@slexisvn/mlfw 0.1.1 → 0.1.2

package/dist/index.browser.js

@@ -0,0 +1,284 @@
+var SS=Object.defineProperty;var l=(r,e)=>SS(r,"name",{value:e,configurable:!0});var M=(r,e)=>()=>(r&&(e=r(r=0)),e);var nt=(r,e)=>{for(var t in e)SS(r,t,{get:e[t],enumerable:!0})};function oh(r,e){let t=e||"mlc",n=r.indexOf("->"),s=n>=0?r.substring(0,n).trim():r.trim(),o=n>=0?r.substring(n+2).trim():"",i=s.indexOf("("),a=s.lastIndexOf(")"),u,p="",c=s.substring(0,i).trim(),h=c.indexOf(".");h>=0?(u=c.substring(0,h),p=c.substring(h+1)):u=c;let f=s.substring(i+1,a).trim(),d=f.length>0?ME(f):[],m=o.length>0?IE(o):[{kind:"Tensor"}];return new sb(t,u,p,d,m)}function ME(r){return kS(r,",").map(t=>{let n=t.trim(),s=n.indexOf("="),o=n,i;s>=0&&(o=n.substring(0,s).trim(),i=n.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 h=ob.get(p)||"Scalar";return new rb(c,h,i,a)})}function IE(r){let e=r.trim();if(e.startsWith("(")&&e.endsWith(")")){let n=e.substring(1,e.length-1);return kS(n,",").map(o=>{let i=o.trim();return{kind:ob.get(i)||"Tensor"}})}return[{kind:ob.get(e)||"Tensor"}]}function kS(r,e){let t=[],n=0,s=0;for(let o=0;o<r.length;o++){let i=r[o];i==="("||i==="["?n++:i===")"||i==="]"?n--:i===e&&n===0&&(t.push(r.substring(s,o)),s=o+1)}return t.push(r.substring(s)),t}var RE,rb,sb,ob,ib=M(()=>{"use strict";RE=new Set(["Tensor","Tensor[]"]),rb=class{static{l(this,"SchemaArg")}name;kind;defaultValue;isOut;constructor(e,t,n,s){this.name=e,this.kind=t,this.defaultValue=n??null,this.isOut=s??!1}get isTensor(){return RE.has(this.kind)}},sb=class{static{l(this,"OperatorSchema")}namespace;name;overload;args;returns;_key;_tensorArgIndices;constructor(e,t,n,s,o){this.namespace=e,this.name=t,this.overload=n||"",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}},ob=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(oh,"parseSchema");l(ME,"_parseArgs");l(IE,"_parseReturns");l(kS,"_splitTopLevel")});function vS(r){return typeof r=="object"&&r!==null&&"_impl"in r}function NS(r){return r instanceof Dr?r:vS(r)?Dr.tensor(r):typeof r=="number"?Number.isInteger(r)?Dr.int(r):Dr.float(r):typeof r=="boolean"?Dr.bool(r):typeof r=="string"?Dr.string(r):Array.isArray(r)?r.length>0&&vS(r[0])?Dr.tensorList(r):Dr.intList(r):(r==null,Dr.none())}var Dr,wr,nc=M(()=>{"use strict";Dr=class r{static{l(this,"IValue")}tag;value;constructor(e,t){this.tag=e,this.value=t}static tensor(e){return new r(0,e)}static int(e){return new r(1,e)}static float(e){return new r(2,e)}static bool(e){return new r(3,e)}static intList(e){return new r(4,e)}static tensorList(e){return new r(5,e)}static string(e){return new r(6,e)}static none(){return new r(7,null)}static device(e){return new r(8,e)}static dtype(e){return new r(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}},wr=class r{static{l(this,"KernelFunction")}_boxed;_unboxed;constructor(e,t){this._boxed=e||null,this._unboxed=t||null}static fromBoxed(e){return new r(e)}static fromUnboxed(e){return new r(void 0,e)}static fromBoth(e,t){return new r(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 n=t.map(i=>NS(i)),s=this._boxed;if(!s)return;let o=s(e,n);return Array.isArray(o)?o.length===1?o[0].value:o.map(i=>i.value):o instanceof Dr?o.value:o}_callUnboxedAsBoxed(e,t){let n=t.map(i=>i.value),s=this._unboxed;if(!s)return[];let o=s(e,...n);return o==null?[]:[NS(o)]}};l(vS,"hasImpl");l(NS,"_toIValue")});function AS(r){_o=r}var _o,Xt,gs=M(()=>{"use strict";ib();nc();_o=null;l(AS,"_setDispatcher");Xt=class{static{l(this,"Library")}_namespace;_kind;_registrations;constructor(e,t){this._namespace=e,this._kind=t,this._registrations=[]}def(e){let t=oh(e,this._namespace);return _o&&_o.registerOp(t),this._registrations.push({type:"def",schema:t}),this}impl(e,t,n){let s=n instanceof wr?n:wr.fromUnboxed(n);return _o&&_o.registerKernel(`${this._namespace}::${e}`,t,s),this._registrations.push({type:"impl",name:e,key:t,kernelFn:s}),this}implBoxed(e,t,n){let s=wr.fromBoxed(n);return _o&&_o.registerKernel(`${this._namespace}::${e}`,t,s),this._registrations.push({type:"impl",name:e,key:t,kernelFn:s}),this}fallback(e,t){let n=t instanceof wr?t:wr.fromBoxed(t);return _o&&_o.registerFallback(e,n),this._registrations.push({type:"fallback",key:e,kernelFn:n}),this}replay(e){for(let t of this._registrations)t.type==="def"?e.registerOp(t.schema):t.type==="impl"?e.registerKernel(`${this._namespace}::${t.name}`,t.key,t.kernelFn):t.type==="fallback"&&e.registerFallback(t.key,t.kernelFn)}}});function ja(r){let e=EE[r];if(e===void 0)throw new Error(`No backend key for device: ${r}`);return e}function RS(r){return LE[r]??40}function OS(r){return r=r-(r>>1&1431655765),r=(r&858993459)+(r>>2&858993459),(r+(r>>4)&252645135)*16843009>>24}var tr,BS,EE,LE,Pr,rc,cG,sc,pG,Nn=M(()=>{"use strict";tr=(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))(tr||{}),BS=new Array(49).fill(null);for(let[r,e]of Object.entries(tr))r!=="NUM_KEYS"&&typeof e=="number"&&e<49&&(BS[e]=r);EE=Object.freeze({cpu:0,gpu:1,wasm:2,webgpu:5,meta:3,lazy:4}),LE=Object.freeze({0:41,1:42,2:43});l(ja,"backendKeyForDevice");l(RS,"autogradKeyForBackend");Pr=class r{static{l(this,"DispatchKeySet")}_lo;_hi;constructor(e,t){this._lo=e|0,this._hi=t|0}static fromKey(e){return e<32?new r(1<<e,0):new r(0,1<<e-32)}static fromKeys(...e){let t=0,n=0;for(let s=0;s<e.length;s++){let o=e[s];o<32?t|=1<<o:n|=1<<o-32}return new r(t,n)}add(e){return e<32?new r(this._lo|1<<e,this._hi):new r(this._lo,this._hi|1<<e-32)}remove(e){return e<32?new r(this._lo&~(1<<e),this._hi):new r(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 r(this._lo|e._lo,this._hi|e._hi)}intersect(e){return new r(this._lo&e._lo,this._hi&e._hi)}subtract(e){return new r(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 OS(this._lo)+OS(this._hi)}*[Symbol.iterator](){let e=this._hi;for(;e!==0;){let n=31-Math.clz32(e);yield n+32,e&=~(1<<n)}let t=this._lo;for(;t!==0;){let n=31-Math.clz32(t);yield n,t&=~(1<<n)}}toString(){let e=[];for(let t of this)e.push(BS[t]||String(t));return`DispatchKeySet(${e.join(", ")})`}};l(OS,"_popcount32");rc=new Pr(0,0),cG=Pr.fromKeys(0,1,2,3,4,5,6,7,8),sc=Pr.fromKeys(40,41,42,43),pG=Pr.fromKeys(20,24,28,32,40,41,42,43,48)});var ih,MS=M(()=>{"use strict";Nn();ih=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 n=this._kernels[t];if(n)return{key:t,kernel:n}}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 ah,IS=M(()=>{"use strict";ah=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 uh,ES=M(()=>{"use strict";Nn();uh=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 CE(r){return typeof r=="object"&&r!==null&&"then"in r&&typeof r.then=="function"}function LS(r,e){Pi.push(void 0,r);let t;try{t=e()}catch(n){throw Pi.pop(),n}return CE(t)?t.then(n=>(Pi.pop(),n),n=>{throw Pi.pop(),n}):(Pi.pop(),t)}var ab,Pi,ub=M(()=>{"use strict";Nn();ab=class{static{l(this,"GuardStack")}_frames;constructor(){this._frames=[]}push(e,t){this._frames.push({exclude:e||rc,include:t||rc})}pop(){return this._frames.pop()}apply(e){let t=e;for(let n=this._frames.length-1;n>=0;n--){let s=this._frames[n];t=t.subtract(s.exclude),t=t.union(s.include)}return t}get depth(){return this._frames.length}clear(){this._frames.length=0}},Pi=new ab;l(CE,"isThenable");l(LS,"withIncludedKeys")});function DE(r){return typeof r=="object"&&r!==null&&"dispatchKeySet"in r}function cb(r,e){if(!e)return r;if(DE(e)&&e.dispatchKeySet)return r.union(e.dispatchKeySet);if(Array.isArray(e))for(let t=0;t<e.length;t++)r=cb(r,e[t]);return r}function pb(r,e){let t=rc;if(e){let n=e.tensorArgIndices;for(let s=0;s<n.length;s++){let o=n[s];o<r.length&&(t=cb(t,r[o]))}}else for(let n=0;n<r.length;n++)t=cb(t,r[n]);return t}var lb,Nt,$i=M(()=>{"use strict";Nn();MS();IS();ES();ub();gs();ib();l(DE,"hasDispatchKeySet");lb=class{static{l(this,"Dispatcher")}_entries;_handles;_fallbacks;constructor(){this._entries=new Map,this._handles=new Map,this._fallbacks=new uh}registerOp(e){let t=e.key(),n=this._handles.get(t);if(n)return n;let s=new ih(e);this._entries.set(t,s);let o=new ah(s,e);return this._handles.set(t,o),o}findOp(e){let t=e.includes("::")?e:`mlc::${e}`;return this._handles.get(t)||null}findOrRegisterOp(e){let t=this.findOp(e);if(!t){let n=oh(`${e}() -> Tensor`,"mlc");t=this.registerOp(n)}return t}registerKernel(e,t,n){let s=e.includes("::")?e:`mlc::${e}`,o=this._entries.get(s);if(!o)throw new Error(`Op '${s}' not registered`);o.registerKernel(t,n)}registerFallback(e,t){this._fallbacks.register(e,t)}dispatch(e,t,...n){let s=Pi.apply(t);return this._dispatchInternal(e,s,n)}redispatch(e,t,...n){return this._dispatchInternal(e,t,n)}_dispatchInternal(e,t,n){let s=t.highestPriority();if(s<0)throw new Error(`No dispatch key found for op '${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 ${s}`);let i=t.without(s);return o.callUnboxed(i,...n)}callOp(e,...t){let n=this.findOp(e);if(!n)throw new Error(`Op '${e}' not found`);let s=pb(t,n.schema);return this.dispatch(n,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(cb,"_unionArg");l(pb,"computeKeySet");Nt=new lb;AS(Nt)});function qa(r){return $E+r}var PE,$E,Ce,Vs=M(()=>{"use strict";PE={ceildiv:"ceilDiv"},$E="_sym_";l(qa,"symVarName");Ce=class r{static{l(this,"SymInt")}constructor(e,t=null,n=[]){this.type=e,this.name=t,this.args=n}static var(e){return new r("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:r.equals(e,t)?r.mul(2,e):new r("add",null,[e,t])}static sub(e,t){return typeof e=="number"&&typeof t=="number"?e-t:t===0?e:r.equals(e,t)?0:new r("sub",null,[e,t])}static neg(e){return typeof e=="number"?-e:e instanceof r&&e.type==="neg"?e.args[0]:new r("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 r("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:r.equals(e,t)?1:new r("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||r.equals(e,t)?0:new r("mod",null,[e,t])}static max(e,t){return typeof e=="number"&&typeof t=="number"?Math.max(e,t):e===t||r.equals(e,t)?e:new r("max",null,[e,t])}static min(e,t){return typeof e=="number"&&typeof t=="number"?Math.min(e,t):e===t||r.equals(e,t)?e:new r("min",null,[e,t])}static ceilDiv(e,t){return typeof e=="number"&&typeof t=="number"?Math.ceil(e/t):new r("ceildiv",null,[e,t])}static equals(e,t){if(e===t)return!0;if(typeof e=="number"||typeof t=="number"||!(e instanceof r)||!(t instanceof r)||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 n=r.equals(e.args[0],t.args[0])&&r.equals(e.args[1],t.args[1]),s=r.equals(e.args[0],t.args[1])&&r.equals(e.args[1],t.args[0]);return n||s}for(let n=0;n<e.args.length;n++)if(!r.equals(e.args[n],t.args[n]))return!1;return!0}static substitute(e,t,n){if(typeof e=="number"||!(e instanceof r))return e;if(e.type==="var")return e.name===t?n:e;let s=e.args.map(i=>r.substitute(i,t,n)),o=r[PE[e.type]||e.type];return o&&s.length===2?o(s[0],s[1]):o&&s.length===1?o(s[0]):new r(e.type,e.name,s)}static evaluate(e,t){if(typeof e=="number"||!(e instanceof r))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 n=e.args.map(s=>r.evaluate(s,t));switch(e.type){case"add":return n[0]+n[1];case"sub":return n[0]-n[1];case"mul":return n[0]*n[1];case"div":return Math.floor(n[0]/n[1]);case"mod":return(n[0]%n[1]+n[1])%n[1];case"max":return Math.max(n[0],n[1]);case"min":return Math.min(n[0],n[1]);case"neg":return-n[0];case"ceildiv":return Math.ceil(n[0]/n[1]);default:throw new Error(`Unknown SymInt op: ${e.type}`)}}static freeVars(e,t=new Set){if(typeof e=="number"||!(e instanceof r))return t;if(e.type==="var")return t.add(e.name),t;for(let n of e.args)r.freeVars(n,t);return t}static isConst(e){return typeof e=="number"}static toConst(e){return typeof e=="number"?e:null}toString(){return this.type==="var"?this.name:this.type==="add"?`(${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 lh(r,e){return r===e?!0:r instanceof Ce&&e instanceof Ce?Ce.equals(r,e):!1}function oc(r,e){let t=1;for(let n=0;n<r.length;n++){let s=r[n];if(typeof s!="number"||s<0)return e;t*=s}return t}function mb(r){let e=1;for(let t=0;t<r.length;t++){let n=r[t];if(typeof n=="number"){if(n<0)return X;e=Ce.mul(e,n)}else if(n instanceof Ce)e=Ce.mul(e,n);else return X}return e}function Gs(r){let e=CS[r];if(e===void 0)throw new Error(`Unknown dtype: ${r}`);return e}function Lt(r){return FE.has(r)}function gb(r){return zE.has(r)}function VE(r,e){return lh(r,e)?r:r===1?e:e===1?r:r===X?e===X?X:e:e===X?r:r instanceof Ce||e instanceof Ce?X:null}function db(r,e){return r===e?!0:r&&typeof r.equals=="function"?r.equals(e):!1}function Rn(r){if(r instanceof I)return`tensor<${r.shape.map(t=>t===X?"?":String(t)).join("x")}x${r.dtype}>`;if(r instanceof Fi)return`tuple<${r.elements.map(Rn).join(", ")}>`;if(r instanceof fb)return"token";if(r instanceof hb){let e=r.inputs.map(Rn).join(", "),t=r.outputs.map(Rn).join(", ");return`(${e}) -> (${t})`}return"unknown"}var zn,CS,FE,zE,X,Et,I,Fi,fb,hb,Oe=M(()=>{"use strict";Vs();l(lh,"dimEquals");zn=(h=>(h.F16="f16",h.BF16="bf16",h.F32="f32",h.F64="f64",h.I8="i8",h.I16="i16",h.I32="i32",h.I64="i64",h.UI8="ui8",h.BOOL="bool",h.INDEX="index",h))(zn||{}),CS=Object.freeze({f16:2,bf16:2,f32:4,f64:8,i8:1,i16:2,i32:4,i64:8,ui8:1,bool:1,index:4}),FE=new Set(["f16","bf16","f32","f64"]),zE=new Set(["i8","i16","i32","i64","ui8"]),X=-1;l(oc,"shapeProduct");l(mb,"symbolicShapeProduct");l(Gs,"scalarBytes");l(Lt,"isFloatType");l(gb,"isIntType");Et=class r{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 n=0;n<e;n++)t[n]=n;return new r(t)}static columnMajor(e){let t=new Array(e);for(let n=0;n<e;n++)t[n]=e-1-n;return new r(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 r(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 n=0;n<this.order.length;n++)t[n]=this.order[e.order[n]];return new r(t)}computeStrides(e){let t=e.length,n=new Array(t),s=1;for(let o=t-1;o>=0;o--){let i=this.order[o];n[i]=s,e[i]===X||e[i]instanceof Ce?s=X:s!==X&&(s*=e[i])}return n}equals(e){if(this===e)return!0;if(!(e instanceof r)||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(VE,"broadcastDim");I=class r{static{l(this,"TensorType")}shape;dtype;layout;_hash;constructor(e,t,n=null){this.shape=Object.freeze([...e]),this.dtype=t,this.layout=n||Et.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 Ce)}numel(){return oc(this.shape,X)}symbolicNumel(){return mb(this.shape)}sizeInBytes(){let e=this.numel();return e===X?X:e*CS[this.dtype]}strides(){return this.layout.computeStrides(this.shape)}withShape(e){return new r(e,this.dtype,this.layout)}withDtype(e){return new r(this.shape,e,this.layout)}withLayout(e){return new r(this.shape,this.dtype,e)}equals(e){if(this===e)return!0;if(!(e instanceof r)||this.dtype!==e.dtype||this.shape.length!==e.shape.length)return!1;for(let t=0;t<this.shape.length;t++)if(!lh(this.shape[t],e.shape[t]))return!1;return this.layout.equals(e.layout)}shapeEquals(e){if(!(e instanceof r)||this.dtype!==e.dtype||this.shape.length!==e.shape.length)return!1;for(let t=0;t<this.shape.length;t++)if(!lh(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 n=this.shape[t],s=e.shape[t];if(!(n===X||s===X)&&!lh(n,s)&&typeof n=="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 n=this.shape[t],s=typeof n=="number"?n&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 n=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=VE(o,a[u]);if(p===null)return null;o=p}n[s]=o}return n}static broadcastCompatible(...e){return r.broadcastShape(...e)!==null}},Fi=class r{static{l(this,"TupleType")}elements;constructor(e){this.elements=Object.freeze([...e])}equals(e){if(this===e)return!0;if(!(e instanceof r)||this.elements.length!==e.elements.length)return!1;for(let t=0;t<this.elements.length;t++)if(!db(this.elements[t],e.elements[t]))return!1;return!0}},fb=class r{static{l(this,"TokenType")}equals(e){return e instanceof r}},hb=class r{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 r)||this.inputs.length!==e.inputs.length||this.outputs.length!==e.outputs.length)return!1;for(let t=0;t<this.inputs.length;t++)if(!db(this.inputs[t],e.inputs[t]))return!1;for(let t=0;t<this.outputs.length;t++)if(!db(this.outputs[t],e.outputs[t]))return!1;return!0}};l(db,"typeEquals");l(Rn,"typeToString")});var GE,ic,Wo,ac,ch=M(()=>{"use strict";GE=0,ic=class{static{l(this,"UseLink")}constructor(e,t){this.user=e,this.operandIndex=t,this.prev=null,this.next=null}},Wo=class{static{l(this,"Value")}constructor(e,t=null,n=0){this.type=e,this.definingOp=t,this.resultIndex=n,this.id=GE++,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,n=this._useHead;for(;n;)n.user.operands[n.operandIndex]=e,n=n.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}},ac=class extends Wo{static{l(this,"BlockArgument")}constructor(e,t,n){super(e,null,0),this.ownerBlock=t,this.argIndex=n}_owningFunction(){return this.ownerBlock?this.ownerBlock._owningFunction():null}isBlockArgument(){return!0}}});var yn,Ct,zi=M(()=>{"use strict";ch();yn=class{static{l(this,"Block")}constructor(e=[]){this.parentRegion=null,this._head=null,this._tail=null,this._size=0,this.arguments=e.map((t,n)=>new ac(t,this,n))}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 ac(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 ${s} still has uses`);let n=[];for(let s=0;s<this.arguments.length;s++)t.has(s)||n.push(this.arguments[s]);this.arguments=n;for(let s=0;s<n.length;s++)n[s].argIndex=s;return this._notifyMutation(),this}pushOp(e){if(e.parentBlock!==null)throw new Error(`pushOp: operation '${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()}},Ct=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 an,H,Z,ph,_n=M(()=>{"use strict";an=Object.freeze({NONE:0,READ:1,WRITE:2,ALLOCATE:4,CONTROL:8}),H=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||an.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(H.COMMUTATIVE)}get isAssociative(){return this.traits.has(H.ASSOCIATIVE)}get isElementwise(){return this.traits.has(H.ELEMENTWISE)}get isTerminator(){return this.traits.has(H.TERMINATOR)}get isConstant(){return this.traits.has(H.CONSTANT)}get isReduction(){return this.traits.has(H.REDUCTION)}get isBroadcast(){return this.traits.has(H.BROADCAST)}get isInjective(){return this.traits.has(H.INJECTIVE)}get isOutEWiseFusable(){return this.traits.has(H.OUT_EWISE_FUSABLE)}get isOpaque(){return this.traits.has(H.OPAQUE)}get hasSideEffects(){return this.sideEffects!==an.NONE}},ph=class{static{l(this,"OpRegistry")}constructor(){this._defs=new Map}register(e){if(this._defs.has(e.name))throw new Error(`Op '${e.name}' already registered`);this._defs.set(e.name,e)}registerOpAttr(e,t,n){let s=this._defs.get(e);if(!s)throw new Error(`registerOpAttr: op '${e}' not registered`);return s.setAttr(t,n),s}registerTrait(e,t){let n=this._defs.get(e);if(!n)throw new Error(`registerTrait: op '${e}' not registered`);return n.addTrait(t),n}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 ut,Ua,uc=M(()=>{"use strict";ut=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}},Ua=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,n)=>n.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 n=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?n[i++]=t[s++]:n[i++]=this._generic[o++];for(;s<t.length;)n[i++]=t[s++];for(;o<this._generic.length;)n[i++]=this._generic[o++];return n}hasPatterns(){return this.patterns.length>0}}});function ys(r){return(lc[r]||hh).wasm}function dh(r){return(lc[r]||hh).load}function $S(r){return(lc[r]||hh).store}function cc(r){return(lc[r]||hh).bytes}function yb(r){return PS[r]||null}function _s(r,e){let t=PS[r];return t&&t[e]||null}function Ho(r){return DS[r]||jE}function mh(r){return Ho(r).js}function $r(r){return Ho(r).c}function _b(r){return Ho(r).cPtr}function bb(r){return Ho(r).suffix}function qE(r){return Ho(r).mathSuffix}function bn(r){return Ho(r).bytes}function rr(r){return Ho(r).isFloat}function nr(r){return Ho(r).isInt}function FS(r,e){return r==="sum"||r==="mean"?0:r==="prod"?1:r==="max"?nr(e)?(fh[e]||fh.i32)[0]:-1/0:r==="min"?nr(e)?(fh[e]||fh.i32)[1]:1/0:0}function gh(r,e){let t=UE[r];return t?t+qE(e):r}function Ka(r){let e=KE[r];if(!e)throw new Error(`unsupported compare direction '${r}'`);return e}function zS(r){let e=WE[r];if(!e)throw new Error(`unsupported compare direction '${r}'`);return e}function VS(r){return HE.has(r)}function Mn(r){return(wb[r]||GS).wgsl}function bs(r){return(wb[r]||GS).bytes}function qS(r){return jS[r]||r}function US(r){return r in jS}var DS,lc,hh,PS,jE,fh,UE,KE,WE,HE,wb,GS,jS,Vn=M(()=>{"use strict";DS={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}},lc={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}},hh=lc.f32,PS={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(ys,"wasmType");l(dh,"wasmLoad");l($S,"wasmStore");l(cc,"wasmBytes");l(yb,"wasmSimdEntry");l(_s,"wasmVecOp");jE=DS.f32;l(Ho,"dtypeInfo");l(mh,"jsTypedArray");l($r,"cType");l(_b,"cPtrType");l(bb,"cLiteralSuffix");l(qE,"cMathFuncSuffix");l(bn,"dtypeBytes");l(rr,"isDtypeFloat");l(nr,"isDtypeInt");fh={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(FS,"reduceInitValue");UE={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(gh,"cMathFunc");KE={eq:"==",ne:"!=",lt:"<",le:"<=",gt:">",ge:">="},WE={eq:"===",ne:"!==",lt:"<",le:"<=",gt:">",ge:">="};l(Ka,"cCompareOp");l(zS,"jsCompareOp");HE=new Set(["exp","log","sqrt","tanh","abs","ceil","floor","sin","cos","max","min","pow","round","sign","log2","log10"]);l(VS,"isJSMathFunc");wb={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}},GS=wb.f32;l(Mn,"wgslType");l(bs,"wgslBytes");jS={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(qS,"wgslMathFunc");l(US,"hasWgslMathFunc")});function Ha(r,e){let t={};return r.match(e,t)?t:null}var yh,Tb,xb,Wa,ws,KS=M(()=>{"use strict";yh=class{static{l(this,"DFPattern")}match(e,t){return!1}},Tb=class extends yh{static{l(this,"AnyPattern")}match(){return!0}},xb=class extends yh{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 n=0;n<this.operandPatterns.length;n++){let s=e.getOperand(n).definingOp;if(!this.operandPatterns[n].match(s,t))return!1}return!0}},Wa=l(()=>new Tb,"wildcard"),ws=l((r,...e)=>new xb(r,e),"isOp");l(Ha,"matchPattern")});function js(r,e){return r&&r.opName==="constant"&&r.getAttr("value")===e}function kb(r){return new Sb(r)}var YE,JE,ZE,QE,XE,_h,Ya,bh,wh,Th,xh,Sb,Ja,Za,Qa,Xa,eu,tu,nu,Sh,kh,vh,Nh,Ah,Oh,Bh,Rh,Mh,Ih,ru=M(()=>{"use strict";uc();Oe();Vn();KS();YE=ws("transpose",ws("transpose",Wa())),JE=ws("reshape",ws("reshape",Wa())),ZE=ws("neg",ws("neg",Wa())),QE=ws("exp",ws("log",Wa())),XE=ws("log",ws("exp",Wa()));l(js,"isConstantVal");_h=class extends ut{static{l(this,"FoldTrivialReshape")}constructor(){super("fold_trivial_reshape",10),this.rootOpName="reshape"}match(e){let t=e.getOperand(0).type,n=e.getResult(0).type;return t instanceof I&&n instanceof I&&t.shapeEquals(n)}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},Ya=class extends ut{static{l(this,"ReshapeReshape")}constructor(){super("reshape_reshape",10),this.rootOpName="reshape"}match(e){return Ha(JE,e)!==null}rewrite(e,t){let n=e.getOperand(0).definingOp.getOperand(0),s=e.getAttr("new_shape"),o=t.reshape(n,s);return e.replaceAllResultsWith([o.getResult(0)]),e.erase(),!0}},bh=class extends ut{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 n=0;n<t.length;n++)if(t[n]!==n)return!1;return!0}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},wh=class extends ut{static{l(this,"TransposeTranspose")}constructor(){super("transpose_transpose",10),this.rootOpName="transpose"}match(e){return Ha(YE,e)!==null}rewrite(e,t){let n=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(n,i);return e.replaceAllResultsWith([a.getResult(0)]),e.erase(),!0}},Th=class extends ut{static{l(this,"FoldTrivialPad")}constructor(){super("fold_trivial_pad",10),this.rootOpName="pad"}match(e){let t=e.getAttr("low"),n=e.getAttr("high"),s=e.getAttr("interior")||[];return!(t&&t.some(o=>o!==0)||n&&n.some(o=>o!==0)||s&&s.some(o=>o!==0))}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},xh=class extends ut{static{l(this,"FoldTrivialSlice")}constructor(){super("fold_trivial_slice",10),this.rootOpName="slice"}match(e){let t=e.getOperand(0).type,n=e.getResult(0).type;if(!(t instanceof I)||!(n 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(n)}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},Sb=class extends ut{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,n=e.getOperand(1).definingOp;return t&&t.opName==="constant"&&!(n&&n.opName==="constant")}rewrite(e,t){let n=e.getOperand(0),s=e.getOperand(1);return e.replaceOperand(0,s),e.replaceOperand(1,n),!0}};l(kb,"commutativeConstantRightFor");Ja=class extends ut{static{l(this,"AddZero")}constructor(){super("add_zero",5),this.rootOpName="add"}match(e){return js(e.getOperand(1).definingOp,0)||js(e.getOperand(0).definingOp,0)}rewrite(e,t){let n=js(e.getOperand(1).definingOp,0)?e.getOperand(0):e.getOperand(1);return n.type.equals(e.getResult(0).type)?(e.replaceAllResultsWith([n]),e.erase(),!0):!1}},Za=class extends ut{static{l(this,"SubZero")}constructor(){super("sub_zero",5),this.rootOpName="sub"}match(e){return js(e.getOperand(1).definingOp,0)}rewrite(e,t){let n=e.getOperand(0);return n.type.equals(e.getResult(0).type)?(e.replaceAllResultsWith([n]),e.erase(),!0):!1}},Qa=class extends ut{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:nr(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}},Xa=class extends ut{static{l(this,"MulOne")}constructor(){super("mul_one",5),this.rootOpName="mul"}match(e){return js(e.getOperand(1).definingOp,1)||js(e.getOperand(0).definingOp,1)}rewrite(e,t){let n=js(e.getOperand(1).definingOp,1)?e.getOperand(0):e.getOperand(1);return n.type.equals(e.getResult(0).type)?(e.replaceAllResultsWith([n]),e.erase(),!0):!1}},eu=class extends ut{static{l(this,"MulZero")}constructor(e=!1){super("mul_zero",5),this.rootOpName="mul",this.fastMath=e}match(e){return!nr(e.getResult(0).type.dtype)&&!this.fastMath?!1:js(e.getOperand(1).definingOp,0)||js(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}},tu=class extends ut{static{l(this,"DivOne")}constructor(){super("div_one",5),this.rootOpName="div"}match(e){return js(e.getOperand(1).definingOp,1)}rewrite(e,t){let n=e.getOperand(0);return n.type.equals(e.getResult(0).type)?(e.replaceAllResultsWith([n]),e.erase(),!0):!1}},nu=class extends ut{static{l(this,"DoubleNeg")}constructor(){super("double_neg",5),this.rootOpName="neg"}match(e){return Ha(ZE,e)!==null}rewrite(e,t){let n=e.getOperand(0).definingOp.getOperand(0);return e.replaceAllResultsWith([n]),e.erase(),!0}},Sh=class extends ut{static{l(this,"ExpLog")}constructor(e=!1){super("exp_log",5),this.rootOpName="exp",this.fastMath=e}match(e){return this.fastMath&&Ha(QE,e)!==null}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0).definingOp.getOperand(0)]),e.erase(),!0}},kh=class extends ut{static{l(this,"LogExp")}constructor(e=!1){super("log_exp",5),this.rootOpName="log",this.fastMath=e}match(e){return this.fastMath&&Ha(XE,e)!==null}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0).definingOp.getOperand(0)]),e.erase(),!0}},vh=class extends ut{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}},Nh=class extends ut{static{l(this,"MulNegNeg")}constructor(){super("mul_neg_neg",4),this.rootOpName="mul"}match(e){let t=e.getOperand(0).definingOp,n=e.getOperand(1).definingOp;return t&&t.opName==="neg"&&n&&n.opName==="neg"}rewrite(e,t){let n=e.getOperand(0).definingOp.getOperand(0),s=e.getOperand(1).definingOp.getOperand(0),o=t.mul(n,s);return e.replaceAllResultsWith([o.getResult(0)]),e.erase(),!0}},Ah=class extends ut{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 n=e.getOperand(0),s=e.getOperand(1).definingOp.getOperand(0),o=t.sub(n,s);return e.replaceAllResultsWith([o.getResult(0)]),e.erase(),!0}},Oh=class extends ut{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 n=e.getOperand(0),s=e.getOperand(1).definingOp.getOperand(0),o=t.add(n,s);return e.replaceAllResultsWith([o.getResult(0)]),e.erase(),!0}},Bh=class extends ut{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 n=t.getOperand(0).type.dtype,s=e.getAttr("target_dtype");return n===s}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0).definingOp.getOperand(0)]),e.erase(),!0}},Rh=class extends ut{static{l(this,"LayoutTransformIdentity")}constructor(){super("layout_transform_identity",10),this.rootOpName="layout_transform"}match(e){let t=e.getAttr("src_layout"),n=e.getAttr("dst_layout");if(!t||!n||t.length!==n.length)return!1;for(let s=0;s<t.length;s++)if(t[s]!==n[s])return!1;return!0}rewrite(e,t){return e.replaceAllResultsWith([e.getOperand(0)]),e.erase(),!0}},Mh=class extends ut{static{l(this,"FoldTransposeIntoDot")}constructor(){super("fold_transpose_into_dot",10),this.rootOpName="dot"}match(e){for(let t=0;t<2;t++){let n=e.getOperand(t).definingOp;if(!n||n.opName!=="transpose")continue;let s=n.getAttr("permutation");if(!(!s||s.length!==2)&&!(s[0]!==1||s[1]!==0))return!0}return!1}rewrite(e,t){let n=[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=n[p].definingOp;if(!c||c.opName!=="transpose")continue;let h=c.getAttr("permutation");if(!h||h.length!==2||h[0]!==1||h[1]!==0)continue;n[p]=c.getOperand(0);let f=p===0?s:o,d=p===0?i:a;for(let m=0;m<f.length;m++)f[m]=f[m]===0?1:0;for(let m=0;m<d.length;m++)d[m]=d[m]===0?1:0}let u=t.dot(n[0],n[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}},Ih=class extends ut{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 n=e.getOperand(0).definingOp,s=n.getAttr("src_layout"),o=n.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=n.getOperand(0),p=t._inferAndBuild("layout_transform",[u],{src_layout:s,dst_layout:a});return e.replaceAllResultsWith([p.getResult(0)]),e.erase(),!0}}});function Vi(r){if(r.length!==2)return null;let e=r[0],t=r[1];if(!(e instanceof I)||!(t instanceof I)||e.dtype!==t.dtype)return null;let n=I.broadcastShape(e.shape,t.shape);return n?[new I(n,e.dtype)]:null}function Eh(r){if(r.length!==1)return null;let e=r[0];return e instanceof I?[new I(e.shape,e.dtype)]:null}function WS(r){if(r.length!==2)return null;let e=r[0],t=r[1];if(!(e instanceof I)||!(t instanceof I))return null;let n=I.broadcastShape(e.shape,t.shape);return n?[new I(n,"bool")]:null}function vb(r){if(r.length!==1)return null;let e=r[0];return!(e instanceof I)||!Lt(e.dtype)?null:[new I(e.shape,e.dtype)]}function Gi(r){let e=[];if(r.numOperands!==2)return e.push(`${r.opName} expects 2 operands, got ${r.numOperands}`),e;let t=r.getOperand(0).type,n=r.getOperand(1).type;return t instanceof I||e.push(`${r.opName} operand 0 is not tensor`),n instanceof I||e.push(`${r.opName} operand 1 is not tensor`),t instanceof I&&n instanceof I&&t.dtype!==n.dtype&&e.push(`${r.opName} dtype mismatch: ${t.dtype} vs ${n.dtype}`),e}function fc(r){let e=[];return r.numOperands!==1?(e.push(`${r.opName} expects 1 operand, got ${r.numOperands}`),e):(r.getOperand(0).type instanceof I||e.push(`${r.opName} operand is not tensor`),e)}function Nb(r){let e=fc(r);if(e.length===0){let t=r.getOperand(0).type.dtype;Lt(t)||e.push(`${r.opName} requires float type, got ${t}`)}return e}var su,Ab,Lh=M(()=>{"use strict";Oe();_n();l(Vi,"inferBinaryElementwise");l(Eh,"inferUnaryElementwise");l(WS,"inferCompare");l(vb,"inferUnaryFloat");l(Gi,"verifyBinaryElementwise");l(fc,"verifyUnaryElementwise");l(Nb,"verifyUnaryFloat");su=[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],Ab=[...su,H.COMMUTATIVE,H.ASSOCIATIVE]});function Ch(r){return e=>{if(!(typeof e[0]!="number"||typeof e[1]!="number"))return r(e[0],e[1])}}function eL(r){return e=>typeof e[0]=="number"?r(e[0]):void 0}function HS(r){r.register(new Z({name:"add",numOperands:2,numResults:1,traits:Ab,inferResultTypes:Vi,verify:Gi,getCanonicalizationPatterns(){return[kb("add"),new Ja]},fold:Ch((e,t)=>e+t)})),r.register(new Z({name:"mul",numOperands:2,numResults:1,traits:Ab,inferResultTypes:Vi,verify:Gi,getCanonicalizationPatterns(){return[kb("mul"),new Xa,new eu]},fold:Ch((e,t)=>e*t)})),r.register(new Z({name:"sub",numOperands:2,numResults:1,traits:su,inferResultTypes:Vi,verify:Gi,getCanonicalizationPatterns(){return[new Za,new Qa]},fold:Ch((e,t)=>e-t)})),r.register(new Z({name:"div",numOperands:2,numResults:1,traits:su,inferResultTypes:Vi,verify:Gi,getCanonicalizationPatterns(){return[new tu]},fold:Ch((e,t)=>e/t)}));for(let e of["rem","pow"])r.register(new Z({name:e,numOperands:2,numResults:1,traits:su,inferResultTypes:Vi,verify:Gi}));r.register(new Z({name:"neg",numOperands:1,numResults:1,traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Eh,verify:fc,getCanonicalizationPatterns(){return[new nu]},fold:eL(e=>-e)}));for(let e of["maximum","minimum"])r.register(new Z({name:e,numOperands:2,numResults:1,traits:[...su,H.COMMUTATIVE],inferResultTypes:Vi,verify:Gi}))}var YS=M(()=>{"use strict";_n();ru();Lh();l(Ch,"scalarBinaryFold");l(eL,"scalarUnaryFold");l(HS,"register")});function JS(r){for(let e of["abs","floor","ceil","round","sign","square","reciprocal"])r.register(new Z({name:e,numOperands:1,numResults:1,traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Eh,verify:fc}));r.register(new Z({name:"exp",numOperands:1,numResults:1,traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:vb,fold(e){return typeof e[0]=="number"?Math.exp(e[0]):void 0},verify:Nb}));for(let e of["log","sqrt","rsqrt","tanh","sin","cos","erf","erfc","lgamma","gamma","log2","log10","exp2"])r.register(new Z({name:e,numOperands:1,numResults:1,traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:vb,verify:Nb}))}var ZS=M(()=>{"use strict";_n();Lh();l(JS,"register")});function tL(r){if(r.length!==1)return null;let e=r[0];return!(e instanceof I)||e.dtype!=="bool"?null:[new I(e.shape,"bool")]}function nL(r){if(r.length!==2)return null;let[e,t]=r;if(!(e instanceof I)||e.dtype!=="bool"||!(t instanceof I)||t.dtype!=="bool")return null;let n=I.broadcastShape(e.shape,t.shape);return n?[new I(n,"bool")]:null}function QS(r){r.register(new Z({name:"compare",numOperands:2,numResults:1,attrs:[{name:"direction",type:"string",required:!0}],traits:[H.ELEMENTWISE],inferResultTypes:WS,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 n=e.getAttr("direction");rL.has(n)||t.push(`compare invalid direction: ${n}`)}return t}})),r.register(new Z({name:"select",numOperands:3,numResults:1,traits:[H.ELEMENTWISE],inferResultTypes(e){if(e.length!==3)return null;let[t,n,s]=e;if(!(t instanceof I)||t.dtype!=="bool"||!(n instanceof I)||!(s instanceof I)||n.dtype!==s.dtype)return null;let o=I.broadcastShape(t.shape,n.shape,s.shape);return o?[new I(o,n.dtype)]:null}})),r.register(new Z({name:"where",numOperands:3,numResults:1,traits:[H.ELEMENTWISE],inferResultTypes(e){if(e.length!==3)return null;let[t,n,s]=e;if(!(t instanceof I)||!(n instanceof I)||!(s instanceof I)||n.dtype!==s.dtype)return null;let o=I.broadcastShape(t.shape,n.shape,s.shape);return o?[new I(o,n.dtype)]:null}})),r.register(new Z({name:"logical_not",numOperands:1,numResults:1,traits:[H.ELEMENTWISE],inferResultTypes:tL}));for(let e of["logical_and","logical_or"])r.register(new Z({name:e,numOperands:2,numResults:1,traits:[H.ELEMENTWISE],inferResultTypes:nL}));r.register(new Z({name:"clamp",numOperands:3,numResults:1,traits:[H.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 rL,XS=M(()=>{"use strict";_n();Oe();Lh();l(tL,"inferUnaryBool");l(nL,"inferBinaryBool");rL=new Set(["eq","ne","lt","le","gt","ge"]);l(QS,"register")});function ek(r){r.register(new Z({name:"stop_gradient",numOperands:1,numResults:1,traits:[H.VIEW],inferResultTypes(e){return e.length!==1?null:[e[0]]},propagateSymbolicShapes(e,t){let n=t.get(e.getOperand(0));return n?[n]:null}})),r.register(new Z({name:"reverse",numOperands:1,numResults:1,traits:[H.INJECTIVE],attrs:[{name:"dimensions",type:"array",required:!0}],inferResultTypes(e){return e.length!==1?null:[e[0]]},propagateSymbolicShapes(e,t){let n=t.get(e.getOperand(0));return n?[n]:null}})),r.register(new Z({name:"broadcast_in_dim",numOperands:1,numResults:1,traits:[H.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 n=e[0];if(!(n instanceof I))return null;let s=t.get?t.get("result_shape"):t.result_shape;return s?[new I(s,n.dtype)]:null},propagateSymbolicShapes(e,t){let n=t.get(e.getOperand(0));if(!n)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 n[a]!="number"||i[s[a]]===null)&&(i[s[a]]=n[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 n=e.getAttr("broadcast_dimensions"),s=e.getAttr("result_shape"),o=e.getOperand(0).type;n.length!==o.rank&&t.push(`broadcast_dimensions length ${n.length} != input rank ${o.rank}`);for(let a=0;a<n.length;a++)n[a]<0||n[a]>=s.length?t.push(`broadcast_dimensions[${a}]=${n[a]} out of range for result rank ${s.length}`):o instanceof I&&o.shape[a]!==X&&o.shape[a]!==1&&s[n[a]]!==X&&o.shape[a]!==s[n[a]]&&t.push(`broadcast_in_dim: input dim ${a} size ${o.shape[a]} incompatible with result dim ${n[a]} size ${s[n[a]]}`);let i=new Set;for(let a of n)i.has(a)&&t.push(`broadcast_dimensions has duplicate: ${a}`),i.add(a)}return t}})),r.register(new Z({name:"reshape",numOperands:1,numResults:1,traits:[H.VIEW],attrs:[{name:"new_shape",type:"array",required:!0}],inferResultTypes(e,t){if(e.length!==1)return null;let n=e[0];if(!(n instanceof I))return null;let s=t.get?t.get("new_shape"):t.new_shape;return s?[new I(s,n.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 n=e.getOperand(0).type,s=e.getAttr("new_shape");if(n instanceof I&&n.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=n.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 n=t.get(e.getOperand(0));if(!n)return null;let s=e.getAttr("new_shape"),o=[];for(let i of s)if(i===-1){let a=n.find(u=>typeof u!="number");o.push(a||-1)}else o.push(i);return[o]},getCanonicalizationPatterns(){return[new _h,new Ya]},fold(e){return e[0]}})),r.register(new Z({name:"transpose",numOperands:1,numResults:1,traits:[H.VIEW],attrs:[{name:"permutation",type:"array",required:!0}],inferResultTypes(e,t){if(e.length!==1)return null;let n=e[0];if(!(n instanceof I))return null;let s=t.get?t.get("permutation"):t.permutation;if(!s)return null;let o=s.map(i=>n.shape[i]);return[new I(o,n.dtype)]},propagateSymbolicShapes(e,t){let n=t.get(e.getOperand(0));if(!n)return null;let s=e.getAttr("permutation");return s?[s.map(o=>n[o])]:null},getCanonicalizationPatterns(){return[new bh]},fold(e,t,n){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=n&&n[0]?n[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]),h=s.length,f=new Array(h),d=new Array(o.length).fill(0);for(let m=0;m<h;m++){let g=0;for(let y=0;y<o.length;y++)g+=d[y]*u[o[y]];f[m]=s[g];for(let y=o.length-1;y>=0&&!(++d[y]<c[y]);y--)d[y]=0}return f},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 n=e.getAttr("permutation"),s=e.getOperand(0).type;s instanceof I&&n.length!==s.rank&&t.push(`transpose permutation length ${n.length} != input rank ${s.rank}`);let o=new Set;for(let i of n)o.has(i)&&t.push(`transpose duplicate in permutation: ${i}`),o.add(i);return t}})),r.register(new Z({name:"slice",numOperands:1,numResults:1,traits:[H.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 n=e[0];if(!(n 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,n.dtype)]},getCanonicalizationPatterns(){return[new xh]}})),r.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 n=e[0];if(!(n instanceof I))return null;let s=t.get?t.get("dimension"):t.dimension;if(s===void 0)return null;let o=[...n.shape];for(let i=1;i<e.length;i++){let a=e[i];if(!(a instanceof I)||a.dtype!==n.dtype||a.rank!==n.rank)return null;o[s]===X||a.shape[s]===X?o[s]=X:o[s]+=a.shape[s]}return[new I(o,n.dtype)]}})),r.register(new Z({name:"pad",numOperands:2,numResults:1,traits:[H.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 n=e[0];if(!(n 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<n.rank;u++)n.shape[u]===X?a.push(X):a.push(s[u]+n.shape[u]+(n.shape[u]-1)*i[u]+o[u]);return[new I(a,n.dtype)]},getCanonicalizationPatterns(){return[new Th]}})),r.register(new Z({name:"gather",numOperands:2,numResults:1,traits:[H.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 n=e[0],s=e[1];if(!(n 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 h=new Set(o),f=[],d=0,m=0,g=p.length+c.length;for(let y=0;y<g;y++)h.has(y)?f.push(c[m++]):f.push(p[d++]);return[new I(f,n.dtype)]}})),r.register(new Z({name:"scatter",numOperands:3,numResults:1,traits:[H.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}})),r.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 n=e[0];if(!(n 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=[...n.shape];return a[s]=i,new I(a,n.dtype)})}}))}var tk=M(()=>{"use strict";_n();Oe();ru();l(ek,"register")});function nk(r){r.register(new Z({name:"reduce",numOperands:2,numResults:1,attrs:[{name:"dimensions",type:"array",required:!0},{name:"reduce_type",type:"string",required:!0}],traits:[H.REDUCTION],hasRegions:!0,numRegions:1,inferResultTypes(t,n){if(t.length<1)return null;let s=t[0];if(!(s instanceof I))return null;let o=n.get?n.get("dimensions"):n.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,n){let s=n.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 n=[];if(t.hasAttr("dimensions")||n.push("reduce missing dimensions"),!t.hasAttr("reduce_type"))n.push("reduce missing reduce_type");else{let s=t.getAttr("reduce_type");sL.has(s)||n.push(`reduce invalid reduce_type: ${s}`)}return n}}));function e(t,n){if(t.length<1)return null;let s=t[0];if(!(s instanceof I))return null;let o=n.get?n.get("axis"):n.axis;if(o===void 0)return null;let i=(n.get?n.get("keep_dims"):n.keep_dims)||!1,a=[];for(let u=0;u<s.rank;u++)u===o?i&&a.push(1):a.push(s.shape[u]);return[new I(a,"i32")]}l(e,"inferArgReduceTypes"),r.register(new Z({name:"argmax",numOperands:1,numResults:1,attrs:[{name:"axis",type:"number",required:!0},{name:"keep_dims",type:"boolean",required:!1}],traits:[H.REDUCTION],inferResultTypes:e})),r.register(new Z({name:"argmin",numOperands:1,numResults:1,attrs:[{name:"axis",type:"number",required:!0},{name:"keep_dims",type:"boolean",required:!1}],traits:[H.REDUCTION],inferResultTypes:e}))}var sL,rk=M(()=>{"use strict";_n();Oe();sL=new Set(["sum","max","min","prod","mean","and","or"]);l(nk,"register")});function oL(r,e){let t=r.getAttr("scale"),n=e.getAttr("scale"),s=r.getAttr("zero_point"),o=e.getAttr("zero_point"),i=r.getAttr("scheme"),a=e.getAttr("scheme");if(i!==a)return!1;if(typeof t=="number"&&typeof n=="number"){if(t!==n)return!1}else return!1;if(typeof s=="number"&&typeof o=="number"){if(s!==o)return!1}else return!1;return!0}var Dh,Ph,$h,Ob=M(()=>{"use strict";uc();Oe();l(oL,"quantParamsMatch");Dh=class extends ut{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:oL(e,t)}rewrite(e,t){let n=e.getOperand(0).definingOp.getOperand(0);if(!n.type||!e.getResult(0).type)return!1;let s=n.type,o=e.getResult(0).type;return s instanceof I&&o instanceof I&&s.dtype===o.dtype?(e.replaceAllResultsWith([n]),e.erase(),!0):!1}},Ph=class extends ut{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=Gs(a)*8,p=a==="ui8",c=p?0:-(1<<u-1),h=p?(1<<u)-1:(1<<u-1)-1,f=Math.max(c,Math.min(h,Math.round(s/o+i))),d=e.getResult(0).type,m=t.constant(f,d);return e.replaceAllResultsWith([m.getResult(0)]),e.erase(),!0}},$h=class extends ut{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,n=e.getOperand(1).definingOp;return t&&t.opName==="dequantize"&&n&&n.opName==="dequantize"}rewrite(e,t){let n=e.getOperand(0).definingOp,s=e.getOperand(1).definingOp,o=n.getOperand(0),i=s.getOperand(0),a=n.getAttr("scale"),u=s.getAttr("scale"),p=n.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 h=a*u,f=e.getAttr("lhs_contracting"),d=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:f,rhs_contracting:d,lhs_batch:m,rhs_batch:g,lhs_scale:a,lhs_zero_point:p,rhs_scale:u,rhs_zero_point:c,output_scale:h,output_zero_point:0}),b=t._inferAndBuild("dequantize",[_.getResult(0)],{scale:h,zero_point:0,scheme:n.getAttr("scheme"),target_dtype:e.getResult(0).type.dtype});return e.replaceAllResultsWith([b.getResult(0)]),e.erase(),!0}}});function ok(r){r.register(new Z({name:"dot",numOperands:2,numResults:1,opAttrs:{gpuCapable:!0},traits:[H.OPAQUE,H.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,n=e.getOperand(1).type;if(!(t instanceof I)||!(n 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:Fh,getCanonicalizationPatterns(){return[new Mh,new $h]},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 n=e.getOperand(0).type,s=e.getOperand(1).type;if(n instanceof I&&s instanceof I){n.dtype!==s.dtype&&t.push(`dot dtype mismatch: ${n.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 ${i.length}`);else for(let a=0;a<o.length;a++){let u=o[a],p=i[a];u>=n.rank&&t.push(`dot lhs_contracting[${a}]=${u} out of range (rank ${n.rank})`),p>=s.rank&&t.push(`dot rhs_contracting[${a}]=${p} out of range (rank ${s.rank})`);let c=n.shape[u],h=s.shape[p];u<n.rank&&p<s.rank&&typeof c=="number"&&c!==X&&typeof h=="number"&&h!==X&&c!==h&&t.push(`dot contracting dim size mismatch at [${a}]: lhs dim ${u} size ${c} vs rhs dim ${p} size ${h}`)}}return t}})),r.register(new Z({name:"cublas_gemm",numOperands:2,numResults:1,traits:[H.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:Fh,getFlops(e){let t=e.getOperand(0).type,n=e.getResult(0).type;if(!(t instanceof I)||!(n 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=n.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)}})),r.register(new Z({name:"conv",numOperands:2,numResults:1,opAttrs:{gpuCapable:!0},traits:[H.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,n=e.getOperand(1).type;if(!(t instanceof I)||!(n instanceof I))return 0;let s=t.numel(),o=n.numel();return s===X||o===X?0:2*s*o/(n.shape[0]||1)},inferResultTypes:Bb,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 Fh(r,e,t){if(r.length!==2)return null;let n=r[0],s=r[1];if(!(n instanceof I)||!(s instanceof I))return null;let o=t&&!Array.isArray(t)?t:{};if(!o.allowMixedDtype&&n.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 h=0;h<n.rank;h++)u.has(h)&&c.push(n.shape[h]);for(let h=0;h<n.rank;h++)!i.has(h)&&!u.has(h)&&c.push(n.shape[h]);for(let h=0;h<s.rank;h++)!a.has(h)&&!p.has(h)&&c.push(s.shape[h]);return[new I(c,o.outputDtype||n.dtype)]}function Bb(r,e,t){if(r.length!==2)return null;let n=r[0],s=r[1];if(!(n instanceof I)||!(s instanceof I))return null;let o=t&&!Array.isArray(t)?t:{};if(!o.allowMixedDtype&&n.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=n.shape[0],h=s.shape[0],f=[];for(let d=0;d<p;d++){let m=n.shape[d+2],g=s.shape[d+2],y=a[d][0]+a[d][1];if(m===X||g===X)f.push(X);else{let _=(g-1)*u[d]+1;f.push(Math.floor((m+y-_)/i[d])+1)}}return[new I([c,h,...f],o.outputDtype||n.dtype)]}var Rb=M(()=>{"use strict";_n();Oe();ru();Ob();l(ok,"register");l(Fh,"inferDotResultTypes");l(Bb,"inferConvResultTypes")});function ik(r){let e={numOperands:0,numResults:1,attrs:[{name:"value",type:"any",required:!0},{name:"tensor_type",type:"object",required:!0}],traits:[H.CONSTANT],inferResultTypes(t,n){let s=n.get?n.get("tensor_type"):n.tensor_type;return s?[s]:null}};for(let t of["constant","scalar_constant"])r.register(new Z({name:t,...e}));r.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,n){let s=n.get?n.get("tensor_type"):n.tensor_type;return s?[s]:null}})),r.register(new Z({name:"tuple",numOperands:-1,numResults:1,inferResultTypes(t){return[new Fi(t)]}})),r.register(new Z({name:"get_tuple_element",numOperands:1,numResults:1,attrs:[{name:"index",type:"number",required:!0}],inferResultTypes(t,n){if(t.length!==1)return null;let s=t[0];if(!s||!s.elements)return null;let o=n.get?n.get("index"):n.index;return o===void 0||o<0||o>=s.elements.length?null:[s.elements[o]]}})),r.register(new Z({name:"convert",numOperands:1,numResults:1,attrs:[{name:"target_dtype",type:"string",required:!0}],traits:[H.ELEMENTWISE],inferResultTypes(t,n){if(t.length!==1)return null;let s=t[0];if(!(s instanceof I))return null;let o=n.get?n.get("target_dtype"):n.target_dtype;return[new I(s.shape,o)]}})),r.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,n){if(t.length<1)return null;let s=t[0];if(!(s instanceof I))return null;let o=n.get?n.get("depth"):n.depth,i=(n.get?n.get("axis"):n.axis)??-1,a=(n.get?n.get("dtype"):n.dtype)||"f32",u=[...s.shape],p=i<0?u.length+1+i:i;return u.splice(p,0,o),[new I(u,a)]}}))}var ak=M(()=>{"use strict";_n();Oe();l(ik,"register")});function uk(r){r.register(new Z({name:"return",numOperands:-1,numResults:0,traits:[H.TERMINATOR]})),r.register(new Z({name:"yield",numOperands:-1,numResults:0,traits:[H.TERMINATOR]})),r.register(new Z({name:"if",numOperands:1,numResults:-1,hasRegions:!0,numRegions:2,sideEffects:an.CONTROL,inferResultTypes(e,t,n){return n||null}})),r.register(new Z({name:"while",numOperands:-1,numResults:-1,hasRegions:!0,numRegions:2,sideEffects:an.CONTROL,inferResultTypes(e){return[...e]}})),r.register(new Z({name:"scan",numOperands:-1,numResults:-1,hasRegions:!0,numRegions:1,sideEffects:an.CONTROL,attrs:[{name:"num_carry",type:"number",required:!0},{name:"num_xs",type:"number",required:!0}],inferResultTypes(e,t,n){return n||null}})),r.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:an.WRITE,inferResultTypes(e,t,n){return n||null}})),r.register(new Z({name:"fused_dot_epilogue",numOperands:-1,numResults:1,traits:[H.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,n){return n||null}})),r.register(new Z({name:"fusion",numOperands:-1,numResults:-1,hasRegions:!0,numRegions:1,attrs:[{name:"fusion_kind",type:"string",required:!1}],inferResultTypes(e,t,n){return n||null}}))}var lk=M(()=>{"use strict";_n();l(uk,"register")});function ck(r){r.register(new Z({name:"layout_transform",numOperands:1,numResults:1,traits:[H.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 n=e[0];if(!(n instanceof I))return null;let s=t.get?t.get("dst_layout"):t.dst_layout;return s?[new I(n.shape,n.dtype,new Et(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 n=e.getOperand(0).type,s=e.getAttr("src_layout"),o=e.getAttr("dst_layout");n instanceof I&&(s.length!==n.rank&&t.push(`src_layout length ${s.length} != input rank ${n.rank}`),o.length!==n.rank&&t.push(`dst_layout length ${o.length} != input rank ${n.rank}`))}return t},getCanonicalizationPatterns(){return[new Rh,new Ih]}}))}var pk=M(()=>{"use strict";_n();Oe();ru();l(ck,"register")});function fk(r){r.register(new Z({name:"quantize",numOperands:1,numResults:1,traits:[H.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 n=e[0];if(!(n instanceof I))return null;let s=t.get?t.get("target_dtype"):t.target_dtype;return s?[new I(n.shape,s)]:null},verify(e){let t=[];if(e.numOperands!==1)return t.push("quantize expects 1 operand"),t;let n=e.getOperand(0).type;n instanceof I&&!Lt(n.dtype)&&t.push(`quantize input must be float, got ${n.dtype}`);let s=e.getAttr("target_dtype");s&&!Mb.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 n=t.get?t.get("scale"):t.scale,s=t.get?t.get("zero_point"):t.zero_point;if(!(typeof n!="number"||typeof s!="number"))return Math.round(e[0]/n+s)},getCanonicalizationPatterns(){return[new Dh,new Ph]}})),r.register(new Z({name:"dequantize",numOperands:1,numResults:1,traits:[H.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 n=e[0];if(!(n instanceof I))return null;let s=t.get?t.get("target_dtype"):t.target_dtype;return s?[new I(n.shape,s)]:null},verify(e){let t=[];if(e.numOperands!==1)return t.push("dequantize expects 1 operand"),t;let n=e.getOperand(0).type;n instanceof I&&!iL.has(n.dtype)&&t.push("dequantize input must be i8, ui8, or i32, got "+n.dtype);let s=e.getAttr("target_dtype");return s&&!Lt(s)&&t.push(`dequantize target_dtype must be float, got ${s}`),t},fold(e,t){if(typeof e[0]!="number")return;let n=t.get?t.get("scale"):t.scale,s=t.get?t.get("zero_point"):t.zero_point;if(!(typeof n!="number"||typeof s!="number"))return(e[0]-s)*n}})),r.register(new Z({name:"quantized_dot",numOperands:2,numResults:1,traits:[H.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,n=e.getOperand(1).type;if(!(t instanceof I)||!(n 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 Fh(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 n=e.getOperand(0).type,s=e.getOperand(1).type;return n instanceof I&&!Mb.has(n.dtype)&&t.push(`quantized_dot lhs must be i8/ui8, got ${n.dtype}`),s instanceof I&&!Mb.has(s.dtype)&&t.push(`quantized_dot rhs must be i8/ui8, got ${s.dtype}`),t}})),r.register(new Z({name:"quantized_conv",numOperands:2,numResults:1,traits:[H.OPAQUE],attrs:[{name:"strides",type:"array",required:!0},{name:"padding",type:"array",required:!0},{name:"dilation",type:"array",required:!1},{name:"groups",type:"number",required:!1},{name:"input_layout",type:"string",required:!0},{name:"kernel_layout",type:"string",required:!0},{name:"input_scale",type:"number",required:!0},{name:"input_zero_point",type:"number",required:!0},{name:"kernel_scale",type:"number",required:!0},{name:"kernel_zero_point",type:"number",required:!0},{name:"output_scale",type:"number",required:!0},{name:"output_zero_point",type:"number",required:!0}],inferResultTypes(e,t){return Bb(e,t,{outputDtype:"i32",allowMixedDtype:!0})},verify(e){let t=[];return e.numOperands!==2?(t.push("quantized_conv expects 2 operands"),t):(e.hasAttr("strides")||t.push("quantized_conv missing strides"),e.hasAttr("padding")||t.push("quantized_conv missing padding"),e.hasAttr("input_layout")||t.push("quantized_conv missing input_layout"),e.hasAttr("kernel_layout")||t.push("quantized_conv missing kernel_layout"),e.hasAttr("input_scale")||t.push("quantized_conv missing input_scale"),e.hasAttr("kernel_scale")||t.push("quantized_conv missing kernel_scale"),t)}}))}var Mb,iL,hk=M(()=>{"use strict";_n();Oe();Ob();Rb();Mb=new Set(["i8","ui8"]),iL=new Set(["i8","ui8","i32"]);l(fk,"register")});function aL(r){if(r.length<1)return null;let e=r[0];return e instanceof I?[new I(e.shape,e.dtype)]:null}function Ts(r){if(r.length<1)return null;let e=r[0];return!(e instanceof I)||!Lt(e.dtype)?null:[new I(e.shape,e.dtype)]}function bo(r){let e=[];if(r.numOperands<1)return e.push(`${r.opName} expects at least 1 operand`),e;let t=r.getOperand(0).type;return t instanceof I&&!Lt(t.dtype)&&e.push(`${r.opName} requires float input, got ${t.dtype}`),e}function dk(r){r.register(new Z({name:"all_reduce",numOperands:1,numResults:1,traits:[H.OPAQUE],attrs:[{name:"reduce_op",type:"string",required:!1},{name:"mesh_axis",type:"number",required:!1}],inferResultTypes:aL})),r.register(new Z({name:"all_gather",numOperands:1,numResults:1,traits:[H.OPAQUE],attrs:[{name:"mesh_axis",type:"number",required:!1},{name:"gather_dim",type:"number",required:!1}],inferResultTypes(e,t){let n=e[0];if(!(n 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=[...n.shape];return i[o]=i[o]*i[s],[new I(i,n.dtype)]}})),r.register(new Z({name:"scaled_dot_product_attention",numOperands:3,numResults:1,traits:[H.OPAQUE],attrs:[{name:"scale",type:"number",required:!0},{name:"causal",type:"boolean",required:!1}],inferResultTypes(e){let t=e[0],n=e[2];return!(t instanceof I)||!(n instanceof I)?null:[new I([...t.shape.slice(0,t.rank-1),n.shape[n.rank-1]],t.dtype)]}})),r.register(new Z({name:"softmax",numOperands:1,numResults:1,opAttrs:{gpuCapable:!0},attrs:[{name:"axis",type:"number",required:!0}],inferResultTypes:Ts,verify:bo})),r.register(new Z({name:"log_softmax",numOperands:1,numResults:1,attrs:[{name:"axis",type:"number",required:!0}],inferResultTypes:Ts,verify:bo})),r.register(new Z({name:"gelu",numOperands:1,numResults:1,traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Ts,verify:bo})),r.register(new Z({name:"sigmoid",numOperands:1,numResults:1,traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Ts,verify:bo})),r.register(new Z({name:"silu",numOperands:1,numResults:1,traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Ts,verify:bo})),r.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:Ts,verify(e){let t=[];if(e.numOperands!==3)return t.push("layer_norm expects 3 operands (input, gamma, beta)"),t;let n=e.getOperand(0).type;return n instanceof I&&!Lt(n.dtype)&&t.push(`layer_norm requires float input, got ${n.dtype}`),t}})),r.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:Ts,verify(e){let t=[];if(e.numOperands!==5)return t.push("batch_norm expects 5 operands (input, gamma, beta, mean, var)"),t;let n=e.getOperand(0).type;return n instanceof I&&!Lt(n.dtype)&&t.push(`batch_norm requires float input, got ${n.dtype}`),t}}));for(let e of["selu","mish","hardswish","hardsigmoid"])r.register(new Z({name:e,numOperands:1,numResults:1,traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Ts,verify:bo}));r.register(new Z({name:"elu",numOperands:1,numResults:1,attrs:[{name:"alpha",type:"number"}],traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Ts,verify:bo})),r.register(new Z({name:"leaky_relu",numOperands:1,numResults:1,attrs:[{name:"negative_slope",type:"number"}],traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Ts,verify:bo})),r.register(new Z({name:"celu",numOperands:1,numResults:1,attrs:[{name:"alpha",type:"number"}],traits:[H.ELEMENTWISE,H.SAME_OPERAND_AND_RESULT_TYPE],inferResultTypes:Ts,verify:bo})),r.register(new Z({name:"embedding",numOperands:2,numResults:1,inferResultTypes(e){if(e.length<2)return null;let t=e[0],n=e[1];if(!(t instanceof I)||!(n instanceof I))return null;let s=[...n.shape,t.shape[t.rank-1]];return[new I(s,t.dtype)]}}))}var mk=M(()=>{"use strict";_n();Oe();l(aL,"inferSameAsInput");l(Ts,"inferSameAsInputFloat");l(bo,"verifyUnaryFloat");l(dk,"register")});function gk(r){r.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:an.READ|an.WRITE,traits:[H.INJECTIVE],inferResultTypes(e){return e.length!==1?null:[e[0]]},getFlops(){return 0}}))}var yk=M(()=>{"use strict";_n();l(gk,"register")});function _k(r,e,t,n,s,o){let i=r+n+s-e;return(o?Math.ceil(i/t):Math.floor(i/t))+1}function bk(r){r.register(new Z({name:"pool2d",numOperands:1,numResults:1,traits:[H.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 n=e[0];if(!(n instanceof I)||n.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=_k(n.shape[2],s[0],o[0],i[0][0],i[0][1],a),p=_k(n.shape[3],s[1],o[1],i[1][0],i[1][1],a);return[new I([n.shape[0],n.shape[1],u,p],n.dtype)]}}))}var wk=M(()=>{"use strict";_n();Oe();l(_k,"computePoolOutputDim");l(bk,"register")});function Tk(r){r.register(new Z({name:"resize",numOperands:1,numResults:1,traits:[H.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 n=e[0];if(!(n instanceof I)||n.rank!==4)return null;let s=t.get?t.get("output_size"):t.output_size;return[new I([n.shape[0],n.shape[1],s[0],s[1]],n.dtype)]}}))}var xk=M(()=>{"use strict";_n();Oe();l(Tk,"register")});function lL(r=uL){let e=new ph;for(let t of r)t(e);return e}var uL,ne,At=M(()=>{"use strict";_n();YS();ZS();XS();tk();rk();Rb();ak();lk();pk();hk();mk();yk();wk();xk();uL=[HS,JS,QS,ek,nk,ok,ik,uk,ck,fk,dk,gk,bk,Tk];l(lL,"buildRegistry");ne=lL()});function wo(r,e,t="ignore"){let n=[],s=new Map;for(let o of r){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),n.push(u),i.pop()}}return n}function zh(r,e="throw"){let t=Array.isArray(r)?r:[...r],n=new Set(t);return wo(t,s=>n.has(s),e)}var ji=M(()=>{"use strict";l(wo,"topoSortByOperands");l(zh,"topoSortOpSet")});function Sk(r){return Array.isArray(r)?r.map(Sk):r}function kk(r){if(typeof r=="number")return r*2654435761&2147483647;if(typeof r=="string"){let e=0;for(let t=0;t<r.length;t++)e=(e<<5)-e+r.charCodeAt(t)&2147483647;return e}if(typeof r=="boolean")return r?1:0;if(Array.isArray(r)){let e=2654435769;for(let t=0;t<r.length;t++)e=(e^kk(r[t]))*16777619&2147483647;return e}if(ArrayBuffer.isView(r)&&r.buffer instanceof ArrayBuffer){let e=new Uint8Array(r.buffer,r.byteOffset,r.byteLength),t=2654435769;t=(t^e.length)*16777619&2147483647;let n=e.length>256?Math.ceil(e.length/256):1;for(let s=0;s<e.length;s+=n)t=(t^e[s])*16777619&2147483647;return t}return typeof r=="object"&&r!==null&&typeof r.hash=="function"?r.hash():0}function vk(r,e){if(r===e)return!0;if(typeof r!=typeof e)return!1;if(Array.isArray(r)){if(!Array.isArray(e)||r.length!==e.length)return!1;for(let t=0;t<r.length;t++)if(!vk(r[t],e[t]))return!1;return!0}return typeof r=="object"&&r!==null&&typeof r.equals=="function"?r.equals(e):!1}function pL(r,e=new Map){let t=new Ct;for(let n of r.blocks){let s=n.arguments.map(p=>p.type),o=new n.constructor(s);for(let p=0;p<n.arguments.length;p++)e.set(n.arguments[p],o.arguments[p]);let i=n.opsArray(),a=new Set(i),u=new Map;for(let p of wo(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 cL,Re,sr=M(()=>{"use strict";ch();zi();At();ji();cL=0,Re=class r{static{l(this,"Operation")}constructor(e,t,n,s=null,o=null){if(this.opName=e,this.id=cL++,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 ic(this,i);t[i].addUse(a),this._operandLinks[i]=a}this.results=new Array(n.length);for(let i=0;i<n.length;i++)this.results[i]=new Wo(n[i],this,i);if(this.regions=[],o)for(let i=0;i<o.length;i++){let a=o[i]instanceof Ct?o[i]:new Ct;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}' (${this.operands.length} operands)`);if(!(t instanceof Wo))throw new Error(`replaceOperand: new operand for '${this.opName}' is not a Value`);let n=this.operands[e];if(n===t)return;n.removeUse(this._operandLinks[e]),this.operands[e]=t;let s=new ic(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),n=this.regions.map(i=>pL(i,e)),s=new Map;for(let[i,a]of this.attributes)s.set(i,Sk(a));let o=new r(this.opName,t,this.results.map(i=>i.type),s,n);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,n]of this.attributes){for(let s=0;s<t.length;s++)e=(e^t.charCodeAt(s))*16777619&2147483647;e=(e^kk(n))*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,n]of this.attributes)if(!e.attributes.has(t)||!vk(n,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(Sk,"cloneAttrValue");l(kk,"hashAttrValue");l(vk,"attrValueEquals");l(pL,"cloneRegion")});function*Nk(r){if(!(!r.regions||r.regions.length===0))for(let e of r.regions)for(let t of e.blocks)for(let n of t.ops())yield n,yield*Nk(n)}function*Ak(r){if(!(!r.regions||r.regions.length===0))for(let e of r.regions)for(let t of e.blocks){yield t;for(let n of t.ops())yield*Ak(n)}}function fL(r){let e=r.opsArray(),t=new Set(e);return wo(e,n=>t.has(n),"ignore")}function Vh(r){let e=new un(r.name,r.inputTypes,r.outputTypes),t=new Map,n=r.entryBlock,s=e.entryBlock;for(let i=0;i<n.arguments.length;i++)t.set(n.arguments[i],s.arguments[i]);let o=new Map;for(let i of fL(n))o.set(i,i.clone(t));for(let i of n)s.pushOp(o.get(i));return e._version=r._version,e}var un,xs=M(()=>{"use strict";zi();ji();l(Nk,"opsInRegions");l(Ak,"blocksInRegions");un=class{static{l(this,"GraphFunction")}constructor(e,t,n){this.name=e,this.inputTypes=Object.freeze([...t]),this.outputTypes=Object.freeze([...n]),this.body=new Ct;let s=new yn(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*Nk(e)}*blocksRecursive(){for(let e of this.body){yield e;for(let t of e)yield*Ak(t)}}opsArray(){let e=[];for(let t of this.body)for(let n of t)e.push(n);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 n of this.ops())e(n)&&t.push(n);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 ${this.outputTypes.length} outputs`):e.push("Function body has no return op"),e}};l(fL,"topoOrderTopLevel");l(Vh,"cloneGraphFunction")});function Ok(r){let e=new wn(r.name);for(let t of r)e.addFunction(Vh(t));return e._version=r._version,e}var wn,Yo=M(()=>{"use strict";xs();wn=class{static{l(this,"GraphModule")}constructor(e="module"){this.name=e,this._functions=new Map,this._version=0}get version(){return this._version}addFunction(e){return this._functions.set(e.name,e),e._module=this,this._version++,e}getFunction(e){return this._functions.get(e)||null}hasFunction(e){return this._functions.has(e)}removeFunction(e){let t=this._functions.delete(e);return t&&this._version++,t}get functionCount(){return this._functions.size}functionNames(){return[...this._functions.keys()]}*functions(){yield*this._functions.values()}*[Symbol.iterator](){yield*this._functions.values()}verify(){let e=[];this._functions.size===0&&e.push("Module has no functions");for(let t of this._functions.values()){let n=t.verify();for(let s=0;s<n.length;s++)e.push(`${t.name}: ${n[s]}`)}return e}};l(Ok,"cloneGraphModule")});function Bk(r){return r?r instanceof Fi?`tuple(${r.types.map(Bk).join(", ")})`:r.shape!==void 0&&r.dtype!==void 0?`[${r.shape.join(",")}]:${r.dtype}`:String(r):"?"}function Rk(r,e,t){let n=r.rank,s=e<0?n+e:e,o=r.shape.map((a,u)=>u===s?1:a),i=[];for(let a=0;a<n;a++)a!==s&&i.push(a<s?a:a-1+t);return{offsetDims:i,collapsedSliceDims:[s],startIndexMap:[s],indexVectorDim:t,sliceSizes:o}}function hL(r,e){let t=Math.max(r.length,e.length),n=new Array(t);for(let s=0;s<t;s++){let o=s<r.length?r[r.length-1-s]:1,i=s<e.length?e[e.length-1-s]:1;n[t-1-s]=o===1?i:o}return n}function To(r,e){let t=[];for(let n=0;n<r;n++)n!==e&&t.push(n);return t}function dL(r){if(!r)return new Map;if(r instanceof Map)return r;let e=new Map;for(let t of Object.keys(r))e.set(t,r[t]);return e}function hc(r,e,t,n){let s=new un(r,e,t),o=new An(s);return n(o,s.args),s}var An,es=M(()=>{"use strict";Oe();sr();zi();xs();Yo();At();l(Bk,"describeType");l(Rk,"indexSelectGatherOpts");l(hL,"bcastBatchDims");An=class r{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,n,s=null,o=null){let i=new Re(e,t,n,s,o);return this._insert(i)}_inferAndBuild(e,t,n=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=n instanceof Map?n:dL(n);a=i.inferResultTypes(u,p,o)}if(!a){let u=t.map(c=>Bk(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,n,s)}constant(e,t){return this._buildOp("constant",[],[t],{value:e,tensor_type:t})}scalarConstant(e,t="f32"){let n=new I([],t);return this.constant(e,n)}tensorConstant(e,t,n="f32"){let s=new I(t,n);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,n,s,o=!1){return this._inferAndBuild("scaled_dot_product_attention",[e,t,n],{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,n){return this._inferAndBuild("compare",[e,t],{direction:n})}select(e,t,n){return this._inferAndBuild("select",[e,t,n])}clamp(e,t,n){return this._inferAndBuild("clamp",[e,t,n])}broadcast(e,t,n){return this._inferAndBuild("broadcast_in_dim",[e],{result_shape:t,broadcast_dimensions:n})}reshape(e,t){return this._inferAndBuild("reshape",[e],{new_shape:t})}transpose(e,t){return this._inferAndBuild("transpose",[e],{permutation:t})}slice(e,t,n,s=null){let o={starts:t,limits:n};return s&&(o.strides=s),this._inferAndBuild("slice",[e],o)}concat(e,t){return this._inferAndBuild("concat",e,{dimension:t})}pad(e,t,n,s,o=null){let i={low:n,high:s};return o&&(i.interior=o),this._inferAndBuild("pad",[e,t],i)}reduce(e,t,n,s){let o=new I([],e.type.dtype),i=new Ct,a=new yn([o,o]);return i.addBlock(a),this._inferAndBuild("reduce",[e,t],{dimensions:n,reduce_type:s},[i])}dot(e,t,n,s,o=[],i=[]){return this._inferAndBuild("dot",[e,t],{lhs_contracting:n,rhs_contracting:s,lhs_batch:o,rhs_batch:i})}matmul(e,t){let n=e.type.rank,s=t.type.rank;if(n===1&&s===1)return this.dot(e,t,[0],[0]);let o=e,i=t,a=!1,u=!1;n===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,h=i.type.shape.slice(0,c-2),f;if(h.length===0)f=this.dot(o,i,[p-1],[0]);else{let g=o.type.shape.slice(0,p-2),y=hL(g,h),_=y.length,b=this._broadcastBatch(o,g,y),w=this._broadcastBatch(i,h,y),T=Array.from({length:_},(S,x)=>x);f=this.dot(b,w,[_+1],[_],T,T)}if(!a&&!u)return f;let d=f.getResult(0).type.shape,m=new Set;return u&&m.add(d.length-1),a&&m.add(d.length-2),this.reshape(f.getResult(0),d.filter((g,y)=>!m.has(y)))}_broadcastBatch(e,t,n){let s=e.type.shape.slice(e.type.rank-2),o=[...n,...s];if(t.length===n.length&&t.every((u,p)=>u===n[p]))return e;let i=n.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,n,s,o={}){return this._inferAndBuild("conv",[e,t],{strides:n,padding:s,dilation:o.dilation||n.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,n,s=null){let o={call_target_name:e};return s&&(o.backend_config=s),this._buildOp("custom_call",t,n,o)}fusion(e,t,n,s){let o=new Ct,i=new yn(e.map(u=>u.type));o.addBlock(i);let a=this._buildOp("fusion",e,t,{fusion_kind:n},[o]);if(s){let u=new r(this.func);u.block=i,s(u,i.arguments)}return a}ifOp(e,t,n,s){let o=new Ct,i=new yn([]);o.addBlock(i);let a=new Ct,u=new yn([]);a.addBlock(u);let p=this._buildOp("if",[e],t,null,[o,a]);if(n){let c=new r(this.func);c.block=i,n(c)}if(s){let c=new r(this.func);c.block=u,s(c)}return p}whileOp(e,t,n){let s=e.map(c=>c.type),o=new Ct,i=new yn(s);o.addBlock(i);let a=new Ct,u=new yn(s);a.addBlock(u);let p=this._buildOp("while",e,s,null,[o,a]);if(t){let c=new r(this.func);c.block=i,t(c,i.arguments)}if(n){let c=new r(this.func);c.block=u,n(c,u.arguments)}return p}scanOp(e,t,n){let s=e.map(g=>g.type.withShape(g.type.shape.slice(1))),o=t.map(g=>g.type),i=new Ct,a=new yn([...s,...o]);i.addBlock(a);let u=new r(this.func);u.block=a;let p=a.arguments.slice(0,s.length),c=a.arguments.slice(s.length),[h,f]=n(u,p,c);if(u.yieldOp([...h,...f]),e.length===0)throw new Error("scanOp requires at least one xs input");let d=e[0].type.shape[0];if(typeof d!="number"||d<0)throw new Error(`scanOp requires a static, non-negative leading dim on xs, got ${d}`);for(let g=1;g<e.length;g++)if(e[g].type.shape[0]!==d)throw new Error("scanOp requires all xs inputs to share the same leading length");let m=f.map(g=>g.type.withShape([d,...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),n=this.broadcast(t.getResult(0),e.type.shape,[]);return this.maximum(e,n.getResult(0))}softmax(e,t=-1){let n=t<0?e.type.rank+t:t;return this._inferAndBuild("softmax",[e],{axis:n})}logSoftmax(e,t=-1){let n=t<0?e.type.rank+t:t;return this._inferAndBuild("log_softmax",[e],{axis:n})}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,n,s=-1,o=1e-5){let i=s<0?e.type.rank+s:s;return this._inferAndBuild("layer_norm",[e,t,n],{axis:i,epsilon:o})}batchnorm(e,t,n,s,o,i=1,a=1e-5){return this._inferAndBuild("batch_norm",[e,t,n,s,o],{axis:i,epsilon:a})}where(e,t,n){return this._inferAndBuild("where",[e,t,n])}split(e,t,n){return this._inferAndBuild("split",[e],{dimension:t,split_sizes:n})}oneHot(e,t,n={}){return this._inferAndBuild("one_hot",[e],{depth:t,axis:n.axis??-1,on_value:n.onValue??1,off_value:n.offValue??0,dtype:n.dtype||e.type.dtype})}embedding(e,t){return this._inferAndBuild("embedding",[e,t])}gather(e,t,n){return this._inferAndBuild("gather",[e,t],{offset_dims:n.offsetDims,collapsed_slice_dims:n.collapsedSliceDims,start_index_map:n.startIndexMap,index_vector_dim:n.indexVectorDim,slice_sizes:n.sliceSizes})}scatter(e,t,n,s){let o=new I([],e.type.dtype),i=new Ct;return i.addBlock(new yn([o,o])),this._inferAndBuild("scatter",[e,t,n],{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,n,s){return this.scatter(e,t,n,s)}_dimCoordIndices(e,t,n){let s=n.length,o=e.type.dtype==="i32"?e:this.convert(e,"i32").getResult(0),i=[...n,1],a=[];for(let u=0;u<s;u++){let p=u===t?o:this.iota(u,new I(n,"i32")).getResult(0);a.push(this.reshape(p,i).getResult(0))}return this.concat(a,s).getResult(0)}gatherDim(e,t,n){let s=e.type.rank,o=n<0?s+n:n,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,n,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,n,{updateWindowDims:[],insertedWindowDims:a,scatterDimsToOperandDims:a,indexVectorDim:o})}argmax(e,t,n=!1){let s=t<0?e.type.rank+t:t;return this._inferAndBuild("argmax",[e],{axis:s,keep_dims:n})}argmin(e,t,n=!1){let s=t<0?e.type.rank+t:t;return this._inferAndBuild("argmin",[e],{axis:s,keep_dims:n})}pool2d(e,t,n,s,o,i={}){return this._inferAndBuild("pool2d",[e],{pool_type:t,kernel_size:n,strides:s,padding:o,ceil_mode:i.ceilMode||!1,count_include_pad:i.countIncludePad||!1,layout:i.layout||"NCHW"})}resize(e,t,n,s={}){return this._inferAndBuild("resize",[e],{output_size:t,method:n,coordinate_mode:s.coordinateMode||"asymmetric",layout:s.layout||"NCHW"})}};l(To,"broadcastDimsExcluding");l(dL,"toMap");l(hc,"buildFunction")});function In(r,e,t){if(t==null)return new C(r,e,t);let n=e instanceof L,s=t instanceof L;if(n&&s){let o=e.value,i=t.value;switch(r){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((r==="+"||r==="-")&&o===0||r==="*"&&o===1)return e;if(r==="*"&&o===0)return new L(0);if(r==="//"&&o===1)return e;if(r==="%"&&o===1)return new L(0)}if(n){let o=e.value;if(r==="+"&&o===0||r==="*"&&o===1)return t;if(r==="*"&&o===0)return new L(0)}return new C(r,e,t)}var F,Tt,Jo,de,pe,Ib,qi,Y,q,xe,Ye,Ve,te,Zo,ou,On,iu,ke,C,qe,rt,Ne,L,fe,De=M(()=>{"use strict";F=Object.freeze({SERIAL:"serial",PARALLEL:"parallel",VECTORIZED:"vectorized",UNROLLED:"unrolled",THREAD_BINDING:"thread_binding",RECURRENCE:"recurrence"}),Tt=class r{static{l(this,"TensorNode")}constructor(){this.type=this.constructor.name.replace(/^_+/,""),this._parent=null,this._parentKey=null,this._parentIdx=-1}_setChild(e,t,n=-1){t instanceof r&&(t._parent=this,t._parentKey=e,t._parentIdx=n)}_setChildren(e,t){if(t)for(let n=0;n<t.length;n++)t[n]instanceof r&&(t[n]._parent=this,t[n]._parentKey=e,t[n]._parentIdx=n)}replaceWith(e){let t=this._parent;return t?(this._parentIdx>=0?t[this._parentKey][this._parentIdx]=e:t[this._parentKey]=e,e instanceof r&&(e._parent=t,e._parentKey=this._parentKey,e._parentIdx=this._parentIdx),this._parent=null,this._parentKey=null,this._parentIdx=-1,!0):!1}},Jo=class extends Tt{static{l(this,"PrimFunc")}constructor(e,t,n,s=new Map,o=[],i=null){super(),this.name=e,this.params=t,this.body=n,this.bufferMap=s,this.shapeParams=o,this.shapeParamMap=i||new Map,this._setChild("body",n)}},de=class extends Tt{static{l(this,"ForNode")}constructor(e,t,n,s,o,i=null){super(),this.loopVar=e,this.min=t,this.extent=n,this.kind=s,this.body=o,this.threadTag=i,this._setChild("body",o)}},pe=class extends Tt{static{l(this,"BlockNode")}constructor(e,t,n,s,o,i=null){super(),this.name=e,this.iterVars=t,this.reads=n,this.writes=s,this.body=o,this.initBody=i,this._setChild("body",o),this._setChild("initBody",i)}},Ib=Object.freeze({DATA_PAR:"DataPar",COMM_REDUCE:"CommReduce"}),qi=class extends Tt{static{l(this,"BlockRealizeNode")}constructor(e,t,n=Ib.DATA_PAR){super(),this.iterVar=e,this.binding=t,this.kind=n}},Y=class extends Tt{static{l(this,"BufferStoreNode")}constructor(e,t,n){super(),this.buffer=e,this.indices=t,this.value=n}},q=class extends Tt{static{l(this,"BufferLoadNode")}constructor(e,t){super(),this.buffer=e,this.indices=t}},xe=class extends Tt{static{l(this,"IfThenElseNode")}constructor(e,t,n=null){super(),this.condition=e,this.thenBody=t,this.elseBody=n,this._setChild("thenBody",t),this._setChild("elseBody",n)}},Ye=class extends Tt{static{l(this,"LetStmtNode")}constructor(e,t,n){super(),this.variable=e,this.value=t,this.body=n,this._setChild("body",n)}},Ve=class extends Tt{static{l(this,"AllocateNode")}constructor(e,t,n){super(),this.buffer=e,this.scope=t,this.body=n,this._setChild("body",n)}},te=class extends Tt{static{l(this,"SeqNode")}constructor(e){super(),this.stmts=e,this._setChildren("stmts",e)}},Zo=class extends Tt{static{l(this,"WhileNode")}constructor(e,t,n){super(),this.condVar=e,this.condBody=t,this.loopBody=n,this._setChild("condBody",t),this._setChild("loopBody",n)}},ou=class extends Tt{static{l(this,"EvaluateNode")}constructor(e){super(),this.value=e}},On=class extends Tt{static{l(this,"SyncThreadsNode")}constructor(){super()}},iu=class extends Tt{static{l(this,"VecCopyNode")}constructor(e,t,n,s,o){super(),this.dstBuffer=e,this.dstIndex=t,this.srcBuffer=n,this.srcIndex=s,this.width=o}},ke=class extends Tt{static{l(this,"CallExternNode")}constructor(e,t,n){super(),this.externName=e,this.args=t,this.dtype=n}},C=class extends Tt{static{l(this,"MathOpNode")}constructor(e,t,n=null){super(),this.op=e,this.a=t,this.b=n}},qe=class extends Tt{static{l(this,"CompareNode")}constructor(e,t,n){super(),this.direction=e,this.a=t,this.b=n}},rt=class extends Tt{static{l(this,"CastNode")}constructor(e,t,n){super(),this.expr=e,this.fromDtype=t,this.toDtype=n}},Ne=class extends Tt{static{l(this,"VariableNode")}constructor(e,t){super(),this.name=e,this.dtype=t}},L=class extends Tt{static{l(this,"IntImmNode")}constructor(e){super(),this.value=e}},fe=class extends Tt{static{l(this,"FloatImmNode")}constructor(e){super(),this.value=e}};l(In,"mathOp")});var au,Eb=M(()=>{"use strict";au=Object.freeze({GLOBAL:"global",SHARED:"shared",LOCAL:"local",REGISTER:"register"})});var me,xo=M(()=>{"use strict";Vn();Oe();me=class{static{l(this,"Buffer")}constructor(e,t,n,s,o=null,i=0,a=64){if(this.name=e,this.shape=t,this.dtype=n,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 oc(this.shape,-1)}symbolicNumel(){return mb(this.shape)}sizeInBytes(){let e=this.numel();return e<0?-1:e*bn(this.dtype)}}});function mL(r,e,t){switch(r){case"add":return new C("+",e,t);case"sub":return new C("-",e,t);case"mul":return new C("*",e,t);case"div":return new C("//",e,t);case"mod":return new C("%",e,t);case"neg":return new C("-",e);case"ceildiv":return new C("//",new C("-",new C("+",e,t),new L(1)),t);default:throw new Error(`symIntToNode: unsupported op '${r}' in extent/index context`)}}function uu(r,e){if(typeof r=="number")return new L(r);if(!(r instanceof Ce))return new L(r);if(r.type==="var")return e(r.name);let t=uu(r.args[0],e),n=r.args.length>1?uu(r.args[1],e):null;return mL(r.type,t,n)}var Lb=M(()=>{"use strict";Vs();De();l(mL,"symOpToNode");l(uu,"symIntToNode")});function ln(r){let e=ne.get(r);return e!==null&&e.isConstant}function qt(r){let e=ne.get(r);return e!==null&&e.isTerminator}function lu(r){let e=ne.get(r);return e!==null&&e.isBroadcast}var Tr=M(()=>{"use strict";At();l(ln,"isConstantOp");l(qt,"isTerminatorOp");l(lu,"isBroadcastOp")});function qh(r,{name:e,compute:t,plevel:n=10,targetKind:s=null}){let o=Cb.get(r);return o||(o=new dc(r),Cb.set(r,o)),o.addImplementation(e,t,n,s),Gh.clear(),o}function Uh(r,e=null){let t=Cb.get(r);if(!t)return null;let n=e?e.kind:null,s=`${r}|${n===null?"":n}`;if(Gh.has(s))return Gh.get(s);let o=new dc(r);for(let a of t.implementations)(a.targetKind===null||a.targetKind===n)&&o.implementations.push(a);let i=o.implementations.length>0?o:null;return Gh.set(s,i),i}function Db(r,e=null){let t=Uh(r,e);return t?t.best():null}var jh,dc,Cb,Gh,Pb=M(()=>{"use strict";jh=class{static{l(this,"OpImplementation")}constructor(e,t,n=10,s=null){this.name=e,this.compute=t,this.plevel=n,this.targetKind=s}},dc=class{static{l(this,"OpStrategy")}constructor(e){this.opName=e,this.implementations=[]}addImplementation(e,t,n=10,s=null){let o=new jh(e,t,n,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}},Cb=new Map,Gh=new Map;l(qh,"registerOpStrategy");l(Uh,"getOpStrategy");l(Db,"selectImplementation")});function Fb(r,e){return Uh(r,e)?!0:ln(r)}function Me(r,e,t=gL){qh(r,{name:`${r}.generic`,compute:e,plevel:t,targetKind:null})}function Ik(r,e,t,n=yL){qh(r,{name:`${r}.${e}`,compute:t,plevel:n,targetKind:e})}function Qo(r,e,t=null){if(t){let s=t.getLoweringRule(r);if(s)return s}let n=Db(r,e);return n?n.compute:void 0}function xt(r,e,t){let n=e.length,s=r.allocVarArray("i",n),o=r.allocBindArray("v",s),i=new Array(n);for(let u=0;u<n;u++)i[u]=o[u].iterVar;let a=t?r.extentNodes(e,t):null;return{loopVars:s,loopBinds:o,indices:i,extentNodes:a}}function gc(r,e,t){let n=r;for(let s=e.length-1;s>=0;s--)n=new de(e[s],new L(0),t[s],F.SERIAL,n);return n}function _L(r,e,t){let n=r;for(let s=e.length-1;s>=0;s--)n=new de(e[s],new L(0),new L(t[s]),F.SERIAL,n);return n}function St(r,e,t,n){return n?gc(r,e,n):_L(r,e,t)}function yc(r,e,t,n,s,{prefix:o,initBlockName:i,accBlockName:a,initVal:u,accLeaf:p}){let c=bL(r,e,t,n),h=or(r,o,Array.from({length:s.shape.length},(A,N)=>N),s.shape,s),f=new Y(s,h.indices,u()),d=new pe(r.blockName(i),h.ivs,[],[{buffer:s}],f),m=h.wrap(d),g=new q(t,c.lhsIdx),y=new q(n,c.rhsIdx),_=e.getAttr("lhs_prologue_cast"),b=e.getAttr("rhs_prologue_cast");_&&(g=new rt(g,t.dtype,_)),b&&(y=new rt(y,n.dtype,b));let w=p(g,y),T=new C("+",new q(s,c.outIdx),w),S=new Y(s,c.outIdx,T),x=new pe(r.blockName(a),c.allIvs,[{buffer:t},{buffer:n}],[{buffer:s}],S),k=c.wrapAccBody(x);return{geo:c,initBody:m,accBody:k}}function Hh(r,e,t,n,s,{prefix:o,blockPrefix:i,initVal:a,guardFill:u,leafBuilder:p}){let c=e.getAttr("strides"),h=e.getAttr("padding"),f=e.getAttr("dilation")||c.map(()=>1),d=e.getAttr("groups")||1,m=Wh(e.getAttr("input_layout")),g=Wh(e.getAttr("kernel_layout")),y=c.length,_=t.shape[m.N],b=n.shape[g.O],w=n.shape[g.I],T=s.shape,S=or(r,o+"i",Array.from({length:T.length},(ye,He)=>He),T,s),x=new Y(s,S.indices,a()),k=new pe(r.blockName(i+"_init"),S.ivs,[],[{buffer:s}],x),A=S.wrap(k),N=r.allocVar(o+"n"),v=r.allocVar(o+"oc"),O=r.allocVar(o+"ic"),R=r.allocVarArray(o+"o",y),B=r.allocVarArray(o+"k",y),E=[N,v,...R,O,...B],z=r.allocBindArray(o+"v",E),D=z[0].iterVar,J=z[1].iterVar,j=z.slice(2,2+y),U=z[2+y].iterVar,P=z.slice(3+y),V=new Array(T.length);V[m.N]=D,V[m.C]=J;let he=Object.keys(m).filter(ye=>ye!=="N"&&ye!=="C").sort();for(let ye=0;ye<y;ye++)V[m[he[ye]]]=j[ye].iterVar;let le=new Array(t.shape.length);le[m.N]=D;let Ie=Math.floor(b/d);d>1?le[m.C]=new C("+",new C("*",new C("//",J,new L(Ie)),new L(w)),U):le[m.C]=U;let G=new Array(n.shape.length);G[g.O]=J,G[g.I]=U;let K=null;for(let ye=0;ye<y;ye++){let He=he[ye],gn=He.toUpperCase(),jt=In("+",In("*",j[ye].iterVar,new L(c[ye])),In("+",In("*",P[ye].iterVar,new L(f[ye])),new L(-h[ye][0])));if(le[m[He]]=jt,G[g[gn]]=P[ye].iterVar,h[ye][0]!==0||h[ye][1]!==0){let Zt=new qe("ge",jt,new L(0)),Qt=new qe("lt",jt,new L(t.shape[m[He]])),on=new C("*",Zt,Qt);K=K?new C("*",K,on):on}}let ie=p(le,G),ze=K?new xe(K,ie,u()):ie,Se=new q(s,V),Qe=new C("+",Se,ze),It=new Y(s,V,Qe),Jt=new pe(r.blockName(i+"_acc"),z,[{buffer:t},{buffer:n}],[{buffer:s}],It),Gt=new Array(y);for(let ye=0;ye<y;ye++){let He=he[ye].toUpperCase();Gt[ye]=n.shape[g[He]]}let tt=Jt;for(let ye=y-1;ye>=0;ye--){let He=he[ye].toUpperCase();tt=new de(B[ye],new L(0),r.extentNode(Gt[ye],n,g[He]),F.SERIAL,tt)}tt=new de(O,new L(0),r.extentNode(w,n,g.I),F.SERIAL,tt);for(let ye=y-1;ye>=0;ye--){let He=m[he[ye]];tt=new de(R[ye],new L(0),r.extentNode(T[He],s,He),F.SERIAL,tt)}return tt=new de(v,new L(0),r.extentNode(b,n,g.O),F.SERIAL,tt),tt=new de(N,new L(0),r.extentNode(_,t,m.N),F.SERIAL,tt),new te([A,tt])}function or(r,e,t,n,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]=r.allocVar(`${e}${t[c]}`),a[c]=new qi(r.allocVar(`${e}v${t[c]}`),i[c]),u[c]=a[c].iterVar,p[c]=r.extentNode(n[t[c]],s,t[c]);return{vars:i,ivs:a,indices:u,extentNodes:p,wrap(c){return gc(c,i,p)}}}function So(r,e,t){let n=r.shape.length;if(r.broadcastDims){let a=r.broadcastDims,u=new Array(n);for(let p=0;p<n;p++)u[p]=r.shape[p]===1?new L(0):t[a[p]];return u}let o=e.shape.length-n,i=new Array(n);for(let a=0;a<n;a++)i[a]=r.shape[a]===1?new L(0):t[o+a];return i}function cu(r){let e=new Array(r.length);for(let t=0;t<r.length;t++)e[t]={buffer:r[t]};return e}function mc(){let r=0;for(let n=0;n<arguments.length;n++)r+=arguments[n].length;let e=new Array(r),t=0;for(let n=0;n<arguments.length;n++){let s=arguments[n];for(let o=0;o<s.length;o++)e[t++]=s[o]}return e}function $b(r){let e=new Array(r.length);for(let t=0;t<r.length;t++)e[t]=r[t].iterVar;return e}function qs(r,e,t,n,s){let o=n[0],{loopVars:i,loopBinds:a,indices:u,extentNodes:p}=xt(r,o.shape,o),c=new Array(t.length);for(let m=0;m<t.length;m++){let g=So(t[m],o,u);c[m]=new q(t[m],g)}let h=s(e,c,o.dtype),f=new Y(o,u,h),d=new pe(r.blockName(`${e.opName}_block`),a,cu(t),[{buffer:o}],f);return St(d,i,o.shape,p)}function pu(r,e){let t=e.getResult(0),n=e.getAttr("value"),s=t.type&&t.type.shape||[];if(typeof n=="number"&&s.length>0&&!r.bufferMap.has(t)){let g=t.type&&t.type.dtype||"f32",y=new Array(s.length).fill(1),_=new me(`buf_${r.varCounter++}`,y,g,au.GLOBAL);_.broadcastDims=Array.from({length:s.length},(w,T)=>T),r.bufferMap.set(t,_);let b=nr(g)?new L(n):new fe(n);return new Y(_,y.map(()=>new L(0)),b)}let o=r.getOrAllocBuffer(t),i=nr(o.dtype),a=l(g=>i?new L(g):new fe(g),"imm");if(n&&typeof n!="number"&&typeof n.length=="number"){if(o.shape.length===0)return new Y(o,[],a(n[0]));let g=new Array(o.shape.length),y=1;for(let b=o.shape.length-1;b>=0;b--)g[b]=y,y*=o.shape[b];let _=[];for(let b=0;b<n.length;b++){let w=new Array(o.shape.length);for(let T=0;T<o.shape.length;T++)w[T]=new L(Math.floor(b/g[T])%o.shape[T]);_.push(new Y(o,w,a(n[b])))}return new te(_)}let u=a(typeof n=="number"?n:0);if(o.shape.length===0)return new Y(o,[],u);let{loopVars:p,loopBinds:c,indices:h,extentNodes:f}=xt(r,o.shape,o),d=new Y(o,h,u),m=new pe(r.blockName(`${e.opName}_block`),c,[],[{buffer:o}],d);return St(m,p,o.shape,f)}function Wh(r){let e={};for(let t=0;t<r.length;t++)e[r[t]]=t;return e}function Mk(r,e){if(!r.broadcastDims)return e;let t=r.broadcastDims,n=new Array(r.shape.length);for(let s=0;s<r.shape.length;s++)n[s]=r.shape[s]===1?new L(0):e[t[s]];return n}function bL(r,e,t,n){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),h=new Set(i),f=new Set(o),d=new Set(a),m=[];for(let D=0;D<u.length;D++)!c.has(D)&&!h.has(D)&&m.push(D);let g=[];for(let D=0;D<p.length;D++)!f.has(D)&&!d.has(D)&&g.push(D);let y=r.allocVarArray("b",i.length),_=r.allocVarArray("ls",m.length),b=r.allocVarArray("rs",g.length),w=r.allocVarArray("c",s.length),T=r.allocBindArray("vb",y),S=r.allocBindArray("vls",_),x=r.allocBindArray("vrs",b),k=r.allocBindArray("vc",w),A=mc($b(T),$b(S),$b(x)),N=new Array(u.length);for(let D=0;D<i.length;D++)N[i[D]]=T[D].iterVar;for(let D=0;D<m.length;D++)N[m[D]]=S[D].iterVar;for(let D=0;D<s.length;D++)N[s[D]]=k[D].iterVar;let v=new Array(p.length);for(let D=0;D<a.length;D++)v[a[D]]=T[D].iterVar;for(let D=0;D<g.length;D++)v[g[D]]=x[D].iterVar;for(let D=0;D<o.length;D++)v[o[D]]=k[D].iterVar;let O=Mk(t,N),R=Mk(n,v),B=mc(T,S,x,k),E=[{vars:y,dims:i,shape:u,buf:t},{vars:_,dims:m,shape:u,buf:t},{vars:b,dims:g,shape:p,buf:n},{vars:w,dims:s,shape:u,buf:t}];function z(D){let J=D;for(let j=E.length-1;j>=0;j--){let{vars:U,dims:P,shape:V,buf:he}=E[j];for(let le=U.length-1;le>=0;le--){let Ie=he.broadcastDims?new L(V[P[le]]):r.extentNode(V[P[le]],he,P[le]);J=new de(U[le],new L(0),Ie,F.SERIAL,J)}}return J}return l(z,"wrapAccBody"),{outIdx:A,lhsIdx:O,rhsIdx:R,allIvs:B,wrapAccBody:z}}var gL,yL,Kh,Fr=M(()=>{"use strict";Oe();Vs();Eb();xo();Vn();De();Lb();Tr();Pb();Pb();gL=10,yL=20;l(Fb,"hasLoweringRule");l(Me,"registerLoweringRule");l(Ik,"registerTargetLoweringRule");l(Qo,"getLoweringRule");Kh=class{static{l(this,"LoweringContext")}constructor(){this.bufferMap=new Map,this.varCounter=0,this.shapeParams=new Map,this.symbolToVar=new Map,this.symVars=new Map,this._blockCounter=0}blockName(e){return`${e}_${this._blockCounter++}`}allocVar(e,t="int32"){return new Ne(`${e}_${this.varCounter++}`,t)}getOrAllocBuffer(e){let t=this.bufferMap.get(e);if(t)return t;let n=e.type,s=n.shape||[],o=n.dtype||"f32",i=n.layout?n.layout.computeStrides(s):null;return t=new me(`buf_${this.varCounter++}`,s,o,au.GLOBAL,i),e.symbolicShape&&(t.symbolicShape=e.symbolicShape),this.bufferMap.set(e,t),this._registerDynamicDims(t),t}allocFreshBuffer(e){let t=e.type,n=t.shape||[],s=t.dtype||"f32",o=t.layout?t.layout.computeStrides(n):null,i=new me(`buf_${this.varCounter++}`,n,s,au.GLOBAL,o);return e.symbolicShape&&(i.symbolicShape=e.symbolicShape),this._registerDynamicDims(i),i}_registerDynamicDims(e){for(let t=0;t<e.shape.length;t++){let n=e.shape[t];n===X?this.extentNode(X,e,t):n instanceof Ce&&this._registerSymIntDim(e,t,n)}}_symVarNode(e){let t=this.symVars.get(e);return t||(t=new Ne(qa(e),"int32"),this.symVars.set(e,t)),t}_registerSymIntDim(e,t,n){for(let s of Ce.freeVars(n))this._symVarNode(s);if(n.type==="var"){let s=`${e.name}:${t}`;this.shapeParams.has(s)||this.shapeParams.set(s,this._symVarNode(n.name))}}symIntToExtentNode(e){return uu(e,t=>this._symVarNode(t))}_shapeParamVar(e,t){let n=t>=0?`${e.name}:${t}`:`${e.name}:dyn`,s=this.shapeParams.get(n);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(n,s),s}extentNode(e,t,n=-1){return e instanceof Ce?this.symIntToExtentNode(e):e!==X?new L(e):this._shapeParamVar(t,n)}extentNodes(e,t){let n=new Array(e.length);for(let s=0;s<e.length;s++)n[s]=this.extentNode(e[s],t,s);return n}allocVarArray(e,t){let n=new Array(t);for(let s=0;s<t;s++)n[s]=this.allocVar(`${e}${s}`);return n}allocBindArray(e,t){let n=new Array(t.length);for(let s=0;s<t.length;s++)n[s]=new qi(this.allocVar(`${e}${s}`),t[s]);return n}};l(xt,"makeLoopNest");l(gc,"wrapLoopsWithNodes");l(_L,"wrapLoops");l(St,"wrapInLoops");l(yc,"emitMatmulInitAcc");l(Hh,"buildConvNest");l(or,"buildSpatialNest");l(So,"computeBroadcastIndices");l(cu,"bufRefs");l(mc,"concatIterVars");l($b,"extractIterVars");l(qs,"lowerPointwise");l(pu,"lowerConstant");l(Wh,"parseLayout");l(Mk,"physicalDotIndices");l(bL,"buildDotGeometry")});function kL(r){let e=ne.get(r);return e?e.getAttr(zb):null}function _c(){let r=[];for(let e of ne.allOps())e.hasAttr(zb)&&r.push(e.name);return r}function Vb(r,e,t){let n=kL(r);return n?Ek[r]?Ek[r](e):e.length===2&&wL.has(n)?new C(n,e[0],e[1]):e.length===2&&TL.has(n)?new C(n,e[0],e[1]):e.length===1&&xL.has(n)?new C(n,e[0]):e.length===1&&n==="-"?new C("-",e[0]):new ke(n,e,t):null}function Lk(){for(let r of _c())Me(r,(e,t,n,s)=>qs(e,t,n,s,(o,i,a)=>Vb(o.opName,i,a)));Me("compare",(r,e,t,n)=>qs(r,e,t,n,(s,o)=>new qe(s.getAttr("direction")||"eq",o[0],o[1]))),Me("select",(r,e,t,n)=>qs(r,e,t,n,(s,o)=>new xe(o[0],o[1],o[2]))),Me("clamp",(r,e,t,n)=>qs(r,e,t,n,(s,o,i)=>new ke("min",[new ke("max",[o[1],o[0]],i),o[2]],i))),Me("convert",(r,e,t,n)=>qs(r,e,t,n,(s,o)=>new rt(o[0],t[0].dtype,n[0].dtype))),Me("copy_to_device",(r,e,t,n)=>qs(r,e,t,n,(s,o)=>o[0]))}var wL,TL,xL,Ek,zb,SL,Yh=M(()=>{"use strict";De();At();Fr();wL=new Set(["+","-","*","/"]),TL=new Set(["&&","||"]),xL=new Set(["!"]),Ek={square:l(r=>new C("*",r[0],r[0]),"square"),reciprocal:l(r=>new C("/",new fe(1),r[0]),"reciprocal")},zb="elementwiseScalarOp",SL={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[r,e]of Object.entries(SL))ne.has(r)&&ne.registerOpAttr(r,zb,e);l(kL,"elementwiseScalarOp");l(_c,"elementwiseOpNames");l(Vb,"buildElementwiseExpr");l(Lk,"register")});function Dk(){Me("broadcast_in_dim",Ck),Me("broadcast",Ck),Me("transpose",(r,e,t,n)=>{let s=e.getAttr("permutation"),o=t[0],i=n[0],{loopVars:a,loopBinds:u,indices:p,extentNodes:c}=xt(r,i.shape,i),h=new Array(o.shape.length);for(let m=0;m<s.length;m++)h[s[m]]=p[m];let f=new Y(i,p,new q(o,h)),d=new pe(r.blockName("transpose_block"),u,[{buffer:o}],[{buffer:i}],f);return St(d,a,i.shape,c)}),Me("reverse",(r,e,t,n)=>{let s=new Set(e.getAttr("dimensions")),o=t[0],i=n[0],{loopVars:a,loopBinds:u,indices:p,extentNodes:c}=xt(r,i.shape,i),h=new Array(o.shape.length);for(let m=0;m<o.shape.length;m++)h[m]=s.has(m)?new C("-",new C("-",c[m],new L(1)),p[m]):p[m];let f=new Y(i,p,new q(o,h)),d=new pe(r.blockName("reverse_block"),u,[{buffer:o}],[{buffer:i}],f);return St(d,a,i.shape,c)}),Me("reshape",(r,e,t,n)=>{let s=t[0],o=n[0],{loopVars:i,loopBinds:a,indices:u,extentNodes:p}=xt(r,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=In("+",m,In("*",u[_],new L(g)));c=new Array(s.shape.length);let y=m;for(let _=s.shape.length-1;_>=0;_--)_===0?c[_]=y:(c[_]=In("%",y,new L(s.shape[_])),y=In("//",y,new L(s.shape[_])))}let f=new Y(o,u,new q(s,c)),d=new pe(r.blockName("reshape_block"),a,[{buffer:s}],[{buffer:o}],f);return St(d,i,o.shape,p)}),Me("slice",(r,e,t,n)=>{let s=t[0],o=n[0],i=e.getAttr("starts"),a=e.getAttr("strides")||i.map(()=>1),{loopVars:u,loopBinds:p,indices:c,extentNodes:h}=xt(r,o.shape,o),f=new Array(s.shape.length);for(let g=0;g<s.shape.length;g++){let y=new L(i[g]);a[g]===1?f[g]=new C("+",y,c[g]):f[g]=new C("+",y,new C("*",c[g],new L(a[g])))}let d=new Y(o,c,new q(s,f)),m=new pe(r.blockName("slice_block"),p,[{buffer:s}],[{buffer:o}],d);return St(m,u,o.shape,h)}),Me("pad",(r,e,t,n)=>{let s=t[0],o=t[1],i=n[0],a=e.getAttr("low"),u=e.getAttr("interior")||a.map(()=>0),{loopVars:p,loopBinds:c,indices:h,extentNodes:f}=xt(r,i.shape,i),d=new Array(s.shape.length),m=new L(1);for(let T=0;T<s.shape.length;T++){let S=new C("+",h[T],new L(-a[T]));if(u[T]>0){let A=u[T]+1,N=new C("%",S,new L(A)),v=new qe("eq",N,new L(0));m=new C("*",m,v),d[T]=new C("//",S,new L(A))}else d[T]=S;let x=new qe("ge",d[T],new L(0)),k=new qe("lt",d[T],new L(s.shape[T]));m=new C("*",m,new C("*",x,k))}let g=new q(s,d),y=new q(o,[]),_=new xe(m,g,y),b=new Y(i,h,_),w=new pe(r.blockName("pad_block"),c,[{buffer:s},{buffer:o}],[{buffer:i}],b);return St(w,p,i.shape,f)}),Me("concat",(r,e,t,n)=>{let s=n[0],o=e.getAttr("dimension"),i=[],a=0;for(let u=0;u<t.length;u++){let p=t[u],{loopVars:c,loopBinds:h,indices:f,extentNodes:d}=xt(r,p.shape,p),m=new Array(p.shape.length);for(let _=0;_<p.shape.length;_++)m[_]=_===o&&a>0?new C("+",f[_],new L(a)):f[_];let g=new Y(s,m,new q(p,f)),y=new pe(r.blockName("concat"),h,[{buffer:p}],[{buffer:s}],g);i.push(St(y,c,p.shape,d)),a+=p.shape[o]}return i.length===1?i[0]:new te(i)}),Me("iota",(r,e,t,n)=>{let s=n[0],o=e.getAttr("iota_dimension"),{loopVars:i,loopBinds:a,indices:u,extentNodes:p}=xt(r,s.shape,s),c=new rt(u[o],"index",s.dtype),h=new Y(s,u,c),f=new pe(r.blockName("iota_block"),a,[],[{buffer:s}],h);return St(f,i,s.shape,p)})}function Ck(r,e,t,n){let s=t[0],o=n[0],i=e.getAttr("broadcast_dimensions")||[],{loopVars:a,loopBinds:u,indices:p,extentNodes:c}=xt(r,o.shape,o),h=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);h[m]=s.shape[m]===1?new L(0):p[g]}let f=new Y(o,p,new q(s,h)),d=new pe(r.blockName("broadcast_block"),u,[{buffer:s}],[{buffer:o}],f);return St(d,a,o.shape,c)}var Pk=M(()=>{"use strict";De();Fr();l(Dk,"register");l(Ck,"lowerBroadcast")});function AL(r,e){if(nr(r)){let t=e?vL[r]??-2147483648:NL[r]??2147483647;return new L(t)}return new fe(e?-1/0:1/0)}function $k(){Me("reduce",(e,t,n,s)=>{let o=n[0],i=n[1],a=s[0],u=t.getAttr("dimensions")||[],p=t.getAttr("reduce_type")||"sum",c=new Set(u),h=[],f=[];for(let B=0;B<o.shape.length;B++)(c.has(B)?f:h).push(B);let d=or(e,"si",h,o.shape,o),m=new Y(a,d.indices,new q(i,[])),g=new pe(e.blockName("reduce_init"),d.ivs,[{buffer:i}],[{buffer:a}],m),y=h.length>0?d.wrap(g):g,_=or(e,"sa",h,o.shape,o),b=e.allocVarArray("r",f.length),w=e.allocBindArray("rv",b);for(let B of w)B.kind=Ib.COMM_REDUCE;let T=new Array(o.shape.length);for(let B=0;B<h.length;B++)T[h[B]]=_.ivs[B].iterVar;for(let B=0;B<f.length;B++)T[f[B]]=w[B].iterVar;let S=new q(a,_.indices),x=new q(o,T),k=OL[p];if(!k)throw new Error(`reduction lowering: unsupported reduce_type '${p}'`);let A=new Y(a,_.indices,k(S,x,a.dtype)),N=new Array(f.length);for(let B=0;B<f.length;B++)N[B]=e.extentNode(o.shape[f[B]],o,f[B]);let v=new pe(e.blockName("reduce_acc"),mc(_.ivs,w),[{buffer:o}],[{buffer:a}],A),O=gc(v,b,N);O=_.wrap(O);let R=[y,O];if(p==="mean"){let B=1,E=[];for(let P=0;P<f.length;P++){let V=o.shape[f[P]];V===X?E.push(e.extentNode(X,o,f[P])):B*=V}let z=or(e,"sm",h,o.shape,o),D=new q(a,z.indices),J;if(E.length===0)J=new C("*",D,new fe(1/B));else{let P=new L(B);for(let V of E)P=new C("*",P,V);J=new C("/",D,P)}let j=new Y(a,z.indices,J),U=new pe(e.blockName("mean_div"),z.ivs,[{buffer:a}],[{buffer:a}],j);R.push(h.length>0?z.wrap(U):U)}return new te(R)});function r(e,t){Me(e,(n,s,o,i)=>{let a=o[0],u=i[0],p=s.getAttr("axis"),c=s.getAttr("keep_dims")||!1,h=new Set([p]),f=[],d=p;for(let V=0;V<a.shape.length;V++)h.has(V)||f.push(V);let m=l(V=>{if(!c)return V.indices;let he=new Array(a.shape.length);for(let le=0;le<f.length;le++)he[f[le]]=V.indices[le];return he[d]=new L(0),he},"outIndicesFor"),g=new me("_argval_"+n.varCounter,f.map(V=>a.shape[V]),a.dtype,"global");n.varCounter++;let y=or(n,"ai",f,a.shape,a),_=new Y(g,y.indices,AL(a.dtype,t==="gt")),b=new Y(u,m(y),new L(0)),w=new pe(n.blockName("arg_init"),y.ivs,[],[{buffer:g},{buffer:u}],new te([_,b])),T=f.length>0?y.wrap(w):w,S=or(n,"as",f,a.shape,a),x=n.allocVar("ar"),k=n.allocBindArray("arv",[x]),A=new Array(a.shape.length);for(let V=0;V<f.length;V++)A[f[V]]=S.ivs[V].iterVar;A[d]=k[0].iterVar;let N=new q(a,A),v=new q(g,S.indices),O=new qe(t,N,v),R=new xe(O,N,v),B=m(S),E=new q(u,B),z=new xe(O,k[0].iterVar,E),D=new Y(u,B,z),J=new Y(g,S.indices,R),j=new pe(n.blockName("arg_acc"),mc(S.ivs,k),[{buffer:a},{buffer:g}],[{buffer:g},{buffer:u}],new te([D,J])),U=n.extentNode(a.shape[d],a,d),P=gc(j,[x],[U]);return P=S.wrap(P),new te([T,P])})}l(r,"registerArgReduce"),r("argmax","gt"),r("argmin","lt")}var vL,NL,OL,Fk=M(()=>{"use strict";De();xo();Oe();Fr();Vn();vL={i8:-128,i16:-32768,i32:-2147483648,i64:-2147483648,ui8:0,ui16:0,ui32:0,bool:0},NL={i8:127,i16:32767,i32:2147483647,i64:2147483647,ui8:255,ui16:65535,ui32:4294967295,bool:1};l(AL,"argReduceSentinel");OL={sum:l((r,e)=>new C("+",r,e),"sum"),mean:l((r,e)=>new C("+",r,e),"mean"),prod:l((r,e)=>new C("*",r,e),"prod"),max:l((r,e,t)=>new ke("max",[r,e],t),"max"),min:l((r,e,t)=>new ke("min",[r,e],t),"min"),and:l((r,e)=>new C("*",r,e),"and"),or:l((r,e,t)=>new ke("max",[r,e],t),"or")};l($k,"register")});function zk(r,e){return Lt(e)?new rt(r,e,"i32"):r}function Vk(){Me("dot",(r,e,t,n)=>{let{initBody:s,accBody:o}=yc(r,e,t[0],t[1],n[0],{prefix:"di",initBlockName:"matmul_init",accBlockName:"matmul",initVal:l(()=>new fe(0),"initVal"),accLeaf:l((i,a)=>new C("*",i,a),"accLeaf")});return new te([s,o])}),Me("conv",(r,e,t,n)=>{let s=t[0],o=t[1];return Hh(r,e,s,o,n[0],{prefix:"c",blockPrefix:"conv",initVal:l(()=>new fe(0),"initVal"),guardFill:l(()=>new fe(0),"guardFill"),leafBuilder:l((i,a)=>new C("*",new q(s,i),new q(o,a)),"leafBuilder")})}),Me("gather",(r,e,t,n)=>{let s=t[0],o=t[1],i=n[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:h,loopBinds:f,indices:d,extentNodes:m}=xt(r,i.shape,i),g=[],y=[];for(let S=0;S<i.shape.length;S++)a.has(S)?y.push(d[S]):g.push(d[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 x=new Array(o.shape.length),k=0;for(let v=0;v<o.shape.length;v++)v===c?x[v]=new L(S):x[v]=g[k++];let A=zk(new q(o,x),o.dtype),N=p[S];_[N]=new C("+",_[N],A)}let w=new Y(i,d,new q(s,_)),T=new pe(r.blockName("gather_block"),f,[{buffer:s},{buffer:o}],[{buffer:i}],w);return St(T,h,i.shape,m)}),Me("scatter",(r,e,t,n)=>{let s=t[0],o=t[1],i=t[2],a=n[0],u=new Set(e.getAttr("inserted_window_dims")),p=e.getAttr("scatter_dims_to_operand_dims"),c=e.getAttr("index_vector_dim"),h=new Set(e.getAttr("update_window_dims")),f=xt(r,s.shape,s),d=new Y(a,f.indices,new q(s,f.indices)),m=new pe(r.blockName("scatter_copy"),f.loopBinds,[{buffer:s}],[{buffer:a}],d),g=St(m,f.loopVars,s.shape,f.extentNodes),{loopVars:y,loopBinds:_,indices:b,extentNodes:w}=xt(r,i.shape,i),T=[],S=[];for(let E=0;E<i.shape.length;E++)h.has(E)?S.push(b[E]):T.push(b[E]);let x=new Array(s.shape.length),k=0;for(let E=0;E<s.shape.length;E++)u.has(E)?x[E]=new L(0):x[E]=S[k++];for(let E=0;E<p.length;E++){let z=new Array(o.shape.length),D=0;for(let U=0;U<o.shape.length;U++)U===c?z[U]=new L(E):z[U]=T[D++];let J=zk(new q(o,z),o.dtype),j=p[E];x[j]=new C("+",x[j],J)}let A=new q(i,b),N=new q(a,x),v=new C("+",N,A),O=new Y(a,x,v),R=new pe(r.blockName("scatter_update"),_,[{buffer:i},{buffer:o}],[{buffer:a}],O),B=St(R,y,i.shape,w);return new te([g,B])}),Me("fused_dot_epilogue",(r,e,t,n)=>{let s=e.getAttr("num_dot_operands")||2,o=t[0],i=t[1],a=t.slice(s),u=n[0],p=e.getAttr("epilogue_tags")||[],{initBody:c,accBody:h}=yc(r,e,o,i,u,{prefix:"ei",initBlockName:"matmul_init",accBlockName:"matmul_acc",initVal:l(()=>new fe(0),"initVal"),accLeaf:l((T,S)=>new C("*",T,S),"accLeaf")});if(p.length===0)return new te([c,h]);let f=or(r,"ep",Array.from({length:u.shape.length},(T,S)=>S),u.shape,u),d=f.indices,m=new q(u,d),g={v:0};for(let T of p){let S=zr.get(T);S&&(m=S(m,a,g,u,d))}let y=cu([u,...a]),_=new Y(u,d,m),b=new pe(r.blockName("epilogue"),f.ivs,y,[{buffer:u}],_),w=f.wrap(b);return new te([c,h,w])}),Me("cublas_gemm",Qo("dot"))}var zr,Gk=M(()=>{"use strict";De();Oe();Fr();l(zk,"asIndexValue");zr=new Map;zr.set("bias",(r,e,t,n,s)=>{let o=e[t.v++];return new C("+",r,new q(o,So(o,n,s)))});zr.set("residual_add",(r,e,t,n,s)=>{let o=e[t.v++];return new C("+",r,new q(o,So(o,n,s)))});zr.set("scale",(r,e,t,n,s)=>{let o=e[t.v++];return new C("*",r,new q(o,So(o,n,s)))});zr.set("relu",(r,e,t,n)=>new ke("max",[r,new fe(0)],n.dtype));zr.set("clamp",(r,e,t,n,s)=>{let o=e[t.v++],i=e[t.v++];return new ke("min",[new ke("max",[r,new q(o,So(o,n,s))],n.dtype),new q(i,So(i,n,s))],n.dtype)});zr.set("neg",r=>new C("-",r));zr.set("exp",(r,e,t,n)=>new ke("exp",[r],n.dtype));zr.set("tanh",(r,e,t,n)=>new ke("tanh",[r],n.dtype));zr.set("sqrt",(r,e,t,n)=>new ke("sqrt",[r],n.dtype));zr.set("abs",(r,e,t,n)=>new ke("abs",[r],n.dtype));zr.set("log",(r,e,t,n)=>new ke("log",[r],n.dtype));zr.set("activation",r=>r);l(Vk,"register")});function fu(r,e,t){let{loopVars:n,loopBinds:s,indices:o,extentNodes:i}=xt(r,t.shape,t),a=new Y(t,o,new q(e,o)),u=new pe(r.blockName("cf_copy"),s,[{buffer:e}],[{buffer:t}],a);return St(u,n,t.shape,i)}function BL(r,e,t,n){let{loopVars:s,loopBinds:o,indices:i,extentNodes:a}=xt(r,t.shape,t),u=new Y(t,i,new q(e,[n,...i])),p=new pe(r.blockName("scan_in"),o,[{buffer:e}],[{buffer:t}],u);return St(p,s,t.shape,a)}function RL(r,e,t,n){let{loopVars:s,loopBinds:o,indices:i,extentNodes:a}=xt(r,e.shape,e),u=new Y(t,[n,...i],new q(e,i)),p=new pe(r.blockName("scan_out"),o,[{buffer:e}],[{buffer:t}],u);return St(p,s,e.shape,a)}function bc(r,e,t){let n=e.entryBlock,s=new Map;for(let i=0;i<n.arguments.length;i++)s.set(n.arguments[i],t[i]),t[i]&&r.bufferMap.set(n.arguments[i],t[i]);let o=[];for(let i of n.ops()){if(i.opName==="yield"){let f=new Array(i.numOperands);for(let d=0;d<i.numOperands;d++)f[d]=s.get(i.getOperand(d))||r.getOrAllocBuffer(i.getOperand(d));return{stmts:o,yieldBuffers:f}}if(ln(i.opName)){o.push(pu(r,i));continue}let a=new Array(i.numOperands);for(let f=0;f<i.numOperands;f++)a[f]=s.get(i.getOperand(f))||i.getOperand(f);let u=new Array(a.length);for(let f=0;f<a.length;f++)u[f]=a[f]instanceof me?a[f]:r.getOrAllocBuffer(a[f]);let p=new Array(i.numResults);for(let f=0;f<i.numResults;f++){let d={type:i.getResult(f).type};p[f]=r.getOrAllocBuffer(d),s.set(i.getResult(f),p[f]),r.bufferMap.set(i.getResult(f),p[f])}let c=Qo(i.opName);if(!c)throw new Error(`No lowering rule for op '${i.opName}' inside region`);let h=c(r,i,u,p);h&&o.push(h)}return{stmts:o,yieldBuffers:[]}}function jk(){Me("if",(r,e,t,n)=>{let s=t[0],o=new q(s,[]),i=e.regions[0],a=e.regions[1],u=new Array(e.numResults);for(let d=0;d<e.numResults;d++)u[d]=n[d]||r.getOrAllocBuffer(e.getResult(d)),r.bufferMap.set(e.getResult(d),u[d]);let p=bc(r,i,[]),c=p.stmts.slice();for(let d=0;d<u.length&&d<p.yieldBuffers.length;d++){let m=p.yieldBuffers[d];m&&m!==u[d]&&c.push(fu(r,m,u[d]))}let h=c.length===1?c[0]:new te(c),f=null;if(a&&a.entryBlock){let d=bc(r,a,[]),m=d.stmts.slice();for(let g=0;g<u.length&&g<d.yieldBuffers.length;g++){let y=d.yieldBuffers[g];y&&y!==u[g]&&m.push(fu(r,y,u[g]))}m.length>0&&(f=m.length===1?m[0]:new te(m))}return new xe(o,h,f)}),Me("while",(r,e,t,n)=>{let s=e.regions[0],o=e.regions[1],i=new Array(t.length),a=[];for(let y=0;y<t.length;y++){let _=n[y]||r.getOrAllocBuffer({type:e.getResult(y).type});i[y]=_,a.push(fu(r,t[y],_))}for(let y=0;y<e.numResults;y++)r.bufferMap.set(e.getResult(y),i[y]);let u=new me(`_wcond_${r.varCounter++}`,[],"bool",au.GLOBAL),p=bc(r,s,i),c=p.stmts.slice(),h=p.yieldBuffers[0];c.push(new Y(u,[],new q(h,[])));let f=c.length===1?c[0]:new te(c),d=bc(r,o,i),m=d.stmts.slice();for(let y=0;y<i.length&&y<d.yieldBuffers.length;y++){let _=d.yieldBuffers[y];_&&_!==i[y]&&m.push(fu(r,_,i[y]))}let g=m.length===1?m[0]:new te(m);return new te([...a,new Zo(u,f,g)])}),Me("scan",(r,e,t,n)=>{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=n[b]||r.getOrAllocBuffer({type:e.getResult(b).type});u[b]=w,p.push(fu(r,a[b],w)),r.bufferMap.set(e.getResult(b),w)}let c=n.length-o,h=new Array(c);for(let b=0;b<c;b++){let w=n[o+b]||r.getOrAllocBuffer({type:e.getResult(o+b).type});h[b]=w,r.bufferMap.set(e.getResult(o+b),w)}let f=r.allocVar("t"),d=r.extentNode(i[0].shape[0],i[0],0),m=[],g=new Array(s);for(let b=0;b<s;b++){let w=r.getOrAllocBuffer({type:{shape:i[b].shape.slice(1),dtype:i[b].dtype}});g[b]=w,m.push(BL(r,i[b],w,f))}let y=bc(r,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(fu(r,w,u[b]))}for(let b=0;b<c;b++)m.push(RL(r,y.yieldBuffers[o+b],h[b],f));m.push(new On);let _=m.length===1?m[0]:new te(m);return new te([...p,new de(f,new L(0),d,F.RECURRENCE,_)])})}var qk=M(()=>{"use strict";De();xo();Eb();Fr();l(fu,"copyBuffer");l(BL,"sliceCopyIn");l(RL,"sliceCopyOut");l(bc,"lowerRegionBody");l(jk,"register")});function Uk(){Me("layout_transform",(r,e,t,n)=>{let s=t[0],o=n[0],i=e.getAttr("src_layout"),a=e.getAttr("dst_layout"),{loopVars:u,loopBinds:p,indices:c,extentNodes:h}=xt(r,o.shape,o),f=new Y(o,c,new q(s,c)),d=new pe(r.blockName("layout_transform_block"),p,[{buffer:s}],[{buffer:o}],f);return St(d,u,o.shape,h)})}var Kk=M(()=>{"use strict";De();Fr();l(Uk,"register")});function ML(r){let e=r==="ui8",t=8;return{cMin:e?0:-(1<<t-1),cMax:e?(1<<t)-1:(1<<t-1)-1}}function Jh(r,{scale:e,zeroPoint:t,targetDtype:n}){let s=n||"i8",{cMin:o,cMax:i}=ML(s),a=new C("/",r,new fe(e)),u=t!==0?new C("+",a,new fe(t)):a,p=new ke("round",[u],"f32"),c=new ke("min",[new ke("max",[p,new fe(o)],"f32"),new fe(i)],"f32");return new rt(c,"f32",s)}function Zh(r,{scale:e,zeroPoint:t,srcDtype:n,targetDtype:s}){let o=s||"f32",i=new rt(r,n||"i8",o),a=t!==0?new C("-",i,new fe(t)):i;return new C("*",a,new fe(e))}var Gb=M(()=>{"use strict";De();l(ML,"clampRange");l(Jh,"buildQuantizeExpr");l(Zh,"buildDequantizeExpr")});function Wk(){Me("quantize",(r,e,t,n)=>{let s=e.getAttr("scale"),o=e.getAttr("zero_point"),i=e.getAttr("target_dtype")||"i8";return qs(r,e,t,n,(a,u)=>Jh(u[0],{scale:s,zeroPoint:o,targetDtype:i}))}),Me("dequantize",(r,e,t,n)=>{let s=e.getAttr("scale"),o=e.getAttr("zero_point"),i=t[0].dtype||"i8",a=e.getAttr("target_dtype")||"f32";return qs(r,e,t,n,(u,p)=>Zh(p[0],{scale:s,zeroPoint:o,srcDtype:i,targetDtype:a}))}),Me("quantized_dot",(r,e,t,n)=>{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}=yc(r,e,s,o,n[0],{prefix:"qdi",initBlockName:"qmatmul_init",accBlockName:"qmatmul_acc",initVal:l(()=>new L(0),"initVal"),accLeaf:l((c,h)=>{let f=new rt(c,s.dtype,"i32"),d=new rt(h,o.dtype,"i32"),m=i!==0?new C("-",f,new L(i)):f,g=a!==0?new C("-",d,new L(a)):d;return new C("*",m,g)},"accLeaf")});return new te([u,p])}),Me("quantized_conv",(r,e,t,n)=>{let s=t[0],o=t[1],i=e.getAttr("input_zero_point")||0,a=e.getAttr("kernel_zero_point")||0;return Hh(r,e,s,o,n[0],{prefix:"qc",blockPrefix:"qconv",initVal:l(()=>new L(0),"initVal"),guardFill:l(()=>new L(0),"guardFill"),leafBuilder:l((u,p)=>{let c=new rt(new q(s,u),s.dtype,"i32"),h=new rt(new q(o,p),o.dtype,"i32"),f=i!==0?new C("-",c,new L(i)):c,d=a!==0?new C("-",h,new L(a)):h;return new C("*",f,d)},"leafBuilder")})})}var Hk=M(()=>{"use strict";De();Fr();Gb();l(Wk,"register")});function ko(r){return Vr.has(r)}function IL(){for(let r of _c())Vr.set(r,(e,t,n)=>Vb(e.opName,t,n));Vr.set("compare",(r,e)=>new qe(r.getAttr("direction")||"eq",e[0],e[1])),Vr.set("select",(r,e)=>new xe(e[0],e[1],e[2])),Vr.set("clamp",(r,e,t)=>new ke("min",[new ke("max",[e[1],e[0]],t),e[2]],t)),Vr.set("convert",(r,e)=>new rt(e[0],r.getOperand(0).type.dtype,r.getAttr("target_dtype")||r.getResult(0).type.dtype)),Vr.set("broadcast_in_dim",(r,e)=>e[0]),Vr.set("broadcast",(r,e)=>e[0]),Vr.set("iota",()=>{throw new Error("iota fusion must be handled by the index-aware path in lowerFusion")}),Vr.set("quantize",(r,e)=>Jh(e[0],{scale:r.getAttr("scale"),zeroPoint:r.getAttr("zero_point"),targetDtype:r.getAttr("target_dtype")||"i8"})),Vr.set("dequantize",(r,e)=>Zh(e[0],{scale:r.getAttr("scale"),zeroPoint:r.getAttr("zero_point"),srcDtype:r.getOperand(0).type?.dtype||"i8",targetDtype:r.getAttr("target_dtype")||"f32"}))}function Yk(r,e){let t=e.numOperands,n=e.numResults,s=new Array(t);for(let N=0;N<t;N++)s[N]=r.getOrAllocBuffer(e.getOperand(N));let o=new Array(n);for(let N=0;N<n;N++)o[N]=r.getOrAllocBuffer(e.getResult(N));let i=o[0],{loopVars:a,loopBinds:u,indices:p,extentNodes:c}=xt(r,i.shape,i),h=new Map,f=e.regions[0].entryBlock,d=new Map,m=[...f.ops()];for(let N=m.length-1;N>=0;N--){let v=m[N];if(lu(v.opName)){let O=v.getAttr("broadcast_dimensions");O&&O.length>0&&d.set(v.getOperand(0),O);continue}if(!(v.opName==="yield"||ln(v.opName)))for(let O=0;O<v.numResults;O++){let R=d.get(v.getResult(O));if(R)for(let B=0;B<v.numOperands;B++)d.has(v.getOperand(B))||d.set(v.getOperand(B),R)}}let g=new Map,y=f.arguments;for(let N=0;N<y.length;N++){let v=d.get(y[N]);v&&g.set(N,v)}for(let N=0;N<f.arguments.length;N++){let v=g.get(N),O;if(v){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[v[B]]}else O=So(s[N],i,p);h.set(f.arguments[N],new q(s[N],O))}let _=new Map;for(let N of f.ops())for(let v=0;v<N.numOperands;v++){let O=N.getOperand(v);_.set(O,(_.get(O)||0)+1)}let b=new Map,w=0,T=[];function S(N){let v=h.get(N);if(v===void 0)throw new Error(`Fusion lowering: unmapped operand from '${N.definingOp?N.definingOp.opName:"unknown"}'`);if((_.get(N)||0)>1&&!EL.has(v.type)){if(!b.has(N)){let O=v.type==="CompareNode"?"i32":v.type==="CastNode"?v.toDtype:N.type&&N.type.dtype?N.type.dtype:i.dtype,R=r.allocVar(`cse${w++}`,O);b.set(N,R),T.push({variable:R,value:v}),h.set(N,R)}return b.get(N)}return v}l(S,"getExpr");let x=[];for(let N of f.ops()){if(N.opName==="yield"){for(let B=0;B<N.numOperands;B++)x.push(new Y(o[B],p,S(N.getOperand(B))));break}if(ln(N.opName)){let B=N.getAttr("value");h.set(N.getResult(0),new fe(typeof B=="number"?B:0));continue}if(N.opName==="iota"){let B=N.getAttr("iota_dimension")??N.getAttr("dimension")??0;h.set(N.getResult(0),p[B]);continue}let v=Vr.get(N.opName);if(!v)throw new Error(`Fusion lowering: unsupported op '${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;h.set(N.getResult(0),v(N,O,R))}let k=x.length===1?x[0]:new te(x);for(let N=T.length-1;N>=0;N--)k=new Ye(T[N].variable,T[N].value,k);let A=new pe(r.blockName("fusion_block"),u,cu(s),cu(o),k);return St(A,a,i.shape,c)}function LL(r,e){if(r.length!==e.length)return!1;for(let t=0;t<r.length;t++)if(r[t]!==e[t])return!1;return!0}function Jk(r){let e=r.regions[0];if(!e)return!1;for(let t of e.entryBlock.ops())if(t.opName!=="yield"){if(ln(t.opName)){if(typeof t.getAttr("value")=="number")continue;return!1}if(!Vr.has(t.opName))return!1}if(r.numResults>1){let t=r.getResult(0).type.shape;for(let n=1;n<r.numResults;n++)if(!LL(r.getResult(n).type.shape,t))return!1}return!0}function Zk(r,e,t){let n=e.regions[0].entryBlock,s=new Map;for(let i=0;i<n.arguments.length;i++)s.set(n.arguments[i],e.getOperand(i));let o=new Map;for(let i of n.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 n.ops()){if(i.opName==="yield")continue;let a=new Array(i.numOperands);for(let f=0;f<i.numOperands;f++)a[f]=s.get(i.getOperand(f))||i.getOperand(f);let u=new Array(a.length);for(let f=0;f<a.length;f++)u[f]=r.getOrAllocBuffer(a[f]);let p=new Array(i.numResults);for(let f=0;f<i.numResults;f++){let d=i.getResult(f),m=o.get(d);if(m){let g=r.getOrAllocBuffer(m);p[f]=g,s.set(d,m)}else{let g={type:d.type};p[f]=r.getOrAllocBuffer(g),s.set(d,g)}}if(ln(i.opName)){t.push(pu(r,i));continue}let c=Qo(i.opName);if(!c)throw new Error(`Fusion lowering: no lowering rule for op '${i.opName}' inside fusion body`);let h=c(r,i,u,p);h&&t.push(h)}}function Qk(){IL()}var Vr,EL,Xk=M(()=>{"use strict";De();Fr();Yh();Gb();Tr();Vr=new Map;l(ko,"canInlineFuse");l(IL,"initBuiltinFusionBuilders");EL=new Set(["BufferLoadNode","VariableNode","IntImmNode","FloatImmNode"]);l(Yk,"lowerFusion");l(LL,"shapesEqual");l(Jk,"canLowerAsElementwiseFusion");l(Zk,"lowerFusionAsIndividualOps");l(Qk,"register")});function ev(r,e,t,n,s){let o=new C("-",new C("*",r,new L(e)),new L(n)),i=new C("+",o,new L(t)),a=new ke("max",[o,new L(0)],"index"),u=new ke("min",[i,new L(s)],"index"),p=new C("-",u,a);return new ke("max",[p,new L(0)],"index")}function CL(r,e,t,n,s,o,i,a,u,p,c){let h=ev(r,t,s,i,u),f=ev(e,n,o,a,p),d=new C("*",h,f);return new rt(d,"index",c)}function tv(){Me("pool2d",(r,e,t,n)=>{let s=t[0],o=n[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,h=Wh(e.getAttr("layout")||"NCHW"),f=h.N,d=h.C,m=h.H,g=h.W,y=s.shape[f],_=s.shape[d],b=s.shape[m],w=s.shape[g],T=o.shape[m],S=o.shape[g],x=a[0],k=a[1],A=u[0],N=u[1],v=p[0][0],O=p[1][0],R=i==="max",B=R?new fe(-1/0):new fe(0),E=or(r,"pi",[0,1,2,3],o.shape,o),z=new Y(o,E.indices,B),D=new pe(r.blockName("pool_init"),E.ivs,[],[{buffer:o}],z),J=E.wrap(D),j=r.allocVar("pn"),U=r.allocVar("pc"),P=r.allocVar("poh"),V=r.allocVar("pow"),he=r.allocVar("pkh"),le=r.allocVar("pkw"),Ie=[j,U,P,V,he,le],G=r.allocBindArray("pv",Ie),K=G[0].iterVar,ie=G[1].iterVar,ze=G[2].iterVar,Se=G[3].iterVar,Qe=G[4].iterVar,It=G[5].iterVar,Jt=new C("+",new C("*",ze,new L(A)),new C("-",Qe,new L(v))),Gt=new C("+",new C("*",Se,new L(N)),new C("-",It,new L(O))),tt=new qe("ge",Jt,new L(0)),ye=new qe("lt",Jt,new L(b)),He=new qe("ge",Gt,new L(0)),gn=new qe("lt",Gt,new L(w)),jt=new C("*",new C("*",tt,ye),new C("*",He,gn)),Zt=new Array(4);Zt[f]=K,Zt[d]=ie,Zt[m]=ze,Zt[g]=Se;let Qt=new Array(4);Qt[f]=K,Qt[d]=ie,Qt[m]=Jt,Qt[g]=Gt;let on=new q(s,Qt),Jn=new q(o,Zt),Bn;if(R){let mt=new xe(jt,on,new fe(-1/0));Bn=new ke("max",[Jn,mt],o.dtype)}else{let mt=new xe(jt,on,new fe(0));Bn=new C("+",Jn,mt)}let yr=new Y(o,Zt,Bn),pt=new pe(r.blockName("pool_acc"),G,[{buffer:s}],[{buffer:o}],yr);pt=new de(le,new L(0),new L(k),F.SERIAL,pt),pt=new de(he,new L(0),new L(x),F.SERIAL,pt),pt=new de(V,new L(0),new L(S),F.SERIAL,pt),pt=new de(P,new L(0),new L(T),F.SERIAL,pt),pt=new de(U,new L(0),new L(_),F.SERIAL,pt),pt=new de(j,new L(0),new L(y),F.SERIAL,pt);let Qn=[J,pt];if(!R){let mt=or(r,"pd",[0,1,2,3],o.shape,o),Or=mt.indices[m],hs=mt.indices[g],$s=c?new fe(x*k):CL(Or,hs,A,N,x,k,v,O,b,w,o.dtype),Br=new q(o,mt.indices),Xn=c?new C("*",Br,new fe(1/(x*k))):new C("/",Br,$s),_r=new Y(o,mt.indices,Xn),fo=new pe(r.blockName("pool_div"),mt.ivs,[{buffer:o}],[{buffer:o}],_r);Qn.push(mt.wrap(fo))}return new te(Qn)})}var nv=M(()=>{"use strict";De();Fr();l(ev,"spatialCount");l(CL,"avgPoolDivisorExpr");l(tv,"register")});function rv(){Me("resize",(r,e,t,n)=>{let s=t[0],o=n[0],i=e.getAttr("method"),a=s.shape[2],u=s.shape[3],p=o.shape[2],c=o.shape[3],h=or(r,"rz",[0,1,2,3],o.shape,o),[f,d,m,g]=h.indices,y=a/p,_=u/c,b;if(i==="nearest"){let S=new ke("floor",[new C("*",m,new fe(y))],"f32"),x=new ke("floor",[new C("*",g,new fe(_))],"f32"),k=new rt(new ke("min",[new ke("max",[S,new fe(0)],"f32"),new fe(a-1)],"f32"),"f32","i32"),A=new rt(new ke("min",[new ke("max",[x,new fe(0)],"f32"),new fe(u-1)],"f32"),"f32","i32");b=new q(s,[f,d,k,A])}else{let S=new C("*",m,new fe(y)),x=new C("*",g,new fe(_)),k=new ke("floor",[S],"f32"),A=new ke("floor",[x],"f32"),N=new ke("min",[new C("+",k,new fe(1)),new fe(a-1)],"f32"),v=new ke("min",[new C("+",A,new fe(1)),new fe(u-1)],"f32"),O=new C("-",S,k),R=new C("-",x,A),B=new ke("max",[k,new fe(0)],"f32"),E=new ke("max",[A,new fe(0)],"f32"),z=new rt(B,"f32","i32"),D=new rt(E,"f32","i32"),J=new rt(N,"f32","i32"),j=new rt(v,"f32","i32"),U=new q(s,[f,d,z,D]),P=new q(s,[f,d,z,j]),V=new q(s,[f,d,J,D]),he=new q(s,[f,d,J,j]),le=new C("-",new fe(1),R),Ie=new C("-",new fe(1),O),G=new C("+",new C("*",U,le),new C("*",P,R)),K=new C("+",new C("*",V,le),new C("*",he,R));b=new C("+",new C("*",G,Ie),new C("*",K,O))}let w=new Y(o,h.indices,b),T=new pe(r.blockName("resize_block"),h.ivs,[{buffer:s}],[{buffer:o}],w);return h.wrap(T)})}var sv=M(()=>{"use strict";De();Fr();l(rv,"register")});function jb(r,e){let[t,n,s]=e;return{Q:t,K:n,V:s,O:null,scale:r.getAttr("scale"),causal:!!r.getAttr("causal"),dtype:t.dtype,B:t.shape[0],H:t.shape[1],Lq:t.shape[2],Dk:t.shape[3],Lk:n.shape[2],Dv:s.shape[3]}}function ov(r,e,t,n){let{Q:s,K:o,V:i,scale:a,causal:u,dtype:p,B:c,H:h,Lq:f,Dk:d,Lk:m,Dv:g}=jb(e,t),y=n[0],_=new me(r.blockName("fa_m"),[1],p,"local"),b=new me(r.blockName("fa_l"),[1],p,"local"),w=new me(r.blockName("fa_s"),[1],p,"local"),T=new me(r.blockName("fa_mn"),[1],p,"local"),S=new me(r.blockName("fa_p"),[1],p,"local"),x=new me(r.blockName("fa_cr"),[1],p,"local"),k=new me(r.blockName("fa_o"),[g],p,"local"),A=new Ne(r.blockName("fa_b"),"int32"),N=new Ne(r.blockName("fa_h"),"int32"),v=new Ne(r.blockName("fa_i"),"i32"),O=new Ne(r.blockName("fa_j"),"int32"),R=new Ne(r.blockName("fa_d"),"int32"),B=new Ne(r.blockName("fa_di"),"int32"),E=new Ne(r.blockName("fa_do"),"int32"),z=new Ne(r.blockName("fa_dw"),"int32"),D=new pe(r.blockName("fa_initm"),[],[],[{buffer:_}],ct(_,[_e],new fe(-1/0))),J=new pe(r.blockName("fa_initl"),[],[],[{buffer:b}],ct(b,[_e],new fe(0))),j=xr(B,g,new pe(r.blockName("fa_inito"),[],[],[{buffer:k}],ct(k,[B],new fe(0)))),U=new pe(r.blockName("fa_sinit"),[],[],[{buffer:w}],ct(w,[_e],new fe(0))),P=xr(R,d,new pe(r.blockName("fa_sred"),[],[{buffer:s},{buffer:o},{buffer:w}],[{buffer:w}],ct(w,[_e],st("+",Ee(w,[_e]),st("*",Ee(s,[A,N,v,R]),Ee(o,[A,N,O,R])))))),V=new te([U,P]),he=new pe(r.blockName("fa_supd"),[],[{buffer:w},{buffer:_},{buffer:b}],[{buffer:w},{buffer:T},{buffer:S},{buffer:x},{buffer:b}],new te([ct(w,[_e],st("*",Ee(w,[_e]),new fe(a))),ct(T,[_e],hu("max",Ee(_,[_e]),Ee(w,[_e]))),ct(S,[_e],hu("exp",st("-",Ee(w,[_e]),Ee(T,[_e])))),ct(x,[_e],hu("exp",st("-",Ee(_,[_e]),Ee(T,[_e])))),ct(b,[_e],st("+",st("*",Ee(b,[_e]),Ee(x,[_e])),Ee(S,[_e])))])),le=xr(E,g,new pe(r.blockName("fa_oupd"),[],[{buffer:k},{buffer:x},{buffer:S},{buffer:i}],[{buffer:k}],ct(k,[E],st("+",st("*",Ee(k,[E]),Ee(x,[_e])),st("*",Ee(S,[_e]),Ee(i,[A,N,O,E])))))),Ie=new pe(r.blockName("fa_mupd"),[],[{buffer:T}],[{buffer:_}],ct(_,[_e],Ee(T,[_e]))),G=new te([V,he,le,Ie]),K=m-f,ie=u?new xe(new qe("le",O,st("+",v,Xo(K))),G):G,ze=xr(O,m,ie),Se=xr(z,g,new pe(r.blockName("fa_finw"),[],[{buffer:k},{buffer:b}],[{buffer:y}],ct(y,[A,N,v,z],st("/",Ee(k,[z]),Ee(b,[_e]))))),Qe=new te([D,J,j,ze,Se]);for(let ye of[k,x,S,T,w,b,_])Qe=new Ve(ye,ye.scope,Qe);let It=Math.min(f,256),Jt=Math.ceil(f/It),Gt;if(Jt===1)Gt=new de(v,_e,new L(f),F.THREAD_BINDING,Qe,"threadIdx.x");else{let ye=new Ne(r.blockName("fa_ii"),"i32"),He=new Ne(r.blockName("fa_io"),"i32"),gn=st("+",st("*",He,new L(It)),ye),jt=Jt*It===f?Qe:new xe(new qe("lt",v,new L(f)),Qe),Zt=new de(ye,_e,new L(It),F.THREAD_BINDING,new Ye(v,gn,jt),"threadIdx.x");Gt=new de(He,_e,new L(Jt),F.THREAD_BINDING,Zt,"blockIdx.z")}let tt=new de(N,_e,new L(h),F.THREAD_BINDING,Gt,"blockIdx.y");return new de(A,_e,new L(c),F.THREAD_BINDING,tt,"blockIdx.x")}function DL(r,e,t,n,s){let{Q:o,K:i,V:a,scale:u,causal:p,dtype:c,B:h,H:f,Lq:d,Dk:m,Lk:g,Dv:y}=jb(e,t),_=n[0],b=Math.ceil(g/s),w=Math.ceil(d/s),T=g-d,S=new me(r.blockName("fa_Ks"),[s,m],c,"shared"),x=new me(r.blockName("fa_Vs"),[s,y],c,"shared"),k=new me(r.blockName("fa_q"),[m],c,"local"),A=new me(r.blockName("fa_m"),[1],c,"local"),N=new me(r.blockName("fa_l"),[1],c,"local"),v=new me(r.blockName("fa_s"),[1],c,"local"),O=new me(r.blockName("fa_mn"),[1],c,"local"),R=new me(r.blockName("fa_p"),[1],c,"local"),B=new me(r.blockName("fa_cr"),[1],c,"local"),E=new me(r.blockName("fa_o"),[y],c,"local"),z=new Ne(r.blockName("fa_b"),"int32"),D=new Ne(r.blockName("fa_h"),"int32"),J=new Ne(r.blockName("fa_qb"),"int32"),j=new Ne(r.blockName("fa_t"),"int32"),U=new Ne(r.blockName("fa_i"),"i32"),P=new Ne(r.blockName("fa_kt"),"int32"),V=new Ne(r.blockName("fa_jj"),"int32"),he=new Ne(r.blockName("fa_dq"),"int32"),le=new Ne(r.blockName("fa_dk"),"int32"),Ie=new Ne(r.blockName("fa_dv"),"int32"),G=new Ne(r.blockName("fa_ds"),"int32"),K=new Ne(r.blockName("fa_doo"),"int32"),ie=new Ne(r.blockName("fa_dw"),"int32"),ze=new Ne(r.blockName("fa_din"),"int32"),Se=st("+",st("*",J,Xo(s)),j),Qe=st("+",st("*",P,Xo(s)),j),It=st("+",st("*",P,Xo(s)),V),Jt=new qe("lt",U,Xo(d)),Gt=new qe("lt",Qe,Xo(g)),tt=xr(he,m,ct(k,[he],Ee(o,[z,D,U,he]))),ye=ct(A,[_e],new fe(-1/0)),He=ct(N,[_e],new fe(0)),gn=xr(ze,y,ct(E,[ze],new fe(0))),jt=new xe(Jt,new te([tt,ye,He,gn])),Zt=new xe(Gt,xr(le,m,ct(S,[j,le],Ee(i,[z,D,Qe,le])))),Qt=new xe(Gt,xr(Ie,y,ct(x,[j,Ie],Ee(a,[z,D,Qe,Ie])))),on=ct(v,[_e],new fe(0)),Jn=xr(G,m,ct(v,[_e],st("+",Ee(v,[_e]),st("*",Ee(k,[G]),Ee(S,[V,G]))))),Bn=new te([ct(v,[_e],st("*",Ee(v,[_e]),new fe(u))),ct(O,[_e],hu("max",Ee(A,[_e]),Ee(v,[_e]))),ct(R,[_e],hu("exp",st("-",Ee(v,[_e]),Ee(O,[_e])))),ct(B,[_e],hu("exp",st("-",Ee(A,[_e]),Ee(O,[_e])))),ct(N,[_e],st("+",st("*",Ee(N,[_e]),Ee(B,[_e])),Ee(R,[_e])))]),yr=xr(K,y,ct(E,[K],st("+",st("*",Ee(E,[K]),Ee(B,[_e])),st("*",Ee(R,[_e]),Ee(x,[V,K]))))),Zn=ct(A,[_e],Ee(O,[_e])),pt=new te([on,Jn,Bn,yr,Zn]),Qn=new qe("lt",It,Xo(g)),mt=p?new xe(new qe("le",It,st("+",U,Xo(T))),pt):pt,Or=xr(V,s,new xe(Qn,mt)),hs=new xe(Jt,Or),$s=new te([Zt,Qt,new On,hs,new On]),Br=xr(P,b,$s),Xn=new xe(Jt,xr(ie,y,ct(_,[z,D,U,ie],st("/",Ee(E,[ie]),Ee(N,[_e]))))),_r=new Ye(U,Se,new te([jt,Br,Xn]));for(let Qr of[E,B,R,O,v,N,A,k])_r=new Ve(Qr,Qr.scope,_r);let fo=Qh(j,"threadIdx.x",s,_r),Rr=new Ve(S,"shared",new Ve(x,"shared",fo)),Mr=Qh(J,"blockIdx.z",w,Rr),ds=Qh(D,"blockIdx.y",f,Mr);return Qh(z,"blockIdx.x",h,ds)}function iv(){Me("scaled_dot_product_attention",ov),Ik("scaled_dot_product_attention","cuda",(r,e,t,n)=>{let{Lq:s,Dk:o,Dv:i,dtype:a}=jb(e,t),u=r.target,p=u&&u.sharedMemoryBytes||0,c=u&&u.maxThreadsPerBlock||1024,h=Math.min(c,s,Math.floor(p/((o+i)*bn(a)*2)));return h<8?ov(r,e,t,n):DL(r,e,t,n,h)})}var _e,Xo,Ee,ct,st,hu,xr,Qh,av=M(()=>{"use strict";De();xo();Fr();Vn();_e=new L(0),Xo=l(r=>new L(r),"I"),Ee=l((r,e)=>new q(r,e),"ld"),ct=l((r,e,t)=>new Y(r,e,t),"st"),st=l((r,e,t)=>new C(r,e,t),"mop"),hu=l((r,...e)=>new ke(r,e),"ext"),xr=l((r,e,t,n=F.SERIAL)=>new de(r,_e,new L(e),n,t),"forL"),Qh=l((r,e,t,n)=>new de(r,_e,new L(t),F.THREAD_BINDING,n,e),"forT");l(jb,"dims");l(ov,"buildNaive");l(DL,"buildTiled");l(iv,"register")});function uv(r,e,t=new Set){let n=ne.get(e.opName);if(!n||!n.getAttr("broadcastViewSafe"))return!1;if(e.opName!=="fusion")return!0;let s=e.regions[0];if(!s)return!1;let o=s.entryBlock;for(let i=0;i<e.numOperands;i++){if(e.getOperand(i)!==r)continue;let a=o.arguments[i];for(let u of a.getUsers()){let p=`${a.id}:${u.id}`;if(!t.has(p)&&(t.add(p),!uv(a,u,t)))return!1}}return!0}function $L(r){return zh([...r.ops()],"ignore")}function Xh(r,e=null,t=null){let n=new Kh;n.target=e;let s=[],o=new Map;for(let g of r.args){let y=n.allocVar("arg");s.push(y),o.set(y,n.getOrAllocBuffer(g))}let i=r.getReturnOp(),a=new Set;for(let g of r.args)a.add(n.getOrAllocBuffer(g));let u=[],p=new Set;for(let g=0;g<i.numOperands;g++){let y=n.allocVar("ret");s.push(y);let _=n.getOrAllocBuffer(i.getOperand(g));if(a.has(_)||p.has(_)){let b=n.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 h=[];for(let g of r.ops())ln(g.opName)&&h.push(pu(n,g));for(let g of $L(r)){if(qt(g.opName)||ln(g.opName))continue;if(g.opName==="fusion"){Jk(g)?h.push(Yk(n,g)):Zk(n,g,h);continue}if(lu(g.opName)&&!c.has(g.getResult(0))&&g.getOperand(0).getUsers().length===1&&g.getResult(0).getUsers().every(T=>uv(g.getResult(0),T))){let T=n.getOrAllocBuffer(g.getOperand(0)),S=g.getResult(0).type.shape,x=g.getAttr("broadcast_dimensions"),k;if(x&&x.length>0)k=x;else{let A=S.length-T.shape.length;k=Array.from({length:T.shape.length},(N,v)=>v+A)}T.broadcastDims=k,n.bufferMap.set(g.getResult(0),T);continue}let y=Qo(g.opName,e,t);if(!y)throw new Error(`No lowering rule defined for op: ${g.opName}`);let _=new Array(g.numOperands);for(let T=0;T<g.numOperands;T++)_[T]=n.getOrAllocBuffer(g.getOperand(T));let b=new Array(g.numResults);for(let T=0;T<g.numResults;T++)b[T]=n.getOrAllocBuffer(g.getResult(T));let w=y(n,g,_,b);w&&h.push(w)}for(let{src:g,dst:y}of u){let{loopVars:_,loopBinds:b,indices:w,extentNodes:T}=xt(n,g.shape,g),S=new q(g,w),x=new Y(y,w,S),k=new pe(n.blockName("copy_block"),b,[{buffer:g}],[{buffer:y}],x);h.push(St(k,_,g.shape,T))}let f=[],d=new Set;for(let g of n.shapeParams.values())d.has(g.name)||(d.add(g.name),f.push(g));for(let[g,y]of n.symVars)if(!d.has(y.name))throw new Error(`Symbolic dimension '${g}' has no input dimension to bind it to at runtime`);for(let g of f)s.push(g);let m=new Jo(r.name,s,h.length===1?h[0]:new te(h),o,f,new Map(n.shapeParams));return r._partitionTarget&&(m._partitionTarget=r._partitionTarget),m}var PL,Ui=M(()=>{"use strict";De();ji();At();Fr();Tr();Yh();Pk();Fk();Gk();qk();Kk();Hk();Xk();nv();sv();av();Yh();PL=["compare","select","clamp","convert","copy_to_device","dot","fusion"];for(let r of[..._c(),...PL])ne.has(r)&&ne.registerOpAttr(r,"broadcastViewSafe",!0);l(uv,"broadcastViewSafeForUser");Lk();Dk();$k();Vk();jk();Uk();Wk();Qk();tv();rv();iv();l($L,"topologicalOps");l(Xh,"lowerGraphToPrimFunc")});var lv={};nt(lv,{CPUTarget:()=>Us,CUDATarget:()=>ed,TargetFeatures:()=>Ki,TargetKind:()=>Ot,WasmTarget:()=>td,WebGPUTarget:()=>nd});var Ot,Ki,Us,ed,td,nd,Ks=M(()=>{"use strict";Ot=Object.freeze({CPU:"cpu",CUDA:"cuda",WEBGPU:"webgpu",WASM:"wasm",ACCELERATOR:"accelerator"}),Ki=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===Ot.CUDA||this.kind===Ot.WEBGPU}isWebGPU(){return this.kind===Ot.WEBGPU}isCPU(){return this.kind===Ot.CPU}isWasm(){return this.kind===Ot.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)}},Us=l((r={})=>new Ki({kind:Ot.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,...r}),"CPUTarget"),ed=l((r={})=>new Ki({kind:Ot.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,...r}),"CUDATarget"),td=l((r={})=>new Ki({kind:Ot.WASM,name:"wasm_generic",vectorWidth:4,numCores:1,cacheLineSizeBytes:64,memoryBandwidthGBs:10,computeTFLOPs:.1,supportsInt8:!0,simd:!0,...r}),"WasmTarget"),nd=l((r={})=>new Ki({kind:Ot.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,...r}),"WebGPUTarget")});function du(r,e,t,n,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=r.strides[i];u===1?o.push(a):typeof u=="number"&&u>=0?o.push(`${a} * ${u}`):o.push(`${a} * ${n(r,i)}`)}return o.length===0?"0":o.join(" + ")}var rd=M(()=>{"use strict";l(du,"flattenRowMajorIndex")});function ei(r){return zL[r]||"f32"}function ir(r){if(!r)return"f32";if(r._dtype)return r._dtype;switch(r.type){case"IntImmNode":return"i32";case"FloatImmNode":return"f32";case"LIRFlatLoadNode":return ei(r.dtype);case"BufferLoadNode":return ei(r.buffer.dtype);case"CastNode":return ei(r.toDtype);case"CallExternNode":return ei(r.dtype||"f32");case"CompareNode":return"i32";case"VariableNode":return ei(r.dtype||"i32");case"MathOpNode":{let e=ir(r.a);if(rr(e))return e;if(r.b){let t=ir(r.b);if(rr(t))return t}return e}case"IfThenElseNode":return ir(r.thenBody);default:return"f32"}}function No(r){!r||typeof r!="object"||!r.type||(r._dtype=ir(r))}function cv(r){return FL.has(r)}var vo,sd,wc,Tc,od,mu,id,FL,zL,xc=M(()=>{"use strict";Vn();vo=class r{static{l(this,"LIRNode")}constructor(){this.type=this.constructor.name.replace(/^_+/,""),this._parent=null,this._parentKey=null,this._parentIdx=-1}_setChild(e,t,n=-1){(t instanceof r||t&&t._parent!==void 0)&&(t._parent=this,t._parentKey=e,t._parentIdx=n)}_setChildren(e,t){if(t)for(let n=0;n<t.length;n++){let s=t[n];(s instanceof r||s&&s._parent!==void 0)&&(s._parent=this,s._parentKey=e,s._parentIdx=n)}}replaceWith(e){let t=this._parent;return t?(this._parentIdx>=0?t[this._parentKey][this._parentIdx]=e:t[this._parentKey]=e,e&&(e instanceof r||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}},sd=class extends vo{static{l(this,"LIRFunc")}constructor(e,t,n,s,o,i,a){super(),this.name=e,this.params=t,this.body=n,this.bufferMap=s,this.shapeParams=o,this.shapeParamMap=i,this.metadata=a,this._setChild("body",n)}},wc=class extends vo{static{l(this,"LIRFlatLoadNode")}constructor(e,t,n){super(),this.buffer=e,this.offsetExpr=t,this.dtype=n,this._setChild("offsetExpr",t)}},Tc=class extends vo{static{l(this,"LIRFlatStoreNode")}constructor(e,t,n,s){super(),this.buffer=e,this.offsetExpr=t,this.value=n,this.dtype=s,this._setChild("offsetExpr",t),this._setChild("value",n)}},od=class extends vo{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)}},mu=class extends vo{static{l(this,"LIRBindingsNode")}constructor(e,t){super(),this.bindings=e,this.body=t,this._setChild("body",t)}},id=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}},FL=new Set(["sqrt","abs","ceil","floor","min","max"]),zL={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(ei,"normalizeDtype");l(ir,"inferDtype");l(No,"annotateDtype");l(cv,"isWasmNativeOp")});function Gn(r){return r instanceof Tt||r instanceof vo}function ve(r,e,t={}){return{key:r,kind:e,array:!!t.array,poly:!!t.poly,bind:!!t.bind,param:!!t.param,region:!!t.region,iterVarBinding:!!t.iterVarBinding,bindingsExpr:!!t.bindingsExpr}}function Ub(r){return{kinds:r.kinds||"both",descendParams:r.descendParams===!0,bindVars:r.bindVars!==!1}}function VL(r,e){return e.iterVarBinding?{read(){return(r.iterVars||[]).map(t=>t&&t.binding).filter(Gn)},write(t){let n=0;for(let s of r.iterVars||[])s&&Gn(s.binding)&&(s.binding=t[n++])}}:e.bindingsExpr?{read(){return(r.bindings||[]).map(t=>t&&t.expr).filter(Gn)},write(t){let n=0;for(let s of r.bindings||[])s&&Gn(s.expr)&&(s.expr=t[n++])}}:e.poly?{read(){let t=r[e.key];return Array.isArray(t)?t.filter(Gn):Gn(t)?[t]:[]},write(t){let n=r[e.key];if(Array.isArray(n)){let s=0;r[e.key]=n.map(o=>Gn(o)?t[s++]:o)}else t.length&&(r[e.key]=t[0])}}:e.array?{read(){let t=r[e.key];return Array.isArray(t)?t.filter(Gn):[]},write(t){let n=r[e.key];if(!Array.isArray(n))return;let s=0;r[e.key]=n.map(o=>Gn(o)?t[s++]:o)}}:{read(){return Gn(r[e.key])?[r[e.key]]:[]},write(t){t.length&&Gn(r[e.key])&&(r[e.key]=t[0])}}}function Kb(r,e={}){let t=Ub(e),n=qb[r.type];if(n===void 0)throw new Error(`ir_visitor: no child schema for node type '${r.type}'`);let s=[];for(let o of n)o.region||o.param&&!t.descendParams||o.bind&&!t.bindVars||t.kinds!=="both"&&o.kind!==t.kinds||s.push(VL(r,o));return s}function en(r,e={}){if(!Gn(r)||qb[r.type]===void 0)return[];let t=[];for(let n of Kb(r,e)){let s=n.read();for(let o=0;o<s.length;o++)t.push(s[o])}return t}function pv(r,e,t,n){r._parent=e,r._parentKey=t,r._parentIdx=n}function GL(r){let e=qb[r.type];if(e)for(let t of e){if(t.iterVarBinding||t.bindingsExpr||t.region)continue;let n=r[t.key];if(Array.isArray(n))for(let s=0;s<n.length;s++)Gn(n[s])&&pv(n[s],r,t.key,s);else Gn(n)&&pv(n,r,t.key,-1)}}function fv(r,e,t,n,s,o){if(o.stop||!Gn(r))return;let i={parent:n,depth:s},a=!0;if(e.pre){let u=e.pre(r,i);if(u===ad){o.stop=!0;return}u===!1&&(a=!1)}if(a)for(let u of Kb(r,t)){let p=u.read();for(let c=0;c<p.length;c++)if(fv(p[c],e,t,r,s+1,o),o.stop)return}e.post&&e.post(r,i)}function Ss(r,e,t={}){fv(r,typeof e=="function"?{pre:e}:e||{},Ub(t),null,0,{stop:!1})}function gu(r,e,t={}){let n=[];return Ss(r,s=>{e(s)&&n.push(s)},t),n}function Sc(r,e,t={}){let n=!1;return Ss(r,s=>{if(e(s))return n=!0,ad},t),n}function ti(r,e,t={}){let n=Ub(t);return hv(r,e,n)}function hv(r,e,t){if(!Gn(r))return r;let n=!1;for(let o of Kb(r,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=hv(i[p],e,t);c!==i[p]&&(a=!0),u[p]=c}a&&(o.write(u),n=!0)}n&&GL(r);let s=e(r);return s??r}var ad,qb,Gr=M(()=>{"use strict";De();xc();ad=Symbol("ir-visitor-stop");l(Gn,"isIRNode");l(ve,"field");qb={VariableNode:[],IntImmNode:[],FloatImmNode:[],SyncThreadsNode:[],MathOpNode:[ve("a","expr"),ve("b","expr")],CompareNode:[ve("a","expr"),ve("b","expr")],CastNode:[ve("expr","expr")],CallExternNode:[ve("args","expr",{array:!0})],BufferLoadNode:[ve("indices","expr",{array:!0})],IfThenElseNode:[ve("condition","expr"),ve("thenBody","stmt"),ve("elseBody","stmt")],BlockRealizeNode:[ve("binding","expr",{poly:!0}),ve("iterVar","expr",{bind:!0})],LIRFlatLoadNode:[ve("offsetExpr","expr")],PrimFunc:[ve("params","expr",{array:!0,param:!0}),ve("body","stmt")],ForNode:[ve("min","expr"),ve("extent","expr"),ve("loopVar","expr",{bind:!0}),ve("body","stmt")],BlockNode:[ve("iterVars","expr",{iterVarBinding:!0}),ve("reads","expr",{region:!0,array:!0}),ve("writes","expr",{region:!0,array:!0}),ve("initBody","stmt"),ve("body","stmt")],SeqNode:[ve("stmts","stmt",{array:!0})],VecCopyNode:[ve("dstIndex","expr"),ve("srcIndex","expr")],LetStmtNode:[ve("value","expr"),ve("variable","expr",{bind:!0}),ve("body","stmt")],AllocateNode:[ve("body","stmt")],WhileNode:[ve("condBody","stmt"),ve("condVar","expr",{bind:!0}),ve("loopBody","stmt")],EvaluateNode:[ve("value","expr")],BufferStoreNode:[ve("indices","expr",{array:!0}),ve("value","expr")],LIRFunc:[ve("body","stmt")],LIRFlatStoreNode:[ve("offsetExpr","expr"),ve("value","expr")],LIRAccumulatorNode:[ve("loopVar","expr",{bind:!0}),ve("extent","expr"),ve("initLoad","stmt"),ve("initBody","stmt"),ve("body","expr"),ve("flushStore","stmt")],LIRBindingsNode:[ve("bindings","expr",{bindingsExpr:!0}),ve("body","stmt")]};l(Ub,"normOpts");l(VL,"buildAccessor");l(Kb,"childAccessors");l(en,"irChildNodes");l(pv,"setParent");l(GL,"relink");l(fv,"walkInner");l(Ss,"walk");l(gu,"collect");l(Sc,"some");l(ti,"transform");l(hv,"transformInner")});function jL(r,e,t,n){switch(r){case"add":return`(${e} + ${t})`;case"sub":return`(${e} - ${t})`;case"mul":return`(${e} * ${t})`;case"neg":return`(-${e})`;case"div":return n==="js"?`((${e} / ${t}) | 0)`:`(${e} / ${t})`;case"mod":return n==="js"?`((${e} % ${t} + ${t}) % ${t})`:`(${e} % ${t})`;case"ceildiv":return n==="js"?`(((${e} + ${t} - 1) / ${t}) | 0)`:`((${e} + ${t} - 1) / ${t})`;case"max":return n==="js"?`Math.max(${e}, ${t})`:`max(${e}, ${t})`;case"min":return n==="js"?`Math.min(${e}, ${t})`:`min(${e}, ${t})`;default:throw new Error(`emitSymInt: unsupported op '${r}'`)}}function Wb(r,e,t="c"){if(typeof r=="number"||!(r instanceof Ce))return String(r);if(r.type==="var")return e({name:qa(r.name)});if(t==="wat")throw new Error("emitSymInt: compound symbolic expressions are not supported on the WASM backend");let n=Wb(r.args[0],e,t),s=r.args.length>1?Wb(r.args[1],e,t):null;return jL(r.type,n,s,t)}function Wi(r){let e=r.indexOf(".");if(e<0)return null;let t=r.charCodeAt(e+1)-120;if(t<0||t>2)return null;let n=r.substring(0,e);return n==="threadIdx"?{space:"thread",axis:t}:n==="blockIdx"?{space:"block",axis:t}:null}function ud(r,e){let t=[e];for(;t.length>0;){let n=t.pop();if(n)switch(n.type){case"SeqNode":for(let s=n.stmts.length-1;s>=0;s--)t.push(n.stmts[s]);continue;case"AllocateNode":r._visitAllocateNode(n),t.push(n.body);continue;case"ForNode":r._visitForNode(n);continue;case"BlockNode":r._visitBlockNode(n);continue;case"IfThenElseNode":r._visitIfStmt(n);continue;case"LetStmtNode":r._visitLetStmtNode(n);continue;case"BufferStoreNode":r._visitBufferStoreNode(n);continue;case"LIRFlatStoreNode":r._visitLIRFlatStore(n);continue;case"LIRBindingsNode":r._visitLIRBindings(n);continue;case"LIRAccumulatorNode":r._visitLIRAccumulator(n);continue;case"WhileNode":r._visitWhileNode(n);continue;case"SyncThreadsNode":r._emitSync();continue;case"VecCopyNode":r._visitVecCopyNode(n);continue;case"EvaluateNode":continue;default:throw new Error(`${r.constructor.name}: unhandled statement node '${n.type}'`)}}}function yu(r,e){let t=[r];for(;t.length>0;){let n=t.pop();if(!(!n||typeof n!="object")&&e(n)!==!1)for(let s of qL){let o=n[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 ld(r){let e=r;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 ni(r,e,t,n,s,o="c"){let i=e.shape[t];if(i instanceof Ce)return Wb(i,n,o);if(r&&r.shapeParamMap){let a=r.shapeParamMap.get(`${e.name}:${t}`);if(a)return n(a)}throw new Error(`${s} codegen: missing shape param for ${e.name}:${t}`)}function cd(r){let e=1;for(let t of r.shape)typeof t=="number"&&t>0&&(e*=t);return e}function ri(r,e,t){let n=[];for(let s=e;s<r.shape.length;s++){let o=r.shape[s];typeof o=="number"&&o>=0?n.push(String(o)):n.push(t(r,s))}return n.length===0?"1":n.join(" * ")}function pd(r,e){let t=r.get(e);if(!t)return 0;let n=0;for(let s of t)s.extent>n&&(n=s.extent);return n}var qL,kc=M(()=>{"use strict";Vs();l(jL,"symOpToString");l(Wb,"emitSymInt");l(Wi,"parseThreadAxis");l(ud,"visitStatements");qL=["body","stmts","thenBody","elseBody","loopBody","condBody","initBody"];l(yu,"walkStmtTree");l(ld,"isZeroFillBody");l(ni,"resolveShapeParam");l(cd,"estimateBufferSize");l(ri,"dynamicDimProduct");l(pd,"maxBindingExtent")});function bu(r){let e=r>=0?1:-1,t=Math.abs(r),n=1/(1+vc*t),s=0;for(let o=_u.length-1;o>=0;o--)s=s*n+_u[o];return e*(1-s*n*Math.exp(-t*t))}function wu(r){return 1-bu(r)}function Tn(r){if(r<.5)return Math.log(Math.PI/Math.abs(Math.sin(Math.PI*r)))-Tn(1-r);let e=r-1,t=Hi[0];for(let s=1;s<Hi.length;s++)t+=Hi[s]/(e+s);let n=e+7+.5;return .5*Math.log(2*Math.PI)+(e+.5)*Math.log(n)-n+Math.log(t)}function Tu(r){return r<.5?Math.PI/(Math.sin(Math.PI*r)*Tu(1-r)):Math.exp(Tn(r))}function Ac(r){let e=0,t=r;for(;t<Yi;)e-=1/t,t+=1;let n=1/t,s=n*n;e+=Math.log(t)-.5*n;let o=s;for(let i of Nc)e+=i*o,o*=s;return e}var Hi,_u,vc,Yi,Nc,Ji=M(()=>{"use strict";Hi=[.9999999999998099,676.5203681218851,-1259.1392167224028,771.3234287776531,-176.6150291621406,12.507343278686905,-.13857109526572012,9984369578019572e-21,15056327351493116e-23],_u=[.254829592,-.284496736,1.421413741,-1.453152027,1.061405429],vc=.3275911,Yi=6,Nc=[-1/12,1/120,-1/252,1/240,-1/132];l(bu,"erfScalar");l(wu,"erfcScalar");l(Tn,"lgammaScalar");l(Tu,"gammaScalar");l(Ac,"digammaScalar")});function gv(r){return Ao[0]=(r&32767)<<13,si[0]*=mv,si[0]>=65536&&(Ao[0]|=2139095040),Ao[0]|=(r&32768)<<16,si[0]}function yv(r){si[0]=r;let e=Ao[0]&2147483648,t=(Ao[0]^e)>>>0,n;if(t>=1199570944)n=t>2139095040?32256:31744;else if(t<947912704)Ao[0]=t,si[0]+=.5,n=Ao[0]-1056964608&65535;else{let s=t>>13&1;t=t+dv>>>0,t=t+s>>>0,n=t>>13&65535}return(n|e>>>16)&65535}function fd(r){return Ao[0]=r<<16>>>0,si[0]}function hd(r){si[0]=r;let e=Ao[0];if((e>>>23&255)===255)return e>>>16&65535;let t=32767+(e>>>16&1)>>>0;return e+t>>>0>>>16&65535}function dd(r,e){return r==="f16"?yv(Number(e)):r==="bf16"?hd(Number(e)):r==="i64"?typeof e=="bigint"?e:BigInt(Math.trunc(e)):e}function md(r,e){return r==="f16"?gv(Number(e)):r==="bf16"?fd(Number(e)):r==="i64"&&typeof e=="bigint"?Number(e):e}var si,Ao,dv,mv,Hb,gd,xu=M(()=>{"use strict";si=new Float32Array(1),Ao=new Uint32Array(si.buffer),dv=(-112<<23>>>0)+4095>>>0,mv=5192296858534828e18;l(gv,"f16ToF32");l(yv,"f32ToF16");l(fd,"bf16ToF32");l(hd,"f32ToBf16");l(dd,"coerceForStorage");l(md,"readFromStorage");Hb={F16_ADD_BIAS:dv|0,F16_MAGIC_MUL:mv},gd=globalThis;gd.__mlfw_f16_to_f32=gv;gd.__mlfw_f32_to_f16=yv;gd.__mlfw_bf16_to_f32=fd;gd.__mlfw_f32_to_bf16=hd});function bv(r){return`((x_erf => { ${KL} })(${r}))`}function HL(r){return`((x_lg => { const lg = ${wv}; return x_lg < 0.5 ? Math.log(Math.PI / Math.abs(Math.sin(Math.PI * x_lg))) - lg(1 - x_lg) : lg(x_lg); })(${r}))`}function YL(r){return`((x_g => { const lg = ${wv}; return x_g < 0.5 ? Math.PI / (Math.sin(Math.PI * x_g) * Math.exp(lg(1 - x_g))) : Math.exp(lg(x_g)); })(${r}))`}var Yb,_v,UL,KL,WL,wv,yd,Tv=M(()=>{"use strict";De();Vn();rd();Gr();kc();Ji();xu();Yb={f16:"__mlfw_f16_to_f32",bf16:"__mlfw_bf16_to_f32"},_v={f16:"__mlfw_f32_to_f16",bf16:"__mlfw_f32_to_bf16"},UL=_u.slice().reverse().reduce((r,e)=>`(${e} + t * ${r})`,"0"),KL=`const t = 1.0 / (1.0 + ${vc} * Math.abs(x_erf)); const p = t * ${UL}; return (x_erf >= 0 ? 1 : -1) * (1.0 - p * Math.exp(-x_erf * x_erf));`;l(bv,"_erfExpr");WL=Hi.map((r,e)=>e===0?`${r}`:`${r} / (zz + ${e})`).join(" + "),wv=`(zg => { const zz = zg - 1; const s = ${WL}; const t = zz + ${7} + 0.5; return ${.5*Math.log(2*Math.PI)} + (zz + 0.5) * Math.log(t) - t + Math.log(s); })`;l(HL,"_lgammaExpr");l(YL,"_gammaExpr");yd=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",n=new Set,s=[];for(let[,c]of e.bufferMap)s.push(c.name),n.add(c.name);for(let c of e.shapeParams)s.push(c.name);this._paramBuffers=n,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,n),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 h=c.numel();h>0&&(p=Math.max(p,c.poolByteOffset+h*bn(c.dtype)))}p>0&&this._emit(`const _mem_pool = new ArrayBuffer(${p});`);for(let[c,h]of o)if(!a.has(c)&&!u.has(c)&&!n.has(c)&&!i.has(c)){let f=h.numel();if(f>0)this._emit(`const ${c} = ${this._allocRhs(h,f)};`);else if(f<0){let d=[];for(let m=0;m<h.shape.length;m++)d.push(typeof h.shape[m]=="number"&&h.shape[m]>=0?String(h.shape[m]):this._resolveShapeParam(h,m));this._emit(`const ${c} = new ${mh(h.dtype)}(${d.join(" * ")});`)}}for(let[,c]of o)n.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 n=mh(e.dtype);return e.poolByteOffset!==void 0&&e.poolByteOffset!==null&&t>0?`new ${n}(_mem_pool, ${e.poolByteOffset}, ${t})`:`new ${n}(${t})`}_wrapLoad(e,t){let n=Yb[e];return n?`${n}(${t})`:t}_wrapStoreVal(e,t){let n=_v[e];return n?`${n}(${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 n=0;n<t.stmts.length-1;n++)this._visitNode(t.stmts[n]);t=t.stmts[t.stmts.length-1];continue;case"AllocateNode":{let n=t.buffer,s=n.numel();s>0?this._emit(`const ${n.name} = ${this._allocRhs(n,s)};`):s<0&&this._emit(`const ${n.name} = new ${mh(n.dtype)}(${this._dynamicNumel(n)});`),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 n=this._exprToJS(e.extent);if(e.kind===F.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+" < "+n+"; "+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+" < "+n+"; "+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 n=this._exprToJS(t.expr);this._aliases.set(t.name,n)}this._visitNode(e.body)}_visitLIRAccumulator(e){let t=e.localName,n=this._exprToJS(e.initLoad);this._emit("let "+t+" = "+n+";");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 n=t.body;if(!n||n.type!=="BufferStoreNode")return null;let s=n,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 n=this._exprToJS(t.binding);this._aliases.set(t.iterVar.name,n)}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=[],n=[{node:e,phase:0}];for(;n.length>0;){let s=n[n.length-1],o=s.node;if(!o){n.pop(),t.push("0");continue}switch(o.type){case"IntImmNode":n.pop(),t.push(String(o.value));continue;case"FloatImmNode":n.pop(),t.push(String(o.value));continue;case"VariableNode":n.pop(),t.push(this._aliases.get(o.name)||o.name);continue;case"BufferLoadNode":{if(n.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(${i})`:i)}else this._accTarget&&o.buffer.name===this._accTarget.bufName&&this._flatIndex(o.buffer,o.indices)===this._accTarget.idxExpr?t.push(this._accVar):t.push(this._wrapLoad(o.buffer.dtype,o.buffer.name+"["+this._flatIndex(o.buffer,o.indices)+"]"));continue}case"LIRFlatLoadNode":{if(n.pop(),this._zeroBuffers&&this._zeroBuffers.has(o.buffer.name))t.push(this._zeroLit(o.dtype||o.buffer.dtype));else if(this._constantBuffers&&this._constantBuffers.has(o.buffer.name)){let i=String(this._constantBuffers.get(o.buffer.name));t.push((o.dtype||o.buffer.dtype)==="i64"?`BigInt(${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,n.push({node:o.a,phase:0});else if(s.phase===1&&o.b)s.phase=2,n.push({node:o.b,phase:0});else if(n.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} ${i})`)}continue;case"CompareNode":if(s.phase===0)s.phase=1,n.push({node:o.a,phase:0});else if(s.phase===1)s.phase=2,n.push({node:o.b,phase:0});else{n.pop();let i=t.pop(),a=t.pop();t.push(`(${a} ${zS(o.direction)} ${i})`)}continue;case"IfThenElseNode":if(s.phase===0)s.phase=1,n.push({node:o.condition,phase:0});else if(s.phase===1)s.phase=2,n.push({node:o.thenBody,phase:0});else if(s.phase===2)s.phase=3,n.push({node:o.elseBody,phase:0});else{n.pop();let i=t.pop(),a=t.pop(),u=t.pop();t.push(`(${u} ? ${a} : ${i})`)}continue;case"CastNode":if(s.phase===0)s.phase=1,n.push({node:o.expr,phase:0});else{n.pop();let i=t.pop();o.toDtype==="bool"?t.push(`(${i} ? 1 : 0)`):nr(o.toDtype)?t.push(`(${i} | 0)`):Yb[o.toDtype]?t.push(`${Yb[o.toDtype]}(${_v[o.toDtype]}(${i}))`):t.push(`(+${i})`)}continue;case"CallExternNode":if(s.phase<o.args.length){let i=s.phase;s.phase++,n.push({node:o.args[i],phase:0})}else{n.pop();let i=[];for(let u=0;u<o.args.length;u++)i.unshift(t.pop());let a=i.join(", ");if(VS(o.externName))t.push(`Math.${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(bv(a));else if(o.externName==="erfc")t.push(`(1.0 - ${bv(a)})`);else if(o.externName==="lgamma")t.push(HL(a));else if(o.externName==="gamma")t.push(YL(a));else if(o.externName==="log10")t.push(`(Math.log(${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 ri(e,0,(t,n)=>this._resolveShapeParam(t,n))}_flatIndex(e,t){return du(e,t,n=>this._exprToJS(n),(n,s)=>this._computeDynamicStride(n,s),!0)}_computeDynamicStride(e,t){return ri(e,t+1,(n,s)=>this._resolveShapeParam(n,s))}_resolveShapeParam(e,t){return ni(this._primFunc,e,t,n=>n.name,"CPU","js")}_cleanupSource(e){let t=e.split(`
+`),n=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*$/);a&&(n[i]=a[1]);let u=t[i].match(/[A-Za-z_]\w*/g);if(u)for(let p of u)s.set(p,(s.get(p)||0)+1)}let o=[];for(let i=0;i<t.length;i++){if(n[i]!==null&&s.get(n[i])===1)continue;let a=t[i];if(/^\s*\}\s*$/.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 n=t.value;if(!(n.type==="FloatImmNode"&&n.value===0||n.type==="IntImmNode"&&n.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 ld(e)}_scanTree(e,t,n,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&&n.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 en(i))o.push(a)}}}_findZeroOnlyBuffers(e,t){let n=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)||(n.has(a)||n.set(a,[]),n.get(a).push(i.value))}for(let a of en(i))s.push(a)}}let o=new Set;this._constantBuffers=new Map;for(let[i,a]of n){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 Sv(r,e){xv.set(r,e)}function kv(r){return xv.get(r)||null}var xv,vv=M(()=>{"use strict";xv=new Map;l(Sv,"registerCudaIntrin");l(kv,"getCudaIntrin");Sv("wmma_16x16x16_f16f16f32",(r,e)=>{let{M:t,N:n,K:s,a:o,b:i,c:a}=e;r._blockDim=[32,1,1],r._gridDim=[Math.ceil(t/16),Math.ceil(n/16),1],r._emit("const int warpM = blockIdx.x;"),r._emit("const int warpN = blockIdx.y;"),r._emit("fragment<accumulator, 16, 16, 16, float> cf;"),r._emit("fill_fragment(cf, 0.0f);"),r._emit(`for (int kk = 0; kk < ${s}; kk += 16) {`),r._indent++,r._emit("fragment<matrix_a, 16, 16, 16, half, row_major> af;"),r._emit("fragment<matrix_b, 16, 16, 16, half, row_major> bf;"),r._emit(`load_matrix_sync(af, ${o} + warpM * 16 * ${s} + kk, ${s});`),r._emit(`load_matrix_sync(bf, ${i} + kk * ${n} + warpN * 16, ${n});`),r._emit("mma_sync(cf, af, bf, cf);"),r._indent--,r._emit("}"),r._emit(`store_matrix_sync(${a} + warpM * 16 * ${n} + warpN * 16, cf, ${n}, mem_row_major);`)});Sv("gemm_pipelined_f32",(r,e)=>{let{M:t,N:n,K:s,a:o,b:i,c:a,tile:u=16}=e;r._blockDim=[u,u,1],r._gridDim=[Math.ceil(n/u),Math.ceil(t/u),1];let p=u;r._emit(`__shared__ float As[2][${p}][${p}];`),r._emit(`__shared__ float Bs[2][${p}][${p}];`),r._emit(`const int row = blockIdx.y * ${p} + threadIdx.y;`),r._emit(`const int col = blockIdx.x * ${p} + threadIdx.x;`),r._emit("float acc = 0.0f;"),r._emit(`const int nTiles = ${s} / ${p};`),r._emit(`__pipeline_memcpy_async(&As[0][threadIdx.y][threadIdx.x], &${o}[row * ${s} + threadIdx.x], sizeof(float));`),r._emit(`__pipeline_memcpy_async(&Bs[0][threadIdx.y][threadIdx.x], &${i}[threadIdx.y * ${n} + col], sizeof(float));`),r._emit("__pipeline_commit();"),r._emit("for (int t = 0; t < nTiles; t++) {"),r._indent++,r._emit("const int cur = t & 1, nxt = (t + 1) & 1;"),r._emit("if (t + 1 < nTiles) {"),r._indent++,r._emit(`__pipeline_memcpy_async(&As[nxt][threadIdx.y][threadIdx.x], &${o}[row * ${s} + (t + 1) * ${p} + threadIdx.x], sizeof(float));`),r._emit(`__pipeline_memcpy_async(&Bs[nxt][threadIdx.y][threadIdx.x], &${i}[((t + 1) * ${p} + threadIdx.y) * ${n} + col], sizeof(float));`),r._emit("__pipeline_commit();"),r._indent--,r._emit("}"),r._emit("__pipeline_wait_prior(t + 1 < nTiles ? 1 : 0);"),r._emit("__syncthreads();"),r._emit(`for (int kk = 0; kk < ${p}; kk++) acc += As[cur][threadIdx.y][kk] * Bs[cur][kk][threadIdx.x];`),r._emit("__syncthreads();"),r._indent--,r._emit("}"),r._emit(`${a}[row * ${n} + col] = acc;`)})});var Nv,Zb,_d,Av=M(()=>{"use strict";De();Vn();rd();Gr();kc();vv();Nv=256*1024,Zb=class{static{l(this,"CUDAKernel")}constructor(e,t,n,s,o,i,a,u){this.name=e,this.source=t,this.blockDim=n,this.gridDim=s,this.sharedMemBytes=o,this.params=i,this.outputIndices=a,this.scratch=u||[]}},_d=class{static{l(this,"CUDACodegen")}constructor(e){this.target=e,this._indent=0,this._lines=[],this._threadBindings=new Map,this._sharedBuffers=[],this._blockDim=[1,1,1],this._gridDim=[1,1,1],this._defaultDtype="f32",this._storeBuffers=new Set,this._promotedBuffers=new Set,this._promotedBufferDecls=[],this._declaredLocals=new Set,this._needsBarriers=!1,this._globalScratch=[],this._scratchNames=new Set,this._serializeThreads=!1}generate(e){if(this._indent=0,this._lines=[],this._threadBindings.clear(),this._sharedBuffers=[],this._blockDim=[1,1,1],this._gridDim=[1,1,1],this._didParallelReduce=!1,this._primFunc=e,this._storeBuffers=new Set,this._promotedBuffers=new Set,this._promotedBufferDecls=[],this._declaredLocals=new Set,this._needsBarriers=!1,this._globalScratch=[],this._scratchNames=new Set,this._serializeThreads=!1,e.type==="LIRFunc"){for(let[m,g]of e.metadata.threadBindings){this._threadBindings.set(m,g);for(let y of g)y.isDynamic||this._applyBindingDim(m,y.extent)}this._sharedBuffers=e.metadata.sharedBuffers}else this._scanBindings(e.body);this._scanStoreTargets(e.body),this._analyzeSharing(e),this._collectGlobalScratch(e);let n=[],s=[],o=[],i=0;for(let[,m]of e.bufferMap)s.push(m.name),n.push(`${_b(m.dtype)} ${m.name}`),this._defaultDtype=m.dtype,this._storeBuffers.has(m.name)&&o.push(i),i++;for(let m of this._globalScratch)s.push(m.name),n.push(`${_b(m.dtype)} ${m.name}`);for(let m of e.shapeParams)s.push(m.name),n.push(`int ${m.name}`);this._emit(`__global__ void ${e.name}(${n.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__ ${y}${$r(m.dtype)} ${m.name}[${g>0?g:1}];`)}for(let m of this._promotedBufferDecls)this._emit(`__shared__ ${$r(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} = ${m};`),u.add(y.varName));if(this._emitMissingLocalDecls(e),e._tensorIntrin){let m=kv(e._tensorIntrin.name);if(!m)throw new Error(`CUDA codegen: unknown tensor intrinsic '${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)],h=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)],f=c[0]*c[1]*c[2];if(f>p.maxThreadsPerBlock)throw new Error(`[codegen] kernel '${e.name}' block ${c.join("x")} = ${f} threads exceeds maxThreadsPerBlock ${p.maxThreadsPerBlock}`);let d=this._sharedBuffers.reduce((m,g)=>m+Math.max(g.sizeInBytes(),0),0)+this._promotedBufferDecls.reduce((m,g)=>m+Math.max(g.size,0)*bn(g.dtype),0);if(d>p.sharedMemoryBytes)throw new Error(`[codegen] kernel '${e.name}' shared memory ${d} bytes exceeds device limit ${p.sharedMemoryBytes}`);return new Zb(e.name,this._lines.join(`
+`),c,h,d,s,o,this._globalScratch)}_scanBindings(e){let t=[e];for(;t.length>0;){let n=t.pop();if(n){if(n.type==="ForNode"&&n.kind===F.THREAD_BINDING&&n.threadTag){let s=n.extent.type==="IntImmNode"?n.extent.value:0,o=n.extent.type!=="IntImmNode",i={varName:n.loopVar.name,extent:s,isDynamic:o,extentNode:n.extent};this._threadBindings.has(n.threadTag)?this._threadBindings.get(n.threadTag).push(i):this._threadBindings.set(n.threadTag,[i]),o||this._applyBindingDim(n.threadTag,s)}if(n.type==="AllocateNode"&&n.scope==="shared"&&this._sharedBuffers.push(n.buffer),n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody)}}}_applyBindingDim(e,t){let n=Wi(e);n&&(n.space==="thread"?this._blockDim[n.axis]=Math.max(this._blockDim[n.axis],t):this._gridDim[n.axis]=Math.max(this._gridDim[n.axis],t))}_emit(e){this._lines.push("  ".repeat(this._indent)+e)}_visitNode(e){ud(this,e)}_emitSync(){this._emit("__syncthreads();")}_matchFullReduction(e){let t=[],n=e;for(;n&&n.type==="ForNode";){if(n.kind!==F.SERIAL)return null;let f=n.extent.type==="IntImmNode"?n.extent.value:0;if(f<=0)return null;t.push({extC:this._exprToC(n.extent),extVal:f,varName:n.loopVar.name}),n=n.body}if(!n||n.type!=="BlockNode")return null;let s=n,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(f=>this._exprToC(f)).join(","),u=l(f=>f&&f.type==="BufferLoadNode"&&f.buffer===o.buffer&&f.indices.map(d=>this._exprToC(d)).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(f=>f.varName));if(a.split(/[^A-Za-z0-9_]/).some(f=>c.has(f)))return null;let h=t.reduce((f,d)=>f*d.extVal,1);return h<2048?null:{loops:t,block:s,store:o,valExpr:p,total:h}}_emitParallelReduction(e,t){this._blockDim=[256,1,1],this._didParallelReduce=!0;let s=$r(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===F.THREAD_BINDING&&this._serializeThreads){let s=e.loopVar.name,o=this._exprToC(e.extent);this._emit(`for (int ${s} = 0; ${s} < ${o}; ${s}++) {`),this._indent++,this._visitNode(e.body),this._indent--,this._emit("}");return}if(e.kind===F.THREAD_BINDING){let s=e.extent.type==="IntImmNode"?e.extent.value:0,o=e.threadTag,i=this._getMaxBindingExtent(o);s>0&&i>0&&s<i?(this._emit(`if (${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,n=this._exprToC(e.extent);(e.kind===F.UNROLLED||e.kind===F.VECTORIZED)&&this._emit("#pragma unroll"),this._emit(`for (int ${t} = 0; ${t} < ${n}; ${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} = ${this._exprToC(t.binding)};`);e.initBody&&this._visitNode(e.initBody),this._visitNode(e.body)}_visitAllocateNode(e){if(e.scope!=="shared"){if(this._promotedBuffers.has(e.buffer.name)||this._scratchNames.has(e.buffer.name)||this._declaredLocals.has(e.buffer.name))return;this._declaredLocals.add(e.buffer.name);let t=e.buffer.numel();if(t>0){let n=e.buffer.align16?"__align__(16) ":"";this._emit(`${n}${$r(e.buffer.dtype)} ${e.buffer.name}[${t}];`)}else this._emit(`${$r(e.buffer.dtype)}* ${e.buffer.name} = (${$r(e.buffer.dtype)}*)alloca(${this._dynamicNumel(e.buffer)} * sizeof(${$r(e.buffer.dtype)}));`)}}_visitIfStmt(e){this._emit(`if (${this._exprToC(e.condition)}) {`),this._indent++,this._visitNode(e.thenBody),this._indent--,e.elseBody&&(this._emit("} else {"),this._indent++,this._visitNode(e.elseBody),this._indent--),this._emit("}")}_visitLetStmtNode(e){let t=e.variable.dtype||this._defaultDtype;this._emit(`${$r(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)}] = ${this._exprToC(e.value)};`)}_visitVecCopyNode(e){let t=`${$r(e.dstBuffer.dtype)}${e.width}`,n=`${e.dstBuffer.name}[${this._flatIndex(e.dstBuffer,[e.dstIndex])}]`,s=`${e.srcBuffer.name}[${this._flatIndex(e.srcBuffer,[e.srcIndex])}]`;this._emit(`*reinterpret_cast<${t}*>(&${n}) = *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} = ${this._exprToC(t.expr)};`);this._visitNode(e.body)}_visitLIRAccumulator(e){let t=e.localName,n=e.dtype||this._defaultDtype;this._emit(`${$r(n)} ${t} = ${this._exprToC(e.initLoad)};`);let s=e.loopVar.name,o=this._exprToC(e.extent);this._emit(`for (int ${s} = 0; ${s} < ${o}; ${s}++) {`),this._indent++;let i=this._exprToC(e.body),a=e.op||"+",u;a==="max"||a==="min"?nr(e.dtype)?u=`((${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}${t})`;let n=this._exprToC(e.b);return e.op==="//"?`(${t} / ${n})`:`(${t} ${e.op} ${n})`}case"CompareNode":return`(${this._exprToC(e.a)} ${Ka(e.direction)} ${this._exprToC(e.b)})`;case"IfThenElseNode":return`(${this._exprToC(e.condition)} ? ${this._exprToC(e.thenBody)} : ${this._exprToC(e.elseBody)})`;case"CastNode":return`((${$r(e.toDtype)})(${this._exprToC(e.expr)}))`;case"CallExternNode":return this._emitExternCall(e);default:throw new Error(`CUDA codegen: unhandled expr node '${e.type}'`)}}_emitFloatLiteral(e){if(e===1/0)return"INFINITY";if(e===-1/0)return"(-INFINITY)";let t=bb(this._defaultDtype),n=String(e);return`${/[.eEnN]/.test(n)?n:n+".0"}${t}`}_emitExternCall(e){let t=e.args.length,n=new Array(t);for(let a=0;a<t;a++)n[a]=this._exprToC(e.args[a]);let s=n.join(", "),o=e.dtype||this._defaultDtype;if(e.externName==="rsqrt")return`${gh("rsqrt",o)}(${s})`;if(e.externName==="sign"){let a=n[0],u=`0.0${bb(o)}`;return`((${a} > ${u}) - (${a} < ${u}))`}if(e.externName==="min"||e.externName==="max"){if(nr(o)){let a=e.externName==="min"?"<":">";return`((${n[0]}) ${a} (${n[1]}) ? (${n[0]}) : (${n[1]}))`}return`${gh(e.externName,o)}(${s})`}return`${gh(e.externName,o)}(${s})`}_flatIndex(e,t){return du(e,t,n=>this._exprToC(n),(n,s)=>this._computeDynamicStride(n,s),!1)}_computeDynamicStride(e,t){return ri(e,t+1,(n,s)=>this._resolveShapeParam(n,s))}_dynamicNumel(e){return ri(e,0,(t,n)=>this._resolveShapeParam(t,n))}_getMaxBindingExtent(e){return pd(this._threadBindings,e)}_scanStoreTargets(e){let t=[e];for(;t.length>0;){let n=t.pop();if(n){(n.type==="BufferStoreNode"||n.type==="LIRFlatStoreNode")&&this._storeBuffers.add(n.buffer.name),n.type==="VecCopyNode"&&this._storeBuffers.add(n.dstBuffer.name),n.type==="LIRAccumulatorNode"&&n.flushStore&&this._storeBuffers.add(n.flushStore.buffer.name);for(let s of en(n))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 n=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===F.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(n,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 en(a))i(c,p)},"walk");i(e.body,[]);for(let[a,u]of n){let p=s.get(a);if(!p)continue;let c=new Set(u);for(let h of p)c.add(h);if(c.size>1)return!0}return!1}_findThreadPrivateAllocs(e,t){let n=l((s,o)=>{if(!s||typeof s!="object")return;let i=o;if(s.type==="ForNode"&&s.kind===F.THREAD_BINDING&&s.threadTag){let a=Wi(s.threadTag);a&&a.space==="thread"&&(i=!0)}i&&s.type==="AllocateNode"&&s.buffer&&t.add(s.buffer.name);for(let a of en(s))n(a,i)},"walk");n(e,!1)}_findCrossThreadBuffers(e){let t=new Set;for(let[,d]of e.bufferMap)t.add(d.name);let n=new Set(this._sharedBuffers.map(d=>d.name)),s=new Set;this._findThreadPrivateAllocs(e.body,s);let o=l(d=>d&&!t.has(d.name)&&!n.has(d.name)&&!s.has(d.name),"isIntermediate"),i=l(d=>typeof d.numel!="function"||d.numel()!==1,"isMulti"),a=new Set,u=new Set,p=new Set,c=new Set,h=l((d,m)=>{if(!d||typeof d!="object")return;let g=m;if(d.type==="ForNode"&&d.kind===F.THREAD_BINDING&&d.threadTag){let y=d.extent&&d.extent.type==="IntImmNode"?d.extent.value:0,_=this._getMaxBindingExtent(d.threadTag);y>0&&_>0&&y<_&&(g=!0)}(d.type==="BufferStoreNode"||d.type==="LIRFlatStoreNode")&&o(d.buffer)&&(i(d.buffer)?a.add(d.buffer.name):g&&p.add(d.buffer.name)),(d.type==="BufferLoadNode"||d.type==="LIRFlatLoadNode")&&o(d.buffer)&&(i(d.buffer)?u.add(d.buffer.name):c.add(d.buffer.name));for(let y of en(d))h(y,g)},"walk");h(e.body,!1);let f=new Set;for(let d of a)u.has(d)&&f.add(d);for(let d of p)c.has(d)&&f.add(d);return f}_crossThreadBuffersAreBlockLocal(e,t){let n=!0,s=l((o,i)=>{if(!o||typeof o!="object")return;let a=i;if(o.type==="ForNode"&&o.kind===F.THREAD_BINDING&&o.threadTag){let u=Wi(o.threadTag);u&&u.space==="block"&&(a=!0)}a&&(o.type==="BufferStoreNode"||o.type==="LIRFlatStoreNode")&&o.buffer&&t.has(o.buffer.name)&&(n=!1);for(let u of en(o))s(u,a)},"walk");return s(e.body,!1),n}_promoteCrossThreadToShared(e,t){let n=new Set;for(let[,p]of e.bufferMap)n.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(n.has(p)||this._promotedBuffers.has(p)||i.has(p))continue;let c=s.get(p);if(!c)return!1;let h=c.numel(),f=h>0?h:this._estimateBufferSize(c);if(f<=0||(a+=f*bn(c.dtype),a>o))return!1;u.push({name:p,dtype:c.dtype,size:f})}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 n=!1;for(let[u,p]of this._threadBindings){let c=new Set;for(let f of p)f.extent>0&&c.add(f.extent);if(c.size<=1)continue;let h=Wi(u);if(h&&h.space==="block"){this._serializeThreads=!0,this._needsBarriers=!1;return}n=!0}if(!n)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 en(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(),h=c>0?c:this._estimateBufferSize(p);h>0&&(this._promotedBuffers.add(u),this._promotedBufferDecls.push({name:u,dtype:p.dtype,size:h}))}}_emitMissingLocalDecls(e){let t=new Set;for(let[,o]of e.bufferMap)t.add(o.name);let n=new Set;this._scanAllocateNodes(e.body,n);let s=new Map;this._scanBufferRefs(e.body,s);for(let[o,i]of s){if(t.has(o)||n.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(`${$r(i.dtype)} ${o}[${u}];`))}}_collectGlobalScratch(e){let t=this._serializeThreads;if(this._threadBindings.size>0&&!t)return;let n=32768,s=new Set;for(let[,f]of e.bufferMap)s.add(f.name);let o=[],i=new Set,a=l((f,d)=>{if(!d||s.has(f)||this._promotedBuffers.has(f)||this._scratchNames.has(f)||i.has(f))return;let m=typeof d.numel=="function"?d.numel():0,g=m>0?m:this._estimateBufferSize(d);g<=0||(i.add(f),o.push({name:f,dtype:d.dtype,size:g,bytes:g*bn(d.dtype)}))},"consider"),u=[e.body];for(;u.length>0;){let f=u.pop();if(f){f.type==="AllocateNode"&&f.scope!=="shared"&&a(f.buffer.name,f.buffer);for(let d of en(f))u.push(d)}}let p=new Map;this._scanBufferRefs(e.body,p);for(let[f,d]of p)a(f,d);let c=l(f=>{this._scratchNames.add(f.name),this._globalScratch.push({name:f.name,dtype:f.dtype,size:f.size})},"offload");if(!t){for(let f of o)f.size>n&&c(f);return}let h=o.reduce((f,d)=>f+d.bytes,0);if(!(h<=Nv))for(let f of[...o].sort((d,m)=>m.bytes-d.bytes)){if(h<=Nv)break;c(f),h-=f.bytes}}_scanAllocateNodes(e,t){let n=[e];for(;n.length>0;){let s=n.pop();if(s){s.type==="AllocateNode"&&t.add(s.buffer.name);for(let o of en(s))n.push(o)}}}_scanBufferRefs(e,t){let n=[e];for(;n.length>0;){let s=n.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 en(s))n.push(o)}}}_estimateBufferSize(e){return cd(e)}_resolveShapeParam(e,t){return ni(this._primFunc,e,t,n=>n.name,"CUDA","c")}}});var Su,bd,Ov=M(()=>{"use strict";De();Vn();xc();xu();Gr();kc();Su=new Set(["f16","bf16"]),bd=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,h]of e.metadata.memoryLayout.bufferOffsets)this._bufferOffsets.set(c,h);this._totalMemBytes=e.metadata.memoryLayout.totalBytes;for(let[c,h]of e.metadata.externCalls){let f=this._mathImportSig(c,h.argCount);this._imports.set(c,f)}for(let[c,h]of e.metadata.locals)this._ensureLocal(c,ys(h))}else this._layoutBuffers(e),this._scanMathImports(e.body);this._scanParallel(e.body);let n=[];for(let[,c]of e.bufferMap)n.push(c.name);for(let c of e.shapeParams)n.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} ${o})`);for(let[c,h]of this._imports)this._emit(`(import "math" "${c}" (func $math_${c} ${h}))`);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"),n.push("_par_start"),n.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,h]of this._locals)u.push("(local $"+c+" "+h+")");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 $_par_end")}return this._visitNode(e.body),this._indent--,this._emit(")"),this._indent--,this._emit(")"),{name:e.name,wat:this._lines.join(`
+`),memoryPages:s,bufferOffsets:new Map(this._bufferOffsets),imports:this._imports,params:n,parallel:this._hasParallel?{extent:this._parallelExtent,outputIndices:this._findOutputIndices(e),poolSafe:this._isParallelSafe(e)}:null}}_isParallelSafe(e){let t=[],n=[e.body];for(;n.length>0;){let p=n.pop();if(p){if(p.type==="ForNode"&&p.kind===F.PARALLEL&&t.push(p),p.body&&n.push(p.body),p.stmts)for(let c of p.stmts)n.push(c);p.thenBody&&n.push(p.thenBody),p.elseBody&&n.push(p.elseBody),p.loopBody&&n.push(p.loopBody)}}if(t.length!==1)return!1;let s=t[0];if(!(e.body&&e.body.stmts?e.body.stmts:[e.body]).includes(s))return!1;let i=l((p,c)=>{let h=[p];for(;h.length>0;){let f=h.pop();if(f){if((f.type==="BufferStoreNode"||f.type==="LIRFlatStoreNode")&&c.add(f),f.type==="LIRAccumulatorNode"&&f.flushStore&&c.add(f.flushStore),f.body&&h.push(f.body),f.stmts)for(let d of f.stmts)h.push(d);f.thenBody&&h.push(f.thenBody),f.elseBody&&h.push(f.elseBody),f.loopBody&&h.push(f.loopBody)}}},"collectStores"),a=new Set,u=new Set;i(e.body,a),i(s.body,u);for(let p of a)if(!u.has(p))return!1;return!0}_ensureLocal(e,t){this._locals.has(e)||this._locals.set(e,t)}_emit(e){this._lines.push("  ".repeat(this._indent)+e)}_emitLoadOp(e){if(Su.has(e)){this._emitHalfDecode(e);return}this._emit(dh(e))}_emitStoreOp(e){if(Su.has(e)){this._emitHalfEncode(e);return}this._emit($S(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 "+Hb.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 "+Hb.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[,n]of e.bufferMap)if(Su.has(n.dtype)){t=!0;break}if(!t&&e.metadata&&e.metadata.locals){for(let[,n]of e.metadata.locals)if(Su.has(n)){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 n=t.pop();if(!(!n||typeof n!="object")){if((n.type==="BufferLoadNode"||n.type==="BufferStoreNode")&&n.buffer&&Su.has(n.buffer.dtype)||(n.type==="LIRFlatLoadNode"||n.type==="LIRFlatStoreNode")&&Su.has(n.dtype))return!0;if(n.body&&t.push(n.body),n.value&&typeof n.value=="object"&&t.push(n.value),n.stmts)for(let s of n.stmts)t.push(s);if(n.a&&t.push(n.a),n.b&&t.push(n.b),n.expr&&t.push(n.expr),n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),n.initBody&&t.push(n.initBody),n.condition&&t.push(n.condition),n.offsetExpr&&t.push(n.offsetExpr),n.args)for(let s of n.args)t.push(s);if(n.indices)for(let s of n.indices)t.push(s)}}return!1}_visitNode(e){let t=e;for(;t;)switch(t.type){case"SeqNode":for(let n=0;n<t.stmts.length-1;n++)this._visitNode(t.stmts[n]);t=t.stmts[t.stmts.length-1];continue;case"AllocateNode":t=t.body;continue;case"LetStmtNode":{let n=this._vectorMode;if(n&&n.simd&&this._dependsOnVecVar(t.value)){let o=t.variable.name+"_vlet";this._ensureLocal(o,"v128"),n.vecLets||(n.vecLets=new Set),n.vecLets.add(t.variable.name),this._emitVecExpr(t.value),this._emit("local.set $"+o),t=t.body;continue}let s=ir(t.value)||t.variable.dtype||this._defaultDtype;this._locals.set(t.variable.name,ys(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,n=[e.body];for(;n.length>0;){let i=n.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&&n.push(i.body),i.stmts)for(let a of i.stmts)n.push(a);i.thenBody&&n.push(i.thenBody),i.elseBody&&n.push(i.elseBody),i.loopBody&&n.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 n=t.pop();if(n){if(n.type==="ForNode"&&n.kind===F.PARALLEL){this._hasParallel=!0,this._parallelExtent=n.extent&&n.extent.type==="IntImmNode"?n.extent.value:0;return}if(n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody)}}}_visitFor(e){let t=e.loopVar.name,n=this._constExtent(e.extent);if(e.kind===F.PARALLEL){if(n!==null&&this._parallelExtent&&n!==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===F.VECTORIZED&&this.target.supportsSimd()&&n!==null){this._visitVectorizedFor(e);return}if((e.kind===F.UNROLLED||e.kind===F.VECTORIZED)&&n!==null&&n<=32&&!this._isZeroFillBody(e.body)){for(let o=0;o<n;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,ys(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 n=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==="*"&&(n(u.a)||n(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(n(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(rr(t))return e;throw new Error(`wasm accumulator: integer ${e} reduction not supported (dtype ${t})`)}return"add"}_visitLIRAccumulator(e){let t=e.localName,n=e.dtype;this._ensureLocal(t,ys(n));let s=this._constExtent(e.extent),o=e.op||"+",i=o==="+"&&s!==null&&e.loopKind===F.VECTORIZED&&this.target.supportsSimd()?yb(n):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(n)),this._emit(this._numPrefix(n)+"."+this._accumInstr(o,n)),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,n,s){let o=e.localName,i=e.dtype,a=e.loopVar.name,u=rr(i),p=u?"f32.add":"i32.add",c=_s(i,"add"),h=Math.floor(s/n)*n,f=h,d=o+"_vec";this._ensureLocal(d,"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 $"+d);let m=this._wasmAcc;this._wasmAcc={local:o,bufName:e.flushStore.buffer.name},this._vectorMode={dtype:i,lanes:n,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 "+h+")"),this._emit("i32.ge_s"),this._emit("br_if $vbreak_"+a),this._emit("(local.get $"+d+")"),this._emitVecExpr(e.body),this._emit(c),this._emit("local.set $"+d),this._emit("(local.get $"+a+")"),this._emit("(i32.const "+n+")"),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<n;g++)this._emit("(local.get $"+d+")"),this._emit(t.extractLane+" "+g),this._emit(p);this._emit("local.set $"+o),f<s&&(this._emit("(i32.const "+f+")"),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 n=this._bufferOffsets.get(e.name)||0,s=cc(e.dtype||"f32");if(!t||t.type==="IntImmNode"&&t.value===0){this._emit(`(i32.const ${n})`);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+")"),n>0&&(this._emit(`(i32.const ${n})`),this._emit("i32.add"));return}this._emitExpr(t),this._emit(`(i32.const ${s})`),this._emit("i32.mul"),n>0&&(this._emit(`(i32.const ${n})`),this._emit("i32.add"))}_emitForLoop(e,t,n){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(n),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 n=this._vectorMode;n&&n.laneVars&&this._dependsOnVecVar(t.binding)&&n.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 n=this._bufferOffsets.get(e.name)||0,s=cc(e.dtype);if(t.length===0){this._emit(`(i32.const ${n})`);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+")"),n>0&&(this._emit(`(i32.const ${n})`),this._emit("i32.add"));return}this._emitFlatIndex(e,t),this._emit(`(i32.const ${s})`),this._emit("i32.mul"),n>0&&(this._emit(`(i32.const ${n})`),this._emit("i32.add"))}_emitFlatIndex(e,t){if(t.length===1){this._emitExpr(t[0]);return}let n=!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")),n||this._emit("i32.add"),n=!1}}_emitDynamicStride(e,t){let n=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})`)}n>0&&this._emit("i32.mul"),n++}n===0&&this._emit("(i32.const 1)")}_resolveShapeParam(e,t){return ni(this._primFunc,e,t,n=>n.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||ir(e.thenBody),n=this._numPrefix(t);this._emitExpr(e.condition),rr(this._wasmExprDtype(e.condition))&&(this._emit("(f32.const 0)"),this._emit("f32.ne")),this._emit("(if (result "+n+")"),this._indent++,this._emit("(then"),this._indent++,this._emitCoercedTo(e.thenBody,n),this._indent--,this._emit(")"),this._emit("(else"),this._indent++,this._emitCoercedTo(e.elseBody,n),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":rr(e)?"f32":"i32"}_exprPrefix(e){return this._numPrefix(e&&e._dtype||ir(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),n=e.b?this._exprPrefix(e.b):t,s=this._joinPrefix(t,n);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),n=this._exprPrefix(e.b),s=this._joinPrefix(t,n),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")&&!rr(e.dtype)){this._emitIntMinMax(e);return}if(e.externName==="abs"&&!rr(e.dtype)){this._emitIntAbs(e);return}let t=new Set(["sqrt","abs","ceil","floor","min","max","rsqrt"]),n=e.externName==="abs"||t.has(e.externName)?this._numPrefix(e.dtype)==="i32"?"f32":this._numPrefix(e.dtype):"f32";if(e.externName==="rsqrt"){this._emit("("+n+".const 1)");for(let s of e.args)this._emitCoercedTo(s,n);this._emit(n+".sqrt"),this._emit(n+".div");return}for(let s of e.args)this._emitCoercedTo(s,n);switch(e.externName){case"sqrt":this._emit(n+".sqrt");break;case"abs":this._emit(n+".abs");break;case"ceil":this._emit(n+".ceil");break;case"floor":this._emit(n+".floor");break;case"min":this._emit(n+".min");break;case"max":this._emit(n+".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,n="_iabs"+t;this._intMinMaxEmitDepth=t+1,this._emitExpr(e.args[0]),this._emit("local.set $"+n),this._intMinMaxEmitDepth=t,this._emit("(i32.const 0)"),this._emit("(local.get $"+n+")"),this._emit("i32.sub"),this._emit("(local.get $"+n+")"),this._emit("(local.get $"+n+")"),this._emit("(i32.const 0)"),this._emit("i32.lt_s"),this._emit("select")}_emitIntMinMax(e){let t=this._intMinMaxEmitDepth||0,n="_immm_a"+t,s="_immm_b"+t;this._intMinMaxEmitDepth=t+1,this._emitExpr(e.args[0]),this._emit("local.set $"+n),this._emitExpr(e.args[1]),this._emit("local.set $"+s),this._intMinMaxEmitDepth=t,this._emit("(local.get $"+n+")"),this._emit("(local.get $"+s+")"),this._emit("(local.get $"+n+")"),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 n=this._intDivEmitDepth||0,s="_idiv_a"+n,o="_idiv_b"+n;this._intDivEmitDepth=n+1,this._emitCoercedTo(e.a,t),this._emit("local.set $"+s),this._emitCoercedTo(e.b,t),this._emit("local.set $"+o),this._intDivEmitDepth=n;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 n=this._intDivEmitDepth||0,s="_idiv_a"+n,o="_idiv_b"+n;this._intDivEmitDepth=n+1,this._emitCoercedTo(e.a,t),this._emit("local.set $"+s),this._emitCoercedTo(e.b,t),this._emit("local.set $"+o),this._intDivEmitDepth=n,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((n,s,o)=>{if(!n||typeof n!="object")return;let i=s,a=o;if(n.type==="CallExternNode"&&(n.externName==="min"||n.externName==="max")&&!rr(n.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(n.type==="CallExternNode"&&n.externName==="abs"&&!rr(n.dtype))this._ensureLocal("_iabs"+s,"i32"),s+1>this._intMinMaxDepth&&(this._intMinMaxDepth=s+1),i=s+1;else if(this._isIntDivNode(n)){let u=ys(this._joinPrefix(this._exprPrefix(n.a),this._exprPrefix(n.b)));this._ensureLocal("_idiv_a"+o,u),this._ensureLocal("_idiv_b"+o,u),a=o+1}if(n.body&&t(n.body,i,a),n.value&&typeof n.value=="object"&&t(n.value,i,a),n.a&&t(n.a,i,a),n.b&&t(n.b,i,a),n.expr&&t(n.expr,i,a),n.condition&&t(n.condition,i,a),n.offsetExpr&&t(n.offsetExpr,i,a),n.thenBody&&t(n.thenBody,i,a),n.elseBody&&t(n.elseBody,i,a),n.initBody&&t(n.initBody,i,a),n.stmts)for(let u of n.stmts)t(u,i,a);if(n.args)for(let u of n.args)t(u,i,a);if(n.indices)for(let u of n.indices)t(u,i,a);if(n.bindings)for(let u of n.bindings)t(u.expr,i,a);if(n.iterVars)for(let u of n.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 ld(e)}_collectBindings(e,t){let n=[e];for(;n.length>0;){let s=n.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&&n.push(s.body),s.stmts)for(let o of s.stmts)n.push(o);s.thenBody&&n.push(s.thenBody),s.elseBody&&n.push(s.elseBody),s.initBody&&n.push(s.initBody),s.loopBody&&n.push(s.loopBody)}}}_computeLaneVars(e){let t=new Set([e.loopVar.name]),n=[];this._collectBindings(e.body,n);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 n)t.has(i.name)||s(i.expr).some(a=>t.has(a))&&(t.add(i.name),o=!0)}return t}_vecLoadsContiguous(e,t){let n=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==="*"&&(n(u.a)||n(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(n(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 n=[e];for(;n.length>0;){let s=n.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&&n.push(s.body),s.stmts)for(let i of s.stmts)n.push(i);s.thenBody&&n.push(s.thenBody),s.elseBody&&n.push(s.elseBody),s.loopBody&&n.push(s.loopBody)}return!0}_visitVectorizedFor(e){let t=e.loopVar.name,n=this._constExtent(e.extent),s=this._inferBodyDtype(e.body)||this._defaultDtype,o=this.target.vectorWidth,i=yb(s);if(!i||n<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(n/o)*o,p=n-u,c=cc(s),d=this._countBufAccesses(e.body)>=2?"_vaddr_"+t:null;if(u>0&&(this._vectorMode={dtype:s,lanes:o,loopVar:t,simd:i,addrLocal:d,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<n;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,n=[e];for(;n.length>0;){let s=n.pop();if(!(!s||typeof s!="object")){if((s.type==="BufferLoadNode"||s.type==="BufferStoreNode"||s.type==="LIRFlatLoadNode"||s.type==="LIRFlatStoreNode")&&t++,s.body&&n.push(s.body),s.value&&typeof s.value=="object"&&n.push(s.value),s.a&&n.push(s.a),s.b&&n.push(s.b),s.stmts)for(let o of s.stmts)n.push(o);if(s.args)for(let o of s.args)n.push(o);if(s.indices)for(let o of s.indices)n.push(o);s.thenBody&&n.push(s.thenBody),s.elseBody&&n.push(s.elseBody),s.expr&&n.push(s.expr),s.condition&&n.push(s.condition),s.offsetExpr&&n.push(s.offsetExpr)}}return t}_inferBodyDtype(e){let t=[e];for(;t.length>0;){let n=t.pop();if(n){if(n.type==="BufferStoreNode"&&n.buffer)return n.buffer.dtype;if(n.type==="LIRFlatStoreNode")return n.dtype||this._defaultDtype;if(n.type==="BufferLoadNode"&&n.buffer)return n.buffer.dtype;if(n.type==="LIRFlatLoadNode")return n.dtype||this._defaultDtype;if(n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),n.value&&typeof n.value=="object"&&t.push(n.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 n=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||n&&n.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,n=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":rr(n)?(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+")"),rr(n)&&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),rr(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,n=t.dtype;if(!e.b){if(e.op==="-"){let u=_s(n,"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(_s(n,"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=_s(n,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=_s(n,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,n=t.dtype,s=_s(n,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,n=t.dtype,s=_s(n,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=_s(n,"sqrt");if(o){this._emit("(f32.const 1)"),this._emit(t.simd.splat),this._emitVecExpr(e.args[0]),this._emit(o),this._emit(_s(n,"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,n=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<n;p++){let c="_vl_"+a+"_"+p;this._ensureLocal(c,ys(t.dtype)),u.push(c)}this._emitVecExpr(e.args[0]),this._emit("local.set $"+a);for(let p=0;p<n;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 h=this._vectorMode;this._vectorMode=t,this._emitVecExpr(e.args[1]),this._vectorMode=h,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[n-1]+")"),this._emit(i);for(let p=n-2;p>=0;p--)this._emit("(local.get $"+u[p]+")"),this._emit(o+" "+p)}_prescanVecLocalsAll(e){let t=[e];for(;t.length>0;){let n=t.pop();if(n){if(n.type==="ForNode"&&n.kind===F.VECTORIZED&&(this._prescanVecLocals(n.body),this._countBufAccesses(n.body)>=2&&this._ensureLocal("_vaddr_"+n.loopVar.name,"i32"),this._prescanVecLets(n)),n.type==="LIRAccumulatorNode"&&n.loopKind===F.VECTORIZED&&(this._ensureLocal(n.localName+"_vec","v128"),this._prescanVecLocals(n.body)),n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),n.initBody&&t.push(n.initBody),n.loopBody&&t.push(n.loopBody)}}}_prescanVecLets(e){let t=this._computeLaneVars(e),n=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&&n(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 n=t.pop();if(n){if(n.type==="CallExternNode"&&n.externName&&!_s(this._defaultDtype,n.externName)&&n.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,ys(this._defaultDtype));n.args.length>1&&this._ensureLocal("_vtmp2_"+o,"v128"),this._vecTmpCounter++}if(n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);if(n.value&&typeof n.value=="object"&&n.value.type&&t.push(n.value),n.a&&typeof n.a=="object"&&t.push(n.a),n.b&&typeof n.b=="object"&&t.push(n.b),n.args)for(let s of n.args)typeof s=="object"&&t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody)}}}_layoutBuffers(e){let t=0,n=16;this._dynamicBuffers=new Set;let s=l(i=>{t=Math.ceil(t/n)*n,this._bufferOffsets.set(i.name,t);let a=i.numel();!i.shape.some(p=>typeof p!="number"||p<0)&&a>0?t+=a*cc(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 n=[e];for(;n.length>0;){let s=n.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 en(s))n.push(o)}}}_scanMathImports(e){let t=[e];for(;t.length>0;){let n=t.pop();if(!(!n||typeof n!="object")){if(n.type==="CallExternNode"&&n.externName){let s=n.externName;if(s!=="sqrt"&&s!=="min"&&s!=="max"){let o=this._mathImportSig(s,n.args.length);this._imports.set(s,o)}}if(n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);if(n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),n.initBody&&t.push(n.initBody),n.value&&typeof n.value=="object"&&n.value.type&&t.push(n.value),n.a&&typeof n.a=="object"&&t.push(n.a),n.b&&typeof n.b=="object"&&t.push(n.b),n.expr&&typeof n.expr=="object"&&t.push(n.expr),n.args)for(let s of n.args)typeof s=="object"&&t.push(s);if(n.indices)for(let s of n.indices)typeof s=="object"&&t.push(s);n.condition&&typeof n.condition=="object"&&t.push(n.condition)}}}_prescanLocals(e){this._waccCounter=0;let t=[e];for(;t.length>0;){let n=t.pop();if(n){if(n.type==="ForNode"){this._ensureLocal(n.loopVar.name,"i32");let s=this._accPatternDtype(n);s&&this._ensureLocal("_wacc_"+ ++this._waccCounter,ys(s)),n.kind===F.VECTORIZED&&this.target.supportsSimd()&&this._prescanVecLocals(n.body)}if(n.type==="BlockNode")for(let s of n.iterVars)s.iterVar&&this._ensureLocal(s.iterVar.name,"i32");if(n.type==="LetStmtNode"&&n.variable){let s=ir(n.value)||n.variable.dtype||this._defaultDtype;this._ensureLocal(n.variable.name,ys(s))}if(n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),n.initBody&&t.push(n.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 ir(e)}_fixLetStmtLocals(e){let t=[e];for(;t.length>0;){let n=t.pop();if(n){if(n.type==="LetStmtNode"&&n.variable&&n.value){let s=ir(n.value);s&&this._locals.set(n.variable.name,ys(s))}if(n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),n.initBody&&t.push(n.initBody)}}}_accPatternDtype(e){let t=e.body;if(!t||t.type!=="BlockNode")return null;let n=t.body;if(!n||n.type!=="BufferStoreNode")return null;let s=n.value;return!s||s.type!=="MathOpNode"||s.op!=="+"?null:s.a&&s.a.type==="BufferLoadNode"&&s.a.buffer.name===n.buffer.name||s.b&&s.b.type==="BufferLoadNode"&&s.b.buffer.name===n.buffer.name?n.buffer.dtype:null}_inferDtype(e){return e&&e._dtype||ir(e)}_detectWasmAcc(e){let t=e.body;if(!t||t.type!=="BlockNode")return null;let n=t.body;if(!n||n.type!=="BufferStoreNode")return null;let s=n,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 h of t.iterVars)if(h.iterVar&&h.iterVar.name===c.name&&h.binding)return h.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 n=[];for(let s=0;s<t.length;s++)n.push(this._exprKey(t[s]));return e.name+":"+n.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 ku,Qb=M(()=>{"use strict";ku=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 n=t.length-1;for(;n>0;){let s=n-1>>1;if(this._compare(t[n],t[s])>=0)break;[t[n],t[s]]=[t[s],t[n]],n=s}}pop(){let e=this._items;if(e.length===0)return null;let t=e[0],n=e.pop();if(e.length>0){e[0]=n;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 oi(r){let e=String(r);return/[.e]/.test(e)?e:`${e}.0`}function Bv(r){let e=`(1.0 / (1.0 + ${oi(vc)} * abs(${r})))`,t=_u.slice().reverse().reduce((n,s)=>`(${oi(s)} + ${e} * ${n})`,"0.0");return`((select(-1.0, 1.0, ${r} >= 0.0)) * (1.0 - ${e} * ${t} * exp(-${r} * ${r})))`}function wd(r){let e=`(${r} - 1.0)`,t=Hi.map((s,o)=>o===0?oi(s):`${oi(s)} / (${e} + ${oi(o)})`).join(" + "),n=`(${e} + ${oi(7+.5)})`;return`(${oi(.5*Math.log(2*Math.PI))} + (${e} + 0.5) * log(${n}) - ${n} + log(${t}))`}function QL(r){let e=`(log(${Td} / abs(sin(${Td} * ${r}))) - ${wd(`(1.0 - ${r})`)})`;return`(select(${wd(r)}, ${e}, ${r} < 0.5))`}function XL(r){let e=`(${Td} / (sin(${Td} * ${r}) * exp(${wd(`(1.0 - ${r})`)})))`;return`(select(exp(${wd(r)}), ${e}, ${r} < 0.5))`}function Rv(r,e){for(let t of e2)r[t]&&e(r[t]);if(r.stmts)for(let t of r.stmts)e(t);if(r.indices)for(let t of r.indices)e(t);if(r.args)for(let t of r.args)e(t);r.initLoad&&e(r.initLoad),r.flushStore&&e(r.flushStore)}var JL,ZL,Td,e2,t2,Xb,xd,Mv=M(()=>{"use strict";De();Vn();rd();Qb();Gr();kc();Ji();JL=new Set(["!","&&","||"]),ZL=new Set(["<",">","<=",">=","==","!="]);l(oi,"_wgslFloat");l(Bv,"_wgslErf");l(wd,"_wgslLanczosCore");Td=oi(Math.PI);l(QL,"_wgslLgamma");l(XL,"_wgslGamma");e2=["body","loopBody","condBody","initBody","thenBody","elseBody","value","a","b","condition","expr","offsetExpr"];l(Rv,"walkFullChildren");t2={"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"},Xb=class{static{l(this,"WebGPUKernel")}constructor(e,t,n,s,o,i,a){this.name=e,this.source=t,this.workgroupSize=n,this.dispatchSize=s,this.sharedMemBytes=o,this.params=i,this.bindings=a}},xd=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[x,k]of e.metadata.threadBindings){this._threadBindings.set(x,k);for(let A of k)A.isDynamic||this._applyBindingDim(x,A.extent)}this._sharedBuffers=e.metadata.sharedBuffers}else this._scanBindings(e.body);this._scanStoreTargets(e.body),this._analyzeSharing(e);let n=[],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(x=>x.numel()>0);this._packedMode=p,this._packedOffsets=null;let c=new Map;{let x=0;for(let[,k]of e.bufferMap)c.set(k.name,x++)}{let x=null;for(let[,k]of e.bufferMap)if(this._storeBuffers.has(k.name)){x=k.dtype;break}this._defaultDtype=x||(u.length?u[0].dtype:"f32")}let h={f32:"f32",i32:"i32",u32:"u32",f16:"f16"};if(p){this._packedOffsets=new Map;let x=l(N=>Math.ceil(N/4)*4,"align4"),k=new Map;for(let[,N]of e.bufferMap){n.push(N.name);let v=this._storeBuffers.has(N.name),O=Mn(N.dtype),R=(v?"w:":"r:")+O,B=k.get(R);B||(B={isWrite:v,wt:O,dtype:h[O]||N.dtype,bufs:[],off:0},k.set(R,B));let E=N.numel(),z=(v?"_pw_":"_pr_")+O;this._packedOffsets.set(N.name,{storage:z,offset:B.off}),B.bufs.push({name:N.name,offset:B.off,size:E,dtype:N.dtype,argIndex:c.get(N.name)}),B.off+=x(E)}let A=[...k.values()].sort((N,v)=>N.isWrite===v.isWrite?N.wt.localeCompare(v.wt):N.isWrite?1:-1);for(let N of A){let v=(N.isWrite?"_pw_":"_pr_")+N.wt,O=N.isWrite?"storage, read_write":"storage, read";s.push({index:o,name:v,mode:N.isWrite?"read_write":"read",packed:N.bufs,packedSize:N.off,dtype:N.dtype}),this._emit(`@group(0) @binding(${o}) var<${O}> ${v}: array<${N.wt}>;`),o++}}else for(let[,x]of e.bufferMap){n.push(x.name);let k=this._storeBuffers.has(x.name)?"read_write":"read",A=k==="read_write"?"storage, read_write":"storage, read";s.push({index:o,name:x.name,mode:k,dtype:x.dtype,argIndex:c.get(x.name)}),this._emit(`@group(0) @binding(${o}) var<${A}> ${x.name}: array<${Mn(x.dtype)}>;`),o++}if(e.shapeParams.length>0){this._emit(""),this._emit("struct ShapeParams {"),this._indent++;for(let x of e.shapeParams)n.push(x.name),this._emit(`${x.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 x of this._sharedBuffers){let k=x.numel();this._emit(`var<workgroup> ${x.name}: array<${Mn(x.dtype)}, ${k>0?k:1}>;`)}for(let x of this._promotedBufferDecls)this._emit(`var<workgroup> ${x.name}: array<${Mn(x.dtype)}, ${x.size}>;`);for(let x of this._wgPoolDecls)this._emit(`var<workgroup> ${x.pool}: array<${Mn(x.dtype)}, ${x.size>0?x.size:1}>;`);(this._sharedBuffers.length>0||this._promotedBufferDecls.length>0||this._wgPoolDecls.length>0)&&this._emit("");let f=[],d=!this._serializeThreads&&this._hasBindingPrefix("threadIdx"),m=!this._serializeThreads&&this._hasBindingPrefix("blockIdx");d&&f.push("@builtin(local_invocation_id) _lid: vec3u"),m&&f.push("@builtin(workgroup_id) _wid: vec3u"),!d&&!m&&f.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 x=0;this._workgroupSize[1]>this._workgroupSize[x]&&(x=1),this._workgroupSize[2]>this._workgroupSize[x]&&(x=2),this._workgroupSize[x]=Math.max(1,this._workgroupSize[x]>>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}(${f.join(", ")}) {`),this._indent++,!this._serializeThreads){let x=new Set;for(let[k,A]of this._threadBindings)if(t2[k]){for(let v of A)if(!x.has(v.varName)){let O=this._wgslBuiltinAccess(k);this._emit(`let ${v.varName}: i32 = i32(${O});`),x.add(v.varName)}}}this._assignLocalSlots(e),this._emitMissingLocalDecls(e),this._visitNode(e.body),this._indent--,this._emit("}");let w=this.target,T=[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 Xb(e.name,this._lines.join(`
+`),T,S,this._sharedBuffers.reduce((x,k)=>x+Math.max(k.sizeInBytes(),0),0),n,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 n=e.substring(0,t),s=e.substring(t+1);return n==="threadIdx"?`_lid.${s}`:n==="blockIdx"?`_wid.${s}`:`_gid.${s}`}_scanStoreTargets(e){let t=[e];for(;t.length>0;){let n=t.pop();if(n){(n.type==="BufferStoreNode"||n.type==="LIRFlatStoreNode")&&this._storeBuffers.add(n.buffer.name),n.type==="LIRAccumulatorNode"&&n.flushStore&&this._storeBuffers.add(n.flushStore.buffer.name);for(let s of en(n))t.push(s)}}}_scanBindings(e){yu(e,t=>{if(t.type==="ForNode"&&t.kind===F.THREAD_BINDING&&t.threadTag){let n=t.extent.type==="IntImmNode"?t.extent.value:0,s=t.extent.type!=="IntImmNode",o={varName:t.loopVar.name,extent:n,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,n)}t.type==="AllocateNode"&&t.scope==="shared"&&this._sharedBuffers.push(t.buffer)})}_applyBindingDim(e,t){let n=Wi(e);n&&(n.space==="thread"?this._workgroupSize[n.axis]=Math.max(this._workgroupSize[n.axis],t):this._dispatchSize[n.axis]=Math.max(this._dispatchSize[n.axis],t))}_getMaxBindingExtent(e){return pd(this._threadBindings,e)}_hasRecurrence(e){let t=!1;return yu(e.body,n=>{if(n.type==="SyncThreadsNode")return t=!0,!1}),t}_analyzeSharing(e){let t=new Set;for(let[,h]of this._threadBindings)for(let f of h)f.extent>0&&t.add(f.extent);let n=this._threadBindings.size>0?this._findCrossThreadBuffers(e):new Set,s=this._threadBindings.size>0?this._findCrossExtentBuffers(e):new Set;this._crossThread=n,this._crossExtent=s;let o=this._hasRecurrence(e);if(!o&&t.size<=1&&n.size===0&&s.size===0)return;let i=this.target.sharedMemoryBytes||16384,a=this._sharedBuffers.reduce((h,f)=>h+Math.max(f.sizeInBytes(),0),0),u=new Set;for(let[,h]of e.bufferMap)u.add(h.name);let p=this._collectPromotionCandidates(e,u);p.sort((h,f)=>(n.has(f.name)?1:0)-(n.has(h.name)?1:0));let c=this._dispatchSize[0]*this._dispatchSize[1]*this._dispatchSize[2];if(o){let h=this.target.maxThreadsPerBlock||256,f=this._workgroupSize[0]*this._workgroupSize[1]*this._workgroupSize[2],d=this._packWorkgroupPool(e,p),m=a+d.bytes<=i;if(c===1&&f<=h&&m){this._needsBarriers=!0,this._wgPoolOffsets=d.offsets,this._wgPoolDecls=d.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=(n.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 h of p){let f=h.size*(bs(h.dtype)||4);a+f<=i&&(this._promotedBuffers.add(h.name),this._promotedBufferDecls.push(h),a+=f)}}_collectPromotionCandidates(e,t){let n=[];yu(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&&n.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(n.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&&(n.push({name:a,dtype:u.dtype,size:c}),i.add(a))}return n}_findRecurrenceBody(e){let t=null;return yu(e,n=>{if(!t&&n.type==="ForNode"&&n.kind===F.RECURRENCE)return t=n.body,!1}),t}_namesTouchedOutside(e,t,n){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&&n.has(i.buffer.name)&&s.add(i.buffer.name),i.type==="LIRAccumulatorNode"&&(i.flushStore&&i.flushStore.buffer&&n.has(i.flushStore.buffer.name)&&s.add(i.flushStore.buffer.name),i.initLoad&&i.initLoad.buffer&&n.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 n=new Set(t.map(f=>f.name)),s=this._findRecurrenceBody(e.body),{minPos:o,maxPos:i}=s?this._livenessWalk(s,n):{minPos:new Map,maxPos:new Map},a=s?this._namesTouchedOutside(e.body,s,n):n,u=new Map;for(let f of t){u.has(f.dtype)||u.set(f.dtype,[]);let d=a.has(f.name)||!o.has(f.name),m=d?0:o.get(f.name),g=d?Number.MAX_SAFE_INTEGER:i.get(f.name);u.get(f.dtype).push({name:f.name,size:f.size,first:m,last:g})}let p=new Map,c=[],h=0;for(let[f,d]of u){d.sort((_,b)=>_.first-b.first||b.size-_.size);let m=[],g=`_wg_${Mn(f)}`,y=0;for(let _ of d){let b=m.filter(T=>T.first<=_.last&&_.first<=T.last).map(T=>[T.offset,T.offset+T.size]).sort((T,S)=>T[0]-S[0]),w=0;for(let[T,S]of b){if(w+_.size<=T)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:f,size:y}),h+=y*(bs(f)||4)}return{offsets:p,decls:c,bytes:h}}_findCrossThreadBuffers(e){let t=new Set;for(let[,a]of e.bufferMap)t.add(a.name);let n=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!==F.THREAD_BINDING||a.type==="WhileNode"||a.type==="LIRAccumulatorNode")&&a.loopVar&&(p=[...u,a.loopVar.name]),a.type==="LIRBindingsNode"&&a.bindings)for(let h of a.bindings)!p.includes(h.name)&&o(h.expr,p)&&(p=[...p,h.name]);if(a.type==="BlockNode"&&a.iterVars)for(let h of a.iterVars)h.iterVar&&h.binding&&!p.includes(h.iterVar.name)&&o(h.binding,p)&&(p=[...p,h.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)&&n.add(a.buffer.name),a.type==="LIRFlatLoadNode"&&a.buffer&&!t.has(a.buffer.name)&&a.offsetExpr&&s([a.offsetExpr],p)&&n.add(a.buffer.name),Rv(a,h=>i(h,p))},"walk");return i(e.body,[]),n}_findCrossExtentBuffers(e){let t=new Set;for(let[,c]of e.bufferMap)t.add(c.name);let n=new Map,s=new Map,o=new Set,i=l(c=>c&&(c.type==="FloatImmNode"||c.type==="IntImmNode"),"isConst"),a=l((c,h,f)=>{if(t.has(h))return;let d=c.get(h);d||(d=new Set,c.set(h,d)),d.add(f)},"record"),u=l((c,h)=>{if(!c)return;let f=h;if(c.type==="ForNode"&&c.kind===F.THREAD_BINDING){let d=c.extent&&c.extent.type==="IntImmNode"?c.extent.value:0;d>0&&(f=h*d)}c.type==="BufferStoreNode"&&c.buffer&&(a(n,c.buffer.name,f),i(c.value)||o.add(c.buffer.name)),c.type==="LIRFlatStoreNode"&&c.buffer&&(a(n,c.buffer.name,f),i(c.value)||o.add(c.buffer.name)),c.type==="LIRAccumulatorNode"&&c.flushStore&&c.flushStore.buffer&&(a(n,c.flushStore.buffer.name,f),o.add(c.flushStore.buffer.name)),c.type==="BufferLoadNode"&&c.buffer&&a(s,c.buffer.name,f),c.type==="LIRFlatLoadNode"&&c.buffer&&a(s,c.buffer.name,f),Rv(c,d=>u(d,f))},"walk");u(e.body,1);let p=new Set;for(let[c,h]of s){if(!o.has(c))continue;let f=n.get(c);if(f){for(let d of h)if(!f.has(d)){p.add(c);break}}}return p}_collectVarNames(e,t){let n=[e];for(;n.length>0;){let s=n.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]&&n.push(s[o]);if(s.indices)for(let o of s.indices)n.push(o);if(s.args)for(let o of s.args)n.push(o)}}}_emit(e){this._lines.push("  ".repeat(this._indent)+e)}_visitNode(e){ud(this,e)}_emitSync(){this._needsBarriers&&(this._emit("storageBarrier();"),this._emit("workgroupBarrier();"))}_visitForNode(e){if(e.kind===F.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,n=this._exprToWGSL(e.extent);this._emit(`for (var ${t}: i32 = 0; ${t} < ${n}; ${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}: ${Mn(e.dtype)};`):this._emit(`var ${e.name}: array<${Mn(e.dtype)}, ${e.size}>;`)}_collectLocalBuffers(e){let t=new Set;for(let[,a]of e.bufferMap)t.add(a.name);let n=new Set;for(let a of this._sharedBuffers)n.add(a.name);let s=new Map,o=l(a=>!t.has(a)&&!n.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 yu(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:n,maxPos:s}=this._computeBufferLiveness(e,t),o=[...t.keys()].sort((d,m)=>n.get(d)-n.get(m)||s.get(d)-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,h=this._crossExtent||new Set,f=l((d,m)=>!this._serializeThreads&&this._threadBindings.size>0&&!c.has(d)&&!h.has(d)&&m.numel()>1&&m.numel()<=p,"scalarEligible");for(let d of o){let m=t.get(d),g=m.numel()>0?m.numel():this._estimateBufferSize(m),y=Math.max(g,1),_=n.get(d),b=s.get(d);if(f(d,m)){let x=`_s${u++}`;this._slotDecls.push({name:x,dtype:m.dtype,size:1,scalar:!0}),this._localSlots.set(d,x),this._scalarSlotNames.add(x);continue}let w=i.get(m.dtype);w||(w=new ku((x,k)=>x.freeAt-k.freeAt),i.set(m.dtype,w));let T=null,S=w.peek();if(S&&S.freeAt<_&&(T=w.pop()),T)y>T.decl.size&&(T.decl.size=y);else{let x={name:`_lt${a++}`,dtype:m.dtype,size:y};this._slotDecls.push(x),T={decl:x}}T.freeAt=b,this._localSlots.set(d,T.decl.name),w.push(T)}}_computeBufferLiveness(e,t){return this._livenessWalk(e.body,t)}_livenessWalk(e,t){let n=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,h=l(d=>{t.has(d)&&(n.has(d)||n.set(d,a),s.set(d,a),u.length>0&&p.add(d))},"touch"),f=l(d=>{if(!d)return;a++;let m=o.has(d.type),g=!1;m&&(u.length===0&&(p=new Set,c=a,g=!0),u.push(d)),(d.type==="BufferLoadNode"||d.type==="BufferStoreNode"||d.type==="LIRFlatLoadNode"||d.type==="LIRFlatStoreNode")&&d.buffer&&h(d.buffer.name),d.type==="LIRAccumulatorNode"&&(d.flushStore&&d.flushStore.buffer&&h(d.flushStore.buffer.name),d.initLoad&&d.initLoad.buffer&&h(d.initLoad.buffer.name));for(let y of i){let _=d[y];if(_)if(Array.isArray(_))for(let b of _)f(b);else typeof _=="object"&&f(_)}if(m&&(u.pop(),g)){let y=a;for(let _ of p)n.set(_,Math.min(n.get(_),c)),s.set(_,Math.max(s.get(_),y));p=null}},"walk");return f(e),{minPos:n,maxPos:s}}_scanAllocateNodes(e,t){let n=[e];for(;n.length>0;){let s=n.pop();if(s){s.type==="AllocateNode"&&t.add(s.buffer.name);for(let o of en(s))n.push(o)}}}_scanBufferRefs(e,t){let n=[e];for(;n.length>0;){let s=n.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 en(s))n.push(o)}}}_estimateBufferSize(e){return cd(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(),n=t>0?t:this._estimateBufferSize(e.buffer);n>0&&this._emit(`var ${e.buffer.name}: array<${Mn(e.buffer.dtype)}, ${n}>;`)}}_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,n=Mn(t),s=this._numericExpr(e.value,t);this._emit(`var ${e.variable.name}: ${n} = ${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),n=this._packedBufAccess(e.buffer.name,t),s=this._numericExpr(e.value,e.buffer.dtype);this._emit(`${n} = ${s};`)}_visitLIRFlatStore(e){let t=this._exprToWGSL(e.offsetExpr),n=this._packedBufAccess(e.buffer.name,t),s=this._numericExpr(e.value,e.buffer.dtype);this._emit(`${n} = ${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,n=Mn(e.dtype||this._defaultDtype);this._emit(`var ${t}: ${n} = ${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"&&(JL.has(e.op)||ZL.has(e.op))):!1}_numericExpr(e,t){if(this._isBoolExpr(e)){let n=t?Mn(t):"i32";return n==="f32"?`select(0.0, 1.0, ${this._exprToWGSL(e)})`:`select(${n}(0), ${n}(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),n=this._numericExpr(e.b);return e.op==="//"?`(${t} / ${n})`:e.op==="%"?`(${t} % ${n})`:`(${t} ${e.op} ${n})`}case"CompareNode":return`(${this._numericExpr(e.a)} ${Ka(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`${Mn(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,n=new Array(t);for(let i=0;i<t;i++)n[i]=this._exprToWGSL(e.args[i]);let s=n.join(", ");if(e.externName==="fmod")return`(${n[0]} % ${n[1]})`;if(e.externName==="erf")return Bv(n[0]);if(e.externName==="erfc")return`(1.0 - ${Bv(n[0])})`;if(e.externName==="lgamma")return QL(n[0]);if(e.externName==="gamma")return XL(n[0]);if(e.externName==="log10")return`(log(${n[0]}) * ${1/Math.LN10})`;let o=qS(e.externName);if(o===e.externName&&!US(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 n=this._wgPoolOffsets.get(e);return n.offset===0?`${n.pool}[${t}]`:`${n.pool}[${n.offset}u + u32(${t})]`}if(this._packedMode&&this._packedOffsets&&this._packedOffsets.has(e)){let n=this._packedOffsets.get(e);return n.offset===0?`${n.storage}[${t}]`:`${n.storage}[${n.offset}u + u32(${t})]`}if(this._localSlots&&this._localSlots.has(e)){let n=this._localSlots.get(e);return this._scalarSlotNames.has(n)?n:`${n}[${t}]`}return`${e}[${t}]`}_flatIndex(e,t){return du(e,t,n=>this._exprToWGSL(n),(n,s)=>this._computeDynamicStride(n,s),!1)}_computeDynamicStride(e,t){return ri(e,t+1,(n,s)=>this._resolveShapeParam(n,s))}_resolveShapeParam(e,t){return ni(this._primFunc,e,t,n=>`i32(_shapes.${n.name})`,"WebGPU","wgsl")}}});function Iv(r){let{source:e,name:t,metadata:n}=r,s=[];for(let i=0;i<n.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 Ev=M(()=>{"use strict";l(Iv,"buildSnippet")});function re(r){let e=[];do{let t=r&127;r>>>=7,r&&(t|=128),e.push(t)}while(r);return e}function p2(r){let e=[],t=!0;for(;t;){let n=r&127;r>>=7,r===0&&!(n&64)||r===-1&&n&64?t=!1:n|=128,e.push(n)}return e}function f2(r){r=BigInt(r);let e=[],t=!0;for(;t;){let n=Number(r&0x7fn);r>>=7n,r===0n&&!(n&64)||r===-1n&&n&64?t=!1:n|=128,e.push(n)}return e}function Sd(r){let e=new TextEncoder().encode(r);return[...re(e.length),...e]}function h2(r){let e=new ArrayBuffer(4);return new Float32Array(e)[0]=r,[...new Uint8Array(e)]}function d2(r){let e=new ArrayBuffer(8);return new Float64Array(e)[0]=r,[...new Uint8Array(e)]}function Ws(r,e){for(let t=0;t<e.length;t++)r.push(e[t]);return r}function vu(r,e){let t=[r];return Ws(t,re(e.length)),Ws(t,e),t}function Zi(r){let e=[];Ws(e,re(r.length));for(let t of r)Ws(e,t);return e}function m2(r){let e=[],t=0;for(;t<r.length;){let n=r[t];if(n<=" "){t++;continue}if(n===";"&&r[t+1]===";"){t=r.indexOf(`
+`,t),t<0&&(t=r.length);continue}if(n==="("&&r[t+1]===";"){t=r.indexOf(";)",t+2),t=t<0?r.length:t+2;continue}if(n==="("||n===")"){e.push(n),t++;continue}if(n==='"'){let o=t+1;for(;o<r.length&&r[o]!=='"';)o++;e.push(r.substring(t,o+1)),t=o+1;continue}let s=t;for(;s<r.length&&r[s]>" "&&r[s]!=="("&&r[s]!==")";)s++;e.push(r.substring(t,s)),t=s}return e}function g2(r){let e=0,t=l(()=>r[e++],"eat"),n=l(()=>r[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=[],h=[],f=[],d=-1,m=-1;for(;n()!==")";){if(n()!=="("){e++;continue}e++;let y=t();if(y==="memory"){for(;n()==="(";)g();n()!==")"&&(i=parseInt(t())||1),n()!==")"&&(a=parseInt(t())||256),s(")")}else if(y==="import"){let _=t().replace(/"/g,""),b=t().replace(/"/g,"");s("("),s("func"),n().startsWith("$")&&e++;let w=[],T=[];for(;n()==="(";){e++;let S=t();if(S==="param")for(;n()!==")";){let x=t();x==="f32"?w.push(Rc):x==="f64"?w.push(Bc):x==="i32"?w.push(Oc):x==="v128"&&w.push(Mc)}else if(S==="result")for(;n()!==")";){let x=t();x==="f32"?T.push(Rc):x==="f64"?T.push(Bc):x==="i32"?T.push(Oc):x==="v128"&&T.push(Mc)}s(")")}s(")"),s(")"),o.push({module:_,name:b,params:w,results:T})}else if(y==="func"){for(;n()==="(";){let b=e;e++;let w=t();if(w==="export")u=t().replace(/"/g,""),s(")");else if(w==="param"){let T="";for(;n()!==")";){let S=t();if(S.startsWith("$")){T=S.replace("$","");continue}let x=null;S==="i32"?x=Oc:S==="i64"?x=nw:S==="f32"?x=Rc:S==="f64"?x=Bc:S==="v128"&&(x=Mc),x!==null&&(p.push(x),c.push(T),T="")}s(")")}else if(w==="result"){for(;n()!==")";)t();s(")")}else if(w==="local"){for(;n()!==")";){let T=t();T.startsWith("$")?f.push(T.replace("$","")):T==="i32"?h.push(Oc):T==="i64"?h.push(nw):T==="f32"?h.push(Rc):T==="f64"?h.push(Bc):T==="v128"&&h.push(Mc)}s(")")}else{e=b;break}}d=e;let _=1;for(;_>0;){let b=t();b==="("?_++:b===")"&&_--}m=e-1}else g()}function g(){if(n()!=="("){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:h,funcLocalNames:f,bodyTokens:r.slice(d,m)}}function y2(r,e,t){let n=[],s=0,o=l(()=>r[s],"peek"),i=l(()=>r[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 h(y,_){let b=y.indexOf(_);return b>=0?y.substring(b+_.length):""}l(h,"extractName");function f(){for(;s<r.length&&o()!==")";){let y=o();if(y==="("){s++;let _=i();if(_==="i32.const")n.push(65),n.push(...p2(parseInt(i()))),m(")");else if(_==="i64.const")n.push(66),n.push(...f2(i())),m(")");else if(_==="f32.const")n.push(67),n.push(...h2(parseFloat(i()))),m(")");else if(_==="f64.const")n.push(68),n.push(...d2(parseFloat(i()))),m(")");else if(_==="local.get")n.push(32),n.push(...re(p(i()))),m(")");else if(_==="local.set")n.push(33),n.push(...re(p(i()))),m(")");else if(_==="block"){let b=i().replace("$","");a.push(b),n.push(2,ew),f(),n.push(11),a.pop(),m(")")}else if(_==="loop"){let b=i().replace("$","");a.push(b),n.push(3,ew),f(),n.push(11),a.pop(),m(")")}else if(_==="if"){let b=ew;if(o()==="("){let w=s;if(s++,o()==="result"){i();let T=i();b=T==="f32"?Rc:T==="f64"?Bc:T==="v128"?Mc:T==="i64"?nw:Oc,m(")")}else s=w}for(n.push(4,b),a.push("_if");s<r.length&&o()!==")"&&o()==="(";){let w=s;s++;let T=i();if(T==="then")f(),m(")");else if(T==="else")n.push(5),f(),m(")");else{s=w;break}}n.push(11),a.pop(),m(")")}else s-=2,g();continue}if(s++,y==="br_if")n.push(13),n.push(...re(d(c())));else if(y==="br")n.push(12),n.push(...re(d(c())));else if(y.startsWith("call")){n.push(16);let _=h(y,"math_")||(o()&&o().startsWith("$")?h(i(),"math_"):""),b=t.get(_);if(b===void 0)throw new Error(`WAT encoder: call to unresolved import '${_}'`);n.push(...re(b))}else y==="local.get"?(n.push(32),n.push(...re(p(c())))):y==="local.set"?(n.push(33),n.push(...re(p(c())))):y.endsWith(".extract_lane")||y.endsWith(".replace_lane")?(n.push(...tw.get(y)),n.push(parseInt(i(),10))):tw.has(y)&&n.push(...tw.get(y))}}l(f,"emitBlock");function d(y){for(let _=a.length-1;_>=0;_--)if(a[_]===y)return a.length-1-_;return 0}l(d,"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"),f(),n}function Nu(r){let e=m2(r),t=g2(e),n=new Map,s=t.funcParams.length;for(let x=0;x<t.funcParamNames.length;x++)t.funcParamNames[x]&&n.set(t.funcParamNames[x],x);for(let x=0;x<t.funcLocalNames.length;x++)n.set(t.funcLocalNames[x],s+x);let o=new Map;for(let x=0;x<t.imports.length;x++)o.set(t.imports[x].name,x);let i=[],a=new Map;function u(x,k){let A=x.join(",")+">"+k.join(",");if(a.has(A))return a.get(A);let N=i.length;return i.push({params:x,results:k}),a.set(A,N),N}l(u,"getType");for(let x of t.imports)x.typeIdx=u(x.params,x.results);let p=u(t.funcParams,[]),c=vu(s2,Zi(i.map(x=>[c2,...re(x.params.length),...x.params,...re(x.results.length),...x.results]))),h=[];t.imports.length>0&&(h=vu(o2,Zi(t.imports.map(x=>[...Sd(x.module),...Sd(x.name),0,...re(x.typeIdx)]))));let f=vu(i2,Zi([[...re(p)]])),d=vu(a2,Zi([[1,...re(t.memMin),...re(t.memMax)]])),m=t.imports.length,g=vu(u2,Zi([[...Sd("memory"),2,...re(0)],[...Sd(t.funcExportName),0,...re(m)]])),y=[];if(t.funcLocals.length>0){let x=[],k=t.funcLocals[0],A=1;for(let N=1;N<t.funcLocals.length;N++)t.funcLocals[N]===k?A++:(x.push([...re(A),k]),k=t.funcLocals[N],A=1);x.push([...re(A),k]),Ws(y,Zi(x))}else y.push(0);let _=y2(t.bodyTokens,n,o),b=[];Ws(b,y),Ws(b,_),b.push(11);let w=[];Ws(w,re(b.length)),Ws(w,b);let T=vu(l2,Zi([w])),S=[];for(let x of[n2,r2,c,h,f,d,g,T])Ws(S,x);return new Uint8Array(S)}var n2,r2,s2,o2,i2,a2,u2,l2,Oc,nw,Bc,Rc,Mc,c2,ew,tw,kd=M(()=>{"use strict";n2=[0,97,115,109],r2=[1,0,0,0],s2=1,o2=2,i2=3,a2=5,u2=7,l2=10,Oc=127,nw=126,Bc=124,Rc=125,Mc=123,c2=96,ew=64;l(re,"uleb");l(p2,"sleb");l(f2,"slebBig");l(Sd,"encStr");l(h2,"encF32");l(d2,"encF64");l(Ws,"pushAll");l(vu,"section");l(Zi,"vec");tw=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(m2,"tokenize");l(g2,"parseModule");l(y2,"encodeBody");l(Nu,"encodeWat")});function Lv(r){let{source:e,name:t,metadata:n}=r,{bufferOffsets:s,imports:o}=n,i=[...s.entries()],a=Nu(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,h]of i)u.push(`// ${c}: new Float32Array(memory.buffer, ${h}, 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 Cv=M(()=>{"use strict";kd();l(Lv,"buildSnippet")});function Dv(r){let{source:e,name:t,metadata:n}=r,{bindings:s,dispatchSize:o}=n,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"?_2(i,a):a.packed?b2(i,a):a.mode==="read_write"?w2(i,a):T2(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(""),x2(i,s),i.push(""),i.push("gpuBuffers.forEach(b => b.destroy());"),i.push("device.destroy();"),i.push("})();"),i.join(`
+`)}function _2(r,e){r.push(`// binding ${e.index}: uniform (_shapes)`),r.push("{"),r.push("  const shapeData = new Uint32Array([/* shape values */]);"),r.push("  const size = Math.max(Math.ceil(shapeData.byteLength / 16) * 16, 16);"),r.push("  const buf = device.createBuffer({ size, usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST });"),r.push("  device.queue.writeBuffer(buf, 0, shapeData);"),r.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),r.push("  gpuBuffers.push(buf);"),r.push("}")}function b2(r,e){let t=e.mode==="read_write",n=t?"GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST":"GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST";if(r.push(`// binding ${e.index}: packed ${e.name} (${e.packed.length} buffers, ${e.packedSize} elems)`),r.push("{"),r.push(`  const buf = device.createBuffer({ size: ${e.packedSize*4}, usage: ${n}, mappedAtCreation: true });`),r.push("  const mapped = new Float32Array(buf.getMappedRange());"),t)r.push("  mapped.fill(0);");else for(let s of e.packed)r.push(`  // mapped.set(${s.name}_data, ${s.offset}); // size=${s.size}`);r.push("  buf.unmap();"),r.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),r.push("  gpuBuffers.push(buf);"),r.push("}")}function w2(r,e){r.push(`// binding ${e.index}: output ${e.name}`),r.push("{"),r.push("  const size = N * 4; // N = output element count"),r.push("  const buf = device.createBuffer({ size, usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST, mappedAtCreation: true });"),r.push("  new Float32Array(buf.getMappedRange()).fill(0);"),r.push("  buf.unmap();"),r.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),r.push("  gpuBuffers.push(buf);"),r.push("}")}function T2(r,e){r.push(`// binding ${e.index}: input ${e.name}`),r.push("{"),r.push(`  const data = ${e.name}_data; // Float32Array`),r.push("  const buf = device.createBuffer({ size: Math.max(data.byteLength, 4), usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST, mappedAtCreation: true });"),r.push("  new Float32Array(buf.getMappedRange()).set(data);"),r.push("  buf.unmap();"),r.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),r.push("  gpuBuffers.push(buf);"),r.push("}")}function x2(r,e){if(e.filter(n=>n.mode==="read_write").length===0){r.push("device.queue.submit([encoder.finish()]);");return}for(let n=0;n<e.length;n++){let s=e[n];if(s.mode==="read_write")if(s.packed)for(let o of s.packed)r.push("{"),r.push(`  const rb = device.createBuffer({ size: ${o.size*4}, usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST });`),r.push(`  encoder.copyBufferToBuffer(gpuBuffers[${n}], ${o.offset*4}, rb, 0, ${o.size*4});`),r.push("  device.queue.submit([encoder.finish()]);"),r.push("  await rb.mapAsync(GPUMapMode.READ);"),r.push(`  console.log("${o.name}:", new Float32Array(rb.getMappedRange()));`),r.push("  rb.unmap(); rb.destroy();"),r.push("}");else r.push("{"),r.push(`  const src = gpuBuffers[${n}];`),r.push("  const rb = device.createBuffer({ size: src.size, usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST });"),r.push("  encoder.copyBufferToBuffer(src, 0, rb, 0, src.size);"),r.push("  device.queue.submit([encoder.finish()]);"),r.push("  await rb.mapAsync(GPUMapMode.READ);"),r.push(`  console.log("${s.name}:", new Float32Array(rb.getMappedRange()));`),r.push("  rb.unmap(); rb.destroy();"),r.push("}")}}var Pv=M(()=>{"use strict";l(Dv,"buildSnippet");l(_2,"emitShapeBinding");l(b2,"emitPackedBinding");l(w2,"emitOutputBinding");l(T2,"emitInputBinding");l(x2,"emitReadback")});function $v(r){let{source:e,name:t,metadata:n}=r,{blockDim:s,gridDim:o,sharedMemBytes:i,params:a}=n,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 Fv=M(()=>{"use strict";l($v,"buildSnippet")});function vd(r,e){zv.set(r,e),e.runtimeKind&&e.snippet&&Vv.set(e.runtimeKind,e.snippet)}function Gv(r){return zv.get(r)||null}function jv(r){return Vv.get(r)||null}var zv,Vv,qv=M(()=>{"use strict";Ks();Tv();Av();Ov();Mv();Ev();Cv();Pv();Fv();zv=new Map,Vv=new Map;l(vd,"registerCodegen");l(Gv,"getCodegenEntry");l(jv,"getSnippetBuilder");vd(Ot.CPU,{runtimeKind:"js",snippet:Iv,compile(r,e){return{source:new yd(e).generate(r),metadata:{kind:"js",paramCount:r.params.length}}}});vd(Ot.WASM,{runtimeKind:"wasm",snippet:Lv,compile(r,e){let t=new bd(e).generate(r),n={kind:"wasm",memoryPages:t.memoryPages,bufferOffsets:t.bufferOffsets,imports:t.imports,params:t.params,bufferMap:r.bufferMap};return t.parallel&&(n.parallel=t.parallel),{source:t.wat,metadata:n}}});vd(Ot.WEBGPU,{runtimeKind:"webgpu",snippet:Dv,compile(r,e){let t=new xd(e).generate(r);return{source:t.source,metadata:{kind:"webgpu",workgroupSize:t.workgroupSize,dispatchSize:t.dispatchSize,sharedMemBytes:t.sharedMemBytes,params:t.params,bindings:t.bindings}}}});vd(Ot.CUDA,{runtimeKind:"cuda",snippet:$v,compile(r,e,t){if(t&&t.matmulBackend==="cublas"&&r.cublasInfo)return{source:"",metadata:{kind:"cuda",cublas:r.cublasInfo,outputIndices:[r.cublasInfo.cIdx]}};let n=new _d(e).generate(r);return{source:n.source,metadata:{kind:"cuda",blockDim:n.blockDim,gridDim:n.gridDim,sharedMemBytes:n.sharedMemBytes,params:n.params,outputIndices:n.outputIndices,scratch:n.scratch}}}})});function Uv(r){let e=[],t=[r.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 n=null;for(let m of e)if(m.name.includes("matmul"))m.reads.length>=2&&m.writes.length>=1&&(n=m);else return null;if(!n)return null;let s=n.reads[0].buffer,o=n.reads[1].buffer,i=n.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 r.bufferMap)c.push(m.name);let h=c.indexOf(s.name),f=c.indexOf(o.name),d=c.indexOf(i.name);return h<0||f<0||d<0?null:{M:a,N:u,K:p,aIdx:h,bIdx:f,cIdx:d}}var Ic,Hs,Ec=M(()=>{"use strict";qv();l(Uv,"detectPureMatmul");Ic=class{static{l(this,"CompiledKernel")}constructor(e,t,n,s={}){this.name=e,this.source=t,this.target=n,this.metadata=s}snippet(){let e=jv(this.metadata.kind);if(!e)throw new Error(`No snippet for kind: ${this.metadata.kind}`);return e(this)}},Hs=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)||Gv(this.target.kind);if(!t)throw new Error(`Unsupported target kind: ${this.target.kind}`);let{source:n,metadata:s}=t.compile(e,this.target,this);return new Ic(e.name,n,this.target,s)}compileAll(e){return e.map(t=>this.compile(t))}}});function Wv(r,e){Kv.set(r,e)}function Hv(r){return Kv.get(r)||null}var Kv,rw=M(()=>{"use strict";Kv=new Map;l(Wv,"registerMeasurer");l(Hv,"getMeasurer")});var Yv={};nt(Yv,{runWasmParallel:()=>S2});async function S2(r,e,t,n,s){let{exports:o,memory:i,bufferOffsets:a}=r,u=o[e],p=[...a.values()],c=Math.min(p.length,t.length);for(let f=0;f<c;f++){let d=t[f];ArrayBuffer.isView(d)&&new d.constructor(i.buffer,p[f],d.length).set(d)}let h=p.slice(0,c);if(n)for(let f of n)h.push(f);h.push(0,s.extent),u(...h);for(let f=0;f<c;f++){let d=t[f];ArrayBuffer.isView(d)&&d.set(new d.constructor(i.buffer,p[f],d.length))}}var Jv=M(()=>{"use strict";l(S2,"runWasmParallel")});var Nd={};nt(Nd,{deviceClipGradNorm:()=>A2,measureCudaKernel:()=>v2,runCudaKernel:()=>k2,runCudaPlan:()=>N2});function k2(){throw new Error("CUDA runtime is not available in the browser")}function v2(){throw new Error("CUDA runtime is not available in the browser")}function N2(){throw new Error("CUDA runtime is not available in the browser")}function A2(){throw new Error("CUDA runtime is not available in the browser")}var Ad=M(()=>{"use strict";l(k2,"runCudaKernel");l(v2,"measureCudaKernel");l(N2,"runCudaPlan");l(A2,"deviceClipGradNorm")});function ii(){return Zv}function O2(r){Zv=r}var Qi,ks,Od,sw,Ys,Au,Zv,ar=M(()=>{"use strict";Nn();Qi=class{static{l(this,"Device")}type;index;constructor(e,t=0){this.type=e,this.index=t}dispatchKey(){return ja(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}`}},ks=new Qi("cpu"),Od=new Qi("gpu"),sw=new Qi("wasm"),Ys=new Qi("webgpu"),Au=new Qi("meta"),Zv=ks;l(ii,"getDefaultDevice");l(O2,"setDefaultDevice")});function Xv(r){let e=Qv.get(r)??null;return e||(e=Nt.findOp(r),e&&Qv.set(r,e)),e||null}var Qv,eN=M(()=>{"use strict";$i();Qv=new Map;l(Xv,"getHandle")});function vs(r,e){return r<0?e+r:r}function xn(r){let e=r.length,t=new Array(e),n=1;for(let s=e-1;s>=0;s--)t[s]=n,n*=r[s];return t}function cn(r){let e=1;for(let t=0;t<r.length;t++)e*=r[t];return e}function ow(r,e){let t=r.length;if(t===0)return!0;let n=1;for(let s=t-1;s>=0;s--){let o=r[s];if(o===0)return!0;if(o!==1&&e[s]!==n)return!1;n*=o}return!0}function Ou(r,e){let t=r.length,n=e.length,s=Math.max(t,n),o=new Array(s);for(let i=0;i<s;i++){let a=i<t?r[t-1-i]:1,u=i<n?e[n-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 Bd(r,e){let t=r.length,n=e.length;if(t===1&&n===1)return[];let s=t===1?[1,r[0]]:[...r],o=n===1?[e[0],1]:[...e],i=s.length,a=o.length,u=Ou(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 n===1&&c.add(p.length-1),t===1&&c.add(p.length-2),p.filter((h,f)=>!c.has(f))}function tN(r,e,t){let n=cn(r),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]=n/o|0,i[s]*o!==n))return null}else if(cn(i)!==n)return null;return ow(r,e)?{sizes:i,strides:xn(i),needsCopy:!1}:{sizes:i,strides:xn(i),needsCopy:!0}}var jr=M(()=>{"use strict";l(vs,"normalizeAxis");l(xn,"computeStrides");l(cn,"computeNumel");l(ow,"isContiguous");l(Ou,"broadcastShapes");l(Bd,"matmulOutputShape");l(tN,"inferReshape")});function nN(r,e){if(r.length<4)return!1;let t=r.length,n=1,s=t===4?B2:R2;if(!s)return!1;for(let o=t-1;o>=0;o--){let i=s[o];if(i===void 0||r[i]===0||e[i]!==n)return!1;n*=r[i]}return!0}var B2,R2,iw=M(()=>{"use strict";B2=Object.freeze([0,2,3,1]),R2=Object.freeze([0,2,3,4,1]);l(nN,"isChannelsLast")});var Bt,Oo=M(()=>{"use strict";Nn();jr();iw();iw();Bt=class{static{l(this,"TensorImpl")}_storage;_storageOffset;_sizes;_strides;_dtype;_device;_numel;_keySet;_autogradMeta;_version;_contiguousCache;constructor(e,t,n,s,o,i){this._storage=e,this._storageOffset=t,this._sizes=Object.freeze([...n]),this._strides=Object.freeze(s?[...s]:xn(n)),this._dtype=o,this._device=i,this._numel=cn(n),this._keySet=Pr.fromKey(ja(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=ow(this._sizes,this._strides)),this._contiguousCache):e==="channels_last"?nN(this._sizes,this._strides):!1}setSizesAndStrides(e,t){this._sizes=Object.freeze([...e]),this._strides=Object.freeze(t?[...t]:xn(e)),this._numel=cn(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=Pr.fromKey(ja(this._device.type));if(this._autogradMeta&&this._autogradMeta.requiresGrad){let t=ja(this._device.type);e=e.add(RS(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 Bu,aw=M(()=>{"use strict";Bu=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 sN(r,e,t,n,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]=md(o,r[n+p*t[s]]);else for(let p=0;p<i;p++)u[p]=r[n+p*t[s]];return u}let a=new Array(i);for(let u=0;u<i;u++)a[u]=sN(r,e,t,n+u*t[s],s+1,o);return a}function M2(){if(!uw)throw new Error("Autograd engine not initialized. Import autograd/engine.js first.");return uw}function oN(r){uw=r}var rN,We,uw,jn=M(()=>{"use strict";Oo();aw();xu();rN=new Set(["f16","bf16","i64"]),We=class r{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,n=this._impl.storageOffset;return t&&n===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 Bt(this._impl.storage,this._impl.storageOffset,this._impl.sizes(),this._impl.strides(),this._impl.dtype,this._impl.device);return new r(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 rN.has(this._impl.dtype)?md(this._impl.dtype,t):t}toArray(){let e=this.shape,t=this.strides,n=this._impl.storage.data,s=this._impl.storageOffset,o=rN.has(this._impl.dtype)?this._impl.dtype:null;return e.length===0?o?md(o,n[s]):n[s]:sN(n,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 n=this._impl.sizes(),s=this._impl.strides(),o=[],i=[];for(let p=0;p<n.length;p++)p!==e&&(o.push(n[p]),i.push(s[p]));let a=this._impl.storageOffset+t*s[e],u=new Bt(this._impl.storage,a,o,i,this._impl.dtype,this._impl.device);return new r(u)}_ensureAutogradMeta(){if(!this._impl.autogradMeta){let e=new Bu;e.versionAtCreation=this._impl.version,this._impl.setAutogradMeta(e)}return this._impl.autogradMeta}backward(e){let{backward:t}=M2();t(this,e)}};l(sN,"_toNestedArray");uw=null;l(M2,"require_autograd_engine");l(oN,"setAutogradEngine")});function Dt(r){return E2[r]||Float32Array}function Rd(r,e){if(r===e)return r;let t=Ru.get(r)??0,n=Ru.get(e)??0;return t>=n?r:e}function Xe(r,e){let t=uN.get(r);t||(t=new Set,uN.set(r,t)),t.add(e)}function Xi(r){return I2(r)}var W,I2,E2,iN,aN,Ru,uN,ts=M(()=>{"use strict";Oe();W=zn,I2=Gs,E2=Object.freeze({[W.F16]:Uint16Array,[W.BF16]:Uint16Array,[W.F32]:Float32Array,[W.F64]:Float64Array,[W.I8]:Int8Array,[W.I16]:Int16Array,[W.I32]:Int32Array,[W.I64]:BigInt64Array,[W.UI8]:Uint8Array,[W.BOOL]:Uint8Array,[W.INDEX]:Int32Array});l(Dt,"typedArrayCtor");iN=[W.F16,W.BF16,W.F32,W.F64],aN=[W.UI8,W.I8,W.I16,W.I32,W.I64],Ru=new Map;for(let r=0;r<iN.length;r++)Ru.set(iN[r],100+r);for(let r=0;r<aN.length;r++)Ru.set(aN[r],r);Ru.set(W.BOOL,-1);Ru.set(W.INDEX,50);l(Rd,"resultDtype");uN=new Map;l(Xe,"_addCastPair");Xe(W.BOOL,W.I8);Xe(W.BOOL,W.I16);Xe(W.BOOL,W.I32);Xe(W.BOOL,W.I64);Xe(W.BOOL,W.F16);Xe(W.BOOL,W.F32);Xe(W.BOOL,W.F64);Xe(W.UI8,W.I16);Xe(W.UI8,W.I32);Xe(W.UI8,W.I64);Xe(W.UI8,W.F16);Xe(W.UI8,W.F32);Xe(W.UI8,W.F64);Xe(W.I8,W.I16);Xe(W.I8,W.I32);Xe(W.I8,W.I64);Xe(W.I8,W.F16);Xe(W.I8,W.F32);Xe(W.I8,W.F64);Xe(W.I16,W.I32);Xe(W.I16,W.I64);Xe(W.I16,W.F32);Xe(W.I16,W.F64);Xe(W.I32,W.I64);Xe(W.I32,W.F64);Xe(W.F16,W.F32);Xe(W.F16,W.F64);Xe(W.BF16,W.F32);Xe(W.BF16,W.F64);Xe(W.BOOL,W.BF16);Xe(W.UI8,W.BF16);Xe(W.I8,W.BF16);Xe(W.F32,W.F64);l(Xi,"dtypeSize")});function lw(r){return lN.get(r)||lN.get("cpu")}var Md,Id,lN,Lc,cN=M(()=>{"use strict";ts();ar();Md=class{static{l(this,"CPUAllocator")}allocate(e,t){let n=Dt(t);return new n(e)}free(e){}},Id=class{static{l(this,"MetaAllocator")}allocate(){return null}free(e){}},lN=new Map([["cpu",new Md],["wasm",new Md],["meta",new Id],["lazy",new Id]]);l(lw,"getAllocator");Lc=class r{static{l(this,"StorageImpl")}static#e=null;_data;_nbytes;_device;_allocator;_refCount;static setHostReadHook(e){r.#e=e}constructor(e,t,n,s){this._data=e,this._nbytes=t,this._device=n,this._allocator=s,this._refCount=1}static allocate(e,t,n){let s=lw(n.type),o=Xi(t),i=Math.max(Math.ceil(e/o),1),a=s.allocate(i,t);return new r(a,e,n,s)}static fromData(e,t){let n=e?e.byteLength:0,s=lw(t.type);return new r(e,n,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 r.#e&&this._data&&r.#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 n=Xi(t),s=Math.max(Math.ceil(e/n),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=lw(this._device.type),t=null;if(this._data){r.#e&&r.#e(this._data);let n=this._data.constructor;t=new n(this._data.length),t.set(this._data)}return new r(t,this._nbytes,this._device,e)}}});var En,ea=M(()=>{"use strict";cN();En=class r{static{l(this,"Storage")}_impl;constructor(e){this._impl=e}static allocate(e,t,n){return new r(Lc.allocate(e,t,n))}static fromData(e,t){return new r(Lc.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 r(this._impl.clone())}resize(e,t){this._impl.resize(e,t)}}});function Pt(r,e){let t=e?.dtype??W.F32,n=e?.device??ii(),s=e?.requiresGrad??!1;if(ArrayBuffer.isView(r)&&"length"in r)return L2(r,e?.shape,t,n,s);if(Array.isArray(r)){let{flat:o,shape:i}=D2(r),a=e?.shape??i;return C2(o,a,t,n,s)}if(typeof r=="number")return pN(r,t,n,s);throw new Error("Unsupported data type for tensor()")}function ta(r,e,t,n){let s=n?.device??ii(),o=n?.requiresGrad??!1,i=xn(e),a=En.fromData(r,s),u=n?.offset??0,p=new Bt(a,u,e,i,t,s),c=new We(p);return o&&c.requiresGrad_(!0),c}function Dc(r,e){let t=e?.dtype??W.F32,n=e?.device??ii();return pN(r,t,n,e?.requiresGrad??!1)}function pN(r,e,t,n){let s=Dt(e),o=new s(1),i=o;i[0]=cw.has(e)?dd(e,r):r;let a=En.fromData(o,t),u=new Bt(a,0,[],[],e,t),p=new We(u);return n&&p.requiresGrad_(!0),p}function L2(r,e,t,n,s){let o=Dt(t),i=new o(r.length),a=i;if(cw.has(t))for(let d=0;d<r.length;d++)a[d]=dd(t,r[d]);else for(let d=0;d<r.length;d++)a[d]=r[d];let u=e??[r.length],p=xn(u),c=En.fromData(i,n),h=new Bt(c,0,u,p,t,n),f=new We(h);return s&&f.requiresGrad_(!0),f}function C2(r,e,t,n,s){let o=Dt(t),i=new o(r.length),a=i;if(cw.has(t))for(let f=0;f<r.length;f++)a[f]=dd(t,r[f]);else for(let f=0;f<r.length;f++)a[f]=r[f];let u=xn(e),p=En.fromData(i,n),c=new Bt(p,0,e,u,t,n),h=new We(c);return s&&h.requiresGrad_(!0),h}function D2(r){let e=[],t=r;for(;Array.isArray(t);)e.push(t.length),t=t[0];let n=[];return fN(r,n,e,0),{flat:n,shape:e}}function fN(r,e,t,n){if(n===t.length-1){for(let s=0;s<r.length;s++)e.push(r[s]);return}for(let s=0;s<r.length;s++)fN(r[s],e,t,n+1)}var cw,qr=M(()=>{"use strict";jn();Oo();ea();ts();ar();jr();xu();l(Pt,"tensor");l(ta,"fromBuffer");l(Dc,"scalar");cw=new Set(["f16","bf16","i64"]);l(pN,"_fromScalar");l(L2,"_fromTypedArray");l(C2,"_fromFlatArray");l(D2,"_flattenNested");l(fN,"_recurFlatten")});var Ft={};nt(Ft,{_dispatch:()=>ot,abs:()=>Pc,add:()=>ae,argmax:()=>Ud,argmin:()=>kN,argsort:()=>LN,batch_norm:()=>fw,broadcast_in_dim:()=>ON,cat:()=>sa,ceil:()=>wN,chunk:()=>BN,clamp:()=>Vd,clone:()=>AN,contiguous:()=>ci,conv2d:()=>hw,cos:()=>mN,cumsum:()=>IN,div:()=>$t,dot:()=>NN,embedding:()=>z2,eq:()=>$c,erf:()=>gN,erfc:()=>yN,exp:()=>Kr,expand:()=>Lu,fill:()=>F2,flip:()=>MN,floor:()=>bN,gamma:()=>Cd,gather:()=>jd,ge:()=>zd,gelu:()=>Dd,gt:()=>Iu,index_select:()=>li,layer_norm:()=>pw,le:()=>Fd,lgamma:()=>_N,log:()=>Zs,log_softmax:()=>Hd,lt:()=>xN,matmul:()=>lt,max:()=>Qs,maximum:()=>ai,mean:()=>lr,min:()=>SN,minimum:()=>Ld,mul:()=>ee,narrow:()=>Un,ne:()=>TN,neg:()=>Ur,one_hot:()=>Gd,pad:()=>na,permute:()=>Eu,pool2d:()=>Yd,pow:()=>Js,prod:()=>vN,relu:()=>Mu,remainder:()=>$2,repeat:()=>Du,reshape:()=>Cn,roll:()=>RN,rsqrt:()=>hN,scatter:()=>qd,scatter_add:()=>ra,select:()=>Ut,sigmoid:()=>ur,sign:()=>$d,silu:()=>Pd,sin:()=>dN,slice:()=>Kd,softmax:()=>Ro,sort:()=>EN,split:()=>Bo,sqrt:()=>ui,squeeze:()=>Cu,stack:()=>Wr,sub:()=>Pe,sum:()=>ht,tanh:()=>Ns,tile:()=>Wd,topk:()=>CN,transpose:()=>gt,unsqueeze:()=>Sr,where:()=>qn});function P2(r){return typeof r=="object"&&r!==null&&"_impl"in r}function Ln(r,e){return P2(r)?r:Dc(r,{dtype:e.dtype,device:e.device})}function ot(r,...e){let t=Xv(r);if(!t)return Nt.callOp(r,...e);let n=pb(e,t.schema);return Nt.dispatch(t,n,...e)}function se(r,...e){return ot(r,...e)}function Ed(r,...e){return ot(r,...e)}function ae(r,e){return se("add",r,Ln(e,r))}function Pe(r,e){return se("sub",r,Ln(e,r))}function ee(r,e){return se("mul",r,Ln(e,r))}function $t(r,e){return se("div",r,Ln(e,r))}function Ur(r){return se("neg",r)}function Js(r,e){return se("pow",r,Ln(e,r))}function $2(r,e){return se("rem",r,Ln(e,r))}function ai(r,e){return se("maximum",r,Ln(e,r))}function Ld(r,e){return se("minimum",r,Ln(e,r))}function Kr(r){return se("exp",r)}function Zs(r){return se("log",r)}function ui(r){return se("sqrt",r)}function hN(r){return se("rsqrt",r)}function Pc(r){return se("abs",r)}function dN(r){return se("sin",r)}function mN(r){return se("cos",r)}function Ns(r){return se("tanh",r)}function gN(r){return se("erf",r)}function yN(r){return se("erfc",r)}function _N(r){return se("lgamma",r)}function Cd(r){return se("gamma",r)}function ur(r){return se("sigmoid",r)}function Mu(r){return se("relu",r)}function Dd(r){return se("gelu",r)}function Pd(r){return se("silu",r)}function $d(r){return se("sign",r)}function bN(r){return se("floor",r)}function wN(r){return se("ceil",r)}function $c(r,e){return se("eq",r,Ln(e,r))}function TN(r,e){return se("ne",r,Ln(e,r))}function xN(r,e){return se("lt",r,Ln(e,r))}function Fd(r,e){return se("le",r,Ln(e,r))}function Iu(r,e){return se("gt",r,Ln(e,r))}function zd(r,e){return se("ge",r,Ln(e,r))}function qn(r,e,t){return se("where",r,e,t)}function Vd(r,e,t){return se("clamp",r,Ln(e,r),Ln(t,r))}function na(r,e,t,n=0){return se("pad",r,Ln(n,r),e,t)}function Gd(r,e){return se("one_hot",r,e)}function li(r,e,t){return se("index_select",r,t,e)}function jd(r,e,t){return se("gather",r,t,e)}function ra(r,e,t,n){return se("scatter_add",r,t,n,e)}function qd(r,e,t,n){return se("scatter",r,e,t,n)}function ht(r,e,t){return se("sum",r,e,t)}function lr(r,e,t){return se("mean",r,e,t)}function Qs(r,e,t){return se("max",r,e,t)}function SN(r,e,t){return se("min",r,e,t)}function Ud(r,e,t){return se("argmax",r,e,t)}function kN(r,e,t){return se("argmin",r,e,t)}function vN(r,e,t){return se("prod",r,e,t)}function lt(r,e){let t=dw();if(t){let n=t(r,e);if(n!==null)return n}return se("matmul",r,e)}function NN(r,e){return se("dot",r,e)}function sa(r,e){return se("cat",r,e)}function Wr(r,e){return se("stack",r,e)}function AN(r){return se("clone",r)}function F2(r,e){return se("fill",r,e)}function Cn(r,e){return se("reshape",r,e)}function gt(r,e,t){return se("transpose",r,e,t)}function Eu(r,e){return se("permute",r,e)}function ON(r,e,t){return se("broadcast_in_dim",r,e,t)}function Lu(r,e){return se("expand",r,e)}function Kd(r,e,t,n=null,s=1){return se("slice",r,e,t,n,s)}function Sr(r,e){return se("unsqueeze",r,e)}function Cu(r,e=null){return se("squeeze",r,e)}function Un(r,e,t,n){return se("narrow",r,e,t,n)}function Ut(r,e,t){return se("select",r,e,t)}function ci(r){return se("contiguous",r)}function Du(r,e){return se("repeat",r,e)}function Wd(r,e){return se("tile",r,e)}function Bo(r,e,t=0){if(typeof e!="number")return Ed("split",r,e,t);let n=r.shape.length,s=t<0?n+t:t,o=r.shape[s],i=[];for(let a=0;a<o;a+=e)i.push(Math.min(e,o-a));return Ed("split",r,i,t)}function BN(r,e,t=0){return Ed("chunk",r,e,t)}function RN(r,e,t=0){return se("roll",r,e,t)}function MN(r,e){return se("flip",r,Array.isArray(e)?e:[e])}function IN(r,e=0){return se("cumsum",r,e)}function EN(r,e=-1,t=!1){return se("sort",r,e,t)}function LN(r,e=-1,t=!1){return se("argsort",r,e,t)}function CN(r,e,t=-1,n=!0){return Ed("topk",r,e,t,n)}function Ro(r,e){return se("softmax",r,e)}function Hd(r,e){return se("log_softmax",r,e)}function pw(r,e,t,n,s){return se("layer_norm",r,e,t,n,s)}function fw(r,e,t,n,s,o,i){return se("batch_norm",r,e,t,n,s,o,i)}function hw(r,e,t,n,s,o){return se("conv2d",r,e,t,n,s,o)}function Yd(r,e,t,n,s){return se("pool2d",r,e,t,n,s)}function z2(r,e){return se("embedding",r,e)}var be=M(()=>{"use strict";$i();eN();qr();Kn();l(P2,"hasTensorImpl");l(Ln,"_asTensor");l(ot,"_dispatch");l(se,"_dispatchTensor");l(Ed,"_dispatchTensorArray");l(ae,"add");l(Pe,"sub");l(ee,"mul");l($t,"div");l(Ur,"neg");l(Js,"pow");l($2,"remainder");l(ai,"maximum");l(Ld,"minimum");l(Kr,"exp");l(Zs,"log");l(ui,"sqrt");l(hN,"rsqrt");l(Pc,"abs");l(dN,"sin");l(mN,"cos");l(Ns,"tanh");l(gN,"erf");l(yN,"erfc");l(_N,"lgamma");l(Cd,"gamma");l(ur,"sigmoid");l(Mu,"relu");l(Dd,"gelu");l(Pd,"silu");l($d,"sign");l(bN,"floor");l(wN,"ceil");l($c,"eq");l(TN,"ne");l(xN,"lt");l(Fd,"le");l(Iu,"gt");l(zd,"ge");l(qn,"where");l(Vd,"clamp");l(na,"pad");l(Gd,"one_hot");l(li,"index_select");l(jd,"gather");l(ra,"scatter_add");l(qd,"scatter");l(ht,"sum");l(lr,"mean");l(Qs,"max");l(SN,"min");l(Ud,"argmax");l(kN,"argmin");l(vN,"prod");l(lt,"matmul");l(NN,"dot");l(sa,"cat");l(Wr,"stack");l(AN,"clone");l(F2,"fill");l(Cn,"reshape");l(gt,"transpose");l(Eu,"permute");l(ON,"broadcast_in_dim");l(Lu,"expand");l(Kd,"slice");l(Sr,"unsqueeze");l(Cu,"squeeze");l(Un,"narrow");l(Ut,"select");l(ci,"contiguous");l(Du,"repeat");l(Wd,"tile");l(Bo,"split");l(BN,"chunk");l(RN,"roll");l(MN,"flip");l(IN,"cumsum");l(EN,"sort");l(LN,"argsort");l(CN,"topk");l(Ro,"softmax");l(Hd,"log_softmax");l(pw,"layer_norm");l(fw,"batch_norm");l(hw,"conv2d");l(Yd,"pool2d");l(z2,"embedding")});var $N={};nt($N,{create:()=>DN,default:()=>V2,globals:()=>PN});var DN,PN,V2,FN=M(()=>{DN=l(()=>{throw new Error('mlfw: the "webgpu" npm package is not available in the browser; use navigator.gpu instead')},"create"),PN=void 0,V2={create:DN,globals:PN}});var XN={};nt(XN,{ensureWebGPUEager:()=>H2,flushWebGPUEager:()=>ZN,instantiateWebGPU:()=>q2,prewarmPipelines:()=>YN,recordWebGPUEager:()=>Gc,resetDevice:()=>G2,runWebGPUKernel:()=>U2,runWebGPUPlan:()=>K2,webgpuEagerOp:()=>Z2,webgpuEagerReady:()=>Y2,webgpuRNN:()=>tC});function $u(r){switch(Mn(r)){case"i32":return Int32Array;case"u32":return Uint32Array;case"f16":return Uint16Array;default:return Float32Array}}function Qd(r,e,t,n){if(t==="bf16")for(let s=0;s<e.length;s++)r[n+s]=fd(e[s]);else if(t==="i64")for(let s=0;s<e.length;s++)r[n+s]=Number(BigInt.asIntN(32,e[s]));else r.set(e,n)}function bw(r,e,t,n,s){if(t==="bf16")for(let o=0;o<s;o++)r[o]=hd(e[n+o]);else if(t==="i64")for(let o=0;o<s;o++)r[o]=BigInt(e[n+o]);else r.set(e.subarray(n,n+s))}function eo(r){return Math.ceil(r/4)*4}async function ww(){return kt||Fc||(Fc=(async()=>{let r=typeof navigator<"u"&&navigator.gpu;if(!r)try{let i=await Promise.resolve().then(()=>(FN(),$N));qN=i,r=i.create([]),i.globals&&(zc=i.globals.GPUBufferUsage,Jd=i.globals.GPUMapMode,Zd=i.globals.GPUShaderStage)}catch{throw new Error('WebGPU not available: install the "webgpu" npm package or run in a browser with WebGPU support')}!zc&&typeof GPUBufferUsage<"u"&&(zc=GPUBufferUsage,Jd=GPUMapMode,Zd=GPUShaderStage);let e=await r.requestAdapter();if(!e)throw new Error("WebGPU: no adapter found");let t=e.limits||{},n={},s=["maxStorageBuffersPerShaderStage","maxStorageBufferBindingSize","maxBufferSize","maxBindingsPerBindGroup"];for(let i of s)t[i]!==void 0&&(n[i]=t[i]);let o=[];return e.features&&e.features.has("shader-f16")&&o.push("shader-f16"),kt=await e.requestDevice({requiredLimits:n,requiredFeatures:o}),!zN&&typeof process<"u"&&process.on&&(zN=!0,process.on("exit",()=>{kt&&(kt.destroy(),kt=null)})),kt})(),Fc)}function Fu(){return zc}function Tw(){return Jd.READ}function G2(){kt&&(kt.destroy(),kt=null),Fc=null,zc=null,Jd=null,Zd=null,qN=null}function j2(r){return r.name==="_shapes"?"uniform":r.mode==="read_write"?"storage":"read-only-storage"}function UN(r,e){let t=r.createShaderModule({code:e.source}),n=[];for(let i of e.metadata.bindings)n.push({binding:i.index,visibility:Zd.COMPUTE,buffer:{type:j2(i)}});let s=r.createBindGroupLayout({entries:n}),o=r.createPipelineLayout({bindGroupLayouts:[s]});return{shaderModule:t,bindGroupLayout:s,pipelineLayout:o}}function KN(r,e){let{shaderModule:t,bindGroupLayout:n,pipelineLayout:s}=UN(r,e);return{pipeline:r.createComputePipeline({layout:s,compute:{module:t,entryPoint:e.name}}),bindGroupLayout:n}}function WN(r,e){let t=Xd.get(e);return t||(t=KN(r,e),Xd.set(e,t)),t}async function q2(r){let e=await ww(),{pipeline:t,bindGroupLayout:n}=KN(e,r);return{device:e,pipeline:t,bindGroupLayout:n,kernel:r,workgroupSize:r.metadata.workgroupSize,dispatchSize:r.metadata.dispatchSize,bindings:r.metadata.bindings}}function HN(r){let e=new Map,t=0;for(let n of r)if(n.name!=="_shapes")if(n.packed)for(let s of n.packed)e.set(s.name,s.argIndex!==void 0?s.argIndex:t++);else e.set(n.name,n.argIndex!==void 0?n.argIndex:t++);return e}async function U2(r,e,t){let{device:n,pipeline:s,bindGroupLayout:o,bindings:i,dispatchSize:a}=r,u=Fu(),p=[],c=[],h=HN(i);for(let y=0;y<i.length;y++){let _=i[y];if(_.name==="_shapes"){let v=new Uint32Array(t||[]),O=Math.max(Math.ceil(v.byteLength/16)*16,16),R=n.createBuffer({size:O,usage:u.UNIFORM|u.COPY_DST});n.queue.writeBuffer(R,0,v),p.push(R),c.push({binding:_.index,resource:{buffer:R}});continue}let b=bs(_.dtype),w=$u(_.dtype);if(_.packed){let v=eo(_.packedSize*b),R=_.mode==="read_write"?u.STORAGE|u.COPY_SRC|u.COPY_DST:u.STORAGE|u.COPY_DST,B=n.createBuffer({size:Math.max(v,4),usage:R,mappedAtCreation:!0}),E=new w(B.getMappedRange());for(let z of _.packed){let D=h.get(z.name),J=e[D];J&&Qd(E,J,z.dtype,z.offset)}B.unmap(),p.push(B),c.push({binding:_.index,resource:{buffer:B}});continue}let T=h.get(_.name),S=e[T],x=eo(S.length*b),A=_.mode==="read_write"?u.STORAGE|u.COPY_SRC|u.COPY_DST:u.STORAGE|u.COPY_DST,N=n.createBuffer({size:Math.max(x,4),usage:A,mappedAtCreation:!0});Qd(new w(N.getMappedRange()),S,_.dtype,0),N.unmap(),p.push(N),c.push({binding:_.index,resource:{buffer:N}})}let f=n.createBindGroup({layout:o,entries:c}),d=n.createCommandEncoder(),m=d.beginComputePass();m.setPipeline(s),m.setBindGroup(0,f),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=bs(_.dtype),T=$u(_.dtype);if(_.packed)for(let S of _.packed){let x=S.offset*w,k=eo(S.size*w),A=n.createBuffer({size:k,usage:u.MAP_READ|u.COPY_DST});d.copyBufferToBuffer(b,x,A,0,k),g.push({readBuf:A,tensorIdx:h.get(S.name),size:S.size,dtype:S.dtype,ViewCtor:T})}else{let S=b.size,x=n.createBuffer({size:S,usage:u.MAP_READ|u.COPY_DST});d.copyBufferToBuffer(b,0,x,0,S),g.push({readBuf:x,tensorIdx:h.get(_.name),size:e[h.get(_.name)].length,dtype:_.dtype,ViewCtor:T})}}n.queue.submit([d.finish()]);for(let y of g){await y.readBuf.mapAsync(Tw());let _=new y.ViewCtor(y.readBuf.getMappedRange());bw(e[y.tensorIdx],_,y.dtype,0,y.size),y.readBuf.unmap(),y.readBuf.destroy()}for(let y of p)y.destroy()}async function YN(r,e){if(typeof r.createComputePipelineAsync!="function")return;let t=[],n=new Set;for(let s of e){if(!s||n.has(s)||Xd.has(s))continue;n.add(s);let{shaderModule:o,bindGroupLayout:i,pipelineLayout:a}=UN(r,s);t.push(r.createComputePipelineAsync({layout:a,compute:{module:o,entryPoint:s.name}}).then(u=>Xd.set(s,{pipeline:u,bindGroupLayout:i}),()=>{}))}t.length>0&&await Promise.all(t)}async function K2(r,e,t){let n=await ww(),s=Fu();await YN(n,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(r.numSlots).fill(null),a=new Array(r.numSlots).fill("f32");for(let b=0;b<r.numSlots;b++){let w=e[b];if(!w)continue;let T=w.dtype||"f32";a[b]=T;let S=Math.max(eo(w.data.length*bs(T)),4),x=!o.has(b),k=n.createBuffer({size:S,usage:s.STORAGE|s.COPY_DST|s.COPY_SRC,mappedAtCreation:x});x&&(Qd(new($u(T))(k.getMappedRange()),w.data,T,0),k.unmap()),i[b]=k}let u=32,p=[],c=[],h={encoder:n.createCommandEncoder(),pending:0},f=l(()=>{++h.pending>=u&&(n.queue.submit([h.encoder.finish()]),h.encoder=n.createCommandEncoder(),h.pending=0)},"maybeFlush"),d=l(b=>{let w=h.encoder,{pipeline:T,bindGroupLayout:S}=WN(n,b.kernel),x=b.inputSlots.concat(b.outputSlots),k=[],A=[],N=0;for(let B of b.kernel.metadata.bindings)if(B.name==="_shapes"){let E=new Uint32Array(b.shapeValues||[]),z=Math.max(Math.ceil(E.byteLength/16)*16,16),D=n.createBuffer({size:z,usage:s.UNIFORM|s.COPY_DST});n.queue.writeBuffer(D,0,E),p.push(D),k.push({binding:B.index,resource:{buffer:D}})}else if(B.packed){let E=bs(B.dtype),z=n.createBuffer({size:Math.max(eo(B.packedSize*E),4),usage:s.STORAGE|s.COPY_DST|s.COPY_SRC});c.push(z);let D=B.mode==="read_write";for(let J of B.packed){let j=x[J.argIndex],U=eo(J.size*E);D?A.push({slot:j,src:z,srcOff:J.offset*E,bytes:U}):w.copyBufferToBuffer(i[j],0,z,J.offset*E,U)}k.push({binding:B.index,resource:{buffer:z}})}else k.push({binding:B.index,resource:{buffer:i[x[N++]]}});let v=n.createBindGroup({layout:S,entries:k}),O=w.beginComputePass();O.setPipeline(T),O.setBindGroup(0,v);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);f()},"encodeStep"),m=r.scanLoops?[...r.scanLoops].sort((b,w)=>b.loopStart-w.loopStart):r.scanLoop?[r.scanLoop]:null;if(m&&m.length){let b=0;for(let w of m){for(;b<w.loopStart;b++)d(t[b]);for(let T of w.carry)h.encoder.copyBufferToBuffer(i[T.initSlot],0,i[T.a],0,T.bytes);for(let T=0;T<w.T;T++){for(let S of w.xs)h.encoder.copyBufferToBuffer(i[S.xsSlot],T*S.stepBytes,i[S.xtSlot],0,S.stepBytes);for(let S=w.loopStart;S<w.loopEnd;S++)d(t[S]);for(let S of w.ys)h.encoder.copyBufferToBuffer(i[S.ytSlot],0,i[S.ysSlot],T*S.stepBytes,S.stepBytes);for(let S of w.carry){let x=i[S.a];i[S.a]=i[S.b],i[S.b]=x}}for(let T of w.carry)h.encoder.copyBufferToBuffer(i[T.a],0,i[T.finalSlot],0,T.bytes);b=w.loopEnd}for(;b<t.length;b++)d(t[b])}else for(let b of t)d(b);let g=h.encoder,y=new Set(r.argSlots),_=[];for(let b=0;b<r.numSlots;b++){if(!i[b]||!o.has(b)||!y.has(b))continue;let w=e[b],T=Math.max(eo(w.data.length*bs(a[b])),4),S=n.createBuffer({size:T,usage:s.MAP_READ|s.COPY_DST});g.copyBufferToBuffer(i[b],0,S,0,T),_.push({rb:S,dtype:a[b],size:w.data.length,dst:w.data})}n.queue.submit([g.finish()]);for(let b of _)await b.rb.mapAsync(Tw()),bw(b.dst,new($u(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 JN(r){let e=_w.get(r);if(e&&e.length)return e.pop();let t=Fu();return kt.createBuffer({size:r,usage:t.STORAGE|t.COPY_SRC|t.COPY_DST})}function VN(r,e){let t=_w.get(e);t||(t=[],_w.set(e,t)),t.push(r)}async function H2(){return await ww(),As||(As=kt.createCommandEncoder()),kt}function Y2(){return!!kt&&!!As}function GN(r,e,t){let n=Vc.get(r);if(n)return n.buf;let s=Fu(),o=Math.max(eo(r.length*bs(e)),4);if(t){let a=kt.createBuffer({size:o,usage:s.STORAGE|s.COPY_SRC|s.COPY_DST,mappedAtCreation:!0});return Qd(new($u(e))(a.getMappedRange()),r,e,0),a.unmap(),Vc.set(r,{buf:a,bytes:o,dtype:e,persistent:t}),a}let i=JN(o);return Vc.set(r,{buf:i,bytes:o,dtype:e,persistent:!1}),i}function Gc(r,e,t){let{pipeline:n,bindGroupLayout:s}=WN(kt,r),{bindings:o,dispatchSize:i}=r.metadata,a=HN(o),u=Fu(),p=[],c=[];for(let d of o){if(d.name==="_shapes"){let y=new Uint32Array(t||[]),_=Math.max(Math.ceil(y.byteLength/16)*16,16),b=kt.createBuffer({size:_,usage:u.UNIFORM|u.COPY_DST});kt.queue.writeBuffer(b,0,y),gw.push(b),p.push({binding:d.index,resource:{buffer:b}});continue}if(d.packed){let y=bs(d.dtype),_=Math.max(eo(d.packedSize*y),4),b=JN(_);yw.push({buf:b,bytes:_});let w=d.mode==="read_write";for(let T of d.packed){let S=GN(e[a.get(T.name)],T.dtype,!w),x=eo(T.size*y);w?c.push({pbuf:b,off:T.offset*y,sub:S,bytes:x}):As.copyBufferToBuffer(S,0,b,T.offset*y,x)}p.push({binding:d.index,resource:{buffer:b}});continue}let m=e[a.get(d.name)],g=GN(m,d.dtype,d.mode!=="read_write");p.push({binding:d.index,resource:{buffer:g}})}let h=kt.createBindGroup({layout:s,entries:p}),f=As.beginComputePass();f.setPipeline(n),f.setBindGroup(0,h),f.dispatchWorkgroups(i[0],i[1],i[2]),f.end();for(let d of c)As.copyBufferToBuffer(d.pbuf,d.off,d.sub,0,d.bytes);++mw>=W2&&(kt.queue.submit([As.finish()]),As=kt.createCommandEncoder(),mw=0)}async function ZN(){if(!kt)return;let r=Fu(),e=[];for(let[t,n]of Vc){if(n.persistent)continue;let s=kt.createBuffer({size:n.bytes,usage:r.MAP_READ|r.COPY_DST});As.copyBufferToBuffer(n.buf,0,s,0,n.bytes),e.push({rb:s,ha:t,buf:n.buf,bytes:n.bytes,dtype:n.dtype})}kt.queue.submit([As.finish()]),As=kt.createCommandEncoder(),mw=0;for(let t of gw)t.destroy();gw.length=0,await Promise.all(e.map(t=>t.rb.mapAsync(Tw())));for(let t of e)bw(t.ha,new($u(t.dtype))(t.rb.getMappedRange()),t.dtype,0,t.ha.length),t.rb.unmap(),t.rb.destroy();for(let t of yw)VN(t.buf,t.bytes);yw.length=0;for(let t of e)VN(t.buf,t.bytes),Vc.delete(t.ha)}function J2(r,e,t,n,s){let o=`${r.join(",")}|${e.join(",")}|${t}|${n}`,i=jN.get(o);if(i)return i;let a=Mn(n),u="";for(let c=r.length-1;c>=0;c--)u+=`  let idx${c} = rem % ${r[c]}u; rem = rem / ${r[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:n,mode:"read"},{index:1,name:"outp",dtype:n,mode:"read_write"}],dispatchSize:[Math.ceil(s/64),1,1]}},jN.set(o,i),i}function Pu(r){let e=r._impl.storage.rawData;if(r.isContiguous&&r._impl.storageOffset===0&&e.length===r.numel)return e;let t=new e.constructor(r.numel);return Gc(J2(r.shape,r.strides,r._impl.storageOffset,r.dtype,r.numel),[e,t],void 0),t}function Z2(r,e,t){let n=e.map(Pu);n.push(t),Gc(r,n,void 0)}function Q2(r,e,t){let n=`lstm|${r}|${e}|${t}`,s=em.get(n);if(s)return s;let o=4*e,i=o*r,a=o*r+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(${Xs})
+fn rnn(@builtin(local_invocation_id) lid : vec3<u32>) {
+  let tid = lid.x;
+  for (var p = tid; p < ${e}u; p = p + ${Xs}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 + ${Xs}u) {
+      var acc = w[${a}u + j] + w[${u}u + j];
+      let xb = t * ${r}u;
+      let wxb = j * ${r}u;
+      for (var e = 0u; e < ${r}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 + ${Xs}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 + ${Xs}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]}},em.set(n,s),s}function QN(r,e,t,n){let s=`${e}|${t}|${n}`;if(r._webgpuPacked&&r._webgpuPackedTag===s)return r._webgpuPacked;let o=l(h=>h.contiguous()._impl.storage.rawData,"raw"),i=o(r.x2h.weight),a=o(r.h2h.weight),u=r.x2h.bias?o(r.x2h.bias):new Float32Array(n),p=r.h2h.bias?o(r.h2h.bias):new Float32Array(n),c=new Float32Array(n*e+n*t+n+n);return c.set(i,0),c.set(a,n*e),c.set(u,n*e+n*t),c.set(p,n*e+n*t+n),r._webgpuPacked=c,r._webgpuPackedTag=s,c}function X2(r,e,t){return QN(r,e,t,4*t)}function eC(r,e,t){let n=`gru|${r}|${e}|${t}`,s=em.get(n);if(s)return s;let o=3*e,i=o*r,a=o*r+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(${Xs})
+fn rnn(@builtin(local_invocation_id) lid : vec3<u32>) {
+  let tid = lid.x;
+  for (var p = tid; p < ${e}u; p = p + ${Xs}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 + ${Xs}u) {
+      var ax = w[${a}u + j];
+      var ah = w[${u}u + j];
+      let xb = t * ${r}u;
+      let wxb = j * ${r}u;
+      for (var e = 0u; e < ${r}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 + ${Xs}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 + ${Xs}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]}},em.set(n,s),s}function tC(r,e,t,n,s){if(t.kind==="gru")return nC(r,e,t,n);if(t.batch!==1)return null;let o=t.hiddenSize,i=t.seqLen,a=e.length,u=Pu(r),p=[],c=[];for(let d=0;d<a;d++){let m=d===0?t.inputSize:o,g=X2(e[d],m,o),y,_;n!=null&&s!=null?(y=Pu(Ut(n,0,d)),_=Pu(Ut(s,0,d))):(y=new Float32Array(o),_=new Float32Array(o));let b=new Float32Array(i*o),w=new Float32Array(o),T=new Float32Array(o);Gc(Q2(m,o,i),[u,g,y,_,b,w,T],void 0),u=b,p.push(w),c.push(T)}let h=tn(u,[i,1,o],"f32",Ys),f=l(d=>a===1?tn(d[0],[1,1,o],"f32",Ys):Wr(d.map(m=>tn(m,[1,o],"f32",Ys)),0),"wrapState");return[h,f(p),f(c)]}function nC(r,e,t,n){if(t.batch!==1)return null;let s=t.hiddenSize,o=t.seqLen,i=e.length,a=Pu(r),u=[];for(let h=0;h<i;h++){let f=h===0?t.inputSize:s,d=QN(e[h],f,s,3*s),m=n!=null?Pu(Ut(n,0,h)):new Float32Array(s),g=new Float32Array(o*s),y=new Float32Array(s);Gc(eC(f,s,o),[a,d,m,g,y],void 0),a=g,u.push(y)}let p=tn(a,[o,1,s],"f32",Ys),c=i===1?tn(u[0],[1,1,s],"f32",Ys):Wr(u.map(h=>tn(h,[1,s],"f32",Ys)),0);return[p,c]}var kt,Fc,zc,Jd,Zd,qN,zN,Xd,As,mw,W2,Vc,gw,yw,_w,jN,Xs,em,xw=M(()=>{"use strict";Vn();xu();Kn();ar();be();be();kt=null,Fc=null,zc=null,Jd=null,Zd=null,qN=null,zN=!1;l($u,"wgslViewCtor");l(Qd,"packTensorInto");l(bw,"unpackTensorFrom");l(eo,"align4");l(ww,"ensureDevice");l(Fu,"bufUsage");l(Tw,"mapModeRead");l(G2,"resetDevice");l(j2,"bindingBufferType");l(UN,"pipelineParts");l(KN,"createPipeline");Xd=new WeakMap;l(WN,"pipelineFor");l(q2,"instantiateWebGPU");l(HN,"buildParamIndex");l(U2,"runWebGPUKernel");l(YN,"prewarmPipelines");l(K2,"runWebGPUPlan");As=null,mw=0,W2=64,Vc=new Map,gw=[],yw=[],_w=new Map;l(JN,"acquireStorage");l(VN,"releaseStorage");l(H2,"ensureWebGPUEager");l(Y2,"webgpuEagerReady");l(GN,"eagerBufferFor");l(Gc,"recordWebGPUEager");l(ZN,"flushWebGPUEager");jN=new Map;l(J2,"contigKernel");l(Pu,"webgpuEagerInput");l(Z2,"webgpuEagerOp");Xs=256,em=new Map;l(Q2,"lstmKernelWGSL");l(QN,"rnnPackedWeights");l(X2,"lstmPackedWeights");l(eC,"gruKernelWGSL");l(tC,"webgpuRNN");l(nC,"webgpuGRU")});var Bw={};nt(Bw,{getBackend:()=>Ow,hasBackend:()=>rC,preloadCudaRuntime:()=>eA,preloadWebGPU:()=>qc,registerBackend:()=>Uc});function Uc(r,e){Aw.set(r,e)}function Ow(r){return Aw.get(r)||null}function rC(r){return Aw.has(r)}function sC(){return Sw||(Sw=Promise.resolve().then(()=>(Jv(),Yv))),Sw}function Nw(){return kw||(kw=Promise.resolve().then(()=>(Ad(),Nd))),kw}async function eA(){if(jc)return jc;let r=await Nw();return r.preloadCublas&&await r.preloadCublas(),jc=r,r}async function qc(){if(vw)return vw;let r=await Promise.resolve().then(()=>(xw(),XN)),{setWebGPUEagerFn:e,setWebgpuRNN:t}=await Promise.resolve().then(()=>(Kn(),nA));return await r.ensureWebGPUEager(),e(r.webgpuEagerOp),t(r.webgpuRNN),vw=r,r}function tA(r){let e;try{e=Nu(r.source)}catch(o){throw new Error("encodeWat: "+o.message+`
+`+r.source)}let t;try{t=new WebAssembly.Module(e)}catch(o){throw new Error("WASM: "+o.message+`
+`+r.source)}let n={};if(r.metadata.imports)for(let[o]of r.metadata.imports)n[o]=oC[o]||Math[o]||(i=>i);let s=new WebAssembly.Instance(t,{math:n});return{exports:s.exports,memory:s.exports.memory,bufferOffsets:r.metadata.bufferOffsets,funcName:r.name,binary:e,parallel:r.metadata.parallel||null,mathNames:r.metadata.imports?[...r.metadata.imports.keys()]:[]}}function tm(r,e,t,n,s){let{exports:o,memory:i,bufferOffsets:a,funcName:u}=r,p=o[u],c=[...a.values()],h=Math.min(c.length,e.length);for(let d=0;d<h;d++){let m=e[d];ArrayBuffer.isView(m)&&new m.constructor(i.buffer,c[d],m.length).set(m)}let f=c.slice(0,h);if(t)for(let d of t)f.push(d);n!==void 0&&s!==void 0&&f.push(n,s),p(...f);for(let d=0;d<h;d++){let m=e[d];ArrayBuffer.isView(m)&&m.set(new m.constructor(i.buffer,c[d],m.length))}}function iC(r,e,t=[],n={}){let s=n.warmup??5,o=n.repeat??30,i=tA(r),a=e.map(h=>new Float32Array(Math.max(1,Math.ceil(h/4)))),u=t||[],p=l(()=>{i.parallel?tm(i,a,u,0,i.parallel.extent):tm(i,a,u)},"once");for(let h=0;h<s;h++)p();let c=[];for(let h=0;h<o;h++){let f=performance.now();p(),c.push(performance.now()-f)}return c}var Aw,Sw,kw,jc,vw,oC,Kc=M(()=>{"use strict";kd();Ji();rw();Aw=new Map;l(Uc,"registerBackend");l(Ow,"getBackend");l(rC,"hasBackend");Sw=null;l(sC,"getWasmPool");kw=null;l(Nw,"getCudaRuntime");jc=null;l(eA,"preloadCudaRuntime");vw=null;l(qc,"preloadWebGPU");oC={exp:Math.exp,log:Math.log,sin:Math.sin,cos:Math.cos,tan:Math.tan,tanh:Math.tanh,pow:Math.pow,fmod:l((r,e)=>r%e,"fmod"),rsqrt:l(r=>1/Math.sqrt(r),"rsqrt"),sign:Math.sign,round:Math.round,erf:bu,erfc:wu,lgamma:Tn,gamma:Tu,digamma:Ac};l(tA,"instantiateWasm");l(tm,"runWasmKernel");Uc("js",{instantiate(r){return new Function("return "+r.source)()},runSync(r,e,t){let n=t?[...e,...t]:e;return r(...n)},runAsync(r,e,t){let n=t?[...e,...t]:e;return r(...n)},isAsync(){return!1}});Uc("wasm",{instantiate(r){return tA(r)},runSync(r,e,t){r.parallel?tm(r,e,t,0,r.parallel.extent):tm(r,e,t)},async runAsync(r,e,t){if(r.parallel&&r.parallel.poolSafe){let{runWasmParallel:n}=await sC();await n(r,r.funcName,e,t,r.parallel,r.mathNames);return}this.runSync(r,e,t)},isAsync(r){return!!(r&&r.parallel&&r.parallel.poolSafe)}});l(iC,"measureWasm");Wv("wasm",iC);Uc("webgpu",{instantiate(r){return qc().then(e=>e.instantiateWebGPU(r))},runSync(){throw new Error("WebGPU kernel requires async execution \u2014 use runAsync()")},async runAsync(r,e,t){let{runWebGPUKernel:n}=await qc();await n(r,e,t)},async runPlan(r,e,t){let{runWebGPUPlan:n}=await qc();await n(r,e,t)},isAsync(){return!0}});Uc("cuda",{instantiate(r){return{kernel:r}},runSync(r,e,t){if(!jc)throw new Error("CUDA sync runtime not preloaded \u2014 call preloadCudaRuntime() before synchronous execution");jc.runCudaKernelResident(r.kernel,e,t)},async runAsync(r,e,t){let{runCudaKernel:n}=await Nw();await n(r.kernel,e,t)},async runPlan(r,e,t,n){let{runCudaPlan:s}=await Nw();await s(r,e,t,n)},isAsync(){return!0}})});function Rw(r){return aC[r]||Float32Array}function uC(r){return r instanceof Float64Array?"f64":r instanceof Int32Array?"i32":r instanceof Int16Array?"i16":r instanceof Int8Array?"i8":r instanceof Uint16Array?"ui16":r instanceof Uint8Array?"ui8":r instanceof BigInt64Array?"i64":"f32"}function rA(r,e){if(!(!r.returnFixups||r.returnFixups.length===0))for(let t of r.returnFixups){let n=e[r.argSlots[t.pos]];if(!(!n||!n.data))if(t.kind==="copy"){let s=e[t.srcSlot];s&&s.data&&n.data.set(s.data.subarray(0,n.data.length))}else t.kind==="const"&&n.data.fill(t.value)}}var aC,to,Mw,zu,nm=M(()=>{"use strict";Kc();Ec();aC={f16:Uint16Array,bf16:Uint16Array,f32:Float32Array,f64:Float64Array,i8:Int8Array,i16:Int16Array,i32:Int32Array,i64:BigInt64Array,ui8:Uint8Array,bool:Uint8Array,index:Int32Array};l(Rw,"typedArrayCtor");l(uC,"dtypeOfTypedArray");l(rA,"_applyReturnFixups");to=class r{static{l(this,"RuntimeTensor")}data;shape;dtype;strides;resident;constructor(e,t,n,s=null){this.data=e,this.shape=[...t],this.dtype=n,this.strides=s?[...s]:r.defaultStrides(t)}static defaultStrides(e){let t=new Array(e.length),n=1;for(let s=e.length-1;s>=0;s--)t[s]=n,n*=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 n=1;for(let s=0;s<e.length;s++)n*=e[s];return n=Math.max(n,1),new r(new(Rw(t))(n),e,t)}static fromArray(e,t,n="f32"){let s=Rw(n);return new r(new s(e),t,n)}get(e){let t=0;for(let n=0;n<e.length;n++)t+=e[n]*this.strides[n];return this.data[t]}set(e,t){let n=0;for(let s=0;s<e.length;s++)n+=e[s]*this.strides[s];this.data[n]=t}},Mw=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()]}},zu=class r{static{l(this,"RuntimeModule")}name;kernels;_instances;_shapeParamMaps;_bufferMaps;constructor(e){this.name=e,this.kernels=new Mw,this._instances=new Map}addCompiledKernel(e){this.kernels.register(e.name,e);let t=Ow(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,n){this._shapeParamMaps||(this._shapeParamMaps=new Map),this._shapeParamMaps.set(e,t),n&&(this._bufferMaps||(this._bufferMaps=new Map),this._bufferMaps.set(e,n))}_prepareArgs(e,t){let n=[],s=new Map;for(let a=0;a<t.length;a++){let u=t[a];u instanceof to?(n.push(u.data),s.set(a,u.shape)):n.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=r._extractShapeParams(o,s,t,a)}return{tensorArgs:n,shapeValues:i}}run(e,...t){let n=this._instances.get(e);if(!n)throw new Error("Kernel '"+e+"' not found or not executable");if(n.instance instanceof Promise)throw new Error("Kernel '"+e+"' requires async execution \u2014 use runAsync()");let{tensorArgs:s,shapeValues:o}=this._prepareArgs(e,t);return n.backend.runSync(n.instance,s,o)}async runAsync(e,...t){let n=this._instances.get(e);if(!n)throw new Error("Kernel '"+e+"' not found or not executable");let{tensorArgs:s,shapeValues:o}=this._prepareArgs(e,t),i=await n.instance;return n.backend.runAsync(i,s,o)}isAsync(e){let t=this._instances.get(e);if(!t)return!1;let n=t.instance instanceof Promise?null:t.instance;return t.backend.isAsync(n)}async runPlanAsync(e,t,n){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 to?u:new to(u,[u.length],uC(u))}for(let a of e.intermediates){let u=1;for(let p of a.shape)u*=p;s[a.slot]=new to(new(Rw(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 h of u.inputSlots)p.push(s[h]);for(let h of u.outputSlots)p.push(s[h]);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,n),rA(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)}rA(e,s)}_uniformPlanBackend(e){let t=null;for(let n of e.steps){let s=this._instances.get(n.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,n,s){let o=new Map;if(s){let u=0;for(let[p,c]of s){let h=typeof p=="string"?p:c&&c.name;typeof h=="string"&&o.set(h,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(":"),h=u.substring(0,c),f=parseInt(u.substring(c+1),10),d=null;if(o.has(h)){let m=t.get(o.get(h));m&&f<m.length&&m[f]>0&&(d=m[f])}if(d===null){for(let[,m]of t)if(f<m.length&&m[f]>0){d=m[f];break}}a.push(d!==null?d: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 n=this.kernels.get(t);e.push({name:n.name,source:n.source,target:n.target.name,metadata:n.metadata})}return{name:this.name,kernels:e}}static deserialize(e){let t=new r(e.name);for(let n of e.kernels)t.addCompiledKernel(new Ic(n.name,n.source,{name:n.target},n.metadata));return t}}});var rm,sA=M(()=>{"use strict";rm=class{static{l(this,"AnalysisManager")}constructor(){this._cache=new WeakMap}getAnalysis(e,t){let n=this._cache.get(t);n||(n=new Map,this._cache.set(t,n));let s=n.get(e);if(!s||s.version!==t.version){let o=this._resolveDeps(e,t,n);s={data:e.compute(t,o),version:t.version},n.set(e,s)}return s.data}_resolveDeps(e,t,n){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 n=this._cache.get(e);if(!n)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,h=!1;if(c){for(let f of c)if(!s(f)||i(f)){h=!0;break}}return o.set(u,h),h},"isStale"),a=[];for(let u of n.keys())if(!s(u)||i(u))a.push(u);else{let p=n.get(u);p&&(p.version=e.version)}for(let u of a)n.delete(u)}invalidateFunctions(e,t=null){for(let n of e)this.invalidate(n,t)}invalidateAll(){this._cache=new WeakMap}}});var oe,sm,om,Ge,im,nn=M(()=>{"use strict";oe=Object.freeze({UNCHANGED:0,CHANGED:1,FAILED:2}),sm=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")}},om=class{static{l(this,"PassContext")}constructor({optLevel:e=1/0,disabledPasses:t=[],requiredPasses:n=[],config:s={}}={}){this.optLevel=e,this.disabledPasses=t instanceof Set?t:new Set(t),this.requiredPasses=n instanceof Set?n:new Set(n),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)}},Ge=class extends sm{static{l(this,"FunctionPass")}run(e,t){throw new Error("Not implemented")}},im=class extends sm{static{l(this,"ModulePass")}run(e,t){throw new Error("Not implemented")}}});var ue,Dn,lC,am,Kt=M(()=>{"use strict";nn();ue=Object.freeze({SILENT:0,INFO:1,VERBOSE:2,DEBUG:3}),Dn=class{static{l(this,"CompilationError")}constructor(e,t,n,s){this.phase=e,this.funcName=t,this.message=n,this.passName=s||null}toString(){let e="["+this.phase+"]";return this.funcName&&(e+=" "+this.funcName),this.passName&&(e+=" ("+this.passName+")"),e+=": "+this.message,e}},lC=l(()=>{},"NOOP"),am=class{static{l(this,"TraceLog")}constructor(e={}){this.level=e.level??ue.SILENT,this.sink=typeof e.sink=="function"?e.sink:lC,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:ue.INFO})}phaseEnd(e,t){this.emit({type:"phase",action:"end",phase:e,durationMs:t,level:ue.INFO})}passRun(e,t,n,s,o){this.emit({type:"pass",passName:e,changed:t===oe.CHANGED,durationMs:n,opCountBefore:s,opCountAfter:o,level:ue.VERBOSE})}functionEvent(e,t,n){this.emit({type:"function",phase:e,funcName:t,...n,level:ue.INFO})}irDump(e,t){this.emit({type:"ir_snapshot",label:e,text:t,level:ue.DEBUG})}memoryStats(e,t){this.emit({type:"memory",funcName:e,...t,level:ue.VERBOSE})}autotuneStats(e,t){this.emit({type:"autotune",funcName:e,...t,level:ue.VERBOSE})}codegenStats(e,t){this.emit({type:"codegen",funcName:e,...t,level:ue.VERBOSE})}errorEvent(e,t,n,s){this.emit({type:"error",phase:e,funcName:t,message:n,passName:s||null,level:ue.INFO})}warn(e,t,n,s){this.emit({type:"warning",phase:e,funcName:t,message:n,detail:s||null,level:ue.INFO})}explain(e,t,n,s,o){this.emit({type:"explain",category:e,subject:t,decision:n,reason:s,...o||{},level:ue.DEBUG})}get explainsEnabled(){return this.level>=ue.DEBUG}shouldSnapshot(e){return this.level>=ue.DEBUG&&!!this.irSnapshot[e]}}});function Wc(r){if(typeof r.numOps=="function")return r.numOps();if(typeof r[Symbol.iterator]=="function"){let e=0;for(let t of r)typeof t.numOps=="function"&&(e+=t.numOps());return e}return-1}var Vu,oa,um=M(()=>{"use strict";sA();nn();Kt();Kt();l(Wc,"countOps");Vu=class{static{l(this,"FixedPointGroup")}constructor(e,t,n=8){this.name=e,this.passes=t,this.maxIterations=n}},oa=class{static{l(this,"PassManager")}constructor(){this.passes=[],this.analysisManager=new rm,this.trace=null,this.verifyHook=null,this.instruments=[]}addPass(e){this.passes.push(e)}addInstrument(e){this.instruments.push(e)}_notify(e,t,n,s){for(let o of this.instruments)typeof o[e]=="function"&&o[e](t,n,s)}setTrace(e){this.trace=e}setVerifyHook(e){this.verifyHook=e}_verifyAfter(e,t,n){if(!this.verifyHook)return null;let s=this.verifyHook(t,n);if(!s||s.length===0)return null;let o=n?t.name||"<module>":t.name;return new Dn("verification",o,`pass '${e.name}' produced invalid IR: ${s.join("; ")}`,e.name)}_applyPass(e,t,n,s){this.trace&&(e.trace=this.trace),this.instruments.length&&this._notify("runBeforePass",e,t,null);let o=n.verbose,i=n.resilient,a=!1,u=!1;if(e instanceof im){let p=o?Wc(t):-1,c=o?performance.now():0,h;try{h=e.run(t,this.analysisManager)}catch(f){if(!i)throw f;return this.analysisManager.invalidateAll(),s.push(oe.FAILED),n.errors.push(new Dn("graphPasses",t.name||"<module>",f.message,e.name)),{changed:a,fatal:!1}}if(s.push(h),o&&this.trace.passRun(e.name,h,performance.now()-c,p,Wc(t)),h===oe.CHANGED){a=!0,n.anyChanged=!0,this.analysisManager.invalidateFunctions(t,e.preservedAnalyses);let f=this._verifyAfter(e,t,!0);f&&(n.errors.push(f),i||(u=!0))}else h===oe.FAILED&&(this.analysisManager.invalidateAll(),n.errors.push(new Dn("graphPasses",t.name||"<module>",`pass '${e.name}' failed`,e.name)),i||(u=!0))}else if(e instanceof Ge){let p=!1;for(let c of t){if(n.failedFunctions.has(c.name))continue;let h=o?Wc(c):-1,f=o?performance.now():0;if(i)try{for(let m of e.requiredAnalyses)this.analysisManager.getAnalysis(m,c);let d=e.run(c,this.analysisManager);if(o&&this.trace.passRun(e.name,d,performance.now()-f,h,Wc(c)),d===oe.CHANGED){p=!0,n.anyChanged=!0,c.bumpVersion(),this.analysisManager.invalidate(c,e.preservedAnalyses);let m=this._verifyAfter(e,c,!1);m&&(n.errors.push(m),n.failedFunctions.add(c.name))}else d===oe.FAILED&&(this.analysisManager.invalidate(c),n.errors.push(new Dn("graphPasses",c.name,`pass '${e.name}' failed`,e.name)),n.failedFunctions.add(c.name))}catch(d){n.errors.push(new Dn("graphPasses",c.name,d.message,e.name)),n.failedFunctions.add(c.name)}else{for(let m of e.requiredAnalyses)this.analysisManager.getAnalysis(m,c);let d=e.run(c,this.analysisManager);if(o&&this.trace.passRun(e.name,d,performance.now()-f,h,Wc(c)),d===oe.CHANGED){p=!0,n.anyChanged=!0,c.bumpVersion(),this.analysisManager.invalidate(c,e.preservedAnalyses);let m=this._verifyAfter(e,c,!1);if(m){n.errors.push(m),n.failedFunctions.add(c.name),u=!0;break}}else if(d===oe.FAILED){this.analysisManager.invalidate(c),n.errors.push(new Dn("graphPasses",c.name,`pass '${e.name}' failed`,e.name)),n.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,n,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(!n.passContext.shouldRun(u))continue;let{changed:p,fatal:c}=this._applyPass(u,t,n,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 n={verbose:this.trace&&this.trace.level>=ue.VERBOSE,resilient:t.errorMode==="resilient",errors:[],failedFunctions:new Set,anyChanged:!1,passContext:t.passContext||new om},s=[];for(let o of this.passes){if(!(o instanceof Vu)&&!n.passContext.shouldRun(o))continue;if(o instanceof Vu?this._runGroup(o,e,n,s):this._applyPass(o,e,n,s).fatal)return{changed:n.anyChanged,results:s,errors:n.errors,failedFunctions:n.failedFunctions.size>0?n.failedFunctions:null}}return{changed:n.anyChanged,results:s,errors:n.errors.length>0?n.errors:null,failedFunctions:n.failedFunctions.size>0?n.failedFunctions:null}}}});function rn(r,e){Iw.set(r,e)}function Ew(r,e,t,n){let s=r.neg(e).getResult(0),o=r.exp(s).getResult(0),i=r.broadcast(r.scalarConstant(1,t).getResult(0),n,[]).getResult(0),a=r.add(i,o).getResult(0);return r.div(i,a).getResult(0)}function pn(r,e,t,n){return r.broadcast(r.scalarConstant(e,t).getResult(0),n,[]).getResult(0)}var Iw,Gu,Lw=M(()=>{"use strict";nn();es();Oe();Kt();Iw=new Map;l(rn,"registerDecomposition");Gu=class extends Ge{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())Iw.has(o.opName)&&this._shouldDecompose(o)&&t.push(o);if(t.length===0)return oe.UNCHANGED;let n=new An(e),s=[];for(let o of t){if(!o.parentBlock)continue;let i=Iw.get(o.opName);n.block=o.parentBlock,n.setInsertionPoint(o),s.push(o.opName),i(o,n)}if(this.trace&&this.trace.level>=ue.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:ue.DEBUG})}return oe.CHANGED}};rn("stop_gradient",r=>{r.replaceAllResultsWith([r.getOperand(0)]),r.erase()});rn("all_reduce",(r,e)=>{let t=r.getOperand(0),n=r.getAttr("mesh_axis")??0,s=r.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),[n],s).getResult(0),p=[];for(let h=0;h<o.length;h++)h!==n&&p.push(h);let c=e.broadcast(u,o,p).getResult(0);r.replaceAllResultsWith([c]),r.erase()});rn("all_gather",(r,e)=>{let t=r.getOperand(0),n=t.type.shape,s=r.getAttr("mesh_axis")??0,o=r.getAttr("gather_dim")??1,i=n[s],a=n.filter((m,g)=>g!==s),u=o<s?o:o-1,p=[];for(let m=0;m<i;m++){let g=n.map((b,w)=>w===s?m:0),y=n.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),h=[...n];h[o]=i*n[o];let f=[];for(let m=0;m<h.length;m++)m!==s&&f.push(m);let d=e.broadcast(c,h,f).getResult(0);r.replaceAllResultsWith([d]),r.erase()});rn("softmax",(r,e)=>{let t=r.getOperand(0),n=r.getAttr("axis"),s=t.type.rank,o=t.type.dtype,i=t.type.shape,a=To(s,n),u=e.reduce(t,e.scalarConstant(-1/0,o).getResult(0),[n],"max"),p=e.broadcast(u.getResult(0),i,a),c=e.sub(t,p.getResult(0)),h=e.exp(c.getResult(0)),f=e.reduce(h.getResult(0),e.scalarConstant(0,o).getResult(0),[n],"sum"),d=e.broadcast(f.getResult(0),i,a),m=e.div(h.getResult(0),d.getResult(0));r.replaceAllResultsWith([m.getResult(0)]),r.erase()});rn("log_softmax",(r,e)=>{let t=r.getOperand(0),n=r.getAttr("axis"),s=t.type.rank,o=t.type.dtype,i=t.type.shape,a=To(s,n),u=e.reduce(t,e.scalarConstant(-1/0,o).getResult(0),[n],"max"),p=e.broadcast(u.getResult(0),i,a),c=e.sub(t,p.getResult(0)),h=e.exp(c.getResult(0)),f=e.reduce(h.getResult(0),e.scalarConstant(0,o).getResult(0),[n],"sum"),d=e.log(f.getResult(0)),m=e.broadcast(d.getResult(0),i,a),g=e.sub(c.getResult(0),m.getResult(0));r.replaceAllResultsWith([g.getResult(0)]),r.erase()});l(Ew,"emitSigmoid");rn("sigmoid",(r,e)=>{let t=r.getOperand(0),n=Ew(e,t,t.type.dtype,t.type.shape);r.replaceAllResultsWith([n]),r.erase()});rn("gelu",(r,e)=>{let t=r.getOperand(0),n=t.type.dtype,s=t.type.shape,o=e.broadcast(e.scalarConstant(1.702,n).getResult(0),s,[]).getResult(0),i=e.mul(o,t).getResult(0),a=Ew(e,i,n,s),u=e.mul(t,a).getResult(0);r.replaceAllResultsWith([u]),r.erase()});rn("silu",(r,e)=>{let t=r.getOperand(0),n=Ew(e,t,t.type.dtype,t.type.shape),s=e.mul(t,n).getResult(0);r.replaceAllResultsWith([s]),r.erase()});rn("layer_norm",(r,e)=>{let t=r.getOperand(0),n=r.getOperand(1),s=r.getOperand(2),o=r.getAttr("axis"),i=r.getAttr("epsilon"),a=t.type.rank,u=t.type.dtype,p=t.type.shape,c=To(a,o),h=e.reduce(t,e.scalarConstant(0,u).getResult(0),[o],"mean"),f=e.broadcast(h.getResult(0),p,c),d=e.sub(t,f.getResult(0)),m=e.mul(d.getResult(0),d.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),T=e.mul(d.getResult(0),w.getResult(0)),S=e.broadcast(n,p,[o]),x=e.mul(T.getResult(0),S.getResult(0)),k=e.broadcast(s,p,[o]),A=e.add(x.getResult(0),k.getResult(0));r.replaceAllResultsWith([A.getResult(0)]),r.erase()});rn("batch_norm",(r,e)=>{let t=r.getOperand(0),n=r.getOperand(1),s=r.getOperand(2),o=r.getOperand(3),i=r.getOperand(4),a=r.getAttr("axis"),u=r.getAttr("epsilon"),p=t.type.rank,c=t.type.dtype,h=t.type.shape,f=e.broadcast(e.scalarConstant(u,c).getResult(0),i.type.shape,[]),d=e.add(i,f.getResult(0)),m=e.rsqrt(d.getResult(0)),g=e.broadcast(o,h,[a]),y=e.sub(t,g.getResult(0)),_=e.broadcast(m.getResult(0),h,[a]),b=e.mul(y.getResult(0),_.getResult(0)),w=e.broadcast(n,h,[a]),T=e.mul(b.getResult(0),w.getResult(0)),S=e.broadcast(s,h,[a]),x=e.add(T.getResult(0),S.getResult(0));r.replaceAllResultsWith([x.getResult(0)]),r.erase()});rn("where",(r,e)=>{let t=r.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 n=e.select(t,r.getOperand(1),r.getOperand(2));r.replaceAllResultsWith([n.getResult(0)]),r.erase()});rn("split",(r,e)=>{let t=r.getOperand(0),n=r.getAttr("dimension"),s=r.getAttr("split_sizes"),o=t.type.shape,i=[],a=0;for(let u of s){let p=o.map((h,f)=>f===n?a:0),c=o.map((h,f)=>f===n?a+u:h);i.push(e.slice(t,p,c).getResult(0)),a+=u}r.replaceAllResultsWith(i),r.erase()});rn("one_hot",(r,e)=>{let t=r.getOperand(0),n=r.getAttr("depth"),s=r.getAttr("axis")??-1,o=r.getAttr("on_value")??1,i=r.getAttr("off_value")??0,a=r.getResult(0).type,u=a.dtype,p=a.shape,c=s<0?t.type.rank+1+s:s,h=new I(p,"i32"),f=e._inferAndBuild("iota",[],{iota_dimension:c,tensor_type:h}),d=t.type.rank,m=[];for(let S=0;S<d;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),f.getResult(0),"eq"),b=e.broadcast(e.scalarConstant(o,u).getResult(0),p,[]),w=e.broadcast(e.scalarConstant(i,u).getResult(0),p,[]),T=e.select(_.getResult(0),b.getResult(0),w.getResult(0));r.replaceAllResultsWith([T.getResult(0)]),r.erase()});l(pn,"bcast");rn("elu",(r,e)=>{let t=r.getOperand(0),n=r.getAttr("alpha")??1,s=t.type.dtype,o=t.type.shape,i=pn(e,0,s,o),a=e.compare(t,i,"gt").getResult(0),u=pn(e,1,s,o),p=e.exp(t).getResult(0),c=e.sub(p,u).getResult(0),h=pn(e,n,s,o),f=e.mul(h,c).getResult(0),d=e.select(a,t,f);r.replaceAllResultsWith([d.getResult(0)]),r.erase()});rn("leaky_relu",(r,e)=>{let t=r.getOperand(0),n=r.getAttr("negative_slope")??.01,s=t.type.dtype,o=t.type.shape,i=pn(e,0,s,o),a=e.compare(t,i,"gt").getResult(0),u=pn(e,n,s,o),p=e.mul(u,t).getResult(0),c=e.select(a,t,p);r.replaceAllResultsWith([c.getResult(0)]),r.erase()});rn("celu",(r,e)=>{let t=r.getOperand(0),n=r.getAttr("alpha")??1,s=t.type.dtype,o=t.type.shape,i=pn(e,0,s,o),a=e.maximum(t,i).getResult(0),u=pn(e,n,s,o),p=e.div(t,u).getResult(0),c=e.exp(p).getResult(0),h=pn(e,1,s,o),f=e.sub(c,h).getResult(0),d=e.mul(u,f).getResult(0),m=e.minimum(i,d).getResult(0),g=e.add(a,m);r.replaceAllResultsWith([g.getResult(0)]),r.erase()});rn("selu",(r,e)=>{let t=r.getOperand(0),n=t.type.dtype,s=t.type.shape,o=1.0507009873554805,i=1.6732632423543772,a=pn(e,0,n,s),u=e.compare(t,a,"gt").getResult(0),p=pn(e,1,n,s),c=e.exp(t).getResult(0),h=e.sub(c,p).getResult(0),f=pn(e,i,n,s),d=e.mul(f,h).getResult(0),m=e.select(u,t,d).getResult(0),g=pn(e,o,n,s),y=e.mul(g,m);r.replaceAllResultsWith([y.getResult(0)]),r.erase()});rn("mish",(r,e)=>{let t=r.getOperand(0),n=t.type.dtype,s=t.type.shape,o=pn(e,1,n,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);r.replaceAllResultsWith([c.getResult(0)]),r.erase()});rn("hardswish",(r,e)=>{let t=r.getOperand(0),n=t.type.dtype,s=t.type.shape,o=pn(e,3,n,s),i=pn(e,0,n,s),a=pn(e,6,n,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),h=e.mul(t,c);r.replaceAllResultsWith([h.getResult(0)]),r.erase()});rn("hardsigmoid",(r,e)=>{let t=r.getOperand(0),n=t.type.dtype,s=t.type.shape,o=pn(e,6,n,s),i=pn(e,.5,n,s),a=pn(e,0,n,s),u=pn(e,1,n,s),p=e.div(t,o).getResult(0),c=e.add(p,i).getResult(0),h=e.minimum(e.maximum(c,a).getResult(0),u);r.replaceAllResultsWith([h.getResult(0)]),r.erase()});rn("embedding",(r,e)=>{let t=r.getOperand(0),n=r.getOperand(1),s=t.type.shape[t.type.rank-1],o=n.type.rank,i=e._inferAndBuild("gather",[t,n],{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});r.replaceAllResultsWith([i.getResult(0)]),r.erase()})});var ju,Cw=M(()=>{"use strict";nn();es();Kt();ju=class{static{l(this,"PatternApplicator")}constructor(e){this.patternSet=e}applyPatterns(e,t=10,n=null){let s=new An(e),o=0,i=[...e.opsRecursive()],a=0,u=new Set(i),p=l(d=>{!d||!d.parentBlock||u.has(d)||(u.add(d),i.push(d))},"enqueue"),c=Math.max(t,1)*Math.max(i.length,1)*4+1e3,h=0,f=!1;for(;a<i.length;){if(++h>c){f=!0;break}let d=i[a++];if(u.delete(d),!d.parentBlock)continue;let m=this.patternSet.getForOp(d.opName);for(let g of m){if(!g.match(d))continue;let y=d.parentBlock,_=d._prev,b=d._next,w=[];for(let x=0;x<d.numResults;x++)for(let k of d.getResult(x).getUsers())w.push(k);for(let x=0;x<d.numOperands;x++){let k=d.getOperand(x).definingOp;k&&w.push(k)}if(s.block=y,s.setInsertionPoint(d),!g.rewrite(d,s))continue;o++;for(let x of w)p(x);let T=_?_._next:y._head,S=y._size+2;for(;T&&T!==b&&S-- >0;)p(T),T=T._next;p(d);break}}return n&&(f&&n.level>=ue.INFO&&n.emit({type:"pass_detail",passName:"PatternApplicator",message:`pattern rewriting hit safety budget (${c}) without converging`,totalRewrites:o,level:ue.INFO}),n.level>=ue.DEBUG&&o>0&&n.emit({type:"pass_detail",passName:"PatternApplicator",totalRewrites:o,patternCount:this.patternSet.patterns.length,level:ue.DEBUG})),o>0?oe.CHANGED:oe.UNCHANGED}}});function cC(){if(Hc)return Hc;Hc=new Ua;for(let r of ne.allOps())if(r.getCanonicalizationPatterns){let e=r.getCanonicalizationPatterns();if(e)for(let t of e)Hc.add(t)}return Hc}var Hc,ia,Dw=M(()=>{"use strict";nn();uc();Cw();At();Hc=null;l(cC,"getCanonicalizationPatterns");ia=class extends Ge{static{l(this,"CanonicalizePass")}constructor(){super("canonicalize")}run(e,t){let n=cC();return new ju(n).applyPatterns(e,10,this.trace)}}});var lm,Pw,qu,oA=M(()=>{"use strict";At();_n();lm=class{static{l(this,"MemoryEffect")}constructor(e,t){this.kind=e,this.value=t}},Pw=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 n=[];for(let s=0;s<t.length;s++)t[s].effect.kind===an.READ&&n.push(t[s].op);return n}getWritersOf(e){let t=this._valueEffects.get(e);if(!t)return[];let n=[];for(let s=0;s<t.length;s++)t[s].effect.kind===an.WRITE&&n.push(t[s].op);return n}},qu=class{static get name(){return"memory_effect"}static get depKey(){return"memoryEffect"}static get dependencies(){return[]}static compute(e){let t=new Map,n=new Map,s=l((i,a,u)=>{let p=n.get(i);p||(p=[],n.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=an.NONE;if(a&&a.sideEffects?p=a.sideEffects:i.hasSideEffects&&i.hasSideEffects()&&(p=an.WRITE),p!==an.NONE)if(a&&a.getMemoryEffects){let c=a.getMemoryEffects(i);for(let h of c)u.push(h),s(h.value,i,h)}else{if(p&an.READ)for(let c=0;c<i.numOperands;c++){let h=new lm(an.READ,i.getOperand(c));u.push(h),s(i.getOperand(c),i,h)}if(p&an.WRITE)for(let c=0;c<i.numResults;c++){let h=new lm(an.WRITE,i.getResult(c));u.push(h),s(i.getResult(c),i,h)}}t.set(i,u)}return new Pw(t,n)}}});var no,$w=M(()=>{"use strict";nn();oA();Tr();Kt();no=class extends Ge{static{l(this,"DCEPass")}constructor(){super("dce"),this.preservedAnalyses=new Set,this.requiredAnalyses=[qu]}run(e,t){let n=!1,s=t?t.getAnalysis(qu,e):qu.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(),n=!0,i++;for(let p of u)p.parentBlock&&this._isDead(p,s)&&o.push(p)}return this.trace&&this.trace.level>=ue.DEBUG&&i>0&&this.trace.emit({type:"pass_detail",passName:this.name,erasedCount:i,level:ue.DEBUG}),n?oe.CHANGED:oe.UNCHANGED}_isDead(e,t){if(qt(e.opName)||e.regions&&e.regions.length>0)return!1;for(let n=0;n<e.numResults;n++)if(e.getResult(n).hasUses)return!1;return!t.hasSideEffect(e)}}});var cm,pm,iA=M(()=>{"use strict";cm=class r{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 r(e.primitive,e.args)}},pm=class r{static{l(this,"ScheduleTrace")}constructor(){this.steps=[]}record(e,t){let n=new cm(e,t);return this.steps.push(n),n}serialize(){return this.steps.map(e=>e.serialize())}static deserialize(e){let t=new r;for(let n of e)t.steps.push(cm.deserialize(n));return t}replay(e){for(let t of this.steps){let n=e[t.primitive];if(typeof n!="function")throw new Error(`Unknown schedule primitive: ${t.primitive}`);e._replaying=!0;try{n.call(e,...t.args)}finally{e._replaying=!1}}}get length(){return this.steps.length}clear(){this.steps.length=0}}});function aa(r){return Number.isFinite(r)}function aA(r){return Number.isNaN(r)?lA:r}function uA(r){return Number.isNaN(r)?cA:r}var lA,cA,Hr,Uu,Os,Ku=M(()=>{"use strict";Vs();lA=-1/0,cA=1/0;l(aa,"isFinite2");l(aA,"boundLo");l(uA,"boundHi");Hr=class{static{l(this,"IntBound")}constructor(e,t){this.min=e,this.max=t}isConst(){return this.min===this.max&&aa(this.min)}},Uu=new Hr(lA,cA),Os=class{static{l(this,"Analyzer")}constructor(){this._varBounds=new Map}bind(e,t,n){return this._varBounds.set(e,new Hr(t,n)),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,n]of e)typeof n=="number"&&n>0&&this.bind(t,0,n-1);return this}constIntBound(e){if(typeof e=="number")return new Hr(e,e);if(!(e instanceof Ce))return Uu;if(e.type==="var")return this._varBounds.get(e.name)||Uu;let t=e.args.length>0?this.constIntBound(e.args[0]):null,n=e.args.length>1?this.constIntBound(e.args[1]):null;switch(e.type){case"add":return new Hr(aA(t.min+n.min),uA(t.max+n.max));case"sub":return new Hr(aA(t.min-n.max),uA(t.max-n.min));case"neg":return new Hr(-t.max,-t.min);case"mul":return this._mulBound(t,n);case"max":return new Hr(Math.max(t.min,n.min),Math.max(t.max,n.max));case"min":return new Hr(Math.min(t.min,n.min),Math.min(t.max,n.max));case"div":return this._divBound(t,n,Math.floor);case"ceildiv":return this._divBound(t,n,Math.ceil);case"mod":return this._modBound(t,n);default:return Uu}}_mulBound(e,t){if(!aa(e.min)||!aa(e.max)||!aa(t.min)||!aa(t.max))return Uu;let n=[e.min*t.min,e.min*t.max,e.max*t.min,e.max*t.max];return new Hr(Math.min(...n),Math.max(...n))}_divBound(e,t,n){if(t.isConst()&&t.min>0){let s=aa(e.min)?n(e.min/t.min):e.min,o=aa(e.max)?n(e.max/t.min):e.max;return new Hr(s,o)}return Uu}_modBound(e,t){return t.isConst()&&t.min>0?new Hr(0,t.min-1):Uu}canProveGreaterEqual(e,t){return this.constIntBound(Ce.sub(e,t)).min>=0}canProveLess(e,t){return this.constIntBound(Ce.sub(e,t)).max<0}canProveNonNegative(e){return this.canProveGreaterEqual(e,0)}canProveEqual(e,t){if(Ce.equals(e,t))return!0;let n=this.constIntBound(Ce.sub(e,t));return n.min===0&&n.max===0}}});function pi(r){if(r==null)return null;if(typeof r=="number")return r;switch(r.type){case"IntImmNode":return r.value;case"VariableNode":return Ce.var(r.name);case"MathOpNode":{if(r.b===null||r.b===void 0){if(r.op==="-"){let s=pi(r.a);return s===null?null:Ce.neg(s)}return null}let e=pC[r.op];if(!e)return null;let t=pi(r.a);if(t===null)return null;let n=pi(r.b);return n===null||(r.op==="//"||r.op==="%")&&(typeof n!="number"||n<=0)?null:Ce[e](t,n)}case"CallExternNode":{if(r.args.length!==2||r.externName!=="max"&&r.externName!=="min")return null;let e=pi(r.args[0]);if(e===null)return null;let t=pi(r.args[1]);return t===null?null:r.externName==="max"?Ce.max(e,t):Ce.min(e,t)}default:return null}}function la(r,e){let t=pi(e);return t===null?null:r.constIntBound(t)}function fC(r,e,t){let n=pi(e);if(n===null)return null;let s=pi(t);return s===null?null:r.constIntBound(Ce.sub(n,s))}function fm(r,e,t,n,s){let o=s[e];if(!o)return!1;let i=fC(r,t,n);return i===null?!1:o(i)}function Mo(r,e){if(e==null)return!1;if(e.type==="IntImmNode")return e.value!==0;if(e.type==="CompareNode")return fm(r,e.direction,e.a,e.b,pA);if(e.type==="MathOpNode"){if(e.op==="*"&&e.b)return Mo(r,e.a)&&Mo(r,e.b);if(Fw.has(e.op))return fm(r,mA[e.op],e.a,e.b,pA)}return!1}function ua(r,e){if(e==null)return!1;if(e.type==="IntImmNode")return e.value===0;if(e.type==="CompareNode")return fm(r,e.direction,e.a,e.b,fA);if(e.type==="MathOpNode"){if(e.op==="*"&&e.b)return ua(r,e.a)||ua(r,e.b);if(Fw.has(e.op))return fm(r,mA[e.op],e.a,e.b,fA)}return!1}function hA(r){return!r||r.type!=="MathOpNode"||r.op!=="*"?null:r.b&&r.b.type==="IntImmNode"&&r.b.value>0?{factor:r.a,c:r.b.value}:r.a&&r.a.type==="IntImmNode"&&r.a.value>0?{factor:r.b,c:r.a.value}:null}function dA(r,e,t,n){let s=la(r,e);return s===null?!1:s.min>=t&&s.max<=n}var pC,Fw,mA,pA,fA,hm,Yc=M(()=>{"use strict";Ku();Vs();De();pC={"+":"add","-":"sub","*":"mul","//":"div","%":"mod"},Fw=new Set(["<","<=",">",">=","==","!="]),mA={"<":"lt","<=":"le",">":"gt",">=":"ge","==":"eq","!=":"ne"};l(pi,"irToSymInt");l(la,"irBound");l(fC,"diffBound");pA={lt:l(r=>r.max<0,"lt"),le:l(r=>r.max<=0,"le"),gt:l(r=>r.min>0,"gt"),ge:l(r=>r.min>=0,"ge"),eq:l(r=>r.min===0&&r.max===0,"eq"),ne:l(r=>r.min>0||r.max<0,"ne")},fA={lt:l(r=>r.min>=0,"lt"),le:l(r=>r.min>0,"le"),gt:l(r=>r.max<=0,"gt"),ge:l(r=>r.max<0,"ge"),eq:l(r=>r.min>0||r.max<0,"eq"),ne:l(r=>r.min===0&&r.max===0,"ne")};l(fm,"proveCompare");l(Mo,"proveTrue");l(ua,"proveFalse");l(hA,"asScaledVar");l(dA,"boundWithin");hm=class{static{l(this,"RewriteSimplify")}constructor(e=new Os){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),n=this.simplify(e.b);return Mo(this.analyzer,new qe(e.direction,t,n))?new L(1):ua(this.analyzer,new qe(e.direction,t,n))?new L(0):t===e.a&&n===e.b?e:new qe(e.direction,t,n)}_simplifyMathOp(e){if(e.b===null||e.b===void 0){let o=this.simplify(e.a);return o===e.a?e:new C(e.op,o)}let t=this.simplify(e.a),n=this.simplify(e.b),s=In(e.op,t,n);if(!s||s.type!=="MathOpNode")return s;if(s.op==="//"&&s.b.type==="IntImmNode"&&s.b.value>0){let o=s.b.value,i=hA(s.a);if(i&&i.c===o)return i.factor;if(dA(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=hA(s.a);if(i&&i.c===o)return new L(0);if(dA(this.analyzer,s.a,0,o-1))return s.a}if(Fw.has(s.op)){if(Mo(this.analyzer,s))return new L(1);if(ua(this.analyzer,s))return new L(0)}return s}}});function gA(r,e,t){if(typeof t!="number"||t<0)return"unknown";let n=la(r,e);return n===null?"unknown":n.min>=0&&n.max<=t-1?"in":n.min>t-1||n.max<0?"oob":"unknown"}function yt(r,e){if(!(!r||typeof r!="object"))switch(r.type){case"VariableNode":r.name&&e.add(r.name);return;case"BufferLoadNode":if(r.indices)for(let t of r.indices)yt(t,e);return;case"BufferStoreNode":if(r.indices)for(let t of r.indices)yt(t,e);yt(r.value,e);return;case"MathOpNode":yt(r.a,e),r.b&&yt(r.b,e);return;case"CompareNode":yt(r.a,e),yt(r.b,e);return;case"CastNode":yt(r.expr,e);return;case"CallExternNode":for(let t of r.args)yt(t,e);return;case"IfThenElseNode":yt(r.condition,e),yt(r.thenBody,e),r.elseBody&&yt(r.elseBody,e);return;case"SeqNode":for(let t of r.stmts)yt(t,e);return;case"LetStmtNode":yt(r.value,e),yt(r.body,e);return;default:return}}function dm(r,e){let t=new Set((r.writes||[]).map(s=>s.buffer&&s.buffer.name)),n=[r.body,r.initBody];for(;n.length>0;){let s=n.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)yt(o,e);if(s.type==="BufferStoreNode"){s.value&&n.push(s.value);continue}if(s.type==="SeqNode"){for(let o of s.stmts)n.push(o);continue}if(s.type==="IfThenElseNode"){n.push(s.thenBody),s.elseBody&&n.push(s.elseBody);continue}if(s.type==="LetStmtNode"){n.push(s.body);continue}}}}function hC(r,e){let t=new Set;if(yt(e.body,t),e.initBody&&yt(e.initBody,t),!t.has(r))return!1;let n=new Set;return dm(e,n),!n.has(r)}function ca(r,e){if(!(!r||typeof r!="object"))switch(r.type){case"BlockNode":e.push(r),ca(r.body,e),r.initBody&&ca(r.initBody,e);return;case"ForNode":case"AllocateNode":case"LetStmtNode":ca(r.body,e);return;case"SeqNode":for(let t of r.stmts)ca(t,e);return;case"IfThenElseNode":ca(r.thenBody,e),r.elseBody&&ca(r.elseBody,e);return;default:return}}function zw(r){let e=r.loopVar.name,t=[];ca(r.body,t);for(let n of t)if(hC(e,n))return n.name;return null}var Vw=M(()=>{"use strict";Ku();Vs();Yc();l(gA,"classifyBufferIndex");l(yt,"collectVarsUsed");l(dm,"collectWriteIndexVars");l(hC,"loopVarIsReductionOf");l(ca,"collectBlocksUnder");l(zw,"loopCarriesReduction")});var Wu,Gw=M(()=>{"use strict";De();Vw();Ku();Yc();Wu=class r{static{l(this,"ScheduleValidator")}static validate(e){let t=[],n={boundVars:new Set,threadBindings:new Map,parallelExtents:new Map,parLoops:[],innermostLoopVar:null,loopStack:[],analyzer:new Os,condDepth:0,errors:t};for(let s of e.params)n.boundVars.add(s.name);return r._visitNode(e.body,n),r._checkPartitionConsistency(n),t}static _checkPartitionConsistency(e){if(e.parallelExtents.size<=1)return;let t=[...e.parallelExtents.entries()].map(([n,s])=>`'${s}'(extent ${n})`);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":r._visitFor(e,t);break;case"BlockNode":r._visitBlock(e,t);break;case"SeqNode":for(let n of e.stmts)r._visitNode(n,t);break;case"IfThenElseNode":r._visitExpr(e.condition,t),t.condDepth++,r._visitNode(e.thenBody,t),e.elseBody&&r._visitNode(e.elseBody,t),t.condDepth--;break;case"AllocateNode":r._visitNode(e.body,t);break;case"LetStmtNode":t.boundVars.add(e.variable.name),r._visitNode(e.body,t),t.boundVars.delete(e.variable.name);break;case"BufferStoreNode":r._validateBufferAccess(e,t);break;case"EvaluateNode":break;default:break}}static _visitFor(e,t){let n=e.loopVar.name;if(t.boundVars.has(n)&&t.errors.push(`Duplicate loop variable: ${n}`),e.kind===F.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,n)):t.errors.push(`Thread-bound loop '${n}' missing threadTag`)),e.kind===F.VECTORIZED){let a=e.extent;a.type==="IntImmNode"&&a.value<=0&&t.errors.push(`Vectorized loop '${n}' has non-positive extent ${a.value}`)}if(e.kind===F.PARALLEL){let a=e.extent;a&&a.type==="IntImmNode"&&!t.parallelExtents.has(a.value)&&t.parallelExtents.set(a.value,n)}(e.kind===F.PARALLEL||e.kind===F.THREAD_BINDING)&&r._checkNoNestedParallel(e.body,n,t);let s=e.kind===F.PARALLEL||e.kind===F.VECTORIZED;s&&t.parLoops.push({varName:n,kind:e.kind}),t.boundVars.add(n);let o=t.analyzer.getVarBound(n);e.extent&&e.extent.type==="IntImmNode"&&e.extent.value>0?t.analyzer.bind(n,0,e.extent.value-1):t.analyzer.setVarBound(n,null);let i=t.innermostLoopVar;t.innermostLoopVar=n,t.loopStack.push(n),r._visitNode(e.body,t),t.loopStack.pop(),t.innermostLoopVar=i,t.boundVars.delete(n),t.analyzer.setVarBound(n,o),s&&t.parLoops.pop(),e.kind===F.THREAD_BINDING&&e.threadTag&&t.threadBindings.delete(e.threadTag)}static _checkNoNestedParallel(e,t,n){e&&e.type==="ForNode"&&e.kind===F.PARALLEL&&n.errors.push(`Parallel loop '${e.loopVar.name}' nested inside parallel/thread-bound loop '${t}'`)}static _visitBlock(e,t){let n=[];for(let s of e.iterVars)s.iterVar&&(t.boundVars.add(s.iterVar.name),n.push([s.iterVar.name,t.analyzer.getVarBound(s.iterVar.name)]),t.analyzer.setVarBound(s.iterVar.name,s.binding?la(t.analyzer,s.binding):null));if(e.initBody&&t.innermostLoopVar){let s=new Set;dm(e,s);let o=new Set;yt(e.body,o),yt(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;yt(e.body,s),e.initBody&&yt(e.initBody,s);let o=new Set;dm(e,o);for(let i of t.parLoops)if(s.has(i.varName)&&!o.has(i.varName)){let a=i.kind===F.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`)}}r._visitNode(e.body,t),e.initBody&&r._visitNode(e.initBody,t);for(let s of e.iterVars)s.iterVar&&t.boundVars.delete(s.iterVar.name);for(let[s,o]of n)t.analyzer.setVarBound(s,o)}static _validateBufferAccess(e,t){if(e.buffer?(r._checkRank(e,t),r._checkBounds(e,t)):t.errors.push("BufferStore with null buffer"),e.indices)for(let n of e.indices)r._visitExpr(n,t);e.value&&r._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 n=0;n<e.indices.length;n++){let s=e.buffer.shape[n];gA(t.analyzer,e.indices[n],s)==="oob"&&t.errors.push(`Buffer '${e.buffer.name}' access is out of bounds on axis ${n}: index is provably outside [0, ${s-1}]`)}}static _visitExpr(e,t){if(!(!e||typeof e!="object"))switch(e.type){case"BufferLoadNode":if(e.buffer?(r._checkRank(e,t),r._checkBounds(e,t)):t.errors.push("BufferLoad with null buffer"),e.indices)for(let n of e.indices)r._visitExpr(n,t);break;case"BufferStoreNode":r._validateBufferAccess(e,t);break;case"MathOpNode":case"CompareNode":r._visitExpr(e.a,t),r._visitExpr(e.b,t);break;case"CastNode":r._visitExpr(e.expr,t);break;case"CallExternNode":for(let n of e.args)r._visitExpr(n,t);break;case"IfThenElseNode":r._visitExpr(e.condition,t),t.condDepth++,r._visitNode(e.thenBody,t),e.elseBody&&r._visitNode(e.elseBody,t),t.condDepth--;break}}}});var jw,Hu,qw=M(()=>{"use strict";De();jw=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 n=t.pop();if(n.isBlock)e.push(n);else for(let s of n.children)t.push(s)}return e}childLoops(){return this.children.filter(e=>e.isLoop)}},Hu=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 n=t.pop();this._nodeToSRef.delete(n.node),n.isLoop?this._loopSRefs.delete(n):n.isBlock&&(this._blockSRefs.delete(n),this._blockNameToSRef.get(n.node.name)===n&&this._blockNameToSRef.delete(n.node.name));for(let s of n.children)t.push(s)}}_buildSubtree(e,t){let n=[],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 jw(o,i);this._register(u),a?n.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 n}_build(e,t){let n=this._buildSubtree(e,t);if(t)for(let s of n)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(n=>n.node.kind!==F.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 n=this._nodeToSRef.get(e);if(!n)return!1;let s=n.parent,o=this.root===n;this._unregisterSubtree(n);let i=this._buildSubtree(t,s);if(s){let a=s.children.indexOf(n);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 n=t.parent,s=this.root===t;if(this._unregisterSubtree(t),n){let o=n.children.indexOf(t);o>=0&&n.children.splice(o,1)}else s&&(this.root=null);return!0}}});var Uw,Kw,mm,yA=M(()=>{"use strict";qw();De();Uw=class{static{l(this,"LoopBinding")}constructor(e,t,n,s,o){this.loopSRef=e,this.iterVar=t,this.extent=n,this.kind=s,this.threadTag=o}},Kw=class{static{l(this,"BlockBinding")}constructor(e,t,n,s){this.blockSRef=e,this.iterVars=t,this.readBuffers=n,this.writeBuffers=s}},mm=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 Hu(this.primFunc),this._loopBindings=new Map,this._blockBindings=new Map;for(let e of this._tree.allLoops()){let t=e.node,n=t.extent&&t.extent.type==="IntImmNode"?t.extent.value:null;this._loopBindings.set(t.loopVar.name,new Uw(e,t.loopVar.name,n,t.kind,t.threadTag))}for(let e of this._tree.allBlocks()){let t=e.node,n=[];for(let i of t.iterVars)i.iterVar&&n.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 Kw(e,n,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===F.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 gm,_A=M(()=>{"use strict";Gr();gm=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 n=t.stmts.indexOf(e);if(n>=0){t.stmts.splice(n,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,n){e&&Ss(e,s=>{s.type==="BufferLoadNode"&&s.buffer===t&&(s.buffer=n)})}redirectBuffer(e,t,n){e&&Ss(e,s=>{(s.type==="BufferLoadNode"||s.type==="BufferStoreNode")&&s.buffer===t&&(s.buffer=n)})}}});function ym(r,e,t){if(!r||typeof r!="object")return r;if(Array.isArray(r))return r.map(e);let n=Object.create(Object.getPrototypeOf(r));switch(n.type=r.type,n._parent=null,n._parentKey=null,n._parentIdx=-1,r.type){case"ForNode":return n.loopVar=r.loopVar,n.min=e(r.min),n.extent=e(r.extent),n.kind=r.kind,n.body=e(r.body),n.threadTag=r.threadTag,n._setChild("body",n.body),n;case"BlockNode":return n.name=r.name,n.iterVars=r.iterVars.map(e),n.reads=r.reads,n.writes=r.writes,n.body=e(r.body),n.initBody=r.initBody?e(r.initBody):null,n._setChild("body",n.body),n._setChild("initBody",n.initBody),n;case"SeqNode":return n.stmts=r.stmts.map(e),n._setChildren("stmts",n.stmts),n;case"IfThenElseNode":return n.condition=e(r.condition),n.thenBody=e(r.thenBody),n.elseBody=r.elseBody?e(r.elseBody):null,n._setChild("thenBody",n.thenBody),n._setChild("elseBody",n.elseBody),n;case"BufferStoreNode":return n.buffer=r.buffer,n.indices=r.indices.map(e),n.value=e(r.value),n;case"BufferLoadNode":return n.buffer=r.buffer,n.indices=r.indices.map(e),n;case"MathOpNode":return n.op=r.op,n.a=e(r.a),n.b=e(r.b),n;case"CompareNode":return n.direction=r.direction,n.a=e(r.a),n.b=e(r.b),n;case"CastNode":return n.expr=e(r.expr),n.fromDtype=r.fromDtype,n.toDtype=r.toDtype,n;case"CallExternNode":return n.externName=r.externName,n.args=r.args.map(e),n.dtype=r.dtype,n;default:return t(r,n,e)}}var Ww=M(()=>{"use strict";l(ym,"cloneIRShared")});function fi(r,e,t){return ti(r,n=>n.type==="VariableNode"&&n.name===e?t():n,{bindVars:!1})}function mC(r,e,t,n){return ti(r,s=>s.type==="BufferLoadNode"&&s.buffer&&s.buffer.name===e?(n.n++,t(s)):s,{bindVars:!1})}function Hw(r,e){return Sc(r,t=>t.type==="BufferLoadNode"&&t.buffer&&t.buffer.name===e)}function Ju(r,e){if(!r||typeof r!="object")return!1;if((r.type==="BufferLoadNode"||r.type==="BufferStoreNode")&&r.indices){for(let t of r.indices)if(Hw(t,e))return!0}for(let t of["a","b","expr","value","condition","thenBody","elseBody","body","initBody"])if(r[t]&&Ju(r[t],e))return!0;if(r.args){for(let t of r.args)if(Ju(t,e))return!0}if(r.indices){for(let t of r.indices)if(Ju(t,e))return!0}if(r.stmts){for(let t of r.stmts)if(Ju(t,e))return!0}return!1}function Wt(r){return ym(r,Wt,(e,t,n)=>{if(e.type==="BlockRealizeNode")return t.iterVar=e.iterVar,t.binding=n(e.binding),t;for(let s of Object.keys(e))s==="_parent"||s==="_parentKey"||s==="_parentIdx"||(t[s]=e[s]);return t})}function _m(r,e){if(r){if(r.type==="BufferStoreNode"){e.push(r);return}if(r.type==="SeqNode"){for(let t of r.stmts)_m(t,e);return}(r.type==="ForNode"||r.type==="BlockNode")&&_m(r.body,e)}}function bA(r){return r.indices.map(e=>e&&e.type==="VariableNode"?e.name:null)}function wA(r,e,t){let n=e.indices.map(i=>i.name),s=n.map((i,a)=>`__inl_${t}_${a}`),o={n:0};return mC(r,e.buffer.name,i=>{let a=Wt(e.value);for(let u=0;u<n.length;u++)a=fi(a,n[u],()=>new Ne(s[u],"int32"));for(let u=0;u<n.length;u++)a=fi(a,s[u],()=>Wt(i.indices[u]));return a},o),o.n}function TA(r,e){let t=[r];for(;t.length>0;){let n=t.pop();if(!(!n||typeof n!="object")){if(n.type==="BlockNode"&&n.reads){let s=n.reads.filter(o=>!(o.buffer&&o.buffer.name===e));s.length!==n.reads.length&&(n.reads=s)}if(n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),n.initBody&&t.push(n.initBody)}}}function pa(r){return r.type==="IntImmNode"?r.value:null}function Yu(r,e="int32"){return new Ne(`${r}_${gC++}`,e)}function Jc(r,e,t){if(!(!r||typeof r!="object")){r.type==="BufferLoadNode"&&r.buffer&&e.set(r.buffer.name,r.buffer),r.type==="BufferStoreNode"&&r.buffer&&t.set(r.buffer.name,r.buffer);for(let n of["a","b","value","expr","condition","thenBody","elseBody","body","initBody"])r[n]&&Jc(r[n],e,t);if(r.indices)for(let n of r.indices)Jc(n,e,t);if(r.args)for(let n of r.args)Jc(n,e,t);if(r.stmts)for(let n of r.stmts)Jc(n,e,t)}}var dC,gC,Sn,fa=M(()=>{"use strict";De();xo();iA();Gw();yA();_A();qw();Vw();Ww();Gr();dC=new Set(["+","*","min","max"]);l(fi,"substituteVar");l(mC,"replaceBufferLoads");l(Hw,"loadsBuffer");l(Ju,"usesBufferInIndex");l(Wt,"cloneExprTree");l(_m,"collectStores");l(bA,"storeIndexNames");l(wA,"inlineStoreValue");l(TA,"dropBufferReads");l(pa,"getConstExtent");gC=0;l(Yu,"freshVar");Sn=class{static{l(this,"Schedule")}constructor(e){this.func=e,this.trace=new pm,this.state=new mm(e),this._replaying=!1,this._srefTree=new Hu(e),this.mutator=new gm(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,n=l(s=>{if(!(!s||typeof s!="object"||t)){if(s.type==="ForNode"&&s.loopVar&&s.loopVar.name===e){t=s;return}if(s.body&&n(s.body),s.initBody&&n(s.initBody),s.stmts)for(let o of s.stmts)n(o);s.thenBody&&n(s.thenBody),s.elseBody&&n(s.elseBody)}},"walk");return n(this.func.body),t||e}split(e,t){e=this._resolveLoop(e);let n=pa(e.extent);if(n===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(n/t),o=Yu(`${e.loopVar.name}_o`),i=Yu(`${e.loopVar.name}_i`),a=e.loopVar.name,u=Wt(e.body),p=new de(i,new L(0),new L(t),e.kind,u,e.threadTag);if(n%t!==0){let f=new C("+",new C("*",o,new L(t)),i),d=new C("<",f,new L(n)),m=new xe(d,p.body);p.body=m,p._setChild("body",m)}let h=new de(o,new L(0),new L(s),e.kind,p,e.threadTag);return fi(p.body,a,()=>new C("+",new C("*",o,new L(t)),i)),this.mutator.replaceNode(e,h),this._replaceInTree(e,h),this._replaying||this.trace.record("split",[e.loopVar.name,t]),[h,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 n=null,s=1/0,o=new Map,i=l((c,h)=>{if(c){if(c.type==="ForNode"){t.has(c)&&(o.set(c,h),h<s&&(s=h,n=c)),i(c.body,h+1);return}if(c.type==="SeqNode")for(let f of c.stmts)i(f,h);else c.type==="IfThenElseNode"?i(c.thenBody,h):(c.type==="BlockNode"||c.type==="AllocateNode"||c.type==="LetStmtNode")&&i(c.body,h)}},"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,n);this.mutator.replaceNode(n,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 h=c<e.length-1?e[c+1]:p;e[c].body=h,e[c]._setChild("body",h)}this._replaceInTree(n,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 n=new Set(e),s=[],o=t,i=!1,a=null;for(;o;)if(o.type==="ForNode"){if(e.has(o)){if(n.delete(o),i=!0,n.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(n.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 n=pa(e.extent),s=pa(t.extent);if(n===null||s===null)throw new Error("Cannot fuse loops with non-constant extents");let o=n*s,i=Yu(`${e.loopVar.name}_${t.loopVar.name}_fused`),a=e.loopVar.name,u=t.loopVar.name,p=new de(i,new L(0),new L(o),e.kind,t.body);return fi(p.body,a,()=>new C("//",i,new L(s))),fi(p.body,u,()=>new C("%",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,n){if(t.length!==n.length)throw new Error("tile: loopIndices and tileSizes must have same length");let s=this.getLoops(e),o=t.map(h=>{if(h>=s.length)throw new Error(`tile: loop index ${h} out of range`);return s[h]}),i=[],a=[];for(let h=0;h<o.length;h++){let d=this.getLoops(e).find(y=>y.loopVar.name===o[h].loopVar.name||y===o[h]);if(!d)throw new Error(`tile: lost track of loop at index ${h}`);let[m,g]=this.split(d,n[h]);i.push(m),a.push(g)}let u=this.getLoops(e),p=[],c=[];for(let h of u)i.some(f=>f.loopVar.name===h.loopVar.name)?p.push(h):a.some(f=>f.loopVar.name===h.loopVar.name)&&c.push(h);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(pa(e.extent)===null)throw new Error("Cannot vectorize loop with non-constant extent");let n=zw(e);if(n!==null)throw new Error(`Cannot vectorize reduction loop '${e.loopVar.name}' (loop-carried dependency in block '${n}')`);e.kind=F.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=F.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=zw(e);if(t!==null)throw new Error(`Cannot parallelize reduction loop '${e.loopVar.name}' (loop-carried dependency in block '${t}')`);e.kind=F.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 n=["blockIdx.x","blockIdx.y","blockIdx.z","threadIdx.x","threadIdx.y","threadIdx.z"];if(!n.includes(t))throw new Error(`Invalid thread tag: ${t}. Must be one of: ${n.join(", ")}`);e.kind=F.THREAD_BINDING,e.threadTag=t,this.state.invalidate(),this._replaying||this.trace.record("bindThread",[e.loopVar.name,t])}rfactor(e,t,n){let s=this.getBlock(e),o=this.getLoops(e),i=o.find(D=>D.loopVar.name===t);if(!i)throw new Error(`rfactor: reduction loop '${t}' not found for block '${e}'`);let a=pa(i.extent);if(a===null)throw new Error(`rfactor: reduction loop '${t}' has non-constant extent`);if(!Number.isInteger(n)||n<=1||n>=a||a%n!==0)throw new Error(`rfactor: factor ${n} 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,h=u.value.op,f=l(D=>D&&D.type==="BufferLoadNode"&&D.buffer===p,"isAccLoad"),d;if(f(u.value.a))d=u.value.b;else if(f(u.value.b))d=u.value.a;else throw new Error(`rfactor: accumulator load not found in block '${e}' body`);if(!dC.has(h))throw new Error(`rfactor: op '${h}' 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(D=>D.loopVar.name!==t),y=a/n,_=new me(`${p.name}_rf`,[n,...p.shape],p.dtype,p.scope),b=Yu(`${t}_rfi`),w=Yu(`${t}_rfo`),T=Yu(`${t}_rfp`),S=l(D=>[D,...c.map(Wt)],"cfIdx"),x=fi(Wt(d),t,()=>new C("+",new C("*",w,new L(n)),b)),k=new Y(_,S(b),new C(h,new q(_,S(b)),x)),A=new Y(_,S(b),Wt(m)),N=new pe(`${e}_rf_p`,[],s.reads.map(D=>({buffer:D.buffer})),[{buffer:_}],k,A),v=new de(w,new L(0),new L(y),F.SERIAL,N);v=new de(b,new L(0),new L(n),F.SERIAL,v);for(let D=g.length-1;D>=0;D--)v=new de(g[D].loopVar,new L(0),Wt(g[D].extent),F.SERIAL,v);let O=new Y(p,c.map(Wt),new C(h,new q(p,c.map(Wt)),new q(_,S(T)))),R=new Y(p,c.map(Wt),Wt(m)),B=new pe(`${e}_rf_c`,[],[{buffer:_}],[{buffer:p}],O,R),E=new de(T,new L(0),new L(n),F.SERIAL,B);for(let D=g.length-1;D>=0;D--)E=new de(g[D].loopVar,new L(0),Wt(g[D].extent),F.SERIAL,E);let z=new te([v,E]);return this.mutator.replaceNode(o[0],z),this._replaceInTree(o[0],z),this._replaying||this.trace.record("rfactor",[e,t,n]),_}decomposeReduction(e){let t=this.getBlock(e);if(!t.initBody)throw new Error(`decomposeReduction: block '${e}' has no initBody`);let n=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)yt(g,i);let a=n.filter(g=>i.has(g.loopVar.name));if(n.filter(g=>!i.has(g.loopVar.name)).length===0)throw new Error(`decomposeReduction: block '${e}' has no reduction loop`);let p=new Y(o,s.indices.map(Wt),Wt(t.initBody.value)),h=new pe(`${e}_init`,[],[],[{buffer:o}],p);for(let g=a.length-1;g>=0;g--)h=new de(a[g].loopVar,new L(0),Wt(a[g].extent),F.SERIAL,h);let d=new pe(`${e}_upd`,[],t.reads.map(g=>({buffer:g.buffer})),[{buffer:o}],Wt(s));for(let g=n.length-1;g>=0;g--)d=new de(n[g].loopVar,new L(0),Wt(n[g].extent),F.SERIAL,d);let m=new te([h,d]);this.mutator.replaceNode(n[0],m),this._replaceInTree(n[0],m),this._replaying||this.trace.record("decomposeReduction",[e])}cacheWrite(e,t,n="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,n);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 Ne(`${u.name}_o${_}`,"int32")),c=new Y(a,p,new q(u,p)),f=new pe(`${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--)f=new de(p[y],new L(0),new L(a.shape[y]),F.SERIAL,f);let d=o[0],m=new te([]),g=new Ve(u,n,m);this.mutator.replaceNode(d,g),m.stmts.push(d,f),this._replaceInTree(d,g),this._replaying||this.trace.record("cacheWrite",[e,t,n])}setScope(e,t,n){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=n,this.state.invalidate(),this._replaying||this.trace.record("setScope",[e,t,n])}storageAlign(e,t,n,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:n,factor:s,offset:o||0},this.state.invalidate(),this._replaying||this.trace.record("storageAlign",[e,t,n,s,o||0])}_removeBlockNest(e,t){let n=this.getLoops(e),s=n.length>0?n[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 n=[];_m(t.body,n);let s=n[0];if(!s)throw new Error("computeInline: producer has no single store to inline");let o=s.buffer;if(bA(s).some(i=>i===null))throw new Error("computeInline: producer indices must be simple loop variables");if(Hw(s.value,o.name))throw new Error("computeInline: producer is self-referential (recurrence), cannot inline");if(Ju(this.func.body,o.name))throw new Error(`computeInline: buffer '${o.name}' is used inside an index expression (indirect), cannot safely inline`);if(wA(this.func.body,s,e)===0)throw new Error(`computeInline: buffer '${o.name}' has no consumers to inline into`);TA(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 n=[];if(_m(t.body,n),n.length===0)throw new Error("computeInlineBlock: producer has no stores to inline");let s=new Set(n.map(i=>i.buffer.name));if(s.size!==n.length)throw new Error("computeInlineBlock: buffer written more than once in block");for(let i of n){if(bA(i).some(a=>a===null))throw new Error("computeInlineBlock: producer indices must be simple loop variables");for(let a of s)if(Hw(i.value,a))throw new Error("computeInlineBlock: producer store depends on a co-produced buffer");if(Ju(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 n)o+=wA(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)TA(this.func.body,i);this._removeBlockNest(e,t),this._replaying||this.trace.record("computeInlineBlock",[e])}_relocateBlockToLoop(e,t,n){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=pa(a.extent),p=pa(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=Wt(a.body);fi(c,a.loopVar.name,()=>o.loopVar);let h=new te([]);this.mutator.replaceNode(a,h);let f=o.body;if(f&&f.type==="SeqNode")n?f.stmts.unshift(c):f.stmts.push(c);else{let d=n?new te([c,f]):new te([f,c]);o.body=d,o._setChild("body",d)}return this._replaceInTree(a,h),this._replaceInTree(o,o),o.loopVar.name}computeAt(e,t){let n=this._relocateBlockToLoop(e,t,!0);this._replaying||this.trace.record("computeAt",[e,n])}reverseComputeAt(e,t){let n=this._relocateBlockToLoop(e,t,!1);this._replaying||this.trace.record("reverseComputeAt",[e,n])}cacheRead(e,t,n="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,n),p=a.shape.map((y,_)=>new Ne(`${u.name}_i${_}`,"int32")),c=new Y(u,p,new q(a,p)),f=new pe(`${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--)f=new de(p[y],new L(0),new L(a.shape[y]),F.SERIAL,f);this.mutator.redirectReads(s.body,a,u),s.initBody&&this.mutator.redirectReads(s.initBody,a,u),i.buffer=u;let d=o[0],m=new te([f]),g=new Ve(u,n,m);this.mutator.replaceNode(d,g),m.stmts.push(d),this._replaceInTree(d,g),this._replaying||this.trace.record("cacheRead",[e,t,n])}fuseConsumer(e,t){let n=this.getBlock(e),s=this.getBlock(t),o=this.getLoops(e),i=this.getLoops(t);if(!n.body||n.body.type!=="BufferStoreNode")throw new Error(`fuseConsumer: producer '${e}' body is not a store`);let a=new Set;for(let g of n.body.indices)yt(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=Wt(p.body),h=Wt(s.body);for(let g=0;g<i.length;g++){let y=u[g].loopVar.name;h=fi(h,i[g].loopVar.name,()=>new Ne(y,"int32"))}let f=new pe(`${t}_fused`,[],s.reads.map(g=>({buffer:g.buffer})),s.writes.map(g=>({buffer:g.buffer})),h),d=new te([c,f]);for(let g=u.length-1;g>=0;g--){let y=u[g];d=new de(y.loopVar,new L(0),Wt(y.extent),y.kind,d,y.threadTag)}let m=i[0];this.mutator.replaceNode(o[0],d),this.mutator.removeNode(m),this._replaceInTree(o[0],d),this._removeFromTree(m),this._replaying||this.trace.record("fuseConsumer",[e,t])}annotate(e,t,n){if(e=this._resolveLoop(e),e.type!=="ForNode")throw new Error("annotate expects ForNode");e.annotations||(e.annotations={}),e.annotations[t]=n,this.state.invalidate(),this._replaying||this.trace.record("annotate",[e.loopVar.name,t,n])}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 n=this._resolveLoop(e);if(!n||n.type!=="ForNode")throw new Error("blockize expects a loop");let s=new Map,o=new Map;Jc(n,s,o);let i=t||`blockized_${n.loopVar.name}`,a=new pe(i,[],[...s.values()].map(u=>({buffer:u})),[...o.values()].map(u=>({buffer:u})),new te([]));return this.mutator.replaceNode(n,a),a.body=n,a._setChild("body",n),this._replaceInTree(n,a),this._replaying||this.trace.record("blockize",[n.loopVar.name]),a}getTrace(){return this.trace}verify(){return Wu.validate(this.func)}};l(Jc,"collectBufferAccess")});function dt(r,e){let t=bm.get(r);return t&&t.has(e)?t.get(e):(t||(t=new Map,bm.set(r,t)),_C(r.body,t,[]),t.has(e)?t.get(e):null)}function s0(r){r&&bm.delete(r)}function yC(r,e){let t=bm.get(r);t&&t.delete(e)}function xA(r,e){let t=[r];for(;t.length>0;){let n=t.pop();if(!(!n||typeof n!="object")){if(n.type==="VariableNode"){e.add(n.name);continue}for(let s of["a","b","condition","thenBody","elseBody","expr","value","offsetExpr","extent"])n[s]&&t.push(n[s]);if(n.indices)for(let s of n.indices)t.push(s);if(n.args)for(let s of n.args)t.push(s)}}}function SA(r){let e=new Set((r.writes||[]).map(o=>o.buffer&&o.buffer.name)),t=new Set,n=[r.body,r.initBody];for(;n.length>0;){let o=n.pop();if(o){if(o.type==="BufferStoreNode"&&o.buffer&&e.has(o.buffer.name))for(let i of o.indices)xA(i,t);if(o.body&&n.push(o.body),o.stmts)for(let i of o.stmts)n.push(i);o.thenBody&&n.push(o.thenBody),o.elseBody&&n.push(o.elseBody),o.value&&n.push(o.value)}}let s=new Set;for(let o of r.iterVars||[]){let i=new Set;if(xA(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 _C(r,e,t){let n=[{node:r,loops:[...t]}];for(;n.length>0;){let{node:s,loops:o}=n.pop();if(s)if(s.type==="ForNode")n.push({node:s.body,loops:[...o,s]});else if(s.type==="BlockNode"){if(!e.has(s.name)){let i=SA(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]})}n.push({node:s.body,loops:o})}else if(s.type==="SeqNode")for(let i=s.stmts.length-1;i>=0;i--)n.push({node:s.stmts[i],loops:o});else s.type==="IfThenElseNode"?(s.elseBody&&n.push({node:s.elseBody,loops:o}),n.push({node:s.thenBody,loops:o})):s.type==="AllocateNode"?n.push({node:s.body,loops:o}):s.type==="LetStmtNode"&&n.push({node:s.body,loops:o})}}function kA(r){return r.hasReduction&&r.readBuffers.length===2&&r.writeBuffers.length===1&&r.loopCount>=3}function bC(r,e){let t=dt(r,e);return t?t.loops.some(n=>n.extent&&n.extent.type!=="IntImmNode"):!1}function vA(r){let e=0,t=[r.body];for(;t.length>0;){let n=t.pop();if(n){if(n.type==="BlockNode"&&(e++,e>1))return!0;if(n.type==="ForNode"&&t.push(n.body),n.type==="SeqNode")for(let s of n.stmts)t.push(s)}}return!1}function o0(r){return Sc(r.body,e=>e.type==="BlockNode"&&(e.initBody!==null||SA(e).size>0),{kinds:"stmt"})}function i0(r){return Sc(r.body,e=>e.type==="ForNode"&&e.kind===F.RECURRENCE,{kinds:"stmt"})}function a0(r,e,t){let n=Math.min(t.maxThreadsPerBlock,256),s=e.extent;if(s.type==="IntImmNode"&&s.value>n){let[o,i]=r.split(e,n);r.bindThread(i,"threadIdx.x"),i0(r.func)||r.bindThread(o,"blockIdx.x")}else r.bindThread(e,"threadIdx.x")}function Zu(r,e){return e.reductionLoopVars?e.reductionLoopVars.has(r.loopVar.name):!1}function u0(r,e){return r.body===e}function wC(r){return gu(r,e=>e.type==="BlockNode",{kinds:"stmt"}).map(e=>e.name)}var Bs,bm,Yw,Jw,Zw,Qw,Xw,e0,t0,n0,r0,Rs,ro=M(()=>{"use strict";Ks();De();Gr();Bs=class{static{l(this,"ScheduleRule")}constructor(e){this.name=e}matches(e,t,n){throw new Error("ScheduleRule.matches must be implemented")}apply(e,t,n){throw new Error("ScheduleRule.apply must be implemented")}},bm=new WeakMap;l(dt,"classifyBlock");l(s0,"invalidateClassifyCache");l(yC,"invalidateClassifyBlock");l(xA,"collectVarNames");l(SA,"computeReductionLoopVars");l(_C,"collectBlockInfo");l(kA,"isMatmulShape");l(bC,"blockHasNonConstExtent");l(vA,"hasMultipleBlocks");Yw=class extends Bs{static{l(this,"ElementwiseCPURule")}constructor(){super("elementwise_cpu")}matches(e,t,n){if(n.kind!==Ot.CPU)return!1;let s=dt(e,t);if(!s||s.hasReduction||s.loopCount<1||s.loops.length>0&&vA(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>=n.numCores*n.vectorWidth}apply(e,t,n){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>=n.vectorWidth*2){let[u,p]=e.split(s[0],n.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>=n.vectorWidth&&i.value%n.vectorWidth===0){let[a,u]=e.split(o,n.vectorWidth);e.vectorize(u)}}};l(o0,"primFuncHasReduction");l(i0,"primFuncHasRecurrence");l(a0,"bindFusedSpatialGPU");Jw=class extends Bs{static{l(this,"ElementwiseGPURule")}constructor(){super("elementwise_gpu")}matches(e,t,n){if(!n.isGPU())return!1;let s=dt(e,t);return s?!s.hasReduction&&s.loopCount>=1:!1}apply(e,t,n){let s=e.getLoops(t);if(s.length===0)return;let o=s[0];for(let p=1;p<s.length;p++){let h=e.getLoops(t).find(f=>f.loopVar.name===s[p].loopVar.name);h&&u0(o,h)&&(o=e.fuseLoops(o,h))}let i=o.extent;if(i.type!=="IntImmNode"){e.bindThread(o,"threadIdx.x");return}let a=i.value,u=Math.min(n.maxThreadsPerBlock,1024);if(o0(e.func)&&a<=u){e.bindThread(o,"threadIdx.x");return}a0(e,o,n)}},Zw=class extends Bs{static{l(this,"ReductionCPURule")}constructor(){super("reduction_cpu")}matches(e,t,n){if(n.kind!==Ot.CPU)return!1;let s=dt(e,t);return s?s.hasReduction:!1}apply(e,t,n){let s=e.getLoops(t);if(s.length===0)return;let o=[],i=[],a=dt(e.func,t);for(let u of s)a&&Zu(u,a)?i.push(u):o.push(u);o.length>0&&e.parallelize(o[0])}},Qw=class extends Bs{static{l(this,"ReductionGPURule")}constructor(){super("reduction_gpu")}matches(e,t,n){if(!n.isGPU())return!1;let s=dt(e,t);return s?s.hasReduction:!1}apply(e,t,n){let s=e.getLoops(t);if(s.length===0)return;let o=[],i=dt(e.func,t);for(let u of s)(!i||!Zu(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(h=>h.loopVar.name===o[u].loopVar.name);c&&u0(a,c)&&(a=e.fuseLoops(a,c))}a0(e,a,n)}},Xw=class extends Bs{static{l(this,"MatmulTiledCPURule")}constructor(){super("matmul_tiled_cpu")}matches(e,t,n){if(n.kind!==Ot.CPU)return!1;let s=dt(e,t);if(!s||!kA(s))return!1;let o=n.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,n){let s=e.getLoops(t);if(s.length<3)return;let o=n.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 h=s[c].extent.type==="IntImmNode"?s[c].extent.value:null;h&&h>=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])}},e0=class extends Bs{static{l(this,"MatmulTiledGPURule")}constructor(){super("matmul_tiled_gpu")}matches(e,t,n){if(!n.isGPU())return!1;let s=dt(e,t);if(!s||!kA(s))return!1;let o=n.sharedMemoryBytes||49152,a=Math.max(16,Math.min(128,Math.floor(Math.sqrt(o/8))));return s.loops.reduce((p,c)=>{let h=c.extent&&c.extent.type==="IntImmNode"?c.extent.value:0;return h>p?h:p},0)>=a}apply(e,t,n){let s=e.getLoops(t);if(s.length<3)return;let o=dt(e.func,t),i=s.filter(u=>!o||!Zu(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(h=>h.loopVar.name===i[u].loopVar.name);c&&u0(a,c)&&(a=e.fuseLoops(a,c))}a0(e,a,n)}},t0=class extends Bs{static{l(this,"ElementwiseWasmRule")}constructor(){super("elementwise_wasm")}matches(e,t,n){if(n.kind!==Ot.WASM)return!1;let s=dt(e,t);return!(!s||s.hasReduction||s.loopCount<1||s.loops.length>0&&vA(s.loops[0]))}apply(e,t,n){let s=e.getLoops(t);if(s.length===0)return;let o=n.numCores||1,i=n.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&&n.supportsSimd&&n.supportsSimd()&&c>=i*2&&c%i===0){let[h,f]=e.split(s[0],i);e.parallelize(h),e.vectorize(f);return}if(e.parallelize(s[0]),s.length>1){let h=s[s.length-1],f=h.extent;if(f.type==="IntImmNode"&&f.value>=i&&f.value%i===0){let[,d]=e.split(h,i);e.vectorize(d)}}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)}},n0=class extends Bs{static{l(this,"ReductionWasmRule")}constructor(){super("reduction_wasm")}matches(e,t,n){if(n.kind!==Ot.WASM)return!1;let s=n.numCores>1,o=n.supportsSimd&&n.supportsSimd();if(!s&&!o)return!1;let i=dt(e,t);return i?i.hasReduction&&i.loopCount>=2:!1}apply(e,t,n){let s=e.getLoops(t);if(s.length<2)return;let o=dt(e.func,t),i=[],a=[];for(let u of s)!o||!Zu(u,o)?i.push(u):a.push(u);if(i.length>0){let u=i[0].extent;(u&&u.type==="IntImmNode"?u.value:0)>=(n.numCores||1)*4&&e.parallelize(i[0])}if(n.supportsSimd&&n.supportsSimd()&&a.length>0){let u=a[a.length-1];(u.extent&&u.extent.type==="IntImmNode"?u.extent.value:0)>=n.vectorWidth*2&&e.vectorize(u)}}},r0=class extends Bs{static{l(this,"FallbackRule")}constructor(){super("fallback")}matches(){return!0}apply(e,t,n){let s=e.getLoops(t);s.length!==0&&n.isCPU()&&s.length>=1&&e.parallelize(s[0])}};l(Zu,"isReductionLoop");l(u0,"findDirectChild");Rs=class r{static{l(this,"SchedulePolicy")}constructor(e,t=null,n=null){this.target=e,this.rules=t||r.defaultRules(),this.trace=n}static defaultRules(){return[new Xw,new e0,new Zw,new Qw,new n0,new Yw,new Jw,new t0,new r0]}selectRule(e,t){for(let n of this.rules)if(n.matches(e,t,this.target))return n;return null}applyToBlock(e,t){if(this.target.isGPU()&&bC(e.func,t))return this._explain(t,"none","block has dynamic loop extents; runs sequentially (no dynamic grid)"),null;let n=this.selectRule(e.func,t);return n?(n.apply(e,t,this.target),yC(e.func,t),this._explain(t,n.name,`matched rule '${n.name}' for ${this.target.name}`),n.name):(this._explain(t,"none","no schedule rule matched; runs sequentially"),null)}_explain(e,t,n){this.trace&&this.trace.explainsEnabled&&this.trace.explain("schedule",e,t,n,{target:this.target.name})}applyToAllBlocks(e){s0(e.func);let t=wC(e.func.body),n=new Set,s=new Map;for(let o of t){if(n.has(o))continue;n.add(o);let i=this.applyToBlock(e,o);i&&s.set(o,i)}return s}};l(wC,"collectAllBlockNames")});function wm(r,e,t,n){if(TC[e])return xC(r,e,t,n);let s=kC[e];return s?s(r,t,n||{}):typeof r[e]=="function"?SC(r,e,t,n):r._inferAndBuild(e,t,n||null)}function xC(r,e,t,n){let s=t[0].type.rank,o=n?.dim,i=o!=null?(Array.isArray(o)?o:[o]).map(h=>h<0?s+h:h):Array.from({length:s},(h,f)=>f),a=r.scalarConstant(FS(e,t[0].type.dtype),t[0].type.dtype),u=r.reduce(t[0],a.getResult(0),i,e);if(!n?.keepdim)return u;let p=new Set(i),c=t[0].type.shape.map((h,f)=>p.has(f)?1:h);return r.reshape(u.getResult(0),c)}function SC(r,e,t,n){return t.length===1?Le(r,e,t[0]):t.length===2?Le(r,e,t[0],t[1]):t.length===3?Le(r,e,t[0],t[1],t[2]):r._inferAndBuild(e,t,n||null)}function Le(r,e,...t){return r[e].call(r,...t)}function Io(r,e){return r<0?e+r:r}function vC(r,e){let t=1,n=-1,s=1;for(let i of e)s*=i;let o=[...r];for(let i=0;i<o.length;i++)o[i]===-1?n=i:t*=o[i];return n>=0&&(o[n]=t===0?0:s/t),o}function NA(r,e){return r<0?e+1+r:r}function NC(r,e,t,n,s){let o=r.length,i=Io(e,o),a=r[i],u=t??0,p=n??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 h=new Array(o).fill(0),f=[...r],d=new Array(o).fill(1);return h[i]=u,f[i]=p,d[i]=c,{starts:h,limits:f,strides:d}}var TC,kC,l0=M(()=>{"use strict";es();Oe();Vn();TC=Object.freeze({sum:"sum",mean:"mean",max:"max",min:"min",prod:"prod"});l(wm,"buildMappedOp");l(xC,"buildReduce");l(SC,"callBuilder");l(Le,"callMethod");kC=Object.freeze({matmul:l((r,e)=>Le(r,"matmul",e[0],e[1]),"matmul"),dot:l((r,e)=>Le(r,"dot",e[0],e[1],[e[0].type.rank-1],[0]),"dot"),clone:l((r,e)=>r._inferAndBuild("add",[e[0],r.scalarConstant(0,e[0].type.dtype).getResult(0)],null),"clone"),relu:l((r,e)=>Le(r,"relu",e[0]),"relu"),sigmoid:l((r,e)=>Le(r,"sigmoid",e[0]),"sigmoid"),gelu:l((r,e)=>Le(r,"gelu",e[0]),"gelu"),silu:l((r,e)=>Le(r,"silu",e[0]),"silu"),softmax:l((r,e,t)=>Le(r,"softmax",e[0],t?.dim??-1),"softmax"),log_softmax:l((r,e,t)=>Le(r,"logSoftmax",e[0],t?.dim??-1),"log_softmax"),layer_norm:l((r,e,t)=>Le(r,"layernorm",e[0],e[1],e[2],t?.axis??-1,t?.eps??1e-5),"layer_norm"),batch_norm:l((r,e,t)=>Le(r,"batchnorm",e[0],e[1],e[2],e[3],e[4],t?.axis??1,t?.eps??1e-5),"batch_norm"),conv2d:l((r,e,t)=>Le(r,"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((r,e,t)=>Le(r,"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((r,e)=>Le(r,"embedding",e[0],e[1]),"embedding"),argmax:l((r,e,t)=>Le(r,"argmax",e[0],t?.dim??0,t?.keepdim??!1),"argmax"),argmin:l((r,e,t)=>Le(r,"argmin",e[0],t?.dim??0,t?.keepdim??!1),"argmin"),maximum:l((r,e)=>Le(r,"maximum",e[0],e[1]),"maximum"),minimum:l((r,e)=>Le(r,"minimum",e[0],e[1]),"minimum"),eq:l((r,e)=>Le(r,"compare",e[0],e[1],"eq"),"eq"),ne:l((r,e)=>Le(r,"compare",e[0],e[1],"ne"),"ne"),lt:l((r,e)=>Le(r,"compare",e[0],e[1],"lt"),"lt"),le:l((r,e)=>Le(r,"compare",e[0],e[1],"le"),"le"),gt:l((r,e)=>Le(r,"compare",e[0],e[1],"gt"),"gt"),ge:l((r,e)=>Le(r,"compare",e[0],e[1],"ge"),"ge"),clamp:l((r,e)=>Le(r,"clamp",e[1],e[0],e[2]),"clamp"),pad:l((r,e,t)=>Le(r,"pad",e[0],e[1],t.low,t.high),"pad"),one_hot:l((r,e,t)=>Le(r,"oneHot",e[0],t.depth,{dtype:"f32"}),"one_hot"),index_select:l((r,e,t)=>Le(r,"gather",e[0],e[1],Rk(e[0].type,t?.dim??0,e[1].type.rank)),"index_select"),gather:l((r,e,t)=>Le(r,"gatherDim",e[0],e[1],t?.dim??0),"gather"),scatter_add:l((r,e,t)=>Le(r,"scatterAddDim",e[0],e[1],e[2],t?.dim??0),"scatter_add"),cat:l((r,e,t)=>Le(r,"concat",e,Io(t?.dim??0,e[0].type.rank)),"cat"),stack:l((r,e,t)=>{let n=NA(t?.dim??0,e[0].type.rank),s=e.map(o=>{let i=[...o.type.shape];return i.splice(n,0,1),r.reshape(o,i).getResult(0)});return Le(r,"concat",s,n)},"stack"),reshape:l((r,e,t)=>r.reshape(e[0],vC(t.shape??t.new_shape,e[0].type.shape)),"reshape"),transpose:l((r,e,t)=>{let n=e[0].type.rank,s=Io(t?.dim0??0,n),o=Io(t?.dim1??1,n),i=Array.from({length:n},(a,u)=>u);return i[s]=o,i[o]=s,r.transpose(e[0],i)},"transpose"),permute:l((r,e,t)=>Le(r,"transpose",e[0],t.dims),"permute"),broadcast_in_dim:l((r,e,t)=>Le(r,"broadcast",e[0],t.result_shape,t.broadcast_dimensions),"broadcast_in_dim"),expand:l((r,e,t)=>{let n=e[0].type.rank,s=t.shape,o=s.length-n,i=s.map((u,p)=>u===-1?e[0].type.shape[p-o]:u),a=Array.from({length:n},(u,p)=>p+o);return Le(r,"broadcast",e[0],i,a)},"expand"),slice:l((r,e,t)=>{let n=NC(e[0].type.shape,t.dim,t.start,t.end,t.step);return Le(r,"slice",e[0],n.starts,n.limits,n.strides)},"slice"),unsqueeze:l((r,e,t)=>{let n=[...e[0].type.shape];return n.splice(NA(t.dim,n.length),0,1),r.reshape(e[0],n)},"unsqueeze"),squeeze:l((r,e,t)=>{let n=[...e[0].type.shape];if(t.dim===void 0||t.dim===null)return r.reshape(e[0],n.filter(o=>o!==1));let s=Io(t.dim,n.length);return n[s]===1&&n.splice(s,1),r.reshape(e[0],n)},"squeeze"),narrow:l((r,e,t)=>{let n=Io(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[n]=t.start,o[n]=t.start+t.length,Le(r,"slice",e[0],s,o,i)},"narrow"),select:l((r,e,t)=>{let n=e[0].type.rank,s=Io(t.dim,n),o=t.index,i=o<0?e[0].type.shape[s]+o:o,a=new Array(n).fill(0),u=[...e[0].type.shape];a[s]=i,u[s]=i+1;let p=Le(r,"slice",e[0],a,u,new Array(n).fill(1)).getResult(0);return r.reshape(p,e[0].type.shape.filter((c,h)=>h!==s))},"select"),split:l((r,e,t)=>Le(r,"split",e[0],Io(t.dim??0,e[0].type.rank),t.sizes),"split"),chunk:l((r,e,t)=>{let n=Io(t.dim??0,e[0].type.rank),s=e[0].type.shape[n],o=Math.ceil(s/t.chunks),i=[];for(let a=0;a<s;a+=o)i.push(Math.min(o,s-a));return Le(r,"split",e[0],n,i)},"chunk")});l(Io,"normalizeDim");l(vC,"resolveShape");l(NA,"normalizeInsertDim");l(NC,"sliceAttrs")});function OC(r,e,t,n){let s=r;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+="|"+n.name,s}function BC(r){let e=OA.get(r);return e||(e=new zu("jit_"+r),OA.set(r,e)),e}function RC(r){let e=1;for(let t of r.shape)e*=typeof t=="number"&&t>0?t:1;return Math.max(e,1)}function MC(r,e,t){let n=[];for(let[,s]of t.bufferMap)n.push(new(Dt(s.dtype))(RC(s)));r.run(e.name,...n)}function IC(r,e,t,n){try{let s=BA(r,e);if(s.shapeParams&&s.shapeParams.length>0)return null;new Rs(e).applyToAllBlocks(new Sn(s));let o=t.compile(s);return n.addCompiledKernel(o),e.isWebGPU()||MC(n,o,s),o}catch{return null}}function LC(r,e,t){let n=e.map(a=>new I(a.shape,a.dtype)),s=r+"_jit_"+EC++,o=hc(s,n,[],(a,u)=>{let p;p=wm(a,r,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 RA(r,e,t=null,n){let s=OC(r,e,t,n),o=AA.get(s);if(o)return o;let i=LC(r,e,t),a=new wn(r+"_jit_mod");a.addFunction(i);let u=new oa;u.addPass(new Gu),u.addPass(new ia),u.addPass(new no),u.run(a);let p=BC(n.name),c=new AC(n),f=typeof n.isGPU=="function"&&n.isGPU()?IC(i,n,c,p):null;f||(f=c.compile(BA(i,n)),p.addCompiledKernel(f));let d=i.getReturnOp(),m=d&&d.operands.length>0?d.operands[0].type.dtype:null;return o={funcName:f.name,runtime:p,numInputs:e.length,outDtype:m,compiled:f},AA.set(s,o),o}var AA,OA,BA,AC,EC,MA=M(()=>{"use strict";es();Yo();Oe();Ui();Ec();nm();um();Lw();Dw();$w();fa();ro();ts();l0();AA=new Map,OA=new Map,BA=Xh,AC=Hs;l(OC,"_cacheKey");l(BC,"_getRuntime");l(RC,"_bufferNumel");l(MC,"_trialLaunch");l(IC,"_compileScheduledGPU");EC=0;l(LC,"_buildGraphFunc");l(RA,"jitCompile")});function Zc(){return CC}function EA(r){DC=r}function c0(){IA&&IA()}var CC,IA,DC,Tm=M(()=>{"use strict";CC=!1,IA=null,DC=!1;l(Zc,"isEagerDeferred");l(EA,"setCudaGraphArmed");l(c0,"eagerFlush")});function xm(r){return r.slice(0,r.indexOf("(")).split(".")[0].trim()}function LA(){return Sm.map(r=>r.schema)}function km(r){return $C.get(r)||null}function FC(r){let e=r.indexOf("(")+1,t=r.lastIndexOf(")"),n=r.slice(e,t).trim();if(!n)return[];let s=[];for(let o of CA(n)){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 CA(r){let e=[],t=0,n=0;for(let s=0;s<r.length;s++){let o=r[s];o==="("||o==="["?t++:o===")"||o==="]"?t--:o===","&&t===0&&(e.push(r.slice(n,s)),n=s+1)}return e.push(r.slice(n)),e}function zC(r){let t=r.lastIndexOf("->"),n=t>=0?r.slice(t+2).trim():"Tensor";return n==="Tensor"?{kind:"tensor",arity:1}:n==="Tensor[]"?{kind:"tensor_list"}:n.startsWith("(")&&n.endsWith(")")?{kind:"tuple",arity:CA(n.slice(1,-1)).length}:{kind:"value",arity:1}}var Sm,PC,$C,h5,vm=M(()=>{"use strict";Sm=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"}]),PC=new Map(Sm.filter(r=>r.scalarArgs).map(r=>[xm(r.schema),r.scalarArgs])),$C=new Map(Sm.map(r=>[xm(r.schema),PC.get(xm(r.schema))||FC(r.schema)])),h5=new Map(Sm.map(r=>{let e=xm(r.schema);return[e,Object.freeze({...r,name:e,returns:r.returns||zC(r.schema)})]}));l(xm,"opNameFromSchema");l(LA,"tensorOpSchemas");l(km,"scalarArgNames");l(FC,"scalarArgNamesFromSchema");l(CA,"splitTopLevel");l(zC,"returnsFromSchema")});var nA={};nt(nA,{getCudnnGRU:()=>d0,getCudnnLSTM:()=>h0,getGpuAdamFn:()=>m0,getGpuContiguousFn:()=>QC,getGpuMatmul:()=>dw,getWebgpuRNN:()=>Qc,gpuContiguousArray:()=>oD,registerJITKernels:()=>y0,setCudnnGRU:()=>tD,setCudnnLSTM:()=>eD,setGpuAdamFn:()=>rD,setGpuConcatFn:()=>XC,setGpuContiguousFn:()=>ZC,setGpuMatmul:()=>sD,setWebGPUEagerFn:()=>KC,setWebgpuRNN:()=>nD,tensorToContiguous:()=>Je,wrapResult:()=>tn});function KC(r){PA=r}function p0(r){return typeof r=="object"&&r!==null&&"_impl"in r}function WC(r,e){let t=[],n={},s=km(r),o=0;for(let i=0;i<e.length;i++){let a=e[i];if(p0(a))t.push(a);else if(Array.isArray(a)&&a.length>0&&p0(a[0]))for(let u of a)p0(u)&&t.push(u);else a!=null&&(s&&o<s.length&&(n[s[o]]=a),o++)}return{tensors:t,scalars:n}}function ha(r,e){return typeof r=="number"?r:e}function Qu(r,e){return Array.isArray(r)?r:e}function DA(r){return Array.isArray(r)?r:[[0,0],[0,0]]}function $A(r,e,t){if(e.length===0)return[];if(new Set(["sum","mean","max","min","prod","argmax","argmin"]).has(r)){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(r==="matmul")return Bd(e[0].shape,e[1].shape)||[];if(r==="dot")return[];if(r==="transpose"){let o=[...e[0].shape],i=ha(t.dim0,0),a=ha(t.dim1,1),u=o[i];return o[i]=o[a],o[a]=u,o}if(r==="conv2d"){let o=e[0].shape,i=e[1].shape,a=Qu(t.strides,[1,1]),u=DA(t.padding),p=Qu(t.dilation,[1,1]),c=[];for(let h=0;h<2;h++){let f=u[h][0]+u[h][1],d=(i[h+2]-1)*p[h]+1;c.push(Math.floor((o[h+2]+f-d)/a[h])+1)}return[o[0],i[0],...c]}if(r==="pool2d"){let o=e[0].shape,i=Qu(t.kernel_size,[2,2]),a=Qu(t.strides,i),u=DA(t.padding),p=[];for(let c=0;c<2;c++){let h=u[c][0]+u[c][1];p.push(Math.floor((o[c+2]+h-i[c])/a[c])+1)}return[o[0],o[1],...p]}if(r==="clamp"){let o=[...e[0].shape];for(let i=1;i<e.length;i++)o=Ou(o,e[i].shape)||o;return o}if(r==="pad"){let o=e[0].shape,i=Qu(t.low,[]),a=Qu(t.high,[]);return o.map((u,p)=>u+(i[p]||0)+(a[p]||0))}if(r==="one_hot")return[...e[0].shape,ha(t.depth,0)];if(r==="cat"){let o=e[0].shape.length,i=ha(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(r==="stack"){let o=e[0].shape.length,i=ha(t.dim,0),a=i<0?o+1+i:i,u=[...e[0].shape];return u.splice(a,0,e.length),u}if(r==="index_select"){let o=[...e[0].shape],i=o.length,a=ha(t.dim,0),u=a<0?i+a:a;return o[u]=e[1].shape.reduce((p,c)=>p*c,1),o}if(r==="gather")return[...e[1].shape];if(r==="scatter_add")return[...e[0].shape];if(r==="softmax"||r==="log_softmax")return[...e[0].shape];if(r==="layer_norm"||r==="batch_norm")return[...e[0].shape];if(r==="embedding"){let o=e[0].shape;return[...e[1].shape,o[1]]}return e.length===1?[...e[0].shape]:Ou(e[0].shape,e[1].shape)||[...e[0].shape]}function Je(r){let e=r._impl.storage.data,t=r._impl.storageOffset,n=r.numel;if(r.isContiguous&&t===0&&e.length===n)return r.data||e;let s=r.shape,o=r.strides,i=e.constructor,a=new i(n),u=a,p=s.length,c=new Int32Array(p),h=t;for(let f=0;f<n;f++){u[f]=e[h];for(let d=p-1;d>=0;d--){if(c[d]++,c[d]<s[d]){h+=o[d];break}h-=(s[d]-1)*o[d],c[d]=0}}return a}function tn(r,e,t,n){let s=xn(e),o=En.fromData(r,n),i=new Bt(o,0,e,s,t,n);return new We(i)}function Nm(r,e,t){let n=1;for(let s=e;s<t;s++)n*=r[s];return n}function HC(r,e,t,n){let s=e.length,o=t<0?s+1+t:t,i=Nm(e,0,o),a=Nm(e,o,s),u=r.length;for(let p=0;p<i;p++)for(let c=0;c<u;c++)n.set(r[c].subarray(p*a,(p+1)*a),(p*u+c)*a)}function YC(r,e,t,n){let s=e[0].length,o=t<0?s+t:t,i=Nm(e[0],0,o),a=Nm(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 h=0;h<r.length;h++){let f=e[h][o]*a;n.set(r[h].subarray(p*f,(p+1)*f),c),c+=f}}}function JC(r,e,t){let n=$A(r,e,t),s=e[0].dtype,o=Dt(s),i=new o(Math.max(cn(n),1)),a=ha(t.dim,0);if(Zc()&&f0){let p=e.map(c=>g0(c));return f0(r,p,e.map(c=>[...c.shape]),a,n,i,s),tn(i,n,s,e[0].device)}let u=e.map(p=>Je(p));return r==="stack"?HC(u,e[0].shape,a,i):YC(u,e.map(p=>p.shape),a,i),tn(i,n,s,e[0].device)}function ZC(r){Am=r}function QC(){return Am}function XC(r){f0=r}function eD(r){FA=r}function h0(){return FA}function tD(r){zA=r}function d0(){return zA}function nD(r){VA=r}function Qc(){return VA}function rD(r){GA=r}function m0(){return GA}function sD(r){jA=r}function dw(){return jA}function g0(r){if(r.isContiguous&&r._impl.storageOffset===0){let e=r._impl.storage.rawData;if(e&&e.length===r.numel)return e}return Zc()&&Am?Am(r._impl.storage.rawData,r.shape,r.strides,r._impl.storageOffset,r.dtype):Je(r)}function oD(r){return g0(r)}function iD(r,e){let t=UC[e];if(!t)return null;let n=e===1,s=e===5,o=n&&(r==="stack"||r==="cat");return(i,...a)=>{let{tensors:u,scalars:p}=WC(r,a);if(u.length===0)throw new Error(`JIT dispatch: no tensor args for op '${r}'`);if(o)return JC(r,u,p);let c=t(),h=RA(r,u,p,c),f=$A(r,u,p),d=h.outDtype||Rd(u[0].dtype,u.length>1?u[1].dtype:u[0].dtype),m=cn(f),g=Dt(d),y=new g(Math.max(m,1));if(s)return PA(h.compiled,u,y),tn(y,f,d,u[0].device);let _=u.map(b=>n?g0(b):Je(b));return _.push(y),h.runtime.run(h.funcName,..._),tn(y,f,d,u[0].device)}}function y0(){let r=Nt.listOps(),e=[0,1,2,5];for(let t of r){let n=Nt.findOp(t);if(!n)continue;let s=n.name;for(let o of e){if(n.entry.hasKernel(o))continue;let i=iD(s,o);i&&n.entry.registerKernel(o,wr.fromUnboxed(i))}}}var VC,GC,jC,qC,UC,PA,Am,f0,FA,zA,VA,GA,jA,Kn=M(()=>{"use strict";Nn();nc();$i();MA();Tm();Ks();jn();Oo();ea();jr();ts();vm();UC={0:()=>VC??=Us(),1:()=>GC??=ed(),2:()=>jC??=td(),5:()=>qC??=nd()},PA=null;l(KC,"setWebGPUEagerFn");l(p0,"hasTensorImpl");l(WC,"_extractTensorsAndScalars");l(ha,"numberScalar");l(Qu,"numberArrayScalar");l(DA,"paddingScalar");l($A,"_inferOutputShape");l(Je,"tensorToContiguous");l(tn,"wrapResult");l(Nm,"_prod");l(HC,"_hostStack");l(YC,"_hostCat");l(JC,"_runHostConcatLike");Am=null;l(ZC,"setGpuContiguousFn");l(QC,"getGpuContiguousFn");f0=null;l(XC,"setGpuConcatFn");FA=null;l(eD,"setCudnnLSTM");l(h0,"getCudnnLSTM");zA=null;l(tD,"setCudnnGRU");l(d0,"getCudnnGRU");VA=null;l(nD,"setWebgpuRNN");l(Qc,"getWebgpuRNN");GA=null;l(rD,"setGpuAdamFn");l(m0,"getGpuAdamFn");jA=null;l(sD,"setGpuMatmul");l(dw,"getGpuMatmul");l(g0,"_gpuInputArray");l(oD,"gpuContiguousArray");l(iD,"_wrapOpForJIT");l(y0,"registerJITKernels")});var o1={};nt(o1,{arange:()=>n1,empty:()=>fn,eye:()=>Rm,full:()=>et,linspace:()=>s1,ones:()=>Ht,randn:()=>Bm,randperm:()=>r1,zeros:()=>$e});function Om(r){return{dtype:r?.dtype??W.F32,device:r?.device??ii(),requiresGrad:r?.requiresGrad??!1}}function Xc(r,e,t,n){let s=xn(r),i=cn(r)*Xi(e),a=En.allocate(i,e,t),u=new Bt(a,0,r,s,e,t),p=new We(u);return n&&p.requiresGrad_(!0),p}function fn(r,e){let{dtype:t,device:n,requiresGrad:s}=Om(e);return Xc(r,t,n,s)}function b0(r,e,t){let n=fn(r,e),s=n.data;if(s)return s.fill(t),n;if(n.device&&n.device.type==="meta"){let o=fn(r,{dtype:e?.dtype});return o.data.fill(t),o}return n}function $e(r,e){return b0(r,e,0)}function Ht(r,e){return b0(r,e,1)}function et(r,e,t){return b0(r,t,e)}function Bm(r,e){let t=fn(r,e),n=t.data;if(n){let s=n.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;n[o]=u*Math.cos(p),o+1<s&&(n[o+1]=u*Math.sin(p))}}return t}function n1(r,e,t,n){let s,o,i;e===void 0&&t===void 0?(s=0,o=r,i=1):t===void 0?(s=r,o=e,i=1):(s=r,o=e,i=t);let a=Math.max(0,Math.ceil((o-s)/i)),{dtype:u,device:p,requiresGrad:c}=Om(n),h=Xc([a],u,p,c),f=h.data;if(f)for(let d=0;d<a;d++)f[d]=s+d*i;return h}function Rm(r,e,t){let n=e??r,{dtype:s,device:o,requiresGrad:i}=Om(t),a=Xc([r,n],s,o,i),u=a.data;if(u){u.fill(0);let p=Math.min(r,n);for(let c=0;c<p;c++)u[c*n+c]=1}return a}function r1(r,e){let t=e?.dtype??W.I32,n=e?.device??ii(),s=Xc([r],t,n,e?.requiresGrad??!1),o=s.data;for(let i=0;i<r;i++)o[i]=i;for(let i=r-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 s1(r,e,t,n){let{dtype:s,device:o,requiresGrad:i}=Om(n),a=Xc([t],s,o,i),u=a.data;if(u&&t>0)if(t===1)u[0]=r;else{let p=(e-r)/(t-1);for(let c=0;c<t;c++)u[c]=r+c*p}return a}var _t=M(()=>{"use strict";jn();Oo();ea();ts();ar();jr();l(Om,"_defaultOpts");l(Xc,"_makeTensor");l(fn,"empty");l(b0,"_filled");l($e,"zeros");l(Ht,"ones");l(et,"full");l(Bm,"randn");l(n1,"arange");l(Rm,"eye");l(r1,"randperm");l(s1,"linspace")});var _O={};nt(_O,{GradMode:()=>rl,enableGrad:()=>yO,noGrad:()=>gO});function gO(r){let e=hi;hi=!1;try{return r()}finally{hi=e}}function yO(r){let e=hi;hi=!0;try{return r()}finally{hi=e}}var hi,rl,sp=M(()=>{"use strict";hi=!0,rl={isEnabled(){return hi},setEnabled(r){hi=r}};l(gO,"noGrad");l(yO,"enableGrad")});var YP,ft,ga=M(()=>{"use strict";jn();Oo();ea();Nn();ar();jr();YP=Pr.fromKey(48),ft=class extends We{static{l(this,"SymbolicTensor")}_irValue;_tracer;_symbolicShape;constructor(e,t,n,s,o){let i=xn(t),a=En.allocate(0,n,Au),u=new Bt(a,0,t,i,n,Au);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(YP)}get isSymbolic(){return!0}}});function JP(r){let e=[...r.entries()];return e.sort((t,n)=>t[0]<n[0]?-1:t[0]>n[0]?1:0),e}function Tg(r){return r==null?"null":typeof r=="number"?Number.isFinite(r)?String(r):r===1/0?"inf":r===-1/0?"-inf":"nan":typeof r=="boolean"?String(r):typeof r=="string"?`"${r}"`:r instanceof I?Rn(r):Array.isArray(r)?r.length===0?"[]":Array.isArray(r[0])?"["+r.map(e=>Tg(e)).join(", ")+"]":"["+r.map(e=>Tg(e)).join(", ")+"]":typeof r=="object"&&r.constructor===Object?"{"+Object.entries(r).sort((t,n)=>t[0].localeCompare(n[0])).map(([t,n])=>`${t}: ${Tg(n)}`).join(", ")+"}":String(r)}function ZP(r){return new cp().printModule(r)}var cp,P0=M(()=>{"use strict";Oe();cp=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 n=!0;for(let s of e)n||t.push(""),n=!1,this.printFunction(s,t);return this.indent--,t.push("}"),t.join(`
+`)}printFunction(e,t=null){let n=t===null;n&&(t=[]),this.valueNames.clear(),this._nextValueId=0;let s=e.args.map((i,a)=>`${this._nameValue(i)}: ${Rn(i.type)}`),o=e.outputTypes.map(i=>Rn(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()}}`),n?t.join(`
+`):void 0}printBlock(e,t,n=!1){if(!n&&e.arguments.length>0){let s=e.arguments.map(o=>`${this._nameValue(o)}: ${Rn(o.type)}`);t.push(`${this._indentPrefix()}^bb(${s.join(", ")}):`)}for(let s of e)this.printOperation(s,t)}printOperation(e,t=null){let n=t===null;n&&(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 JP(e.attributes))i.push(`${a} = ${Tg(u)}`);o+=" {"+i.join(", ")+"}"}if(e.numResults>0){let i=[];for(let a=0;a<e.numResults;a++)i.push(Rn(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)}: ${Rn(c.type)}`);t.push(`${this._indentPrefix()}^bb(${p.join(", ")}):`)}for(let p of u)this.printOperation(p,t)}this.indent--,t.push(`${s}}`)}return n?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(JP,"sortedEntries");l(Tg,"formatAttrValue");l(ZP,"printModule")});var g$,Ig,CO=M(()=>{"use strict";Oe();Vs();g$={eq:l((r,e)=>r===e,"eq"),ne:l((r,e)=>r!==e,"ne"),gt:l((r,e)=>r>e,"gt"),ge:l((r,e)=>r>=e,"ge"),lt:l((r,e)=>r<e,"lt"),le:l((r,e)=>r<=e,"le")},Ig=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,n){let s=`s${this._nextId++}`;return this._symbols.set(s,{hint:n,inputIdx:e,dimIdx:t}),s}produceShapeSpec(e,t,n){let s=new Array(t.length),o=new Array(t.length);for(let i=0;i<t.length;i++)if(n&&n.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,n){this._guards.push({lhs:e,op:t,rhs:n})}guardDivisible(e,t){this._guards.push({type:"divisible",sym:e,divisor:t})}bindInputShapes(e){this._bindings.clear();for(let[t,n]of this._symbols)this._bindings.set(t,e[n.inputIdx].shape[n.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 n=this._resolve(t.lhs),s=this._resolve(t.rhs);if(!g$[t.op](n,s))return{passed:!1,failedGuard:t}}return{passed:!0,failedGuard:null}}resolveSymbolicShape(e){let t=new Array(e.length);for(let n=0;n<e.length;n++)t[n]=this._resolve(e[n]);return t}_resolve(e){return typeof e=="number"?e:typeof e=="string"?this._bindings.get(e):e instanceof Ce?Ce.evaluate(e,this._bindings):e}get symbols(){return this._symbols}get guards(){return this._guards}get bindings(){return this._bindings}}});function ss(){return Eg}var Eg,Lg,ya=M(()=>{"use strict";es();Yo();xs();Oe();At();ga();CO();l0();Eg=null;l(ss,"getActiveTracer");Lg=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 Ig,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,n){let s=this._inputTypes.length,{irShape:o,symShape:i}=this._shapeEnv.produceShapeSpec(s,e,n);if(n)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 un(this._name,this._inputTypes,[]),this._func.inputTypes=[...this._func.inputTypes],this._builder=new An(this._func),this._module=new wn(this._name);let e=[],t=this._func.args;for(let n=0;n<t.length;n++){let s=t[n],o=this._inputTypes[n];s.symbolicShape=this._inputSymShapes[n];let i=new ft(s,o.shape,o.dtype,this,this._inputSymShapes[n]);e.push(i)}return this._inputs=e,e}recordOp(e,t,n){let s=[];for(let u of t)u instanceof ft&&s.push(u.irValue);let o=this._requireBuilder(),i=wm(o,e,s,n),a=[];for(let u=0;u<i.numResults;u++){let p=i.getResult(u),c=p.type,h=this._propagateSymbolicShape(e,i,t,c,u);p.symbolicShape=h,a.push(new ft(p,c.shape,c.dtype,this,h))}return a.length===1?a[0]:a}_propagateSymbolicShape(e,t,n,s,o=0){let i=n.filter(c=>c instanceof ft),a=ne.get(t.opName||e);if(a&&a.propagateSymbolicShapes){let c=new Map;for(let f of i)c.set(f.irValue,f.symbolicShape);let h=a.propagateSymbolicShapes(t,c);if(h&&h[o])return h[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 h=null;for(let f of i){let d=f.symbolicShape;if(!d)continue;let m=u.length-d.length,g=c-m;if(g>=0&&g<d.length&&typeof d[g]=="string"){h=d[g];break}}p[c]=h!==null?h:X}return p}scan(e,t,n){let s=l(f=>f instanceof ft?f.irValue:this.captureConstant(f).irValue,"toIr"),o=e.map(s),i=t.map(s),u=this._requireBuilder().scanOp(o,i,(f,d,m)=>{let g=this._requireBuilder();this._builder=f;try{let y=l(w=>new ft(w,w.type.shape,w.type.dtype,this,[...w.type.shape]),"wrap"),[_,b]=n(m.map(y),d.map(y));return[_.map(w=>w.irValue),b.map(w=>w.irValue)]}finally{this._builder=g}}),p=t.length,c=[],h=[];for(let f=0;f<u.numResults;f++){let d=u.getResult(f),m=new ft(d,d.type.shape,d.type.dtype,this,[...d.type.shape]);f<p?c.push(m):h.push(m)}return[c,h]}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),h=new ft(c,[],e.dtype,this,[]);return this._capturedParams.set(e,h),h}let n=new I(e.shape,e.dtype),s=this._requireFunc();s.inputTypes.push(n);let i=s.entryBlock.addArgument(n),a=new ft(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 ft&&(this._requireBuilder().returnOp([e.irValue]),this._outputSymShapes=[e.symbolicShape]),this._outputTypes=[new I(e.shape,e.dtype)]}markOutputs(e){let t=e.map(n=>n.irValue);this._requireBuilder().returnOp(t),this._outputTypes=e.map(n=>new I(n.shape,n.dtype)),this._outputSymShapes=e.map(n=>n.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(){Eg=this}deactivate(){Eg===this&&(Eg=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 YO(){return[...HO]}function Gg(r,e,t,n=HO){return n.filter(s=>s.phase===r).sort((s,o)=>s.priority-o.priority).map(s=>s.factory(e,t)).filter(Boolean)}var HO,uT=M(()=>{"use strict";HO=[];l(YO,"snapshotGraphPasses");l(Gg,"graphPassesForPhase")});function JO(r){return r instanceof Map?r:r&&typeof r=="object"?new Map(Object.entries(r)):new Map}var jg,ZO=M(()=>{"use strict";uT();l(JO,"toMap");jg=class{static{l(this,"CompilerContext")}constructor({loweringRules:e=null,codegenEntries:t=null,graphPasses:n=null}={}){this.loweringRules=JO(e),this.codegenEntries=JO(t),this.graphPasses=n||YO()}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,n){return Gg(e,t,n,this.graphPasses)}}});function qg(r){return new lT().print(r)}var lT,cT=M(()=>{"use strict";Vn();lT=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,n]of e.bufferMap)this.push(`${n.name} = buffer_map(${t.name}, shape=[${n.shape.join(",")}], dtype=${n.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} `,n=e.threadTag?`[${e.threadTag}] `:"";this.push(`for ${e.loopVar.name} in 0..`),this.visit(e.extent),this.push(` ${t}${n}{`),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(` ${Ka(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(qg,"printTensorIR")});var Ug,QO=M(()=>{"use strict";Kt();cT();Ug=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 n of this.passes)this._runPass(n,e,t);return e}_runPass(e,t,n){let s=n.trace;e.trace=s,s.phaseStart(e.phase);let o=performance.now();e.begin(n);for(let i=0;i<t.length;i++){let a=t[i];if(!n.failed.has(a.name))try{let u=e.run(a,n);u&&u!==a&&(t[i]=u)}catch(u){if(n.errors.push(new Dn(e.phase,a.name,u.message)),n.failed.add(a.name),s.errorEvent(e.phase,a.name,u.message),!n.resilient)break}}if(e.end(n),s.phaseEnd(e.phase,performance.now()-o),e.snapshotPoint&&s.shouldSnapshot(e.snapshotPoint))for(let i of t)n.failed.has(i.name)||s.irDump(e.snapshotPoint+":"+i.name,qg(i));this.verifyHook&&this._verifyFuncs(t,n),e.trace=null}_verifyFuncs(e,t){for(let n of e){if(t.failed.has(n.name))continue;let s=this.verifyHook(n);if(s&&s.length>0){let o=s.join("; ");if(t.resilient)t.errors.push(new Dn("verification",n.name,o)),t.failed.add(n.name);else throw new Error("TensorIR verification failed for "+n.name+": "+o)}}}}});function XO(r){let e=new Ua;return e.add(new Ja),e.add(new Za),e.add(new Qa(r)),e.add(new Xa),e.add(new eu(r)),e.add(new tu),e.add(new nu),e.add(new wh),e.add(new Ya),e.add(new Nh),e.add(new Ah),e.add(new Oh),e.add(new Bh),r&&(e.add(new vh(r)),e.add(new Sh(r)),e.add(new kh(r))),e}var L$,C$,Kg,eB=M(()=>{"use strict";nn();uc();Cw();ru();l(XO,"buildAlgebraicPatterns");L$=XO(!1),C$=XO(!0),Kg=class extends Ge{static{l(this,"AlgebraicSimplificationPass")}constructor(e={}){super("algebraic_simplify"),this.preservedAnalyses=new Set,this.patterns=e.fastMath?C$:L$}run(e,t){return new ju(this.patterns).applyPatterns(e,10,this.trace)}}});function tB(r){let e=ne.get(r);return!!(e&&e.isConstant)}function D$(r,e){return!gb(e)||typeof r!="number"?!0:Number.isInteger(r)&&Number.isSafeInteger(r)}function nB(r,e,t){if(t.has(r))return t.get(r);let n=P$(r,e,t);return t.set(r,n),n}function P$(r,e,t){let n=r.definingOp;if(!n)return;if(tB(n.opName))return n.getAttr("value");if(e.has(n))return;e.add(n);let s=ne.get(n.opName);if(!s||!s.fold||s.hasSideEffects||n.regions.length>0)return;let o=new Array(n.numOperands),i=new Array(n.numOperands);for(let a=0;a<n.numOperands;a++){let u=nB(n.getOperand(a),e,t);if(u===void 0)return;o[a]=u,i[a]=n.getOperand(a).definingOp}try{return s.fold(o,n.attributes,i)}catch{return}}var Wg,rB=M(()=>{"use strict";nn();es();At();Kt();Oe();l(tB,"isConstantProducer");l(D$,"isFoldResultRepresentable");l(nB,"resolveConstantValue");l(P$,"computeConstantValue");Wg=class extends Ge{static{l(this,"ConstantFoldPass")}constructor(){super("constant_fold")}run(e,t){let n=!1,s=0,o=new An(e),i=new Map;for(let a of[...e.opsRecursive()]){if(!a.parentBlock||tB(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),h=!0;for(let f=0;f<a.numOperands;f++){let d=nB(a.getOperand(f),new Set,i);if(d===void 0){h=!1;break}p[f]=d,c[f]=a.getOperand(f).definingOp}if(h)try{let f=u.fold(p,a.attributes,c);if(f===void 0||!D$(f,a.getResult(0).type.dtype))continue;o.block=a.parentBlock,o.setInsertionPoint(a);let d=o.constant(f,a.getResult(0).type);a.replaceAllResultsWith([d.getResult(0)]),a.erase(),n=!0,s++}catch(f){this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,foldError:a.opName,message:f.message,level:ue.DEBUG})}}return this.trace&&this.trace.level>=ue.DEBUG&&s>0&&this.trace.emit({type:"pass_detail",passName:this.name,foldedCount:s,level:ue.DEBUG}),n?oe.CHANGED:oe.UNCHANGED}}});var Hg,sB=M(()=>{"use strict";nn();At();Kt();Hg=class extends Ge{static{l(this,"CSEPass")}constructor(){super("cse"),this.preservedAnalyses=new Set}run(e,t){let n=!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 h=a.get(c),f=!1;for(let d of h)if(d.parentBlock&&d.structuralEquals(u)){let m=[];for(let g=0;g<d.numResults;g++)m.push(d.getResult(g));u.replaceAllResultsWith(m),u.erase(),n=!0,s++,f=!0;break}f||h.push(u)}}return this.trace&&this.trace.level>=ue.DEBUG&&s>0&&this.trace.emit({type:"pass_detail",passName:this.name,eliminated:s,level:ue.DEBUG}),n?oe.CHANGED:oe.UNCHANGED}}});function fT(r,e){let t=pT[r],n=pT[e];return t===void 0||n===void 0?!1:r===je.REDUCTION?e===je.ELEMENTWISE:e===je.REDUCTION?t<=pT[je.INJECTIVE]:!0}function bl(r){let e=!1,t=!1,n=!1,s=!1;for(let o of r){let i=ne.get(o.opName);if(!i||i.isOpaque){n=!0;continue}i.isReduction?e=!0:i.isInjective?t=!0:i.isBroadcast?s=!0:i.isElementwise||(n=!0)}return n?je.OPAQUE:e?je.REDUCTION:t?je.INJECTIVE:s?je.BROADCAST:je.ELEMENTWISE}function Aa(r){let e=ne.get(r.opName);return!e||e.isOpaque?je.OPAQUE:e.isReduction?je.REDUCTION:e.isInjective?je.INJECTIVE:e.isBroadcast?je.BROADCAST:e.isElementwise?je.ELEMENTWISE:je.OPAQUE}function oB(r,e){if(r.length!==e.length)return!1;for(let t=0;t<r.length;t++)if(!(r[t]===X||e[t]===X)&&r[t]!==e[t])return!1;return!0}var je,pT,yl,_l,Oa=M(()=>{"use strict";Oe();At();Ui();Tr();je=Object.freeze({ELEMENTWISE:"kElementwise",BROADCAST:"kBroadcast",REDUCTION:"kReduction",INJECTIVE:"kInjective",OPAQUE:"kOpaque",HORIZONTAL:"kHorizontal"}),pT={[je.ELEMENTWISE]:0,[je.BROADCAST]:1,[je.INJECTIVE]:2,[je.REDUCTION]:3};l(fT,"canFusePatterns");l(bl,"classifyFusionKind");l(Aa,"classifyOpPattern");yl=Object.freeze({legal:!0}),_l=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=qt(e)||ko(e)||Fb(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 n=ne.get(e.opName),s=ne.get(t.opName);if(!n)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(n.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=Aa(e),i=Aa(t);return fT(o,i)?(o===je.REDUCTION||i===je.REDUCTION)&&!this.allowReductionFusion?{legal:!1,reason:"reduction fusion disabled by target"}:o===je.ELEMENTWISE&&i===je.ELEMENTWISE?this._checkElementwisePair(e,t):(o===je.BROADCAST||o===je.REDUCTION)&&i===je.ELEMENTWISE?yl: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 n=0;for(let s of e.ops){let o=ne.get(s.opName);if(o&&o.isReduction&&n++,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&&n++,o&&o.isOpaque)return{legal:!1,reason:"opaque op in merge (use EpilogueFusionPass)"}}return n>1?{legal:!1,reason:"merged group would contain multiple reductions"}:yl}_checkElementwisePair(e,t){let n=this._getOutputShape(e),s=this._getOutputShape(t);return n&&s&&!oB(n,s)?{legal:!1,reason:`elementwise shape mismatch: [${n}] vs [${s}]`}:yl}_checkProducerConsumerShapes(e,t){let n=this._getOutputShape(e);if(!n)return yl;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 oB(n,o.shape)?yl:{legal:!1,reason:`shape mismatch on data edge: [${n}] vs [${o.shape}]`}}return yl}_getOutputShape(e){for(let t=0;t<e.numResults;t++){let n=e.getResult(t).type;if(n instanceof I)return n.shape}return null}};l(oB,"shapesCompatible")});function Ba(r){return zh(r,"throw")}function nf(r,{labelOf:e,sameLabel:t=l((i,a)=>i===a,"sameLabel"),canMerge:n=l(()=>!0,"canMerge"),onAttach:s=l(()=>{},"onAttach"),sort:o=Ba}){let i=o(r),a=new Map,u=new Map,p=[],c=0,h=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 T of w)y.push(T)}return!1},"isUpstreamOf"),f=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"),d=l((m,g)=>{for(let y of f(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)||!n(w,m,g))continue;let T=!1;for(let S of f(m))if(S!==w&&h(w,S)){T=!0;break}if(!T){w.ops.push(m),w.opSet.add(m),s(w,m),a.set(m,w),d(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,_),d(m,_)}}return{partitions:p,opToPart:a,preds:u}}function _i(r,e,{pullConstants:t=!1,isConstant:n=l(()=>!1,"isConstant")}={}){let s=[],o=new Set,i=[],a=new Set,u=[],p=new Set;for(let c of e){for(let h=0;h<c.numOperands;h++){let f=c.getOperand(h),d=f.definingOp;if(!(d&&r.has(d))){if(t&&d&&n(d)){p.has(d)||(p.add(d),u.push(d));continue}o.has(f)||(o.add(f),s.push(f))}}for(let h=0;h<c.numResults;h++){let f=c.getResult(h);if(a.has(f))continue;let d=!1;for(let m of f.uses())if(!r.has(m.user)){d=!0;break}d&&(a.add(f),i.push(f))}}return{inputs:s,outputs:i,constDefs:u}}function hT(r,e){let t=new Map,n=new Map;for(let a of r)t.set(a,0),n.set(a,[]);for(let a of r){let u=e.get(a);if(u)for(let p of u)n.has(p)&&(n.get(p).push(a),t.set(a,t.get(a)+1))}let s=[];for(let a of r)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 n.get(a)){let p=t.get(u)-1;t.set(u,p),p===0&&s.push(u)}}return o.length===r.length?o:null}var rf=M(()=>{"use strict";ji();l(Ba,"topoSortOps");l(nf,"buildPartitions");l(_i,"computePartitionIO");l(hT,"topoSortPartitions")});function $$(r){for(let e=0;e<r.numResults;e++){let t=r.getResult(e).type;if(t instanceof I)return t.shape.join(",")}return null}var Eo,Yg,Jg=M(()=>{"use strict";Oe();At();Oa();Ui();rf();Eo=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}=_i(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=bl(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)&&!ko(e.opName))return!1}return!0}};l($$,"outputShapeKey");Yg=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 n=[],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 h=ne.get(c.opName);if(!h||h.isConstant||h.isReduction)continue;let f=s.get(i),d=s.get(c);if(f&&f===d||!this.legality.canFuse(c,i).legal)continue;let g=this._topoIndex.get(c);if(!f&&!d){let y=new Eo(this._nextId++);y.addOp(c,g),y.addOp(i,u),s.set(c,y),s.set(i,y)}else if(f&&!d)f.size<this.legality.maxFusionSize&&!this._wouldCreateCycle(f,c)&&(f.addOp(c,g),s.set(c,f));else if(!f&&d)d.size<this.legality.maxFusionSize&&!this._wouldCreateCycle(d,i)&&(d.addOp(i,u),s.set(i,d));else if(this.legality.canMergeGroups(f,d).legal&&!this._mergeWouldCreateCycle(f,d)){f.merge(d);for(let _ of d.ops)s.set(_,f)}}}let o=new Set;for(let i of s.values())o.has(i)||i.size<2||(o.add(i),i.classifyKind(),n.push(i));return n}_wouldCreateCycle(e,t){for(let n=0;n<t.numOperands;n++){let s=t.getOperand(n).definingOp;if(!(!s||e.hasOp(s)||this._topoIndex.get(s)<e.minTopoPos)&&this._dependsOnGroup(s,e))return!0}return!1}_dependsOnOps(e,t,n){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)<n||(s.add(u),o.push(u))}}return!1}_dependsOnGroup(e,t){return this._dependsOnOps(e,t.opSet,t.minTopoPos)}_mergeWouldCreateCycle(e,t){let n=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<n||p>s)&&this._dependsOnOps(u,o,n))return!0}return!1}_bucketable(e,t){return t&&!t.isConstant&&!t.isTerminator&&!t.isOpaque&&e.numResults>0&&$$(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 n=[],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 h=p.getResult(0).type;a++,i.set(p,a);let f=new Eo(this._nextId++);f.addOp(p,u),s.set(p,f);let d=Math.min(t.length,u+1+o);for(let m=u+1;m<d;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(f.size>=o||s.has(g))continue;let _=ne.get(g.opName);this._bucketable(g,_)&&h.equals(g.getResult(0).type)&&(this._sharesInput(p,g)||c.isElementwise&&_.isElementwise)&&(f.addOp(g,m),s.set(g,f),i.set(g,a))}f.size>=2&&(f.kind=je.HORIZONTAL,n.push(f))}return n}buildAllGroups(e){let t=this.buildProducerConsumerGroups(e),n=this.buildHorizontalGroups(e),s=new Map;for(let c of t)for(let h of c.ops)s.set(h,c);let o=new Set(s.keys()),i=[];for(let c of n){let h=!1;for(let f of c.ops)if(o.has(f)){h=!0;break}h||i.push(c)}let a=[...e.ops()],u=[];for(let c of a)for(let h=0;h<c.numOperands;h++){let f=c.getOperand(h).definingOp;f&&u.push(c,f)}for(let c of i)for(let h of c.ops)s.set(h,c);if(!this._condensedHasCycle(a,u,s))return[...t,...i];for(let c of i)for(let h of c.ops)s.delete(h);let p=[...t];for(let c of i){for(let h of c.ops)s.set(h,c);if(this._condensedHasCycle(a,u,s)){for(let h of c.ops)s.delete(h);continue}p.push(c)}return p}_condensedHasCycle(e,t,n){let s=l(h=>n.get(h)||h,"repOf"),o=new Map,i=new Set;for(let h of e)i.add(s(h));for(let h=0;h<t.length;h+=2){let f=s(t[h]),d=s(t[h+1]);if(d===f)continue;i.add(d),i.add(f);let m=o.get(d);m||(m=new Set,o.set(d,m)),m.add(f)}let a=0,u=1,p=2,c=new Map;for(let h of i)c.set(h,a);for(let h of i){if(c.get(h)!==a)continue;let f=[h];for(;f.length>0;){let d=f[f.length-1],m=c.get(d);if(m===a){c.set(d,u);let g=o.get(d);if(g)for(let y of g){let _=c.get(y);if(_===u)return!0;_===a&&f.push(y)}}else m===u&&c.set(d,p),f.pop()}}return!1}_sharesInput(e,t){if(e.numOperands===0||t.numOperands===0)return!1;let n=new Set;for(let s=0;s<e.numOperands;s++)n.add(e.getOperand(s));for(let s=0;s<t.numOperands;s++)if(n.has(t.getOperand(s)))return!0;return!1}_hasDependency(e,t){for(let n=0;n<e.numResults;n++){let s=e.getResult(n);for(let o=0;o<t.numOperands;o++)if(t.getOperand(o)===s)return!0}return!1}}});var F$,bi,Zg=M(()=>{"use strict";Oe();At();F$=Object.freeze({memory:1,launch:1e3}),bi=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={...F$,...e.benefitWeights||{}}}edgeBenefit(e){let t=this.benefitWeights;return t.launch*this.launchOverheadUs+t.memory*e}estimateOpCost(e){let t=this.estimateFLOPs(e),n=this.estimateBytes(e);return{flops:t,bytes:n,arithmeticIntensity:n>0?t/n:0}}estimateFLOPs(e){let t=ne.get(e.opName);if(t&&t.getFlops)return t.getFlops(e);let n=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&&(n=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 n}estimateBytes(e){let t=0;for(let n=0;n<e.numOperands;n++){let s=e.getOperand(n).type;if(s instanceof I){let o=s.sizeInBytes();o!==X&&(t+=o)}}for(let n=0;n<e.numResults;n++){let s=e.getResult(n).type;if(s instanceof I){let o=s.sizeInBytes();o!==X&&(t+=o)}}return t}estimateGroupCost(e){let t=e.opSet,n=new Map,s=0,o=0,i=0,a=0,u=!1,p=!1,c=0,h=0,f=0,d=0,m=new Set;for(let T of e.ops){let S=this.estimateFLOPs(T);n.set(T,S),s+=S,o+=this.estimateBytes(T),this.libraryOps.has(T.opName)&&a++;let x=ne.get(T.opName);if(x){let k=this._outputElements(T);x.isReduction&&(u=!0,c+=k),x.isElementwise&&(p=!0,h+=k)}for(let k=0;k<T.numResults;k++){let A=T.getResult(k);m.add(A);let N=0,v=!1;for(let O of A.uses())t.has(O.user)?N++:v=!0;N>1&&(i+=(N-1)*S),(N>0||v)&&f++}for(let k=0;k<T.numOperands;k++){let A=T.getOperand(k);if(!m.has(A))continue;let N=!1;for(let v of A.uses())if(t.has(v.user)&&v.user!==T){N=!0;break}N||f--}f>d&&(d=f)}let g=e.getInputValues(),y=e.getOutputValues(),_=0;for(let T of g)if(T.type instanceof I){let S=T.type.sizeInBytes();S!==X&&(_+=S)}for(let T of y)if(T.type instanceof I){let S=T.type.sizeInBytes();S!==X&&(_+=S)}let b=0;for(let T of e.ops)for(let S=0;S<T.numResults;S++){let x=T.getResult(S),k=0;for(let A of x.uses())if(t.has(A.user)&&(k++,k>1))break;if(k>1){let A=x.type;if(A instanceof I){let N=A.sizeInBytes();N!==X&&(b+=N)}}}let w=0;return u&&p&&h>0&&c>0&&(w=Math.abs(h-c)),{unfusedFLOPs:s,unfusedBytes:o,fusedFLOPs:s+i,fusedBytes:_,recomputeCost:i,memorySaved:o-_,launchSaved:(e.size-1)*this.launchOverheadUs,registerPressure:d*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 n=this.policy.shouldFuse(e,this);if(n)return n}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 n=t.memorySaved+t.launchSaved*1e3;if(t.parallelismLoss>n*this.minBenefitRatio)return{fuse:!1,reason:"parallelism loss outweighs fusion benefit",cost:t}}return{fuse:!0,reason:`saves ${t.memorySaved} bytes, ${t.launchSaved}us launch`,cost:t}}_outputElements(e){for(let t=0;t<e.numResults;t++){let n=e.getResult(t).type;if(n instanceof I){let s=n.numel();if(s!==X)return s}}return 0}}});function wl(r){let e=null;for(let t of r.ops())e=t;return e&&e.opName==="yield"?e:null}function Tl(r){let e=0,t=r.regions[0]?.entryBlock;if(!t)return 0;for(let n of t.ops())n.opName!=="yield"&&e++;return e}function xl(r){let e=0,t=r.regions[0]?.entryBlock;if(!t)return 0;for(let n of t.ops()){let s=ne.get(n.opName);s&&s.isReduction&&e++}return e}function Sl(r){let e=r.regions[0]?.entryBlock;if(!e)return!1;for(let t of e.ops())if(t.opName!=="yield"&&!ko(t.opName))return!1;return!0}function z$(r){let e=new Map,t=0;for(let n=r.firstOp;n;n=n._next)e.set(n,t++);return e}function dT(r){let e=z$(r);return(t,n)=>{let s=e.get(t),o=e.get(n);return s!==void 0&&o!==void 0&&s<o}}function kl(r,e){let t=wo(r.ops,g=>r.hasOp(g),"null");if(t===null||t.length===0)return null;r._inputValues=null,r._outputValues=null;let n=r.getInputValues(),s=r.getOutputValues(),o=n.map(g=>g.type),i=s.map(g=>g.type),a=new Ct,u=new yn(o);a.addBlock(u);let p=new Map;for(let g=0;g<n.length;g++)p.set(n[g],u.arguments[g]);for(let g of t)u.pushOp(g.clone(p));let c=s.map(g=>{let y=p.get(g);if(y===void 0)throw new Error("Fusion materialization: output value not found in valueMap");return y});u.pushOp(new Re("yield",c,[]));let h=new Re("fusion",n,i,{fusion_kind:r.kind||e},[a]),f=t[0].parentBlock;if(!f)return null;let d=dT(f),m=null;for(let g of n){let y=g.definingOp;!y||r.hasOp(y)||(!m||!d(y,m))&&(m=y)}m&&m.parentBlock===f?f.insertAfter(h,m):f.insertBefore(h,t[0]);for(let g=0;g<s.length;g++)s[g].replaceAllUsesWith(h.getResult(g));for(let g of t)g.dropAllOperands(),g.parentBlock&&g.parentBlock.removeOp(g);return h}var vl=M(()=>{"use strict";At();Ui();sr();zi();ji();l(wl,"getYieldOp");l(Tl,"countInnerOps");l(xl,"countReductions");l(Sl,"allInnerOpsFusable");l(z$,"blockPositionIndex");l(dT,"makeComesBefore");l(kl,"materializeFusionGroup")});var Qg,iB=M(()=>{"use strict";nn();Oa();Jg();Zg();vl();Kt();Qg=class extends Ge{static{l(this,"FusionPass")}constructor(e={}){super("FusionPass");let t=e.target||{};this.legality=new _l({maxFusionSize:t.maxFusionSize||e.maxFusionSize,maxSharedMemory:t.sharedMemoryBytes||e.maxSharedMemory,libraryOps:t.libraryOps||e.libraryOps,allowReductionFusion:e.allowReductionFusion}),this.costModel=new bi({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 Yg(this.legality)}run(e,t){let n=this.groupBuilder.buildAllGroups(e),s=[];for(let i of n)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)kl(i,je.ELEMENTWISE);return oe.CHANGED}_traceDecision(e,t,n){if(!this.trace||this.trace.level<ue.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:n||null,level:ue.DEBUG}),this.trace.explain("fusion",s.join("+"),t?"fused":"not-fused",n||null,{groupSize:s.length})}_createsCycle(e,t){let n=t.opSet,s=t.getOutputValues(),o=new Set,i=[];for(let u of s)for(let p of u.uses())n.has(p.user)||(i.push(p.user),o.add(p.user));let a=0;for(;a<i.length;){let u=i[a++];if(n.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 Xg(r){return lu(r.opName)||ln(r.opName)}function aB(r){if(Xg(r))return!0;let e=ne.get(r.opName);return e!==null&&e.isElementwise}function G$(r,e){let t=r.getOperand(0).definingOp;return t&&e.has(t)&&!Xg(t)?r.getOperand(1).definingOp:t}function j$(r,e){if(!r.parentBlock||r.parentBlock!==e.parentBlock)return!1;let t=r.parentBlock.firstOp;for(;t;){if(t===r)return!0;if(t===e)return!1;t=t._next}return!1}function q$(r,e){for(let t of r)if(t!==e){for(let n=0;n<t.numResults;n++)for(let s of t.getResult(n).uses())if(!r.has(s.user))return!0}return!1}function U$(r,e){let t=ne.get(r.opName),n=t&&t.getAttr("epilogueTag");return n?n(r,e):"activation"}function K$(r){let e=[],t=new Set,n=new Set;function s(f){if(!(n.has(f)||f===r)){n.add(f);for(let d=0;d<f.numOperands;d++){let m=f.getOperand(d).definingOp;m&&m!==r&&!n.has(m)&&Xg(m)&&s(m)}e.push(f),t.add(f)}}l(s,"absorb");let o=[],i=r.getResult(0);for(let f of i.uses())aB(f.user)&&o.push(f.user);for(;o.length>0;){let f=o.pop();if(!n.has(f)){s(f);for(let d=0;d<f.numResults;d++)for(let m of f.getResult(d).uses())!n.has(m.user)&&aB(m.user)&&o.push(m.user)}}let a=[];for(let f of e)Xg(f)||a.push(U$(f,t));let u=e[e.length-1];for(let f=e.length-1;f>=0;f--){let d=e[f];for(let m=0;m<d.numResults;m++){for(let g of d.getResult(m).uses())if(!t.has(g.user)){u=d,f=-1;break}if(f<0)break}}let p=new Set;for(let f=0;f<r.numOperands;f++)p.add(r.getOperand(f));let c=[],h=new Set;for(let f of e)for(let d=0;d<f.numOperands;d++){let m=f.getOperand(d);if(h.has(m))continue;h.add(m);let g=m.definingOp;g!==r&&(g&&t.has(g)||p.has(m)||c.push(m))}return{chain:e,chainSet:t,tags:a,lastOp:u,extras:c}}function W$(r){let e=null,t=null,n=null,s=null,o=new Set;for(let i=0;i<2;i++){let a=r.getOperand(i),u=a.definingOp;if(!u||u.opName!=="convert")continue;let p=!1;for(let h of a.uses())if(h.user!==r){p=!0;break}if(p)continue;let c=u.getAttr("target_dtype")||u.getResult(0).type.dtype;i===0?(e=c,n=u.getOperand(0)):(t=c,s=u.getOperand(0)),o.add(u)}return{lhsCast:e,rhsCast:t,lhsInput:n,rhsInput:s,removed:o}}var V$,ey,uB=M(()=>{"use strict";nn();sr();At();Tr();Kt();l(Xg,"isPassthrough");l(aB,"isEpilogueCandidate");V$=new Map([["add",(r,e)=>e.has(G$(r,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(G$,"resolveOtherOperand");for(let[r,e]of V$)ne.has(r)&&ne.registerOpAttr(r,"epilogueTag",e);l(j$,"comesBefore");l(q$,"hasEscapingUse");l(U$,"classifyTag");l(K$,"collectChainAndAnalyze");l(W$,"collectPrologue");ey=class extends Ge{static{l(this,"EpilogueFusionPass")}constructor(e={}){super("EpilogueFusionPass"),this.maxEpilogueOps=e.maxEpilogueOps||16,this.target=e.target||null}run(e){if(this.target&&!this.target.enableEpilogueFusion)return oe.UNCHANGED;let t=!1,n=[];for(let s of e.ops()){let o=ne.get(s.opName);o&&o.isOutEWiseFusable&&n.push(s)}for(let s of n){let o=K$(s);if(o.chain.length>this.maxEpilogueOps)continue;let i=W$(s),a=o.chain.length>0,u=i.lhsCast!==null||i.rhsCast!==null;if(!a&&!u)continue;let{chain:p,tags:c,extras:h}=o,f=0;for(let k of c)k==="bias"||k==="residual_add"||k==="scale"?f++:k==="clamp"&&(f+=2);if(f!==h.length)continue;let d=i.lhsInput||s.getOperand(0),m=i.rhsInput||s.getOperand(1),g=[d,m,...h],y=a?o.lastOp:s,_=y.getResult(0).type,b=new Map(s.attributes);b.set("epilogue_ops",p.map(k=>k.opName)),b.set("epilogue_tags",c),b.set("num_dot_operands",2),b.set("num_extra_inputs",h.length),i.lhsCast&&b.set("lhs_prologue_cast",i.lhsCast),i.rhsCast&&b.set("rhs_prologue_cast",i.rhsCast);let w=new Re("fused_dot_epilogue",g,[_],b),T=s.parentBlock;if(!T)continue;let S=new Set(p);S.add(s);for(let k of i.removed)S.add(k);if(q$(S,y))continue;let x=null;for(let k of g){let A=k.definingOp;!A||S.has(A)||(!x||!j$(A,x))&&(x=A)}x&&x.parentBlock===T?T.insertAfter(w,x):T.insertBefore(w,s),y.getResult(0).replaceAllUsesWith(w.getResult(0));for(let k=p.length-1;k>=0;k--)p[k].dropAllOperands(),p[k].parentBlock&&p[k].parentBlock.removeOp(p[k]);s.dropAllOperands(),s.parentBlock&&s.parentBlock.removeOp(s);for(let k of i.removed)k.dropAllOperands(),k.parentBlock&&k.parentBlock.removeOp(k);t=!0}return this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,dotsFound:n.length,changed:t,level:ue.DEBUG}),t?oe.CHANGED:oe.UNCHANGED}}});var ty,lB=M(()=>{"use strict";nn();sr();zi();Oe();At();Kt();Oa();vl();ty=class extends Ge{static{l(this,"FusionMergerPass")}constructor(e={}){super("FusionMergerPass"),this.maxFusionSize=e.maxFusionSize||512,this.maxReductions=e.maxReductions||1,this.launchOverheadUs=e.launchOverheadUs||5,this.minMemorySavings=e.minMemorySavings||0}run(e){let t=[];for(let u of e.ops())u.opName==="fusion"&&t.push(u);if(t.length<2)return oe.UNCHANGED;let n=new Set(t),s=this._buildProducerConsumerEdges(t,n);if(s.length===0)return oe.UNCHANGED;let o=!1,i=0,a=new Set;for(let{producer:u,consumer:p,sharedResults:c}of s)a.has(u)||a.has(p)||this._canMerge(u,p)&&this._shouldMerge(u,p,c)&&(this._mergeCreatesCycle(u,p,n)||(this._merge(u,p,c),a.add(u),a.add(p),o=!0,i++));return this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,fusionOps:t.length,edges:s.length,mergeCount:i,level:ue.DEBUG}),o?oe.CHANGED:oe.UNCHANGED}_buildProducerConsumerEdges(e,t){let n=[],s=new Map;for(let o of e)for(let i=0;i<o.numResults;i++){let a=o.getResult(i);for(let u of a.uses()){if(!t.has(u.user))continue;let p=u.user;if(p===o)continue;let c=`${o.id}|${p.id}`,h=s.get(c);h||(h={producer:o,consumer:p,sharedResults:new Map},s.set(c,h),n.push(h)),h.sharedResults.set(i,u.operandIndex)}}return n}_canMerge(e,t){let n=Tl(e),s=Tl(t);if(n+s>this.maxFusionSize)return!1;let o=0;return o+=xl(e),o+=xl(t),!(o>this.maxReductions||!Sl(e)||!Sl(t))}_shouldMerge(e,t,n){let s=0;for(let[u]of n){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,n){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,n){let s=e.regions[0].entryBlock,o=t.regions[0].entryBlock,i=wl(s),a=wl(o);if(!i||!a)return;let u=new Map;for(let[N,v]of n)u.set(v,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 h=new Map;for(let N=0;N<t.numOperands;N++){if(u.has(N))continue;let v=t.getOperand(N),O=-1;for(let R=0;R<p.length;R++)if(p[R]===v){O=R;break}O>=0?h.set(N,O):(h.set(N,p.length),p.push(v))}let f=p.map(N=>N.type),d=new Ct,m=new yn(f);d.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 v=i.getOperand(N),O=g.get(v)||v;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 v=h.get(N);v!==void 0&&g.set(o.arguments[N],m.arguments[v])}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 v=!1;for(let O of e.getResult(N).uses())if(O.user!==t){v=!0;break}if(v){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 v=a.getOperand(N),O=g.get(v)||v;y.push(O),_.push(t.getResult(N).type)}let w=new Re("yield",y,[]);m.pushOp(w);let T=[];for(let N of m.ops())N.opName!=="yield"&&T.push(N);let S=bl(T),x=new Re("fusion",p,_,{fusion_kind:S},[d]),k=t;if(!k.parentBlock)return;k.parentBlock.insertBefore(x,k);let A=0;for(let{resultIdx:N}of b)e.getResult(N).replaceAllUsesWith(x.getResult(A)),A++;for(let N=0;N<t.numResults;N++)t.getResult(N).replaceAllUsesWith(x.getResult(A)),A++;e.dropAllOperands(),e.parentBlock&&e.parentBlock.removeOp(e),t.dropAllOperands(),t.parentBlock&&t.parentBlock.removeOp(t)}_estimateRecomputeCost(e){let t=0,n=e.regions[0]?.entryBlock;if(!n)return 0;for(let s of n.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 H$(r,e){let t=Math.min(r.id,e.id),n=Math.max(r.id,e.id);return`${t}|${n}`}var sf,cB=M(()=>{"use strict";nn();sr();zi();Oe();Oa();Kt();vl();sf=class extends Ge{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 n=this._findCandidates(t);if(n.length===0)return oe.UNCHANGED;let s=!1,o=0,i=new Set;for(let{left:a,right:u,sharedInputs:p,sharedBytes:c}of n)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>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,fusionOps:t.length,candidates:n.length,mergeCount:o,level:ue.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 n=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],h=i[p];if(c===h)continue;let f=H$(c,h),d=n.get(f);d||(d={left:c,right:h,sharedInputs:new Set,sharedBytes:0},n.set(f,d)),d.sharedInputs.add(o)}}let s=[];for(let o of n.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 h=p.getOperand(c);if(h.id===a&&h.type instanceof I){let f=h.type.sizeInBytes();f!==X&&(i+=f);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 n=Tl(e),s=Tl(t);if(n+s>this.maxFusionSize||e.numResults+t.numResults>this.maxOutputs)return!1;let i=0;return i+=xl(e),i+=xl(t),!(i>this.maxReductions||!Sl(e)||!Sl(t))}_mergeCreatesCycle(e,t){let n=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&&!n.has(u.user)&&(n.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())n.has(u.user)||(n.add(u.user),s.push(u.user))}n.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&&!n.has(u.user)&&(n.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())n.has(u.user)||(n.add(u.user),s.push(u.user))}return!1}_mergeMultiOutput(e,t,n){let s=e.regions[0].entryBlock,o=t.regions[0].entryBlock,i=wl(s),a=wl(o);if(!i||!a)return;let u=[],p=new Map,c=new Map;for(let v=0;v<e.numOperands;v++){let O=e.getOperand(v),R=O.id,B=p.get(R);B===void 0&&(B=u.length,p.set(R,B),u.push(O)),c.set(v,B)}let h=new Map;for(let v=0;v<t.numOperands;v++){let O=t.getOperand(v),R=O.id,B=p.get(R);B===void 0&&(B=u.length,p.set(R,B),u.push(O)),h.set(v,B)}let f=u.map(v=>v.type),d=new Ct,m=new yn(f);d.addBlock(m);let g=new Map;for(let v=0;v<s.arguments.length;v++)g.set(s.arguments[v],m.arguments[c.get(v)]);for(let v of s.ops())v.opName!=="yield"&&m.pushOp(v.clone(g));for(let v=0;v<o.arguments.length;v++)g.set(o.arguments[v],m.arguments[h.get(v)]);for(let v of o.ops())v.opName!=="yield"&&m.pushOp(v.clone(g));let y=[],_=[];for(let v=0;v<i.numOperands;v++){let O=i.getOperand(v);y.push(g.get(O)||O),_.push(e.getResult(v).type)}for(let v=0;v<a.numOperands;v++){let O=a.getOperand(v);y.push(g.get(O)||O),_.push(t.getResult(v).type)}let b=new Re("yield",y,[]);m.pushOp(b);let w=[];for(let v of m.ops())v.opName!=="yield"&&w.push(v);let T=bl(w),S=new Re("fusion",u,_,{fusion_kind:T},[d]),x=e.parentBlock;if(!x)return;let k=dT(x),A=null;for(let v of u){let O=v.definingOp;!O||O===e||O===t||(!A||!k(O,A))&&(A=O)}A&&A.parentBlock===x?x.insertAfter(S,A):x.insertBefore(S,e);for(let v=0;v<e.numResults;v++)e.getResult(v).replaceAllUsesWith(S.getResult(v));let N=e.numResults;for(let v=0;v<t.numResults;v++)t.getResult(v).replaceAllUsesWith(S.getResult(N+v));e.dropAllOperands(),e.parentBlock&&e.parentBlock.removeOp(e),t.dropAllOperands(),t.parentBlock&&t.parentBlock.removeOp(t)}_hasProducerConsumerEdge(e,t){for(let n=0;n<e.numResults;n++)for(let s of e.getResult(n).uses())if(s.user===t)return!0;return!1}};l(H$,"pairKey")});var mT,at,Es=M(()=>{"use strict";mT=class{static{l(this,"UseDefResult")}constructor(e,t,n,s,o){this.topologicalOrder=e,this.valueToOp=t,this.opUsers=n,this.depth=s,this.height=o}},at=class{static get name(){return"use_def"}static get depKey(){return"useDef"}static get dependencies(){return[]}static compute(e){let t=[],n=new Map,s=new Map,o=new Set,i=new Set,a=l(c=>{if(o.has(c))return;i.add(c);let h=[{op:c,i:0}];for(;h.length>0;){let f=h[h.length-1],d=f.op;if(f.i<d.numOperands){let m=d.getOperand(f.i).definingOp;if(f.i++,m&&!o.has(m)){if(i.has(m))throw new Error("Cycle detected in UseDefAnalysis");i.add(m),h.push({op:m,i:0})}continue}i.delete(d),o.add(d),t.push(d),h.pop()}},"visit");for(let c of e.ops()){for(let h=0;h<c.numResults;h++)n.set(c.getResult(h),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 h=0;for(let f=0;f<c.numOperands;f++){let d=c.getOperand(f).definingOp;if(d){s.get(d).add(c);let m=u.get(d)||0;m+1>h&&(h=m+1)}}u.set(c,h)}for(let c=t.length-1;c>=0;c--){let h=t[c],f=0,d=s.get(h)||new Set;for(let m of d){let g=p.get(m)||0;g+1>f&&(f=g+1)}p.set(h,f)}return new mT(t,n,s,u,p)}}});var gT,ny,pB=M(()=>{"use strict";Es();Tr();gT=class{static{l(this,"DominanceResult")}constructor(e){this.idom=e}postDominates(e,t){let n=t;for(;n;){if(n===e)return!0;n=this.idom.get(n)}return!1}immediatePDom(e){return this.idom.get(e)||null}pathToPDom(e){let t=[],n=this.idom.get(e);for(;n&&(t.push(n),n!==this.idom.get(n));)n=this.idom.get(n);return t}},ny=class{static get name(){return"post_dominance"}static get depKey(){return"postDominance"}static get dependencies(){return[at]}static compute(e,t={}){let s=(t.useDef||at.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)qt(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,h=new Map;c.set(a,0),h.set(a,new Array(u).fill(a));let f=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[_]=h.get(y[_-1])[_-1];h.set(m,y)},"link"),d=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=h.get(m)[_]);if(m===g)return m;for(let _=u-1;_>=0;_--)h.get(m)[_]!==h.get(g)[_]&&(m=h.get(m)[_],g=h.get(g)[_]);return h.get(m)[0]},"lca");for(let m=s.length-1;m>=0;m--){let g=s[m];if(i.has(g)){f(g,a);continue}let y=null;for(let _ of o.get(g))p.has(_)&&(y=y===null?_:d(y,_));y!==null&&f(g,y)}for(let[m,g]of p)g===a&&p.delete(m);return new gT(p,s)}}});function fB(r){return qt(r)||ln(r)}var ry,hB=M(()=>{"use strict";nn();At();Oa();Jg();Zg();pB();Kt();Es();vl();Tr();l(fB,"isSkipOp");ry=class extends Ge{static{l(this,"DominatorFusionPass")}constructor(e={}){super("DominatorFusionPass"),this.requiredAnalyses=[at];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 bi({memoryBandwidthGBs:t.memoryBandwidthGBs,computeTFLOPs:t.computeTFLOPs,maxRegistersPerThread:t.registersPerThread,maxSharedMemory:t.sharedMemoryBytes,libraryOps:t.libraryOps,...e.cost})}run(e,t){let n=t?t.getAnalysis(at,e):at.compute(e),s=ny.compute(e,{useDef:n}),o=n.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)kl(u,je.ELEMENTWISE);return this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,groupsBuilt:i.length,groupsFused:a.length,level:ue.DEBUG}),oe.CHANGED}_explain(e,t,n){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",n||null,{groupSize:s.length,strategy:"dominator"})}_buildGroups(e,t){let n=new Map,s=[],o=0;for(let p=e.length-1;p>=0;p--){let c=e[p];if(fB(c.opName))continue;let h=Aa(c);if(h===je.OPAQUE||this.libraryOps.has(c.opName))continue;let f=t.immediatePDom(c);if(!f||fB(f.opName))continue;let d=Aa(f);if(d===je.OPAQUE||this.libraryOps.has(f.opName)||!this._canFusePatterns(h,d)||!this._pathAllFusable(c,f,e,t))continue;let m=n.get(f);if(m){if(m.size+1>this.maxFusionSize||!this._checkReductionLimit(m,c))continue;m.addOp(c),n.set(c,m),this._absorbIntermediates(c,f,m,n,e,t)}else{let g=n.get(c);if(g){if(g.size+1>this.maxFusionSize||!this._checkReductionLimit(g,f))continue;g.addOp(f),n.set(f,g),this._absorbIntermediates(c,f,g,n,e,t)}else{let y=new Eo(o++);if(y.addOp(c),y.addOp(f),!this._checkGroupReductions(y))continue;n.set(c,y),n.set(f,y),s.push(y),this._absorbIntermediates(c,f,y,n,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 n.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 n=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(!n.has(o)){if(n.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 fT(e,t)}_pathAllFusable(e,t,n,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),Aa(c)===je.OPAQUE||this.libraryOps.has(c.opName))return!1;i.push(c)}}return!0}_absorbIntermediates(e,t,n,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 h of p.getResult(c).uses()){let f=h.user;f===t||a.has(f)||!i.postDominates(t,f)||(a.add(f),Aa(f)===je.OPAQUE||this.libraryOps.has(f.opName))||(n.size<this.maxFusionSize&&this._checkReductionLimit(n,f)&&(n.addOp(f),s.set(f,n)),u.push(f))}}}_checkGroupReductions(e){let t=0;for(let n of e.ops){let s=ne.get(n.opName);s&&s.isReduction&&t++}return t<=this.maxReductions}_checkReductionLimit(e,t){let n=ne.get(t.opName);if(!n||!n.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 sy,dB=M(()=>{"use strict";sy=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 n=0;n<e;n++)this._parent[n]=n,this._rank[n]=n,this._nodeAtRank[n]=n,this._out[n]=new Set,this._in[n]=new Set;if(t)for(let[n,s]of t)n!==s&&(this._out[n].add(s),this._in[s].add(n))}find(e){let t=this._parent,n=e;for(;t[n]!==n;)n=t[n];for(;t[e]!==n;){let s=t[e];t[e]=n,e=s}return n}wouldCreateCycle(e,t){let n=this.find(e),s=this.find(t);if(n===s)return!1;let o=this._rank[n]<this._rank[s]?n:s,i=o===n?s:n;return this._hasIntermediatePath(o,i)}_hasIntermediatePath(e,t){let n=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]>=n||s.has(u)||(s.add(u),o.push(u))}}}return!1}merge(e,t){let n=this.find(e),s=this.find(t);if(n===s)return n;let o=this._rank[n]<this._rank[s]?n:s,i=o===n?s:n,a=this._rank[o],u=this._rank[i],p=this._out[n].size+this._in[n].size,c=this._out[s].size+this._in[s].size,h=p>=c?n:s,f=h===n?s:n,d=this._out[h],m=this._in[h];for(let g of this._in[f]){let y=this.find(g);y===h||y===f||(this._out[y].delete(f),this._out[y].add(h),m.add(y))}for(let g of this._out[f]){let y=this.find(g);y===h||y===f||(this._in[y].delete(f),this._in[y].add(h),d.add(y))}return d.delete(f),m.delete(f),d.delete(h),m.delete(h),this._parent[f]=h,this._nodeAtRank[this._rank[f]]=-1,this._reorder(a,u),h}_reorder(e,t){let n=[],s=[];for(let c=e;c<=t;c++){let h=this._nodeAtRank[c];if(!(h<0)){if(this.find(h)!==h){this._nodeAtRank[c]=-1;continue}n.push(c),s.push(h)}}if(s.length<=1){for(let c=0;c<s.length;c++)this._rank[s[c]]=n[c],this._nodeAtRank[n[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 h of this._out[c]){let f=this.find(h);f!==c&&o.has(f)&&i.set(f,i.get(f)+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 h of this._out[c]){let f=this.find(h);if(f!==c&&o.has(f)){let d=i.get(f)-1;i.set(f,d),d===0&&a.push(f)}}}for(let c=0;c<u.length;c++){let h=u[c],f=n[c];this._rank[h]=f,this._nodeAtRank[f]=h}}}});var oy,mB=M(()=>{"use strict";oy=class{static{l(this,"MaxHeap")}constructor(){this._items=[]}get size(){return this._items.length}isEmpty(){return this._items.length===0}push(e,t){let n=this._items;n.push({priority:e,value:t});let s=n.length-1;for(;s>0;){let o=s-1>>1;if(n[o].priority>=n[s].priority)break;let i=n[o];n[o]=n[s],n[s]=i,s=o}}pop(){let e=this._items,t=e.length;if(t===0)return;let n=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 n.value}}});function Y$(r){let e=ne.get(r.opName);return!e||e.isConstant||e.isTerminator||e.isOpaque?!1:e.isReduction?!0:ko(r.opName)}var iy,gB=M(()=>{"use strict";nn();At();Oe();Oa();Jg();Zg();Es();dB();mB();vl();Ui();Kt();l(Y$,"isFusibleOp");iy=class extends Ge{static{l(this,"PriorityFusionPass")}constructor(e={}){super("PriorityFusionPass"),this.requiredAnalyses=[at];let t=e.target||{};this.maxReductions=e.maxReductions??1,this.legality=new _l({maxFusionSize:t.maxFusionSize||e.maxFusionSize,maxSharedMemory:t.sharedMemoryBytes||e.maxSharedMemory,libraryOps:t.libraryOps||e.libraryOps,allowReductionFusion:e.allowReductionFusion});let n=t.getAttr&&t.getAttr("fusionBenefitWeights")||e.benefitWeights;this.costModel=new bi({memoryBandwidthGBs:t.memoryBandwidthGBs,computeTFLOPs:t.computeTFLOPs,maxRegistersPerThread:t.registersPerThread,maxSharedMemory:t.sharedMemoryBytes,libraryOps:t.libraryOps,policy:t.getAttr?t.getAttr("fusionPolicy"):null,benefitWeights:n,...e.cost})}run(e,t){let s=(t?t.getAnalysis(at,e):at.compute(e)).topologicalOrder,o=s.length;if(o===0)return oe.UNCHANGED;let i=new Map;for(let x=0;x<o;x++)i.set(s[x],x);let a=[];for(let x=0;x<o;x++){let k=s[x];for(let A=0;A<k.numOperands;A++){let N=k.getOperand(A).definingOp;N&&i.has(N)&&a.push([i.get(N),x])}}let u=new sy(o,a),p=new Int32Array(o),c=new Map;for(let x=0;x<o;x++)if(Y$(s[x])){let k=new Eo(x);k.addOp(s[x]),c.set(x,k)}let h=new oy,f=l((x,k)=>{let A=new Eo(-1);for(let N of x.ops)A.addOp(N);for(let N of k.ops)A.addOp(N);return A},"mergedGroup"),d=l(x=>{if(x.size<2||!x.allOpsInlineFusable())return!1;let k=0;for(let A of x.ops){let N=ne.get(A.opName);N&&N.isReduction&&k++}return k>this.maxReductions||(x.classifyKind(),x.kind===je.OPAQUE)?!1:this.costModel.shouldFuse(x).fuse===!0},"legalMerge"),m=l((x,k)=>{let A=0;for(let N=0;N<k.numOperands;N++){let v=k.getOperand(N);if(v.definingOp===x&&v.type instanceof I){let O=v.type.sizeInBytes();O!==X&&(A+=O)}}return A},"edgeBytes"),g=l((x,k)=>{let A=u.find(i.get(x)),N=u.find(i.get(k));if(A===N||!c.has(A)||!c.has(N)||!this.legality.canFuse(x,k).legal)return;let v=this.costModel.edgeBenefit(m(x,k));h.push(v,{a:A,b:N,va:p[A],vb:p[N],prodOp:x,consOp:k})},"pushCandidate"),y=l((x,k)=>{let A=new Set;for(let N of x.ops){for(let v=0;v<N.numResults;v++)for(let O of N.getResult(v).uses()){let R=i.get(O.user);if(R===void 0)continue;let B=u.find(R);if(B===k||!c.has(B))continue;let E=k*o+B;A.has(E)||(A.add(E),g(N,O.user))}for(let v=0;v<N.numOperands;v++){let O=N.getOperand(v).definingOp;if(!O)continue;let R=i.get(O);if(R===void 0)continue;let B=u.find(R);if(B===k||!c.has(B))continue;let E=B*o+k;A.has(E)||(A.add(E),g(O,N))}}},"reEval"),_=new Set;for(let x=0;x<o;x++){if(!c.has(x))continue;let k=s[x];for(let A=0;A<k.numOperands;A++){let N=k.getOperand(A).definingOp;if(!N)continue;let v=i.get(N);if(v===void 0||!c.has(v))continue;let O=v*o+x;_.has(O)||(_.add(O),g(N,k))}}let b=0;for(;!h.isEmpty();){let x=h.pop(),k=u.find(x.a),A=u.find(x.b);if(k===A||p[k]!==x.va||p[A]!==x.vb)continue;let N=c.get(k),v=c.get(A);if(!N||!v||!this.legality.canFuse(x.prodOp,x.consOp).legal)continue;let O=f(N,v);if(!d(O)||u.wouldCreateCycle(k,A))continue;let R=u.merge(k,A),B=N.size>=v.size?N:v,E=B===N?v:N;B.merge(E),c.delete(k),c.delete(A),c.set(R,B),p[R]=(p[k]>p[A]?p[k]:p[A])+1,b++,y(B,R)}if(b===0)return oe.UNCHANGED;let w=[],T=new Set;for(let x of c.values())T.has(x)||x.size<2||(T.add(x),w.push(x));let S=0;for(let x of w)x.classifyKind(),this._explain(x),kl(x,je.ELEMENTWISE)&&S++;return this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,ops:o,merges:b,fusions:S,level:ue.DEBUG}),S>0?oe.CHANGED:oe.UNCHANGED}_explain(e){if(!this.trace||!this.trace.explainsEnabled)return;let t=e.ops.map(n=>n.opName);this.trace.explain("fusion",t.join("+"),"fused",null,{groupSize:t.length,strategy:"priority"})}}});function J$(r,e){return r===e?!0:!r||!e?!1:typeof r.equals=="function"?r.equals(e):!1}var wi,ay,yB=M(()=>{"use strict";Oe();wi=class{static{l(this,"LayoutPreference")}constructor(e,t,n=0){this.inputs=e,this.outputs=t,this.cost=n}},ay=class{static{l(this,"LayoutPolicy")}constructor(e){this.target=e,this._rules=new Map,this._initDefaultRules()}registerRule(e,t){this._rules.set(e,t)}getPreference(e){let t=this._rules.get(e.opName);return t?t(e,this.target):null}estimateConversionCost(e,t,n){if(!(n instanceof I)||J$(e,t))return 0;let s=n.numel();return s<0?1024:s*2}estimateBenefit(e,t,n){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*n;if(o==="reduce")return s*2*n;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 wi([t.preferredConvLayout,null],[t.preferredConvLayout]);if(t.isGPU()&&s===4){let o=new Et([0,2,3,1]);return new wi([o,null],[o])}if(t.isCPU()&&s===4){let o=new Et([0,2,3,1]);return new wi([o,null],[o])}return null}),this._rules.set("dot",(e,t)=>{let n=e.getOperand(0).type,s=e.getOperand(1).type;if(!n||!s)return null;let o=Et.rowMajor(n.rank);if(t.isCPU()&&s.rank===2){let a=Et.columnMajor(s.rank);return new wi([o,a],[o])}let i=Et.rowMajor(s.rank);return new wi([o,i],[o])}),this._rules.set("reduce",(e,t)=>{let n=e.getResult(0).type;return n?new wi([null],[Et.rowMajor(n.rank)]):null})}};l(J$,"layoutEquals")});function _B(r,e){let t=new Map,n=null,s=0;for(let i=0;i<r.numOperands;i++){let a=r.getOperand(i),u=e.get(a);if(!u)continue;let p=u.hash?u.hash():0,c=(t.get(p)||0)+1;t.set(p,c),c>s&&(s=c,n=u)}if(n)return n;let o=r.getResult(0);return o&&o.type instanceof I?Et.rowMajor(o.type.rank):Et.rowMajor(1)}function Z$(r,e){return r===e?!0:!r||!e?!1:typeof r.equals=="function"?r.equals(e):!1}function bB(r){if(r instanceof Et)return r;if(r&&typeof r.toLayout=="function")try{return r.toLayout()}catch{}return r&&r.baseOrder&&!r.isBlocked?.()?new Et(r.baseOrder):r&&r.order?new Et(r.order):null}var yT,uy,wB=M(()=>{"use strict";Es();Oe();At();_n();Tr();yT=class{static{l(this,"LayoutAnalysisResult")}constructor(e,t,n){this.assignments=e,this.conversions=t,this.totalCost=n}},uy=class{static get name(){return"layout"}static get depKey(){return"layout"}static get dependencies(){return[at]}static compute(e,t,n){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||Et.rowMajor(p.type.rank));for(let p=0;p<i.length;p++){let c=i[p];if(qt(c.opName))continue;let h=n?n.getPreference(c):null;if(h&&h.outputs.length>0){for(let m=0;m<c.numResults;m++){let g=c.getResult(m);if(!(g.type instanceof I))continue;let y=h.outputs[m]||null;y?o.set(g,y):o.set(g,_B(c,o))}continue}let f=ne.get(c.opName),d=f&&f.hasTrait(H.ELEMENTWISE);for(let m=0;m<c.numResults;m++){let g=c.getResult(m);g.type instanceof I&&(d?o.set(g,_B(c,o)):o.set(g,Et.rowMajor(g.type.rank)))}}let a=[],u=0;for(let p=0;p<i.length;p++){let c=i[p],h=n?n.getPreference(c):null;if(h)for(let f=0;f<c.numOperands;f++){let d=c.getOperand(f);if(!(d.type instanceof I))continue;let m=o.get(d);if(!m)continue;let g=h.inputs[f];if(!g||Z$(m,g))continue;let y=bB(m),_=bB(g);!y||!_||y.equals(_)||(a.push({value:d,consumer:c,operandIdx:f,from:y,to:_}),n&&(u+=n.estimateConversionCost(y,_,d.type)))}}return new yT(o,a,u)}};l(_B,"resolveFromInputs");l(Z$,"layoutEquals");l(bB,"toPlainLayout")});function Q$(r,e,t){let n=r.id,s=e.hash?e.hash():0,o=t.hash?t.hash():0;return`${n}:${s}:${o}`}var ly,TB=M(()=>{"use strict";nn();sr();Oe();yB();wB();Es();Kt();ly=class extends Ge{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 ay(this.target));let n=t?t.getAnalysis(at,e):at.compute(e),s=uy.compute(e,{useDef:n},this._policy);if(s.conversions.length===0)return oe.UNCHANGED;let o=new Map;for(let p of s.conversions){let{value:c,consumer:h,operandIdx:f,from:d,to:m}=p,g=Q$(c,d,m),y=o.get(g);y||(y={value:c,from:d,to:m,consumers:[],cost:this._policy.estimateConversionCost(d,m,c.type),benefit:0},o.set(g,y)),y.consumers.push({consumer:h,operandIdx:f});let _=this.target.layoutAwareOps&&this.target.layoutAwareOps.has(h.opName);y.benefit+=_?this._policy.estimateBenefit(h,c.type,1):0}let i=0,a=0,u=[];for(let p of o.values())p.benefit<p.cost||(u.push(p),i+=p.cost,a+=p.benefit);if(u.length===0||i>a)return oe.UNCHANGED;for(let p of u){let c=p.from instanceof Et?p.from.order:Array.from({length:p.value.type.rank},(y,_)=>_),h=p.to instanceof Et?p.to.order:Array.from({length:p.value.type.rank},(y,_)=>_),f=new I(p.value.type.shape,p.value.type.dtype,p.to),d=new Re("layout_transform",[p.value],[f],{src_layout:[...c],dst_layout:[...h]}),m=p.value.definingOp;m&&m.parentBlock?m.parentBlock.insertAfter(d,m):p.consumers[0].consumer.parentBlock&&p.consumers[0].consumer.parentBlock.insertBefore(d,p.consumers[0].consumer);let g=d.getResult(0);for(let y of p.consumers)y.consumer.replaceOperand(y.operandIdx,g)}return this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,conversions:s.conversions.length,uniqueTransforms:u.length,level:ue.DEBUG}),oe.CHANGED}};l(Q$,"valueLayoutKey")});function xB(r,e){if(r===e)return!0;if(typeof r=="number"&&typeof e=="number")return r===e;if(typeof r!=typeof e||r.length!==e.length)return!1;for(let t=0;t<r.length;t++)if(r[t]!==e[t])return!1;return!0}function SB(r){let e=0;for(let t=0;t<r.length;t++)e=(e<<5)-e+r.charCodeAt(t)&2147483647;return e}function tF(r){if(typeof r=="number"){let t=new Float64Array([r]),n=new Uint32Array(t.buffer);return(n[0]^n[1])&2147483647}let e=r.length;for(let t=0;t<Math.min(r.length,8);t++){let n=new Float64Array([r[t]]),s=new Uint32Array(n.buffer);e=(e^(s[0]^s[1]))*16777619&2147483647}return e}var kr,X$,eF,hr,_T=M(()=>{"use strict";Oe();kr=Object.freeze({PER_TENSOR_SYMMETRIC:"per_tensor_symmetric",PER_TENSOR_ASYMMETRIC:"per_tensor_asymmetric",PER_CHANNEL:"per_channel",PER_GROUP:"per_group"}),X$=new Set(Object.values(kr)),eF=new Set(["i8","ui8"]),hr=class r{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||Gs(this.dtype)*8,this._hash=null}clampRange(){if(this.isSymmetric()){let n=2**(this.numBits-1)-1;return[-n,n]}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,n]=this.clampRange(),s=Math.round(e/this.getScalarScale()+this.getScalarZeroPoint());return Math.max(t,Math.min(n,s))}dequantize(e){return(e-this.getScalarZeroPoint())*this.getScalarScale()}quantizeArray(e){let[t,n]=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(n,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 n=this.getScalarScale(),s=this.getScalarZeroPoint();for(let o=0;o<e.length;o++)t[o]=(e[o]-s)*n;return t}_channelStride(e){let t=1;for(let n=this.axis+1;n<e.length;n++)t*=e[n];return t}quantizeArrayPerChannel(e,t){let[n,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),h=this.getZeroPointForChannel(p);a[u]=Math.max(n,Math.min(s,Math.round(e[u]/c+h)))}return a}dequantizeArrayPerChannel(e,t){let n=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/n)%s;o[i]=(e[i]-this.getZeroPointForChannel(a))*this.getScaleForChannel(a)}return o}quantizeArrayPerGroup(e){let[t,n]=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(n,Math.round(e[o]/a+u)))}return s}dequantizeArrayPerGroup(e){let t=new Array(e.length);for(let n=0;n<e.length;n++){let s=Math.floor(n/this.groupSize);t[n]=(e[n]-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===kr.PER_CHANNEL}isPerGroup(){return this.scheme===kr.PER_GROUP}isSymmetric(){return this.scheme===kr.PER_TENSOR_SYMMETRIC}equals(e){return this===e?!0:!(!(e instanceof r)||this.scheme!==e.scheme||this.dtype!==e.dtype||this.numBits!==e.numBits||this.axis!==e.axis||this.groupSize!==e.groupSize||!xB(this.scale,e.scale)||!xB(this.zeroPoint,e.zeroPoint))}hash(){if(this._hash!==null)return this._hash;let e=2166136261;return e=(e^SB(this.scheme))*16777619&2147483647,e=(e^SB(this.dtype))*16777619&2147483647,e=(e^this.numBits)*16777619&2147483647,e=(e^tF(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 r(e)}static fromRange(e,t,n,s="i8",o=8){if(n===kr.PER_TENSOR_SYMMETRIC){let f=Math.max(Math.abs(e),Math.abs(t)),d=2**(o-1)-1,m=f/d||1e-10;return new r({scheme:n,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),h=Math.max(i,Math.min(a,c));return new r({scheme:n,scale:p,zeroPoint:h,dtype:s,numBits:o})}static fromRangePerChannel(e,t,n,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 h=Math.max(Math.abs(e[c]),Math.abs(t[c]));a[c]=h/p||1e-10,u[c]=0}return new r({scheme:kr.PER_CHANNEL,scale:a,zeroPoint:u,axis:n,dtype:s,numBits:o})}static defaultForActivation(e,t="i8",n=8){return r.fromRange(-6,6,e,t,n)}static fromConstantArray(e,t,n="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),r.fromRange(o,i,t,n,s)}static fromConstantArrayPerChannel(e,t,n,s="i8",o=8){let i=t[n],a=1;for(let c=n+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 h=Math.floor(c/a)%i,f=e[c];f<u[h]&&(u[h]=f),f>p[h]&&(p[h]=f)}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 r.fromRangePerChannel(u,p,n,s,o)}static fromConstantArrayPerGroup(e,t,n="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,h=Math.min(c+t,e.length),f=0;for(let d=c;d<h;d++){let m=Math.abs(e[d]);m>f&&(f=m)}i[p]=f/u||1e-10,a[p]=0}return new r({scheme:kr.PER_GROUP,scale:i,zeroPoint:a,groupSize:t,dtype:n,numBits:s})}static isQuantizableDtype(e){return eF.has(e)}static isValidScheme(e){return X$.has(e)}};l(xB,"scaleEquals");l(SB,"hashStr");l(tF,"hashScaleValue")});function oF(r){for(let e=0;e<r.numOperands;e++){let t=r.getOperand(e).definingOp;if(t&&t.opName==="constant")return!0}return!1}function iF(r,e,t){for(let n=0;n<r.numOperands;n++){let s=r.getOperand(n);if(!e.has(s)&&(!(s.type instanceof I)||!Lt(s.type.dtype)))return!1}return r.numOperands>0}var nF,rF,sF,cy,py,kB=M(()=>{"use strict";nn();sr();Oe();At();Es();_T();Kt();Tr();nF=new Set(["softmax","sqrt","div","rsqrt","log","exp","tanh"]),rF=new Set(["dot","conv","add","mul","sub"]),sF=new Map([["dot","quantized_dot"],["conv","quantized_conv"]]);for(let[r,e]of sF)ne.has(r)&&ne.registerOpAttr(r,"quantizedVariant",e);cy=class{static{l(this,"QuantizationConfig")}constructor(e={}){this.scheme=e.scheme||kr.PER_TENSOR_SYMMETRIC,this.calibration=e.calibration||null,this.targetDtype=e.targetDtype||"i8",this.excludeOps=e.excludeOps||nF,this.quantizableOps=e.quantizableOps||rF,this.sensitivityThreshold=e.sensitivityThreshold||0,this.sensitivityResult=e.sensitivityResult||null,this.weightOnly=e.weightOnly||!1,this.target=e.target||null}},py=class extends Ge{static{l(this,"QuantizationPass")}constructor(e={}){super("QuantizationPass"),this.requiredAnalyses=[at],this.config=e instanceof cy?e:new cy(e)}run(e,t){let s=(t?t.getAnalysis(at,e):at.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(qt(c.opName))continue;if(i.excludeOps.has(c.opName)||!i.quantizableOps.has(c.opName)){for(let d=0;d<c.numOperands;d++)o.has(c.getOperand(d))&&(a=this._insertDequantBefore(c,d,c.getOperand(d),i)||a);continue}if(i.sensitivityResult&&i.sensitivityThreshold>0&&i.sensitivityResult.isSensitive(c,i.sensitivityThreshold)||i.weightOnly&&!oF(c))continue;let h=ne.get(c.opName),f=h?h.getAttr("quantizedVariant"):null;if(f&&iF(c,o,i)){i.scheme===kr.PER_CHANNEL&&this._canPerChannelDot(c,o)?a=this._replacePerChannelDot(c,i)||a:a=this._replaceWithNativeQuantized(c,f,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>=ue.DEBUG&&a&&this.trace.emit({type:"pass_detail",passName:this.name,opsProcessed:s.length,changed:a,level:ue.DEBUG}),a?oe.CHANGED:oe.UNCHANGED}_getQuantParams(e,t){let n=Gs(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 hr.fromRange(-Math.abs(o)||-1,Math.abs(o)||1,t.scheme,t.targetDtype,n);if(o&&typeof o.length=="number")return hr.fromConstantArray(o,t.scheme,t.targetDtype,n)}return e.type instanceof I&&Lt(e.type.dtype)?hr.defaultForActivation(t.scheme,t.targetDtype,n):null}_resolveQuantParams(e,t){let n=this._paramsByValue.get(e);if(n)return n;let s=e.definingOp;return s&&s.opName==="quantize"?new hr({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,n){let s=e.getResult(t),o=this._getQuantParams(s,n);if(!o)return null;let i=new I(s.type.shape,n.targetDtype),a=new Re("quantize",[s],[i],{scale:o.getScalarScale(),zero_point:o.getScalarZeroPoint(),scheme:n.scheme,target_dtype:n.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,n,s){let o=this._resolveQuantParams(n,s);if(!o)return!1;let i="f32",a=new I(n.type.shape,i),u=new Re("dequantize",[n],[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,n){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,n)||s)}for(let o=0;o<e.numResults;o++){let i=e.getResult(o);if(!(i.type instanceof I)||!Lt(i.type.dtype))continue;let a=this._insertQuantizeAfter(e,o,n);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,n,s){let o=[],i={};for(let[h,f]of e.attributes||[])i[h]=f;for(let h=0;h<e.numOperands;h++){let f=e.getOperand(h);if(n.has(f)){o.push(f);let _=this._resolveQuantParams(f,s);if(_){let b=h===0?"lhs":"rhs";if(t==="quantized_dot")i[`${b}_scale`]=_.getScalarScale(),i[`${b}_zero_point`]=_.getScalarZeroPoint();else{let w=h===0?"input":"kernel";i[`${w}_scale`]=_.getScalarScale(),i[`${w}_zero_point`]=_.getScalarZeroPoint()}}continue}let d=this._getQuantParams(f,s);if(!d)return!1;let m=new I(f.type.shape,s.targetDtype),g=new Re("quantize",[f],[m],{scale:d.getScalarScale(),zero_point:d.getScalarZeroPoint(),scheme:s.scheme,target_dtype:s.targetDtype});e.parentBlock&&e.parentBlock.insertBefore(g,e),o.push(g.getResult(0));let y=h===0?t==="quantized_dot"?"lhs":"input":t==="quantized_dot"?"rhs":"kernel";i[`${y}_scale`]=d.getScalarScale(),i[`${y}_zero_point`]=d.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 Re(t,o,[p],i);if(e.parentBlock){e.parentBlock.insertBefore(c,e);let h=new I(e.getResult(0).type.shape,e.getResult(0).type.dtype),f=i.output_scale||1,d=new Re("dequantize",[c.getResult(0)],[h],{scale:f,zero_point:i.output_zero_point||0,scheme:s.scheme,target_dtype:e.getResult(0).type.dtype});e.parentBlock.insertBefore(d,e),e.replaceAllResultsWith([d.getResult(0)]),e.erase()}return!0}_canPerChannelDot(e,t){if(e.opName!=="dot")return!1;let n=e.getOperand(0),s=e.getOperand(1);if(t.has(n)||t.has(s)||!(n.type instanceof I)||n.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 n=Gs(t.targetDtype)*8,s=kr.PER_TENSOR_SYMMETRIC;return t.calibration&&t.calibration.hasData(e)?t.calibration.getQuantParams(e,s,t.targetDtype):hr.defaultForActivation(s,t.targetDtype,n)}_replacePerChannelDot(e,t){let n=e.getOperand(0),s=e.getOperand(1),o=s.type.shape,i=s.definingOp.getAttr("value"),a=Gs(t.targetDtype)*8,p=e.getAttr("rhs_contracting")[0]===0?1:0,c=hr.fromConstantArrayPerChannel([...i],o,p,t.targetDtype,a),h=c.quantizeArrayPerChannel([...i],o),f=this._activationParams(n,t),d=f.getScalarScale(),m=f.getScalarZeroPoint(),g=e.parentBlock;if(!g)return!1;let y=new I(o,t.targetDtype),_=new Re("constant",[],[y],{value:h,tensor_type:y});g.insertBefore(_,e);let b=new I(n.type.shape,t.targetDtype),w=new Re("quantize",[n],[b],{scale:d,zero_point:m,scheme:kr.PER_TENSOR_SYMMETRIC,target_dtype:t.targetDtype});g.insertBefore(w,e);let T=e.getResult(0).type.shape,S={};for(let[U,P]of e.attributes||[])S[U]=P;S.lhs_scale=d,S.lhs_zero_point=m,S.rhs_scale=1,S.rhs_zero_point=0,S.output_scale=1,S.output_zero_point=0;let x=new I(T,"i32"),k=new Re("quantized_dot",[w.getResult(0),_.getResult(0)],[x],S);g.insertBefore(k,e);let A=new I(T,"f32"),N=new Re("convert",[k.getResult(0)],[A],{target_dtype:"f32"});g.insertBefore(N,e);let v=o[p],O=new Array(v);for(let U=0;U<v;U++)O[U]=d*c.getScaleForChannel(U);let R=new Set(e.getAttr("lhs_contracting")||[]),B=0;for(let U=0;U<n.type.shape.length;U++)R.has(U)||B++;let E=B,z=new I([v],"f32"),D=new Re("constant",[],[z],{value:O,tensor_type:z});g.insertBefore(D,e);let J=new Re("broadcast_in_dim",[D.getResult(0)],[A],{broadcast_dimensions:[E],result_shape:T});g.insertBefore(J,e);let j=new Re("mul",[N.getResult(0),J.getResult(0)],[A],{});return g.insertBefore(j,e),e.replaceAllResultsWith([j.getResult(0)]),e.erase(),!0}};l(oF,"hasConstantOperand");l(iF,"allOperandsCanQuantize")});var bT,fy,vB=M(()=>{"use strict";Es();Oe();bT=class{static{l(this,"LivenessResult")}constructor(e,t,n,s,o,i,a){this.liveIn=e,this.liveOut=t,this.intervals=n,this.opIndex=s,this.peakPressure=o,this.peakOp=i,this.pressureAtOp=a}interfere(e,t){if(e===t)return!0;let n=this.intervals.get(e),s=this.intervals.get(t);return n&&s?n.start<=s.end&&s.start<=n.end:!1}liveAtOp(e){return this.liveIn.get(e)||new Set}intervalOf(e){return this.intervals.get(e)||null}},fy=class r{static get name(){return"liveness"}static get depKey(){return"liveness"}static get dependencies(){return[at]}static buildIntervals(e,t){let n=new Map;for(let o=0;o<t.length;o++)n.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:n}}static compute(e,t={}){let n=t.useDef||at.compute(e),s=n.topologicalOrder,{intervals:o,opIndex:i}=r.buildIntervals(e,s),a=new Map,u=new Map;for(let f of s)a.set(f,new Set),u.set(f,new Set);for(let f=s.length-1;f>=0;f--){let d=s[f],m=u.get(d),g=n.opUsers.get(d);if(g)for(let _ of g){let b=a.get(_);if(b)for(let w of b)m.add(w)}let y=a.get(d);for(let _ of m)y.add(_);for(let _=0;_<d.numResults;_++)y.delete(d.getResult(_));for(let _=0;_<d.numOperands;_++)y.add(d.getOperand(_))}for(let[f,d]of i){let m=a.get(f);for(let g of m){let y=o.get(g);y&&y.end<d&&(y.end=d)}}let p=0,c=null,h=new Map;for(let f=0;f<s.length;f++){let d=s[f],m=a.get(d),g=0;for(let y of m)if(y.type instanceof I){let _=y.type.sizeInBytes();_!==X&&(g+=_)}h.set(d,g),g>p&&(p=g,c=d)}return new bT(a,u,o,i,p,c,h)}}});var hy,dy,NB=M(()=>{"use strict";nn();sr();Oe();At();Kt();Es();vB();Tr();hy=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}},dy=class extends Ge{static{l(this,"RematerializationPass")}constructor(e={}){super("RematerializationPass"),this.config=e instanceof hy?e:new hy(e)}run(e,t){if(this.config.memoryBudget===1/0)return oe.UNCHANGED;let n=!1,s=0,o=0;for(;s<this.config.maxIterations;){let i=at.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,h)=>h.score-c.score),this._rematerialize(e,u[0],p),n=!0,s++}return this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,iterations:s,peakPressure:o,budget:this.config.memoryBudget,changed:n,level:ue.DEBUG}),n?oe.CHANGED:oe.UNCHANGED}_analyzeIntervalPressure(e,t){let n=t.topologicalOrder,{intervals:s,opIndex:o}=fy.buildIntervals(e,n),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,h=new Set,f=0;for(let m=-1;m<=n.length;m++){for(;f<i.length&&i[f].idx<=m;)a+=i[f].delta,i[f].delta>0&&i[f].value&&h.add(i[f].value),i[f].delta<0&&i[f].value&&h.delete(i[f].value),f++;if(a>u){u=a,p=m,c.clear();for(let g of h)c.add(g)}}let d=[];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&&d.push({value:m,definingOp:g,memorySaved:y,recomputeCost:_,score:y/_})}return{peakPressure:u,peakIdx:p,candidates:d,opIndex:o}}_canRematerialize(e){if(e.isBlockArgument())return!1;let t=e.definingOp;if(!t||qt(t.opName)||ln(t.opName)||this.config.excludeOps.has(t.opName)||t.regions.length>0||t.hasSideEffects()||e.useCount<=1)return!1;for(let n=0;n<t.numOperands;n++){let s=t.getOperand(n);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,n){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=n.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 n=t.getFlops(e);if(n>0)return n}if(t.isElementwise||t.isBroadcast||t.isView){let n=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){n=i;break}}}return n||1}if(t.isReduction)for(let n=0;n<e.numOperands;n++){let s=e.getOperand(n).type;if(s instanceof I){let o=s.numel();if(o!==X)return o}}return 1/0}_rematerialize(e,t,n){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,h)=>(n.get(c.user)||0)-(n.get(h.user)||0));let a=n.get(i[i.length-1].user)||0,u=this._operandExtensionCost(o,a,n),p=this._computeMemorySaved(s);if(!(u>=p))for(let c=1;c<i.length;c++){let h=i[c],f=h.user;if(!f.parentBlock)continue;let d=new Re(o.opName,[...o.operands],o.results.map(m=>m.type),new Map(o.attributes));f.parentBlock.insertBefore(d,f),f.replaceOperand(h.operandIndex,d.getResult(s.resultIndex))}}}});function oo(r){return ln(r.opName)}function AB(r){let e=0;for(let t=0;t<r.numResults;t++){let n=r.getResult(t).type;if(!n||!n.isFullyStatic)continue;let s=n.sizeInBytes();s>e&&(e=s)}return e}function uF(r){let e=r.definingOp;if(e&&e.opName==="broadcast"){let t=e.getOperand(0);e=t&&t.definingOp}if(e&&oo(e)){let t=e.getAttr("value");if(typeof t=="number")return t}return 0}function wT(r){if(r.opName!=="dot"&&r.opName!=="cublas_gemm")return null;let e=r.getOperand(0).type,t=r.getOperand(1).type,n=r.getResult(0).type;if(!e||!t||!n||e.dtype!=="f32"||t.dtype!=="f32"||n.dtype!=="f32")return null;let s=r.getAttr("lhs_batch")||[],o=r.getAttr("rhs_batch")||[];if(s.length>0||o.length>0)return null;let i=r.getAttr("lhs_contracting")||[],a=r.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||!n.isFullyStatic)return null;let u=r.getOperand(0).definingOp,p=r.getOperand(1).definingOp;if(u&&oo(u)||p&&oo(p))return null;let c=a[0]===1,h=e.shape[e.rank-1];if((c?t.shape[1]:t.shape[0])!==h)return null;let d=1;for(let g=0;g<e.rank-1;g++)d*=e.shape[g];let m=c?t.shape[0]:t.shape[1];return d<=0||m<=0||h<=0?null:{M:d,N:m,K:h,transB:c}}function OB(r){return{labelOf:l(e=>r.get(e),"labelOf"),canMerge:l((e,t)=>Math.max(e.maxBuf||0,AB(t))<=aF,"canMerge"),onAttach:l((e,t)=>{e.maxBuf=Math.max(e.maxBuf||0,AB(t))},"onAttach")}}function my(r,e,t){let n=r.opSet,s=Ba(r.ops),{inputs:o,outputs:i,constDefs:a}=_i(n,s,{pullConstants:!0,isConstant:oo});for(let f of o)if(!f.type||!f.type.isFullyStatic)return null;for(let f of i)if(!f.type||!f.type.isFullyStatic)return null;let u=new un(e,o.map(f=>f.type),i.map(f=>f.type)),p=new Map;for(let f=0;f<o.length;f++)p.set(o[f],u.args[f]);for(let f of a)u.entryBlock.pushOp(f.clone(p));for(let f of s)u.entryBlock.pushOp(f.clone(p));let c=i.map(f=>p.get(f));if(c.some(f=>f===void 0))return null;u.entryBlock.pushOp(new Re("return",c,[]));let h=r.ops.length===1&&t.has(r.ops[0])?r.ops[0]:null;return{part:r,subFunc:u,inputs:o,outputs:i,dotOp:h}}function of(r){if(lF.has(r.opName))return!0;if(r.regions)for(let e of r.regions){let t=e.entryBlock;if(t){for(let n of t.ops())if(of(n))return!0}}return!1}function cF(r){let e=0;for(let t=0;t<r.numResults;t++){let n=r.getResult(t).type;if(!n||!n.isFullyStatic)continue;let s=1;for(let o of n.shape)s*=o;s>e&&(e=s)}return e}function BB(r,e=256){if(r.functionCount!==1)return!1;let t=r.functions().next().value,n=new Map,s=l(o=>{if(!o)return 0;let i=n.get(o);if(i!==void 0)return i;n.set(o,0);let a=of(o)?cF(o):0;for(let u=0;u<o.numOperands;u++){let p=s(o.getOperand(u).definingOp);p>a&&(a=p)}return n.set(o,a),a},"maxBoundaryInSubtree");for(let o of t.ops())if(!(qt(o.opName)||!of(o))){for(let i=0;i<o.numOperands;i++)if(s(o.getOperand(i).definingOp)>e)return!0}return!1}function RB(r,e,t){let n=new Map,s=0,o=l(d=>{let m=n.get(d);return m===void 0&&(m=s++,n.set(d,m)),m},"getSlot");for(let d of r.args)o(d);for(let d of t)for(let m of d.outputs)o(m);let i=[];for(let d of r.args)i.push(o(d));let a=[],u=new Set;for(let d=0;d<e.numOperands;d++){let m=e.getOperand(d),g=i.length,y=m.isBlockArgument&&m.isBlockArgument(),_=m.definingOp&&oo(m.definingOp);if(!y&&!_&&n.has(m)){let b=n.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:uF(m)});continue}return null}let p=[];for(let d of t){let m=[];for(let y of d.inputs){let _=n.get(y);if(_===void 0)return null;m.push(_)}let g=d.outputs.map(y=>n.get(y));p.push({name:d.subFunc.name,inputSlots:m,outputSlots:g})}let c=new Set(i),h=[],f=new Set;for(let[d,m]of n)if(!(c.has(m)||f.has(m))){if(f.add(m),!d.type||!d.type.isFullyStatic)return null;h.push({slot:m,shape:[...d.type.shape],dtype:d.type.dtype})}return{plan:{numSlots:s,argSlots:i,intermediates:h,steps:p,returnFixups:a}}}function Ti(r,e=2){if(r.functionCount!==1)return null;let t=r.functions().next().value,n=t.getReturnOp();if(!n)return null;let s=[],o=new Map,i=0;for(let g of t.ops())qt(g.opName)||oo(g)||(of(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}=nf(s,OB(o));if(a.length<2)return null;let p=hT(a,u);if(!p)return null;let c=t.name,h=[],f=new Map,d=0;for(let g of p){let y=my(g,c+"_p"+d++,f);if(!y)return null;h.push(y)}let m=RB(t,n,h);if(!m)return null;r.removeFunction(t.name);for(let g of h)r.addFunction(g.subFunc);return{plan:m.plan}}function MB(r){if(r.functionCount!==1)return null;let e=r.functions().next().value,t=e.getReturnOp();if(!t)return null;let n=[],s=new Map,o=new Map,i=0,a=0;for(let y of e.ops()){if(qt(y.opName)||oo(y))continue;let _=wT(y);_?(s.set(y,"cublas#"+i),o.set(y,_),i++):of(y)?s.set(y,"boundary#"+a++):s.set(y,"native"),n.push(y)}if(i+a===0||n.length===0)return null;let{partitions:u,preds:p}=nf(n,OB(s));if(u.length<2)return null;let c=hT(u,p);if(!c)return null;let h=e.name,f=[],d=0;for(let y of c){let _=my(y,h+"_p"+d++,o);if(!_)return null;f.push(_)}let m=RB(e,t,f);if(!m)return null;let g=new Map;for(let y of f){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)),T=y.outputs.indexOf(y.dotOp.getResult(0));b<0||w<0||T<0||g.set(y.subFunc.name,{M:_.M,N:_.N,K:_.K,transB:_.transB,aIdx:b,bIdx:w,cIdx:y.inputs.length+T})}if(g.size===0)return null;r.removeFunction(e.name);for(let y of f)r.addFunction(y.subFunc);return{plan:m.plan,cublasInfos:g}}var aF,lF,gy=M(()=>{"use strict";xs();sr();rf();Tr();l(oo,"isConstantOp");aF=32*1024;l(AB,"maxResultBytes");l(uF,"constScalarOf");l(wT,"cublasDotInfo");l(OB,"bufferLimitedConfig");l(my,"materializePartition");lF=new Set(["dot","cublas_gemm","reduce","conv","quantized_conv","scaled_dot_product_attention"]);l(of,"containsBoundaryOp");l(cF,"maxBoundaryResultBytes");l(BB,"hasDependentBoundaries");l(RB,"buildExecutionPlan");l(Ti,"splitGraphForNative");l(MB,"splitGraphForCublas")});function EB(r){return IB.push(r),r}var IB,yy,LB=M(()=>{"use strict";nn();IB=[];l(EB,"registerBackendOpRewrite");yy=class extends Ge{static{l(this,"BackendOpRewritePass")}constructor(e={}){super(e.name||"BackendOpRewritePass"),this.config=e,this.rewrites=e.rewrites||IB}run(e){let t=!1;for(let n of[...e.ops()]){let s=n.parentBlock;if(s){for(let o of this.rewrites)if(o.match(n,this.config)){o.build(n,s,this.config),t=!0;break}}}return t?oe.CHANGED:oe.UNCHANGED}}});var pF,_y,CB=M(()=>{"use strict";sr();gy();LB();pF=EB({name:"dot->cublas_gemm",match:l(r=>r.opName==="dot"&&r.numOperands===2&&wT(r)!==null,"match"),build:l((r,e)=>{let t=r.results.map(s=>s.type),n=new Re("cublas_gemm",[r.getOperand(0),r.getOperand(1)],t,new Map(r.attributes),null);e.insertBefore(n,r),r.replaceAllResultsWith(n.results),r.erase()},"build")}),_y=class extends yy{static{l(this,"CublasRewritePass")}constructor(e={}){super({...e,name:"CublasRewritePass",rewrites:[pF]})}}});function DB(r,e,{cudaMatmulChain:t=!1,context:n=null}={}){let s=n?a=>n.passesForPhase(a,r,e):a=>Gg(a,r,e),o=[];for(let a of s("pre"))o.push(a);if(o.push(new Gu(e)),o.push(new Vu("canonicalize",[new ia,new Kg({fastMath:r.optimization.fastMath}),new Wg,new Hg,new no],r.optimization.maxSimplifyIterations)),r.optimization.layout&&e&&(o.push(new ly({target:e})),o.push(new no)),r.quantization.enabled&&(o.push(new py({...r.quantization,target:e})),o.push(new ia),o.push(new no)),r.matmulBackend!=="cublas"&&!t&&(r.fusion.epilogue!==void 0?r.fusion.epilogue:e&&e.enableEpilogueFusion)&&(o.push(new ey({target:e})),o.push(new no)),r.fusion.enabled){let a=r.fusion,u=a.launchOverheadUs??fF;a.strategy==="dominator"?o.push(new ry({target:e,...a})):a.strategy==="priority"?(o.push(new iy({target:e,cost:{launchOverheadUs:u},...a})),o.push(new sf({maxFusionSize:e?.maxFusionSize,...a}))):(o.push(new Qg({target:e,cost:{launchOverheadUs:u},...a})),o.push(new ty({maxFusionSize:e?.maxFusionSize,...a})),o.push(new sf({maxFusionSize:e?.maxFusionSize,...a}))),o.push(new no)}if(r.matmulBackend==="cublas"&&o.push(new _y),r.optimization.rematerialization){let a={...r.optimization.rematConfig};a.memoryBudget===void 0&&e&&e.memoryBudgetBytes>0&&(a.memoryBudget=e.memoryBudgetBytes),o.push(new dy(a))}for(let a of s("post"))o.push(a);return o}var fF,PB=M(()=>{"use strict";um();Dw();eB();rB();sB();$w();iB();uB();lB();cB();hB();gB();TB();kB();Lw();NB();CB();uT();fF=5;l(DB,"buildGraphPipeline")});var Wn,xi=M(()=>{"use strict";Wn=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 by,TT,Ra,xT=M(()=>{"use strict";De();by=["iterCount","depth","parallelLoops","vectorizedLoops","unrolledLoops","threadBoundLoops","serialLoops","threadBlockSize","gridSize","underReduction","numMathOps","numExternCalls","numReads","numWrites","stride1Accesses","stridedAccesses","reuseCount","touchedBytes","arithmeticIntensity","vectorized","parallelized","innermostExtent"],TT=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}},Ra=class r{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};r._visitIterative(e.body,t);let n=t.totalBufferBytes,s=t.numMathOps+t.numExternCalls;return t.arithmeticIntensity=n>0?s/n:0,new TT(t)}static extractStatements(e){let t=[],n=[],s=0,o=[{node:e.body,action:"enter"}];for(;o.length>0;){let{node:i,action:a}=o.pop();if(i){if(a==="leaveFor"){n.pop();continue}if(a==="leaveBlock"){s--;continue}switch(i.type){case"ForNode":n.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(r._statementVector(i,n,s));break}}}return t}static _statementVector(e,t,n){let s=1,o=0,i=0,a=0,u=0,p=0,c=1,h=1;for(let x of t){let k=x.extent&&x.extent.type==="IntImmNode"?x.extent.value:1;switch(s*=k,x.kind){case F.PARALLEL:o++;break;case F.VECTORIZED:i++;break;case F.UNROLLED:a++;break;case F.THREAD_BINDING:u++,x.threadTag&&x.threadTag.startsWith("threadIdx")?c*=k:x.threadTag&&x.threadTag.startsWith("blockIdx")&&(h*=k);break;default:p++;break}}let f={math:0,extern:0};r._countExpr(e.value,f);let d=t.map(x=>x.loopVar.name),m=[{buffer:e.buffer,indices:e.indices}];r._collectLoads(e.value,m);let g=0,y=0,_=0,b=0;for(let x of m){let k=x.buffer&&x.buffer.sizeInBytes?x.buffer.sizeInBytes():0;k>0&&(b+=k);let A=x.indices&&x.indices.length>0?x.indices[x.indices.length-1]:null;A&&A.type==="VariableNode"?g++:y++;let N=new Set;if(x.indices)for(let v of x.indices)r._collectVars(v,N);for(let v of d)N.has(v)||_++}let w=t.length>0?t[t.length-1]:null,T=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:h,underReduction:n>0?1:0,numMathOps:f.math,numExternCalls:f.extern,numReads:m.length-1,numWrites:1,stride1Accesses:g,stridedAccesses:y,reuseCount:_,touchedBytes:b,arithmeticIntensity:b>0?(f.math+f.extern)/b:0,vectorized:i>0?1:0,parallelized:o+u>0?1:0,innermostExtent:T};return by.map(x=>S[x]||0)}static _countExpr(e,t){if(!(!e||typeof e!="object"))switch(e.type){case"MathOpNode":t.math++,r._countExpr(e.a,t),e.b&&r._countExpr(e.b,t);break;case"CompareNode":t.math++,r._countExpr(e.a,t),r._countExpr(e.b,t);break;case"CallExternNode":t.extern++;for(let n of e.args)r._countExpr(n,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&&r._collectLoads(e.a,t),e.b&&r._collectLoads(e.b,t),e.args)for(let n of e.args)r._collectLoads(n,t)}}static _collectVars(e,t){if(!(!e||typeof e!="object")){if(e.type==="VariableNode"){t.add(e.name);return}if(e.a&&r._collectVars(e.a,t),e.b&&r._collectVars(e.b,t),e.args)for(let n of e.args)r._collectVars(n,t);if(e.indices)for(let n of e.indices)r._collectVars(n,t)}}static _visitIterative(e,t){let n=[{node:e,action:"enter"}];for(;n.length>0;){let{node:s,action:o}=n.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 F.PARALLEL:t.numParallelLoops++;break;case F.VECTORIZED:t.numVectorizedLoops++;break;case F.UNROLLED:t.numUnrolledLoops++;break;case F.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,n.push({node:null,action:"leave_for"}),n.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)r._visitBuffer(i.buffer,t);for(let i of s.writes)r._visitBuffer(i.buffer,t);n.push({node:s.body,action:"enter"}),s.initBody&&n.push({node:s.initBody,action:"enter"});break;case"SeqNode":for(let i=s.stmts.length-1;i>=0;i--)n.push({node:s.stmts[i],action:"enter"});break;case"AllocateNode":r._visitBuffer(s.buffer,t),n.push({node:s.body,action:"enter"});break;case"IfThenElseNode":s.elseBody&&n.push({node:s.elseBody,action:"enter"}),n.push({node:s.thenBody,action:"enter"});break;case"LetStmtNode":n.push({node:s.body,action:"enter"});break;case"BufferStoreNode":t.numBufferWrites++,r._visitBuffer(s.buffer,t),r._checkStride(s.buffer,s.indices,t),r._visitExpr(s.value,t);break}}}}static _visitBuffer(e,t){if(!e||t.buffersSeen.has(e))return;t.buffersSeen.add(e);let n=e.sizeInBytes();n>0&&(t.totalBufferBytes+=n)}static _visitExpr(e,t){if(e)switch(e.type){case"MathOpNode":t.numMathOps++,r._visitExpr(e.a,t),e.b&&r._visitExpr(e.b,t);break;case"CallExternNode":t.numExternCalls++;for(let n of e.args)r._visitExpr(n,t);break;case"BufferLoadNode":t.numBufferReads++,r._checkStride(e.buffer,e.indices,t);break;case"CompareNode":t.numMathOps++,r._visitExpr(e.a,t),r._visitExpr(e.b,t);break;default:break}}static _checkStride(e,t,n){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"?n.strideOneAccesses++:n.nonStrideOneAccesses++}}});function $B(r,e){if(r.length===0)return 0;let t=0;for(let n of r)t+=e[n];return t/r.length}function ST(r,e,t,n,s,o,i){let a=t.length;if(n>=s||a<o*2)return{leaf:$B(t,e)};let u=r[0].length,p=0,c=0,h=new Set(t);for(let g of t){let y=e[g];p+=y,c+=y*y}let f=null,d=null;for(let g=0;g<u;g++){let y=i[g].filter(w=>h.has(w)),_=0,b=0;for(let w=0;w<y.length-1;w++){let T=e[y[w]];_+=T,b+=T*T;let S=r[y[w]][g],x=r[y[w+1]][g];if(S===x)continue;let k=w+1,A=a-k;if(k<o||A<o)continue;let N=p-_,v=c-b,O=b-_*_/k+(v-N*N/A);(!f||O<f.cost)&&(f={f:g,thr:(S+x)/2,sIdx:w,cost:O},d=y)}}if(!f)return{leaf:$B(t,e)};let m=d;return{f:f.f,thr:f.thr,left:ST(r,e,m.slice(0,f.sIdx+1),n+1,s,o,i),right:ST(r,e,m.slice(f.sIdx+1),n+1,s,o,i)}}function hF(r){let e=r.length,t=e>0?r[0].length:0,n=new Array(t),s=[];for(let o=0;o<e;o++)s.push(o);for(let o=0;o<t;o++)n[o]=s.slice().sort((i,a)=>r[i][o]-r[a][o]);return n}function FB(r,e){for(;r&&r.leaf===void 0;)r=e[r.f]<=r.thr?r.left:r.right;return r?r.leaf:0}var af,zB=M(()=>{"use strict";l($B,"leafValue");l(ST,"buildTree");l(hF,"buildFeatureOrder");l(FB,"predictTree");af=class r{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 n=e.length;if(n===0)return;let s=0;for(let p of t)s+=p;s/=n;let o=new Array(n).fill(s),i=[];for(let p=0;p<n;p++)i.push(p);let a=hF(e),u=[];for(let p=0;p<this.numTrees;p++){let c=new Array(n);for(let f=0;f<n;f++)c[f]=t[f]-o[f];let h=ST(e,c,i,0,this.maxDepth,this.minSamples,a);for(let f=0;f<n;f++)o[f]+=this.lr*FB(h,e[f]);u.push(h)}this.trees=u,this.base=s}predict(e){if(!this.trees)return 0;let t=this.base;for(let n of this.trees)t+=this.lr*FB(n,e);return t}serialize(){return{trees:this.trees,base:this.base,lr:this.lr}}static deserialize(e){let t=new r({lr:e.lr});return t.trees=e.trees,t.base=e.base,t}}});function VB(r){let e=r[0].length,t=new Array(e+1).fill(0);for(let n of r)for(let s=0;s<e;s++){let o=n[s]||0;gF.has(s)?o>t[s]&&(t[s]=o):t[s]+=o}for(let n of yF)n<e&&(t[n]/=r.length);return t[e]=r.length,t}var dF,mF,gF,yF,kT,_F,wy,Nl,Al,vT=M(()=>{"use strict";xT();zB();dF=new Set(["depth","threadBlockSize","gridSize","underReduction","vectorized","parallelized","innermostExtent"]),mF=new Set(["arithmeticIntensity"]),gF=new Set(by.map((r,e)=>dF.has(r)?e:-1).filter(r=>r>=0)),yF=new Set(by.map((r,e)=>mF.has(r)?e:-1).filter(r=>r>=0));l(VB,"aggregateStatements");kT=class{static{l(this,"CostEstimate")}constructor(e,t){this.score=e,this.breakdown=t}},_F={parallelism:2,vectorization:1.5,memoryCoalescing:2,occupancy:1,arithmeticIntensity:1,loopOverhead:-.5,codeSize:-.3},wy=class{static{l(this,"AnalyticalCostModel")}constructor(e,t={}){this.target=e,this._weights={..._F,...e&&e.costModelWeights?e.costModelWeights:{},...t.weights||{}}}estimate(e){let t=Ra.extract(e);return this.estimateFromFeatures(t)}score(e){return this.estimate(e).score}estimateFromFeatures(e){let t=this._scoreParallelism(e),n=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*n+p.memoryCoalescing*s+p.occupancy*o+p.arithmeticIntensity*i+p.loopOverhead*a+p.codeSize*u;return new kT(c,{parallelism:t,vectorization:n,memoryCoalescing:s,occupancy:o,arithmeticIntensity:i,loopOverhead:a,codeSize:u})}_scoreParallelism(e){if(this.target.isGPU()){let n=e.threadBlockSize*e.gridSize,s=this.target.maxParallelism();return Math.min(1,n/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,n=Math.ceil(e.threadBlockSize/t),s=Math.floor(this.target.maxThreadsPerBlock/t);return Math.min(1,n/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}},Nl=class r{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?af.deserialize(e):null,this._X=[],this._Y=[]}addSample(e,t){!e||e.length===0||Number.isFinite(t)&&(this._X.push(VB(e)),this._Y.push(t))}train(){if(this._X.length===0)return;let e=new af(this.opts);e.fit(this._X,this._Y),this._gbt=e}predict(e){return!this._gbt||!e||e.length===0?0:this._gbt.predict(VB(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 r(e&&e.gbt?e.gbt:null)}},Al=class{static{l(this,"GuidedCostModel")}constructor(e,t,n={}){this.analytical=e,this.learned=t,this.confidenceSamples=n.confidenceSamples??8}_learnedConfident(){return this.learned&&this.learned.trained&&this.learned.sampleCount>=this.confidenceSamples}score(e){return this._learnedConfident()?this.learned.predict(Ra.extractStatements(e)):this.analytical.score(e)}}});function*GB(r){let e=[r];for(;e.length>0;){let t=e.pop();if(t){if(yield t,t.body&&e.push(t.body),t.stmts)for(let n of t.stmts)e.push(n);t.thenBody&&e.push(t.thenBody),t.elseBody&&e.push(t.elseBody),t.initBody&&e.push(t.initBody)}}}function uf(r,e){for(let t of GB(r))if(t.type==="BlockNode"&&t.name===e)return t;return null}function io(r){let e=[];for(let t of GB(r))t.type==="BlockNode"&&e.push(t.name);return e}function lf(r,e){let t=dt(r,e);if(!t)return{spatial:0,reduction:0,reads:0,hasReduction:!1};let n=0,s=0;for(let o of t.loops)t.reductionLoopVars.has(o.loopVar.name)?s++:n++;return{spatial:n,reduction:s,reads:t.readBuffers.length,hasReduction:t.hasReduction}}var Ol=M(()=>{"use strict";ro();l(GB,"walkStmts");l(uf,"findBlock");l(io,"collectAllBlockNames");l(lf,"analyzeBlockStructure")});function qB(r){return r.isGPU()?wF:bF}function NT(r){let e=0,t=0;for(let[n,s]of r.order)n==="S"?e=Math.max(e,s+1):t=Math.max(t,s+1);return{spatialLevels:e,reductionLevels:t}}var bF,wF,jB,AT=M(()=>{"use strict";bF={name:"mlt_cpu",order:[["S",0],["S",1],["S",2],["S",3],["R",0]],roles:{S0:"parallelize",S3:"vectorize"}},wF={name:"mlt_gpu",order:[["S",0],["S",1],["S",2],["R",0]],roles:{S0:"blockIdx",S1:"threadIdx",S2:"unroll"}},jB={name:"ssrsrs_cpu",order:[["S",0],["S",1],["R",0],["S",2],["R",1],["S",3]],roles:{S0:"parallelize",S3:"vectorize",R1:"unroll"}};l(qB,"getTileStructure");l(NT,"levelCounts")});var os,vr,Ty=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)]}},vr=class{static{l(this,"ScheduleSketch")}constructor(e,t,n){this.name=e,this.variables=t,this._apply=n}instantiate(e){return(t,n,s)=>{this._apply(t,n,s,e)}}sampleParams(e){let t={};for(let n of this.variables)t[n.name]=n.sample(e);return t}}});function TF(r){let e=[];for(let t=1;t*t<=r;t++){if(r%t!==0)continue;e.push(t);let n=r/t;n!==t&&e.push(n)}return e.sort((t,n)=>t-n),e}function xF(r,e){let t=Math.min(r.length,e.length);for(let n=0;n<t;n++)if(r[n]!==e[n])return r[n]-e[n];return r.length-e.length}function SF(r,e){let t=new Set,n=[];r.sort(xF);for(let i of r){let a=i.join(",");t.has(a)||(t.add(a),n.push(i))}if(n.length<=e)return n;let s=new Set,o=[];for(let i=0;i<e;i++){let a=Math.floor(i*(n.length-1)/(e-1));s.has(a)||(s.add(a),o.push(n[a]))}return o}function OT(r,e,t=48){if(!Number.isFinite(r)||r<1||e<=1)return[[r]];let n=[],s=t*8,o=l((i,a,u)=>{if(!(n.length>=s)){if(a===e-1){n.push([...u,i]);return}for(let p of TF(i))if(u.push(p),o(i/p,a+1,u),u.pop(),n.length>=s)return}},"rec");return o(r,0,[]),SF(n,t)}var UB=M(()=>{"use strict";l(TF,"divisorsOf");l(xF,"cmpTuple");l(SF,"selectDiverse");l(OT,"enumerateFactorizations")});function xy(r){return r.extent&&r.extent.type==="IntImmNode"?r.extent.value:null}function WB(r,e,t,n){let s=[],o=t;for(let i=0;i<n.length-1;i++){let a=r.getLoops(e).find(h=>h.loopVar.name===o);if(!a)return null;let u=1;for(let h=i+1;h<n.length;h++)u*=n[h];let[p,c]=r.split(a,u);s.push(p.loopVar.name),o=c.loopVar.name}return s.push(o),s}function kF(r,e,t,n,s){let o=l(i=>r.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"?n:s;if(u==="parallelize"){let c=p[0]&&a<p[0].length?o(p[0][a]):null;c&&r.parallelize(c)}else if(u==="vectorize"){let c=p[p.length-1],h=c&&a<c.length?o(c[a]):null;h&&r.vectorize(h)}else if(u==="unroll")for(let c of p){let h=a<c.length?o(c[a]):null;h&&r.unroll(h)}else(u==="blockIdx"||u==="threadIdx")&&p.forEach((c,h)=>{if(h>=KB.length)return;let f=a<c.length?o(c[a]):null;f&&r.bindThread(f,`${u}.${KB[h]}`)})}}function HB(r,e,t,n,s,o){let i=[];for(let h=0;h<n.length;h++){let f=WB(r,e,n[h],o[`s${h}`]);if(!f)return;i.push(f)}let a=[];for(let h=0;h<s.length;h++){let f=WB(r,e,s[h],o[`r${h}`]);if(!f)return;a.push(f)}let u=[];for(let[h,f]of t.order){let d=h==="S"?i:a;for(let m of d)f<m.length&&u.push(m[f])}let p=new Map(r.getLoops(e).map(h=>[h.loopVar.name,h])),c=u.map(h=>p.get(h)).filter(Boolean);c.length>=2&&r.reorder(...c),kF(r,e,t,i,a)}function YB(r){let e=r.loops.filter(n=>!r.reductionLoopVars.has(n.loopVar.name)),t=r.loops.filter(n=>r.reductionLoopVars.has(n.loopVar.name));return{spatialLoops:e,reductionLoops:t}}function JB(r,e,t,n){let s=[];return r.forEach((o,i)=>s.push(new os(`s${i}`,OT(xy(o),t)))),e.forEach((o,i)=>s.push(new os(`r${i}`,OT(xy(o),n)))),s}function ZB(r,e){let{spatialLevels:t,reductionLevels:n}=NT(e),{spatialLoops:s,reductionLoops:o}=YB(r);if(s.length===0)return null;for(let p of[...s,...o])if(xy(p)===null)return null;let i=JB(s,o,t,n),a=s.map(p=>p.loopVar.name),u=o.map(p=>p.loopVar.name);return new vr(e.name,i,(p,c,h,f)=>{HB(p,c,e,a,u,f)})}function QB(r,e){let{spatialLevels:t,reductionLevels:n}=NT(e),{spatialLoops:s,reductionLoops:o}=YB(r);if(s.length===0||o.length===0)return null;for(let p of[...s,...o])if(xy(p)===null)return null;let i=JB(s,o,t,n),a=s.map(p=>p.loopVar.name),u=o.map(p=>p.loopVar.name);return new vr(e.name,i,(p,c,h,f)=>{p.decomposeReduction(c),HB(p,`${c}_upd`,e,a,u,f)})}var KB,XB=M(()=>{"use strict";Ty();UB();AT();KB=["x","y","z"];l(xy,"staticExtent");l(WB,"multiLevelSplit");l(kF,"applyRoles");l(HB,"tileBlock");l(YB,"splitLoops");l(JB,"tilingVariables");l(ZB,"createMultiLevelTilingSketch");l(QB,"createSSRSRSTilingSketch")});function tR(r){return Math.min(r&&r.maxThreadsPerBlock||256,256)}function NF(r){let e=new Set;for(let t=2;t*t<=r;t++){if(r%t!==0)continue;e.add(t);let n=r/t;n>1&&n<r&&e.add(n)}return[...e].sort((t,n)=>t-n)}function nR(r){return new vr("fused",[],(e,t,n,s)=>{e.fuseConsumer(t,r);let o=e.getLoops(t);o.length>0&&e.parallelize(o[0])})}function rR(r){let e=r.loops.filter(i=>r.reductionLoopVars.has(i.loopVar.name));if(e.length!==1)return null;let t=e[0],n=t.extent&&t.extent.type==="IntImmNode"?t.extent.value:null;if(n===null)return null;let s=NF(n);if(s.length===0)return null;let o=t.loopVar.name;return new vr("rfactor",[new os("rf_factor",s)],(i,a,u,p)=>{i.rfactor(a,o,p.rf_factor);let c=i.getLoops(`${a}_rf_p`);c.length>0&&i.parallelize(c[0])})}function sR(){return new vr("elementwise_cpu",[new os("vector_width",vF)],(r,e,t,n)=>{let s=r.getLoops(e);if(s.length===0)return;if(s.length===1){let a=s[0].extent;if(a.type==="IntImmNode"&&a.value>=n.vector_width*2){let[u,p]=r.split(s[0],n.vector_width);r.parallelize(u),r.vectorize(p)}else r.parallelize(s[0]);return}r.parallelize(s[0]);let o=s[s.length-1],i=o.extent;if(i.type==="IntImmNode"&&i.value>=n.vector_width){let[,a]=r.split(o,n.vector_width);r.vectorize(a)}})}function oR(){return new vr("elementwise_gpu",[new os("block_size",eR)],(r,e,t,n)=>{let s=r.getLoops(e);if(s.length===0)return;let o=s[0];for(let c=1;c<s.length;c++){let f=r.getLoops(e).find(d=>d.loopVar.name===s[c].loopVar.name);f&&o.body===f&&(o=r.fuseLoops(o,f))}let i=o.extent;if(i.type!=="IntImmNode"){r.bindThread(o,"threadIdx.x");return}let a=i.value,u=Math.min(t.maxThreadsPerBlock,1024);if(o0(r.func)&&a<=u){r.bindThread(o,"threadIdx.x");return}let p=Math.min(n.block_size,tR(t));if(a>p){let[c,h]=r.split(o,p);r.bindThread(c,"blockIdx.x"),r.bindThread(h,"threadIdx.x")}else r.bindThread(o,"threadIdx.x")})}function iR(){return new vr("reduction_cpu",[],(r,e,t,n)=>{let s=r.getLoops(e);s.length>0&&r.parallelize(s[0])})}function aR(){return new vr("reduction_gpu",[new os("block_size",eR)],(r,e,t,n)=>{let s=r.getLoops(e);if(s.length===0)return;let o=dt(r.func,e),i=s.filter(c=>!o||!Zu(c,o));if(i.length===0)return;let a=i[0];for(let c=1;c<i.length;c++){let f=r.getLoops(e).find(d=>d.loopVar.name===i[c].loopVar.name);f&&a.body===f&&(a=r.fuseLoops(a,f))}let u=a.extent;if(u.type!=="IntImmNode"){r.bindThread(a,"threadIdx.x");return}let p=Math.min(n.block_size,tR(t));if(u.value>p){let[c,h]=r.split(a,p);r.bindThread(c,"blockIdx.x"),r.bindThread(h,"threadIdx.x")}else r.bindThread(a,"threadIdx.x")})}var eR,vF,uR=M(()=>{"use strict";ro();Ty();eR=[32,64,128,256,512,1024],vF=[1,2,4,8,16];l(tR,"gpuThreadCap");l(NF,"reductionDivisors");l(nR,"createFusedTilingSketch");l(rR,"createRfactorSketch");l(sR,"createElementwiseCPUSketch");l(oR,"createElementwiseGPUSketch");l(iR,"createReductionCPUSketch");l(aR,"createReductionGPUSketch")});function Sy(r){return ym(r,Sy,(e,t,n)=>{switch(e.type){case"PrimFunc":return t.name=e.name,t.params=e.params,t.body=n(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=n(e.body),t._setChild("body",t.body),t;case"LetStmtNode":return t.variable=e.variable,t.value=n(e.value),t.body=n(e.body),t._setChild("body",t.body),t;case"WhileNode":return t.condVar=e.condVar,t.condBody=n(e.condBody),t.loopBody=n(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(n):typeof o=="object"&&o!==null&&o.type?t[s]=n(o):t[s]=o}return t}})}function Lo(r){return Sy(r)}function Ma(r){return Sy(r)}function BT(r){return Sy(r)}function lR(r,e,t){let n=t.get(e);if(!n)return null;let s=[],o=n._parent;for(;o&&o!==r;)o.type==="ForNode"&&s.push(o),o=o._parent;s.reverse();let i=BT(n);for(let p=s.length-1;p>=0;p--){let c=s[p],h=new de(c.loopVar,BT(c.min),BT(c.extent),c.kind,i,c.threadTag);h._setChild("body",i),i=h}let a=new Map;for(let p of n.reads)a.set(p.buffer.name,p.buffer);for(let p of n.writes)a.set(p.buffer.name,p.buffer);let u=[];for(let p of r.params)a.has(p.name)&&u.push(p);return new Jo("__tune_"+e,u,i,a,[])}var cf=M(()=>{"use strict";De();Ww();l(Sy,"deepCloneIR");l(Lo,"clonePrimFunc");l(Ma,"cloneTensorIR");l(BT,"cloneBlockSubtree");l(lR,"extractBlockMini")});function ET(r,e){let t=[];for(let n=1;n<=e;n*=2)n>=r&&t.push(n);return t}function hf(r){if(!r)return null;let e=[];for(let t of r){if(!t||t.type!=="VariableNode")return null;e.push(t.name)}return e}function AF(r,e){let t=[r];for(;t.length>0;){let n=t.pop();if(!(!n||typeof n!="object")){if(n.type==="BufferStoreNode"&&n.buffer&&n.buffer.name===e&&n.value&&n.value.type==="MathOpNode"&&n.value.op==="+")return n;if(n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody)}}return null}function ff(r){if(!r||r.broadcastDims||r.offset&&r.offset!==0)return!1;let e=r.shape,t=r.strides;if(!t||t.length!==e.length)return!1;let n=1;for(let s=e.length-1;s>=0;s--){if(typeof e[s]!="number"||e[s]<=0||t[s]!==n)return!1;n*=e[s]}return!0}function LT(r,e){if(!dt(r,e))return null;let n=uf(r.body,e);if(!n||n.reads.length<2||n.writes.length<1)return null;let s=n.writes[0].buffer;if(!s||s.shape.length<2)return null;let o=s.shape.length,i=AF(n.body,s.name);if(!i)return null;let a=hf(i.indices);if(!a||a.length!==o)return null;let u=i.value,p=l(x=>x&&x.type==="BufferLoadNode"&&x.buffer&&x.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 h=[c.a,c.b];if(!h.every(x=>x&&x.type==="BufferLoadNode"&&x.buffer))return null;let f=a[o-1],d=a.slice(0,o-1),m=null,g=null;for(let x of h){let k=hf(x.indices);if(!k||k.length!==o)continue;let A=!0;for(let N=0;N<o-1;N++)if(k[N]!==d[N]){A=!1;break}if(A){m=x,g=k[o-1];break}}if(!m||g==null)return null;let y=h[0]===m?h[1]:h[0],_=hf(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 x=0;x<o-1;x++)if(b.shape[x]!==s.shape[x])return null;let T=s.shape[o-1],S=b.shape[o-1];if(![T,S].every(x=>typeof x=="number"&&x>0))return null;if(_.length===2&&w.shape.length===2){let x;if(_[0]===g&&_[1]===f)x=!1;else if(_[0]===f&&_[1]===g)x=!0;else return null;if(x){if(w.shape[0]!==T||w.shape[1]!==S)return null}else if(w.shape[0]!==S||w.shape[1]!==T)return null;let k=1;for(let A=0;A<o-1;A++){let N=s.shape[A];if(typeof N!="number"||N<=0)return null;k*=N}return o===2?{A:b,B:w,C:s,M:k,N:T,K:S,transB:x,batch:1}:!ff(b)||!ff(s)?null:{A:new me(b.name,[k,S],b.dtype,b.scope),B:w,C:new me(s.name,[k,T],s.dtype,s.scope),M:k,N:T,K:S,transB:x,batch:1}}if(o>=3&&_.length===o&&w.shape.length===o){let x=a.slice(0,o-2);for(let B=0;B<o-2;B++)if(_[B]!==x[B])return null;let k;if(_[o-2]===g&&_[o-1]===f)k=!1;else if(_[o-2]===f&&_[o-1]===g)k=!0;else return null;for(let B=0;B<o-2;B++)if(w.shape[B]!==s.shape[B])return null;if(k){if(w.shape[o-2]!==T||w.shape[o-1]!==S)return null}else if(w.shape[o-2]!==S||w.shape[o-1]!==T)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(!ff(b)||!ff(w)||!ff(s))return null;let v=new me(b.name,[N,A,S],b.dtype,b.scope),O=new me(w.name,k?[N,T,S]:[N,S,T],w.dtype,w.scope),R=new me(s.name,[N,A,T],s.dtype,s.scope);return{A:v,B:O,C:R,M:A,N:T,K:S,transB:k,batch:N}}return null}function dR(r,e,t=32){let n=r.maxThreadsPerBlock||1024,s=r.warpSize||32,o=r.sharedMemoryBytes||49152,i=r.registersPerThread||255,a=4,{M:u,N:p,K:c}=e,h=ET(2,8),f=ET(2,s/2),d=ET(s/4,s/2),m=2*s,g=Math.min(n,8*s),y=4*s,_=new Set,b=[];for(let w of h)for(let T of h)if(!(w*T>y)&&!(w*T+w+T+s>i))for(let S of f)for(let x of f){let k=x*S;if(k<m||k>g||k%s!==0)continue;let A=w*S,N=T*x;if(!(A>u||N>p||A>y||N>y))for(let v of d){if(v>c||(A*v+v*N)*a>o)continue;let R=`${A}_${N}_${v}_${w}_${T}`;_.has(R)||(_.add(R),b.push({BM:A,BN:N,BK:v,TM:w,TN:T,threads:k}))}}return b.sort((w,T)=>fR(T,s)-fR(w,s)),b.slice(0,t)}function fR(r,e){let t=r.TM*r.TN,n=-Math.abs(r.TM-r.TN),s=-Math.abs(Math.log2(r.BM)-Math.log2(r.BN)),o=r.threads>=4*e&&r.threads<=8*e?1:0,i=r.BK===e/4?1:0;return o*100+t*4+n*6+s*4+i}function ky(r,e){if(!e)return null;let t=dR(r,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 OF(r,e,t,n){let s=Ma(r.storeValue);return ti(s,o=>o.type==="BufferLoadNode"&&o.buffer?o.buffer.name===r.inputName?e:r.scalarConsts.has(o.buffer.name)?Ma(r.scalarConsts.get(o.buffer.name)):o:o.type==="VariableNode"?o.name===r.iv0?Ma(t):o.name===r.iv1?Ma(n):o:o)}function CT(r,e,t=null){let{A:n,B:s,C:o,M:i,N:a,K:u,transB:p}=r,c=r.batch||1,{BM:h,BN:f,BK:d,TM:m,TN:g}=e,y=f/g,_=h/m,b=y*_,w=Math.ceil(u/d),T=h*d,S=d*f,x=Math.ceil(T/b),k=Math.ceil(S/b),A=i%h!==0,N=a%f!==0,v=u%d!==0,O=T%b!==0,R=S%b!==0,B=new me("rb_As",[T],"f32","shared"),E=new me("rb_Bs",[S],"f32","shared"),z=new me("rb_acc",[m*g],"f32","local"),D=new me("rb_af",[m],"f32","local"),J=new me("rb_bf",[g],"f32","local"),j=Vt("rb_bx"),U=Vt("rb_by"),P=Vt("rb_tx"),V=Vt("rb_ty"),he=Vt("rb_tid"),le=Vt("rb_brow"),Ie=Vt("rb_bcol"),G=Vt("rb_k0"),K=Vt("rb_bz"),ie=c>1,ze=l((wt,Fs)=>ie?[K,wt,Fs]:[wt,Fs],"gIdx"),Se=l((wt,Fs)=>Pn(Nr(wt,Ze(g)),Fs),"accIdx"),Qe=Vt("rb_im"),It=Vt("rb_in"),Jt=Co(Qe,m,Co(It,g,new Y(z,[Se(Qe,It)],RT()))),Gt=Vt("rb_la"),tt=Vt("rb_aidx"),ye=Pn(le,cR(tt,Ze(d))),He=Pn(G,pR(tt,Ze(d))),gn=new q(n,ze(ye,He));if(A||v){let wt=A?Ls(ye,Ze(i)):null;v&&(wt=wt?MT(wt,Ls(He,Ze(u))):Ls(He,Ze(u))),gn=new xe(wt,gn,RT())}let jt=new Y(B,[tt],gn);O&&(jt=new xe(Ls(tt,Ze(T)),jt));let Zt=IT(Gt,x,new Ye(tt,Pn(he,Nr(Gt,Ze(b))),jt)),Qt=Vt("rb_lb"),on=Vt("rb_bidx"),Jn=Pn(G,cR(on,Ze(f))),Bn=Pn(Ie,pR(on,Ze(f))),yr=new q(s,p?ze(Bn,Jn):ze(Jn,Bn));if(v||N){let wt=v?Ls(Jn,Ze(u)):null;N&&(wt=wt?MT(wt,Ls(Bn,Ze(a))):Ls(Bn,Ze(a))),yr=new xe(wt,yr,RT())}let Zn=new Y(E,[on],yr);R&&(Zn=new xe(Ls(on,Ze(S)),Zn));let pt=IT(Qt,k,new Ye(on,Pn(he,Nr(Qt,Ze(b))),Zn)),Qn=Vt("rb_kk"),mt=Vt("rb_fi"),Or=Vt("rb_fj"),hs=Co(mt,m,new Y(D,[mt],new q(B,[Pn(Nr(Pn(Nr(V,Ze(m)),mt),Ze(d)),Qn)]))),$s=Co(Or,g,new Y(J,[Or],new q(E,[Pn(Nr(Qn,Ze(f)),Pn(Nr(P,Ze(g)),Or))]))),Br=Vt("rb_mi"),Xn=Vt("rb_ni"),_r=Co(Br,m,Co(Xn,g,new Y(z,[Se(Br,Xn)],Pn(new q(z,[Se(Br,Xn)]),Nr(new q(D,[Br]),new q(J,[Xn])))))),fo=Co(Qn,d,new te([hs,$s,_r])),Rr=Vt("rb_kt"),Mr=new Ye(G,Nr(Rr,Ze(d)),new te([Zt,pt,new On,fo,new On])),ds=IT(Rr,w,Mr),Qr=Vt("rb_wm"),Mi=Vt("rb_wn"),Vo=Pn(Pn(le,Nr(V,Ze(m))),Qr),Go=Pn(Pn(Ie,Nr(P,Ze(g))),Mi),Ii=new q(z,[Se(Qr,Mi)]),ec=t?t.outBuffer:o,Fa=t?OF(t,Ii,Vo,Go):Ii,Ei=new Y(ec,ze(Vo,Go),Fa);if(A||N){let wt=A?Ls(Vo,Ze(i)):null;N&&(wt=wt?MT(wt,Ls(Go,Ze(a))):Ls(Go,Ze(a))),Ei=new xe(wt,Ei)}let Li=Co(Qr,m,Co(Mi,g,Ei)),ho=new te([Jt,ds,Li]),Ci=new Ve(z,"local",new Ve(D,"local",new Ve(J,"local",ho))),jo=new Ye(he,Pn(Nr(V,Ze(y)),P),new Ye(le,Nr(U,Ze(h)),new Ye(Ie,Nr(j,Ze(f)),Ci))),mo=Math.ceil(a/f),qo=Math.ceil(i/h),Di=pf(U,"blockIdx.y",qo,pf(j,"blockIdx.x",mo,pf(V,"threadIdx.y",_,pf(P,"threadIdx.x",y,jo))));return ie&&(Di=pf(K,"blockIdx.z",c,Di)),new Ve(B,"shared",new Ve(E,"shared",Di))}function BF(r){let e=new os("config_index",r.map((n,s)=>s)),t=new vr("matmul_register_block_gpu",[e],(n,s,o,i)=>{let a=LT(n.func,s);if(!a)return;let u=r[i.config_index];if(!u)return;let p=CT(a,u);n.func.body=p,n.func._setChild&&n.func._setChild("body",p),n.func.gpuRegisterBlocked=!0});return t.configs=r,t.enumerate=()=>r.map((n,s)=>({config_index:s})),t}function RF(r){let e=io(r.body),t=null;for(let s of e){let o=lf(r,s);if(o.hasReduction&&o.spatial>=2&&o.reads>=2){if(t)return null;t=s}}if(!t)return null;let n=LT(r,t);if(!n)return null;for(let s of e){if(s===t)continue;let o=dt(r,s);if(!o||o.hasReduction||o.readBuffers.length>0)return null;for(let i of o.writeBuffers)if(i!==n.C.name)return null}return{reductionBlock:t,dims:n}}function MF(r){let e=new Map,t=new Map;Ss(r,n=>{if(n.type==="BufferStoreNode"&&n.buffer){let s=n.buffer.name;t.set(s,(t.get(s)||0)+1),(!n.indices||n.indices.length===0)&&n.value&&(n.value.type==="FloatImmNode"||n.value.type==="IntImmNode")&&e.set(s,n.value)}});for(let n of[...e.keys()])(t.get(n)||0)!==1&&e.delete(n);return e}function IF(r,e){let t=null;return Ss(r,n=>{if(n.type==="BufferStoreNode"&&n.buffer&&n.buffer.name===e)return t=n,ad}),t}function EF(r,e){let t=null,n=0;return Ss(r,s=>{s.type==="BufferLoadNode"&&s.buffer&&s.buffer.name===e&&(n++,t=s)}),n===1?t:null}function mR(r){let e=io(r.body),t=null;for(let y of e){let _=lf(r,y);if(_.hasReduction&&_.spatial>=2&&_.reads>=2){if(t)return null;t=y}}if(!t)return null;let n=LT(r,t);if(!n)return null;let s=n.C.name,o=new Set;for(let y of e){let _=dt(r,y);if(!_)return null;for(let b of _.writeBuffers)o.add(b)}let i=[];for(let y of e){if(y===t)continue;let _=dt(r,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:n,epilogue:null};if(i.length!==1||(n.batch||1)!==1)return null;let a=i[0];if(a.info.writeBuffers.length!==1)return null;let u=a.info.writeBuffers[0],p=MF(r.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 h=uf(r.body,a.name);if(!h)return null;let f=IF(h.body,u);if(!f)return null;let d=hf(f.indices);if(!d||d.length!==2)return null;let m=EF(f.value,s);if(!m)return null;let g=hf(m.indices);return!g||g.length!==d.length||g.some((y,_)=>y!==d[_])?null:{reductionBlock:t,dims:n,epilogue:{outBuffer:f.buffer,storeValue:f.value,inputName:s,iv0:d[0],iv1:d[1],scalarConsts:p}}}function gR(r,e,t){let n=RF(r);if(!n)return null;let s=hR.get(r);if(s===void 0){let o=dR(t,n.dims);s=o.length>0?BF(o):null,hR.set(r,s)}return s?e===n.reductionBlock?[s]:[]:null}var Ze,RT,Vt,Pn,Nr,cR,pR,Ls,MT,IT,Co,pf,hR,vy=M(()=>{"use strict";De();xo();ro();Ty();Ol();Gr();cf();Ze=l(r=>new L(r),"I"),RT=l(()=>new fe(0),"FZERO"),Vt=l(r=>new Ne(r,"i32"),"IV"),Pn=l((r,e)=>In("+",r,e),"ADD"),Nr=l((r,e)=>In("*",r,e),"MUL"),cR=l((r,e)=>In("//",r,e),"DIV"),pR=l((r,e)=>In("%",r,e),"MOD"),Ls=l((r,e)=>new qe("lt",r,e),"LT"),MT=l((r,e)=>new C("&&",r,e),"AND"),IT=l((r,e,t)=>new de(r,Ze(0),Ze(e),F.SERIAL,t),"forS"),Co=l((r,e,t)=>new de(r,Ze(0),Ze(e),F.UNROLLED,t),"forU"),pf=l((r,e,t,n)=>new de(r,Ze(0),Ze(t),F.THREAD_BINDING,n,e),"forT");l(ET,"pow2Range");l(hf,"plainVars");l(AF,"findAccStore");l(ff,"isContiguousRowMajor");l(LT,"matmulTileDims");l(dR,"enumerateRegisterBlockConfigs");l(fR,"goodness");l(ky,"pickFixedConfig");l(OF,"foldEpilogue");l(CT,"buildRegisterBlockedMatmul");l(BF,"createMatmulRegisterBlockGPUSketch");l(RF,"analyzePureMatmul");l(MF,"collectScalarConstBuffers");l(IF,"findEpilogueStore");l(EF,"singleBufferLoad");l(mR,"analyzeMatmulEpilogue");hR=new WeakMap;l(gR,"richMatmulSketches")});function LF(r,e){let t=[r];for(;t.length;){let n=t.pop();if(!(!n||typeof n!="object")&&(n.type==="BlockNode"&&e(n),n.body&&t.push(n.body),n.initBody&&t.push(n.initBody),n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),Array.isArray(n.stmts)))for(let s of n.stmts)t.push(s)}}function mf(r){let e=[];LF(r.body,n=>{e.push({name:n.name,node:n,reads:(n.reads||[]).map(s=>s.buffer&&s.buffer.name).filter(Boolean),writes:(n.writes||[]).map(s=>s.buffer&&s.buffer.name).filter(Boolean)})});let t=new Map;for(let n of e)for(let s of n.reads)t.has(s)||t.set(s,[]),t.get(s).push(n.name);return{blocks:e,consumersByBuf:t,byName:new Map(e.map(n=>[n.name,n]))}}function DT(r){let e=[];for(let t of r){if(!t||t.type!=="VariableNode")return null;e.push(t.name)}return e}function df(r,e,t){if(!r||typeof r!="object")return t;r.type==="BufferLoadNode"&&r.buffer&&r.buffer.name===e&&t.push(r);for(let n of["a","b","expr","value","condition","thenBody","elseBody","body","initBody"])r[n]&&df(r[n],e,t);if(Array.isArray(r.args))for(let n of r.args)df(n,e,t);if(Array.isArray(r.indices))for(let n of r.indices)df(n,e,t);if(Array.isArray(r.stmts))for(let n of r.stmts)df(n,e,t);return t}function Bl(r,e){if(!r||typeof r!="object")return e;r.type==="BufferLoadNode"&&r.buffer&&r.buffer.name&&e.add(r.buffer.name);for(let t of["a","b","expr","value","condition","thenBody","elseBody","body","initBody"])r[t]&&Bl(r[t],e);if(Array.isArray(r.args))for(let t of r.args)Bl(t,e);if(Array.isArray(r.indices))for(let t of r.indices)Bl(t,e);if(Array.isArray(r.stmts))for(let t of r.stmts)Bl(t,e);return e}function Ny(r,e,t,n){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=n(r,t),p=n(r,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 h=new Set(s.reads);if(Bl(s.node.body,h),Bl(s.node.initBody,h),h.has(c.writes[0]))return null;let f=s.node.body&&s.node.body.type==="BufferStoreNode"?s.node.body:null,d=c.node.body&&c.node.body.type==="BufferStoreNode"?c.node.body:null;if(!f||!d)return null;let m=u.loops.filter(w=>!u.reductionLoopVars.has(w.loopVar.name)).map(w=>w.loopVar.name),g=DT(f.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 _=DT(d.indices);if(!_||_.join(",")!==y.join(","))return null;let b=df(d.value,o,[]);if(b.length===0)return null;for(let w of b){let T=DT(w.indices);if(!T||T.join(",")!==_.join(","))return null}return a}var PT=M(()=>{"use strict";l(LF,"walkBlocks");l(mf,"buildBlockDAG");l(DT,"indexVarNames");l(df,"collectLoads");l(Bl,"collectLoadedBuffers");l(Ny,"findFusibleConsumer")});function $T(r){return r.isGPU()?aR():iR()}function CF(r){return r.isGPU()?oR():sR()}function DF(r,e,t,n){let s=dt(r,e);if(!s)return[$T(t)];let o=[],i=ZB(s,qB(t));if(i&&o.push(i),t.kind===Ot.CPU){let a=QB(s,jB);a&&o.push(a);let u=rR(s);u&&o.push(u);let p=n?Ny(r,n,e,dt):null;p&&o.push(nR(p))}return o.push($T(t)),o}function FT(r,{priority:e=100}={}){Ay.some(t=>t.derive===r.derive&&t.matches===r.matches)||(Ay.push({matches:r.matches,derive:r.derive,priority:e}),Ay.sort((t,n)=>t.priority-n.priority))}function yR(r,e,t,n={}){if(n.richGpu&&t.isGPU()){let i=gR(r,e,t);if(i!==null)return i}if(t.kind!==Ot.CPU&&!t.isGPU())return[];let s=lf(r,e),o=n.dag||mf(r);for(let i of Ay)if(i.matches(s,t))return i.derive(r,e,t,o);return[]}var Ay,_R=M(()=>{"use strict";Ks();ro();Ol();AT();XB();uR();vy();PT();l($T,"reductionSketch");l(CF,"elementwiseSketch");l(DF,"deriveMultiLevel");Ay=[];l(FT,"registerSketchRule");FT({matches:l(r=>r.hasReduction&&r.spatial>=1&&r.reads>=2,"matches"),derive:DF},{priority:10});FT({matches:l(r=>r.hasReduction,"matches"),derive:l((r,e,t)=>[$T(t)],"derive")},{priority:20});FT({matches:l(()=>!0,"matches"),derive:l((r,e,t)=>[CF(t)],"derive")},{priority:30});l(yR,"deriveSketches")});function zT(r,e,t,n,s={}){return yR(r,e,t,s)}var bR=M(()=>{"use strict";_R();l(zT,"getSketchesForBlock")});function Oy(r){let e=new Map,t=[r];for(;t.length>0;){let n=t.pop();if(n){if(n.type==="BlockNode"&&e.set(n.name,n),n.body&&t.push(n.body),n.stmts)for(let s of n.stmts)t.push(s);n.thenBody&&t.push(n.thenBody),n.elseBody&&t.push(n.elseBody),n.initBody&&t.push(n.initBody)}}return e}function By(r,e,t,n=null){let s=n||Oy(r.body),o=s.get(e)||null,i=[];if(o){let a=[];for(let p of o.reads)a.push(`${p.buffer.shape.join("x")}:${p.buffer.dtype}`);for(let p of o.writes)a.push(`${p.buffer.shape.join("x")}:${p.buffer.dtype}`);i.push(a.join(","));let u=[];if(Yt(o.body,u),o.initBody&&Yt(o.initBody,u),i.push(u.join(";")),o.writes.length===1){let p=o.writes[0].buffer.name,c=[];for(let h of PF(s).get(p)||[]){if(h===o)continue;let f=[];Yt(h.body,f),h.initBody&&Yt(h.initBody,f),c.push(f.join(";"))}c.length>0&&(c.sort(),i.push(`consumers:${c.join("|")}`))}}return i.push(t.name),i.push(t.kind),$F(i.join("|"))}function PF(r){if(r.__readersByBuffer)return r.__readersByBuffer;let e=new Map;for(let t of r.values()){if(!t.reads)continue;let n=new Set;for(let s of t.reads)s.buffer&&n.add(s.buffer.name);for(let s of n){let o=e.get(s);o||(o=[],e.set(s,o)),o.push(t)}}return r.__readersByBuffer=e,e}function Yt(r,e){if(!(!r||typeof r!="object"))switch(r.type){case"BufferStoreNode":if(e.push("store"),r.indices)for(let t of r.indices)Yt(t,e);Yt(r.value,e);return;case"BufferLoadNode":if(e.push(`load:${r.buffer.shape.join("x")}:${r.buffer.dtype}`),r.indices)for(let t of r.indices)Yt(t,e);return;case"MathOpNode":e.push(`math:${r.op}`),Yt(r.a,e),r.b&&Yt(r.b,e);return;case"CallExternNode":e.push(`call:${r.externName}`);for(let t of r.args)Yt(t,e);return;case"CompareNode":e.push(`cmp:${r.direction}`),Yt(r.a,e),Yt(r.b,e);return;case"CastNode":e.push(`cast:${r.fromDtype}->${r.toDtype}`),Yt(r.expr,e);return;case"IfThenElseNode":Yt(r.condition,e),Yt(r.thenBody,e),r.elseBody&&Yt(r.elseBody,e);return;case"SeqNode":for(let t of r.stmts)Yt(t,e);return;case"ForNode":Yt(r.body,e);return;case"BlockNode":r.initBody&&Yt(r.initBody,e),Yt(r.body,e);return;case"LetStmtNode":Yt(r.value,e),Yt(r.body,e);return;default:return}}function $F(r){let e=2166136261;for(let t=0;t<r.length;t++)e^=r.charCodeAt(t),e=e*16777619>>>0;return e.toString(16).padStart(8,"0")}var VT=M(()=>{"use strict";l(Oy,"buildBlockMap");l(By,"computeWorkloadKey");l(PF,"readersByBuffer");l(Yt,"collectBlockOps");l($F,"fnv1a")});function FF(r,e){let t=r.medianMs!=null?1:0,n=e.medianMs!=null?1:0;return t!==n?n-t:t===1?r.medianMs-e.medianMs:e.score-r.score}var wR,gf,yf,TR=M(()=>{"use strict";VT();wR="mlfw-codegen-1",gf=class{static{l(this,"TuningRecord")}constructor(e,t,n,s,o,i){this.workloadKey=e,this.sketchName=t,this.params=n,this.score=s,this.traceData=o,this.version=i,this.timestamp=Date.now(),this.medianMs=null,this.minMs=null}};l(FF,"rankRecords");yf=class r{static{l(this,"TuningDatabase")}constructor(e=1){this.version=e,this._records=new Map}computeWorkloadKey(e,t,n,s=null){return By(e,t,n,s)}store(e,t){let n=this._records.get(e);n||(n=[],this._records.set(e,n)),n.push(t),n.sort(FF),n.length>10&&(n.length=10)}lookup(e){let t=this._records.get(e);return!t||t.length===0?null:t[0]}lookupTopK(e,t=5){let n=this._records.get(e);return n?n.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,n]of this._records)for(let s of n)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:wR,entries:e}}static deserialize(e){let t=new r(e.version);if(e.codegenVersion!==void 0&&e.codegenVersion!==wR)return t;for(let n of e.entries){let s=new gf(n.workloadKey,n.sketchName,n.params,n.score,n.traceData,n.version);s.timestamp=n.timestamp,s.medianMs=n.medianMs??null,s.minMs=n.minMs??null,t.store(n.workloadKey,s)}return t}saveToFile(e,t){return t.writeFile(e,JSON.stringify(this.serialize())),e}static loadFromFile(e,t){return t.exists(e)?r.deserialize(JSON.parse(t.readFile(e))):new r}clear(){this._records.clear()}}});function Si(r,e){if(!r)return 0;if(r.type==="ForNode"){let t=r.extent&&r.extent.type==="IntImmNode"?r.extent.value:1,n=r.threadTag?e:e*t;return Si(r.body,n)}if(r.type==="BlockNode")return Math.max(e,Si(r.body,e),r.initBody?Si(r.initBody,e):0);if(r.type==="SeqNode"){let t=e;for(let n of r.stmts)t=Math.max(t,Si(n,e));return t}return r.type==="IfThenElseNode"?Math.max(Si(r.thenBody,e),r.elseBody?Si(r.elseBody,e):e):r.type==="AllocateNode"||r.type==="LetStmtNode"?Si(r.body,e):e}function xR(r,e=.1){let t=r.slice().sort((g,y)=>g-y),n=t.length;if(n===0)return{median:0,min:0,trimmedMean:0,cv:0};let s=t[n>>1],o=t[0],i=Math.floor(n*e),a=i,u=n-i>a?n-i:n,p=0,c=0;for(let g=a;g<u;g++)p+=t[g],c++;let h=c>0?p/c:s,f=0;for(let g=a;g<u;g++){let y=t[g]-h;f+=y*y}let d=c>1?Math.sqrt(f/(c-1)):0,m=h>0?d/h:0;return{median:s,min:o,trimmedMean:h,cv:m}}var zF,Ry,My,SR=M(()=>{"use strict";Ec();zF=1e6;l(Si,"maxSerialTripCount");Ry=class{static{l(this,"BenchmarkResult")}constructor(e,t,n,s,o=null,i=0){this.medianMs=e,this.minMs=t,this.trimmedMeanMs=o??e,this.cv=i,this.samples=n,this._totalBytes=s}get throughputGBs(){return this.minMs<=0||!this._totalBytes?0:this._totalBytes/(this.minMs*1e6)}};l(xR,"robustStats");My=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,n=[];for(let[,i]of e.bufferMap){let a=Math.max(i.numel(),1);n.push(a);let u=i.sizeInBytes();u>0&&(t+=u)}let s=n.join(","),o=this._bufferCache.get(s);o||(o=n.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 Hs(this.target),n;try{n=t.compile(e)}catch(p){return this._record("benchmark-compile",p),null}if(n.metadata.kind!=="js")return null;let s;try{s=new Function("return "+n.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=xR(a),!(this.maxCv<=0||u.cv<=this.maxCv));p++);return new Ry(u.median,u.min,a,i,u.trimmedMean,u.cv)}_runMeasured(e){if(e.shapeParams&&e.shapeParams.length>0||Si(e.body,1)>zF)return null;let t;try{t=new Hs(this.target).compile(e)}catch(a){return this._record("measured-compile",a),null}let n=[],s=0;for(let[,a]of e.bufferMap){let u=Math.max(a.sizeInBytes(),1);n.push(u),s+=u}let o;try{o=this.measurer(t,n,[],{warmup:this.warmup,repeat:this.repeat})}catch(a){return this._record("measurer",a),null}if(!o||o.length===0)return null;let i=xR(o);return new Ry(i.median,i.min,o,s,i.trimmedMean,i.cv)}_collect(e,t,n){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;n.push(u),s+=u}}}});var Iy,kR=M(()=>{"use strict";Iy=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 vR(r){let e=null;for(let w of r.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"),n=e.getAttr("kernel_layout");if(t&&t!=="NCHW"||n&&n!=="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],h=s.shape[1],f=s.shape[2],d=s.shape[3],m=o.shape[0],g=o.shape[2],y=o.shape[3],_=i.shape[2],b=i.shape[3];return![c,h,f,d,m,g,y,_,b].every(w=>typeof w=="number"&&w>0)||o.shape[1]!==h||i.shape[1]!==m||i.shape[0]!==c?null:{N:c,Cin:h,H:f,W:d,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 VF(r,e,t){let{weight:n,input:s,output:o}=r,{N:i,Cin:a,H:u,W:p,O:c,Kh:h,Kw:f,Oh:d,Ow:m,sH:g,sW:y,pH:_,pW:b,dH:w,dW:T}=e,{BM:S,BN:x,BK:k,TM:A,TN:N}=t,v=c,O=i*d*m,R=a*h*f,B=h*f,E=d*m,z=x/N,D=S/A,J=z*D,j=Math.ceil(R/k),U=S*k,P=k*x,V=Math.ceil(U/J),he=Math.ceil(P/J),le=new me("ig_As",[U],"f32","shared"),Ie=new me("ig_Bs",[P],"f32","shared"),G=new me("ig_acc",[A*N],"f32","local"),K=new me("ig_af",[A],"f32","local"),ie=new me("ig_bf",[N],"f32","local"),ze=Ae("ig_bx"),Se=Ae("ig_by"),Qe=Ae("ig_tx"),It=Ae("ig_ty"),Jt=Ae("ig_tid"),Gt=Ae("ig_brow"),tt=Ae("ig_bcol"),ye=Ae("ig_k0"),He=l((Cr,X_)=>we(Te(Cr,$(N)),X_),"accIdx"),gn=Ae("ig_im"),jt=Ae("ig_in"),Zt=mn(gn,A,mn(jt,N,new Y(G,[He(gn,jt)],_f()))),Qt=v%S!==0,on=R%k!==0,Jn=O%x!==0,Bn=U%J!==0,yr=P%J!==0,Zn=Ae("ig_la"),pt=Ae("ig_aidx"),Qn=we(Gt,kn(pt,$(k))),mt=we(ye,sn(pt,$(k))),Or=kn(mt,$(B)),hs=sn(mt,$(B)),$s=kn(hs,$(f)),Br=sn(hs,$(f)),Xn=new q(n,[Qn,Or,$s,Br]);if(Qt||on){let Cr=Qt?dr(Qn,$(v)):null;on&&(Cr=Cr?ao(Cr,dr(mt,$(R))):dr(mt,$(R))),Xn=new xe(Cr,Xn,_f())}let _r=new Y(le,[pt],Xn);Bn&&(_r=new xe(dr(pt,$(U)),_r));let fo=Ey(Zn,V,new Ye(pt,we(Jt,Te(Zn,$(J))),_r)),Rr=Ae("ig_lb"),Mr=Ae("ig_bidx"),ds=we(ye,kn(Mr,$(x))),Qr=we(tt,sn(Mr,$(x))),Mi=kn(ds,$(B)),Vo=sn(ds,$(B)),Go=kn(Vo,$(f)),Ii=sn(Vo,$(f)),ec=kn(Qr,$(E)),Fa=sn(Qr,$(E)),Ei=kn(Fa,$(m)),Li=sn(Fa,$(m)),ho=Ly(we(Te(Ei,$(g)),Te(Go,$(w))),$(_)),Ci=Ly(we(Te(Li,$(y)),Te(Ii,$(T))),$(b)),jo=new q(s,[ec,Mi,ho,Ci]),mo=ao(ao(Cy(ho,$(0)),dr(ho,$(u))),ao(Cy(Ci,$(0)),dr(Ci,$(p))));on&&(mo=ao(mo,dr(ds,$(R)))),Jn&&(mo=ao(mo,dr(Qr,$(O)))),jo=new xe(mo,jo,_f());let qo=new Y(Ie,[Mr],jo);yr&&(qo=new xe(dr(Mr,$(P)),qo));let Di=Ey(Rr,he,new Ye(Mr,we(Jt,Te(Rr,$(J))),qo)),wt=Ae("ig_kk"),Fs=Ae("ig_fi"),tc=Ae("ig_fj"),U_=mn(Fs,A,new Y(K,[Fs],new q(le,[we(Te(we(Te(It,$(A)),Fs),$(k)),wt)]))),K_=mn(tc,N,new Y(ie,[tc],new q(Ie,[we(Te(wt,$(x)),we(Te(Qe,$(N)),tc))]))),za=Ae("ig_mi"),Va=Ae("ig_ni"),W_=mn(za,A,mn(Va,N,new Y(G,[He(za,Va)],we(new q(G,[He(za,Va)]),Te(new q(K,[za]),new q(ie,[Va])))))),H_=mn(wt,k,new te([U_,K_,W_])),ms=Ae("ig_kt"),Xr=new Ye(ye,Te(ms,$(k)),new te([fo,Di,new On,H_,new On])),br=Ey(ms,j,Xr),er=Ae("ig_wm"),Ir=Ae("ig_wn"),Uo=we(we(Gt,Te(It,$(A))),er),zs=we(we(tt,Te(Qe,$(N))),Ir),Er=kn(zs,$(E)),Lr=sn(zs,$(E)),go=kn(Lr,$(m)),Fn=sn(Lr,$(m)),Ko=new Y(o,[Er,Uo,go,Fn],new q(G,[He(er,Ir)]));if(Qt||Jn){let Cr=Qt?dr(Uo,$(v)):null;Jn&&(Cr=Cr?ao(Cr,dr(zs,$(O))):dr(zs,$(O))),Ko=new xe(Cr,Ko)}let yo=mn(er,A,mn(Ir,N,Ko)),Y_=new te([Zt,br,yo]),rh=new Ve(G,"local",new Ve(K,"local",new Ve(ie,"local",Y_))),J_=new Ye(Jt,we(Te(It,$(z)),Qe),new Ye(Gt,Te(Se,$(S)),new Ye(tt,Te(ze,$(x)),rh))),Z_=Math.ceil(O/x),Q_=Math.ceil(v/S),sh=ki(Se,"blockIdx.y",Q_,ki(ze,"blockIdx.x",Z_,ki(It,"threadIdx.y",D,ki(Qe,"threadIdx.x",z,J_))));return new Ve(le,"shared",new Ve(Ie,"shared",sh))}function NR(r,e,t){if(!AR(r,{M:e.O,N:e.N*e.Oh*e.Ow,K:e.Cin*e.Kh*e.Kw},t))return!1;let{BM:n,BN:s,BK:o,TM:i,TN:a}=t,u=e.O,p=e.N*e.Oh*e.Ow,c=e.Cin*e.Kh*e.Kw;if(u%n!==0||p%s!==0||c%o!==0||[n,s,o,i,a].some(f=>f%GT!==0))return!1;let h=n/i*(s/a);return!(n*o%(GT*h)!==0||o*s%h!==0)}function GF(r,e,t){let{weight:n,input:s,output:o}=r,{N:i,Cin:a,H:u,W:p,O:c,Kh:h,Kw:f,Oh:d,Ow:m,sH:g,sW:y,pH:_,pW:b,dH:w,dW:T}=e,{BM:S,BN:x,BK:k,TM:A,TN:N}=t,v=GT,O=c,R=i*d*m,B=a*h*f,E=h*f,z=d*m,D=u*p,J=a*u*p,j=x/N,U=S/A,P=j*U,V=B/k,he=S*k,le=k*x,Ie=he/P,G=le/P,K=Ie/v,ie=k/v,ze=(d-1)*g+(h-1)*w-_,Se=(m-1)*y+(f-1)*T-b,Qe=_>0||ze>=u,It=b>0||Se>=p,Jt=new me(n.name,[O*B>0?O*B:1],n.dtype,n.scope),Gt=new me(s.name,[i*J>0?i*J:1],s.dtype,s.scope),tt=new me("iv_As",[2*he],"f32","shared");tt.align16=!0;let ye=new me("iv_Bs",[2*le],"f32","shared");ye.align16=!0;let He=new me("iv_acc",[A*N],"f32","local"),gn=new me("iv_af",[A],"f32","local");gn.align16=!0;let jt=new me("iv_bf",[N],"f32","local");jt.align16=!0;let Zt=new me("iv_ra",[Ie],"f32","local");Zt.align16=!0;let Qt=new me("iv_rb",[G],"f32","local"),on=Ae("iv_bx"),Jn=Ae("iv_by"),Bn=Ae("iv_tx"),yr=Ae("iv_ty"),Zn=Ae("iv_tid"),pt=Ae("iv_brow"),Qn=Ae("iv_bcol"),mt=l((ms,Xr)=>we(Te(ms,$(N)),Xr),"accIdx"),Or=0,hs=Ae("iv_im"),$s=Ae("iv_in"),Br=mn(hs,A,mn($s,N,new Y(He,[mt(hs,$s)],_f()))),Xn=l(ms=>{let Xr=Or++,br=Ae("iv_la"+Xr),er=Ae("iv_c"+Xr),Ir=kn(er,$(ie)),Uo=Te(sn(er,$(ie)),$(v)),zs=mn(br,K,new Ye(er,we(Zn,Te(br,$(P))),new iu(Zt,Te(br,$(v)),Jt,we(Te(we(pt,Ir),$(B)),we(ms,Uo)),v))),Er=Ae("iv_lb"+Xr),Lr=Ae("iv_e"+Xr),go=kn(Lr,$(x)),Fn=sn(Lr,$(x)),Ko=we(Qn,Fn),yo=we(ms,go),Y_=kn(yo,$(E)),rh=sn(yo,$(E)),J_=kn(rh,$(f)),Z_=sn(rh,$(f)),Q_=kn(Ko,$(z)),sh=sn(Ko,$(z)),Cr=kn(sh,$(m)),X_=sn(sh,$(m)),eb=Ly(we(Te(Cr,$(g)),Te(J_,$(w))),$(_)),tb=Ly(we(Te(X_,$(y)),Te(Z_,$(T))),$(b)),OE=we(we(we(Te(Q_,$(J)),Te(Y_,$(D))),Te(eb,$(p))),tb),nb=new q(Gt,[OE]),Ga=null;if(Qe&&(Ga=ao(Cy(eb,$(0)),dr(eb,$(u)))),It){let xS=ao(Cy(tb,$(0)),dr(tb,$(p)));Ga=Ga?ao(Ga,xS):xS}Ga&&(nb=new xe(Ga,nb,_f()));let BE=mn(Er,G,new Ye(Lr,we(Zn,Te(Er,$(P))),new Y(Qt,[Er],nb)));return new te([zs,BE])},"prefetch"),_r=l((ms,Xr)=>{let br=Or++,er=Ae("iv_la"+br),Ir=Ae("iv_c"+br),Uo=kn(Ir,$(ie)),zs=Te(sn(Ir,$(ie)),$(v)),Er=[];for(let yo=0;yo<v;yo++)Er.push(new Y(tt,[we(ms,we(Te(we(zs,$(yo)),$(S)),Uo))],new q(Zt,[we(Te(er,$(v)),$(yo))])));let Lr=mn(er,K,new Ye(Ir,we(Zn,Te(er,$(P))),new te(Er))),go=Ae("iv_lb"+br),Fn=Ae("iv_e"+br),Ko=mn(go,G,new Ye(Fn,we(Zn,Te(go,$(P))),new Y(ye,[we(Xr,Fn)],new q(Qt,[go]))));return new te([Lr,Ko])},"commit"),fo=l((ms,Xr)=>{let br=Or++,er=Ae("iv_kk"+br),Ir=[],Uo=we(ms,we(Te(er,$(S)),Te(yr,$(A))));for(let Fn=0;Fn<A/v;Fn++)Ir.push(new iu(gn,$(Fn*v),tt,we(Uo,$(Fn*v)),v));let zs=we(Xr,we(Te(er,$(x)),Te(Bn,$(N))));for(let Fn=0;Fn<N/v;Fn++)Ir.push(new iu(jt,$(Fn*v),ye,we(zs,$(Fn*v)),v));let Er=Ae("iv_mi"+br),Lr=Ae("iv_ni"+br),go=mn(Er,A,mn(Lr,N,new Y(He,[mt(Er,Lr)],we(new q(He,[mt(Er,Lr)]),Te(new q(gn,[Er]),new q(jt,[Lr]))))));return mn(er,k,new te([...Ir,go]))},"computeMMA"),Rr=Ae("iv_kt"),Mr=Ae("iv_p"),ds=Ae("iv_pN"),Qr=Te(Mr,$(he)),Mi=Te(Mr,$(le)),Vo=Te(ds,$(he)),Go=Te(ds,$(le)),Ii=dr(Rr,$(V-1)),ec=new te([Xn($(0)),_r($(0),$(0)),new On]),Fa=new Ye(Mr,sn(Rr,$(2)),new Ye(ds,sn(we(Rr,$(1)),$(2)),new te([new xe(Ii,Xn(Te(we(Rr,$(1)),$(k)))),fo(Qr,Mi),new xe(Ii,new te([_r(Vo,Go),new On]))]))),Ei=Ey(Rr,V,Fa),Li=Ae("iv_wm"),ho=Ae("iv_wn"),Ci=we(we(pt,Te(yr,$(A))),Li),jo=we(we(Qn,Te(Bn,$(N))),ho),mo=kn(jo,$(z)),qo=sn(jo,$(z)),Di=kn(qo,$(m)),wt=sn(qo,$(m)),Fs=new Y(o,[mo,Ci,Di,wt],new q(He,[mt(Li,ho)])),tc=mn(Li,A,mn(ho,N,Fs)),U_=new te([Br,ec,Ei,tc]),K_=new Ve(He,"local",new Ve(gn,"local",new Ve(jt,"local",new Ve(Zt,"local",new Ve(Qt,"local",U_))))),za=new Ye(Zn,we(Te(yr,$(j)),Bn),new Ye(pt,Te(Jn,$(S)),new Ye(Qn,Te(on,$(x)),K_))),Va=R/x,W_=O/S,H_=ki(Jn,"blockIdx.y",W_,ki(on,"blockIdx.x",Va,ki(yr,"threadIdx.y",U,ki(Bn,"threadIdx.x",j,za))));return new Ve(tt,"shared",new Ve(ye,"shared",H_))}function AR(r,e,t){let{BM:n,BN:s,BK:o,TM:i,TN:a}=t;if(![n,s,o,i,a].every(m=>typeof m=="number"&&m>0)||n%i!==0||s%a!==0)return!1;let u=s/a,p=n/i,c=u*p,h=r.warpSize||32;return!(c%h!==0||c>(r.maxThreadsPerBlock||1024)||(n*o+o*s)*4*2>(r.sharedMemoryBytes||49152)||i*a+i+a+h>(r.registersPerThread||255))}function qF(r,e){for(let t of jF)if(NR(r,e,t))return t;return null}function OR(r,e,t){let n=r.func,s=n.convInfo;if(!s)return!1;let i=io(n.body).find(_=>/^conv_acc_/.test(_));if(!i)return!1;let a=uf(n.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 h=s.O,f=s.N*s.Oh*s.Ow,d=s.Cin*s.Kh*s.Kw;if(d<128||f<64)return!1;let m=t&&t.convConfig,g=t&&t.convNoVec,y=null;if(!g){let _=m?NR(e,s,m)?m:null:qF(e,s);_&&(y=GF({weight:p,input:u,output:c},s,_))}if(!y){let _=m?AR(e,{M:h,N:f,K:d},m)?m:null:ky(e,{M:h,N:f,K:d});if(!_)return!1;y=VF({weight:p,input:u,output:c},s,_)}return r.func.body=y,r.func._setChild&&r.func._setChild("body",y),r.func.gpuRegisterBlocked=!0,!0}var $,_f,Ae,we,Ly,Te,kn,sn,dr,Cy,ao,Ey,mn,ki,GT,jF,jT=M(()=>{"use strict";De();xo();Ol();vy();$=l(r=>new L(r),"I"),_f=l(()=>new fe(0),"FZERO"),Ae=l(r=>new Ne(r,"i32"),"IV"),we=l((r,e)=>new C("+",r,e),"ADD"),Ly=l((r,e)=>new C("-",r,e),"SUB"),Te=l((r,e)=>new C("*",r,e),"MUL"),kn=l((r,e)=>new C("//",r,e),"DIV"),sn=l((r,e)=>new C("%",r,e),"MOD"),dr=l((r,e)=>new qe("lt",r,e),"LT"),Cy=l((r,e)=>new qe("ge",r,e),"GE"),ao=l((r,e)=>new C("&&",r,e),"AND"),Ey=l((r,e,t)=>new de(r,$(0),$(e),F.SERIAL,t),"forS"),mn=l((r,e,t)=>new de(r,$(0),$(e),F.UNROLLED,t),"forU"),ki=l((r,e,t,n)=>new de(r,$(0),$(t),F.THREAD_BINDING,n,e),"forT");l(vR,"detectPureConv");l(VF,"buildImplicitGemmConv");GT=4;l(NR,"vectorizableConvConfig");l(GF,"buildVectorizedImplicitGemmConv");l(AR,"validConvConfig");jF=[{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(qF,"pickVectorizedConvConfig");l(OR,"applyImplicitGemmConv")});function BR(r,e,t,n,s,o,i,a){let u=new me(r,[t,n],e.dtype,"shared"),p=new Y(u,[s,o],new q(e,[i,a]));return{tile:u,fill:p}}function WF(r,e,t){let{A:n,B:s,C:o,M:i,N:a,K:u,transB:p}=r,c=u/t,h=Rl("ts_by"),f=Rl("ts_bx"),d=Rl("ts_ty"),m=Rl("ts_tx"),g=Rl("ts_ko"),y=Rl("ts_ki"),_=new me("ts_acc",[1],"f32","local"),b=Ml(Dy(h,Jr(e)),d),w=Ml(Dy(f,Jr(e)),m),T=Dy(g,Jr(t)),S=BR("ts_As",n,e,t,d,m,b,Ml(T,m)),x=BR("ts_Bs",s,t,e,d,m,p?w:Ml(T,d),p?Ml(T,d):w),k=KF(y,t,new Y(_,[Jr(0)],Ml(new q(_,[Jr(0)]),Dy(new q(S.tile,[d,y]),new q(x.tile,[y,m]))))),A=new te([S.fill,x.fill,new On,k,new On]),N=UF(g,c,A),v=new Y(_,[Jr(0)],new fe(0)),O=new Y(o,[b,w],new q(_,[Jr(0)])),R=new Ve(_,"local",new te([v,N,O]));return Py(h,"blockIdx.y",i/e,Py(f,"blockIdx.x",a/e,new Ve(S.tile,"shared",new Ve(x.tile,"shared",Py(d,"threadIdx.y",e,Py(m,"threadIdx.x",e,R))))))}function HF(r,e){let{M:t,N:n,K:s}=e,o=16,i=16;return o*o>(r.maxThreadsPerBlock||1024)||(o*i+i*o)*4>(r.sharedMemoryBytes||49152)||t%o!==0||n%o!==0||s%i!==0||t<o||n<o||s<i?null:{BS:o,BK:i}}function YF(r,e){let n=io(r.func.body).filter(o=>/^q?conv_(init|acc)_/.test(o));if(n.length===0)return!1;let s=new Rs(e);for(let o of n)s.applyToBlock(r,o);return!0}function JF(r,e,t={}){if(!e.isGPU())return!1;let n=mR(r.func);if(!n)return!1;let s=n.dims;if(t&&t.primitiveMatmul&&(s.batch||1)===1&&!n.epilogue){let a=HF(e,s);if(a){let u=WF(s,a.BS,a.BK);return r.func.body=u,r.func._setChild&&r.func._setChild("body",u),r.func.gpuRegisterBlocked=!0,!0}}let o=ky(e,s);if(!o)return!1;let i=CT(s,o,n.epilogue);return r.func.body=i,r.func._setChild&&r.func._setChild("body",i),r.func.gpuRegisterBlocked=!0,!0}function $y(r,e,t={}){if(!e.isGPU()||e.isWebGPU&&e.isWebGPU())return!1;let n=JF(r,e,t);return n||(n=OR(r,e,t)),n||(n=YF(r,e)),n}var Jr,Rl,Ml,Dy,UF,KF,Py,qT=M(()=>{"use strict";vy();Ol();jT();ro();De();xo();Jr=l(r=>new L(r),"I"),Rl=l(r=>new Ne(r,"i32"),"IV"),Ml=l((r,e)=>new C("+",r,e),"ADD"),Dy=l((r,e)=>new C("*",r,e),"MUL"),UF=l((r,e,t)=>new de(r,Jr(0),Jr(e),F.SERIAL,t),"forS"),KF=l((r,e,t)=>new de(r,Jr(0),Jr(e),F.UNROLLED,t),"forU"),Py=l((r,e,t,n)=>new de(r,Jr(0),Jr(t),F.THREAD_BINDING,n,e),"forT");l(BR,"stageTileToShared");l(WF,"buildTiledSharedMatmul");l(HF,"pickTiledConfig");l(YF,"applyDeterministicGpuConv");l(JF,"applyDeterministicGpuMatmul");l($y,"applyDeterministicGpuSchedule")});function UT(r){return r*1664525+1013904223&2147483647}function QF(r,e,t){let n={...r};for(let s of e.variables)t.rngFloat()<t.mutationRate&&(n[s.name]=s.sample(o=>t.rng(o)));return n}function RR(r={}){return r.strategy==="random"?new KT({numTrials:r.numTrials,seed:r.seed,deadline:r.deadline}):new WT({populationSize:r.populationSize,numGenerations:r.numGenerations,mutationRate:r.mutationRate,eliteRatio:r.eliteRatio,seed:r.seed,deadline:r.deadline})}var Fy,KT,WT,ZF,MR=M(()=>{"use strict";l(UT,"nextLcg");Fy=class{static{l(this,"SearchCandidate")}constructor(e,t,n){this.sketchName=e,this.params=t,this.score=n}},KT=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=UT(this._rngState),this._rngState%e}_expired(){return this.deadline?this.deadline.expired:!1}search(e,t){let n=[];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&&n.push(new Fy(s.name,i,a.score))}}return n.sort((s,o)=>o.score-s.score),{candidates:n,population:null}}},WT=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=UT(this._rngState),this._rngState%e}_rngFloat(){return this._rngState=UT(this._rngState),this._rngState/2147483647}_expired(){return this.deadline?this.deadline.expired:!1}search(e,t,n=null){let s=n&&n.length?n: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 h=t(u,p);return o.set(c,h),h},"evalMemo");for(let u=0;u<this.numGenerations&&!this._expired();u++){let p=[];for(let d of s){let m=i(d.sketch,d.params);m&&p.push({...d,score:m.score})}if(p.sort((d,m)=>m.score-d.score),p.length===0)break;let c=Math.max(1,Math.floor(p.length*this.eliteRatio)),h=p.slice(0,c),f=h.map(d=>({sketch:d.sketch,params:{...d.params}}));for(;f.length<this.populationSize;){let d=h[this._rng(h.length)],m=h[this._rng(h.length)];if(d.sketch!==m.sketch){let g=this._rngFloat()<.5?d:m;f.push(this._mutate(g.sketch,{...g.params}))}else f.push(this._mutate(d.sketch,this._crossover(d,m)))}s=f}let a=[];for(let u of s){let p=i(u.sketch,u.params);p&&a.push(new Fy(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 n=0;n<this.populationSize;n++){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 n={};for(let s of e.sketch.variables)n[s.name]=this._rngFloat()<.5?e.params[s.name]:t.params[s.name];return n}_mutate(e,t){let n={rngFloat:l(()=>this._rngFloat(),"rngFloat"),rng:l(i=>this._rng(i),"rng"),mutationRate:this.mutationRate},s=[QF,...ZF],o=t;for(let i of s)o=i(o,e,n);return{sketch:e,params:o}}},ZF=[];l(QF,"defaultResampleMutator");l(RR,"createSearchStrategy")});function Vy(r){let e={"threadIdx.x":1,"threadIdx.y":1,"threadIdx.z":1},t=new Set,n=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(n):i&&typeof i=="object"&&n(i)}}},"visit");return n(r.body),XF.reduce((s,o)=>s*e[o],1)}var XF,zy,IR=M(()=>{"use strict";fa();Gw();xT();cf();MR();vT();XF=["threadIdx.x","threadIdx.y","threadIdx.z"];l(Vy,"gpuThreadBlockSize");zy=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 Nl,this.costModel=new Al(e.costModel.analytical,this.learnedModel)):(this.costModel=e.costModel,this.learnedModel=e.learnedModel);let n=t?null:lR(e.primFunc,e.blockName,e.blockMap);this.evalFunc=n||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=RR({...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 n=this._best?this._best.measuredScore:-1/0;return Math.max(0,n-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 Sn(Lo(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:n}=this.strategy.search(this.sketches,e,this.population);return this.population=n,t}_evaluate(e,t){try{let n=Lo(this.evalFunc),s=new Sn(n);if(e.instantiate(t)(s,this.evalBlockName,this.target),Wu.validate(n).length>0)return null;let i=this.target.maxThreadsPerBlock;return this.target.isGPU&&this.target.isGPU()&&i&&Vy(n)>i?null:{score:this.costModel.score(n)}}catch(n){return this._warnedEvalSketches.has(e.name)||(this._warnedEvalSketches.add(e.name),this._warn("evaluate-candidate",this.blockName,n)),null}}_measureAndLearn(e){let t=e.slice(0,this.config.topKForBenchmark);for(let n of t){if(this.deadline&&this.deadline.expired)break;let s=this._measure(n);if(!s)continue;let o=-s.result.medianMs;this.learnedModel.addSample(s.features,o),this._consider({sketchName:n.sketchName,params:n.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 n,s;try{n=Lo(this.primFunc),t.instantiate(e.params)(new Sn(n),this.blockName,this.target),s=Lo(this.evalFunc),t.instantiate(e.params)(new Sn(s),this.evalBlockName,this.target)}catch(i){return this._warn("measure-candidate",this.blockName,i),null}let o=this.benchmarkRunner.run(n);return o?{result:o,features:Ra.extractStatements(s)}:null}_consider(e){(!this._best||e.measuredScore>this._best.measuredScore)&&(this._best=e)}}});var HT,Gy,ER=M(()=>{"use strict";HT=class{static{l(this,"GradientSchedulerPolicy")}pick(e){let t=e.filter(i=>!i.plateaued);if(t.length===0)return null;let n=t.filter(i=>i.rounds===0);if(n.length>0)return n[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}},Gy=class{static{l(this,"TaskScheduler")}constructor(e=null){this.policy=e||new HT}run(e,t,n={}){let s=n.maxRoundsPerTask??8,o=n.plateauPatience??2,i=n.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 ez(r){if(r.isCPU())return null;let e=Hv(r.kind);if(!e)throw new Error("hardwareMeasure requested for target '"+r.kind+"' but no measurer is registered for it; the corresponding runtime must be loaded (Node: import '#io/cuda_runtime') before compiling");return e}var jy,qy,LR=M(()=>{"use strict";fa();vT();bR();TR();SR();kR();VT();Ol();PT();ro();qT();IR();cf();De();ER();rw();l(ez,"resolveMeasurer");jy=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}},qy=class{static{l(this,"Autotuner")}constructor(e,t={},n=null){this.target=e,this.config=t instanceof jy?t:new jy(t),this.trace=n,this._funcName=null,this.config.hardwareMeasure&&(this.config.measurer=ez(e)),this.analyticalModel=new wy(e),this.learnedModel=new Nl,this.costModel=new Al(this.analyticalModel,this.learnedModel),this.db=this.config.tuningDB instanceof yf?this.config.tuningDB:new yf;let s=l((o,i,a)=>this._warn(o,i,a),"warn");this.benchmarkRunner=this.config.enableBenchmark?new My(e,{warmup:this.config.benchmarkWarmup,repeat:this.config.benchmarkRepeat,maxCv:this.config.benchmarkMaxCv,measurer:this.config.measurer,warn:s}):null,this.scheduler=new Gy(this.config.schedulerPolicy)}_warn(e,t,n){let s=n&&n.message?n.message:String(n);if(this.config.onWarning)try{this.config.onWarning({stage:e,func:this._funcName,block:t||null,message:s,error:n})}catch(o){}this.trace&&this.trace.warn("autotune",this._funcName,`${e}${t?" ["+t+"]":""}: ${s}`)}tune(e,t=null){this._funcName=e.name;let n=t?[t]:io(e.body),s=Oy(e.body),o=mf(e),i=new Iy(this.config.timeBudgetMs,this.config.clock),a=new Map,u=new Map;for(let h of n){let f=By(e,h,this.target,s);u.set(h,f);let d=a.get(f);if(d){d.weight++;continue}if(this.config.useTuningDB&&this.db.has(f)){a.set(f,{key:f,kind:"cache",cached:this.db.lookup(f),weight:1});continue}let m=zT(e,h,this.target,s,{richGpu:this.config.richGpu??!!this.config.measurer,dag:o});if(m.length===0){a.set(f,{key:f,kind:"empty",weight:1});continue}let g=new zy({target:this.target,primFunc:e,blockName:h,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(f,{key:f,kind:"session",session:g,weight:1})}let p=[...a.values()].filter(h=>h.kind==="session");p.length>0&&this.scheduler.run(p,i,this.config);let c=new Map;for(let h of n){let f=a.get(u.get(h));if(f.kind==="cache"){c.set(h,{sketchName:f.cached.sketchName,params:f.cached.params,score:f.cached.score,fromCache:!0});continue}if(f.kind==="empty")continue;let d=f.session.best();if(d){if(this.config.useTuningDB&&!f.stored){let m=new gf(f.key,d.sketchName,d.params,d.score,f.session.bestTrace(),this.db.version);m.medianMs=d.medianMs||null,m.minMs=d.minMs||null,this.db.store(f.key,m),f.stored=!0}c.set(h,{sketchName:d.sketchName,params:d.params,score:d.score,fromCache:!1,medianMs:d.medianMs,minMs:d.minMs})}}return c}tuneAndApply(e,t=null){let n=this.tune(e,t);if(n.size>0){let s=this._applyBestSchedule(e,n);if(s)return{func:s.func,results:n,applied:!0}}return{func:e,results:n,applied:!1}}_applyBestSchedule(e,t){let n=this._buildDefaultSchedule(e),s=!!n&&this._scheduleIsValid(n),o=s&&this._isStrongBackendSchedule(n),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,n),{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 n=Lo(e),s=new Sn(n),o=Oy(n.body),i=mf(n),a=new Set,u=[];for(let c of t)if(c[1].sketchName==="fused"){let h=Ny(n,i,c[0],dt);h&&a.add(h),u.unshift(c)}else u.push(c);let p=new Set;for(let[c,h]of u)if(!(a.has(c)||p.has(h))&&(p.add(h),!(!h.sketchName||!h.params)))try{let d=zT(n,c,this.target,o,{richGpu:this.config.richGpu??!!this.config.measurer,dag:i}).find(m=>m.name===h.sketchName);d&&this._fitsThreadBlock(n,c,d,h.params)&&d.instantiate(h.params)(s,c,this.target)}catch(f){this._warn("apply-tuned-block",c,f);continue}return this._scheduleResidualBlocks(s,a),n}catch(n){return this._warn("build-tuned-schedule",null,n),null}}_scheduleResidualBlocks(e,t){let n=null;for(let s of io(e.func.body))if(!(t.has(s)||this._blockIsParallelized(e,s))){n||(n=new Rs(this.target));try{n.applyToBlock(e,s)}catch(o){this._warn("residual-block",s,o)}}}_blockIsParallelized(e,t){let n;try{n=e.getLoops(t)}catch(s){return this._warn("block-loops",t,s),!0}for(let s of n)if(s.kind===F.THREAD_BINDING||s.kind===F.PARALLEL||s.kind===F.VECTORIZED)return!0;return!1}_scheduleIsValid(e){return!this.target.isGPU||!this.target.isGPU()||!this.target.maxThreadsPerBlock?!0:Vy(e)<=this.target.maxThreadsPerBlock}_buildDefaultSchedule(e){try{let t=Lo(e),n=new Sn(t);return $y(n,this.target,this.config)||new Rs(this.target).applyToAllBlocks(n),t}catch(t){return this._warn("build-default-schedule",null,t),null}}_adoptSchedule(e,t){Object.assign(e,t),e._setChild("body",e.body)}_fitsThreadBlock(e,t,n,s){if(!this.target.isGPU||!this.target.isGPU()||!this.target.maxThreadsPerBlock)return!0;try{let o=Lo(e);return n.instantiate(s)(new Sn(o),t,this.target),Vy(o)<=this.target.maxThreadsPerBlock}catch(o){return this._warn("fits-thread-block",t,o),!1}}}});var Uy,CR=M(()=>{"use strict";xi();fa();ro();LR();qT();Uy=class extends Wn{static{l(this,"SchedulePass")}constructor(e){super("SchedulePass","scheduling"),this.config=e,this.target=e.target,this.snapshotPoint="afterScheduling",this._autotuner=null,this._policy=null}begin(e){let t=this.config.scheduling;t.autotune?this._autotuner=new qy(this.target,t,e.trace):t.enabled&&(this._policy=new Rs(this.target,null,e.trace))}run(e,t){let n=this.config.scheduling,s=t.trace;if(n.autotune){if(e.cublasInfo||e._tensorIntrin)return;let o=performance.now(),i=this._autotuner.tuneAndApply(e),a=performance.now()-o,u=0,p=0;if(i&&i.results){p=i.results.size;for(let[c,h]of i.results)h.fromCache&&u++,s.explainsEnabled&&s.explain("schedule",c,h.sketchName,`autotuned: best of search${h.fromCache?" (cached)":""}, score ${h.score!=null?h.score.toFixed(3):"n/a"}`,{target:this.target.name,params:h.params})}s.autotuneStats(e.name,{durationMs:a,blockCount:p,applied:!!(i&&i.applied),cacheHits:u})}else if(n.enabled||n.gpuTiling){if(e.cublasInfo||e._tensorIntrin)return;let o=performance.now(),i=new Sn(e);!$y(i,this.target,n)&&n.enabled&&this._policy.applyToAllBlocks(i),s.functionEvent("scheduling",e.name,{durationMs:performance.now()-o})}}}});function DR(r){let e={analyzer:new Os,simp:null};e.simp=new hm(e.analyzer);let t=Zr(r.body,e);return r.body=t,r._setChild("body",t),r}function tz(r,e,t){let n=r.analyzer.getVarBound(e);return t&&t.type==="IntImmNode"&&t.value>0?r.analyzer.bind(e,0,t.value-1):r.analyzer.setVarBound(e,null),n}function Zr(r,e){if(!r||typeof r!="object")return r;switch(r.type){case"ForNode":{let t=tz(e,r.loopVar.name,r.extent),n=Zr(r.body,e);e.analyzer.setVarBound(r.loopVar.name,t);let s=new de(r.loopVar,r.min,r.extent,r.kind,n,r.threadTag);return r.annotations&&(s.annotations=r.annotations),s}case"BlockNode":{let t=[];for(let i of r.iterVars)i.iterVar&&(t.push([i.iterVar.name,e.analyzer.getVarBound(i.iterVar.name)]),e.analyzer.setVarBound(i.iterVar.name,i.binding?la(e.analyzer,i.binding):null));let n=r.iterVars.map(nz(e)),s=Zr(r.body,e),o=r.initBody?Zr(r.initBody,e):null;for(let[i,a]of t)e.analyzer.setVarBound(i,a);return new pe(r.name,n,r.reads,r.writes,s,o)}case"SeqNode":return new te(r.stmts.map(t=>Zr(t,e)));case"IfThenElseNode":{let t=Hn(r.condition,e);return Mo(e.analyzer,t)?Zr(r.thenBody,e):ua(e.analyzer,t)?r.elseBody?Zr(r.elseBody,e):new te([]):new xe(t,Zr(r.thenBody,e),r.elseBody?Zr(r.elseBody,e):null)}case"BufferStoreNode":return new Y(r.buffer,r.indices.map(t=>Hn(t,e)),Hn(r.value,e));case"LetStmtNode":return new Ye(r.variable,Hn(r.value,e),Zr(r.body,e));case"AllocateNode":return new Ve(r.buffer,r.scope,Zr(r.body,e));case"WhileNode":return new Zo(r.condVar,Zr(r.condBody,e),Zr(r.loopBody,e));case"EvaluateNode":return new ou(Hn(r.value,e));default:return r}}function nz(r){return e=>{if(!e.iterVar||!e.binding)return e;let t=Hn(e.binding,r);return new qi(e.iterVar,t,e.kind)}}function Hn(r,e){if(!r||typeof r!="object"||!r.type)return r;switch(r.type){case"IntImmNode":case"FloatImmNode":case"VariableNode":return r;case"BufferLoadNode":return new q(r.buffer,r.indices.map(t=>Hn(t,e)));case"MathOpNode":{let t=Hn(r.a,e),n=r.b?Hn(r.b,e):null;return e.simp.simplify(new C(r.op,t,n))}case"CompareNode":{let t=Hn(r.a,e),n=Hn(r.b,e);return e.simp.simplify(new qe(r.direction,t,n))}case"CastNode":return new rt(Hn(r.expr,e),r.fromDtype,r.toDtype);case"CallExternNode":return new ke(r.externName,r.args.map(t=>Hn(t,e)),r.dtype);case"IfThenElseNode":{let t=Hn(r.condition,e),n=Hn(r.thenBody,e),s=r.elseBody?Hn(r.elseBody,e):null;return Mo(e.analyzer,t)?n:s!==null&&ua(e.analyzer,t)?s:new xe(t,n,s)}default:return r}}var PR=M(()=>{"use strict";Ku();Yc();De();l(DR,"simplifyPrimFunc");l(tz,"bindLoopVar");l(Zr,"simplifyStmt");l(nz,"simplifyIterVar");l(Hn,"simplifyExpr")});var Ky,$R=M(()=>{"use strict";xi();PR();Ky=class extends Wn{static{l(this,"SimplifyPass")}constructor(){super("SimplifyPass","simplify")}run(e,t){let n=performance.now();return DR(e),t.trace.functionEvent("simplify",e.name,{durationMs:performance.now()-n}),e}}});function FR(r){return!!r&&typeof r=="object"&&typeof r.name=="string"&&r.dtype!==void 0&&r.shape!==void 0&&r.type===void 0}var YT,JT,rz,Wy,zR=M(()=>{"use strict";YT=class{static{l(this,"BufferInterval")}constructor(e,t,n,s){this.buffer=e,this.firstUse=t,this.lastUse=n,this.scope=s}get size(){return this.buffer.sizeInBytes()}overlaps(e){return this.firstUse<=e.lastUse&&e.firstUse<=this.lastUse}},JT=class{static{l(this,"BufferLivenessResult")}constructor(e,t,n){this.intervals=e,this.stmtOrder=t,this.paramBuffers=n}isParam(e){return this.paramBuffers.has(e)}getTemporaries(){let e=[];for(let[t,n]of this.intervals)this.paramBuffers.has(t)||e.push(n);return e}interfere(e,t){let n=this.intervals.get(e),s=this.intervals.get(t);return!n||!s?!1:n.overlaps(s)}},rz=new Set(["_parent","_parentKey","_parentIdx"]);l(FR,"isBuffer");Wy=class{static{l(this,"BufferLiveness")}static analyze(e){let t=new Map,n=[],s=0,o=new Set;for(let[,h]of e.bufferMap)o.add(h);let i=[],a=l(h=>{if(!h)return;let f=t.get(h);f?(s<f.firstUse&&(f.firstUse=s),s>f.lastUse&&(f.lastUse=s)):(f=new YT(h,s,s,h.scope),t.set(h,f)),i.push(h)},"touch"),u=l((h,f)=>{if(!(!h||typeof h!="object"||f.has(h))){if(f.add(h),FR(h)){a(h);return}FR(h.buffer)&&a(h.buffer);for(let d of Object.keys(h)){if(rz.has(d)||d==="buffer")continue;let m=h[d];if(!(!m||typeof m!="object"))if(Array.isArray(m))for(let g of m)u(g,f);else u(m,f)}}},"touchAll"),p=l((h,f)=>{for(let d=h;d<i.length;d++){let m=t.get(i[d]);m&&f>m.lastUse&&(m.lastUse=f)}},"extendRegion"),c=l(h=>{if(h)switch(h.type){case"SeqNode":for(let f of h.stmts)c(f);break;case"ForNode":{let f=s,d=i.length;u(h.min,new Set),u(h.extent,new Set),c(h.body);let m=s>f?s-1:f;p(d,m);break}case"WhileNode":{let f=s,d=i.length;u(h.condVar,new Set),c(h.condBody),c(h.loopBody);let m=s>f?s-1:f;p(d,m);break}case"BlockNode":n.push({idx:s,node:h});for(let f of h.reads)a(f.buffer);for(let f of h.writes)a(f.buffer);u(h.body,new Set),h.initBody&&u(h.initBody,new Set),s++;break;case"AllocateNode":a(h.buffer),c(h.body);break;case"IfThenElseNode":{let f=s,d=i.length;u(h.condition,new Set),c(h.thenBody),h.elseBody&&c(h.elseBody);let m=s>f?s-1:f;p(d,m);break}case"LetStmtNode":u(h.value,new Set),c(h.body);break;case"EvaluateNode":u(h.value,new Set);break;default:u(h,new Set);break}},"walk");return c(e.body),new JT(t,n,o)}}});function sz(r,e){if(r.shape.length!==e.shape.length)return!1;for(let t=0;t<r.shape.length;t++)if(r.shape[t]!==e.shape[t])return!1;return!0}function Il(r,e){if(r===e)return!0;if(!r||!e||typeof r!="object"||typeof e!="object"||r.type!==e.type)return!1;switch(r.type){case"VariableNode":return r.name===e.name;case"IntImmNode":return r.value===e.value;case"FloatImmNode":return r.value===e.value;case"MathOpNode":return r.op===e.op&&Il(r.a,e.a)&&Il(r.b,e.b);case"CompareNode":return r.direction===e.direction&&Il(r.a,e.a)&&Il(r.b,e.b);case"CastNode":return r.toDtype===e.toDtype&&Il(r.expr,e.expr);case"BufferLoadNode":return r.buffer===e.buffer&&QT(r.indices,e.indices);default:return!1}}function QT(r,e){if(r.length!==e.length)return!1;for(let t=0;t<r.length;t++)if(!Il(r[t],e[t]))return!1;return!0}function VR(r,e){r&&Ss(r,e)}function oz(r,e,t){let n=[],s=[];for(let p of[r.body,r.initBody])p&&VR(p,c=>{c.type==="BufferStoreNode"&&c.buffer===t?n.push(c):c.type==="BufferLoadNode"&&c.buffer===e&&s.push(c)});if(n.length!==1||s.length===0)return!1;let o=n[0],i=o.indices;for(let p of s)if(!QT(p.indices,i))return!1;let a=new Set;for(let p of[o.value,...o.indices])p&&VR(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&&QT(p.indices,i)))return!1}return!0}function iz(r,e){if(r)for(let t of gu(r,n=>n.type==="BlockNode",{kinds:"stmt"}))e.push(t)}var ZT,Hy,GR=M(()=>{"use strict";Gr();ZT=class{static{l(this,"InplaceCandidate")}constructor(e,t,n){this.srcBuffer=e,this.dstBuffer=t,this.reason=n}},Hy=class{static{l(this,"InplaceAnalysis")}static analyze(e,t,n=new Set){let s=[],o=[];iz(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 h=a.get(c.buffer);(h===void 0||p.idx>h)&&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 h of p.writes){let f=h.buffer;if(!(t.isParam(f)&&!n.has(f)))for(let d of p.reads){let m=d.buffer;if(m===f||u.has(m)||t.isParam(m)&&!n.has(m)||!sz(m,f)||m.dtype!==f.dtype||m.scope!==f.scope)continue;let g=t.intervals.get(m),y=t.intervals.get(f);if(!(!g||!y)&&g.lastUse<=y.firstUse){let _=a.get(m);if(_===void 0||_<=c){if(!oz(p,m,f))continue;s.push(new ZT(m,f,`${m.name} last used at ${g.lastUse}, ${f.name} first used at ${y.firstUse}`)),u.add(m);break}}}}}return s}};l(sz,"shapesMatch");l(Il,"exprEqual");l(QT,"indexListEqual");l(VR,"walkNodes");l(oz,"isInplaceComputeSafe");l(iz,"collectBlocks")});function jR(r,e){let t=0,n=r.length;for(;t<n;){let s=t+n>>1;r[s].offset<=e.offset?t=s+1:n=s}r.splice(t,0,e)}function qR(r){return{best:null,consider(e,t,n){return t<n?null:r==="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 XT,ex,Yy,UR=M(()=>{"use strict";l(jR,"insertByOffset");l(qR,"gapSelector");XT=class{static{l(this,"MemoryBlock")}constructor(e,t,n){this.offset=e,this.size=t,this.buffer=n}get end(){return this.offset+this.size}overlaps(e){return this.offset<e.end&&e.offset<this.end}},ex=class{static{l(this,"MemoryPool")}constructor(e,t=64,n="best-fit"){this.scope=e,this.alignment=t,this.strategy=n,this.blocks=[],this.peakUsage=0}allocate(e,t){let n=this._align(e),s=this._findFreeOffset(n);return this.placeAt(s,n,t)}placeAt(e,t,n){let s=this._align(t),o=new XT(e,s,n);jR(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,n=0,s=qR(this.strategy);for(let o of t){let i=this._align(n),a=o.offset-i,u=s.consider(i,a,e);if(u!==null)return u;o.end>n&&(n=o.end)}return s.result(this._align(n))}fragmentation(){if(this.peakUsage===0)return 0;let e=this.blocks.reduce((t,n)=>t+n.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)}},Yy=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=[],n=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,h]of this.inplaceMap)!o.has(c)||!o.has(h)||(a.has(h)||a.set(h,[]),a.get(h).push(c));if(a.size>0){let c=new Map;for(let h of e){let f=h.buffer;if(c.get(f)===1)continue;let d=[f];for(;d.length;){let m=d[d.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&&d.push(y);continue}if(d.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,h)=>{let f=c.size,d=h.size,m=f>0,g=d>0;if(m&&g){let y=d-f;if(y!==0)return y}else if(m!==g)return m?-1:1;return c.firstUse-h.firstUse}),p=new Map;for(let c of u){let h=c.buffer;if(this.inplaceMap.has(h)){let w=this.inplaceMap.get(h),T=this.assignments.get(w);if(T){this.assignments.set(h,{offset:T.offset,size:c.size,scope:c.scope,pool:T.pool,inplaceOf:w});continue}}let f=c.size;if(f===0)continue;if(f<0){this.assignments.set(h,{offset:0,size:0,scope:c.scope,pool:c.scope,inplaceOf:null,isDynamic:!0});continue}let d=c.scope,m=this.pools.get(d);m||(m=new ex(d,n),this.pools.set(d,m)),p.has(d)||p.set(d,[]);let g=p.get(d),y=i.get(h)??c.lastUse,_=this._interferenceOffset(g,c.firstUse,y,m._align(f),n,s),b=m.placeAt(_,f,h);jR(g,{firstUse:c.firstUse,lastUseEff:y,offset:b.offset,size:b.size}),this.assignments.set(h,{offset:b.offset,size:b.size,scope:d,pool:d,inplaceOf:null})}return this}_interferenceOffset(e,t,n,s,o,i){let a=[];for(let c of e)c.firstUse<=n&&t<=c.lastUseEff&&a.push([c.offset,c.offset+c.size]);let u=0,p=qR(i);for(let[c,h]of a){let f=Math.ceil(u/o)*o,d=c-f,m=p.consider(f,d,s);if(m!==null)return m;h>u&&(u=h)}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 n=this.pools.get(e);return n?n.peakUsage:0}let t=0;for(let[,n]of this.pools)t+=n.peakUsage;return t}}});function Zy(r,e,t){if(!r||typeof r!="object"||t.has(r))return!1;if(t.add(r),r.type==="BufferLoadNode"&&(!e||r.buffer===e))return!0;for(let n of Object.keys(r)){if(n==="_parent"||n==="_parentKey"||n==="_parentIdx")continue;let s=r[n];if(!(!s||typeof s!="object")){if(Array.isArray(s)){for(let o of s)if(Zy(o,e,t))return!0}else if(Zy(s,e,t))return!0}}return!1}function KR(r){let e=new Set,t=new Map,n=new Set,s=[r.body];for(;s.length>0;){let o=s.pop();if(!(!o||typeof o!="object"||n.has(o))){if(n.add(o),o.type==="BufferStoreNode"&&o.buffer){let i=o.buffer;for(let u of o.indices)if(Zy(u,null,new Set)){e.add(i);break}Zy(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 az(r,e){let t=new Set,n=[r];for(;n.length>0;){let s=n.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"&&n.push(u)}continue}i&&typeof i=="object"&&n.push(i)}}}}var tx,Jy,WR=M(()=>{"use strict";zR();GR();UR();De();Qb();tx=class{static{l(this,"MemoryPlan")}constructor(e,t,n){this.assignment=e,this.liveness=t,this.inplaceCandidates=n,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,n=this.inplaceCandidates.length;return{peakMemory:this.assignment.peakMemory(),scopeBreakdown:e,totalTemporaries:t,totalInplace:n,materializedReuse:this.aliasMap.size,assignments:this.assignment.assignments}}},Jy=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=Wy.analyze(e),n=t.getTemporaries(),s=[];this.enableInplace&&(s=Hy.analyze(e,t));let o=new Yy;return o.assign(n,s,this.alignment,this.allocStrategy),new tx(o,t,s)}planAndRewrite(e){let t=this.plan(e);return{func:this._insertAllocations(e,t),plan:t}}_insertAllocations(e,t){let n=t.liveness.getTemporaries();if(n.length===0)return e;let s=new Map;this.poolAllocation?this._assignPoolOffsets(e,t,n):(s=this._buildReuseAliases(n,t,e),s.size>0&&az(e.body,s)),t.aliasMap=s;let o=[...n].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 Ve(p,c.isDynamic?"dynamic":c.scope,i)))}return e.body=i,e._setChild("body",i),e}_assignPoolOffsets(e,t,n){let s=KR(e);for(let o of n){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,n){let s=KR(n),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 h=c.buffer,f=t.assignment.getAssignment(h);if(!f||f.inplaceOf||f.isDynamic||o.has(h)||h.numel()<=0||s.has(h))continue;let d=`${h.scope}|${h.dtype}|${h.shape.join(",")}|${h.strides.join(",")}`,m=u.get(d);m||(m=[],u.set(d,m)),m.push(c)}let p=new Map;for(let c of u.values()){if(c.length<2)continue;c.sort((f,d)=>f.firstUse-d.firstUse||a(f)-a(d));let h=new ku((f,d)=>f.lastUse-d.lastUse);for(let f of c){let d=h.peek();d&&d.lastUse<f.firstUse?(h.pop(),d.lastUse=a(f),h.push(d),p.set(f.buffer,d.rep)):h.push({rep:f.buffer,lastUse:a(f)})}}return p}};l(Zy,"exprLoadsBuffer");l(KR,"collectFreshZeroDependent");l(az,"rewriteBufferAliases")});var Qy,HR=M(()=>{"use strict";xi();WR();Qy=class extends Wn{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 Jy({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 n=performance.now(),{plan:s}=this._planner.planAndRewrite(e),o=s.getReport();t.trace.memoryStats(e.name,{durationMs:performance.now()-n,peakMemory:o.peakMemory,totalTemporaries:o.totalTemporaries,totalInplace:o.totalInplace})}}});function El(r){return r&&r.type==="IntImmNode"?r.value:null}function uz(r){if(!r||r.type!=="ForNode"||!YR.has(r.kind)||El(r.min)!==0)return null;let e=El(r.extent);if(e===null)return null;let t=r.body;if(!t||t.type!=="ForNode"||!YR.has(t.kind)||El(t.min)!==0)return null;let n=El(t.extent);if(n===null||n<=0)return null;let s=t.body;if(!s||s.type!=="IfThenElseNode"||s.elseBody)return null;let o=lz(s.condition,r.loopVar.name,t.loopVar.name,n);if(o===null||o<=0)return null;let i=Math.floor(o/n),a=o-i*n;if(a===0||i<1||e!==i+1)return null;let u=new Os;if(u.bind(r.loopVar.name,0,i-1),u.bind(t.loopVar.name,0,n-1),!Mo(u,s.condition))return null;let p=s.thenBody,c=new de(t.loopVar,new L(0),new L(n),t.kind,p,t.threadTag),h=new de(r.loopVar,new L(0),new L(i),r.kind,c,r.threadTag),f=ti(Ma(p),g=>g.type==="VariableNode"&&g.name===r.loopVar.name?new L(i):g),d=new Ne(t.loopVar.name,t.loopVar.dtype),m=new de(d,new L(0),new L(a),t.kind,f,t.threadTag);return new te([h,m])}function lz(r,e,t,n){if(!r)return null;let s,o;if(r.type==="MathOpNode"&&r.op==="<")s=r.a,o=r.b;else if(r.type==="CompareNode"&&r.direction==="lt")s=r.a,o=r.b;else return null;let i=El(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||El(a.b)!==n?null:i}var YR,Xy,JR=M(()=>{"use strict";xi();De();Gr();cf();Ku();Yc();YR=new Set([F.SERIAL,F.PARALLEL,F.VECTORIZED,F.UNROLLED]),Xy=class extends Wn{static{l(this,"LoopPartitionPass")}constructor(){super("LoopPartitionPass","loopPartition")}run(e,t){let n=ti(e.body,s=>uz(s)||s);return n&&n!==e.body&&(e.body=n,e._setChild("body",n)),e}};l(El,"intImm");l(uz,"partitionLoop");l(lz,"matchFlatGuard")});function e_(r){let e=r.body;if(!e||e.type!=="BlockNode")return null;let t=e.body;if(!t||t.type!=="BufferStoreNode")return null;let n=t,s=n.value;if(!s||s.type!=="MathOpNode"||!cz.has(s.op))return null;let o=null,i=null;if(s.a&&s.a.type==="BufferLoadNode"&&s.a.buffer===n.buffer?(o=s.a,i=s.b):s.b&&s.b.type==="BufferLoadNode"&&s.b.buffer===n.buffer&&(o=s.b,i=s.a),!o)return null;let a=nx(n.indices),u=nx(o.indices);if(a!==u||a.includes("?"))return null;let p=n.indices.map(f=>{if(f.type!=="VariableNode")return f;for(let d of e.iterVars)if(d.iterVar&&d.iterVar.name===f.name&&d.binding)return d.binding;return f}),c=r.loopVar.name,h=nx(p);return h.includes("?")||h.includes("$"+c)?null:{store:n,loadSide:o,valueSide:i,outerIndices:p,block:e,op:n.value.op}}function nx(r){return r.map(rx).join(",")}function rx(r){return r?r.type==="VariableNode"?"$"+r.name:r.type==="IntImmNode"?String(r.value):r.type==="MathOpNode"?"("+rx(r.a)+r.op+(r.b?rx(r.b):"")+")":"?":"?"}var cz,sx=M(()=>{"use strict";cz=new Set(["+","*","max","min"]);l(e_,"detectAccumulator");l(nx,"indicesKey");l(rx,"exprKey")});function uo(r){if(!(!r||typeof r!="object"||!r.type))switch(r.type){case"ForNode":{let e=e_(r);r.accumulator=e,e||uo(r.body);return}case"BlockNode":uo(r.body),r.initBody&&uo(r.initBody);return;case"SeqNode":for(let e of r.stmts)uo(e);return;case"LetStmtNode":case"AllocateNode":uo(r.body);return;case"IfThenElseNode":uo(r.thenBody),r.elseBody&&uo(r.elseBody);return;case"WhileNode":uo(r.condBody),uo(r.loopBody);return;default:return}}var t_,ZR=M(()=>{"use strict";xi();sx();t_=class extends Wn{static{l(this,"AccumulatorDetectionPass")}constructor(){super("AccumulatorDetectionPass","accumulatorDetect")}run(e,t){return uo(e.body),e}};l(uo,"annotateStmt")});function fz(r){let e=[],t=[r.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 n=null;for(let s of e){if(!s.name.includes("matmul"))return null;s.reads.length>=2&&s.writes.length>=1&&(n=s)}return n}function hz(r){let e=fz(r);if(!e)return null;let t=e.reads[0].buffer,n=e.reads[1].buffer,s=e.writes[0].buffer;if(!QR.has(t.dtype)||!QR.has(n.dtype)||s.dtype!=="f32"||t.shape.length!==2||n.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%pz===0))return null;let u=new Set;for(let[,p]of r.bufferMap)u.add(p.name);return!u.has(t.name)||!u.has(n.name)||!u.has(s.name)?null:{M:o,N:i,K:a,a:t.name,b:n.name,c:s.name}}var pz,QR,n_,XR=M(()=>{"use strict";xi();fa();pz=16,QR=new Set(["f16","bf16"]);l(fz,"findMatmulBlock");l(hz,"detectWmmaMatmul");n_=class extends Wn{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 n=hz(e);n&&(new Sn(e).tensorize("wmma_16x16x16_f16f16f32",n),t&&t.trace&&t.trace.explainsEnabled&&t.trace.explain("tensorize",e.name,"wmma_16x16x16_f16f16f32",`auto-tensorized ${n.M}x${n.N}x${n.K} f16 GEMM`,{target:this.target.name}))}}});function dz(r){let e=[],t=new Map,n=new Map,s=[{node:r,block:null}];for(;s.length>0;){let{node:o,block:i}=s.pop();if(o.type==="BufferLoadNode"&&o.buffer)t.set(o.buffer.name,(t.get(o.buffer.name)||0)+1),i&&i.reads.add(o.buffer.name);else if(o.type==="BufferStoreNode"&&o.buffer&&i){i.writes.add(o.buffer.name);let u=n.get(o.buffer.name);u||(u=new Set,n.set(o.buffer.name,u)),u.add(i.name)}let a=i;o.type==="BlockNode"&&(a={name:o.name,writes:new Set,reads:new Set,hasInit:o.initBody!=null},e.push(a));for(let u of en(o))s.push({node:u,block:a})}return{blocks:e,loadCount:t,storeWriters:n}}var r_,eM=M(()=>{"use strict";xi();fa();ro();Gr();l(dz,"analyzeFunc");r_=class extends Wn{static{l(this,"InlineReindexPass")}constructor(e){super("InlineReindexPass","scheduling"),this.config=e,this.target=e.target,this.snapshotPoint="afterInlineReindex"}run(e,t){if(!this.target.isGPU()||this.target.isWebGPU()||e.cublasInfo||e._tensorIntrin||i0(e))return;let n=this.config.scheduling;if(!(n.enabled||n.gpuTiling||n.autotune))return;let s=new Set;for(let[,c]of e.bufferMap)s.add(c.name);let{blocks:o,loadCount:i,storeWriters:a}=dz(e.body),u=new Sn(e),p=!1;for(let c of o){if(c.hasInit||c.writes.size===0)continue;let h=[...c.writes];if(h.every(f=>!s.has(f)&&a.get(f).size===1)&&h.some(f=>(i.get(f)||0)>0)&&[...c.reads].every(f=>(a.get(f)?a.get(f).size:0)<=1))try{u.computeInlineBlock(c.name),p=!0}catch{}}p&&s0(e)}}});function ox(r,e,t=mz){return t.filter(n=>n.phase===r).sort((n,s)=>n.priority-s.priority).map(n=>n.factory(e)).filter(Boolean)}var mz,tM=M(()=>{"use strict";mz=[];l(ox,"tirPassesForPhase")});function nM(r){let e=[];for(let t of ox("pre",r))e.push(t);e.push(new r_(r)),r.optimization.tensorize&&e.push(new n_(r)),e.push(new Uy(r)),r.optimization.loopPartition&&e.push(new Xy),e.push(new Ky),e.push(new Qy(r)),r.optimization.detectAccumulators&&e.push(new t_);for(let t of ox("post",r))e.push(t);return e}var rM=M(()=>{"use strict";CR();$R();HR();JR();ZR();XR();eM();tM();l(nM,"buildTirPipeline")});var bf,sM=M(()=>{"use strict";De();bf=class{static{l(this,"TensorVerifier")}constructor(){this.errors=[],this.boundVars=new Set}verify(e){if(this.errors=[],this.boundVars.clear(),!(e instanceof Jo))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 n=t.pop();if(n)switch(n.type){case"SeqNode":for(let s=n.stmts.length-1;s>=0;s--)t.push(n.stmts[s]);break;case"ForNode":this.boundVars.has(n.loopVar.name)&&this.errors.push(`Loop variable ${n.loopVar.name} already bound`),this.boundVars.add(n.loopVar.name),t.push({type:"_unbind",name:n.loopVar.name}),t.push(n.body);break;case"BlockNode":for(let s of n.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=n.iterVars.length-1;s>=0;s--)n.iterVars[s].iterVar&&t.push({type:"_unbind",name:n.iterVars[s].iterVar.name});t.push(n.body),n.initBody&&t.push(n.initBody);break;case"AllocateNode":n.buffer||this.errors.push("Allocate missing buffer"),t.push(n.body);break;case"LetStmtNode":this._visitExpr(n.value),this.boundVars.add(n.variable.name),t.push({type:"_unbind",name:n.variable.name}),t.push(n.body);break;case"IfThenElseNode":this._visitExpr(n.condition),n.elseBody&&t.push(n.elseBody),t.push(n.thenBody);break;case"WhileNode":n.condVar||this.errors.push("WhileNode missing condition variable"),t.push(n.loopBody),t.push(n.condBody);break;case"BufferStoreNode":if(n.buffer||this.errors.push("BufferStore missing buffer"),(!n.indices||n.indices.length!==n.buffer.shape.length)&&this.errors.push(`BufferStore rank mismatch for ${n.buffer?n.buffer.name:"unknown"}`),n.indices)for(let s of n.indices)this._visitExpr(s);this._visitExpr(n.value);break;case"EvaluateNode":this._visitExpr(n.value);break;case"_unbind":this.boundVars.delete(n.name);break;default:this._visitExpr(n);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 ix(r){let e=[];r.functionCount===0&&e.push(new vt("Module has no functions"));for(let t of r)wf(t,e);return e}function wf(r,e=[]){if(!r.entryBlock)return e.push(new vt("Function has no entry block",null,r)),e;r.entryBlock.arguments.length!==r.inputTypes.length&&e.push(new vt(`Entry block has ${r.entryBlock.arguments.length} args, expected ${r.inputTypes.length}`,null,r));for(let s=0;s<r.entryBlock.arguments.length;s++){let o=r.entryBlock.arguments[s],i=r.inputTypes[s];i&&!o.type.equals(i)&&e.push(new vt(`Block arg ${s} type ${Rn(o.type)} != expected ${Rn(i)}`,null,r))}let t=new Set;for(let s of r.entryBlock.arguments)t.add(s);for(let s of r.body)oM(s,t);for(let s of r.body)iM(s,r,t,e);let n=r.getReturnOp();if(!n)e.push(new vt("Missing return op",null,r));else{n.numOperands!==r.outputTypes.length&&e.push(new vt(`Return has ${n.numOperands} operands, function declares ${r.outputTypes.length} outputs`,n,r));for(let s=0;s<Math.min(n.numOperands,r.outputTypes.length);s++){let o=n.getOperand(s).type,i=r.outputTypes[s];i instanceof I&&o instanceof I&&(o.dtype!==i.dtype&&e.push(new vt(`Return operand ${s} dtype ${o.dtype} != expected ${i.dtype}`,n,r)),o.shapeCompatible(i)||e.push(new vt(`Return operand ${s} shape incompatible: ${Rn(o)} vs ${Rn(i)}`,n,r)))}}return e}function oM(r,e){for(let t of r.arguments)e.add(t);for(let t of r)for(let n=0;n<t.numResults;n++){let s=t.getResult(n);s&&e.add(s)}}function gz(r,e,t){let n=new Set;for(let u of r)n.add(u);let s=1,o=2,i=new Map,a=new Set;for(let u of r){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],h=c.op;if(c.i<h.numOperands){let f=h.getOperand(c.i);c.i++;let d=f&&f.definingOp;if(!d||!n.has(d))continue;let m=i.get(d);m===s?a.has(d)||(a.add(d),t.push(new vt("participates in a value dependency cycle",d,e))):m===void 0&&(i.set(d,s),p.push({op:d,i:0}));continue}i.set(h,o),p.pop()}}}function iM(r,e,t,n){gz(r,e,n);for(let s of r)yz(s,e,t,n);if(r.size>0){let s=r.lastOp;if(r.parentRegion&&r.parentRegion.parentOp){let o=r.parentRegion.parentOp.opName,i=ne.get(o);i&&i.hasRegions&&(s.isTerminator()||n.push(new vt(`Block in ${o} region must end with terminator, got '${s.opName}'`,s,e)))}}}function yz(r,e,t,n){for(let o=0;o<r.numOperands;o++){let i=r.getOperand(o);if(!i){n.push(new vt(`Operand ${o} is null`,r,e));continue}if(!(i instanceof Wo)){n.push(new vt(`Operand ${o} is not a Value`,r,e));continue}t.has(i)||n.push(new vt(`Operand ${o} used before definition`,r,e))}for(let o=0;o<r.numResults;o++){let i=r.getResult(o);i||n.push(new vt(`Result ${o} is null`,r,e)),i&&i.definingOp!==r&&n.push(new vt(`Result ${o} definingOp mismatch`,r,e))}let s=ne.get(r.opName);if(!s){!["return","yield"].includes(r.opName)&&!ne.has(r.opName)&&n.push(new vt(`Unknown op '${r.opName}'`,r,e));return}s.numOperands>=0&&r.numOperands!==s.numOperands&&n.push(new vt(`'${r.opName}' expects ${s.numOperands} operands, got ${r.numOperands}`,r,e)),s.numResults>=0&&r.numResults!==s.numResults&&n.push(new vt(`'${r.opName}' expects ${s.numResults} results, got ${r.numResults}`,r,e));for(let o of s.attrs)o.required&&!r.hasAttr(o.name)&&n.push(new vt(`'${r.opName}' missing required attribute '${o.name}'`,r,e));if(s.hasRegions&&s.numRegions>0&&r.numRegions!==s.numRegions&&n.push(new vt(`'${r.opName}' expects ${s.numRegions} regions, got ${r.numRegions}`,r,e)),s.verify){let o=s.verify(r);if(o)for(let i of o)n.push(new vt(i,r,e))}if(s.inferResultTypes&&r.numResults>0){let o=[];for(let a=0;a<r.numOperands;a++)o.push(r.getOperand(a).type);let i=s.inferResultTypes(o,r.attributes,r.results.map(a=>a.type));if(i)for(let a=0;a<Math.min(i.length,r.numResults);a++){let u=r.getResult(a).type,p=i[a];p instanceof I&&u instanceof I&&(u.dtype!==p.dtype&&n.push(new vt(`Result ${a} dtype ${u.dtype} != inferred ${p.dtype}`,r,e)),u.shapeCompatible(p)||n.push(new vt(`Result ${a} shape ${Rn(u)} incompatible with inferred ${Rn(p)}`,r,e)))}}for(let o of r.regions){let i=new Set(t);for(let a of o)oM(a,i);for(let a of o)iM(a,e,i,n)}}var vt,aM=M(()=>{"use strict";Oe();ch();At();vt=class{static{l(this,"VerificationError")}constructor(e,t=null,n=null){this.message=e,this.op=t,this.func=n}toString(){let e="";return this.func&&(e+=`[${this.func.name}] `),this.op&&(e+=`op '${this.op.opName}' (id=${this.op.id}): `),e+this.message}};l(ix,"verifyModule");l(wf,"verifyFunction");l(oM,"collectScopeDefs");l(gz,"detectCycles");l(iM,"verifyBlock");l(yz,"verifyOperation")});var s_,ax,Ll,ux,lx=M(()=>{"use strict";_T();Oe();s_=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 n=e[t];n<this.min&&(this.min=n),n>this.max&&(this.max=n)}this.count+=e.length,this.histogram&&this.histogram.update(e)}enableHistogram(e=2048){this.histogram=new ax(e)}},ax=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 n=this.numBins/t;for(let s=0;s<e.length;s++){let o=Math.min(this.numBins-1,Math.floor((e[s]-this.rangeMin)*n));this.bins[o]++}this.totalCount+=e.length}_initRange(e){let t=1/0,n=-1/0;for(let s=0;s<e.length;s++)e[s]<t&&(t=e[s]),e[s]>n&&(n=e[s]);t===n&&(n=t+1),this.rangeMin=t,this.rangeMax=n,this.initialized=!0}_expandAndRebucket(e){let t=this.rangeMin,n=this.rangeMax;for(let c=0;c<e.length;c++)e[c]<t&&(t=e[c]),e[c]>n&&(n=e[c]);let s=new Float64Array(this.bins),o=this.rangeMin,a=this.rangeMax-o;this.rangeMin=t,this.rangeMax=n;let u=n-t,p=this.numBins/u;if(this.bins.fill(0),a>0){let c=a/this.numBins;for(let h=0;h<this.numBins;h++){if(s[h]===0)continue;let f=o+(h+.5)*c,d=Math.min(this.numBins-1,Math.floor((f-t)*p));this.bins[d]+=s[h]}}for(let c=0;c<e.length;c++){let h=Math.min(this.numBins-1,Math.floor((e[c]-t)*p));this.bins[h]++}this.totalCount+=e.length}computePercentileThreshold(e){let t=this.totalCount*e,n=0,s=(this.rangeMax-this.rangeMin)/this.numBins;for(let o=0;o<this.numBins;o++)if(n+=this.bins[o],n>=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 n=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 h=Math.floor(c*a),f=Math.min(i,Math.floor((c+1)*a)),d=0;for(let g=h;g<f;g++)d+=t[g];let m=f-h;if(m>0){let g=d/m;for(let y=h;y<f;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,n=i)}let o=(this.rangeMax-this.rangeMin)/this.numBins;return this.rangeMin+n*o}},Ll=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 n=0;n<t.numResults;n++){let s=t.getResult(n);if(s.type instanceof I&&Lt(s.type.dtype)){let o=new s_;(this.mode==="entropy"||this.mode==="percentile")&&o.enableHistogram(),this.observers.set(s,o)}}for(let t of e.args)if(t.type instanceof I&&Lt(t.type.dtype)){let n=new s_;(this.mode==="entropy"||this.mode==="percentile")&&n.enableHistogram(),this.observers.set(t,n)}}observe(e,t){let n=this.observers.get(e);n&&n.update(t)}getResult(){return new ux(this.observers,this.mode)}},ux=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,n){let s=this.getRange(e);if(!s)return null;let o=t===kr.PER_TENSOR_ASYMMETRIC;if(this._mode==="percentile"&&this._observers.get(e).histogram){if(o)return hr.fromRange(s.min,s.max,t,n);let a=this._observers.get(e).histogram.computePercentileThreshold(.999);return hr.fromRange(-a,a,t,n)}if(this._mode==="entropy"&&this._observers.get(e).histogram){if(o)return hr.fromRange(s.min,s.max,t,n);let i=this._observers.get(e).histogram,a=n==="ui8"?256:255,u=i.computeEntropyThreshold(a);return hr.fromRange(-u,u,t,n)}return hr.fromRange(s.min,s.max,t,n)}hasData(e){let t=this._observers.get(e);return t&&t.count>0}values(){return this._observers.keys()}}});function bz(r,e){let t=[],n=new Set;for(let s of r.ops())if(e.has(s.opName))for(let o=0;o<s.numOperands;o++){let i=s.getOperand(o);if(n.has(i)||!(i.type instanceof I)||!Lt(i.type.dtype))continue;let a=i.definingOp;a&&a.opName==="constant"||(n.add(i),t.push(i))}return t}function wz(r,e){let t=e.map(d=>d.type),n=new un(r.name,r.inputTypes,[...r.outputTypes,...t]),s=new Map,o=r.entryBlock.arguments,i=n.entryBlock.arguments;for(let d=0;d<o.length;d++)s.set(o[d],i[d]);let a=r.entryBlock.opsArray(),u=new Set(a),p=wo(a,d=>u.has(d),"ignore"),c=new Map;for(let d of p)d.opName!=="return"&&c.set(d,d.clone(s));for(let d of a)d.opName!=="return"&&n.entryBlock.pushOp(c.get(d));let h=r.getReturnOp(),f=[];if(h)for(let d of h.operands)f.push(s.get(d)||d);for(let d of e)f.push(s.get(d));return n.entryBlock.pushOp(new Re("return",f,[],{})),n}function uM(r,e,t,n={}){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=n.quantizableOps||_z,o=n.mode||"minmax",i=bz(r,s),a=new Ll(o);a.attach(r);let u=a.getResult();if(i.length===0)return u;let p=wz(r,i),c=new wn("__calibrate__");c.addFunction(p);let h=r.outputTypes.length,d=(n.compileFn||Tz)(c,e);for(let m of t){let g=Array.isArray(m)?m:[m],y=p.outputTypes.map(_=>new Float32Array(Math.max(1,_.numel())));d.run(r.name,...g,...y);for(let _=0;_<i.length;_++)a.observe(i[_],y[h+_])}return u}function Tz(r,e){throw new Error("collectCalibration: opts.compileFn is required (pass the compileModule function to avoid a circular import)")}var _z,lM=M(()=>{"use strict";xs();Yo();sr();Oe();ji();lx();_z=new Set(["dot","conv","add","mul","sub"]);l(bz,"activationTargets");l(wz,"buildCaptureFunction");l(uM,"collectCalibration");l(Tz,"defaultCompileFn")});function px(r){let e=0;for(let t=0;t<r.numResults;t++){let n=r.getResult(t).type;n instanceof I&&n.isFullyStatic&&(e+=n.sizeInBytes())}return e}var cx,Cl,fx,o_,cM=M(()=>{"use strict";At();Oe();rf();cx=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+=px(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}=_i(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(px,"estimateOpMemory");Cl=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||{}}}},fx=class{static{l(this,"PartitionResult")}constructor(e,t,n){this.partitions=e,this.opToPartition=t,this.transferEdges=n}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}},o_=class{static{l(this,"GraphPartitioner")}constructor(e){this.config=e instanceof Cl?e:new Cl(e),this._supportCache=new Map,this._buildSupportMap()}partition(e){let t=this._collectPartitionableOps(e),n=this._assignTargets(t),s=this._buildPartitions(t,n),o=this._mergeSmallPartitions(s,n),i=this._computeTransferEdges(o),a=new Map;for(let u of o)for(let p of u.ops)a.set(p,u);return new fx(o,a,i)}_buildSupportMap(){for(let e of this.config.targets){let t=new Set;for(let n of ne.names())this._targetSupportsOp(e,n)&&t.add(n);this._supportCache.set(e,t)}}_targetSupportsOp(e,t){if(e.hasLibraryOp(t))return!0;let n=ne.get(t);return n?n.isConstant||n.isTerminator?!0:e.isGPU()?n.isElementwise||n.isReduction||n.isBroadcast||n.isInjective||n.getAttr("gpuCapable")===!0:e.isCPU()?!0:e.isWasm()?!n.isOpaque&&t!=="custom_call":!1:!1}_collectPartitionableOps(e){let t=[];for(let n of e.ops()){let s=ne.get(n.opName);!s||s.isTerminator||t.push(n)}return t}_assignTargets(e){let t=new Map;for(let n of e){let s=n.getAttr("device");if(s){let a=this._resolveDeviceAttr(s);if(a){t.set(n,a);continue}}let o=this.config.opTargetOverrides.get(n.opName);if(o){t.set(n,o);continue}let i=this._selectBestTarget(n);t.set(n,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],n=-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>n&&(n=i,t=s)}return t}_scoreTargetForOp(e,t){let n=ne.get(t.opName);if(!n)return 0;let s=0;if(e.hasLibraryOp(t.opName)&&(s+=100),n.isReduction||n.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:n}=nf(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)+px(a)},"onAttach")}),s=new Map;for(let i of n){let a=new cx(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,n){let s=this.config.memoryLimits.get(n.name);if(!s)return!0;let o=px(t);return e.memoryBytes+o<=s}_mergeSmallPartitions(e,t){if(e.length<=1)return e;let n=new Map;for(let d of e)for(let m of d.ops)n.set(m,d);let s=new Set,o=new Map,i=new Map;for(let d of e)s.has(d)||o.set(d,new Set);for(let d of e){if(s.has(d))continue;let m=o.get(d);for(let g of d.ops)for(let y=0;y<g.numResults;y++)for(let _ of g.getResult(y).uses()){let b=n.get(_.user);b&&b!==d&&m.add(b)}}let a=l((d,m)=>{let g=o.get(d);for(let y of o.get(m))y!==d&&g.add(y);g.delete(m);for(let[y,_]of o)_.has(m)&&(_.delete(m),y!==d&&_.add(d));o.delete(m),i=new Map},"mergeSucc"),u=l(d=>{let m=i.get(d);if(m)return m;m=new Set;let g=[...o.get(d)];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(d,m),m},"reachOf"),p=l((d,m)=>{for(let g of o.get(d))if(g!==m&&u(g).has(m))return!0;return!1},"pathThroughIntermediate"),c=l((d,m)=>p(d,m)||p(m,d),"mergeCreatesCycle"),h=[];for(let d=0;d<e.length;d++){let m=e[d];if(s.has(m))continue;if(m.size>=this.config.minPartitionSize){h.push(m);continue}let g=-1,y=-1/0;for(let _=0;_<e.length;_++){if(d===_)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)n.set(b,_);s.add(m)}else h.push(m)}let f=new Set;for(let d of h)for(let m of d.ops)f.add(m);for(let d of e){let m=!1;for(let g of d.ops)if(!f.has(g)){m=!0;break}if(m){h.push(d);for(let g of d.ops)f.add(g)}}return h}_mergeScore(e,t){let n=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))&&n++;return n}_computeTransferEdges(e){let t=[],n=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),h=`${p.id}|${o.id}|${c.id}`;n.has(h)||(n.add(h),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),n=new Map,s=new Map;for(let u of e)n.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),n.set(u,n.get(u)+1))}let o=[];for(let u of e)n.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=n.get(p)-1;n.set(p,c),c===0&&o.push(p)}}return i}}});var i_,a_,pM=M(()=>{"use strict";nn();sr();xs();cM();rf();Kt();i_=class extends Ge{static{l(this,"GraphPartitionPass")}constructor(e={}){super("GraphPartitionPass"),this.partitionerConfig=new Cl(e),this.partitionResult=null}run(e,t){if(this.partitionerConfig.targets.length<2)return oe.UNCHANGED;let n=new o_(this.partitionerConfig);return this.partitionResult=n.partition(e),this.partitionResult.numPartitions<=1?oe.UNCHANGED:(this._annotateOps(),this._insertTransferOps(e),this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,numPartitions:this.partitionResult.numPartitions,transferEdges:this.partitionResult.transferEdges.length,level:ue.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,n=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 Re("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 h=s.get(i.dst),f=h&&h.get(a)||o.get(i.dst)||null,d=a.definingOp;if(f&&d&&n.has(d)&&n.has(f)&&n.get(d)>=n.get(f))t.insertAfter(c,d);else if(f)t.insertBefore(c,f);else if(d&&n.has(d))t.insertAfter(c,d);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))],n=new Map,s=new Map;for(let o of t)n.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=n.get(i);for(let u=0;u<o.numOperands;u++){let p=o.getOperand(u);a.has(p)||a.set(p,o)}}return{useMap:n,firstInPart:s}}_buildOrderIndex(e){let t=new Map,n=0;for(let s of e.ops())t.set(s,n++);return t}},a_=class extends Ge{static{l(this,"PartitionMaterializationPass")}constructor(e={}){super("PartitionMaterializationPass"),this.targets=e.targets||[]}run(e,t){let n=this._collectPartitions(e);if(n.size<=1)return oe.UNCHANGED;let s=this._getModule(e);if(!s)return oe.UNCHANGED;let o=this._materializePartitions(e,n);for(let i of o)s.addFunction(i);return this._rewriteOriginalFunction(e,o,n),this.trace&&this.trace.level>=ue.DEBUG&&this.trace.emit({type:"pass_detail",passName:this.name,partitions:n.size,subFunctions:o.length,level:ue.DEBUG}),e.bumpVersion(),oe.CHANGED}_collectPartitions(e){let t=new Map;for(let n of e.ops()){let s=n.getAttr("partition_id");s!==void 0&&(t.has(s)||t.set(s,{id:s,target:n.getAttr("partition_target"),ops:[]}),t.get(s).ops.push(n))}return t}_getModule(e){return e._module||null}_materializePartitions(e,t){let n=[];for(let[s,o]of t){let i=new Set(o.ops),{inputs:a,outputs:u}=_i(i,o.ops),p=a.map(_=>_.type),c=u.map(_=>_.type),h=`${e.name}_partition_${s}`,f=new un(h,p,c);f._partitionTarget=o.target;let d=new Map;for(let _=0;_<a.length;_++)d.set(a[_],f.args[_]);let m=Ba(o.ops);for(let _ of m)f.entryBlock.pushOp(_.clone(d));let g=u.map(_=>d.get(_)||_),y=new Re("return",g,[]);f.entryBlock.pushOp(y),n.push(f)}return n}_rewriteOriginalFunction(e,t,n){for(let s of t)e.setAttr?.(`sub_${s.name}`,s._partitionTarget)}}});function Tf(r){return oc(r,-1)}function xz(r,e,t){let n=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 h=Tf(c.shape);if(h<0)return!0;let f=h*bn(c.dtype);f>s&&(s=f)}let o=r.getAttr("num_carry"),i=r.getAttr("num_xs"),a=0;for(let u=0;u<o;u++){let p=r.getOperand(i+u).type,c=p&&p.shape?Tf(p.shape):-1;c>0&&(a+=c*bn(p.dtype))}return 3*s+2*a>n}function Sz(r,e){let n=r.regions[0].entryBlock,s=r.getAttr("num_carry"),o=[],i=null;for(let b of n.ops()){if(b.opName==="yield"){i=b;continue}o.push(b)}if(!i)return null;let a=new Set(o),u=new Set(n.arguments),p=[],c=new Set,h=[],f=new Set;for(let b of o)for(let w=0;w<b.numOperands;w++){let T=b.getOperand(w),S=T.definingOp;if(!u.has(T)&&!(S&&a.has(S))){if(S&&oo(S)){f.has(S)||(f.add(S),h.push(S));continue}c.has(T)||(c.add(T),p.push(T))}}let d=[...n.arguments,...p];for(let b of d)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 un(e,d.map(b=>b.type),m.map(b=>b.type)),y=new Map;for(let b=0;b<d.length;b++)y.set(d[b],g.args[b]);for(let b of h)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 Re("return",_,[])),{bodyFunc:g,captured:p,numCarry:s,numYs:m.length-s})}function kz(r,e,t,n,s){let{getSlot:o,newSlot:i,steps:a,addedFuncs:u}=n,p=new wn(r.name+"_mod");p.addFunction(r);let c=Ti(p,s);if(c){let h=c.plan,f=e.length,d=new Map,m=!0;for(let g of h.returnFixups||[]){if(g.kind!=="copy"){m=!1;break}d.set(g.pos,g.srcSlot)}if(m){let g=l(T=>{let S=f+T;return d.has(S)?d.get(S):h.argSlots[S]},"retSlot"),y=new Map;for(let T=0;T<f;T++)y.set(h.argSlots[T],o(e[T]));for(let T of h.intermediates)y.set(T.slot,i(T.shape,T.dtype));for(let T=0;T<t.length;T++){let S=g(T);y.has(S)||y.set(S,o(t[T]))}let _=l(T=>{let S=y.get(T);return S===void 0?null:S},"mapSlot"),b=[],w=!0;for(let T of h.steps){let S=T.inputSlots.map(_),x=T.outputSlots.map(_);if(S.includes(null)||x.includes(null)){w=!1;break}b.push({name:T.name,inputSlots:S,outputSlots:x})}if(w){for(let T of b)a.push(T);for(let T of p.functions())u.push(T);return!0}}}return a.push({name:r.name,inputSlots:e.map(o),outputSlots:t.map(o)}),u.push(r),!0}function fM(r,e,t){let n=my({ops:r,opSet:new Set(r)},e,new Map);return n?kz(n.subFunc,n.inputs,n.outputs,t,1):!1}function vz(r,e,t){let{getSlot:n,newSlot:s,steps:o,scanLoops:i,addedFuncs:a}=t,u=r.regions[0];if(!u||!u.entryBlock)return!1;let p=Sz(r,e);if(!p)return!1;let{bodyFunc:c,captured:h,numCarry:f,numYs:d}=p,m=r.getAttr("num_xs"),g=new wn(e+"_mod");g.addFunction(c);let y=Ti(g,2),_=y?y.plan:null,b=[],w=[];for(let P=0;P<f;P++){let V=r.getOperand(m+P).type;b.push(V.shape),w.push(V.dtype)}let T=[],S=[];for(let P=0;P<m;P++){let V=r.getOperand(P).type;T.push(V.shape.slice(1)),S.push(V.dtype)}let x=[],k=[];for(let P=0;P<d;P++){let V=r.getResult(f+P).type;x.push(V.shape.slice(1)),k.push(V.dtype)}let A=b.map((P,V)=>s(P,w[V])),N=b.map((P,V)=>s(P,w[V])),v=T.map((P,V)=>s(P,S[V])),O=[],R=[],B=[],E=[];for(let P=0;P<f;P++)R.push(n(r.getOperand(m+P)));for(let P=0;P<f;P++)B.push(n(r.getResult(P)));for(let P of h)E.push(n(P));let z=[],D=[];for(let P=0;P<m;P++)z.push(n(r.getOperand(P)));for(let P=0;P<d;P++)D.push(n(r.getResult(f+P)));let J=o.length;if(!_)O=x.map((P,V)=>s(P,k[V])),o.push({name:c.name,inputSlots:[...v,...A,...E],outputSlots:[...N,...O]}),a.push(c);else{let P=m+f+h.length,V=new Map;for(let G of _.returnFixups||[]){if(G.kind!=="copy")return!1;V.set(G.pos,G.srcSlot)}let he=l(G=>{let K=P+G;return V.has(K)?V.get(K):_.argSlots[K]},"bodyReturnSlot"),le=new Map;for(let G=0;G<m;G++)le.set(_.argSlots[G],v[G]);for(let G=0;G<f;G++)le.set(_.argSlots[m+G],A[G]);for(let G=0;G<h.length;G++)le.set(_.argSlots[m+f+G],E[G]);for(let G of _.intermediates)le.set(G.slot,s(G.shape,G.dtype));for(let G=0;G<f;G++){let K=he(G);le.has(K)||le.set(K,N[G])}for(let G=0;G<d;G++){let K=he(f+G),ie=le.get(K);ie===void 0&&(ie=s(x[G],k[G]),le.set(K,ie)),O.push(ie)}let Ie=l(G=>{let K=le.get(G);return K===void 0?null:K},"mapSlot");for(let G of _.steps){let K=G.inputSlots.map(Ie),ie=G.outputSlots.map(Ie);if(K.includes(null)||ie.includes(null))return!1;o.push({name:G.name,inputSlots:K,outputSlots:ie})}for(let G of g.functions())a.push(G)}let j=o.length,U=r.getOperand(0).type.shape[0];return typeof U!="number"||U<0?!1:(i.push({T:U,loopStart:J,loopEnd:j,carry:b.map((P,V)=>({a:A[V],b:N[V],initSlot:R[V],finalSlot:B[V],bytes:Tf(P)*bn(w[V])})),xs:v.map((P,V)=>({xtSlot:P,xsSlot:z[V],stepBytes:Tf(T[V])*bn(S[V])})),ys:O.map((P,V)=>({ytSlot:P,ysSlot:D[V],stepBytes:Tf(x[V])*bn(k[V])}))}),!0)}function hx(r,e,t=!1){if(!e||typeof e.isWebGPU!="function"||!e.isWebGPU()||r.functionCount!==1)return null;let n=r.functions().next().value,s=n.getReturnOp();if(!s)return null;let o=[];for(let R of n.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&&xz(B,E,e)){R=!0;break}}if(!R)return null}let u=[];for(let R of i)oo(R)||qt(R.opName)||u.push(R);let p=Ba(u),c=new Map,h=0,f=l(R=>{let B=c.get(R);return B===void 0&&(B=h++,c.set(R,B)),B},"getSlot");for(let R of n.args)f(R);let d=[],m=l((R,B)=>{let E=h++;return d.push({slot:E,shape:[...R],dtype:B}),E},"newSlot"),g=[],y=[],_=[],b={getSlot:f,newSlot:m,steps:g,scanLoops:y,addedFuncs:_},w=new Set(o),T=[],S=0,x=0;for(let R of p)if(w.has(R)){if(T.length&&!fM(T,n.name+"_seg"+S++,b)||(T=[],!vz(R,n.name+"_scan"+x++,b)))return null}else T.push(R);if(T.length&&!fM(T,n.name+"_seg"+S++,b))return null;let k=n.args.map(f),A=[],N=new Set;for(let R=0;R<s.numOperands;R++){let B=s.getOperand(R),E=k.length;if(c.has(B)&&!(B.isBlockArgument&&B.isBlockArgument())){let z=f(B);N.has(z)?(k.push(h++),A.push({pos:E,kind:"copy",srcSlot:z})):(N.add(z),k.push(z))}else if(B.isBlockArgument&&B.isBlockArgument())k.push(h++),A.push({pos:E,kind:"copy",srcSlot:f(B)});else return null}let v=new Set(k),O=new Set;for(let[R,B]of c)if(!(v.has(B)||O.has(B))){if(O.add(B),!R.type||!R.type.isFullyStatic)return null;d.some(E=>E.slot===B)||d.push({slot:B,shape:[...R.type.shape],dtype:R.type.dtype})}r.removeFunction(n.name);for(let R of _)r.addFunction(R);return{plan:{numSlots:h,argSlots:k,intermediates:d,steps:g,returnFixups:A,scanLoops:y}}}var hM=M(()=>{"use strict";sr();xs();Yo();gy();Tr();Vn();Oe();l(Tf,"numel");l(xz,"isScanOversized");l(Sz,"buildScanBodyFunction");l(kz,"inlineNativeSplit");l(fM,"emitSegment");l(vz,"emitScanLoop");l(hx,"splitGraphForScan")});function xf(r){dx.push(r),dx.sort((e,t)=>(e.priority??100)-(t.priority??100))}function Nz(r){for(let e of dx)if(e.applies(r))return e;return null}function dM(r,e){let t=Nz(e);return t?t.run(r,e):null}var dx,mM=M(()=>{"use strict";gy();hM();dx=[];l(xf,"registerGraphSplitStrategy");l(Nz,"selectGraphSplitStrategy");l(dM,"splitGraph");xf({name:"cublas",priority:10,applies:l(r=>r.config.matmulBackend==="cublas","applies"),run:l(r=>MB(r),"run")});xf({name:"cuda-attention",priority:15,applies:l(r=>r.cudaAttention,"applies"),run:l(r=>Ti(r,1),"run")});xf({name:"cuda-native-matmul-chain",priority:20,applies:l(r=>r.cudaMatmulChain,"applies"),run:l(r=>Ti(r),"run")});xf({name:"cuda-native-conv-chain",priority:25,applies:l(r=>r.cudaConvChain,"applies"),run:l(r=>Ti(r),"run")});xf({name:"webgpu",priority:30,applies:l(r=>r.isWebGPU,"applies"),run:l((r,e)=>{let t=hx(r,e.target);return!t&&BB(r,e.target.maxThreadsPerBlock||256)&&(t=hx(r,e.target,!0),t||(t=Ti(r,2))),t},"run")})});function gM(r,e){let t=new id;for(let[,n]of r.bufferMap)t.paramBuffers.add(n.name);return Az(r.body,t,e),Bz(r,t,e),Rz(r.body,t),t}function Az(r,e,t){let n=[r];for(;n.length>0;){let s=n.pop();if(!(!s||typeof s!="object")){switch(s.type){case"ForNode":if(s.loopVar&&e.locals.set(s.loopVar.name,"i32"),s.kind===F.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,ei(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&&!cv(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,ei(o.iterVar.dtype)),o.binding&&typeof o.binding=="object"&&o.binding.type&&n.push(o.binding);break}if(s.body&&n.push(s.body),s.value&&typeof s.value=="object"&&s.value.type&&n.push(s.value),s.stmts)for(let o of s.stmts)n.push(o);if(s.thenBody&&n.push(s.thenBody),s.elseBody&&n.push(s.elseBody),s.initBody&&n.push(s.initBody),s.condBody&&n.push(s.condBody),s.loopBody&&n.push(s.loopBody),s.condition&&typeof s.condition=="object"&&s.condition.type&&n.push(s.condition),s.a&&typeof s.a=="object"&&s.a.type&&n.push(s.a),s.b&&typeof s.b=="object"&&s.b.type&&n.push(s.b),s.expr&&typeof s.expr=="object"&&s.expr.type&&n.push(s.expr),s.args)for(let o of s.args)typeof o=="object"&&o!==null&&o.type&&n.push(o);if(s.indices)for(let o of s.indices)typeof o=="object"&&o!==null&&o.type&&n.push(o)}}}function Bz(r,e,t){let n=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*bn(i.dtype);let p=1;for(let c of i.shape)p*=typeof c=="number"&&c>0?c:1;return Math.max(Oz,p*bn(i.dtype))},"bufBytes");for(let[,i]of r.bufferMap)s=Math.ceil(s/n)*n,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/n)*n,e.memoryLayout.bufferOffsets.set(i,s),s+=o(a));e.memoryLayout.totalBytes=s}function Rz(r,e){let t=new Map;for(let n of gu(r,s=>s.type==="BufferStoreNode"&&s.buffer&&!e.paramBuffers.has(s.buffer.name))){let s=n.buffer.name;t.has(s)||t.set(s,[]),t.get(s).push(n.value)}for(let[n,s]of t)if(s.every(i=>i&&i.type==="IntImmNode"&&i.value===0||i&&i.type==="FloatImmNode"&&i.value===0)&&e.zeroBuffers.add(n),s.length===1&&s[0]){let i=s[0];(i.type==="IntImmNode"||i.type==="FloatImmNode")&&e.constantBuffers.set(n,i.value)}}var Oz,yM=M(()=>{"use strict";xc();Vn();De();Gr();l(gM,"scanMetadata");l(Az,"walkTree");Oz=65536;l(Bz,"computeMemoryLayout");l(Rz,"detectZeroBuffers")});function Sf(r,e,t){let n=typeof r.offset=="number"?r.offset:0;if(e.length===0)return new L(n);if(e.length===1)return n===0?e[0]:new C("+",e[0],new L(n));let s=[];n!==0&&s.push(new L(n));for(let o=0;o<e.length;o++){let i=e[o];if(i.type==="IntImmNode"&&i.value===0)continue;let a=r.strides[o];if(a===1)s.push(i);else if(typeof a=="number"&&a>=0)s.push(new C("*",i,new L(a)));else{let u=Mz(r,o,t);s.push(new C("*",i,u))}}return s.length===0?new L(0):s.reduce((o,i)=>new C("+",o,i))}function Mz(r,e,t){let n=[];for(let s=e+1;s<r.shape.length;s++){let o=r.shape[s];typeof o=="number"&&o>=0?n.push(new L(o)):n.push(Iz(r,s,t))}return n.length===0?new L(1):n.reduce((s,o)=>new C("*",s,o))}function Iz(r,e,t){let n=r.shape[e];if(n instanceof Ce)return uu(n,s=>new Ne(qa(s),"index"));if(t){let s=`${r.name}:${e}`,o=t.get(s);if(o)return new Ne(o.name,"index")}return new L(1)}var _M=M(()=>{"use strict";De();Vs();Lb();l(Sf,"flattenIndex");l(Mz,"computeDynamicStride");l(Iz,"resolveShapeParam")});function bM(r,e){let t=gM(r,e),n={shapeParamMap:r.shapeParamMap,accCounter:0,metadata:t},s=is(r.body,n);return new sd(r.name,r.params,s,r.bufferMap,r.shapeParams,r.shapeParamMap,t)}function is(r,e){if(!r||typeof r!="object")return r;switch(r.type){case"ForNode":return Ez(r,e);case"BlockNode":return Lz(r,e);case"SeqNode":return Cz(r,e);case"BufferStoreNode":return Dz(r,e);case"LetStmtNode":return Pz(r,e);case"AllocateNode":return $z(r,e);case"IfThenElseNode":return Fz(r,e);case"WhileNode":return zz(r,e);case"EvaluateNode":return new ou(mr(r.value,e));case"SyncThreadsNode":return r;default:return r}}function Ez(r,e){let t=r.accumulator!==void 0?r.accumulator:e_(r);if(t)return Vz(r,t,e);let n=is(r.body,e);return new de(r.loopVar,r.min,r.extent,r.kind,n,r.threadTag)}function Lz(r,e){let t=[];for(let o of r.iterVars)o.iterVar&&o.binding&&t.push({name:o.iterVar.name,dtype:o.iterVar.dtype,expr:mr(o.binding,e)});let n=is(r.body,e),s=r.initBody?is(r.initBody,e):null;return t.length>0?s?(n=new te([new mu(t,s),new mu(t,n)]),s=null):n=new mu(t,n):s&&(n=new te([s,n])),n}function Cz(r,e){let t=[];for(let n of r.stmts)t.push(is(n,e));return new te(t)}function Dz(r,e){let t=Sf(r.buffer,r.indices,e.shapeParamMap),n=mr(r.value,e),s=r.buffer.dtype||ir(r.value);return new Tc(r.buffer,t,n,s)}function Pz(r,e){return new Ye(r.variable,mr(r.value,e),is(r.body,e))}function $z(r,e){return new Ve(r.buffer,r.scope,is(r.body,e))}function Fz(r,e){return new xe(mr(r.condition,e),is(r.thenBody,e),r.elseBody?is(r.elseBody,e):null)}function zz(r,e){return new Zo(r.condVar,is(r.condBody,e),is(r.loopBody,e))}function mr(r,e){if(!r||typeof r!="object"||!r.type)return r;switch(r.type){case"BufferLoadNode":{let t=Sf(r.buffer,r.indices,e.shapeParamMap),n=new wc(r.buffer,t,r.buffer.dtype);return No(n),n}case"MathOpNode":{let t=mr(r.a,e),n=r.b?mr(r.b,e):null,s={...r,a:t,b:n};return Object.setPrototypeOf(s,Object.getPrototypeOf(r)),No(s),s}case"CompareNode":{let t=mr(r.a,e),n=mr(r.b,e),s={...r,a:t,b:n};return Object.setPrototypeOf(s,Object.getPrototypeOf(r)),No(s),s}case"CastNode":{let t=mr(r.expr,e),n={...r,expr:t};return Object.setPrototypeOf(n,Object.getPrototypeOf(r)),No(n),n}case"CallExternNode":{let t=r.args.map(s=>mr(s,e)),n={...r,args:t};return Object.setPrototypeOf(n,Object.getPrototypeOf(r)),No(n),n}case"IfThenElseNode":{let t=mr(r.condition,e),n=mr(r.thenBody,e),s=r.elseBody?mr(r.elseBody,e):null,o={...r,condition:t,thenBody:n,elseBody:s};return Object.setPrototypeOf(o,Object.getPrototypeOf(r)),No(o),o}default:return No(r),r}}function Vz(r,e,t){let n=`_acc_${t.accCounter++}`,s=ir(e.loadSide);t.metadata.locals.set(n,s);let o=new Map;for(let d of e.block.iterVars)d.iterVar&&d.binding&&o.set(d.iterVar.name,d.binding);let i=o.size>0?Ar(e.valueSide,o):e.valueSide,a=Sf(e.store.buffer,e.outerIndices,t.shapeParamMap),u=new wc(e.store.buffer,a,e.store.buffer.dtype);No(u);let p=mr(i,t),c=Sf(e.store.buffer,e.outerIndices,t.shapeParamMap),h=new Tc(e.store.buffer,c,null,e.store.buffer.dtype),f=e.block.initBody?o.size>0?kf(e.block.initBody,o):e.block.initBody:null;return new od({localName:n,dtype:s,op:e.op,initLoad:u,loopVar:r.loopVar,extent:r.extent,loopKind:r.kind,body:p,flushStore:h,initBody:f?is(f,t):null})}function Ar(r,e){if(!r||typeof r!="object"||!r.type)return r;if(r.type==="VariableNode"&&e.has(r.name))return e.get(r.name);if(r.type==="BufferLoadNode"){let t=r.indices.map(o=>Ar(o,e));if(!t.some((o,i)=>o!==r.indices[i]))return r;let s={...r,indices:t};return Object.setPrototypeOf(s,Object.getPrototypeOf(r)),s}if(r.type==="MathOpNode"){let t=Ar(r.a,e),n=r.b?Ar(r.b,e):null;if(t===r.a&&n===r.b)return r;let s={...r,a:t,b:n};return Object.setPrototypeOf(s,Object.getPrototypeOf(r)),s}if(r.type==="CompareNode"){let t=Ar(r.a,e),n=Ar(r.b,e);if(t===r.a&&n===r.b)return r;let s={...r,a:t,b:n};return Object.setPrototypeOf(s,Object.getPrototypeOf(r)),s}if(r.type==="CastNode"){let t=Ar(r.expr,e);if(t===r.expr)return r;let n={...r,expr:t};return Object.setPrototypeOf(n,Object.getPrototypeOf(r)),n}if(r.type==="CallExternNode"){let t=r.args.map(o=>Ar(o,e));if(!t.some((o,i)=>o!==r.args[i]))return r;let s={...r,args:t};return Object.setPrototypeOf(s,Object.getPrototypeOf(r)),s}if(r.type==="IfThenElseNode"){let t=Ar(r.condition,e),n=Ar(r.thenBody,e),s=r.elseBody?Ar(r.elseBody,e):null;if(t===r.condition&&n===r.thenBody&&s===r.elseBody)return r;let o={...r,condition:t,thenBody:n,elseBody:s};return Object.setPrototypeOf(o,Object.getPrototypeOf(r)),o}return r}function kf(r,e){if(!r||typeof r!="object"||!r.type)return r;switch(r.type){case"BufferStoreNode":{let t=r.indices.map(o=>Ar(o,e)),n=Ar(r.value,e);if(t.every((o,i)=>o===r.indices[i])&&n===r.value)return r;let s={...r,indices:t,value:n};return Object.setPrototypeOf(s,Object.getPrototypeOf(r)),s}case"SeqNode":{let t=r.stmts.map(s=>kf(s,e)),n={...r,stmts:t};return Object.setPrototypeOf(n,Object.getPrototypeOf(r)),n}case"ForNode":{let t=kf(r.body,e);if(t===r.body)return r;let n={...r,body:t};return Object.setPrototypeOf(n,Object.getPrototypeOf(r)),n}case"IfThenElseNode":{let t=Ar(r.condition,e),n=kf(r.thenBody,e),s=r.elseBody?kf(r.elseBody,e):null,o={...r,condition:t,thenBody:n,elseBody:s};return Object.setPrototypeOf(o,Object.getPrototypeOf(r)),o}default:return r}}var wM=M(()=>{"use strict";xc();yM();_M();De();sx();l(bM,"lowerToLIR");l(is,"lowerStmt");l(Ez,"lowerForNode");l(Lz,"lowerBlockNode");l(Cz,"lowerSeqNode");l(Dz,"lowerBufferStore");l(Pz,"lowerLetStmt");l($z,"lowerAllocate");l(Fz,"lowerIfThenElse");l(zz,"lowerWhile");l(mr,"lowerExpr");l(Vz,"lowerAccumulator");l(Ar,"substituteVars");l(kf,"substituteVarsStmt")});function TM(r){let e=[],t={errors:e,boundVars:new Set,path:[],bufferNames:new Set};if(!r||r.type!=="LIRFunc")return e.push(new lo("root must be LIRFunc")),e;for(let[,n]of r.bufferMap)t.bufferNames.add(n.name);for(let n of r.shapeParams)t.boundVars.add(n.name);if(r.metadata)for(let[n]of r.metadata.locals)t.boundVars.add(n);return as(r.body,t),e}function as(r,e){if(!(!r||typeof r!="object")){switch(e.path.push(r.type),r.type){case"ForNode":Gz(r,e);break;case"SeqNode":for(let t of r.stmts)as(t,e);break;case"LIRFlatStoreNode":jz(r,e);break;case"LIRAccumulatorNode":qz(r,e);break;case"LIRBindingsNode":Uz(r,e);break;case"LetStmtNode":{vn(r.value,e);let t=e.boundVars.has(r.variable.name);e.boundVars.add(r.variable.name),as(r.body,e),t||e.boundVars.delete(r.variable.name);break}case"AllocateNode":{let t=r.buffer?e.bufferNames.has(r.buffer.name):!0;r.buffer&&e.bufferNames.add(r.buffer.name),as(r.body,e),r.buffer&&!t&&e.bufferNames.delete(r.buffer.name);break}case"IfThenElseNode":vn(r.condition,e),as(r.thenBody,e),r.elseBody&&as(r.elseBody,e);break;case"WhileNode":as(r.condBody,e),as(r.loopBody,e);break;case"EvaluateNode":vn(r.value,e);break;default:break}e.path.pop()}}function Gz(r,e){r.extent?vn(r.extent,e):e.errors.push(new lo("ForNode missing extent",[...e.path]));let t=!0;r.loopVar?(t=e.boundVars.has(r.loopVar.name),e.boundVars.add(r.loopVar.name)):e.errors.push(new lo("ForNode missing loopVar",[...e.path])),as(r.body,e),r.loopVar&&!t&&e.boundVars.delete(r.loopVar.name)}function jz(r,e){r.buffer||e.errors.push(new lo("LIRFlatStoreNode missing buffer",[...e.path])),vn(r.offsetExpr,e),vn(r.value,e)}function qz(r,e){r.localName||e.errors.push(new lo("LIRAccumulatorNode missing localName",[...e.path])),r.dtype||e.errors.push(new lo("LIRAccumulatorNode missing dtype",[...e.path]));let t=r.localName?e.boundVars.has(r.localName):!0;r.localName&&e.boundVars.add(r.localName),vn(r.initLoad,e),vn(r.body,e),as(r.flushStore,e),r.initBody&&as(r.initBody,e),r.localName&&!t&&e.boundVars.delete(r.localName)}function Uz(r,e){let t=[];for(let n of r.bindings)vn(n.expr,e),e.boundVars.has(n.name)||t.push(n.name),e.boundVars.add(n.name);as(r.body,e);for(let n of t)e.boundVars.delete(n)}function vn(r,e){if(!(!r||typeof r!="object"||!r.type))switch(r.type){case"LIRFlatLoadNode":r.buffer||e.errors.push(new lo("LIRFlatLoadNode missing buffer",[...e.path])),vn(r.offsetExpr,e);break;case"MathOpNode":vn(r.a,e),r.b&&vn(r.b,e);break;case"CompareNode":vn(r.a,e),vn(r.b,e);break;case"CastNode":vn(r.expr,e);break;case"CallExternNode":if(r.args)for(let t of r.args)vn(t,e);break;case"IfThenElseNode":vn(r.condition,e),vn(r.thenBody,e),r.elseBody&&vn(r.elseBody,e);break;case"VariableNode":r.name!==void 0&&!e.boundVars.has(r.name)&&e.errors.push(new lo(`unbound variable '${r.name}'`,[...e.path]));break;case"IntImmNode":case"FloatImmNode":break;default:break}}var lo,xM=M(()=>{"use strict";lo=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(TM,"verifyLIR");l(as,"verifyStmt");l(Gz,"verifyForNode");l(jz,"verifyFlatStore");l(qz,"verifyAccumulator");l(Uz,"verifyBindings");l(vn,"verifyExpr")});function SM(r,e,t={}){return new Do({target:e,...t}).compileFunction(r)}var u_,mx,Do,l_=M(()=>{"use strict";Yo();xs();ZO();um();QO();PB();rM();Ui();Ec();nm();sM();aM();lx();lM();pM();mM();jT();Kt();P0();cT();wM();xM();Kt();u_=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??ue.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}},mx=class{static{l(this,"CompilationResult")}constructor(e,t,n){this.module=e,this.trace=t,this.errors=n||[]}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()}},Do=class r{static{l(this,"Compiler")}constructor(e){if(this.config=e instanceof u_?e:new u_(e),!this.config.target)throw new Error("Compiler requires a target");this.context=new jg({loweringRules:this.config.loweringRules,codegenEntries:this.config.codegenEntries})}compile(e){let t=new am(this.config.trace),n=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:n,original:e,working:n?Ok(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),!n&&s.length>0)throw new Error(s[0].toString());return new mx(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,n=typeof t.target.isWebGPU=="function"&&t.target.isWebGPU(),s=typeof t.target.isGPU=="function"&&t.target.isGPU()&&!n,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=dM(e.working,{config:t,target:t.target,cudaMatmulChain:e.cudaMatmulChain,cudaConvChain:a,cudaAttention:u,isWebGPU:n})},"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:Uv(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 wn("single");return t.addFunction(e),this.compile(t)}calibrate(e,t="minmax"){let n=new Ll(t);for(let s of e)n.attach(s);return n}_runCalibration(e,t){let n=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,h)=>new r({target:h,verify:this.config.verify}).compile(c),"compileFn"),p=uM(o,a,n.calibrationData,{mode:n.calibrationMode||"minmax",quantizableOps:n.quantizableOps,compileFn:u});this.config.quantization={...n,calibration:p},t.phaseEnd("calibrate",performance.now()-i)}_runGraphPasses(e,t,n,s,o,i){let a=new oa,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),h=u>=c;for(let m of DB(this.config,this.config.target,{cudaMatmulChain:h,context:this.context}))a.addPass(m);a.setTrace(n),this.config.verifyMode==="full"&&a.setVerifyHook((m,g)=>(g?ix(m):wf(m)).map(_=>_.toString())),n.phaseStart("graphPasses");let f=performance.now(),d=a.run(e,{errorMode:i?"resilient":"strict",passContext:this.config.passContext});if(d.errors){for(let m of d.errors)s.push(m),n.errorEvent(m.phase,m.funcName,m.message,m.passName);if(d.failedFunctions){for(let m of d.failedFunctions)if(o.add(m),i&&t&&t!==e){let g=t.getFunction(m);g&&e.addFunction(Vh(g))}}}if(n.phaseEnd("graphPasses",performance.now()-f),n.shouldSnapshot("afterGraphPasses")){let m=new cp;n.irDump("afterGraphPasses",m.printModule(e))}return h}_runPartitioning(e,t){let n=new oa;n.addPass(new i_(this.config.partition)),n.addPass(new a_({targets:this.config.partition.targets})),n.setTrace(t),t.phaseStart("partition");let s=performance.now();n.run(e),t.phaseEnd("partition",performance.now()-s)}_eachFunc(e,t,n,s,o,i,a){for(let u of e)if(!o.has(u.name))try{a(u)}catch(p){if(s.push(new Dn(t,u.name,p.message)),o.add(u.name),n.errorEvent(t,u.name,p.message),!i)break}}_lowerAll(e,t,n,s,o){t.phaseStart("lowering");let i=performance.now(),a=[];return this._eachFunc(e,"lowering",t,n,s,o,u=>{let p=performance.now(),c=Xh(u,this.config.target,this.context);this.config.target.isGPU&&this.config.target.isGPU()&&!(this.config.target.isWebGPU&&this.config.target.isWebGPU())&&(c.convInfo=vR(u)),t.functionEvent("lowering",u.name,{durationMs:performance.now()-p}),a.push(c),t.shouldSnapshot("afterLowering")&&t.irDump("afterLowering:"+u.name,qg(c))}),t.phaseEnd("lowering",performance.now()-i),a}_runTirPasses(e){let t=new Ug;for(let n of nM(this.config))t.addPass(n);if(t.setTrace(e.trace),this.config.verifyMode==="full"){let n=new bf;t.setVerifyHook(s=>n.verify(s))}t.run(e.primFuncs,{trace:e.trace,errors:e.errors,failed:e.failed,resilient:e.resilient})}_verifyGraph(e,t,n,s,o,i){if(i){for(let u of e){if(o.has(u.name))continue;let p=wf?wf(u):[];if(p.length>0){let c=p.map(h=>h.toString()).join("; ");s.push(new Dn("verification",u.name,c)),o.add(u.name),n.errorEvent("verification",u.name,c)}}return}let a=ix(e);if(a.length>0)throw new Error("Graph verification failed ("+t+"): "+a.map(u=>u.toString()).join("; "))}_verifyAll(e,t,n,s){let o=new bf;for(let i of e){if(n.has(i.name))continue;let a=o.verify(i);if(a.length>0){let u=a.join("; ");if(s)t.push(new Dn("verification",i.name,u)),n.add(i.name);else throw new Error("TensorIR verification failed for "+i.name+": "+u)}}}_lowerToLIR(e,t,n,s,o){t.phaseStart("lirLowering");let i=performance.now(),a=[];return this._eachFunc(e,"lirLowering",t,n,s,o,u=>{let p=performance.now(),c=bM(u,this.config.target);if(u.cublasInfo&&(c.cublasInfo=u.cublasInfo),u.gpuRegisterBlocked&&(c.gpuRegisterBlocked=!0),this.config.verifyMode==="full"){let h=TM(c);if(h.length>0)throw new Error("LIR verification failed: "+h.map(f=>f.toString()).join("; "))}t.functionEvent("lirLowering",u.name,{durationMs:performance.now()-p}),a.push(c)}),t.phaseEnd("lirLowering",performance.now()-i),a}_codegen(e,t,n,s,o){t.phaseStart("codegen");let i=performance.now(),a=new zu("compiled"),u=this.config.usePartition,p={matmulBackend:this.config.matmulBackend,context:this.context},c=new Map,h=l(d=>(c.has(d.name)||c.set(d.name,new Hs(d,p)),c.get(d.name)),"getBackend"),f=u?null:new Hs(this.config.target,p);return this._eachFunc(e,"codegen",t,n,s,o,d=>{let m=performance.now(),g;if(u){let _=d._partitionTarget,b=_?this.config.partition.targets.find(w=>w.name===_):this.config.target;g=h(b||this.config.target)}else g=f;let y=g.compile(d);a.addCompiledKernel(y),d.shapeParamMap&&d.shapeParamMap.size>0&&a.setShapeParamMap(d.name,d.shapeParamMap,d.bufferMap),t.codegenStats(d.name,{durationMs:performance.now()-m,sourceSize:y.source.length,targetName:y.target.name})}),t.phaseEnd("codegen",performance.now()-i),a}};l(SM,"compileGraph")});function Wz(r){return typeof r=="object"&&r!==null}function Hz(r){return Wz(r)&&"_impl"in r}function Yz(r){return(e,...t)=>{let n=ss();if(!n)throw new Error(`TRACING dispatch key active but no tracer is set for op '${r}'`);if(Kz.has(r)){let c=Nt.findOp(r);return Nt.redispatch(c,e,...t)}let s=[],o={},i=km(r),a=0,u=l(c=>c instanceof ft||Hz(c),"isTensor"),p=l(c=>{c instanceof ft||c.isSymbolic?s.push(c):s.push(n.captureConstant(c))},"pushTensor");for(let c of t)if(Array.isArray(c)&&c.length>0&&u(c[0]))for(let h of c)p(h);else u(c)?p(c):i&&(c!=null&&a<i.length&&(o[i[a]]=c),a++);return n.recordOp(r,s,o)}}function vM(){let r=Nt.listOps();for(let e of r){let t=Nt.findOp(e);if(!t)continue;let n=t.name,s=wr.fromUnboxed(Yz(n));t.entry.registerKernel(48,s)}}var Kz,NM=M(()=>{"use strict";Nn();nc();$i();ya();ga();vm();Kz=new Set(["scatter","repeat","tile","split","chunk","roll","flip","cumsum","sort","argsort","topk"]);l(Wz,"_isObject");l(Hz,"_hasImpl");l(Yz,"_tracingKernel");l(vM,"registerTracingDispatch")});function Zz(r){return r&&r.shape&&r.shape.length>=2&&Jz.has(r.dtype)}function AM(r,e,t=Zz){let n=r.graph.functions().next().value;if(!n)return r;let s=n.entryBlock,o=r.numUserInputs,i=r.capturedParams,a=[];for(let h=0;h<i.length;h++){let f=o+h,d=s.getArgument(f);d&&t(i[h],h,d)&&a.push({j:h,argIndex:f,arg:d,param:i[h]})}if(a.length===0)return r;for(let h of a){let f=e(h.param),d=h.arg.type,m=Re,g=new m("constant",[],[d],{value:f,tensor_type:d}),y=s.firstOp;y?s.insertBefore(g,y):s.pushOp(g),h.arg.replaceAllUsesWith(g.getResult(0))}let u=new Set(a.map(h=>h.argIndex));s.removeArguments(u),n.inputTypes=Object.freeze(n.inputTypes.filter((h,f)=>!u.has(f)));let p=new Set(a.map(h=>h.j)),c=i.filter((h,f)=>!p.has(f));return{...r,capturedParams:c}}var Jz,OM=M(()=>{"use strict";sr();Jz=new Set(["f16","f32","f64"]);l(Zz,"defaultWeightPredicate");l(AM,"foldWeightParams")});function vf(r,e,t){let n=t.get(e.id);if(n)return n;let s=r.scalarConstant(0,e.type.dtype).getResult(0);return r.broadcast(s,e.type.shape,[]).getResult(0)}var co,Nf=M(()=>{"use strict";l(vf,"gradOrZero");co=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 n=this._pending.get(e);n||(n=[],this._pending.set(e,n)),n.push(t),this._reduced.delete(e)}_treeReduce(e){let t=e;for(;t.length>1;){let n=[];for(let s=0;s<t.length;s+=2)s+1<t.length?n.push(this._builder.add(t[s],t[s+1]).getResult(0)):n.push(t[s]);t=n}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 n=this._treeReduce(t);return this._reduced.set(e,n),n}has(e){let t=this._pending.get(e);return!!t&&t.length>0}}});function ce(r,e){gx.set(r,e)}function Ia(r){return gx.get(r)||null}function h_(r){let e=gx.get(r);if(e)return e;if(yx.has(r))return null;throw new Error(`autodiff: op '${r}' is on the gradient path but has no VJP rule and is not a registered gradient barrier. Register one with registerVJPRule('${r}', ...) or registerGradientBarrier('${r}').`)}function _x(r,e){BM.set(r,e)}function d_(r){return BM.get(r)||null}function bx(r){yx.add(r)}function m_(r){return yx.has(r)}var gx,yx,BM,ls=M(()=>{"use strict";gx=new Map,yx=new Set;l(ce,"registerVJPRule");l(Ia,"getVJPRule");l(h_,"requireVJPRuleOrBarrier");BM=new Map;l(_x,"registerRegionVJP");l(d_,"getRegionVJP");l(bx,"registerGradientBarrier");l(m_,"isGradientBarrier")});var Po,wx=M(()=>{"use strict";Po=new Set(["scan","if"])});function Af(r){let e=new Set(r.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(r);let n=new Set,s=[],o=l(i=>{for(let a of i.ops()){for(let u of a.operands)e.has(u.id)||n.has(u.id)||u.definingOp&&u.definingOp.opName==="constant"||(n.add(u.id),s.push(u));for(let u of a.regions||[])for(let p of u.blocks)o(p)}},"scan");return o(r),s}function g_(r,e){let t=r.scalarConstant(0,e.type.dtype).getResult(0);return r.broadcast(t,e.type.shape,[]).getResult(0)}function RM(r,e,t){let n=e.type.shape,s=n.map((a,u)=>u===0?t:0),o=n.map((a,u)=>u===0?t+1:a),i=r.slice(e,s,o).getResult(0);return r.reshape(i,n.slice(1)).getResult(0)}function Xz(r,e,t){let n=t.slice(1),s=e.map(o=>r.reshape(o,[1,...n]).getResult(0));return s.length===1?r.reshape(s[0],t).getResult(0):r.concat(s,0).getResult(0)}function y_(r,e,t,n,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 n)u.set(_,b);for(let _ of a)n.has(_.id)&&p.set(_,n.get(_.id));let c=[],h=null;for(let _ of e.ops())_.opName==="yield"?h=_:c.push(_);let f=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=r._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(Po.has(_.opName)){for(let S of _.operands)f(S),p.set(S,u.get(S.id)??S);let T=_.clone(p);r.block.pushOp(T);for(let S=0;S<_.numResults;S++)u.set(_.getResult(S).id,T.getResult(S)),p.set(_.getResult(S),T.getResult(S));continue}let b=_.operands.map(f),w=r._buildOp(_.opName,b,_.results.map(T=>T.type),new Map(_.attributes),null);for(let T=0;T<_.numResults;T++)u.set(_.getResult(T).id,w.getResult(T)),p.set(_.getResult(T),w.getResult(T))}let d=h.operands.map(_=>u.get(_.id));if(o)return{forwardYields:d};let m=new co(r);for(let _=0;_<h.operands.length;_++)s[_]&&m.accumulate(h.operands[_].id,s[_]);for(let _=c.length-1;_>=0;_--){let b=c[_];if(b.opName==="constant")continue;let w=b.results.map(k=>m.get(k.id));if(w.every(k=>k===null))continue;if(Po.has(b.opName)){let k=d_(b.opName);k&&k(b,{accumulator:m,builder:r,materialize:f,needsGrad:Qz,scanCheckpoint:null});continue}let T=h_(b.opName);if(!T)continue;let S={builder:r,op:b,operands:b.operands.map(f),results:b.results.map(k=>u.get(k.id)),gradOutputs:w,attrs:b.attributes,full:l((k,A)=>r.broadcast(r.scalarConstant(k,A.dtype).getResult(0),A.shape,[]).getResult(0),"full")},x=T(S);if(x)for(let k=0;k<b.numOperands;k++)x[k]&&m.accumulate(b.getOperand(k).id,Tx(r,x[k],b.getOperand(k).type.shape))}let g=e.arguments.map(_=>m.get(_.id)),y=new Map;for(let _ of n.keys())y.set(_,m.get(_));return{forwardYields:d,gradArgs:g,gradFree:y}}function e3(r,e,t,n,s){let o=r.regions[0].blocks[0],i=r.regions[1].blocks[0],a=n(r.getOperand(0)),u=[];for(let y=0;y<r.numResults;y++)u.push(e.get(r.getResult(y).id));let p=Af(o),c=Af(i),h=new Map(p.map(y=>[y.id,n(y)])),f=new Map(c.map(y=>[y.id,n(y)])),{gradFree:d}=y_(t,o,[],h,u,!1,new Map,p),{gradFree:m}=y_(t,i,[],f,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=d.get(y),w=m.get(y);if(!b&&!w)continue;let T=g_(t,_),S=t.broadcast(a,_.type.shape,[]).getResult(0);e.accumulate(y,t.select(S,b??T,w??T).getResult(0))}}function t3(r,e){if(!r||e<=1)return null;if(r==="sqrt"||r===!0)return Math.max(1,Math.ceil(Math.sqrt(e)));if(typeof r=="number"&&r>=1){let t=Math.floor(r);return t>=e?null:t}return null}function n3(r,e,t,n,s,o=null){let i=r.regions[0].blocks[0],a=r.getAttr("num_carry"),u=r.getAttr("num_xs"),p=r.numResults-a,c=[];for(let v=0;v<u;v++)c.push(r.getOperand(v));let h=[];for(let v=0;v<a;v++)h.push(r.getOperand(u+v));let f=c[0].type.shape[0],d=Af(i),m=c.map(n),g=h.map(n),y=new Map(d.map(v=>[v.id,n(v)])),_=new Map,b=l(v=>m.map(O=>RM(t,O,v)),"sliceX"),w=l((v,O)=>y_(t,i,[...v,...O],y,null,!0,_,d).forwardYields.slice(0,a),"stepForward"),T=[];for(let v=0;v<p;v++)T.push(e.get(r.getResult(a+v).id));let S=[];for(let v=0;v<a;v++){let O=e.get(r.getResult(v).id);S.push(O??g_(t,g[v]))}let x=new Map,k=m.map(()=>new Array(f)),A=l((v,O,R)=>{let B=[...O,...R],E=T.map(j=>j===null?null:RM(t,j,v)),z=[...S,...E],{gradArgs:D,gradFree:J}=y_(t,i,B,y,z,!1,_,d);for(let j=0;j<u;j++)k[j][v]=D[j]??g_(t,O[j]);S=[];for(let j=0;j<a;j++)S.push(D[u+j]??g_(t,R[j]));for(let[j,U]of J){if(!U)continue;let P=x.get(j);x.set(j,P?t.add(P,U).getResult(0):U)}},"backwardStep"),N=t3(o,f);if(N){let v=Math.ceil(f/N),O=new Array(v),R=g;O[0]=R;for(let B=0;B<f;B++){R=w(b(B),R);let E=(B+1)/N;Number.isInteger(E)&&E<v&&(O[E]=R)}for(let B=v-1;B>=0;B--){let E=B*N,z=Math.min(E+N,f),D=[],J=[O[B]],j=O[B];for(let U=E;U<z;U++){let P=b(U);D.push(P),j=w(P,j),J.push(j)}for(let U=z-1;U>=E;U--)A(U,D[U-E],J[U-E])}}else{let v=g,O=[v],R=[];for(let B=0;B<f;B++){let E=b(B);R.push(E),v=w(E,v),O.push(v)}for(let B=f-1;B>=0;B--)A(B,R[B],O[B])}for(let v=0;v<u;v++)s.has(c[v].id)&&e.accumulate(c[v].id,Xz(t,k[v],m[v].type.shape));for(let v=0;v<a;v++)s.has(h[v].id)&&e.accumulate(h[v].id,S[v]);for(let[v,O]of x)O&&s.has(v)&&e.accumulate(v,O)}var Qz,MM=M(()=>{"use strict";Nf();ls();Of();wx();Qz={has:l(()=>!0,"has")};_x("scan",(r,e)=>n3(r,e.accumulator,e.builder,e.materialize,e.needsGrad,e.scanCheckpoint));_x("if",(r,e)=>e3(r,e.accumulator,e.builder,e.materialize,e.needsGrad));l(Af,"regionFreeVars");l(g_,"zeroLike");l(RM,"sliceStep");l(Xz,"stackSteps");l(y_,"diffBodyStep");l(e3,"buildCondBackward");l(t3,"resolveSegmentLength");l(n3,"buildScanBackward")});function Bf(r,{accumulator:e,builder:t,needsGrad:n,resolveValue:s,handleRegionOp:o=null}){for(let i=r.length-1;i>=0;i--){let a=r[i];if(a.opName==="return"||a.opName==="constant"||!a.results.some(g=>n.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=h_(a.opName);if(!c)continue;let h=new Array(a.numOperands);for(let g=0;g<a.numOperands;g++)h[g]=s(a.getOperand(g));let f=new Array(a.numResults);for(let g=0;g<a.numResults;g++)f[g]=s(a.getResult(g));let d=l((g,y)=>t.broadcast(t.scalarConstant(g,y.dtype).getResult(0),y.shape,[]).getResult(0),"full"),m=c({builder:t,op:a,operands:h,results:f,gradOutputs:p,attrs:a.attributes,full:d});if(m)for(let g=0;g<a.numOperands;g++){if(g>=m.length||!m[g])continue;let y=a.getOperand(g);n.has(y.id)&&e.accumulate(y.id,Tx(t,m[g],y.type.shape))}}}function IM(r){let e=[];for(let t of r.regions)t.blocks[0]&&e.push(...Af(t.blocks[0]));return e}function Tx(r,e,t){let n=e.type.shape;if(n.length===t.length&&n.every((u,p)=>u===t[p]))return e;let s=n.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&&n[s+u]!==1&&o.push(s+u);let i=e;if(o.length>0){let u=r.scalarConstant(0,e.type.dtype).getResult(0);i=r.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=r.reshape(i,t).getResult(0)),i}var r3,Rf,Of=M(()=>{"use strict";xs();es();Es();Nf();ls();MM();wx();r3=new Set(["neg","abs","sign","floor","ceil"]);l(Bf,"backpropOps");l(IM,"regionControlFlowFreeVars");l(Tx,"reduceGradToOperandShape");Rf=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 n=at.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,n),{savedValues:u,savedValueIndices:p}=this._identifySavedValues(n,a,i),c=o.map(x=>x.type),h=u.map(x=>x.type),f=[...c,...h],d=[];for(let x=0;x<i.length;x++)a.has(i[x].id)&&d.push(i[x].type);let m=new un(`backward_${e.name}`,f,d),g=new An(m),y=m.args,_=y.slice(0,c.length),b=y.slice(c.length),w=new Map;for(let x=0;x<u.length;x++)w.set(u[x].id,b[x]);for(let x=0;x<i.length;x++)p.has(i[x].id)&&w.set(i[x].id,b[p.get(i[x].id)]);let T=new co(g);for(let x=0;x<o.length;x++){let k=o[x];T.accumulate(k.id,_[x])}Bf(n,{accumulator:T,builder:g,needsGrad:a,resolveValue:l(x=>this._materialize(x,w,g),"resolveValue"),handleRegionOp:l(x=>{let k=d_(x.opName);return k?(k(x,{accumulator:T,builder:g,materialize:l(A=>this._materialize(A,w,g),"materialize"),needsGrad:a,scanCheckpoint:this._scanCheckpoint}),!0):!1},"handleRegionOp")});let S=[];for(let x=0;x<i.length;x++)a.has(i[x].id)&&S.push(vf(g,i[x],T));return g.returnOp(S),{backwardFunc:m,savedValues:u,gradInputIndices:this._getGradInputIndices(i,a)}}_materialize(e,t,n){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 f=u.getOperand(i.i);i.i++,f.definingOp&&!t.has(f.id)&&!s.has(f.id)&&(s.add(f.id),o.push({val:f,i:0}));continue}let p=new Array(u.numOperands);for(let f=0;f<u.numOperands;f++){let d=u.getOperand(f);p[f]=t.has(d.id)?t.get(d.id):d}let c=u.results.map(f=>f.type),h=n._buildOp(u.opName,p,c,new Map(u.attributes),null);for(let f=0;f<u.numResults;f++)t.set(u.getResult(f).id,h.getResult(f));s.delete(a.id),o.pop()}return t.has(e.id)?t.get(e.id):e}_computeGradReachability(e,t){let n=new Set,s=e.getReturnOp();for(let o of s.operands)n.add(o.id);for(let o=t.length-1;o>=0;o--){let i=t[o];if(!(i.opName==="return"||!i.results.some(u=>n.has(u.id)))){if(Po.has(i.opName)){for(let u=0;u<i.numOperands;u++)n.add(i.getOperand(u).id);for(let u of IM(i))n.add(u.id);continue}if(Ia(i.opName)&&!m_(i.opName))for(let u=0;u<i.numOperands;u++)n.add(i.getOperand(u).id)}}return n}_identifySavedValues(e,t,n){let s=[],o=new Map,i=new Set(n.map(h=>h.id));for(let h of e)if(!(h.opName==="return"||h.opName==="constant"||!Ia(h.opName)||!h.results.some(m=>t.has(m.id)))){for(let m=0;m<h.numOperands;m++){let g=h.getOperand(m);i.has(g.id)&&!o.has(g.id)&&(o.set(g.id,s.length),s.push(g))}for(let m=0;m<h.numResults;m++){let g=h.getResult(m);t.has(g.id)&&!o.has(g.id)&&this._shouldSaveResult(h)&&(o.set(g.id,s.length),s.push(g))}}let a=new Set(o.keys()),u=new Map(n.map(h=>[h.id,h])),p=new Set,c=l(h=>{if(a.has(h.id)||p.has(h.id))return;p.add(h.id);let f=h.definingOp;if(!f){i.has(h.id)&&!o.has(h.id)&&(o.set(h.id,s.length),s.push(u.get(h.id)));return}for(let d=0;d<f.numOperands;d++)c(f.getOperand(d))},"collect");for(let h of e)if(!(h.opName==="return"||h.opName==="constant")&&h.results.some(f=>t.has(f.id))){if(Po.has(h.opName)){for(let f of h.operands)c(f);for(let f of IM(h))c(f);continue}if(Ia(h.opName)){for(let f=0;f<h.numOperands;f++)c(h.getOperand(f));for(let f=0;f<h.numResults;f++)c(h.getResult(f))}}return{savedValues:s,savedValueIndices:o}}_shouldSaveResult(e){return this._rematPolicy?!this._rematPolicy.shouldRematerialize(e):!r3.has(e.opName)}_getGradInputIndices(e,t){let n=[];for(let s=0;s<e.length;s++)t.has(e[s].id)&&n.push(s);return n}_buildCheckpointed(e){let n=at.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,n),u=this._checkpointPolicy.segment(n,e);for(let O of u)for(let R of O.ops)if(Po.has(R.opName))throw new Error(`Checkpointed backward does not support region control-flow op '${R.opName}'; build the backward without a checkpointPolicy, which differentiates scan/if via buildScanBackward/buildCondBackward.`);let p=new Set,c=[],h=new Map,f=new Map;for(let O of n)for(let R=0;R<O.numResults;R++){let B=O.getResult(R);f.set(B.id,B)}let d=new Set(i.map(O=>O.id));for(let O of i)a.has(O.id)&&!p.has(O.id)&&(p.add(O.id),h.set(O.id,c.length),c.push(O));for(let O of u){for(let R of O.boundaryInputs)if(!d.has(R)&&!p.has(R)){p.add(R);let B=f.get(R)||null;B&&(h.set(R,c.length),c.push(B))}for(let R of O.boundaryOutputs)if(!p.has(R)){p.add(R);let B=f.get(R)||null;B&&(h.set(R,c.length),c.push(B))}}for(let O of o)p.has(O.id)||(p.add(O.id),h.set(O.id,c.length),c.push(O));let m=o.map(O=>O.type),g=c.map(O=>O.type),y=[...m,...g],_=[];for(let O=0;O<i.length;O++)a.has(i[O].id)&&_.push(i[O].type);let b=new un(`backward_${e.name}`,y,_),w=new An(b),T=b.args,S=T.slice(0,m.length),x=T.slice(m.length),k=new Map;for(let O=0;O<c.length;O++)k.set(c[O].id,x[O]);let A=new co(w);for(let O=0;O<o.length;O++)A.accumulate(o[O].id,S[O]);let N=new Map;for(let O of n)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 z=new Array(E.numOperands);for(let j=0;j<E.numOperands;j++){let U=E.getOperand(j),P=B.get(U.id)||k.get(U.id)||N.get(U.id);z[j]=P||U}let D=E.results.map(j=>j.type),J=w._buildOp(E.opName,z,D,new Map(E.attributes),null);for(let j=0;j<E.numResults;j++)B.set(E.getResult(j).id,J.getResult(j))}Bf(R.ops,{accumulator:A,builder:w,needsGrad:a,resolveValue:l(E=>B.get(E.id)||k.get(E.id)||N.get(E.id)||E,"resolveValue")})}let v=[];for(let O=0;O<i.length;O++)a.has(i[O].id)&&v.push(vf(w,i[O],A));return w.returnOp(v),{backwardFunc:b,savedValues:c,gradInputIndices:this._getGradInputIndices(i,a)}}}});var Ea,__=M(()=>{"use strict";Ea=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 n=0;n<e.numOperands;n++){let s=e.getOperand(n).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 Mf,xx=M(()=>{"use strict";xs();es();Es();Nf();ls();__();Of();Mf=class{static{l(this,"JointGraphBuilder")}constructor(e={}){this._rematPolicy=e.rematPolicy||new Ea(e.remat||{}),this._checkpointPolicy=e.checkpointPolicy||null}build(e){if(this._checkpointPolicy)return this._buildCheckpointed(e);let t=this._buildScaffold(e);return Bf(t.topoOrder,{accumulator:t.accumulator,builder:t.builder,needsGrad:t.needsGrad,resolveValue:l(n=>t.valueMap.get(n.id)||n,"resolveValue")}),this._finish(t)}_buildCheckpointed(e){let t=this._buildScaffold(e),n=this._checkpointPolicy.segment(t.topoOrder,e);for(let s=n.length-1;s>=0;s--){let o=n[s],i=new Map;for(let a of o.ops){let u=new Array(a.numOperands);for(let h=0;h<a.numOperands;h++){let f=a.getOperand(h);u[h]=i.get(f.id)||t.valueMap.get(f.id)||f}let p=a.results.map(h=>h.type),c=t.builder._buildOp(a.opName,u,p,new Map(a.attributes),null);for(let h=0;h<a.numResults;h++)i.set(a.getResult(h).id,c.getResult(h))}Bf(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 n=at.compute(e).topologicalOrder;this._assertNoRegionControlFlow(n);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 un(`joint_${e.name}`,u,p),h=new An(c),f=c.args,d=f.slice(0,e.inputTypes.length),m=f.slice(e.inputTypes.length),g=new Map;for(let w=0;w<i.length;w++)g.set(i[w].id,d[w]);for(let w of n){if(w.opName==="return")continue;let T=new Array(w.numOperands);for(let k=0;k<w.numOperands;k++){let A=w.getOperand(k);T[k]=g.get(A.id)||A}let S=w.results.map(k=>k.type),x=h._buildOp(w.opName,T,S,new Map(w.attributes),null);for(let k=0;k<w.numResults;k++)g.set(w.getResult(k).id,x.getResult(k))}let y=o.map(w=>g.get(w.id)),_=this._computeGradReachability(e,n),b=new co(h);for(let w=0;w<o.length;w++)b.accumulate(o[w].id,m[w]);return{topoOrder:n,forwardInputs:i,forwardOutputs:o,fwdOutputValues:y,valueMap:g,builder:h,needsGrad:_,accumulator:b,jointFunc:c}}_finish(e){let t=[];for(let n=0;n<e.forwardInputs.length;n++)t.push(vf(e.builder,e.forwardInputs[n],e.accumulator));return e.builder.returnOp([...e.fwdOutputValues,...t]),{jointFunc:e.jointFunc,numForwardOutputs:e.forwardOutputs.length,numGradInputs:t.length}}_computeGradReachability(e,t){let n=new Set,s=e.getReturnOp();for(let o of s.operands)n.add(o.id);for(let o=t.length-1;o>=0;o--){let i=t[o];if(!(i.opName==="return"||!i.results.some(u=>n.has(u.id)))&&Ia(i.opName)&&!m_(i.opName))for(let u=0;u<i.numOperands;u++)n.add(i.getOperand(u).id)}return n}_assertNoRegionControlFlow(e){for(let t of e)if(Po.has(t.opName))throw new Error(`JointGraphBuilder does not support region control-flow op '${t.opName}'; use BackwardGraphBuilder (separate mode) without a checkpointPolicy, which differentiates scan/if.`)}}});var EM=M(()=>{"use strict"});function LM(r,e){let t=r.gradOutputs[0],[n,s]=r.operands,o=r.full(0,n.type),i=r.builder.compare(n,s,e).getResult(0),a=r.builder.select(i,t,o).getResult(0),u=r.builder.select(i,o,t).getResult(0);return[a,u]}function CM(r){let e=r.gradOutputs[0],[t]=r.operands,n=r.full(0,e.type),s=r.builder.where(t,e,n).getResult(0),o=r.builder.where(t,n,e).getResult(0);return[null,s,o]}var DM=M(()=>{"use strict";ls();ce("add",r=>{let e=r.gradOutputs[0];return[e,e]});ce("sub",r=>{let e=r.gradOutputs[0],t=r.builder.neg(e).getResult(0);return[e,t]});l(LM,"_minMaxVJP");ce("maximum",r=>LM(r,"ge"));ce("minimum",r=>LM(r,"le"));ce("clamp",r=>{let e=r.gradOutputs[0],[t,n,s]=r.operands,o=r.full(0,n.type),i=r.builder.compare(n,t,"ge").getResult(0),a=r.builder.where(i,e,o).getResult(0),u=r.builder.compare(n,s,"le").getResult(0);return[null,r.builder.where(u,a,o).getResult(0),null]});l(CM,"_whereVJP");ce("where",CM);ce("select",CM);ce("mul",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=r.builder.mul(e,n).getResult(0),o=r.builder.mul(e,t).getResult(0);return[s,o]});ce("div",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=r.builder.div(e,n).getResult(0),o=r.builder.mul(n,n).getResult(0),i=r.builder.neg(t).getResult(0),a=r.builder.mul(e,i).getResult(0),u=r.builder.div(a,o).getResult(0);return[s,u]});ce("neg",r=>{let e=r.gradOutputs[0];return[r.builder.neg(e).getResult(0)]});ce("pow",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=r.full(1,t.type),o=r.builder.sub(n,s).getResult(0),i=r.builder.pow(t,o).getResult(0),a=r.builder.mul(n,i).getResult(0),u=r.builder.mul(e,a).getResult(0),p=r.builder.pow(t,n).getResult(0),c=r.builder.log(t).getResult(0),h=r.builder.mul(p,c).getResult(0),f=r.builder.mul(e,h).getResult(0);return[u,f]})});function PM(r,e){let t=r.full(2/Math.sqrt(Math.PI),e.type),n=r.builder.mul(e,e).getResult(0),s=r.builder.neg(n).getResult(0),o=r.builder.exp(s).getResult(0);return r.builder.mul(t,o).getResult(0)}function $M(r,e){let t=r.builder,n=r.full(1,e.type),s=t.add(e,r.full(Yi,e.type)).getResult(0),o=t.div(n,s).getResult(0),i=t.log(s).getResult(0),a=t.sub(i,t.mul(r.full(.5,e.type),o).getResult(0)).getResult(0),u=t.mul(o,o).getResult(0),p=u;for(let c of Nc)a=t.add(a,t.mul(r.full(c,e.type),p).getResult(0)).getResult(0),p=t.mul(p,u).getResult(0);for(let c=0;c<Yi;c++){let h=t.add(e,r.full(c,e.type)).getResult(0);a=t.sub(a,t.div(n,h).getResult(0)).getResult(0)}return a}var FM=M(()=>{"use strict";ls();Ji();ce("exp",r=>{let e=r.gradOutputs[0],t=r.results[0];return[r.builder.mul(e,t).getResult(0)]});ce("log",r=>{let e=r.gradOutputs[0],[t]=r.operands;return[r.builder.div(e,t).getResult(0)]});ce("sqrt",r=>{let e=r.gradOutputs[0],t=r.results[0],n=r.full(2,t.type),s=r.builder.mul(n,t).getResult(0);return[r.builder.div(e,s).getResult(0)]});ce("tanh",r=>{let e=r.gradOutputs[0],t=r.results[0],n=r.builder.mul(t,t).getResult(0),s=r.builder.sub(r.full(1,t.type),n).getResult(0);return[r.builder.mul(e,s).getResult(0)]});ce("sigmoid",r=>{let e=r.gradOutputs[0],t=r.results[0],n=r.builder.sub(r.full(1,t.type),t).getResult(0),s=r.builder.mul(t,n).getResult(0);return[r.builder.mul(e,s).getResult(0)]});ce("relu",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.full(0,t.type),s=r.builder.compare(t,n,"gt").getResult(0);return[r.builder.select(s,e,n).getResult(0)]});ce("gelu",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.full(1.702,t.type),s=r.full(1,t.type),o=r.builder.mul(n,t).getResult(0),i=r.builder.sigmoid(o).getResult(0),a=r.builder.sub(s,i).getResult(0),u=r.builder.mul(o,a).getResult(0),p=r.builder.add(s,u).getResult(0),c=r.builder.mul(i,p).getResult(0);return[r.builder.mul(e,c).getResult(0)]});ce("silu",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.builder.sigmoid(t).getResult(0),s=r.builder.sub(r.full(1,t.type),n).getResult(0),o=r.builder.mul(t,s).getResult(0),i=r.builder.add(n,r.builder.mul(n,o).getResult(0)).getResult(0);return[r.builder.mul(e,i).getResult(0)]});ce("sin",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.builder.cos(t).getResult(0);return[r.builder.mul(e,n).getResult(0)]});ce("cos",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.builder.sin(t).getResult(0),s=r.builder.neg(n).getResult(0);return[r.builder.mul(e,s).getResult(0)]});ce("abs",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.builder.sign(t).getResult(0);return[r.builder.mul(e,n).getResult(0)]});l(PM,"_erfDerivIR");l($M,"_digammaIR");ce("erf",r=>{let e=r.gradOutputs[0],[t]=r.operands;return[r.builder.mul(e,PM(r,t)).getResult(0)]});ce("erfc",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.builder.neg(PM(r,t)).getResult(0);return[r.builder.mul(e,n).getResult(0)]});ce("lgamma",r=>{let e=r.gradOutputs[0],[t]=r.operands;return[r.builder.mul(e,$M(r,t)).getResult(0)]});ce("gamma",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.results[0],s=r.builder.mul(n,$M(r,t)).getResult(0);return[r.builder.mul(e,s).getResult(0)]});ce("log2",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.full(Math.LN2,t.type),s=r.builder.mul(t,n).getResult(0);return[r.builder.div(e,s).getResult(0)]});ce("log10",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.full(Math.LN10,t.type),s=r.builder.mul(t,n).getResult(0);return[r.builder.div(e,s).getResult(0)]});ce("exp2",r=>{let e=r.gradOutputs[0],t=r.results[0],n=r.full(Math.LN2,t.type),s=r.builder.mul(t,n).getResult(0);return[r.builder.mul(e,s).getResult(0)]});ce("square",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.full(2,t.type),s=r.builder.mul(n,t).getResult(0);return[r.builder.mul(e,s).getResult(0)]});ce("reciprocal",r=>{let e=r.gradOutputs[0],t=r.results[0],n=r.builder.mul(t,t).getResult(0),s=r.builder.neg(n).getResult(0);return[r.builder.mul(e,s).getResult(0)]});ce("rsqrt",r=>{let e=r.gradOutputs[0],t=r.results[0],n=r.builder.mul(r.builder.mul(t,t).getResult(0),t).getResult(0),s=r.full(-.5,t.type),o=r.builder.mul(s,n).getResult(0);return[r.builder.mul(e,o).getResult(0)]})});var zM=M(()=>{"use strict";ls();ce("dot",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=r.op.getAttr("lhs_contracting"),o=r.op.getAttr("rhs_contracting"),i=r.op.getAttr("lhs_batch")||[],a=r.op.getAttr("rhs_batch")||[],u=t.type.rank,p=n.type.rank,c=e.type.rank,h=[];for(let _=0;_<u;_++)!s.includes(_)&&!i.includes(_)&&h.push(_);let f=[];for(let _=0;_<p;_++)!o.includes(_)&&!a.includes(_)&&f.push(_);let d=[];for(let _=i.length;_<i.length+f.length;_++)d.push(_+h.length);let m=[];for(let _=i.length;_<i.length+h.length;_++)m.push(_);let g=r.builder.dot(e,n,d,f,Array.from({length:i.length},(_,b)=>b),a).getResult(0),y=r.builder.dot(t,e,h,m,i,Array.from({length:i.length},(_,b)=>b)).getResult(0);return[g,y]});ce("matmul",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=l(p=>{let c=Array.from({length:p},(h,f)=>f);return c[p-2]=p-1,c[p-1]=p-2,c},"swapLastTwo"),o=r.builder.transpose(n,s(n.type.rank)).getResult(0),i=r.builder.matmul(e,o).getResult(0),a=r.builder.transpose(t,s(t.type.rank)).getResult(0),u=r.builder.matmul(a,e).getResult(0);return[i,u]});ce("conv",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=r.builder,o=r.op.getAttr("strides"),i=r.op.getAttr("padding"),a=r.op.getAttr("dilation")||o.map(()=>1),u=r.op.getAttr("groups")||1,p=r.op.getAttr("input_layout"),c=r.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 f=n.type.shape,d=f[2],m=f[3],[g,y]=i,_=s.reverse(s.transpose(n,[1,0,2,3]).getResult(0),[2,3]).getResult(0),b=[[d-1-g[0],d-1-g[1]],[m-1-y[0],m-1-y[1]]],w=s.conv(e,_,[1,1],b).getResult(0),T=s.transpose(t,[1,0,2,3]).getResult(0),S=s.transpose(e,[1,0,2,3]).getResult(0),x=s.conv(T,S,[1,1],[g,y]).getResult(0),k=s.transpose(x,[1,0,2,3]).getResult(0);return[w,k]})});var VM=M(()=>{"use strict";ls();ce("reduce",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.op.getAttr("reduce_type"),s=r.op.getAttr("dimensions"),o=t.type.shape,i=t.type.dtype,a=new Set(s),u=o.map((h,f)=>a.has(f)?1:h),p=o.map((h,f)=>f),c=r.builder.reshape(e,u).getResult(0);if(n==="sum")return[r.builder.broadcast(c,o,p).getResult(0),null];if(n==="mean"){let h=1;for(let y=0;y<o.length;y++)a.has(y)&&(h*=o[y]);let f=r.builder.broadcast(c,o,p).getResult(0),d=r.builder.scalarConstant(h,i).getResult(0),m=r.builder.broadcast(d,o,[]).getResult(0);return[r.builder.div(f,m).getResult(0),null]}if(n==="max"||n==="min"){let h=r.builder.reshape(r.results[0],u).getResult(0),f=r.builder.broadcast(h,o,p).getResult(0),d=r.builder.compare(t,f,"eq").getResult(0),m=r.builder.convert(d,i).getResult(0),g=r.builder.broadcast(c,o,p).getResult(0);return[r.builder.mul(g,m).getResult(0),null]}throw n==="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 '${n}' on the gradient path (would silently drop the gradient)`)})});var GM=M(()=>{"use strict";ls();ce("reshape",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=t.type.shape;return[r.builder.reshape(e,n).getResult(0)]});ce("reverse",r=>[r.builder.reverse(r.gradOutputs[0],r.op.getAttr("dimensions")).getResult(0)]);ce("transpose",r=>{let e=r.gradOutputs[0],t=r.op.getAttr("permutation"),n=new Array(t.length);for(let s=0;s<t.length;s++)n[t[s]]=s;return[r.builder.transpose(e,n).getResult(0)]});ce("broadcast_in_dim",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.op.getAttr("broadcast_dimensions"),s=t.type.shape,o=r.results[0].type.shape,i=t.type.dtype,a=[];for(let h=0;h<o.length;h++)if(!n.includes(h))a.push(h);else{let f=n.indexOf(h);s[f]===1&&o[h]!==1&&a.push(h)}if(a.length===0)return[e];let u=r.builder.scalarConstant(0,i).getResult(0),p=r.builder.reduce(e,u,a,"sum").getResult(0),c=p.type.shape;return c.length!==s.length||!c.every((h,f)=>h===s[f])?[r.builder.reshape(p,s).getResult(0)]:[p]});ce("slice",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.op.getAttr("starts"),s=t.type.shape,o=e.type.shape,i=r.op.getAttr("strides")||s.map(()=>1),a=t.type.dtype,u=[...n],p=new Array(s.length),c=new Array(s.length);for(let f=0;f<s.length;f++)c[f]=i[f]-1,p[f]=s[f]-n[f]-(o[f]-1)*i[f]-1;let h=r.builder.scalarConstant(0,a).getResult(0);return[r.builder.pad(e,h,u,p,c).getResult(0)]});ce("concat",r=>{let e=r.gradOutputs[0],t=r.op.getAttr("dimension"),n=r.operands,s=[],o=0;for(let i=0;i<n.length;i++){let a=n[i].type.shape,u=new Array(a.length).fill(0),p=[...e.type.shape];u[t]=o,p[t]=o+a[t],s.push(r.builder.slice(e,u,p).getResult(0)),o+=a[t]}return s});ce("gather",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=r.builder.scalarConstant(0,t.type.dtype).getResult(0),o=r.builder.broadcast(s,t.type.shape,[]).getResult(0);return[r.builder.scatterAdd(o,n,e,{updateWindowDims:r.op.getAttr("offset_dims"),insertedWindowDims:r.op.getAttr("collapsed_slice_dims"),scatterDimsToOperandDims:r.op.getAttr("start_index_map"),indexVectorDim:r.op.getAttr("index_vector_dim")}).getResult(0),null]});ce("scatter",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=r.op.getAttr("inserted_window_dims"),o=t.type.shape.map((a,u)=>s.includes(u)?1:a),i=r.builder.gather(e,n,{offsetDims:r.op.getAttr("update_window_dims"),collapsedSliceDims:s,startIndexMap:r.op.getAttr("scatter_dims_to_operand_dims"),indexVectorDim:r.op.getAttr("index_vector_dim"),sliceSizes:o}).getResult(0);return[e,null,i]});ce("pad",r=>{let e=r.gradOutputs[0],t=r.op.getAttr("low"),n=r.op.getAttr("high"),[s]=r.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[r.builder.slice(e,i,a).getResult(0),null]})});var jM=M(()=>{"use strict";ls();es();ce("softmax",r=>{let e=r.gradOutputs[0],t=r.results[0],n=r.op.getAttr("axis"),s=t.type.dtype,o=t.type.shape,i=r.builder.mul(e,t).getResult(0),a=r.builder.scalarConstant(0,s).getResult(0),u=r.builder.reduce(i,a,[n],"sum").getResult(0),p=To(o.length,n),c=r.builder.broadcast(u,o,p).getResult(0),h=r.builder.sub(e,c).getResult(0);return[r.builder.mul(t,h).getResult(0)]});ce("log_softmax",r=>{let e=r.gradOutputs[0],t=r.results[0],[n]=r.operands,s=r.op.getAttr("axis"),o=t.type.dtype,i=t.type.shape,a=r.builder.exp(t).getResult(0),u=r.builder.scalarConstant(0,o).getResult(0),p=r.builder.reduce(e,u,[s],"sum").getResult(0),c=To(i.length,s),h=r.builder.broadcast(p,i,c).getResult(0),f=r.builder.mul(a,h).getResult(0);return[r.builder.sub(e,f).getResult(0)]});ce("layer_norm",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=r.op.getAttr("axis"),o=r.op.getAttr("epsilon"),i=t.type.dtype,a=t.type.shape,u=[s<0?a.length+s:s],p=a[u[0]],c=To(a.length,u[0]),h=r.builder.scalarConstant(0,i).getResult(0),f=r.builder.reduce(t,h,u,"mean").getResult(0),d=r.builder.broadcast(f,a,c).getResult(0),m=r.builder.sub(t,d).getResult(0),g=r.builder.mul(m,m).getResult(0),y=r.builder.reduce(g,h,u,"mean").getResult(0),_=r.full(o,y.type),b=r.builder.add(y,_).getResult(0),w=r.builder.rsqrt(b).getResult(0),T=r.builder.broadcast(w,a,c).getResult(0),S=[];for(let he=0;he<a.length;he++)u.includes(he)&&S.push(he);let x=r.builder.broadcast(n,a,S).getResult(0),k=r.builder.mul(e,x).getResult(0),A=r.full(p,t.type),N=r.builder.mul(A,k).getResult(0),v=r.builder.reduce(k,h,u,"sum").getResult(0),O=r.builder.broadcast(v,a,c).getResult(0),R=r.builder.mul(m,T).getResult(0),B=r.builder.mul(k,R).getResult(0),E=r.builder.reduce(B,h,u,"sum").getResult(0),z=r.builder.broadcast(E,a,c).getResult(0),D=r.builder.mul(R,z).getResult(0),J=r.builder.sub(N,r.builder.add(O,D).getResult(0)).getResult(0),j=r.builder.div(T,A).getResult(0),U=r.builder.mul(j,J).getResult(0),P=r.builder.reduce(r.builder.mul(e,R).getResult(0),h,c,"sum").getResult(0),V=r.builder.reduce(e,h,c,"sum").getResult(0);return[U,P,V]});ce("scaled_dot_product_attention",r=>{let e=r.gradOutputs[0],[t,n,s]=r.operands,o=r.op.getAttr("scale");if(r.op.getAttr("causal"))throw new Error("causal scaled_dot_product_attention VJP not supported");let i=r.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"),h=i.matmul(t,c(n)).getResult(0),f=i.mul(h,r.full(o,h.type)).getResult(0),d=i.softmax(f,u-1).getResult(0),m=i.matmul(c(d),e).getResult(0),g=i.matmul(e,c(s)).getResult(0),y=i.scalarConstant(0,a).getResult(0),_=i.mul(g,d).getResult(0),b=i.reduce(_,y,[u-1],"sum").getResult(0),w=[];for(let N=0;N<u-1;N++)w.push(N);let T=i.broadcast(b,d.type.shape,w).getResult(0),S=i.mul(d,i.sub(g,T).getResult(0)).getResult(0),x=i.mul(S,r.full(o,S.type)).getResult(0),k=i.matmul(x,n).getResult(0),A=i.matmul(c(x),t).getResult(0);return[k,A,m]});ce("pool2d",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.builder,s=r.op.getAttr("pool_type"),o=r.op.getAttr("kernel_size"),i=r.op.getAttr("strides"),a=r.op.getAttr("padding"),u=r.op.getAttr("layout")||"NCHW",p=a.every(k=>k[0]===0&&k[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[h,f,d,m]=e.type.shape,[g,y]=o,_=t.type.shape,b=l(k=>n.reshape(n.broadcast(k,[h,f,d,g,m,y],[0,1,2,4]).getResult(0),_).getResult(0),"upsample"),w=b(e);if(s==="avg"){let k=r.full(g*y,t.type);return[n.div(w,k).getResult(0)]}let T=b(r.results[0]),S=n.compare(t,T,"eq").getResult(0),x=r.full(0,t.type);return[n.select(S,w,x).getResult(0)]});ce("batch_norm",r=>{let e=r.gradOutputs[0],[t,n,,s,o]=r.operands,i=r.op.getAttr("axis"),a=r.op.getAttr("epsilon"),u=t.type.dtype,p=t.type.shape,c=r.builder,h=To(p.length,i),f=c.scalarConstant(0,u).getResult(0),d=r.full(a,o.type),m=c.rsqrt(c.add(o,d).getResult(0)).getResult(0),g=c.broadcast(m,p,[i]).getResult(0),y=c.broadcast(s,p,[i]).getResult(0),_=c.broadcast(n,p,[i]).getResult(0),b=c.sub(t,y).getResult(0),w=c.mul(b,g).getResult(0),T=c.mul(e,_).getResult(0),S=c.mul(T,g).getResult(0),x=c.reduce(c.mul(e,w).getResult(0),f,h,"sum").getResult(0),k=c.reduce(e,f,h,"sum").getResult(0),A=c.neg(c.reduce(S,f,h,"sum").getResult(0)).getResult(0),N=c.mul(c.mul(g,g).getResult(0),g).getResult(0),v=r.full(-.5,t.type),O=c.mul(c.mul(c.mul(T,b).getResult(0),N).getResult(0),v).getResult(0),R=c.reduce(O,f,h,"sum").getResult(0);return[S,x,k,A,R]});ce("elu",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.results[0],s=r.op.getAttr("alpha")??1,o=r.full(0,t.type),i=r.full(1,t.type),a=r.full(s,t.type),u=r.builder.compare(t,o,"gt").getResult(0),p=r.builder.add(n,a).getResult(0),c=r.builder.select(u,i,p).getResult(0);return[r.builder.mul(e,c).getResult(0)]});ce("leaky_relu",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.op.getAttr("negative_slope")??.01,s=r.full(0,t.type),o=r.full(1,t.type),i=r.full(n,t.type),a=r.builder.compare(t,s,"gt").getResult(0),u=r.builder.select(a,o,i).getResult(0);return[r.builder.mul(e,u).getResult(0)]});ce("celu",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.op.getAttr("alpha")??1,s=r.full(0,t.type),o=r.full(1,t.type),i=r.builder.compare(t,s,"gt").getResult(0),a=r.full(n,t.type),u=r.builder.div(t,a).getResult(0),p=r.builder.exp(u).getResult(0),c=r.builder.select(i,o,p).getResult(0);return[r.builder.mul(e,c).getResult(0)]});ce("selu",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=1.0507009873554805,s=1.6732632423543772,o=r.full(0,t.type),i=r.full(n,t.type),a=r.builder.compare(t,o,"gt").getResult(0),u=r.full(s,t.type),p=r.builder.exp(t).getResult(0),c=r.builder.mul(u,p).getResult(0),h=r.builder.select(a,i,r.builder.mul(i,c).getResult(0)).getResult(0);return[r.builder.mul(e,h).getResult(0)]});ce("hardswish",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.full(-3,t.type),s=r.full(3,t.type),o=r.full(0,t.type),i=r.full(1,t.type),a=r.full(2,t.type),u=r.full(6,t.type),p=r.builder.compare(t,n,"le").getResult(0),c=r.builder.compare(t,s,"ge").getResult(0),h=r.builder.add(r.builder.mul(a,t).getResult(0),s).getResult(0),f=r.builder.div(h,u).getResult(0),d=r.builder.select(p,o,r.builder.select(c,i,f).getResult(0)).getResult(0);return[r.builder.mul(e,d).getResult(0)]});ce("hardsigmoid",r=>{let e=r.gradOutputs[0],[t]=r.operands,n=r.full(-3,t.type),s=r.full(3,t.type),o=r.full(0,t.type),i=r.full(1/6,t.type),a=r.builder.compare(t,n,"le").getResult(0),u=r.builder.compare(t,s,"ge").getResult(0),p=r.builder.select(a,o,r.builder.select(u,o,i).getResult(0)).getResult(0);return[r.builder.mul(e,p).getResult(0)]});ce("embedding",r=>{let e=r.gradOutputs[0],[t,n]=r.operands,s=n.type.rank,o=r.full(0,t.type);return[r.builder.scatter(o,n,e,{updateWindowDims:[s],insertedWindowDims:[0],scatterDimsToOperandDims:[0],indexVectorDim:s}).getResult(0),null]})});var qM=M(()=>{"use strict";ls();bx("stop_gradient");ce("stop_gradient",()=>[null]);for(let r of["compare","logical_not","argmax","argmin","iota"])bx(r)});var UM=M(()=>{"use strict";ls();Nf();Of();xx();__();EM();DM();FM();zM();VM();GM();jM();qM()});var KM={};nt(KM,{compileWithBackward:()=>w_});function b_(r){return typeof r.then=="function"}function w_(r,e,t={}){let n=t.target??Us(),s=t.mode||"separate",o=t.rematPolicy||new Ea(t.remat||{}),i={target:n,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,h=null;function f(v){let O=x_((...B)=>r.forward(...B),v,{name:r.constructor.name||"compiled",dynamicShapes:a}),R=l(B=>{let E=B.graph.functions().next().value,z=s==="joint"?m(E,B,o):d(E,B,o);return z.shapeEnv=B.shapeEnv,z.outputSymShapes=B.outputSymShapes,z},"finish");return b_(O)?O.then(R):R(O)}l(f,"_compile");function d(v,O,R){let B=new Rf({rematPolicy:R}),{backwardFunc:E,savedValues:z,gradInputIndices:D}=B.build(v),J=v.getReturnOp(),j=[...J.operands],U=j.length,P=new Map(v.args.map((Se,Qe)=>[Se.id,Qe])),V=new Map(j.map((Se,Qe)=>[Se.id,Qe])),he=[],le=new Map;for(let Se of z)P.has(Se.id)||V.has(Se.id)||le.has(Se.id)||(le.set(Se.id,U+he.length),he.push(Se));he.length>0&&(J.erase(),new An(v).returnOp([...j,...he]),v.outputTypes=Object.freeze([...j.map(Se=>Se.type),...he.map(Se=>Se.type)]));let Ie=z.map(Se=>P.has(Se.id)?{kind:"arg",index:P.get(Se.id)}:V.has(Se.id)?{kind:"output",index:V.get(Se.id)}:{kind:"output",index:le.get(Se.id)}),G=new wn("forward");G.addFunction(v);let K=new Do(i).compile(G),ie=new wn("backward");ie.addFunction(E);let ze=new Do(i).compile(ie);return{mode:"separate",fwdResult:K,bwdResult:ze,forwardFunc:v,backwardFunc:E,savedValues:z,savedSources:Ie,numRealOutputs:U,gradInputIndices:D,capturedParams:O.capturedParams,numUserInputs:O.numUserInputs,outputTypes:O.outputTypes,shapeEnv:O.shapeEnv,outputSymShapes:O.outputSymShapes}}l(d,"_compileSeparate");function m(v,O,R){let B=new Mf({rematPolicy:R}),{jointFunc:E,numForwardOutputs:z,numGradInputs:D}=B.build(v),J=new wn("joint");return J.addFunction(E),{mode:"joint",result:new Do(i).compile(J),jointFunc:E,numForwardOutputs:z,numGradInputs:D,capturedParams:O.capturedParams,numUserInputs:O.numUserInputs,outputTypes:O.outputTypes,inputTypes:v.inputTypes,shapeEnv:O.shapeEnv,outputSymShapes:O.outputSymShapes}}l(m,"_compileJoint");function g(v,O){return v.outputSymShapes&&v.shapeEnv?v.shapeEnv.resolveSymbolicShape(v.outputSymShapes[O]):[...v.outputTypes[O].shape]}l(g,"_resolveOutputShape");function y(v,O,R){let B=v.module||v;return B.executionPlan?B.runPlanAsync(B.executionPlan,R):v.isAsync(O)?v.runAsync(O,...R):(v.run(O,...R),null)}l(y,"_runK");function _(v,O){let B=v.fwdResult.listKernels()[0],E=O.length>0?O[0].device:"cpu",z=O.map(G=>Je(G)),J=v.capturedParams.map(G=>Je(G)),j=v.forwardFunc.outputTypes,U=v.numRealOutputs,P=new Array(j.length),V=new Array(j.length);for(let G=0;G<j.length;G++){let K=G<U?g(v,G):[...j[G].shape],ie=j[G].dtype,ze=cn(K),Se=Dt(ie);P[G]=new Se(Math.max(ze,1)),V[G]=K}let he=[...z,...J,...P],le=l(()=>({results:U===1?tn(P[0],V[0],j[0].dtype,E):Array.from({length:U},(K,ie)=>tn(P[ie],V[ie],j[ie].dtype,E)),inputArrays:z,paramArrays:J,outputArrays:P,device:E}),"build"),Ie=y(v.fwdResult,B,he);return Ie?Ie.then(le):le()}l(_,"_executeSeparateForward");function b(v,O,R){let E=v.bwdResult.listKernels()[0],z=O.map(ie=>Je(ie)),D=v.savedValues,J=v.savedSources,j=[...R.inputArrays,...R.paramArrays],U=new Array(D.length);for(let ie=0;ie<D.length;ie++){let ze=J[ie];U[ie]=ze.kind==="arg"?j[ze.index]:R.outputArrays[ze.index]}let P=v.backwardFunc,V=P.outputTypes.length,he=new Array(V),le=new Array(V);for(let ie=0;ie<V;ie++){let ze=[...P.outputTypes[ie].shape],Se=P.outputTypes[ie].dtype,Qe=cn(ze),It=Dt(Se);he[ie]=new It(Math.max(Qe,1)),le[ie]=ze}let Ie=[...z,...U,...he],G=l(()=>he.map((ie,ze)=>tn(ie,le[ze],P.outputTypes[ze].dtype,R.device)),"build"),K=y(v.bwdResult,E,Ie);return K?K.then(G):G()}l(b,"_executeSeparateBackward");function w(v){for(let O=0;O<u.length;O++){let R=u[O];R.shapeEnv.bindInputShapes(v);let{passed:B}=R.shapeEnv.evaluateGuards();if(B)return R}return null}l(w,"_findCachedEntry");function T(v,O){if(c=v,v.mode==="joint")return k(v,O);let R=_(v,O);return b_(R)?R.then(B=>(p=B,B.results)):(p=R,R.results)}l(T,"_forwardWith");function S(v){let O=w(v);if(O)return T(O,v);let R=f(v);return b_(R)?R.then(B=>(u.push(B),B.shapeEnv.bindInputShapes(v),T(B,v))):(u.push(R),R.shapeEnv.bindInputShapes(v),T(R,v))}l(S,"_runForward");function x(...v){return h?h.then(()=>S(v)):S(v)}l(x,"compiledForward");function k(v,O){let B=v.result.listKernels()[0],E=O.length>0?O[0].device:"cpu",z=O.map(K=>Je(K)),J=v.capturedParams.map(K=>Je(K)),j=v.jointFunc,U=j.outputTypes.length,P=new Array(U),V=new Array(U);for(let K=0;K<U;K++){let ie=[...j.outputTypes[K].shape],ze=j.outputTypes[K].dtype,Se=cn(ie),Qe=Dt(ze);P[K]=new Qe(Math.max(Se,1)),V[K]=ie}let he=new Array(v.numForwardOutputs);for(let K=0;K<v.numForwardOutputs;K++){let ie=v.outputTypes[K],ze=cn(ie.shape),Se=Dt(ie.dtype);he[K]=new Se(Math.max(ze,1))}p={inputArrays:z,paramArrays:J,gradOutputArrays:he,outputArrays:P,outputShapes:V,device:E,compiled:v};let le=[...z,...J,...he,...P],Ie=l(()=>{let K=[];for(let ie=0;ie<v.numForwardOutputs;ie++)K.push(tn(P[ie],V[ie],j.outputTypes[ie].dtype,E));return K.length===1?K[0]:K},"build"),G=y(v.result,B,le);return G?G.then(Ie):Ie()}l(k,"_executeJointForward");let A=x;A.backward=function(...v){if(!c||!p)throw new Error("Must run forward before backward");return c.mode==="joint"?N(c,v,p):b(c,v,p)};function N(v,O,R){let{inputArrays:B,paramArrays:E,outputArrays:z,outputShapes:D,device:J}=R,j=O.map(K=>Je(K));for(let K=0;K<j.length;K++)R.gradOutputArrays[K].set(j[K]);let U=v.jointFunc,P=U.outputTypes.length,V=new Array(P),he=new Array(P);for(let K=0;K<P;K++){let ie=[...U.outputTypes[K].shape],ze=U.outputTypes[K].dtype,Se=cn(ie),Qe=Dt(ze);V[K]=new Qe(Math.max(Se,1)),he[K]=ie}let le=[...B,...E,...j,...V],Ie=l(()=>{let K=[];for(let ie=v.numForwardOutputs;ie<P;ie++)K.push(tn(V[ie],he[ie],U.outputTypes[ie].dtype,J));return K},"build"),G=y(v.result,v.result.listKernels()[0],le);return G?G.then(Ie):Ie()}if(l(N,"_executeJointBackward"),A.original=r,A.backwardGraph=()=>{if(u.length===0)return null;let v=u[0];return v.mode==="joint"?v.jointFunc:v.backwardFunc},A.forwardGraph=()=>{if(u.length===0)return null;let v=u[0];return v.mode==="joint"?v.jointFunc:v.forwardFunc},A.capturedParams=()=>u.length?u[0].capturedParams:[],e){let v=f(e);b_(v)?h=v.then(O=>(u.push(O),h=null,O)):u.push(v)}return A}var T_=M(()=>{"use strict";Sx();l_();Ks();Yo();Of();es();xx();__();UM();Kn();ts();jr();l(b_,"_isThenable");l(w_,"compileWithBackward")});function La(r){return typeof r.then=="function"}function s3(){WM||(vM(),WM=!0)}function o3(r,e){if(!r)return new Array(e.length).fill(null);let t=new Array(e.length);for(let n=0;n<e.length;n++){let s=r[n];if(s===!0){let o=new Set;for(let i=0;i<e[n].shape.length;i++)o.add(i);t[n]=o}else s instanceof Set?t[n]=s:t[n]=null}return t}function x_(r,e,t){s3();let n=t?.name||r.name||"traced",s=new Lg(n),o=o3(t?.dynamicShapes,e);for(let h=0;h<e.length;h++)s.createInput(e[h].shape,e[h].dtype,o[h]);let i=e.length,a=s._initGraph();function u(h){Array.isArray(h)?s.markOutputs(h):s.markOutput(h),s.deactivate();let f=s.getGraphModule(),d=f.functions().next().value;return{graph:f,capturedParams:[...s.capturedParams],numUserInputs:i,outputTypes:d.outputTypes,shapeEnv:s.shapeEnv,outputSymShapes:s.outputSymShapes}}l(u,"_finalize"),s.activate();let p=Pr.fromKey(48),c=LS(p,()=>r(...a));if(La(c))return c.then(h=>u(h),h=>{throw s.deactivate(),h});try{return u(c)}catch(h){throw s.deactivate(),h}}function HM(r,e,t){let n=x_(r,e,t);return La(n)?n.then(s=>s.graph):n.graph}function i3(r,e,t){let n=r.result.listKernels();if(n.length===0)throw new Error("No kernels compiled");let s=n[0],o=e.length>0?e[0].device:"cpu",i=new Array(e.length);for(let g=0;g<e.length;g++)i[g]=Je(e[g]);let a=r.capturedParams,u=new Array(a.length);for(let g=0;g<a.length;g++)u[g]=Je(a[g]);let p=r.outputTypes,c=r.outputSymShapes,h=new Array(p.length),f=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=cn(y),w=Dt(_);h[g]=new w(Math.max(b,1)),f[g]=[...y]}let d=new Array(i.length+u.length+h.length),m=0;for(let g=0;g<i.length;g++)d[m++]=new to(i[g],e[g].shape,e[g].dtype);for(let g=0;g<u.length;g++){let y=new to(u[g],a[g].shape,a[g].dtype),_=a[g]._impl;_&&(y.resident={key:_.storage.rawData,version:_.version}),d[m++]=y}for(let g=0;g<h.length;g++)d[m++]=new to(h[g],f[g],p[g].dtype);return{funcName:s,device:o,outputTypes:p,outputArrays:h,outputShapes:f,allArgs:d}}function kx(r,e,t,n){if(e.length===1)return tn(t[0],n[0],e[0].dtype,r);let s=new Array(e.length);for(let o=0;o<e.length;o++)s[o]=tn(t[o],n[o],e[o].dtype,r);return s}function a3(r,e,t){let{funcName:n,device:s,outputTypes:o,outputArrays:i,outputShapes:a,allArgs:u}=i3(r,e,t),p=r.result.module.executionPlan;return p?r.result.module.runPlanAsync(p,u,{resident:!0}).then(()=>kx(s,o,i,a)):r.result.isAsync(n)?r.result.runAsync(n,...u).then(()=>kx(s,o,i,a)):(r.result.run(n,...u),kx(s,o,i,a))}function u3(r,e,t={}){if(t?.backward)return w_(r,e,t);let n=t?.target??Us(),s={target:n,verify:!1,...t},o=t?.dynamic_shapes||null,i=t?.shapeBuckets||null,a=t?.foldWeights??t?.quantization?.foldWeights??!1,u=[];function p(w,T,S){if(!w||typeof w!="object"||w.repro)return w;try{w.repro={name:r.constructor?.name||"compiled",phase:S,target:n?.name,inputs:(T||[]).map(x=>({shape:x.shape,dtype:x.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 T=a?AM(w,Je):w;return{result:new Do(s).compile(T.graph),graph:T.graph,capturedParams:T.capturedParams,numUserInputs:T.numUserInputs,outputTypes:T.outputTypes,shapeEnv:T.shapeEnv,outputSymShapes:T.outputSymShapes}}l(c,"_finalize");function h(w,T){try{let S=x_((...x)=>r.forward(...x),w,{name:r.constructor.name||"compiled",dynamicShapes:T});return La(S)?S.then(c,x=>{throw p(x,w,"compile")}):c(S)}catch(S){throw p(S,w,"compile")}}l(h,"_compileWith");function f(w){return h(w,o)}l(f,"_compile");function d(w){return w.map((T,S)=>({shape:T,dtype:e[S].dtype}))}l(d,"_bucketInputs");function m(w){for(let T=0;T<u.length;T++){let S=u[T];S.shapeEnv.bindInputShapes(w);let{passed:x}=S.shapeEnv.evaluateGuards();if(x)return S}return null}l(m,"_findCachedEntry");function g(w,T){try{let S=a3(w,T,w.shapeEnv);return La(S)?S.then(void 0,x=>{throw p(x,T,"run")}):S}catch(S){throw p(S,T,"run")}}l(g,"_execute");function y(...w){let T=m(w);if(!T){let S=f(w);if(La(S))return S.then(x=>(u.push(x),x.shapeEnv.bindInputShapes(w),g(x,w)));T=S,u.push(T),T.shapeEnv.bindInputShapes(w)}return g(T,w)}l(y,"compiledForward");let _=null;if(e){let w=[];if(i)for(let T of i)w.push(h(d(T),null));if(w.push(f(e)),w.some(T=>La(T)))_=Promise.all(w).then(T=>{for(let S of T)u.push(S)});else for(let T of w)La(T)||u.push(T)}let b=y;return b.original=r,b.graph=w=>HM((...S)=>r.forward(...S),w||e,{name:r.constructor.name||"compiled",dynamicShapes:o}),b.source=()=>{if(u.length===0)return null;let w=u[0],T=w.result.listKernels();return T.length>0?w.result.getSource(T[0]):null},b.kernels=()=>u.length===0?[]:u[0].result.listKernels(),b.snippet=()=>{if(u.length===0)return null;let w=u[0],T=w.result.listKernels();return T.length>0?w.result.getSnippet(T[0]):null},b.result=()=>u.length>0?u[0].result:null,b._ready=_,b}var WM,Sx=M(()=>{"use strict";ya();NM();Nn();ub();l_();Ks();Kn();nm();ts();jr();OM();T_();WM=!1;l(La,"_isThenable");l(s3,"_ensureTracing");l(o3,"_normalizeDynamicShapes");l(x_,"_traceCore");l(HM,"trace");l(i3,"_prepareExecution");l(kx,"_wrapOutputs");l(a3,"executeCompiled");l(u3,"compile")});var QM={};nt(QM,{beginEagerCapture:()=>g3,endEagerCapture:()=>y3,replay:()=>_3,syncStream:()=>b3});var S_,g3,y3,_3,b3,XM=M(()=>{"use strict";S_=l(r=>{throw new Error(`mlfw: CUDA eager-graph (${r}) is not available in the browser`)},"unavailable"),g3=l(()=>S_("beginEagerCapture"),"beginEagerCapture"),y3=l(()=>S_("endEagerCapture"),"endEagerCapture"),_3=l(()=>S_("replay"),"replay"),b3=l(()=>S_("syncStream"),"syncStream")});var eI={};nt(eI,{clearCapturePins:()=>k3,deviceBufferDptr:()=>x3,deviceBufferForInput:()=>T3,pinResident:()=>S3,setEagerDeferred:()=>w3});var Ef,w3,T3,x3,S3,k3,tI=M(()=>{"use strict";Ef=l(r=>{throw new Error(`mlfw: CUDA resident memory (${r}) is not available in the browser`)},"unavailable"),w3=l(()=>Ef("setEagerDeferred"),"setEagerDeferred"),T3=l(()=>Ef("deviceBufferForInput"),"deviceBufferForInput"),x3=l(()=>Ef("deviceBufferDptr"),"deviceBufferDptr"),S3=l(()=>Ef("pinResident"),"pinResident"),k3=l(()=>Ef("clearCapturePins"),"clearCapturePins")});var rI={};nt(rI,{copyDeviceToHost:()=>v3,copyHostToDeviceAsync:()=>N3});var nI,v3,N3,sI=M(()=>{"use strict";nI=l(r=>{throw new Error(`mlfw: CUDA memory (${r}) is not available in the browser`)},"unavailable"),v3=l(()=>nI("copyDeviceToHost"),"copyDeviceToHost"),N3=l(()=>nI("copyHostToDeviceAsync"),"copyHostToDeviceAsync")});gs();Nn();Kn();vm();jn();Oo();ea();ar();jr();ts();function Xu(r,e){let t=xn(r),n=En.allocate(0,e,Au),s=new Bt(n,0,r,t,e,Au);return new We(s)}l(Xu,"_metaTensor");function aD(r,e,t){let n=Ou(e.shape,t.shape);if(!n)throw new Error(`Incompatible shapes: [${e.shape}] vs [${t.shape}]`);let s=Rd(e.dtype,t.dtype);return Xu(n,s)}l(aD,"_metaBinary");function uD(r,e){return Xu([...e.shape],e.dtype)}l(uD,"_metaUnary");function lD(r,e,t,n){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 Xu(n?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)?n&&i.push(1):i.push(s[u]);return Xu(i,e.dtype)}l(lD,"_metaReduction");function cD(r,e,t){let n=Bd(e.shape,t.shape);if(n===null)throw new Error("metaMatmul: unsupported shapes");return Xu(n,e.dtype)}l(cD,"metaMatmul");function pD(r,e){return Xu([...e.shape],e.dtype)}l(pD,"metaClone");function _0(r,e){return Object.fromEntries(r.map(t=>[t,e]))}l(_0,"mapKernel");var qA=Object.freeze({..._0(["add","sub","mul","div","pow","rem","maximum","minimum"],aD),..._0(["neg","exp","log","sqrt","rsqrt","abs","sin","cos","tanh","erf","erfc","lgamma","gamma","sigmoid","relu","gelu","silu","sign","floor","ceil"],uD),..._0(["sum","mean","max","min","prod"],lD),matmul:cD,clone:pD});jn();Oo();ea();ts();jr();function UA(r,e,t,n){let s=r.length,o=t<0?s+t:t,i=n<0?s+n:n,a=[...r],u=[...e];return a[o]=r[i],a[i]=r[o],u[o]=e[i],u[i]=e[o],{sizes:a,strides:u}}l(UA,"computeTranspose");function KA(r,e,t){let n=r.length,s=new Array(n),o=new Array(n);for(let i=0;i<n;i++){let a=t[i],u=a<0?n+a:a;s[i]=r[u],o[i]=e[u]}return{sizes:s,strides:o}}l(KA,"computePermute");function WA(r,e,t){let n=r.length,s=t.length,o=s-n,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(r[p]===1&&t[u]!==1)a[u]=t[u],i[u]=0;else if(r[p]===t[u]||t[u]===-1)a[u]=r[p],i[u]=e[p];else throw new Error(`Cannot expand size ${r[p]} to ${t[u]} at dim ${u}`)}return{sizes:a,strides:i}}l(WA,"computeExpand");function HA(r,e,t,n,s,o){let i=r.length,a=t<0?i+t:t,u=r[a],p=n??0,c=s??u,h=o??1;p<0&&(p+=u),c<0&&(c+=u),p=Math.max(0,Math.min(p,u)),c=Math.max(0,Math.min(c,u));let f=Math.max(0,Math.ceil((c-p)/h)),d=p*e[a],m=[...r],g=[...e];return m[a]=f,g[a]=e[a]*h,{sizes:m,strides:g,offsetDelta:d}}l(HA,"computeSlice");function YA(r,e,t){let n=r.length,s=t<0?n+1+t:t,o=[...r],i=[...e],a=s<n?r[s]*e[s]:1;return o.splice(s,0,1),i.splice(s,0,a),{sizes:o,strides:i}}l(YA,"computeUnsqueeze");function JA(r,e,t){if(t!=null){let o=r.length,i=t<0?o+t:t;if(r[i]!==1)return{sizes:[...r],strides:[...e]};let a=[...r],u=[...e];return a.splice(i,1),u.splice(i,1),{sizes:a,strides:u}}let n=[],s=[];for(let o=0;o<r.length;o++)r[o]!==1&&(n.push(r[o]),s.push(e[o]));return{sizes:n,strides:s}}l(JA,"computeSqueeze");function ZA(r,e,t,n,s){let o=r.length,i=t<0?o+t:t,a=[...r];a[i]=s;let u=n*e[i];return{sizes:a,strides:[...e],offsetDelta:u}}l(ZA,"computeNarrow");function QA(r,e,t,n){let s=r.length,o=t<0?s+t:t,i=n<0?r[o]+n:n,a=[],u=[];for(let c=0;c<s;c++)c!==o&&(a.push(r[c]),u.push(e[c]));let p=i*e[o];return{sizes:a,strides:u,offsetDelta:p}}l(QA,"computeSelect");function Ms(r,e,t,n){let s=new Bt(r._impl.storage,r._impl.storageOffset+(n||0),e,t,r._impl.dtype,r._impl.device),o=r._impl.autogradMeta;return o&&s.setAutogradMeta(o),new We(s)}l(Ms,"makeView");function fD(r,e){let t=tN(r.shape,r.strides,e);if(!t)throw new Error(`Cannot reshape tensor of shape [${r.shape}] to [${e}]`);if(!t.needsCopy)return Ms(r,t.sizes,t.strides,0);let n=r.isContiguous?r:e1(r);return Ms(n,t.sizes,t.strides,0)}l(fD,"reshapeKernel");function hD(r,e,t){let{sizes:n,strides:s}=UA(r.shape,r.strides,e,t);return Ms(r,n,s,0)}l(hD,"transposeKernel");function dD(r,e){let{sizes:t,strides:n}=KA(r.shape,r.strides,e);return Ms(r,t,n,0)}l(dD,"permuteKernel");function mD(r,e){let{sizes:t,strides:n}=WA(r.shape,r.strides,e);return Ms(r,t,n,0)}l(mD,"expandKernel");function gD(r,e,t){let n=r.shape,s=r.strides,o=[...e],i=new Array(e.length).fill(0);for(let a=0;a<t.length;a++){let u=t[a];i[u]=n[a]===1&&e[u]!==1?0:s[a]}return Ms(r,o,i,0)}l(gD,"broadcastInDimKernel");function yD(r,e,t,n,s){let{sizes:o,strides:i,offsetDelta:a}=HA(r.shape,r.strides,e,t,n,s);return Ms(r,o,i,a)}l(yD,"sliceKernel");function _D(r,e){let{sizes:t,strides:n}=YA(r.shape,r.strides,e);return Ms(r,t,n,0)}l(_D,"unsqueezeKernel");function bD(r,e){let{sizes:t,strides:n}=JA(r.shape,r.strides,e);return Ms(r,t,n,0)}l(bD,"squeezeKernel");function wD(r,e,t,n){let{sizes:s,strides:o,offsetDelta:i}=ZA(r.shape,r.strides,e,t,n);return Ms(r,s,o,i)}l(wD,"narrowKernel");function TD(r,e,t){let{sizes:n,strides:s,offsetDelta:o}=QA(r.shape,r.strides,e,t);return Ms(r,n,s,o)}l(TD,"selectKernel");var XA=null;function e1(r){let e=r._impl;return r.isContiguous&&e.storageOffset===0&&e.storage.rawData.length===r.numel?r:xD(r)}l(e1,"contiguousKernel");var t1=Object.freeze({reshape:l((r,e,t)=>fD(e,t),"reshape"),transpose:l((r,e,t,n)=>hD(e,t,n),"transpose"),permute:l((r,e,t)=>dD(e,t),"permute"),broadcast_in_dim:l((r,e,t,n)=>gD(e,t,n),"broadcast_in_dim"),expand:l((r,e,t)=>mD(e,t),"expand"),slice:l((r,e,t,n,s,o)=>yD(e,t,n,s,o),"slice"),unsqueeze:l((r,e,t)=>_D(e,t),"unsqueeze"),squeeze:l((r,e,t)=>bD(e,t),"squeeze"),narrow:l((r,e,t,n,s)=>wD(e,t,n,s),"narrow"),select:l((r,e,t,n)=>TD(e,t,n),"select"),contiguous:l((r,e)=>e1(e),"contiguous")});function xD(r){if(XA){let g=XA(r);if(g)return g}let e=r.shape,t=r.strides,n=r._impl.storage.data,s=r._impl.storageOffset,o=r.numel,i=xn(e),a=r._impl.dtype,u=En.allocate(o*Xi(a),a,r._impl.device),p=u.data,c=e.length,h=new Int32Array(c),f=s;for(let g=0;g<o;g++){p[g]=n[f];for(let y=c-1;y>=0&&(h[y]++,f+=t[y],!(h[y]<e[y]));y--)f-=h[y]*t[y],h[y]=0}let d=new Bt(u,0,e,i,r._impl.dtype,r._impl.device),m=r._impl.autogradMeta;return m&&d.setAutogradMeta(m),new We(d)}l(xD,"copyContiguous");be();_t();qr();jr();function SD(r,e,t,n){let s=$e(r.shape,{dtype:r.dtype}),o=ra(s,e,t,n),i=ra(s,e,t,Ht(n.shape,{dtype:r.dtype}));return qn(Iu(i,$e(i.shape,{dtype:i.dtype})),o,r)}l(SD,"scatterKernel");function kD(r,e){let t=r.shape,n=t.length;if(e.length<n)throw new Error("repeat: reps length must be >= tensor rank");let s=e.length-n,o=s>0?[...Array(s).fill(1),...t]:t.slice(),i=[],a=[],u=[];for(let p=0;p<o.length;p++)i.push(1,o[p]),a.push(e[p],o[p]),u.push(e[p]*o[p]);return Cn(Lu(Cn(r,i),a),u)}l(kD,"repeatKernel");function vD(r,e){let t=r.shape.length,n=e.length<t?[...Array(t-e.length).fill(1),...e]:e;return Du(r,n)}l(vD,"tileKernel");function ND(r,e,t){let n=[],s=r.shape.length,o=t<0?s+t:t,i=0;for(let a of e)n.push(Un(r,o,i,a)),i+=a;return n}l(ND,"splitKernel");function AD(r,e,t){let n=r.shape.length,s=t<0?n+t:t,o=r.shape[s],i=Math.ceil(o/e);return Bo(r,i,s)}l(AD,"chunkKernel");function OD(r,e,t=0){let n=r.shape.length,s=vs(t,n),o=r.shape[s],i=(e%o+o)%o;return i===0?r:sa([Un(r,s,o-i,i),Un(r,s,0,o-i)],s)}l(OD,"rollKernel");function BD(r,e){let t=Array.isArray(e)?e:[e],n=r.shape.length,s=r;for(let o of t){let i=vs(o,n),a=s.shape[i],u=new Array(a);for(let p=0;p<a;p++)u[p]=a-1-p;s=li(s,i,Pt(u,{dtype:"i32"}))}return s}l(BD,"flipKernel");function RD(r,e=0){let t=r.shape.length,n=vs(e,t),s=r.shape[n],o=r;for(let i=1;i<s;i*=2){let a=[...o.shape];a[n]=i;let u=sa([$e(a,{dtype:o.dtype}),Un(o,n,0,s-i)],n);o=ae(o,u)}return o}l(RD,"cumsumKernel");function MD(r){let e=1;for(;e<r;)e<<=1;return e}l(MD,"nextPow2");function Mm(r,e,t){let n=r.shape.length,s=n-1,o=r.shape[s],i=MD(o),a=e?-1/0:1/0,u=r;if(i>o){let h=new Array(n).fill(0),f=new Array(n).fill(0);f[s]=i-o,u=na(r,h,f,a)}let p=new Array(n).fill(1);p[s]=i;let c=null;if(t){let h=new Array(i);for(let f=0;f<i;f++)h[f]=f;c=Cn(Pt(h,{dtype:"i32"}),p)}for(let h=2;h<=i;h<<=1)for(let f=h>>1;f>=1;f>>=1){let d=new Array(i),m=new Array(i);for(let T=0;T<i;T++){let S=T^f;d[T]=S;let x=(T&h)===0,k=e?!x:x;S>T?m[T]=k?1:0:m[T]=k?0:1}let g=Pt(d,{dtype:"i32"}),y=li(u,s,g),_=Ld(u,y),b=ai(u,y),w=Cn(Pt(m,{dtype:"f32"}),p);if(t){let T=li(c,s,g),S=qn(w,$c(_,u),$c(b,u));c=qn(S,c,T)}u=qn(w,_,b)}return i>o&&(u=Un(u,s,0,o),t&&(c=Un(c,s,0,o))),t?{values:u,indices:c}:u}l(Mm,"bitonicLastDim");function ID(r,e=-1,t=!1){let n=r.shape.length,s=vs(e,n);if(s===n-1)return Mm(r,t,!1);let o=gt(r,s,n-1),i=Mm(o,t,!1);return gt(i,s,n-1)}l(ID,"sortKernel");function i1(r,e,t){let n=r.shape.length,s=vs(e,n);if(s===n-1)return Mm(r,t,!0);let o=gt(r,s,n-1),i=Mm(o,t,!0);return{values:gt(i.values,s,n-1),indices:gt(i.indices,s,n-1)}}l(i1,"sortWithIndices");function ED(r,e=-1,t=!1){return i1(r,e,t).indices}l(ED,"argsortKernel");function LD(r,e,t=-1,n=!0){let s=r.shape.length,o=vs(t,s),{values:i,indices:a}=i1(r,o,n);return[Un(i,o,0,e),Un(a,o,0,e)]}l(LD,"topkKernel");var a1=Object.freeze({scatter:l((r,e,t,n,s)=>SD(e,t,n,s),"scatter"),repeat:l((r,e,t)=>kD(e,t),"repeat"),tile:l((r,e,t)=>vD(e,t),"tile"),split:l((r,e,t,n)=>ND(e,t,n),"split"),chunk:l((r,e,t,n)=>AD(e,t,n),"chunk"),roll:l((r,e,t,n)=>OD(e,t,n),"roll"),flip:l((r,e,t)=>BD(e,t),"flip"),cumsum:l((r,e,t)=>RD(e,t),"cumsum"),sort:l((r,e,t,n)=>ID(e,t,n),"sort"),argsort:l((r,e,t,n)=>ED(e,t,n),"argsort"),topk:l((r,e,t,n,s)=>LD(e,t,n,s),"topk")});var CD=3,u1=!1,DD=[0,1,2,3,5];function l1(){if(u1)return;u1=!0;let r=new Xt("mlc","DEF");for(let t of LA())r.def(t);let e=new Xt("mlc","IMPL");for(let[t,n]of Object.entries(qA))e.impl(t,CD,n);for(let[t,n]of Object.entries({...t1,...a1}))for(let s of DD)e.impl(t,s,n);y0()}l(l1,"registerNativeOps");gs();Nn();gs();var PD=["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"],c1=!1;function Im(){if(c1)return;c1=!0;let r=new Xt("mlc","DEF");for(let e of PD)r.def(e)}l(Im,"ensureLinalgSchemas");qr();Kn();ar();var $D=new Set(["cpu","wasm"]);function w0(r){let e=new Float64Array(r.length);for(let t=0;t<r.length;t++)e[t]=Number(r[t]);return e}l(w0,"float64From");function p1(r){if(!$D.has(r.device.type))throw new Error(`linalg/ml: host-readable device required (cpu or wasm), got '${r.device.type}'`)}l(p1,"requireHost");function Ue(r){if(r.ndim!==2)throw new Error(`linalg/ml: expected a 2-D matrix, got ${r.ndim}-D`);p1(r);let[e,t]=r.shape;return{data:w0(Je(r)),rows:e,cols:t}}l(Ue,"hostMatrix");function Rt(r){if(p1(r),r.ndim===1)return{data:w0(Je(r)),rows:r.shape[0],cols:1,wasVector:!0};if(r.ndim===2)return{data:w0(Je(r)),rows:r.shape[0],cols:r.shape[1],wasVector:!1};throw new Error(`linalg/ml: expected a 1-D or 2-D right-hand side, got ${r.ndim}-D`)}l(Rt,"hostColumns");function Q(r,e,t,n=ks){return Pt(r,{shape:e,dtype:t,device:n})}l(Q,"toHostTensor");qr();function f1(r,e){let t=Float64Array.from(r),n=new Int32Array(e);for(let o=0;o<e;o++)n[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 h=t[o*e+c];t[o*e+c]=t[i*e+c],t[i*e+c]=h}let p=n[o];n[o]=n[i],n[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 h=o+1;h<e;h++)t[p*e+h]-=c*t[o*e+h]}}return{lu:t,piv:n,sign:s}}l(f1,"luFactor");function FD(r,e,t,n,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]=n[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-=r[i*t+p]*o[p*s+a];o[i*s+a]=u}for(let i=t-1;i>=0;i--){let a=r[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-=r[i*t+c]*o[c*s+u];o[i*s+u]=p/a}}return o}l(FD,"luSolve");function T0(r,e,t,n){let{lu:s,piv:o}=f1(r,e);return FD(s,o,e,t,n)}l(T0,"solveHost");function h1(r,e){let t;try{t=f1(r,e)}catch{return 0}let n=t.sign;for(let s=0;s<e;s++)n*=t.lu[s*e+s];return n}l(h1,"detHost");function d1(r,e){let t=new Float64Array(e*e);for(let n=0;n<e;n++)for(let s=0;s<=n;s++){let o=r[n*e+s];for(let i=0;i<s;i++)o-=t[n*e+i]*t[s*e+i];if(n===s){if(o<=0)throw new Error("linalg.cholesky: matrix is not positive definite");t[n*e+s]=Math.sqrt(o)}else t[n*e+s]=o/t[s*e+s]}return t}l(d1,"choleskyHost");function ep(r,e,t){let n=t?.tol??1e-12,s=t?.maxSweeps??100,o=Float64Array.from(r),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 h=0;for(let f=0;f<e;f++)for(let d=f+1;d<e;d++)h+=o[f*e+d]*o[f*e+d];if(Math.sqrt(h)<n)break;for(let f=0;f<e;f++)for(let d=f+1;d<e;d++){let m=o[f*e+d];if(Math.abs(m)<1e-300)continue;let g=(o[d*e+d]-o[f*e+f])/(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 T=o[w*e+f],S=o[w*e+d];o[w*e+f]=_*T-b*S,o[w*e+d]=b*T+_*S}for(let w=0;w<e;w++){let T=o[f*e+w],S=o[d*e+w];o[f*e+w]=_*T-b*S,o[d*e+w]=b*T+_*S}for(let w=0;w<e;w++){let T=i[w*e+f],S=i[w*e+d];i[w*e+f]=_*T-b*S,i[w*e+d]=b*T+_*S}}}let a=Array.from({length:e},(c,h)=>h).sort((c,h)=>o[c*e+c]-o[h*e+h]),u=new Float64Array(e),p=new Float64Array(e*e);for(let c=0;c<e;c++){let h=a[c];u[c]=o[h*e+h];for(let f=0;f<e;f++)p[f*e+c]=i[f*e+h]}return{values:u,vectors:p}}l(ep,"eighHost");function GD(r,e,t,n){if(n){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+=r[p*t+i]*r[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+=r[o*t+u]*r[i*t+u];s[o*e+i]=a,s[i*e+o]=a}return s}l(GD,"gram");function jD(r,e,t,n,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+=r[i*t+p]*n[a*t+p];o[i*s+a]=u}return o}l(jD,"matmulRows");function da(r,e,t,n,s=GD,o=jD){let i=n?.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:h,vectors:f}=ep(s(r,e,t,!0),t,n),d=new Float64Array(a*t);for(let g=0;g<a;g++){let y=t-1-g;p[g]=Math.sqrt(Math.max(h[y],0));for(let _=0;_<t;_++){let b=f[_*t+y];c[_*a+g]=b,d[g*t+_]=b}}let m=o(r,e,t,d,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:h,vectors:f}=ep(s(r,e,t,!1),e,n),d=new Float64Array(a*e);for(let y=0;y<a;y++){let _=e-1-y;p[y]=Math.sqrt(Math.max(h[_],0));for(let b=0;b<e;b++){let w=f[b*e+_];u[b*a+y]=w,d[y*e+b]=w}}let m=new Float64Array(t*e);for(let y=0;y<e;y++)for(let _=0;_<t;_++)m[_*e+y]=r[y*t+_];let g=o(m,t,e,d,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(da,"svdHost");function g1(r,e,t,n,s,o){let i=o?.rcond??1e-12,{U:a,S:u,V:p,k:c}=da(r,e,t,o),h=i*(u.length?u[0]:0),f=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]*n[b*s+y];f[m*s+y]=g>h?_/g:0}}let d=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+_]*f[_*s+g];d[m*s+g]=y}return d}l(g1,"lstsqHost");function y1(r,e,t,n){let s=n?.rcond??1e-12,{U:o,S:i,V:a,k:u}=da(r,e,t,n),p=s*(i.length?i[0]:0),c=new Float64Array(t*e);for(let h=0;h<t;h++)for(let f=0;f<e;f++){let d=0;for(let m=0;m<u;m++){let g=i[m];g>p&&(d+=a[h*u+m]*(1/g)*o[f*u+m])}c[h*e+f]=d}return c}l(y1,"pinvHost");function _1(r,e,t){let n=new Float64Array(t);for(let i=0;i<t;i++){let a=0;for(let u=0;u<e;u++)a+=r[u*t+i];n[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+=(r[c*t+i]-n[i])*(r[c*t+a]-n[a]);let p=u/s;o[i*t+a]=p,o[a*t+i]=p}return o}l(_1,"covHost");function tp(r,e){let t=Ue(r);if(t.rows!==t.cols)throw new Error(`linalg.${e}: matrix must be square`);return t}l(tp,"squareMatrix");function b1(r,e){let{data:t,rows:n}=tp(e,"cholesky");return Q(d1(t,n),[n,n],e.dtype,e.device)}l(b1,"cpuCholesky");function w1(r,e,t){let{data:n,rows:s}=tp(e,"solve"),o=Rt(t);if(o.rows!==s)throw new Error("linalg.solve: right-hand side rows must match matrix");let i=T0(n,s,o.data,o.cols);return Q(i,o.wasVector?[s]:[s,o.cols],e.dtype,e.device)}l(w1,"cpuSolve");function T1(r,e,t){let{data:n,rows:s,cols:o}=Ue(e),i=Rt(t);if(i.rows!==s)throw new Error("linalg.lstsq: right-hand side rows must match matrix");let a=g1(n,s,o,i.data,i.cols);return Q(a,i.wasVector?[o]:[o,i.cols],e.dtype,e.device)}l(T1,"cpuLstsq");function x1(r,e){let{data:t,rows:n}=tp(e,"inv"),s=new Float64Array(n*n);for(let i=0;i<n;i++)s[i*n+i]=1;let o=T0(t,n,s,n);return Q(o,[n,n],e.dtype,e.device)}l(x1,"cpuInv");function S1(r,e){let{data:t,rows:n,cols:s}=Ue(e);return Q(y1(t,n,s),[s,n],e.dtype,e.device)}l(S1,"cpuPinv");function k1(r,e){let{data:t,rows:n}=tp(e,"det");return Dc(h1(t,n),{dtype:e.dtype})}l(k1,"cpuDet");function v1(r,e){let{data:t,rows:n,cols:s}=Ue(e);return Q(_1(t,n,s),[s,s],e.dtype,e.device)}l(v1,"cpuCov");function N1(r,e){let{data:t,rows:n}=tp(e,"eigh"),{values:s,vectors:o}=ep(t,n);return[Q(s,[n],e.dtype,e.device),Q(o,[n,n],e.dtype,e.device)]}l(N1,"cpuEigh");function Em(r,e){let{data:t,rows:n,cols:s}=Ue(e),{U:o,S:i,V:a,k:u}=da(t,n,s);return[Q(o,[n,u],e.dtype,e.device),Q(i,[u],e.dtype,e.device),Q(a,[s,u],e.dtype,e.device)]}l(Em,"cpuSvd");var qD={svd:Em,eigh:N1,cholesky:b1,inv:x1,pinv:S1,det:k1,cov:v1,solve:w1,lstsq:T1},UD=[0,2],A1=!1;function O1(){if(A1)return;A1=!0,Im();let r=new Xt("mlc","IMPL");for(let[e,t]of Object.entries(qD))for(let n of UD)r.impl(e,n,t)}l(O1,"registerCpuLinalg");gs();Nn();gs();var KD=["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"],B1=!1;function Lm(){if(B1)return;B1=!0;let r=new Xt("mlc","DEF");for(let e of KD)r.def(e)}l(Lm,"ensureMlSchemas");function cr(r){let e=(r??2654435769)>>>0;return l(function(){e|=0,e=e+1831565813|0;let n=Math.imul(e^e>>>15,1|e);return n=n+Math.imul(n^n>>>7,61|n)^n,((n^n>>>14)>>>0)/4294967296},"next")}l(cr,"makeRng");function el(r,e){return Math.floor(r()*e)}l(el,"randInt");function S0(r,e){let t=Array.from({length:r},(n,s)=>s);for(let n=r-1;n>0;n--){let s=el(e,n+1),o=t[n];t[n]=t[s],t[s]=o}return t}l(S0,"shuffledIndices");function k0(r,e,t,n,s){let o=0;for(let i=0;i<s;i++){let a=r[e*s+i]-t[n*s+i];o+=a*a}return o}l(k0,"sqDist");function v0(r,e,t,n,s){let o=new Float64Array(n*t),i=el(s,e);for(let u=0;u<t;u++)o[u]=r[i*t+u];let a=new Float64Array(e).fill(1/0);for(let u=1;u<n;u++){let p=0;for(let f=0;f<e;f++){let d=k0(r,f,o,u-1,t);d<a[f]&&(a[f]=d),p+=a[f]}let c=s()*p,h=e-1;for(let f=0;f<e;f++)if(c-=a[f],c<=0){h=f;break}for(let f=0;f<t;f++)o[u*t+f]=r[h*t+f]}return o}l(v0,"kmeansPlusPlus");function R1(r,e,t,n,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 h=k0(r,a,n,c,t);h<p&&(p=h,u=c)}o[a]=u,i+=p}return i}l(R1,"assign");function N0(r,e,t,n,s,o){let i=new Float64Array(n*t),a=new Int32Array(n);for(let p=0;p<e;p++){let c=s[p];a[c]++;for(let h=0;h<t;h++)i[c*t+h]+=r[p*t+h]}let u=!1;for(let p=0;p<n;p++)if(a[p]!==0)for(let c=0;c<t;c++){let h=i[p*t+c]/a[p];h!==o[p*t+c]&&(u=!0),o[p*t+c]=h}return u}l(N0,"updateCenters");function WD(r,e,t,n,s,o){let i=v0(r,e,t,n,o),a=new Int32Array(e),u=1/0;for(let p=0;p<s&&(u=R1(r,e,t,i,n,a),!!N0(r,e,t,n,a,i));p++);return u=R1(r,e,t,i,n,a),{centers:i,labels:a,inertia:u}}l(WD,"lloyd");function Cm(r,e,t,n,s,o){let{data:i,rows:a,cols:u}=Ue(e),p=null;for(let h=0;h<s;h++){let f=cr(o+h*2654435769),d=WD(i,a,u,t,n,f);(!p||d.inertia<p.inertia)&&(p=d)}let c=new Float64Array(a);for(let h=0;h<a;h++)c[h]=p.labels[h];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(Cm,"cpuKmeans");function Dm(r,e,t){let{data:n,rows:s,cols:o}=Ue(e),i=Ue(t),a=new Float64Array(s);for(let u=0;u<s;u++){let p=0,c=1/0;for(let h=0;h<i.rows;h++){let f=k0(n,u,i.data,h,o);f<c&&(c=f,p=h)}a[u]=p}return Q(a,[s],e.dtype,e.device)}l(Dm,"cpuKmeansPredict");function A0(r,e,t,n){for(let s=0;s<t;s++)e[s]=s;for(let s=0;s<n;s++){let o=s;for(let a=s+1;a<t;a++)r[e[a]]<r[e[o]]&&(o=a);let i=e[s];e[s]=e[o],e[o]=i}}l(A0,"topKSelect");function O0(r,e,t,n){if(n){let o=new Map,i=r[e[0]],a=0;for(let u=0;u<t;u++){let p=r[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+=r[e[o]];return s/t}l(O0,"knnVote");function Pm(r,e,t,n,s,o){let i=Ue(e),a=Ue(n),u=Rt(t),p=i.cols,c=Math.min(s,i.rows),h=new Float64Array(a.rows),f=new Float64Array(i.rows),d=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}f[g]=y}A0(f,d,i.rows,c),h[m]=O0(u.data,d,c,o)}return Q(h,[a.rows],n.dtype,n.device)}l(Pm,"cpuKnnPredict");be();qr();function ns(r,e){let t=Pt(Int32Array.from(e),{shape:[e.length],dtype:"i32",device:r.device});return ot("index_select",r,t,0)}l(ns,"takeRows");function tl(r){return Ue(r)}l(tl,"matrixOf");function zt(r){let e=Rt(r);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(zt,"vectorOf");function np(r,e,t,n){return Q(r,[e,t],n??"f32")}l(np,"matrix");function B0(r,e,t){return Q(r,[e],t??"f32")}l(B0,"vector");function rs(r,e){let t=[],n=new Map,s=new Int32Array(e);for(let o=0;o<e;o++){let i=r[o],a=n.get(i);a===void 0&&(a=t.length,n.set(i,a),t.push(i)),s[o]=a}return{y:s,classes:t}}l(rs,"encodeLabels");var HD=1e-9;function M1(r,e,t){let{data:n,rows:s,cols:o}=Ue(e),i=Rt(t),{y:a,classes:u}=rs(i.data,i.rows),p=u.length,c=new Float64Array(p*o),h=new Float64Array(p*o),f=new Float64Array(p),d=new Int32Array(p);for(let _=0;_<s;_++){let b=a[_];d[b]++;for(let w=0;w<o;w++)c[b*o+w]+=n[_*o+w]}for(let _=0;_<p;_++){for(let b=0;b<o;b++)c[_*o+b]/=d[_];f[_]=d[_]/s}for(let _=0;_<s;_++){let b=a[_];for(let w=0;w<o;w++){let T=n[_*o+w]-c[b*o+w];h[b*o+w]+=T*T}}let m=0;for(let _=0;_<p;_++)for(let b=0;b<o;b++)h[_*o+b]/=d[_],h[_*o+b]>m&&(m=h[_*o+b]);let g=HD*m;for(let _=0;_<p*o;_++)h[_]+=g;let y=new Float64Array(p);for(let _=0;_<p;_++)y[_]=u[_];return[Q(c,[p,o],e.dtype,e.device),Q(h,[p,o],e.dtype,e.device),Q(f,[p],e.dtype,e.device),Q(y,[p],e.dtype,e.device)]}l(M1,"cpuGaussianNbFit");function I1(r,e,t,n,s,o){let{data:i,rows:a,cols:u}=Ue(e),p=Ue(t),c=Ue(n),h=Rt(s),f=Rt(o),d=p.rows,m=new Float64Array(a);for(let g=0;g<a;g++){let y=0,_=-1/0;for(let b=0;b<d;b++){let w=Math.log(h.data[b]);for(let T=0;T<u;T++){let S=c.data[b*u+T],x=i[g*u+T]-p.data[b*u+T];w+=-.5*(Math.log(2*Math.PI*S)+x*x/S)}w>_&&(_=w,y=b)}m[g]=f.data[y]}return Q(m,[a],e.dtype,e.device)}l(I1,"cpuGaussianNbPredict");function YD(r,e){return r>e?r-e:r<-e?r+e:0}l(YD,"softThreshold");function R0(r,e,t,n,s){let o=new Float64Array(t),i=0;if(s){for(let c=0;c<t;c++){let h=0;for(let f=0;f<e;f++)h+=r[f*t+c];o[c]=h/e}for(let c=0;c<e;c++)i+=n[c];i/=e}let a=new Float64Array(e*t);for(let c=0;c<e;c++)for(let h=0;h<t;h++)a[c*t+h]=r[c*t+h]-o[h];let u=new Float64Array(e);for(let c=0;c<e;c++)u[c]=n[c]-i;let p=new Float64Array(t);for(let c=0;c<t;c++){let h=0;for(let f=0;f<e;f++)h+=a[f*t+c]*a[f*t+c];p[c]=h/e}return{xc:a,r:u,z:p,meanX:o,meanY:i}}l(R0,"elasticNetPrep");function M0(r,e,t,n){let s=e;for(let o=0;o<n;o++)s-=r[o]*t[o];return s}l(M0,"elasticNetIntercept");function JD(r,e,t,n,s,o,i,a,u,p){for(let c=0;c<u;c++){let h=0;for(let f=0;f<t;f++){if(s[f]===0)continue;let d=0;for(let _=0;_<e;_++)d+=r[_*t+f]*n[_];let m=d/e+o[f]*s[f],g=YD(m,i)/(s[f]+a),y=g-o[f];if(y!==0){for(let _=0;_<e;_++)n[_]-=y*r[_*t+f];o[f]=g,Math.abs(y)>h&&(h=Math.abs(y))}}if(h<p)break}}l(JD,"coordinateDescent");function $m(r,e,t,n,s,o,i,a){let{data:u,rows:p,cols:c}=Ue(e),h=Rt(t),{xc:f,r:d,z:m,meanX:g,meanY:y}=R0(u,p,c,h.data,a),_=new Float64Array(c),b=n*s,w=n*(1-s);JD(f,p,c,d,m,_,b,w,o,i);let T=M0(g,y,_,c);return[Q(_,[c],e.dtype,e.device),Q(new Float64Array([T]),[1],e.dtype,e.device)]}l($m,"cpuElasticNet");function ZD(r,e,t){if(e<=0||e>=r)return Array.from({length:r},(s,o)=>o);let n=Array.from({length:r},(s,o)=>o);for(let s=r-1;s>r-1-e;s--){let o=Math.floor(t()*(s+1)),i=n[s];n[s]=n[o],n[o]=i}return n.slice(r-e)}l(ZD,"chooseFeatures");function QD(r,e,t,n,s,o,i){let a=t.length,u=null;for(let p of n){let c=t.slice().sort((m,g)=>r[m*e+p]-r[g*e+p]),h=new Float64Array(s),f=new Float64Array(s);for(let m of c)f[o[m]]++;let d=0;for(let m=0;m<a-1;m++){let g=o[c[m]];h[g]++,f[g]--,d++;let y=a-d,_=r[c[m]*e+p],b=r[c[m+1]*e+p];if(_===b||d<i||y<i)continue;let w=1,T=1;for(let x=0;x<s;x++){let k=h[x]/d,A=f[x]/y;w-=k*k,T-=A*A}let S=(d*w+y*T)/a;(!u||S<u.impurity)&&(u={feature:p,threshold:(_+b)/2,impurity:S})}}return u}l(QD,"bestSplitClassify");function XD(r,e,t,n,s,o){let i=t.length,a=null;for(let u of n){let p=t.slice().sort((g,y)=>r[g*e+u]-r[y*e+u]),c=0,h=0;for(let g of p)c+=s[g],h+=s[g]*s[g];let f=0,d=0,m=0;for(let g=0;g<i-1;g++){let y=s[p[g]];f+=y,d+=y*y,m++;let _=i-m,b=r[p[g]*e+u],w=r[p[g+1]*e+u];if(b===w||m<o||_<o)continue;let T=d-f*f/m,S=c-f,k=h-d-S*S/_,A=T+k;(!a||A<a.impurity)&&(a={feature:u,threshold:(b+w)/2,impurity:A})}}return a}l(XD,"bestSplitRegress");function E1(r,e,t,n,s,o,i,a,u){let p=Ue(e),c=p.data,h=p.cols,f=Rt(t),d=cr(u),m=null,g=null,y=0;if(a){let N=rs(f.data,f.rows);m=N.y,g=N.classes,y=g.length}let _=[],b=[],w=[],T=[],S=[],x=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 v=0;for(let O of N)v+=f.data[O];return v/N.length},"leafValue"),k=l(N=>{let v=a?m:f.data,O=v[N[0]];for(let R of N)if(v[R]!==O)return!1;return!0},"isPure"),A=l((N,v)=>{let O=_.length;if(_.push(-1),b.push(0),w.push(-1),T.push(-1),S.push(0),v>=n||N.length<s||k(N))return S[O]=x(N),O;let R=ZD(h,i,d),B=a?QD(c,h,N,R,y,m,o):XD(c,h,N,R,f.data,o);if(!B)return S[O]=x(N),O;let E=[],z=[];for(let D of N)c[D*h+B.feature]<=B.threshold?E.push(D):z.push(D);return _[O]=B.feature,b[O]=B.threshold,w[O]=A(E,v+1),T[O]=A(z,v+1),O},"build");return A(Array.from({length:p.rows},(N,v)=>v),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(T),[T.length],e.dtype,e.device),Q(Float64Array.from(S),[S.length],e.dtype,e.device)]}l(E1,"cpuDecisionTreeFit");function L1(r,e,t,n,s,o,i){let a=Ue(e),u=Rt(t).data,p=Rt(n).data,c=Rt(s).data,h=Rt(o).data,f=Rt(i).data,d=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=h[y];d[g]=f[y]}return Q(d,[a.rows],e.dtype,e.device)}l(L1,"cpuDecisionTreePredict");var eP={kmeans:Cm,kmeans_predict:Dm,knn_predict:Pm,gaussian_nb_fit:M1,gaussian_nb_predict:I1,elastic_net:$m,decision_tree_fit:E1,decision_tree_predict:L1},tP=[0,2],nP=[1,5];function rP(r){return()=>{throw new Error(`ml.${r}: scalar-iterative algorithm runs on CPU/WASM only; GPU/WebGPU not supported (no performance benefit)`)}}l(rP,"unsupported");var C1=!1;function D1(){if(C1)return;C1=!0,Lm();let r=new Xt("mlc","IMPL");for(let[e,t]of Object.entries(eP)){for(let n of tP)r.impl(e,n,t);for(let n of nP)r.impl(e,n,rP(e))}}l(D1,"registerCpuMl");gs();Nn();gs();var sP=["qr(Tensor input) -> (Tensor, Tensor)","fft(Tensor input) -> Tensor","ifft(Tensor input) -> Tensor"],P1=!1;function $1(){if(P1)return;P1=!0;let r=new Xt("mlc","DEF");for(let e of sP)r.def(e)}l($1,"ensureNumericSchemas");Kn();function F1(r,e,t){let n=Math.min(e,t),s=Float64Array.from(r),o=[];for(let u=0;u<n;u++){let p=0;for(let d=u;d<e;d++)p+=s[d*t+u]*s[d*t+u];if(p=Math.sqrt(p),p===0){o.push(null);continue}let c=s[u*t+u]>=0?-p:p,h=new Float64Array(e-u);h[0]=s[u*t+u]-c;for(let d=u+1;d<e;d++)h[d-u]=s[d*t+u];let f=0;for(let d=0;d<h.length;d++)f+=h[d]*h[d];if(f===0){o.push(null);continue}o.push(h);for(let d=u;d<t;d++){let m=0;for(let y=0;y<h.length;y++)m+=h[y]*s[(u+y)*t+d];let g=2*m/f;for(let y=0;y<h.length;y++)s[(u+y)*t+d]-=g*h[y]}}let i=new Float64Array(e*n);for(let u=0;u<n;u++)i[u*n+u]=1;for(let u=n-1;u>=0;u--){let p=o[u];if(!p)continue;let c=0;for(let h=0;h<p.length;h++)c+=p[h]*p[h];for(let h=0;h<n;h++){let f=0;for(let m=0;m<p.length;m++)f+=p[m]*i[(u+m)*n+h];let d=2*f/c;for(let m=0;m<p.length;m++)i[(u+m)*n+h]-=d*p[m]}}let a=new Float64Array(n*t);for(let u=0;u<n;u++)for(let p=u;p<t;p++)a[u*t+p]=s[u*t+p];return{Q:i,R:a,k:n}}l(F1,"qrHost");function oP(r,e){let t=r.length;for(let n=1,s=0;n<t;n++){let o=t>>1;for(;s&o;o>>=1)s^=o;s^=o,n<s&&([r[n],r[s]]=[r[s],r[n]],[e[n],e[s]]=[e[s],e[n]])}}l(oP,"bitReversePermute");function Fm(r,e,t){let n=r.length;oP(r,e);for(let s=2;s<=n;s<<=1){let o=(t?1:-1)*2*Math.PI/s,i=Math.cos(o),a=Math.sin(o);for(let u=0;u<n;u+=s){let p=1,c=0,h=s>>1;for(let f=0;f<h;f++){let d=r[u+f],m=e[u+f],g=r[u+f+h]*p-e[u+f+h]*c,y=r[u+f+h]*c+e[u+f+h]*p;r[u+f]=d+g,e[u+f]=m+y,r[u+f+h]=d-g,e[u+f+h]=m-y;let _=p*i-c*a;c=p*a+c*i,p=_}}}if(t)for(let s=0;s<n;s++)r[s]/=n,e[s]/=n}l(Fm,"fftRadix2");function iP(r){let e=1;for(;e<r;)e<<=1;return e}l(iP,"nextPow2");function aP(r,e,t){let n=r.length,s=iP(2*n-1),o=t?1:-1,i=new Float64Array(n),a=new Float64Array(n);for(let f=0;f<n;f++){let d=f*f%(2*n),m=o*Math.PI*d/n;i[f]=Math.cos(m),a[f]=Math.sin(m)}let u=new Float64Array(s),p=new Float64Array(s);for(let f=0;f<n;f++)u[f]=r[f]*i[f]-e[f]*a[f],p[f]=r[f]*a[f]+e[f]*i[f];let c=new Float64Array(s),h=new Float64Array(s);c[0]=i[0],h[0]=-a[0];for(let f=1;f<n;f++)c[f]=i[f],h[f]=-a[f],c[s-f]=i[f],h[s-f]=-a[f];Fm(u,p,!1),Fm(c,h,!1);for(let f=0;f<s;f++){let d=u[f]*c[f]-p[f]*h[f];p[f]=u[f]*h[f]+p[f]*c[f],u[f]=d}Fm(u,p,!0);for(let f=0;f<n;f++)r[f]=u[f]*i[f]-p[f]*a[f],e[f]=u[f]*a[f]+p[f]*i[f];if(t)for(let f=0;f<n;f++)r[f]/=n,e[f]/=n}l(aP,"bluestein");function z1(r,e,t){let n=r.length;n!==0&&((n&n-1)===0?Fm(r,e,t):aP(r,e,t))}l(z1,"fftHost");function V1(r,e){let{data:t,rows:n,cols:s}=Ue(e),{Q:o,R:i,k:a}=F1(t,n,s);return[Q(o,[n,a],e.dtype,e.device),Q(i,[a,s],e.dtype,e.device)]}l(V1,"cpuQr");function uP(r){let e=Je(r);if(r.ndim===1){let t=r.shape[0];return{re:Float64Array.from(e),im:new Float64Array(t),n:t}}if(r.ndim===2&&r.shape[1]===2){let t=r.shape[0],n=new Float64Array(t),s=new Float64Array(t);for(let o=0;o<t;o++)n[o]=e[2*o],s[o]=e[2*o+1];return{re:n,im:s,n:t}}throw new Error(`fft: expected a 1-D real signal or [n, 2] complex tensor, got shape [${r.shape}]`)}l(uP,"splitComplex");function lP(r,e,t){let n=new Float64Array(2*t);for(let s=0;s<t;s++)n[2*s]=r[s],n[2*s+1]=e[s];return n}l(lP,"interleave");function G1(r,e){let{re:t,im:n,n:s}=uP(r);return z1(t,n,e),Q(lP(t,n,s),[s,2],r.dtype,r.device)}l(G1,"transform");function j1(r,e){return G1(e,!1)}l(j1,"cpuFft");function q1(r,e){return G1(e,!0)}l(q1,"cpuIfft");var cP={qr:V1,fft:j1,ifft:q1},pP=[0,2],U1=!1;function K1(){if(U1)return;U1=!0,$1();let r=new Xt("mlc","IMPL");for(let[e,t]of Object.entries(cP))for(let n of pP)r.impl(e,n,t)}l(K1,"registerCpuNumeric");gs();Nn();kd();var it=8,E0=4,nl=2,fP=nl*it,ma=Math.log2(nl),hP=65536,W1=fP,dP=64*1024*1024;function H1(r,e){return Math.ceil(r/e)*e}l(H1,"alignUp");var I0=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>dP?this._instantiate():this._top=0}_ensure(e){let t=this._memory.buffer.byteLength;e<=t||this._memory.grow(Math.ceil((e-t)/hP))}alloc(e){let t=H1(this._top,W1);return this._top=t+H1(e,W1),this._ensure(this._top),t}allocF64(e){return this.alloc(e*it)}allocI32(e){return this.alloc(e*E0)}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)}},Y1=new Map;function so(r,e,t){let n=Y1.get(r);return n||(n=new I0(new WebAssembly.Module(Nu(e())),t),Y1.set(r,n)),n.reclaim(),n}l(so,"simdModule");var mP=1e308,gP=1,yP=32768;function rp(r,e,t,n){let s=e.map(([i,a])=>`(param $${i} ${a})`).join(" "),o=t.map(([i,a])=>`(local $${i} ${a})`).join(" ");return`(module (memory (export "memory") ${gP} ${yP})
+ (func (export "${r}") ${s}
+  ${o}
+${n}))`}l(rp,"moduleWrap");function J1(r,e,t,n,s,o){return`
+    (local.get $${n}) (i32.const ${ma}) i32.shr_s (i32.const ${ma}) i32.shl local.set $${o.nmain}
+    (f64.const 0) f64x2.splat local.set $${o.acc}
+    (i32.const 0) local.set $${o.j}
+    (block $${r}_ve (loop $${r}_vl
+      (local.get $${o.j}) (local.get $${o.nmain}) i32.ge_s br_if $${r}_ve
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add v128.load local.set $${o.va}
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${it}) 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 $${o.j}) (i32.const ${nl}) i32.add local.set $${o.j}
+      br $${r}_vl))
+    (local.get $${o.acc}) f64x2.extract_lane 0
+    (local.get $${o.acc}) f64x2.extract_lane 1
+    f64.add local.set $${s}
+    (block $${r}_te (loop $${r}_tl
+      (local.get $${o.j}) (local.get $${n}) i32.ge_s br_if $${r}_te
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add f64.load
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add f64.load
+      f64.sub local.set $${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 $${r}_tl))`}l(J1,"sqDistSnippet");var Z1={j:"j",nmain:"nmain",acc:"acc",va:"va",vb:"vb",diff:"diff"},Q1=[["j","i32"],["nmain","i32"],["acc","v128"],["va","v128"],["vb","v128"],["diff","f64"]];function L0(){let r=`
+    (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 ${it}) i32.mul i32.add local.set $xrow
+      (i32.const 0) local.set $best
+      (f64.const ${mP}) 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 ${it}) i32.mul i32.add local.set $crow
+${J1("nc","xrow","crow","d","dist",Z1)}
+        (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 ${E0}) 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 rp("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"],...Q1],r)}l(L0,"nearestCentroidWat");function X1(){let r=`
+    (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 ${it}) i32.mul i32.add local.set $trow
+${J1("dr","q","trow","d","dist",Z1)}
+      (local.get $out) (local.get $t) (i32.const ${it}) i32.mul i32.add (local.get $dist) f64.store
+      (local.get $t) (i32.const 1) i32.add local.set $t
+      br $tloop))`;return rp("dist_row",[["q","i32"],["tr","i32"],["ntr","i32"],["d","i32"],["out","i32"]],[["t","i32"],["trow","i32"],["dist","f64"],...Q1],r)}l(X1,"distRowWat");function C0(r,e,t,n,s,o){return`
+    (local.get $${n}) (i32.const ${ma}) i32.shr_s (i32.const ${ma}) i32.shl local.set $${o.nmain}
+    (f64.const 0) f64x2.splat local.set $${o.acc}
+    (i32.const 0) local.set $${o.j}
+    (block $${r}_ve (loop $${r}_vl
+      (local.get $${o.j}) (local.get $${o.nmain}) i32.ge_s br_if $${r}_ve
+      (local.get $${o.acc})
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add v128.load
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add v128.load
+      f64x2.mul f64x2.add local.set $${o.acc}
+      (local.get $${o.j}) (i32.const ${nl}) i32.add local.set $${o.j}
+      br $${r}_vl))
+    (local.get $${o.acc}) f64x2.extract_lane 0
+    (local.get $${o.acc}) f64x2.extract_lane 1
+    f64.add local.set $${s}
+    (block $${r}_te (loop $${r}_tl
+      (local.get $${o.j}) (local.get $${n}) i32.ge_s br_if $${r}_te
+      (local.get $${s})
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add f64.load
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add f64.load
+      f64.mul f64.add local.set $${s}
+      (local.get $${o.j}) (i32.const 1) i32.add local.set $${o.j}
+      br $${r}_tl))`}l(C0,"reduceDotSnippet");function _P(r,e,t,n,s,o){return`
+    (local.get $${n}) f64x2.splat local.set $${o.sv}
+    (local.get $${s}) (i32.const ${ma}) i32.shr_s (i32.const ${ma}) i32.shl local.set $${o.nmain}
+    (i32.const 0) local.set $${o.j}
+    (block $${r}_ve (loop $${r}_vl
+      (local.get $${o.j}) (local.get $${o.nmain}) i32.ge_s br_if $${r}_ve
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add local.set $${o.ya}
+      (local.get $${o.ya}) v128.load
+      (local.get $${o.sv})
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add v128.load
+      f64x2.mul f64x2.sub local.set $${o.vy}
+      (local.get $${o.ya}) (local.get $${o.vy}) v128.store
+      (local.get $${o.j}) (i32.const ${nl}) i32.add local.set $${o.j}
+      br $${r}_vl))
+    (block $${r}_te (loop $${r}_tl
+      (local.get $${o.j}) (local.get $${s}) i32.ge_s br_if $${r}_te
+      (local.get $${e}) (local.get $${o.j}) (i32.const ${it}) i32.mul i32.add local.set $${o.ya}
+      (local.get $${o.ya})
+      (local.get $${o.ya}) f64.load
+      (local.get $${n})
+      (local.get $${t}) (local.get $${o.j}) (i32.const ${it}) 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 $${r}_tl))`}l(_P,"axpySnippet");function eO(){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 ${it}) 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 ${it}) i32.mul i32.add local.set $rowJ
+${C0("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 ${it}) 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 ${it}) 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 rp("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(eO,"gramSymWat");function tO(){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 ${it}) 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 ${it}) i32.mul i32.add local.set $rowB
+${C0("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 ${it}) 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 rp("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(tO,"matmulRowsWat");function nO(){let r={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 ${it}) 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 ${it}) i32.mul i32.add local.set $col
+${C0("cd","col","r","n","dot",r)}
+          (local.get $w) (local.get $jc) (i32.const ${it}) 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 $delta) (f64.const 0) f64.ne
+          (if (then
+${_P("cd","r","col","delta","n",e)}
+            (local.get $w) (local.get $jc) (i32.const ${it}) 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 rp("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(nO,"coordDescentWat");var bP=5e5;function rO(r,e,t){let n=so("gram_sym",eO,"gram_sym"),s=n.allocF64(e*t),o=n.allocF64(e*e);return n.writeF64(s,r),n.run(s,e,t,o),Float64Array.from(n.f64(o,e*e))}l(rO,"gramSym");function wP(r,e,t,n){if(n){let s=new Float64Array(t*e);for(let o=0;o<e;o++)for(let i=0;i<t;i++)s[i*e+o]=r[o*t+i];return rO(s,t,e)}return rO(r,e,t)}l(wP,"wasmGram");function TP(r,e,t,n,s){let o=so("matmul_rows",tO,"matmul_rows"),i=o.allocF64(e*t),a=o.allocF64(s*t),u=o.allocF64(e*s);return o.writeF64(i,r),o.writeF64(a,n),o.run(i,e,t,a,s,u),Float64Array.from(o.f64(u,e*s))}l(TP,"wasmMatmulRows");function sO(r,e){let{data:t,rows:n,cols:s}=Ue(e);if(Math.max(n,s)*Math.min(n,s)*Math.min(n,s)<bP)return Em(r,e);let{U:i,S:a,V:u,k:p}=da(t,n,s,void 0,wP,TP);return[Q(i,[n,p],e.dtype,e.device),Q(a,[p],e.dtype,e.device),Q(u,[s,p],e.dtype,e.device)]}l(sO,"wasmSvd");var oO={svd:sO},xP={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"},iO=!1;function aO(){let r={enabled:Object.keys(oO),skipped:Object.keys(xP)};if(iO)return r;iO=!0,Im();let e=new Xt("mlc","IMPL");for(let[t,n]of Object.entries(oO))e.impl(t,2,n);return r}l(aO,"registerWasmLinalg");gs();Nn();var SP=8e3,kP=5e4,vP=2654435769;function uO(r,e,t,n,s,o,i,a,u,p){return r.writeF64(t,u),r.run(e,o,i,t,a,n,s),p.set(r.i32(n,o)),r.f64(s,1)[0]}l(uO,"assignWasm");function lO(r,e,t,n,s,o){let{data:i,rows:a,cols:u}=Ue(e);if(a*u*t<SP)return Cm(r,e,t,n,s,o);let p=so("kmeans_assign",L0,"nearest_centroid"),c=null;for(let f=0;f<s;f++){p.reset();let d=p.allocF64(a*u),m=p.allocF64(t*u),g=p.allocI32(a),y=p.allocF64(1);p.writeF64(d,i);let _=cr(o+f*vP),b=v0(i,a,u,t,_),w=new Int32Array(a),T=1/0;for(let S=0;S<n&&(T=uO(p,d,m,g,y,a,u,t,b,w),!!N0(i,a,u,t,w,b));S++);T=uO(p,d,m,g,y,a,u,t,b,w),(!c||T<c.inertia)&&(c={centers:Float64Array.from(b),labels:Int32Array.from(w),inertia:T})}let h=new Float64Array(a);for(let f=0;f<a;f++)h[f]=c.labels[f];return[Q(c.centers,[t,u],e.dtype,e.device),Q(h,[a],e.dtype,e.device),Q(new Float64Array([c.inertia]),[1],e.dtype,e.device)]}l(lO,"wasmKmeans");function cO(r,e,t){let{data:n,rows:s,cols:o}=Ue(e),i=Ue(t);if(s*o*i.rows<kP)return Dm(r,e,t);let a=so("kmeans_assign",L0,"nearest_centroid");a.reset();let u=a.allocF64(s*o),p=a.allocF64(i.rows*o),c=a.allocI32(s),h=a.allocF64(1);a.writeF64(u,n),a.writeF64(p,i.data),a.run(u,s,o,p,i.rows,c,h);let f=new Float64Array(s),d=a.i32(c,s);for(let m=0;m<s;m++)f[m]=d[m];return Q(f,[s],e.dtype,e.device)}l(cO,"wasmKmeansPredict");var NP=2e4;function pO(r,e,t,n,s,o){let i=Ue(e),a=Ue(n),u=Rt(t),p=i.cols;if(a.rows*i.rows*p<NP)return Pm(r,e,t,n,s,o);let c=so("knn_dist",X1,"dist_row");c.reset();let h=c.allocF64(a.rows*p),f=c.allocF64(i.rows*p),d=c.allocF64(i.rows);c.writeF64(h,a.data),c.writeF64(f,i.data);let m=Math.min(s,i.rows),g=new Float64Array(a.rows),y=new Int32Array(i.rows),_=p*it;for(let b=0;b<a.rows;b++){c.run(h+b*_,f,i.rows,p,d);let w=c.f64(d,i.rows);A0(w,y,i.rows,m),g[b]=O0(u.data,y,m,o)}return Q(g,[a.rows],n.dtype,n.device)}l(pO,"wasmKnnPredict");var AP=3e6;function fO(r,e,t,n,s,o,i,a){let{data:u,rows:p,cols:c}=Ue(e);if(p*c*o<AP)return $m(r,e,t,n,s,o,i,a);let h=Rt(t),{xc:f,r:d,z:m,meanX:g,meanY:y}=R0(u,p,c,h.data,a),_=new Float64Array(p*c);for(let O=0;O<c;O++)for(let R=0;R<p;R++)_[O*p+R]=f[R*c+O];let b=n*s,w=n*(1-s),T=so("coord_descent",nO,"coord_descent");T.reset();let S=T.allocF64(p*c),x=T.allocF64(p),k=T.allocF64(c),A=T.allocF64(c);T.writeF64(S,_),T.writeF64(x,d),T.writeF64(A,m),T.f64(k,c).fill(0),T.run(S,p,c,x,k,A,b,w,o,i);let N=Float64Array.from(T.f64(k,c)),v=M0(g,y,N,c);return[Q(N,[c],e.dtype,e.device),Q(new Float64Array([v]),[1],e.dtype,e.device)]}l(fO,"wasmElasticNet");var hO={kmeans:lO,kmeans_predict:cO,knn_predict:pO,elastic_net:fO},dO=!1;function mO(){let r={enabled:Object.keys(hO)};if(dO)return r;dO=!0,Lm();let e=new Xt("mlc","IMPL");for(let[t,n]of Object.entries(hO))e.impl(t,2,n);return r}l(mO,"registerWasmMl");Nn();nc();$i();sp();var OP=0,zm=class{static{l(this,"Edge")}node;inputNr;constructor(e,t){this.node=e,this.inputNr=t}},Be=class{static{l(this,"AutogradNode")}_id;_numInputs;_nextEdges;_savedTensors;_inputMetadata;_opArgs;constructor(e){this._id=OP++,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 zm(e,t))}setNextEdge(e,t,n){for(;this._nextEdges.length<=e;)this._nextEdges.push(null);this._nextEdges[e]=new zm(t,n)}saveTensor(e){this._savedTensors.push(e)}savedTensors(){return this._savedTensors}saveInputMetadata(e,t,n){this._inputMetadata[e]={shape:t,dtype:n}}inputMetadata(e){return this._inputMetadata[e]||null}apply(e){throw new Error(`${this.name()}.apply() not implemented`)}name(){return this.constructor.name}releaseVariables(){this._savedTensors=[],this._inputMetadata=[]}};be();_t();var Vm=class extends Be{static{l(this,"AddBackward")}constructor(){super(2)}apply(e){let t=e[0];return[t,t]}},Gm=class extends Be{static{l(this,"SubBackward")}constructor(){super(2)}apply(e){let t=e[0];return[t,Ur(t)]}},jm=class extends Be{static{l(this,"MulBackward")}constructor(){super(2)}apply(e){let t=e[0],[n,s]=this.savedTensors();return[ee(t,s.detach()),ee(t,n.detach())]}},qm=class extends Be{static{l(this,"DivBackward")}constructor(){super(2)}apply(e){let t=e[0],[n,s]=this.savedTensors(),o=$t(t,s.detach()),i=Ur($t(ee(t,n.detach()),ee(s.detach(),s.detach())));return[o,i]}},Um=class extends Be{static{l(this,"NegBackward")}constructor(){super(1)}apply(e){let t=e[0];return[Ur(t)]}},Km=class extends Be{static{l(this,"PowBackward")}constructor(){super(2)}apply(e){let t=e[0],[n,s]=this.savedTensors(),o=Ht(s.shape,{dtype:s.dtype,device:s.device}),i=ee(t,ee(s.detach(),Js(n.detach(),Pe(s.detach(),o)))),a=ee(t,ee(Js(n.detach(),s.detach()),Zs(n.detach())));return[i,a]}};be();_t();Ji();var BP=2/Math.sqrt(Math.PI);function bO(r){let e=et(r.shape,BP,{dtype:r.dtype,device:r.device});return ee(e,Kr(Ur(ee(r,r))))}l(bO,"_erfDeriv");function wO(r){let e=ae(r,Yi),t=Js(e,-1),n=Pe(Zs(e),ee(t,.5)),s=ee(t,t),o=s;for(let i of Nc)n=ae(n,ee(o,i)),o=ee(o,s);for(let i=0;i<Yi;i++)n=Pe(n,Js(ae(r,i),-1));return n}l(wO,"_digammaTensor");var Wm=class extends Be{static{l(this,"ErfBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[ee(e[0],bO(t.detach()))]}},Hm=class extends Be{static{l(this,"ErfcBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[Ur(ee(e[0],bO(t.detach())))]}},Ym=class extends Be{static{l(this,"LgammaBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[ee(e[0],wO(t.detach()))]}},Jm=class extends Be{static{l(this,"GammaBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),n=t.detach();return[ee(e[0],ee(Cd(n),wO(n)))]}},Zm=class extends Be{static{l(this,"ExpBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[ee(e[0],Kr(t.detach()))]}},Qm=class extends Be{static{l(this,"LogBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors();return[$t(e[0],t.detach())]}},Xm=class extends Be{static{l(this,"SqrtBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),n=t.detach(),s=ui(n),o=et(n.shape,2,{dtype:n.dtype,device:n.device});return[$t(e[0],ee(o,s))]}},eg=class extends Be{static{l(this,"TanhBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),n=t.detach(),s=Ns(n),o=Ht(n.shape,{dtype:n.dtype,device:n.device});return[ee(e[0],Pe(o,ee(s,s)))]}},tg=class extends Be{static{l(this,"SigmoidBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),n=t.detach(),s=ur(n),o=Ht(n.shape,{dtype:n.dtype,device:n.device});return[ee(e[0],ee(s,Pe(o,s)))]}},ng=class extends Be{static{l(this,"SoftmaxBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),n=this.opArgs(),s=n&&n.length>1&&n[1]!=null?n[1]:-1,o=e[0],i=Ro(t.detach(),s),a=ht(ee(o,i),s,!0);return[ee(i,Pe(o,a))]}},rg=class extends Be{static{l(this,"LogSoftmaxBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),n=this.opArgs(),s=n&&n.length>1&&n[1]!=null?n[1]:-1,o=e[0],i=Ro(t.detach(),s),a=ht(o,s,!0);return[Pe(o,ee(i,a))]}},sg=class extends Be{static{l(this,"ReluBackward")}constructor(){super(1)}apply(e){let[t]=this.savedTensors(),n=t.detach();return[ee(e[0],$d(Mu(n)))]}},og=class extends Be{static{l(this,"GeluBackward")}constructor(){super(1)}apply(e){let t=e[0],[n]=this.savedTensors(),s=n.detach(),o=s.shape,i=s.dtype,a=s.device,u=et(o,1.702,{dtype:i,device:a}),p=Ht(o,{dtype:i,device:a}),c=ee(u,s),h=ur(c),f=ee(h,ae(p,ee(c,Pe(p,h))));return[ee(t,f)]}},ig=class extends Be{static{l(this,"SiluBackward")}constructor(){super(1)}apply(e){let t=e[0],[n]=this.savedTensors(),s=n.detach(),o=ur(s),i=Ht(s.shape,{dtype:s.dtype,device:s.device}),a=ee(o,ae(i,ee(s,Pe(i,o))));return[ee(t,a)]}};be();_t();be();function TO(r,e){if(r==null){let n=[];for(let s=0;s<e;s++)n.push(s);return n}return(Array.isArray(r)?r:[r]).map(n=>n<0?n+e:n).sort((n,s)=>n-s)}l(TO,"_normalizeDims");function xO(r,e,t,n){let s=r;if(!n)for(let i of t)s=Sr(s,i);let o=$e(e,{dtype:s.dtype,device:s.device});return ae(o,s)}l(xO,"_unreduce");var ag=class extends Be{static{l(this,"SumBackward")}constructor(){super(1)}apply(e){let t=e[0],s=this.inputMetadata(0).shape,o=this.opArgs(),i=o?o[1]:void 0,a=o?o[2]:!1,u=TO(i,s.length);return[xO(t,s,u,a)]}},ug=class extends Be{static{l(this,"MeanBackward")}constructor(){super(1)}apply(e){let t=e[0],s=this.inputMetadata(0).shape,o=this.opArgs(),i=o?o[1]:void 0,a=o?o[2]:!1,u=TO(i,s.length),p=1;for(let f of u)p*=s[f];let c=xO(t,s,u,a),h=et(s,p,{dtype:t.dtype,device:t.device});return[$t(c,h)]}};be();be();function SO(r,e){let t=r;for(;t.ndim>e.length;)t=ht(t,0,!1);for(let n=0;n<e.length;n++)e[n]===1&&t.shape[n]!==1&&(t=ht(t,n,!0));return t}l(SO,"_sumToShape");var lg=class extends Be{static{l(this,"MatmulBackward")}constructor(){super(2)}apply(e){let t=e[0],[n,s]=this.savedTensors(),o=n.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 h=Sr(t,1),f=Sr(i,0);return[lt(h,f),Cu(lt(gt(o,0,1),h),1)]}if(a===1&&u===2){let h=lt(t,gt(i,0,1)),f=lt(Sr(o,1),Sr(t,0));return[h,f]}let p=lt(t,gt(i,u-2,u-1)),c=lt(gt(o,a-2,a-1),t);return p=SO(p,o.shape),c=SO(c,i.shape),[p,c]}},cg=class extends Be{static{l(this,"DotBackward")}constructor(){super(2)}apply(e){let t=e[0],[n,s]=this.savedTensors();return[ee(t,s.detach()),ee(t,n.detach())]}};be();_t();be();jr();function RP(r,e,t){r[e]+=t}l(RP,"_addAt");var pg=class extends Be{static{l(this,"CatBackward")}constructor(){super(0)}apply(e){let t=e[0],n=this.opArgs(),s=t.shape.length,o=vs(n&&n.length>1?n[1]??0:0,s),i=[],a=0,u=0;for(;this.inputMetadata(u);){let c=this.inputMetadata(u).shape[o];i.push(ci(Un(t,o,a,c))),a+=c,u++}return i}},fg=class extends Be{static{l(this,"StackBackward")}constructor(){super(0)}apply(e){let t=e[0],n=this.opArgs(),s=t.shape.length,o=vs(n&&n.length>1?n[1]??0:0,s),i=[],a=0;for(;this.inputMetadata(a);)i.push(ci(Ut(t,o,a))),a++;return i}},hg=class extends Be{static{l(this,"ClampBackward")}constructor(){super(3)}apply(e){let t=e[0],[n,s,o]=this.savedTensors(),i=$e(t.shape,{dtype:t.dtype,device:t.device}),a=zd(n.detach(),s.detach()),u=qn(a,t,i),p=Fd(n.detach(),o.detach());return[qn(p,u,i),null,null]}},dg=class extends Be{static{l(this,"PadBackward")}constructor(){super(2)}apply(e){let t=e[0],s=this.opArgs()[2],i=this.inputMetadata(0).shape,a=t;for(let u=0;u<i.length;u++){let p=s[u]||0;a=Un(a,u,p,i[u])}return[ci(a),null]}},mg=class r extends Be{static{l(this,"IndexSelectBackward")}static#e=null;static setGpuBackward(e){r.#e=e}constructor(){super(2)}apply(e){let t=e[0],[,n]=this.savedTensors(),o=this.inputMetadata(0).shape,i=o.length,a=this.opArgs(),u=vs(a&&a.length>2?a[2]??0:0,i);if(r.#e){let k=r.#e(t,n,o,u);if(k)return[k]}let p=$e(o,{dtype:t.dtype,device:t.device}),c=p._impl.storage.data,h=p.strides,f=ci(t),d=f._impl.storage.data,m=f._impl.storageOffset,g=f.shape,y=f.strides,_=ci(n),b=_._impl.storage.data,w=_._impl.storageOffset,T=g.length,S=new Int32Array(T),x=m;for(let k=0;k<f.numel;k++){let A=0;for(let N=0;N<T;N++){let v=N===u?Number(b[w+S[N]]):S[N];A+=v*h[N]}RP(c,A,d[x]);for(let N=T-1;N>=0;N--){if(S[N]++,S[N]<g[N]){x+=y[N];break}x-=(g[N]-1)*y[N],S[N]=0}}return[p,null]}},gg=class extends Be{static{l(this,"WhereBackward")}constructor(){super(3)}apply(e){let t=e[0],[n]=this.savedTensors(),s=$e(t.shape,{dtype:t.dtype,device:t.device}),o=qn(n.detach(),t,s),i=qn(n.detach(),s,t);return[null,o,i]}};be();_t();be();function MP(r,e,t){r[e]+=t}l(MP,"_addAt");var sl=class extends Be{static{l(this,"ReshapeBackward")}constructor(){super(1)}apply(e){let t=this.inputMetadata(0);return[Cn(e[0],t.shape)]}},yg=class extends Be{static{l(this,"TransposeBackward")}_dim0;_dim1;constructor(e,t){super(1),this._dim0=e,this._dim1=t}apply(e){return[gt(e[0],this._dim0,this._dim1)]}},op=class extends Be{static{l(this,"SliceBackward")}_dim;_start;_end;_step;constructor(e,t,n,s){super(1),this._dim=e,this._start=t,this._end=n,this._step=s}apply(e){let t=e[0],n=this.inputMetadata(0),s=this._dim,o=n.shape[s],i=this._step||1,a=this._start<0?this._start+o:this._start,u=this._end<0?this._end+o:this._end;if(a=Math.max(0,Math.min(a,o)),u=Math.max(0,Math.min(u,o)),i===1){let w=n.shape.map(()=>0),T=n.shape.map(()=>0);return w[s]=a,T[s]=o-u,[na(t,w,T,0)]}let p=$e(n.shape,{dtype:t.dtype,device:t.device}),c=p._impl.storage.data,h=t._impl.storage.data,f=t._impl.storageOffset,d=t.shape,m=t.strides,g=p.strides,y=d.length,_=new Int32Array(y),b=f;for(let w=0;w<t.numel;w++){let T=0;for(let S=0;S<y;S++){let x=S===s?a+_[S]*i:_[S];T+=x*g[S]}MP(c,T,h[b]);for(let S=y-1;S>=0;S--){if(_[S]++,_[S]<d[S]){b+=m[S];break}b-=(d[S]-1)*m[S],_[S]=0}}return[p]}},_g=class extends Be{static{l(this,"SelectBackward")}_dim;_index;constructor(e,t){super(1),this._dim=e,this._index=t}apply(e){let t=e[0],n=this.inputMetadata(0),s=this._dim,o=n.shape[s],i=this._index<0?this._index+o:this._index,a=Sr(t,s),u=n.shape.map(()=>0),p=n.shape.map(()=>0);return u[s]=i,p[s]=o-1-i,[na(a,u,p,0)]}},ip=class extends Be{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=ht(t,i,!0)),[Cn(u,s)]}},bg=class extends Be{static{l(this,"PermuteBackward")}_dims;constructor(e){super(1),this._dims=e}apply(e){let t=this._dims.length,n=new Array(t);for(let s=0;s<t;s++){let o=this._dims[s]<0?t+this._dims[s]:this._dims[s];n[o]=s}return[Eu(e[0],n)]}};var D0=new Map;function Fe(r,e){D0.set(r,e)}l(Fe,"_register");Fe("add",()=>new Vm);Fe("sub",()=>new Gm);Fe("mul",()=>new jm);Fe("div",()=>new qm);Fe("neg",()=>new Um);Fe("pow",()=>new Km);Fe("exp",()=>new Zm);Fe("log",()=>new Qm);Fe("sqrt",()=>new Xm);Fe("tanh",()=>new eg);Fe("erf",()=>new Wm);Fe("erfc",()=>new Hm);Fe("lgamma",()=>new Ym);Fe("gamma",()=>new Jm);Fe("sigmoid",()=>new tg);Fe("relu",()=>new sg);Fe("softmax",()=>new ng);Fe("log_softmax",()=>new rg);Fe("gelu",()=>new og);Fe("silu",()=>new ig);Fe("sum",()=>new ag);Fe("mean",()=>new ug);Fe("matmul",()=>new lg);Fe("dot",()=>new cg);Fe("cat",()=>new pg);Fe("stack",()=>new fg);Fe("clamp",()=>new hg);Fe("pad",()=>new dg);Fe("index_select",()=>new mg);Fe("where",()=>new gg);Fe("reshape",()=>new sl);Fe("transpose",r=>new yg(r[1],r[2]));Fe("permute",r=>new bg(r[1]));Fe("broadcast_in_dim",()=>new ip);Fe("expand",()=>new ip);Fe("slice",r=>new op(r[1],r[2],r[3],r[4]));Fe("narrow",r=>new op(r[1],r[2],r[2]+r[3],1));Fe("select",r=>new _g(r[1],r[2]));Fe("unsqueeze",()=>new sl);Fe("squeeze",()=>new sl);function kO(r,e=null){let t=D0.get(r);return t?t(e):null}l(kO,"getGradFn");function vO(r){return D0.has(r)}l(vO,"hasGradFn");be();var wg=class extends Be{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],n=this._variable.deref();if(!n)return[];let s=n._impl.autogradMeta;return s?(s.grad===null?s.grad=t:s.grad=ae(s.grad,t),[]):[]}name(){return"GradAccumulator"}};aw();Oo();jn();ar();Tm();jn();_t();be();function NO(r,e){let t=r.gradFn;if(!t)throw new Error("Cannot call backward on a tensor that does not require grad");let n=e;if(!n){if(r.numel!==1)throw new Error("grad must be specified for non-scalar tensors");n=Ht(r.shape,{dtype:r.dtype,device:r.device})}let s=new Map;EP(t,s,new Set);let i=new Map;i.set(t.id,[n]);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 h=c.node,f=s.get(h)-1;s.set(h,f),f===0&&a.push(h)}}}for(let p of u){let c=i.get(p.id);if(!c)continue;let h=p.apply(c),f=h;if(h&&(f=h.map((m,g)=>{if(!m)return m;let y=p.inputMetadata(g);return y?IP(m,y.shape):m})),p.releaseVariables(),!f)continue;let d=p.nextEdges;for(let m=0;m<d.length;m++){let g=d[m];if(!g||!g.node||m>=f.length||!f[m])continue;let y=f[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(NO,"backward");function IP(r,e){let t=r.shape;if(t.length===e.length&&t.every((o,i)=>o===e[i]))return r;let n=r,s=t.length-e.length;for(let o=0;o<s;o++)n=ht(n,0,!1);for(let o=e.length-1;o>=0;o--)e[o]===1&&n.shape[o]!==1&&(n=ht(n,o,!0));return n}l(IP,"_reduceBroadcastGrad");function EP(r,e,t){let n=[r];for(e.set(r,0);n.length>0;){let s=n.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)||n.push(i)}}}}l(EP,"_countDeps");oN({backward:NO});function LP(r){let e=r._impl;if(Zc()&&e.device&&e.device.type==="gpu"){e.storage.retain();let s=new Bt(e.storage,e.storageOffset,e.sizes(),e.strides(),e.dtype,e.device);return new We(s)}let t=e.storage.clone(),n=new Bt(t,e.storageOffset,e.sizes(),e.strides(),e.dtype,e.device);return new We(n)}l(LP,"_snapshotTensor");function ap(r){return typeof r=="object"&&r!==null&&"_impl"in r}l(ap,"_isTensor");function CP(r){for(let e=0;e<r.length;e++){let t=r[e];if(ap(t)&&t.requiresGrad)return!0;if(Array.isArray(t)){for(let n=0;n<t.length;n++)if(ap(t[n])&&t[n].requiresGrad)return!0}}return!1}l(CP,"_anyRequiresGrad");function DP(r){let e=[];for(let t=0;t<r.length;t++){let n=r[t];if(ap(n))e.push(n);else if(Array.isArray(n))for(let s=0;s<n.length;s++)ap(n[s])&&e.push(n[s])}return e}l(DP,"_extractTensors");function PP(r){let e=r._impl.autogradMeta;if(!e)return null;let t=e.getGradAccumulator();return t||(t=new wg(r),e.setGradAccumulator(t)),t}l(PP,"_getOrCreateAccumulator");function $P(r,e){let t=Nt.dispatch.bind(Nt);return(n,...s)=>{let o=n;if(!rl.isEnabled()||!CP(s)){let c=o.subtract(sc);return Nt.redispatch(e,c,...s)}let i=kO(r,s);if(!i){let c=o.subtract(sc);return Nt.redispatch(e,c,...s)}let a=DP(s);i.setOpArgs(s);for(let c=0;c<a.length;c++)i.saveTensor(LP(a[c])),i.saveInputMetadata(c,[...a[c].shape],a[c].dtype);for(let c=0;c<a.length;c++){let h=a[c];if(h.requiresGrad){let f=h.gradFn;if(f){let d=h._impl.autogradMeta?h._impl.autogradMeta.outputNr:0;i.setNextEdge(c,f,d)}else{let d=PP(h);d&&i.setNextEdge(c,d,0)}}}let u=o.subtract(sc),p=Nt.redispatch(e,u,...s);if(ap(p)){p._impl.setAutogradMeta(new Bu);let c=p._impl.autogradMeta;c.setGradFn(i,0),c.requiresGrad=!0,p._impl._updateKeySet()}return p}}l($P,"wrapWithAutograd");function FP(r){return(e,...t)=>{let n=e.subtract(sc);return Nt.redispatch(r,n,...t)}}l(FP,"_makePassthrough");function AO(){let r=Nt.listOps(),e=[40,41,42,43];for(let t of r){let n=Nt.findOp(t);if(!n)continue;let s=n.name,o=vO(s)?$P(s,n):FP(n),i=wr.fromUnboxed(o);for(let a of e)n.entry.registerKernel(a,i)}}l(AO,"registerAutogradKernels");be();qr();var lp=Ft,zP=["neg","exp","log","sqrt","rsqrt","abs","sin","cos","tanh","erf","erfc","lgamma","gamma","sigmoid","relu","gelu","silu","sign","floor","ceil","clone","contiguous"],VP=["add","sub","mul","div","pow","remainder","maximum","minimum","eq","ne","lt","le","gt","ge","matmul","dot","flip","unsqueeze"],GP=["sum","mean","max","min","argmax","argmin","prod"],jP={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 qP(r,e){let t=[];for(let n=0;n<e.length;n++)t[n]=r[n]===void 0?e[n]:r[n];for(let n=e.length;n<r.length;n++)t[n]=r[n];return t}l(qP,"withDefaults");function UP(r){for(let e of zP)r[e]=function(){return lp[e](this)}}l(UP,"installSelfMethods");function KP(r){for(let e of VP)r[e]=function(t){return lp[e](this,t)}}l(KP,"installOneArgMethods");function WP(r){for(let[e,t]of Object.entries(jP))r[e]=function(...n){return lp[e](this,...qP(n,t))}}l(WP,"installDefaultedMethods");function HP(r){for(let e of GP)r[e]=function(t,n){return lp[e](this,t,n)}}l(HP,"installReductionMethods");function up(r){return r.length===1&&Array.isArray(r[0])?r[0]:r}l(up,"arrayArg");function OO(r){let e=r.prototype;e.to=function(t){if(this.device.equals(t))return this;let s=lp.contiguous(this).data.slice(0,this.numel);return ta(s,this.shape,this.dtype,{device:t})},UP(e),KP(e),WP(e),HP(e),e.mm=function(t){return lt(this,t)},e.requires_grad=function(t=!0){return this.requiresGrad_(t)},e.gather=function(t,n){return jd(this,t,n)},e.scatter_add=function(t,n,s){return ra(this,t,n,s)},e.scatter=function(t,n,s){return qd(this,t,n,s)},e.transpose=function(t,n){return gt(this,t,n)},e.slice=function(t,n,s,o){return Kd(this,t,n,s,o)},e.narrow=function(t,n,s){return Un(this,t,n,s)},e.select=function(t,n){return Ut(this,t,n)},e.reshape=function(...t){return Cn(this,up(t))},e.permute=function(...t){return Eu(this,up(t))},e.expand=function(...t){return Lu(this,up(t))},e.repeat=function(...t){return Du(this,up(t))},e.tile=function(...t){return Wd(this,up(t))},e.t=function(){if(this.ndim!==2)throw new Error("t() expects a 2D tensor");return gt(this,0,1)}}l(OO,"installOps");jn();jn();ar();sp();ga();Kt();P0();Kc();xw();$i();_t();_t();function pp(r,e){return{dtype:e?.dtype??r.dtype,device:e?.device??r.device,requiresGrad:e?.requiresGrad??!1}}l(pp,"_likeOpts");function QP(r,e){return fn(r.shape,pp(r,e))}l(QP,"emptyLike");function XP(r,e){return $e(r.shape,pp(r,e))}l(XP,"zerosLike");function e$(r,e){return Ht(r.shape,pp(r,e))}l(e$,"onesLike");function t$(r,e,t){return et(r.shape,e,pp(r,t))}l(t$,"fullLike");function n$(r,e){return Bm(r.shape,pp(r,e))}l(n$,"randnLike");qr();be();var $O={};nt($O,{AdaptiveAvgPool2d:()=>Rp,AvgPool2d:()=>Bp,BCELoss:()=>Lp,BatchNorm1d:()=>ul,BatchNorm2d:()=>Ap,Conv1d:()=>yp,Conv2d:()=>gp,CrossEntropyLoss:()=>Ip,Dropout:()=>pr,ELU:()=>kp,Embedding:()=>Cp,F:()=>PO,Flatten:()=>Vp,GELU:()=>bp,GRU:()=>Pp,GRUCell:()=>ll,GroupNorm:()=>Cg,LSTM:()=>$p,LSTMCell:()=>cl,LayerNorm:()=>Is,LeakyReLU:()=>Sp,Linear:()=>hn,LogSoftmax:()=>Np,MSELoss:()=>Mp,MaxPool2d:()=>Op,Module:()=>ge,ModuleDict:()=>zp,ModuleList:()=>di,MultiheadAttention:()=>mi,NLLLoss:()=>Ep,Parameter:()=>bt,PositionalEncoding:()=>jp,ReLU:()=>_p,Sequential:()=>Fp,SiLU:()=>wp,Sigmoid:()=>Tp,Softmax:()=>vp,Tanh:()=>xp,Transformer:()=>Gp,TransformerDecoder:()=>fl,TransformerDecoderLayer:()=>ba,TransformerEncoder:()=>pl,TransformerEncoderLayer:()=>_a,functional:()=>al,init:()=>Sg});jn();var bt=class extends We{static{l(this,"Parameter")}constructor(e,t=!0){e instanceof We?super(e._impl):super(e),t&&this.requiresGrad_(!0)}get isParameter(){return!0}};var ge=class r{static{l(this,"Module")}_parameters;_buffers;_modules;_training;_detected;constructor(){this._parameters=new Map,this._buffers=new Map,this._modules=new Map,this._training=!0}forward(...e){throw new Error(`${this.constructor.name}.forward() not implemented`)}call(...e){return this.forward(...e)}registerParameter(e,t){if(t!==null&&!(t instanceof bt))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 r))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 n=e?e+".":"";for(let[s,o]of this._parameters)o!==null&&(yield[n+s,o]);if(t)for(let[s,o]of this._modules)o!==null&&(yield*o.namedParameters(n+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[n,s]of this._modules)s!==null&&(yield*s.namedModules(t+n))}stateDict(e=""){this._autoDetect();let t=new Map,n=e?e+".":"";for(let[s,o]of this._parameters)o!==null&&t.set(n+s,o);for(let[s,o]of this._buffers)o!==null&&t.set(n+s,o);for(let[s,o]of this._modules)if(o!==null)for(let[i,a]of o.stateDict(n+s))t.set(i,a);return t}loadStateDict(e){this._autoDetect();for(let[t,n]of this._parameters){let s=t;if(e.has(s)&&n!==null){let o=e.get(s),i=n._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,n]of this._modules)if(n!==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&&n.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,n]of this._parameters){let s=n;n!==null&&typeof s?.to=="function"&&(this._parameters.set(t,new bt(s.to(e),n.requiresGrad)),this[t]=this._parameters.get(t))}for(let[t,n]of this._buffers){let s=n;n!==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 n=this[t];n instanceof bt&&!this._parameters.has(t)?this._parameters.set(t,n):n instanceof r&&!this._modules.has(t)&&this._modules.set(t,n)}}toString(){return this._buildRepr("")}_buildRepr(e){this._autoDetect();let t=this.constructor.name;if(this._modules.size===0)return t+"()";let n=[t+"("];for(let[s,o]of this._modules)o!==null&&n.push(e+"  ("+s+"): "+o._buildRepr(e+"  "));return n.push(e+")"),n.join(`
+`)}};var Sg={};nt(Sg,{_calculateFanInFanOut:()=>ol,constant_:()=>o$,kaiming_normal_:()=>u$,kaiming_uniform_:()=>BO,normal_:()=>fp,ones_:()=>s$,resetLinearParameters:()=>il,uniform_:()=>xg,xavier_normal_:()=>a$,xavier_uniform_:()=>i$,zeros_:()=>r$});function ol(r){let e=r.shape,t=e.length;if(t<2)throw new Error("Fan in/out requires at least 2D tensor");let n=e[1],s=e[0],o=1;for(let i=2;i<t;i++)o*=e[i];return{fanIn:n*o,fanOut:s*o}}l(ol,"_calculateFanInFanOut");function il(r,e){if(BO(r,Math.sqrt(5)),e){let{fanIn:t}=ol(r),n=1/Math.sqrt(t);xg(e,-n,n)}}l(il,"resetLinearParameters");function xg(r,e=0,t=1){let n=r._impl.storage.data;if(!n)return r;let s=t-e;for(let o=0;o<n.length;o++)n[o]=e+Math.random()*s;return r._impl.bumpVersion(),r}l(xg,"uniform_");function fp(r,e=0,t=1){let n=r._impl.storage.data;if(!n)return r;for(let s=0;s<n.length;s+=2){let o=Math.random()||1e-10,i=Math.random(),a=Math.sqrt(-2*Math.log(o)),u=6.283185307179586*i;n[s]=e+t*a*Math.cos(u),s+1<n.length&&(n[s+1]=e+t*a*Math.sin(u))}return r._impl.bumpVersion(),r}l(fp,"normal_");function r$(r){let e=r._impl.storage.data;return e&&e.fill(0),r._impl.bumpVersion(),r}l(r$,"zeros_");function s$(r){let e=r._impl.storage.data;return e&&e.fill(1),r._impl.bumpVersion(),r}l(s$,"ones_");function o$(r,e){let t=r._impl.storage.data;return t&&t.fill(e),r._impl.bumpVersion(),r}l(o$,"constant_");function i$(r,e=1){let{fanIn:t,fanOut:n}=ol(r),s=e*Math.sqrt(2/(t+n)),o=Math.sqrt(3)*s;return xg(r,-o,o)}l(i$,"xavier_uniform_");function a$(r,e=1){let{fanIn:t,fanOut:n}=ol(r),s=e*Math.sqrt(2/(t+n));return fp(r,0,s)}l(a$,"xavier_normal_");function BO(r,e=0,t="fan_in",n="leaky_relu"){let{fanIn:s,fanOut:o}=ol(r),i=t==="fan_in"?s:o,u=RO(n,e)/Math.sqrt(i),p=Math.sqrt(3)*u;return xg(r,-p,p)}l(BO,"kaiming_uniform_");function u$(r,e=0,t="fan_in",n="leaky_relu"){let{fanIn:s,fanOut:o}=ol(r),i=t==="fan_in"?s:o,u=RO(n,e)/Math.sqrt(i);return fp(r,0,u)}l(u$,"kaiming_normal_");function RO(r,e=.01){switch(r){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(RO,"_calculateGain");var al={};nt(al,{elu:()=>q0,gelu:()=>dp,leaky_relu:()=>j0,log_softmax:()=>G0,relu:()=>hp,sigmoid:()=>F0,silu:()=>$0,softmax:()=>V0,tanh:()=>z0});be();_t();function hp(r){return Mu(r)}l(hp,"relu");function dp(r){return Dd(r)}l(dp,"gelu");function $0(r){return Pd(r)}l($0,"silu");function F0(r){return ur(r)}l(F0,"sigmoid");function z0(r){return Ns(r)}l(z0,"tanh");function V0(r,e=-1){return Ro(r,e)}l(V0,"softmax");function G0(r,e=-1){return Hd(r,e)}l(G0,"log_softmax");function j0(r,e=.01){let t=ee(r,et(r.shape,e,{dtype:r.dtype,device:r.device}));return ai(r,t)}l(j0,"leaky_relu");function q0(r,e=1){let t=et(r.shape,0,{dtype:r.dtype,device:r.device}),n=et(r.shape,1,{dtype:r.dtype,device:r.device}),s=et(r.shape,e,{dtype:r.dtype,device:r.device}),o=Iu(r,t),i=ee(s,Pe(Kr(r),n));return qn(o,r,i)}l(q0,"elu");var vg={};nt(vg,{batch_norm:()=>W0,group_norm:()=>K0,layer_norm:()=>U0});be();_t();ga();function U0(r,e,t,n,s=1e-5){if((r instanceof ft||r.isSymbolic)&&t&&n){let m=r.ndim-e.length;return pw(r,t,n,m,s)}let o=r.ndim-e.length,i=[];for(let m=o;m<r.ndim;m++)i.push(m);let a=r;for(let m=i.length-1;m>=0;m--)a=lr(a,i[m],!0);let u=Pe(r,a),c=ee(u,u);for(let m=i.length-1;m>=0;m--)c=lr(c,i[m],!0);let h=et([],s),f=$t(et([],1),ui(ae(c,h))),d=ee(u,f);return t&&(d=ee(d,t)),n&&(d=ae(d,n)),d}l(U0,"layer_norm");function K0(r,e,t,n,s=1e-5){let o=r.shape,i=o[0],a=o[1],u=o.slice(2),p=r.reshape([i,e,a/e,...u]),c=[];for(let w=2;w<p.ndim;w++)c.push(w);let h=p;for(let w=c.length-1;w>=0;w--)h=lr(h,c[w],!0);let f=Pe(p,h),m=ee(f,f);for(let w=c.length-1;w>=0;w--)m=lr(m,c[w],!0);let g=et([],s),y=$t(et([],1),ui(ae(m,g))),_=ee(f,y).reshape(o),b=[1,a,...u.map(()=>1)];return t&&(_=ee(_,t.reshape(b))),n&&(_=ae(_,n.reshape(b))),_}l(K0,"group_norm");var kg=1;function l$(r,e){let t=new Array(r).fill(1);return t[kg]=e,t}l(l$,"channelShape");function MO(r,e){let t=r;for(let n=e.length-1;n>=0;n--)t=lr(t,e[n],!0);return t}l(MO,"reduceMeanOver");function IO(r,e,t){let n=r.data;if(!n)return;let s=e.reshape([r.shape[0]]).toArray();for(let o=0;o<n.length;o++)n[o]=n[o]*(1-t)+s[o]*t}l(IO,"blendRunning");function W0(r,e,t,n,s,o=!0,i=1e-5,a=.1){let u=r instanceof ft||r.isSymbolic;if(!o)return fw(r,n,s,e,t,kg,i);let p=[];for(let y=0;y<r.ndim;y++)y!==kg&&p.push(y);let c=MO(r,p),h=Pe(r,c),f=MO(ee(h,h),p),d=$t(et([],1),ui(ae(f,et([],i)))),m=ee(h,d),g=l$(r.ndim,r.shape[kg]);return n&&(m=ee(m,n.reshape(g))),s&&(m=ae(m,s.reshape(g))),u||(e&&IO(e,c,a),t&&IO(t,f,a)),m}l(W0,"batch_norm");var Y0={};nt(Y0,{linear:()=>H0});be();be();ga();function H0(r,e,t){let n;r instanceof ft||r.isSymbolic?n=gt(e,0,1):n=gt(e,0,1);let s=lt(r,n);return t?ae(s,t):s}l(H0,"linear");var Ag={};nt(Ag,{conv1d:()=>J0,conv2d:()=>Ng});be();be();function Ng(r,e,t,n=[1,1],s=[[0,0],[0,0]],o=[1,1],i=1){let a=Array.isArray(n)?n:[n,n],u=c$(s),p=Array.isArray(o)?o:[o,o],c=hw(r,e,a,u,p,i);if(t){let h=Cn(t,[1,t.shape[0],1,1]);return ae(c,h)}return c}l(Ng,"conv2d");function J0(r,e,t,n=1,s=0,o=1,i=1){let a=Sr(r,2),u=Sr(e,2),p=[1,Array.isArray(n)?n[0]:n],c=[[0,0],p$(s)],h=[1,Array.isArray(o)?o[0]:o],f=Ng(a,u,null,p,c,h,i),d=Cu(f,2);if(t){let m=Cn(t,[1,t.shape[0],1]);return ae(d,m)}return d}l(J0,"conv1d");function c$(r){return typeof r=="number"?[[r,r],[r,r]]:Array.isArray(r)&&typeof r[0]=="number"?r.map(e=>[e,e]):r}l(c$,"_normalizePadding");function p$(r){return typeof r=="number"?[r,r]:r}l(p$,"_normalizePadding1d");var Bg={};nt(Bg,{adaptive_avg_pool2d:()=>Q0,avg_pool2d:()=>Og,max_pool2d:()=>Z0});be();function Z0(r,e,t,n=[[0,0],[0,0]]){let s=Array.isArray(e)?e:[e,e],o=t?Array.isArray(t)?t:[t,t]:s,i=EO(n);return Yd(r,"max",s,o,i)}l(Z0,"max_pool2d");function Og(r,e,t,n=[[0,0],[0,0]]){let s=Array.isArray(e)?e:[e,e],o=t?Array.isArray(t)?t:[t,t]:s,i=EO(n);return Yd(r,"avg",s,o,i)}l(Og,"avg_pool2d");function Q0(r,e){let[t,n]=Array.isArray(e)?e:[e,e],s=r.shape[2],o=r.shape[3],i=Math.floor(s/t),a=Math.floor(o/n);return Og(r,[i,a],[i,a],[[0,0],[0,0]])}l(Q0,"adaptive_avg_pool2d");function EO(r){if(typeof r=="number")return[[r,r],[r,r]];if(Array.isArray(r)&&typeof r[0]=="number"){let e=r.map(t=>[t,t]);return e.length===1?[e[0],e[0]]:e}return r}l(EO,"_normalizePadding");var X0={};nt(X0,{dropout:()=>mp});be();_t();function mp(r,e=.5,t=!0){if(!t||e===0)return r;if(e===1)return et(r.shape,0,{dtype:r.dtype,device:r.device});let n=f$(r.shape,1-e,r.dtype,r.device),s=et(r.shape,1/(1-e),{dtype:r.dtype,device:r.device});return ee(ee(r,n),s)}l(mp,"dropout");function f$(r,e,t,n){let s=fn(r,{dtype:t,device:n}),o=s._impl.storage.data;if(o)for(let i=0;i<o.length;i++)o[i]=Math.random()<e?1:0;return s}l(f$,"_bernoulliMask");var Mg={};nt(Mg,{binary_cross_entropy:()=>nT,cross_entropy:()=>tT,mse_loss:()=>eT,nll_loss:()=>Rg});be();_t();be();jr();function eT(r,e,t="mean"){let n=Pe(r,e),s=ee(n,n);return LO(s,t)}l(eT,"mse_loss");function h$(r,e){let t=d$(r,e),n=Pe(r,t),s=Kr(n),o=m$(s,e),i=Zs(o);return Pe(n,i)}l(h$,"_logSoftmaxAutograd");function d$(r,e){let t=e<0?r.ndim+e:e;return Qs(r,t,!0)}l(d$,"_dimMax");function m$(r,e){let t=e<0?r.ndim+e:e;return ht(r,t,!0)}l(m$,"_dimSum");function Rg(r,e,t="mean",n=null){let s=r.ndim-1,o=r.shape[s],i=Gd(e,o),a=ht(ee(r,i),s),u=a,p=null;if(n!==null){let h=Ut(i,s,n),f=ae(Ur(h),1);u=ee(a,f),p=ht(f)}let c=Ur(ht(u));if(t==="sum")return c;if(t==="mean")return p!==null?$t(c,p):$t(c,cn(e.shape));throw new Error(`nll_loss: unknown reduction '${t}'`)}l(Rg,"nll_loss");function tT(r,e,t="mean",n=null){let s=h$(r,-1);return Rg(s,e,t,n)}l(tT,"cross_entropy");function nT(r,e,t="mean"){let n=et(r.shape,1e-7,{dtype:r.dtype,device:r.device}),s=et(r.shape,1,{dtype:r.dtype,device:r.device}),o=Zs(ae(r,n)),i=Zs(ae(Pe(s,r),n)),a=Ur(ae(ee(e,o),ee(Pe(s,e),i)));return LO(a,t)}l(nT,"binary_cross_entropy");function LO(r,e){if(e==="mean")return lr(r);if(e==="sum")return ht(r);if(e==="none")return r;throw new Error(`Unknown reduction: ${e}`)}l(LO,"_reduce");var sT={};nt(sT,{embedding:()=>rT});be();function rT(r,e){let t=r.shape[r.shape.length-1],n=e.shape.reduce((o,i)=>o*i,1);return li(r,0,e.reshape([n])).reshape([...e.shape,t])}l(rT,"embedding");var iT={};nt(iT,{scaled_dot_product_attention:()=>oT});be();be();ga();ya();_t();function y$(r,e){if(r instanceof ft||r.isSymbolic)return Ro(r,e);let t=e<0?r.ndim+e:e,n=Qs(r,t,!0).detach(),s=Pe(r,n),o=Kr(s),i=ht(o,t,!0);return $t(o,i)}l(y$,"_softmax");function _$(r){let e=r.ndim;return r instanceof ft||r.isSymbolic?gt(r,e-2,e-1):gt(r,e-2,e-1)}l(_$,"_transposeLastTwo");function b$(r,e){let t=fn([r,e]),n=t._impl.storage.data,s=e-r;for(let o=0;o<r;o++)for(let i=0;i<e;i++)n[o*e+i]=i<=o+s?0:-1/0;return t}l(b$,"_generateCausalMask");function oT(r,e,t,n=null,s=0,o=!1,i=!1){let a=r.shape[r.ndim-1],u=r.shape[r.ndim-2],p=e.shape[e.ndim-2],c=ss();if(c&&!n&&!(s>0&&i)&&r instanceof ft&&e instanceof ft&&t instanceof ft&&r.ndim===4)return c.recordOp("scaled_dot_product_attention",[r,e,t],{scale:1/Math.sqrt(a),causal:o});let f=et([],1/Math.sqrt(a)),d=_$(e),m=lt(r,d);m=ee(m,f),o&&(m=ae(m,b$(u,p))),n&&(m=ae(m,n));let g=y$(m,-1);return s>0&&i&&(g=mp(g,s,!0)),lt(g,t)}l(oT,"scaled_dot_product_attention");_t();var hn=class extends ge{static{l(this,"Linear")}inFeatures;outFeatures;weight;bias;constructor(e,t,n=!0){super(),this.inFeatures=e,this.outFeatures=t,this.weight=new bt(fn([t,e])),this.bias=n?new bt($e([t])):null,this._resetParameters()}_resetParameters(){il(this.weight,this.bias)}forward(e){return H0(e,this.weight,this.bias)}};_t();var gp=class extends ge{static{l(this,"Conv2d")}inChannels;outChannels;kernelSize;stride;padding;dilation;groups;weight;bias;constructor(e,t,n,s={}){super(),this.inChannels=e,this.outChannels=t,this.kernelSize=Array.isArray(n)?n:[n,n],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 bt(fn([t,e/this.groups,o,i])),this.bias=s.bias!==!1?new bt($e([t])):null,this._resetParameters()}_resetParameters(){il(this.weight,this.bias)}forward(e){return Ng(e,this.weight,this.bias,this.stride,this.padding,this.dilation,this.groups)}},yp=class extends ge{static{l(this,"Conv1d")}inChannels;outChannels;kernelSize;stride;padding;dilation;groups;weight;bias;constructor(e,t,n,s={}){super(),this.inChannels=e,this.outChannels=t,this.kernelSize=Array.isArray(n)?n[0]:n,this.stride=s.stride??1,this.padding=s.padding??0,this.dilation=s.dilation??1,this.groups=s.groups??1,this.weight=new bt(fn([t,e/this.groups,this.kernelSize])),this.bias=s.bias!==!1?new bt($e([t])):null,il(this.weight,this.bias)}forward(e){return J0(e,this.weight,this.bias,this.stride,this.padding,this.dilation,this.groups)}};var _p=class extends ge{static{l(this,"ReLU")}forward(e){return hp(e)}},bp=class extends ge{static{l(this,"GELU")}forward(e){return dp(e)}},wp=class extends ge{static{l(this,"SiLU")}forward(e){return $0(e)}},Tp=class extends ge{static{l(this,"Sigmoid")}forward(e){return F0(e)}},xp=class extends ge{static{l(this,"Tanh")}forward(e){return z0(e)}},Sp=class extends ge{static{l(this,"LeakyReLU")}negativeSlope;constructor(e=.01){super(),this.negativeSlope=e}forward(e){return j0(e,this.negativeSlope)}},kp=class extends ge{static{l(this,"ELU")}alpha;constructor(e=1){super(),this.alpha=e}forward(e){return q0(e,this.alpha)}},vp=class extends ge{static{l(this,"Softmax")}dim;constructor(e=-1){super(),this.dim=e}forward(e){return V0(e,this.dim)}},Np=class extends ge{static{l(this,"LogSoftmax")}dim;constructor(e=-1){super(),this.dim=e}forward(e){return G0(e,this.dim)}};_t();var Is=class extends ge{static{l(this,"LayerNorm")}normalizedShape;eps;weight;bias;constructor(e,t=1e-5,n=!0){super(),this.normalizedShape=Array.isArray(e)?e:[e],this.eps=t,this.weight=n?new bt(Ht(this.normalizedShape)):null,this.bias=n?new bt($e(this.normalizedShape)):null}forward(e){return U0(e,this.normalizedShape,this.weight,this.bias,this.eps)}},Cg=class extends ge{static{l(this,"GroupNorm")}numGroups;numChannels;eps;weight;bias;constructor(e,t,n=1e-5,s=!0){super(),this.numGroups=e,this.numChannels=t,this.eps=n,this.weight=s?new bt(Ht([t])):null,this.bias=s?new bt($e([t])):null}forward(e){return K0(e,this.numGroups,this.weight,this.bias,this.eps)}},ul=class extends ge{static{l(this,"BatchNorm1d")}numFeatures;eps;weight;bias;runningMean;runningVar;constructor(e,t=1e-5,n=!0){super(),this.numFeatures=e,this.eps=t,this.weight=n?new bt(Ht([e])):null,this.bias=n?new bt($e([e])):null,this.runningMean=$e([e]),this.runningVar=Ht([e]),this.registerBuffer("runningMean",this.runningMean),this.registerBuffer("runningVar",this.runningVar)}forward(e){return W0(e,this.runningMean,this.runningVar,this.weight,this.bias,this.training,this.eps)}},Ap=class extends ul{static{l(this,"BatchNorm2d")}constructor(e,t=1e-5,n=!0){super(e,t,n)}};var Op=class extends ge{static{l(this,"MaxPool2d")}kernelSize;stride;padding;constructor(e,t,n=0){super(),this.kernelSize=Array.isArray(e)?e:[e,e],this.stride=t?Array.isArray(t)?t:[t,t]:this.kernelSize,this.padding=n}forward(e){return Z0(e,this.kernelSize,this.stride,this.padding)}},Bp=class extends ge{static{l(this,"AvgPool2d")}kernelSize;stride;padding;constructor(e,t,n=0){super(),this.kernelSize=Array.isArray(e)?e:[e,e],this.stride=t?Array.isArray(t)?t:[t,t]:this.kernelSize,this.padding=n}forward(e){return Og(e,this.kernelSize,this.stride,this.padding)}},Rp=class extends ge{static{l(this,"AdaptiveAvgPool2d")}outputSize;constructor(e){super(),this.outputSize=Array.isArray(e)?e:[e,e]}forward(e){return Q0(e,this.outputSize)}};var pr=class extends ge{static{l(this,"Dropout")}p;constructor(e=.5){super(),this.p=e}forward(e){return mp(e,this.p,this.training)}};var Mp=class extends ge{static{l(this,"MSELoss")}reduction;constructor(e="mean"){super(),this.reduction=e}forward(e,t){return eT(e,t,this.reduction)}},Ip=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 tT(e,t,this.reduction,this.ignoreIndex===-100?null:this.ignoreIndex)}},Ep=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 Rg(e,t,this.reduction,this.ignoreIndex===-100?null:this.ignoreIndex)}},Lp=class extends ge{static{l(this,"BCELoss")}reduction;constructor(e="mean"){super(),this.reduction=e}forward(e,t){return nT(e,t,this.reduction)}};_t();var Cp=class extends ge{static{l(this,"Embedding")}numEmbeddings;embeddingDim;weight;constructor(e,t){super(),this.numEmbeddings=e,this.embeddingDim=t,this.weight=new bt(fn([e,t])),fp(this.weight)}forward(e){return rT(this.weight,e)}};_t();be();be();ya();be();be();function w$(r){if(Array.isArray(r[0])){let e=r[0],t=new Array(e.length);for(let n=0;n<t.length;n++)t[n]=Wr(r.map(s=>s[n]),0);return t}return Wr(r,0)}l(w$,"stackSteps");function Dp(r,e,t){let n=Array.isArray(e),s=Array.isArray(t),o=s?t:[t],i=n?e:[e],a=ss();if(!a){let f=e,d=[],m=o[0].shape[0];for(let g=0;g<m;g++){let y=s?o.map(w=>Ut(w,0,g)):Ut(o[0],0,g),[_,b]=r(f,y);f=_,d.push(b)}return[f,w$(d)]}let u=!1,p=l((f,d)=>{let[m,g]=r(n?f:f[0],s?d:d[0]);return u=Array.isArray(g),[n?m:[m],u?g:[g]]},"step"),[c,h]=a.scan(o,i,p);return[n?c:c[0],u?h:h[0]]}l(Dp,"scan");ya();Kn();ar();var ll=class extends ge{static{l(this,"GRUCell")}inputSize;hiddenSize;x2h;h2h;constructor(e,t,n=!0){super(),this.inputSize=e,this.hiddenSize=t,this.x2h=new hn(e,3*t,n),this.h2h=new hn(t,3*t,n)}forward(e,t=null){let n=t!==null?t:$e([e.shape[0],this.hiddenSize],{device:e.device}),s=this.x2h.forward(e),o=this.h2h.forward(n),[i,a,u]=Bo(s,this.hiddenSize,-1),[p,c,h]=Bo(o,this.hiddenSize,-1),f=ur(ae(i,p)),d=ur(ae(a,c)),m=Ns(ae(u,ee(f,h)));return ae(m,ee(d,Pe(n,m)))}},Pp=class extends ge{static{l(this,"GRU")}inputSize;hiddenSize;numLayers;batchFirst;cells;constructor(e,t,n=1,s=!1,o=!0){super(),this.inputSize=e,this.hiddenSize=t,this.numLayers=n,this.batchFirst=s,this.cells=[];for(let i=0;i<n;i++){let a=new ll(i===0?e:t,t,o);this.cells.push(a),this.registerModule("cell_"+i,a)}}forward(e,t=null){let n=d0();if(n&&e.device.type==="gpu"&&!ss()){let c=this.batchFirst?e.transpose(0,1):e,h={inputSize:this.inputSize,hiddenSize:this.hiddenSize,seqLen:c.shape[0],batch:c.shape[1]},[f,d]=n(c,this.cells,h,t);return[this.batchFirst?f.transpose(0,1):f,d]}let s=Qc();if(s&&e.device.type==="webgpu"&&!ss()){let c=this.batchFirst?e.transpose(0,1):e,h={kind:"gru",inputSize:this.inputSize,hiddenSize:this.hiddenSize,seqLen:c.shape[0],batch:c.shape[1]},f=s(c,this.cells,h,t,null);if(f){let[d,m]=f;return[this.batchFirst?d.transpose(0,1):d,m]}}let o=this.batchFirst?e.transpose(0,1):e,i=o.shape[1],a=o,u=[];for(let c=0;c<this.numLayers;c++){let h=t!==null?Ut(t,0,c):$e([i,this.hiddenSize],{device:o.device}),f=this.cells[c],[d,m]=Dp((g,y)=>{let _=f.forward(y,g);return[_,_]},h,a);u.push(d),a=m}let p=a;return this.batchFirst&&(p=p.transpose(0,1)),[p,Wr(u,0)]}};_t();be();be();ya();Kn();ar();var cl=class extends ge{static{l(this,"LSTMCell")}inputSize;hiddenSize;x2h;h2h;constructor(e,t,n=!0){super(),this.inputSize=e,this.hiddenSize=t,this.x2h=new hn(e,4*t,n),this.h2h=new hn(t,4*t,n)}forward(e,t=null){let n=t!==null?t[0]:$e([e.shape[0],this.hiddenSize],{device:e.device}),s=t!==null?t[1]:$e([e.shape[0],this.hiddenSize],{device:e.device}),o=ae(this.x2h.forward(e),this.h2h.forward(n)),[i,a,u,p]=Bo(o,this.hiddenSize,-1),c=ur(i),h=ur(a),f=Ns(u),d=ur(p),m=ae(ee(h,s),ee(c,f));return[ee(d,Ns(m)),m]}},$p=class extends ge{static{l(this,"LSTM")}inputSize;hiddenSize;numLayers;batchFirst;cells;constructor(e,t,n=1,s=!1,o=!0){super(),this.inputSize=e,this.hiddenSize=t,this.numLayers=n,this.batchFirst=s,this.cells=[];for(let i=0;i<n;i++){let a=new cl(i===0?e:t,t,o);this.cells.push(a),this.registerModule("cell_"+i,a)}}forward(e,t=null){let n=h0();if(n&&e.device.type==="gpu"&&!ss()){let h=this.batchFirst?e.transpose(0,1):e,f={inputSize:this.inputSize,hiddenSize:this.hiddenSize,seqLen:h.shape[0],batch:h.shape[1]},[d,m,g]=n(h,this.cells,f,t?t[0]:null,t?t[1]:null);return[this.batchFirst?d.transpose(0,1):d,[m,g]]}let s=Qc();if(s&&e.device.type==="webgpu"&&!ss()){let h=this.batchFirst?e.transpose(0,1):e,f={inputSize:this.inputSize,hiddenSize:this.hiddenSize,seqLen:h.shape[0],batch:h.shape[1]},d=s(h,this.cells,f,t?t[0]:null,t?t[1]:null);if(d){let[m,g,y]=d;return[this.batchFirst?m.transpose(0,1):m,[g,y]]}}let o=this.batchFirst?e.transpose(0,1):e,i=o.shape[1],a=o,u=[],p=[];for(let h=0;h<this.numLayers;h++){let f=t!==null?Ut(t[0],0,h):$e([i,this.hiddenSize],{device:o.device}),d=t!==null?Ut(t[1],0,h):$e([i,this.hiddenSize],{device:o.device}),m=this.cells[h],[[g,y],_]=Dp((b,w)=>{let[T,S]=m.forward(w,b);return[[T,S],T]},[f,d],a);u.push(g),p.push(y),a=_}let c=a;return this.batchFirst&&(c=c.transpose(0,1)),[c,[Wr(u,0),Wr(p,0)]]}};var Fp=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 n=0;n<this._length;n++)t=this[String(n)].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}},di=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()")}},zp=class extends ge{static{l(this,"ModuleDict")}_dict;constructor(e){if(super(),this._dict=new Map,e)for(let[t,n]of Object.entries(e))this._dict.set(t,n),this.registerModule(t,n)}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 Vp=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,n=t.length,s=this.startDim<0?n+this.startDim:this.startDim,o=this.endDim<0?n+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<n;u++)a.push(t[u]);return Cn(e,a)}};be();_t();_t();function DO(r){return r==="gelu"?dp:hp}l(DO,"_getActivation");var mi=class extends ge{static{l(this,"MultiheadAttention")}embedDim;numHeads;headDim;batchFirst;dropout;qProj;kProj;vProj;outProj;constructor(e,t,n=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=n,this.qProj=new hn(e,e,s),this.kProj=new hn(o??e,e,s),this.vProj=new hn(i??e,e,s),this.outProj=new hn(e,e,s)}forward(e,t,n,s=null,o=null,i=!1){this.batchFirst||(e=e.transpose(0,1),t=t.transpose(0,1),n=n.transpose(0,1));let a=e.shape[0],u=e.shape[1],p=t.shape[1],c=this.qProj.forward(e),h=this.kProj.forward(t),f=this.vProj.forward(n);if(c=c.reshape([a,u,this.numHeads,this.headDim]).permute([0,2,1,3]),h=h.reshape([a,p,this.numHeads,this.headDim]).permute([0,2,1,3]),f=f.reshape([a,p,this.numHeads,this.headDim]).permute([0,2,1,3]),o){let g=et(o.shape,-1/0),y=et(o.shape,0),_=qn(o,g,y);_=_.unsqueeze(1).unsqueeze(2),s=s?ae(s,_):_}let d=oT(c,h,f,s,this.dropout,i,this.training);d=d.permute([0,2,1,3]).reshape([a,u,this.embedDim]);let m=this.outProj.forward(d);return this.batchFirst||(m=m.transpose(0,1)),m}},_a=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,n=2048,s=.1,o="relu",i=1e-5,a=!0,u=!1){super(),this.selfAttn=new mi(e,t,s,!0,null,null,a),this.linear1=new hn(e,n),this.linear2=new hn(n,e),this.norm1=new Is(e,i),this.norm2=new Is(e,i),this.dropout1=new pr(s),this.dropout2=new pr(s),this.dropoutFFN=new pr(s),this._activation=DO(o),this._activationName=o,this.normFirst=u,this._dModel=e,this._nhead=t,this._dimFeedforward=n,this._dropout=s,this._layerNormEps=i,this._batchFirst=a}forward(e,t=null,n=null,s=!1){return this.normFirst?this._forwardPreNorm(e,t,n,s):this._forwardPostNorm(e,t,n,s)}_forwardPostNorm(e,t,n,s){let o=this.selfAttn.forward(e,e,e,t,n,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,n,s){let o=this.norm1.forward(e),i=this.selfAttn.forward(o,o,o,t,n,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}},ba=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,n=2048,s=.1,o="relu",i=1e-5,a=!0,u=!1){super(),this.selfAttn=new mi(e,t,s,!0,null,null,a),this.crossAttn=new mi(e,t,s,!0,null,null,a),this.linear1=new hn(e,n),this.linear2=new hn(n,e),this.norm1=new Is(e,i),this.norm2=new Is(e,i),this.norm3=new Is(e,i),this.dropout1=new pr(s),this.dropout2=new pr(s),this.dropout3=new pr(s),this.dropoutFFN=new pr(s),this._activation=DO(o),this._activationName=o,this.normFirst=u,this._dModel=e,this._nhead=t,this._dimFeedforward=n,this._dropout=s,this._layerNormEps=i,this._batchFirst=a}forward(e,t,n=null,s=null,o=null,i=null,a=!1){return this.normFirst?this._forwardPreNorm(e,t,n,s,o,i,a):this._forwardPostNorm(e,t,n,s,o,i,a)}_forwardPostNorm(e,t,n,s,o,i,a){let u=this.selfAttn.forward(e,e,e,n,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,n,s,o,i,a){let u=this.norm1.forward(e),p=this.selfAttn.forward(u,u,u,n,o,a);p=ae(e,this.dropout1.forward(p));let c=this.norm2.forward(p),h=this.crossAttn.forward(c,t,t,s,i);p=ae(p,this.dropout2.forward(h));let f=this._activation(this.linear1.forward(this.norm3.forward(p)));return f=this.linear2.forward(this.dropoutFFN.forward(f)),p=ae(p,this.dropout3.forward(f)),p}};function T$(r){return new _a(r._dModel,r._nhead,r._dimFeedforward,r._dropout,r._activationName,r._layerNormEps,r._batchFirst,r.normFirst)}l(T$,"_cloneEncoderLayer");function x$(r){return new ba(r._dModel,r._nhead,r._dimFeedforward,r._dropout,r._activationName,r._layerNormEps,r._batchFirst,r.normFirst)}l(x$,"_cloneDecoderLayer");var pl=class extends ge{static{l(this,"TransformerEncoder")}layers;norm;constructor(e,t,n=null){super(),this.layers=new di(Array.from({length:t},()=>T$(e))),this.norm=n}forward(e,t=null,n=null,s=!1){let o=e;for(let i of this.layers)o=i.forward(o,t,n,s);return this.norm&&(o=this.norm.forward(o)),o}},fl=class extends ge{static{l(this,"TransformerDecoder")}layers;norm;constructor(e,t,n=null){super(),this.layers=new di(Array.from({length:t},()=>x$(e))),this.norm=n}forward(e,t,n=null,s=null,o=null,i=null,a=!1){let u=e;for(let p of this.layers)u=p.forward(u,t,n,s,o,i,a);return this.norm&&(u=this.norm.forward(u)),u}},Gp=class extends ge{static{l(this,"Transformer")}encoder;decoder;dModel;constructor({dModel:e=512,nhead:t=8,numEncoderLayers:n=6,numDecoderLayers:s=6,dimFeedforward:o=2048,dropout:i=.1,activation:a="relu",batchFirst:u=!0,normFirst:p=!1,layerNormEps:c=1e-5}={}){super();let h=new _a(e,t,o,i,a,c,u,p),f=new ba(e,t,o,i,a,c,u,p);this.encoder=new pl(h,n),this.decoder=new fl(f,s),this.dModel=e}forward(e,t,n=null,s=null,o=null,i=null,a=null,u=null){let p=this.encoder.forward(e,n,i);return this.decoder.forward(t,p,s,o,a,u)}static generateSquareSubsequentMask(e){let t=fn([e,e]),n=t._impl.storage.data;for(let s=0;s<e;s++)for(let o=0;o<e;o++)n[s*e+o]=o<=s?0:-1/0;return t}};be();_t();var jp=class extends ge{static{l(this,"PositionalEncoding")}dropoutLayer;pe;constructor(e,t=5e3,n=.1){super(),this.dropoutLayer=new pr(n);let s=$e([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],n=this.pe.narrow(1,0,t);return this.dropoutLayer.forward(ae(e,n))}};var PO={...al,...vg,...Y0,...Ag,...Bg,...X0,...Mg,...sT,...iT};var FO={};nt(FO,{BatchSampler:()=>Sa,DataLoader:()=>Kp,Dataset:()=>wa,MapDataset:()=>Up,RandomSampler:()=>xa,Sampler:()=>gi,SequentialSampler:()=>Ta,TensorDataset:()=>qp,defaultCollate:()=>ka});be();var wa=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)}},qp=class extends wa{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 n=1;n<e.length;n++)if(e[n].shape[0]!==t)throw new Error(`Size mismatch at dim 0: tensor 0 has ${t}, tensor ${n} has ${e[n].shape[0]}`);this._tensors=e,this._length=t}get length(){return this._length}get(e){let t=new Array(this._tensors.length);for(let n=0;n<this._tensors.length;n++)t[n]=Ut(this._tensors[n],0,e);return t}},Up=class extends wa{static{l(this,"MapDataset")}_dataset;_transform;constructor(e,t){super(),this._dataset=e,this._transform=t}get length(){return this._dataset.length}get(e){return this._transform(this._dataset.get(e))}};var gi=class{static{l(this,"Sampler")}*[Symbol.iterator](){throw new Error("Subclass must implement [Symbol.iterator]()")}},Ta=class extends gi{static{l(this,"SequentialSampler")}_dataSource;constructor(e){super(),this._dataSource=e}*[Symbol.iterator](){let e=this._dataSource.length;for(let t=0;t<e;t++)yield t}},xa=class extends gi{static{l(this,"RandomSampler")}_dataSource;constructor(e){super(),this._dataSource=e}*[Symbol.iterator](){let e=this._dataSource.length,t=new Int32Array(e);for(let n=0;n<e;n++)t[n]=n;for(let n=e-1;n>0;n--){let s=Math.floor(Math.random()*(n+1)),o=t[n];t[n]=t[s],t[s]=o}for(let n=0;n<e;n++)yield t[n]}},Sa=class extends gi{static{l(this,"BatchSampler")}_dropLast;_sampler;_batchSize;constructor(e,t,n=!1){super(),this._sampler=e,this._batchSize=t,this._dropLast=n}*[Symbol.iterator](){let e=[];for(let t of this._sampler)e.push(t),e.length===this._batchSize&&(yield e,e=[]);e.length>0&&!this._dropLast&&(yield e)}};jn();qr();ts();function S$(r,e=0){let t=r.length,n=r[0].shape,s=r[0].dtype,o=[...n];o.splice(e,0,t);let i=n.reduce((c,h)=>c*h,1),a=Dt(s),u=new a(t*i),p=u;for(let c=0;c<t;c++){let h=r[c],f=h.data,d=h._impl?h._impl.storageOffset:0;if(h.isContiguous)p.set(f.subarray(d,d+i),c*i);else{let m=h.shape,g=h.strides,y=h._impl.storage.data,_=m.length,b=new Int32Array(_),w=d;for(let T=0;T<i;T++){p[c*i+T]=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 ta(u,o,s)}l(S$,"stackTensors");function ka(r){let e=r[0];if(e instanceof We)return S$(r,0);if(typeof e=="number")return Pt(r);if(Array.isArray(e)){let t=new Array(e.length);for(let n=0;n<e.length;n++){let s=new Array(r.length);for(let o=0;o<r.length;o++)s[o]=r[o][n];t[n]=ka(s)}return t}if(e!==null&&typeof e=="object"&&e.constructor===Object){let t=Object.keys(e),n={};for(let s=0;s<t.length;s++){let o=t[s],i=new Array(r.length);for(let a=0;a<r.length;a++)i[a]=r[a][o];n[o]=ka(i)}return n}throw new Error(`defaultCollate: unsupported element type ${typeof e}`)}l(ka,"defaultCollate");var Kp=class{static{l(this,"DataLoader")}_dataset;_collate;_batchSampler;_batchSize;_dropLast;constructor(e,t={}){if(this._dataset=e,this._collate=t.collate??(n=>ka(n)),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 n=t.batchSize??1,s=t.dropLast??!1;this._batchSize=n,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 xa(e):new Ta(e);this._batchSampler=new Sa(o,n,s)}}get dataset(){return this._dataset}get length(){let e=this._dataset.length;if(this._batchSize==null){let t=0;for(let n 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 n=0;n<e.length;n++)t[n]=this._dataset.get(e[n]);yield this._collate(t)}}};var qO={};nt(qO,{DEFAULT_SPECIALS:()=>Vg,TOKENIZER_FORMAT:()=>ml,TOKENIZER_VERSION:()=>gl,Tokenizer:()=>Wp,Vocab:()=>yi});var yi=class r{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 n=this._tokenToId.get(e);return n===void 0?t:n}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 r,n=new Set;for(let s of e){if(typeof s!="string")throw new Error("mlfw tokenizer: vocab entries must be strings");if(n.has(s))throw new Error(`mlfw tokenizer: duplicate vocab token '${s}'`);n.add(s),t.add(s)}return t}};var Dg=class r{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 r({lowercase:e.lowercase??!1})}};var Pg=class r{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 r({lowercase:e.lowercase??!1})}};function $g(r,e){return r+""+e}l($g,"pairKey");var Fg=class r{static{l(this,"BpeStrategy")}_numMerges;_lowercase;_eow;_ranks;_encodeCache;constructor({numMerges:e=1e3,lowercase:t=!1,endOfWord:n="</w>"}={}){this._numMerges=e,this._lowercase=t,this._eow=n,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 n=new Map;for(let d of e)for(let m of this._pretokenize(d))n.set(m,(n.get(m)||0)+1);let s=new Map;for(let d of n.keys())s.set(d,this._baseSymbols(d));let o=new Map,i=new Map,a=[],u=l((d,m)=>d[0]>m[0]||d[0]===m[0]&&d[1]<m[1],"higher"),p=l((d,m)=>{a.push([d,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 d=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 T=a[_];a[_]=a[g],a[g]=T,g=_}}return d},"popHeap"),h=l(d=>{let m=s.get(d),g=n.get(d);for(let y=0;y+1<m.length;y++){let _=$g(m[y],m[y+1]),b=(o.get(_)||0)+g;o.set(_,b),p(b,_);let w=i.get(_);w||(w=new Set,i.set(_,w)),w.add(d)}},"addWordPairs"),f=l(d=>{let m=s.get(d),g=n.get(d);for(let y=0;y+1<m.length;y++){let _=$g(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(d)}},"removeWordPairs");for(let d of n.keys())h(d);this._ranks=new Map;for(let d=0;d<this._numMerges;d++){let m=null,g=0;for(;a.length>0;){let[S,x]=c(),k=o.get(x);if(k===S&&k>0){m=x,g=k;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,d);let T=i.get(m);if(!T||T.size===0)break;for(let S of[...T]){f(S);let x=s.get(S),k=[];for(let A=0;A<x.length;A++)A+1<x.length&&x[A]===_&&x[A+1]===b?(k.push(w),A++):k.push(x[A]);s.set(S,k),h(S)}}}_encodeWord(e){let t=this._encodeCache.get(e);if(t!==void 0)return t;let n=this._baseSymbols(e);for(;n.length>1;){let s=1/0,o=-1;for(let i=0;i+1<n.length;i++){let a=this._ranks.get($g(n[i],n[i+1]));a!==void 0&&a<s&&(s=a,o=i)}if(o<0)break;n=n.slice(0,o).concat(n[o]+n[o+1],n.slice(o+2))}return this._encodeCache.set(e,n),n}segment(e){let t=[];for(let n of this._pretokenize(e))for(let s of this._encodeWord(n))t.push(s);return t}detokenize(e){let t="";for(let n of e)t+=n;return t.split(this._eow).join(" ").trim()}toJSON(){let e=[...this._ranks.entries()].map(([t])=>{let n=t.indexOf("");return[t.slice(0,n),t.slice(n+1)]});return{numMerges:this._numMerges,lowercase:this._lowercase,endOfWord:this._eow,merges:e}}static fromJSON(e={}){if(!Array.isArray(e.merges))throw new Error("mlfw tokenizer: bpe strategy merges must be an array");let t=new r({numMerges:e.numMerges??e.merges.length,lowercase:e.lowercase??!1,endOfWord:e.endOfWord??"</w>"});t._ranks=new Map;for(let n=0;n<e.merges.length;n++){let s=e.merges[n];if(!Array.isArray(s)||s.length!==2||s.some(o=>typeof o!="string"))throw new Error("mlfw tokenizer: bpe merges must be string pairs");t._ranks.set($g(s[0],s[1]),n)}return t}};qr();var fr=new Map,aT=new Set,dn={readFile(r){if(!fr.has(r))throw new Error("mlfw: file not found in browser memfs: "+r);return fr.get(r)},readBinary(r){if(!fr.has(r))throw new Error("mlfw: file not found in browser memfs: "+r);return fr.get(r)},writeFile(r,e){fr.set(r,e)},writeBinary(r,e){fr.set(r,e)},appendFile(r,e){fr.set(r,(fr.get(r)||"")+e)},exists(r){return fr.has(r)||aT.has(r)},mkdir(r){aT.add(r)},rename(r,e){if(!fr.has(r))throw new Error("mlfw: file not found in browser memfs: "+r);fr.set(e,fr.get(r)),fr.delete(r)},readdir(r){let e=r.endsWith("/")?r:r+"/",t=new Set;for(let n of[...fr.keys(),...aT])if(n.startsWith(e)){let s=n.slice(e.length).split("/")[0];s&&t.add(s)}return[...t]},remove(r){fr.delete(r)}};var ml="mlfw-tokenizer",gl=1,Vg=Object.freeze({pad:"<pad>",unk:"<unk>",bos:"<bos>",eos:"<eos>"}),zO=`${ml}-v${gl}`,zg={word:Dg,char:Pg,bpe:Fg},dl=Object.freeze(["pad","unk","bos","eos"]),Wp=class r{static{l(this,"Tokenizer")}_mode;_specials;_maxVocab;_strategy;_vocab;constructor(e={}){let t=e.mode??"word",n=zg[t];if(!n)throw new Error(`Unknown tokenizer mode '${t}'. Available: ${Object.keys(zg).join(", ")}`);this._mode=t,this._specials=VO(e.specialTokens),GO(this._specials),this._maxVocab=e.vocabSize??null,jO(this._maxVocab,this._specials),this._strategy=new n(e),this._vocab=null}get mode(){return this._mode}get vocabSize(){return this._fitVocab().size}get padId(){return this._specialId("pad")}get unkId(){return this._specialId("unk")}get bosId(){return this._specialId("bos")}get eosId(){return this._specialId("eos")}_ensureFit(){if(!this._vocab)throw new Error("Tokenizer must be fit() on a corpus before use")}_fitVocab(){return this._ensureFit(),this._vocab}_specialId(e){return this._fitVocab().getId(this._specials[e])}fit(e){let t=Array.isArray(e)?e:[e];this._strategy.fit(t,{vocabSize:this._maxVocab});let n=new yi(Object.values(this._specials));if(this._maxVocab==null)for(let s of t)for(let o of this._strategy.segment(s))n.add(o);else{let s=new Map;for(let a of t)for(let u of this._strategy.segment(a))s.set(u,(s.get(u)||0)+1);let o=[...s.entries()].sort((a,u)=>u[1]-a[1]),i=this._maxVocab-n.size;for(let a=0;a<o.length&&a<i;a++)n.add(o[a][0])}return this._vocab=n,this}toJSON(){let e=this._fitVocab(),t=this._strategy.toJSON();return{format:ml,version:gl,mode:this._mode,config:{vocabSize:this._maxVocab,...t},specialTokens:{...this._specials},vocab:e.tokens(),strategy:t}}save(e){if(typeof e!="string")throw new Error("mlfw tokenizer: save(path) requires a file path string");let t=e+".tmp";dn.writeFile(t,v$(this.toJSON())),dn.rename(t,e)}static load(e){if(typeof e!="string")throw new Error("mlfw tokenizer: load(path) requires a file path string");let t=dn.readFile(e),n=typeof t=="string"?t:new TextDecoder().decode(t);return r.fromJSON(N$(n))}static fromJSON(e){k$(e);let t=zg[e.mode],n=new r({mode:e.mode,vocabSize:e.config.vocabSize,specialTokens:e.specialTokens});n._strategy=t.fromJSON(e.strategy),n._vocab=yi.fromTokens(e.vocab);for(let s of Object.values(n._specials))if(!n._vocab.has(s))throw new Error(`mlfw tokenizer: special token '${s}' is missing from vocab`);return n}encode(e,t={}){let n=this._fitVocab(),s=[];t.addBos&&s.push(this.bosId);let o=this.unkId;for(let i of this._strategy.segment(e))s.push(n.getId(i,o));return t.addEos&&s.push(this.eosId),s}decode(e,t={}){let n=this._fitVocab(),s=t.skipSpecial??!0,o=new Set(Object.values(this._specials).map(a=>n.getId(a))),i=[];for(let a of e){if(s&&o.has(a))continue;let u=n.getToken(a);u!==void 0&&i.push(u)}return this._strategy.detokenize(i)}encodeBatch(e,t={}){this._ensureFit();let s=(Array.isArray(e)?e:[e]).map(p=>this.encode(p,t)),o=t.maxLen??s.reduce((p,c)=>Math.max(p,c.length),0),i=t.padId??this.padId,a=s.length,u=new Int32Array(a*o).fill(i);for(let p=0;p<a;p++){let c=s[p],h=Math.min(c.length,o);for(let f=0;f<h;f++)u[p*o+f]=c[f]}return Pt(u,{shape:[a,o],dtype:"i32"})}};function VO(r){if(Array.isArray(r)){let e={...Vg};for(let t=0;t<r.length&&t<dl.length;t++)e[dl[t]]=r[t];return e}return{...Vg,...r??{}}}l(VO,"normalizeSpecials");function GO(r){let e=new Set;for(let t of dl){let n=r[t];if(typeof n!="string"||n.length===0)throw new Error(`mlfw tokenizer: special token '${t}' must be a non-empty string`);if(e.has(n))throw new Error(`mlfw tokenizer: duplicate special token '${n}'`);e.add(n)}}l(GO,"validateSpecials");function jO(r,e){if(r!=null&&(!Number.isInteger(r)||r<Object.keys(e).length))throw new Error(`mlfw tokenizer: vocabSize must be an integer >= ${Object.keys(e).length}`)}l(jO,"validateVocabSize");function k$(r){if(!r||typeof r!="object")throw new Error("mlfw tokenizer: artifact must be an object");let e=r;if(e.format!==ml)throw new Error("mlfw tokenizer: unrecognized tokenizer format");if(e.version!==gl)throw new Error(`mlfw tokenizer: unsupported tokenizer version ${e.version}`);if(typeof e.mode!="string"||!(e.mode in zg))throw new Error(`mlfw tokenizer: unknown tokenizer mode '${e.mode}'`);if(!e.config||typeof e.config!="object")throw new Error("mlfw tokenizer: config must be an object");let t=e.config,n=VO(R$(e.specialTokens));if(GO(n),jO(M$(t.vocabSize),n),!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(k$,"validateTokenizerData");function v$(r){let e=[zO,va("m",[r.mode]),va("z",[r.config.vocabSize==null?"":String(r.config.vocabSize)]),va("s",dl.map(t=>r.specialTokens[t]))];if(r.mode==="bpe"){let t=[];for(let n of r.strategy.merges)t.push(n[0],n[1]);e.push(va("g",[String(r.strategy.lowercase?1:0),String(r.strategy.numMerges),r.strategy.endOfWord])),e.push(va("r",t))}else e.push(va("g",[String(r.strategy.lowercase?1:0)]));return e.push(va("v",r.vocab)),e.join(`
+`)}l(v$,"serializeTokenizer");function N$(r){if(!r.startsWith(zO))throw new Error("mlfw tokenizer: unrecognized tokenizer format");let e=new Map;for(let p of r.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),A$(p.slice(c+1)))}let t=hl(e,"m")[0],n=hl(e,"z")[0],s=hl(e,"s"),o=hl(e,"g"),i=hl(e,"v"),a=Object.fromEntries(dl.map((p,c)=>[p,s[c]])),u={lowercase:o[0]==="1"};if(t==="bpe"){let p=hl(e,"r");if(p.length%2!==0)throw new Error("mlfw tokenizer: bpe merges must be string pairs");u.numMerges=Number(o[1]),u.endOfWord=o[2],u.merges=[];for(let c=0;c<p.length;c+=2)u.merges.push([p[c],p[c+1]])}return{format:ml,version:gl,mode:t,config:{vocabSize:n===""?null:Number(n),...u},specialTokens:a,vocab:i,strategy:u}}l(N$,"parseTokenizer");function va(r,e){return`${r}	${e.map(O$).join("	")}`}l(va,"record");function A$(r){return r.split("	").map(B$)}l(A$,"parseRecord");function O$(r){return String(r).replace(/\\/g,"\\\\").replace(/\t/g,"\\t").replace(/\n/g,"\\n").replace(/\r/g,"\\r")}l(O$,"escapeField");function B$(r){let e="";for(let t=0;t<r.length;t++){let n=r[t];if(n!=="\\"){e+=n;continue}let s=r[++t];s==="t"?e+="	":s==="n"?e+=`
+`:s==="r"?e+="\r":s==="\\"?e+="\\":e+=s??""}return e}l(B$,"unescapeField");function hl(r,e){let t=r.get(e);if(!t)throw new Error(`mlfw tokenizer: missing '${e}' record`);return t}l(hl,"requiredField");function R$(r){if(r==null)return;if(Array.isArray(r))return r.map(String);if(typeof r!="object")return;let e=r,t={};for(let n of dl){let s=e[n];typeof s=="string"&&(t[n]=s)}return t}l(R$,"asSpecialTokensInput");function M$(r){return r==null||typeof r=="number"?r:Number(r)}l(M$,"asNullableNumber");var kM={};nt(kM,{Adam:()=>Yp,AdamW:()=>Jp,CosineAnnealingLR:()=>Qp,FusedAdam:()=>p_,FusedOptimizer:()=>Pl,FusedSGD:()=>c_,GradScaler:()=>f_,LRScheduler:()=>Na,Optimizer:()=>Yr,ReduceLROnPlateau:()=>Xp,SGD:()=>Hp,StepLR:()=>Zp,clipGradNorm_:()=>ef,clipGradValue_:()=>tf});var Yr=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 n=I$(e);for(let s of n)this._addParamGroup(s)}get paramGroups(){return this._paramGroups}get defaults(){return this._defaults}step(){throw new Error(`${this.constructor.name}.step() not implemented`)}zeroGrad(e=!0){for(let t of this._paramGroups)for(let n of t.params)n.grad!==null&&(e?n.grad=null:n.grad._impl.storage.data.fill(0))}stateDict(){let e=new Map;for(let[t,n]of this._state){let s={};for(let o of Object.keys(n)){let i=n[o];s[o]=UO(i)?KO(i):i}e.set(t,s)}return{state:e,paramGroups:this._paramGroups.map(t=>{let n={};for(let s of Object.keys(t)){if(s==="params")continue;let o=t[s];n[s]=Array.isArray(o)?[...o]:o}return n})}}loadStateDict(e){for(let t=0;t<this._paramGroups.length;t++){let n=e.paramGroups[t];for(let s of Object.keys(n))this._paramGroups[t][s]=Array.isArray(n[s])?[...n[s]]:n[s]}this._state=new Map;for(let[t,n]of e.state){let s={};for(let o of Object.keys(n)){let i=n[o];s[o]=UO(i)?KO(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 n=Array.isArray(e.params)?e.params:[...e.params];for(let s of n){if(this._paramIndex.has(s))throw new Error("Parameter appears in more than one parameter group");this._paramIndex.set(s,this._nextId++)}t.params=n,this._paramGroups.push(t)}_getParamId(e){return this._paramIndex.get(e)}_getState(e){let t=this._paramIndex.get(e),n=this._state.get(t);return n||(n={},this._state.set(t,n)),n}};function I$(r){if(r==null)throw new Error("Optimizer requires at least one parameter");let e=Array.isArray(r)?r:[...r];if(e.length===0)throw new Error("Optimizer got an empty parameter list");return E$(e[0])?e:[{params:e}]}l(I$,"normalizeParamGroups");function UO(r){return r instanceof Float32Array||r instanceof Float64Array||r instanceof Int32Array||r instanceof Int16Array||r instanceof Int8Array||r instanceof Uint8Array||r instanceof Uint16Array||r instanceof Uint32Array}l(UO,"isTypedArray");function KO(r){return r.slice()}l(KO,"cloneTypedArray");function E$(r){return typeof r=="object"&&r!==null&&"params"in r}l(E$,"isParamGroupInput");var Hp=class extends Yr{static{l(this,"SGD")}constructor(e,{lr:t=.01,momentum:n=0,dampening:s=0,weightDecay:o=0,nesterov:i=!1}={}){if(i&&(n===0||s!==0))throw new Error("Nesterov momentum requires momentum > 0 and dampening = 0");super(e,{lr:t,momentum:n,dampening:s,weightDecay:o,nesterov:i})}step(){for(let e of this._paramGroups){let t=e.lr,n=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,h=this._getState(a);if(n===0)if(o===0)for(let f=0;f<c;f++)u[f]-=t*p[f];else for(let f=0;f<c;f++)u[f]-=t*(p[f]+o*u[f]);else{let f=h.momentumBuffer;if(f){f=f;for(let d=0;d<c;d++){let m=o!==0?p[d]+o*u[d]:p[d];f[d]=n*f[d]+(1-s)*m}}else{f=new u.constructor(c);for(let d=0;d<c;d++)f[d]=o!==0?p[d]+o*u[d]:p[d];h.momentumBuffer=f}if(i)for(let d=0;d<c;d++){let m=o!==0?p[d]+o*u[d]:p[d];u[d]-=t*(m+n*f[d])}else for(let d=0;d<c;d++)u[d]-=t*f[d]}a._impl.bumpVersion()}}}};Kn();ar();var Yp=class extends Yr{static{l(this,"Adam")}constructor(e,{lr:t=.001,betas:n=[.9,.999],eps:s=1e-8,weightDecay:o=0,amsgrad:i=!1}={}){super(e,{lr:t,betas:[...n],eps:s,weightDecay:o,amsgrad:i})}step(){let e=m0();for(let t of this._paramGroups){let n=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 h=this._getState(c);if(e&&!a&&c.device&&c.device.type==="gpu"){h.step=(h.step||0)+1;let k=h.step,A=1-Math.pow(u,k),N=1-Math.pow(p,k);if(e(c,h,{beta1:u,beta2:p,omb1:1-u,omb2:1-p,eps:o,stepSize:n/A,bc2sqrt:Math.sqrt(N),wd:i,lr:n})){c._impl.bumpVersion();continue}}let f=c._impl.storage.data,d=c.grad._impl.storage.data,m=f.length,g=f.constructor;h.expAvg===void 0&&(h.step===void 0&&(h.step=0),h.expAvg=new g(m),h.expAvgSq=new g(m),a&&(h.maxExpAvgSq=new g(m))),h.step=h.step+1;let y=h.step,_=h.expAvg,b=h.expAvgSq,w=1-Math.pow(u,y),T=1-Math.pow(p,y),S=n/w,x=Math.sqrt(T);if(a){let k=h.maxExpAvgSq;for(let A=0;A<m;A++){let N=i!==0?d[A]+i*f[A]:d[A];_[A]=u*_[A]+(1-u)*N,b[A]=p*b[A]+(1-p)*N*N,b[A]>k[A]&&(k[A]=b[A]),f[A]-=S*_[A]/(Math.sqrt(k[A])/x+o)}}else for(let k=0;k<m;k++){let A=i!==0?d[k]+i*f[k]:d[k];_[k]=u*_[k]+(1-u)*A,b[k]=p*b[k]+(1-p)*A*A,f[k]-=S*_[k]/(Math.sqrt(b[k])/x+o)}c._impl.bumpVersion()}}}};var Jp=class extends Yr{static{l(this,"AdamW")}constructor(e,{lr:t=.001,betas:n=[.9,.999],eps:s=1e-8,weightDecay:o=.01,amsgrad:i=!1}={}){super(e,{lr:t,betas:[...n],eps:s,weightDecay:o,amsgrad:i})}step(){for(let e of this._paramGroups){let t=e.lr,n=e.betas,s=e.eps,o=e.weightDecay,i=e.amsgrad,[a,u]=n;for(let p of e.params){if(p.grad===null)continue;let c=p._impl.storage.data,h=p.grad._impl.storage.data,f=c.length,d=this._getState(p),m=c.constructor;d.step===void 0&&(d.step=0,d.expAvg=new m(f),d.expAvgSq=new m(f),i&&(d.maxExpAvgSq=new m(f))),d.step=d.step+1;let g=d.step,y=d.expAvg,_=d.expAvgSq,b=1-Math.pow(a,g),w=1-Math.pow(u,g),T=t/b,S=Math.sqrt(w);if(o!==0){let x=1-t*o;for(let k=0;k<f;k++)c[k]*=x}if(i){let x=d.maxExpAvgSq;for(let k=0;k<f;k++)y[k]=a*y[k]+(1-a)*h[k],_[k]=u*_[k]+(1-u)*h[k]*h[k],_[k]>x[k]&&(x[k]=_[k]),c[k]-=T*y[k]/(Math.sqrt(x[k])/S+s)}else for(let x=0;x<f;x++)y[x]=a*y[x]+(1-a)*h[x],_[x]=u*_[x]+(1-u)*h[x]*h[x],c[x]-=T*y[x]/(Math.sqrt(_[x])/S+s);p._impl.bumpVersion()}}}};var Na=class{static{l(this,"LRScheduler")}_optimizer;_baseLRs;_lastEpoch;_lastLR;constructor(e,t=-1){this._optimizer=e,this._baseLRs=e.paramGroups.map(n=>n.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 n=0;n<t.length;n++)t[n].lr=e[n]}},Zp=class extends Na{static{l(this,"StepLR")}_stepSize;_gamma;constructor(e,t,n=.1,s=-1){super(e,s),this._stepSize=t,this._gamma=n,this._init()}getLR(){let e=Math.pow(this._gamma,Math.floor(this._lastEpoch/this._stepSize));return this._baseLRs.map(t=>t*e)}},Qp=class extends Na{static{l(this,"CosineAnnealingLR")}_tMax;_etaMin;constructor(e,t,n=0,s=-1){super(e,s),this._tMax=t,this._etaMin=n,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)}},Xp=class{static{l(this,"ReduceLROnPlateau")}_optimizer;_mode;_factor;_patience;_threshold;_thresholdMode;_cooldown;_minLR;_eps;_best;_numBadEpochs;_cooldownCounter;constructor(e,{mode:t="min",factor:n=.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=n,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,n=Math.max(t*this._factor,this._minLR);t-n>this._eps&&(e.lr=n)}}};be();function WO(r){return(Array.isArray(r)?r:[...r]).filter(t=>t.grad!==null)}l(WO,"_gradParams");function ef(r,e,t=2){let n=WO(r);if(n.length===0)return 0;let s;if(t===2){let i=null;for(let a of n){let u=ht(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 n){let u=Qs(Pc(a.grad));i=i===null?u:ai(i,u)}s=Number(i.item())}else{let i=null;for(let a of n){let u=ht(Js(Pc(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 n)i.grad=ee(i.grad,o);return s}l(ef,"clipGradNorm_");function tf(r,e){for(let t of WO(r))t.grad=Vd(t.grad,-e,e)}l(tf,"clipGradValue_");es();Oe();l_();Ks();var Cs="f32",Mt=l(r=>new I([r],Cs),"VEC"),Dl=new I([],Cs);function us(r,e,t){return r.broadcast(e,[t],[]).getResult(0)}l(us,"bcast");var Pl=class extends Yr{static{l(this,"FusedOptimizer")}_target;_kernels;constructor(e,t,n=null){super(e,t),this._target=n||Us(),this._kernels=new Map}_kernel(e){let t=this._kernels.get(e);if(!t){let n=this._buildGraph(e);t=SM(n,this._target,{fusion:{enabled:!0}}),this._kernels.set(e,t)}return t}_buildGraph(e){throw new Error(`${this.constructor.name}._buildGraph() not implemented`)}},c_=class extends Pl{static{l(this,"FusedSGD")}constructor(e,{lr:t=.01,momentum:n=0,dampening:s=0,weightDecay:o=0,nesterov:i=!1}={},a=null){if(i&&(n===0||s!==0))throw new Error("Nesterov momentum requires momentum > 0 and dampening = 0");super(e,{lr:t,momentum:n,dampening:s,weightDecay:o,nesterov:i},a)}_buildGraph(e){let t=this._defaults.momentum,n=this._defaults.dampening,s=this._defaults.weightDecay,o=this._defaults.nesterov,i=t!==0,a=i?[Mt(e),Mt(e),Mt(e),Dl]:[Mt(e),Mt(e),Dl],u=i?[Mt(e),Mt(e)]:[Mt(e)];return hc("sgd_update",a,u,(c,h)=>{let f=h[0],d=h[1],m=i?h[3]:h[2],g=us(c,m,e),y=d;if(s!==0){let k=c.scalarConstant(s,Cs).getResult(0);y=c.add(d,c.mul(us(c,k,e),f).getResult(0)).getResult(0)}if(!i){let k=c.sub(f,c.mul(g,y).getResult(0)).getResult(0);c.returnOp([k]);return}let _=h[2],b=us(c,c.scalarConstant(t,Cs).getResult(0),e),w=us(c,c.scalarConstant(1-n,Cs).getResult(0),e),T=c.add(c.mul(b,_).getResult(0),c.mul(w,y).getResult(0)).getResult(0),S=T;o&&(S=c.add(y,c.mul(b,T).getResult(0)).getResult(0));let x=c.sub(f,c.mul(g,S).getResult(0)).getResult(0);c.returnOp([x,T])})}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 h=this._getState(i),f=h.momentumBuffer;f||(f=new a.constructor(p),h.momentumBuffer=f),f=f,c.run("sgd_update",a,u,f,o,a,f)}else c.run("sgd_update",a,u,o,a);i._impl.bumpVersion()}}}},p_=class extends Pl{static{l(this,"FusedAdam")}constructor(e,{lr:t=.001,betas:n=[.9,.999],eps:s=1e-8,weightDecay:o=0,amsgrad:i=!1}={},a=null){super(e,{lr:t,betas:[...n],eps:s,weightDecay:o,amsgrad:i},a)}_buildGraph(e){let t=this._defaults.betas,n=this._defaults.eps,s=this._defaults.weightDecay,o=this._defaults.amsgrad,[i,a]=t,u=o?[Mt(e),Mt(e),Mt(e),Mt(e),Mt(e),Dl,Dl]:[Mt(e),Mt(e),Mt(e),Mt(e),Dl,Dl],p=o?[Mt(e),Mt(e),Mt(e),Mt(e)]:[Mt(e),Mt(e),Mt(e)];return hc("adam_update",u,p,(c,h)=>{let f=h[0],d=h[1],m=h[2],g=h[3],y=o?h[4]:null,_=o?h[5]:h[4],b=o?h[6]:h[5],w=d;if(s!==0){let J=us(c,c.scalarConstant(s,Cs).getResult(0),e);w=c.add(d,c.mul(J,f).getResult(0)).getResult(0)}let T=us(c,c.scalarConstant(i,Cs).getResult(0),e),S=us(c,c.scalarConstant(a,Cs).getResult(0),e),x=us(c,c.scalarConstant(1-i,Cs).getResult(0),e),k=us(c,c.scalarConstant(1-a,Cs).getResult(0),e),A=us(c,c.scalarConstant(n,Cs).getResult(0),e),N=c.add(c.mul(T,m).getResult(0),c.mul(x,w).getResult(0)).getResult(0),v=c.mul(w,w).getResult(0),O=c.add(c.mul(S,g).getResult(0),c.mul(k,v).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),us(c,b,e)).getResult(0),A).getResult(0),z=c.mul(us(c,_,e),c.div(N,E).getResult(0)).getResult(0),D=c.sub(f,z).getResult(0);c.returnOp(o?[D,N,O,B]:[D,N,O])})}step(){for(let e of this._paramGroups){let t=e.lr,n=e.betas,s=e.amsgrad,[o,i]=n;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,h=this._getState(a);h.step===void 0&&(h.step=0,h.expAvg=new u.constructor(c),h.expAvgSq=new u.constructor(c),s&&(h.maxExpAvgSq=new u.constructor(c))),h.step=h.step+1;let f=h.step,d=1-Math.pow(o,f),m=1-Math.pow(i,f),g=new Float32Array([t/d]),y=new Float32Array([Math.sqrt(m)]),_=this._kernel(c),b=h.expAvg,w=h.expAvgSq;if(s){let T=h.maxExpAvgSq;_.run("adam_update",u,p,b,w,T,g,y,u,b,w,T)}else _.run("adam_update",u,p,b,w,g,y,u,b,w);a._impl.bumpVersion()}}}};var f_=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 n=0;n<t.length;n++)t[n]*=this._scale;return e._impl.bumpVersion&&e._impl.bumpVersion(),e}unscale_(e){if(!this.enabled)return!1;let t=1/this._scale,n=!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)||(n=!0),i[a]=u}}return this._unscaled.add(e),n&&(this._foundInf=!0),n}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)}};Sx();T_();Ks();var dI={};nt(dI,{Accuracy:()=>Uf,CSVLogger:()=>Vf,Callback:()=>$n,CallbackConnector:()=>Vl,ConfusionMatrix:()=>Yf,ConsoleLogger:()=>Ca,EarlyStopping:()=>Pf,EvaluationLoop:()=>Wl,F1Score:()=>Hf,FitLoop:()=>Kl,GradientAccumulationScheduler:()=>zf,HOOKS:()=>vx,LearningRateMonitor:()=>$f,LightningModule:()=>If,Logger:()=>$o,LoggerConnector:()=>Gl,MeanMetric:()=>Gf,Metric:()=>Yn,MetricAccumulator:()=>$l,MetricCollection:()=>qf,ModelCheckpoint:()=>Oi,Precision:()=>Kf,PredictionLoop:()=>Hl,ProgressCallback:()=>Ai,Recall:()=>Wf,SingleDeviceStrategy:()=>zl,Stage:()=>vi,SumMetric:()=>jf,Timer:()=>Ff,Trainer:()=>Df,TrainerState:()=>Fl,TrainingLoop:()=>ql,ValidationLoop:()=>Ul,applyCheckpoint:()=>Mx,deserializeCheckpoint:()=>N_,loadCheckpoint:()=>Rx,serializeCheckpoint:()=>Cf});var If=class extends ge{static{l(this,"LightningModule")}_trainer;_logBuffer;_automaticOptimization;_currentOptimizers;_device;constructor(){super(),this._trainer=null,this._logBuffer=new Map,this._automaticOptimization=!0,this._currentOptimizers=[],this._device=null}get trainer(){return this._trainer}get currentEpoch(){return this._trainer?this._trainer.state.epoch:0}get globalStep(){return this._trainer?this._trainer.state.globalStep:0}get device(){return this._device}get logger(){return this._trainer?this._trainer.logger:null}get loggers(){return this._trainer?this._trainer.loggers:[]}get automaticOptimization(){return this._automaticOptimization}set automaticOptimization(e){this._automaticOptimization=e}get optimizers(){return this._currentOptimizers}trainingStep(e,t){throw new Error(`${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:n=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:n!==null?n:u,onEpoch:s!==null?s:p,reduceFx:o,progBar:i})}logDict(e,t={}){for(let n of Object.keys(e))this.log(n,e[n],t)}manualBackward(e){this._trainer&&this._trainer.strategy?this._trainer.strategy.backward(e):e.backward()}};function JM(r){if(!r)throw new Error("configureOptimizers() returned null/undefined");if(Array.isArray(r)){let t=[],n=[];for(let s=0;s<r.length;s++){let o=r[s];c3(o)?(t.push(o.optimizer),n.push(YM(o.lrScheduler))):(t.push(o),n.push(null))}return{optimizers:t,schedulerConfigs:n}}let e=r;if(e.optimizer)return{optimizers:[e.optimizer],schedulerConfigs:[YM(e.lrScheduler)]};if(l3(r))return{optimizers:[r],schedulerConfigs:[null]};throw new Error("configureOptimizers() returned an unrecognized format")}l(JM,"parseOptimizersConfig");function YM(r){return r?"scheduler"in r&&r.scheduler?{scheduler:r.scheduler,interval:r.interval||"epoch",frequency:r.frequency||1,monitor:r.monitor||null}:ZM(r)?{scheduler:r,interval:"epoch",frequency:1,monitor:null}:null:null}l(YM,"normalizeSchedulerConfig");function ZM(r){return typeof r=="object"&&r!==null&&typeof r.step=="function"}l(ZM,"isSchedulerLike");function l3(r){return ZM(r)}l(l3,"isOptimizerLike");function c3(r){return typeof r=="object"&&r!==null&&"optimizer"in r}l(c3,"isOptimizerConfig");var vi=(o=>(o.IDLE="idle",o.TRAINING="training",o.VALIDATING="validating",o.TESTING="testing",o.PREDICTING="predicting",o))(vi||{}),p3={mean:l(r=>r.count===0?0:r.sum/r.count,"mean"),sum:l(r=>r.sum,"sum"),min:l(r=>r.min,"min"),max:l(r=>r.max,"max"),last:l(r=>r.last,"last")};function f3(){return{sum:0,count:0,min:1/0,max:-1/0,last:0}}l(f3,"createAccumulator");function h3(r,e){let t=typeof e=="number"?e:e.item();r.sum+=t,r.count+=1,t<r.min&&(r.min=t),t>r.max&&(r.max=t),r.last=t}l(h3,"updateAccumulator");var $l=class{static{l(this,"MetricAccumulator")}_accumulators;_reduceFns;constructor(){this._accumulators=new Map,this._reduceFns=new Map}update(e,t,n="mean"){this._accumulators.has(e)||(this._accumulators.set(e,f3()),this._reduceFns.set(e,n)),h3(this._accumulators.get(e),t)}compute(e){let t=this._accumulators.get(e);if(!t)return;let n=p3[this._reduceFns.get(e)];return n(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}},Fl=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 $l,this.epochMetrics=new $l}resetEpochMetrics(){this.epochMetrics.reset()}resetStepMetrics(){this.stepMetrics.reset()}},zl=class{static{l(this,"SingleDeviceStrategy")}device;constructor(){this.device=null}setup(e,t){this.device=t,(t.type==="gpu"||t.type==="webgpu")&&d3(e,"to")&&e.to(t)}toDevice(e){if(!this.device||this.device.type!=="gpu"&&this.device.type!=="webgpu")return e;if(m3(e))return e.to(this.device);if(Array.isArray(e)){let t=new Array(e.length);for(let n=0;n<e.length;n++)t[n]=this.toDevice(e[n]);return t}return e}backward(e){e.backward()}optimizerStep(e){e.step()}};function d3(r,e){return typeof r=="object"&&r!==null&&typeof r[e]=="function"}l(d3,"hasMethod");function m3(r){return typeof r=="object"&&r!==null&&"device"in r&&typeof r.to=="function"}l(m3,"hasDeviceTo");var vx=(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))(vx||{}),Vl=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 n=0;n<this._callbacks.length;n++){let s=this._callbacks[n][e];typeof s=="function"&&s.apply(this._callbacks[n],t)}}remove(e){let t=this._callbacks.indexOf(e);t!==-1&&this._callbacks.splice(t,1)}},Gl=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[n,s]of t){let{value:o,onStep:i,onEpoch:a,reduceFx:u,progBar:p}=s;if(a&&this._state.epochMetrics.update(n,o,u),i&&this._state.stepMetrics.update(n,o,u),p){this._state._progBarMetrics||(this._state._progBarMetrics=new Map);let c=typeof o=="number"?o:o.item();this._state._progBarMetrics.set(n,c)}}t.clear()}}flushStepMetrics(e){let t=this._state.stepMetrics.computeAll();if(Object.keys(t).length===0)return t;for(let n=0;n<this._loggers.length;n++)this._loggers[n].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 n=0;n<this._loggers.length;n++)this._loggers[n].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();Tm();sp();function Ni(r,e){return r==null?e:typeof r=="number"?r>0&&r<=1?Math.max(1,Math.round(r*e)):Math.min(r,e):e}l(Ni,"resolveLimit");async function jl(r){let{GradMode:e}=await Promise.resolve().then(()=>(sp(),_O)),t=e.isEnabled();e.setEnabled(!1);try{await r()}finally{e.setEnabled(t)}}l(jl,"noGradAsync");var ql=class{static{l(this,"TrainingLoop")}async run(e,t,n,s,o){let i=n.state,a=n.callbackConnector,u=n.loggerConnector,p=n.strategy,c=n.accumulateGradBatches,h=Ni(n.limitTrainBatches,t.length);i.numTrainingBatches=h,i.stage="training",e.train(),rl.setEnabled(!0),e.onTrainEpochStart(),a.dispatch("onTrainEpochStart",n,e);let f=0;for(let m of t){if(f>=h||i.shouldStop)break;if(i.maxSteps>0&&i.globalStep>=i.maxSteps){i.shouldStop=!0;break}let g=p.toDevice(m);a.dispatch("onTrainBatchStart",n,e,g,f);let y;e.automaticOptimization?y=await this._automaticStep(e,g,f,n,s,o,p,c,a):y=await Promise.resolve(e.trainingStep(g,f)),u.drain(e),i.globalStep%n.logEveryNSteps===0&&u.flushStepMetrics(i.globalStep),a.dispatch("onTrainBatchEnd",n,e,y,g,f),c0(),i.globalStep++,f++}c0();let d=u.flushEpochMetrics(i.globalStep);return this._stepEpochSchedulers(o,i.epoch),e.onTrainEpochEnd(),a.dispatch("onTrainEpochEnd",n,e),d}async _automaticStep(e,t,n,s,o,i,a,u,p){if(s.compile)return this._compiledStep(e,t,n,s,o,i,u);if(s.cudaGraph)return this._graphedStep(e,t,s,o,i,a);let c=await Promise.resolve(e.trainingStep(t,n)),h=c,f=c;if(oI(c)&&(h=c.loss,f=c),u>1&&(h=$t(h,u)),p.dispatch("onBeforeBackward",s,e,h),a.backward(h),p.dispatch("onAfterBackward",s,e),(n+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 f}async _compiledStep(e,t,n,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),Nx(c)&&(c=await c);else{let{compileWithBackward:_}=await Promise.resolve().then(()=>(T_(),KM)),{CPUTarget:b,CUDATarget:w,WebGPUTarget:T}=await Promise.resolve().then(()=>(Ks(),lv)),S=e._device&&e._device.type,x=S==="webgpu"?T():S==="gpu"?w():b();e.__compiledTrainStep=_({forward:p},u,{target:x,mode:s.compileMode});let k=e.log;e.log=()=>{};try{c=e.__compiledTrainStep(...u),Nx(c)&&(c=await c)}finally{e.log=k}}let h=e.__compiledTrainStep,f=h.capturedParams(),{ones:d}=await Promise.resolve().then(()=>(_t(),o1)),m=h.backward(d(c.shape));Nx(m)&&(m=await m);let g=m,y=g.length-f.length;for(let _=0;_<f.length;_++){let b=g[y+_];b&&(f[_].grad=b)}if((n+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,n,s,o){let i=await Promise.resolve(e.trainingStep(t,0)),a=i;if(oI(i)&&(a=i.loss),s.backward(a),o&&o.gradientClipVal){let{deviceClipGradNorm:u}=await Promise.resolve().then(()=>(Ad(),Nd));u([...e.parameters()],o.gradientClipVal)}for(let u=0;u<n.length;u++)s.optimizerStep(n[u]),n[u].zeroGrad();return a}async _graphedStep(e,t,n,s,o,i){let a=await Promise.resolve().then(()=>(XM(),QM)),u=await Promise.resolve().then(()=>(tI(),eI)),p=await Promise.resolve().then(()=>(sI(),rI)),c=e.__eagerGraphRunner;if(c||(c=e.__eagerGraphRunner={phase:"warmup",seen:0},u.setEagerDeferred(!0),EA(!0)),c.phase==="disabled")return this._eagerTrainStepCore(e,t,s,i,n);if(c.phase==="warmup"&&c.seen<n.cudaGraphWarmupSteps)return c.seen++,this._eagerTrainStepCore(e,t,s,i,n);let h=Ax(t);if(c.phase==="warmup"){c.inputs=h.map(y=>{let _=y._impl.storage.rawData;return u.deviceBufferForInput(_),u.pinResident(_),{dptr:u.deviceBufferDptr(_)}});let d=e.log,m;try{a.beginEagerCapture(),e.log=()=>{};try{m=await this._eagerTrainStepCore(e,t,s,i,n)}finally{e.log=d}c.captured=a.endEagerCapture()}catch(y){try{a.endEagerCapture()}catch{}return e.log=d,u.clearCapturePins(),c.phase="disabled",c.captureError=y&&typeof y=="object"&&"message"in y?y.message:y,this._eagerTrainStepCore(e,t,s,i,n)}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(n,g),this._stepStepSchedulers(o,n.state.globalStep),g}for(let d=0;d<c.inputs.length&&d<h.length;d++)c.inputs[d].dptr&&p.copyHostToDeviceAsync(c.inputs[d].dptr,h[d]._impl.storage.rawData);a.replay(c.exec),a.syncStream(),p.copyDeviceToHost(c.lossScratch,c.lossDptr);let f=c.lossScratch[0];return this._logGraphLoss(n,f),this._stepStepSchedulers(o,n.state.globalStep),f}_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 n=[...e.parameters()];t.gradientClipAlgorithm==="norm"?ef(n,t.gradientClipVal):tf(n,t.gradientClipVal)}_stepStepSchedulers(e,t){if(e)for(let n=0;n<e.length;n++){let s=e[n];!s||s.interval!=="step"||s.monitor||(t+1)%s.frequency===0&&s.scheduler.step()}}_stepEpochSchedulers(e,t){if(e)for(let n=0;n<e.length;n++){let s=e[n];!s||s.interval!=="epoch"||s.monitor||(t+1)%s.frequency===0&&s.scheduler.step()}}};function Ax(r,e=[]){if(r==null)return e;if(A3(r))return e.push(r),e;if(Array.isArray(r)){for(let t of r)Ax(t,e);return e}if(typeof r=="object"){let t=r;for(let n of Object.keys(t))Ax(t[n],e);return e}return e}l(Ax,"_flattenTensors");function Nx(r){return typeof r=="object"&&r!==null&&typeof r.then=="function"}l(Nx,"isThenable");function oI(r){return typeof r=="object"&&r!==null&&!("backward"in r)&&"loss"in r}l(oI,"isLossObject");function A3(r){return typeof r=="object"&&r!==null&&"shape"in r&&typeof r.contiguous=="function"}l(A3,"isFlattenTensor");var Ul=class{static{l(this,"ValidationLoop")}async run(e,t,n,s){let o=n.state,i=n.callbackConnector,a=n.loggerConnector,u=o.stage;o.stage="validating",e.eval(),e.onValidationEpochStart(),i.dispatch("onValidationStart",n,e),i.dispatch("onValidationEpochStart",n,e);let p=Ni(n.limitValBatches,t.length);o.numValBatches=p;let c=0;await jl(async()=>{for(let f of t){if(c>=p)break;i.dispatch("onValidationBatchStart",n,e,f,c);let d=await Promise.resolve(e.validationStep(f,c));a.drain(e),i.dispatch("onValidationBatchEnd",n,e,d,f,c),c++}});let h=a.flushEpochMetrics(o.globalStep);return this._stepPlateauSchedulers(s,h),e.onValidationEpochEnd(),i.dispatch("onValidationEpochEnd",n,e),i.dispatch("onValidationEnd",n,e),e.train(),o.stage=u,h}_stepPlateauSchedulers(e,t){if(e)for(let n=0;n<e.length;n++){let s=e[n];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 Kl=class{static{l(this,"FitLoop")}_trainingLoop;_validationLoop;constructor(){this._trainingLoop=new ql,this._validationLoop=new Ul}get trainingLoop(){return this._trainingLoop}get validationLoop(){return this._validationLoop}async run(e,t,n,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++)n&&this._shouldRunValidation(p,s)&&await this._validationLoop.run(e,n,s,i);u.dispatch("onTrainEnd",s,e)}_shouldRunValidation(e,t){let n=t.checkValEveryNEpoch;return(e+1)%n===0}};var Wl=class{static{l(this,"EvaluationLoop")}async run(e,t,n){let s=n.state,o=n.callbackConnector,i=n.loggerConnector,a=s.stage;s.stage="testing",e.eval(),e.onTestEpochStart(),o.dispatch("onTestStart",n,e),o.dispatch("onTestEpochStart",n,e);let u=Ni(n.limitTestBatches,t.length),p=0;await jl(async()=>{for(let h of t){if(p>=u)break;o.dispatch("onTestBatchStart",n,e,h,p);let f=await Promise.resolve(e.testStep(h,p));i.drain(e),o.dispatch("onTestBatchEnd",n,e,f,h,p),p++}});let c=i.flushEpochMetrics(s.globalStep);return e.onTestEpochEnd(),o.dispatch("onTestEpochEnd",n,e),o.dispatch("onTestEnd",n,e),e.train(),s.stage=a,c}};var Hl=class{static{l(this,"PredictionLoop")}async run(e,t,n){let s=n.state,o=n.callbackConnector,i=n.strategy,a=s.stage;s.stage="predicting",e.eval(),o.dispatch("onPredictStart",n,e);let u=[],p=Ni(n.limitTestBatches,t.length),c=0;return await jl(async()=>{for(let h of t){if(c>=p)break;let f=i.toDevice(h);o.dispatch("onPredictBatchStart",n,e,f,c);let d=await Promise.resolve(e.predictStep(f,c));await n._flushEagerInference(),u.push(d),o.dispatch("onPredictBatchEnd",n,e,d,f,c),c++}}),o.dispatch("onPredictEnd",n,e),e.train(),s.stage=a,u}};var $o=class{static{l(this,"Logger")}_name;_version;constructor({name:e="default",version:t=0}={}){this._name=e,this._version=t}get name(){return this._name}get version(){return this._version}logMetrics(e,t){throw new Error(`${this.constructor.name}.logMetrics() not implemented`)}logHyperparams(e){throw new Error(`${this.constructor.name}.logHyperparams() not implemented`)}finalize(){}};var Ca=class extends $o{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 n=[`[step ${t}]`],s=Object.keys(e).sort();for(let o=0;o<s.length;o++){let i=e[s[o]];n.push(`${s[o]}: ${O3(i)}`)}console.log(n.join(" | "))}logHyperparams(e){let t=["[hyperparams]"],n=Object.keys(e).sort();for(let s=0;s<n.length;s++)t.push(`${n[s]}: ${e[n[s]]}`);console.log(t.join(" | "))}};function O3(r){return typeof r!="number"||Number.isInteger(r)?String(r):Math.abs(r)<.001&&r!==0?r.toExponential(3):r.toFixed(4)}l(O3,"formatNumber");var $n=class{static{l(this,"Callback")}setup(e,t,n){}teardown(e,t,n){}onFitStart(e,t){}onFitEnd(e,t){}onTrainStart(e,t){}onTrainEnd(e,t){}onTrainEpochStart(e,t){}onTrainEpochEnd(e,t){}onTrainBatchStart(e,t,n,s){}onTrainBatchEnd(e,t,n,s,o){}onValidationStart(e,t){}onValidationEnd(e,t){}onValidationEpochStart(e,t){}onValidationEpochEnd(e,t){}onValidationBatchStart(e,t,n,s){}onValidationBatchEnd(e,t,n,s,o){}onTestStart(e,t){}onTestEnd(e,t){}onTestBatchStart(e,t,n,s){}onTestBatchEnd(e,t,n,s,o){}onPredictStart(e,t){}onPredictEnd(e,t){}onPredictBatchStart(e,t,n,s){}onPredictBatchEnd(e,t,n,s,o){}onBeforeBackward(e,t,n){}onAfterBackward(e,t){}onBeforeOptimizerStep(e,t,n){}onBeforeZeroGrad(e,t,n){}onSaveCheckpoint(e,t,n){}onLoadCheckpoint(e,t,n){}};var B3=["\u258F","\u258E","\u258D","\u258C","\u258B","\u258A","\u2589"];function iI(r){globalThis.process.stdout.write(r)}l(iI,"stdoutWrite");var Ai=class extends $n{static{l(this,"ProgressCallback")}_barLength;_trainBatchCount;_valBatchCount;_epochStartTime;_lastLen;_active;constructor({barLength:e=24}={}){super(),this._barLength=e,this._trainBatchCount=0,this._valBatchCount=0,this._epochStartTime=0,this._lastLen=0,this._active=!1}onTrainEpochStart(e,t){this._trainBatchCount=0,this._epochStartTime=Date.now();let n=this._trainTotal(e);n&&this._render("Epoch",e.state.epoch+1,e.state.maxEpochs,0,n,e.state)}onTrainBatchEnd(e,t,n,s,o){this._trainBatchCount++;let i=this._trainTotal(e);i&&this._render("Epoch",e.state.epoch+1,e.state.maxEpochs,this._trainBatchCount,i,e.state)}onTrainEpochEnd(e,t){let n=this._trainTotal(e);n&&this._render("Epoch",e.state.epoch+1,e.state.maxEpochs,n,n,e.state)}onTrainEnd(e,t){this._active&&iI(`
+`),this._active=!1,this._lastLen=0}onValidationEpochStart(e,t){this._valBatchCount=0,this._epochStartTime=Date.now()}onValidationBatchEnd(e,t,n,s,o){this._valBatchCount++;let i=this._valTotal(e);i&&this._render("Validation",null,null,this._valBatchCount,i,e.state)}onValidationEnd(e,t){let n=this._valTotal(e);n&&this._render("Validation",null,null,n,n,e.state)}_trainTotal(e){return e.state.numTrainingBatches??aI(e.limitTrainBatches)}_valTotal(e){return e.state.numValBatches??aI(e.limitValBatches)}_render(e,t,n,s,o,i){let a=o>0?Math.min(s/o,1):0,u=String(Math.round(a*100)).padStart(3," "),p=this._bar(a),c=t!==null?`${e} ${t}/${n}`:e,h=(Date.now()-this._epochStartTime)/1e3,f=h>0?s/h:0,d=f>0?(o-s)/f:0,m=`${uI(h)}<${uI(d)}, ${f.toFixed(2)}it/s`,g=this._formatProgBarMetrics(i),y=`${c}: ${u}%|${p}| ${s}/${o} [${m}${g}]`,_=Math.max(0,this._lastLen-y.length);iI("\r"+y+" ".repeat(_)),this._lastLen=y.length,this._active=!0}_bar(e){let t=this._barLength,n=e*t,s=Math.floor(n),o=Math.round((n-s)*8);if(o===8&&(s+=1,o=0),s>=t)return"\u2588".repeat(t);let i=o>0?B3[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 n=[];for(let[s,o]of t)n.push(`${s}=${R3(o)}`);return", "+n.join(", ")}};function aI(r){return r==null?null:typeof r=="number"&&r>1?r:null}l(aI,"resolveTotal");function uI(r){(!Number.isFinite(r)||r<0)&&(r=0);let e=Math.round(r),t=Math.floor(e/60),n=e%60;return`${String(t).padStart(2,"0")}:${String(n).padStart(2,"0")}`}l(uI,"fmtTime");function R3(r){return typeof r!="number"||Number.isInteger(r)?String(r):r.toFixed(4)}l(R3,"formatNum");function Fo(...r){return r.filter(e=>e!=null&&e!=="").join("/").replace(/\/+/g,"/")}l(Fo,"joinPath");jn();qr();ts();var cI="mlfw-ckpt-v1",Lf=8,M3={Int8Array,Uint8Array,Uint8ClampedArray,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array};function I3(r){return ArrayBuffer.isView(r)&&!(r instanceof DataView)}l(I3,"isTypedArray");function Ox(r){return new Uint8Array(r.buffer,r.byteOffset,r.byteLength)}l(Ox,"byteView");function E3(r){return r.constructor.name}l(E3,"typedArrayName");function L3(r){return"$buf"in r&&typeof r.$buf=="number"}l(L3,"isBufferNode");function C3(r){return"$map"in r&&Array.isArray(r.$map)}l(C3,"isMapNode");function Cf(r){let e=[],t=k_(r,e),n=0;for(let u of e)u.meta.offset=n,u.meta.length=u.bytes.length,n+=u.bytes.length;let s={format:cI,tree:t,buffers:e.map(u=>u.meta)},o=new TextEncoder().encode(JSON.stringify(s)),i=new Uint8Array(Lf+o.length+n);new DataView(i.buffer).setBigUint64(0,BigInt(o.length),!0),i.set(o,Lf);let a=Lf+o.length;for(let u of e)i.set(u.bytes,a),a+=u.bytes.length;return i}l(Cf,"serializeCheckpoint");function N_(r){let e=r instanceof Uint8Array?r:new Uint8Array(r),t=new DataView(e.buffer,e.byteOffset,e.byteLength),n=Number(t.getBigUint64(0,!0)),s=Lf+n,o=JSON.parse(new TextDecoder().decode(e.subarray(Lf,s)));if(o.format!==cI)throw new Error("mlfw: unrecognized checkpoint format: "+o.format);let i=o.buffers.map(a=>D3(a,e,s));return v_(o.tree,i)}l(N_,"deserializeCheckpoint");function k_(r,e){if(r instanceof We)return lI(e,{kind:"tensor",dtype:r.dtype,shape:r.shape},Ox(r._impl.storage.data));if(I3(r))return lI(e,{kind:"array",arrayType:E3(r)},Ox(r));if(r instanceof Map){let t=[];for(let[n,s]of r)t.push([n,k_(s,e)]);return{$map:t}}if(Array.isArray(r))return r.map(t=>k_(t,e));if(r&&typeof r=="object"){let t={},n=r;for(let s of Object.keys(n))t[s]=k_(n[s],e);return t}return r}l(k_,"encodeNode");function lI(r,e,t){let n=r.length;return r.push({meta:e,bytes:t}),{$buf:n}}l(lI,"pushBuffer");function D3(r,e,t){let n=t+r.offset,s=e.subarray(n,n+r.length),o=r.kind==="tensor"?Dt(r.dtype):M3[r.arrayType],i=r.kind==="tensor"?void 0:r.arrayType;if(!o)throw new Error("mlfw: unknown typed array in checkpoint: "+i);let a=new o(r.length/o.BYTES_PER_ELEMENT);return Ox(a).set(s),r.kind==="tensor"?ta(a,r.shape,r.dtype):a}l(D3,"decodeBuffer");function v_(r,e){if(r===null||typeof r!="object")return r;if(Array.isArray(r))return r.map(n=>v_(n,e));if(L3(r))return e[r.$buf];if(C3(r)){let n=new Map;for(let[s,o]of r.$map)n.set(s,v_(o,e));return n}let t={};for(let n of Object.keys(r))t[n]=v_(r[n],e);return t}l(v_,"decodeNode");var Bx=".ckpt",Oi=class extends $n{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:n=null,mode:s="min",saveTopK:o=1,saveLast:i=!0,everyNEpochs:a=1}={}){super(),this._dirpath=e,this._filename=t,this._monitor=n,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 n=e.state;if((n.epoch+1)%this._everyNEpochs!==0)return;this._ensureDir();let s=this._fillTemplate(n);if(this._saveLast){let c=Fo(this._dirpath,"last"+Bx);this._saveCheckpoint(t,e,c),this._lastModelPath=c}if(!this._monitor){let c=Fo(this._dirpath,s+Bx);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=n.epochMetrics.computeAll()[this._monitor];if(i===void 0)return;let a=Fo(this._dirpath,s+Bx),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,n=this._bestK.length;for(;t<n;){let s=t+n>>>1;this._compareFn(e,this._bestK[s].score)<0?n=s:t=s+1}return t}_updateBest(){this._bestK.length>0&&(this._bestModelPath=this._bestK[0].path)}_saveCheckpoint(e,t,n){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=n+".tmp";dn.writeBinary(i,Cf(s)),dn.rename(i,n)}_fillTemplate(e){return this._filename.replace("{epoch}",String(e.epoch)).replace("{step}",String(e.globalStep))}_ensureDir(){dn.exists(this._dirpath)||dn.mkdir(this._dirpath)}_tryDelete(e){try{dn.remove(e)}catch{}}};function Rx(r){return N_(dn.readBinary(r))}l(Rx,"loadCheckpoint");function Mx(r,e,t=[]){let n=r;if(n.modelState&&e.loadStateDict(n.modelState),n.optimizerStates){let s=Math.min(t.length,n.optimizerStates.length);for(let o=0;o<s;o++)t[o].loadStateDict(n.optimizerStates[o])}return r}l(Mx,"applyCheckpoint");ar();var Df=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:n="auto",precision:s="f32",callbacks:o=[],logger:i=!0,enableCheckpointing:a=!1,enableProgress:u=!0,gradientClipVal:p=null,gradientClipAlgorithm:c="norm",accumulateGradBatches:h=1,limitTrainBatches:f=null,limitValBatches:d=null,limitTestBatches:m=null,valCheckInterval:g=1,checkValEveryNEpoch:y=1,logEveryNSteps:_=50,deterministic:b=!1,fastDevRun:w=!1,defaultRootDir:T="./lightning_logs",compile:S=!1,compileMode:x="separate",cudaGraph:k=!1,cudaGraphWarmupSteps:A=3}={}){if(this._state=new Fl,this._state.maxEpochs=e,this._state.maxSteps=t,this._compile=S,this._compileMode=x,this._cudaGraph=k,this._cudaGraphWarmupSteps=A,this._accelerator=n,this._precision=s,this._gradientClipVal=p,this._gradientClipAlgorithm=c,this._accumulateGradBatches=h,this._limitTrainBatches=f,this._limitValBatches=d,this._limitTestBatches=m,this._valCheckInterval=g,this._checkValEveryNEpoch=y,this._logEveryNSteps=_,this._deterministic=b,this._defaultRootDir=T,w!==!1){let v=typeof w=="number"?w:1;this._limitTrainBatches=v,this._limitValBatches=v,this._limitTestBatches=v,this._state.maxEpochs=1}this._loggers=this._resolveLoggers(i),this._strategy=new zl,this._fitLoop=new Kl,this._evaluationLoop=new Wl,this._predictionLoop=new Hl;let N=[...o];u&&!N.some(v=>v instanceof Ai)&&N.push(new Ai),a&&!N.some(v=>v instanceof Oi)&&N.push(new Oi({dirpath:T})),this._callbackConnector=new Vl(N),this._loggerConnector=new Gl(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,n=null){this._model=e,e._trainer=this;let s=this._resolveDevice();this._guardEagerWebGPU(s,"fit",n!=null),this._guardCudaGraph(s,n!=null),e._device=s,await this._prepareDevice(s),this._strategy.setup(e,s);let{optimizers:o,schedulerConfigs:i}=JM(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,n,this,o,i),s===Od){let{teardownAfterFit:a}=await Promise.resolve().then(()=>(Ad(),Nd));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 n=this._resolveDevice();this._guardEagerWebGPU(n,"validate"),e._device=n,await this._prepareDevice(n),this._strategy.setup(e,n),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 n=this._resolveDevice();this._guardEagerWebGPU(n,"test"),e._device=n,await this._prepareDevice(n),this._strategy.setup(e,n),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 n=this._resolveDevice();return e._device=n,await this._prepareDevice(n),this._strategy.setup(e,n),await this._predictionLoop.run(e,t,this)}_resolveDevice(){return this._accelerator==="gpu"?Od:this._accelerator==="wasm"?sw:this._accelerator==="webgpu"?Ys:this._accelerator==="cpu"?ks:ks}_guardEagerWebGPU(e,t,n=!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(n)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"): ${t}() reads scalar metrics synchronously via .item(), which WebGPU's asynchronous readback cannot serve eagerly. Use predict() for eager WebGPU inference, or train via compile=true.`)}}_guardCudaGraph(e,t=!1){if(this._cudaGraph){if(e.type!=="gpu")throw new Error('Trainer(cudaGraph=true) requires accelerator="gpu" (eager CUDA whole-step capture/replay).');if(this._compile)throw new Error("Trainer(cudaGraph=true) is incompatible with compile=true: CUDA graph capture targets the eager training step, not the compiled path.");if(this._gradientClipVal!=null&&this._gradientClipAlgorithm!=="norm")throw new Error('Trainer(cudaGraph=true) supports gradient_clip_algorithm="norm" only; "value" clipping is not yet device-side.');if(this._accumulateGradBatches!==1)throw new Error("Trainer(cudaGraph=true) v1 requires accumulateGradBatches=1.");if(t)throw new Error("Trainer(cudaGraph=true) v1 does not support in-fit validation. Call fit() without a valLoader.")}}async _prepareDevice(e){if(e.type==="gpu"){let{preloadCudaRuntime:t}=await Promise.resolve().then(()=>(Kc(),Bw));await t()}else if(e.type==="webgpu"){let{preloadWebGPU:t}=await Promise.resolve().then(()=>(Kc(),Bw));this._webgpuMod=await t()}}async _flushEagerInference(){this._webgpuMod&&await this._webgpuMod.flushWebGPUEager()}_resolveLoggers(e){return e===!1||e===null?[]:e===!0?[new Ca]:Array.isArray(e)?e:[e]}_extractHyperparams(e,t){let n={maxEpochs:this._state.maxEpochs,maxSteps:this._state.maxSteps,accelerator:this._accelerator,precision:this._precision,accumulateGradBatches:this._accumulateGradBatches};for(let s=0;s<t.length;s++){let o=t[s],i=o.defaults||{},a=t.length>1?`optimizer_${s}_`:"";n[a+"optimizer"]=o.constructor.name,i.lr!==void 0&&(n[a+"lr"]=i.lr),i.weightDecay!==void 0&&(n[a+"weight_decay"]=i.weightDecay)}return n}};var Pf=class extends $n{static{l(this,"EarlyStopping")}_monitor;_patience;_mode;_minDelta;_checkOnTrainEpochEnd;_waitCount;_bestScore;_compareFn;constructor({monitor:e="val_loss",patience:t=3,mode:n="min",minDelta:s=0,checkOnTrainEpochEnd:o=!1}={}){super(),this._monitor=e,this._patience=t,this._mode=n,this._minDelta=s,this._checkOnTrainEpochEnd=o,this._waitCount=0,this._bestScore=null,this._compareFn=n==="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 n=e.state.epochMetrics.computeAll()[this._monitor];if(n!==void 0){if(this._bestScore===null||this._compareFn(n,this._bestScore)){this._bestScore=n,this._waitCount=0;return}this._waitCount++,this._waitCount>=this._patience&&(e.shouldStop=!0)}}reset(){this._waitCount=0,this._bestScore=null}};var $f=class extends $n{static{l(this,"LearningRateMonitor")}_logMomentum;_lrHistory;constructor({logMomentum:e=!1}={}){super(),this._logMomentum=e,this._lrHistory={}}get lrHistory(){return this._lrHistory}onTrainBatchStart(e,t,n,s){let o=t._currentOptimizers;if(o)for(let i=0;i<o.length;i++){let a=o[i].paramGroups;for(let u=0;u<a.length;u++){let p=o.length>1||a.length>1?`lr_opt${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 h=p.replace("lr","momentum");t.log(h,a[u].momentum,{onStep:!0,onEpoch:!1})}}}}};var Ff=class extends $n{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 zf=class extends $n{static{l(this,"GradientAccumulationScheduler")}_scheduling;_sortedEpochs;constructor({scheduling:e}){super(),this._scheduling=new Map;let t=Object.entries(e);for(let n=0;n<t.length;n++)this._scheduling.set(Number(t[n][0]),t[n][1]);this._sortedEpochs=[...this._scheduling.keys()].sort((n,s)=>n-s)}onTrainEpochStart(e,t){let n=e.state.epoch;this._scheduling.has(n)&&(e.accumulateGradBatches=this._scheduling.get(n))}getCurrentAccumulation(e){let t=1;for(let n=0;n<this._sortedEpochs.length&&this._sortedEpochs[n]<=e;n++)t=this._scheduling.get(this._sortedEpochs[n]);return t}};var Vf=class extends $o{static{l(this,"CSVLogger")}_saveDir;_flushInterval;_columns;_columnSet;_buffer;_filePath;_headerWritten;_version;constructor({saveDir:e="./lightning_logs",name:t="default",version:n=null,flushInterval:s=10}={}){super({name:t,version:n||0}),this._saveDir=e,this._flushInterval=s,this._columns=[],this._columnSet=new Set,this._buffer=[],this._filePath=null,this._headerWritten=!1,this._version=n}get logDir(){let e=this._version!==null?this._version:this._resolveVersion();return Fo(this._saveDir,this._name,`version_${e}`)}logMetrics(e,t){let n={step:t},s=Object.keys(e);for(let o=0;o<s.length;o++){let i=s[o];n[i]=e[i],this._columnSet.has(i)||(this._columnSet.add(i),this._columns.push(i),this._headerWritten=!1)}this._buffer.push(n),this._buffer.length>=this._flushInterval&&this._flush()}logHyperparams(e){this._ensureDir();let t=Fo(this.logDir,"hparams.json");dn.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(",");dn.writeFile(e,s+`
+`),this._headerWritten=!0}let t=["step",...this._columns],n=[];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):"")}n.push(i.join(","))}dn.appendFile(e,n.join(`
+`)+`
+`),this._buffer.length=0}_getFilePath(){return this._filePath||(this._filePath=Fo(this.logDir,"metrics.csv")),this._filePath}_ensureDir(){let e=this.logDir;dn.exists(e)||dn.mkdir(e)}_resolveVersion(){if(this._version!==null)return this._version;let e=Fo(this._saveDir,this._name);if(!dn.exists(e))return this._version=0,0;let t=-1;try{let n=dn.readdir(e);for(let s=0;s<n.length;s++){let o=n[s].match(/^version_(\d+)$/);if(o){let i=parseInt(o[1],10);i>t&&(t=i)}}}catch{}return this._version=t+1,this._version}};var Yn=class{static{l(this,"Metric")}_computed;constructor(){this._computed=null}update(e,t){throw new Error(`${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 Gf=class extends Yn{static{l(this,"MeanMetric")}_sum;_count;constructor(){super(),this._sum=0,this._count=0}update(e,t=1){let n=typeof e=="number"?e:e.item();this._sum+=n*t,this._count+=t}compute(){return this._count===0?0:this._sum/this._count}reset(){super.reset(),this._sum=0,this._count=0}},jf=class extends Yn{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 qf=class{static{l(this,"MetricCollection")}_metrics;constructor(e={}){this._metrics=new Map;let t=Object.keys(e);for(let n=0;n<t.length;n++)this._metrics.set(t[n],e[t[n]])}add(e,t){return this._metrics.set(e,t),this}update(e,t){for(let[,n]of this._metrics)n.update(e,t)}compute(){let e={};for(let[t,n]of this._metrics)e[t]=n.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 Uf=class extends Yn{static{l(this,"Accuracy")}_task;_numClasses;_topK;_threshold;_correct;_total;constructor({task:e="multiclass",numClasses:t=null,topK:n=1,threshold:s=.5}={}){super(),this._task=e,this._numClasses=t,this._topK=n,this._threshold=s,this._correct=0,this._total=0}update(e,t){let n=e._impl.storage.data,s=t._impl.storage.data;this._task==="binary"?this._updateBinary(n,s):this._task==="multiclass"?this._updateMulticlass(e,t):this._task==="multilabel"&&this._updateMultilabel(n,s)}compute(){return this._total===0?0:this._correct/this._total}reset(){super.reset(),this._correct=0,this._total=0}_updateBinary(e,t){let n=t.length;for(let s=0;s<n;s++)(e[s]>=this._threshold?1:0)===t[s]&&this._correct++;this._total+=n}_updateMulticlass(e,t){let n=e.shape;if(n.length<2){let u=e._impl.storage.data,p=t._impl.storage.data,c=p.length;for(let h=0;h<c;h++)Math.round(u[h])===p[h]&&this._correct++;this._total+=c;return}let s=n[0],o=n[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 h=1;h<o;h++){let f=i[u*o+h];f>c&&(c=f,p=h)}p===a[u]&&this._correct++}else for(let u=0;u<s;u++){let p=P3(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 n=t.length;for(let s=0;s<n;s++)(e[s]>=this._threshold?1:0)===t[s]&&this._correct++;this._total+=n}};function P3(r,e,t,n){let s=[];for(let o=0;o<t;o++){let i=r[e+o];s.length<n?(s.push({val:i,idx:o}),s.length===n&&$3(s)):i>s[0].val&&(s[0]={val:i,idx:o},pI(s,0))}return s.map(o=>o.idx)}l(P3,"topKIndices");function $3(r){for(let e=(r.length>>>1)-1;e>=0;e--)pI(r,e)}l($3,"heapify");function pI(r,e){let t=r.length;for(;;){let n=e,s=2*e+1,o=2*e+2;if(s<t&&r[s].val<r[n].val&&(n=s),o<t&&r[o].val<r[n].val&&(n=o),n===e)break;let i=r[e];r[e]=r[n],r[n]=i,e=n}}l(pI,"siftDown");var Kf=class extends Yn{static{l(this,"Precision")}_task;_numClasses;_average;_tp;_fp;_support;constructor({task:e="multiclass",numClasses:t=2,average:n="macro"}={}){super(),this._task=e,this._numClasses=t,this._average=n,this._tp=new Int32Array(t),this._fp=new Int32Array(t),this._support=new Int32Array(t)}update(e,t){Ix(e,t,this._tp,this._fp,null,this._support,this._task,this._numClasses)}compute(){return z3(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)}},Wf=class extends Yn{static{l(this,"Recall")}_task;_numClasses;_average;_tp;_fn;_support;constructor({task:e="multiclass",numClasses:t=2,average:n="macro"}={}){super(),this._task=e,this._numClasses=t,this._average=n,this._tp=new Int32Array(t),this._fn=new Int32Array(t),this._support=new Int32Array(t)}update(e,t){Ix(e,t,this._tp,null,this._fn,this._support,this._task,this._numClasses)}compute(){return V3(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)}},Hf=class extends Yn{static{l(this,"F1Score")}_task;_numClasses;_average;_tp;_fp;_fn;_support;constructor({task:e="multiclass",numClasses:t=2,average:n="macro"}={}){super(),this._task=e,this._numClasses=t,this._average=n,this._tp=new Int32Array(t),this._fp=new Int32Array(t),this._fn=new Int32Array(t),this._support=new Int32Array(t)}update(e,t){Ix(e,t,this._tp,this._fp,this._fn,this._support,this._task,this._numClasses)}compute(){let e=fI(this._tp,this._fp,this._numClasses),t=hI(this._tp,this._fn,this._numClasses),n=new Float64Array(this._numClasses);for(let s=0;s<this._numClasses;s++){let o=e[s]+t[s];n[s]=o>0?2*e[s]*t[s]/o:0}return Ex(n,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 Ix(r,e,t,n,s,o,i,a){let u=r._impl.storage.data,p=e._impl.storage.data,c=p.length,h=r.shape.length>=2&&r.shape[1]>1;for(let f=0;f<c;f++){let d=p[f]|0,m;i==="binary"?m=u[f]>=.5?1:0:h?m=F3(u,f,a):m=u[f]|0,o[d]++,m===d?t&&t[d]++:(n&&n[m]++,s&&s[d]++)}}l(Ix,"updateConfusionCounters");function F3(r,e,t){let n=0,s=r[e*t];for(let o=1;o<t;o++){let i=r[e*t+o];i>s&&(s=i,n=o)}return n}l(F3,"argmaxRow");function fI(r,e,t){let n=new Float64Array(t);for(let s=0;s<t;s++){let o=r[s]+e[s];n[s]=o>0?r[s]/o:0}return n}l(fI,"computePrecisionPerClass");function hI(r,e,t){let n=new Float64Array(t);for(let s=0;s<t;s++){let o=r[s]+e[s];n[s]=o>0?r[s]/o:0}return n}l(hI,"computeRecallPerClass");function z3(r,e,t,n,s){let o=fI(r,e,n);return Ex(o,t,n,s)}l(z3,"computePrecision");function V3(r,e,t,n,s){let o=hI(r,e,n);return Ex(o,t,n,s)}l(V3,"computeRecall");function Ex(r,e,t,n){if(n==="none")return[...r];if(n==="micro"){let i=0,a=0;for(let u=0;u<t;u++)i+=r[u]*e[u],a+=e[u];return a>0?i/a:0}if(n==="weighted"){let i=0,a=0;for(let u=0;u<t;u++)a+=r[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+=r[i],s++);return s>0?o/s:0}l(Ex,"aggregate");var Yf=class extends Yn{static{l(this,"ConfusionMatrix")}_numClasses;_matrix;constructor({numClasses:e}){super(),this._numClasses=e,this._matrix=new Int32Array(e*e)}update(e,t){let n=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=G3(n,u,i):c=n[u]|0,this._matrix[p*i+c]++}}compute(){let e=this._numClasses,t=[];for(let n=0;n<e;n++){let s=new Array(e);for(let o=0;o<e;o++)s[o]=this._matrix[n*e+o];t.push(s)}return t}reset(){super.reset(),this._matrix.fill(0)}};function G3(r,e,t){let n=0,s=r[e*t];for(let o=1;o<t;o++){let i=r[e*t+o];i>s&&(s=i,n=o)}return n}l(G3,"argmaxRow");be();var mI={};nt(mI,{cholesky:()=>Dx,cov:()=>K3,det:()=>W3,eigh:()=>j3,inv:()=>q3,lstsq:()=>Zf,pinv:()=>U3,qr:()=>Cx,solve:()=>Jf,svd:()=>Lx});be();function Lx(r){let[e,t,n]=ot("svd",r);return{U:e,S:t,V:n}}l(Lx,"svd");function j3(r){let[e,t]=ot("eigh",r);return{values:e,vectors:t}}l(j3,"eigh");function Cx(r){let[e,t]=ot("qr",r);return{Q:e,R:t}}l(Cx,"qr");var Dx=l(r=>ot("cholesky",r),"cholesky"),q3=l(r=>ot("inv",r),"inv"),U3=l(r=>ot("pinv",r),"pinv"),K3=l(r=>ot("cov",r),"cov"),Jf=l((r,e)=>ot("solve",r,e),"solve"),Zf=l((r,e)=>ot("lstsq",r,e),"lstsq"),W3=l(r=>ot("det",r).item(),"det");var bI={};nt(bI,{DecisionTreeClassifier:()=>Qx,DecisionTreeRegressor:()=>Zx,ElasticNet:()=>A_,GaussianNB:()=>Jx,GradientBoostingClassifier:()=>nS,GradientBoostingRegressor:()=>tS,GridSearchCV:()=>sS,KFold:()=>I_,KMeans:()=>Wx,KNeighborsClassifier:()=>Hx,KNeighborsRegressor:()=>Yx,LabelEncoder:()=>$x,Lasso:()=>jx,LinearRegression:()=>Vx,LogisticRegression:()=>qx,MinMaxScaler:()=>zx,OneHotEncoder:()=>Fx,PCA:()=>Kx,RandomForestClassifier:()=>eS,RandomForestRegressor:()=>Xx,Ridge:()=>Gx,StandardScaler:()=>Px,TimeSeriesSplit:()=>rS,accuracy_score:()=>Ps,confusion_matrix:()=>J3,cross_val_score:()=>_I,mean_absolute_error:()=>Y3,mean_squared_error:()=>H3,r2_score:()=>Ds,train_test_split:()=>X3});var Px=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=tl(e);this.mean_=new Float64Array(t.cols),this.scale_=new Float64Array(t.cols);for(let n=0;n<t.cols;n++){let s=0;for(let a=0;a<t.rows;a++)s+=t.data[a*t.cols+n];s/=t.rows;let o=0;for(let a=0;a<t.rows;a++){let u=t.data[a*t.cols+n]-s;o+=u*u}o/=t.rows;let i=Math.sqrt(o);this.mean_[n]=this.withMean?s:0,this.scale_[n]=this.withStd&&i>0?i:1}return this._cols=t.cols,this}transform(e){let t=tl(e),n=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]-n[a])/s[a];return np(o,t.rows,t.cols,e.dtype)}fit_transform(e){return this.fit(e).transform(e)}inverse_transform(e){let t=tl(e),n=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]+n[a];return np(o,t.rows,t.cols,e.dtype)}},$x=class{static{l(this,"LabelEncoder")}classes_;_lookup;constructor(){this.classes_=null,this._lookup=null}fit(e){let t=zt(e),{classes:n}=rs(t.data,t.n);return this.classes_=n,this._lookup=new Map(n.map((s,o)=>[s,o])),this}transform(e){let t=zt(e),n=this._lookup,s=new Float64Array(t.n);for(let o=0;o<t.n;o++){let i=n.get(t.data[o]);if(i===void 0)throw new Error(`LabelEncoder: unseen label ${t.data[o]}`);s[o]=i}return B0(s,t.n,e.dtype)}fit_transform(e){let t=zt(e),{y:n,classes:s}=rs(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]=n[i];return B0(o,t.n,e.dtype)}inverse_transform(e){let t=zt(e),n=this.classes_;return Array.from({length:t.n},(s,o)=>n[Math.round(t.data[o])])}},Fx=class{static{l(this,"OneHotEncoder")}classes_;constructor(){this.classes_=null}fit(e){let t=zt(e);return this.classes_=rs(t.data,t.n).classes,this}transform(e){let t=zt(e),n=this.classes_,s=new Map(n.map((a,u)=>[a,u])),o=n.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 np(i,t.n,o,e.dtype)}fit_transform(e){return this.fit(e).transform(e)}},zx=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=tl(e);this.dataMin_=new Float64Array(t.cols),this.dataRange_=new Float64Array(t.cols);for(let n=0;n<t.cols;n++){let s=1/0,o=-1/0;for(let i=0;i<t.rows;i++){let a=t.data[i*t.cols+n];a<s&&(s=a),a>o&&(o=a)}this.dataMin_[n]=s,this.dataRange_[n]=o>s?o-s:1}return this}transform(e){let t=tl(e),[n,s]=this.featureRange,o=this.dataMin_,i=this.dataRange_,a=s-n,u=new Float64Array(t.rows*t.cols);for(let p=0;p<t.rows;p++)for(let c=0;c<t.cols;c++){let h=(t.data[p*t.cols+c]-o[c])/i[c];u[p*t.cols+c]=h*a+n}return np(u,t.rows,t.cols,e.dtype)}fit_transform(e){return this.fit(e).transform(e)}};function H3(r,e){let t=zt(r),n=zt(e),s=0;for(let o=0;o<t.n;o++){let i=t.data[o]-n.data[o];s+=i*i}return s/t.n}l(H3,"mean_squared_error");function Y3(r,e){let t=zt(r),n=zt(e),s=0;for(let o=0;o<t.n;o++)s+=Math.abs(t.data[o]-n.data[o]);return s/t.n}l(Y3,"mean_absolute_error");function Ds(r,e){let t=zt(r),n=zt(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]-n.data[a],p=t.data[a]-s;o+=u*u,i+=p*p}return i===0?0:1-o/i}l(Ds,"r2_score");function Ps(r,e){let t=zt(r),n=zt(e),s=0;for(let o=0;o<t.n;o++)t.data[o]===n.data[o]&&s++;return s/t.n}l(Ps,"accuracy_score");function J3(r,e){let t=zt(r),n=zt(e),s=0;for(let a=0;a<t.n;a++)s=Math.max(s,t.data[a],n.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]][n.data[a]]++;return i}l(J3,"confusion_matrix");be();_t();qr();function gI(r,e){return e?sa([r,Ht([r.shape[0],1],{device:r.device,dtype:r.dtype})],1):r}l(gI,"addBias");function Ux(r){return r.ndim===1?r.reshape([r.shape[0],1]):r}l(Ux,"asColumn");var Vx=class{static{l(this,"LinearRegression")}fitIntercept;weight_;constructor({fitIntercept:e=!0}={}){this.fitIntercept=e,this.weight_=null}fit(e,t){let n=gI(e,this.fitIntercept);return this.weight_=Zf(n,Ux(t)),this}predict(e){let t=gI(e,this.fitIntercept);return lt(t,this.weight_).reshape([e.shape[0]])}score(e,t){return Ds(t,this.predict(e))}},Gx=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 n=e.shape[1],s=Ux(t),o=e,i=s,a=null,u=null;this.fitIntercept&&(a=lr(e,[0],!0),u=lr(s,[0],!0),o=Pe(e,a),i=Pe(s,u));let p=o.transpose(0,1),c=ae(lt(p,o),ee(Rm(n,n,{device:e.device,dtype:e.dtype}),this.alpha));return this.coef_=Jf(c,lt(p,i)),this.intercept_=this.fitIntercept?Pe(u,lt(a,this.coef_)):$e([1,1],{device:e.device,dtype:e.dtype}),this}predict(e){return ae(lt(e,this.coef_),this.intercept_).reshape([e.shape[0]])}score(e,t){return Ds(t,this.predict(e))}},A_=class{static{l(this,"ElasticNet")}alpha;l1Ratio;fitIntercept;maxIter;tol;coef_;intercept_;constructor({alpha:e=1,l1Ratio:t=.5,fitIntercept:n=!0,maxIter:s=1e3,tol:o=1e-6}={}){this.alpha=e,this.l1Ratio=t,this.fitIntercept=n,this.maxIter=s,this.tol=o,this.coef_=null,this.intercept_=null}fit(e,t){let[n,s]=ot("elastic_net",e,Ux(t).reshape([e.shape[0]]),this.alpha,this.l1Ratio,this.maxIter,this.tol,this.fitIntercept);return this.coef_=n,this.intercept_=s,this}predict(e){let t=this.coef_.reshape([e.shape[1],1]);return ae(lt(e,t).reshape([e.shape[0]]),this.intercept_)}score(e,t){return Ds(t,this.predict(e))}},jx=class extends A_{static{l(this,"Lasso")}constructor({alpha:e=1,fitIntercept:t=!0,maxIter:n=1e3,tol:s=1e-6}={}){super({alpha:e,l1Ratio:1,fitIntercept:t,maxIter:n,tol:s})}},qx=class{static{l(this,"LogisticRegression")}C;lr;maxIter;W_;b_;classes_;constructor({C:e=1,lr:t=.5,maxIter:n=1e3}={}){this.C=e,this.lr=t,this.maxIter=n,this.W_=null,this.b_=null,this.classes_=null}fit(e,t){let n=e.shape[0],s=e.shape[1],o=zt(t),{y:i,classes:a}=rs(o.data,o.n);this.classes_=a;let u=a.length,p=new Float64Array(n*u);for(let y=0;y<n;y++)p[y*u+i[y]]=1;let c=Pt(p,{shape:[n,u],dtype:e.dtype,device:e.device}),h=$e([s,u],{device:e.device,dtype:e.dtype}),f=$e([1,u],{device:e.device,dtype:e.dtype}),d=1/this.C,m=e.transpose(0,1),g=this.lr/n;for(let y=0;y<this.maxIter;y++){let _=ae(lt(e,h),f),b=Pe(_,Qs(_,1,!0)),w=Kr(b),T=$t(w,ht(w,[1],!0)),S=Pe(T,c),x=ae(lt(m,S),ee(h,d)),k=ht(S,[0],!0);h=Pe(h,ee(x,g)),f=Pe(f,ee(k,g))}return this.W_=h,this.b_=f,this}decisionLogits(e){return ae(lt(e,this.W_),this.b_)}predict(e){let t=Ud(this.decisionLogits(e),1,!1).toArray(),n=new Float64Array(t.length),s=this.classes_;for(let o=0;o<t.length;o++)n[o]=s[t[o]];return Pt(n,{shape:[t.length],dtype:e.dtype,device:e.device})}score(e,t){return Ps(t,this.predict(e))}};be();var Kx=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],n=e.shape[1];this.mean_=lr(e,[0],!0);let s=Pe(e,this.mean_),{S:o,V:i}=Lx(s),a=i.shape[1],u=Math.min(this.nComponents??Math.min(t,n),a);this.components_=i.narrow(1,0,u),this._nc=u;let p=o.toArray(),c=t>1?t-1:1,h=0;for(let f=0;f<p.length;f++)h+=p[f]*p[f]/c;this.explainedVariance_=new Array(u),this.explainedVarianceRatio_=new Array(u);for(let f=0;f<u;f++){let d=p[f]*p[f]/c;this.explainedVariance_[f]=d,this.explainedVarianceRatio_[f]=h>0?d/h:0}return this}transform(e){return lt(Pe(e,this.mean_),this.components_)}fit_transform(e){return this.fit(e).transform(e)}inverse_transform(e){return ae(lt(e,this.components_.transpose(0,1)),this.mean_)}};be();var Wx=class{static{l(this,"KMeans")}nClusters;maxIter;nInit;randomState;clusterCenters_;labels_;inertia_;constructor({nClusters:e=8,maxIter:t=300,nInit:n=10,randomState:s=0}={}){this.nClusters=e,this.maxIter=t,this.nInit=n,this.randomState=s,this.clusterCenters_=null,this.labels_=null,this.inertia_=null}fit(e){let[t,n,s]=ot("kmeans",e,this.nClusters,this.maxIter,this.nInit,this.randomState);return this.clusterCenters_=t,this.labels_=n,this.inertia_=Number(s.item()),this}predict(e){return ot("kmeans_predict",e,this.clusterCenters_)}fit_predict(e){return this.fit(e),this.labels_}};be();var O_=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 ot("knn_predict",this._X,this._y,e,this.nNeighbors,this._classify)}},Hx=class extends O_{static{l(this,"KNeighborsClassifier")}constructor({nNeighbors:e=5}={}){super(e,!0)}score(e,t){return Ps(t,this.predict(e))}},Yx=class extends O_{static{l(this,"KNeighborsRegressor")}constructor({nNeighbors:e=5}={}){super(e,!1)}score(e,t){return Ds(t,this.predict(e))}};be();var Jx=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[n,s,o,i]=ot("gaussian_nb_fit",e,t);return this.means_=n,this.variances_=s,this.priors_=o,this.classes_=i,this}predict(e){return ot("gaussian_nb_predict",e,this.means_,this.variances_,this.priors_,this.classes_)}score(e,t){return Ps(t,this.predict(e))}};be();qr();var yI=1<<30;function M_(r,e,t,n,s){return ot("decision_tree_fit",r,e,t.maxDepth,t.minSamplesSplit,t.minSamplesLeaf,t.maxFeatures,n,s)}l(M_,"fitTree");function Yl(r,e){return ot("decision_tree_predict",r,e[0],e[1],e[2],e[3],e[4])}l(Yl,"treePredict");var B_=class{static{l(this,"BaseTree")}maxDepth;minSamplesSplit;minSamplesLeaf;maxFeatures;randomState;_classify;_nodes;constructor(e,t){this.maxDepth=e.maxDepth??yI,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=M_(e,t,this,this._classify,this.randomState),this}predict(e){return Yl(e,this._nodes)}},Zx=class extends B_{static{l(this,"DecisionTreeRegressor")}constructor(e={}){super(e,!1)}score(e,t){return Ds(t,this.predict(e))}},Qx=class extends B_{static{l(this,"DecisionTreeClassifier")}constructor(e={}){super(e,!0)}score(e,t){return Ps(t,this.predict(e))}};function Z3(r,e,t){return r>0?r:t?Math.max(1,Math.floor(Math.sqrt(e))):Math.max(1,Math.floor(e/3))}l(Z3,"defaultMaxFeatures");var R_=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??yI,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 n=e.shape[0],s=e.shape[1],o=cr(this.randomState),i=Z3(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(n);for(let f=0;f<n;f++)p[f]=Math.floor(o()*n);let c=ns(e,p),h=ns(t,p);this._trees.push(M_(c,h,a,this._classify,this.randomState+u+1))}return this}predict(e){let t=e.shape[0],n=this._trees.map(o=>Yl(e,o).toArray()),s=new Float64Array(t);if(this._classify)for(let o=0;o<t;o++){let i=new Map,a=n[0][o],u=0;for(let p=0;p<n.length;p++){let c=n[p][o],h=(i.get(c)||0)+1;i.set(c,h),h>u&&(u=h,a=c)}s[o]=a}else for(let o=0;o<t;o++){let i=0;for(let a=0;a<n.length;a++)i+=n[a][o];s[o]=i/n.length}return Pt(s,{shape:[t],dtype:e.dtype,device:e.device})}},Xx=class extends R_{static{l(this,"RandomForestRegressor")}constructor(e={}){super(e,!1)}score(e,t){return Ds(t,this.predict(e))}},eS=class extends R_{static{l(this,"RandomForestClassifier")}constructor(e={}){super(e,!0)}score(e,t){return Ps(t,this.predict(e))}},tS=class{static{l(this,"GradientBoostingRegressor")}nEstimators;learningRate;params;randomState;init_;_trees;constructor({nEstimators:e=100,learningRate:t=.1,maxDepth:n=3,minSamplesSplit:s=2,minSamplesLeaf:o=1,randomState:i=0}={}){this.nEstimators=e,this.learningRate=t,this.params={maxDepth:n,minSamplesSplit:s,minSamplesLeaf:o,maxFeatures:0},this.randomState=i,this.init_=0,this._trees=[]}fit(e,t){let n=zt(t),s=n.n,o=0;for(let a=0;a<s;a++)o+=n.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 f=0;f<s;f++)u[f]=n.data[f]-i[f];let p=Pt(u,{shape:[s],dtype:e.dtype,device:e.device}),c=M_(e,p,this.params,!1,this.randomState+a+1);this._trees.push(c);let h=Yl(e,c).toArray();for(let f=0;f<s;f++)i[f]+=this.learningRate*h[f]}return this}predict(e){let t=e.shape[0],n=new Float64Array(t).fill(this.init_);for(let s of this._trees){let o=Yl(e,s).toArray();for(let i=0;i<t;i++)n[i]+=this.learningRate*o[i]}return Pt(n,{shape:[t],dtype:e.dtype,device:e.device})}score(e,t){return Ds(t,this.predict(e))}},nS=class{static{l(this,"GradientBoostingClassifier")}nEstimators;learningRate;params;randomState;classes_;_stages;constructor({nEstimators:e=100,learningRate:t=.1,maxDepth:n=3,minSamplesSplit:s=2,minSamplesLeaf:o=1,randomState:i=0}={}){this.nEstimators=e,this.learningRate=t,this.params={maxDepth:n,minSamplesSplit:s,minSamplesLeaf:o,maxFeatures:0},this.randomState=i,this.classes_=null,this._stages=[]}fit(e,t){let n=zt(t),s=n.n,{y:o,classes:i}=rs(n.data,n.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=Q3(u,s,a),h=[];for(let f=0;f<a;f++){let d=new Float64Array(s);for(let _=0;_<s;_++)d[_]=(o[_]===f?1:0)-c[_*a+f];let m=Pt(d,{shape:[s],dtype:e.dtype,device:e.device}),g=M_(e,m,this.params,!1,this.randomState+p*a+f+1);h.push(g);let y=Yl(e,g).toArray();for(let _=0;_<s;_++)u[_*a+f]+=this.learningRate*y[_]}this._stages.push(h)}return this}predict(e){let t=e.shape[0],n=this.classes_,s=n.length,o=new Float64Array(t*s);for(let a of this._stages)for(let u=0;u<s;u++){let p=Yl(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]=n[u]}return Pt(i,{shape:[t],dtype:e.dtype,device:e.device})}score(e,t){return Ps(t,this.predict(e))}};function Q3(r,e,t){let n=new Float64Array(e*t);for(let s=0;s<e;s++){let o=-1/0;for(let a=0;a<t;a++)r[s*t+a]>o&&(o=r[s*t+a]);let i=0;for(let a=0;a<t;a++){let u=Math.exp(r[s*t+a]-o);n[s*t+a]=u,i+=u}for(let a=0;a<t;a++)n[s*t+a]/=i}return n}l(Q3,"softmaxRows");function X3(r,e,{testSize:t=.25,shuffle:n=!0,randomState:s=0}={}){let o=r.shape[0],i=n?S0(o,cr(s)):Array.from({length:o},(c,h)=>h),a=Math.max(1,Math.round(o*t)),u=i.slice(0,a),p=i.slice(a);return[ns(r,p),ns(r,u),ns(e,p),ns(e,u)]}l(X3,"train_test_split");var I_=class{static{l(this,"KFold")}nSplits;shuffle;randomState;constructor({nSplits:e=5,shuffle:t=!1,randomState:n=0}={}){this.nSplits=e,this.shuffle=t,this.randomState=n}split(e){let t=this.shuffle?S0(e,cr(this.randomState)):Array.from({length:e},(a,u)=>u),n=[],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));n.push({train:c,test:p}),i+=u}return n}},rS=class{static{l(this,"TimeSeriesSplit")}nSplits;constructor({nSplits:e=5}={}){this.nSplits=e}split(e){let t=Math.floor(e/(this.nSplits+1)),n=[];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);n.push({train:a,test:u})}return n}};function _I(r,e,t,{cv:n=5,scoring:s=null,shuffle:o=!0,randomState:i=0}={}){let a=new I_({nSplits:n,shuffle:o,randomState:i}).split(e.shape[0]),u=[];for(let{train:p,test:c}of a){let h=ns(e,p),f=ns(t,p),d=ns(e,c),m=ns(t,c),g=r().fit(h,f);u.push(s?s(m,g.predict(d)):g.score(d,m))}return u}l(_I,"cross_val_score");function eV(r){let e=Object.keys(r),t=[{}];for(let n of e){let s=[];for(let o of t)for(let i of r[n])s.push({...o,[n]:i});t=s}return t}l(eV,"cartesian");var sS=class{static{l(this,"GridSearchCV")}makeEstimator;paramGrid;cv;scoring;bestParams_;bestScore_;bestEstimator_;constructor(e,t,{cv:n=5,scoring:s=null}={}){this.makeEstimator=e,this.paramGrid=t,this.cv=n,this.scoring=s,this.bestParams_=null,this.bestScore_=-1/0,this.bestEstimator_=null}fit(e,t){for(let n of eV(this.paramGrid)){let s=_I(()=>this.makeEstimator(n),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_=n)}return this.bestEstimator_=this.makeEstimator(this.bestParams_).fit(e,t),this}predict(e){return this.bestEstimator_.predict(e)}};var AE={};nt(AE,{Generator:()=>q_,acf:()=>iE,andersonDarling:()=>rE,betaRegularized:()=>Xf,bisect:()=>wI,brentq:()=>E_,chi2:()=>Ql,chi2Gof:()=>JI,chi2Independence:()=>ZI,constrainedMinimize:()=>dS,convolve:()=>gE,correlate:()=>yE,cubicSpline:()=>UI,dagostinoK2:()=>nE,differentialEvolution:()=>cS,digammaScalar:()=>Ac,durbinWatson:()=>lE,erfScalar:()=>bu,erfcScalar:()=>wu,fft:()=>$a,fisherF:()=>II,gammaScalar:()=>Tu,ifft:()=>Xl,jarqueBera:()=>tE,ksTest1Samp:()=>XI,ksTest2Samp:()=>eE,lbfgs:()=>th,lbfgsB:()=>fS,levenbergMarquardt:()=>hS,lgammaScalar:()=>Tn,linearInterp:()=>qI,ljungBox:()=>uE,lowerGammaRegularized:()=>C_,mannWhitneyU:()=>sE,nelderMead:()=>uS,newton:()=>TI,normal:()=>MI,normalCdfScalar:()=>po,normalPdfScalar:()=>eh,normalPpfScalar:()=>Bi,pacf:()=>aE,periodogram:()=>cE,polyfit:()=>kE,polyroots:()=>NE,polyval:()=>vE,qr:()=>Cx,quadrature:()=>VI,rollingMax:()=>SE,rollingMean:()=>bE,rollingMin:()=>xE,rollingStd:()=>wE,rollingSum:()=>_E,simpson:()=>zI,studentT:()=>D_,tTest1Samp:()=>wS,tTestInd:()=>HI,tTestPaired:()=>YI,trapezoid:()=>FI});jn();Kn();function cs(r,e,t){return{root:r,iterations:e,converged:t}}l(cs,"result");function wI(r,e,t,n={}){let s=n.tol??1e-12,o=n.maxIter??100,i=r(e),a=r(t);if(i===0)return cs(e,0,!0);if(a===0)return cs(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 h=1;h<=o;h++){let f=.5*(u+p),d=r(f);if(d===0||.5*(p-u)<s)return cs(f,h,!0);c*d<0?p=f:(u=f,c=d)}return cs(.5*(u+p),o,!1)}l(wI,"bisect");function TI(r,e,t={}){let n=t.tol??1e-12,s=t.maxIter??100,o=t.step??1e-7,i=t.derivative??(u=>(r(u+o)-r(u-o))/(2*o)),a=e;for(let u=1;u<=s;u++){let p=r(a);if(Math.abs(p)<n)return cs(a,u,!0);let c=i(a);if(c===0)break;let h=a-p/c;if(Math.abs(h-a)<n)return cs(h,u,!0);a=h}return cs(a,s,!1)}l(TI,"newton");function E_(r,e,t,n={}){let s=n.tol??1e-12,o=n.maxIter??100,i=r(e),a=r(t);if(i===0)return cs(e,0,!0);if(a===0)return cs(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,h=!0;for(let f=1;f<=o;f++){let d;i!==p&&a!==p?d=e*a*p/((i-a)*(i-p))+t*i*p/((a-i)*(a-p))+u*i*a/((p-i)*(p-a)):d=t-a*(t-e)/(a-i);let m=.25*(3*e+t),g=!(d>Math.min(m,t)&&d<Math.max(m,t)),y=h&&Math.abs(d-t)>=.5*Math.abs(t-u),_=!h&&Math.abs(d-t)>=.5*Math.abs(u-c),b=h&&Math.abs(t-u)<s,w=!h&&Math.abs(u-c)<s;g||y||_||b||w?(d=.5*(e+t),h=!0):h=!1;let T=r(d);if(c=u,u=t,p=a,i*T<0?(t=d,a=T):(e=d,i=T),Math.abs(i)<Math.abs(a)&&([e,t]=[t,e],[i,a]=[a,i]),T===0)return cs(d,f,!0);if(Math.abs(a)<s||Math.abs(t-e)<s)return cs(t,f,!0)}return cs(t,o,!1)}l(E_,"brentq");Ji();var Jl=[-39.69683028665376,220.9460984245205,-275.9285104469687,138.357751867269,-30.66479806614716,2.506628277459239],Qf=[-54.47609879822406,161.5858368580409,-155.6989798598866,66.80131188771972,-13.28068155288572],Zl=[-.007784894002430293,-.3223964580411365,-2.400758277161838,-2.549732539343734,4.374664141464968,2.938163982698783],L_=[.007784695709041462,.3224671290700398,2.445134137142996,3.754408661907416],xI=.02425,oS=300,iS=3e-14,gr=1e-300,tV=2,nV=Math.SQRT2,rV=Math.sqrt(2*Math.PI),ige=1/Math.sqrt(Math.PI);function C_(r,e){if(e<=0)return 0;if(e<r+1){let i=1/r,a=i;for(let u=1;u<oS&&(i*=e/(r+u),a+=i,!(Math.abs(i)<Math.abs(a)*iS));u++);return a*Math.exp(-e+r*Math.log(e)-Tn(r))}let t=e+1-r,n=1/gr,s=1/t,o=s;for(let i=1;i<oS;i++){let a=-i*(i-r);t+=2,s=a*s+t,Math.abs(s)<gr&&(s=gr),n=t+a/n,Math.abs(n)<gr&&(n=gr),s=1/s;let u=s*n;if(o*=u,Math.abs(u-1)<iS)break}return 1-o*Math.exp(-e+r*Math.log(e)-Tn(r))}l(C_,"lowerGammaRegularized");function SI(r,e,t){let n=r+e,s=r+1,o=r-1,i=1,a=1-n*t/s;Math.abs(a)<gr&&(a=gr),a=1/a;let u=a;for(let p=1;p<oS;p++){let c=2*p,h=p*(e-p)*t/((o+c)*(r+c));a=1+h*a,Math.abs(a)<gr&&(a=gr),i=1+h/i,Math.abs(i)<gr&&(i=gr),a=1/a,u*=a*i,h=-(r+p)*(n+p)*t/((r+c)*(s+c)),a=1+h*a,Math.abs(a)<gr&&(a=gr),i=1+h/i,Math.abs(i)<gr&&(i=gr),a=1/a;let f=a*i;if(u*=f,Math.abs(f-1)<iS)break}return u}l(SI,"betacf");function Xf(r,e,t){if(t<=0)return 0;if(t>=1)return 1;let n=Math.exp(Tn(r+e)-Tn(r)-Tn(e)+r*Math.log(t)+e*Math.log(1-t));return t<(r+1)/(r+e+2)?n*SI(r,e,t)/r:1-n*SI(e,r,1-t)/e}l(Xf,"betaRegularized");function po(r){return .5*wu(-r/nV)}l(po,"normalCdfScalar");function eh(r){return Math.exp(-.5*r*r)/rV}l(eh,"normalPdfScalar");function kI(r){if(r<xI){let n=Math.sqrt(-2*Math.log(r));return(((((Zl[0]*n+Zl[1])*n+Zl[2])*n+Zl[3])*n+Zl[4])*n+Zl[5])/((((L_[0]*n+L_[1])*n+L_[2])*n+L_[3])*n+1)}if(r>1-xI)return-kI(1-r);let e=r-.5,t=e*e;return(((((Jl[0]*t+Jl[1])*t+Jl[2])*t+Jl[3])*t+Jl[4])*t+Jl[5])*e/(((((Qf[0]*t+Qf[1])*t+Qf[2])*t+Qf[3])*t+Qf[4])*t+1)}l(kI,"ppfSeed");function Bi(r,e={}){if(r<=0)return-1/0;if(r>=1)return 1/0;let t=e.refineSteps??tV,n=kI(r);for(let s=0;s<t;s++){let i=(po(n)-r)/eh(n);n-=i/(1+n*i/2)}return n}l(Bi,"normalPpfScalar");var sV=3,oV=2,iV=60,aV=1e-11;function ps(r,e){if(r instanceof We){let t=Je(r),n=new Float64Array(t.length);for(let s=0;s<t.length;s++)n[s]=e(Number(t[s]));return Q(n,r.shape,r.dtype,r.device)}return e(r)}l(ps,"mapElementwise");function aS(r,e,t,n,s={}){if(t<=0||t>=1)return t===0?s.lowerLimit??-1/0:t===1?1/0:NaN;let o=s.refineSteps??sV,i=s.tol??aV,a=n;for(let h=0;h<o;h++){let f=r(a)-t,d=e(a);if(!Number.isFinite(a)||d<=0)break;let m=a-f/d;if(a=s.lowerLimit!==void 0&&m<=s.lowerLimit?(a+s.lowerLimit)/2:m,Math.abs(f)<i)return a}if(Math.abs(r(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 h=0;h<iV;h++){if(s.lowerLimit===void 0&&(u=a-c),p=a+c,r(u)-t<0&&r(p)-t>0)return E_(f=>r(f)-t,u,p,{tol:i}).root;c*=oV}return a}l(aS,"invertCdf");var MI={cdf:l((r,e={})=>{let t=e.loc??0,n=e.scale??1;return ps(r,s=>po((s-t)/n))},"cdf"),pdf:l((r,e={})=>{let t=e.loc??0,n=e.scale??1;return ps(r,s=>eh((s-t)/n)/n)},"pdf"),ppf:l((r,e={})=>{let t=e.loc??0,n=e.scale??1;return ps(r,s=>t+n*Bi(s,e))},"ppf")};function vI(r,e){if(r===0)return .5;let t=Xf(e/2,.5,e/(e+r*r));return r>0?1-t/2:t/2}l(vI,"tCdfScalar");function NI(r,e){let t=Tn((e+1)/2)-Tn(e/2)-.5*Math.log(e*Math.PI);return Math.exp(t-(e+1)/2*Math.log(1+r*r/e))}l(NI,"tPdfScalar");var D_={cdf:l((r,e,t={})=>ps(r,n=>vI(n,e??t.df)),"cdf"),pdf:l((r,e,t={})=>ps(r,n=>NI(n,e??t.df)),"pdf"),ppf:l((r,e,t={})=>{let n=e??t.df;return ps(r,s=>aS(o=>vI(o,n),o=>NI(o,n),s,Bi(s),t))},"ppf")};function AI(r,e){return r<=0?0:C_(e/2,r/2)}l(AI,"chi2CdfScalar");function OI(r,e){if(r<=0)return 0;let t=e/2;return Math.exp((t-1)*Math.log(r)-r/2-t*Math.log(2)-Tn(t))}l(OI,"chi2PdfScalar");function uV(r,e){let t=Bi(r),n=2/(9*e),s=1-n+t*Math.sqrt(n);return Math.max(e*s*s*s,Number.EPSILON)}l(uV,"chi2Seed");var Ql={cdf:l((r,e,t={})=>ps(r,n=>AI(n,e??t.df)),"cdf"),pdf:l((r,e,t={})=>ps(r,n=>OI(n,e??t.df)),"pdf"),ppf:l((r,e,t={})=>{let n=e??t.df;return ps(r,s=>aS(o=>AI(o,n),o=>OI(o,n),s,uV(s,n),{...t,lowerLimit:0}))},"ppf")};function BI(r,e,t){return r<=0?0:Xf(e/2,t/2,e*r/(e*r+t))}l(BI,"fCdfScalar");function RI(r,e,t){if(r<=0)return 0;let n=Tn(e/2)+Tn(t/2)-Tn((e+t)/2);return Math.exp(e/2*Math.log(e/t)+(e/2-1)*Math.log(r)-(e+t)/2*Math.log(1+e/t*r)-n)}l(RI,"fPdfScalar");var II={cdf:l((r,e,t,n={})=>ps(r,s=>BI(s,e??n.d1,t??n.d2)),"cdf"),pdf:l((r,e,t,n={})=>ps(r,s=>RI(s,e??n.d1,t??n.d2)),"pdf"),ppf:l((r,e,t,n={})=>{let s=e??n.d1,o=t??n.d2;return ps(r,i=>aS(a=>BI(a,s,o),a=>RI(a,s,o),i,1,{...n,lowerLimit:0}))},"ppf")};function uS(r,e,t={}){let n=e.length,s=t.maxIter??200*n,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,h=t.sigma??.5,f=[e.slice()];for(let _=0;_<n;_++){let b=e.slice();b[_]=b[_]!==0?b[_]*(1+i):a,f.push(b)}let d=f.map(r),m=Array.from({length:n+1},(_,b)=>b),g=0;for(;g<s;g++){m.sort((k,A)=>d[k]-d[A]);let _=m.map(k=>f[k]),b=m.map(k=>d[k]);for(let k=0;k<=n;k++)f[k]=_[k],d[k]=b[k];if(Math.abs(d[n]-d[0])<o)break;let w=new Array(n).fill(0);for(let k=0;k<n;k++)for(let A=0;A<n;A++)w[A]+=f[k][A];for(let k=0;k<n;k++)w[k]/=n;let T=f[n],S=w.map((k,A)=>k+u*(k-T[A])),x=r(S);if(x<d[0]){let k=w.map((N,v)=>N+p*(S[v]-N)),A=r(k);A<x?(f[n]=k,d[n]=A):(f[n]=S,d[n]=x)}else if(x<d[n-1])f[n]=S,d[n]=x;else{let k=w.map((N,v)=>N+c*(T[v]-N)),A=r(k);if(A<d[n])f[n]=k,d[n]=A;else{let N=f[0];for(let v=1;v<=n;v++)f[v]=N.map((O,R)=>O+h*(f[v][R]-O)),d[v]=r(f[v])}}}let y=0;for(let _=1;_<=n;_++)d[_]<d[y]&&(y=_);return{point:f[y],value:d[y],iterations:g,converged:g<s}}l(uS,"nelderMead");function fs(r,e){let t=0;for(let n=0;n<r.length;n++)t+=r[n]*e[n];return t}l(fs,"dot");function Da(r){return Math.sqrt(fs(r,r))}l(Da,"norm");function P_(r,e){let t=new Array(r.length);for(let n=0;n<r.length;n++)t[n]=r[n]-e[n];return t}l(P_,"sub");function EI(r,e){let t=new Array(r.length);for(let n=0;n<r.length;n++)t[n]=r[n]*e;return t}l(EI,"scale");function LI(r){let e=0;for(let t=1;t<r.length;t++)r[t]<r[e]&&(e=t);return e}l(LI,"argmin");function CI(r,e,t){let n=e.length,s=new Array(n),o=e.slice();for(let i=0;i<n;i++){let a=o[i];o[i]=a+t;let u=r(o);o[i]=a-t;let p=r(o);o[i]=a,s[i]=(u-p)/(2*t)}return s}l(CI,"numericGradient");function Pa(r,e){return e?r.map((t,n)=>{let s=e[n];if(!s)return t;let[o,i]=s;return t<o?o:t>i?i:t}):r.slice()}l(Pa,"clampToBounds");var lV=10,cV=5,pV=.8,fV=.9,hV=1e3,dV=1e-10;function lS(r,e,t){let n;do n=el(r,e);while(t.includes(n));return n}l(lS,"pickDistinct");function cS(r,e,t={}){let n=e.length,s=cr(t.seed),o=t.populationSize??Math.max(cV,lV*n),i=t.mutation??pV,a=t.recombination??fV,u=t.maxIter??hV,p=t.tol??dV,c=[];for(let m=0;m<o;m++){let g=new Array(n);for(let y=0;y<n;y++){let[_,b]=e[y];g[y]=_+s()*(b-_)}c.push(g)}let h=c.map(r),f=LI(h),d=0;for(;d<u;d++){for(let y=0;y<o;y++){let _=lS(s,o,[y]),b=lS(s,o,[y,_]),w=lS(s,o,[y,_,b]),T=el(s,n),S=c[y].slice();for(let k=0;k<n;k++)if(s()<a||k===T){let[A,N]=e[k],v=c[_][k]+i*(c[b][k]-c[w][k]);v<A?v=A:v>N&&(v=N),S[k]=v}let x=r(S);x<h[y]&&(c[y]=S,h[y]=x,x<h[f]&&(f=y))}let m=h[0],g=h[0];for(let y=1;y<o;y++)h[y]<m&&(m=h[y]),h[y]>g&&(g=h[y]);if(g-m<p){d++;break}}return{point:c[f].slice(),value:h[f],iterations:d,converged:d<u}}l(cS,"differentialEvolution");var mV=10,gV=1e3,yV=1e-8,_V=1e-12,bV=1e-4,wV=.1,DI=.5,TV=40;function xV(r,e,t,n){let s=r.slice(),o=e.length,i=new Array(o);for(let a=o-1;a>=0;a--){i[a]=n[a]*fs(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=fs(a,u)/fs(u,u);for(let c=0;c<s.length;c++)s[c]*=p}for(let a=0;a<o;a++){let u=n[a]*fs(t[a],s);for(let p=0;p<s.length;p++)s[p]+=(i[a]-u)*e[a][p]}return s}l(xV,"twoLoop");function PI(r,e,t,n){let s=e.length,o=t.memory??mV,i=t.maxIter??gV,a=t.gtol??yV,u=t.ftol??_V,p=t.step??1e-7,c=t.gradient??(T=>CI(r,T,p)),h=l((T,S)=>n?T.map((x,k)=>{let A=n[k];if(!A)return x;let[N,v]=A;return S[k]<=N&&x>0||S[k]>=v&&x<0?0:x}):T,"projectGrad"),f=Pa(e,n),d=r(f),m=c(f),g=h(m,f),y=[],_=[],b=[],w=0;for(;w<i;w++){if(Da(g)<a)return{point:f,value:d,iterations:w,converged:!0};let T=EI(xV(g,y,_,b),-1);if(fs(T,g)>=0)for(let J=0;J<s;J++)T[J]=-g[J];let S=1,x=f,k=d,A=!1,N=!1,v=fs(g,T);for(let J=0;J<TV;J++){let j=f.map((G,K)=>G+S*T[K]),U=Pa(j,n),P=r(U);if(P<=d+bV*S*v){x=U,k=P,A=!0,N=n!=null&&U.some((G,K)=>G!==j[K]);break}let V=2*(P-d-v*S),he=V>0?-v*S*S/V:S*DI,le=wV*S,Ie=DI*S;S=he<le?le:he>Ie?Ie:he}if(!A)return{point:f,value:d,iterations:w,converged:!1};let O=c(x),R=h(O,x),B=P_(x,f),E=N?P_(R,g):P_(O,m),z=fs(B,E);z>Number.EPSILON*Da(B)*Da(E)&&(y.push(B),_.push(E),b.push(1/z),y.length>o&&(y.shift(),_.shift(),b.shift()));let D=Math.abs(k-d);if(f=x,d=k,m=O,g=R,D<u*(1+Math.abs(d))&&Da(B)<u*(1+Da(f)))return w++,{point:f,value:d,iterations:w,converged:!0}}return{point:f,value:d,iterations:w,converged:Da(g)<a}}l(PI,"minimize");function th(r,e,t={}){return PI(r,e,t,null)}l(th,"lbfgs");function fS(r,e,t,n={}){return PI(r,e,n,t)}l(fS,"lbfgsB");var SV=200,kV=1e-10,vV=.001,$I=10,NV=10,AV=1e12;function OV(r,e,t,n){let s=e.length,o=new Array(t);for(let a=0;a<t;a++)o[a]=new Array(s);let i=e.slice();for(let a=0;a<s;a++){let u=i[a];i[a]=u+n;let p=r(i);i[a]=u-n;let c=r(i);i[a]=u;for(let h=0;h<t;h++)o[h][a]=(p[h]-c[h])/(2*n)}return o}l(OV,"numericJacobian");function BV(r,e,t,n){let s=new Float64Array(n*n);for(let a=0;a<n;a++){for(let u=0;u<n;u++)s[a*n+u]=r[a][u];s[a*n+a]+=t*r[a][a]}let o=Float64Array.from(e),i=Jf(Q(s,[n,n],"f64"),Q(o,[n],"f64"));return Array.from(i.toArray())}l(BV,"solveNormal");function hS(r,e,t={}){let n=t.maxIter??SV,s=t.tol??kV,o=t.step??1e-7,i=t.jacobian??((m,g)=>OV(r,m,g,o)),a=e.slice(),u=r(a),p=u.length,c=a.length,h=fs(u,u),f=t.lambda??vV,d=0;for(;d<n;d++){let m=i(a,p),g=new Array(c),y=new Array(c).fill(0);for(let b=0;b<c;b++){g[b]=new Array(c).fill(0);for(let w=0;w<p;w++)y[b]+=m[w][b]*u[w]}for(let b=0;b<c;b++)for(let w=b;w<c;w++){let T=0;for(let S=0;S<p;S++)T+=m[S][b]*m[S][w];g[b][w]=T,g[w][b]=T}let _=!1;for(;f<AV;){let b;try{b=BV(g,y,f,c)}catch{f*=$I;continue}let w=a.map((x,k)=>x-b[k]),T=r(w),S=fs(T,T);if(S<h){a=w,u=T;let x=h-S;if(h=S,f/=NV,_=!0,x<s*(1+h))return{point:a,value:h,iterations:d+1,converged:!0};break}f*=$I}if(!_)return{point:a,value:h,iterations:d,converged:h<s}}return{point:a,value:h,iterations:d,converged:!1}}l(hS,"levenbergMarquardt");var RV=10,MV=10,IV=20,EV=1e-8;function dS(r,e,t={}){let n=t.bounds??null,s=t.inequalities??[],o=t.equalities??[],i=t.outerIter??IV,a=t.ctol??EV,u=t.penaltyGrowth??MV,p=t.inner??th,c=t.innerOpts??{},h=t.penalty??RV,f=Pa(e,n),d=0,m=null;for(let g=0;g<i;g++){m=p(l(b=>{let w=Pa(b,n),T=r(w);for(let S of s){let x=S(w);x>0&&(T+=h*x*x)}for(let S of o){let x=S(w);T+=h*x*x}return T},"penalized"),f,c),f=Pa(m.point,n),d+=m.iterations;let _=0;for(let b of s)_=Math.max(_,b(f));for(let b of o)_=Math.max(_,Math.abs(b(f)));if(_<a)return{point:f,value:r(f),iterations:d,converged:!0};h*=u}return{point:f,value:r(f),iterations:d,converged:!1}}l(dS,"constrainedMinimize");function FI(r,e,t,n={}){let s=n.n??1024,o=(t-e)/s,i=.5*(r(e)+r(t));for(let a=1;a<s;a++)i+=r(e+a*o);return i*o}l(FI,"trapezoid");function zI(r,e,t,n={}){let s=n.n??1024;s%2===1&&(s+=1);let o=(t-e)/s,i=r(e)+r(t);for(let a=1;a<s;a++)i+=(a%2===0?2:4)*r(e+a*o);return i*o/3}l(zI,"simpson");function mS(r,e,t,n,s){return(s-n)/6*(r+4*t+e)}l(mS,"simpsonCell");function gS(r,e,t,n,s,o,i,a,u,p){let c=.5*(e+t),h=.5*(e+c),f=.5*(c+t),d=r(h),m=r(f),g=mS(n,o,d,e,c),y=mS(o,s,m,c,t),_=g+y-i;return u>=p||Math.abs(_)<=15*a?g+y+_/15:gS(r,e,c,n,o,d,g,.5*a,u+1,p)+gS(r,c,t,o,s,m,y,.5*a,u+1,p)}l(gS,"adaptiveStep");function VI(r,e,t,n={}){let s=n.tol??1e-10,o=n.maxDepth??50,i=.5*(e+t),a=r(e),u=r(t),p=r(i),c=mS(a,u,p,e,t);return gS(r,e,t,a,u,p,c,s,0,o)}l(VI,"quadrature");function jI(r,e){let t=1,n=r.length-1;for(;t<n;){let s=t+n>>1;r[s]<e?t=s+1:n=s}return t}l(jI,"upperInterval");function GI(r,e,t){let n=r.length;if(t<=r[0])return e[0];if(t>=r[n-1])return e[n-1];let s=jI(r,t),o=(t-r[s-1])/(r[s]-r[s-1]);return e[s-1]+o*(e[s]-e[s-1])}l(GI,"evalLinear");function qI(r,e,t){return typeof t!="number"?t.map(n=>GI(r,e,n)):GI(r,e,t)}l(qI,"linearInterp");function LV(r,e,t,n){let s=e.length,o=new Array(s),i=new Array(s);o[0]=t[0]/e[0],i[0]=n[0]/e[0];for(let u=1;u<s;u++){let p=e[u]-r[u]*o[u-1];o[u]=t[u]/p,i[u]=(n[u]-r[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(LV,"thomas");function UI(r,e){let t=r.length,n=new Array(t-1);for(let c=0;c<t-1;c++)n[c]=r[c+1]-r[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]=n[c-1],o[c]=2*(n[c-1]+n[c]),i[c]=n[c],a[c]=6*((e[c+1]-e[c])/n[c]-(e[c]-e[c-1])/n[c-1]);let u=LV(s,o,i,a),p=l(c=>{if(c<=r[0])return e[0];if(c>=r[t-1])return e[t-1];let h=jI(r,c),f=n[h-1],d=(r[h]-c)/f,m=(c-r[h-1])/f;return d*e[h-1]+m*e[h]+((d*d*d-d)*u[h-1]+(m*m*m-m)*u[h])*f*f/6},"evalOne");return{xs:r,ys:e,coefficients:u,evaluate:l((c=>typeof c!="number"?c.map(p):p(c)),"evaluate")}}l(UI,"cubicSpline");be();var $a=l(r=>ot("fft",r),"fft"),Xl=l(r=>ot("ifft",r),"ifft");jn();Kn();ar();function KI(r){let e=new Float64Array(r.length);for(let t=0;t<r.length;t++)e[t]=Number(r[t]);return e}l(KI,"float64From");function Ke(r){return r instanceof We?{data:KI(Je(r)),dtype:r.dtype,device:r.device}:typeof r=="number"?{data:Float64Array.of(r),dtype:"f64",device:ks}:{data:Float64Array.from(r),dtype:"f64",device:ks}}l(Ke,"hostVector");function $_(r){if(r instanceof We){if(r.ndim!==2)throw new Error(`numeric: expected a 2-D table, got ${r.ndim}-D`);let[s,o]=r.shape;return{data:KI(Je(r)),rows:s,cols:o,dtype:r.dtype,device:r.device}}let e=r.length,t=r[0].length,n=new Float64Array(e*t);for(let s=0;s<e;s++)for(let o=0;o<t;o++)n[s*t+o]=r[s][o];return{data:n,rows:e,cols:t,dtype:"f64",device:ks}}l($_,"hostGrid");function F_(r){let e=1;for(;e<r;)e<<=1;return e}l(F_,"nextPow2");var CV=100,DV=1e-16,PV=.12,$V=.11,WI=[.75,2.25],yS=[.2,.34,.6],FV=[-13.436,101.14,-223.73],zV=[-8.318,42.796,-59.938],VV=[.9177,-4.279,-1.38],GV=[1.2937,-5.709,.0186];function V_(r){let e=r.length,t=0;for(let o=0;o<e;o++)t+=r[o];let n=t/e,s=0;for(let o=0;o<e;o++){let i=r[o]-n;s+=i*i}return{n:e,mean:n,variance:s/(e-1)}}l(V_,"summary");function bS(r){let e=r.length,t=0;for(let a=0;a<e;a++)t+=r[a];let n=t/e,s=0,o=0,i=0;for(let a=0;a<e;a++){let u=r[a]-n,p=u*u;s+=p,o+=p*u,i+=p*p}return s/=e,o/=e,i/=e,{n:e,mean:n,m2:s,skew:o/Math.pow(s,1.5),kurt:i/(s*s)}}l(bS,"centralMoments");function Ri(r){return Math.min(Math.max(r,0),1)}l(Ri,"clampP");function _S(r,e){return Ri(2*(1-D_.cdf(Math.abs(r),e)))}l(_S,"tPvalue");function G_(r,e){return Ri(1-Ql.cdf(r,e))}l(G_,"chi2Pvalue");function wS(r,e={}){let{data:t}=Ke(r),n=e.popmean??0,{n:s,mean:o,variance:i}=V_(t),a=(o-n)/Math.sqrt(i/s),u=s-1;return{statistic:a,pvalue:_S(a,u),df:u}}l(wS,"tTest1Samp");function HI(r,e,t={}){let n=V_(Ke(r).data),s=V_(Ke(e).data);if(t.equalVar??!0){let p=n.n+s.n-2,c=((n.n-1)*n.variance+(s.n-1)*s.variance)/p,h=(n.mean-s.mean)/Math.sqrt(c*(1/n.n+1/s.n));return{statistic:h,pvalue:_S(h,p),df:p}}let o=n.variance/n.n,i=s.variance/s.n,a=(n.mean-s.mean)/Math.sqrt(o+i),u=(o+i)*(o+i)/(o*o/(n.n-1)+i*i/(s.n-1));return{statistic:a,pvalue:_S(a,u),df:u}}l(HI,"tTestInd");function YI(r,e,t={}){let n=Ke(r).data,s=Ke(e).data;if(n.length!==s.length)throw new Error(`tTestPaired: samples must have equal length, got ${n.length} and ${s.length}`);let o=new Float64Array(n.length);for(let i=0;i<n.length;i++)o[i]=n[i]-s[i];return wS(o,{popmean:t.popmean??0})}l(YI,"tTestPaired");function JI(r,e,t={}){let n=Ke(r).data,s=n.length,o;if(e==null){let u=0;for(let p=0;p<s;p++)u+=n[p];o=new Float64Array(s).fill(u/s)}else o=Ke(e).data;let i=0;for(let u=0;u<s;u++){let p=n[u]-o[u];i+=p*p/o[u]}let a=s-1-(t.ddof??0);return{statistic:i,pvalue:G_(i,a),df:a}}l(JI,"chi2Gof");function ZI(r){let{data:e,rows:t,cols:n}=$_(r),s=new Float64Array(t),o=new Float64Array(n),i=0;for(let p=0;p<t;p++)for(let c=0;c<n;c++){let h=e[p*n+c];s[p]+=h,o[c]+=h,i+=h}let a=0;for(let p=0;p<t;p++)for(let c=0;c<n;c++){let h=s[p]*o[c]/i,f=e[p*n+c]-h;a+=f*f/h}let u=(t-1)*(n-1);return{statistic:a,pvalue:G_(a,u),df:u}}l(ZI,"chi2Independence");function jV(r){if(r<=0)return 1;let e=0;for(let t=1;t<=CV;t++){let n=Math.exp(-2*t*t*r*r);if(e+=t%2===1?n:-n,n<DV)break}return Ri(2*e)}l(jV,"kolmogorovSf");function QI(r,e){let t=Math.sqrt(e);return jV(r*(t+PV+$V/t))}l(QI,"ksPvalue");function XI(r,e,t={}){let{data:n}=Ke(r),s=Float64Array.from(n).sort(),o=s.length,i=t.loc??0,a=t.scale??1,u=typeof e=="function"?e:c=>po((c-i)/a),p=0;for(let c=0;c<o;c++){let h=u(s[c]);p=Math.max(p,(c+1)/o-h,h-c/o)}return{statistic:p,pvalue:QI(p,o)}}l(XI,"ksTest1Samp");function eE(r,e){let t=Float64Array.from(Ke(r).data).sort(),n=Float64Array.from(Ke(e).data).sort(),s=t.length,o=n.length,i=0,a=0,u=0;for(;i<s&&a<o;){let c=Math.min(t[i],n[a]);for(;i<s&&t[i]<=c;)i++;for(;a<o&&n[a]<=c;)a++;u=Math.max(u,Math.abs(i/s-a/o))}let p=s*o/(s+o);return{statistic:u,pvalue:QI(u,p)}}l(eE,"ksTest2Samp");function tE(r){let{n:e,skew:t,kurt:n}=bS(Ke(r).data),s=e/6*(t*t+(n-3)*(n-3)/4);return{statistic:s,pvalue:G_(s,2),df:2}}l(tE,"jarqueBera");function qV(r){let{n:e,skew:t}=bS(r),n=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=n/a;return i*Math.log(u+Math.sqrt(u*u+1))}l(qV,"skewZ");function UV(r){let{n:e,kurt:t}=bS(r),n=3*(e-1)/(e+1),s=24*e*(e-2)*(e-3)/((e+1)*(e+1)*(e+3)*(e+5)),o=(t-n)/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(UV,"kurtosisZ");function nE(r){let{data:e}=Ke(r),t=qV(e),n=UV(e),s=t*t+n*n;return{statistic:s,pvalue:G_(s,2),df:2}}l(nE,"dagostinoK2");function z_(r,e){return r[0]+r[1]*e+r[2]*e*e}l(z_,"poly2");function KV(r){return r>=yS[2]?Ri(Math.exp(z_(GV,r))):r>yS[1]?Ri(Math.exp(z_(VV,r))):r>yS[0]?Ri(1-Math.exp(z_(zV,r))):Ri(1-Math.exp(z_(FV,r)))}l(KV,"adPvalue");function rE(r){let{data:e}=Ke(r),t=Float64Array.from(e).sort(),n=t.length,{mean:s,variance:o}=V_(t),i=Math.sqrt(o),a=0;for(let c=0;c<n;c++){let h=po((t[c]-s)/i),f=po(-(t[n-1-c]-s)/i);a+=(2*c+1)*(Math.log(h)+Math.log(f))}let u=-n-a/n,p=u*(1+WI[0]/n+WI[1]/(n*n));return{statistic:u,pvalue:KV(p)}}l(rE,"andersonDarling");function WV(r){let e=r.length,t=Array.from({length:e},(i,a)=>a);t.sort((i,a)=>r[i]-r[a]);let n=new Float64Array(e),s=0,o=0;for(;o<e;){let i=o;for(;i+1<e&&r[t[i+1]]===r[t[o]];)i++;let a=(o+i+2)/2;for(let p=o;p<=i;p++)n[t[p]]=a;let u=i-o+1;s+=u*u*u-u,o=i+1}return{ranks:n,tieSum:s}}l(WV,"rankData");function sE(r,e){let t=Ke(r).data,n=Ke(e).data,s=t.length,o=n.length,i=new Float64Array(s+o);i.set(t),i.set(n,s);let{ranks:a,tieSum:u}=WV(i),p=0;for(let y=0;y<s;y++)p+=a[y];let c=p-s*(s+1)/2,h=s+o,f=s*o/2,d=Math.sqrt(s*o/12*(h+1-u/(h*(h-1)))),m=c-f,g=(m-.5*Math.sign(m))/d;return{statistic:c,pvalue:Ri(2*(1-po(Math.abs(g))))}}l(sE,"mannWhitneyU");Kn();function zo(r,e){return Number(r[e])}l(zo,"f64");function oE(r){return Math.max(1,Math.min(r-1,Math.floor(10*Math.log10(r))))}l(oE,"defaultNlags");function HV(r,e,t){let n=r.length,s=0;for(let f=0;f<n;f++)s+=r[f];let o=s/n,i=F_(2*n),a=new Float64Array(i);for(let f=0;f<n;f++)a[f]=r[f]-o;let u=Je($a(Q(a,[i],"f64",e))),p=new Float64Array(2*i);for(let f=0;f<i;f++)p[2*f]=zo(u,2*f)*zo(u,2*f)+zo(u,2*f+1)*zo(u,2*f+1);let c=Je(Xl(Q(p,[i,2],"f64",e))),h=new Float64Array(t+1);for(let f=0;f<=t;f++)h[f]=zo(c,2*f)/n;return h}l(HV,"autocovariance");function TS(r,e,t){let n=HV(r,e,t),s=new Float64Array(t+1);for(let o=0;o<=t;o++)s[o]=n[o]/n[0];return s}l(TS,"normalizedAcf");function iE(r,e={}){let{data:t,device:n,dtype:s}=Ke(r),o=e.nlags??oE(t.length),i=TS(t,n,o);return Q(i,[o+1],s,n)}l(iE,"acf");function aE(r,e={}){let{data:t,device:n,dtype:s}=Ke(r),o=e.nlags??oE(t.length),i=TS(t,n,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 h=i[c],f=1;for(let g=1;g<c;g++)h-=u[g]*i[c-g],f-=u[g]*i[g];let d=h/f;p[c]=d;for(let g=1;g<c;g++)p[g]=u[g]-d*u[c-g];a[c]=d;let m=u;u=p,p=m}return Q(a,[o+1],s,n)}l(aE,"pacf");function uE(r,e={}){let{data:t,device:n}=Ke(r),s=t.length,o=e.lags??Math.min(10,s-1),i=TS(t,n,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-Ql.cdf(a,u),0),1),df:u}}l(uE,"ljungBox");function lE(r){let{data:e}=Ke(r),t=0,n=0;for(let s=0;s<e.length;s++)if(n+=e[s]*e[s],s>0){let o=e[s]-e[s-1];t+=o*o}return t/n}l(lE,"durbinWatson");function cE(r,e={}){let{data:t,device:n,dtype:s}=Ke(r),o=t.length,i=t;if(e.detrend!==!1){let c=0;for(let f=0;f<o;f++)c+=t[f];let h=c/o;i=new Float64Array(o);for(let f=0;f<o;f++)i[f]=t[f]-h}let a=Je($a(Q(i,[o],"f64",n))),u=Math.floor(o/2),p=new Float64Array(u+1);for(let c=0;c<=u;c++)p[c]=(zo(a,2*c)*zo(a,2*c)+zo(a,2*c+1)*zo(a,2*c+1))/o;return Q(p,[u+1],s,n)}l(cE,"periodogram");Kn();var YV=200,JV=1e-13,pE=.4,fE=.9;function nh(r,e){return Number(r[e])}l(nh,"f64");function ZV(r){let e=new Float64Array(r.length);for(let t=0;t<r.length;t++)e[t]=Number(r[t]);return e}l(ZV,"float64From");function hE(r,e,t){let n=new Float64Array(e);return n.set(r),Je($a(Q(n,[e],"f64",t)))}l(hE,"fftInterleaved");function QV(r,e,t){let n=r+e-1;if(t==="full")return{offset:0,length:n};let s=Math.max(r,e),o=Math.min(r,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(QV,"modeSlice");function mE(r,e,t,n,s){let o=r.length,i=e.length,a=F_(o+i-1),u=hE(r,a,n),p=hE(e,a,n),c=new Float64Array(2*a);for(let g=0;g<a;g++){let y=nh(u,2*g),_=nh(u,2*g+1),b=nh(p,2*g),w=nh(p,2*g+1);c[2*g]=y*b-_*w,c[2*g+1]=y*w+_*b}let h=Je(Xl(Q(c,[a,2],"f64",n))),{offset:f,length:d}=QV(o,i,t),m=new Float64Array(d);for(let g=0;g<d;g++)m[g]=nh(h,2*(f+g));return Q(m,[d],s,n)}l(mE,"fftConvolve");function gE(r,e,t={}){let n=Ke(r),s=Ke(e);return mE(n.data,s.data,t.mode??"full",n.device,n.dtype)}l(gE,"convolve");function yE(r,e,t={}){let n=Ke(r),s=Ke(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 mE(n.data,o,t.mode??"full",n.device,n.dtype)}l(yE,"correlate");function j_(r,e){let t=Math.floor(r);if(!(t>=1&&t<=e))throw new Error(`rolling: window must be in [1, ${e}], got ${r}`);return t}l(j_,"checkWindow");function _E(r,e){let{data:t,device:n,dtype:s}=Ke(r),o=t.length,i=j_(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,n)}l(_E,"rollingSum");function bE(r,e){let{data:t,device:n,dtype:s}=Ke(r),o=t.length,i=j_(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,n)}l(bE,"rollingMean");function wE(r,e,t={}){let{data:n,device:s,dtype:o}=Ke(r),i=n.length,a=j_(e,i),u=t.ddof??1,p=0;for(let m=0;m<i;m++)p+=n[m];let c=p/i,h=new Float64Array(i+1),f=new Float64Array(i+1);for(let m=0;m<i;m++){let g=n[m]-c;h[m+1]=h[m]+g,f[m+1]=f[m]+g*g}let d=new Float64Array(i-a+1);for(let m=0;m<d.length;m++){let g=h[m+a]-h[m],y=f[m+a]-f[m];d[m]=Math.sqrt(Math.max((y-g*g/a)/(a-u),0))}return Q(d,[d.length],o,s)}l(wE,"rollingStd");function TE(r,e,t){let{data:n,device:s,dtype:o}=Ke(r),i=n.length,a=j_(e,i),u=new Float64Array(i-a+1),p=new Int32Array(i),c=0,h=0;for(let f=0;f<i;f++){for(;h>c&&t(n[f],n[p[h-1]]);)h--;p[h++]=f,p[c]<=f-a&&c++,f>=a-1&&(u[f-a+1]=n[p[c]])}return Q(u,[u.length],o,s)}l(TE,"rollingExtreme");function xE(r,e){return TE(r,e,(t,n)=>t<=n)}l(xE,"rollingMin");function SE(r,e){return TE(r,e,(t,n)=>t>=n)}l(SE,"rollingMax");function kE(r,e,t){let n=Ke(r),s=Ke(e),o=n.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]*n.data[p]}let u=Zf(Q(a,[o,i],"f64",n.device),Q(s.data,[o,1],"f64",n.device));return Q(ZV(Je(u)),[i],"f64",n.device)}l(kE,"polyfit");function dE(r,e){let t=0;for(let n=0;n<r.length;n++)t=t*e+r[n];return t}l(dE,"horner");function vE(r,e){let t=Ke(r).data;if(typeof e=="number")return dE(t,e);let n=Ke(e),s=new Float64Array(n.data.length);for(let o=0;o<s.length;o++)s[o]=dE(t,n.data[o]);return Q(s,[s.length],n.dtype,n.device)}l(vE,"polyval");function NE(r){let{data:e,device:t}=Ke(r),n=0;for(;n<e.length-1&&e[n]===0;)n++;let s=e.length-1-n;if(s<1)return Q(new Float64Array(0),[0,2],"f64",t);let o=new Float64Array(s+1);for(let h=0;h<=s;h++)o[h]=e[n+h]/e[n];let i=new Float64Array(s),a=new Float64Array(s),u=1,p=0;for(let h=0;h<s;h++){let f=u*pE-p*fE;p=u*fE+p*pE,u=f,i[h]=u,a[h]=p}for(let h=0;h<YV;h++){let f=0;for(let d=0;d<s;d++){let m=1,g=0;for(let S=1;S<=s;S++){let x=m*i[d]-g*a[d]+o[S];g=m*a[d]+g*i[d],m=x}let y=1,_=0;for(let S=0;S<s;S++){if(S===d)continue;let x=i[d]-i[S],k=a[d]-a[S],A=y*x-_*k;_=y*k+_*x,y=A}let b=y*y+_*_,w=(m*y+g*_)/b,T=(g*y-m*_)/b;i[d]-=w,a[d]-=T,f=Math.max(f,Math.hypot(w,T))}if(f<JV)break}let c=new Float64Array(2*s);for(let h=0;h<s;h++)c[2*h]=i[h],c[2*h+1]=a[h];return Q(c,[s,2],"f64",t)}l(NE,"polyroots");be();var XV=.0331;function eG(r){let e=Array.isArray(r)?r:[r],t=1;for(let n of e)t*=n;return{shp:e,n:t}}l(eG,"normalizeShape");var q_=class{static{l(this,"Generator")}_next;constructor(e){this._next=cr(e)}_uniformPositive(){let e=this._next();for(;e<=0;)e=this._next();return e}_normalDraw(){return Bi(this._uniformPositive())}_gammaDraw(e){if(e<1)return this._gammaDraw(e+1)*Math.pow(this._uniformPositive(),1/e);let t=e-1/3,n=1/Math.sqrt(9*t);for(;;){let s,o;do s=this._normalDraw(),o=1+n*s;while(o<=0);o=o*o*o;let i=this._uniformPositive();if(i<1-XV*s*s*s*s||Math.log(i)<.5*s*s+t*(1-o+Math.log(o)))return t*o}}_fill(e,t,n){let{shp:s,n:o}=eG(e),i=new Float64Array(o);for(let a=0;a<o;a++)i[a]=n();return Q(i,s,t.dtype??"f64",t.device)}uniform(e,t={}){let n=t.low??0,s=t.high??1;return this._fill(e,t,()=>n+(s-n)*this._next())}normal(e,t={}){let n=t.loc??0,s=t.scale??1;return this._fill(e,t,()=>n+s*this._normalDraw())}standardT(e,t={}){let n=t.df;return this._fill(e,t,()=>{let s=this._normalDraw(),o=2*this._gammaDraw(n/2);return s/Math.sqrt(o/n)})}chi2(e,t={}){let n=t.df;return this._fill(e,t,()=>2*this._gammaDraw(n/2))}exponential(e,t={}){let n=t.scale??1;return this._fill(e,t,()=>-n*Math.log(this._uniformPositive()))}multivariateNormal(e,t,n=1,s={}){let o=Ke(e),i=$_(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=Dx(Q(i.data,[a,a],p,u)),h=this.normal([n,a],{device:u,dtype:p}),f=Q(o.data,[a],p,u);return ae(lt(h,c.transpose(0,1)),f)}};l1();O1();D1();K1();aO();mO();AO();OO(We);export{Uf as Accuracy,Yp as Adam,Jp as AdamW,Rp as AdaptiveAvgPool2d,Bp as AvgPool2d,Lp as BCELoss,ul as BatchNorm1d,Ap as BatchNorm2d,Sa as BatchSampler,Us as CPUTarget,ks as CPU_DEVICE,Vf as CSVLogger,ed as CUDATarget,$n as Callback,Yf as ConfusionMatrix,Ca as ConsoleLogger,yp as Conv1d,gp as Conv2d,Qp as CosineAnnealingLR,Ip as CrossEntropyLoss,Kp as DataLoader,wa as Dataset,pr as Dropout,kp as ELU,Pf as EarlyStopping,Cp as Embedding,PO as F,Hf as F1Score,Vp as Flatten,bp as GELU,Od as GPU_DEVICE,Pp as GRU,ll as GRUCell,rl as GradMode,zf as GradientAccumulationScheduler,Na as LRScheduler,$p as LSTM,cl as LSTMCell,Is as LayerNorm,Sp as LeakyReLU,$f as LearningRateMonitor,If as LightningModule,hn as Linear,Np as LogSoftmax,$o as Logger,Mp as MSELoss,Up as MapDataset,Op as MaxPool2d,Gf as MeanMetric,Yn as Metric,qf as MetricCollection,Oi as ModelCheckpoint,ge as Module,zp as ModuleDict,di as ModuleList,mi as MultiheadAttention,Ep as NLLLoss,Yr as Optimizer,bt as Parameter,jp as PositionalEncoding,Kf as Precision,Ai as ProgressCallback,xa as RandomSampler,_p as ReLU,Wf as Recall,Xp as ReduceLROnPlateau,Hp as SGD,gi as Sampler,Fp as Sequential,Ta as SequentialSampler,wp as SiLU,Tp as Sigmoid,vp as Softmax,Zp as StepLR,jf as SumMetric,ft as SymbolicTensor,xp as Tanh,We as Tensor,qp as TensorDataset,Ff as Timer,Wp as Tokenizer,ue as TraceLevel,Df as Trainer,Gp as Transformer,fl as TransformerDecoder,ba as TransformerDecoderLayer,pl as TransformerEncoder,_a as TransformerEncoderLayer,yi as Vocab,sw as WASM_DEVICE,Ys as WEBGPU_DEVICE,td as WasmTarget,nd as WebGPUTarget,Pc as abs,ae as add,Mx as applyCheckpoint,n1 as arange,Ud as argmax,kN as argmin,LN as argsort,ON as broadcast_in_dim,sa as cat,wN as ceil,BN as chunk,Vd as clamp,ef as clipGradNorm_,tf as clipGradValue_,AN as clone,u3 as compile,w_ as compileWithBackward,ci as contiguous,mN as cos,IN as cumsum,FO as data,ka as defaultCollate,Nt as dispatcher,$t as div,NN as dot,fn as empty,QP as emptyLike,yO as enableGrad,$c as eq,gN as erf,yN as erfc,Kr as exp,Lu as expand,Rm as eye,MN as flip,bN as floor,ZN as flushWebGPUEager,ta as fromBuffer,et as full,t$ as fullLike,Cd as gamma,jd as gather,zd as ge,Dd as gelu,ii as getDefaultDevice,Iu as gt,li as index_select,Sg as init,Fd as le,_N as lgamma,dI as lightning,mI as linalg,s1 as linspace,Rx as loadCheckpoint,Zs as log,Hd as log_softmax,xN as lt,lt as matmul,Qs as max,ai as maximum,lr as mean,dn as memfs,SN as min,Ld as minimum,bI as ml,ee as mul,Un as narrow,TN as ne,Ur as neg,$O as nn,gO as noGrad,AE as numeric,Gd as one_hot,Ht as ones,e$ as onesLike,Ft as ops,kM as optim,na as pad,Eu as permute,Js as pow,eA as preloadCudaRuntime,qc as preloadWebGPU,ZP as printModule,vN as prod,Bm as randn,n$ as randnLike,r1 as randperm,Mu as relu,Du as repeat,Cn as reshape,RN as roll,hN as rsqrt,Dc as scalar,Dp as scan,qd as scatter,ra as scatter_add,Ut as select,Cf as serializeCheckpoint,O2 as setDefaultDevice,ur as sigmoid,$d as sign,Pd as silu,dN as sin,Kd as slice,Ro as softmax,EN as sort,Bo as split,ui as sqrt,Cu as squeeze,Wr as stack,Pe as sub,ht as sum,Ns as tanh,Pt as tensor,Wd as tile,qO as tokenizer,CN as topk,HM as trace,gt as transpose,Sr as unsqueeze,qn as where,$e as zeros,XP as zerosLike};