[_*t+d]*r[_];let m=h/e+o[d]*s[d],g=PV(m,i)/(s[d]+a),y=g-o[d];if(y!==0){for(let _=0;_<e;_++)r[_]-=y*n[_*t+d];o[d]=g,Math.abs(y)>f&&(f=Math.abs(y))}}if(f<p)break}}l($V,"coordinateDescent");function Uy(n,e,t,r,s,o,i,a){let{data:u,rows:p,cols:c}=Ye(e),f=$t(t),{xc:d,r:h,z:m,meanX:g,meanY:y}=SS(u,p,c,f.data,a),_=new Float64Array(c),b=r*s,w=r*(1-s);$V(d,p,c,h,m,_,b,w,o,i);let x=vS(g,y,_,c);return[Q(_,[c],e.dtype,e.device),Q(new Float64Array([x]),[1],e.dtype,e.device)]}l(Uy,"cpuElasticNet");function FV(n,e,t){if(e<=0||e>=n)return Array.from({length:n},(s,o)=>o);let r=Array.from({length:n},(s,o)=>o);for(let s=n-1;s>n-1-e;s--){let o=Math.floor(t()*(s+1)),i=r[s];r[s]=r[o],r[o]=i}return r.slice(n-e)}l(FV,"chooseFeatures");function zV(n,e,t,r,s,o,i){let a=t.length,u=null;for(let p of r){let c=t.slice().sort((m,g)=>n[m*e+p]-n[g*e+p]),f=new Float64Array(s),d=new Float64Array(s);for(let m of c)d[o[m]]++;let h=0;for(let m=0;m<a-1;m++){let g=o[c[m]];f[g]++,d[g]--,h++;let y=a-h,_=n[c[m]*e+p],b=n[c[m+1]*e+p];if(_===b||h<i||y<i)continue;let w=1,x=1;for(let T=0;T<s;T++){let v=f[T]/h,A=d[T]/y;w-=v*v,x-=A*A}let S=(h*w+y*x)/a;(!u||S<u.impurity)&&(u={feature:p,threshold:(_+b)/2,impurity:S})}}return u}l(zV,"bestSplitClassify");function VV(n,e,t,r,s,o){let i=t.length,a=null;for(let u of r){let p=t.slice().sort((g,y)=>n[g*e+u]-n[y*e+u]),c=0,f=0;for(let g of p)c+=s[g],f+=s[g]*s[g];let d=0,h=0,m=0;for(let g=0;g<i-1;g++){let y=s[p[g]];d+=y,h+=y*y,m++;let _=i-m,b=n[p[g]*e+u],w=n[p[g+1]*e+u];if(b===w||m<o||_<o)continue;let x=h-d*d/m,S=c-d,v=f-h-S*S/_,A=x+v;(!a||A<a.impurity)&&(a={feature:u,threshold:(b+w)/2,impurity:A})}}return a}l(VV,"bestSplitRegress");function pI(n,e,t,r,s,o,i,a,u){let p=Ye(e),c=p.data,f=p.cols,d=$t(t),h=xr(u),m=null,g=null,y=0;if(a){let N=Ns(d.data,d.rows);m=N.y,g=N.classes,y=g.length}let _=[],b=[],w=[],x=[],S=[],T=l(N=>{if(a){let O=new Float64Array(y);for(let B of N)O[m[B]]++;let R=0;for(let B=1;B<y;B++)O[B]>O[R]&&(R=B);return g[R]}let k=0;for(let O of N)k+=d.data[O];return k/N.length},"leafValue"),v=l(N=>{let k=a?m:d.data,O=k[N[0]];for(let R of N)if(k[R]!==O)return!1;return!0},"isPure"),A=l((N,k)=>{let O=_.length;if(_.push(-1),b.push(0),w.push(-1),x.push(-1),S.push(0),k>=r||N.length<s||v(N))return S[O]=T(N),O;let R=FV(f,i,h),B=a?zV(c,f,N,R,y,m,o):VV(c,f,N,R,d.data,o);if(!B)return S[O]=T(N),O;let E=[],F=[];for(let C of N)c[C*f+B.feature]<=B.threshold?E.push(C):F.push(C);return _[O]=B.feature,b[O]=B.threshold,w[O]=A(E,k+1),x[O]=A(F,k+1),O},"build");return A(Array.from({length:p.rows},(N,k)=>k),0),[Q(Float64Array.from(_),[_.length],e.dtype,e.device),Q(Float64Array.from(b),[b.length],e.dtype,e.device),Q(Float64Array.from(w),[w.length],e.dtype,e.device),Q(Float64Array.from(x),[x.length],e.dtype,e.device),Q(Float64Array.from(S),[S.length],e.dtype,e.device)]}l(pI,"cpuDecisionTreeFit");function fI(n,e,t,r,s,o,i){let a=Ye(e),u=$t(t).data,p=$t(r).data,c=$t(s).data,f=$t(o).data,d=$t(i).data,h=new Float64Array(a.rows),m=a.cols;for(let g=0;g<a.rows;g++){let y=0;for(;u[y]!==-1;)a.data[g*m+u[y]]<=p[y]?y=c[y]:y=f[y];h[g]=d[y]}return Q(h,[a.rows],e.dtype,e.device)}l(fI,"cpuDecisionTreePredict");var GV={kmeans:Vy,kmeans_predict:Gy,knn_predict:jy,gaussian_nb_fit:lI,gaussian_nb_predict:cI,elastic_net:Uy,decision_tree_fit:pI,decision_tree_predict:fI},jV=[0,2],UV=[1,5];function qV(n){return()=>{throw new Error(`ml.${n}: scalar-iterative algorithm runs on CPU/WASM only; GPU/WebGPU not supported (no performance benefit)`)}}l(qV,"unsupported");var dI=!1;function hI(){if(dI)return;dI=!0,zy();let n=new Vt("mlc","IMPL");for(let[e,t]of Object.entries(GV)){for(let r of jV)n.impl(e,r,t);for(let r of UV)n.impl(e,r,qV(e))}}l(hI,"registerCpuMl");ys();An();ys();var WV=["qr(Tensor input) -> (Tensor, Tensor)","fft(Tensor input) -> Tensor","ifft(Tensor input) -> Tensor"],mI=!1;function gI(){if(mI)return;mI=!0;let n=new Vt("mlc","DEF");for(let e of WV)n.def(e)}l(gI,"ensureNumericSchemas");wn();function yI(n,e,t){let r=Math.min(e,t),s=Float64Array.from(n),o=[];for(let u=0;u<r;u++){let p=0;for(let h=u;h<e;h++)p+=s[h*t+u]*s[h*t+u];if(p=Math.sqrt(p),p===0){o.push(null);continue}let c=s[u*t+u]>=0?-p:p,f=new Float64Array(e-u);f[0]=s[u*t+u]-c;for(let h=u+1;h<e;h++)f[h-u]=s[h*t+u];let d=0;for(let h=0;h<f.length;h++)d+=f[h]*f[h];if(d===0){o.push(null);continue}o.push(f);for(let h=u;h<t;h++){let m=0;for(let y=0;y<f.length;y++)m+=f[y]*s[(u+y)*t+h];let g=2*m/d;for(let y=0;y<f.length;y++)s[(u+y)*t+h]-=g*f[y]}}let i=new Float64Array(e*r);for(let u=0;u<r;u++)i[u*r+u]=1;for(let u=r-1;u>=0;u--){let p=o[u];if(!p)continue;let c=0;for(let f=0;f<p.length;f++)c+=p[f]*p[f];for(let f=0;f<r;f++){let d=0;for(let m=0;m<p.length;m++)d+=p[m]*i[(u+m)*r+f];let h=2*d/c;for(let m=0;m<p.length;m++)i[(u+m)*r+f]-=h*p[m]}}let a=new Float64Array(r*t);for(let u=0;u<r;u++)for(let p=u;p<t;p++)a[u*t+p]=s[u*t+p];return{Q:i,R:a,k:r}}l(yI,"qrHost");function KV(n,e){let t=n.length;for(let r=1,s=0;r<t;r++){let o=t>>1;for(;s&o;o>>=1)s^=o;s^=o,r<s&&([n[r],n[s]]=[n[s],n[r]],[e[r],e[s]]=[e[s],e[r]])}}l(KV,"bitReversePermute");function qy(n,e,t){let r=n.length;KV(n,e);for(let s=2;s<=r;s<<=1){let o=(t?1:-1)*2*Math.PI/s,i=Math.cos(o),a=Math.sin(o);for(let u=0;u<r;u+=s){let p=1,c=0,f=s>>1;for(let d=0;d<f;d++){let h=n[u+d],m=e[u+d],g=n[u+d+f]*p-e[u+d+f]*c,y=n[u+d+f]*c+e[u+d+f]*p;n[u+d]=h+g,e[u+d]=m+y,n[u+d+f]=h-g,e[u+d+f]=m-y;let _=p*i-c*a;c=p*a+c*i,p=_}}}if(t)for(let s=0;s<r;s++)n[s]/=r,e[s]/=r}l(qy,"fftRadix2");function HV(n){let e=1;for(;e<n;)e<<=1;return e}l(HV,"nextPow2");function YV(n,e,t){let r=n.length,s=HV(2*r-1),o=t?1:-1,i=new Float64Array(r),a=new Float64Array(r);for(let d=0;d<r;d++){let h=d*d%(2*r),m=o*Math.PI*h/r;i[d]=Math.cos(m),a[d]=Math.sin(m)}let u=new Float64Array(s),p=new Float64Array(s);for(let d=0;d<r;d++)u[d]=n[d]*i[d]-e[d]*a[d],p[d]=n[d]*a[d]+e[d]*i[d];let c=new Float64Array(s),f=new Float64Array(s);c[0]=i[0],f[0]=-a[0];for(let d=1;d<r;d++)c[d]=i[d],f[d]=-a[d],c[s-d]=i[d],f[s-d]=-a[d];qy(u,p,!1),qy(c,f,!1);for(let d=0;d<s;d++){let h=u[d]*c[d]-p[d]*f[d];p[d]=u[d]*f[d]+p[d]*c[d],u[d]=h}qy(u,p,!0);for(let d=0;d<r;d++)n[d]=u[d]*i[d]-p[d]*a[d],e[d]=u[d]*a[d]+p[d]*i[d];if(t)for(let d=0;d<r;d++)n[d]/=r,e[d]/=r}l(YV,"bluestein");function _I(n,e,t){let r=n.length;r!==0&&((r&r-1)===0?qy(n,e,t):YV(n,e,t))}l(_I,"fftHost");function bI(n,e){let{data:t,rows:r,cols:s}=Ye(e),{Q:o,R:i,k:a}=yI(t,r,s);return[Q(o,[r,a],e.dtype,e.device),Q(i,[a,s],e.dtype,e.device)]}l(bI,"cpuQr");function JV(n){let e=Qe(n);if(n.ndim===1){let t=n.shape[0];return{re:Float64Array.from(e),im:new Float64Array(t),n:t}}if(n.ndim===2&&n.shape[1]===2){let t=n.shape[0],r=new Float64Array(t),s=new Float64Array(t);for(let o=0;o<t;o++)r[o]=e[2*o],s[o]=e[2*o+1];return{re:r,im:s,n:t}}throw new Error(`fft: expected a 1-D real signal or [n, 2] complex tensor, got shape [${n.shape}]`)}l(JV,"splitComplex");function ZV(n,e,t){let r=new Float64Array(2*t);for(let s=0;s<t;s++)r[2*s]=n[s],r[2*s+1]=e[s];return r}l(ZV,"interleave");function wI(n,e){let{re:t,im:r,n:s}=JV(n);return _I(t,r,e),Q(ZV(t,r,s),[s,2],n.dtype,n.device)}l(wI,"transform");function TI(n,e){return wI(e,!1)}l(TI,"cpuFft");function xI(n,e){return wI(e,!0)}l(xI,"cpuIfft");var QV={qr:bI,fft:TI,ifft:xI},XV=[0,2],SI=!1;function vI(){if(SI)return;SI=!0,gI();let n=new Vt("mlc","IMPL");for(let[e,t]of Object.entries(QV))for(let r of XV)n.impl(e,r,t)}l(vI,"registerCpuNumeric");ys();An();ty();Wm();var pt=8,NS=4,oc=2,eG=oc*pt,gu=Math.log2(oc),tG=65536,kI=eG,nG=64*1024*1024;function NI(n,e){return Math.ceil(n/e)*e}l(NI,"alignUp");var kS=class{static{l(this,"SimdModule")}constructor(e,t){this._module=e,this._exportName=t,this._instantiate()}_instantiate(){this._instance=new WebAssembly.Instance(this._module,{}),this._memory=this._instance.exports.memory,this._fn=this._instance.exports[this._exportName],this._top=0}reset(){this._top=0}reclaim(){this._memory.buffer.byteLength>nG?this._instantiate():this._top=0}_ensure(e){let t=this._memory.buffer.byteLength;e<=t||this._memory.grow(Math.ceil((e-t)/tG))}alloc(e){let t=NI(this._top,kI);return this._top=t+NI(e,kI),this._ensure(this._top),t}allocF64(e){return this.alloc(e*pt)}allocI32(e){return this.alloc(e*NS)}f64(e,t){return new Float64Array(this._memory.buffer,e,t)}i32(e,t){return new Int32Array(this._memory.buffer,e,t)}writeF64(e,t){this.f64(e,t.length).set(t)}run(...e){this._fn(...e)}},AI=new Map;function Po(n,e,t){let r=AI.get(n);return r||(r=new kS(new WebAssembly.Module(Sl(e())),t),AI.set(n,r)),r.reclaim(),r}l(Po,"simdModule");var rG=1e308,sG=1,oG=32768;function Af(n,e,t,r){let s=e.map(([i,a])=>`(param $${i} ${a})`).join(" "),o=t.map(([i,a])=>`(local $${i} ${a})`).join(" ");return`(module (memory (export "memory") ${sG} ${oG})
+ (func (export "${n}") ${s}
+  ${o}
+${r}))`}l(Af,"moduleWrap");function OI(n,e,t,r,s,o){return`
+    (local.get $${r}) (i32.const ${gu}) i32.shr_s (i32.const ${gu}) i32.shl local.set $${o.nmain}
+    (f64.const 0) f64x2.splat local.set $${o.acc}
+    (i32.const 0) local.set $${o.j}
+    (block $${n}_ve (loop $${n}_vl
+      (local.get $${o.j}) (local.get $${o.nmain}) i32.ge_s br_if $${n}_ve
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add v128.load local.set $${o.va}
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add v128.load local.set $${o.vb}
+      (local.get $${o.acc})
+      (local.get $${o.va}) (local.get $${o.vb}) f64x2.sub local.set $${o.va}
+      (local.get $${o.va}) (local.get $${o.va}) f64x2.mul
+      f64x2.add local.set $${o.acc}
+      (local.get $${o.j}) (i32.const ${oc}) i32.add local.set $${o.j}
+      br $${n}_vl))
+    (local.get $${o.acc}) f64x2.extract_lane 0
+    (local.get $${o.acc}) f64x2.extract_lane 1
+    f64.add local.set $${s}
+    (block $${n}_te (loop $${n}_tl
+      (local.get $${o.j}) (local.get $${r}) i32.ge_s br_if $${n}_te
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add f64.load
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add f64.load
+      f64.sub local.set $${o.diff}
+      (local.get $${s}) (local.get $${o.diff}) (local.get $${o.diff}) f64.mul f64.add local.set $${s}
+      (local.get $${o.j}) (i32.const 1) i32.add local.set $${o.j}
+      br $${n}_tl))`}l(OI,"sqDistSnippet");var BI={j:"j",nmain:"nmain",acc:"acc",va:"va",vb:"vb",diff:"diff"},RI=[["j","i32"],["nmain","i32"],["acc","v128"],["va","v128"],["vb","v128"],["diff","f64"]];function AS(){let n=`
+    (f64.const 0) local.set $inertiaAcc
+    (i32.const 0) local.set $i
+    (block $iend (loop $iloop
+      (local.get $i) (local.get $n) i32.ge_s br_if $iend
+      (local.get $x) (local.get $i) (local.get $d) i32.mul (i32.const ${pt}) i32.mul i32.add local.set $xrow
+      (i32.const 0) local.set $best
+      (f64.const ${rG}) local.set $bestd
+      (i32.const 0) local.set $cc
+      (block $cend (loop $cloop
+        (local.get $cc) (local.get $k) i32.ge_s br_if $cend
+        (local.get $c) (local.get $cc) (local.get $d) i32.mul (i32.const ${pt}) i32.mul i32.add local.set $crow
+${OI("nc","xrow","crow","d","dist",BI)}
+        (local.get $dist) (local.get $bestd) f64.lt
+        (if (then
+          (local.get $dist) local.set $bestd
+          (local.get $cc) local.set $best))
+        (local.get $cc) (i32.const 1) i32.add local.set $cc
+        br $cloop))
+      (local.get $labels) (local.get $i) (i32.const ${NS}) i32.mul i32.add (local.get $best) i32.store
+      (local.get $inertiaAcc) (local.get $bestd) f64.add local.set $inertiaAcc
+      (local.get $i) (i32.const 1) i32.add local.set $i
+      br $iloop))
+    (local.get $inertia) (local.get $inertiaAcc) f64.store`;return Af("nearest_centroid",[["x","i32"],["n","i32"],["d","i32"],["c","i32"],["k","i32"],["labels","i32"],["inertia","i32"]],[["i","i32"],["cc","i32"],["best","i32"],["bestd","f64"],["dist","f64"],["xrow","i32"],["crow","i32"],["inertiaAcc","f64"],...RI],n)}l(AS,"nearestCentroidWat");function MI(){let n=`
+    (i32.const 0) local.set $t
+    (block $tend (loop $tloop
+      (local.get $t) (local.get $ntr) i32.ge_s br_if $tend
+      (local.get $tr) (local.get $t) (local.get $d) i32.mul (i32.const ${pt}) i32.mul i32.add local.set $trow
+${OI("dr","q","trow","d","dist",BI)}
+      (local.get $out) (local.get $t) (i32.const ${pt}) i32.mul i32.add (local.get $dist) f64.store
+      (local.get $t) (i32.const 1) i32.add local.set $t
+      br $tloop))`;return Af("dist_row",[["q","i32"],["tr","i32"],["ntr","i32"],["d","i32"],["out","i32"]],[["t","i32"],["trow","i32"],["dist","f64"],...RI],n)}l(MI,"distRowWat");function OS(n,e,t,r,s,o){return`
+    (local.get $${r}) (i32.const ${gu}) i32.shr_s (i32.const ${gu}) i32.shl local.set $${o.nmain}
+    (f64.const 0) f64x2.splat local.set $${o.acc}
+    (i32.const 0) local.set $${o.j}
+    (block $${n}_ve (loop $${n}_vl
+      (local.get $${o.j}) (local.get $${o.nmain}) i32.ge_s br_if $${n}_ve
+      (local.get $${o.acc})
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add v128.load
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add v128.load
+      f64x2.mul f64x2.add local.set $${o.acc}
+      (local.get $${o.j}) (i32.const ${oc}) i32.add local.set $${o.j}
+      br $${n}_vl))
+    (local.get $${o.acc}) f64x2.extract_lane 0
+    (local.get $${o.acc}) f64x2.extract_lane 1
+    f64.add local.set $${s}
+    (block $${n}_te (loop $${n}_tl
+      (local.get $${o.j}) (local.get $${r}) i32.ge_s br_if $${n}_te
+      (local.get $${s})
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add f64.load
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add f64.load
+      f64.mul f64.add local.set $${s}
+      (local.get $${o.j}) (i32.const 1) i32.add local.set $${o.j}
+      br $${n}_tl))`}l(OS,"reduceDotSnippet");function iG(n,e,t,r,s,o){return`
+    (local.get $${r}) f64x2.splat local.set $${o.sv}
+    (local.get $${s}) (i32.const ${gu}) i32.shr_s (i32.const ${gu}) i32.shl local.set $${o.nmain}
+    (i32.const 0) local.set $${o.j}
+    (block $${n}_ve (loop $${n}_vl
+      (local.get $${o.j}) (local.get $${o.nmain}) i32.ge_s br_if $${n}_ve
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add local.set $${o.ya}
+      (local.get $${o.ya}) v128.load
+      (local.get $${o.sv})
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add v128.load
+      f64x2.mul f64x2.sub local.set $${o.vy}
+      (local.get $${o.ya}) (local.get $${o.vy}) v128.store
+      (local.get $${o.j}) (i32.const ${oc}) i32.add local.set $${o.j}
+      br $${n}_vl))
+    (block $${n}_te (loop $${n}_tl
+      (local.get $${o.j}) (local.get $${s}) i32.ge_s br_if $${n}_te
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add local.set $${o.ya}
+      (local.get $${o.ya})
+      (local.get $${o.ya}) f64.load
+      (local.get $${r})
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${pt}) i32.mul i32.add f64.load
+      f64.mul f64.sub f64.store
+      (local.get $${o.j}) (i32.const 1) i32.add local.set $${o.j}
+      br $${n}_tl))`}l(iG,"axpySnippet");function II(){let e=`
+    (i32.const 0) local.set $i
+    (block $ie (loop $il
+      (local.get $i) (local.get $m) i32.ge_s br_if $ie
+      (local.get $mat) (local.get $i) (local.get $len) i32.mul (i32.const ${pt}) i32.mul i32.add local.set $rowI
+      (local.get $i) local.set $j
+      (block $je (loop $jl
+        (local.get $j) (local.get $m) i32.ge_s br_if $je
+        (local.get $mat) (local.get $j) (local.get $len) i32.mul (i32.const ${pt}) i32.mul i32.add local.set $rowJ
+${OS("gs","rowI","rowJ","len","d",{j:"kk",nmain:"nm",acc:"acc"})}
+        (local.get $g) (local.get $i) (local.get $m) i32.mul (local.get $j) i32.add (i32.const ${pt}) i32.mul i32.add (local.get $d) f64.store
+        (local.get $g) (local.get $j) (local.get $m) i32.mul (local.get $i) i32.add (i32.const ${pt}) i32.mul i32.add (local.get $d) f64.store
+        (local.get $j) (i32.const 1) i32.add local.set $j
+        br $jl))
+      (local.get $i) (i32.const 1) i32.add local.set $i
+      br $il))`;return Af("gram_sym",[["mat","i32"],["m","i32"],["len","i32"],["g","i32"]],[["i","i32"],["j","i32"],["rowI","i32"],["rowJ","i32"],["d","f64"],["kk","i32"],["nm","i32"],["acc","v128"]],e)}l(II,"gramSymWat");function EI(){let e=`
+    (i32.const 0) local.set $i
+    (block $ie (loop $il
+      (local.get $i) (local.get $m) i32.ge_s br_if $ie
+      (local.get $a) (local.get $i) (local.get $len) i32.mul (i32.const ${pt}) i32.mul i32.add local.set $rowA
+      (i32.const 0) local.set $c
+      (block $ce (loop $cl
+        (local.get $c) (local.get $p) i32.ge_s br_if $ce
+        (local.get $b) (local.get $c) (local.get $len) i32.mul (i32.const ${pt}) i32.mul i32.add local.set $rowB
+${OS("mm","rowA","rowB","len","d",{j:"kk",nmain:"nm",acc:"acc"})}
+        (local.get $out) (local.get $i) (local.get $p) i32.mul (local.get $c) i32.add (i32.const ${pt}) i32.mul i32.add (local.get $d) f64.store
+        (local.get $c) (i32.const 1) i32.add local.set $c
+        br $cl))
+      (local.get $i) (i32.const 1) i32.add local.set $i
+      br $il))`;return Af("matmul_rows",[["a","i32"],["m","i32"],["len","i32"],["b","i32"],["p","i32"],["out","i32"]],[["i","i32"],["c","i32"],["rowA","i32"],["rowB","i32"],["d","f64"],["kk","i32"],["nm","i32"],["acc","v128"]],e)}l(EI,"matmulRowsWat");function LI(){let n={j:"j",nmain:"nmain",acc:"acc"},e={j:"j2",nmain:"nmain2",sv:"sv",ya:"ya",vy:"vy"},t=`
+    (local.get $n) f64.convert_i32_s local.set $nf
+    (i32.const 0) local.set $iter
+    (block $ie (loop $il
+      (local.get $iter) (local.get $maxIter) i32.ge_s br_if $ie
+      (f64.const 0) local.set $maxChange
+      (i32.const 0) local.set $jc
+      (block $je (loop $jl
+        (local.get $jc) (local.get $d) i32.ge_s br_if $je
+        (local.get $z) (local.get $jc) (i32.const ${pt}) i32.mul i32.add f64.load local.set $zj
+        (local.get $zj) (f64.const 0) f64.ne
+        (if (then
+          (local.get $xc) (local.get $jc) (local.get $n) i32.mul (i32.const ${pt}) i32.mul i32.add local.set $col
+${OS("cd","col","r","n","dot",n)}
+          (local.get $w) (local.get $jc) (i32.const ${pt}) i32.mul i32.add f64.load local.set $wold
+          (local.get $dot) (local.get $nf) f64.div (local.get $wold) (local.get $zj) f64.mul f64.add local.set $rho
+          (local.get $rho) (local.get $l1) f64.gt
+          (if (then
+            (local.get $rho) (local.get $l1) f64.sub local.set $thr
+          ) (else
+            (local.get $rho) (f64.const 0) (local.get $l1) f64.sub f64.lt
+            (if (then
+              (local.get $rho) (local.get $l1) f64.add local.set $thr
+            ) (else
+              (f64.const 0) local.set $thr))))
+          (local.get $thr) (local.get $zj) (local.get $l2) f64.add f64.div local.set $wj
+          (local.get $wj) (local.get $wold) f64.sub local.set $delta
+          (local.get $delta) (f64.const 0) f64.ne
+          (if (then
+${iG("cd","r","col","delta","n",e)}
+            (local.get $w) (local.get $jc) (i32.const ${pt}) i32.mul i32.add (local.get $wj) f64.store
+            (local.get $delta) f64.abs local.set $absd
+            (local.get $absd) (local.get $maxChange) f64.gt
+            (if (then (local.get $absd) local.set $maxChange))))))
+        (local.get $jc) (i32.const 1) i32.add local.set $jc
+        br $jl))
+      (local.get $maxChange) (local.get $tol) f64.lt br_if $ie
+      (local.get $iter) (i32.const 1) i32.add local.set $iter
+      br $il))`;return Af("coord_descent",[["xc","i32"],["n","i32"],["d","i32"],["r","i32"],["w","i32"],["z","i32"],["l1","f64"],["l2","f64"],["maxIter","i32"],["tol","f64"]],[["iter","i32"],["jc","i32"],["nf","f64"],["maxChange","f64"],["zj","f64"],["col","i32"],["dot","f64"],["wold","f64"],["rho","f64"],["thr","f64"],["wj","f64"],["delta","f64"],["absd","f64"],["j","i32"],["nmain","i32"],["acc","v128"],["j2","i32"],["nmain2","i32"],["sv","v128"],["ya","i32"],["vy","v128"]],t)}l(LI,"coordDescentWat");var aG=5e5;function CI(n,e,t){let r=Po("gram_sym",II,"gram_sym"),s=r.allocF64(e*t),o=r.allocF64(e*e);return r.writeF64(s,n),r.run(s,e,t,o),Float64Array.from(r.f64(o,e*e))}l(CI,"gramSym");function uG(n,e,t,r){if(r){let s=new Float64Array(t*e);for(let o=0;o<e;o++)for(let i=0;i<t;i++)s[i*e+o]=n[o*t+i];return CI(s,t,e)}return CI(n,e,t)}l(uG,"wasmGram");function lG(n,e,t,r,s){let o=Po("matmul_rows",EI,"matmul_rows"),i=o.allocF64(e*t),a=o.allocF64(s*t),u=o.allocF64(e*s);return o.writeF64(i,n),o.writeF64(a,r),o.run(i,e,t,a,s,u),Float64Array.from(o.f64(u,e*s))}l(lG,"wasmMatmulRows");function DI(n,e){let{data:t,rows:r,cols:s}=Ye(e);if(Math.max(r,s)*Math.min(r,s)*Math.min(r,s)<aG)return Fy(n,e);let{U:i,S:a,V:u,k:p}=mu(t,r,s,void 0,uG,lG);return[Q(i,[r,p],e.dtype,e.device),Q(a,[p],e.dtype,e.device),Q(u,[s,p],e.dtype,e.device)]}l(DI,"wasmSvd");var PI={svd:DI},cG={eigh:"Jacobi rotation sweeps use strided column access; not SIMD-friendly",cholesky:"sequential dependent updates; not vectorizable at f64x2 width",solve:"triangular back-substitution is inherently sequential",cov:"single O(n*d) pass dominated by the eigen-free reduction"},$I=!1;function FI(){let n={enabled:Object.keys(PI),skipped:Object.keys(cG)};if($I)return n;$I=!0,$l();let e=new Vt("mlc","IMPL");for(let[t,r]of Object.entries(PI))e.impl(t,2,r);return n}l(FI,"registerWasmLinalg");ys();An();var pG=8e3,fG=5e4,dG=2654435769;function zI(n,e,t,r,s,o,i,a,u,p){return n.writeF64(t,u),n.run(e,o,i,t,a,r,s),p.set(n.i32(r,o)),n.f64(s,1)[0]}l(zI,"assignWasm");function VI(n,e,t,r,s,o){let{data:i,rows:a,cols:u}=Ye(e);if(a*u*t<pG)return Vy(n,e,t,r,s,o);let p=Po("kmeans_assign",AS,"nearest_centroid"),c=null;for(let d=0;d<s;d++){p.reset();let h=p.allocF64(a*u),m=p.allocF64(t*u),g=p.allocI32(a),y=p.allocF64(1);p.writeF64(h,i);let _=xr(o+d*dG),b=_S(i,a,u,t,_),w=new Int32Array(a),x=1/0;for(let S=0;S<r&&(x=zI(p,h,m,g,y,a,u,t,b,w),!!bS(i,a,u,t,w,b));S++);x=zI(p,h,m,g,y,a,u,t,b,w),(!c||x<c.inertia)&&(c={centers:Float64Array.from(b),labels:Int32Array.from(w),inertia:x})}let f=new Float64Array(a);for(let d=0;d<a;d++)f[d]=c.labels[d];return[Q(c.centers,[t,u],e.dtype,e.device),Q(f,[a],e.dtype,e.device),Q(new Float64Array([c.inertia]),[1],e.dtype,e.device)]}l(VI,"wasmKmeans");function GI(n,e,t){let{data:r,rows:s,cols:o}=Ye(e),i=Ye(t);if(s*o*i.rows<fG)return Gy(n,e,t);let a=Po("kmeans_assign",AS,"nearest_centroid");a.reset();let u=a.allocF64(s*o),p=a.allocF64(i.rows*o),c=a.allocI32(s),f=a.allocF64(1);a.writeF64(u,r),a.writeF64(p,i.data),a.run(u,s,o,p,i.rows,c,f);let d=new Float64Array(s),h=a.i32(c,s);for(let m=0;m<s;m++)d[m]=h[m];return Q(d,[s],e.dtype,e.device)}l(GI,"wasmKmeansPredict");var hG=2e4;function jI(n,e,t,r,s,o){let i=Ye(e),a=Ye(r),u=$t(t),p=i.cols;if(a.rows*i.rows*p<hG)return jy(n,e,t,r,s,o);let c=Po("knn_dist",MI,"dist_row");c.reset();let f=c.allocF64(a.rows*p),d=c.allocF64(i.rows*p),h=c.allocF64(i.rows);c.writeF64(f,a.data),c.writeF64(d,i.data);let m=Math.min(s,i.rows),g=new Float64Array(a.rows),y=new Int32Array(i.rows),_=p*pt;for(let b=0;b<a.rows;b++){c.run(f+b*_,d,i.rows,p,h);let w=c.f64(h,i.rows);wS(w,y,i.rows,m),g[b]=TS(u.data,y,m,o)}return Q(g,[a.rows],r.dtype,r.device)}l(jI,"wasmKnnPredict");var mG=3e6;function UI(n,e,t,r,s,o,i,a){let{data:u,rows:p,cols:c}=Ye(e);if(p*c*o<mG)return Uy(n,e,t,r,s,o,i,a);let f=$t(t),{xc:d,r:h,z:m,meanX:g,meanY:y}=SS(u,p,c,f.data,a),_=new Float64Array(p*c);for(let O=0;O<c;O++)for(let R=0;R<p;R++)_[O*p+R]=d[R*c+O];let b=r*s,w=r*(1-s),x=Po("coord_descent",LI,"coord_descent");x.reset();let S=x.allocF64(p*c),T=x.allocF64(p),v=x.allocF64(c),A=x.allocF64(c);x.writeF64(S,_),x.writeF64(T,h),x.writeF64(A,m),x.f64(v,c).fill(0),x.run(S,p,c,T,v,A,b,w,o,i);let N=Float64Array.from(x.f64(v,c)),k=vS(g,y,N,c);return[Q(N,[c],e.dtype,e.device),Q(new Float64Array([k]),[1],e.dtype,e.device)]}l(UI,"wasmElasticNet");var qI={kmeans:VI,kmeans_predict:GI,knn_predict:jI,elastic_net:UI},WI=!1;function KI(){let n={enabled:Object.keys(qI)};if(WI)return n;WI=!0,zy();let e=new Vt("mlc","IMPL");for(let[t,r]of Object.entries(qI))e.impl(t,2,r);return n}l(KI,"registerWasmMl");An();sp();Ma();ou();Ao();be();kt();var Wy=class extends ve{static{l(this,"AddBackward")}constructor(){super(2)}apply(e){let t=e[0];return[t,t]}},Ky=class extends ve{static{l(this,"SubBackward")}constructor(){super(2)}apply(e){let t=e[0];return[t,is(t)]}},Hy=class extends ve{static{l(this,"MulBackward")}constructor(){super(2)}apply(e){let t=e[0],[r,s]=this.savedTensors();return[ee(t,s.detach()),ee(t,r.detach())]}},Yy=class extends ve{static{l(this,"DivBackward")}constructor(){super(2)}apply(e){let t=e[0],[r,s]=this.savedTensors(),o=Kt(t,s.detach()),i=is(Kt(ee(t,r.detach()),ee(s.detach(),s.detach())));return[o,i]}},Jy=class extends ve{static{l(this,"NegBackward")}constructor(){super(1)}apply(e){let t=e[0];return[is(t)]}},Zy=class extends ve{static{l(this,"PowBackward")}constructor(){super(2)}apply(e){let t=e[0],[r,s]=this.savedTensors(),o=tn(s.shape,{dtype:s.dtype,device:s.device}),i=ee(t,ee(s.detach(),vo(r.detach(),ze(s.detach(),o)))),a=ee(t,ee(vo(r.detach(),s.detach()),ko(r.detach())));return[i,a]}};Ao();be();kt();Yi();var gG=2/Math.sqrt(Math.PI);function HI(n){let e=ot(n.shape,gG,{dtype:n.dtype,device:n.device});return ee(e,as(is(ee(n,n))))}l(HI,"_erfDeriv");function YI(n){let e=ae(n,ii),t=vo(e,-1),r=ze(ko(e),ee(t,.5)),s=ee(t,t),o=s;for(let i of Ga)r=ae(r,ee(o,i)),o=ee(o,s);for(let i=0;i<ii;i++)r=ze(r,vo(ae(n,i),-1));return r}l(YI,"_digammaTensor");var Qy=class extends ve{static{l(this,"ErfBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[ee(e[0],HI(t.detach()))]}},Xy=class extends ve{static{l(this,"ErfcBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[is(ee(e[0],HI(t.detach())))]}},e_=class extends ve{static{l(this,"LgammaBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[ee(e[0],YI(t.detach()))]}},t_=class extends ve{static{l(this,"GammaBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),r=t.detach();return[ee(e[0],ee(kg(r),YI(r)))]}},n_=class extends ve{static{l(this,"ExpBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[ee(e[0],as(t.detach()))]}},r_=class extends ve{static{l(this,"LogBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[Kt(e[0],t.detach())]}},s_=class extends ve{static{l(this,"SqrtBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),r=t.detach(),s=na(r),o=ot(r.shape,2,{dtype:r.dtype,device:r.device});return[Kt(e[0],ee(o,s))]}},o_=class extends ve{static{l(this,"TanhBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),r=t.detach(),s=Hs(r),o=tn(r.shape,{dtype:r.dtype,device:r.device});return[ee(e[0],ze(o,ee(s,s)))]}},i_=class extends ve{static{l(this,"SigmoidBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),r=t.detach(),s=wr(r),o=tn(r.shape,{dtype:r.dtype,device:r.device});return[ee(e[0],ee(s,ze(o,s)))]}},a_=class extends ve{static{l(this,"SoftmaxBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),r=this.opArgs(),s=r&&r.length>1&&r[1]!=null?r[1]:-1,o=e[0],i=mi(t.detach(),s),a=wt(ee(o,i),s,!0);return[ee(i,ze(o,a))]}},u_=class extends ve{static{l(this,"LogSoftmaxBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),r=this.opArgs(),s=r&&r.length>1&&r[1]!=null?r[1]:-1,o=e[0],i=mi(t.detach(),s),a=wt(o,s,!0);return[ze(o,ee(i,a))]}},l_=class extends ve{static{l(this,"ReluBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),r=t.detach();return[ee(e[0],Og(Ml(r)))]}},c_=class extends ve{static{l(this,"GeluBackward")}constructor(){super(1)}apply(e){let t=e[0],[r]=this.savedTensors(),s=r.detach(),o=s.shape,i=s.dtype,a=s.device,u=ot(o,1.702,{dtype:i,device:a}),p=tn(o,{dtype:i,device:a}),c=ee(u,s),f=wr(c),d=ee(f,ae(p,ee(c,ze(p,f))));return[ee(t,d)]}},p_=class extends ve{static{l(this,"SiluBackward")}constructor(){super(1)}apply(e){let t=e[0],[r]=this.savedTensors(),s=r.detach(),o=wr(s),i=tn(s.shape,{dtype:s.dtype,device:s.device}),a=ee(o,ae(i,ee(s,ze(i,o))));return[ee(t,a)]}};Ao();be();kt();be();function JI(n,e){if(n==null){let r=[];for(let s=0;s<e;s++)r.push(s);return r}return(Array.isArray(n)?n:[n]).map(r=>r<0?r+e:r).sort((r,s)=>r-s)}l(JI,"_normalizeDims");function ZI(n,e,t,r){let s=n;if(!r)for(let i of t)s=Lr(s,i);let o=Ve(e,{dtype:s.dtype,device:s.device});return ae(o,s)}l(ZI,"_unreduce");var f_=class extends ve{static{l(this,"SumBackward")}constructor(){super(1)}apply(e){let t=e[0],s=this.inputMetadata(0).shape,o=this.opArgs(),i=o?o[1]:void 0,a=o?o[2]:!1,u=JI(i,s.length);return[ZI(t,s,u,a)]}},d_=class extends ve{static{l(this,"MeanBackward")}constructor(){super(1)}apply(e){let t=e[0],s=this.inputMetadata(0).shape,o=this.opArgs(),i=o?o[1]:void 0,a=o?o[2]:!1,u=JI(i,s.length),p=1;for(let d of u)p*=s[d];let c=ZI(t,s,u,a),f=ot(s,p,{dtype:t.dtype,device:t.device});return[Kt(c,f)]}};Ao();be();be();function QI(n,e){let t=n;for(;t.ndim>e.length;)t=wt(t,0,!1);for(let r=0;r<e.length;r++)e[r]===1&&t.shape[r]!==1&&(t=wt(t,r,!0));return t}l(QI,"_sumToShape");var h_=class extends ve{static{l(this,"MatmulBackward")}constructor(){super(2)}apply(e){let t=e[0],[r,s]=this.savedTensors(),o=r.detach(),i=s.detach(),a=o.ndim,u=i.ndim;if(a===1&&u===1)return[ee(t,i),ee(t,o)];if(a===2&&u===1){let f=Lr(t,1),d=Lr(i,0);return[tt(f,d),Cl(tt(vt(o,0,1),f),1)]}if(a===1&&u===2){let f=tt(t,vt(i,0,1)),d=tt(Lr(o,1),Lr(t,0));return[f,d]}let p=tt(t,vt(i,u-2,u-1)),c=tt(vt(o,a-2,a-1),t);return p=QI(p,o.shape),c=QI(c,i.shape),[p,c]}},m_=class extends ve{static{l(this,"DotBackward")}constructor(){super(2)}apply(e){let t=e[0],[r,s]=this.savedTensors();return[ee(t,s.detach()),ee(t,r.detach())]}};UT();Ao();be();kt();be();function yG(n,e,t){n[e]+=t}l(yG,"_addAt");var ic=class extends ve{static{l(this,"ReshapeBackward")}constructor(){super(1)}apply(e){let t=this.inputMetadata(0);return[Hn(e[0],t.shape)]}},g_=class extends ve{static{l(this,"TransposeBackward")}_dim0;_dim1;constructor(e,t){super(1),this._dim0=e,this._dim1=t}apply(e){return[vt(e[0],this._dim0,this._dim1)]}},Of=class extends ve{static{l(this,"SliceBackward")}_dim;_start;_end;_step;constructor(e,t,r,s){super(1),this._dim=e,this._start=t,this._end=r,this._step=s}apply(e){let t=e[0],r=this.inputMetadata(0),s=this._dim,o=r.shape[s],i=this._step||1,a=this._start<0?this._start+o:this._start,u=this._end<0?this._end+o:this._end;if(a=Math.max(0,Math.min(a,o)),u=Math.max(0,Math.min(u,o)),i===1){let w=r.shape.map(()=>0),x=r.shape.map(()=>0);return w[s]=a,x[s]=o-u,[tu(t,w,x,0)]}let p=Ve(r.shape,{dtype:t.dtype,device:t.device}),c=p._impl.storage.data,f=t._impl.storage.data,d=t._impl.storageOffset,h=t.shape,m=t.strides,g=p.strides,y=h.length,_=new Int32Array(y),b=d;for(let w=0;w<t.numel;w++){let x=0;for(let S=0;S<y;S++){let T=S===s?a+_[S]*i:_[S];x+=T*g[S]}yG(c,x,f[b]);for(let S=y-1;S>=0;S--){if(_[S]++,_[S]<h[S]){b+=m[S];break}b-=(h[S]-1)*m[S],_[S]=0}}return[p]}},y_=class extends ve{static{l(this,"SelectBackward")}_dim;_index;constructor(e,t){super(1),this._dim=e,this._index=t}apply(e){let t=e[0],r=this.inputMetadata(0),s=this._dim,o=r.shape[s],i=this._index<0?this._index+o:this._index,a=Lr(t,s),u=r.shape.map(()=>0),p=r.shape.map(()=>0);return u[s]=i,p[s]=o-1-i,[tu(a,u,p,0)]}},Bf=class extends ve{static{l(this,"ExpandBackward")}constructor(){super(1)}apply(e){let t=e[0],s=this.inputMetadata(0).shape,o=t.shape,i=[],a=o.length-s.length;for(let p=0;p<a;p++)i.push(p);for(let p=0;p<s.length;p++)s[p]===1&&o[p+a]!==1&&i.push(p+a);let u=t;return i.length>0&&(u=wt(t,i,!0)),[Hn(u,s)]}},__=class extends ve{static{l(this,"PermuteBackward")}_dims;constructor(e){super(1),this._dims=e}apply(e){let t=this._dims.length,r=new Array(t);for(let s=0;s<t;s++){let o=this._dims[s]<0?t+this._dims[s]:this._dims[s];r[o]=s}return[El(e[0],r)]}};var BS=new Map;function Ge(n,e){BS.set(n,e)}l(Ge,"_register");Ge("add",()=>new Wy);Ge("sub",()=>new Ky);Ge("mul",()=>new Hy);Ge("div",()=>new Yy);Ge("neg",()=>new Jy);Ge("pow",()=>new Zy);Ge("exp",()=>new n_);Ge("log",()=>new r_);Ge("sqrt",()=>new s_);Ge("tanh",()=>new o_);Ge("erf",()=>new Qy);Ge("erfc",()=>new Xy);Ge("lgamma",()=>new e_);Ge("gamma",()=>new t_);Ge("sigmoid",()=>new i_);Ge("relu",()=>new l_);Ge("softmax",()=>new a_);Ge("log_softmax",()=>new u_);Ge("gelu",()=>new c_);Ge("silu",()=>new p_);Ge("sum",()=>new f_);Ge("mean",()=>new d_);Ge("matmul",()=>new h_);Ge("dot",()=>new m_);Ge("cat",()=>new Ug);Ge("stack",()=>new qg);Ge("clamp",()=>new Wg);Ge("pad",()=>new Kg);Ge("index_select",()=>new Pl);Ge("where",()=>new Hg);Ge("reshape",()=>new ic);Ge("transpose",n=>new g_(n[1],n[2]));Ge("permute",n=>new __(n[1]));Ge("broadcast_in_dim",()=>new Bf);Ge("expand",()=>new Bf);Ge("slice",n=>new Of(n[1],n[2],n[3],n[4]));Ge("narrow",n=>new Of(n[1],n[2],n[2]+n[3],1));Ge("select",n=>new y_(n[1],n[2]));Ge("unsqueeze",()=>new ic);Ge("squeeze",()=>new ic);function XI(n,e=null){let t=BS.get(n);return t?t(e):null}l(XI,"getGradFn");function eE(n){return BS.has(n)}l(eE,"hasGradFn");Xg();Wp();di();sr();qn();fi();sr();kt();be();function tE(n,e){let t=n.gradFn;if(!t)throw new Error("Cannot call backward on a tensor that does not require grad");let r=e;if(!r){if(n.numel!==1)throw new Error("grad must be specified for non-scalar tensors");r=tn(n.shape,{dtype:n.dtype,device:n.device})}let s=new Map;bG(t,s,new Set);let i=new Map;i.set(t.id,[r]);let a=[];for(let[p,c]of s)c===0&&a.push(p);let u=[];for(;a.length>0;){let p=a.pop();if(p){u.push(p);for(let c of p.nextEdges){if(!c||!c.node)continue;let f=c.node,d=s.get(f)-1;s.set(f,d),d===0&&a.push(f)}}}for(let p of u){let c=i.get(p.id);if(!c)continue;let f=p.apply(c),d=f;if(f&&(d=f.map((m,g)=>{if(!m)return m;let y=p.inputMetadata(g);return y?_G(m,y.shape):m})),p.releaseVariables(),!d)continue;let h=p.nextEdges;for(let m=0;m<h.length;m++){let g=h[m];if(!g||!g.node||m>=d.length||!d[m])continue;let y=d[m],_=g.node.id,b=i.get(_);if(b){let w=b[g.inputNr];b[g.inputNr]=b[g.inputNr]?ae(w,y):y}else{let w=[];w[g.inputNr]=y,i.set(_,w)}}}}l(tE,"backward");function _G(n,e){let t=n.shape;if(t.length===e.length&&t.every((o,i)=>o===e[i]))return n;let r=n,s=t.length-e.length;for(let o=0;o<s;o++)r=wt(r,0,!1);for(let o=e.length-1;o>=0;o--)e[o]===1&&r.shape[o]!==1&&(r=wt(r,o,!0));return r}l(_G,"_reduceBroadcastGrad");function bG(n,e,t){let r=[n];for(e.set(n,0);r.length>0;){let s=r.shift();if(!t.has(s.id)){t.add(s.id);for(let o of s.nextEdges){if(!o||!o.node)continue;let i=o.node;e.has(i)||e.set(i,0),e.set(i,e.get(i)+1),t.has(i.id)||r.push(i)}}}}l(bG,"_countDeps");WO({backward:tE});function wG(n){let e=n._impl;if(In()&&e.device&&e.device.type==="gpu"){e.storage.retain();let s=new Pt(e.storage,e.storageOffset,e.sizes(),e.strides(),e.dtype,e.device);return new Ze(s)}let t=e.storage.clone(),r=new Pt(t,e.storageOffset,e.sizes(),e.strides(),e.dtype,e.device);return new Ze(r)}l(wG,"_snapshotTensor");function Rf(n){return typeof n=="object"&&n!==null&&"_impl"in n}l(Rf,"_isTensor");function TG(n){for(let e=0;e<n.length;e++){let t=n[e];if(Rf(t)&&t.requiresGrad)return!0;if(Array.isArray(t)){for(let r=0;r<t.length;r++)if(Rf(t[r])&&t[r].requiresGrad)return!0}}return!1}l(TG,"_anyRequiresGrad");function xG(n){let e=[];for(let t=0;t<n.length;t++){let r=n[t];if(Rf(r))e.push(r);else if(Array.isArray(r))for(let s=0;s<r.length;s++)Rf(r[s])&&e.push(r[s])}return e}l(xG,"_extractTensors");function SG(n){let e=n._impl.autogradMeta;if(!e)return null;let t=e.getGradAccumulator();return t||(t=new Xp(n),e.setGradAccumulator(t)),t}l(SG,"_getOrCreateAccumulator");function vG(n,e){let t=Lt.dispatch.bind(Lt);return(r,...s)=>{let o=r;if(!Oo.isEnabled()||!TG(s)){let c=o.subtract(ip);return Lt.redispatch(e,c,...s)}let i=XI(n,s);if(!i){let c=o.subtract(ip);return Lt.redispatch(e,c,...s)}let a=xG(s);i.setOpArgs(s);for(let c=0;c<a.length;c++)i.saveTensor(wG(a[c])),i.saveInputMetadata(c,[...a[c].shape],a[c].dtype);for(let c=0;c<a.length;c++){let f=a[c];if(f.requiresGrad){let d=f.gradFn;if(d){let h=f._impl.autogradMeta?f._impl.autogradMeta.outputNr:0;i.setNextEdge(c,d,h)}else{let h=SG(f);h&&i.setNextEdge(c,h,0)}}}let u=o.subtract(ip),p=Lt.redispatch(e,u,...s);if(Rf(p)){p._impl.setAutogradMeta(new So);let c=p._impl.autogradMeta;c.setGradFn(i,0),c.requiresGrad=!0,p._impl._updateKeySet()}return p}}l(vG,"wrapWithAutograd");function kG(n){return(e,...t)=>{let r=e.subtract(ip);return Lt.redispatch(n,r,...t)}}l(kG,"_makePassthrough");function nE(){let n=Lt.listOps(),e=[40,41,42,43];for(let t of n){let r=Lt.findOp(t);if(!r)continue;let s=r.name,o=eE(s)?vG(s,r):kG(r),i=Mr.fromUnboxed(o);for(let a of e)r.entry.registerKernel(a,i)}}l(nE,"registerAutogradKernels");be();os();var If=Ht,NG=["neg","exp","log","sqrt","rsqrt","abs","sin","cos","tanh","erf","erfc","lgamma","gamma","sigmoid","relu","gelu","silu","sign","floor","ceil","clone","contiguous"],AG=["add","sub","mul","div","pow","remainder","maximum","minimum","eq","ne","lt","le","gt","ge","matmul","dot","flip","unsqueeze"],OG=["sum","mean","max","min","argmax","argmin","prod"],BG={softmax:[-1],log_softmax:[-1],roll:[void 0,0],cumsum:[0],sort:[-1,!1],argsort:[-1,!1],topk:[void 0,-1,!0],split:[void 0,0],chunk:[void 0,0],squeeze:[null]};function RG(n,e){let t=[];for(let r=0;r<e.length;r++)t[r]=n[r]===void 0?e[r]:n[r];for(let r=e.length;r<n.length;r++)t[r]=n[r];return t}l(RG,"withDefaults");function MG(n){for(let e of NG)n[e]=function(){return If[e](this)}}l(MG,"installSelfMethods");function IG(n){for(let e of AG)n[e]=function(t){return If[e](this,t)}}l(IG,"installOneArgMethods");function EG(n){for(let[e,t]of Object.entries(BG))n[e]=function(...r){return If[e](this,...RG(r,t))}}l(EG,"installDefaultedMethods");function LG(n){for(let e of OG)n[e]=function(t,r){return If[e](this,t,r)}}l(LG,"installReductionMethods");function Mf(n){return n.length===1&&Array.isArray(n[0])?n[0]:n}l(Mf,"arrayArg");function rE(n){let e=n.prototype;e.to=function(t){if(this.device.equals(t))return this;let s=If.contiguous(this).data.slice(0,this.numel);return eu(s,this.shape,this.dtype,{device:t})},MG(e),IG(e),EG(e),LG(e),e.mm=function(t){return tt(this,t)},e.requires_grad=function(t=!0){return this.requiresGrad_(t)},e.gather=function(t,r){return Eg(this,t,r)},e.scatter_add=function(t,r,s){return nu(this,t,r,s)},e.scatter=function(t,r,s){return Lg(this,t,r,s)},e.transpose=function(t,r){return vt(this,t,r)},e.slice=function(t,r,s,o){return Dg(this,t,r,s,o)},e.narrow=function(t,r,s){return ir(this,t,r,s)},e.select=function(t,r){return en(this,t,r)},e.reshape=function(...t){return Hn(this,Mf(t))},e.permute=function(...t){return El(this,Mf(t))},e.expand=function(...t){return Ll(this,Mf(t))},e.repeat=function(...t){return Dl(this,Mf(t))},e.tile=function(...t){return Pg(this,Mf(t))},e.t=function(){if(this.ndim!==2)throw new Error("t() expects a 2D tensor");return vt(this,0,1)}}l(rE,"installOps");sr();sr();qn();ou();yu();nn();RS();_f();yx();Ma();kt();kt();function Lf(n,e){return{dtype:e?.dtype??n.dtype,device:e?.device??n.device,requiresGrad:e?.requiresGrad??!1}}l(Lf,"_likeOpts");function $G(n,e){return Tn(n.shape,Lf(n,e))}l($G,"emptyLike");function FG(n,e){return Ve(n.shape,Lf(n,e))}l(FG,"zerosLike");function zG(n,e){return tn(n.shape,Lf(n,e))}l(zG,"onesLike");function VG(n,e,t){return ot(n.shape,e,Lf(n,t))}l(VG,"fullLike");function GG(n,e){return Gg(n.shape,Lf(n,e))}l(GG,"randnLike");os();be();var dE={};at(dE,{AdaptiveAvgPool2d:()=>Xf,AvgPool2d:()=>Qf,BCELoss:()=>rd,BatchNorm1d:()=>cc,BatchNorm2d:()=>Jf,Conv1d:()=>zf,Conv2d:()=>Ff,CrossEntropyLoss:()=>td,Dropout:()=>Sr,ELU:()=>Kf,Embedding:()=>sd,F:()=>fE,Flatten:()=>cd,GELU:()=>Gf,GRU:()=>id,GRUCell:()=>pc,GroupNorm:()=>E_,LSTM:()=>ad,LSTMCell:()=>fc,LayerNorm:()=>no,LeakyReLU:()=>Wf,Linear:()=>Sn,LogSoftmax:()=>Yf,MSELoss:()=>ed,MaxPool2d:()=>Zf,Module:()=>ge,ModuleDict:()=>ld,ModuleList:()=>aa,MultiheadAttention:()=>ua,NLLLoss:()=>nd,Parameter:()=>At,PositionalEncoding:()=>fd,ReLU:()=>Vf,Sequential:()=>ud,SiLU:()=>jf,Sigmoid:()=>Uf,Softmax:()=>Hf,Tanh:()=>qf,Transformer:()=>pd,TransformerDecoder:()=>hc,TransformerDecoderLayer:()=>wu,TransformerEncoder:()=>dc,TransformerEncoderLayer:()=>bu,functional:()=>lc,init:()=>T_});sr();var At=class extends Ze{static{l(this,"Parameter")}constructor(e,t=!0){e instanceof Ze?super(e._impl):super(e),t&&this.requiresGrad_(!0)}get isParameter(){return!0}};var ge=class n{static{l(this,"Module")}_parameters;_buffers;_modules;_training;_detected;constructor(){this._parameters=new Map,this._buffers=new Map,this._modules=new Map,this._training=!0}forward(...e){throw new Error(`${this.constructor.name}.forward() not implemented`)}call(...e){return this.forward(...e)}registerParameter(e,t){if(t!==null&&!(t instanceof At))throw new Error("Expected Parameter instance");this._parameters.set(e,t)}registerBuffer(e,t){this._buffers.set(e,t)}registerModule(e,t){if(t!==null&&!(t instanceof n))throw new Error("Expected Module instance");this._modules.set(e,t)}*parameters(e=!0){this._autoDetect();for(let[,t]of this._parameters)t!==null&&(yield t);if(e)for(let[,t]of this._modules)t!==null&&(yield*t.parameters(!0))}*namedParameters(e="",t=!0){this._autoDetect();let r=e?e+".":"";for(let[s,o]of this._parameters)o!==null&&(yield[r+s,o]);if(t)for(let[s,o]of this._modules)o!==null&&(yield*o.namedParameters(r+s,!0))}*buffers(e=!0){for(let[,t]of this._buffers)t!==null&&(yield t);if(e)for(let[,t]of this._modules)t!==null&&(yield*t.buffers(!0))}*children(){this._autoDetect();for(let[,e]of this._modules)e!==null&&(yield e)}*namedChildren(){this._autoDetect();for(let[e,t]of this._modules)t!==null&&(yield[e,t])}*modules(){yield this,this._autoDetect();for(let[,e]of this._modules)e!==null&&(yield*e.modules())}*namedModules(e=""){yield[e,this],this._autoDetect();let t=e?e+".":"";for(let[r,s]of this._modules)s!==null&&(yield*s.namedModules(t+r))}stateDict(e=""){this._autoDetect();let t=new Map,r=e?e+".":"";for(let[s,o]of this._parameters)o!==null&&t.set(r+s,o);for(let[s,o]of this._buffers)o!==null&&t.set(r+s,o);for(let[s,o]of this._modules)if(o!==null)for(let[i,a]of o.stateDict(r+s))t.set(i,a);return t}loadStateDict(e){this._autoDetect();for(let[t,r]of this._parameters){let s=t;if(e.has(s)&&r!==null){let o=e.get(s),i=r._impl.storage.data,a=o._impl?o._impl.storage.data:o.data;for(let u=0;u<i.length;u++)i[u]=a[u]}}for(let[t,r]of this._modules)if(r!==null){let s=new Map,o=t+".";for(let[i,a]of e)i.startsWith(o)&&s.set(i.substring(o.length),a);s.size>0&&r.loadStateDict(s)}}train(e=!0){this._autoDetect(),this._training=e;for(let[,t]of this._modules)t!==null&&t.train(e);return this}eval(){return this.train(!1)}get training(){return this._training}to(e){this._autoDetect();for(let[t,r]of this._parameters){let s=r;r!==null&&typeof s?.to=="function"&&(this._parameters.set(t,new At(s.to(e),r.requiresGrad)),this[t]=this._parameters.get(t))}for(let[t,r]of this._buffers){let s=r;r!==null&&typeof s?.to=="function"&&(this._buffers.set(t,s.to(e)),this[t]=this._buffers.get(t))}for(let[,t]of this._modules)t!==null&&t.to(e);return this}apply(e){this._autoDetect();for(let[,t]of this._modules)t!==null&&t.apply(e);return e(this),this}zeroGrad(){for(let e of this.parameters())if(e.grad){let t=e.grad._impl.storage.data;t&&t.fill(0)}return this}_autoDetect(){if(this._detected)return;this._detected=!0;let e=Object.keys(this);for(let t of e){if(t.startsWith("_"))continue;let r=this[t];r instanceof At&&!this._parameters.has(t)?this._parameters.set(t,r):r instanceof n&&!this._modules.has(t)&&this._modules.set(t,r)}}toString(){return this._buildRepr("")}_buildRepr(e){this._autoDetect();let t=this.constructor.name;if(this._modules.size===0)return t+"()";let r=[t+"("];for(let[s,o]of this._modules)o!==null&&r.push(e+"  ("+s+"): "+o._buildRepr(e+"  "));return r.push(e+")"),r.join(`
+`)}};var T_={};at(T_,{_calculateFanInFanOut:()=>ac,constant_:()=>qG,kaiming_normal_:()=>HG,kaiming_uniform_:()=>sE,normal_:()=>Cf,ones_:()=>UG,resetLinearParameters:()=>uc,uniform_:()=>w_,xavier_normal_:()=>KG,xavier_uniform_:()=>WG,zeros_:()=>jG});function ac(n){let e=n.shape,t=e.length;if(t<2)throw new Error("Fan in/out requires at least 2D tensor");let r=e[1],s=e[0],o=1;for(let i=2;i<t;i++)o*=e[i];return{fanIn:r*o,fanOut:s*o}}l(ac,"_calculateFanInFanOut");function uc(n,e){if(sE(n,Math.sqrt(5)),e){let{fanIn:t}=ac(n),r=1/Math.sqrt(t);w_(e,-r,r)}}l(uc,"resetLinearParameters");function w_(n,e=0,t=1){let r=n._impl.storage.data;if(!r)return n;let s=t-e;for(let o=0;o<r.length;o++)r[o]=e+Math.random()*s;return n._impl.bumpVersion(),n}l(w_,"uniform_");function Cf(n,e=0,t=1){let r=n._impl.storage.data;if(!r)return n;for(let s=0;s<r.length;s+=2){let o=Math.random()||1e-10,i=Math.random(),a=Math.sqrt(-2*Math.log(o)),u=6.283185307179586*i;r[s]=e+t*a*Math.cos(u),s+1<r.length&&(r[s+1]=e+t*a*Math.sin(u))}return n._impl.bumpVersion(),n}l(Cf,"normal_");function jG(n){let e=n._impl.storage.data;return e&&e.fill(0),n._impl.bumpVersion(),n}l(jG,"zeros_");function UG(n){let e=n._impl.storage.data;return e&&e.fill(1),n._impl.bumpVersion(),n}l(UG,"ones_");function qG(n,e){let t=n._impl.storage.data;return t&&t.fill(e),n._impl.bumpVersion(),n}l(qG,"constant_");function WG(n,e=1){let{fanIn:t,fanOut:r}=ac(n),s=e*Math.sqrt(2/(t+r)),o=Math.sqrt(3)*s;return w_(n,-o,o)}l(WG,"xavier_uniform_");function KG(n,e=1){let{fanIn:t,fanOut:r}=ac(n),s=e*Math.sqrt(2/(t+r));return Cf(n,0,s)}l(KG,"xavier_normal_");function sE(n,e=0,t="fan_in",r="leaky_relu"){let{fanIn:s,fanOut:o}=ac(n),i=t==="fan_in"?s:o,u=oE(r,e)/Math.sqrt(i),p=Math.sqrt(3)*u;return w_(n,-p,p)}l(sE,"kaiming_uniform_");function HG(n,e=0,t="fan_in",r="leaky_relu"){let{fanIn:s,fanOut:o}=ac(n),i=t==="fan_in"?s:o,u=oE(r,e)/Math.sqrt(i);return Cf(n,0,u)}l(HG,"kaiming_normal_");function oE(n,e=.01){switch(n){case"linear":case"sigmoid":return 1;case"tanh":return 5/3;case"relu":return Math.sqrt(2);case"leaky_relu":return Math.sqrt(2/(1+e*e));default:return 1}}l(oE,"_calculateGain");var lc={};at(lc,{elu:()=>PS,gelu:()=>Pf,leaky_relu:()=>DS,log_softmax:()=>CS,relu:()=>Df,sigmoid:()=>IS,silu:()=>MS,softmax:()=>LS,tanh:()=>ES});be();kt();function Df(n){return Ml(n)}l(Df,"relu");function Pf(n){return Ng(n)}l(Pf,"gelu");function MS(n){return Ag(n)}l(MS,"silu");function IS(n){return wr(n)}l(IS,"sigmoid");function ES(n){return Hs(n)}l(ES,"tanh");function LS(n,e=-1){return mi(n,e)}l(LS,"softmax");function CS(n,e=-1){return $g(n,e)}l(CS,"log_softmax");function DS(n,e=.01){let t=ee(n,ot(n.shape,e,{dtype:n.dtype,device:n.device}));return ta(n,t)}l(DS,"leaky_relu");function PS(n,e=1){let t=ot(n.shape,0,{dtype:n.dtype,device:n.device}),r=ot(n.shape,1,{dtype:n.dtype,device:n.device}),s=ot(n.shape,e,{dtype:n.dtype,device:n.device}),o=Il(n,t),i=ee(s,ze(as(n),r));return or(o,n,i)}l(PS,"elu");var S_={};at(S_,{batch_norm:()=>zS,group_norm:()=>FS,layer_norm:()=>$S});be();kt();yu();function $S(n,e,t,r,s=1e-5){if((n instanceof bt||n.isSymbolic)&&t&&r){let m=n.ndim-e.length;return PT(n,t,r,m,s)}let o=n.ndim-e.length,i=[];for(let m=o;m<n.ndim;m++)i.push(m);let a=n;for(let m=i.length-1;m>=0;m--)a=Tr(a,i[m],!0);let u=ze(n,a),c=ee(u,u);for(let m=i.length-1;m>=0;m--)c=Tr(c,i[m],!0);let f=ot([],s),d=Kt(ot([],1),na(ae(c,f))),h=ee(u,d);return t&&(h=ee(h,t)),r&&(h=ae(h,r)),h}l($S,"layer_norm");function FS(n,e,t,r,s=1e-5){let o=n.shape,i=o[0],a=o[1],u=o.slice(2),p=n.reshape([i,e,a/e,...u]),c=[];for(let w=2;w<p.ndim;w++)c.push(w);let f=p;for(let w=c.length-1;w>=0;w--)f=Tr(f,c[w],!0);let d=ze(p,f),m=ee(d,d);for(let w=c.length-1;w>=0;w--)m=Tr(m,c[w],!0);let g=ot([],s),y=Kt(ot([],1),na(ae(m,g))),_=ee(d,y).reshape(o),b=[1,a,...u.map(()=>1)];return t&&(_=ee(_,t.reshape(b))),r&&(_=ae(_,r.reshape(b))),_}l(FS,"group_norm");var x_=1;function YG(n,e){let t=new Array(n).fill(1);return t[x_]=e,t}l(YG,"channelShape");function iE(n,e){let t=n;for(let r=e.length-1;r>=0;r--)t=Tr(t,e[r],!0);return t}l(iE,"reduceMeanOver");function aE(n,e,t){let r=n.data;if(!r)return;let s=e.reshape([n.shape[0]]).toArray();for(let o=0;o<r.length;o++)r[o]=r[o]*(1-t)+s[o]*t}l(aE,"blendRunning");function zS(n,e,t,r,s,o=!0,i=1e-5,a=.1){let u=n instanceof bt||n.isSymbolic;if(!o)return $T(n,r,s,e,t,x_,i);let p=[];for(let y=0;y<n.ndim;y++)y!==x_&&p.push(y);let c=iE(n,p),f=ze(n,c),d=iE(ee(f,f),p),h=Kt(ot([],1),na(ae(d,ot([],i)))),m=ee(f,h),g=YG(n.ndim,n.shape[x_]);return r&&(m=ee(m,r.reshape(g))),s&&(m=ae(m,s.reshape(g))),u||(e&&aE(e,c,a),t&&aE(t,d,a)),m}l(zS,"batch_norm");var GS={};at(GS,{linear:()=>VS});be();be();yu();function VS(n,e,t){let r;n instanceof bt||n.isSymbolic?r=vt(e,0,1):r=vt(e,0,1);let s=tt(n,r);return t?ae(s,t):s}l(VS,"linear");var k_={};at(k_,{conv1d:()=>jS,conv2d:()=>v_});be();be();function v_(n,e,t,r=[1,1],s=[[0,0],[0,0]],o=[1,1],i=1){let a=Array.isArray(r)?r:[r,r],u=JG(s),p=Array.isArray(o)?o:[o,o],c=FT(n,e,a,u,p,i);if(t){let f=Hn(t,[1,t.shape[0],1,1]);return ae(c,f)}return c}l(v_,"conv2d");function jS(n,e,t,r=1,s=0,o=1,i=1){let a=Lr(n,2),u=Lr(e,2),p=[1,Array.isArray(r)?r[0]:r],c=[[0,0],ZG(s)],f=[1,Array.isArray(o)?o[0]:o],d=v_(a,u,null,p,c,f,i),h=Cl(d,2);if(t){let m=Hn(t,[1,t.shape[0],1]);return ae(h,m)}return h}l(jS,"conv1d");function JG(n){return typeof n=="number"?[[n,n],[n,n]]:Array.isArray(n)&&typeof n[0]=="number"?n.map(e=>[e,e]):n}l(JG,"_normalizePadding");function ZG(n){return typeof n=="number"?[n,n]:n}l(ZG,"_normalizePadding1d");var A_={};at(A_,{adaptive_avg_pool2d:()=>qS,avg_pool2d:()=>N_,max_pool2d:()=>US});be();function US(n,e,t,r=[[0,0],[0,0]]){let s=Array.isArray(e)?e:[e,e],o=t?Array.isArray(t)?t:[t,t]:s,i=uE(r);return Fg(n,"max",s,o,i)}l(US,"max_pool2d");function N_(n,e,t,r=[[0,0],[0,0]]){let s=Array.isArray(e)?e:[e,e],o=t?Array.isArray(t)?t:[t,t]:s,i=uE(r);return Fg(n,"avg",s,o,i)}l(N_,"avg_pool2d");function qS(n,e){let[t,r]=Array.isArray(e)?e:[e,e],s=n.shape[2],o=n.shape[3],i=Math.floor(s/t),a=Math.floor(o/r);return N_(n,[i,a],[i,a],[[0,0],[0,0]])}l(qS,"adaptive_avg_pool2d");function uE(n){if(typeof n=="number")return[[n,n],[n,n]];if(Array.isArray(n)&&typeof n[0]=="number"){let e=n.map(t=>[t,t]);return e.length===1?[e[0],e[0]]:e}return n}l(uE,"_normalizePadding");var WS={};at(WS,{dropout:()=>$f});be();kt();function $f(n,e=.5,t=!0){if(!t||e===0)return n;if(e===1)return ot(n.shape,0,{dtype:n.dtype,device:n.device});let r=QG(n.shape,1-e,n.dtype,n.device),s=ot(n.shape,1/(1-e),{dtype:n.dtype,device:n.device});return ee(ee(n,r),s)}l($f,"dropout");function QG(n,e,t,r){let s=Tn(n,{dtype:t,device:r}),o=s._impl.storage.data;if(o)for(let i=0;i<o.length;i++)o[i]=Math.random()<e?1:0;return s}l(QG,"_bernoulliMask");var B_={};at(B_,{binary_cross_entropy:()=>YS,cross_entropy:()=>HS,mse_loss:()=>KS,nll_loss:()=>O_});be();kt();be();ss();function KS(n,e,t="mean"){let r=ze(n,e),s=ee(r,r);return lE(s,t)}l(KS,"mse_loss");function XG(n,e){let t=e6(n,e),r=ze(n,t),s=as(r),o=t6(s,e),i=ko(o);return ze(r,i)}l(XG,"_logSoftmaxAutograd");function e6(n,e){let t=e<0?n.ndim+e:e;return No(n,t,!0)}l(e6,"_dimMax");function t6(n,e){let t=e<0?n.ndim+e:e;return wt(n,t,!0)}l(t6,"_dimSum");function O_(n,e,t="mean",r=null){let s=n.ndim-1,o=n.shape[s],i=Ig(e,o),a=wt(ee(n,i),s),u=a,p=null;if(r!==null){let f=en(i,s,r),d=ae(is(f),1);u=ee(a,d),p=wt(d)}let c=is(wt(u));if(t==="sum")return c;if(t==="mean")return p!==null?Kt(c,p):Kt(c,bn(e.shape));throw new Error(`nll_loss: unknown reduction '${t}'`)}l(O_,"nll_loss");function HS(n,e,t="mean",r=null){let s=XG(n,-1);return O_(s,e,t,r)}l(HS,"cross_entropy");function YS(n,e,t="mean"){let r=ot(n.shape,1e-7,{dtype:n.dtype,device:n.device}),s=ot(n.shape,1,{dtype:n.dtype,device:n.device}),o=ko(ae(n,r)),i=ko(ae(ze(s,n),r)),a=is(ae(ee(e,o),ee(ze(s,e),i)));return lE(a,t)}l(YS,"binary_cross_entropy");function lE(n,e){if(e==="mean")return Tr(n);if(e==="sum")return wt(n);if(e==="none")return n;throw new Error(`Unknown reduction: ${e}`)}l(lE,"_reduce");var ZS={};at(ZS,{embedding:()=>JS});be();function JS(n,e){let t=n.shape[n.shape.length-1],r=e.shape.reduce((o,i)=>o*i,1);return ra(n,0,e.reshape([r])).reshape([...e.shape,t])}l(JS,"embedding");var XS={};at(XS,{scaled_dot_product_attention:()=>QS});be();be();yu();_u();kt();function r6(n,e){if(n instanceof bt||n.isSymbolic)return mi(n,e);let t=e<0?n.ndim+e:e,r=No(n,t,!0).detach(),s=ze(n,r),o=as(s),i=wt(o,t,!0);return Kt(o,i)}l(r6,"_softmax");function s6(n){let e=n.ndim;return n instanceof bt||n.isSymbolic?vt(n,e-2,e-1):vt(n,e-2,e-1)}l(s6,"_transposeLastTwo");function o6(n,e){let t=Tn([n,e]),r=t._impl.storage.data,s=e-n;for(let o=0;o<n;o++)for(let i=0;i<e;i++)r[o*e+i]=i<=o+s?0:-1/0;return t}l(o6,"_generateCausalMask");function QS(n,e,t,r=null,s=0,o=!1,i=!1){let a=n.shape[n.ndim-1],u=n.shape[n.ndim-2],p=e.shape[e.ndim-2],c=As();if(c&&!r&&!(s>0&&i)&&n instanceof bt&&e instanceof bt&&t instanceof bt&&n.ndim===4)return c.recordOp("scaled_dot_product_attention",[n,e,t],{scale:1/Math.sqrt(a),causal:o});let d=ot([],1/Math.sqrt(a)),h=s6(e),m=tt(n,h);m=ee(m,d),o&&(m=ae(m,o6(u,p))),r&&(m=ae(m,r));let g=r6(m,-1);return s>0&&i&&(g=$f(g,s,!0)),tt(g,t)}l(QS,"scaled_dot_product_attention");kt();var Sn=class extends ge{static{l(this,"Linear")}inFeatures;outFeatures;weight;bias;constructor(e,t,r=!0){super(),this.inFeatures=e,this.outFeatures=t,this.weight=new At(Tn([t,e])),this.bias=r?new At(Ve([t])):null,this._resetParameters()}_resetParameters(){uc(this.weight,this.bias)}forward(e){return VS(e,this.weight,this.bias)}};kt();var Ff=class extends ge{static{l(this,"Conv2d")}inChannels;outChannels;kernelSize;stride;padding;dilation;groups;weight;bias;constructor(e,t,r,s={}){super(),this.inChannels=e,this.outChannels=t,this.kernelSize=Array.isArray(r)?r:[r,r],this.stride=s.stride?Array.isArray(s.stride)?s.stride:[s.stride,s.stride]:[1,1],this.padding=s.padding??0,this.dilation=s.dilation?Array.isArray(s.dilation)?s.dilation:[s.dilation,s.dilation]:[1,1],this.groups=s.groups??1;let o=this.kernelSize[0],i=this.kernelSize[1];this.weight=new At(Tn([t,e/this.groups,o,i])),this.bias=s.bias!==!1?new At(Ve([t])):null,this._resetParameters()}_resetParameters(){uc(this.weight,this.bias)}forward(e){return v_(e,this.weight,this.bias,this.stride,this.padding,this.dilation,this.groups)}},zf=class extends ge{static{l(this,"Conv1d")}inChannels;outChannels;kernelSize;stride;padding;dilation;groups;weight;bias;constructor(e,t,r,s={}){super(),this.inChannels=e,this.outChannels=t,this.kernelSize=Array.isArray(r)?r[0]:r,this.stride=s.stride??1,this.padding=s.padding??0,this.dilation=s.dilation??1,this.groups=s.groups??1,this.weight=new At(Tn([t,e/this.groups,this.kernelSize])),this.bias=s.bias!==!1?new At(Ve([t])):null,uc(this.weight,this.bias)}forward(e){return jS(e,this.weight,this.bias,this.stride,this.padding,this.dilation,this.groups)}};var Vf=class extends ge{static{l(this,"ReLU")}forward(e){return Df(e)}},Gf=class extends ge{static{l(this,"GELU")}forward(e){return Pf(e)}},jf=class extends ge{static{l(this,"SiLU")}forward(e){return MS(e)}},Uf=class extends ge{static{l(this,"Sigmoid")}forward(e){return IS(e)}},qf=class extends ge{static{l(this,"Tanh")}forward(e){return ES(e)}},Wf=class extends ge{static{l(this,"LeakyReLU")}negativeSlope;constructor(e=.01){super(),this.negativeSlope=e}forward(e){return DS(e,this.negativeSlope)}},Kf=class extends ge{static{l(this,"ELU")}alpha;constructor(e=1){super(),this.alpha=e}forward(e){return PS(e,this.alpha)}},Hf=class extends ge{static{l(this,"Softmax")}dim;constructor(e=-1){super(),this.dim=e}forward(e){return LS(e,this.dim)}},Yf=class extends ge{static{l(this,"LogSoftmax")}dim;constructor(e=-1){super(),this.dim=e}forward(e){return CS(e,this.dim)}};kt();var no=class extends ge{static{l(this,"LayerNorm")}normalizedShape;eps;weight;bias;constructor(e,t=1e-5,r=!0){super(),this.normalizedShape=Array.isArray(e)?e:[e],this.eps=t,this.weight=r?new At(tn(this.normalizedShape)):null,this.bias=r?new At(Ve(this.normalizedShape)):null}forward(e){return $S(e,this.normalizedShape,this.weight,this.bias,this.eps)}},E_=class extends ge{static{l(this,"GroupNorm")}numGroups;numChannels;eps;weight;bias;constructor(e,t,r=1e-5,s=!0){super(),this.numGroups=e,this.numChannels=t,this.eps=r,this.weight=s?new At(tn([t])):null,this.bias=s?new At(Ve([t])):null}forward(e){return FS(e,this.numGroups,this.weight,this.bias,this.eps)}},cc=class extends ge{static{l(this,"BatchNorm1d")}numFeatures;eps;weight;bias;runningMean;runningVar;constructor(e,t=1e-5,r=!0){super(),this.numFeatures=e,this.eps=t,this.weight=r?new At(tn([e])):null,this.bias=r?new At(Ve([e])):null,this.runningMean=Ve([e]),this.runningVar=tn([e]),this.registerBuffer("runningMean",this.runningMean),this.registerBuffer("runningVar",this.runningVar)}forward(e){return zS(e,this.runningMean,this.runningVar,this.weight,this.bias,this.training,this.eps)}},Jf=class extends cc{static{l(this,"BatchNorm2d")}constructor(e,t=1e-5,r=!0){super(e,t,r)}};var Zf=class extends ge{static{l(this,"MaxPool2d")}kernelSize;stride;padding;constructor(e,t,r=0){super(),this.kernelSize=Array.isArray(e)?e:[e,e],this.stride=t?Array.isArray(t)?t:[t,t]:this.kernelSize,this.padding=r}forward(e){return US(e,this.kernelSize,this.stride,this.padding)}},Qf=class extends ge{static{l(this,"AvgPool2d")}kernelSize;stride;padding;constructor(e,t,r=0){super(),this.kernelSize=Array.isArray(e)?e:[e,e],this.stride=t?Array.isArray(t)?t:[t,t]:this.kernelSize,this.padding=r}forward(e){return N_(e,this.kernelSize,this.stride,this.padding)}},Xf=class extends ge{static{l(this,"AdaptiveAvgPool2d")}outputSize;constructor(e){super(),this.outputSize=Array.isArray(e)?e:[e,e]}forward(e){return qS(e,this.outputSize)}};var Sr=class extends ge{static{l(this,"Dropout")}p;constructor(e=.5){super(),this.p=e}forward(e){return $f(e,this.p,this.training)}};var ed=class extends ge{static{l(this,"MSELoss")}reduction;constructor(e="mean"){super(),this.reduction=e}forward(e,t){return KS(e,t,this.reduction)}},td=class extends ge{static{l(this,"CrossEntropyLoss")}reduction;ignoreIndex;constructor(e="mean",t=-100){super(),this.reduction=e,this.ignoreIndex=t}forward(e,t){return HS(e,t,this.reduction,this.ignoreIndex===-100?null:this.ignoreIndex)}},nd=class extends ge{static{l(this,"NLLLoss")}reduction;ignoreIndex;constructor(e="mean",t=-100){super(),this.reduction=e,this.ignoreIndex=t}forward(e,t){return O_(e,t,this.reduction,this.ignoreIndex===-100?null:this.ignoreIndex)}},rd=class extends ge{static{l(this,"BCELoss")}reduction;constructor(e="mean"){super(),this.reduction=e}forward(e,t){return YS(e,t,this.reduction)}};kt();var sd=class extends ge{static{l(this,"Embedding")}numEmbeddings;embeddingDim;weight;constructor(e,t){super(),this.numEmbeddings=e,this.embeddingDim=t,this.weight=new At(Tn([e,t])),Cf(this.weight)}forward(e){return JS(this.weight,e)}};kt();be();be();_u();be();be();function i6(n){if(Array.isArray(n[0])){let e=n[0],t=new Array(e.length);for(let r=0;r<t.length;r++)t[r]=us(n.map(s=>s[r]),0);return t}return us(n,0)}l(i6,"stackSteps");function od(n,e,t){let r=Array.isArray(e),s=Array.isArray(t),o=s?t:[t],i=r?e:[e],a=As();if(!a){let d=e,h=[],m=o[0].shape[0];for(let g=0;g<m;g++){let y=s?o.map(w=>en(w,0,g)):en(o[0],0,g),[_,b]=n(d,y);d=_,h.push(b)}return[d,i6(h)]}let u=!1,p=l((d,h)=>{let[m,g]=n(r?d:d[0],s?h:h[0]);return u=Array.isArray(g),[r?m:[m],u?g:[g]]},"step"),[c,f]=a.scan(o,i,p);return[r?c:c[0],u?f:f[0]]}l(od,"scan");_u();wn();qn();var pc=class extends ge{static{l(this,"GRUCell")}inputSize;hiddenSize;x2h;h2h;constructor(e,t,r=!0){super(),this.inputSize=e,this.hiddenSize=t,this.x2h=new Sn(e,3*t,r),this.h2h=new Sn(t,3*t,r)}forward(e,t=null){let r=t!==null?t:Ve([e.shape[0],this.hiddenSize],{device:e.device}),s=this.x2h.forward(e),o=this.h2h.forward(r),[i,a,u]=hi(s,this.hiddenSize,-1),[p,c,f]=hi(o,this.hiddenSize,-1),d=wr(ae(i,p)),h=wr(ae(a,c)),m=Hs(ae(u,ee(d,f)));return ae(m,ee(h,ze(r,m)))}},id=class extends ge{static{l(this,"GRU")}inputSize;hiddenSize;numLayers;batchFirst;cells;constructor(e,t,r=1,s=!1,o=!0){super(),this.inputSize=e,this.hiddenSize=t,this.numLayers=r,this.batchFirst=s,this.cells=[];for(let i=0;i<r;i++){let a=new pc(i===0?e:t,t,o);this.cells.push(a),this.registerModule("cell_"+i,a)}}forward(e,t=null){let r=uS();if(r&&e.device.type==="gpu"&&!As()){let c=this.batchFirst?e.transpose(0,1):e,f={inputSize:this.inputSize,hiddenSize:this.hiddenSize,seqLen:c.shape[0],batch:c.shape[1]},[d,h]=r(c,this.cells,f,t);return[this.batchFirst?d.transpose(0,1):d,h]}let s=Sf();if(s&&e.device.type==="webgpu"&&!As()){let c=this.batchFirst?e.transpose(0,1):e,f={kind:"gru",inputSize:this.inputSize,hiddenSize:this.hiddenSize,seqLen:c.shape[0],batch:c.shape[1]},d=s(c,this.cells,f,t,null);if(d){let[h,m]=d;return[this.batchFirst?h.transpose(0,1):h,m]}}let o=this.batchFirst?e.transpose(0,1):e,i=o.shape[1],a=o,u=[];for(let c=0;c<this.numLayers;c++){let f=t!==null?en(t,0,c):Ve([i,this.hiddenSize],{device:o.device}),d=this.cells[c],[h,m]=od((g,y)=>{let _=d.forward(y,g);return[_,_]},f,a);u.push(h),a=m}let p=a;return this.batchFirst&&(p=p.transpose(0,1)),[p,us(u,0)]}};kt();be();be();_u();wn();qn();var fc=class extends ge{static{l(this,"LSTMCell")}inputSize;hiddenSize;x2h;h2h;constructor(e,t,r=!0){super(),this.inputSize=e,this.hiddenSize=t,this.x2h=new Sn(e,4*t,r),this.h2h=new Sn(t,4*t,r)}forward(e,t=null){let r=t!==null?t[0]:Ve([e.shape[0],this.hiddenSize],{device:e.device}),s=t!==null?t[1]:Ve([e.shape[0],this.hiddenSize],{device:e.device}),o=ae(this.x2h.forward(e),this.h2h.forward(r)),[i,a,u,p]=hi(o,this.hiddenSize,-1),c=wr(i),f=wr(a),d=Hs(u),h=wr(p),m=ae(ee(f,s),ee(c,d));return[ee(h,Hs(m)),m]}},ad=class extends ge{static{l(this,"LSTM")}inputSize;hiddenSize;numLayers;batchFirst;cells;constructor(e,t,r=1,s=!1,o=!0){super(),this.inputSize=e,this.hiddenSize=t,this.numLayers=r,this.batchFirst=s,this.cells=[];for(let i=0;i<r;i++){let a=new fc(i===0?e:t,t,o);this.cells.push(a),this.registerModule("cell_"+i,a)}}forward(e,t=null){let r=aS();if(r&&e.device.type==="gpu"&&!As()){let f=this.batchFirst?e.transpose(0,1):e,d={inputSize:this.inputSize,hiddenSize:this.hiddenSize,seqLen:f.shape[0],batch:f.shape[1]},[h,m,g]=r(f,this.cells,d,t?t[0]:null,t?t[1]:null);return[this.batchFirst?h.transpose(0,1):h,[m,g]]}let s=Sf();if(s&&e.device.type==="webgpu"&&!As()){let f=this.batchFirst?e.transpose(0,1):e,d={inputSize:this.inputSize,hiddenSize:this.hiddenSize,seqLen:f.shape[0],batch:f.shape[1]},h=s(f,this.cells,d,t?t[0]:null,t?t[1]:null);if(h){let[m,g,y]=h;return[this.batchFirst?m.transpose(0,1):m,[g,y]]}}let o=this.batchFirst?e.transpose(0,1):e,i=o.shape[1],a=o,u=[],p=[];for(let f=0;f<this.numLayers;f++){let d=t!==null?en(t[0],0,f):Ve([i,this.hiddenSize],{device:o.device}),h=t!==null?en(t[1],0,f):Ve([i,this.hiddenSize],{device:o.device}),m=this.cells[f],[[g,y],_]=od((b,w)=>{let[x,S]=m.forward(w,b);return[[x,S],x]},[d,h],a);u.push(g),p.push(y),a=_}let c=a;return this.batchFirst&&(c=c.transpose(0,1)),[c,[us(u,0),us(p,0)]]}};var ud=class extends ge{static{l(this,"Sequential")}_length;constructor(...e){super();for(let t=0;t<e.length;t++)this[String(t)]=e[t],this.registerModule(String(t),e[t]);this._length=e.length}forward(e){let t=e;for(let r=0;r<this._length;r++)t=this[String(r)].forward(t);return t}get length(){return this._length}*[Symbol.iterator](){for(let e=0;e<this._length;e++)yield this[String(e)]}push(e){let t=this._length;return this[String(t)]=e,this.registerModule(String(t),e),this._length++,this}},aa=class extends ge{static{l(this,"ModuleList")}_list;constructor(e){if(super(),this._list=[],e)for(let t=0;t<e.length;t++)this._list.push(e[t]),this.registerModule(String(t),e[t])}get length(){return this._list.length}get(e){return this._list[e]}push(e){let t=this._list.length;return this._list.push(e),this.registerModule(String(t),e),this}*[Symbol.iterator](){for(let e of this._list)yield e}forward(){throw new Error("ModuleList does not implement forward()")}},ld=class extends ge{static{l(this,"ModuleDict")}_dict;constructor(e){if(super(),this._dict=new Map,e)for(let[t,r]of Object.entries(e))this._dict.set(t,r),this.registerModule(t,r)}get(e){return this._dict.get(e)}set(e,t){return this._dict.set(e,t),this.registerModule(e,t),this}has(e){return this._dict.has(e)}get size(){return this._dict.size}*keys(){yield*this._dict.keys()}*values(){yield*this._dict.values()}*[Symbol.iterator](){yield*this._dict.entries()}forward(){throw new Error("ModuleDict does not implement forward()")}};be();var cd=class extends ge{static{l(this,"Flatten")}startDim;endDim;constructor(e=1,t=-1){super(),this.startDim=e,this.endDim=t}forward(e){let t=e.shape,r=t.length,s=this.startDim<0?r+this.startDim:this.startDim,o=this.endDim<0?r+this.endDim:this.endDim,i=1;for(let u=s;u<=o;u++)i*=t[u];let a=[];for(let u=0;u<s;u++)a.push(t[u]);a.push(i);for(let u=o+1;u<r;u++)a.push(t[u]);return Hn(e,a)}};be();kt();kt();function pE(n){return n==="gelu"?Pf:Df}l(pE,"_getActivation");var ua=class extends ge{static{l(this,"MultiheadAttention")}embedDim;numHeads;headDim;batchFirst;dropout;qProj;kProj;vProj;outProj;constructor(e,t,r=0,s=!0,o=null,i=null,a=!0){super(),this.embedDim=e,this.numHeads=t,this.headDim=Math.floor(e/t),this.batchFirst=a,this.dropout=r,this.qProj=new Sn(e,e,s),this.kProj=new Sn(o??e,e,s),this.vProj=new Sn(i??e,e,s),this.outProj=new Sn(e,e,s)}forward(e,t,r,s=null,o=null,i=!1){this.batchFirst||(e=e.transpose(0,1),t=t.transpose(0,1),r=r.transpose(0,1));let a=e.shape[0],u=e.shape[1],p=t.shape[1],c=this.qProj.forward(e),f=this.kProj.forward(t),d=this.vProj.forward(r);if(c=c.reshape([a,u,this.numHeads,this.headDim]).permute([0,2,1,3]),f=f.reshape([a,p,this.numHeads,this.headDim]).permute([0,2,1,3]),d=d.reshape([a,p,this.numHeads,this.headDim]).permute([0,2,1,3]),o){let g=ot(o.shape,-1/0),y=ot(o.shape,0),_=or(o,g,y);_=_.unsqueeze(1).unsqueeze(2),s=s?ae(s,_):_}let h=QS(c,f,d,s,this.dropout,i,this.training);h=h.permute([0,2,1,3]).reshape([a,u,this.embedDim]);let m=this.outProj.forward(h);return this.batchFirst||(m=m.transpose(0,1)),m}},bu=class extends ge{static{l(this,"TransformerEncoderLayer")}selfAttn;linear1;linear2;norm1;norm2;dropout1;dropout2;dropoutFFN;_activation;_activationName;normFirst;_dModel;_nhead;_dimFeedforward;_dropout;_layerNormEps;_batchFirst;constructor(e,t,r=2048,s=.1,o="relu",i=1e-5,a=!0,u=!1){super(),this.selfAttn=new ua(e,t,s,!0,null,null,a),this.linear1=new Sn(e,r),this.linear2=new Sn(r,e),this.norm1=new no(e,i),this.norm2=new no(e,i),this.dropout1=new Sr(s),this.dropout2=new Sr(s),this.dropoutFFN=new Sr(s),this._activation=pE(o),this._activationName=o,this.normFirst=u,this._dModel=e,this._nhead=t,this._dimFeedforward=r,this._dropout=s,this._layerNormEps=i,this._batchFirst=a}forward(e,t=null,r=null,s=!1){return this.normFirst?this._forwardPreNorm(e,t,r,s):this._forwardPostNorm(e,t,r,s)}_forwardPostNorm(e,t,r,s){let o=this.selfAttn.forward(e,e,e,t,r,s);o=this.norm1.forward(ae(e,this.dropout1.forward(o)));let i=this._activation(this.linear1.forward(o));return i=this.linear2.forward(this.dropoutFFN.forward(i)),o=this.norm2.forward(ae(o,this.dropout2.forward(i))),o}_forwardPreNorm(e,t,r,s){let o=this.norm1.forward(e),i=this.selfAttn.forward(o,o,o,t,r,s);i=ae(e,this.dropout1.forward(i));let a=this._activation(this.linear1.forward(this.norm2.forward(i)));return a=this.linear2.forward(this.dropoutFFN.forward(a)),i=ae(i,this.dropout2.forward(a)),i}},wu=class extends ge{static{l(this,"TransformerDecoderLayer")}selfAttn;crossAttn;linear1;linear2;norm1;norm2;norm3;dropout1;dropout2;dropout3;dropoutFFN;_activation;_activationName;normFirst;_dModel;_nhead;_dimFeedforward;_dropout;_layerNormEps;_batchFirst;constructor(e,t,r=2048,s=.1,o="relu",i=1e-5,a=!0,u=!1){super(),this.selfAttn=new ua(e,t,s,!0,null,null,a),this.crossAttn=new ua(e,t,s,!0,null,null,a),this.linear1=new Sn(e,r),this.linear2=new Sn(r,e),this.norm1=new no(e,i),this.norm2=new no(e,i),this.norm3=new no(e,i),this.dropout1=new Sr(s),this.dropout2=new Sr(s),this.dropout3=new Sr(s),this.dropoutFFN=new Sr(s),this._activation=pE(o),this._activationName=o,this.normFirst=u,this._dModel=e,this._nhead=t,this._dimFeedforward=r,this._dropout=s,this._layerNormEps=i,this._batchFirst=a}forward(e,t,r=null,s=null,o=null,i=null,a=!1){return this.normFirst?this._forwardPreNorm(e,t,r,s,o,i,a):this._forwardPostNorm(e,t,r,s,o,i,a)}_forwardPostNorm(e,t,r,s,o,i,a){let u=this.selfAttn.forward(e,e,e,r,o,a);u=this.norm1.forward(ae(e,this.dropout1.forward(u)));let p=this.crossAttn.forward(u,t,t,s,i);u=this.norm2.forward(ae(u,this.dropout2.forward(p)));let c=this._activation(this.linear1.forward(u));return c=this.linear2.forward(this.dropoutFFN.forward(c)),u=this.norm3.forward(ae(u,this.dropout3.forward(c))),u}_forwardPreNorm(e,t,r,s,o,i,a){let u=this.norm1.forward(e),p=this.selfAttn.forward(u,u,u,r,o,a);p=ae(e,this.dropout1.forward(p));let c=this.norm2.forward(p),f=this.crossAttn.forward(c,t,t,s,i);p=ae(p,this.dropout2.forward(f));let d=this._activation(this.linear1.forward(this.norm3.forward(p)));return d=this.linear2.forward(this.dropoutFFN.forward(d)),p=ae(p,this.dropout3.forward(d)),p}};function a6(n){return new bu(n._dModel,n._nhead,n._dimFeedforward,n._dropout,n._activationName,n._layerNormEps,n._batchFirst,n.normFirst)}l(a6,"_cloneEncoderLayer");function u6(n){return new wu(n._dModel,n._nhead,n._dimFeedforward,n._dropout,n._activationName,n._layerNormEps,n._batchFirst,n.normFirst)}l(u6,"_cloneDecoderLayer");var dc=class extends ge{static{l(this,"TransformerEncoder")}layers;norm;constructor(e,t,r=null){super(),this.layers=new aa(Array.from({length:t},()=>a6(e))),this.norm=r}forward(e,t=null,r=null,s=!1){let o=e;for(let i of this.layers)o=i.forward(o,t,r,s);return this.norm&&(o=this.norm.forward(o)),o}},hc=class extends ge{static{l(this,"TransformerDecoder")}layers;norm;constructor(e,t,r=null){super(),this.layers=new aa(Array.from({length:t},()=>u6(e))),this.norm=r}forward(e,t,r=null,s=null,o=null,i=null,a=!1){let u=e;for(let p of this.layers)u=p.forward(u,t,r,s,o,i,a);return this.norm&&(u=this.norm.forward(u)),u}},pd=class extends ge{static{l(this,"Transformer")}encoder;decoder;dModel;constructor({dModel:e=512,nhead:t=8,numEncoderLayers:r=6,numDecoderLayers:s=6,dimFeedforward:o=2048,dropout:i=.1,activation:a="relu",batchFirst:u=!0,normFirst:p=!1,layerNormEps:c=1e-5}={}){super();let f=new bu(e,t,o,i,a,c,u,p),d=new wu(e,t,o,i,a,c,u,p);this.encoder=new dc(f,r),this.decoder=new hc(d,s),this.dModel=e}forward(e,t,r=null,s=null,o=null,i=null,a=null,u=null){let p=this.encoder.forward(e,r,i);return this.decoder.forward(t,p,s,o,a,u)}static generateSquareSubsequentMask(e){let t=Tn([e,e]),r=t._impl.storage.data;for(let s=0;s<e;s++)for(let o=0;o<e;o++)r[s*e+o]=o<=s?0:-1/0;return t}};be();kt();var fd=class extends ge{static{l(this,"PositionalEncoding")}dropoutLayer;pe;constructor(e,t=5e3,r=.1){super(),this.dropoutLayer=new Sr(r);let s=Ve([1,t,e]),o=s._impl.storage.data;for(let i=0;i<t;i++)for(let a=0;a<e;a+=2){let u=i*Math.exp(-(a*Math.log(1e4))/e);o[i*e+a]=Math.sin(u),a+1<e&&(o[i*e+a+1]=Math.cos(u))}this.pe=s,this.registerBuffer("pe",s)}forward(e){let t=e.shape[1],r=this.pe.narrow(1,0,t);return this.dropoutLayer.forward(ae(e,r))}};var fE={...lc,...S_,...GS,...k_,...A_,...WS,...B_,...ZS,...XS};var hE={};at(hE,{BatchSampler:()=>vu,DataLoader:()=>md,Dataset:()=>Tu,MapDataset:()=>hd,RandomSampler:()=>Su,Sampler:()=>la,SequentialSampler:()=>xu,TensorDataset:()=>dd,defaultCollate:()=>ku});be();var Tu=class{static{l(this,"Dataset")}get length(){throw new Error("Subclass must implement get length()")}get(e){throw new Error("Subclass must implement get(index)")}*[Symbol.iterator](){let e=this.length;for(let t=0;t<e;t++)yield this.get(t)}},dd=class extends Tu{static{l(this,"TensorDataset")}_tensors;_length;constructor(...e){if(super(),e.length===0)throw new Error("TensorDataset requires at least one tensor");let t=e[0].shape[0];for(let r=1;r<e.length;r++)if(e[r].shape[0]!==t)throw new Error(`Size mismatch at dim 0: tensor 0 has ${t}, tensor ${r} has ${e[r].shape[0]}`);this._tensors=e,this._length=t}get length(){return this._length}get(e){let t=new Array(this._tensors.length);for(let r=0;r<this._tensors.length;r++)t[r]=en(this._tensors[r],0,e);return t}},hd=class extends Tu{static{l(this,"MapDataset")}_dataset;_transform;constructor(e,t){super(),this._dataset=e,this._transform=t}get length(){return this._dataset.length}get(e){return this._transform(this._dataset.get(e))}};var la=class{static{l(this,"Sampler")}*[Symbol.iterator](){throw new Error("Subclass must implement [Symbol.iterator]()")}},xu=class extends la{static{l(this,"SequentialSampler")}_dataSource;constructor(e){super(),this._dataSource=e}*[Symbol.iterator](){let e=this._dataSource.length;for(let t=0;t<e;t++)yield t}},Su=class extends la{static{l(this,"RandomSampler")}_dataSource;constructor(e){super(),this._dataSource=e}*[Symbol.iterator](){let e=this._dataSource.length,t=new Int32Array(e);for(let r=0;r<e;r++)t[r]=r;for(let r=e-1;r>0;r--){let s=Math.floor(Math.random()*(r+1)),o=t[r];t[r]=t[s],t[s]=o}for(let r=0;r<e;r++)yield t[r]}},vu=class extends la{static{l(this,"BatchSampler")}_dropLast;_sampler;_batchSize;constructor(e,t,r=!1){super(),this._sampler=e,this._batchSize=t,this._dropLast=r}*[Symbol.iterator](){let e=[];for(let t of this._sampler)e.push(t),e.length===this._batchSize&&(yield e,e=[]);e.length>0&&!this._dropLast&&(yield e)}};sr();os();rs();function l6(n,e=0){let t=n.length,r=n[0].shape,s=n[0].dtype,o=[...r];o.splice(e,0,t);let i=r.reduce((c,f)=>c*f,1),a=St(s),u=new a(t*i),p=u;for(let c=0;c<t;c++){let f=n[c],d=f.data,h=f._impl?f._impl.storageOffset:0;if(f.isContiguous)p.set(d.subarray(h,h+i),c*i);else{let m=f.shape,g=f.strides,y=f._impl.storage.data,_=m.length,b=new Int32Array(_),w=h;for(let x=0;x<i;x++){p[c*i+x]=y[w];for(let S=_-1;S>=0;S--){if(b[S]++,b[S]<m[S]){w+=g[S];break}w-=(m[S]-1)*g[S],b[S]=0}}}}return eu(u,o,s)}l(l6,"stackTensors");function ku(n){let e=n[0];if(e instanceof Ze)return l6(n,0);if(typeof e=="number")return Wt(n);if(Array.isArray(e)){let t=new Array(e.length);for(let r=0;r<e.length;r++){let s=new Array(n.length);for(let o=0;o<n.length;o++)s[o]=n[o][r];t[r]=ku(s)}return t}if(e!==null&&typeof e=="object"&&e.constructor===Object){let t=Object.keys(e),r={};for(let s=0;s<t.length;s++){let o=t[s],i=new Array(n.length);for(let a=0;a<n.length;a++)i[a]=n[a][o];r[o]=ku(i)}return r}throw new Error(`defaultCollate: unsupported element type ${typeof e}`)}l(ku,"defaultCollate");var md=class{static{l(this,"DataLoader")}_dataset;_collate;_batchSampler;_batchSize;_dropLast;constructor(e,t={}){if(this._dataset=e,this._collate=t.collate??(r=>ku(r)),t.batchSampler!=null){if(t.batchSize!=null||t.shuffle||t.sampler!=null||t.dropLast)throw new Error("batchSampler is mutually exclusive with batchSize, shuffle, sampler, and dropLast");this._batchSampler=t.batchSampler,this._batchSize=null,this._dropLast=null}else{let r=t.batchSize??1,s=t.dropLast??!1;this._batchSize=r,this._dropLast=s;let o;if(t.sampler!=null){if(t.shuffle)throw new Error("sampler and shuffle are mutually exclusive");o=t.sampler}else o=t.shuffle?new Su(e):new xu(e);this._batchSampler=new vu(o,r,s)}}get dataset(){return this._dataset}get length(){let e=this._dataset.length;if(this._batchSize==null){let t=0;for(let r of this._batchSampler)t++;return t}return this._dropLast?Math.floor(e/this._batchSize):Math.ceil(e/this._batchSize)}*[Symbol.iterator](){for(let e of this._batchSampler){let t=new Array(e.length);for(let r=0;r<e.length;r++)t[r]=this._dataset.get(e[r]);yield this._collate(t)}}};var TE={};at(TE,{DEFAULT_SPECIALS:()=>F_,TOKENIZER_FORMAT:()=>yc,TOKENIZER_VERSION:()=>_c,Tokenizer:()=>yd,Vocab:()=>ca});var ca=class n{static{l(this,"Vocab")}_tokenToId;_idToToken;constructor(e=[]){this._tokenToId=new Map,this._idToToken=[];for(let t of e)this.add(t)}add(e){let t=this._tokenToId.get(e);return t===void 0&&(t=this._idToToken.length,this._tokenToId.set(e,t),this._idToToken.push(e)),t}getId(e,t=-1){let r=this._tokenToId.get(e);return r===void 0?t:r}getToken(e){return e>=0&&e<this._idToToken.length?this._idToToken[e]:void 0}has(e){return this._tokenToId.has(e)}get size(){return this._idToToken.length}tokens(){return this._idToToken.slice()}static fromTokens(e){if(!Array.isArray(e))throw new Error("mlfw tokenizer: vocab must be an array");let t=new n,r=new Set;for(let s of e){if(typeof s!="string")throw new Error("mlfw tokenizer: vocab entries must be strings");if(r.has(s))throw new Error(`mlfw tokenizer: duplicate vocab token '${s}'`);r.add(s),t.add(s)}return t}};var L_=class n{static{l(this,"WordStrategy")}_lowercase;constructor({lowercase:e=!1}={}){this._lowercase=e}fit(e,t){}segment(e){return(this._lowercase?String(e).toLowerCase():String(e)).split(/\s+/).filter(Boolean)}detokenize(e){return e.join(" ")}toJSON(){return{lowercase:this._lowercase}}static fromJSON(e={}){return new n({lowercase:e.lowercase??!1})}};var C_=class n{static{l(this,"CharStrategy")}_lowercase;constructor({lowercase:e=!1}={}){this._lowercase=e}fit(e,t){}segment(e){let t=this._lowercase?String(e).toLowerCase():String(e);return Array.from(t)}detokenize(e){return e.join("")}toJSON(){return{lowercase:this._lowercase}}static fromJSON(e={}){return new n({lowercase:e.lowercase??!1})}};function D_(n,e){return n+""+e}l(D_,"pairKey");var P_=class n{static{l(this,"BpeStrategy")}_numMerges;_lowercase;_eow;_ranks;_encodeCache;constructor({numMerges:e=1e3,lowercase:t=!1,endOfWord:r="</w>"}={}){this._numMerges=e,this._lowercase=t,this._eow=r,this._ranks=new Map,this._encodeCache=new Map}_pretokenize(e){return(this._lowercase?String(e).toLowerCase():String(e)).split(/\s+/).filter(Boolean)}_baseSymbols(e){let t=Array.from(e);return t.push(this._eow),t}fit(e,t){this._encodeCache=new Map;let r=new Map;for(let h of e)for(let m of this._pretokenize(h))r.set(m,(r.get(m)||0)+1);let s=new Map;for(let h of r.keys())s.set(h,this._baseSymbols(h));let o=new Map,i=new Map,a=[],u=l((h,m)=>h[0]>m[0]||h[0]===m[0]&&h[1]<m[1],"higher"),p=l((h,m)=>{a.push([h,m]);let g=a.length-1;for(;g>0;){let y=g-1>>1;if(u(a[g],a[y])){let _=a[y];a[y]=a[g],a[g]=_,g=y}else break}},"pushHeap"),c=l(()=>{let h=a[0],m=a.pop();if(a.length>0){a[0]=m;let g=0,y=a.length;for(;;){let _=g,b=2*g+1,w=2*g+2;if(b<y&&u(a[b],a[_])&&(_=b),w<y&&u(a[w],a[_])&&(_=w),_===g)break;let x=a[_];a[_]=a[g],a[g]=x,g=_}}return h},"popHeap"),f=l(h=>{let m=s.get(h),g=r.get(h);for(let y=0;y+1<m.length;y++){let _=D_(m[y],m[y+1]),b=(o.get(_)||0)+g;o.set(_,b),p(b,_);let w=i.get(_);w||(w=new Set,i.set(_,w)),w.add(h)}},"addWordPairs"),d=l(h=>{let m=s.get(h),g=r.get(h);for(let y=0;y+1<m.length;y++){let _=D_(m[y],m[y+1]),b=(o.get(_)||0)-g;b<=0?o.delete(_):(o.set(_,b),p(b,_));let w=i.get(_);w&&w.delete(h)}},"removeWordPairs");for(let h of r.keys())f(h);this._ranks=new Map;for(let h=0;h<this._numMerges;h++){let m=null,g=0;for(;a.length>0;){let[S,T]=c(),v=o.get(T);if(v===S&&v>0){m=T,g=v;break}}if(m===null||g<=0)break;let y=m.indexOf(""),_=m.slice(0,y),b=m.slice(y+1),w=_+b;this._ranks.set(m,h);let x=i.get(m);if(!x||x.size===0)break;for(let S of[...x]){d(S);let T=s.get(S),v=[];for(let A=0;A<T.length;A++)A+1<T.length&&T[A]===_&&T[A+1]===b?(v.push(w),A++):v.push(T[A]);s.set(S,v),f(S)}}}_encodeWord(e){let t=this._encodeCache.get(e);if(t!==void 0)return t;let r=this._baseSymbols(e);for(;r.length>1;){let s=1/0,o=-1;for(let i=0;i+1<r.length;i++){let a=this._ranks.get(D_(r[i],r[i+1]));a!==void 0&&a<s&&(s=a,o=i)}if(o<0)break;r=r.slice(0,o).concat(r[o]+r[o+1],r.slice(o+2))}return this._encodeCache.set(e,r),r}segment(e){let t=[];for(let r of this._pretokenize(e))for(let s of this._encodeWord(r))t.push(s);return t}detokenize(e){let t="";for(let r of e)t+=r;return t.split(this._eow).join(" ").trim()}toJSON(){let e=[...this._ranks.entries()].map(([t])=>{let r=t.indexOf("");return[t.slice(0,r),t.slice(r+1)]});return{numMerges:this._numMerges,lowercase:this._lowercase,endOfWord:this._eow,merges:e}}static fromJSON(e={}){if(!Array.isArray(e.merges))throw new Error("mlfw tokenizer: bpe strategy merges must be an array");let t=new n({numMerges:e.numMerges??e.merges.length,lowercase:e.lowercase??!1,endOfWord:e.endOfWord??"</w>"});t._ranks=new Map;for(let r=0;r<e.merges.length;r++){let s=e.merges[r];if(!Array.isArray(s)||s.length!==2||s.some(o=>typeof o!="string"))throw new Error("mlfw tokenizer: bpe merges must be string pairs");t._ranks.set(D_(s[0],s[1]),r)}return t}};os();import{writeFileSync as mE,appendFileSync as c6,readFileSync as gE,existsSync as p6,mkdirSync as f6,readdirSync as d6,unlinkSync as h6,renameSync as m6}from"node:fs";import{resolve as gd}from"node:path";var vn={readFile(n){return gE(gd(n),"utf8")},readBinary(n){return gE(gd(n))},writeFile(n,e){mE(n,e)},writeBinary(n,e){mE(gd(n),e)},appendFile(n,e){c6(n,e)},exists(n){return p6(n)},mkdir(n){f6(n,{recursive:!0})},readdir(n){return d6(n)},remove(n){h6(n)},rename(n,e){m6(gd(n),gd(e))}};var yc="mlfw-tokenizer",_c=1,F_=Object.freeze({pad:"<pad>",unk:"<unk>",bos:"<bos>",eos:"<eos>"}),yE=`${yc}-v${_c}`,$_={word:L_,char:C_,bpe:P_},gc=Object.freeze(["pad","unk","bos","eos"]),yd=class n{static{l(this,"Tokenizer")}_mode;_specials;_maxVocab;_strategy;_vocab;constructor(e={}){let t=e.mode??"word",r=$_[t];if(!r)throw new Error(`Unknown tokenizer mode '${t}'. Available: ${Object.keys($_).join(", ")}`);this._mode=t,this._specials=_E(e.specialTokens),bE(this._specials),this._maxVocab=e.vocabSize??null,wE(this._maxVocab,this._specials),this._strategy=new r(e),this._vocab=null}get mode(){return this._mode}get vocabSize(){return this._fitVocab().size}get padId(){return this._specialId("pad")}get unkId(){return this._specialId("unk")}get bosId(){return this._specialId("bos")}get eosId(){return this._specialId("eos")}_ensureFit(){if(!this._vocab)throw new Error("Tokenizer must be fit() on a corpus before use")}_fitVocab(){return this._ensureFit(),this._vocab}_specialId(e){return this._fitVocab().getId(this._specials[e])}fit(e){let t=Array.isArray(e)?e:[e];this._strategy.fit(t,{vocabSize:this._maxVocab});let r=new ca(Object.values(this._specials));if(this._maxVocab==null)for(let s of t)for(let o of this._strategy.segment(s))r.add(o);else{let s=new Map;for(let a of t)for(let u of this._strategy.segment(a))s.set(u,(s.get(u)||0)+1);let o=[...s.entries()].sort((a,u)=>u[1]-a[1]),i=this._maxVocab-r.size;for(let a=0;a<o.length&&a<i;a++)r.add(o[a][0])}return this._vocab=r,this}toJSON(){let e=this._fitVocab(),t=this._strategy.toJSON();return{format:yc,version:_c,mode:this._mode,config:{vocabSize:this._maxVocab,...t},specialTokens:{...this._specials},vocab:e.tokens(),strategy:t}}save(e){if(typeof e!="string")throw new Error("mlfw tokenizer: save(path) requires a file path string");let t=e+".tmp";vn.writeFile(t,y6(this.toJSON())),vn.rename(t,e)}static load(e){if(typeof e!="string")throw new Error("mlfw tokenizer: load(path) requires a file path string");let t=vn.readFile(e),r=typeof t=="string"?t:new TextDecoder().decode(t);return n.fromJSON(_6(r))}static fromJSON(e){g6(e);let t=$_[e.mode],r=new n({mode:e.mode,vocabSize:e.config.vocabSize,specialTokens:e.specialTokens});r._strategy=t.fromJSON(e.strategy),r._vocab=ca.fromTokens(e.vocab);for(let s of Object.values(r._specials))if(!r._vocab.has(s))throw new Error(`mlfw tokenizer: special token '${s}' is missing from vocab`);return r}encode(e,t={}){let r=this._fitVocab(),s=[];t.addBos&&s.push(this.bosId);let o=this.unkId;for(let i of this._strategy.segment(e))s.push(r.getId(i,o));return t.addEos&&s.push(this.eosId),s}decode(e,t={}){let r=this._fitVocab(),s=t.skipSpecial??!0,o=new Set(Object.values(this._specials).map(a=>r.getId(a))),i=[];for(let a of e){if(s&&o.has(a))continue;let u=r.getToken(a);u!==void 0&&i.push(u)}return this._strategy.detokenize(i)}encodeBatch(e,t={}){this._ensureFit();let s=(Array.isArray(e)?e:[e]).map(p=>this.encode(p,t)),o=t.maxLen??s.reduce((p,c)=>Math.max(p,c.length),0),i=t.padId??this.padId,a=s.length,u=new Int32Array(a*o).fill(i);for(let p=0;p<a;p++){let c=s[p],f=Math.min(c.length,o);for(let d=0;d<f;d++)u[p*o+d]=c[d]}return Wt(u,{shape:[a,o],dtype:"i32"})}};function _E(n){if(Array.isArray(n)){let e={...F_};for(let t=0;t<n.length&&t<gc.length;t++)e[gc[t]]=n[t];return e}return{...F_,...n??{}}}l(_E,"normalizeSpecials");function bE(n){let e=new Set;for(let t of gc){let r=n[t];if(typeof r!="string"||r.length===0)throw new Error(`mlfw tokenizer: special token '${t}' must be a non-empty string`);if(e.has(r))throw new Error(`mlfw tokenizer: duplicate special token '${r}'`);e.add(r)}}l(bE,"validateSpecials");function wE(n,e){if(n!=null&&(!Number.isInteger(n)||n<Object.keys(e).length))throw new Error(`mlfw tokenizer: vocabSize must be an integer >= ${Object.keys(e).length}`)}l(wE,"validateVocabSize");function g6(n){if(!n||typeof n!="object")throw new Error("mlfw tokenizer: artifact must be an object");let e=n;if(e.format!==yc)throw new Error("mlfw tokenizer: unrecognized tokenizer format");if(e.version!==_c)throw new Error(`mlfw tokenizer: unsupported tokenizer version ${e.version}`);if(typeof e.mode!="string"||!(e.mode in $_))throw new Error(`mlfw tokenizer: unknown tokenizer mode '${e.mode}'`);if(!e.config||typeof e.config!="object")throw new Error("mlfw tokenizer: config must be an object");let t=e.config,r=_E(x6(e.specialTokens));if(bE(r),wE(S6(t.vocabSize),r),!Array.isArray(e.vocab))throw new Error("mlfw tokenizer: vocab must be an array");if(!e.strategy||typeof e.strategy!="object")throw new Error("mlfw tokenizer: strategy must be an object")}l(g6,"validateTokenizerData");function y6(n){let e=[yE,Nu("m",[n.mode]),Nu("z",[n.config.vocabSize==null?"":String(n.config.vocabSize)]),Nu("s",gc.map(t=>n.specialTokens[t]))];if(n.mode==="bpe"){let t=[];for(let r of n.strategy.merges)t.push(r[0],r[1]);e.push(Nu("g",[String(n.strategy.lowercase?1:0),String(n.strategy.numMerges),n.strategy.endOfWord])),e.push(Nu("r",t))}else e.push(Nu("g",[String(n.strategy.lowercase?1:0)]));return e.push(Nu("v",n.vocab)),e.join(`
+`)}l(y6,"serializeTokenizer");function _6(n){if(!n.startsWith(yE))throw new Error("mlfw tokenizer: unrecognized tokenizer format");let e=new Map;for(let p of n.split(/\r?\n/).slice(1)){if(!p)continue;let c=p.indexOf("	");if(c<0)throw new Error("mlfw tokenizer: malformed tokenizer artifact");e.set(p.slice(0,c),b6(p.slice(c+1)))}let t=mc(e,"m")[0],r=mc(e,"z")[0],s=mc(e,"s"),o=mc(e,"g"),i=mc(e,"v"),a=Object.fromEntries(gc.map((p,c)=>[p,s[c]])),u={lowercase:o[0]==="1"};if(t==="bpe"){let p=mc(e,"r");if(p.length%2!==0)throw new Error("mlfw tokenizer: bpe merges must be string pairs");u.numMerges=Number(o[1]),u.endOfWord=o[2],u.merges=[];for(let c=0;c<p.length;c+=2)u.merges.push([p[c],p[c+1]])}return{format:yc,version:_c,mode:t,config:{vocabSize:r===""?null:Number(r),...u},specialTokens:a,vocab:i,strategy:u}}l(_6,"parseTokenizer");function Nu(n,e){return`${n}	${e.map(w6).join("	")}`}l(Nu,"record");function b6(n){return n.split("	").map(T6)}l(b6,"parseRecord");function w6(n){return String(n).replace(/\\/g,"\\\\").replace(/\t/g,"\\t").replace(/\n/g,"\\n").replace(/\r/g,"\\r")}l(w6,"escapeField");function T6(n){let e="";for(let t=0;t<n.length;t++){let r=n[t];if(r!=="\\"){e+=r;continue}let s=n[++t];s==="t"?e+="	":s==="n"?e+=`
+`:s==="r"?e+="\r":s==="\\"?e+="\\":e+=s??""}return e}l(T6,"unescapeField");function mc(n,e){let t=n.get(e);if(!t)throw new Error(`mlfw tokenizer: missing '${e}' record`);return t}l(mc,"requiredField");function x6(n){if(n==null)return;if(Array.isArray(n))return n.map(String);if(typeof n!="object")return;let e=n,t={};for(let r of gc){let s=e[r];typeof s=="string"&&(t[r]=s)}return t}l(x6,"asSpecialTokensInput");function S6(n){return n==null||typeof n=="number"?n:Number(n)}l(S6,"asNullableNumber");var tC={};at(tC,{Adam:()=>bd,AdamW:()=>wd,CosineAnnealingLR:()=>xd,FusedAdam:()=>lw,FusedOptimizer:()=>Fc,FusedSGD:()=>uw,GradScaler:()=>cw,LRScheduler:()=>Au,Optimizer:()=>fs,ReduceLROnPlateau:()=>Sd,SGD:()=>_d,StepLR:()=>Td,clipGradNorm_:()=>vd,clipGradValue_:()=>kd});var fs=class{static{l(this,"Optimizer")}_defaults;_paramGroups;_state;_paramIndex;_nextId;constructor(e,t){this._defaults={...t},this._paramGroups=[],this._state=new Map,this._paramIndex=new WeakMap,this._nextId=0;let r=v6(e);for(let s of r)this._addParamGroup(s)}get paramGroups(){return this._paramGroups}get defaults(){return this._defaults}step(){throw new Error(`${this.constructor.name}.step() not implemented`)}zeroGrad(e=!0){for(let t of this._paramGroups)for(let r of t.params)r.grad!==null&&(e?r.grad=null:r.grad._impl.storage.data.fill(0))}stateDict(){let e=new Map;for(let[t,r]of this._state){let s={};for(let o of Object.keys(r)){let i=r[o];s[o]=xE(i)?SE(i):i}e.set(t,s)}return{state:e,paramGroups:this._paramGroups.map(t=>{let r={};for(let s of Object.keys(t)){if(s==="params")continue;let o=t[s];r[s]=Array.isArray(o)?[...o]:o}return r})}}loadStateDict(e){for(let t=0;t<this._paramGroups.length;t++){let r=e.paramGroups[t];for(let s of Object.keys(r))this._paramGroups[t][s]=Array.isArray(r[s])?[...r[s]]:r[s]}this._state=new Map;for(let[t,r]of e.state){let s={};for(let o of Object.keys(r)){let i=r[o];s[o]=xE(i)?SE(i):i}this._state.set(t,s)}}_addParamGroup(e){let t={...this._defaults,params:[]};for(let s of Object.keys(e))s!=="params"&&(t[s]=e[s]);let r=Array.isArray(e.params)?e.params:[...e.params];for(let s of r){if(this._paramIndex.has(s))throw new Error("Parameter appears in more than one parameter group");this._paramIndex.set(s,this._nextId++)}t.params=r,this._paramGroups.push(t)}_getParamId(e){return this._paramIndex.get(e)}_getState(e){let t=this._paramIndex.get(e),r=this._state.get(t);return r||(r={},this._state.set(t,r)),r}};function v6(n){if(n==null)throw new Error("Optimizer requires at least one parameter");let e=Array.isArray(n)?n:[...n];if(e.length===0)throw new Error("Optimizer got an empty parameter list");return k6(e[0])?e:[{params:e}]}l(v6,"normalizeParamGroups");function xE(n){return n instanceof Float32Array||n instanceof Float64Array||n instanceof Int32Array||n instanceof Int16Array||n instanceof Int8Array||n instanceof Uint8Array||n instanceof Uint16Array||n instanceof Uint32Array}l(xE,"isTypedArray");function SE(n){return n.slice()}l(SE,"cloneTypedArray");function k6(n){return typeof n=="object"&&n!==null&&"params"in n}l(k6,"isParamGroupInput");var _d=class extends fs{static{l(this,"SGD")}constructor(e,{lr:t=.01,momentum:r=0,dampening:s=0,weightDecay:o=0,nesterov:i=!1}={}){if(i&&(r===0||s!==0))throw new Error("Nesterov momentum requires momentum > 0 and dampening = 0");super(e,{lr:t,momentum:r,dampening:s,weightDecay:o,nesterov:i})}step(){for(let e of this._paramGroups){let t=e.lr,r=e.momentum,s=e.dampening,o=e.weightDecay,i=e.nesterov;for(let a of e.params){if(a.grad===null)continue;let u=a._impl.storage.data,p=a.grad._impl.storage.data,c=u.length,f=this._getState(a);if(r===0)if(o===0)for(let d=0;d<c;d++)u[d]-=t*p[d];else for(let d=0;d<c;d++)u[d]-=t*(p[d]+o*u[d]);else{let d=f.momentumBuffer;if(d){d=d;for(let h=0;h<c;h++){let m=o!==0?p[h]+o*u[h]:p[h];d[h]=r*d[h]+(1-s)*m}}else{d=new u.constructor(c);for(let h=0;h<c;h++)d[h]=o!==0?p[h]+o*u[h]:p[h];f.momentumBuffer=d}if(i)for(let h=0;h<c;h++){let m=o!==0?p[h]+o*u[h]:p[h];u[h]-=t*(m+r*d[h])}else for(let h=0;h<c;h++)u[h]-=t*d[h]}a._impl.bumpVersion()}}}};wn();qn();var bd=class extends fs{static{l(this,"Adam")}constructor(e,{lr:t=.001,betas:r=[.9,.999],eps:s=1e-8,weightDecay:o=0,amsgrad:i=!1}={}){super(e,{lr:t,betas:[...r],eps:s,weightDecay:o,amsgrad:i})}step(){let e=lS();for(let t of this._paramGroups){let r=t.lr,s=t.betas,o=t.eps,i=t.weightDecay,a=t.amsgrad,[u,p]=s;for(let c of t.params){if(c.grad===null)continue;let f=this._getState(c);if(e&&!a&&c.device&&c.device.type==="gpu"){f.step=(f.step||0)+1;let v=f.step,A=1-Math.pow(u,v),N=1-Math.pow(p,v);if(e(c,f,{beta1:u,beta2:p,omb1:1-u,omb2:1-p,eps:o,stepSize:r/A,bc2sqrt:Math.sqrt(N),wd:i,lr:r})){c._impl.bumpVersion();continue}}let d=c._impl.storage.data,h=c.grad._impl.storage.data,m=d.length,g=d.constructor;f.expAvg===void 0&&(f.step===void 0&&(f.step=0),f.expAvg=new g(m),f.expAvgSq=new g(m),a&&(f.maxExpAvgSq=new g(m))),f.step=f.step+1;let y=f.step,_=f.expAvg,b=f.expAvgSq,w=1-Math.pow(u,y),x=1-Math.pow(p,y),S=r/w,T=Math.sqrt(x);if(a){let v=f.maxExpAvgSq;for(let A=0;A<m;A++){let N=i!==0?h[A]+i*d[A]:h[A];_[A]=u*_[A]+(1-u)*N,b[A]=p*b[A]+(1-p)*N*N,b[A]>v[A]&&(v[A]=b[A]),d[A]-=S*_[A]/(Math.sqrt(v[A])/T+o)}}else for(let v=0;v<m;v++){let A=i!==0?h[v]+i*d[v]:h[v];_[v]=u*_[v]+(1-u)*A,b[v]=p*b[v]+(1-p)*A*A,d[v]-=S*_[v]/(Math.sqrt(b[v])/T+o)}c._impl.bumpVersion()}}}};var wd=class extends fs{static{l(this,"AdamW")}constructor(e,{lr:t=.001,betas:r=[.9,.999],eps:s=1e-8,weightDecay:o=.01,amsgrad:i=!1}={}){super(e,{lr:t,betas:[...r],eps:s,weightDecay:o,amsgrad:i})}step(){for(let e of this._paramGroups){let t=e.lr,r=e.betas,s=e.eps,o=e.weightDecay,i=e.amsgrad,[a,u]=r;for(let p of e.params){if(p.grad===null)continue;let c=p._impl.storage.data,f=p.grad._impl.storage.data,d=c.length,h=this._getState(p),m=c.constructor;h.step===void 0&&(h.step=0,h.expAvg=new m(d),h.expAvgSq=new m(d),i&&(h.maxExpAvgSq=new m(d))),h.step=h.step+1;let g=h.step,y=h.expAvg,_=h.expAvgSq,b=1-Math.pow(a,g),w=1-Math.pow(u,g),x=t/b,S=Math.sqrt(w);if(o!==0){let T=1-t*o;for(let v=0;v<d;v++)c[v]*=T}if(i){let T=h.maxExpAvgSq;for(let v=0;v<d;v++)y[v]=a*y[v]+(1-a)*f[v],_[v]=u*_[v]+(1-u)*f[v]*f[v],_[v]>T[v]&&(T[v]=_[v]),c[v]-=x*y[v]/(Math.sqrt(T[v])/S+s)}else for(let T=0;T<d;T++)y[T]=a*y[T]+(1-a)*f[T],_[T]=u*_[T]+(1-u)*f[T]*f[T],c[T]-=x*y[T]/(Math.sqrt(_[T])/S+s);p._impl.bumpVersion()}}}};var Au=class{static{l(this,"LRScheduler")}_optimizer;_baseLRs;_lastEpoch;_lastLR;constructor(e,t=-1){this._optimizer=e,this._baseLRs=e.paramGroups.map(r=>r.lr),this._lastEpoch=t,this._lastLR=null}_init(){this.step()}getLR(){throw new Error(`${this.constructor.name}.getLR() not implemented`)}getLastLR(){return this._lastLR}step(){this._lastEpoch++;let e=this.getLR();this._lastLR=e;let t=this._optimizer.paramGroups;for(let r=0;r<t.length;r++)t[r].lr=e[r]}},Td=class extends Au{static{l(this,"StepLR")}_stepSize;_gamma;constructor(e,t,r=.1,s=-1){super(e,s),this._stepSize=t,this._gamma=r,this._init()}getLR(){let e=Math.pow(this._gamma,Math.floor(this._lastEpoch/this._stepSize));return this._baseLRs.map(t=>t*e)}},xd=class extends Au{static{l(this,"CosineAnnealingLR")}_tMax;_etaMin;constructor(e,t,r=0,s=-1){super(e,s),this._tMax=t,this._etaMin=r,this._init()}getLR(){let e=(1+Math.cos(Math.PI*this._lastEpoch/this._tMax))/2;return this._baseLRs.map(t=>this._etaMin+(t-this._etaMin)*e)}},Sd=class{static{l(this,"ReduceLROnPlateau")}_optimizer;_mode;_factor;_patience;_threshold;_thresholdMode;_cooldown;_minLR;_eps;_best;_numBadEpochs;_cooldownCounter;constructor(e,{mode:t="min",factor:r=.1,patience:s=10,threshold:o=1e-4,thresholdMode:i="rel",cooldown:a=0,minLR:u=0,eps:p=1e-8}={}){this._optimizer=e,this._mode=t,this._factor=r,this._patience=s,this._threshold=o,this._thresholdMode=i,this._cooldown=a,this._minLR=u,this._eps=p,this._best=t==="min"?1/0:-1/0,this._numBadEpochs=0,this._cooldownCounter=0}step(e){if(e===void 0)throw new Error("ReduceLROnPlateau.step() requires a metric value");this._cooldownCounter>0&&(this._cooldownCounter--,this._numBadEpochs=0),this._isBetter(e)?(this._best=e,this._numBadEpochs=0):this._numBadEpochs++,this._numBadEpochs>this._patience&&(this._reduceAllLRs(),this._cooldownCounter=this._cooldown,this._numBadEpochs=0)}_isBetter(e){return this._mode==="min"?this._thresholdMode==="rel"?e<this._best*(1-this._threshold):e<this._best-this._threshold:this._thresholdMode==="rel"?e>this._best*(1+this._threshold):e>this._best+this._threshold}_reduceAllLRs(){for(let e of this._optimizer.paramGroups){let t=e.lr,r=Math.max(t*this._factor,this._minLR);t-r>this._eps&&(e.lr=r)}}};be();function vE(n){return(Array.isArray(n)?n:[...n]).filter(t=>t.grad!==null)}l(vE,"_gradParams");function vd(n,e,t=2){let r=vE(n);if(r.length===0)return 0;let s;if(t===2){let i=null;for(let a of r){let u=wt(ee(a.grad,a.grad));i=i===null?u:ae(i,u)}s=Math.sqrt(Number(i.item()))}else if(t===1/0){let i=null;for(let a of r){let u=No(Hp(a.grad));i=i===null?u:ta(i,u)}s=Number(i.item())}else{let i=null;for(let a of r){let u=wt(vo(Hp(a.grad),t));i=i===null?u:ae(i,u)}s=Math.pow(Number(i.item()),1/t)}let o=e/(s+1e-6);if(o<1)for(let i of r)i.grad=ee(i.grad,o);return s}l(vd,"clipGradNorm_");function kd(n,e){for(let t of vE(n))t.grad=Mg(t.grad,-e,e)}l(kd,"clipGradValue_");_s();Re();aw();yo();var oo="f32",Ft=l(n=>new I([n],oo),"VEC"),$c=new I([],oo);function Ms(n,e,t){return n.broadcast(e,[t],[]).getResult(0)}l(Ms,"bcast");var Fc=class extends fs{static{l(this,"FusedOptimizer")}_target;_kernels;constructor(e,t,r=null){super(e,t),this._target=r||go(),this._kernels=new Map}_kernel(e){let t=this._kernels.get(e);if(!t){let r=this._buildGraph(e);t=eC(r,this._target,{fusion:{enabled:!0}}),this._kernels.set(e,t)}return t}_buildGraph(e){throw new Error(`${this.constructor.name}._buildGraph() not implemented`)}},uw=class extends Fc{static{l(this,"FusedSGD")}constructor(e,{lr:t=.01,momentum:r=0,dampening:s=0,weightDecay:o=0,nesterov:i=!1}={},a=null){if(i&&(r===0||s!==0))throw new Error("Nesterov momentum requires momentum > 0 and dampening = 0");super(e,{lr:t,momentum:r,dampening:s,weightDecay:o,nesterov:i},a)}_buildGraph(e){let t=this._defaults.momentum,r=this._defaults.dampening,s=this._defaults.weightDecay,o=this._defaults.nesterov,i=t!==0,a=i?[Ft(e),Ft(e),Ft(e),$c]:[Ft(e),Ft(e),$c],u=i?[Ft(e),Ft(e)]:[Ft(e)];return mp("sgd_update",a,u,(c,f)=>{let d=f[0],h=f[1],m=i?f[3]:f[2],g=Ms(c,m,e),y=h;if(s!==0){let v=c.scalarConstant(s,oo).getResult(0);y=c.add(h,c.mul(Ms(c,v,e),d).getResult(0)).getResult(0)}if(!i){let v=c.sub(d,c.mul(g,y).getResult(0)).getResult(0);c.returnOp([v]);return}let _=f[2],b=Ms(c,c.scalarConstant(t,oo).getResult(0),e),w=Ms(c,c.scalarConstant(1-r,oo).getResult(0),e),x=c.add(c.mul(b,_).getResult(0),c.mul(w,y).getResult(0)).getResult(0),S=x;o&&(S=c.add(y,c.mul(b,x).getResult(0)).getResult(0));let T=c.sub(d,c.mul(g,S).getResult(0)).getResult(0);c.returnOp([T,x])})}step(){for(let e of this._paramGroups){let t=e.lr,s=e.momentum!==0,o=new Float32Array([t]);for(let i of e.params){if(i.grad===null)continue;let a=i._impl.storage.data,u=i.grad._impl.storage.data,p=a.length,c=this._kernel(p);if(s){let f=this._getState(i),d=f.momentumBuffer;d||(d=new a.constructor(p),f.momentumBuffer=d),d=d,c.run("sgd_update",a,u,d,o,a,d)}else c.run("sgd_update",a,u,o,a);i._impl.bumpVersion()}}}},lw=class extends Fc{static{l(this,"FusedAdam")}constructor(e,{lr:t=.001,betas:r=[.9,.999],eps:s=1e-8,weightDecay:o=0,amsgrad:i=!1}={},a=null){super(e,{lr:t,betas:[...r],eps:s,weightDecay:o,amsgrad:i},a)}_buildGraph(e){let t=this._defaults.betas,r=this._defaults.eps,s=this._defaults.weightDecay,o=this._defaults.amsgrad,[i,a]=t,u=o?[Ft(e),Ft(e),Ft(e),Ft(e),Ft(e),$c,$c]:[Ft(e),Ft(e),Ft(e),Ft(e),$c,$c],p=o?[Ft(e),Ft(e),Ft(e),Ft(e)]:[Ft(e),Ft(e),Ft(e)];return mp("adam_update",u,p,(c,f)=>{let d=f[0],h=f[1],m=f[2],g=f[3],y=o?f[4]:null,_=o?f[5]:f[4],b=o?f[6]:f[5],w=h;if(s!==0){let H=Ms(c,c.scalarConstant(s,oo).getResult(0),e);w=c.add(h,c.mul(H,d).getResult(0)).getResult(0)}let x=Ms(c,c.scalarConstant(i,oo).getResult(0),e),S=Ms(c,c.scalarConstant(a,oo).getResult(0),e),T=Ms(c,c.scalarConstant(1-i,oo).getResult(0),e),v=Ms(c,c.scalarConstant(1-a,oo).getResult(0),e),A=Ms(c,c.scalarConstant(r,oo).getResult(0),e),N=c.add(c.mul(x,m).getResult(0),c.mul(T,w).getResult(0)).getResult(0),k=c.mul(w,w).getResult(0),O=c.add(c.mul(S,g).getResult(0),c.mul(v,k).getResult(0)).getResult(0),R=O,B=null;o&&(B=c.maximum(y,O).getResult(0),R=B);let E=c.add(c.div(c.sqrt(R).getResult(0),Ms(c,b,e)).getResult(0),A).getResult(0),F=c.mul(Ms(c,_,e),c.div(N,E).getResult(0)).getResult(0),C=c.sub(d,F).getResult(0);c.returnOp(o?[C,N,O,B]:[C,N,O])})}step(){for(let e of this._paramGroups){let t=e.lr,r=e.betas,s=e.amsgrad,[o,i]=r;for(let a of e.params){if(a.grad===null)continue;let u=a._impl.storage.data,p=a.grad._impl.storage.data,c=u.length,f=this._getState(a);f.step===void 0&&(f.step=0,f.expAvg=new u.constructor(c),f.expAvgSq=new u.constructor(c),s&&(f.maxExpAvgSq=new u.constructor(c))),f.step=f.step+1;let d=f.step,h=1-Math.pow(o,d),m=1-Math.pow(i,d),g=new Float32Array([t/h]),y=new Float32Array([Math.sqrt(m)]),_=this._kernel(c),b=f.expAvg,w=f.expAvgSq;if(s){let x=f.maxExpAvgSq;_.run("adam_update",u,p,b,w,x,g,y,u,b,w,x)}else _.run("adam_update",u,p,b,w,g,y,u,b,w);a._impl.bumpVersion()}}}};var cw=class{static{l(this,"GradScaler")}enabled;_scale;_growthFactor;_backoffFactor;_growthInterval;_growthTracker;_foundInf;_unscaled;constructor(e={}){this.enabled=e.enabled!==!1,this._scale=e.initScale??65536,this._growthFactor=e.growthFactor??2,this._backoffFactor=e.backoffFactor??.5,this._growthInterval=e.growthInterval??2e3,this._growthTracker=0,this._foundInf=!1,this._unscaled=new WeakSet}getScale(){return this.enabled?this._scale:1}get growthTracker(){return this._growthTracker}scale(e){if(!this.enabled)return e;let t=e._impl.storage.data;for(let r=0;r<t.length;r++)t[r]*=this._scale;return e._impl.bumpVersion&&e._impl.bumpVersion(),e}unscale_(e){if(!this.enabled)return!1;let t=1/this._scale,r=!1;for(let s of e.paramGroups)for(let o of s.params){if(o.grad===null||o.grad===void 0)continue;let i=o.grad._impl.storage.data;for(let a=0;a<i.length;a++){let u=i[a]*t;Number.isFinite(u)||(r=!0),i[a]=u}}return this._unscaled.add(e),r&&(this._foundInf=!0),r}step(e){return this.enabled?(this._unscaled.has(e)||this.unscale_(e),this._foundInf?!1:(e.step(),!0)):(e.step(),!0)}update(e){this.enabled&&(e!==void 0?(this._scale=e,this._growthTracker=0):this._foundInf?(this._scale*=this._backoffFactor,this._growthTracker=0):(this._growthTracker++,this._growthTracker>=this._growthInterval&&(this._scale*=this._growthFactor,this._growthTracker=0)),this._foundInf=!1,this._unscaled=new WeakSet)}};mk();bw();yo();var $C={};at($C,{Accuracy:()=>dh,CSVLogger:()=>lh,Callback:()=>Zn,CallbackConnector:()=>jc,ConfusionMatrix:()=>yh,ConsoleLogger:()=>Du,EarlyStopping:()=>oh,EvaluationLoop:()=>Yc,F1Score:()=>gh,FitLoop:()=>Hc,GradientAccumulationScheduler:()=>uh,HOOKS:()=>yk,LearningRateMonitor:()=>ih,LightningModule:()=>th,Logger:()=>ki,LoggerConnector:()=>Uc,MeanMetric:()=>ch,Metric:()=>lr,MetricAccumulator:()=>zc,MetricCollection:()=>fh,ModelCheckpoint:()=>Ta,Precision:()=>hh,PredictionLoop:()=>Jc,ProgressCallback:()=>wa,Recall:()=>mh,SingleDeviceStrategy:()=>Gc,Stage:()=>_a,SumMetric:()=>ph,Timer:()=>ah,Trainer:()=>sh,TrainerState:()=>Vc,TrainingLoop:()=>Wc,ValidationLoop:()=>Kc,applyCheckpoint:()=>Sk,deserializeCheckpoint:()=>Sw,loadCheckpoint:()=>xk,serializeCheckpoint:()=>rh});var th=class extends ge{static{l(this,"LightningModule")}_trainer;_logBuffer;_automaticOptimization;_currentOptimizers;_device;constructor(){super(),this._trainer=null,this._logBuffer=new Map,this._automaticOptimization=!0,this._currentOptimizers=[],this._device=null}get trainer(){return this._trainer}get currentEpoch(){return this._trainer?this._trainer.state.epoch:0}get globalStep(){return this._trainer?this._trainer.state.globalStep:0}get device(){return this._device}get logger(){return this._trainer?this._trainer.logger:null}get loggers(){return this._trainer?this._trainer.loggers:[]}get automaticOptimization(){return this._automaticOptimization}set automaticOptimization(e){this._automaticOptimization=e}get optimizers(){return this._currentOptimizers}trainingStep(e,t){throw new Error(`${this.constructor.name}.trainingStep() not implemented`)}validationStep(e,t){}testStep(e,t){}predictStep(e,t){return this.forward(Array.isArray(e)&&e.length===1?e[0]:e)}configureOptimizers(){throw new Error(`${this.constructor.name}.configureOptimizers() not implemented`)}onTrainEpochStart(){}onTrainEpochEnd(){}onValidationEpochStart(){}onValidationEpochEnd(){}onTestEpochStart(){}onTestEpochEnd(){}log(e,t,{onStep:r=null,onEpoch:s=null,reduceFx:o="mean",progBar:i=!1}={}){let a=this._trainer?this._trainer.state.stage:"training",u=a==="training",p=a!=="training";this._logBuffer.set(e,{value:t,onStep:r!==null?r:u,onEpoch:s!==null?s:p,reduceFx:o,progBar:i})}logDict(e,t={}){for(let r of Object.keys(e))this.log(r,e[r],t)}manualBackward(e){this._trainer&&this._trainer.strategy?this._trainer.strategy.backward(e):e.backward()}};function AC(n){if(!n)throw new Error("configureOptimizers() returned null/undefined");if(Array.isArray(n)){let t=[],r=[];for(let s=0;s<n.length;s++){let o=n[s];nU(o)?(t.push(o.optimizer),r.push(NC(o.lrScheduler))):(t.push(o),r.push(null))}return{optimizers:t,schedulerConfigs:r}}let e=n;if(e.optimizer)return{optimizers:[e.optimizer],schedulerConfigs:[NC(e.lrScheduler)]};if(tU(n))return{optimizers:[n],schedulerConfigs:[null]};throw new Error("configureOptimizers() returned an unrecognized format")}l(AC,"parseOptimizersConfig");function NC(n){return n?"scheduler"in n&&n.scheduler?{scheduler:n.scheduler,interval:n.interval||"epoch",frequency:n.frequency||1,monitor:n.monitor||null}:OC(n)?{scheduler:n,interval:"epoch",frequency:1,monitor:null}:null:null}l(NC,"normalizeSchedulerConfig");function OC(n){return typeof n=="object"&&n!==null&&typeof n.step=="function"}l(OC,"isSchedulerLike");function tU(n){return OC(n)}l(tU,"isOptimizerLike");function nU(n){return typeof n=="object"&&n!==null&&"optimizer"in n}l(nU,"isOptimizerConfig");var _a=(o=>(o.IDLE="idle",o.TRAINING="training",o.VALIDATING="validating",o.TESTING="testing",o.PREDICTING="predicting",o))(_a||{}),rU={mean:l(n=>n.count===0?0:n.sum/n.count,"mean"),sum:l(n=>n.sum,"sum"),min:l(n=>n.min,"min"),max:l(n=>n.max,"max"),last:l(n=>n.last,"last")};function sU(){return{sum:0,count:0,min:1/0,max:-1/0,last:0}}l(sU,"createAccumulator");function oU(n,e){let t=typeof e=="number"?e:e.item();n.sum+=t,n.count+=1,t<n.min&&(n.min=t),t>n.max&&(n.max=t),n.last=t}l(oU,"updateAccumulator");var zc=class{static{l(this,"MetricAccumulator")}_accumulators;_reduceFns;constructor(){this._accumulators=new Map,this._reduceFns=new Map}update(e,t,r="mean"){this._accumulators.has(e)||(this._accumulators.set(e,sU()),this._reduceFns.set(e,r)),oU(this._accumulators.get(e),t)}compute(e){let t=this._accumulators.get(e);if(!t)return;let r=rU[this._reduceFns.get(e)];return r(t)}computeAll(){let e={};for(let[t]of this._accumulators)e[t]=this.compute(t);return e}reset(){this._accumulators.clear(),this._reduceFns.clear()}has(e){return this._accumulators.has(e)}get size(){return this._accumulators.size}},Vc=class{static{l(this,"TrainerState")}stage;epoch;globalStep;maxEpochs;maxSteps;shouldStop;stepMetrics;epochMetrics;numTrainingBatches;numValBatches;_progBarMetrics;constructor(){this.stage="idle",this.epoch=0,this.globalStep=0,this.maxEpochs=0,this.maxSteps=-1,this.shouldStop=!1,this.stepMetrics=new zc,this.epochMetrics=new zc}resetEpochMetrics(){this.epochMetrics.reset()}resetStepMetrics(){this.stepMetrics.reset()}},Gc=class{static{l(this,"SingleDeviceStrategy")}device;constructor(){this.device=null}setup(e,t){this.device=t,(t.type==="gpu"||t.type==="webgpu")&&iU(e,"to")&&e.to(t)}toDevice(e){if(!this.device||this.device.type!=="gpu"&&this.device.type!=="webgpu")return e;if(aU(e))return e.to(this.device);if(Array.isArray(e)){let t=new Array(e.length);for(let r=0;r<e.length;r++)t[r]=this.toDevice(e[r]);return t}return e}backward(e){e.backward()}optimizerStep(e){e.step()}};function iU(n,e){return typeof n=="object"&&n!==null&&typeof n[e]=="function"}l(iU,"hasMethod");function aU(n){return typeof n=="object"&&n!==null&&"device"in n&&typeof n.to=="function"}l(aU,"hasDeviceTo");var yk=(E=>(E.ON_FIT_START="onFitStart",E.ON_FIT_END="onFitEnd",E.ON_TRAIN_START="onTrainStart",E.ON_TRAIN_END="onTrainEnd",E.ON_TRAIN_EPOCH_START="onTrainEpochStart",E.ON_TRAIN_EPOCH_END="onTrainEpochEnd",E.ON_TRAIN_BATCH_START="onTrainBatchStart",E.ON_TRAIN_BATCH_END="onTrainBatchEnd",E.ON_VALIDATION_START="onValidationStart",E.ON_VALIDATION_END="onValidationEnd",E.ON_VALIDATION_EPOCH_START="onValidationEpochStart",E.ON_VALIDATION_EPOCH_END="onValidationEpochEnd",E.ON_VALIDATION_BATCH_START="onValidationBatchStart",E.ON_VALIDATION_BATCH_END="onValidationBatchEnd",E.ON_TEST_START="onTestStart",E.ON_TEST_END="onTestEnd",E.ON_TEST_BATCH_START="onTestBatchStart",E.ON_TEST_BATCH_END="onTestBatchEnd",E.ON_PREDICT_START="onPredictStart",E.ON_PREDICT_END="onPredictEnd",E.ON_PREDICT_BATCH_START="onPredictBatchStart",E.ON_PREDICT_BATCH_END="onPredictBatchEnd",E.SETUP="setup",E.TEARDOWN="teardown",E.ON_BEFORE_BACKWARD="onBeforeBackward",E.ON_AFTER_BACKWARD="onAfterBackward",E.ON_BEFORE_OPTIMIZER_STEP="onBeforeOptimizerStep",E.ON_BEFORE_ZERO_GRAD="onBeforeZeroGrad",E.ON_SAVE_CHECKPOINT="onSaveCheckpoint",E.ON_LOAD_CHECKPOINT="onLoadCheckpoint",E))(yk||{}),jc=class{static{l(this,"CallbackConnector")}_callbacks;constructor(e=[]){this._callbacks=e}get callbacks(){return this._callbacks}add(e){this._callbacks.push(e)}dispatch(e,...t){for(let r=0;r<this._callbacks.length;r++){let s=this._callbacks[r][e];typeof s=="function"&&s.apply(this._callbacks[r],t)}}remove(e){let t=this._callbacks.indexOf(e);t!==-1&&this._callbacks.splice(t,1)}},Uc=class{static{l(this,"LoggerConnector")}_loggers;_state;constructor(e=[],t){this._loggers=Array.isArray(e)?e:[e],this._state=t}drain(e){let t=e._logBuffer;if(t.size!==0){for(let[r,s]of t){let{value:o,onStep:i,onEpoch:a,reduceFx:u,progBar:p}=s;if(a&&this._state.epochMetrics.update(r,o,u),i&&this._state.stepMetrics.update(r,o,u),p){this._state._progBarMetrics||(this._state._progBarMetrics=new Map);let c=typeof o=="number"?o:o.item();this._state._progBarMetrics.set(r,c)}}t.clear()}}flushStepMetrics(e){let t=this._state.stepMetrics.computeAll();if(Object.keys(t).length===0)return t;for(let r=0;r<this._loggers.length;r++)this._loggers[r].logMetrics(t,e);return this._state.stepMetrics.reset(),t}flushEpochMetrics(e){let t=this._state.epochMetrics.computeAll();if(Object.keys(t).length===0)return t;for(let r=0;r<this._loggers.length;r++)this._loggers[r].logMetrics(t,e);return this._state.epochMetrics.reset(),t}logHyperparams(e){for(let t=0;t<this._loggers.length;t++)this._loggers[t].logHyperparams(e)}};be();fi();ou();function ba(n,e){return n==null?e:typeof n=="number"?n>0&&n<=1?Math.max(1,Math.round(n*e)):Math.min(n,e):e}l(ba,"resolveLimit");async function qc(n){let{GradMode:e}=await Promise.resolve().then(()=>(ou(),HB)),t=e.isEnabled();e.setEnabled(!1);try{await n()}finally{e.setEnabled(t)}}l(qc,"noGradAsync");var Wc=class{static{l(this,"TrainingLoop")}async run(e,t,r,s,o){let i=r.state,a=r.callbackConnector,u=r.loggerConnector,p=r.strategy,c=r.accumulateGradBatches,f=ba(r.limitTrainBatches,t.length);i.numTrainingBatches=f,i.stage="training",e.train(),Oo.setEnabled(!0),e.onTrainEpochStart(),a.dispatch("onTrainEpochStart",r,e);let d=0;for(let m of t){if(d>=f||i.shouldStop)break;if(i.maxSteps>0&&i.globalStep>=i.maxSteps){i.shouldStop=!0;break}let g=p.toDevice(m);a.dispatch("onTrainBatchStart",r,e,g,d);let y;e.automaticOptimization?y=await this._automaticStep(e,g,d,r,s,o,p,c,a):y=await Promise.resolve(e.trainingStep(g,d)),u.drain(e),i.globalStep%r.logEveryNSteps===0&&u.flushStepMetrics(i.globalStep),a.dispatch("onTrainBatchEnd",r,e,y,g,d),iT(),i.globalStep++,d++}iT();let h=u.flushEpochMetrics(i.globalStep);return this._stepEpochSchedulers(o,i.epoch),e.onTrainEpochEnd(),a.dispatch("onTrainEpochEnd",r,e),h}async _automaticStep(e,t,r,s,o,i,a,u,p){if(s.compile)return this._compiledStep(e,t,r,s,o,i,u);if(s.cudaGraph)return this._graphedStep(e,t,s,o,i,a);let c=await Promise.resolve(e.trainingStep(t,r)),f=c,d=c;if(BC(c)&&(f=c.loss,d=c),u>1&&(f=Kt(f,u)),p.dispatch("onBeforeBackward",s,e,f),a.backward(f),p.dispatch("onAfterBackward",s,e),(r+1)%u===0){for(let m=0;m<o.length;m++)this._clipGradients(e,s),p.dispatch("onBeforeOptimizerStep",s,e,o[m]),a.optimizerStep(o[m]),p.dispatch("onBeforeZeroGrad",s,e,o[m]),o[m].zeroGrad();this._stepStepSchedulers(i,s.state.globalStep)}return d}async _compiledStep(e,t,r,s,o,i,a){let u=Array.isArray(t)?t:[t],p=l((..._)=>e.trainingStep(Array.isArray(t)?_:_[0],0),"callForward"),c;if(e.__compiledTrainStep)c=e.__compiledTrainStep(...u),_k(c)&&(c=await c);else{let{compileWithBackward:_}=await Promise.resolve().then(()=>(bw(),SC)),{CPUTarget:b,CUDATarget:w,WebGPUTarget:x}=await Promise.resolve().then(()=>(yo(),t1)),S=e._device&&e._device.type,T=S==="webgpu"?x():S==="gpu"?w():b();e.__compiledTrainStep=_({forward:p},u,{target:T,mode:s.compileMode});let v=e.log;e.log=()=>{};try{c=e.__compiledTrainStep(...u),_k(c)&&(c=await c)}finally{e.log=v}}let f=e.__compiledTrainStep,d=f.capturedParams(),{ones:h}=await Promise.resolve().then(()=>(kt(),EB)),m=f.backward(h(c.shape));_k(m)&&(m=await m);let g=m,y=g.length-d.length;for(let _=0;_<d.length;_++){let b=g[y+_];b&&(d[_].grad=b)}if((r+1)%a===0){for(let _=0;_<o.length;_++)this._clipGradients(e,s),o[_].step(),o[_].zeroGrad();this._stepStepSchedulers(i,s.state.globalStep)}return e.log&&e.log("train_loss",c),c}async _eagerTrainStepCore(e,t,r,s,o){let i=await Promise.resolve(e.trainingStep(t,0)),a=i;if(BC(i)&&(a=i.loss),s.backward(a),o&&o.gradientClipVal){let{deviceClipGradNorm:u}=await Promise.resolve().then(()=>(iy(),oy));u([...e.parameters()],o.gradientClipVal)}for(let u=0;u<r.length;u++)s.optimizerStep(r[u]),r[u].zeroGrad();return a}async _graphedStep(e,t,r,s,o,i){let a=await Promise.resolve().then(()=>(ug(),gO)),u=await Promise.resolve().then(()=>(qp(),EO)),p=await Promise.resolve().then(()=>(pi(),iO)),c=e.__eagerGraphRunner;if(c||(c=e.__eagerGraphRunner={phase:"warmup",seen:0},u.setEagerDeferred(!0),Nl(!0)),c.phase==="disabled")return this._eagerTrainStepCore(e,t,s,i,r);if(c.phase==="warmup"&&c.seen<r.cudaGraphWarmupSteps)return c.seen++,this._eagerTrainStepCore(e,t,s,i,r);let f=bk(t);if(c.phase==="warmup"){c.inputs=f.map(y=>{let _=y._impl.storage.rawData;return u.deviceBufferForInput(_),u.pinResident(_),{dptr:u.deviceBufferDptr(_)}});let h=e.log,m;try{a.beginEagerCapture(),e.log=()=>{};try{m=await this._eagerTrainStepCore(e,t,s,i,r)}finally{e.log=h}c.captured=a.endEagerCapture()}catch(y){try{a.endEagerCapture()}catch{}return e.log=h,u.clearCapturePins(),c.phase="disabled",c.captureError=y&&typeof y=="object"&&"message"in y?y.message:y,this._eagerTrainStepCore(e,t,s,i,r)}c.exec=c.captured.exec,c.lossDptr=u.deviceBufferDptr(m._impl.storage.rawData),c.lossScratch=new Float32Array(1),c.phase="replay",a.replay(c.exec),a.syncStream(),p.copyDeviceToHost(c.lossScratch,c.lossDptr);let g=c.lossScratch[0];return this._logGraphLoss(r,g),this._stepStepSchedulers(o,r.state.globalStep),g}for(let h=0;h<c.inputs.length&&h<f.length;h++)c.inputs[h].dptr&&p.copyHostToDeviceAsync(c.inputs[h].dptr,f[h]._impl.storage.rawData);a.replay(c.exec),a.syncStream(),p.copyDeviceToHost(c.lossScratch,c.lossDptr);let d=c.lossScratch[0];return this._logGraphLoss(r,d),this._stepStepSchedulers(o,r.state.globalStep),d}_logGraphLoss(e,t){e.state.stepMetrics.update("train_loss",t),e.state.epochMetrics.update("train_loss",t),e.state._progBarMetrics||(e.state._progBarMetrics=new Map),e.state._progBarMetrics.set("train_loss",t)}_clipGradients(e,t){if(!t.gradientClipVal)return;let r=[...e.parameters()];t.gradientClipAlgorithm==="norm"?vd(r,t.gradientClipVal):kd(r,t.gradientClipVal)}_stepStepSchedulers(e,t){if(e)for(let r=0;r<e.length;r++){let s=e[r];!s||s.interval!=="step"||s.monitor||(t+1)%s.frequency===0&&s.scheduler.step()}}_stepEpochSchedulers(e,t){if(e)for(let r=0;r<e.length;r++){let s=e[r];!s||s.interval!=="epoch"||s.monitor||(t+1)%s.frequency===0&&s.scheduler.step()}}};function bk(n,e=[]){if(n==null)return e;if(uU(n))return e.push(n),e;if(Array.isArray(n)){for(let t of n)bk(t,e);return e}if(typeof n=="object"){let t=n;for(let r of Object.keys(t))bk(t[r],e);return e}return e}l(bk,"_flattenTensors");function _k(n){return typeof n=="object"&&n!==null&&typeof n.then=="function"}l(_k,"isThenable");function BC(n){return typeof n=="object"&&n!==null&&!("backward"in n)&&"loss"in n}l(BC,"isLossObject");function uU(n){return typeof n=="object"&&n!==null&&"shape"in n&&typeof n.contiguous=="function"}l(uU,"isFlattenTensor");var Kc=class{static{l(this,"ValidationLoop")}async run(e,t,r,s){let o=r.state,i=r.callbackConnector,a=r.loggerConnector,u=o.stage;o.stage="validating",e.eval(),e.onValidationEpochStart(),i.dispatch("onValidationStart",r,e),i.dispatch("onValidationEpochStart",r,e);let p=ba(r.limitValBatches,t.length);o.numValBatches=p;let c=0;await qc(async()=>{for(let d of t){if(c>=p)break;i.dispatch("onValidationBatchStart",r,e,d,c);let h=await Promise.resolve(e.validationStep(d,c));a.drain(e),i.dispatch("onValidationBatchEnd",r,e,h,d,c),c++}});let f=a.flushEpochMetrics(o.globalStep);return this._stepPlateauSchedulers(s,f),e.onValidationEpochEnd(),i.dispatch("onValidationEpochEnd",r,e),i.dispatch("onValidationEnd",r,e),e.train(),o.stage=u,f}_stepPlateauSchedulers(e,t){if(e)for(let r=0;r<e.length;r++){let s=e[r];if(!s||!s.monitor)continue;let o=s.scheduler;if(typeof o.step=="function"&&o.step.length>0){let i=t[s.monitor];i!==void 0&&o.step(i)}}}};var Hc=class{static{l(this,"FitLoop")}_trainingLoop;_validationLoop;constructor(){this._trainingLoop=new Wc,this._validationLoop=new Kc}get trainingLoop(){return this._trainingLoop}get validationLoop(){return this._validationLoop}async run(e,t,r,s,o,i){let a=s.state,u=s.callbackConnector;u.dispatch("onTrainStart",s,e);for(let p=0;p<a.maxEpochs&&!(a.shouldStop||a.maxSteps>0&&a.globalStep>=a.maxSteps||(a.epoch=p,a.resetEpochMetrics(),await this._trainingLoop.run(e,t,s,o,i),a.shouldStop));p++)r&&this._shouldRunValidation(p,s)&&await this._validationLoop.run(e,r,s,i);u.dispatch("onTrainEnd",s,e)}_shouldRunValidation(e,t){let r=t.checkValEveryNEpoch;return(e+1)%r===0}};var Yc=class{static{l(this,"EvaluationLoop")}async run(e,t,r){let s=r.state,o=r.callbackConnector,i=r.loggerConnector,a=s.stage;s.stage="testing",e.eval(),e.onTestEpochStart(),o.dispatch("onTestStart",r,e),o.dispatch("onTestEpochStart",r,e);let u=ba(r.limitTestBatches,t.length),p=0;await qc(async()=>{for(let f of t){if(p>=u)break;o.dispatch("onTestBatchStart",r,e,f,p);let d=await Promise.resolve(e.testStep(f,p));i.drain(e),o.dispatch("onTestBatchEnd",r,e,d,f,p),p++}});let c=i.flushEpochMetrics(s.globalStep);return e.onTestEpochEnd(),o.dispatch("onTestEpochEnd",r,e),o.dispatch("onTestEnd",r,e),e.train(),s.stage=a,c}};var Jc=class{static{l(this,"PredictionLoop")}async run(e,t,r){let s=r.state,o=r.callbackConnector,i=r.strategy,a=s.stage;s.stage="predicting",e.eval(),o.dispatch("onPredictStart",r,e);let u=[],p=ba(r.limitTestBatches,t.length),c=0;return await qc(async()=>{for(let f of t){if(c>=p)break;let d=i.toDevice(f);o.dispatch("onPredictBatchStart",r,e,d,c);let h=await Promise.resolve(e.predictStep(d,c));await r._flushEagerInference(),u.push(h),o.dispatch("onPredictBatchEnd",r,e,h,d,c),c++}}),o.dispatch("onPredictEnd",r,e),e.train(),s.stage=a,u}};var ki=class{static{l(this,"Logger")}_name;_version;constructor({name:e="default",version:t=0}={}){this._name=e,this._version=t}get name(){return this._name}get version(){return this._version}logMetrics(e,t){throw new Error(`${this.constructor.name}.logMetrics() not implemented`)}logHyperparams(e){throw new Error(`${this.constructor.name}.logHyperparams() not implemented`)}finalize(){}};var Du=class extends ki{static{l(this,"ConsoleLogger")}_logFrequency;_callCount;constructor(e={}){super(e),this._logFrequency=e.logFrequency||1,this._callCount=0}logMetrics(e,t){if(this._callCount++,this._callCount%this._logFrequency!==0)return;let r=[`[step ${t}]`],s=Object.keys(e).sort();for(let o=0;o<s.length;o++){let i=e[s[o]];r.push(`${s[o]}: ${lU(i)}`)}console.log(r.join(" | "))}logHyperparams(e){let t=["[hyperparams]"],r=Object.keys(e).sort();for(let s=0;s<r.length;s++)t.push(`${r[s]}: ${e[r[s]]}`);console.log(t.join(" | "))}};function lU(n){return typeof n!="number"||Number.isInteger(n)?String(n):Math.abs(n)<.001&&n!==0?n.toExponential(3):n.toFixed(4)}l(lU,"formatNumber");var Zn=class{static{l(this,"Callback")}setup(e,t,r){}teardown(e,t,r){}onFitStart(e,t){}onFitEnd(e,t){}onTrainStart(e,t){}onTrainEnd(e,t){}onTrainEpochStart(e,t){}onTrainEpochEnd(e,t){}onTrainBatchStart(e,t,r,s){}onTrainBatchEnd(e,t,r,s,o){}onValidationStart(e,t){}onValidationEnd(e,t){}onValidationEpochStart(e,t){}onValidationEpochEnd(e,t){}onValidationBatchStart(e,t,r,s){}onValidationBatchEnd(e,t,r,s,o){}onTestStart(e,t){}onTestEnd(e,t){}onTestBatchStart(e,t,r,s){}onTestBatchEnd(e,t,r,s,o){}onPredictStart(e,t){}onPredictEnd(e,t){}onPredictBatchStart(e,t,r,s){}onPredictBatchEnd(e,t,r,s,o){}onBeforeBackward(e,t,r){}onAfterBackward(e,t){}onBeforeOptimizerStep(e,t,r){}onBeforeZeroGrad(e,t,r){}onSaveCheckpoint(e,t,r){}onLoadCheckpoint(e,t,r){}};var cU=["\u258F","\u258E","\u258D","\u258C","\u258B","\u258A","\u2589"];function RC(n){globalThis.process.stdout.write(n)}l(RC,"stdoutWrite");var wa=class extends Zn{static{l(this,"ProgressCallback")}_barLength;_trainBatchCount;_valBatchCount;_epochStartTime;_lastLen;_active;constructor({barLength:e=24}={}){super(),this._barLength=e,this._trainBatchCount=0,this._valBatchCount=0,this._epochStartTime=0,this._lastLen=0,this._active=!1}onTrainEpochStart(e,t){this._trainBatchCount=0,this._epochStartTime=Date.now();let r=this._trainTotal(e);r&&this._render("Epoch",e.state.epoch+1,e.state.maxEpochs,0,r,e.state)}onTrainBatchEnd(e,t,r,s,o){this._trainBatchCount++;let i=this._trainTotal(e);i&&this._render("Epoch",e.state.epoch+1,e.state.maxEpochs,this._trainBatchCount,i,e.state)}onTrainEpochEnd(e,t){let r=this._trainTotal(e);r&&this._render("Epoch",e.state.epoch+1,e.state.maxEpochs,r,r,e.state)}onTrainEnd(e,t){this._active&&RC(`
+`),this._active=!1,this._lastLen=0}onValidationEpochStart(e,t){this._valBatchCount=0,this._epochStartTime=Date.now()}onValidationBatchEnd(e,t,r,s,o){this._valBatchCount++;let i=this._valTotal(e);i&&this._render("Validation",null,null,this._valBatchCount,i,e.state)}onValidationEnd(e,t){let r=this._valTotal(e);r&&this._render("Validation",null,null,r,r,e.state)}_trainTotal(e){return e.state.numTrainingBatches??MC(e.limitTrainBatches)}_valTotal(e){return e.state.numValBatches??MC(e.limitValBatches)}_render(e,t,r,s,o,i){let a=o>0?Math.min(s/o,1):0,u=String(Math.round(a*100)).padStart(3," "),p=this._bar(a),c=t!==null?`${e} ${t}/${r}`:e,f=(Date.now()-this._epochStartTime)/1e3,d=f>0?s/f:0,h=d>0?(o-s)/d:0,m=`${IC(f)}<${IC(h)}, ${d.toFixed(2)}it/s`,g=this._formatProgBarMetrics(i),y=`${c}: ${u}%|${p}| ${s}/${o} [${m}${g}]`,_=Math.max(0,this._lastLen-y.length);RC("\r"+y+" ".repeat(_)),this._lastLen=y.length,this._active=!0}_bar(e){let t=this._barLength,r=e*t,s=Math.floor(r),o=Math.round((r-s)*8);if(o===8&&(s+=1,o=0),s>=t)return"\u2588".repeat(t);let i=o>0?cU[o-1]:"",a=t-s-(i?1:0);return"\u2588".repeat(s)+i+" ".repeat(a)}_formatProgBarMetrics(e){let t=e._progBarMetrics;if(!t||t.size===0)return"";let r=[];for(let[s,o]of t)r.push(`${s}=${pU(o)}`);return", "+r.join(", ")}};function MC(n){return n==null?null:typeof n=="number"&&n>1?n:null}l(MC,"resolveTotal");function IC(n){(!Number.isFinite(n)||n<0)&&(n=0);let e=Math.round(n),t=Math.floor(e/60),r=e%60;return`${String(t).padStart(2,"0")}:${String(r).padStart(2,"0")}`}l(IC,"fmtTime");function pU(n){return typeof n!="number"||Number.isInteger(n)?String(n):n.toFixed(4)}l(pU,"formatNum");function Ni(...n){return n.filter(e=>e!=null&&e!=="").join("/").replace(/\/+/g,"/")}l(Ni,"joinPath");sr();os();rs();var LC="mlfw-ckpt-v1",nh=8,fU={Int8Array,Uint8Array,Uint8ClampedArray,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array};function dU(n){return ArrayBuffer.isView(n)&&!(n instanceof DataView)}l(dU,"isTypedArray");function wk(n){return new Uint8Array(n.buffer,n.byteOffset,n.byteLength)}l(wk,"byteView");function hU(n){return n.constructor.name}l(hU,"typedArrayName");function mU(n){return"$buf"in n&&typeof n.$buf=="number"}l(mU,"isBufferNode");function gU(n){return"$map"in n&&Array.isArray(n.$map)}l(gU,"isMapNode");function rh(n){let e=[],t=Tw(n,e),r=0;for(let u of e)u.meta.offset=r,u.meta.length=u.bytes.length,r+=u.bytes.length;let s={format:LC,tree:t,buffers:e.map(u=>u.meta)},o=new TextEncoder().encode(JSON.stringify(s)),i=new Uint8Array(nh+o.length+r);new DataView(i.buffer).setBigUint64(0,BigInt(o.length),!0),i.set(o,nh);let a=nh+o.length;for(let u of e)i.set(u.bytes,a),a+=u.bytes.length;return i}l(rh,"serializeCheckpoint");function Sw(n){let e=n instanceof Uint8Array?n:new Uint8Array(n),t=new DataView(e.buffer,e.byteOffset,e.byteLength),r=Number(t.getBigUint64(0,!0)),s=nh+r,o=JSON.parse(new TextDecoder().decode(e.subarray(nh,s)));if(o.format!==LC)throw new Error("mlfw: unrecognized checkpoint format: "+o.format);let i=o.buffers.map(a=>yU(a,e,s));return xw(o.tree,i)}l(Sw,"deserializeCheckpoint");function Tw(n,e){if(n instanceof Ze)return EC(e,{kind:"tensor",dtype:n.dtype,shape:n.shape},wk(n._impl.storage.data));if(dU(n))return EC(e,{kind:"array",arrayType:hU(n)},wk(n));if(n instanceof Map){let t=[];for(let[r,s]of n)t.push([r,Tw(s,e)]);return{$map:t}}if(Array.isArray(n))return n.map(t=>Tw(t,e));if(n&&typeof n=="object"){let t={},r=n;for(let s of Object.keys(r))t[s]=Tw(r[s],e);return t}return n}l(Tw,"encodeNode");function EC(n,e,t){let r=n.length;return n.push({meta:e,bytes:t}),{$buf:r}}l(EC,"pushBuffer");function yU(n,e,t){let r=t+n.offset,s=e.subarray(r,r+n.length),o=n.kind==="tensor"?St(n.dtype):fU[n.arrayType],i=n.kind==="tensor"?void 0:n.arrayType;if(!o)throw new Error("mlfw: unknown typed array in checkpoint: "+i);let a=new o(n.length/o.BYTES_PER_ELEMENT);return wk(a).set(s),n.kind==="tensor"?eu(a,n.shape,n.dtype):a}l(yU,"decodeBuffer");function xw(n,e){if(n===null||typeof n!="object")return n;if(Array.isArray(n))return n.map(r=>xw(r,e));if(mU(n))return e[n.$buf];if(gU(n)){let r=new Map;for(let[s,o]of n.$map)r.set(s,xw(o,e));return r}let t={};for(let r of Object.keys(n))t[r]=xw(n[r],e);return t}l(xw,"decodeNode");var Tk=".ckpt",Ta=class extends Zn{static{l(this,"ModelCheckpoint")}_dirpath;_filename;_monitor;_mode;_saveTopK;_saveLast;_everyNEpochs;_bestK;_recent;_compareFn;_bestModelPath;_lastModelPath;constructor({dirpath:e="./lightning_logs/checkpoints",filename:t="epoch={epoch}-step={step}",monitor:r=null,mode:s="min",saveTopK:o=1,saveLast:i=!0,everyNEpochs:a=1}={}){super(),this._dirpath=e,this._filename=t,this._monitor=r,this._mode=s,this._saveTopK=o,this._saveLast=i,this._everyNEpochs=a,this._bestK=[],this._recent=[],this._compareFn=s==="min"?(u,p)=>u-p:(u,p)=>p-u,this._bestModelPath=null,this._lastModelPath=null}get bestModelPath(){return this._bestModelPath}get lastModelPath(){return this._lastModelPath}get bestKModels(){return this._bestK}onTrainEpochEnd(e,t){let r=e.state;if((r.epoch+1)%this._everyNEpochs!==0)return;this._ensureDir();let s=this._fillTemplate(r);if(this._saveLast){let c=Ni(this._dirpath,"last"+Tk);this._saveCheckpoint(t,e,c),this._lastModelPath=c}if(!this._monitor){let c=Ni(this._dirpath,s+Tk);if(this._saveCheckpoint(t,e,c),this._saveTopK>=0)for(this._recent.push(c);this._recent.length>this._saveTopK;)this._tryDelete(this._recent.shift());return}let i=r.epochMetrics.computeAll()[this._monitor];if(i===void 0)return;let a=Ni(this._dirpath,s+Tk),u={score:i,path:a};if(this._saveTopK<0){this._saveCheckpoint(t,e,a),this._bestK.push(u),this._updateBest();return}let p=this._findInsertIndex(i);if(p<this._saveTopK){if(this._saveCheckpoint(t,e,a),this._bestK.splice(p,0,u),this._bestK.length>this._saveTopK){let c=this._bestK.pop();this._tryDelete(c.path)}this._updateBest()}}_findInsertIndex(e){let t=0,r=this._bestK.length;for(;t<r;){let s=t+r>>>1;this._compareFn(e,this._bestK[s].score)<0?r=s:t=s+1}return t}_updateBest(){this._bestK.length>0&&(this._bestModelPath=this._bestK[0].path)}_saveCheckpoint(e,t,r){let s={epoch:t.state.epoch,globalStep:t.state.globalStep,modelState:e.stateDict()},o=e._currentOptimizers;o&&o.length>0&&(s.optimizerStates=o.map(a=>a.stateDict())),t.callbackConnector.dispatch("onSaveCheckpoint",t,e,s);let i=r+".tmp";vn.writeBinary(i,rh(s)),vn.rename(i,r)}_fillTemplate(e){return this._filename.replace("{epoch}",String(e.epoch)).replace("{step}",String(e.globalStep))}_ensureDir(){vn.exists(this._dirpath)||vn.mkdir(this._dirpath)}_tryDelete(e){try{vn.remove(e)}catch{}}};function xk(n){return Sw(vn.readBinary(n))}l(xk,"loadCheckpoint");function Sk(n,e,t=[]){let r=n;if(r.modelState&&e.loadStateDict(r.modelState),r.optimizerStates){let s=Math.min(t.length,r.optimizerStates.length);for(let o=0;o<s;o++)t[o].loadStateDict(r.optimizerStates[o])}return n}l(Sk,"applyCheckpoint");qn();var sh=class{static{l(this,"Trainer")}_state;_compile;_compileMode;_cudaGraph;_cudaGraphWarmupSteps;_accelerator;_precision;_gradientClipVal;_gradientClipAlgorithm;_accumulateGradBatches;_limitTrainBatches;_limitValBatches;_limitTestBatches;_valCheckInterval;_checkValEveryNEpoch;_logEveryNSteps;_deterministic;_defaultRootDir;_loggers;_strategy;_fitLoop;_evaluationLoop;_predictionLoop;_callbackConnector;_loggerConnector;_model;_webgpuMod;constructor({maxEpochs:e=10,maxSteps:t=-1,accelerator:r="auto",precision:s="f32",callbacks:o=[],logger:i=!0,enableCheckpointing:a=!1,enableProgress:u=!0,gradientClipVal:p=null,gradientClipAlgorithm:c="norm",accumulateGradBatches:f=1,limitTrainBatches:d=null,limitValBatches:h=null,limitTestBatches:m=null,valCheckInterval:g=1,checkValEveryNEpoch:y=1,logEveryNSteps:_=50,deterministic:b=!1,fastDevRun:w=!1,defaultRootDir:x="./lightning_logs",compile:S=!1,compileMode:T="separate",cudaGraph:v=!1,cudaGraphWarmupSteps:A=3}={}){if(this._state=new Vc,this._state.maxEpochs=e,this._state.maxSteps=t,this._compile=S,this._compileMode=T,this._cudaGraph=v,this._cudaGraphWarmupSteps=A,this._accelerator=r,this._precision=s,this._gradientClipVal=p,this._gradientClipAlgorithm=c,this._accumulateGradBatches=f,this._limitTrainBatches=d,this._limitValBatches=h,this._limitTestBatches=m,this._valCheckInterval=g,this._checkValEveryNEpoch=y,this._logEveryNSteps=_,this._deterministic=b,this._defaultRootDir=x,w!==!1){let k=typeof w=="number"?w:1;this._limitTrainBatches=k,this._limitValBatches=k,this._limitTestBatches=k,this._state.maxEpochs=1}this._loggers=this._resolveLoggers(i),this._strategy=new Gc,this._fitLoop=new Hc,this._evaluationLoop=new Yc,this._predictionLoop=new Jc;let N=[...o];u&&!N.some(k=>k instanceof wa)&&N.push(new wa),a&&!N.some(k=>k instanceof Ta)&&N.push(new Ta({dirpath:x})),this._callbackConnector=new jc(N),this._loggerConnector=new Uc(this._loggers,this._state),this._model=null,this._webgpuMod=null}get state(){return this._state}get strategy(){return this._strategy}get callbackConnector(){return this._callbackConnector}get loggerConnector(){return this._loggerConnector}get fitLoop(){return this._fitLoop}get gradientClipVal(){return this._gradientClipVal}get gradientClipAlgorithm(){return this._gradientClipAlgorithm}get compile(){return this._compile}get compileMode(){return this._compileMode}get cudaGraph(){return this._cudaGraph}get cudaGraphWarmupSteps(){return this._cudaGraphWarmupSteps}get accumulateGradBatches(){return this._accumulateGradBatches}set accumulateGradBatches(e){this._accumulateGradBatches=e}get limitTrainBatches(){return this._limitTrainBatches}get limitValBatches(){return this._limitValBatches}get limitTestBatches(){return this._limitTestBatches}get checkValEveryNEpoch(){return this._checkValEveryNEpoch}get logEveryNSteps(){return this._logEveryNSteps}get shouldStop(){return this._state.shouldStop}set shouldStop(e){this._state.shouldStop=e}get currentEpoch(){return this._state.epoch}get globalStep(){return this._state.globalStep}get logger(){return this._loggers[0]||null}get loggers(){return this._loggers}get callbacks(){return this._callbackConnector.callbacks}get model(){return this._model}get defaultRootDir(){return this._defaultRootDir}async fit(e,t,r=null){this._model=e,e._trainer=this;let s=this._resolveDevice();this._guardEagerWebGPU(s,"fit",r!=null),this._guardCudaGraph(s,r!=null),e._device=s,await this._prepareDevice(s),this._strategy.setup(e,s);let{optimizers:o,schedulerConfigs:i}=AC(await Promise.resolve(e.configureOptimizers()));if(this._cudaGraph&&i&&i.some(a=>a&&a.scheduler))throw new Error("Trainer(cudaGraph=true) v1 requires a constant learning rate: LR schedulers change lr, but lr is baked into the captured graph. Remove the scheduler or disable cudaGraph.");if(e._currentOptimizers=o,this._loggerConnector.logHyperparams(this._extractHyperparams(e,o)),this._callbackConnector.dispatch("setup",this,e,"training"),this._callbackConnector.dispatch("onFitStart",this,e),this._state.shouldStop=!1,await this._fitLoop.run(e,t,r,this,o,i),s===bg){let{teardownAfterFit:a}=await Promise.resolve().then(()=>(iy(),oy));a(e,o)}this._callbackConnector.dispatch("onFitEnd",this,e),this._callbackConnector.dispatch("teardown",this,e,"training");for(let a=0;a<this._loggers.length;a++)this._loggers[a].finalize()}async validate(e,t){e._trainer=this,this._model=e;let r=this._resolveDevice();this._guardEagerWebGPU(r,"validate"),e._device=r,await this._prepareDevice(r),this._strategy.setup(e,r),this._callbackConnector.dispatch("setup",this,e,"validating");let s=await this._fitLoop.validationLoop.run(e,t,this,null);this._callbackConnector.dispatch("teardown",this,e,"validating");for(let o=0;o<this._loggers.length;o++)this._loggers[o].finalize();return s}async test(e,t){e._trainer=this,this._model=e;let r=this._resolveDevice();this._guardEagerWebGPU(r,"test"),e._device=r,await this._prepareDevice(r),this._strategy.setup(e,r),this._callbackConnector.dispatch("setup",this,e,"testing");let s=await this._evaluationLoop.run(e,t,this);this._callbackConnector.dispatch("teardown",this,e,"testing");for(let o=0;o<this._loggers.length;o++)this._loggers[o].finalize();return s}async predict(e,t){e._trainer=this,this._model=e;let r=this._resolveDevice();return e._device=r,await this._prepareDevice(r),this._strategy.setup(e,r),await this._predictionLoop.run(e,t,this)}_resolveDevice(){return this._accelerator==="gpu"?bg:this._accelerator==="wasm"?RT:this._accelerator==="webgpu"?xo:this._accelerator==="cpu"?Ws:Ws}_guardEagerWebGPU(e,t,r=!1){if(e.type==="webgpu"){if(t==="fit"){if(!this._compile)throw new Error('Trainer(accelerator="webgpu"): eager WebGPU is inference-only (CUSTOM_0 dispatch has no autograd key). Pass compile=true to train on WebGPU, or use predict() for eager inference.');if(r)throw new Error(`Trainer(accelerator="webgpu"): in-fit validation is unsupported \u2014 validationStep runs eagerly and reads scalar metrics via .item(), which WebGPU's asynchronous readback cannot serve, and there is no compiled validation path. Call fit() without a valLoader on WebGPU.`);return}throw new Error(`Trainer(accelerator="webgpu"): ${t}() reads scalar metrics synchronously via .item(), which WebGPU's asynchronous readback cannot serve eagerly. Use predict() for eager WebGPU inference, or train via compile=true.`)}}_guardCudaGraph(e,t=!1){if(this._cudaGraph){if(e.type!=="gpu")throw new Error('Trainer(cudaGraph=true) requires accelerator="gpu" (eager CUDA whole-step capture/replay).');if(this._compile)throw new Error("Trainer(cudaGraph=true) is incompatible with compile=true: CUDA graph capture targets the eager training step, not the compiled path.");if(this._gradientClipVal!=null&&this._gradientClipAlgorithm!=="norm")throw new Error('Trainer(cudaGraph=true) supports gradient_clip_algorithm="norm" only; "value" clipping is not yet device-side.');if(this._accumulateGradBatches!==1)throw new Error("Trainer(cudaGraph=true) v1 requires accumulateGradBatches=1.");if(t)throw new Error("Trainer(cudaGraph=true) v1 does not support in-fit validation. Call fit() without a valLoader.")}}async _prepareDevice(e){if(e.type==="gpu"){let{preloadCudaRuntime:t}=await Promise.resolve().then(()=>(_f(),vx));await t()}else if(e.type==="webgpu"){let{preloadWebGPU:t}=await Promise.resolve().then(()=>(_f(),vx));this._webgpuMod=await t()}}async _flushEagerInference(){this._webgpuMod&&await this._webgpuMod.flushWebGPUEager()}_resolveLoggers(e){return e===!1||e===null?[]:e===!0?[new Du]:Array.isArray(e)?e:[e]}_extractHyperparams(e,t){let r={maxEpochs:this._state.maxEpochs,maxSteps:this._state.maxSteps,accelerator:this._accelerator,precision:this._precision,accumulateGradBatches:this._accumulateGradBatches};for(let s=0;s<t.length;s++){let o=t[s],i=o.defaults||{},a=t.length>1?`optimizer_${s}_`:"";r[a+"optimizer"]=o.constructor.name,i.lr!==void 0&&(r[a+"lr"]=i.lr),i.weightDecay!==void 0&&(r[a+"weight_decay"]=i.weightDecay)}return r}};var oh=class extends Zn{static{l(this,"EarlyStopping")}_monitor;_patience;_mode;_minDelta;_checkOnTrainEpochEnd;_waitCount;_bestScore;_compareFn;constructor({monitor:e="val_loss",patience:t=3,mode:r="min",minDelta:s=0,checkOnTrainEpochEnd:o=!1}={}){super(),this._monitor=e,this._patience=t,this._mode=r,this._minDelta=s,this._checkOnTrainEpochEnd=o,this._waitCount=0,this._bestScore=null,this._compareFn=r==="min"?(i,a)=>i<a-s:(i,a)=>i>a+s}get monitor(){return this._monitor}get patience(){return this._patience}get bestScore(){return this._bestScore}get waitCount(){return this._waitCount}onValidationEnd(e,t){this._checkOnTrainEpochEnd||this._check(e)}onTrainEpochEnd(e,t){this._checkOnTrainEpochEnd&&this._check(e)}_check(e){let r=e.state.epochMetrics.computeAll()[this._monitor];if(r!==void 0){if(this._bestScore===null||this._compareFn(r,this._bestScore)){this._bestScore=r,this._waitCount=0;return}this._waitCount++,this._waitCount>=this._patience&&(e.shouldStop=!0)}}reset(){this._waitCount=0,this._bestScore=null}};var ih=class extends Zn{static{l(this,"LearningRateMonitor")}_logMomentum;_lrHistory;constructor({logMomentum:e=!1}={}){super(),this._logMomentum=e,this._lrHistory={}}get lrHistory(){return this._lrHistory}onTrainBatchStart(e,t,r,s){let o=t._currentOptimizers;if(o)for(let i=0;i<o.length;i++){let a=o[i].paramGroups;for(let u=0;u<a.length;u++){let p=o.length>1||a.length>1?`lr_opt${i}_group${u}`:"lr",c=a[u].lr;if(t.log(p,c,{onStep:!0,onEpoch:!1,progBar:!1}),this._lrHistory[p]||(this._lrHistory[p]=[]),this._lrHistory[p].push({step:e.state.globalStep,lr:c}),this._logMomentum&&a[u].momentum!==void 0){let f=p.replace("lr","momentum");t.log(f,a[u].momentum,{onStep:!0,onEpoch:!1})}}}}};var ah=class extends Zn{static{l(this,"Timer")}_fitStartTime;_epochStartTime;_epochDurations;_validationDurations;_totalTrainingTime;_valStartTime;constructor(){super(),this._fitStartTime=0,this._epochStartTime=0,this._epochDurations=[],this._validationDurations=[],this._totalTrainingTime=0,this._valStartTime=0}get epochDurations(){return this._epochDurations}get validationDurations(){return this._validationDurations}get totalTrainingTime(){return this._totalTrainingTime}onFitStart(e,t){this._fitStartTime=performance.now()}onFitEnd(e,t){this._totalTrainingTime=(performance.now()-this._fitStartTime)/1e3}onTrainEpochStart(e,t){this._epochStartTime=performance.now()}onTrainEpochEnd(e,t){this._epochDurations.push((performance.now()-this._epochStartTime)/1e3)}onValidationStart(e,t){this._valStartTime=performance.now()}onValidationEnd(e,t){this._validationDurations.push((performance.now()-this._valStartTime)/1e3)}};var uh=class extends Zn{static{l(this,"GradientAccumulationScheduler")}_scheduling;_sortedEpochs;constructor({scheduling:e}){super(),this._scheduling=new Map;let t=Object.entries(e);for(let r=0;r<t.length;r++)this._scheduling.set(Number(t[r][0]),t[r][1]);this._sortedEpochs=[...this._scheduling.keys()].sort((r,s)=>r-s)}onTrainEpochStart(e,t){let r=e.state.epoch;this._scheduling.has(r)&&(e.accumulateGradBatches=this._scheduling.get(r))}getCurrentAccumulation(e){let t=1;for(let r=0;r<this._sortedEpochs.length&&this._sortedEpochs[r]<=e;r++)t=this._scheduling.get(this._sortedEpochs[r]);return t}};var lh=class extends ki{static{l(this,"CSVLogger")}_saveDir;_flushInterval;_columns;_columnSet;_buffer;_filePath;_headerWritten;_version;constructor({saveDir:e="./lightning_logs",name:t="default",version:r=null,flushInterval:s=10}={}){super({name:t,version:r||0}),this._saveDir=e,this._flushInterval=s,this._columns=[],this._columnSet=new Set,this._buffer=[],this._filePath=null,this._headerWritten=!1,this._version=r}get logDir(){let e=this._version!==null?this._version:this._resolveVersion();return Ni(this._saveDir,this._name,`version_${e}`)}logMetrics(e,t){let r={step:t},s=Object.keys(e);for(let o=0;o<s.length;o++){let i=s[o];r[i]=e[i],this._columnSet.has(i)||(this._columnSet.add(i),this._columns.push(i),this._headerWritten=!1)}this._buffer.push(r),this._buffer.length>=this._flushInterval&&this._flush()}logHyperparams(e){this._ensureDir();let t=Ni(this.logDir,"hparams.json");vn.writeFile(t,JSON.stringify(e,null,2))}finalize(){this._buffer.length>0&&this._flush()}_flush(){this._ensureDir();let e=this._getFilePath();if(!this._headerWritten){let s=["step",...this._columns].join(",");vn.writeFile(e,s+`
+`),this._headerWritten=!0}let t=["step",...this._columns],r=[];for(let s=0;s<this._buffer.length;s++){let o=this._buffer[s],i=[];for(let a=0;a<t.length;a++){let u=o[t[a]];i.push(u!==void 0?String(u):"")}r.push(i.join(","))}vn.appendFile(e,r.join(`
+`)+`
+`),this._buffer.length=0}_getFilePath(){return this._filePath||(this._filePath=Ni(this.logDir,"metrics.csv")),this._filePath}_ensureDir(){let e=this.logDir;vn.exists(e)||vn.mkdir(e)}_resolveVersion(){if(this._version!==null)return this._version;let e=Ni(this._saveDir,this._name);if(!vn.exists(e))return this._version=0,0;let t=-1;try{let r=vn.readdir(e);for(let s=0;s<r.length;s++){let o=r[s].match(/^version_(\d+)$/);if(o){let i=parseInt(o[1],10);i>t&&(t=i)}}}catch{}return this._version=t+1,this._version}};var lr=class{static{l(this,"Metric")}_computed;constructor(){this._computed=null}update(e,t){throw new Error(`${this.constructor.name}.update() not implemented`)}compute(){throw new Error(`${this.constructor.name}.compute() not implemented`)}reset(){this._computed=null}forward(e,t){return this.update(e,t),this._computed=this.compute(),this._computed}get value(){return this._computed}};var ch=class extends lr{static{l(this,"MeanMetric")}_sum;_count;constructor(){super(),this._sum=0,this._count=0}update(e,t=1){let r=typeof e=="number"?e:e.item();this._sum+=r*t,this._count+=t}compute(){return this._count===0?0:this._sum/this._count}reset(){super.reset(),this._sum=0,this._count=0}},ph=class extends lr{static{l(this,"SumMetric")}_sum;constructor(){super(),this._sum=0}update(e){this._sum+=typeof e=="number"?e:e.item()}compute(){return this._sum}reset(){super.reset(),this._sum=0}};var fh=class{static{l(this,"MetricCollection")}_metrics;constructor(e={}){this._metrics=new Map;let t=Object.keys(e);for(let r=0;r<t.length;r++)this._metrics.set(t[r],e[t[r]])}add(e,t){return this._metrics.set(e,t),this}update(e,t){for(let[,r]of this._metrics)r.update(e,t)}compute(){let e={};for(let[t,r]of this._metrics)e[t]=r.compute();return e}reset(){for(let[,e]of this._metrics)e.reset()}forward(e,t){return this.update(e,t),this.compute()}get(e){return this._metrics.get(e)}has(e){return this._metrics.has(e)}get size(){return this._metrics.size}[Symbol.iterator](){return this._metrics.entries()}};var dh=class extends lr{static{l(this,"Accuracy")}_task;_numClasses;_topK;_threshold;_correct;_total;constructor({task:e="multiclass",numClasses:t=null,topK:r=1,threshold:s=.5}={}){super(),this._task=e,this._numClasses=t,this._topK=r,this._threshold=s,this._correct=0,this._total=0}update(e,t){let r=e._impl.storage.data,s=t._impl.storage.data;this._task==="binary"?this._updateBinary(r,s):this._task==="multiclass"?this._updateMulticlass(e,t):this._task==="multilabel"&&this._updateMultilabel(r,s)}compute(){return this._total===0?0:this._correct/this._total}reset(){super.reset(),this._correct=0,this._total=0}_updateBinary(e,t){let r=t.length;for(let s=0;s<r;s++)(e[s]>=this._threshold?1:0)===t[s]&&this._correct++;this._total+=r}_updateMulticlass(e,t){let r=e.shape;if(r.length<2){let u=e._impl.storage.data,p=t._impl.storage.data,c=p.length;for(let f=0;f<c;f++)Math.round(u[f])===p[f]&&this._correct++;this._total+=c;return}let s=r[0],o=r[1],i=e._impl.storage.data,a=t._impl.storage.data;if(this._topK===1)for(let u=0;u<s;u++){let p=0,c=i[u*o];for(let f=1;f<o;f++){let d=i[u*o+f];d>c&&(c=d,p=f)}p===a[u]&&this._correct++}else for(let u=0;u<s;u++){let p=_U(i,u*o,o,this._topK);for(let c=0;c<p.length;c++)if(p[c]===a[u]){this._correct++;break}}this._total+=s}_updateMultilabel(e,t){let r=t.length;for(let s=0;s<r;s++)(e[s]>=this._threshold?1:0)===t[s]&&this._correct++;this._total+=r}};function _U(n,e,t,r){let s=[];for(let o=0;o<t;o++){let i=n[e+o];s.length<r?(s.push({val:i,idx:o}),s.length===r&&bU(s)):i>s[0].val&&(s[0]={val:i,idx:o},CC(s,0))}return s.map(o=>o.idx)}l(_U,"topKIndices");function bU(n){for(let e=(n.length>>>1)-1;e>=0;e--)CC(n,e)}l(bU,"heapify");function CC(n,e){let t=n.length;for(;;){let r=e,s=2*e+1,o=2*e+2;if(s<t&&n[s].val<n[r].val&&(r=s),o<t&&n[o].val<n[r].val&&(r=o),r===e)break;let i=n[e];n[e]=n[r],n[r]=i,e=r}}l(CC,"siftDown");var hh=class extends lr{static{l(this,"Precision")}_task;_numClasses;_average;_tp;_fp;_support;constructor({task:e="multiclass",numClasses:t=2,average:r="macro"}={}){super(),this._task=e,this._numClasses=t,this._average=r,this._tp=new Int32Array(t),this._fp=new Int32Array(t),this._support=new Int32Array(t)}update(e,t){vk(e,t,this._tp,this._fp,null,this._support,this._task,this._numClasses)}compute(){return TU(this._tp,this._fp,this._support,this._numClasses,this._average)}reset(){super.reset(),this._tp.fill(0),this._fp.fill(0),this._support.fill(0)}},mh=class extends lr{static{l(this,"Recall")}_task;_numClasses;_average;_tp;_fn;_support;constructor({task:e="multiclass",numClasses:t=2,average:r="macro"}={}){super(),this._task=e,this._numClasses=t,this._average=r,this._tp=new Int32Array(t),this._fn=new Int32Array(t),this._support=new Int32Array(t)}update(e,t){vk(e,t,this._tp,null,this._fn,this._support,this._task,this._numClasses)}compute(){return xU(this._tp,this._fn,this._support,this._numClasses,this._average)}reset(){super.reset(),this._tp.fill(0),this._fn.fill(0),this._support.fill(0)}},gh=class extends lr{static{l(this,"F1Score")}_task;_numClasses;_average;_tp;_fp;_fn;_support;constructor({task:e="multiclass",numClasses:t=2,average:r="macro"}={}){super(),this._task=e,this._numClasses=t,this._average=r,this._tp=new Int32Array(t),this._fp=new Int32Array(t),this._fn=new Int32Array(t),this._support=new Int32Array(t)}update(e,t){vk(e,t,this._tp,this._fp,this._fn,this._support,this._task,this._numClasses)}compute(){let e=DC(this._tp,this._fp,this._numClasses),t=PC(this._tp,this._fn,this._numClasses),r=new Float64Array(this._numClasses);for(let s=0;s<this._numClasses;s++){let o=e[s]+t[s];r[s]=o>0?2*e[s]*t[s]/o:0}return kk(r,this._support,this._numClasses,this._average)}reset(){super.reset(),this._tp.fill(0),this._fp.fill(0),this._fn.fill(0),this._support.fill(0)}};function vk(n,e,t,r,s,o,i,a){let u=n._impl.storage.data,p=e._impl.storage.data,c=p.length,f=n.shape.length>=2&&n.shape[1]>1;for(let d=0;d<c;d++){let h=p[d]|0,m;i==="binary"?m=u[d]>=.5?1:0:f?m=wU(u,d,a):m=u[d]|0,o[h]++,m===h?t&&t[h]++:(r&&r[m]++,s&&s[h]++)}}l(vk,"updateConfusionCounters");function wU(n,e,t){let r=0,s=n[e*t];for(let o=1;o<t;o++){let i=n[e*t+o];i>s&&(s=i,r=o)}return r}l(wU,"argmaxRow");function DC(n,e,t){let r=new Float64Array(t);for(let s=0;s<t;s++){let o=n[s]+e[s];r[s]=o>0?n[s]/o:0}return r}l(DC,"computePrecisionPerClass");function PC(n,e,t){let r=new Float64Array(t);for(let s=0;s<t;s++){let o=n[s]+e[s];r[s]=o>0?n[s]/o:0}return r}l(PC,"computeRecallPerClass");function TU(n,e,t,r,s){let o=DC(n,e,r);return kk(o,t,r,s)}l(TU,"computePrecision");function xU(n,e,t,r,s){let o=PC(n,e,r);return kk(o,t,r,s)}l(xU,"computeRecall");function kk(n,e,t,r){if(r==="none")return[...n];if(r==="micro"){let i=0,a=0;for(let u=0;u<t;u++)i+=n[u]*e[u],a+=e[u];return a>0?i/a:0}if(r==="weighted"){let i=0,a=0;for(let u=0;u<t;u++)a+=n[u]*e[u],i+=e[u];return i>0?a/i:0}let s=0,o=0;for(let i=0;i<t;i++)e[i]>0&&(o+=n[i],s++);return s>0?o/s:0}l(kk,"aggregate");var yh=class extends lr{static{l(this,"ConfusionMatrix")}_numClasses;_matrix;constructor({numClasses:e}){super(),this._numClasses=e,this._matrix=new Int32Array(e*e)}update(e,t){let r=e._impl.storage.data,s=t._impl.storage.data,o=s.length,i=this._numClasses,a=e.shape.length>=2&&e.shape[1]>1;for(let u=0;u<o;u++){let p=s[u]|0,c;a?c=SU(r,u,i):c=r[u]|0,this._matrix[p*i+c]++}}compute(){let e=this._numClasses,t=[];for(let r=0;r<e;r++){let s=new Array(e);for(let o=0;o<e;o++)s[o]=this._matrix[r*e+o];t.push(s)}return t}reset(){super.reset(),this._matrix.fill(0)}};function SU(n,e,t){let r=0,s=n[e*t];for(let o=1;o<t;o++){let i=n[e*t+o];i>s&&(s=i,r=o)}return r}l(SU,"argmaxRow");be();var FC={};at(FC,{cholesky:()=>Ok,cov:()=>AU,det:()=>OU,eigh:()=>vU,inv:()=>kU,lstsq:()=>bh,pinv:()=>NU,qr:()=>Ak,solve:()=>_h,svd:()=>Nk});be();function Nk(n){let[e,t,r]=ct("svd",n);return{U:e,S:t,V:r}}l(Nk,"svd");function vU(n){let[e,t]=ct("eigh",n);return{values:e,vectors:t}}l(vU,"eigh");function Ak(n){let[e,t]=ct("qr",n);return{Q:e,R:t}}l(Ak,"qr");var Ok=l(n=>ct("cholesky",n),"cholesky"),kU=l(n=>ct("inv",n),"inv"),NU=l(n=>ct("pinv",n),"pinv"),AU=l(n=>ct("cov",n),"cov"),_h=l((n,e)=>ct("solve",n,e),"solve"),bh=l((n,e)=>ct("lstsq",n,e),"lstsq"),OU=l(n=>ct("det",n).item(),"det");var jC={};at(jC,{DecisionTreeClassifier:()=>Uk,DecisionTreeRegressor:()=>jk,ElasticNet:()=>vw,GaussianNB:()=>Gk,GradientBoostingClassifier:()=>Hk,GradientBoostingRegressor:()=>Kk,GridSearchCV:()=>Jk,KFold:()=>Bw,KMeans:()=>Fk,KNeighborsClassifier:()=>zk,KNeighborsRegressor:()=>Vk,LabelEncoder:()=>Rk,Lasso:()=>Ck,LinearRegression:()=>Ek,LogisticRegression:()=>Dk,MinMaxScaler:()=>Ik,OneHotEncoder:()=>Mk,PCA:()=>$k,RandomForestClassifier:()=>Wk,RandomForestRegressor:()=>qk,Ridge:()=>Lk,StandardScaler:()=>Bk,TimeSeriesSplit:()=>Yk,accuracy_score:()=>ao,confusion_matrix:()=>MU,cross_val_score:()=>GC,mean_absolute_error:()=>RU,mean_squared_error:()=>BU,r2_score:()=>io,train_test_split:()=>LU});var Bk=class{static{l(this,"StandardScaler")}withMean;withStd;mean_;scale_;_cols;constructor({withMean:e=!0,withStd:t=!0}={}){this.withMean=e,this.withStd=t,this.mean_=null,this.scale_=null}fit(e){let t=sc(e);this.mean_=new Float64Array(t.cols),this.scale_=new Float64Array(t.cols);for(let r=0;r<t.cols;r++){let s=0;for(let a=0;a<t.rows;a++)s+=t.data[a*t.cols+r];s/=t.rows;let o=0;for(let a=0;a<t.rows;a++){let u=t.data[a*t.cols+r]-s;o+=u*u}o/=t.rows;let i=Math.sqrt(o);this.mean_[r]=this.withMean?s:0,this.scale_[r]=this.withStd&&i>0?i:1}return this._cols=t.cols,this}transform(e){let t=sc(e),r=this.mean_,s=this.scale_,o=new Float64Array(t.rows*t.cols);for(let i=0;i<t.rows;i++)for(let a=0;a<t.cols;a++)o[i*t.cols+a]=(t.data[i*t.cols+a]-r[a])/s[a];return Nf(o,t.rows,t.cols,e.dtype)}fit_transform(e){return this.fit(e).transform(e)}inverse_transform(e){let t=sc(e),r=this.mean_,s=this.scale_,o=new Float64Array(t.rows*t.cols);for(let i=0;i<t.rows;i++)for(let a=0;a<t.cols;a++)o[i*t.cols+a]=t.data[i*t.cols+a]*s[a]+r[a];return Nf(o,t.rows,t.cols,e.dtype)}},Rk=class{static{l(this,"LabelEncoder")}classes_;_lookup;constructor(){this.classes_=null,this._lookup=null}fit(e){let t=Yt(e),{classes:r}=Ns(t.data,t.n);return this.classes_=r,this._lookup=new Map(r.map((s,o)=>[s,o])),this}transform(e){let t=Yt(e),r=this._lookup,s=new Float64Array(t.n);for(let o=0;o<t.n;o++){let i=r.get(t.data[o]);if(i===void 0)throw new Error(`LabelEncoder: unseen label ${t.data[o]}`);s[o]=i}return xS(s,t.n,e.dtype)}fit_transform(e){let t=Yt(e),{y:r,classes:s}=Ns(t.data,t.n);this.classes_=s,this._lookup=new Map(s.map((i,a)=>[i,a]));let o=new Float64Array(t.n);for(let i=0;i<t.n;i++)o[i]=r[i];return xS(o,t.n,e.dtype)}inverse_transform(e){let t=Yt(e),r=this.classes_;return Array.from({length:t.n},(s,o)=>r[Math.round(t.data[o])])}},Mk=class{static{l(this,"OneHotEncoder")}classes_;constructor(){this.classes_=null}fit(e){let t=Yt(e);return this.classes_=Ns(t.data,t.n).classes,this}transform(e){let t=Yt(e),r=this.classes_,s=new Map(r.map((a,u)=>[a,u])),o=r.length,i=new Float64Array(t.n*o);for(let a=0;a<t.n;a++){let u=s.get(t.data[a]);if(u===void 0)throw new Error(`OneHotEncoder: unseen label ${t.data[a]}`);i[a*o+u]=1}return Nf(i,t.n,o,e.dtype)}fit_transform(e){return this.fit(e).transform(e)}},Ik=class{static{l(this,"MinMaxScaler")}featureRange;min_;dataMin_;dataRange_;constructor({featureRange:e=[0,1]}={}){this.featureRange=e,this.min_=null,this.dataMin_=null,this.dataRange_=null}fit(e){let t=sc(e);this.dataMin_=new Float64Array(t.cols),this.dataRange_=new Float64Array(t.cols);for(let r=0;r<t.cols;r++){let s=1/0,o=-1/0;for(let i=0;i<t.rows;i++){let a=t.data[i*t.cols+r];a<s&&(s=a),a>o&&(o=a)}this.dataMin_[r]=s,this.dataRange_[r]=o>s?o-s:1}return this}transform(e){let t=sc(e),[r,s]=this.featureRange,o=this.dataMin_,i=this.dataRange_,a=s-r,u=new Float64Array(t.rows*t.cols);for(let p=0;p<t.rows;p++)for(let c=0;c<t.cols;c++){let f=(t.data[p*t.cols+c]-o[c])/i[c];u[p*t.cols+c]=f*a+r}return Nf(u,t.rows,t.cols,e.dtype)}fit_transform(e){return this.fit(e).transform(e)}};function BU(n,e){let t=Yt(n),r=Yt(e),s=0;for(let o=0;o<t.n;o++){let i=t.data[o]-r.data[o];s+=i*i}return s/t.n}l(BU,"mean_squared_error");function RU(n,e){let t=Yt(n),r=Yt(e),s=0;for(let o=0;o<t.n;o++)s+=Math.abs(t.data[o]-r.data[o]);return s/t.n}l(RU,"mean_absolute_error");function io(n,e){let t=Yt(n),r=Yt(e),s=0;for(let a=0;a<t.n;a++)s+=t.data[a];s/=t.n;let o=0,i=0;for(let a=0;a<t.n;a++){let u=t.data[a]-r.data[a],p=t.data[a]-s;o+=u*u,i+=p*p}return i===0?0:1-o/i}l(io,"r2_score");function ao(n,e){let t=Yt(n),r=Yt(e),s=0;for(let o=0;o<t.n;o++)t.data[o]===r.data[o]&&s++;return s/t.n}l(ao,"accuracy_score");function MU(n,e){let t=Yt(n),r=Yt(e),s=0;for(let a=0;a<t.n;a++)s=Math.max(s,t.data[a],r.data[a]);let o=s+1,i=Array.from({length:o},()=>new Array(o).fill(0));for(let a=0;a<t.n;a++)i[t.data[a]][r.data[a]]++;return i}l(MU,"confusion_matrix");be();kt();os();function zC(n,e){return e?ru([n,tn([n.shape[0],1],{device:n.device,dtype:n.dtype})],1):n}l(zC,"addBias");function Pk(n){return n.ndim===1?n.reshape([n.shape[0],1]):n}l(Pk,"asColumn");var Ek=class{static{l(this,"LinearRegression")}fitIntercept;weight_;constructor({fitIntercept:e=!0}={}){this.fitIntercept=e,this.weight_=null}fit(e,t){let r=zC(e,this.fitIntercept);return this.weight_=bh(r,Pk(t)),this}predict(e){let t=zC(e,this.fitIntercept);return tt(t,this.weight_).reshape([e.shape[0]])}score(e,t){return io(t,this.predict(e))}},Lk=class{static{l(this,"Ridge")}alpha;fitIntercept;coef_;intercept_;constructor({alpha:e=1,fitIntercept:t=!0}={}){this.alpha=e,this.fitIntercept=t,this.coef_=null,this.intercept_=null}fit(e,t){let r=e.shape[1],s=Pk(t),o=e,i=s,a=null,u=null;this.fitIntercept&&(a=Tr(e,[0],!0),u=Tr(s,[0],!0),o=ze(e,a),i=ze(s,u));let p=o.transpose(0,1),c=ae(tt(p,o),ee(jg(r,r,{device:e.device,dtype:e.dtype}),this.alpha));return this.coef_=_h(c,tt(p,i)),this.intercept_=this.fitIntercept?ze(u,tt(a,this.coef_)):Ve([1,1],{device:e.device,dtype:e.dtype}),this}predict(e){return ae(tt(e,this.coef_),this.intercept_).reshape([e.shape[0]])}score(e,t){return io(t,this.predict(e))}},vw=class{static{l(this,"ElasticNet")}alpha;l1Ratio;fitIntercept;maxIter;tol;coef_;intercept_;constructor({alpha:e=1,l1Ratio:t=.5,fitIntercept:r=!0,maxIter:s=1e3,tol:o=1e-6}={}){this.alpha=e,this.l1Ratio=t,this.fitIntercept=r,this.maxIter=s,this.tol=o,this.coef_=null,this.intercept_=null}fit(e,t){let[r,s]=ct("elastic_net",e,Pk(t).reshape([e.shape[0]]),this.alpha,this.l1Ratio,this.maxIter,this.tol,this.fitIntercept);return this.coef_=r,this.intercept_=s,this}predict(e){let t=this.coef_.reshape([e.shape[1],1]);return ae(tt(e,t).reshape([e.shape[0]]),this.intercept_)}score(e,t){return io(t,this.predict(e))}},Ck=class extends vw{static{l(this,"Lasso")}constructor({alpha:e=1,fitIntercept:t=!0,maxIter:r=1e3,tol:s=1e-6}={}){super({alpha:e,l1Ratio:1,fitIntercept:t,maxIter:r,tol:s})}},Dk=class{static{l(this,"LogisticRegression")}C;lr;maxIter;W_;b_;classes_;constructor({C:e=1,lr:t=.5,maxIter:r=1e3}={}){this.C=e,this.lr=t,this.maxIter=r,this.W_=null,this.b_=null,this.classes_=null}fit(e,t){let r=e.shape[0],s=e.shape[1],o=Yt(t),{y:i,classes:a}=Ns(o.data,o.n);this.classes_=a;let u=a.length,p=new Float64Array(r*u);for(let y=0;y<r;y++)p[y*u+i[y]]=1;let c=Wt(p,{shape:[r,u],dtype:e.dtype,device:e.device}),f=Ve([s,u],{device:e.device,dtype:e.dtype}),d=Ve([1,u],{device:e.device,dtype:e.dtype}),h=1/this.C,m=e.transpose(0,1),g=this.lr/r;for(let y=0;y<this.maxIter;y++){let _=ae(tt(e,f),d),b=ze(_,No(_,1,!0)),w=as(b),x=Kt(w,wt(w,[1],!0)),S=ze(x,c),T=ae(tt(m,S),ee(f,h)),v=wt(S,[0],!0);f=ze(f,ee(T,g)),d=ze(d,ee(v,g))}return this.W_=f,this.b_=d,this}decisionLogits(e){return ae(tt(e,this.W_),this.b_)}predict(e){let t=Cg(this.decisionLogits(e),1,!1).toArray(),r=new Float64Array(t.length),s=this.classes_;for(let o=0;o<t.length;o++)r[o]=s[t[o]];return Wt(r,{shape:[t.length],dtype:e.dtype,device:e.device})}score(e,t){return ao(t,this.predict(e))}};be();var $k=class{static{l(this,"PCA")}nComponents;components_;mean_;explainedVariance_;explainedVarianceRatio_;_nc;constructor({nComponents:e=null}={}){this.nComponents=e,this.components_=null,this.mean_=null,this.explainedVariance_=null,this.explainedVarianceRatio_=null}fit(e){let t=e.shape[0],r=e.shape[1];this.mean_=Tr(e,[0],!0);let s=ze(e,this.mean_),{S:o,V:i}=Nk(s),a=i.shape[1],u=Math.min(this.nComponents??Math.min(t,r),a);this.components_=i.narrow(1,0,u),this._nc=u;let p=o.toArray(),c=t>1?t-1:1,f=0;for(let d=0;d<p.length;d++)f+=p[d]*p[d]/c;this.explainedVariance_=new Array(u),this.explainedVarianceRatio_=new Array(u);for(let d=0;d<u;d++){let h=p[d]*p[d]/c;this.explainedVariance_[d]=h,this.explainedVarianceRatio_[d]=f>0?h/f:0}return this}transform(e){return tt(ze(e,this.mean_),this.components_)}fit_transform(e){return this.fit(e).transform(e)}inverse_transform(e){return ae(tt(e,this.components_.transpose(0,1)),this.mean_)}};be();var Fk=class{static{l(this,"KMeans")}nClusters;maxIter;nInit;randomState;clusterCenters_;labels_;inertia_;constructor({nClusters:e=8,maxIter:t=300,nInit:r=10,randomState:s=0}={}){this.nClusters=e,this.maxIter=t,this.nInit=r,this.randomState=s,this.clusterCenters_=null,this.labels_=null,this.inertia_=null}fit(e){let[t,r,s]=ct("kmeans",e,this.nClusters,this.maxIter,this.nInit,this.randomState);return this.clusterCenters_=t,this.labels_=r,this.inertia_=Number(s.item()),this}predict(e){return ct("kmeans_predict",e,this.clusterCenters_)}fit_predict(e){return this.fit(e),this.labels_}};be();var kw=class{static{l(this,"BaseKNN")}nNeighbors;_classify;_X;_y;constructor(e,t){this.nNeighbors=e,this._classify=t,this._X=null,this._y=null}fit(e,t){return this._X=e,this._y=t,this}predict(e){return ct("knn_predict",this._X,this._y,e,this.nNeighbors,this._classify)}},zk=class extends kw{static{l(this,"KNeighborsClassifier")}constructor({nNeighbors:e=5}={}){super(e,!0)}score(e,t){return ao(t,this.predict(e))}},Vk=class extends kw{static{l(this,"KNeighborsRegressor")}constructor({nNeighbors:e=5}={}){super(e,!1)}score(e,t){return io(t,this.predict(e))}};be();var Gk=class{static{l(this,"GaussianNB")}means_;variances_;priors_;classes_;constructor(){this.means_=null,this.variances_=null,this.priors_=null,this.classes_=null}fit(e,t){let[r,s,o,i]=ct("gaussian_nb_fit",e,t);return this.means_=r,this.variances_=s,this.priors_=o,this.classes_=i,this}predict(e){return ct("gaussian_nb_predict",e,this.means_,this.variances_,this.priors_,this.classes_)}score(e,t){return ao(t,this.predict(e))}};be();os();var VC=1<<30;function Ow(n,e,t,r,s){return ct("decision_tree_fit",n,e,t.maxDepth,t.minSamplesSplit,t.minSamplesLeaf,t.maxFeatures,r,s)}l(Ow,"fitTree");function Zc(n,e){return ct("decision_tree_predict",n,e[0],e[1],e[2],e[3],e[4])}l(Zc,"treePredict");var Nw=class{static{l(this,"BaseTree")}maxDepth;minSamplesSplit;minSamplesLeaf;maxFeatures;randomState;_classify;_nodes;constructor(e,t){this.maxDepth=e.maxDepth??VC,this.minSamplesSplit=e.minSamplesSplit??2,this.minSamplesLeaf=e.minSamplesLeaf??1,this.maxFeatures=e.maxFeatures??0,this.randomState=e.randomState??0,this._classify=t,this._nodes=null}fit(e,t){return this._nodes=Ow(e,t,this,this._classify,this.randomState),this}predict(e){return Zc(e,this._nodes)}},jk=class extends Nw{static{l(this,"DecisionTreeRegressor")}constructor(e={}){super(e,!1)}score(e,t){return io(t,this.predict(e))}},Uk=class extends Nw{static{l(this,"DecisionTreeClassifier")}constructor(e={}){super(e,!0)}score(e,t){return ao(t,this.predict(e))}};function IU(n,e,t){return n>0?n:t?Math.max(1,Math.floor(Math.sqrt(e))):Math.max(1,Math.floor(e/3))}l(IU,"defaultMaxFeatures");var Aw=class{static{l(this,"BaseForest")}nEstimators;maxDepth;minSamplesSplit;minSamplesLeaf;maxFeatures;randomState;_classify;_trees;constructor(e,t){this.nEstimators=e.nEstimators??100,this.maxDepth=e.maxDepth??VC,this.minSamplesSplit=e.minSamplesSplit??2,this.minSamplesLeaf=e.minSamplesLeaf??1,this.maxFeatures=e.maxFeatures??0,this.randomState=e.randomState??0,this._classify=t,this._trees=[]}fit(e,t){let r=e.shape[0],s=e.shape[1],o=xr(this.randomState),i=IU(this.maxFeatures,s,this._classify),a={maxDepth:this.maxDepth,minSamplesSplit:this.minSamplesSplit,minSamplesLeaf:this.minSamplesLeaf,maxFeatures:i};this._trees=[];for(let u=0;u<this.nEstimators;u++){let p=new Array(r);for(let d=0;d<r;d++)p[d]=Math.floor(o()*r);let c=ks(e,p),f=ks(t,p);this._trees.push(Ow(c,f,a,this._classify,this.randomState+u+1))}return this}predict(e){let t=e.shape[0],r=this._trees.map(o=>Zc(e,o).toArray()),s=new Float64Array(t);if(this._classify)for(let o=0;o<t;o++){let i=new Map,a=r[0][o],u=0;for(let p=0;p<r.length;p++){let c=r[p][o],f=(i.get(c)||0)+1;i.set(c,f),f>u&&(u=f,a=c)}s[o]=a}else for(let o=0;o<t;o++){let i=0;for(let a=0;a<r.length;a++)i+=r[a][o];s[o]=i/r.length}return Wt(s,{shape:[t],dtype:e.dtype,device:e.device})}},qk=class extends Aw{static{l(this,"RandomForestRegressor")}constructor(e={}){super(e,!1)}score(e,t){return io(t,this.predict(e))}},Wk=class extends Aw{static{l(this,"RandomForestClassifier")}constructor(e={}){super(e,!0)}score(e,t){return ao(t,this.predict(e))}},Kk=class{static{l(this,"GradientBoostingRegressor")}nEstimators;learningRate;params;randomState;init_;_trees;constructor({nEstimators:e=100,learningRate:t=.1,maxDepth:r=3,minSamplesSplit:s=2,minSamplesLeaf:o=1,randomState:i=0}={}){this.nEstimators=e,this.learningRate=t,this.params={maxDepth:r,minSamplesSplit:s,minSamplesLeaf:o,maxFeatures:0},this.randomState=i,this.init_=0,this._trees=[]}fit(e,t){let r=Yt(t),s=r.n,o=0;for(let a=0;a<s;a++)o+=r.data[a];o/=s,this.init_=o;let i=new Float64Array(s).fill(o);this._trees=[];for(let a=0;a<this.nEstimators;a++){let u=new Float64Array(s);for(let d=0;d<s;d++)u[d]=r.data[d]-i[d];let p=Wt(u,{shape:[s],dtype:e.dtype,device:e.device}),c=Ow(e,p,this.params,!1,this.randomState+a+1);this._trees.push(c);let f=Zc(e,c).toArray();for(let d=0;d<s;d++)i[d]+=this.learningRate*f[d]}return this}predict(e){let t=e.shape[0],r=new Float64Array(t).fill(this.init_);for(let s of this._trees){let o=Zc(e,s).toArray();for(let i=0;i<t;i++)r[i]+=this.learningRate*o[i]}return Wt(r,{shape:[t],dtype:e.dtype,device:e.device})}score(e,t){return io(t,this.predict(e))}},Hk=class{static{l(this,"GradientBoostingClassifier")}nEstimators;learningRate;params;randomState;classes_;_stages;constructor({nEstimators:e=100,learningRate:t=.1,maxDepth:r=3,minSamplesSplit:s=2,minSamplesLeaf:o=1,randomState:i=0}={}){this.nEstimators=e,this.learningRate=t,this.params={maxDepth:r,minSamplesSplit:s,minSamplesLeaf:o,maxFeatures:0},this.randomState=i,this.classes_=null,this._stages=[]}fit(e,t){let r=Yt(t),s=r.n,{y:o,classes:i}=Ns(r.data,r.n);this.classes_=i;let a=i.length,u=new Float64Array(s*a);this._stages=[];for(let p=0;p<this.nEstimators;p++){let c=EU(u,s,a),f=[];for(let d=0;d<a;d++){let h=new Float64Array(s);for(let _=0;_<s;_++)h[_]=(o[_]===d?1:0)-c[_*a+d];let m=Wt(h,{shape:[s],dtype:e.dtype,device:e.device}),g=Ow(e,m,this.params,!1,this.randomState+p*a+d+1);f.push(g);let y=Zc(e,g).toArray();for(let _=0;_<s;_++)u[_*a+d]+=this.learningRate*y[_]}this._stages.push(f)}return this}predict(e){let t=e.shape[0],r=this.classes_,s=r.length,o=new Float64Array(t*s);for(let a of this._stages)for(let u=0;u<s;u++){let p=Zc(e,a[u]).toArray();for(let c=0;c<t;c++)o[c*s+u]+=this.learningRate*p[c]}let i=new Float64Array(t);for(let a=0;a<t;a++){let u=0;for(let p=1;p<s;p++)o[a*s+p]>o[a*s+u]&&(u=p);i[a]=r[u]}return Wt(i,{shape:[t],dtype:e.dtype,device:e.device})}score(e,t){return ao(t,this.predict(e))}};function EU(n,e,t){let r=new Float64Array(e*t);for(let s=0;s<e;s++){let o=-1/0;for(let a=0;a<t;a++)n[s*t+a]>o&&(o=n[s*t+a]);let i=0;for(let a=0;a<t;a++){let u=Math.exp(n[s*t+a]-o);r[s*t+a]=u,i+=u}for(let a=0;a<t;a++)r[s*t+a]/=i}return r}l(EU,"softmaxRows");function LU(n,e,{testSize:t=.25,shuffle:r=!0,randomState:s=0}={}){let o=n.shape[0],i=r?gS(o,xr(s)):Array.from({length:o},(c,f)=>f),a=Math.max(1,Math.round(o*t)),u=i.slice(0,a),p=i.slice(a);return[ks(n,p),ks(n,u),ks(e,p),ks(e,u)]}l(LU,"train_test_split");var Bw=class{static{l(this,"KFold")}nSplits;shuffle;randomState;constructor({nSplits:e=5,shuffle:t=!1,randomState:r=0}={}){this.nSplits=e,this.shuffle=t,this.randomState=r}split(e){let t=this.shuffle?gS(e,xr(this.randomState)):Array.from({length:e},(a,u)=>u),r=[],s=Math.floor(e/this.nSplits),o=e%this.nSplits,i=0;for(let a=0;a<this.nSplits;a++){let u=s+(o>0?1:0);o>0&&o--;let p=t.slice(i,i+u),c=t.slice(0,i).concat(t.slice(i+u));r.push({train:c,test:p}),i+=u}return r}},Yk=class{static{l(this,"TimeSeriesSplit")}nSplits;constructor({nSplits:e=5}={}){this.nSplits=e}split(e){let t=Math.floor(e/(this.nSplits+1)),r=[];for(let s=1;s<=this.nSplits;s++){let o=t*s,i=s===this.nSplits?e:t*(s+1),a=Array.from({length:o},(p,c)=>c),u=Array.from({length:i-o},(p,c)=>o+c);r.push({train:a,test:u})}return r}};function GC(n,e,t,{cv:r=5,scoring:s=null,shuffle:o=!0,randomState:i=0}={}){let a=new Bw({nSplits:r,shuffle:o,randomState:i}).split(e.shape[0]),u=[];for(let{train:p,test:c}of a){let f=ks(e,p),d=ks(t,p),h=ks(e,c),m=ks(t,c),g=n().fit(f,d);u.push(s?s(m,g.predict(h)):g.score(h,m))}return u}l(GC,"cross_val_score");function CU(n){let e=Object.keys(n),t=[{}];for(let r of e){let s=[];for(let o of t)for(let i of n[r])s.push({...o,[r]:i});t=s}return t}l(CU,"cartesian");var Jk=class{static{l(this,"GridSearchCV")}makeEstimator;paramGrid;cv;scoring;bestParams_;bestScore_;bestEstimator_;constructor(e,t,{cv:r=5,scoring:s=null}={}){this.makeEstimator=e,this.paramGrid=t,this.cv=r,this.scoring=s,this.bestParams_=null,this.bestScore_=-1/0,this.bestEstimator_=null}fit(e,t){for(let r of CU(this.paramGrid)){let s=GC(()=>this.makeEstimator(r),e,t,{cv:this.cv,scoring:this.scoring}),o=s.reduce((i,a)=>i+a,0)/s.length;o>this.bestScore_&&(this.bestScore_=o,this.bestParams_=r)}return this.bestEstimator_=this.makeEstimator(this.bestParams_).fit(e,t),this}predict(e){return this.bestEstimator_.predict(e)}};var ZD={};at(ZD,{Generator:()=>Vw,acf:()=>RD,andersonDarling:()=>AD,betaRegularized:()=>Th,bisect:()=>UC,brentq:()=>Rw,chi2:()=>ep,chi2Gof:()=>wD,chi2Independence:()=>TD,constrainedMinimize:()=>iN,convolve:()=>zD,correlate:()=>VD,cubicSpline:()=>mD,dagostinoK2:()=>ND,differentialEvolution:()=>nN,digammaScalar:()=>ja,durbinWatson:()=>ED,erfScalar:()=>Wi,erfcScalar:()=>Ki,fft:()=>Fu,fisherF:()=>nD,gammaScalar:()=>Hi,ifft:()=>tp,jarqueBera:()=>kD,ksTest1Samp:()=>SD,ksTest2Samp:()=>vD,lbfgs:()=>Sh,lbfgsB:()=>sN,levenbergMarquardt:()=>oN,lgammaScalar:()=>cn,linearInterp:()=>hD,ljungBox:()=>ID,lowerGammaRegularized:()=>Iw,mannWhitneyU:()=>OD,nelderMead:()=>eN,newton:()=>qC,normal:()=>tD,normalCdfScalar:()=>Uo,normalPdfScalar:()=>xh,normalPpfScalar:()=>xa,pacf:()=>MD,periodogram:()=>LD,polyfit:()=>HD,polyroots:()=>JD,polyval:()=>YD,qr:()=>Ak,quadrature:()=>pD,rollingMax:()=>KD,rollingMean:()=>jD,rollingMin:()=>WD,rollingStd:()=>UD,rollingSum:()=>GD,simpson:()=>cD,studentT:()=>Ew,tTest1Samp:()=>fN,tTestInd:()=>_D,tTestPaired:()=>bD,trapezoid:()=>lD});sr();wn();function Es(n,e,t){return{root:n,iterations:e,converged:t}}l(Es,"result");function UC(n,e,t,r={}){let s=r.tol??1e-12,o=r.maxIter??100,i=n(e),a=n(t);if(i===0)return Es(e,0,!0);if(a===0)return Es(t,0,!0);if(i*a>0)throw new Error("bisect: f(a) and f(b) must have opposite signs");let u=e,p=t,c=i;for(let f=1;f<=o;f++){let d=.5*(u+p),h=n(d);if(h===0||.5*(p-u)<s)return Es(d,f,!0);c*h<0?p=d:(u=d,c=h)}return Es(.5*(u+p),o,!1)}l(UC,"bisect");function qC(n,e,t={}){let r=t.tol??1e-12,s=t.maxIter??100,o=t.step??1e-7,i=t.derivative??(u=>(n(u+o)-n(u-o))/(2*o)),a=e;for(let u=1;u<=s;u++){let p=n(a);if(Math.abs(p)<r)return Es(a,u,!0);let c=i(a);if(c===0)break;let f=a-p/c;if(Math.abs(f-a)<r)return Es(f,u,!0);a=f}return Es(a,s,!1)}l(qC,"newton");function Rw(n,e,t,r={}){let s=r.tol??1e-12,o=r.maxIter??100,i=n(e),a=n(t);if(i===0)return Es(e,0,!0);if(a===0)return Es(t,0,!0);if(i*a>0)throw new Error("brentq: f(a) and f(b) must have opposite signs");Math.abs(i)<Math.abs(a)&&([e,t]=[t,e],[i,a]=[a,i]);let u=e,p=i,c=u,f=!0;for(let d=1;d<=o;d++){let h;i!==p&&a!==p?h=e*a*p/((i-a)*(i-p))+t*i*p/((a-i)*(a-p))+u*i*a/((p-i)*(p-a)):h=t-a*(t-e)/(a-i);let m=.25*(3*e+t),g=!(h>Math.min(m,t)&&h<Math.max(m,t)),y=f&&Math.abs(h-t)>=.5*Math.abs(t-u),_=!f&&Math.abs(h-t)>=.5*Math.abs(u-c),b=f&&Math.abs(t-u)<s,w=!f&&Math.abs(u-c)<s;g||y||_||b||w?(h=.5*(e+t),f=!0):f=!1;let x=n(h);if(c=u,u=t,p=a,i*x<0?(t=h,a=x):(e=h,i=x),Math.abs(i)<Math.abs(a)&&([e,t]=[t,e],[i,a]=[a,i]),x===0)return Es(h,d,!0);if(Math.abs(a)<s||Math.abs(t-e)<s)return Es(t,d,!0)}return Es(t,o,!1)}l(Rw,"brentq");Yi();var Qc=[-39.69683028665376,220.9460984245205,-275.9285104469687,138.357751867269,-30.66479806614716,2.506628277459239],wh=[-54.47609879822406,161.5858368580409,-155.6989798598866,66.80131188771972,-13.28068155288572],Xc=[-.007784894002430293,-.3223964580411365,-2.400758277161838,-2.549732539343734,4.374664141464968,2.938163982698783],Mw=[.007784695709041462,.3224671290700398,2.445134137142996,3.754408661907416],WC=.02425,Zk=300,Qk=3e-14,Ar=1e-300,DU=2,PU=Math.SQRT2,$U=Math.sqrt(2*Math.PI),aSe=1/Math.sqrt(Math.PI);function Iw(n,e){if(e<=0)return 0;if(e<n+1){let i=1/n,a=i;for(let u=1;u<Zk&&(i*=e/(n+u),a+=i,!(Math.abs(i)<Math.abs(a)*Qk));u++);return a*Math.exp(-e+n*Math.log(e)-cn(n))}let t=e+1-n,r=1/Ar,s=1/t,o=s;for(let i=1;i<Zk;i++){let a=-i*(i-n);t+=2,s=a*s+t,Math.abs(s)<Ar&&(s=Ar),r=t+a/r,Math.abs(r)<Ar&&(r=Ar),s=1/s;let u=s*r;if(o*=u,Math.abs(u-1)<Qk)break}return 1-o*Math.exp(-e+n*Math.log(e)-cn(n))}l(Iw,"lowerGammaRegularized");function KC(n,e,t){let r=n+e,s=n+1,o=n-1,i=1,a=1-r*t/s;Math.abs(a)<Ar&&(a=Ar),a=1/a;let u=a;for(let p=1;p<Zk;p++){let c=2*p,f=p*(e-p)*t/((o+c)*(n+c));a=1+f*a,Math.abs(a)<Ar&&(a=Ar),i=1+f/i,Math.abs(i)<Ar&&(i=Ar),a=1/a,u*=a*i,f=-(n+p)*(r+p)*t/((n+c)*(s+c)),a=1+f*a,Math.abs(a)<Ar&&(a=Ar),i=1+f/i,Math.abs(i)<Ar&&(i=Ar),a=1/a;let d=a*i;if(u*=d,Math.abs(d-1)<Qk)break}return u}l(KC,"betacf");function Th(n,e,t){if(t<=0)return 0;if(t>=1)return 1;let r=Math.exp(cn(n+e)-cn(n)-cn(e)+n*Math.log(t)+e*Math.log(1-t));return t<(n+1)/(n+e+2)?r*KC(n,e,t)/n:1-r*KC(e,n,1-t)/e}l(Th,"betaRegularized");function Uo(n){return .5*Ki(-n/PU)}l(Uo,"normalCdfScalar");function xh(n){return Math.exp(-.5*n*n)/$U}l(xh,"normalPdfScalar");function HC(n){if(n<WC){let r=Math.sqrt(-2*Math.log(n));return(((((Xc[0]*r+Xc[1])*r+Xc[2])*r+Xc[3])*r+Xc[4])*r+Xc[5])/((((Mw[0]*r+Mw[1])*r+Mw[2])*r+Mw[3])*r+1)}if(n>1-WC)return-HC(1-n);let e=n-.5,t=e*e;return(((((Qc[0]*t+Qc[1])*t+Qc[2])*t+Qc[3])*t+Qc[4])*t+Qc[5])*e/(((((wh[0]*t+wh[1])*t+wh[2])*t+wh[3])*t+wh[4])*t+1)}l(HC,"ppfSeed");function xa(n,e={}){if(n<=0)return-1/0;if(n>=1)return 1/0;let t=e.refineSteps??DU,r=HC(n);for(let s=0;s<t;s++){let i=(Uo(r)-n)/xh(r);r-=i/(1+r*i/2)}return r}l(xa,"normalPpfScalar");var FU=3,zU=2,VU=60,GU=1e-11;function Ls(n,e){if(n instanceof Ze){let t=Qe(n),r=new Float64Array(t.length);for(let s=0;s<t.length;s++)r[s]=e(Number(t[s]));return Q(r,n.shape,n.dtype,n.device)}return e(n)}l(Ls,"mapElementwise");function Xk(n,e,t,r,s={}){if(t<=0||t>=1)return t===0?s.lowerLimit??-1/0:t===1?1/0:NaN;let o=s.refineSteps??FU,i=s.tol??GU,a=r;for(let f=0;f<o;f++){let d=n(a)-t,h=e(a);if(!Number.isFinite(a)||h<=0)break;let m=a-d/h;if(a=s.lowerLimit!==void 0&&m<=s.lowerLimit?(a+s.lowerLimit)/2:m,Math.abs(d)<i)return a}if(Math.abs(n(a)-t)<i)return a;let u=s.lowerLimit!==void 0?s.lowerLimit+Number.EPSILON:a,p=a,c=Math.max(1,Math.abs(a));for(let f=0;f<VU;f++){if(s.lowerLimit===void 0&&(u=a-c),p=a+c,n(u)-t<0&&n(p)-t>0)return Rw(d=>n(d)-t,u,p,{tol:i}).root;c*=zU}return a}l(Xk,"invertCdf");var tD={cdf:l((n,e={})=>{let t=e.loc??0,r=e.scale??1;return Ls(n,s=>Uo((s-t)/r))},"cdf"),pdf:l((n,e={})=>{let t=e.loc??0,r=e.scale??1;return Ls(n,s=>xh((s-t)/r)/r)},"pdf"),ppf:l((n,e={})=>{let t=e.loc??0,r=e.scale??1;return Ls(n,s=>t+r*xa(s,e))},"ppf")};function YC(n,e){if(n===0)return .5;let t=Th(e/2,.5,e/(e+n*n));return n>0?1-t/2:t/2}l(YC,"tCdfScalar");function JC(n,e){let t=cn((e+1)/2)-cn(e/2)-.5*Math.log(e*Math.PI);return Math.exp(t-(e+1)/2*Math.log(1+n*n/e))}l(JC,"tPdfScalar");var Ew={cdf:l((n,e,t={})=>Ls(n,r=>YC(r,e??t.df)),"cdf"),pdf:l((n,e,t={})=>Ls(n,r=>JC(r,e??t.df)),"pdf"),ppf:l((n,e,t={})=>{let r=e??t.df;return Ls(n,s=>Xk(o=>YC(o,r),o=>JC(o,r),s,xa(s),t))},"ppf")};function ZC(n,e){return n<=0?0:Iw(e/2,n/2)}l(ZC,"chi2CdfScalar");function QC(n,e){if(n<=0)return 0;let t=e/2;return Math.exp((t-1)*Math.log(n)-n/2-t*Math.log(2)-cn(t))}l(QC,"chi2PdfScalar");function jU(n,e){let t=xa(n),r=2/(9*e),s=1-r+t*Math.sqrt(r);return Math.max(e*s*s*s,Number.EPSILON)}l(jU,"chi2Seed");var ep={cdf:l((n,e,t={})=>Ls(n,r=>ZC(r,e??t.df)),"cdf"),pdf:l((n,e,t={})=>Ls(n,r=>QC(r,e??t.df)),"pdf"),ppf:l((n,e,t={})=>{let r=e??t.df;return Ls(n,s=>Xk(o=>ZC(o,r),o=>QC(o,r),s,jU(s,r),{...t,lowerLimit:0}))},"ppf")};function XC(n,e,t){return n<=0?0:Th(e/2,t/2,e*n/(e*n+t))}l(XC,"fCdfScalar");function eD(n,e,t){if(n<=0)return 0;let r=cn(e/2)+cn(t/2)-cn((e+t)/2);return Math.exp(e/2*Math.log(e/t)+(e/2-1)*Math.log(n)-(e+t)/2*Math.log(1+e/t*n)-r)}l(eD,"fPdfScalar");var nD={cdf:l((n,e,t,r={})=>Ls(n,s=>XC(s,e??r.d1,t??r.d2)),"cdf"),pdf:l((n,e,t,r={})=>Ls(n,s=>eD(s,e??r.d1,t??r.d2)),"pdf"),ppf:l((n,e,t,r={})=>{let s=e??r.d1,o=t??r.d2;return Ls(n,i=>Xk(a=>XC(a,s,o),a=>eD(a,s,o),i,1,{...r,lowerLimit:0}))},"ppf")};function eN(n,e,t={}){let r=e.length,s=t.maxIter??200*r,o=t.tol??1e-10,i=t.initialStep??.05,a=t.zeroStep??25e-5,u=t.alpha??1,p=t.gamma??2,c=t.rho??.5,f=t.sigma??.5,d=[e.slice()];for(let _=0;_<r;_++){let b=e.slice();b[_]=b[_]!==0?b[_]*(1+i):a,d.push(b)}let h=d.map(n),m=Array.from({length:r+1},(_,b)=>b),g=0;for(;g<s;g++){m.sort((v,A)=>h[v]-h[A]);let _=m.map(v=>d[v]),b=m.map(v=>h[v]);for(let v=0;v<=r;v++)d[v]=_[v],h[v]=b[v];if(Math.abs(h[r]-h[0])<o)break;let w=new Array(r).fill(0);for(let v=0;v<r;v++)for(let A=0;A<r;A++)w[A]+=d[v][A];for(let v=0;v<r;v++)w[v]/=r;let x=d[r],S=w.map((v,A)=>v+u*(v-x[A])),T=n(S);if(T<h[0]){let v=w.map((N,k)=>N+p*(S[k]-N)),A=n(v);A<T?(d[r]=v,h[r]=A):(d[r]=S,h[r]=T)}else if(T<h[r-1])d[r]=S,h[r]=T;else{let v=w.map((N,k)=>N+c*(x[k]-N)),A=n(v);if(A<h[r])d[r]=v,h[r]=A;else{let N=d[0];for(let k=1;k<=r;k++)d[k]=N.map((O,R)=>O+f*(d[k][R]-O)),h[k]=n(d[k])}}}let y=0;for(let _=1;_<=r;_++)h[_]<h[y]&&(y=_);return{point:d[y],value:h[y],iterations:g,converged:g<s}}l(eN,"nelderMead");function Cs(n,e){let t=0;for(let r=0;r<n.length;r++)t+=n[r]*e[r];return t}l(Cs,"dot");function Pu(n){return Math.sqrt(Cs(n,n))}l(Pu,"norm");function Lw(n,e){let t=new Array(n.length);for(let r=0;r<n.length;r++)t[r]=n[r]-e[r];return t}l(Lw,"sub");function rD(n,e){let t=new Array(n.length);for(let r=0;r<n.length;r++)t[r]=n[r]*e;return t}l(rD,"scale");function sD(n){let e=0;for(let t=1;t<n.length;t++)n[t]<n[e]&&(e=t);return e}l(sD,"argmin");function oD(n,e,t){let r=e.length,s=new Array(r),o=e.slice();for(let i=0;i<r;i++){let a=o[i];o[i]=a+t;let u=n(o);o[i]=a-t;let p=n(o);o[i]=a,s[i]=(u-p)/(2*t)}return s}l(oD,"numericGradient");function $u(n,e){return e?n.map((t,r)=>{let s=e[r];if(!s)return t;let[o,i]=s;return t<o?o:t>i?i:t}):n.slice()}l($u,"clampToBounds");var UU=10,qU=5,WU=.8,KU=.9,HU=1e3,YU=1e-10;function tN(n,e,t){let r;do r=rc(n,e);while(t.includes(r));return r}l(tN,"pickDistinct");function nN(n,e,t={}){let r=e.length,s=xr(t.seed),o=t.populationSize??Math.max(qU,UU*r),i=t.mutation??WU,a=t.recombination??KU,u=t.maxIter??HU,p=t.tol??YU,c=[];for(let m=0;m<o;m++){let g=new Array(r);for(let y=0;y<r;y++){let[_,b]=e[y];g[y]=_+s()*(b-_)}c.push(g)}let f=c.map(n),d=sD(f),h=0;for(;h<u;h++){for(let y=0;y<o;y++){let _=tN(s,o,[y]),b=tN(s,o,[y,_]),w=tN(s,o,[y,_,b]),x=rc(s,r),S=c[y].slice();for(let v=0;v<r;v++)if(s()<a||v===x){let[A,N]=e[v],k=c[_][v]+i*(c[b][v]-c[w][v]);k<A?k=A:k>N&&(k=N),S[v]=k}let T=n(S);T<f[y]&&(c[y]=S,f[y]=T,T<f[d]&&(d=y))}let m=f[0],g=f[0];for(let y=1;y<o;y++)f[y]<m&&(m=f[y]),f[y]>g&&(g=f[y]);if(g-m<p){h++;break}}return{point:c[d].slice(),value:f[d],iterations:h,converged:h<u}}l(nN,"differentialEvolution");var JU=10,ZU=1e3,QU=1e-8,XU=1e-12,eq=1e-4,tq=.1,iD=.5,nq=40;function rq(n,e,t,r){let s=n.slice(),o=e.length,i=new Array(o);for(let a=o-1;a>=0;a--){i[a]=r[a]*Cs(e[a],s);for(let u=0;u<s.length;u++)s[u]-=i[a]*t[a][u]}if(o>0){let a=e[o-1],u=t[o-1],p=Cs(a,u)/Cs(u,u);for(let c=0;c<s.length;c++)s[c]*=p}for(let a=0;a<o;a++){let u=r[a]*Cs(t[a],s);for(let p=0;p<s.length;p++)s[p]+=(i[a]-u)*e[a][p]}return s}l(rq,"twoLoop");function aD(n,e,t,r){let s=e.length,o=t.memory??JU,i=t.maxIter??ZU,a=t.gtol??QU,u=t.ftol??XU,p=t.step??1e-7,c=t.gradient??(x=>oD(n,x,p)),f=l((x,S)=>r?x.map((T,v)=>{let A=r[v];if(!A)return T;let[N,k]=A;return S[v]<=N&&T>0||S[v]>=k&&T<0?0:T}):x,"projectGrad"),d=$u(e,r),h=n(d),m=c(d),g=f(m,d),y=[],_=[],b=[],w=0;for(;w<i;w++){if(Pu(g)<a)return{point:d,value:h,iterations:w,converged:!0};let x=rD(rq(g,y,_,b),-1);if(Cs(x,g)>=0)for(let H=0;H<s;H++)x[H]=-g[H];let S=1,T=d,v=h,A=!1,N=!1,k=Cs(g,x);for(let H=0;H<nq;H++){let G=d.map((j,W)=>j+S*x[W]),U=$u(G,r),D=n(U);if(D<=h+eq*S*k){T=U,v=D,A=!0,N=r!=null&&U.some((j,W)=>j!==G[W]);break}let V=2*(D-h-k*S),ue=V>0?-k*S*S/V:S*iD,ce=tq*S,Le=iD*S;S=ue<ce?ce:ue>Le?Le:ue}if(!A)return{point:d,value:h,iterations:w,converged:!1};let O=c(T),R=f(O,T),B=Lw(T,d),E=N?Lw(R,g):Lw(O,m),F=Cs(B,E);F>Number.EPSILON*Pu(B)*Pu(E)&&(y.push(B),_.push(E),b.push(1/F),y.length>o&&(y.shift(),_.shift(),b.shift()));let C=Math.abs(v-h);if(d=T,h=v,m=O,g=R,C<u*(1+Math.abs(h))&&Pu(B)<u*(1+Pu(d)))return w++,{point:d,value:h,iterations:w,converged:!0}}return{point:d,value:h,iterations:w,converged:Pu(g)<a}}l(aD,"minimize");function Sh(n,e,t={}){return aD(n,e,t,null)}l(Sh,"lbfgs");function sN(n,e,t,r={}){return aD(n,e,r,t)}l(sN,"lbfgsB");var sq=200,oq=1e-10,iq=.001,uD=10,aq=10,uq=1e12;function lq(n,e,t,r){let s=e.length,o=new Array(t);for(let a=0;a<t;a++)o[a]=new Array(s);let i=e.slice();for(let a=0;a<s;a++){let u=i[a];i[a]=u+r;let p=n(i);i[a]=u-r;let c=n(i);i[a]=u;for(let f=0;f<t;f++)o[f][a]=(p[f]-c[f])/(2*r)}return o}l(lq,"numericJacobian");function cq(n,e,t,r){let s=new Float64Array(r*r);for(let a=0;a<r;a++){for(let u=0;u<r;u++)s[a*r+u]=n[a][u];s[a*r+a]+=t*n[a][a]}let o=Float64Array.from(e),i=_h(Q(s,[r,r],"f64"),Q(o,[r],"f64"));return Array.from(i.toArray())}l(cq,"solveNormal");function oN(n,e,t={}){let r=t.maxIter??sq,s=t.tol??oq,o=t.step??1e-7,i=t.jacobian??((m,g)=>lq(n,m,g,o)),a=e.slice(),u=n(a),p=u.length,c=a.length,f=Cs(u,u),d=t.lambda??iq,h=0;for(;h<r;h++){let m=i(a,p),g=new Array(c),y=new Array(c).fill(0);for(let b=0;b<c;b++){g[b]=new Array(c).fill(0);for(let w=0;w<p;w++)y[b]+=m[w][b]*u[w]}for(let b=0;b<c;b++)for(let w=b;w<c;w++){let x=0;for(let S=0;S<p;S++)x+=m[S][b]*m[S][w];g[b][w]=x,g[w][b]=x}let _=!1;for(;d<uq;){let b;try{b=cq(g,y,d,c)}catch{d*=uD;continue}let w=a.map((T,v)=>T-b[v]),x=n(w),S=Cs(x,x);if(S<f){a=w,u=x;let T=f-S;if(f=S,d/=aq,_=!0,T<s*(1+f))return{point:a,value:f,iterations:h+1,converged:!0};break}d*=uD}if(!_)return{point:a,value:f,iterations:h,converged:f<s}}return{point:a,value:f,iterations:h,converged:!1}}l(oN,"levenbergMarquardt");var pq=10,fq=10,dq=20,hq=1e-8;function iN(n,e,t={}){let r=t.bounds??null,s=t.inequalities??[],o=t.equalities??[],i=t.outerIter??dq,a=t.ctol??hq,u=t.penaltyGrowth??fq,p=t.inner??Sh,c=t.innerOpts??{},f=t.penalty??pq,d=$u(e,r),h=0,m=null;for(let g=0;g<i;g++){m=p(l(b=>{let w=$u(b,r),x=n(w);for(let S of s){let T=S(w);T>0&&(x+=f*T*T)}for(let S of o){let T=S(w);x+=f*T*T}return x},"penalized"),d,c),d=$u(m.point,r),h+=m.iterations;let _=0;for(let b of s)_=Math.max(_,b(d));for(let b of o)_=Math.max(_,Math.abs(b(d)));if(_<a)return{point:d,value:n(d),iterations:h,converged:!0};f*=u}return{point:d,value:n(d),iterations:h,converged:!1}}l(iN,"constrainedMinimize");function lD(n,e,t,r={}){let s=r.n??1024,o=(t-e)/s,i=.5*(n(e)+n(t));for(let a=1;a<s;a++)i+=n(e+a*o);return i*o}l(lD,"trapezoid");function cD(n,e,t,r={}){let s=r.n??1024;s%2===1&&(s+=1);let o=(t-e)/s,i=n(e)+n(t);for(let a=1;a<s;a++)i+=(a%2===0?2:4)*n(e+a*o);return i*o/3}l(cD,"simpson");function aN(n,e,t,r,s){return(s-r)/6*(n+4*t+e)}l(aN,"simpsonCell");function uN(n,e,t,r,s,o,i,a,u,p){let c=.5*(e+t),f=.5*(e+c),d=.5*(c+t),h=n(f),m=n(d),g=aN(r,o,h,e,c),y=aN(o,s,m,c,t),_=g+y-i;return u>=p||Math.abs(_)<=15*a?g+y+_/15:uN(n,e,c,r,o,h,g,.5*a,u+1,p)+uN(n,c,t,o,s,m,y,.5*a,u+1,p)}l(uN,"adaptiveStep");function pD(n,e,t,r={}){let s=r.tol??1e-10,o=r.maxDepth??50,i=.5*(e+t),a=n(e),u=n(t),p=n(i),c=aN(a,u,p,e,t);return uN(n,e,t,a,u,p,c,s,0,o)}l(pD,"quadrature");function dD(n,e){let t=1,r=n.length-1;for(;t<r;){let s=t+r>>1;n[s]<e?t=s+1:r=s}return t}l(dD,"upperInterval");function fD(n,e,t){let r=n.length;if(t<=n[0])return e[0];if(t>=n[r-1])return e[r-1];let s=dD(n,t),o=(t-n[s-1])/(n[s]-n[s-1]);return e[s-1]+o*(e[s]-e[s-1])}l(fD,"evalLinear");function hD(n,e,t){return typeof t!="number"?t.map(r=>fD(n,e,r)):fD(n,e,t)}l(hD,"linearInterp");function mq(n,e,t,r){let s=e.length,o=new Array(s),i=new Array(s);o[0]=t[0]/e[0],i[0]=r[0]/e[0];for(let u=1;u<s;u++){let p=e[u]-n[u]*o[u-1];o[u]=t[u]/p,i[u]=(r[u]-n[u]*i[u-1])/p}let a=new Array(s);a[s-1]=i[s-1];for(let u=s-2;u>=0;u--)a[u]=i[u]-o[u]*a[u+1];return a}l(mq,"thomas");function mD(n,e){let t=n.length,r=new Array(t-1);for(let c=0;c<t-1;c++)r[c]=n[c+1]-n[c];let s=new Array(t).fill(0),o=new Array(t).fill(1),i=new Array(t).fill(0),a=new Array(t).fill(0);for(let c=1;c<t-1;c++)s[c]=r[c-1],o[c]=2*(r[c-1]+r[c]),i[c]=r[c],a[c]=6*((e[c+1]-e[c])/r[c]-(e[c]-e[c-1])/r[c-1]);let u=mq(s,o,i,a),p=l(c=>{if(c<=n[0])return e[0];if(c>=n[t-1])return e[t-1];let f=dD(n,c),d=r[f-1],h=(n[f]-c)/d,m=(c-n[f-1])/d;return h*e[f-1]+m*e[f]+((h*h*h-h)*u[f-1]+(m*m*m-m)*u[f])*d*d/6},"evalOne");return{xs:n,ys:e,coefficients:u,evaluate:l((c=>typeof c!="number"?c.map(p):p(c)),"evaluate")}}l(mD,"cubicSpline");be();var Fu=l(n=>ct("fft",n),"fft"),tp=l(n=>ct("ifft",n),"ifft");sr();wn();qn();function gD(n){let e=new Float64Array(n.length);for(let t=0;t<n.length;t++)e[t]=Number(n[t]);return e}l(gD,"float64From");function Je(n){return n instanceof Ze?{data:gD(Qe(n)),dtype:n.dtype,device:n.device}:typeof n=="number"?{data:Float64Array.of(n),dtype:"f64",device:Ws}:{data:Float64Array.from(n),dtype:"f64",device:Ws}}l(Je,"hostVector");function Cw(n){if(n instanceof Ze){if(n.ndim!==2)throw new Error(`numeric: expected a 2-D table, got ${n.ndim}-D`);let[s,o]=n.shape;return{data:gD(Qe(n)),rows:s,cols:o,dtype:n.dtype,device:n.device}}let e=n.length,t=n[0].length,r=new Float64Array(e*t);for(let s=0;s<e;s++)for(let o=0;o<t;o++)r[s*t+o]=n[s][o];return{data:r,rows:e,cols:t,dtype:"f64",device:Ws}}l(Cw,"hostGrid");function Dw(n){let e=1;for(;e<n;)e<<=1;return e}l(Dw,"nextPow2");var gq=100,yq=1e-16,_q=.12,bq=.11,yD=[.75,2.25],lN=[.2,.34,.6],wq=[-13.436,101.14,-223.73],Tq=[-8.318,42.796,-59.938],xq=[.9177,-4.279,-1.38],Sq=[1.2937,-5.709,.0186];function $w(n){let e=n.length,t=0;for(let o=0;o<e;o++)t+=n[o];let r=t/e,s=0;for(let o=0;o<e;o++){let i=n[o]-r;s+=i*i}return{n:e,mean:r,variance:s/(e-1)}}l($w,"summary");function pN(n){let e=n.length,t=0;for(let a=0;a<e;a++)t+=n[a];let r=t/e,s=0,o=0,i=0;for(let a=0;a<e;a++){let u=n[a]-r,p=u*u;s+=p,o+=p*u,i+=p*p}return s/=e,o/=e,i/=e,{n:e,mean:r,m2:s,skew:o/Math.pow(s,1.5),kurt:i/(s*s)}}l(pN,"centralMoments");function Sa(n){return Math.min(Math.max(n,0),1)}l(Sa,"clampP");function cN(n,e){return Sa(2*(1-Ew.cdf(Math.abs(n),e)))}l(cN,"tPvalue");function Fw(n,e){return Sa(1-ep.cdf(n,e))}l(Fw,"chi2Pvalue");function fN(n,e={}){let{data:t}=Je(n),r=e.popmean??0,{n:s,mean:o,variance:i}=$w(t),a=(o-r)/Math.sqrt(i/s),u=s-1;return{statistic:a,pvalue:cN(a,u),df:u}}l(fN,"tTest1Samp");function _D(n,e,t={}){let r=$w(Je(n).data),s=$w(Je(e).data);if(t.equalVar??!0){let p=r.n+s.n-2,c=((r.n-1)*r.variance+(s.n-1)*s.variance)/p,f=(r.mean-s.mean)/Math.sqrt(c*(1/r.n+1/s.n));return{statistic:f,pvalue:cN(f,p),df:p}}let o=r.variance/r.n,i=s.variance/s.n,a=(r.mean-s.mean)/Math.sqrt(o+i),u=(o+i)*(o+i)/(o*o/(r.n-1)+i*i/(s.n-1));return{statistic:a,pvalue:cN(a,u),df:u}}l(_D,"tTestInd");function bD(n,e,t={}){let r=Je(n).data,s=Je(e).data;if(r.length!==s.length)throw new Error(`tTestPaired: samples must have equal length, got ${r.length} and ${s.length}`);let o=new Float64Array(r.length);for(let i=0;i<r.length;i++)o[i]=r[i]-s[i];return fN(o,{popmean:t.popmean??0})}l(bD,"tTestPaired");function wD(n,e,t={}){let r=Je(n).data,s=r.length,o;if(e==null){let u=0;for(let p=0;p<s;p++)u+=r[p];o=new Float64Array(s).fill(u/s)}else o=Je(e).data;let i=0;for(let u=0;u<s;u++){let p=r[u]-o[u];i+=p*p/o[u]}let a=s-1-(t.ddof??0);return{statistic:i,pvalue:Fw(i,a),df:a}}l(wD,"chi2Gof");function TD(n){let{data:e,rows:t,cols:r}=Cw(n),s=new Float64Array(t),o=new Float64Array(r),i=0;for(let p=0;p<t;p++)for(let c=0;c<r;c++){let f=e[p*r+c];s[p]+=f,o[c]+=f,i+=f}let a=0;for(let p=0;p<t;p++)for(let c=0;c<r;c++){let f=s[p]*o[c]/i,d=e[p*r+c]-f;a+=d*d/f}let u=(t-1)*(r-1);return{statistic:a,pvalue:Fw(a,u),df:u}}l(TD,"chi2Independence");function vq(n){if(n<=0)return 1;let e=0;for(let t=1;t<=gq;t++){let r=Math.exp(-2*t*t*n*n);if(e+=t%2===1?r:-r,r<yq)break}return Sa(2*e)}l(vq,"kolmogorovSf");function xD(n,e){let t=Math.sqrt(e);return vq(n*(t+_q+bq/t))}l(xD,"ksPvalue");function SD(n,e,t={}){let{data:r}=Je(n),s=Float64Array.from(r).sort(),o=s.length,i=t.loc??0,a=t.scale??1,u=typeof e=="function"?e:c=>Uo((c-i)/a),p=0;for(let c=0;c<o;c++){let f=u(s[c]);p=Math.max(p,(c+1)/o-f,f-c/o)}return{statistic:p,pvalue:xD(p,o)}}l(SD,"ksTest1Samp");function vD(n,e){let t=Float64Array.from(Je(n).data).sort(),r=Float64Array.from(Je(e).data).sort(),s=t.length,o=r.length,i=0,a=0,u=0;for(;i<s&&a<o;){let c=Math.min(t[i],r[a]);for(;i<s&&t[i]<=c;)i++;for(;a<o&&r[a]<=c;)a++;u=Math.max(u,Math.abs(i/s-a/o))}let p=s*o/(s+o);return{statistic:u,pvalue:xD(u,p)}}l(vD,"ksTest2Samp");function kD(n){let{n:e,skew:t,kurt:r}=pN(Je(n).data),s=e/6*(t*t+(r-3)*(r-3)/4);return{statistic:s,pvalue:Fw(s,2),df:2}}l(kD,"jarqueBera");function kq(n){let{n:e,skew:t}=pN(n),r=t*Math.sqrt((e+1)*(e+3)/(6*(e-2))),s=3*(e*e+27*e-70)*(e+1)*(e+3)/((e-2)*(e+5)*(e+7)*(e+9)),o=-1+Math.sqrt(2*(s-1)),i=1/Math.sqrt(.5*Math.log(o)),a=Math.sqrt(2/(o-1)),u=r/a;return i*Math.log(u+Math.sqrt(u*u+1))}l(kq,"skewZ");function Nq(n){let{n:e,kurt:t}=pN(n),r=3*(e-1)/(e+1),s=24*e*(e-2)*(e-3)/((e+1)*(e+1)*(e+3)*(e+5)),o=(t-r)/Math.sqrt(s),i=6*(e*e-5*e+2)/((e+7)*(e+9))*Math.sqrt(6*(e+3)*(e+5)/(e*(e-2)*(e-3))),a=6+8/i*(2/i+Math.sqrt(1+4/(i*i))),u=1-2/(9*a),p=1+o*Math.sqrt(2/(a-4));return(u-Math.cbrt((1-2/a)/p))/Math.sqrt(2/(9*a))}l(Nq,"kurtosisZ");function ND(n){let{data:e}=Je(n),t=kq(e),r=Nq(e),s=t*t+r*r;return{statistic:s,pvalue:Fw(s,2),df:2}}l(ND,"dagostinoK2");function Pw(n,e){return n[0]+n[1]*e+n[2]*e*e}l(Pw,"poly2");function Aq(n){return n>=lN[2]?Sa(Math.exp(Pw(Sq,n))):n>lN[1]?Sa(Math.exp(Pw(xq,n))):n>lN[0]?Sa(1-Math.exp(Pw(Tq,n))):Sa(1-Math.exp(Pw(wq,n)))}l(Aq,"adPvalue");function AD(n){let{data:e}=Je(n),t=Float64Array.from(e).sort(),r=t.length,{mean:s,variance:o}=$w(t),i=Math.sqrt(o),a=0;for(let c=0;c<r;c++){let f=Uo((t[c]-s)/i),d=Uo(-(t[r-1-c]-s)/i);a+=(2*c+1)*(Math.log(f)+Math.log(d))}let u=-r-a/r,p=u*(1+yD[0]/r+yD[1]/(r*r));return{statistic:u,pvalue:Aq(p)}}l(AD,"andersonDarling");function Oq(n){let e=n.length,t=Array.from({length:e},(i,a)=>a);t.sort((i,a)=>n[i]-n[a]);let r=new Float64Array(e),s=0,o=0;for(;o<e;){let i=o;for(;i+1<e&&n[t[i+1]]===n[t[o]];)i++;let a=(o+i+2)/2;for(let p=o;p<=i;p++)r[t[p]]=a;let u=i-o+1;s+=u*u*u-u,o=i+1}return{ranks:r,tieSum:s}}l(Oq,"rankData");function OD(n,e){let t=Je(n).data,r=Je(e).data,s=t.length,o=r.length,i=new Float64Array(s+o);i.set(t),i.set(r,s);let{ranks:a,tieSum:u}=Oq(i),p=0;for(let y=0;y<s;y++)p+=a[y];let c=p-s*(s+1)/2,f=s+o,d=s*o/2,h=Math.sqrt(s*o/12*(f+1-u/(f*(f-1)))),m=c-d,g=(m-.5*Math.sign(m))/h;return{statistic:c,pvalue:Sa(2*(1-Uo(Math.abs(g))))}}l(OD,"mannWhitneyU");wn();function Ai(n,e){return Number(n[e])}l(Ai,"f64");function BD(n){return Math.max(1,Math.min(n-1,Math.floor(10*Math.log10(n))))}l(BD,"defaultNlags");function Bq(n,e,t){let r=n.length,s=0;for(let d=0;d<r;d++)s+=n[d];let o=s/r,i=Dw(2*r),a=new Float64Array(i);for(let d=0;d<r;d++)a[d]=n[d]-o;let u=Qe(Fu(Q(a,[i],"f64",e))),p=new Float64Array(2*i);for(let d=0;d<i;d++)p[2*d]=Ai(u,2*d)*Ai(u,2*d)+Ai(u,2*d+1)*Ai(u,2*d+1);let c=Qe(tp(Q(p,[i,2],"f64",e))),f=new Float64Array(t+1);for(let d=0;d<=t;d++)f[d]=Ai(c,2*d)/r;return f}l(Bq,"autocovariance");function dN(n,e,t){let r=Bq(n,e,t),s=new Float64Array(t+1);for(let o=0;o<=t;o++)s[o]=r[o]/r[0];return s}l(dN,"normalizedAcf");function RD(n,e={}){let{data:t,device:r,dtype:s}=Je(n),o=e.nlags??BD(t.length),i=dN(t,r,o);return Q(i,[o+1],s,r)}l(RD,"acf");function MD(n,e={}){let{data:t,device:r,dtype:s}=Je(n),o=e.nlags??BD(t.length),i=dN(t,r,o),a=new Float64Array(o+1);a[0]=1;let u=new Float64Array(o+1),p=new Float64Array(o+1);for(let c=1;c<=o;c++){let f=i[c],d=1;for(let g=1;g<c;g++)f-=u[g]*i[c-g],d-=u[g]*i[g];let h=f/d;p[c]=h;for(let g=1;g<c;g++)p[g]=u[g]-h*u[c-g];a[c]=h;let m=u;u=p,p=m}return Q(a,[o+1],s,r)}l(MD,"pacf");function ID(n,e={}){let{data:t,device:r}=Je(n),s=t.length,o=e.lags??Math.min(10,s-1),i=dN(t,r,o),a=0;for(let p=1;p<=o;p++)a+=i[p]*i[p]/(s-p);a*=s*(s+2);let u=o-(e.modelDf??0);return{statistic:a,pvalue:Math.min(Math.max(1-ep.cdf(a,u),0),1),df:u}}l(ID,"ljungBox");function ED(n){let{data:e}=Je(n),t=0,r=0;for(let s=0;s<e.length;s++)if(r+=e[s]*e[s],s>0){let o=e[s]-e[s-1];t+=o*o}return t/r}l(ED,"durbinWatson");function LD(n,e={}){let{data:t,device:r,dtype:s}=Je(n),o=t.length,i=t;if(e.detrend!==!1){let c=0;for(let d=0;d<o;d++)c+=t[d];let f=c/o;i=new Float64Array(o);for(let d=0;d<o;d++)i[d]=t[d]-f}let a=Qe(Fu(Q(i,[o],"f64",r))),u=Math.floor(o/2),p=new Float64Array(u+1);for(let c=0;c<=u;c++)p[c]=(Ai(a,2*c)*Ai(a,2*c)+Ai(a,2*c+1)*Ai(a,2*c+1))/o;return Q(p,[u+1],s,r)}l(LD,"periodogram");wn();var Rq=200,Mq=1e-13,CD=.4,DD=.9;function vh(n,e){return Number(n[e])}l(vh,"f64");function Iq(n){let e=new Float64Array(n.length);for(let t=0;t<n.length;t++)e[t]=Number(n[t]);return e}l(Iq,"float64From");function PD(n,e,t){let r=new Float64Array(e);return r.set(n),Qe(Fu(Q(r,[e],"f64",t)))}l(PD,"fftInterleaved");function Eq(n,e,t){let r=n+e-1;if(t==="full")return{offset:0,length:r};let s=Math.max(n,e),o=Math.min(n,e);if(t==="same")return{offset:o-1>>1,length:s};if(t==="valid")return{offset:o-1,length:s-o+1};throw new Error(`convolve: unknown mode '${t}', expected 'full', 'same', or 'valid'`)}l(Eq,"modeSlice");function FD(n,e,t,r,s){let o=n.length,i=e.length,a=Dw(o+i-1),u=PD(n,a,r),p=PD(e,a,r),c=new Float64Array(2*a);for(let g=0;g<a;g++){let y=vh(u,2*g),_=vh(u,2*g+1),b=vh(p,2*g),w=vh(p,2*g+1);c[2*g]=y*b-_*w,c[2*g+1]=y*w+_*b}let f=Qe(tp(Q(c,[a,2],"f64",r))),{offset:d,length:h}=Eq(o,i,t),m=new Float64Array(h);for(let g=0;g<h;g++)m[g]=vh(f,2*(d+g));return Q(m,[h],s,r)}l(FD,"fftConvolve");function zD(n,e,t={}){let r=Je(n),s=Je(e);return FD(r.data,s.data,t.mode??"full",r.device,r.dtype)}l(zD,"convolve");function VD(n,e,t={}){let r=Je(n),s=Je(e),o=new Float64Array(s.data.length);for(let i=0;i<s.data.length;i++)o[i]=s.data[s.data.length-1-i];return FD(r.data,o,t.mode??"full",r.device,r.dtype)}l(VD,"correlate");function zw(n,e){let t=Math.floor(n);if(!(t>=1&&t<=e))throw new Error(`rolling: window must be in [1, ${e}], got ${n}`);return t}l(zw,"checkWindow");function GD(n,e){let{data:t,device:r,dtype:s}=Je(n),o=t.length,i=zw(e,o),a=new Float64Array(o+1);for(let p=0;p<o;p++)a[p+1]=a[p]+t[p];let u=new Float64Array(o-i+1);for(let p=0;p<u.length;p++)u[p]=a[p+i]-a[p];return Q(u,[u.length],s,r)}l(GD,"rollingSum");function jD(n,e){let{data:t,device:r,dtype:s}=Je(n),o=t.length,i=zw(e,o),a=new Float64Array(o+1);for(let p=0;p<o;p++)a[p+1]=a[p]+t[p];let u=new Float64Array(o-i+1);for(let p=0;p<u.length;p++)u[p]=(a[p+i]-a[p])/i;return Q(u,[u.length],s,r)}l(jD,"rollingMean");function UD(n,e,t={}){let{data:r,device:s,dtype:o}=Je(n),i=r.length,a=zw(e,i),u=t.ddof??1,p=0;for(let m=0;m<i;m++)p+=r[m];let c=p/i,f=new Float64Array(i+1),d=new Float64Array(i+1);for(let m=0;m<i;m++){let g=r[m]-c;f[m+1]=f[m]+g,d[m+1]=d[m]+g*g}let h=new Float64Array(i-a+1);for(let m=0;m<h.length;m++){let g=f[m+a]-f[m],y=d[m+a]-d[m];h[m]=Math.sqrt(Math.max((y-g*g/a)/(a-u),0))}return Q(h,[h.length],o,s)}l(UD,"rollingStd");function qD(n,e,t){let{data:r,device:s,dtype:o}=Je(n),i=r.length,a=zw(e,i),u=new Float64Array(i-a+1),p=new Int32Array(i),c=0,f=0;for(let d=0;d<i;d++){for(;f>c&&t(r[d],r[p[f-1]]);)f--;p[f++]=d,p[c]<=d-a&&c++,d>=a-1&&(u[d-a+1]=r[p[c]])}return Q(u,[u.length],o,s)}l(qD,"rollingExtreme");function WD(n,e){return qD(n,e,(t,r)=>t<=r)}l(WD,"rollingMin");function KD(n,e){return qD(n,e,(t,r)=>t>=r)}l(KD,"rollingMax");function HD(n,e,t){let r=Je(n),s=Je(e),o=r.data.length;if(s.data.length!==o)throw new Error(`polyfit: x and y must have equal length, got ${o} and ${s.data.length}`);let i=t+1,a=new Float64Array(o*i);for(let p=0;p<o;p++){a[p*i+t]=1;for(let c=t-1;c>=0;c--)a[p*i+c]=a[p*i+c+1]*r.data[p]}let u=bh(Q(a,[o,i],"f64",r.device),Q(s.data,[o,1],"f64",r.device));return Q(Iq(Qe(u)),[i],"f64",r.device)}l(HD,"polyfit");function $D(n,e){let t=0;for(let r=0;r<n.length;r++)t=t*e+n[r];return t}l($D,"horner");function YD(n,e){let t=Je(n).data;if(typeof e=="number")return $D(t,e);let r=Je(e),s=new Float64Array(r.data.length);for(let o=0;o<s.length;o++)s[o]=$D(t,r.data[o]);return Q(s,[s.length],r.dtype,r.device)}l(YD,"polyval");function JD(n){let{data:e,device:t}=Je(n),r=0;for(;r<e.length-1&&e[r]===0;)r++;let s=e.length-1-r;if(s<1)return Q(new Float64Array(0),[0,2],"f64",t);let o=new Float64Array(s+1);for(let f=0;f<=s;f++)o[f]=e[r+f]/e[r];let i=new Float64Array(s),a=new Float64Array(s),u=1,p=0;for(let f=0;f<s;f++){let d=u*CD-p*DD;p=u*DD+p*CD,u=d,i[f]=u,a[f]=p}for(let f=0;f<Rq;f++){let d=0;for(let h=0;h<s;h++){let m=1,g=0;for(let S=1;S<=s;S++){let T=m*i[h]-g*a[h]+o[S];g=m*a[h]+g*i[h],m=T}let y=1,_=0;for(let S=0;S<s;S++){if(S===h)continue;let T=i[h]-i[S],v=a[h]-a[S],A=y*T-_*v;_=y*v+_*T,y=A}let b=y*y+_*_,w=(m*y+g*_)/b,x=(g*y-m*_)/b;i[h]-=w,a[h]-=x,d=Math.max(d,Math.hypot(w,x))}if(d<Mq)break}let c=new Float64Array(2*s);for(let f=0;f<s;f++)c[2*f]=i[f],c[2*f+1]=a[f];return Q(c,[s,2],"f64",t)}l(JD,"polyroots");be();var Lq=.0331;function Cq(n){let e=Array.isArray(n)?n:[n],t=1;for(let r of e)t*=r;return{shp:e,n:t}}l(Cq,"normalizeShape");var Vw=class{static{l(this,"Generator")}_next;constructor(e){this._next=xr(e)}_uniformPositive(){let e=this._next();for(;e<=0;)e=this._next();return e}_normalDraw(){return xa(this._uniformPositive())}_gammaDraw(e){if(e<1)return this._gammaDraw(e+1)*Math.pow(this._uniformPositive(),1/e);let t=e-1/3,r=1/Math.sqrt(9*t);for(;;){let s,o;do s=this._normalDraw(),o=1+r*s;while(o<=0);o=o*o*o;let i=this._uniformPositive();if(i<1-Lq*s*s*s*s||Math.log(i)<.5*s*s+t*(1-o+Math.log(o)))return t*o}}_fill(e,t,r){let{shp:s,n:o}=Cq(e),i=new Float64Array(o);for(let a=0;a<o;a++)i[a]=r();return Q(i,s,t.dtype??"f64",t.device)}uniform(e,t={}){let r=t.low??0,s=t.high??1;return this._fill(e,t,()=>r+(s-r)*this._next())}normal(e,t={}){let r=t.loc??0,s=t.scale??1;return this._fill(e,t,()=>r+s*this._normalDraw())}standardT(e,t={}){let r=t.df;return this._fill(e,t,()=>{let s=this._normalDraw(),o=2*this._gammaDraw(r/2);return s/Math.sqrt(o/r)})}chi2(e,t={}){let r=t.df;return this._fill(e,t,()=>2*this._gammaDraw(r/2))}exponential(e,t={}){let r=t.scale??1;return this._fill(e,t,()=>-r*Math.log(this._uniformPositive()))}multivariateNormal(e,t,r=1,s={}){let o=Je(e),i=Cw(t);if(i.rows!==i.cols||i.rows!==o.data.length)throw new Error(`multivariateNormal: mean of length ${o.data.length} incompatible with ${i.rows}x${i.cols} covariance`);let a=i.rows,u=s.device??i.device,p=s.dtype??"f64",c=Ok(Q(i.data,[a,a],p,u)),f=this.normal([r,a],{device:u,dtype:p}),d=Q(o.data,[a],p,u);return ae(tt(f,c.transpose(0,1)),d)}};jM();iI();hI();vI();FI();KI();nE();rE(Ze);export{dh as Accuracy,bd as Adam,wd as AdamW,Xf as AdaptiveAvgPool2d,Qf as AvgPool2d,rd as BCELoss,cc as BatchNorm1d,Jf as BatchNorm2d,vu as BatchSampler,go as CPUTarget,Ws as CPU_DEVICE,lh as CSVLogger,xm as CUDATarget,Zn as Callback,yh as ConfusionMatrix,Du as ConsoleLogger,zf as Conv1d,Ff as Conv2d,xd as CosineAnnealingLR,td as CrossEntropyLoss,md as DataLoader,Tu as Dataset,Sr as Dropout,Kf as ELU,oh as EarlyStopping,sd as Embedding,fE as F,gh as F1Score,cd as Flatten,Gf as GELU,bg as GPU_DEVICE,id as GRU,pc as GRUCell,Oo as GradMode,uh as GradientAccumulationScheduler,Au as LRScheduler,ad as LSTM,fc as LSTMCell,no as LayerNorm,Wf as LeakyReLU,ih as LearningRateMonitor,th as LightningModule,Sn as Linear,Yf as LogSoftmax,ki as Logger,ed as MSELoss,hd as MapDataset,Zf as MaxPool2d,ch as MeanMetric,lr as Metric,fh as MetricCollection,Ta as ModelCheckpoint,ge as Module,ld as ModuleDict,aa as ModuleList,ua as MultiheadAttention,nd as NLLLoss,fs as Optimizer,At as Parameter,fd as PositionalEncoding,hh as Precision,wa as ProgressCallback,Su as RandomSampler,Vf as ReLU,mh as Recall,Sd as ReduceLROnPlateau,_d as SGD,la as Sampler,ud as Sequential,xu as SequentialSampler,jf as SiLU,Uf as Sigmoid,Hf as Softmax,Td as StepLR,ph as SumMetric,bt as SymbolicTensor,qf as Tanh,Ze as Tensor,dd as TensorDataset,ah as Timer,yd as Tokenizer,le as TraceLevel,sh as Trainer,pd as Transformer,hc as TransformerDecoder,wu as TransformerDecoderLayer,dc as TransformerEncoder,bu as TransformerEncoderLayer,ca as Vocab,RT as WASM_DEVICE,xo as WEBGPU_DEVICE,Sm as WasmTarget,vm as WebGPUTarget,Hp as abs,ae as add,Sk as applyCheckpoint,RB as arange,Cg as argmax,iB as argmin,gB as argsort,cB as broadcast_in_dim,ru as cat,nB as ceil,pB as chunk,Mg as clamp,vd as clipGradNorm_,kd as clipGradValue_,lB as clone,eU as compile,_w as compileWithBackward,ls as contiguous,ZO as cos,hB as cumsum,hE as data,ku as defaultCollate,Lt as dispatcher,Kt as div,uB as dot,Tn as empty,$G as emptyLike,KB as enableGrad,Yp as eq,QO as erf,XO as erfc,as as exp,Ll as expand,jg as eye,dB as flip,tB as floor,HR as flushWebGPUEager,eu as fromBuffer,ot as full,VG as fullLike,kg as gamma,Eg as gather,Rg as ge,Ng as gelu,ea as getDefaultDevice,Il as gt,ra as index_select,T_ as init,Bg as le,eB as lgamma,$C as lightning,FC as linalg,IB as linspace,xk as loadCheckpoint,ko as log,$g as log_softmax,sB as lt,tt as matmul,No as max,ta as maximum,Tr as mean,vn as memfs,oB as min,vg as minimum,jC as ml,ee as mul,ir as narrow,rB as ne,is as neg,dE as nn,WB as noGrad,ZD as numeric,Ig as one_hot,tn as ones,zG as onesLike,Ht as ops,tC as optim,tu as pad,El as permute,vo as pow,ZR as preloadCudaRuntime,gf as preloadWebGPU,PG as printModule,aB as prod,Gg as randn,GG as randnLike,MB as randperm,Ml as relu,Dl as repeat,Hn as reshape,fB as roll,YO as rsqrt,Kp as scalar,od as scan,Lg as scatter,nu as scatter_add,en as select,rh as serializeCheckpoint,GF as setDefaultDevice,wr as sigmoid,Og as sign,Ag as silu,JO as sin,Dg as slice,mi as softmax,mB as sort,hi as split,na as sqrt,Cl as squeeze,us as stack,ze as sub,wt as sum,Hs as tanh,Wt as tensor,Pg as tile,TE as tokenizer,yB as topk,kC as trace,vt as transpose,Lr as unsqueeze,or as where,Ve as zeros,FG as zerosLike};