@huggingface/transformers 3.3.1 → 3.3.3

package/dist/transformers.js

@@ -775,18 +775,19 @@ function parse(tokens) {
     return left;
   }
   function parseCallMemberExpression() {
-    const member = parseMemberExpression();
+    const member = parseMemberExpression(parsePrimaryExpression());
     if (is(TOKEN_TYPES.OpenParen)) {
       return parseCallExpression(member);
     }
     return member;
   }
   function parseCallExpression(callee) {
-    let callExpression = new CallExpression(callee, parseArgs());
+    let expression = new CallExpression(callee, parseArgs());
+    expression = parseMemberExpression(expression);
     if (is(TOKEN_TYPES.OpenParen)) {
-      callExpression = parseCallExpression(callExpression);
+      expression = parseCallExpression(expression);
     }
-    return callExpression;
+    return expression;
   }
   function parseArgs() {
     expect(TOKEN_TYPES.OpenParen, "Expected opening parenthesis for arguments list");
@@ -840,8 +841,7 @@ function parse(tokens) {
     }
     return slices[0];
   }
-  function parseMemberExpression() {
-    let object = parsePrimaryExpression();
+  function parseMemberExpression(object) {
     while (is(TOKEN_TYPES.Dot) || is(TOKEN_TYPES.OpenSquareBracket)) {
       const operator = tokens[current];
       ++current;
@@ -1066,6 +1066,41 @@ var StringValue = class extends RuntimeValue {
       new FunctionValue(() => {
         return new StringValue(this.value.trimStart());
       })
+    ],
+    [
+      "split",
+      // follows Python's `str.split(sep=None, maxsplit=-1)` function behavior
+      // https://docs.python.org/3.13/library/stdtypes.html#str.split
+      new FunctionValue((args) => {
+        const sep = args[0] ?? new NullValue();
+        if (!(sep instanceof StringValue || sep instanceof NullValue)) {
+          throw new Error("sep argument must be a string or null");
+        }
+        const maxsplit = args[1] ?? new NumericValue(-1);
+        if (!(maxsplit instanceof NumericValue)) {
+          throw new Error("maxsplit argument must be a number");
+        }
+        let result = [];
+        if (sep instanceof NullValue) {
+          const text = this.value.trimStart();
+          for (const { 0: match, index } of text.matchAll(/\S+/g)) {
+            if (maxsplit.value !== -1 && result.length >= maxsplit.value && index !== void 0) {
+              result.push(match + text.slice(index + match.length));
+              break;
+            }
+            result.push(match);
+          }
+        } else {
+          if (sep.value === "") {
+            throw new Error("empty separator");
+          }
+          result = this.value.split(sep.value);
+          if (maxsplit.value !== -1 && result.length > maxsplit.value) {
+            result.push(result.splice(maxsplit.value).join(sep.value));
+          }
+        }
+        return new ArrayValue(result.map((part) => new StringValue(part)));
+      })
     ]
   ]);
 };
@@ -1402,6 +1437,8 @@ var Interpreter = class {
                 }
               })
             );
+          case "join":
+            return new StringValue(operand.value.map((x) => x.value).join(""));
           default:
             throw new Error(`Unknown ArrayValue filter: ${filter.value}`);
         }
@@ -1428,6 +1465,7 @@ var Interpreter = class {
                 )
               ).join("\n")
             );
+          case "join":
           case "string":
             return operand;
           default:
@@ -1466,6 +1504,21 @@ var Interpreter = class {
           throw new Error("If set, indent must be a number");
         }
         return new StringValue(toJSON(operand, indent.value));
+      } else if (filterName === "join") {
+        let value;
+        if (operand instanceof StringValue) {
+          value = Array.from(operand.value);
+        } else if (operand instanceof ArrayValue) {
+          value = operand.value.map((x) => x.value);
+        } else {
+          throw new Error(`Cannot apply filter "${filterName}" to type: ${operand.type}`);
+        }
+        const [args, kwargs] = this.evaluateArguments(filter.args, environment);
+        const separator = args.at(0) ?? kwargs.get("separator") ?? new StringValue("");
+        if (!(separator instanceof StringValue)) {
+          throw new Error("separator must be a string");
+        }
+        return new StringValue(value.join(separator.value));
       }
       if (operand instanceof ArrayValue) {
         switch (filterName) {
@@ -1927,8 +1980,10 @@ var Template = class {
       throw new Error(args);
     });
     env.set("range", range);
-    for (const [key, value] of Object.entries(items)) {
-      env.set(key, value);
+    if (items) {
+      for (const [key, value] of Object.entries(items)) {
+        env.set(key, value);
+      }
     }
     const interpreter = new Interpreter(env);
     const result = interpreter.run(this.parsed);
@@ -3899,32 +3954,32 @@ const version = '1.20.1';
 __webpack_require__.r(__webpack_exports__);
 /* harmony export */ __webpack_require__.d(__webpack_exports__, {
 /* harmony export */   InferenceSession: () => (/* binding */ Fp),
-/* harmony export */   TRACE: () => (/* binding */ wr),
-/* harmony export */   TRACE_FUNC_BEGIN: () => (/* binding */ Ue),
-/* harmony export */   TRACE_FUNC_END: () => (/* binding */ De),
-/* harmony export */   Tensor: () => (/* binding */ He),
+/* harmony export */   TRACE: () => (/* binding */ _r),
+/* harmony export */   TRACE_FUNC_BEGIN: () => (/* binding */ Ne),
+/* harmony export */   TRACE_FUNC_END: () => (/* binding */ Be),
+/* harmony export */   Tensor: () => (/* binding */ Fe),
 /* harmony export */   "default": () => (/* binding */ $1),
-/* harmony export */   env: () => (/* binding */ ve),
-/* harmony export */   registerBackend: () => (/* binding */ St)
+/* harmony export */   env: () => (/* binding */ _e),
+/* harmony export */   registerBackend: () => (/* binding */ Tt)
 /* harmony export */ });
 /*!
- * ONNX Runtime Web v1.21.0-dev.20250114-228dd16893
+ * ONNX Runtime Web v1.21.0-dev.20250206-d981b153d3
  * Copyright (c) Microsoft Corporation. All rights reserved.
  * Licensed under the MIT License.
  */
-var Un=Object.defineProperty;var Vp=Object.getOwnPropertyDescriptor;var Wp=Object.getOwnPropertyNames;var Lp=Object.prototype.hasOwnProperty;var Nn=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,r)=>(typeof require<"u"?require:t)[r]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')});var U=(e,t)=>()=>(e&&(t=e(e=0)),t);var Ft=(e,t)=>{for(var r in t)Un(e,r,{get:t[r],enumerable:!0})},Gp=(e,t,r,n)=>{if(t&&typeof t=="object"||typeof t=="function")for(let o of Wp(t))!Lp.call(e,o)&&o!==r&&Un(e,o,{get:()=>t[o],enumerable:!(n=Vp(t,o))||n.enumerable});return e};var br=e=>Gp(Un({},"__esModule",{value:!0}),e);var yr,xt,St,Hp,Ji,Vn=U(()=>{"use strict";yr=new Map,xt=[],St=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let n=yr.get(e);if(n===void 0)yr.set(e,{backend:t,priority:r});else{if(n.priority>r)return;if(n.priority===r&&n.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${r}`)}if(r>=0){let o=xt.indexOf(e);o!==-1&&xt.splice(o,1);for(let i=0;i<xt.length;i++)if(yr.get(xt[i]).priority<=r){xt.splice(i,0,e);return}xt.push(e)}return}throw new TypeError("not a valid backend")},Hp=async e=>{let t=yr.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(n){return r||(t.error=`${n}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},Ji=async e=>{let t=e.executionProviders||[],r=t.map(l=>typeof l=="string"?l:l.name),n=r.length===0?xt:r,o,i=[],a=new Set;for(let l of n){let p=await Hp(l);typeof p=="string"?i.push({name:l,err:p}):(o||(o=p),o===p&&a.add(l))}if(!o)throw new Error(`no available backend found. ERR: ${i.map(l=>`[${l.name}] ${l.err}`).join(", ")}`);for(let{name:l,err:p}of i)r.includes(l)&&console.warn(`removing requested execution provider "${l}" from session options because it is not available: ${p}`);let d=t.filter(l=>a.has(typeof l=="string"?l:l.name));return[o,new Proxy(e,{get:(l,p)=>p==="executionProviders"?d:Reflect.get(l,p)})]}});var ea=U(()=>{"use strict";Vn()});var ta,ra=U(()=>{"use strict";ta="1.21.0-dev.20241212-1f88284f96"});var na,Re,Wn=U(()=>{"use strict";ra();na="warning",Re={wasm:{},webgl:{},webgpu:{},versions:{common:ta},set logLevel(e){if(e!==void 0){if(typeof e!="string"||["verbose","info","warning","error","fatal"].indexOf(e)===-1)throw new Error(`Unsupported logging level: ${e}`);na=e}},get logLevel(){return na}};Object.defineProperty(Re,"logLevel",{enumerable:!0})});var ve,oa=U(()=>{"use strict";Wn();ve=Re});var ia,aa,sa=U(()=>{"use strict";ia=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let n=r.getContext("2d");if(n!=null){let o,i;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],i=e.dims[3]):(o=e.dims[3],i=e.dims[2]);let a=t?.format!==void 0?t.format:"RGB",d=t?.norm,l,p;d===void 0||d.mean===void 0?l=[255,255,255,255]:typeof d.mean=="number"?l=[d.mean,d.mean,d.mean,d.mean]:(l=[d.mean[0],d.mean[1],d.mean[2],0],d.mean[3]!==void 0&&(l[3]=d.mean[3])),d===void 0||d.bias===void 0?p=[0,0,0,0]:typeof d.bias=="number"?p=[d.bias,d.bias,d.bias,d.bias]:(p=[d.bias[0],d.bias[1],d.bias[2],0],d.bias[3]!==void 0&&(p[3]=d.bias[3]));let m=i*o,u=0,h=m,_=m*2,y=-1;a==="RGBA"?(u=0,h=m,_=m*2,y=m*3):a==="RGB"?(u=0,h=m,_=m*2):a==="RBG"&&(u=0,_=m,h=m*2);for(let g=0;g<i;g++)for(let x=0;x<o;x++){let $=(e.data[u++]-p[0])*l[0],v=(e.data[h++]-p[1])*l[1],S=(e.data[_++]-p[2])*l[2],T=y===-1?255:(e.data[y++]-p[3])*l[3];n.fillStyle="rgba("+$+","+v+","+S+","+T+")",n.fillRect(x,g,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},aa=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let o,i,a;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],i=e.dims[1],a=e.dims[3]):(o=e.dims[3],i=e.dims[2],a=e.dims[1]);let d=t!==void 0&&t.format!==void 0?t.format:"RGB",l=t?.norm,p,m;l===void 0||l.mean===void 0?p=[255,255,255,255]:typeof l.mean=="number"?p=[l.mean,l.mean,l.mean,l.mean]:(p=[l.mean[0],l.mean[1],l.mean[2],255],l.mean[3]!==void 0&&(p[3]=l.mean[3])),l===void 0||l.bias===void 0?m=[0,0,0,0]:typeof l.bias=="number"?m=[l.bias,l.bias,l.bias,l.bias]:(m=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(m[3]=l.bias[3]));let u=i*o;if(t!==void 0&&(t.format!==void 0&&a===4&&t.format!=="RGBA"||a===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let h=4,_=0,y=1,g=2,x=3,$=0,v=u,S=u*2,T=-1;d==="RGBA"?($=0,v=u,S=u*2,T=u*3):d==="RGB"?($=0,v=u,S=u*2):d==="RBG"&&($=0,S=u,v=u*2),n=r.createImageData(o,i);for(let A=0;A<i*o;_+=h,y+=h,g+=h,x+=h,A++)n.data[_]=(e.data[$++]-m[0])*p[0],n.data[y]=(e.data[v++]-m[1])*p[1],n.data[g]=(e.data[S++]-m[2])*p[2],n.data[x]=T===-1?255:(e.data[T++]-m[3])*p[3]}else throw new Error("Can not access image data");return n}});var Ln,ua,da,la,ca,pa,ma=U(()=>{"use strict";_r();Ln=(e,t)=>{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:n}=t,o=t.norm??{mean:255,bias:0},i,a;typeof o.mean=="number"?i=[o.mean,o.mean,o.mean,o.mean]:i=[o.mean[0],o.mean[1],o.mean[2],o.mean[3]??255],typeof o.bias=="number"?a=[o.bias,o.bias,o.bias,o.bias]:a=[o.bias[0],o.bias[1],o.bias[2],o.bias[3]??0];let d=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",p=r*n,m=l==="RGBA"?new Float32Array(p*4):new Float32Array(p*3),u=4,h=0,_=1,y=2,g=3,x=0,$=p,v=p*2,S=-1;d==="RGB"&&(u=3,h=0,_=1,y=2,g=-1),l==="RGBA"?S=p*3:l==="RBG"?(x=0,v=p,$=p*2):l==="BGR"&&(v=0,$=p,x=p*2);for(let A=0;A<p;A++,h+=u,y+=u,_+=u,g+=u)m[x++]=(e[h]+a[0])/i[0],m[$++]=(e[_]+a[1])/i[1],m[v++]=(e[y]+a[2])/i[2],S!==-1&&g!==-1&&(m[S++]=(e[g]+a[3])/i[3]);return l==="RGBA"?new ze("float32",m,[1,4,r,n]):new ze("float32",m,[1,3,r,n])},ua=async(e,t)=>{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,n=typeof ImageData<"u"&&e instanceof ImageData,o=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,i=typeof e=="string",a,d=t??{},l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},p=m=>typeof HTMLCanvasElement<"u"&&m instanceof HTMLCanvasElement||m instanceof OffscreenCanvas?m.getContext("2d"):null;if(r){let m=l();m.width=e.width,m.height=e.height;let u=p(m);if(u!=null){let h=e.height,_=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(h=t.resizedHeight,_=t.resizedWidth),t!==void 0){if(d=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");d.tensorFormat="RGBA",d.height=h,d.width=_}else d.tensorFormat="RGBA",d.height=h,d.width=_;u.drawImage(e,0,0),a=u.getImageData(0,0,_,h).data}else throw new Error("Can not access image data")}else if(n){let m,u;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(m=t.resizedHeight,u=t.resizedWidth):(m=e.height,u=e.width),t!==void 0&&(d=t),d.format="RGBA",d.height=m,d.width=u,t!==void 0){let h=l();h.width=u,h.height=m;let _=p(h);if(_!=null)_.putImageData(e,0,0),a=_.getImageData(0,0,u,m).data;else throw new Error("Can not access image data")}else a=e.data}else if(o){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let m=l();m.width=e.width,m.height=e.height;let u=p(m);if(u!=null){let h=e.height,_=e.width;return u.drawImage(e,0,0,_,h),a=u.getImageData(0,0,_,h).data,d.height=h,d.width=_,Ln(a,d)}else throw new Error("Can not access image data")}else{if(i)return new Promise((m,u)=>{let h=l(),_=p(h);if(!e||!_)return u();let y=new Image;y.crossOrigin="Anonymous",y.src=e,y.onload=()=>{h.width=y.width,h.height=y.height,_.drawImage(y,0,0,h.width,h.height);let g=_.getImageData(0,0,h.width,h.height);d.height=h.height,d.width=h.width,m(Ln(g.data,d))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(a!==void 0)return Ln(a,d);throw new Error("Input data provided is not supported - aborted tensor creation")},da=(e,t)=>{let{width:r,height:n,download:o,dispose:i}=t,a=[1,n,r,4];return new ze({location:"texture",type:"float32",texture:e,dims:a,download:o,dispose:i})},la=(e,t)=>{let{dataType:r,dims:n,download:o,dispose:i}=t;return new ze({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:n,download:o,dispose:i})},ca=(e,t)=>{let{dataType:r,dims:n,download:o,dispose:i}=t;return new ze({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:n,download:o,dispose:i})},pa=(e,t,r)=>new ze({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})});var Tt,qt,fa,ha,ga=U(()=>{"use strict";Tt=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),qt=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),fa=!1,ha=()=>{if(!fa){fa=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;e&&(Tt.set("int64",BigInt64Array),qt.set(BigInt64Array,"int64")),t&&(Tt.set("uint64",BigUint64Array),qt.set(BigUint64Array,"uint64")),r?(Tt.set("float16",Float16Array),qt.set(Float16Array,"float16")):Tt.set("float16",Uint16Array)}}});var ba,ya,_a=U(()=>{"use strict";_r();ba=e=>{let t=1;for(let r=0;r<e.length;r++){let n=e[r];if(typeof n!="number"||!Number.isSafeInteger(n))throw new TypeError(`dims[${r}] must be an integer, got: ${n}`);if(n<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${n}`);t*=n}return t},ya=(e,t)=>{switch(e.location){case"cpu":return new ze(e.type,e.data,t);case"cpu-pinned":return new ze({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new ze({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new ze({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new ze({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}});var ze,_r=U(()=>{"use strict";sa();ma();ga();_a();ze=class{constructor(t,r,n){ha();let o,i;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,o=t.type,i=t.dims,t.location){case"cpu-pinned":{let d=Tt.get(o);if(!d)throw new TypeError(`unsupported type "${o}" to create tensor from pinned buffer`);if(!(t.data instanceof d))throw new TypeError(`buffer should be of type ${d.name}`);this.cpuData=t.data;break}case"texture":{if(o!=="float32")throw new TypeError(`unsupported type "${o}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break}case"gpu-buffer":{if(o!=="float32"&&o!=="float16"&&o!=="int32"&&o!=="int64"&&o!=="uint32"&&o!=="uint8"&&o!=="bool"&&o!=="uint4"&&o!=="int4")throw new TypeError(`unsupported type "${o}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}case"ml-tensor":{if(o!=="float32"&&o!=="float16"&&o!=="int32"&&o!=="int64"&&o!=="uint32"&&o!=="uint64"&&o!=="int8"&&o!=="uint8"&&o!=="bool"&&o!=="uint4"&&o!=="int4")throw new TypeError(`unsupported type "${o}" to create tensor from MLTensor`);this.mlTensorData=t.mlTensor,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let d,l;if(typeof t=="string")if(o=t,l=n,t==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");d=r}else{let p=Tt.get(t);if(p===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(r)){if(t==="float16"&&p===Uint16Array||t==="uint4"||t==="int4")throw new TypeError(`Creating a ${t} tensor from number array is not supported. Please use ${p.name} as data.`);t==="uint64"||t==="int64"?d=p.from(r,BigInt):d=p.from(r)}else if(r instanceof p)d=r;else if(r instanceof Uint8ClampedArray)if(t==="uint8")d=Uint8Array.from(r);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else throw new TypeError(`A ${o} tensor's data must be type of ${p}`)}else if(l=r,Array.isArray(t)){if(t.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let p=typeof t[0];if(p==="string")o="string",d=t;else if(p==="boolean")o="bool",d=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${p}.`)}else if(t instanceof Uint8ClampedArray)o="uint8",d=Uint8Array.from(t);else{let p=qt.get(t.constructor);if(p===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);o=p,d=t}if(l===void 0)l=[d.length];else if(!Array.isArray(l))throw new TypeError("A tensor's dims must be a number array");i=l,this.cpuData=d,this.dataLocation="cpu"}let a=ba(i);if(this.cpuData&&a!==this.cpuData.length&&!((o==="uint4"||o==="int4")&&Math.ceil(a/2)===this.cpuData.length))throw new Error(`Tensor's size(${a}) does not match data length(${this.cpuData.length}).`);this.type=o,this.dims=i,this.size=a}static async fromImage(t,r){return ua(t,r)}static fromTexture(t,r){return da(t,r)}static fromGpuBuffer(t,r){return la(t,r)}static fromMLTensor(t,r){return ca(t,r)}static fromPinnedBuffer(t,r,n){return pa(t,r,n)}toDataURL(t){return ia(this,t)}toImageData(t){return aa(this,t)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw new Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(t){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let r=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=r,t&&this.disposer&&(this.disposer(),this.disposer=void 0),r}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return ya(this,t)}}});var He,Gn=U(()=>{"use strict";_r();He=ze});var wr,wa,Ue,De,Hn=U(()=>{"use strict";Wn();wr=(e,t)=>{(typeof Re.trace>"u"?!Re.wasm.trace:!Re.trace)||console.timeStamp(`${e}::ORT::${t}`)},wa=(e,t)=>{let r=new Error().stack?.split(/\r\n|\r|\n/g)||[],n=!1;for(let o=0;o<r.length;o++){if(n&&!r[o].includes("TRACE_FUNC")){let i=`FUNC_${e}::${r[o].trim().split(" ")[1]}`;t&&(i+=`::${t}`),wr("CPU",i);return}r[o].includes("TRACE_FUNC")&&(n=!0)}},Ue=e=>{(typeof Re.trace>"u"?!Re.wasm.trace:!Re.trace)||wa("BEGIN",e)},De=e=>{(typeof Re.trace>"u"?!Re.wasm.trace:!Re.trace)||wa("END",e)}});var vr,va=U(()=>{"use strict";Vn();Gn();Hn();vr=class e{constructor(t){this.handler=t}async run(t,r,n){Ue();let o={},i={};if(typeof t!="object"||t===null||t instanceof He||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let a=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof He)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");a=!1;for(let p of r){if(typeof p!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(p)===-1)throw new RangeError(`'fetches' contains invalid output name: ${p}.`);o[p]=null}if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let p=!1,m=Object.getOwnPropertyNames(r);for(let u of this.outputNames)if(m.indexOf(u)!==-1){let h=r[u];(h===null||h instanceof He)&&(p=!0,a=!1,o[u]=h)}if(p){if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else i=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let p of this.inputNames)if(typeof t[p]>"u")throw new Error(`input '${p}' is missing in 'feeds'.`);if(a)for(let p of this.outputNames)o[p]=null;let d=await this.handler.run(t,o,i),l={};for(let p in d)if(Object.hasOwnProperty.call(d,p)){let m=d[p];m instanceof He?l[p]=m:l[p]=new He(m.type,m.data,m.dims)}return De(),l}async release(){return this.handler.dispose()}static async create(t,r,n,o){Ue();let i,a={};if(typeof t=="string"){if(i=t,typeof r=="object"&&r!==null)a=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof Uint8Array){if(i=t,typeof r=="object"&&r!==null)a=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof ArrayBuffer||typeof SharedArrayBuffer<"u"&&t instanceof SharedArrayBuffer){let m=t,u=0,h=t.byteLength;if(typeof r=="object"&&r!==null)a=r;else if(typeof r=="number"){if(u=r,!Number.isSafeInteger(u))throw new RangeError("'byteOffset' must be an integer.");if(u<0||u>=m.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${m.byteLength}).`);if(h=t.byteLength-u,typeof n=="number"){if(h=n,!Number.isSafeInteger(h))throw new RangeError("'byteLength' must be an integer.");if(h<=0||u+h>m.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${m.byteLength-u}].`);if(typeof o=="object"&&o!==null)a=o;else if(typeof o<"u")throw new TypeError("'options' must be an object.")}else if(typeof n<"u")throw new TypeError("'byteLength' must be a number.")}else if(typeof r<"u")throw new TypeError("'options' must be an object.");i=new Uint8Array(m,u,h)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[d,l]=await Ji(a),p=await d.createInferenceSessionHandler(i,l);return De(),new e(p)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}});var Fp,$a=U(()=>{"use strict";va();Fp=vr});var xa=U(()=>{"use strict"});var Sa=U(()=>{"use strict"});var Ta=U(()=>{"use strict"});var Ia=U(()=>{"use strict"});var Fn={};Ft(Fn,{InferenceSession:()=>Fp,TRACE:()=>wr,TRACE_FUNC_BEGIN:()=>Ue,TRACE_FUNC_END:()=>De,Tensor:()=>He,env:()=>ve,registerBackend:()=>St});var We=U(()=>{"use strict";ea();oa();$a();Gn();xa();Sa();Hn();Ta();Ia()});var $r=U(()=>{"use strict"});var Ea={};Ft(Ea,{default:()=>qp});var Aa,ka,qp,Pa=U(()=>{"use strict";qn();gt();xr();Aa="ort-wasm-proxy-worker",ka=globalThis.self?.name===Aa;ka&&(self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":Sr(r.wasm).then(()=>{Tr(r).then(()=>{postMessage({type:t})},n=>{postMessage({type:t,err:n})})},n=>{postMessage({type:t,err:n})});break;case"init-ep":{let{epName:n,env:o}=r;Ir(o,n).then(()=>{postMessage({type:t})},i=>{postMessage({type:t,err:i})});break}case"copy-from":{let{buffer:n}=r,o=Kt(n);postMessage({type:t,out:o});break}case"create":{let{model:n,options:o}=r;Cr(n,o).then(i=>{postMessage({type:t,out:i})},i=>{postMessage({type:t,err:i})});break}case"release":Ar(r),postMessage({type:t});break;case"run":{let{sessionId:n,inputIndices:o,inputs:i,outputIndices:a,options:d}=r;kr(n,o,i,a,new Array(a.length).fill(null),d).then(l=>{l.some(p=>p[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:l},Pr([...i,...l]))},l=>{postMessage({type:t,err:l})});break}case"end-profiling":Er(r),postMessage({type:t});break;default:}}catch(n){postMessage({type:t,err:n})}});qp=ka?null:e=>new Worker(e??Ne,{type:"module",name:Aa})});var Oa={};Ft(Oa,{default:()=>Kp});var Kn,za,Kp,Da=U(()=>{"use strict";za=(Kn=import.meta.url,async function(e={}){function t(){return se.buffer!=J.buffer&&ye(),J}function r(){return se.buffer!=J.buffer&&ye(),ne}function n(){return se.buffer!=J.buffer&&ye(),be}function o(){return se.buffer!=J.buffer&&ye(),Oe}function i(){return se.buffer!=J.buffer&&ye(),$e}function a(){return se.buffer!=J.buffer&&ye(),le}function d(){return se.buffer!=J.buffer&&ye(),W}function l(){return se.buffer!=J.buffer&&ye(),Ge}var p,m,u=Object.assign({},e),h=new Promise((s,c)=>{p=s,m=c}),_=typeof window=="object",y=typeof importScripts=="function",g=y&&self.name=="em-pthread";u.mountExternalData=(s,c)=>{s.startsWith("./")&&(s=s.substring(2)),(u.Fb||(u.Fb=new Map)).set(s,c)},u.unmountExternalData=()=>{delete u.Fb};var x=globalThis.SharedArrayBuffer??new WebAssembly.Memory({initial:0,maximum:0,shared:!0}).buffer.constructor;let $=()=>{let s=(f,b,w)=>(...I)=>{let O=Xe,B=b?.();I=f(...I);let L=b?.();return B!==L&&(f=L,w(B),b=w=null),Xe!=O?new Promise((H,X)=>{En={resolve:H,reject:X}}):I},c=f=>async(...b)=>{try{if(u.Gb)throw Error("Session already started");let w=u.Gb={hc:b[0],errors:[]},I=await f(...b);if(u.Gb!==w)throw Error("Session mismatch");u.Hb?.flush();let O=w.errors;if(0<O.length){let B=await Promise.all(O);if(B=B.filter(L=>L),0<B.length)throw Error(B.join(`
-`))}return I}finally{u.Gb=null}};u._OrtCreateSession=s(u._OrtCreateSession,()=>u._OrtCreateSession,f=>u._OrtCreateSession=f),u._OrtRun=c(s(u._OrtRun,()=>u._OrtRun,f=>u._OrtRun=f)),u._OrtRunWithBinding=c(s(u._OrtRunWithBinding,()=>u._OrtRunWithBinding,f=>u._OrtRunWithBinding=f)),u._OrtBindInput=s(u._OrtBindInput,()=>u._OrtBindInput,f=>u._OrtBindInput=f),$=void 0};u.jsepInit=(s,c)=>{if($?.(),s==="webgpu"){[u.Hb,u.Vb,u.Zb,u.Ob,u.Yb,u.kb,u.$b,u.cc,u.Wb,u.Xb,u.ac]=c;let f=u.Hb;u.jsepRegisterBuffer=(b,w,I,O)=>f.registerBuffer(b,w,I,O),u.jsepGetBuffer=b=>f.getBuffer(b),u.jsepCreateDownloader=(b,w,I)=>f.createDownloader(b,w,I),u.jsepOnCreateSession=b=>{f.onCreateSession(b)},u.jsepOnReleaseSession=b=>{f.onReleaseSession(b)},u.jsepOnRunStart=b=>f.onRunStart(b),u.dc=(b,w)=>{f.upload(b,w)}}else if(s==="webnn"){[u.Hb,u.bc,u.Pb,u.jsepEnsureTensor,u.ec,u.jsepDownloadTensor]=c,u.jsepReleaseTensorId=u.Pb;let f=u.Hb;u.jsepOnRunStart=b=>f.onRunStart(b),u.jsepRegisterMLContext=(b,w)=>{f.registerMLContext(b,w)},u.jsepOnReleaseSession=b=>{f.onReleaseSession(b)},u.jsepCreateMLTensorDownloader=(b,w)=>f.createMLTensorDownloader(b,w),u.jsepRegisterMLTensor=(b,w,I)=>f.registerMLTensor(b,w,I),u.jsepCreateMLContext=b=>f.createMLContext(b),u.jsepRegisterMLConstant=(b,w,I,O,B)=>f.registerMLConstant(b,w,I,O,B,u.Fb)}};var v,S,T=Object.assign({},u),A="./this.program",k=(s,c)=>{throw c},P="";(_||y)&&(y?P=self.location.href:typeof document<"u"&&document.currentScript&&(P=document.currentScript.src),Kn&&(P=Kn),P=P.startsWith("blob:")?"":P.substr(0,P.replace(/[?#].*/,"").lastIndexOf("/")+1),y&&(S=s=>{var c=new XMLHttpRequest;return c.open("GET",s,!1),c.responseType="arraybuffer",c.send(null),new Uint8Array(c.response)}),v=(s,c,f)=>{var b=new XMLHttpRequest;b.open("GET",s,!0),b.responseType="arraybuffer",b.onload=()=>{b.status==200||b.status==0&&b.response?c(b.response):f()},b.onerror=f,b.send(null)});var D,R=console.log.bind(console),G=console.error.bind(console),K=R,j=G;if(Object.assign(u,T),T=null,g){let s=function(c){try{var f=c.data,b=f.cmd;if(b==="load"){let w=[];self.onmessage=I=>w.push(I),self.startWorker=()=>{postMessage({cmd:"loaded"});for(let I of w)s(I);self.onmessage=s};for(let I of f.handlers)u[I]&&!u[I].proxy||(u[I]=(...O)=>{postMessage({Nb:"callHandler",pc:I,args:O})},I=="print"&&(K=u[I]),I=="printErr"&&(j=u[I]));se=f.wasmMemory,ye(),V(f.wasmModule)}else if(b==="run"){Dn(f.pthread_ptr,0,0,1,0,0),An(f.pthread_ptr),xc(),Ho(),Q||(Wi(),Q=!0);try{Sc(f.start_routine,f.arg)}catch(w){if(w!="unwind")throw w}}else b==="cancel"?Mt()&&hr(-1):f.target!=="setimmediate"&&(b==="checkMailbox"?Q&&ir():b&&(j(`worker: received unknown command ${b}`),j(f)))}catch(w){throw Li(),w}};var wg=s,V,Q=!1;j=function(...c){c=c.join(" "),console.error(c)},self.alert=function(...c){postMessage({Nb:"alert",text:c.join(" "),rc:Mt()})},u.instantiateWasm=(c,f)=>new Promise(b=>{V=w=>{w=new WebAssembly.Instance(w,No()),f(w),b()}}),self.onunhandledrejection=c=>{throw c.reason||c},self.onmessage=s}u.wasmBinary&&(D=u.wasmBinary);var se,Y,ee,J,ne,be,Oe,$e,le,W,q,he,Ge,we=!1;function ye(){var s=se.buffer;u.HEAP8=J=new Int8Array(s),u.HEAP16=be=new Int16Array(s),u.HEAPU8=ne=new Uint8Array(s),u.HEAPU16=Oe=new Uint16Array(s),u.HEAP32=$e=new Int32Array(s),u.HEAPU32=le=new Uint32Array(s),u.HEAPF32=W=new Float32Array(s),u.HEAPF64=Ge=new Float64Array(s),u.HEAP64=q=new BigInt64Array(s),u.HEAPU64=he=new BigUint64Array(s)}if(!g){if(!((se=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0})).buffer instanceof x))throw j("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),Error("bad memory");ye()}var Ye=[],Lt=[],fn=[],Gt=0,hn=null,Ht=null;function Do(){if(--Gt==0&&(hn!==null&&(clearInterval(hn),hn=null),Ht)){var s=Ht;Ht=null,s()}}function ct(s){throw j(s="Aborted("+s+")"),we=!0,ee=1,s=new WebAssembly.RuntimeError(s+". Build with -sASSERTIONS for more info."),m(s),s}var gn,Bo=s=>s.startsWith("data:application/octet-stream;base64,"),Mo=s=>s.startsWith("file://");function Ro(s){if(s==gn&&D)return new Uint8Array(D);if(S)return S(s);throw"both async and sync fetching of the wasm failed"}function Uo(s,c,f){return function(b){if(!D&&(_||y)){if(typeof fetch=="function"&&!Mo(b))return fetch(b,{credentials:"same-origin"}).then(w=>{if(!w.ok)throw`failed to load wasm binary file at '${b}'`;return w.arrayBuffer()}).catch(()=>Ro(b));if(v)return new Promise((w,I)=>{v(b,O=>w(new Uint8Array(O)),I)})}return Promise.resolve().then(()=>Ro(b))}(s).then(b=>WebAssembly.instantiate(b,c)).then(f,b=>{j(`failed to asynchronously prepare wasm: ${b}`),ct(b)})}function No(){return{a:{O:$c,Aa:vc,b:Ic,aa:jo,B:Qo,qa:Xo,Y:ei,_:ti,ra:ri,oa:ni,ha:oi,na:ii,L:ai,Z:si,W:ui,pa:di,X:li,va:Cc,F:kc,Q:Ec,P:zc,E:Dc,u:Bc,q:Mc,G:Rc,A:Hc,R:Fc,ua:qc,ka:Kc,U:jc,ba:Yc,H:Zc,ja:An,ta:Qc,t:Xc,Ba:Jc,x:rp,n:np,l:ip,c:In,o:ap,j:dp,w:lp,p:cp,f:pp,s:mp,m:fp,e:hp,k:gp,i:bp,h:yp,d:_p,ea:wp,fa:vp,ga:$p,ca:Si,da:Ti,T:xp,g:Sp,D:Tp,I:Ip,M:Cp,y:Ap,sa:kp,V:Ep,v:Ci,z:Pp,N:zp,S:Op,za:Dp,ya:Bp,la:Ei,ma:Pi,$:vn,C:zi,K:Oi,ia:Di,J:Bi,a:se,xa:wn,wa:Ui,r:Up}}}var bn={913700:(s,c,f,b,w)=>{if(u===void 0||!u.Fb)return 1;if((s=Ce(Number(s>>>0))).startsWith("./")&&(s=s.substring(2)),!(s=u.Fb.get(s)))return 2;if(c=Number(c>>>0),f=Number(f>>>0),b=Number(b>>>0),c+f>s.byteLength)return 3;try{let I=s.subarray(c,c+f);switch(w){case 0:r().set(I,b>>>0);break;case 1:u.dc(b,I);break;default:return 4}return 0}catch{return 4}},914415:(s,c,f)=>{u.ec(s,r().subarray(c>>>0,c+f>>>0))},914478:()=>u.bc(),914519:s=>{u.Pb(s)},914555:()=>{u.Wb()},914586:()=>{u.Xb()},914615:()=>{u.ac()},914640:s=>u.Vb(s),914673:s=>u.Zb(s),914705:(s,c,f)=>{u.Ob(Number(s),Number(c),Number(f),!0)},914768:(s,c,f)=>{u.Ob(Number(s),Number(c),Number(f))},914825:()=>typeof wasmOffsetConverter<"u",914882:s=>{u.kb("Abs",s,void 0)},914933:s=>{u.kb("Neg",s,void 0)},914984:s=>{u.kb("Floor",s,void 0)},915037:s=>{u.kb("Ceil",s,void 0)},915089:s=>{u.kb("Reciprocal",s,void 0)},915147:s=>{u.kb("Sqrt",s,void 0)},915199:s=>{u.kb("Exp",s,void 0)},915250:s=>{u.kb("Erf",s,void 0)},915301:s=>{u.kb("Sigmoid",s,void 0)},915356:(s,c,f)=>{u.kb("HardSigmoid",s,{alpha:c,beta:f})},915435:s=>{u.kb("Log",s,void 0)},915486:s=>{u.kb("Sin",s,void 0)},915537:s=>{u.kb("Cos",s,void 0)},915588:s=>{u.kb("Tan",s,void 0)},915639:s=>{u.kb("Asin",s,void 0)},915691:s=>{u.kb("Acos",s,void 0)},915743:s=>{u.kb("Atan",s,void 0)},915795:s=>{u.kb("Sinh",s,void 0)},915847:s=>{u.kb("Cosh",s,void 0)},915899:s=>{u.kb("Asinh",s,void 0)},915952:s=>{u.kb("Acosh",s,void 0)},916005:s=>{u.kb("Atanh",s,void 0)},916058:s=>{u.kb("Tanh",s,void 0)},916110:s=>{u.kb("Not",s,void 0)},916161:(s,c,f)=>{u.kb("Clip",s,{min:c,max:f})},916230:s=>{u.kb("Clip",s,void 0)},916282:(s,c)=>{u.kb("Elu",s,{alpha:c})},916340:s=>{u.kb("Gelu",s,void 0)},916392:s=>{u.kb("Relu",s,void 0)},916444:(s,c)=>{u.kb("LeakyRelu",s,{alpha:c})},916508:(s,c)=>{u.kb("ThresholdedRelu",s,{alpha:c})},916578:(s,c)=>{u.kb("Cast",s,{to:c})},916636:s=>{u.kb("Add",s,void 0)},916687:s=>{u.kb("Sub",s,void 0)},916738:s=>{u.kb("Mul",s,void 0)},916789:s=>{u.kb("Div",s,void 0)},916840:s=>{u.kb("Pow",s,void 0)},916891:s=>{u.kb("Equal",s,void 0)},916944:s=>{u.kb("Greater",s,void 0)},916999:s=>{u.kb("GreaterOrEqual",s,void 0)},917061:s=>{u.kb("Less",s,void 0)},917113:s=>{u.kb("LessOrEqual",s,void 0)},917172:(s,c,f,b,w)=>{u.kb("ReduceMean",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},917347:(s,c,f,b,w)=>{u.kb("ReduceMax",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},917521:(s,c,f,b,w)=>{u.kb("ReduceMin",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},917695:(s,c,f,b,w)=>{u.kb("ReduceProd",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},917870:(s,c,f,b,w)=>{u.kb("ReduceSum",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},918044:(s,c,f,b,w)=>{u.kb("ReduceL1",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},918217:(s,c,f,b,w)=>{u.kb("ReduceL2",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},918390:(s,c,f,b,w)=>{u.kb("ReduceLogSum",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},918567:(s,c,f,b,w)=>{u.kb("ReduceSumSquare",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},918747:(s,c,f,b,w)=>{u.kb("ReduceLogSumExp",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},918927:s=>{u.kb("Where",s,void 0)},918980:(s,c,f)=>{u.kb("Transpose",s,{perm:c?Array.from(i().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},919104:(s,c,f,b)=>{u.kb("DepthToSpace",s,{blocksize:c,mode:Ce(f),format:b?"NHWC":"NCHW"})},919237:(s,c,f,b)=>{u.kb("DepthToSpace",s,{blocksize:c,mode:Ce(f),format:b?"NHWC":"NCHW"})},919370:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z,de)=>{u.kb("ConvTranspose",s,{format:L?"NHWC":"NCHW",autoPad:c,dilations:[f],group:b,kernelShape:[w],pads:[I,O],strides:[B],wIsConst:()=>!!t()[H>>>0],outputPadding:X?Array.from(i().subarray(Number(X)>>>0,Number(ce)>>>0)):[],outputShape:ge?Array.from(i().subarray(Number(ge)>>>0,Number(z)>>>0)):[],activation:Ce(de)})},919803:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z)=>{u.kb("ConvTranspose",s,{format:B?"NHWC":"NCHW",autoPad:c,dilations:Array.from(i().subarray(Number(f)>>>0,2+(Number(f)>>>0)>>>0)),group:b,kernelShape:Array.from(i().subarray(Number(w)>>>0,2+(Number(w)>>>0)>>>0)),pads:Array.from(i().subarray(Number(I)>>>0,4+(Number(I)>>>0)>>>0)),strides:Array.from(i().subarray(Number(O)>>>0,2+(Number(O)>>>0)>>>0)),wIsConst:()=>!!t()[L>>>0],outputPadding:H?Array.from(i().subarray(Number(H)>>>0,Number(X)>>>0)):[],outputShape:ce?Array.from(i().subarray(Number(ce)>>>0,Number(ge)>>>0)):[],activation:Ce(z)})},920464:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z,de)=>{u.kb("ConvTranspose",s,{format:L?"NHWC":"NCHW",autoPad:c,dilations:[f],group:b,kernelShape:[w],pads:[I,O],strides:[B],wIsConst:()=>!!t()[H>>>0],outputPadding:X?Array.from(i().subarray(Number(X)>>>0,Number(ce)>>>0)):[],outputShape:ge?Array.from(i().subarray(Number(ge)>>>0,Number(z)>>>0)):[],activation:Ce(de)})},920897:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z)=>{u.kb("ConvTranspose",s,{format:B?"NHWC":"NCHW",autoPad:c,dilations:Array.from(i().subarray(Number(f)>>>0,2+(Number(f)>>>0)>>>0)),group:b,kernelShape:Array.from(i().subarray(Number(w)>>>0,2+(Number(w)>>>0)>>>0)),pads:Array.from(i().subarray(Number(I)>>>0,4+(Number(I)>>>0)>>>0)),strides:Array.from(i().subarray(Number(O)>>>0,2+(Number(O)>>>0)>>>0)),wIsConst:()=>!!t()[L>>>0],outputPadding:H?Array.from(i().subarray(Number(H)>>>0,Number(X)>>>0)):[],outputShape:ce?Array.from(i().subarray(Number(ce)>>>0,Number(ge)>>>0)):[],activation:Ce(z)})},921558:(s,c)=>{u.kb("GlobalAveragePool",s,{format:c?"NHWC":"NCHW"})},921649:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z)=>{u.kb("AveragePool",s,{format:z?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:w,dilations:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],kernel_shape:B?Array.from(i().subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:H?Array.from(i().subarray(Number(H)>>>0,Number(X)>>>0)):[],strides:ce?Array.from(i().subarray(Number(ce)>>>0,Number(ge)>>>0)):[]})},922128:(s,c)=>{u.kb("GlobalAveragePool",s,{format:c?"NHWC":"NCHW"})},922219:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z)=>{u.kb("AveragePool",s,{format:z?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:w,dilations:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],kernel_shape:B?Array.from(i().subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:H?Array.from(i().subarray(Number(H)>>>0,Number(X)>>>0)):[],strides:ce?Array.from(i().subarray(Number(ce)>>>0,Number(ge)>>>0)):[]})},922698:(s,c)=>{u.kb("GlobalMaxPool",s,{format:c?"NHWC":"NCHW"})},922785:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z)=>{u.kb("MaxPool",s,{format:z?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:w,dilations:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],kernel_shape:B?Array.from(i().subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:H?Array.from(i().subarray(Number(H)>>>0,Number(X)>>>0)):[],strides:ce?Array.from(i().subarray(Number(ce)>>>0,Number(ge)>>>0)):[]})},923260:(s,c)=>{u.kb("GlobalMaxPool",s,{format:c?"NHWC":"NCHW"})},923347:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z)=>{u.kb("MaxPool",s,{format:z?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:w,dilations:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],kernel_shape:B?Array.from(i().subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:H?Array.from(i().subarray(Number(H)>>>0,Number(X)>>>0)):[],strides:ce?Array.from(i().subarray(Number(ce)>>>0,Number(ge)>>>0)):[]})},923822:(s,c,f,b,w)=>{u.kb("Gemm",s,{alpha:c,beta:f,transA:b,transB:w})},923926:s=>{u.kb("MatMul",s,void 0)},923980:(s,c,f,b)=>{u.kb("ArgMax",s,{keepDims:!!c,selectLastIndex:!!f,axis:b})},924088:(s,c,f,b)=>{u.kb("ArgMin",s,{keepDims:!!c,selectLastIndex:!!f,axis:b})},924196:(s,c)=>{u.kb("Softmax",s,{axis:c})},924259:(s,c)=>{u.kb("Concat",s,{axis:c})},924319:(s,c,f,b,w)=>{u.kb("Split",s,{axis:c,numOutputs:f,splitSizes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},924475:s=>{u.kb("Expand",s,void 0)},924529:(s,c)=>{u.kb("Gather",s,{axis:Number(c)})},924600:(s,c)=>{u.kb("GatherElements",s,{axis:Number(c)})},924679:(s,c)=>{u.kb("GatherND",s,{batch_dims:Number(c)})},924758:(s,c,f,b,w,I,O,B,L,H,X)=>{u.kb("Resize",s,{antialias:c,axes:f?Array.from(i().subarray(Number(f)>>>0,Number(b)>>>0)):[],coordinateTransformMode:Ce(w),cubicCoeffA:I,excludeOutside:O,extrapolationValue:B,keepAspectRatioPolicy:Ce(L),mode:Ce(H),nearestMode:Ce(X)})},925120:(s,c,f,b,w,I,O)=>{u.kb("Slice",s,{starts:c?Array.from(i().subarray(Number(c)>>>0,Number(f)>>>0)):[],ends:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[],axes:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[]})},925384:s=>{u.kb("Tile",s,void 0)},925436:(s,c,f)=>{u.kb("InstanceNormalization",s,{epsilon:c,format:f?"NHWC":"NCHW"})},925550:(s,c,f)=>{u.kb("InstanceNormalization",s,{epsilon:c,format:f?"NHWC":"NCHW"})},925664:s=>{u.kb("Range",s,void 0)},925717:(s,c)=>{u.kb("Einsum",s,{equation:Ce(c)})},925798:(s,c,f,b,w)=>{u.kb("Pad",s,{mode:c,value:f,pads:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},925941:(s,c,f,b,w,I)=>{u.kb("BatchNormalization",s,{epsilon:c,momentum:f,spatial:!!w,trainingMode:!!b,format:I?"NHWC":"NCHW"})},926110:(s,c,f,b,w,I)=>{u.kb("BatchNormalization",s,{epsilon:c,momentum:f,spatial:!!w,trainingMode:!!b,format:I?"NHWC":"NCHW"})},926279:(s,c,f)=>{u.kb("CumSum",s,{exclusive:Number(c),reverse:Number(f)})},926376:(s,c,f)=>{u.kb("DequantizeLinear",s,{axis:c,blockSize:f})},926466:(s,c,f,b,w)=>{u.kb("GridSample",s,{align_corners:c,mode:Ce(f),padding_mode:Ce(b),format:w?"NHWC":"NCHW"})},926636:(s,c,f,b,w)=>{u.kb("GridSample",s,{align_corners:c,mode:Ce(f),padding_mode:Ce(b),format:w?"NHWC":"NCHW"})},926806:(s,c,f,b,w,I,O,B,L)=>{u.kb("Attention",s,{numHeads:c,isUnidirectional:f,maskFilterValue:b,scale:w,doRotary:I,qkvHiddenSizes:O?Array.from(i().subarray(Number(B)>>>0,Number(B)+O>>>0)):[],pastPresentShareBuffer:!!L})},927078:s=>{u.kb("BiasAdd",s,void 0)},927133:s=>{u.kb("BiasSplitGelu",s,void 0)},927194:s=>{u.kb("FastGelu",s,void 0)},927250:(s,c,f,b,w,I,O,B,L,H,X,ce,ge,z,de,Te)=>{u.kb("Conv",s,{format:ce?"NHWC":"NCHW",auto_pad:c,dilations:f?Array.from(i().subarray(Number(f)>>>0,Number(b)>>>0)):[],group:w,kernel_shape:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],pads:B?Array.from(i().subarray(Number(B)>>>0,Number(L)>>>0)):[],strides:H?Array.from(i().subarray(Number(H)>>>0,Number(X)>>>0)):[],w_is_const:()=>!!t()[Number(ge)>>>0],activation:Ce(z),activation_params:de?Array.from(d().subarray(Number(de)>>>0,Number(Te)>>>0)):[]})},927834:s=>{u.kb("Gelu",s,void 0)},927886:(s,c,f,b,w,I,O,B,L)=>{u.kb("GroupQueryAttention",s,{numHeads:c,kvNumHeads:f,scale:b,softcap:w,doRotary:I,rotaryInterleaved:O,smoothSoftmax:B,localWindowSize:L})},928103:(s,c,f,b)=>{u.kb("LayerNormalization",s,{axis:c,epsilon:f,simplified:!!b})},928214:(s,c,f,b)=>{u.kb("LayerNormalization",s,{axis:c,epsilon:f,simplified:!!b})},928325:(s,c,f,b,w,I)=>{u.kb("MatMulNBits",s,{k:c,n:f,accuracyLevel:b,bits:w,blockSize:I})},928452:(s,c,f,b,w,I)=>{u.kb("MultiHeadAttention",s,{numHeads:c,isUnidirectional:f,maskFilterValue:b,scale:w,doRotary:I})},928611:(s,c)=>{u.kb("QuickGelu",s,{alpha:c})},928675:(s,c,f,b,w)=>{u.kb("RotaryEmbedding",s,{interleaved:!!c,numHeads:f,rotaryEmbeddingDim:b,scale:w})},928814:(s,c,f)=>{u.kb("SkipLayerNormalization",s,{epsilon:c,simplified:!!f})},928916:(s,c,f)=>{u.kb("SkipLayerNormalization",s,{epsilon:c,simplified:!!f})},929018:(s,c,f,b)=>{u.kb("GatherBlockQuantized",s,{gatherAxis:c,quantizeAxis:f,blockSize:b})},929139:s=>{u.$b(s)},929173:(s,c)=>u.cc(Number(s),Number(c),u.Gb.hc,u.Gb.errors)};function vc(s,c,f){return _i(async()=>{await u.Yb(Number(s),Number(c),Number(f))})}function $c(){return typeof wasmOffsetConverter<"u"}function yn(s){this.name="ExitStatus",this.message=`Program terminated with exit(${s})`,this.status=s}var _n=s=>{s.terminate(),s.onmessage=()=>{}},Vo=s=>{pt.length==0&&(qo(),Fo(pt[0]));var c=pt.pop();if(!c)return 6;vt.push(c),Ze[s.Bb]=c,c.Bb=s.Bb;var f={cmd:"run",start_routine:s.ic,arg:s.Rb,pthread_ptr:s.Bb};return c.postMessage(f,s.nc),0},wt=0,xe=(s,c,...f)=>{for(var b=2*f.length,w=Rn(),I=Mn(8*b),O=I>>>3,B=0;B<f.length;B++){var L=f[B];typeof L=="bigint"?(q[O+2*B]=1n,q[O+2*B+1]=L):(q[O+2*B]=0n,l()[O+2*B+1>>>0]=L)}return s=Gi(s,0,b,I,c),gr(w),s};function wn(s){if(g)return xe(0,1,s);if(ee=s,!(0<wt)){for(var c of vt)_n(c);for(c of pt)_n(c);pt=[],vt=[],Ze=[],we=!0}k(s,new yn(s))}function Wo(s){if(g)return xe(1,0,s);vn(s)}var vn=s=>{if(ee=s,g)throw Wo(s),"unwind";wn(s)},pt=[],vt=[],Lo=[],Ze={},Go=s=>{var c=s.Bb;delete Ze[c],pt.push(s),vt.splice(vt.indexOf(s),1),s.Bb=0,Bn(c)};function Ho(){Lo.forEach(s=>s())}var Fo=s=>new Promise(c=>{s.onmessage=w=>{var I=(w=w.data).cmd;if(w.targetThread&&w.targetThread!=Mt()){var O=Ze[w.targetThread];O?O.postMessage(w,w.transferList):j(`Internal error! Worker sent a message "${I}" to target pthread ${w.targetThread}, but that thread no longer exists!`)}else I==="checkMailbox"?ir():I==="spawnThread"?Vo(w):I==="cleanupThread"?Go(Ze[w.thread]):I==="killThread"?(w=w.thread,I=Ze[w],delete Ze[w],_n(I),Bn(w),vt.splice(vt.indexOf(I),1),I.Bb=0):I==="cancelThread"?Ze[w.thread].postMessage({cmd:"cancel"}):I==="loaded"?(s.loaded=!0,c(s)):I==="alert"?alert(`Thread ${w.threadId}: ${w.text}`):w.target==="setimmediate"?s.postMessage(w):I==="callHandler"?u[w.handler](...w.args):I&&j(`worker sent an unknown command ${I}`)},s.onerror=w=>{throw j(`worker sent an error! ${w.filename}:${w.lineno}: ${w.message}`),w};var f,b=[];for(f of[])u.hasOwnProperty(f)&&b.push(f);s.postMessage({cmd:"load",handlers:b,wasmMemory:se,wasmModule:Y})});function qo(){var s=new Worker(import.meta.url.startsWith("file:")?new URL(/* asset import */ __webpack_require__(/*! ort.bundle.min.mjs */ "./node_modules/onnxruntime-web/dist/ort.bundle.min.mjs?46eb"), __webpack_require__.b):new URL(import.meta.url),{type:"module",workerData:"em-pthread",name:"em-pthread"});pt.push(s)}var or=s=>{for(;0<s.length;)s.shift()(u)},xc=()=>{var s=Mt(),c=a()[s+52>>>2>>>0];s=a()[s+56>>>2>>>0],Fi(c,c-s),gr(c)},Sc=(s,c)=>{wt=0,s=qi(s,c),0<wt?ee=s:hr(s)};class Tc{constructor(c){this.Kb=c-24}}function Ic(s,c,f){var b=new Tc(s>>>=0);throw c>>>=0,f>>>=0,a()[b.Kb+16>>>2>>>0]=0,a()[b.Kb+4>>>2>>>0]=c,a()[b.Kb+8>>>2>>>0]=f,s}function Ko(s,c,f,b){return g?xe(2,1,s,c,f,b):jo(s,c,f,b)}function jo(s,c,f,b){if(s>>>=0,c>>>=0,f>>>=0,b>>>=0,x===void 0)return j("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var w=[];return g&&w.length===0?Ko(s,c,f,b):(s={ic:f,Bb:s,Rb:b,nc:w},g?(s.Nb="spawnThread",postMessage(s,w),0):Vo(s))}var Yo=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Zo=(s,c,f)=>{var b=(c>>>=0)+f;for(f=c;s[f]&&!(f>=b);)++f;if(16<f-c&&s.buffer&&Yo)return Yo.decode(s.buffer instanceof x?s.slice(c,f):s.subarray(c,f));for(b="";c<f;){var w=s[c++];if(128&w){var I=63&s[c++];if((224&w)==192)b+=String.fromCharCode((31&w)<<6|I);else{var O=63&s[c++];65536>(w=(240&w)==224?(15&w)<<12|I<<6|O:(7&w)<<18|I<<12|O<<6|63&s[c++])?b+=String.fromCharCode(w):(w-=65536,b+=String.fromCharCode(55296|w>>10,56320|1023&w))}}else b+=String.fromCharCode(w)}return b},Ce=(s,c)=>(s>>>=0)?Zo(r(),s,c):"";function Qo(s,c,f){return g?xe(3,1,s,c,f):0}function Xo(s,c){if(g)return xe(4,1,s,c)}var $n=s=>{for(var c=0,f=0;f<s.length;++f){var b=s.charCodeAt(f);127>=b?c++:2047>=b?c+=2:55296<=b&&57343>=b?(c+=4,++f):c+=3}return c},Jo=(s,c,f,b)=>{if(!(0<b))return 0;var w=f>>>=0;b=f+b-1;for(var I=0;I<s.length;++I){var O=s.charCodeAt(I);if(55296<=O&&57343>=O&&(O=65536+((1023&O)<<10)|1023&s.charCodeAt(++I)),127>=O){if(f>=b)break;c[f++>>>0]=O}else{if(2047>=O){if(f+1>=b)break;c[f++>>>0]=192|O>>6}else{if(65535>=O){if(f+2>=b)break;c[f++>>>0]=224|O>>12}else{if(f+3>=b)break;c[f++>>>0]=240|O>>18,c[f++>>>0]=128|O>>12&63}c[f++>>>0]=128|O>>6&63}c[f++>>>0]=128|63&O}}return c[f>>>0]=0,f-w},Ot=(s,c,f)=>Jo(s,r(),c,f);function ei(s,c){if(g)return xe(5,1,s,c)}function ti(s,c,f){if(g)return xe(6,1,s,c,f)}function ri(s,c,f){return g?xe(7,1,s,c,f):0}function ni(s,c){if(g)return xe(8,1,s,c)}function oi(s,c,f){if(g)return xe(9,1,s,c,f)}function ii(s,c,f,b){if(g)return xe(10,1,s,c,f,b)}function ai(s,c,f,b){if(g)return xe(11,1,s,c,f,b)}function si(s,c,f,b){if(g)return xe(12,1,s,c,f,b)}function ui(s){if(g)return xe(13,1,s)}function di(s,c){if(g)return xe(14,1,s,c)}function li(s,c,f){if(g)return xe(15,1,s,c,f)}var ci,mt,Cc=()=>{ct("")},Qe=s=>{for(var c="";r()[s>>>0];)c+=ci[r()[s++>>>0]];return c},xn={},Sn={},Ac={};function st(s,c,f={}){if(!("argPackAdvance"in c))throw new TypeError("registerType registeredInstance requires argPackAdvance");return function(b,w,I={}){var O=w.name;if(!b)throw new mt(`type "${O}" must have a positive integer typeid pointer`);if(Sn.hasOwnProperty(b)){if(I.Tb)return;throw new mt(`Cannot register type '${O}' twice`)}Sn[b]=w,delete Ac[b],xn.hasOwnProperty(b)&&(w=xn[b],delete xn[b],w.forEach(B=>B()))}(s,c,f)}var pi=(s,c,f)=>{switch(c){case 1:return f?b=>t()[b>>>0]:b=>r()[b>>>0];case 2:return f?b=>n()[b>>>1>>>0]:b=>o()[b>>>1>>>0];case 4:return f?b=>i()[b>>>2>>>0]:b=>a()[b>>>2>>>0];case 8:return f?b=>q[b>>>3]:b=>he[b>>>3];default:throw new TypeError(`invalid integer width (${c}): ${s}`)}};function kc(s,c,f){f>>>=0,st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:b=>b,toWireType:function(b,w){if(typeof w!="bigint"&&typeof w!="number")throw w=w===null?"null":(b=typeof w)=="object"||b==="array"||b==="function"?w.toString():""+w,new TypeError(`Cannot convert "${w}" to ${this.name}`);return typeof w=="number"&&(w=BigInt(w)),w},argPackAdvance:ft,readValueFromPointer:pi(c,f,c.indexOf("u")==-1),Eb:null})}var ft=8;function Ec(s,c,f,b){st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:function(w){return!!w},toWireType:function(w,I){return I?f:b},argPackAdvance:ft,readValueFromPointer:function(w){return this.fromWireType(r()[w>>>0])},Eb:null})}var Tn=[],ut=[];function In(s){9<(s>>>=0)&&--ut[s+1]==0&&(ut[s]=void 0,Tn.push(s))}var Me=s=>{if(!s)throw new mt("Cannot use deleted val. handle = "+s);return ut[s]},Ve=s=>{switch(s){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let c=Tn.pop()||ut.length;return ut[c]=s,ut[c+1]=1,c}};function Cn(s){return this.fromWireType(a()[s>>>2>>>0])}var Pc={name:"emscripten::val",fromWireType:s=>{var c=Me(s);return In(s),c},toWireType:(s,c)=>Ve(c),argPackAdvance:ft,readValueFromPointer:Cn,Eb:null};function zc(s){return st(s>>>0,Pc)}var Oc=(s,c)=>{switch(c){case 4:return function(f){return this.fromWireType(d()[f>>>2>>>0])};case 8:return function(f){return this.fromWireType(l()[f>>>3>>>0])};default:throw new TypeError(`invalid float width (${c}): ${s}`)}};function Dc(s,c,f){f>>>=0,st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:b=>b,toWireType:(b,w)=>w,argPackAdvance:ft,readValueFromPointer:Oc(c,f),Eb:null})}function Bc(s,c,f,b,w){if(s>>>=0,f>>>=0,c=Qe(c>>>0),w===-1&&(w=4294967295),w=B=>B,b===0){var I=32-8*f;w=B=>B<<I>>>I}var O=c.includes("unsigned")?function(B,L){return L>>>0}:function(B,L){return L};st(s,{name:c,fromWireType:w,toWireType:O,argPackAdvance:ft,readValueFromPointer:pi(c,f,b!==0),Eb:null})}function Mc(s,c,f){function b(I){var O=a()[I>>>2>>>0];return I=a()[I+4>>>2>>>0],new w(t().buffer,I,O)}var w=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][c];st(s>>>=0,{name:f=Qe(f>>>0),fromWireType:b,argPackAdvance:ft,readValueFromPointer:b},{Tb:!0})}function Rc(s,c){s>>>=0;var f=(c=Qe(c>>>0))==="std::string";st(s,{name:c,fromWireType:function(b){var w=a()[b>>>2>>>0],I=b+4;if(f)for(var O=I,B=0;B<=w;++B){var L=I+B;if(B==w||r()[L>>>0]==0){if(O=Ce(O,L-O),H===void 0)var H=O;else H+=String.fromCharCode(0),H+=O;O=L+1}}else{for(H=Array(w),B=0;B<w;++B)H[B]=String.fromCharCode(r()[I+B>>>0]);H=H.join("")}return Je(b),H},toWireType:function(b,w){w instanceof ArrayBuffer&&(w=new Uint8Array(w));var I=typeof w=="string";if(!(I||w instanceof Uint8Array||w instanceof Uint8ClampedArray||w instanceof Int8Array))throw new mt("Cannot pass non-string to std::string");var O=f&&I?$n(w):w.length,B=fr(4+O+1),L=B+4;if(a()[B>>>2>>>0]=O,f&&I)Ot(w,L,O+1);else if(I)for(I=0;I<O;++I){var H=w.charCodeAt(I);if(255<H)throw Je(L),new mt("String has UTF-16 code units that do not fit in 8 bits");r()[L+I>>>0]=H}else for(I=0;I<O;++I)r()[L+I>>>0]=w[I];return b!==null&&b.push(Je,B),B},argPackAdvance:ft,readValueFromPointer:Cn,Eb(b){Je(b)}})}var mi=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,Uc=(s,c)=>{for(var f=s>>1,b=f+c/2;!(f>=b)&&o()[f>>>0];)++f;if(32<(f<<=1)-s&&mi)return mi.decode(r().slice(s,f));for(f="",b=0;!(b>=c/2);++b){var w=n()[s+2*b>>>1>>>0];if(w==0)break;f+=String.fromCharCode(w)}return f},Nc=(s,c,f)=>{if(f??=2147483647,2>f)return 0;var b=c;f=(f-=2)<2*s.length?f/2:s.length;for(var w=0;w<f;++w){var I=s.charCodeAt(w);n()[c>>>1>>>0]=I,c+=2}return n()[c>>>1>>>0]=0,c-b},Vc=s=>2*s.length,Wc=(s,c)=>{for(var f=0,b="";!(f>=c/4);){var w=i()[s+4*f>>>2>>>0];if(w==0)break;++f,65536<=w?(w-=65536,b+=String.fromCharCode(55296|w>>10,56320|1023&w)):b+=String.fromCharCode(w)}return b},Lc=(s,c,f)=>{if(c>>>=0,f??=2147483647,4>f)return 0;var b=c;f=b+f-4;for(var w=0;w<s.length;++w){var I=s.charCodeAt(w);if(55296<=I&&57343>=I&&(I=65536+((1023&I)<<10)|1023&s.charCodeAt(++w)),i()[c>>>2>>>0]=I,(c+=4)+4>f)break}return i()[c>>>2>>>0]=0,c-b},Gc=s=>{for(var c=0,f=0;f<s.length;++f){var b=s.charCodeAt(f);55296<=b&&57343>=b&&++f,c+=4}return c};function Hc(s,c,f){if(s>>>=0,c>>>=0,f=Qe(f>>>=0),c===2)var b=Uc,w=Nc,I=Vc,O=B=>o()[B>>>1>>>0];else c===4&&(b=Wc,w=Lc,I=Gc,O=B=>a()[B>>>2>>>0]);st(s,{name:f,fromWireType:B=>{for(var L,H=a()[B>>>2>>>0],X=B+4,ce=0;ce<=H;++ce){var ge=B+4+ce*c;ce!=H&&O(ge)!=0||(X=b(X,ge-X),L===void 0?L=X:(L+=String.fromCharCode(0),L+=X),X=ge+c)}return Je(B),L},toWireType:(B,L)=>{if(typeof L!="string")throw new mt(`Cannot pass non-string to C++ string type ${f}`);var H=I(L),X=fr(4+H+c);return a()[X>>>2>>>0]=H/c,w(L,X+4,H+c),B!==null&&B.push(Je,X),X},argPackAdvance:ft,readValueFromPointer:Cn,Eb(B){Je(B)}})}function Fc(s,c){st(s>>>=0,{Ub:!0,name:c=Qe(c>>>0),argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})}var qc=()=>1;function Kc(s){Dn(s>>>0,!y,1,!_,131072,!1),Ho()}var fi=s=>{if(!we)try{if(s(),!(0<wt))try{g?hr(ee):vn(ee)}catch(c){c instanceof yn||c=="unwind"||k(1,c)}}catch(c){c instanceof yn||c=="unwind"||k(1,c)}};function An(s){s>>>=0,typeof Atomics.oc=="function"&&(Atomics.oc(i(),s>>>2,s).value.then(ir),s+=128,Atomics.store(i(),s>>>2,1))}var ir=()=>{var s=Mt();s&&(An(s),fi(Hi))};function jc(s,c){(s>>>=0)==c>>>0?setTimeout(ir):g?postMessage({targetThread:s,cmd:"checkMailbox"}):(s=Ze[s])&&s.postMessage({cmd:"checkMailbox"})}var kn=[];function Yc(s,c,f,b,w){for(c>>>=0,b/=2,kn.length=b,f=w>>>0>>>3,w=0;w<b;w++)kn[w]=q[f+2*w]?q[f+2*w+1]:l()[f+2*w+1>>>0];return(c?bn[c]:Np[s])(...kn)}function Zc(s){s>>>=0,g?postMessage({cmd:"cleanupThread",thread:s}):Go(Ze[s])}function Qc(s){}var ar=(s,c)=>{var f=Sn[s];if(f===void 0)throw s=Vi(s),f=Qe(s),Je(s),new mt(`${c} has unknown type ${f}`);return f},hi=(s,c,f)=>{var b=[];return s=s.toWireType(b,f),b.length&&(a()[c>>>2>>>0]=Ve(b)),s};function Xc(s,c,f){return c>>>=0,f>>>=0,s=Me(s>>>0),c=ar(c,"emval::as"),hi(c,f,s)}function Jc(s,c){return c>>>=0,s=Me(s>>>0),(c=ar(c,"emval::as")).toWireType(null,s)}var sr=s=>{try{s()}catch(c){ct(c)}},ht=0,Xe=null,gi=0,ur=[],bi={},yi={},ep=0,En=null,tp=[];function _i(s){return function(c){if(!we){if(ht===0){var f=!1,b=!1;c((w=0)=>{if(!we&&(gi=w,f=!0,b)){ht=2,sr(()=>Yi(Xe)),typeof Browser<"u"&&Browser.Lb.Sb&&Browser.Lb.resume(),w=!1;try{var I=function(){var L=i()[Xe+8>>>2>>>0];return L=Z[yi[L]],--wt,L()}()}catch(L){I=L,w=!0}var O=!1;if(!Xe){var B=En;B&&(En=null,(w?B.reject:B.resolve)(I),O=!0)}if(w&&!O)throw I}}),b=!0,f||(ht=1,Xe=function(){var w=fr(65548),I=w+12;a()[w>>>2>>>0]=I,a()[w+4>>>2>>>0]=I+65536,I=ur[0];var O=bi[I];return O===void 0&&(O=ep++,bi[I]=O,yi[O]=I),I=O,i()[w+8>>>2>>>0]=I,w}(),typeof Browser<"u"&&Browser.Lb.Sb&&Browser.Lb.pause(),sr(()=>Ki(Xe)))}else ht===2?(ht=0,sr(Zi),Je(Xe),Xe=null,tp.forEach(fi)):ct(`invalid state: ${ht}`);return gi}}(c=>{s().then(c)})}function rp(s){return s>>>=0,_i(()=>(s=Me(s)).then(Ve))}var dr=[];function np(s,c,f,b){return f>>>=0,b>>>=0,(s=dr[s>>>0])(null,c=Me(c>>>0),f,b)}var op={},lr=s=>{var c=op[s];return c===void 0?Qe(s):c};function ip(s,c,f,b,w){return f>>>=0,b>>>=0,w>>>=0,(s=dr[s>>>0])(c=Me(c>>>0),c[f=lr(f)],b,w)}var wi=()=>typeof globalThis=="object"?globalThis:Function("return this")();function ap(s){return(s>>>=0)==0?Ve(wi()):(s=lr(s),Ve(wi()[s]))}var sp=s=>{var c=dr.length;return dr.push(s),c},up=(s,c)=>{for(var f=Array(s),b=0;b<s;++b)f[b]=ar(a()[c+4*b>>>2>>>0],"parameter "+b);return f},vi=(s,c)=>Object.defineProperty(c,"name",{value:s});function dp(s,c,f){var b=(c=up(s,c>>>0)).shift();s--;var w=`return function (obj, func, destructorsRef, args) {
-`,I=0,O=[];f===0&&O.push("obj");for(var B=["retType"],L=[b],H=0;H<s;++H)O.push("arg"+H),B.push("argType"+H),L.push(c[H]),w+=`  var arg${H} = argType${H}.readValueFromPointer(args${I?"+"+I:""});
-`,I+=c[H].argPackAdvance;return w+=`  var rv = ${f===1?"new func":"func.call"}(${O.join(", ")});
-`,b.Ub||(B.push("emval_returnValue"),L.push(hi),w+=`  return emval_returnValue(retType, destructorsRef, rv);
+var Rn=Object.defineProperty;var Vp=Object.getOwnPropertyDescriptor;var Wp=Object.getOwnPropertyNames;var Lp=Object.prototype.hasOwnProperty;var Un=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,r)=>(typeof require<"u"?require:t)[r]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')});var U=(e,t)=>()=>(e&&(t=e(e=0)),t);var Ht=(e,t)=>{for(var r in t)Rn(e,r,{get:t[r],enumerable:!0})},Gp=(e,t,r,n)=>{if(t&&typeof t=="object"||typeof t=="function")for(let o of Wp(t))!Lp.call(e,o)&&o!==r&&Rn(e,o,{get:()=>t[o],enumerable:!(n=Vp(t,o))||n.enumerable});return e};var gr=e=>Gp(Rn({},"__esModule",{value:!0}),e);var br,St,Tt,Hp,Xi,Nn=U(()=>{"use strict";br=new Map,St=[],Tt=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let n=br.get(e);if(n===void 0)br.set(e,{backend:t,priority:r});else{if(n.priority>r)return;if(n.priority===r&&n.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${r}`)}if(r>=0){let o=St.indexOf(e);o!==-1&&St.splice(o,1);for(let i=0;i<St.length;i++)if(br.get(St[i]).priority<=r){St.splice(i,0,e);return}St.push(e)}return}throw new TypeError("not a valid backend")},Hp=async e=>{let t=br.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(n){return r||(t.error=`${n}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},Xi=async e=>{let t=e.executionProviders||[],r=t.map(l=>typeof l=="string"?l:l.name),n=r.length===0?St:r,o,i=[],a=new Set;for(let l of n){let p=await Hp(l);typeof p=="string"?i.push({name:l,err:p}):(o||(o=p),o===p&&a.add(l))}if(!o)throw new Error(`no available backend found. ERR: ${i.map(l=>`[${l.name}] ${l.err}`).join(", ")}`);for(let{name:l,err:p}of i)r.includes(l)&&console.warn(`removing requested execution provider "${l}" from session options because it is not available: ${p}`);let d=t.filter(l=>a.has(typeof l=="string"?l:l.name));return[o,new Proxy(e,{get:(l,p)=>p==="executionProviders"?d:Reflect.get(l,p)})]}});var Ji=U(()=>{"use strict";Nn()});var ea,ta=U(()=>{"use strict";ea="1.21.0-dev.20250206-d981b153d3"});var ra,Ue,Vn=U(()=>{"use strict";ta();ra="warning",Ue={wasm:{},webgl:{},webgpu:{},versions:{common:ea},set logLevel(e){if(e!==void 0){if(typeof e!="string"||["verbose","info","warning","error","fatal"].indexOf(e)===-1)throw new Error(`Unsupported logging level: ${e}`);ra=e}},get logLevel(){return ra}};Object.defineProperty(Ue,"logLevel",{enumerable:!0})});var _e,na=U(()=>{"use strict";Vn();_e=Ue});var oa,ia,aa=U(()=>{"use strict";oa=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let n=r.getContext("2d");if(n!=null){let o,i;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],i=e.dims[3]):(o=e.dims[3],i=e.dims[2]);let a=t?.format!==void 0?t.format:"RGB",d=t?.norm,l,p;d===void 0||d.mean===void 0?l=[255,255,255,255]:typeof d.mean=="number"?l=[d.mean,d.mean,d.mean,d.mean]:(l=[d.mean[0],d.mean[1],d.mean[2],0],d.mean[3]!==void 0&&(l[3]=d.mean[3])),d===void 0||d.bias===void 0?p=[0,0,0,0]:typeof d.bias=="number"?p=[d.bias,d.bias,d.bias,d.bias]:(p=[d.bias[0],d.bias[1],d.bias[2],0],d.bias[3]!==void 0&&(p[3]=d.bias[3]));let m=i*o,u=0,h=m,_=m*2,y=-1;a==="RGBA"?(u=0,h=m,_=m*2,y=m*3):a==="RGB"?(u=0,h=m,_=m*2):a==="RBG"&&(u=0,_=m,h=m*2);for(let g=0;g<i;g++)for(let x=0;x<o;x++){let $=(e.data[u++]-p[0])*l[0],v=(e.data[h++]-p[1])*l[1],S=(e.data[_++]-p[2])*l[2],T=y===-1?255:(e.data[y++]-p[3])*l[3];n.fillStyle="rgba("+$+","+v+","+S+","+T+")",n.fillRect(x,g,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},ia=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let o,i,a;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],i=e.dims[1],a=e.dims[3]):(o=e.dims[3],i=e.dims[2],a=e.dims[1]);let d=t!==void 0&&t.format!==void 0?t.format:"RGB",l=t?.norm,p,m;l===void 0||l.mean===void 0?p=[255,255,255,255]:typeof l.mean=="number"?p=[l.mean,l.mean,l.mean,l.mean]:(p=[l.mean[0],l.mean[1],l.mean[2],255],l.mean[3]!==void 0&&(p[3]=l.mean[3])),l===void 0||l.bias===void 0?m=[0,0,0,0]:typeof l.bias=="number"?m=[l.bias,l.bias,l.bias,l.bias]:(m=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(m[3]=l.bias[3]));let u=i*o;if(t!==void 0&&(t.format!==void 0&&a===4&&t.format!=="RGBA"||a===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let h=4,_=0,y=1,g=2,x=3,$=0,v=u,S=u*2,T=-1;d==="RGBA"?($=0,v=u,S=u*2,T=u*3):d==="RGB"?($=0,v=u,S=u*2):d==="RBG"&&($=0,S=u,v=u*2),n=r.createImageData(o,i);for(let A=0;A<i*o;_+=h,y+=h,g+=h,x+=h,A++)n.data[_]=(e.data[$++]-m[0])*p[0],n.data[y]=(e.data[v++]-m[1])*p[1],n.data[g]=(e.data[S++]-m[2])*p[2],n.data[x]=T===-1?255:(e.data[T++]-m[3])*p[3]}else throw new Error("Can not access image data");return n}});var Wn,sa,ua,da,la,ca,pa=U(()=>{"use strict";yr();Wn=(e,t)=>{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:n}=t,o=t.norm??{mean:255,bias:0},i,a;typeof o.mean=="number"?i=[o.mean,o.mean,o.mean,o.mean]:i=[o.mean[0],o.mean[1],o.mean[2],o.mean[3]??255],typeof o.bias=="number"?a=[o.bias,o.bias,o.bias,o.bias]:a=[o.bias[0],o.bias[1],o.bias[2],o.bias[3]??0];let d=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",p=r*n,m=l==="RGBA"?new Float32Array(p*4):new Float32Array(p*3),u=4,h=0,_=1,y=2,g=3,x=0,$=p,v=p*2,S=-1;d==="RGB"&&(u=3,h=0,_=1,y=2,g=-1),l==="RGBA"?S=p*3:l==="RBG"?(x=0,v=p,$=p*2):l==="BGR"&&(v=0,$=p,x=p*2);for(let A=0;A<p;A++,h+=u,y+=u,_+=u,g+=u)m[x++]=(e[h]+a[0])/i[0],m[$++]=(e[_]+a[1])/i[1],m[v++]=(e[y]+a[2])/i[2],S!==-1&&g!==-1&&(m[S++]=(e[g]+a[3])/i[3]);return l==="RGBA"?new ze("float32",m,[1,4,r,n]):new ze("float32",m,[1,3,r,n])},sa=async(e,t)=>{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,n=typeof ImageData<"u"&&e instanceof ImageData,o=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,i=typeof e=="string",a,d=t??{},l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},p=m=>typeof HTMLCanvasElement<"u"&&m instanceof HTMLCanvasElement||m instanceof OffscreenCanvas?m.getContext("2d"):null;if(r){let m=l();m.width=e.width,m.height=e.height;let u=p(m);if(u!=null){let h=e.height,_=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(h=t.resizedHeight,_=t.resizedWidth),t!==void 0){if(d=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");d.tensorFormat="RGBA",d.height=h,d.width=_}else d.tensorFormat="RGBA",d.height=h,d.width=_;u.drawImage(e,0,0),a=u.getImageData(0,0,_,h).data}else throw new Error("Can not access image data")}else if(n){let m,u;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(m=t.resizedHeight,u=t.resizedWidth):(m=e.height,u=e.width),t!==void 0&&(d=t),d.format="RGBA",d.height=m,d.width=u,t!==void 0){let h=l();h.width=u,h.height=m;let _=p(h);if(_!=null)_.putImageData(e,0,0),a=_.getImageData(0,0,u,m).data;else throw new Error("Can not access image data")}else a=e.data}else if(o){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let m=l();m.width=e.width,m.height=e.height;let u=p(m);if(u!=null){let h=e.height,_=e.width;return u.drawImage(e,0,0,_,h),a=u.getImageData(0,0,_,h).data,d.height=h,d.width=_,Wn(a,d)}else throw new Error("Can not access image data")}else{if(i)return new Promise((m,u)=>{let h=l(),_=p(h);if(!e||!_)return u();let y=new Image;y.crossOrigin="Anonymous",y.src=e,y.onload=()=>{h.width=y.width,h.height=y.height,_.drawImage(y,0,0,h.width,h.height);let g=_.getImageData(0,0,h.width,h.height);d.height=h.height,d.width=h.width,m(Wn(g.data,d))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(a!==void 0)return Wn(a,d);throw new Error("Input data provided is not supported - aborted tensor creation")},ua=(e,t)=>{let{width:r,height:n,download:o,dispose:i}=t,a=[1,n,r,4];return new ze({location:"texture",type:"float32",texture:e,dims:a,download:o,dispose:i})},da=(e,t)=>{let{dataType:r,dims:n,download:o,dispose:i}=t;return new ze({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:n,download:o,dispose:i})},la=(e,t)=>{let{dataType:r,dims:n,download:o,dispose:i}=t;return new ze({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:n,download:o,dispose:i})},ca=(e,t,r)=>new ze({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})});var It,Ft,ma,fa,ha=U(()=>{"use strict";It=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),Ft=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),ma=!1,fa=()=>{if(!ma){ma=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;e&&(It.set("int64",BigInt64Array),Ft.set(BigInt64Array,"int64")),t&&(It.set("uint64",BigUint64Array),Ft.set(BigUint64Array,"uint64")),r?(It.set("float16",Float16Array),Ft.set(Float16Array,"float16")):It.set("float16",Uint16Array)}}});var ga,ba,ya=U(()=>{"use strict";yr();ga=e=>{let t=1;for(let r=0;r<e.length;r++){let n=e[r];if(typeof n!="number"||!Number.isSafeInteger(n))throw new TypeError(`dims[${r}] must be an integer, got: ${n}`);if(n<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${n}`);t*=n}return t},ba=(e,t)=>{switch(e.location){case"cpu":return new ze(e.type,e.data,t);case"cpu-pinned":return new ze({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new ze({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new ze({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new ze({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}});var ze,yr=U(()=>{"use strict";aa();pa();ha();ya();ze=class{constructor(t,r,n){fa();let o,i;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,o=t.type,i=t.dims,t.location){case"cpu-pinned":{let d=It.get(o);if(!d)throw new TypeError(`unsupported type "${o}" to create tensor from pinned buffer`);if(!(t.data instanceof d))throw new TypeError(`buffer should be of type ${d.name}`);this.cpuData=t.data;break}case"texture":{if(o!=="float32")throw new TypeError(`unsupported type "${o}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break}case"gpu-buffer":{if(o!=="float32"&&o!=="float16"&&o!=="int32"&&o!=="int64"&&o!=="uint32"&&o!=="uint8"&&o!=="bool"&&o!=="uint4"&&o!=="int4")throw new TypeError(`unsupported type "${o}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}case"ml-tensor":{if(o!=="float32"&&o!=="float16"&&o!=="int32"&&o!=="int64"&&o!=="uint32"&&o!=="uint64"&&o!=="int8"&&o!=="uint8"&&o!=="bool"&&o!=="uint4"&&o!=="int4")throw new TypeError(`unsupported type "${o}" to create tensor from MLTensor`);this.mlTensorData=t.mlTensor,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let d,l;if(typeof t=="string")if(o=t,l=n,t==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");d=r}else{let p=It.get(t);if(p===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(r)){if(t==="float16"&&p===Uint16Array||t==="uint4"||t==="int4")throw new TypeError(`Creating a ${t} tensor from number array is not supported. Please use ${p.name} as data.`);t==="uint64"||t==="int64"?d=p.from(r,BigInt):d=p.from(r)}else if(r instanceof p)d=r;else if(r instanceof Uint8ClampedArray)if(t==="uint8")d=Uint8Array.from(r);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else throw new TypeError(`A ${o} tensor's data must be type of ${p}`)}else if(l=r,Array.isArray(t)){if(t.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let p=typeof t[0];if(p==="string")o="string",d=t;else if(p==="boolean")o="bool",d=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${p}.`)}else if(t instanceof Uint8ClampedArray)o="uint8",d=Uint8Array.from(t);else{let p=Ft.get(t.constructor);if(p===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);o=p,d=t}if(l===void 0)l=[d.length];else if(!Array.isArray(l))throw new TypeError("A tensor's dims must be a number array");i=l,this.cpuData=d,this.dataLocation="cpu"}let a=ga(i);if(this.cpuData&&a!==this.cpuData.length&&!((o==="uint4"||o==="int4")&&Math.ceil(a/2)===this.cpuData.length))throw new Error(`Tensor's size(${a}) does not match data length(${this.cpuData.length}).`);this.type=o,this.dims=i,this.size=a}static async fromImage(t,r){return sa(t,r)}static fromTexture(t,r){return ua(t,r)}static fromGpuBuffer(t,r){return da(t,r)}static fromMLTensor(t,r){return la(t,r)}static fromPinnedBuffer(t,r,n){return ca(t,r,n)}toDataURL(t){return oa(this,t)}toImageData(t){return ia(this,t)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw new Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(t){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let r=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=r,t&&this.disposer&&(this.disposer(),this.disposer=void 0),r}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return ba(this,t)}}});var Fe,Ln=U(()=>{"use strict";yr();Fe=ze});var _r,_a,Ne,Be,Gn=U(()=>{"use strict";Vn();_r=(e,t)=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||console.timeStamp(`${e}::ORT::${t}`)},_a=(e,t)=>{let r=new Error().stack?.split(/\r\n|\r|\n/g)||[],n=!1;for(let o=0;o<r.length;o++){if(n&&!r[o].includes("TRACE_FUNC")){let i=`FUNC_${e}::${r[o].trim().split(" ")[1]}`;t&&(i+=`::${t}`),_r("CPU",i);return}r[o].includes("TRACE_FUNC")&&(n=!0)}},Ne=e=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||_a("BEGIN",e)},Be=e=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||_a("END",e)}});var wr,wa=U(()=>{"use strict";Nn();Ln();Gn();wr=class e{constructor(t){this.handler=t}async run(t,r,n){Ne();let o={},i={};if(typeof t!="object"||t===null||t instanceof Fe||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let a=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof Fe)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");a=!1;for(let p of r){if(typeof p!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(p)===-1)throw new RangeError(`'fetches' contains invalid output name: ${p}.`);o[p]=null}if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let p=!1,m=Object.getOwnPropertyNames(r);for(let u of this.outputNames)if(m.indexOf(u)!==-1){let h=r[u];(h===null||h instanceof Fe)&&(p=!0,a=!1,o[u]=h)}if(p){if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else i=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let p of this.inputNames)if(typeof t[p]>"u")throw new Error(`input '${p}' is missing in 'feeds'.`);if(a)for(let p of this.outputNames)o[p]=null;let d=await this.handler.run(t,o,i),l={};for(let p in d)if(Object.hasOwnProperty.call(d,p)){let m=d[p];m instanceof Fe?l[p]=m:l[p]=new Fe(m.type,m.data,m.dims)}return Be(),l}async release(){return this.handler.dispose()}static async create(t,r,n,o){Ne();let i,a={};if(typeof t=="string"){if(i=t,typeof r=="object"&&r!==null)a=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof Uint8Array){if(i=t,typeof r=="object"&&r!==null)a=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof ArrayBuffer||typeof SharedArrayBuffer<"u"&&t instanceof SharedArrayBuffer){let m=t,u=0,h=t.byteLength;if(typeof r=="object"&&r!==null)a=r;else if(typeof r=="number"){if(u=r,!Number.isSafeInteger(u))throw new RangeError("'byteOffset' must be an integer.");if(u<0||u>=m.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${m.byteLength}).`);if(h=t.byteLength-u,typeof n=="number"){if(h=n,!Number.isSafeInteger(h))throw new RangeError("'byteLength' must be an integer.");if(h<=0||u+h>m.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${m.byteLength-u}].`);if(typeof o=="object"&&o!==null)a=o;else if(typeof o<"u")throw new TypeError("'options' must be an object.")}else if(typeof n<"u")throw new TypeError("'byteLength' must be a number.")}else if(typeof r<"u")throw new TypeError("'options' must be an object.");i=new Uint8Array(m,u,h)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[d,l]=await Xi(a),p=await d.createInferenceSessionHandler(i,l);return Be(),new e(p)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}});var Fp,va=U(()=>{"use strict";wa();Fp=wr});var $a=U(()=>{"use strict"});var xa=U(()=>{"use strict"});var Sa=U(()=>{"use strict"});var Ta=U(()=>{"use strict"});var Hn={};Ht(Hn,{InferenceSession:()=>Fp,TRACE:()=>_r,TRACE_FUNC_BEGIN:()=>Ne,TRACE_FUNC_END:()=>Be,Tensor:()=>Fe,env:()=>_e,registerBackend:()=>Tt});var Le=U(()=>{"use strict";Ji();na();va();Ln();$a();xa();Gn();Sa();Ta()});var vr=U(()=>{"use strict"});var ka={};Ht(ka,{default:()=>qp});var Ca,Aa,qp,Ea=U(()=>{"use strict";Fn();gt();$r();Ca="ort-wasm-proxy-worker",Aa=globalThis.self?.name===Ca;Aa&&(self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":xr(r.wasm).then(()=>{Sr(r).then(()=>{postMessage({type:t})},n=>{postMessage({type:t,err:n})})},n=>{postMessage({type:t,err:n})});break;case"init-ep":{let{epName:n,env:o}=r;Tr(o,n).then(()=>{postMessage({type:t})},i=>{postMessage({type:t,err:i})});break}case"copy-from":{let{buffer:n}=r,o=qt(n);postMessage({type:t,out:o});break}case"create":{let{model:n,options:o}=r;Ir(n,o).then(i=>{postMessage({type:t,out:i})},i=>{postMessage({type:t,err:i})});break}case"release":Cr(r),postMessage({type:t});break;case"run":{let{sessionId:n,inputIndices:o,inputs:i,outputIndices:a,options:d}=r;Ar(n,o,i,a,new Array(a.length).fill(null),d).then(l=>{l.some(p=>p[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:l},Er([...i,...l]))},l=>{postMessage({type:t,err:l})});break}case"end-profiling":kr(r),postMessage({type:t});break;default:}}catch(n){postMessage({type:t,err:n})}});qp=Aa?null:e=>new Worker(e??Ve,{type:"module",name:Ca})});var za={};Ht(za,{default:()=>Kp});var qn,Pa,Kp,Oa=U(()=>{"use strict";Pa=(qn=import.meta.url,async function(e={}){function t(){return K.buffer!=Q.buffer&&pe(),Q}function r(){return K.buffer!=Q.buffer&&pe(),ie}function n(){return K.buffer!=Q.buffer&&pe(),te}function o(){return K.buffer!=Q.buffer&&pe(),be}function i(){return K.buffer!=Q.buffer&&pe(),Oe}function a(){return K.buffer!=Q.buffer&&pe(),ve}function d(){return K.buffer!=Q.buffer&&pe(),de}function l(){return K.buffer!=Q.buffer&&pe(),he}var p,m,u=Object.assign({},e),h=new Promise((s,c)=>{p=s,m=c}),_=typeof window=="object",y=typeof importScripts=="function",g=y&&self.name=="em-pthread";u.mountExternalData=(s,c)=>{s.startsWith("./")&&(s=s.substring(2)),(u.Fb||(u.Fb=new Map)).set(s,c)},u.unmountExternalData=()=>{delete u.Fb};var x=globalThis.SharedArrayBuffer??new WebAssembly.Memory({initial:0,maximum:0,shared:!0}).buffer.constructor;let $=()=>{let s=(f,b,w)=>(...I)=>{let O=Xe,B=b?.();I=f(...I);let G=b?.();return B!==G&&(f=G,w(B),b=w=null),Xe!=O?new Promise((F,X)=>{kn={resolve:F,reject:X}}):I},c=f=>async(...b)=>{try{if(u.Gb)throw Error("Session already started");let w=u.Gb={hc:b[0],errors:[]},I=await f(...b);if(u.Gb!==w)throw Error("Session mismatch");u.Hb?.flush();let O=w.errors;if(0<O.length){let B=await Promise.all(O);if(B=B.filter(G=>G),0<B.length)throw Error(B.join(`
+`))}return I}finally{u.Gb=null}};u._OrtCreateSession=s(u._OrtCreateSession,()=>u._OrtCreateSession,f=>u._OrtCreateSession=f),u._OrtRun=c(s(u._OrtRun,()=>u._OrtRun,f=>u._OrtRun=f)),u._OrtRunWithBinding=c(s(u._OrtRunWithBinding,()=>u._OrtRunWithBinding,f=>u._OrtRunWithBinding=f)),u._OrtBindInput=s(u._OrtBindInput,()=>u._OrtBindInput,f=>u._OrtBindInput=f),$=void 0};u.jsepInit=(s,c)=>{if($?.(),s==="webgpu"){[u.Hb,u.Vb,u.Zb,u.Ob,u.Yb,u.kb,u.$b,u.cc,u.Wb,u.Xb,u.ac]=c;let f=u.Hb;u.jsepRegisterBuffer=(b,w,I,O)=>f.registerBuffer(b,w,I,O),u.jsepGetBuffer=b=>f.getBuffer(b),u.jsepCreateDownloader=(b,w,I)=>f.createDownloader(b,w,I),u.jsepOnCreateSession=b=>{f.onCreateSession(b)},u.jsepOnReleaseSession=b=>{f.onReleaseSession(b)},u.jsepOnRunStart=b=>f.onRunStart(b),u.dc=(b,w)=>{f.upload(b,w)}}else if(s==="webnn"){[u.Hb,u.bc,u.Pb,u.jsepEnsureTensor,u.ec,u.jsepDownloadTensor]=c,u.jsepReleaseTensorId=u.Pb;let f=u.Hb;u.jsepOnRunStart=b=>f.onRunStart(b),u.jsepRegisterMLContext=(b,w)=>{f.registerMLContext(b,w)},u.jsepOnReleaseSession=b=>{f.onReleaseSession(b)},u.jsepCreateMLTensorDownloader=(b,w)=>f.createMLTensorDownloader(b,w),u.jsepRegisterMLTensor=(b,w,I)=>f.registerMLTensor(b,w,I),u.jsepCreateMLContext=b=>f.createMLContext(b),u.jsepRegisterMLConstant=(b,w,I,O,B)=>f.registerMLConstant(b,w,I,O,B,u.Fb)}};var v,S,T=Object.assign({},u),A=(s,c)=>{throw c},C="";(_||y)&&(y?C=self.location.href:typeof document<"u"&&document.currentScript&&(C=document.currentScript.src),qn&&(C=qn),C=C.startsWith("blob:")?"":C.substr(0,C.replace(/[?#].*/,"").lastIndexOf("/")+1),y&&(S=s=>{var c=new XMLHttpRequest;return c.open("GET",s,!1),c.responseType="arraybuffer",c.send(null),new Uint8Array(c.response)}),v=(s,c,f)=>{var b=new XMLHttpRequest;b.open("GET",s,!0),b.responseType="arraybuffer",b.onload=()=>{b.status==200||b.status==0&&b.response?c(b.response):f()},b.onerror=f,b.send(null)});var P,D=console.log.bind(console),R=console.error.bind(console),H=D,L=R;if(Object.assign(u,T),T=null,g){let s=function(c){try{var f=c.data,b=f.cmd;if(b==="load"){let w=[];self.onmessage=I=>w.push(I),self.startWorker=()=>{postMessage({cmd:"loaded"});for(let I of w)s(I);self.onmessage=s};for(let I of f.handlers)u[I]&&!u[I].proxy||(u[I]=(...O)=>{postMessage({Nb:"callHandler",pc:I,args:O})},I=="print"&&(H=u[I]),I=="printErr"&&(L=u[I]));K=f.wasmMemory,pe(),re(f.wasmModule)}else if(b==="run"){On(f.pthread_ptr,0,0,1,0,0),Cn(f.pthread_ptr),xc(),Go(),V||(Vi(),V=!0);try{Sc(f.start_routine,f.arg)}catch(w){if(w!="unwind")throw w}}else b==="cancel"?Rt()&&fr(-1):f.target!=="setimmediate"&&(b==="checkMailbox"?V&&or():b&&(L(`worker: received unknown command ${b}`),L(f)))}catch(w){throw Wi(),w}};var wg=s,re,V=!1;L=function(...c){c=c.join(" "),console.error(c)},self.alert=function(...c){postMessage({Nb:"alert",text:c.join(" "),rc:Rt()})},u.instantiateWasm=(c,f)=>new Promise(b=>{re=w=>{w=new WebAssembly.Instance(w,Uo()),f(w),b()}}),self.onunhandledrejection=c=>{throw c.reason||c},self.onmessage=s}u.wasmBinary&&(P=u.wasmBinary);var K,we,j,Q,ie,te,be,Oe,ve,de,W,Y,he,De=!1;function pe(){var s=K.buffer;u.HEAP8=Q=new Int8Array(s),u.HEAP16=te=new Int16Array(s),u.HEAPU8=ie=new Uint8Array(s),u.HEAPU16=be=new Uint16Array(s),u.HEAP32=Oe=new Int32Array(s),u.HEAPU32=ve=new Uint32Array(s),u.HEAPF32=de=new Float32Array(s),u.HEAPF64=he=new Float64Array(s),u.HEAP64=W=new BigInt64Array(s),u.HEAPU64=Y=new BigUint64Array(s)}if(!g){if(!((K=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0})).buffer instanceof x))throw L("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),Error("bad memory");pe()}var Ie=[],He=[],mn=[],wt=0,fn=null,Gt=null;function Oo(){if(--wt==0&&(fn!==null&&(clearInterval(fn),fn=null),Gt)){var s=Gt;Gt=null,s()}}function ct(s){throw L(s="Aborted("+s+")"),De=!0,j=1,s=new WebAssembly.RuntimeError(s+". Build with -sASSERTIONS for more info."),m(s),s}var hn,Do=s=>s.startsWith("data:application/octet-stream;base64,"),Bo=s=>s.startsWith("file://");function Mo(s){if(s==hn&&P)return new Uint8Array(P);if(S)return S(s);throw"both async and sync fetching of the wasm failed"}function Ro(s,c,f){return function(b){if(!P&&(_||y)){if(typeof fetch=="function"&&!Bo(b))return fetch(b,{credentials:"same-origin"}).then(w=>{if(!w.ok)throw`failed to load wasm binary file at '${b}'`;return w.arrayBuffer()}).catch(()=>Mo(b));if(v)return new Promise((w,I)=>{v(b,O=>w(new Uint8Array(O)),I)})}return Promise.resolve().then(()=>Mo(b))}(s).then(b=>WebAssembly.instantiate(b,c)).then(f,b=>{L(`failed to asynchronously prepare wasm: ${b}`),ct(b)})}function Uo(){return{a:{O:$c,Aa:vc,b:Ic,aa:Ko,B:Zo,qa:Qo,Y:Jo,_:ei,ra:ti,oa:ri,ha:ni,na:oi,L:ii,Z:ai,W:si,pa:ui,X:di,va:Cc,F:kc,Q:Ec,P:zc,E:Dc,u:Bc,q:Mc,G:Rc,A:Hc,R:Fc,ua:qc,ka:Kc,U:jc,ba:Yc,H:Zc,ja:Cn,ta:Qc,t:Xc,Ba:Jc,x:rp,o:np,m:ip,c:Tn,n:ap,k:dp,w:lp,p:cp,f:pp,s:mp,l:fp,e:hp,j:gp,i:bp,g:yp,d:_p,ea:wp,fa:vp,ga:$p,ca:xi,da:Si,T:xp,h:Sp,D:Tp,I:Ip,M:Cp,y:Ap,sa:kp,V:Ep,v:Ii,z:Pp,N:zp,S:Op,za:Dp,ya:Bp,la:ki,ma:Ei,$:wn,C:Pi,K:zi,ia:Oi,J:Di,a:K,xa:_n,wa:Ri,r:Up}}}var gn={916868:(s,c,f,b,w)=>{if(u===void 0||!u.Fb)return 1;if((s=Ce(Number(s>>>0))).startsWith("./")&&(s=s.substring(2)),!(s=u.Fb.get(s)))return 2;if(c=Number(c>>>0),f=Number(f>>>0),b=Number(b>>>0),c+f>s.byteLength)return 3;try{let I=s.subarray(c,c+f);switch(w){case 0:r().set(I,b>>>0);break;case 1:u.dc(b,I);break;default:return 4}return 0}catch{return 4}},917583:(s,c,f)=>{u.ec(s,r().subarray(c>>>0,c+f>>>0))},917646:()=>u.bc(),917687:s=>{u.Pb(s)},917723:()=>{u.Wb()},917754:()=>{u.Xb()},917783:()=>{u.ac()},917808:s=>u.Vb(s),917841:s=>u.Zb(s),917873:(s,c,f)=>{u.Ob(Number(s),Number(c),Number(f),!0)},917936:(s,c,f)=>{u.Ob(Number(s),Number(c),Number(f))},917993:()=>typeof wasmOffsetConverter<"u",918050:s=>{u.kb("Abs",s,void 0)},918101:s=>{u.kb("Neg",s,void 0)},918152:s=>{u.kb("Floor",s,void 0)},918205:s=>{u.kb("Ceil",s,void 0)},918257:s=>{u.kb("Reciprocal",s,void 0)},918315:s=>{u.kb("Sqrt",s,void 0)},918367:s=>{u.kb("Exp",s,void 0)},918418:s=>{u.kb("Erf",s,void 0)},918469:s=>{u.kb("Sigmoid",s,void 0)},918524:(s,c,f)=>{u.kb("HardSigmoid",s,{alpha:c,beta:f})},918603:s=>{u.kb("Log",s,void 0)},918654:s=>{u.kb("Sin",s,void 0)},918705:s=>{u.kb("Cos",s,void 0)},918756:s=>{u.kb("Tan",s,void 0)},918807:s=>{u.kb("Asin",s,void 0)},918859:s=>{u.kb("Acos",s,void 0)},918911:s=>{u.kb("Atan",s,void 0)},918963:s=>{u.kb("Sinh",s,void 0)},919015:s=>{u.kb("Cosh",s,void 0)},919067:s=>{u.kb("Asinh",s,void 0)},919120:s=>{u.kb("Acosh",s,void 0)},919173:s=>{u.kb("Atanh",s,void 0)},919226:s=>{u.kb("Tanh",s,void 0)},919278:s=>{u.kb("Not",s,void 0)},919329:(s,c,f)=>{u.kb("Clip",s,{min:c,max:f})},919398:s=>{u.kb("Clip",s,void 0)},919450:(s,c)=>{u.kb("Elu",s,{alpha:c})},919508:s=>{u.kb("Gelu",s,void 0)},919560:s=>{u.kb("Relu",s,void 0)},919612:(s,c)=>{u.kb("LeakyRelu",s,{alpha:c})},919676:(s,c)=>{u.kb("ThresholdedRelu",s,{alpha:c})},919746:(s,c)=>{u.kb("Cast",s,{to:c})},919804:s=>{u.kb("Add",s,void 0)},919855:s=>{u.kb("Sub",s,void 0)},919906:s=>{u.kb("Mul",s,void 0)},919957:s=>{u.kb("Div",s,void 0)},920008:s=>{u.kb("Pow",s,void 0)},920059:s=>{u.kb("Equal",s,void 0)},920112:s=>{u.kb("Greater",s,void 0)},920167:s=>{u.kb("GreaterOrEqual",s,void 0)},920229:s=>{u.kb("Less",s,void 0)},920281:s=>{u.kb("LessOrEqual",s,void 0)},920340:(s,c,f,b,w)=>{u.kb("ReduceMean",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},920515:(s,c,f,b,w)=>{u.kb("ReduceMax",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},920689:(s,c,f,b,w)=>{u.kb("ReduceMin",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},920863:(s,c,f,b,w)=>{u.kb("ReduceProd",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},921038:(s,c,f,b,w)=>{u.kb("ReduceSum",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},921212:(s,c,f,b,w)=>{u.kb("ReduceL1",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},921385:(s,c,f,b,w)=>{u.kb("ReduceL2",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},921558:(s,c,f,b,w)=>{u.kb("ReduceLogSum",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},921735:(s,c,f,b,w)=>{u.kb("ReduceSumSquare",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},921915:(s,c,f,b,w)=>{u.kb("ReduceLogSumExp",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},922095:s=>{u.kb("Where",s,void 0)},922148:(s,c,f)=>{u.kb("Transpose",s,{perm:c?Array.from(i().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},922272:(s,c,f,b)=>{u.kb("DepthToSpace",s,{blocksize:c,mode:Ce(f),format:b?"NHWC":"NCHW"})},922405:(s,c,f,b)=>{u.kb("DepthToSpace",s,{blocksize:c,mode:Ce(f),format:b?"NHWC":"NCHW"})},922538:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z,ue)=>{u.kb("ConvTranspose",s,{format:G?"NHWC":"NCHW",autoPad:c,dilations:[f],group:b,kernelShape:[w],pads:[I,O],strides:[B],wIsConst:()=>!!t()[F>>>0],outputPadding:X?Array.from(i().subarray(Number(X)>>>0,Number(le)>>>0)):[],outputShape:ge?Array.from(i().subarray(Number(ge)>>>0,Number(z)>>>0)):[],activation:Ce(ue)})},922971:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z)=>{u.kb("ConvTranspose",s,{format:B?"NHWC":"NCHW",autoPad:c,dilations:Array.from(i().subarray(Number(f)>>>0,2+(Number(f)>>>0)>>>0)),group:b,kernelShape:Array.from(i().subarray(Number(w)>>>0,2+(Number(w)>>>0)>>>0)),pads:Array.from(i().subarray(Number(I)>>>0,4+(Number(I)>>>0)>>>0)),strides:Array.from(i().subarray(Number(O)>>>0,2+(Number(O)>>>0)>>>0)),wIsConst:()=>!!t()[G>>>0],outputPadding:F?Array.from(i().subarray(Number(F)>>>0,Number(X)>>>0)):[],outputShape:le?Array.from(i().subarray(Number(le)>>>0,Number(ge)>>>0)):[],activation:Ce(z)})},923632:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z,ue)=>{u.kb("ConvTranspose",s,{format:G?"NHWC":"NCHW",autoPad:c,dilations:[f],group:b,kernelShape:[w],pads:[I,O],strides:[B],wIsConst:()=>!!t()[F>>>0],outputPadding:X?Array.from(i().subarray(Number(X)>>>0,Number(le)>>>0)):[],outputShape:ge?Array.from(i().subarray(Number(ge)>>>0,Number(z)>>>0)):[],activation:Ce(ue)})},924065:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z)=>{u.kb("ConvTranspose",s,{format:B?"NHWC":"NCHW",autoPad:c,dilations:Array.from(i().subarray(Number(f)>>>0,2+(Number(f)>>>0)>>>0)),group:b,kernelShape:Array.from(i().subarray(Number(w)>>>0,2+(Number(w)>>>0)>>>0)),pads:Array.from(i().subarray(Number(I)>>>0,4+(Number(I)>>>0)>>>0)),strides:Array.from(i().subarray(Number(O)>>>0,2+(Number(O)>>>0)>>>0)),wIsConst:()=>!!t()[G>>>0],outputPadding:F?Array.from(i().subarray(Number(F)>>>0,Number(X)>>>0)):[],outputShape:le?Array.from(i().subarray(Number(le)>>>0,Number(ge)>>>0)):[],activation:Ce(z)})},924726:(s,c)=>{u.kb("GlobalAveragePool",s,{format:c?"NHWC":"NCHW"})},924817:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z)=>{u.kb("AveragePool",s,{format:z?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:w,dilations:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],kernel_shape:B?Array.from(i().subarray(Number(B)>>>0,Number(G)>>>0)):[],pads:F?Array.from(i().subarray(Number(F)>>>0,Number(X)>>>0)):[],strides:le?Array.from(i().subarray(Number(le)>>>0,Number(ge)>>>0)):[]})},925296:(s,c)=>{u.kb("GlobalAveragePool",s,{format:c?"NHWC":"NCHW"})},925387:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z)=>{u.kb("AveragePool",s,{format:z?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:w,dilations:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],kernel_shape:B?Array.from(i().subarray(Number(B)>>>0,Number(G)>>>0)):[],pads:F?Array.from(i().subarray(Number(F)>>>0,Number(X)>>>0)):[],strides:le?Array.from(i().subarray(Number(le)>>>0,Number(ge)>>>0)):[]})},925866:(s,c)=>{u.kb("GlobalMaxPool",s,{format:c?"NHWC":"NCHW"})},925953:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z)=>{u.kb("MaxPool",s,{format:z?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:w,dilations:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],kernel_shape:B?Array.from(i().subarray(Number(B)>>>0,Number(G)>>>0)):[],pads:F?Array.from(i().subarray(Number(F)>>>0,Number(X)>>>0)):[],strides:le?Array.from(i().subarray(Number(le)>>>0,Number(ge)>>>0)):[]})},926428:(s,c)=>{u.kb("GlobalMaxPool",s,{format:c?"NHWC":"NCHW"})},926515:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z)=>{u.kb("MaxPool",s,{format:z?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:w,dilations:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],kernel_shape:B?Array.from(i().subarray(Number(B)>>>0,Number(G)>>>0)):[],pads:F?Array.from(i().subarray(Number(F)>>>0,Number(X)>>>0)):[],strides:le?Array.from(i().subarray(Number(le)>>>0,Number(ge)>>>0)):[]})},926990:(s,c,f,b,w)=>{u.kb("Gemm",s,{alpha:c,beta:f,transA:b,transB:w})},927094:s=>{u.kb("MatMul",s,void 0)},927148:(s,c,f,b)=>{u.kb("ArgMax",s,{keepDims:!!c,selectLastIndex:!!f,axis:b})},927256:(s,c,f,b)=>{u.kb("ArgMin",s,{keepDims:!!c,selectLastIndex:!!f,axis:b})},927364:(s,c)=>{u.kb("Softmax",s,{axis:c})},927427:(s,c)=>{u.kb("Concat",s,{axis:c})},927487:(s,c,f,b,w)=>{u.kb("Split",s,{axis:c,numOutputs:f,splitSizes:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},927643:s=>{u.kb("Expand",s,void 0)},927697:(s,c)=>{u.kb("Gather",s,{axis:Number(c)})},927768:(s,c)=>{u.kb("GatherElements",s,{axis:Number(c)})},927847:(s,c)=>{u.kb("GatherND",s,{batch_dims:Number(c)})},927926:(s,c,f,b,w,I,O,B,G,F,X)=>{u.kb("Resize",s,{antialias:c,axes:f?Array.from(i().subarray(Number(f)>>>0,Number(b)>>>0)):[],coordinateTransformMode:Ce(w),cubicCoeffA:I,excludeOutside:O,extrapolationValue:B,keepAspectRatioPolicy:Ce(G),mode:Ce(F),nearestMode:Ce(X)})},928288:(s,c,f,b,w,I,O)=>{u.kb("Slice",s,{starts:c?Array.from(i().subarray(Number(c)>>>0,Number(f)>>>0)):[],ends:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[],axes:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[]})},928552:s=>{u.kb("Tile",s,void 0)},928604:(s,c,f)=>{u.kb("InstanceNormalization",s,{epsilon:c,format:f?"NHWC":"NCHW"})},928718:(s,c,f)=>{u.kb("InstanceNormalization",s,{epsilon:c,format:f?"NHWC":"NCHW"})},928832:s=>{u.kb("Range",s,void 0)},928885:(s,c)=>{u.kb("Einsum",s,{equation:Ce(c)})},928966:(s,c,f,b,w)=>{u.kb("Pad",s,{mode:c,value:f,pads:b?Array.from(i().subarray(Number(b)>>>0,Number(w)>>>0)):[]})},929109:(s,c,f,b,w,I)=>{u.kb("BatchNormalization",s,{epsilon:c,momentum:f,spatial:!!w,trainingMode:!!b,format:I?"NHWC":"NCHW"})},929278:(s,c,f,b,w,I)=>{u.kb("BatchNormalization",s,{epsilon:c,momentum:f,spatial:!!w,trainingMode:!!b,format:I?"NHWC":"NCHW"})},929447:(s,c,f)=>{u.kb("CumSum",s,{exclusive:Number(c),reverse:Number(f)})},929544:(s,c,f)=>{u.kb("DequantizeLinear",s,{axis:c,blockSize:f})},929634:(s,c,f,b,w)=>{u.kb("GridSample",s,{align_corners:c,mode:Ce(f),padding_mode:Ce(b),format:w?"NHWC":"NCHW"})},929804:(s,c,f,b,w)=>{u.kb("GridSample",s,{align_corners:c,mode:Ce(f),padding_mode:Ce(b),format:w?"NHWC":"NCHW"})},929974:(s,c,f,b,w,I,O,B,G)=>{u.kb("Attention",s,{numHeads:c,isUnidirectional:f,maskFilterValue:b,scale:w,doRotary:I,qkvHiddenSizes:O?Array.from(i().subarray(Number(B)>>>0,Number(B)+O>>>0)):[],pastPresentShareBuffer:!!G})},930246:s=>{u.kb("BiasAdd",s,void 0)},930301:s=>{u.kb("BiasSplitGelu",s,void 0)},930362:s=>{u.kb("FastGelu",s,void 0)},930418:(s,c,f,b,w,I,O,B,G,F,X,le,ge,z,ue,Se)=>{u.kb("Conv",s,{format:le?"NHWC":"NCHW",auto_pad:c,dilations:f?Array.from(i().subarray(Number(f)>>>0,Number(b)>>>0)):[],group:w,kernel_shape:I?Array.from(i().subarray(Number(I)>>>0,Number(O)>>>0)):[],pads:B?Array.from(i().subarray(Number(B)>>>0,Number(G)>>>0)):[],strides:F?Array.from(i().subarray(Number(F)>>>0,Number(X)>>>0)):[],w_is_const:()=>!!t()[Number(ge)>>>0],activation:Ce(z),activation_params:ue?Array.from(d().subarray(Number(ue)>>>0,Number(Se)>>>0)):[]})},931002:s=>{u.kb("Gelu",s,void 0)},931054:(s,c,f,b,w,I,O,B,G)=>{u.kb("GroupQueryAttention",s,{numHeads:c,kvNumHeads:f,scale:b,softcap:w,doRotary:I,rotaryInterleaved:O,smoothSoftmax:B,localWindowSize:G})},931271:(s,c,f,b)=>{u.kb("LayerNormalization",s,{axis:c,epsilon:f,simplified:!!b})},931382:(s,c,f,b)=>{u.kb("LayerNormalization",s,{axis:c,epsilon:f,simplified:!!b})},931493:(s,c,f,b,w,I)=>{u.kb("MatMulNBits",s,{k:c,n:f,accuracyLevel:b,bits:w,blockSize:I})},931620:(s,c,f,b,w,I)=>{u.kb("MultiHeadAttention",s,{numHeads:c,isUnidirectional:f,maskFilterValue:b,scale:w,doRotary:I})},931779:(s,c)=>{u.kb("QuickGelu",s,{alpha:c})},931843:(s,c,f,b,w)=>{u.kb("RotaryEmbedding",s,{interleaved:!!c,numHeads:f,rotaryEmbeddingDim:b,scale:w})},931982:(s,c,f)=>{u.kb("SkipLayerNormalization",s,{epsilon:c,simplified:!!f})},932084:(s,c,f)=>{u.kb("SkipLayerNormalization",s,{epsilon:c,simplified:!!f})},932186:(s,c,f,b)=>{u.kb("GatherBlockQuantized",s,{gatherAxis:c,quantizeAxis:f,blockSize:b})},932307:s=>{u.$b(s)},932341:(s,c)=>u.cc(Number(s),Number(c),u.Gb.hc,u.Gb.errors)};function vc(s,c,f){return yi(async()=>{await u.Yb(Number(s),Number(c),Number(f))})}function $c(){return typeof wasmOffsetConverter<"u"}function bn(s){this.name="ExitStatus",this.message=`Program terminated with exit(${s})`,this.status=s}var yn=s=>{s.terminate(),s.onmessage=()=>{}},No=s=>{pt.length==0&&(Fo(),Ho(pt[0]));var c=pt.pop();if(!c)return 6;$t.push(c),Ze[s.Bb]=c,c.Bb=s.Bb;var f={cmd:"run",start_routine:s.ic,arg:s.Rb,pthread_ptr:s.Bb};return c.postMessage(f,s.nc),0},vt=0,$e=(s,c,...f)=>{for(var b=2*f.length,w=Mn(),I=Bn(8*b),O=I>>>3,B=0;B<f.length;B++){var G=f[B];typeof G=="bigint"?(W[O+2*B]=1n,W[O+2*B+1]=G):(W[O+2*B]=0n,l()[O+2*B+1>>>0]=G)}return s=Li(s,0,b,I,c),hr(w),s};function _n(s){if(g)return $e(0,1,s);if(j=s,!(0<vt)){for(var c of $t)yn(c);for(c of pt)yn(c);pt=[],$t=[],Ze=[],De=!0}A(0,new bn(s))}function Vo(s){if(g)return $e(1,0,s);wn(s)}var wn=s=>{if(j=s,g)throw Vo(s),"unwind";_n(s)},pt=[],$t=[],Wo=[],Ze={},Lo=s=>{var c=s.Bb;delete Ze[c],pt.push(s),$t.splice($t.indexOf(s),1),s.Bb=0,Dn(c)};function Go(){Wo.forEach(s=>s())}var Ho=s=>new Promise(c=>{s.onmessage=w=>{var I=(w=w.data).cmd;if(w.targetThread&&w.targetThread!=Rt()){var O=Ze[w.targetThread];O?O.postMessage(w,w.transferList):L(`Internal error! Worker sent a message "${I}" to target pthread ${w.targetThread}, but that thread no longer exists!`)}else I==="checkMailbox"?or():I==="spawnThread"?No(w):I==="cleanupThread"?Lo(Ze[w.thread]):I==="killThread"?(w=w.thread,I=Ze[w],delete Ze[w],yn(I),Dn(w),$t.splice($t.indexOf(I),1),I.Bb=0):I==="cancelThread"?Ze[w.thread].postMessage({cmd:"cancel"}):I==="loaded"?(s.loaded=!0,c(s)):I==="alert"?alert(`Thread ${w.threadId}: ${w.text}`):w.target==="setimmediate"?s.postMessage(w):I==="callHandler"?u[w.handler](...w.args):I&&L(`worker sent an unknown command ${I}`)},s.onerror=w=>{throw L(`worker sent an error! ${w.filename}:${w.lineno}: ${w.message}`),w};var f,b=[];for(f of[])u.hasOwnProperty(f)&&b.push(f);s.postMessage({cmd:"load",handlers:b,wasmMemory:K,wasmModule:we})});function Fo(){var s=new Worker(import.meta.url.startsWith("file:")?new URL(/* asset import */ __webpack_require__(/*! ort.bundle.min.mjs */ "./node_modules/onnxruntime-web/dist/ort.bundle.min.mjs?46eb"), __webpack_require__.b):new URL(import.meta.url),{type:"module",workerData:"em-pthread",name:"em-pthread"});pt.push(s)}var nr=s=>{for(;0<s.length;)s.shift()(u)},xc=()=>{var s=Rt(),c=a()[s+52>>>2>>>0];s=a()[s+56>>>2>>>0],Hi(c,c-s),hr(c)},Sc=(s,c)=>{vt=0,s=Fi(s,c),0<vt?j=s:fr(s)};class Tc{constructor(c){this.Kb=c-24}}function Ic(s,c,f){var b=new Tc(s>>>=0);throw c>>>=0,f>>>=0,a()[b.Kb+16>>>2>>>0]=0,a()[b.Kb+4>>>2>>>0]=c,a()[b.Kb+8>>>2>>>0]=f,s}function qo(s,c,f,b){return g?$e(2,1,s,c,f,b):Ko(s,c,f,b)}function Ko(s,c,f,b){if(s>>>=0,c>>>=0,f>>>=0,b>>>=0,x===void 0)return L("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var w=[];return g&&w.length===0?qo(s,c,f,b):(s={ic:f,Bb:s,Rb:b,nc:w},g?(s.Nb="spawnThread",postMessage(s,w),0):No(s))}var jo=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Yo=(s,c,f)=>{var b=(c>>>=0)+f;for(f=c;s[f]&&!(f>=b);)++f;if(16<f-c&&s.buffer&&jo)return jo.decode(s.buffer instanceof x?s.slice(c,f):s.subarray(c,f));for(b="";c<f;){var w=s[c++];if(128&w){var I=63&s[c++];if((224&w)==192)b+=String.fromCharCode((31&w)<<6|I);else{var O=63&s[c++];65536>(w=(240&w)==224?(15&w)<<12|I<<6|O:(7&w)<<18|I<<12|O<<6|63&s[c++])?b+=String.fromCharCode(w):(w-=65536,b+=String.fromCharCode(55296|w>>10,56320|1023&w))}}else b+=String.fromCharCode(w)}return b},Ce=(s,c)=>(s>>>=0)?Yo(r(),s,c):"";function Zo(s,c,f){return g?$e(3,1,s,c,f):0}function Qo(s,c){if(g)return $e(4,1,s,c)}var vn=s=>{for(var c=0,f=0;f<s.length;++f){var b=s.charCodeAt(f);127>=b?c++:2047>=b?c+=2:55296<=b&&57343>=b?(c+=4,++f):c+=3}return c},Xo=(s,c,f,b)=>{if(!(0<b))return 0;var w=f>>>=0;b=f+b-1;for(var I=0;I<s.length;++I){var O=s.charCodeAt(I);if(55296<=O&&57343>=O&&(O=65536+((1023&O)<<10)|1023&s.charCodeAt(++I)),127>=O){if(f>=b)break;c[f++>>>0]=O}else{if(2047>=O){if(f+1>=b)break;c[f++>>>0]=192|O>>6}else{if(65535>=O){if(f+2>=b)break;c[f++>>>0]=224|O>>12}else{if(f+3>=b)break;c[f++>>>0]=240|O>>18,c[f++>>>0]=128|O>>12&63}c[f++>>>0]=128|O>>6&63}c[f++>>>0]=128|63&O}}return c[f>>>0]=0,f-w},Dt=(s,c,f)=>Xo(s,r(),c,f);function Jo(s,c){if(g)return $e(5,1,s,c)}function ei(s,c,f){if(g)return $e(6,1,s,c,f)}function ti(s,c,f){return g?$e(7,1,s,c,f):0}function ri(s,c){if(g)return $e(8,1,s,c)}function ni(s,c,f){if(g)return $e(9,1,s,c,f)}function oi(s,c,f,b){if(g)return $e(10,1,s,c,f,b)}function ii(s,c,f,b){if(g)return $e(11,1,s,c,f,b)}function ai(s,c,f,b){if(g)return $e(12,1,s,c,f,b)}function si(s){if(g)return $e(13,1,s)}function ui(s,c){if(g)return $e(14,1,s,c)}function di(s,c,f){if(g)return $e(15,1,s,c,f)}var li,mt,Cc=()=>{ct("")},Qe=s=>{for(var c="";r()[s>>>0];)c+=li[r()[s++>>>0]];return c},$n={},xn={},Ac={};function st(s,c,f={}){if(!("argPackAdvance"in c))throw new TypeError("registerType registeredInstance requires argPackAdvance");return function(b,w,I={}){var O=w.name;if(!b)throw new mt(`type "${O}" must have a positive integer typeid pointer`);if(xn.hasOwnProperty(b)){if(I.Tb)return;throw new mt(`Cannot register type '${O}' twice`)}xn[b]=w,delete Ac[b],$n.hasOwnProperty(b)&&(w=$n[b],delete $n[b],w.forEach(B=>B()))}(s,c,f)}var ci=(s,c,f)=>{switch(c){case 1:return f?b=>t()[b>>>0]:b=>r()[b>>>0];case 2:return f?b=>n()[b>>>1>>>0]:b=>o()[b>>>1>>>0];case 4:return f?b=>i()[b>>>2>>>0]:b=>a()[b>>>2>>>0];case 8:return f?b=>W[b>>>3]:b=>Y[b>>>3];default:throw new TypeError(`invalid integer width (${c}): ${s}`)}};function kc(s,c,f){f>>>=0,st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:b=>b,toWireType:function(b,w){if(typeof w!="bigint"&&typeof w!="number")throw w=w===null?"null":(b=typeof w)=="object"||b==="array"||b==="function"?w.toString():""+w,new TypeError(`Cannot convert "${w}" to ${this.name}`);return typeof w=="number"&&(w=BigInt(w)),w},argPackAdvance:ft,readValueFromPointer:ci(c,f,c.indexOf("u")==-1),Eb:null})}var ft=8;function Ec(s,c,f,b){st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:function(w){return!!w},toWireType:function(w,I){return I?f:b},argPackAdvance:ft,readValueFromPointer:function(w){return this.fromWireType(r()[w>>>0])},Eb:null})}var Sn=[],ut=[];function Tn(s){9<(s>>>=0)&&--ut[s+1]==0&&(ut[s]=void 0,Sn.push(s))}var Re=s=>{if(!s)throw new mt("Cannot use deleted val. handle = "+s);return ut[s]},We=s=>{switch(s){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let c=Sn.pop()||ut.length;return ut[c]=s,ut[c+1]=1,c}};function In(s){return this.fromWireType(a()[s>>>2>>>0])}var Pc={name:"emscripten::val",fromWireType:s=>{var c=Re(s);return Tn(s),c},toWireType:(s,c)=>We(c),argPackAdvance:ft,readValueFromPointer:In,Eb:null};function zc(s){return st(s>>>0,Pc)}var Oc=(s,c)=>{switch(c){case 4:return function(f){return this.fromWireType(d()[f>>>2>>>0])};case 8:return function(f){return this.fromWireType(l()[f>>>3>>>0])};default:throw new TypeError(`invalid float width (${c}): ${s}`)}};function Dc(s,c,f){f>>>=0,st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:b=>b,toWireType:(b,w)=>w,argPackAdvance:ft,readValueFromPointer:Oc(c,f),Eb:null})}function Bc(s,c,f,b,w){if(s>>>=0,f>>>=0,c=Qe(c>>>0),w===-1&&(w=4294967295),w=B=>B,b===0){var I=32-8*f;w=B=>B<<I>>>I}var O=c.includes("unsigned")?function(B,G){return G>>>0}:function(B,G){return G};st(s,{name:c,fromWireType:w,toWireType:O,argPackAdvance:ft,readValueFromPointer:ci(c,f,b!==0),Eb:null})}function Mc(s,c,f){function b(I){var O=a()[I>>>2>>>0];return I=a()[I+4>>>2>>>0],new w(t().buffer,I,O)}var w=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][c];st(s>>>=0,{name:f=Qe(f>>>0),fromWireType:b,argPackAdvance:ft,readValueFromPointer:b},{Tb:!0})}function Rc(s,c){s>>>=0;var f=(c=Qe(c>>>0))==="std::string";st(s,{name:c,fromWireType:function(b){var w=a()[b>>>2>>>0],I=b+4;if(f)for(var O=I,B=0;B<=w;++B){var G=I+B;if(B==w||r()[G>>>0]==0){if(O=Ce(O,G-O),F===void 0)var F=O;else F+=String.fromCharCode(0),F+=O;O=G+1}}else{for(F=Array(w),B=0;B<w;++B)F[B]=String.fromCharCode(r()[I+B>>>0]);F=F.join("")}return Je(b),F},toWireType:function(b,w){w instanceof ArrayBuffer&&(w=new Uint8Array(w));var I=typeof w=="string";if(!(I||w instanceof Uint8Array||w instanceof Uint8ClampedArray||w instanceof Int8Array))throw new mt("Cannot pass non-string to std::string");var O=f&&I?vn(w):w.length,B=mr(4+O+1),G=B+4;if(a()[B>>>2>>>0]=O,f&&I)Dt(w,G,O+1);else if(I)for(I=0;I<O;++I){var F=w.charCodeAt(I);if(255<F)throw Je(G),new mt("String has UTF-16 code units that do not fit in 8 bits");r()[G+I>>>0]=F}else for(I=0;I<O;++I)r()[G+I>>>0]=w[I];return b!==null&&b.push(Je,B),B},argPackAdvance:ft,readValueFromPointer:In,Eb(b){Je(b)}})}var pi=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,Uc=(s,c)=>{for(var f=s>>1,b=f+c/2;!(f>=b)&&o()[f>>>0];)++f;if(32<(f<<=1)-s&&pi)return pi.decode(r().slice(s,f));for(f="",b=0;!(b>=c/2);++b){var w=n()[s+2*b>>>1>>>0];if(w==0)break;f+=String.fromCharCode(w)}return f},Nc=(s,c,f)=>{if(f??=2147483647,2>f)return 0;var b=c;f=(f-=2)<2*s.length?f/2:s.length;for(var w=0;w<f;++w){var I=s.charCodeAt(w);n()[c>>>1>>>0]=I,c+=2}return n()[c>>>1>>>0]=0,c-b},Vc=s=>2*s.length,Wc=(s,c)=>{for(var f=0,b="";!(f>=c/4);){var w=i()[s+4*f>>>2>>>0];if(w==0)break;++f,65536<=w?(w-=65536,b+=String.fromCharCode(55296|w>>10,56320|1023&w)):b+=String.fromCharCode(w)}return b},Lc=(s,c,f)=>{if(c>>>=0,f??=2147483647,4>f)return 0;var b=c;f=b+f-4;for(var w=0;w<s.length;++w){var I=s.charCodeAt(w);if(55296<=I&&57343>=I&&(I=65536+((1023&I)<<10)|1023&s.charCodeAt(++w)),i()[c>>>2>>>0]=I,(c+=4)+4>f)break}return i()[c>>>2>>>0]=0,c-b},Gc=s=>{for(var c=0,f=0;f<s.length;++f){var b=s.charCodeAt(f);55296<=b&&57343>=b&&++f,c+=4}return c};function Hc(s,c,f){if(s>>>=0,c>>>=0,f=Qe(f>>>=0),c===2)var b=Uc,w=Nc,I=Vc,O=B=>o()[B>>>1>>>0];else c===4&&(b=Wc,w=Lc,I=Gc,O=B=>a()[B>>>2>>>0]);st(s,{name:f,fromWireType:B=>{for(var G,F=a()[B>>>2>>>0],X=B+4,le=0;le<=F;++le){var ge=B+4+le*c;le!=F&&O(ge)!=0||(X=b(X,ge-X),G===void 0?G=X:(G+=String.fromCharCode(0),G+=X),X=ge+c)}return Je(B),G},toWireType:(B,G)=>{if(typeof G!="string")throw new mt(`Cannot pass non-string to C++ string type ${f}`);var F=I(G),X=mr(4+F+c);return a()[X>>>2>>>0]=F/c,w(G,X+4,F+c),B!==null&&B.push(Je,X),X},argPackAdvance:ft,readValueFromPointer:In,Eb(B){Je(B)}})}function Fc(s,c){st(s>>>=0,{Ub:!0,name:c=Qe(c>>>0),argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})}var qc=()=>1;function Kc(s){On(s>>>0,!y,1,!_,131072,!1),Go()}var mi=s=>{if(!De)try{if(s(),!(0<vt))try{g?fr(j):wn(j)}catch(c){c instanceof bn||c=="unwind"||A(0,c)}}catch(c){c instanceof bn||c=="unwind"||A(0,c)}};function Cn(s){s>>>=0,typeof Atomics.oc=="function"&&(Atomics.oc(i(),s>>>2,s).value.then(or),s+=128,Atomics.store(i(),s>>>2,1))}var or=()=>{var s=Rt();s&&(Cn(s),mi(Gi))};function jc(s,c){(s>>>=0)==c>>>0?setTimeout(or):g?postMessage({targetThread:s,cmd:"checkMailbox"}):(s=Ze[s])&&s.postMessage({cmd:"checkMailbox"})}var An=[];function Yc(s,c,f,b,w){for(c>>>=0,b/=2,An.length=b,f=w>>>0>>>3,w=0;w<b;w++)An[w]=W[f+2*w]?W[f+2*w+1]:l()[f+2*w+1>>>0];return(c?gn[c]:Np[s])(...An)}function Zc(s){s>>>=0,g?postMessage({cmd:"cleanupThread",thread:s}):Lo(Ze[s])}function Qc(s){}var ir=(s,c)=>{var f=xn[s];if(f===void 0)throw s=Ni(s),f=Qe(s),Je(s),new mt(`${c} has unknown type ${f}`);return f},fi=(s,c,f)=>{var b=[];return s=s.toWireType(b,f),b.length&&(a()[c>>>2>>>0]=We(b)),s};function Xc(s,c,f){return c>>>=0,f>>>=0,s=Re(s>>>0),c=ir(c,"emval::as"),fi(c,f,s)}function Jc(s,c){return c>>>=0,s=Re(s>>>0),(c=ir(c,"emval::as")).toWireType(null,s)}var ar=s=>{try{s()}catch(c){ct(c)}},ht=0,Xe=null,hi=0,sr=[],gi={},bi={},ep=0,kn=null,tp=[];function yi(s){return function(c){if(!De){if(ht===0){var f=!1,b=!1;c((w=0)=>{if(!De&&(hi=w,f=!0,b)){ht=2,ar(()=>ji(Xe)),typeof Browser<"u"&&Browser.Lb.Sb&&Browser.Lb.resume(),w=!1;try{var I=function(){var G=i()[Xe+8>>>2>>>0];return G=Z[bi[G]],--vt,G()}()}catch(G){I=G,w=!0}var O=!1;if(!Xe){var B=kn;B&&(kn=null,(w?B.reject:B.resolve)(I),O=!0)}if(w&&!O)throw I}}),b=!0,f||(ht=1,Xe=function(){var w=mr(65548),I=w+12;a()[w>>>2>>>0]=I,a()[w+4>>>2>>>0]=I+65536,I=sr[0];var O=gi[I];return O===void 0&&(O=ep++,gi[I]=O,bi[O]=I),I=O,i()[w+8>>>2>>>0]=I,w}(),typeof Browser<"u"&&Browser.Lb.Sb&&Browser.Lb.pause(),ar(()=>qi(Xe)))}else ht===2?(ht=0,ar(Yi),Je(Xe),Xe=null,tp.forEach(mi)):ct(`invalid state: ${ht}`);return hi}}(c=>{s().then(c)})}function rp(s){return s>>>=0,yi(()=>(s=Re(s)).then(We))}var ur=[];function np(s,c,f,b){return f>>>=0,b>>>=0,(s=ur[s>>>0])(null,c=Re(c>>>0),f,b)}var op={},dr=s=>{var c=op[s];return c===void 0?Qe(s):c};function ip(s,c,f,b,w){return f>>>=0,b>>>=0,w>>>=0,(s=ur[s>>>0])(c=Re(c>>>0),c[f=dr(f)],b,w)}var _i=()=>typeof globalThis=="object"?globalThis:Function("return this")();function ap(s){return(s>>>=0)==0?We(_i()):(s=dr(s),We(_i()[s]))}var sp=s=>{var c=ur.length;return ur.push(s),c},up=(s,c)=>{for(var f=Array(s),b=0;b<s;++b)f[b]=ir(a()[c+4*b>>>2>>>0],"parameter "+b);return f},wi=(s,c)=>Object.defineProperty(c,"name",{value:s});function dp(s,c,f){var b=(c=up(s,c>>>0)).shift();s--;var w=`return function (obj, func, destructorsRef, args) {
+`,I=0,O=[];f===0&&O.push("obj");for(var B=["retType"],G=[b],F=0;F<s;++F)O.push("arg"+F),B.push("argType"+F),G.push(c[F]),w+=`  var arg${F} = argType${F}.readValueFromPointer(args${I?"+"+I:""});
+`,I+=c[F].argPackAdvance;return w+=`  var rv = ${f===1?"new func":"func.call"}(${O.join(", ")});
+`,b.Ub||(B.push("emval_returnValue"),G.push(fi),w+=`  return emval_returnValue(retType, destructorsRef, rv);
 `),B.push(w+`};
-`),s=function(X){var ce=Function;if(!(ce instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof ce} which is not a function`);var ge=vi(ce.name||"unknownFunctionName",function(){});return ge.prototype=ce.prototype,ge=new ge,(X=ce.apply(ge,X))instanceof Object?X:ge}(B)(...L),f=`methodCaller<(${c.map(X=>X.name).join(", ")}) => ${b.name}>`,sp(vi(f,s))}function lp(s){return s=lr(s>>>0),Ve(u[s])}function cp(s,c){return c>>>=0,s=Me(s>>>0),c=Me(c),Ve(s[c])}function pp(s){9<(s>>>=0)&&(ut[s+1]+=1)}function mp(){return Ve([])}function fp(s){s=Me(s>>>0);for(var c=Array(s.length),f=0;f<s.length;f++)c[f]=s[f];return Ve(c)}function hp(s){return Ve(lr(s>>>0))}function gp(){return Ve({})}function bp(s){for(var c=Me(s>>>=0);c.length;){var f=c.pop();c.pop()(f)}In(s)}function yp(s,c,f){c>>>=0,f>>>=0,s=Me(s>>>0),c=Me(c),f=Me(f),s[c]=f}function _p(s,c){return c>>>=0,s=(s=ar(s>>>0,"_emval_take_value")).readValueFromPointer(c),Ve(s)}function wp(s,c){s=-9007199254740992>s||9007199254740992<s?NaN:Number(s),c>>>=0,s=new Date(1e3*s),i()[c>>>2>>>0]=s.getUTCSeconds(),i()[c+4>>>2>>>0]=s.getUTCMinutes(),i()[c+8>>>2>>>0]=s.getUTCHours(),i()[c+12>>>2>>>0]=s.getUTCDate(),i()[c+16>>>2>>>0]=s.getUTCMonth(),i()[c+20>>>2>>>0]=s.getUTCFullYear()-1900,i()[c+24>>>2>>>0]=s.getUTCDay(),s=(s.getTime()-Date.UTC(s.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,i()[c+28>>>2>>>0]=s}var Dt=s=>s%4==0&&(s%100!=0||s%400==0),$i=[0,31,60,91,121,152,182,213,244,274,305,335],xi=[0,31,59,90,120,151,181,212,243,273,304,334];function vp(s,c){s=-9007199254740992>s||9007199254740992<s?NaN:Number(s),c>>>=0,s=new Date(1e3*s),i()[c>>>2>>>0]=s.getSeconds(),i()[c+4>>>2>>>0]=s.getMinutes(),i()[c+8>>>2>>>0]=s.getHours(),i()[c+12>>>2>>>0]=s.getDate(),i()[c+16>>>2>>>0]=s.getMonth(),i()[c+20>>>2>>>0]=s.getFullYear()-1900,i()[c+24>>>2>>>0]=s.getDay();var f=(Dt(s.getFullYear())?$i:xi)[s.getMonth()]+s.getDate()-1|0;i()[c+28>>>2>>>0]=f,i()[c+36>>>2>>>0]=-60*s.getTimezoneOffset(),f=new Date(s.getFullYear(),6,1).getTimezoneOffset();var b=new Date(s.getFullYear(),0,1).getTimezoneOffset();s=0|(f!=b&&s.getTimezoneOffset()==Math.min(b,f)),i()[c+32>>>2>>>0]=s}function $p(s){s>>>=0;var c=new Date(i()[s+20>>>2>>>0]+1900,i()[s+16>>>2>>>0],i()[s+12>>>2>>>0],i()[s+8>>>2>>>0],i()[s+4>>>2>>>0],i()[s>>>2>>>0],0),f=i()[s+32>>>2>>>0],b=c.getTimezoneOffset(),w=new Date(c.getFullYear(),6,1).getTimezoneOffset(),I=new Date(c.getFullYear(),0,1).getTimezoneOffset(),O=Math.min(I,w);return 0>f?i()[s+32>>>2>>>0]=+(w!=I&&O==b):0<f!=(O==b)&&(w=Math.max(I,w),c.setTime(c.getTime()+6e4*((0<f?O:w)-b))),i()[s+24>>>2>>>0]=c.getDay(),f=(Dt(c.getFullYear())?$i:xi)[c.getMonth()]+c.getDate()-1|0,i()[s+28>>>2>>>0]=f,i()[s>>>2>>>0]=c.getSeconds(),i()[s+4>>>2>>>0]=c.getMinutes(),i()[s+8>>>2>>>0]=c.getHours(),i()[s+12>>>2>>>0]=c.getDate(),i()[s+16>>>2>>>0]=c.getMonth(),i()[s+20>>>2>>>0]=c.getYear(),s=c.getTime(),BigInt(isNaN(s)?-1:s/1e3)}function Si(s,c,f,b,w,I,O){return g?xe(16,1,s,c,f,b,w,I,O):-52}function Ti(s,c,f,b,w,I){if(g)return xe(17,1,s,c,f,b,w,I)}function xp(s,c,f,b){s>>>=0,c>>>=0,f>>>=0,b>>>=0;var w=new Date().getFullYear(),I=new Date(w,0,1),O=new Date(w,6,1);w=I.getTimezoneOffset();var B=O.getTimezoneOffset(),L=Math.max(w,B);a()[s>>>2>>>0]=60*L,i()[c>>>2>>>0]=+(w!=B),I=(s=H=>H.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1])(I),O=s(O),B<w?(Ot(I,f,17),Ot(O,b,17)):(Ot(I,b,17),Ot(O,f,17))}var Pn=[],Ii=(s,c)=>{Pn.length=0;for(var f;f=r()[s++>>>0];){var b=f!=105;c+=(b&=f!=112)&&c%8?4:0,Pn.push(f==112?a()[c>>>2>>>0]:f==106?q[c>>>3]:f==105?i()[c>>>2>>>0]:l()[c>>>3>>>0]),c+=b?8:4}return Pn};function Sp(s,c,f){return s>>>=0,c=Ii(c>>>0,f>>>0),bn[s](...c)}function Tp(s,c,f){return s>>>=0,c=Ii(c>>>0,f>>>0),bn[s](...c)}var Ip=()=>{},Cp=()=>Date.now();function Ap(s,c){return j(Ce(s>>>0,c>>>0))}var Ci,kp=()=>{throw wt+=1,"unwind"};function Ep(){return 4294901760}Ci=()=>performance.timeOrigin+performance.now();var Pp=()=>navigator.hardwareConcurrency;function zp(){return ct("Cannot use emscripten_pc_get_function without -sUSE_OFFSET_CONVERTER"),0}function Op(s){s>>>=0;var c=r().length;if(s<=c||4294901760<s)return!1;for(var f=1;4>=f;f*=2){var b=c*(1+.2/f);b=Math.min(b,s+100663296);var w=Math;b=Math.max(s,b);e:{w=(w.min.call(w,4294901760,b+(65536-b%65536)%65536)-se.buffer.byteLength+65535)/65536;try{se.grow(w),ye();var I=1;break e}catch{}I=void 0}if(I)return!0}return!1}var cr=()=>(ct("Cannot use convertFrameToPC (needed by __builtin_return_address) without -sUSE_OFFSET_CONVERTER"),0),Bt={},Ai=s=>{s.forEach(c=>{var f=cr();f&&(Bt[f]=c)})};function Dp(){var s=Error().stack.toString().split(`
-`);return s[0]=="Error"&&s.shift(),Ai(s),Bt.Qb=cr(),Bt.fc=s,Bt.Qb}function Bp(s,c,f){if(s>>>=0,c>>>=0,Bt.Qb==s)var b=Bt.fc;else(b=Error().stack.toString().split(`
-`))[0]=="Error"&&b.shift(),Ai(b);for(var w=3;b[w]&&cr()!=s;)++w;for(s=0;s<f&&b[s+w];++s)i()[c+4*s>>>2>>>0]=cr();return s}var zn,On={},ki=()=>{if(!zn){var s,c={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:A||"./this.program"};for(s in On)On[s]===void 0?delete c[s]:c[s]=On[s];var f=[];for(s in c)f.push(`${s}=${c[s]}`);zn=f}return zn};function Ei(s,c){if(g)return xe(18,1,s,c);s>>>=0,c>>>=0;var f=0;return ki().forEach((b,w)=>{var I=c+f;for(w=a()[s+4*w>>>2>>>0]=I,I=0;I<b.length;++I)t()[w++>>>0]=b.charCodeAt(I);t()[w>>>0]=0,f+=b.length+1}),0}function Pi(s,c){if(g)return xe(19,1,s,c);s>>>=0,c>>>=0;var f=ki();a()[s>>>2>>>0]=f.length;var b=0;return f.forEach(w=>b+=w.length+1),a()[c>>>2>>>0]=b,0}function zi(s){return g?xe(20,1,s):52}function Oi(s,c,f,b){return g?xe(21,1,s,c,f,b):52}function Di(s,c,f,b){return g?xe(22,1,s,c,f,b):70}var Mp=[null,[],[]];function Bi(s,c,f,b){if(g)return xe(23,1,s,c,f,b);c>>>=0,f>>>=0,b>>>=0;for(var w=0,I=0;I<f;I++){var O=a()[c>>>2>>>0],B=a()[c+4>>>2>>>0];c+=8;for(var L=0;L<B;L++){var H=r()[O+L>>>0],X=Mp[s];H===0||H===10?((s===1?K:j)(Zo(X,0)),X.length=0):X.push(H)}w+=B}return a()[b>>>2>>>0]=w,0}var Mi=[31,29,31,30,31,30,31,31,30,31,30,31],Ri=[31,28,31,30,31,30,31,31,30,31,30,31],Rp=(s,c)=>{t().set(s,c>>>0)};function Ui(s,c,f,b){function w(z,de,Te){for(z=typeof z=="number"?z.toString():z||"";z.length<de;)z=Te[0]+z;return z}function I(z,de){return w(z,de,"0")}function O(z,de){function Te(Xi){return 0>Xi?-1:0<Xi?1:0}var $t;return($t=Te(z.getFullYear()-de.getFullYear()))===0&&($t=Te(z.getMonth()-de.getMonth()))===0&&($t=Te(z.getDate()-de.getDate())),$t}function B(z){switch(z.getDay()){case 0:return new Date(z.getFullYear()-1,11,29);case 1:return z;case 2:return new Date(z.getFullYear(),0,3);case 3:return new Date(z.getFullYear(),0,2);case 4:return new Date(z.getFullYear(),0,1);case 5:return new Date(z.getFullYear()-1,11,31);case 6:return new Date(z.getFullYear()-1,11,30)}}function L(z){var de=z.Cb;for(z=new Date(new Date(z.Db+1900,0,1).getTime());0<de;){var Te=z.getMonth(),$t=(Dt(z.getFullYear())?Mi:Ri)[Te];if(!(de>$t-z.getDate())){z.setDate(z.getDate()+de);break}de-=$t-z.getDate()+1,z.setDate(1),11>Te?z.setMonth(Te+1):(z.setMonth(0),z.setFullYear(z.getFullYear()+1))}return Te=new Date(z.getFullYear()+1,0,4),de=B(new Date(z.getFullYear(),0,4)),Te=B(Te),0>=O(de,z)?0>=O(Te,z)?z.getFullYear()+1:z.getFullYear():z.getFullYear()-1}s>>>=0,c>>>=0,f>>>=0,b>>>=0;var H=a()[b+40>>>2>>>0];for(var X in b={lc:i()[b>>>2>>>0],kc:i()[b+4>>>2>>>0],Ib:i()[b+8>>>2>>>0],Mb:i()[b+12>>>2>>>0],Jb:i()[b+16>>>2>>>0],Db:i()[b+20>>>2>>>0],vb:i()[b+24>>>2>>>0],Cb:i()[b+28>>>2>>>0],sc:i()[b+32>>>2>>>0],jc:i()[b+36>>>2>>>0],mc:H?Ce(H):""},f=Ce(f),H={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"})f=f.replace(new RegExp(X,"g"),H[X]);var ce="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),ge="January February March April May June July August September October November December".split(" ");for(X in H={"%a":z=>ce[z.vb].substring(0,3),"%A":z=>ce[z.vb],"%b":z=>ge[z.Jb].substring(0,3),"%B":z=>ge[z.Jb],"%C":z=>I((z.Db+1900)/100|0,2),"%d":z=>I(z.Mb,2),"%e":z=>w(z.Mb,2," "),"%g":z=>L(z).toString().substring(2),"%G":L,"%H":z=>I(z.Ib,2),"%I":z=>((z=z.Ib)==0?z=12:12<z&&(z-=12),I(z,2)),"%j":z=>{for(var de=0,Te=0;Te<=z.Jb-1;de+=(Dt(z.Db+1900)?Mi:Ri)[Te++]);return I(z.Mb+de,3)},"%m":z=>I(z.Jb+1,2),"%M":z=>I(z.kc,2),"%n":()=>`
-`,"%p":z=>0<=z.Ib&&12>z.Ib?"AM":"PM","%S":z=>I(z.lc,2),"%t":()=>"	","%u":z=>z.vb||7,"%U":z=>I(Math.floor((z.Cb+7-z.vb)/7),2),"%V":z=>{var de=Math.floor((z.Cb+7-(z.vb+6)%7)/7);if(2>=(z.vb+371-z.Cb-2)%7&&de++,de)de==53&&((Te=(z.vb+371-z.Cb)%7)==4||Te==3&&Dt(z.Db)||(de=1));else{de=52;var Te=(z.vb+7-z.Cb-1)%7;(Te==4||Te==5&&Dt(z.Db%400-1))&&de++}return I(de,2)},"%w":z=>z.vb,"%W":z=>I(Math.floor((z.Cb+7-(z.vb+6)%7)/7),2),"%y":z=>(z.Db+1900).toString().substring(2),"%Y":z=>z.Db+1900,"%z":z=>{var de=0<=(z=z.jc);return z=Math.abs(z)/60,(de?"+":"-")+("0000"+(z/60*100+z%60)).slice(-4)},"%Z":z=>z.mc,"%%":()=>"%"},f=f.replace(/%%/g,"\0\0"),H)f.includes(X)&&(f=f.replace(new RegExp(X,"g"),H[X](b)));return X=function(z){var de=Array($n(z)+1);return Jo(z,de,0,de.length),de}(f=f.replace(/\0\0/g,"%")),X.length>c?0:(Rp(X,s),X.length-1)}function Up(s,c,f,b){return Ui(s>>>0,c>>>0,f>>>0,b>>>0)}g||function(){for(var s=u.numThreads-1;s--;)qo();Ye.unshift(()=>{Gt++,function(c){g?c():Promise.all(pt.map(Fo)).then(c)}(()=>Do())})}();for(var Ni=Array(256),pr=0;256>pr;++pr)Ni[pr]=String.fromCharCode(pr);ci=Ni,mt=u.BindingError=class extends Error{constructor(s){super(s),this.name="BindingError"}},u.InternalError=class extends Error{constructor(s){super(s),this.name="InternalError"}},ut.push(0,1,void 0,1,null,1,!0,1,!1,1),u.count_emval_handles=()=>ut.length/2-5-Tn.length;var Np=[wn,Wo,Ko,Qo,Xo,ei,ti,ri,ni,oi,ii,ai,si,ui,di,li,Si,Ti,Ei,Pi,zi,Oi,Di,Bi],Z=function(){function s(f,b){return Z=f.exports,Z=function(){var w=Z,I={};for(let[O,B]of Object.entries(w))I[O]=typeof B=="function"?(...L)=>{ur.push(O);try{return B(...L)}finally{we||(ur.pop(),Xe&&ht===1&&ur.length===0&&(ht=0,wt+=1,sr(ji),typeof Fibers<"u"&&Fibers.tc()))}}:B;return I}(),Z=function(){var w=Z,I=B=>L=>B(L)>>>0,O=B=>()=>B()>>>0;return(w=Object.assign({},w)).Da=I(w.Da),w.gb=O(w.gb),w.ib=I(w.ib),w.emscripten_main_runtime_thread_id=O(w.emscripten_main_runtime_thread_id),w.tb=I(w.tb),w.ub=O(w.ub),w}(),Lo.push(Z.jb),Lt.unshift(Z.Ca),Y=b,Do(),Z}var c=No();if(Gt++,u.instantiateWasm)try{return u.instantiateWasm(c,s)}catch(f){j(`Module.instantiateWasm callback failed with error: ${f}`),m(f)}return gn||=u.locateFile?Bo("ort-wasm-simd-threaded.jsep.wasm")?"ort-wasm-simd-threaded.jsep.wasm":u.locateFile?u.locateFile("ort-wasm-simd-threaded.jsep.wasm",P):P+"ort-wasm-simd-threaded.jsep.wasm":new URL(/* asset import */ __webpack_require__(/*! ort-wasm-simd-threaded.jsep.wasm */ "./node_modules/onnxruntime-web/dist/ort-wasm-simd-threaded.jsep.wasm"), __webpack_require__.b).href,function(f,b){var w=gn;return D||typeof WebAssembly.instantiateStreaming!="function"||Bo(w)||Mo(w)||typeof fetch!="function"?Uo(w,f,b):fetch(w,{credentials:"same-origin"}).then(I=>WebAssembly.instantiateStreaming(I,f).then(b,function(O){return j(`wasm streaming compile failed: ${O}`),j("falling back to ArrayBuffer instantiation"),Uo(w,f,b)}))}(c,function(f){s(f.instance,f.module)}).catch(m),{}}(),Vi=s=>(Vi=Z.Da)(s),Wi=()=>(Wi=Z.Ea)();u._OrtInit=(s,c)=>(u._OrtInit=Z.Fa)(s,c),u._OrtGetLastError=(s,c)=>(u._OrtGetLastError=Z.Ga)(s,c),u._OrtCreateSessionOptions=(s,c,f,b,w,I,O,B,L,H)=>(u._OrtCreateSessionOptions=Z.Ha)(s,c,f,b,w,I,O,B,L,H),u._OrtAppendExecutionProvider=(s,c)=>(u._OrtAppendExecutionProvider=Z.Ia)(s,c),u._OrtAddFreeDimensionOverride=(s,c,f)=>(u._OrtAddFreeDimensionOverride=Z.Ja)(s,c,f),u._OrtAddSessionConfigEntry=(s,c,f)=>(u._OrtAddSessionConfigEntry=Z.Ka)(s,c,f),u._OrtReleaseSessionOptions=s=>(u._OrtReleaseSessionOptions=Z.La)(s),u._OrtCreateSession=(s,c,f)=>(u._OrtCreateSession=Z.Ma)(s,c,f),u._OrtReleaseSession=s=>(u._OrtReleaseSession=Z.Na)(s),u._OrtGetInputOutputCount=(s,c,f)=>(u._OrtGetInputOutputCount=Z.Oa)(s,c,f),u._OrtGetInputName=(s,c)=>(u._OrtGetInputName=Z.Pa)(s,c),u._OrtGetOutputName=(s,c)=>(u._OrtGetOutputName=Z.Qa)(s,c),u._OrtFree=s=>(u._OrtFree=Z.Ra)(s),u._OrtCreateTensor=(s,c,f,b,w,I)=>(u._OrtCreateTensor=Z.Sa)(s,c,f,b,w,I),u._OrtGetTensorData=(s,c,f,b,w)=>(u._OrtGetTensorData=Z.Ta)(s,c,f,b,w),u._OrtReleaseTensor=s=>(u._OrtReleaseTensor=Z.Ua)(s),u._OrtCreateRunOptions=(s,c,f,b)=>(u._OrtCreateRunOptions=Z.Va)(s,c,f,b),u._OrtAddRunConfigEntry=(s,c,f)=>(u._OrtAddRunConfigEntry=Z.Wa)(s,c,f),u._OrtReleaseRunOptions=s=>(u._OrtReleaseRunOptions=Z.Xa)(s),u._OrtCreateBinding=s=>(u._OrtCreateBinding=Z.Ya)(s),u._OrtBindInput=(s,c,f)=>(u._OrtBindInput=Z.Za)(s,c,f),u._OrtBindOutput=(s,c,f,b)=>(u._OrtBindOutput=Z._a)(s,c,f,b),u._OrtClearBoundOutputs=s=>(u._OrtClearBoundOutputs=Z.$a)(s),u._OrtReleaseBinding=s=>(u._OrtReleaseBinding=Z.ab)(s),u._OrtRunWithBinding=(s,c,f,b,w)=>(u._OrtRunWithBinding=Z.bb)(s,c,f,b,w),u._OrtRun=(s,c,f,b,w,I,O,B)=>(u._OrtRun=Z.cb)(s,c,f,b,w,I,O,B),u._OrtEndProfiling=s=>(u._OrtEndProfiling=Z.db)(s),u._JsepOutput=(s,c,f)=>(u._JsepOutput=Z.eb)(s,c,f),u._JsepGetNodeName=s=>(u._JsepGetNodeName=Z.fb)(s);var mr,Mt=()=>(Mt=Z.gb)(),Je=u._free=s=>(Je=u._free=Z.hb)(s),fr=u._malloc=s=>(fr=u._malloc=Z.ib)(s),Dn=(s,c,f,b,w,I)=>(Dn=Z.lb)(s,c,f,b,w,I),Li=()=>(Li=Z.mb)(),Gi=(s,c,f,b,w)=>(Gi=Z.nb)(s,c,f,b,w),Bn=s=>(Bn=Z.ob)(s),hr=s=>(hr=Z.pb)(s),Hi=()=>(Hi=Z.qb)(),Fi=(s,c)=>(Fi=Z.rb)(s,c),gr=s=>(gr=Z.sb)(s),Mn=s=>(Mn=Z.tb)(s),Rn=()=>(Rn=Z.ub)(),qi=u.dynCall_ii=(s,c)=>(qi=u.dynCall_ii=Z.wb)(s,c),Ki=s=>(Ki=Z.xb)(s),ji=()=>(ji=Z.yb)(),Yi=s=>(Yi=Z.zb)(s),Zi=()=>(Zi=Z.Ab)();function Qi(){0<Gt||(g?(p(u),g||or(Lt),startWorker(u)):(or(Ye),0<Gt||mr||(mr=!0,u.calledRun=!0,we||(g||or(Lt),p(u),g||or(fn)))))}return u.___start_em_js=929301,u.___stop_em_js=929547,u.stackSave=()=>Rn(),u.stackRestore=s=>gr(s),u.stackAlloc=s=>Mn(s),u.setValue=function(s,c,f="i8"){switch(f.endsWith("*")&&(f="*"),f){case"i1":case"i8":t()[s>>>0]=c;break;case"i16":n()[s>>>1>>>0]=c;break;case"i32":i()[s>>>2>>>0]=c;break;case"i64":q[s>>>3]=BigInt(c);break;case"float":d()[s>>>2>>>0]=c;break;case"double":l()[s>>>3>>>0]=c;break;case"*":a()[s>>>2>>>0]=c;break;default:ct(`invalid type for setValue: ${f}`)}},u.getValue=function(s,c="i8"){switch(c.endsWith("*")&&(c="*"),c){case"i1":case"i8":return t()[s>>>0];case"i16":return n()[s>>>1>>>0];case"i32":return i()[s>>>2>>>0];case"i64":return q[s>>>3];case"float":return d()[s>>>2>>>0];case"double":return l()[s>>>3>>>0];case"*":return a()[s>>>2>>>0];default:ct(`invalid type for getValue: ${c}`)}},u.UTF8ToString=Ce,u.stringToUTF8=Ot,u.lengthBytesUTF8=$n,Ht=function s(){mr||Qi(),mr||(Ht=s)},Qi(),u.PTR_SIZE=4,h}),Kp=za;globalThis.self?.name==="em-pthread"&&za()});var Ra,jp,Ne,Ua,jn,Yp,Zp,Na,Qp,Ba,Va,Ma,Wa,xr=U(()=>{"use strict";$r();Ra= false||typeof location>"u"?void 0:location.origin,jp=()=>{if(true)return import.meta.url?.startsWith("file:")?new URL(new URL(/* asset import */ __webpack_require__(/*! ort.bundle.min.mjs */ "./node_modules/onnxruntime-web/dist/ort.bundle.min.mjs?46eb"), __webpack_require__.b).href,Ra).href:import.meta.url},Ne=jp(),Ua=()=>{if(Ne&&!Ne.startsWith("blob:"))return Ne.substring(0,Ne.lastIndexOf("/")+1)},jn=(e,t)=>{try{let r=t??Ne;return(r?new URL(e,r):new URL(e)).origin===Ra}catch{return!1}},Yp=(e,t)=>{let r=t??Ne;try{return(r?new URL(e,r):new URL(e)).href}catch{return}},Zp=(e,t)=>`${t??"./"}${e}`,Na=async e=>{let r=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(r)},Qp=async e=>(await import(/*webpackIgnore:true*/e)).default,Ba=(Pa(),br(Ea)).default,Va=async()=>{if(!Ne)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(jn(Ne))return[void 0,Ba()];let e=await Na(Ne);return[e,Ba(e)]},Ma=(Da(),br(Oa)).default,Wa=async(e,t,r)=>{if(!e&&!t&&Ma&&Ne&&jn(Ne))return[void 0,Ma];{let n="ort-wasm-simd-threaded.jsep.mjs",o=e??Yp(n,t),i= true&&r&&o&&!jn(o,t),a=i?await Na(o):o??Zp(n,t);return[i?a:void 0,await Qp(a)]}}});var Yn,Zn,zr,La,Xp,Jp,Sr,Ie,gt=U(()=>{"use strict";xr();Zn=!1,zr=!1,La=!1,Xp=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},Jp=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},Sr=async e=>{if(Zn)return Promise.resolve();if(zr)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(La)throw new Error("previous call to 'initializeWebAssembly()' failed.");zr=!0;let t=e.initTimeout,r=e.numThreads;if(!Jp())throw new Error("WebAssembly SIMD is not supported in the current environment.");let n=Xp();r>1&&!n&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+r+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=r=1);let o=e.wasmPaths,i=typeof o=="string"?o:void 0,a=o?.mjs,d=a?.href??a,l=o?.wasm,p=l?.href??l,m=e.wasmBinary,[u,h]=await Wa(d,i,r>1),_=!1,y=[];if(t>0&&y.push(new Promise(g=>{setTimeout(()=>{_=!0,g()},t)})),y.push(new Promise((g,x)=>{let $={numThreads:r};if(m)$.wasmBinary=m;else if(p||i)$.locateFile=v=>p??i+v;else if(d&&d.indexOf("blob:")!==0)$.locateFile=v=>new URL(v,d).href;else if(u){let v=Ua();v&&($.locateFile=S=>v+S)}h($).then(v=>{zr=!1,Zn=!0,Yn=v,g(),u&&URL.revokeObjectURL(u)},v=>{zr=!1,La=!0,x(v)})})),await Promise.race(y),_)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Ie=()=>{if(Zn&&Yn)return Yn;throw new Error("WebAssembly is not initialized yet.")}});var ke,jt,pe,Or=U(()=>{"use strict";gt();ke=(e,t)=>{let r=Ie(),n=r.lengthBytesUTF8(e)+1,o=r._malloc(n);return r.stringToUTF8(e,o,n),t.push(o),o},jt=(e,t,r,n)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([o,i])=>{let a=t?t+o:o;if(typeof i=="object")jt(i,a+".",r,n);else if(typeof i=="string"||typeof i=="number")n(a,i.toString());else if(typeof i=="boolean")n(a,i?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof i}`)})},pe=e=>{let t=Ie(),r=t.stackSave();try{let n=t.PTR_SIZE,o=t.stackAlloc(2*n);t._OrtGetLastError(o,o+n);let i=Number(t.getValue(o,n===4?"i32":"i64")),a=t.getValue(o+n,"*"),d=a?t.UTF8ToString(a):"";throw new Error(`${e} ERROR_CODE: ${i}, ERROR_MESSAGE: ${d}`)}finally{t.stackRestore(r)}}});var Ga,Ha=U(()=>{"use strict";gt();Or();Ga=e=>{let t=Ie(),r=0,n=[],o=e||{};try{if(e?.logSeverityLevel===void 0)o.logSeverityLevel=2;else if(typeof e.logSeverityLevel!="number"||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${e.logSeverityLevel}`);if(e?.logVerbosityLevel===void 0)o.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);e?.terminate===void 0&&(o.terminate=!1);let i=0;return e?.tag!==void 0&&(i=ke(e.tag,n)),r=t._OrtCreateRunOptions(o.logSeverityLevel,o.logVerbosityLevel,!!o.terminate,i),r===0&&pe("Can't create run options."),e?.extra!==void 0&&jt(e.extra,"",new WeakSet,(a,d)=>{let l=ke(a,n),p=ke(d,n);t._OrtAddRunConfigEntry(r,l,p)!==0&&pe(`Can't set a run config entry: ${a} - ${d}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseRunOptions(r),n.forEach(a=>t._free(a)),i}}});var em,tm,rm,nm,Fa,qa=U(()=>{"use strict";gt();Or();em=e=>{switch(e){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${e}`)}},tm=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},rm=e=>{e.extra||(e.extra={}),e.extra.session||(e.extra.session={});let t=e.extra.session;t.use_ort_model_bytes_directly||(t.use_ort_model_bytes_directly="1"),e.executionProviders&&e.executionProviders.some(r=>(typeof r=="string"?r:r.name)==="webgpu")&&(e.enableMemPattern=!1)},nm=(e,t,r)=>{for(let n of t){let o=typeof n=="string"?n:n.name;switch(o){case"webnn":if(o="WEBNN",typeof n!="string"){let d=n?.deviceType;if(d){let l=ke("deviceType",r),p=ke(d,r);Ie()._OrtAddSessionConfigEntry(e,l,p)!==0&&pe(`Can't set a session config entry: 'deviceType' - ${d}.`)}}break;case"webgpu":if(o="JS",typeof n!="string"){let a=n;if(a?.preferredLayout){if(a.preferredLayout!=="NCHW"&&a.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${a.preferredLayout}`);let d=ke("preferredLayout",r),l=ke(a.preferredLayout,r);Ie()._OrtAddSessionConfigEntry(e,d,l)!==0&&pe(`Can't set a session config entry: 'preferredLayout' - ${a.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${o}`)}let i=ke(o,r);Ie()._OrtAppendExecutionProvider(e,i)!==0&&pe(`Can't append execution provider: ${o}.`)}},Fa=e=>{let t=Ie(),r=0,n=[],o=e||{};rm(o);try{let i=em(o.graphOptimizationLevel??"all"),a=tm(o.executionMode??"sequential"),d=typeof o.logId=="string"?ke(o.logId,n):0,l=o.logSeverityLevel??2;if(!Number.isInteger(l)||l<0||l>4)throw new Error(`log serverity level is not valid: ${l}`);let p=o.logVerbosityLevel??0;if(!Number.isInteger(p)||p<0||p>4)throw new Error(`log verbosity level is not valid: ${p}`);let m=typeof o.optimizedModelFilePath=="string"?ke(o.optimizedModelFilePath,n):0;if(r=t._OrtCreateSessionOptions(i,!!o.enableCpuMemArena,!!o.enableMemPattern,a,!!o.enableProfiling,0,d,l,p,m),r===0&&pe("Can't create session options."),o.executionProviders&&nm(r,o.executionProviders,n),o.enableGraphCapture!==void 0){if(typeof o.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${o.enableGraphCapture}`);let u=ke("enableGraphCapture",n),h=ke(o.enableGraphCapture.toString(),n);t._OrtAddSessionConfigEntry(r,u,h)!==0&&pe(`Can't set a session config entry: 'enableGraphCapture' - ${o.enableGraphCapture}.`)}if(o.freeDimensionOverrides)for(let[u,h]of Object.entries(o.freeDimensionOverrides)){if(typeof u!="string")throw new Error(`free dimension override name must be a string: ${u}`);if(typeof h!="number"||!Number.isInteger(h)||h<0)throw new Error(`free dimension override value must be a non-negative integer: ${h}`);let _=ke(u,n);t._OrtAddFreeDimensionOverride(r,_,h)!==0&&pe(`Can't set a free dimension override: ${u} - ${h}.`)}return o.extra!==void 0&&jt(o.extra,"",new WeakSet,(u,h)=>{let _=ke(u,n),y=ke(h,n);t._OrtAddSessionConfigEntry(r,_,y)!==0&&pe(`Can't set a session config entry: ${u} - ${h}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseSessionOptions(r)!==0&&pe("Can't release session options."),n.forEach(a=>t._free(a)),i}}});var Yt,bt,Ct,Dr,Zt,Br,Mr,Qn,te=U(()=>{"use strict";Yt=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw new Error(`unsupported data type: ${e}`)}},bt=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";case 22:return"int4";case 21:return"uint4";default:throw new Error(`unsupported data type: ${e}`)}},Ct=(e,t)=>{let r=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],n=typeof t=="number"?t:t.reduce((o,i)=>o*i,1);return r>0?Math.ceil(n*r):void 0},Dr=e=>{switch(e){case"float16":return typeof Float16Array<"u"&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"bool":return Uint8Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw new Error(`unsupported type: ${e}`)}},Zt=e=>{switch(e){case"verbose":return 0;case"info":return 1;case"warning":return 2;case"error":return 3;case"fatal":return 4;default:throw new Error(`unsupported logging level: ${e}`)}},Br=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Mr=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Qn=e=>{switch(e){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;case"ml-tensor":return 5;default:throw new Error(`unsupported data location: ${e}`)}}});var Qt,Xn=U(()=>{"use strict";$r();Qt=async e=>{if(typeof e=="string")if(false){}else{let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),n=r?parseInt(r,10):0;if(n<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let o=t.body.getReader(),i;try{i=new ArrayBuffer(n)}catch(d){if(d instanceof RangeError){let l=Math.ceil(n/65536);i=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw d}let a=0;for(;;){let{done:d,value:l}=await o.read();if(d)break;let p=l.byteLength;new Uint8Array(i,a,p).set(l),a+=p}return new Uint8Array(i,0,n)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}});var om,im,Ka,ja,Rr,am,ue,et=U(()=>{"use strict";te();om=["V","I","W","E","F"],im=(e,t)=>{console.log(`[${om[e]},${new Date().toISOString()}]${t}`)},Rr=(e,t)=>{Ka=e,ja=t},am=(e,t)=>{let r=Zt(e),n=Zt(Ka);r>=n&&im(r,typeof t=="function"?t():t)},ue=(...e)=>{ja&&am(...e)}});var Ur,Jn=U(()=>{"use strict";te();Ur=(e,t)=>new(Dr(t))(e)});var Nr=U(()=>{"use strict"});var Ya,eo,to,sm,um,Za,no,ro,Xa,Ja=U(()=>{"use strict";et();Nr();Ya=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),eo=[],to=e=>Math.ceil(Number(e)/16)*16,sm=e=>{for(let t=0;t<eo.length;t++){let r=eo[t];if(e<=r)return r}return Math.ceil(e/16)*16},um=1,Za=()=>um++,no=async(e,t,r,n)=>{let o=to(r),i=e.device.createBuffer({size:o,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let a=e.getCommandEncoder();e.endComputePass(),a.copyBufferToBuffer(t,0,i,0,o),e.flush(),await i.mapAsync(GPUMapMode.READ);let d=i.getMappedRange();if(n){let l=n();return l.set(new Uint8Array(d,0,r)),l}else return new Uint8Array(d.slice(0,r))}finally{i.destroy()}},ro=class{constructor(t){this.backend=t;this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersPending=[],this.capturedPendingBuffers=new Map;for(let[r]of Ya)eo.push(r),this.freeBuffers.set(r,[]),this.freeUniformBuffers.set(r,[]);this.sessionCount=0}upload(t,r){let n=r.buffer,o=r.byteOffset,i=r.byteLength,a=to(i),d=this.storageCache.get(t);if(!d)throw new Error("gpu data for uploading does not exist");if(Number(d.originalSize)!==i)throw new Error(`inconsistent data size. gpu data size=${d.originalSize}, data size=${i}`);let l=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),p=l.getMappedRange();new Uint8Array(p).set(new Uint8Array(n,o,i)),l.unmap();let m=this.backend.device.createCommandEncoder();m.copyBufferToBuffer(l,0,d.gpuData.buffer,0,a),this.backend.device.queue.submit([m.finish()]),l.destroy(),ue("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`)}memcpy(t,r){let n=this.storageCache.get(t);if(!n)throw new Error("source gpu data for memcpy does not exist");let o=this.storageCache.get(r);if(!o)throw new Error("destination gpu data for memcpy does not exist");if(n.originalSize!==o.originalSize)throw new Error("inconsistent source and destination gpu data size");let i=to(n.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(n.gpuData.buffer,0,o.gpuData.buffer,0,i)}registerExternalBuffer(t,r,n){let o;if(n){if(o=n[0],t===n[1])return ue("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${o}, buffer is the same, skip.`),o;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
-             Please use the previous external buffer!`)}else o=Za();return this.storageCache.set(o,{gpuData:{id:o,type:0,buffer:t},originalSize:r}),ue("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${o}, registered.`),o}unregisterExternalBuffer(t){t!==void 0&&(this.storageCache.delete(t),ue("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${t}`))}create(t,r=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let n=sm(t),o,i=(r&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(r&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(i||a){let p=(i?this.freeBuffers:this.freeUniformBuffers).get(n);p?p.length>0?o=p.pop():o=this.backend.device.createBuffer({size:n,usage:r}):o=this.backend.device.createBuffer({size:n,usage:r})}else o=this.backend.device.createBuffer({size:n,usage:r});let d={id:Za(),type:0,buffer:o};return this.storageCache.set(d.id,{gpuData:d,originalSize:Number(t)}),ue("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${d.id}`),d}get(t){return this.storageCache.get(t)?.gpuData}release(t){let r=typeof t=="bigint"?Number(t):t,n=this.storageCache.get(r);if(!n){if(this.storageCache.size===0)return 0;throw new Error("releasing data does not exist")}return ue("verbose",()=>`[WebGPU] GpuDataManager.release(id=${r}), gpuDataId=${n.gpuData.id}`),this.storageCache.delete(r),this.buffersPending.push(n.gpuData.buffer),n.originalSize}async download(t,r){let n=this.storageCache.get(Number(t));if(!n)throw new Error("data does not exist");await no(this.backend,n.gpuData.buffer,n.originalSize,r)}refreshPendingBuffers(){if(this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let r=Ya.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let n=this.freeBuffers.get(t.size)||[];r===void 0||n.length>=r?t.destroy():n.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let n=this.freeUniformBuffers.get(t.size)||[];r===void 0||n.length>=r?t.destroy():n.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let r of this.buffersPending)t.push(r);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(t){let r=this.capturedPendingBuffers.get(t);r&&(r.forEach(n=>{n.destroy()}),this.capturedPendingBuffers.delete(t)),this.sessionCount-=1,this.sessionCount===0&&(ue("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(n=>{n.gpuData.buffer.destroy()}),this.storageCache=new Map)}},Xa=(...e)=>new ro(...e)});var oo,re,Se=U(()=>{"use strict";oo=class{constructor(t){Object.assign(this,t)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(t=>`${this[t]}`).join(";")),this.key}},re=e=>new oo(e)});var io,tt,C,At,Vr,es,ts,oe=U(()=>{"use strict";io=class{static calcMatMulShape(t,r){return t[1]!==r[0]?void 0:[t[0],r[1]]}},tt=class{static calcShape(t,r,n=!1){let o=t.length,i=r.length;if(o===0)return r;if(i===0)return t;let a=Math.max(t.length,r.length),d=new Array(a);if(n){if(o<2||i<2)return;let l=io.calcMatMulShape([t[o-2],t[o-1]],[r[i-2],r[i-1]]);if(l===void 0)return;[d[a-2],d[a-1]]=l}for(let l=n?3:1;l<=a;l++){let p=o-l<0?1:t[o-l],m=i-l<0?1:r[i-l];if(p!==m&&p>1&&m>1)return;let u=Math.max(p,m);if(p&&m)d[a-l]=Math.max(p,m);else{if(u>1)return;d[a-l]=0}}return d}static isValidBroadcast(t,r){let n=t.length,o=r.length;if(n>o)return!1;for(let i=1;i<=n;i++)if(t[n-i]!==1&&t[n-i]!==r[o-i])return!1;return!0}},C=class e{static size(t){return e.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let n=t.length;if(n===0)return[];let o=new Array(n),i=n-1;for(;i>=0;){if(t[i]%r===0){o[i]=t[i]/r;break}if(r%t[i]!==0)throw new Error("cannot convert shape");o[i]=1,r/=t[i],i--}for(i--;i>=0;i--)o[i]=t[i];return o}static sizeFromDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,r,t.length)}static sizeToDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,n){let o=1;for(let i=r;i<n;i++){if(t[i]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");o*=Number(t[i])}return o}static computeStrides(t){let r=t.length;if(r===0)return[];if(r===1)return[1];let n=new Array(r);n[r-1]=1,n[r-2]=t[r-1];for(let o=r-3;o>=0;--o)n[o]=n[o+1]*t[o+1];return n}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(n=>this.normalizeAxis(n,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(n=>t[n]):t.slice().reverse()}static padShape(t,r){let n=t.length;return t.map((o,i)=>o+r[i]+r[i+n])}static areEqual(t,r){return t.length!==r.length?!1:t.every((n,o)=>n===r[o])}},At=class e{static adjustPoolAttributes(t,r,n,o,i,a){if(!t&&n.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let d=0;d<r.length-2;d++)d>=n.length?n.push(r[d+2]):n[d]=r[d+2];for(let d=0;d<n.length;d++)if(d<o.length){if(o[d]<0)throw new Error("strides should be greater than or equal to 1")}else o.push(1);for(let d=0;d<n.length;d++)if(d<i.length){if(i[d]<0)throw new Error("dilations should be greater than or equal to 1")}else i.push(1);for(let d=0;d<n.length*2;d++)if(d<a.length){if(a[d]<0)throw new Error("pad should be greater than or equal to 1")}else a.push(0);for(let d=0;d<n.length;d++){if(n[d]<=0)throw new Error("kernel shapes need to be greater than 0");if(a[d]>=n[d]||a[d+n.length]>=n[d])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,n,o,i,a,d){if(d){if(i.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(o.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let l=0;l<t.length-2;l++)e.adjustPadAndReturnShape(t[l+(a?1:2)],r[l],n[l],o[l],i,l,l+t.length-2,d)}}static computePoolOutputShape(t,r,n,o,i,a,d){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let l=[r[0],r[1]];return e.computeShapeHelper(t,r,l,n,o,i,a,d),l}static computeConvOutputShape(t,r,n,o,i,a,d){if(t.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let l=[t[0],r[0]];return e.computeShapeHelper(!1,t,l,n,o,i,a,d),l}static computeShapeHelper(t,r,n,o,i,a,d,l){if(t)for(let p=0;p<r.length-2;p++)n.push(1);else for(let p=0;p<r.length-2;p++)n.push(e.adjustPadAndReturnShape(r[p+2],o[p],i[p],a[p],d,p,p+r.length-2,l))}static adjustPadAndReturnShape(t,r,n,o,i,a,d,l){let p=n*(o-1)+1;if(l&&l!=="NOTSET")switch(l){case"VALID":return i[a]=0,i[d]=0,Math.floor((t-p)/r+1);case"SAME_LOWER":case"SAME_UPPER":if(n!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let u=((t+r-1)/r-1)*r+o-t;return i[a]=Math.floor(l==="SAME_LOWER"?(u+1)/2:u/2),i[d]=u-i[a],Math.floor((t+u-o)/r+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((t+i[a]+i[d]-p)/r+1)}},Vr=class{static getShapeOfGemmResult(t,r,n,o,i){if(t.length!==2||n.length!==2)throw new Error("shape need to be of size 2");let a,d,l;r?(a=t[1],d=t[0]):(a=t[0],d=t[1]);let p=-1;if(o?(l=n[0],p=1):(l=n[1],p=0),n[p]!==d)throw new Error("dimension mismatch");if(a<=0||l<=0||d<=0)throw new Error("invalid shape specified");if(i&&!tt.isValidBroadcast(i,[a,l]))throw new Error("gemm: invalid bias shape for broadcast");return[a,l,d]}},es=-34028234663852886e22,ts=34028234663852886e22});var kt,so,_e,Ee,N,me,uo,Et,Fe,F,Wr,E,M,rs,Lr,ao,ns,ae=U(()=>{"use strict";te();oe();kt=64,so=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${t}<f16>`:"f16";case 1:return t>1?`vec${t}<f32>`:"f32";case 6:return t>1?`vec${t}<i32>`:"i32";case 12:return t>1?`vec${t}<u32>`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];case 22:return"i32";case 21:return"u32";default:throw new Error(`Unknown data type: ${e}`)}},_e=(e,t=1)=>{let r=so(e,t);return typeof r=="string"?r:r[0]},Ee=(e,t=1)=>{let r=so(e,t);return typeof r=="string"?r:r[1]},N=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:C.computeStrides(r)})}),t},me=e=>e%4===0?4:e%2===0?2:1,uo=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,Et=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,Fe=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,F=(e,t,r,n)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?n==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:n==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Wr=(e,t,r,n,o)=>{let i=typeof r=="number",a=i?r:r.length,d=[...new Array(a).keys()],l=a<2?"u32":a<=4?`vec${a}<u32>`:`array<u32, ${a}>`,p=so(t,o),m=typeof p=="string"?p:p[1],u=typeof p=="string"?p:p[0],h={indices:l,value:m,storage:u,tensor:t},_=W=>typeof W=="string"?W:`${W}u`,y={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},g=i?"uniforms.":"",x=`${g}${e}_shape`,$=`${g}${e}_strides`,v="";for(let W=0;W<a-1;W++)v+=`
-    let dim${W} = current / ${F($,W,a)};
-    let rest${W} = current % ${F($,W,a)};
+`),s=function(X){var le=Function;if(!(le instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof le} which is not a function`);var ge=wi(le.name||"unknownFunctionName",function(){});return ge.prototype=le.prototype,ge=new ge,(X=le.apply(ge,X))instanceof Object?X:ge}(B)(...G),f=`methodCaller<(${c.map(X=>X.name).join(", ")}) => ${b.name}>`,sp(wi(f,s))}function lp(s){return s=dr(s>>>0),We(u[s])}function cp(s,c){return c>>>=0,s=Re(s>>>0),c=Re(c),We(s[c])}function pp(s){9<(s>>>=0)&&(ut[s+1]+=1)}function mp(){return We([])}function fp(s){s=Re(s>>>0);for(var c=Array(s.length),f=0;f<s.length;f++)c[f]=s[f];return We(c)}function hp(s){return We(dr(s>>>0))}function gp(){return We({})}function bp(s){for(var c=Re(s>>>=0);c.length;){var f=c.pop();c.pop()(f)}Tn(s)}function yp(s,c,f){c>>>=0,f>>>=0,s=Re(s>>>0),c=Re(c),f=Re(f),s[c]=f}function _p(s,c){return c>>>=0,s=(s=ir(s>>>0,"_emval_take_value")).readValueFromPointer(c),We(s)}function wp(s,c){s=-9007199254740992>s||9007199254740992<s?NaN:Number(s),c>>>=0,s=new Date(1e3*s),i()[c>>>2>>>0]=s.getUTCSeconds(),i()[c+4>>>2>>>0]=s.getUTCMinutes(),i()[c+8>>>2>>>0]=s.getUTCHours(),i()[c+12>>>2>>>0]=s.getUTCDate(),i()[c+16>>>2>>>0]=s.getUTCMonth(),i()[c+20>>>2>>>0]=s.getUTCFullYear()-1900,i()[c+24>>>2>>>0]=s.getUTCDay(),s=(s.getTime()-Date.UTC(s.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,i()[c+28>>>2>>>0]=s}var Bt=s=>s%4==0&&(s%100!=0||s%400==0),vi=[0,31,60,91,121,152,182,213,244,274,305,335],$i=[0,31,59,90,120,151,181,212,243,273,304,334];function vp(s,c){s=-9007199254740992>s||9007199254740992<s?NaN:Number(s),c>>>=0,s=new Date(1e3*s),i()[c>>>2>>>0]=s.getSeconds(),i()[c+4>>>2>>>0]=s.getMinutes(),i()[c+8>>>2>>>0]=s.getHours(),i()[c+12>>>2>>>0]=s.getDate(),i()[c+16>>>2>>>0]=s.getMonth(),i()[c+20>>>2>>>0]=s.getFullYear()-1900,i()[c+24>>>2>>>0]=s.getDay();var f=(Bt(s.getFullYear())?vi:$i)[s.getMonth()]+s.getDate()-1|0;i()[c+28>>>2>>>0]=f,i()[c+36>>>2>>>0]=-60*s.getTimezoneOffset(),f=new Date(s.getFullYear(),6,1).getTimezoneOffset();var b=new Date(s.getFullYear(),0,1).getTimezoneOffset();s=0|(f!=b&&s.getTimezoneOffset()==Math.min(b,f)),i()[c+32>>>2>>>0]=s}function $p(s){s>>>=0;var c=new Date(i()[s+20>>>2>>>0]+1900,i()[s+16>>>2>>>0],i()[s+12>>>2>>>0],i()[s+8>>>2>>>0],i()[s+4>>>2>>>0],i()[s>>>2>>>0],0),f=i()[s+32>>>2>>>0],b=c.getTimezoneOffset(),w=new Date(c.getFullYear(),6,1).getTimezoneOffset(),I=new Date(c.getFullYear(),0,1).getTimezoneOffset(),O=Math.min(I,w);return 0>f?i()[s+32>>>2>>>0]=+(w!=I&&O==b):0<f!=(O==b)&&(w=Math.max(I,w),c.setTime(c.getTime()+6e4*((0<f?O:w)-b))),i()[s+24>>>2>>>0]=c.getDay(),f=(Bt(c.getFullYear())?vi:$i)[c.getMonth()]+c.getDate()-1|0,i()[s+28>>>2>>>0]=f,i()[s>>>2>>>0]=c.getSeconds(),i()[s+4>>>2>>>0]=c.getMinutes(),i()[s+8>>>2>>>0]=c.getHours(),i()[s+12>>>2>>>0]=c.getDate(),i()[s+16>>>2>>>0]=c.getMonth(),i()[s+20>>>2>>>0]=c.getYear(),s=c.getTime(),BigInt(isNaN(s)?-1:s/1e3)}function xi(s,c,f,b,w,I,O){return g?$e(16,1,s,c,f,b,w,I,O):-52}function Si(s,c,f,b,w,I){if(g)return $e(17,1,s,c,f,b,w,I)}function xp(s,c,f,b){s>>>=0,c>>>=0,f>>>=0,b>>>=0;var w=new Date().getFullYear(),I=new Date(w,0,1),O=new Date(w,6,1);w=I.getTimezoneOffset();var B=O.getTimezoneOffset(),G=Math.max(w,B);a()[s>>>2>>>0]=60*G,i()[c>>>2>>>0]=+(w!=B),I=(s=F=>F.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1])(I),O=s(O),B<w?(Dt(I,f,17),Dt(O,b,17)):(Dt(I,b,17),Dt(O,f,17))}var En=[],Ti=(s,c)=>{En.length=0;for(var f;f=r()[s++>>>0];){var b=f!=105;c+=(b&=f!=112)&&c%8?4:0,En.push(f==112?a()[c>>>2>>>0]:f==106?W[c>>>3]:f==105?i()[c>>>2>>>0]:l()[c>>>3>>>0]),c+=b?8:4}return En};function Sp(s,c,f){return s>>>=0,c=Ti(c>>>0,f>>>0),gn[s](...c)}function Tp(s,c,f){return s>>>=0,c=Ti(c>>>0,f>>>0),gn[s](...c)}var Ip=()=>{},Cp=()=>Date.now();function Ap(s,c){return L(Ce(s>>>0,c>>>0))}var Ii,kp=()=>{throw vt+=1,"unwind"};function Ep(){return 4294901760}Ii=()=>performance.timeOrigin+performance.now();var Pp=()=>navigator.hardwareConcurrency;function zp(){return ct("Cannot use emscripten_pc_get_function without -sUSE_OFFSET_CONVERTER"),0}function Op(s){s>>>=0;var c=r().length;if(s<=c||4294901760<s)return!1;for(var f=1;4>=f;f*=2){var b=c*(1+.2/f);b=Math.min(b,s+100663296);var w=Math;b=Math.max(s,b);e:{w=(w.min.call(w,4294901760,b+(65536-b%65536)%65536)-K.buffer.byteLength+65535)/65536;try{K.grow(w),pe();var I=1;break e}catch{}I=void 0}if(I)return!0}return!1}var lr=()=>(ct("Cannot use convertFrameToPC (needed by __builtin_return_address) without -sUSE_OFFSET_CONVERTER"),0),Mt={},Ci=s=>{s.forEach(c=>{var f=lr();f&&(Mt[f]=c)})};function Dp(){var s=Error().stack.toString().split(`
+`);return s[0]=="Error"&&s.shift(),Ci(s),Mt.Qb=lr(),Mt.fc=s,Mt.Qb}function Bp(s,c,f){if(s>>>=0,c>>>=0,Mt.Qb==s)var b=Mt.fc;else(b=Error().stack.toString().split(`
+`))[0]=="Error"&&b.shift(),Ci(b);for(var w=3;b[w]&&lr()!=s;)++w;for(s=0;s<f&&b[s+w];++s)i()[c+4*s>>>2>>>0]=lr();return s}var Pn,zn={},Ai=()=>{if(!Pn){var s,c={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:"./this.program"};for(s in zn)zn[s]===void 0?delete c[s]:c[s]=zn[s];var f=[];for(s in c)f.push(`${s}=${c[s]}`);Pn=f}return Pn};function ki(s,c){if(g)return $e(18,1,s,c);s>>>=0,c>>>=0;var f=0;return Ai().forEach((b,w)=>{var I=c+f;for(w=a()[s+4*w>>>2>>>0]=I,I=0;I<b.length;++I)t()[w++>>>0]=b.charCodeAt(I);t()[w>>>0]=0,f+=b.length+1}),0}function Ei(s,c){if(g)return $e(19,1,s,c);s>>>=0,c>>>=0;var f=Ai();a()[s>>>2>>>0]=f.length;var b=0;return f.forEach(w=>b+=w.length+1),a()[c>>>2>>>0]=b,0}function Pi(s){return g?$e(20,1,s):52}function zi(s,c,f,b){return g?$e(21,1,s,c,f,b):52}function Oi(s,c,f,b){return g?$e(22,1,s,c,f,b):70}var Mp=[null,[],[]];function Di(s,c,f,b){if(g)return $e(23,1,s,c,f,b);c>>>=0,f>>>=0,b>>>=0;for(var w=0,I=0;I<f;I++){var O=a()[c>>>2>>>0],B=a()[c+4>>>2>>>0];c+=8;for(var G=0;G<B;G++){var F=r()[O+G>>>0],X=Mp[s];F===0||F===10?((s===1?H:L)(Yo(X,0)),X.length=0):X.push(F)}w+=B}return a()[b>>>2>>>0]=w,0}var Bi=[31,29,31,30,31,30,31,31,30,31,30,31],Mi=[31,28,31,30,31,30,31,31,30,31,30,31],Rp=(s,c)=>{t().set(s,c>>>0)};function Ri(s,c,f,b){function w(z,ue,Se){for(z=typeof z=="number"?z.toString():z||"";z.length<ue;)z=Se[0]+z;return z}function I(z,ue){return w(z,ue,"0")}function O(z,ue){function Se(Qi){return 0>Qi?-1:0<Qi?1:0}var xt;return(xt=Se(z.getFullYear()-ue.getFullYear()))===0&&(xt=Se(z.getMonth()-ue.getMonth()))===0&&(xt=Se(z.getDate()-ue.getDate())),xt}function B(z){switch(z.getDay()){case 0:return new Date(z.getFullYear()-1,11,29);case 1:return z;case 2:return new Date(z.getFullYear(),0,3);case 3:return new Date(z.getFullYear(),0,2);case 4:return new Date(z.getFullYear(),0,1);case 5:return new Date(z.getFullYear()-1,11,31);case 6:return new Date(z.getFullYear()-1,11,30)}}function G(z){var ue=z.Cb;for(z=new Date(new Date(z.Db+1900,0,1).getTime());0<ue;){var Se=z.getMonth(),xt=(Bt(z.getFullYear())?Bi:Mi)[Se];if(!(ue>xt-z.getDate())){z.setDate(z.getDate()+ue);break}ue-=xt-z.getDate()+1,z.setDate(1),11>Se?z.setMonth(Se+1):(z.setMonth(0),z.setFullYear(z.getFullYear()+1))}return Se=new Date(z.getFullYear()+1,0,4),ue=B(new Date(z.getFullYear(),0,4)),Se=B(Se),0>=O(ue,z)?0>=O(Se,z)?z.getFullYear()+1:z.getFullYear():z.getFullYear()-1}s>>>=0,c>>>=0,f>>>=0,b>>>=0;var F=a()[b+40>>>2>>>0];for(var X in b={lc:i()[b>>>2>>>0],kc:i()[b+4>>>2>>>0],Ib:i()[b+8>>>2>>>0],Mb:i()[b+12>>>2>>>0],Jb:i()[b+16>>>2>>>0],Db:i()[b+20>>>2>>>0],vb:i()[b+24>>>2>>>0],Cb:i()[b+28>>>2>>>0],sc:i()[b+32>>>2>>>0],jc:i()[b+36>>>2>>>0],mc:F?Ce(F):""},f=Ce(f),F={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"})f=f.replace(new RegExp(X,"g"),F[X]);var le="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),ge="January February March April May June July August September October November December".split(" ");for(X in F={"%a":z=>le[z.vb].substring(0,3),"%A":z=>le[z.vb],"%b":z=>ge[z.Jb].substring(0,3),"%B":z=>ge[z.Jb],"%C":z=>I((z.Db+1900)/100|0,2),"%d":z=>I(z.Mb,2),"%e":z=>w(z.Mb,2," "),"%g":z=>G(z).toString().substring(2),"%G":G,"%H":z=>I(z.Ib,2),"%I":z=>((z=z.Ib)==0?z=12:12<z&&(z-=12),I(z,2)),"%j":z=>{for(var ue=0,Se=0;Se<=z.Jb-1;ue+=(Bt(z.Db+1900)?Bi:Mi)[Se++]);return I(z.Mb+ue,3)},"%m":z=>I(z.Jb+1,2),"%M":z=>I(z.kc,2),"%n":()=>`
+`,"%p":z=>0<=z.Ib&&12>z.Ib?"AM":"PM","%S":z=>I(z.lc,2),"%t":()=>"	","%u":z=>z.vb||7,"%U":z=>I(Math.floor((z.Cb+7-z.vb)/7),2),"%V":z=>{var ue=Math.floor((z.Cb+7-(z.vb+6)%7)/7);if(2>=(z.vb+371-z.Cb-2)%7&&ue++,ue)ue==53&&((Se=(z.vb+371-z.Cb)%7)==4||Se==3&&Bt(z.Db)||(ue=1));else{ue=52;var Se=(z.vb+7-z.Cb-1)%7;(Se==4||Se==5&&Bt(z.Db%400-1))&&ue++}return I(ue,2)},"%w":z=>z.vb,"%W":z=>I(Math.floor((z.Cb+7-(z.vb+6)%7)/7),2),"%y":z=>(z.Db+1900).toString().substring(2),"%Y":z=>z.Db+1900,"%z":z=>{var ue=0<=(z=z.jc);return z=Math.abs(z)/60,(ue?"+":"-")+("0000"+(z/60*100+z%60)).slice(-4)},"%Z":z=>z.mc,"%%":()=>"%"},f=f.replace(/%%/g,"\0\0"),F)f.includes(X)&&(f=f.replace(new RegExp(X,"g"),F[X](b)));return X=function(z){var ue=Array(vn(z)+1);return Xo(z,ue,0,ue.length),ue}(f=f.replace(/\0\0/g,"%")),X.length>c?0:(Rp(X,s),X.length-1)}function Up(s,c,f,b){return Ri(s>>>0,c>>>0,f>>>0,b>>>0)}g||function(){for(var s=u.numThreads-1;s--;)Fo();Ie.unshift(()=>{wt++,function(c){g?c():Promise.all(pt.map(Ho)).then(c)}(()=>Oo())})}();for(var Ui=Array(256),cr=0;256>cr;++cr)Ui[cr]=String.fromCharCode(cr);li=Ui,mt=u.BindingError=class extends Error{constructor(s){super(s),this.name="BindingError"}},u.InternalError=class extends Error{constructor(s){super(s),this.name="InternalError"}},ut.push(0,1,void 0,1,null,1,!0,1,!1,1),u.count_emval_handles=()=>ut.length/2-5-Sn.length;var Np=[_n,Vo,qo,Zo,Qo,Jo,ei,ti,ri,ni,oi,ii,ai,si,ui,di,xi,Si,ki,Ei,Pi,zi,Oi,Di],Z=function(){function s(f,b){return Z=f.exports,Z=function(){var w=Z,I={};for(let[O,B]of Object.entries(w))I[O]=typeof B=="function"?(...G)=>{sr.push(O);try{return B(...G)}finally{De||(sr.pop(),Xe&&ht===1&&sr.length===0&&(ht=0,vt+=1,ar(Ki),typeof Fibers<"u"&&Fibers.tc()))}}:B;return I}(),Z=function(){var w=Z,I=B=>G=>B(G)>>>0,O=B=>()=>B()>>>0;return(w=Object.assign({},w)).Da=I(w.Da),w.gb=O(w.gb),w.ib=I(w.ib),w.emscripten_main_runtime_thread_id=O(w.emscripten_main_runtime_thread_id),w.tb=I(w.tb),w.ub=O(w.ub),w}(),Wo.push(Z.jb),He.unshift(Z.Ca),we=b,Oo(),Z}var c=Uo();if(wt++,u.instantiateWasm)try{return u.instantiateWasm(c,s)}catch(f){L(`Module.instantiateWasm callback failed with error: ${f}`),m(f)}return hn||=u.locateFile?Do("ort-wasm-simd-threaded.jsep.wasm")?"ort-wasm-simd-threaded.jsep.wasm":u.locateFile?u.locateFile("ort-wasm-simd-threaded.jsep.wasm",C):C+"ort-wasm-simd-threaded.jsep.wasm":new URL(/* asset import */ __webpack_require__(/*! ort-wasm-simd-threaded.jsep.wasm */ "./node_modules/onnxruntime-web/dist/ort-wasm-simd-threaded.jsep.wasm"), __webpack_require__.b).href,function(f,b){var w=hn;return P||typeof WebAssembly.instantiateStreaming!="function"||Do(w)||Bo(w)||typeof fetch!="function"?Ro(w,f,b):fetch(w,{credentials:"same-origin"}).then(I=>WebAssembly.instantiateStreaming(I,f).then(b,function(O){return L(`wasm streaming compile failed: ${O}`),L("falling back to ArrayBuffer instantiation"),Ro(w,f,b)}))}(c,function(f){s(f.instance,f.module)}).catch(m),{}}(),Ni=s=>(Ni=Z.Da)(s),Vi=()=>(Vi=Z.Ea)();u._OrtInit=(s,c)=>(u._OrtInit=Z.Fa)(s,c),u._OrtGetLastError=(s,c)=>(u._OrtGetLastError=Z.Ga)(s,c),u._OrtCreateSessionOptions=(s,c,f,b,w,I,O,B,G,F)=>(u._OrtCreateSessionOptions=Z.Ha)(s,c,f,b,w,I,O,B,G,F),u._OrtAppendExecutionProvider=(s,c)=>(u._OrtAppendExecutionProvider=Z.Ia)(s,c),u._OrtAddFreeDimensionOverride=(s,c,f)=>(u._OrtAddFreeDimensionOverride=Z.Ja)(s,c,f),u._OrtAddSessionConfigEntry=(s,c,f)=>(u._OrtAddSessionConfigEntry=Z.Ka)(s,c,f),u._OrtReleaseSessionOptions=s=>(u._OrtReleaseSessionOptions=Z.La)(s),u._OrtCreateSession=(s,c,f)=>(u._OrtCreateSession=Z.Ma)(s,c,f),u._OrtReleaseSession=s=>(u._OrtReleaseSession=Z.Na)(s),u._OrtGetInputOutputCount=(s,c,f)=>(u._OrtGetInputOutputCount=Z.Oa)(s,c,f),u._OrtGetInputName=(s,c)=>(u._OrtGetInputName=Z.Pa)(s,c),u._OrtGetOutputName=(s,c)=>(u._OrtGetOutputName=Z.Qa)(s,c),u._OrtFree=s=>(u._OrtFree=Z.Ra)(s),u._OrtCreateTensor=(s,c,f,b,w,I)=>(u._OrtCreateTensor=Z.Sa)(s,c,f,b,w,I),u._OrtGetTensorData=(s,c,f,b,w)=>(u._OrtGetTensorData=Z.Ta)(s,c,f,b,w),u._OrtReleaseTensor=s=>(u._OrtReleaseTensor=Z.Ua)(s),u._OrtCreateRunOptions=(s,c,f,b)=>(u._OrtCreateRunOptions=Z.Va)(s,c,f,b),u._OrtAddRunConfigEntry=(s,c,f)=>(u._OrtAddRunConfigEntry=Z.Wa)(s,c,f),u._OrtReleaseRunOptions=s=>(u._OrtReleaseRunOptions=Z.Xa)(s),u._OrtCreateBinding=s=>(u._OrtCreateBinding=Z.Ya)(s),u._OrtBindInput=(s,c,f)=>(u._OrtBindInput=Z.Za)(s,c,f),u._OrtBindOutput=(s,c,f,b)=>(u._OrtBindOutput=Z._a)(s,c,f,b),u._OrtClearBoundOutputs=s=>(u._OrtClearBoundOutputs=Z.$a)(s),u._OrtReleaseBinding=s=>(u._OrtReleaseBinding=Z.ab)(s),u._OrtRunWithBinding=(s,c,f,b,w)=>(u._OrtRunWithBinding=Z.bb)(s,c,f,b,w),u._OrtRun=(s,c,f,b,w,I,O,B)=>(u._OrtRun=Z.cb)(s,c,f,b,w,I,O,B),u._OrtEndProfiling=s=>(u._OrtEndProfiling=Z.db)(s),u._JsepOutput=(s,c,f)=>(u._JsepOutput=Z.eb)(s,c,f),u._JsepGetNodeName=s=>(u._JsepGetNodeName=Z.fb)(s);var pr,Rt=()=>(Rt=Z.gb)(),Je=u._free=s=>(Je=u._free=Z.hb)(s),mr=u._malloc=s=>(mr=u._malloc=Z.ib)(s),On=(s,c,f,b,w,I)=>(On=Z.lb)(s,c,f,b,w,I),Wi=()=>(Wi=Z.mb)(),Li=(s,c,f,b,w)=>(Li=Z.nb)(s,c,f,b,w),Dn=s=>(Dn=Z.ob)(s),fr=s=>(fr=Z.pb)(s),Gi=()=>(Gi=Z.qb)(),Hi=(s,c)=>(Hi=Z.rb)(s,c),hr=s=>(hr=Z.sb)(s),Bn=s=>(Bn=Z.tb)(s),Mn=()=>(Mn=Z.ub)(),Fi=u.dynCall_ii=(s,c)=>(Fi=u.dynCall_ii=Z.wb)(s,c),qi=s=>(qi=Z.xb)(s),Ki=()=>(Ki=Z.yb)(),ji=s=>(ji=Z.zb)(s),Yi=()=>(Yi=Z.Ab)();function Zi(){0<wt||(g?(p(u),g||nr(He),startWorker(u)):(nr(Ie),0<wt||pr||(pr=!0,u.calledRun=!0,De||(g||nr(He),p(u),g||nr(mn)))))}return u.___start_em_js=932469,u.___stop_em_js=932715,u.stackSave=()=>Mn(),u.stackRestore=s=>hr(s),u.stackAlloc=s=>Bn(s),u.setValue=function(s,c,f="i8"){switch(f.endsWith("*")&&(f="*"),f){case"i1":case"i8":t()[s>>>0]=c;break;case"i16":n()[s>>>1>>>0]=c;break;case"i32":i()[s>>>2>>>0]=c;break;case"i64":W[s>>>3]=BigInt(c);break;case"float":d()[s>>>2>>>0]=c;break;case"double":l()[s>>>3>>>0]=c;break;case"*":a()[s>>>2>>>0]=c;break;default:ct(`invalid type for setValue: ${f}`)}},u.getValue=function(s,c="i8"){switch(c.endsWith("*")&&(c="*"),c){case"i1":case"i8":return t()[s>>>0];case"i16":return n()[s>>>1>>>0];case"i32":return i()[s>>>2>>>0];case"i64":return W[s>>>3];case"float":return d()[s>>>2>>>0];case"double":return l()[s>>>3>>>0];case"*":return a()[s>>>2>>>0];default:ct(`invalid type for getValue: ${c}`)}},u.UTF8ToString=Ce,u.stringToUTF8=Dt,u.lengthBytesUTF8=vn,Gt=function s(){pr||Zi(),pr||(Gt=s)},Zi(),u.PTR_SIZE=4,h}),Kp=Pa;globalThis.self?.name==="em-pthread"&&Pa()});var Ma,jp,Ve,Ra,Kn,Yp,Zp,Ua,Qp,Da,Na,Ba,Va,$r=U(()=>{"use strict";vr();Ma= false||typeof location>"u"?void 0:location.origin,jp=()=>{if(true)return import.meta.url?.startsWith("file:")?new URL(new URL(/* asset import */ __webpack_require__(/*! ort.bundle.min.mjs */ "./node_modules/onnxruntime-web/dist/ort.bundle.min.mjs?46eb"), __webpack_require__.b).href,Ma).href:import.meta.url},Ve=jp(),Ra=()=>{if(Ve&&!Ve.startsWith("blob:"))return Ve.substring(0,Ve.lastIndexOf("/")+1)},Kn=(e,t)=>{try{let r=t??Ve;return(r?new URL(e,r):new URL(e)).origin===Ma}catch{return!1}},Yp=(e,t)=>{let r=t??Ve;try{return(r?new URL(e,r):new URL(e)).href}catch{return}},Zp=(e,t)=>`${t??"./"}${e}`,Ua=async e=>{let r=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(r)},Qp=async e=>(await import(/*webpackIgnore:true*/e)).default,Da=(Ea(),gr(ka)).default,Na=async()=>{if(!Ve)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(Kn(Ve))return[void 0,Da()];let e=await Ua(Ve);return[e,Da(e)]},Ba=(Oa(),gr(za)).default,Va=async(e,t,r)=>{if(!e&&!t&&Ba&&Ve&&Kn(Ve))return[void 0,Ba];{let n="ort-wasm-simd-threaded.jsep.mjs",o=e??Yp(n,t),i= true&&r&&o&&!Kn(o,t),a=i?await Ua(o):o??Zp(n,t);return[i?a:void 0,await Qp(a)]}}});var jn,Yn,Pr,Wa,Xp,Jp,xr,Te,gt=U(()=>{"use strict";$r();Yn=!1,Pr=!1,Wa=!1,Xp=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},Jp=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},xr=async e=>{if(Yn)return Promise.resolve();if(Pr)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Wa)throw new Error("previous call to 'initializeWebAssembly()' failed.");Pr=!0;let t=e.initTimeout,r=e.numThreads;if(!Jp())throw new Error("WebAssembly SIMD is not supported in the current environment.");let n=Xp();r>1&&!n&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+r+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=r=1);let o=e.wasmPaths,i=typeof o=="string"?o:void 0,a=o?.mjs,d=a?.href??a,l=o?.wasm,p=l?.href??l,m=e.wasmBinary,[u,h]=await Va(d,i,r>1),_=!1,y=[];if(t>0&&y.push(new Promise(g=>{setTimeout(()=>{_=!0,g()},t)})),y.push(new Promise((g,x)=>{let $={numThreads:r};if(m)$.wasmBinary=m;else if(p||i)$.locateFile=v=>p??i+v;else if(d&&d.indexOf("blob:")!==0)$.locateFile=v=>new URL(v,d).href;else if(u){let v=Ra();v&&($.locateFile=S=>v+S)}h($).then(v=>{Pr=!1,Yn=!0,jn=v,g(),u&&URL.revokeObjectURL(u)},v=>{Pr=!1,Wa=!0,x(v)})})),await Promise.race(y),_)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Te=()=>{if(Yn&&jn)return jn;throw new Error("WebAssembly is not initialized yet.")}});var ke,Kt,ce,zr=U(()=>{"use strict";gt();ke=(e,t)=>{let r=Te(),n=r.lengthBytesUTF8(e)+1,o=r._malloc(n);return r.stringToUTF8(e,o,n),t.push(o),o},Kt=(e,t,r,n)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([o,i])=>{let a=t?t+o:o;if(typeof i=="object")Kt(i,a+".",r,n);else if(typeof i=="string"||typeof i=="number")n(a,i.toString());else if(typeof i=="boolean")n(a,i?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof i}`)})},ce=e=>{let t=Te(),r=t.stackSave();try{let n=t.PTR_SIZE,o=t.stackAlloc(2*n);t._OrtGetLastError(o,o+n);let i=Number(t.getValue(o,n===4?"i32":"i64")),a=t.getValue(o+n,"*"),d=a?t.UTF8ToString(a):"";throw new Error(`${e} ERROR_CODE: ${i}, ERROR_MESSAGE: ${d}`)}finally{t.stackRestore(r)}}});var La,Ga=U(()=>{"use strict";gt();zr();La=e=>{let t=Te(),r=0,n=[],o=e||{};try{if(e?.logSeverityLevel===void 0)o.logSeverityLevel=2;else if(typeof e.logSeverityLevel!="number"||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${e.logSeverityLevel}`);if(e?.logVerbosityLevel===void 0)o.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);e?.terminate===void 0&&(o.terminate=!1);let i=0;return e?.tag!==void 0&&(i=ke(e.tag,n)),r=t._OrtCreateRunOptions(o.logSeverityLevel,o.logVerbosityLevel,!!o.terminate,i),r===0&&ce("Can't create run options."),e?.extra!==void 0&&Kt(e.extra,"",new WeakSet,(a,d)=>{let l=ke(a,n),p=ke(d,n);t._OrtAddRunConfigEntry(r,l,p)!==0&&ce(`Can't set a run config entry: ${a} - ${d}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseRunOptions(r),n.forEach(a=>t._free(a)),i}}});var em,tm,rm,nm,Ha,Fa=U(()=>{"use strict";gt();zr();em=e=>{switch(e){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${e}`)}},tm=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},rm=e=>{e.extra||(e.extra={}),e.extra.session||(e.extra.session={});let t=e.extra.session;t.use_ort_model_bytes_directly||(t.use_ort_model_bytes_directly="1"),e.executionProviders&&e.executionProviders.some(r=>(typeof r=="string"?r:r.name)==="webgpu")&&(e.enableMemPattern=!1)},nm=(e,t,r)=>{for(let n of t){let o=typeof n=="string"?n:n.name;switch(o){case"webnn":if(o="WEBNN",typeof n!="string"){let d=n?.deviceType;if(d){let l=ke("deviceType",r),p=ke(d,r);Te()._OrtAddSessionConfigEntry(e,l,p)!==0&&ce(`Can't set a session config entry: 'deviceType' - ${d}.`)}}break;case"webgpu":if(o="JS",typeof n!="string"){let a=n;if(a?.preferredLayout){if(a.preferredLayout!=="NCHW"&&a.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${a.preferredLayout}`);let d=ke("preferredLayout",r),l=ke(a.preferredLayout,r);Te()._OrtAddSessionConfigEntry(e,d,l)!==0&&ce(`Can't set a session config entry: 'preferredLayout' - ${a.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${o}`)}let i=ke(o,r);Te()._OrtAppendExecutionProvider(e,i)!==0&&ce(`Can't append execution provider: ${o}.`)}},Ha=e=>{let t=Te(),r=0,n=[],o=e||{};rm(o);try{let i=em(o.graphOptimizationLevel??"all"),a=tm(o.executionMode??"sequential"),d=typeof o.logId=="string"?ke(o.logId,n):0,l=o.logSeverityLevel??2;if(!Number.isInteger(l)||l<0||l>4)throw new Error(`log serverity level is not valid: ${l}`);let p=o.logVerbosityLevel??0;if(!Number.isInteger(p)||p<0||p>4)throw new Error(`log verbosity level is not valid: ${p}`);let m=typeof o.optimizedModelFilePath=="string"?ke(o.optimizedModelFilePath,n):0;if(r=t._OrtCreateSessionOptions(i,!!o.enableCpuMemArena,!!o.enableMemPattern,a,!!o.enableProfiling,0,d,l,p,m),r===0&&ce("Can't create session options."),o.executionProviders&&nm(r,o.executionProviders,n),o.enableGraphCapture!==void 0){if(typeof o.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${o.enableGraphCapture}`);let u=ke("enableGraphCapture",n),h=ke(o.enableGraphCapture.toString(),n);t._OrtAddSessionConfigEntry(r,u,h)!==0&&ce(`Can't set a session config entry: 'enableGraphCapture' - ${o.enableGraphCapture}.`)}if(o.freeDimensionOverrides)for(let[u,h]of Object.entries(o.freeDimensionOverrides)){if(typeof u!="string")throw new Error(`free dimension override name must be a string: ${u}`);if(typeof h!="number"||!Number.isInteger(h)||h<0)throw new Error(`free dimension override value must be a non-negative integer: ${h}`);let _=ke(u,n);t._OrtAddFreeDimensionOverride(r,_,h)!==0&&ce(`Can't set a free dimension override: ${u} - ${h}.`)}return o.extra!==void 0&&Kt(o.extra,"",new WeakSet,(u,h)=>{let _=ke(u,n),y=ke(h,n);t._OrtAddSessionConfigEntry(r,_,y)!==0&&ce(`Can't set a session config entry: ${u} - ${h}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseSessionOptions(r)!==0&&ce("Can't release session options."),n.forEach(a=>t._free(a)),i}}});var jt,bt,At,Or,Yt,Dr,Br,Zn,J=U(()=>{"use strict";jt=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw new Error(`unsupported data type: ${e}`)}},bt=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";case 22:return"int4";case 21:return"uint4";default:throw new Error(`unsupported data type: ${e}`)}},At=(e,t)=>{let r=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],n=typeof t=="number"?t:t.reduce((o,i)=>o*i,1);return r>0?Math.ceil(n*r):void 0},Or=e=>{switch(e){case"float16":return typeof Float16Array<"u"&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"bool":return Uint8Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw new Error(`unsupported type: ${e}`)}},Yt=e=>{switch(e){case"verbose":return 0;case"info":return 1;case"warning":return 2;case"error":return 3;case"fatal":return 4;default:throw new Error(`unsupported logging level: ${e}`)}},Dr=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Br=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Zn=e=>{switch(e){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;case"ml-tensor":return 5;default:throw new Error(`unsupported data location: ${e}`)}}});var Zt,Qn=U(()=>{"use strict";vr();Zt=async e=>{if(typeof e=="string")if(false){}else{let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),n=r?parseInt(r,10):0;if(n<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let o=t.body.getReader(),i;try{i=new ArrayBuffer(n)}catch(d){if(d instanceof RangeError){let l=Math.ceil(n/65536);i=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw d}let a=0;for(;;){let{done:d,value:l}=await o.read();if(d)break;let p=l.byteLength;new Uint8Array(i,a,p).set(l),a+=p}return new Uint8Array(i,0,n)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}});var om,im,qa,Ka,Mr,am,se,et=U(()=>{"use strict";J();om=["V","I","W","E","F"],im=(e,t)=>{console.log(`[${om[e]},${new Date().toISOString()}]${t}`)},Mr=(e,t)=>{qa=e,Ka=t},am=(e,t)=>{let r=Yt(e),n=Yt(qa);r>=n&&im(r,typeof t=="function"?t():t)},se=(...e)=>{Ka&&am(...e)}});var Rr,Xn=U(()=>{"use strict";J();Rr=(e,t)=>new(Or(t))(e)});var Ur=U(()=>{"use strict"});var ja,Jn,eo,sm,um,Ya,ro,to,Qa,Xa=U(()=>{"use strict";et();Ur();ja=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),Jn=[],eo=e=>Math.ceil(Number(e)/16)*16,sm=e=>{for(let t=0;t<Jn.length;t++){let r=Jn[t];if(e<=r)return r}return Math.ceil(e/16)*16},um=1,Ya=()=>um++,ro=async(e,t,r,n)=>{let o=eo(r),i=e.device.createBuffer({size:o,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let a=e.getCommandEncoder();e.endComputePass(),a.copyBufferToBuffer(t,0,i,0,o),e.flush(),await i.mapAsync(GPUMapMode.READ);let d=i.getMappedRange();if(n){let l=n();return l.set(new Uint8Array(d,0,r)),l}else return new Uint8Array(d.slice(0,r))}finally{i.destroy()}},to=class{constructor(t){this.backend=t;this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersPending=[],this.capturedPendingBuffers=new Map;for(let[r]of ja)Jn.push(r),this.freeBuffers.set(r,[]),this.freeUniformBuffers.set(r,[]);this.sessionCount=0}upload(t,r){let n=r.buffer,o=r.byteOffset,i=r.byteLength,a=eo(i),d=this.storageCache.get(t);if(!d)throw new Error("gpu data for uploading does not exist");if(Number(d.originalSize)!==i)throw new Error(`inconsistent data size. gpu data size=${d.originalSize}, data size=${i}`);let l=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),p=l.getMappedRange();new Uint8Array(p).set(new Uint8Array(n,o,i)),l.unmap();let m=this.backend.device.createCommandEncoder();m.copyBufferToBuffer(l,0,d.gpuData.buffer,0,a),this.backend.device.queue.submit([m.finish()]),l.destroy(),se("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`)}memcpy(t,r){let n=this.storageCache.get(t);if(!n)throw new Error("source gpu data for memcpy does not exist");let o=this.storageCache.get(r);if(!o)throw new Error("destination gpu data for memcpy does not exist");if(n.originalSize!==o.originalSize)throw new Error("inconsistent source and destination gpu data size");let i=eo(n.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(n.gpuData.buffer,0,o.gpuData.buffer,0,i)}registerExternalBuffer(t,r,n){let o;if(n){if(o=n[0],t===n[1])return se("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${o}, buffer is the same, skip.`),o;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
+             Please use the previous external buffer!`)}else o=Ya();return this.storageCache.set(o,{gpuData:{id:o,type:0,buffer:t},originalSize:r}),se("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${o}, registered.`),o}unregisterExternalBuffer(t){t!==void 0&&(this.storageCache.delete(t),se("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${t}`))}create(t,r=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let n=sm(t),o,i=(r&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(r&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(i||a){let p=(i?this.freeBuffers:this.freeUniformBuffers).get(n);p?p.length>0?o=p.pop():o=this.backend.device.createBuffer({size:n,usage:r}):o=this.backend.device.createBuffer({size:n,usage:r})}else o=this.backend.device.createBuffer({size:n,usage:r});let d={id:Ya(),type:0,buffer:o};return this.storageCache.set(d.id,{gpuData:d,originalSize:Number(t)}),se("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${d.id}`),d}get(t){return this.storageCache.get(t)?.gpuData}release(t){let r=typeof t=="bigint"?Number(t):t,n=this.storageCache.get(r);if(!n){if(this.storageCache.size===0)return 0;throw new Error("releasing data does not exist")}return se("verbose",()=>`[WebGPU] GpuDataManager.release(id=${r}), gpuDataId=${n.gpuData.id}`),this.storageCache.delete(r),this.buffersPending.push(n.gpuData.buffer),n.originalSize}async download(t,r){let n=this.storageCache.get(Number(t));if(!n)throw new Error("data does not exist");await ro(this.backend,n.gpuData.buffer,n.originalSize,r)}refreshPendingBuffers(){if(this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let r=ja.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let n=this.freeBuffers.get(t.size)||[];r===void 0||n.length>=r?t.destroy():n.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let n=this.freeUniformBuffers.get(t.size)||[];r===void 0||n.length>=r?t.destroy():n.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let r of this.buffersPending)t.push(r);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(t){let r=this.capturedPendingBuffers.get(t);r&&(r.forEach(n=>{n.destroy()}),this.capturedPendingBuffers.delete(t)),this.sessionCount-=1,this.sessionCount===0&&(se("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(n=>{n.gpuData.buffer.destroy()}),this.storageCache=new Map)}},Qa=(...e)=>new to(...e)});var no,ee,xe=U(()=>{"use strict";no=class{constructor(t){Object.assign(this,t)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(t=>`${this[t]}`).join(";")),this.key}},ee=e=>new no(e)});var oo,tt,k,kt,Nr,Ja,es,ne=U(()=>{"use strict";oo=class{static calcMatMulShape(t,r){return t[1]!==r[0]?void 0:[t[0],r[1]]}},tt=class{static calcShape(t,r,n=!1){let o=t.length,i=r.length;if(o===0)return r;if(i===0)return t;let a=Math.max(t.length,r.length),d=new Array(a);if(n){if(o<2||i<2)return;let l=oo.calcMatMulShape([t[o-2],t[o-1]],[r[i-2],r[i-1]]);if(l===void 0)return;[d[a-2],d[a-1]]=l}for(let l=n?3:1;l<=a;l++){let p=o-l<0?1:t[o-l],m=i-l<0?1:r[i-l];if(p!==m&&p>1&&m>1)return;let u=Math.max(p,m);if(p&&m)d[a-l]=Math.max(p,m);else{if(u>1)return;d[a-l]=0}}return d}static isValidBroadcast(t,r){let n=t.length,o=r.length;if(n>o)return!1;for(let i=1;i<=n;i++)if(t[n-i]!==1&&t[n-i]!==r[o-i])return!1;return!0}},k=class e{static size(t){return e.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let n=t.length;if(n===0)return[];let o=new Array(n),i=n-1;for(;i>=0;){if(t[i]%r===0){o[i]=t[i]/r;break}if(r%t[i]!==0)throw new Error("cannot convert shape");o[i]=1,r/=t[i],i--}for(i--;i>=0;i--)o[i]=t[i];return o}static sizeFromDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,r,t.length)}static sizeToDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,n){let o=1;for(let i=r;i<n;i++){if(t[i]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");o*=Number(t[i])}return o}static computeStrides(t){let r=t.length;if(r===0)return[];if(r===1)return[1];let n=new Array(r);n[r-1]=1,n[r-2]=t[r-1];for(let o=r-3;o>=0;--o)n[o]=n[o+1]*t[o+1];return n}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(n=>this.normalizeAxis(n,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(n=>t[n]):t.slice().reverse()}static padShape(t,r){let n=t.length;return t.map((o,i)=>o+r[i]+r[i+n])}static areEqual(t,r){return t.length!==r.length?!1:t.every((n,o)=>n===r[o])}},kt=class e{static adjustPoolAttributes(t,r,n,o,i,a){if(!t&&n.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let d=0;d<r.length-2;d++)d>=n.length?n.push(r[d+2]):n[d]=r[d+2];for(let d=0;d<n.length;d++)if(d<o.length){if(o[d]<0)throw new Error("strides should be greater than or equal to 1")}else o.push(1);for(let d=0;d<n.length;d++)if(d<i.length){if(i[d]<0)throw new Error("dilations should be greater than or equal to 1")}else i.push(1);for(let d=0;d<n.length*2;d++)if(d<a.length){if(a[d]<0)throw new Error("pad should be greater than or equal to 1")}else a.push(0);for(let d=0;d<n.length;d++){if(n[d]<=0)throw new Error("kernel shapes need to be greater than 0");if(a[d]>=n[d]||a[d+n.length]>=n[d])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,n,o,i,a,d){if(d){if(i.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(o.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let l=0;l<t.length-2;l++)e.adjustPadAndReturnShape(t[l+(a?1:2)],r[l],n[l],o[l],i,l,l+t.length-2,d)}}static computePoolOutputShape(t,r,n,o,i,a,d){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let l=[r[0],r[1]];return e.computeShapeHelper(t,r,l,n,o,i,a,d),l}static computeConvOutputShape(t,r,n,o,i,a,d){if(t.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let l=[t[0],r[0]];return e.computeShapeHelper(!1,t,l,n,o,i,a,d),l}static computeShapeHelper(t,r,n,o,i,a,d,l){if(t)for(let p=0;p<r.length-2;p++)n.push(1);else for(let p=0;p<r.length-2;p++)n.push(e.adjustPadAndReturnShape(r[p+2],o[p],i[p],a[p],d,p,p+r.length-2,l))}static adjustPadAndReturnShape(t,r,n,o,i,a,d,l){let p=n*(o-1)+1;if(l&&l!=="NOTSET")switch(l){case"VALID":return i[a]=0,i[d]=0,Math.floor((t-p)/r+1);case"SAME_LOWER":case"SAME_UPPER":if(n!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let u=((t+r-1)/r-1)*r+o-t;return i[a]=Math.floor(l==="SAME_LOWER"?(u+1)/2:u/2),i[d]=u-i[a],Math.floor((t+u-o)/r+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((t+i[a]+i[d]-p)/r+1)}},Nr=class{static getShapeOfGemmResult(t,r,n,o,i){if(t.length!==2||n.length!==2)throw new Error("shape need to be of size 2");let a,d,l;r?(a=t[1],d=t[0]):(a=t[0],d=t[1]);let p=-1;if(o?(l=n[0],p=1):(l=n[1],p=0),n[p]!==d)throw new Error("dimension mismatch");if(a<=0||l<=0||d<=0)throw new Error("invalid shape specified");if(i&&!tt.isValidBroadcast(i,[a,l]))throw new Error("gemm: invalid bias shape for broadcast");return[a,l,d]}},Ja=-34028234663852886e22,es=34028234663852886e22});var Et,ao,ye,Ee,N,me,so,Pt,qe,q,Vr,E,M,ts,Wr,io,rs,ae=U(()=>{"use strict";J();ne();Et=64,ao=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${t}<f16>`:"f16";case 1:return t>1?`vec${t}<f32>`:"f32";case 6:return t>1?`vec${t}<i32>`:"i32";case 12:return t>1?`vec${t}<u32>`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];case 22:return"i32";case 21:return"u32";default:throw new Error(`Unknown data type: ${e}`)}},ye=(e,t=1)=>{let r=ao(e,t);return typeof r=="string"?r:r[0]},Ee=(e,t=1)=>{let r=ao(e,t);return typeof r=="string"?r:r[1]},N=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:k.computeStrides(r)})}),t},me=e=>e%4===0?4:e%2===0?2:1,so=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,Pt=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,qe=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,q=(e,t,r,n)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?n==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:n==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Vr=(e,t,r,n,o)=>{let i=typeof r=="number",a=i?r:r.length,d=[...new Array(a).keys()],l=a<2?"u32":a<=4?`vec${a}<u32>`:`array<u32, ${a}>`,p=ao(t,o),m=typeof p=="string"?p:p[1],u=typeof p=="string"?p:p[0],h={indices:l,value:m,storage:u,tensor:t},_=W=>typeof W=="string"?W:`${W}u`,y={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},g=i?"uniforms.":"",x=`${g}${e}_shape`,$=`${g}${e}_strides`,v="";for(let W=0;W<a-1;W++)v+=`
+    let dim${W} = current / ${q($,W,a)};
+    let rest${W} = current % ${q($,W,a)};
     indices[${W}] = dim${W};
     current = rest${W};
     `;v+=`indices[${a-1}] = current;`;let S=a<2?"":`
@@ -3933,25 +3988,25 @@ var Un=Object.defineProperty;var Vp=Object.getOwnPropertyDescriptor;var Wp=Objec
     var current = offset;
     ${v}
     return indices;
-  }`,T=W=>(y.offsetToIndices=!0,a<2?W:`o2i_${e}(${W})`),A=[];if(a>=2)for(let W=a-1;W>=0;W--)A.push(`${F($,W,a)} * (indices[${W}])`);let k=a<2?"":`
+  }`,T=W=>(y.offsetToIndices=!0,a<2?W:`o2i_${e}(${W})`),A=[];if(a>=2)for(let W=a-1;W>=0;W--)A.push(`${q($,W,a)} * (indices[${W}])`);let C=a<2?"":`
   fn i2o_${e}(indices: ${h.indices}) -> u32 {
     return ${A.join("+")};
-  }`,P=W=>(y.indicesToOffset=!0,a<2?W:`i2o_${e}(${W})`),D=(...W)=>a===0?"0u":`${h.indices}(${W.map(_).join(",")})`,R=(W,q)=>a<2?`${W}`:`${F(W,q,a)}`,G=(W,q,he)=>a<2?`${W}=${he};`:`${F(W,q,a)}=${he};`,K={},j=(W,q)=>{y.broadcastedIndicesToOffset=!0;let he=`${q.name}broadcastedIndicesTo${e}Offset`;if(he in K)return`${he}(${W})`;let Ge=[];for(let we=a-1;we>=0;we--){let ye=q.indicesGet("outputIndices",we+q.rank-a);Ge.push(`${R($,we)} * (${ye} % ${R(x,we)})`)}return K[he]=`fn ${he}(outputIndices: ${q.type.indices}) -> u32 {
-             return ${Ge.length>0?Ge.join("+"):"0u"};
-           }`,`${he}(${W})`},V=(W,q)=>(()=>{if(h.storage===h.value)return`${e}[${W}]=${q};`;if(h.storage==="vec2<u32>"&&h.value==="i32")return`${e}[${W}]=vec2<u32>(u32(${q}), select(0u, 0xFFFFFFFFu, ${q} < 0));`;if(h.storage==="vec2<u32>"&&h.value==="u32")return`${e}[${W}]=vec2<u32>(u32(${q}), 0u);`;if(h.storage==="u32"&&h.value==="vec4<bool>")return`${e}[${W}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${q}));`;throw new Error(`not supported combination of storage type ${h.storage} and value type ${h.value} yet`)})(),Q=W=>(()=>{if(h.storage===h.value)return`${e}[${W}]`;if(h.storage==="vec2<u32>"&&h.value==="i32")return`i32(${e}[${W}].x)`;if(h.storage==="vec2<u32>"&&h.value==="u32")return`u32(${e}[${W}].x)`;if(h.storage==="u32"&&h.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${W}] & 0xFFu), bool(${e}[${W}] & 0xFF00u), bool(${e}[${W}] & 0xFF0000u), bool(${e}[${W}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${h.storage} and value type ${h.value} yet`)})(),se=a<2?"":`
+  }`,P=W=>(y.indicesToOffset=!0,a<2?W:`i2o_${e}(${W})`),D=(...W)=>a===0?"0u":`${h.indices}(${W.map(_).join(",")})`,R=(W,Y)=>a<2?`${W}`:`${q(W,Y,a)}`,H=(W,Y,he)=>a<2?`${W}=${he};`:`${q(W,Y,a)}=${he};`,L={},re=(W,Y)=>{y.broadcastedIndicesToOffset=!0;let he=`${Y.name}broadcastedIndicesTo${e}Offset`;if(he in L)return`${he}(${W})`;let De=[];for(let pe=a-1;pe>=0;pe--){let Ie=Y.indicesGet("outputIndices",pe+Y.rank-a);De.push(`${R($,pe)} * (${Ie} % ${R(x,pe)})`)}return L[he]=`fn ${he}(outputIndices: ${Y.type.indices}) -> u32 {
+             return ${De.length>0?De.join("+"):"0u"};
+           }`,`${he}(${W})`},V=(W,Y)=>(()=>{if(h.storage===h.value)return`${e}[${W}]=${Y};`;if(h.storage==="vec2<u32>"&&h.value==="i32")return`${e}[${W}]=vec2<u32>(u32(${Y}), select(0u, 0xFFFFFFFFu, ${Y} < 0));`;if(h.storage==="vec2<u32>"&&h.value==="u32")return`${e}[${W}]=vec2<u32>(u32(${Y}), 0u);`;if(h.storage==="u32"&&h.value==="vec4<bool>")return`${e}[${W}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${Y}));`;throw new Error(`not supported combination of storage type ${h.storage} and value type ${h.value} yet`)})(),K=W=>(()=>{if(h.storage===h.value)return`${e}[${W}]`;if(h.storage==="vec2<u32>"&&h.value==="i32")return`i32(${e}[${W}].x)`;if(h.storage==="vec2<u32>"&&h.value==="u32")return`u32(${e}[${W}].x)`;if(h.storage==="u32"&&h.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${W}] & 0xFFu), bool(${e}[${W}] & 0xFF00u), bool(${e}[${W}] & 0xFF0000u), bool(${e}[${W}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${h.storage} and value type ${h.value} yet`)})(),we=a<2?"":`
   fn get_${e}ByIndices(indices: ${h.indices}) -> ${m} {
-    return ${Q(`i2o_${e}(indices)`)};
-  }`,Y=a<2?"":(()=>{let W=d.map(he=>`d${he}: u32`).join(", "),q=d.map(he=>`d${he}`).join(", ");return`
+    return ${K(`i2o_${e}(indices)`)};
+  }`,j=a<2?"":(()=>{let W=d.map(he=>`d${he}: u32`).join(", "),Y=d.map(he=>`d${he}`).join(", ");return`
   fn get_${e}(${W}) -> ${m} {
-    return get_${e}ByIndices(${D(q)});
-  }`})(),ee=(...W)=>{if(W.length!==a)throw new Error(`indices length must be ${a}`);let q=W.map(_).join(",");return a===0?Q("0u"):a===1?Q(q[0]):(y.get=!0,y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}(${q})`)},J=W=>a<2?Q(W):(y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}ByIndices(${W})`),ne=a<2?"":`
+    return get_${e}ByIndices(${D(Y)});
+  }`})(),Q=(...W)=>{if(W.length!==a)throw new Error(`indices length must be ${a}`);let Y=W.map(_).join(",");return a===0?K("0u"):a===1?K(Y[0]):(y.get=!0,y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}(${Y})`)},ie=W=>a<2?K(W):(y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}ByIndices(${W})`),te=a<2?"":`
   fn set_${e}ByIndices(indices: ${h.indices}, value: ${m}) {
     ${V(`i2o_${e}(indices)`,"value")}
-  }`,be=a<2?"":(()=>{let W=d.map(he=>`d${he}: u32`).join(", "),q=d.map(he=>`d${he}`).join(", ");return`
+  }`,be=a<2?"":(()=>{let W=d.map(he=>`d${he}: u32`).join(", "),Y=d.map(he=>`d${he}`).join(", ");return`
   fn set_${e}(${W}, value: ${m}) {
-    set_${e}ByIndices(${D(q)}, value);
-  }`})();return{impl:()=>{let W=[],q=!1;return y.offsetToIndices&&(W.push(S),q=!0),y.indicesToOffset&&(W.push(k),q=!0),y.broadcastedIndicesToOffset&&(Object.values(K).forEach(he=>W.push(he)),q=!0),y.set&&(W.push(be),q=!0),y.setByIndices&&(W.push(ne),q=!0),y.get&&(W.push(Y),q=!0),y.getByIndices&&(W.push(se),q=!0),!i&&q&&W.unshift(`const ${x} = ${h.indices}(${r.join(",")});`,`const ${$} = ${h.indices}(${C.computeStrides(r).join(",")});`),W.join(`
-`)},type:h,offsetToIndices:T,indicesToOffset:P,broadcastedIndicesToOffset:j,indices:D,indicesGet:R,indicesSet:G,set:(...W)=>{if(W.length!==a+1)throw new Error(`indices length must be ${a}`);let q=W[a];if(typeof q!="string")throw new Error("value must be string");let he=W.slice(0,a).map(_).join(",");return a===0?V("0u",q):a===1?V(he[0],q):(y.set=!0,y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}(${he}, ${q})`)},setByOffset:V,setByIndices:(W,q)=>a<2?V(W,q):(y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}ByIndices(${W}, ${q});`),get:ee,getByOffset:Q,getByIndices:J,usage:n,name:e,strides:$,shape:x,rank:a}},E=(e,t,r,n=1)=>Wr(e,t,r,"input",n),M=(e,t,r,n=1)=>Wr(e,t,r,"output",n),rs=(e,t,r)=>Wr(e,t,r,"atomicOutput",1),Lr=(e,t,r,n=1)=>Wr(e,t,r,"internal",n),ao=class{constructor(t,r){this.normalizedDispatchGroup=t;this.limits=r;this.internalVariables=[];this.variables=[];this.uniforms=[];this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=kt){let r=typeof t=="number"?t:t[0],n=typeof t=="number"?1:t[1],o=typeof t=="number"?1:t[2];if(r>this.limits.maxComputeWorkgroupSizeX||n>this.limits.maxComputeWorkgroupSizeY||o>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${r}, ${n}, ${o}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(r*n*o>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${r}, ${n}, ${o}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let i=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,a=i?`@builtin(global_invocation_id) global_id : vec3<u32>,
+    set_${e}ByIndices(${D(Y)}, value);
+  }`})();return{impl:()=>{let W=[],Y=!1;return y.offsetToIndices&&(W.push(S),Y=!0),y.indicesToOffset&&(W.push(C),Y=!0),y.broadcastedIndicesToOffset&&(Object.values(L).forEach(he=>W.push(he)),Y=!0),y.set&&(W.push(be),Y=!0),y.setByIndices&&(W.push(te),Y=!0),y.get&&(W.push(j),Y=!0),y.getByIndices&&(W.push(we),Y=!0),!i&&Y&&W.unshift(`const ${x} = ${h.indices}(${r.join(",")});`,`const ${$} = ${h.indices}(${k.computeStrides(r).join(",")});`),W.join(`
+`)},type:h,offsetToIndices:T,indicesToOffset:P,broadcastedIndicesToOffset:re,indices:D,indicesGet:R,indicesSet:H,set:(...W)=>{if(W.length!==a+1)throw new Error(`indices length must be ${a}`);let Y=W[a];if(typeof Y!="string")throw new Error("value must be string");let he=W.slice(0,a).map(_).join(",");return a===0?V("0u",Y):a===1?V(he[0],Y):(y.set=!0,y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}(${he}, ${Y})`)},setByOffset:V,setByIndices:(W,Y)=>a<2?V(W,Y):(y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}ByIndices(${W}, ${Y});`),get:Q,getByOffset:K,getByIndices:ie,usage:n,name:e,strides:$,shape:x,rank:a}},E=(e,t,r,n=1)=>Vr(e,t,r,"input",n),M=(e,t,r,n=1)=>Vr(e,t,r,"output",n),ts=(e,t,r)=>Vr(e,t,r,"atomicOutput",1),Wr=(e,t,r,n=1)=>Vr(e,t,r,"internal",n),io=class{constructor(t,r){this.normalizedDispatchGroup=t;this.limits=r;this.internalVariables=[];this.variables=[];this.uniforms=[];this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=Et){let r=typeof t=="number"?t:t[0],n=typeof t=="number"?1:t[1],o=typeof t=="number"?1:t[2];if(r>this.limits.maxComputeWorkgroupSizeX||n>this.limits.maxComputeWorkgroupSizeY||o>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${r}, ${n}, ${o}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(r*n*o>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${r}, ${n}, ${o}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let i=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,a=i?`@builtin(global_invocation_id) global_id : vec3<u32>,
     @builtin(workgroup_id) workgroup_id : vec3<u32>,
     @builtin(local_invocation_index) local_idx : u32,
     @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
@@ -3969,13 +4024,13 @@ var Un=Object.defineProperty;var Vp=Object.getOwnPropertyDescriptor;var Wp=Objec
       struct Uniforms { ${t.join(", ")} };
       @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(`
 `)+this.internalVariables.map(t=>t.impl()).join(`
-`)}get variablesInfo(){if(this.uniforms.length===0)return;let t=r=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(r)];return this.uniforms.map(r=>[t(r.type),r.length??1])}},ns=(e,t)=>new ao(e,t)});var dm,os,lm,cm,pm,mm,Pe,is,as,dt=U(()=>{"use strict";te();oe();Se();ae();dm=(e,t)=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.");if(t.length!==0&&t.length!==e[0].dims.length)throw new Error(`perm size ${t.length} does not match input rank ${e[0].dims.length}`)},os=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),lm=(e,t)=>C.sortBasedOnPerm(e,os(e.length,t)),cm=(e,t,r,n)=>{let o=`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
-    var a: ${r.type.indices};`;for(let i=0;i<t;++i)o+=`a[${e[i]}]=i[${i}];`;return o+="return a;}"},pm=(e,t)=>{let r=[],n=[];for(let o=0;o<e.length;++o)e[o]!==1&&r.push(e[o]),e[t[o]]!==1&&n.push(t[o]);return{newShape:r,newPerm:n}},mm=(e,t)=>{let r=0;for(let n=0;n<e.length;++n)if(t[e[n]]!==1){if(e[n]<r)return!1;r=e[n]}return!0},Pe=(e,t)=>{let r=e.dataType,n=e.dims.length,o=os(n,t),i=lm(e.dims,o),a=e.dims,d=i,l=n<2||mm(o,e.dims),p;if(l)return p=g=>{let x=E("input",r,a,4),$=M("output",r,d,4);return`
+`)}get variablesInfo(){if(this.uniforms.length===0)return;let t=r=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(r)];return this.uniforms.map(r=>[t(r.type),r.length??1])}},rs=(e,t)=>new io(e,t)});var dm,ns,lm,cm,pm,mm,Pe,os,is,dt=U(()=>{"use strict";J();ne();xe();ae();dm=(e,t)=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.");if(t.length!==0&&t.length!==e[0].dims.length)throw new Error(`perm size ${t.length} does not match input rank ${e[0].dims.length}`)},ns=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),lm=(e,t)=>k.sortBasedOnPerm(e,ns(e.length,t)),cm=(e,t,r,n)=>{let o=`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
+    var a: ${r.type.indices};`;for(let i=0;i<t;++i)o+=`a[${e[i]}]=i[${i}];`;return o+="return a;}"},pm=(e,t)=>{let r=[],n=[];for(let o=0;o<e.length;++o)e[o]!==1&&r.push(e[o]),e[t[o]]!==1&&n.push(t[o]);return{newShape:r,newPerm:n}},mm=(e,t)=>{let r=0;for(let n=0;n<e.length;++n)if(t[e[n]]!==1){if(e[n]<r)return!1;r=e[n]}return!0},Pe=(e,t)=>{let r=e.dataType,n=e.dims.length,o=ns(n,t),i=lm(e.dims,o),a=e.dims,d=i,l=n<2||mm(o,e.dims),p;if(l)return p=g=>{let x=E("input",r,a,4),$=M("output",r,d,4);return`
   ${g.registerUniform("output_size","u32").declareVariables(x,$)}
   ${g.mainStart()}
     ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
     output[global_idx] = input[global_idx];
-  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let g=C.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(g/64/4)},programUniforms:[{type:12,data:Math.ceil(g/4)}]}},getShaderSource:p};let{newShape:m,newPerm:u}=pm(e.dims,o),h=C.areEqual(u,[2,3,1]),_=C.areEqual(u,[3,1,2]);if(m.length===2||h||_){a=h?[m[0],m[1]*m[2]]:_?[m[0]*m[1],m[2]]:m,d=[a[1],a[0]];let g=16;return p=x=>{let $=E("a",r,a.length),v=M("output",r,d.length);return`
+  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let g=k.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(g/64/4)},programUniforms:[{type:12,data:Math.ceil(g/4)}]}},getShaderSource:p};let{newShape:m,newPerm:u}=pm(e.dims,o),h=k.areEqual(u,[2,3,1]),_=k.areEqual(u,[3,1,2]);if(m.length===2||h||_){a=h?[m[0],m[1]*m[2]]:_?[m[0]*m[1],m[2]]:m,d=[a[1],a[0]];let g=16;return p=x=>{let $=E("a",r,a.length),v=M("output",r,d.length);return`
   ${x.registerUniform("output_size","u32").declareVariables($,v)}
   var<workgroup> tile : array<array<${v.type.value}, ${g+1}>, ${g}>;
   ${x.mainStart([g,g,1])}
@@ -3994,7 +4049,7 @@ var Un=Object.defineProperty;var Vp=Object.getOwnPropertyDescriptor;var Wp=Objec
     if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
       ${v.setByIndices(`${v.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
     }
-  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let x=C.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(d[1]/g),y:Math.ceil(d[0]/g)},programUniforms:[{type:12,data:x},...N(a,d)]}},getShaderSource:p}}return p=g=>{let x=E("a",r,a.length),$=M("output",r,d.length);return`
+  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let x=k.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(d[1]/g),y:Math.ceil(d[0]/g)},programUniforms:[{type:12,data:x},...N(a,d)]}},getShaderSource:p}}return p=g=>{let x=E("a",r,a.length),$=M("output",r,d.length);return`
   ${g.registerUniform("output_size","u32").declareVariables(x,$)}
 
   ${cm(o,n,x,$)}
@@ -4006,7 +4061,7 @@ var Un=Object.defineProperty;var Vp=Object.getOwnPropertyDescriptor;var Wp=Objec
     let aIndices = perm(indices);
 
     ${$.setByOffset("global_idx",x.getByIndices("aIndices"))}
-  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let g=C.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:[{type:12,data:g},...N(a,d)]}},getShaderSource:p}},is=(e,t)=>{dm(e.inputs,t.perm),e.compute(Pe(e.inputs[0],t.perm))},as=e=>re({perm:e.perm})});var fm,hm,gm,bm,ym,_m,wm,vm,$m,xm,rt,ss,us,ds,ls,cs,ps,ms,fs,hs,gs,bs=U(()=>{"use strict";te();oe();ae();Gr();dt();fm={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate * candidate",logSumExp:"bestValue + exp(candidate)",l1:"bestValue + abs(candidate)",l2:"bestValue + candidate * candidate",logSum:"bestValue + candidate"},hm={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate",logSumExp:"bestValue + candidate",l1:"bestValue + candidate",l2:"bestValue + candidate",logSum:"bestValue + candidate"},gm={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},bm={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},ym=(e,t)=>{let r=[];for(let n=t-e;n<t;++n)r.push(n);return r},_m=(e,t)=>{let r=[],n=e.length;for(let i=0;i<n;i++)t.indexOf(i)===-1&&r.push(e[i]);let o=t.map(i=>e[i]);return[r,o]},wm=(e,t)=>{let r=e.length+t.length,n=[],o=0;for(let i=0;i<r;i++)t.indexOf(i)===-1?n.push(e[o++]):n.push(1);return n},vm=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},$m=(e,t)=>{let r=[];if(!vm(e,t)){for(let n=0;n<t;++n)e.indexOf(n)===-1&&r.push(n);e.forEach(n=>r.push(n))}return r},xm=(e,t,r,n,o,i,a)=>{let d=r[0].dims,l=C.size(i),p=C.size(a),m=E("_A",r[0].dataType,d),u=M("output",o,i),h=64;l===1&&(h=256);let _=`
+  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let g=k.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:[{type:12,data:g},...N(a,d)]}},getShaderSource:p}},os=(e,t)=>{dm(e.inputs,t.perm),e.compute(Pe(e.inputs[0],t.perm))},is=e=>ee({perm:e.perm})});var fm,hm,gm,bm,ym,_m,wm,vm,$m,xm,rt,as,ss,us,ds,ls,cs,ps,ms,fs,hs,gs=U(()=>{"use strict";J();ne();ae();Lr();dt();fm={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate * candidate",logSumExp:"bestValue + exp(candidate)",l1:"bestValue + abs(candidate)",l2:"bestValue + candidate * candidate",logSum:"bestValue + candidate"},hm={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate",logSumExp:"bestValue + candidate",l1:"bestValue + candidate",l2:"bestValue + candidate",logSum:"bestValue + candidate"},gm={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},bm={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},ym=(e,t)=>{let r=[];for(let n=t-e;n<t;++n)r.push(n);return r},_m=(e,t)=>{let r=[],n=e.length;for(let i=0;i<n;i++)t.indexOf(i)===-1&&r.push(e[i]);let o=t.map(i=>e[i]);return[r,o]},wm=(e,t)=>{let r=e.length+t.length,n=[],o=0;for(let i=0;i<r;i++)t.indexOf(i)===-1?n.push(e[o++]):n.push(1);return n},vm=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},$m=(e,t)=>{let r=[];if(!vm(e,t)){for(let n=0;n<t;++n)e.indexOf(n)===-1&&r.push(n);e.forEach(n=>r.push(n))}return r},xm=(e,t,r,n,o,i,a)=>{let d=r[0].dims,l=k.size(i),p=k.size(a),m=E("_A",r[0].dataType,d),u=M("output",o,i),h=64;l===1&&(h=256);let _=`
           var<workgroup> aBestValues : array<f32, ${h}>;
        `,y=g=>`
         ${g.registerUniform("reduceSize","u32").declareVariables(m,u)}
@@ -4044,11 +4099,11 @@ var Un=Object.defineProperty;var Vp=Object.getOwnPropertyDescriptor;var Wp=Objec
          if (local_idx == 0u) {
           ${u.setByOffset("outputIndex",`${n==="mean"?`${u.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${u.type.storage}(${bm[n]})`}`)};
          }
-        }`;return{name:e,shaderCache:{hint:`${t};${h}`,inputDependencies:["type"]},getShaderSource:y,getRunData:()=>({outputs:[{dims:i,dataType:o}],dispatchGroup:{x:l},programUniforms:[{type:12,data:p}]})}},rt=(e,t,r,n)=>{let o=e.inputs.length===1?r:lo(e.inputs,r),i=o.axes;i.length===0&&!o.noopWithEmptyAxes&&(i=e.inputs[0].dims.map((_,y)=>y));let a=C.normalizeAxes(i,e.inputs[0].dims.length),d=a,l=e.inputs[0],p=$m(d,e.inputs[0].dims.length);p.length>0&&(l=e.compute(Pe(e.inputs[0],p),{inputs:[0],outputs:[-1]})[0],d=ym(d.length,l.dims.length));let[m,u]=_m(l.dims,d),h=m;o.keepDims&&(h=wm(m,a)),e.compute(xm(t,o.cacheKey,[l],n,e.inputs[0].dataType,h,u),{inputs:[l]})},ss=(e,t)=>{rt(e,"ReduceMeanShared",t,"mean")},us=(e,t)=>{rt(e,"ReduceL1Shared",t,"l1")},ds=(e,t)=>{rt(e,"ReduceL2Shared",t,"l2")},ls=(e,t)=>{rt(e,"ReduceLogSumExpShared",t,"logSumExp")},cs=(e,t)=>{rt(e,"ReduceMaxShared",t,"max")},ps=(e,t)=>{rt(e,"ReduceMinShared",t,"min")},ms=(e,t)=>{rt(e,"ReduceProdShared",t,"prod")},fs=(e,t)=>{rt(e,"ReduceSumShared",t,"sum")},hs=(e,t)=>{rt(e,"ReduceSumSquareShared",t,"sumSquare")},gs=(e,t)=>{rt(e,"ReduceLogSumShared",t,"logSum")}});var nt,Sm,Hr,lo,ot,Tm,Im,Cm,Am,km,Em,Pm,zm,Om,Dm,it,ys,_s,ws,vs,$s,xs,Ss,Ts,Is,Cs,Gr=U(()=>{"use strict";te();oe();Se();ae();bs();nt=e=>{if(!e||e.length===0||e.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(e.length===2&&e[1].dims.length!==1)throw new Error("Invalid axes input dims.")},Sm=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Hr=(e,t,r,n,o,i,a=!1,d=!1)=>{let l=[],p=r[0].dims,m=p.length,u=C.normalizeAxes(o,m),h=!d&&u.length===0;p.forEach((x,$)=>{h||u.indexOf($)>=0?a&&l.push(1):l.push(x)});let _=l.length,y=C.size(l);return{name:e,shaderCache:t,getShaderSource:x=>{let $=[],v=E("_A",r[0].dataType,m),S=M("output",i,_),T=n(v,S,u),A=T[2];for(let k=0,P=0;k<m;k++)h||u.indexOf(k)>=0?(a&&P++,A=`for(var j${k}: u32 = 0; j${k} < ${p[k]}; j${k}++) {
-                  ${T[2].includes("last_index")?`let last_index = j${k};`:""}
-                  ${v.indicesSet("input_indices",k,`j${k}`)}
+        }`;return{name:e,shaderCache:{hint:`${t};${h}`,inputDependencies:["type"]},getShaderSource:y,getRunData:()=>({outputs:[{dims:i,dataType:o}],dispatchGroup:{x:l},programUniforms:[{type:12,data:p}]})}},rt=(e,t,r,n)=>{let o=e.inputs.length===1?r:uo(e.inputs,r),i=o.axes;i.length===0&&!o.noopWithEmptyAxes&&(i=e.inputs[0].dims.map((_,y)=>y));let a=k.normalizeAxes(i,e.inputs[0].dims.length),d=a,l=e.inputs[0],p=$m(d,e.inputs[0].dims.length);p.length>0&&(l=e.compute(Pe(e.inputs[0],p),{inputs:[0],outputs:[-1]})[0],d=ym(d.length,l.dims.length));let[m,u]=_m(l.dims,d),h=m;o.keepDims&&(h=wm(m,a)),e.compute(xm(t,o.cacheKey,[l],n,e.inputs[0].dataType,h,u),{inputs:[l]})},as=(e,t)=>{rt(e,"ReduceMeanShared",t,"mean")},ss=(e,t)=>{rt(e,"ReduceL1Shared",t,"l1")},us=(e,t)=>{rt(e,"ReduceL2Shared",t,"l2")},ds=(e,t)=>{rt(e,"ReduceLogSumExpShared",t,"logSumExp")},ls=(e,t)=>{rt(e,"ReduceMaxShared",t,"max")},cs=(e,t)=>{rt(e,"ReduceMinShared",t,"min")},ps=(e,t)=>{rt(e,"ReduceProdShared",t,"prod")},ms=(e,t)=>{rt(e,"ReduceSumShared",t,"sum")},fs=(e,t)=>{rt(e,"ReduceSumSquareShared",t,"sumSquare")},hs=(e,t)=>{rt(e,"ReduceLogSumShared",t,"logSum")}});var nt,Sm,Gr,uo,ot,Tm,Im,Cm,Am,km,Em,Pm,zm,Om,Dm,it,bs,ys,_s,ws,vs,$s,xs,Ss,Ts,Is,Lr=U(()=>{"use strict";J();ne();xe();ae();gs();nt=e=>{if(!e||e.length===0||e.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(e.length===2&&e[1].dims.length!==1)throw new Error("Invalid axes input dims.")},Sm=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Gr=(e,t,r,n,o,i,a=!1,d=!1)=>{let l=[],p=r[0].dims,m=p.length,u=k.normalizeAxes(o,m),h=!d&&u.length===0;p.forEach((x,$)=>{h||u.indexOf($)>=0?a&&l.push(1):l.push(x)});let _=l.length,y=k.size(l);return{name:e,shaderCache:t,getShaderSource:x=>{let $=[],v=E("_A",r[0].dataType,m),S=M("output",i,_),T=n(v,S,u),A=T[2];for(let C=0,P=0;C<m;C++)h||u.indexOf(C)>=0?(a&&P++,A=`for(var j${C}: u32 = 0; j${C} < ${p[C]}; j${C}++) {
+                  ${T[2].includes("last_index")?`let last_index = j${C};`:""}
+                  ${v.indicesSet("input_indices",C,`j${C}`)}
                   ${A}
-                }`):($.push(`${v.indicesSet("input_indices",k,S.indicesGet("output_indices",P))};`),P++);return`
+                }`):($.push(`${v.indicesSet("input_indices",C,S.indicesGet("output_indices",P))};`),P++);return`
 
         ${x.registerUniform("output_size","u32").declareVariables(v,S)}
 
@@ -4065,19 +4120,19 @@ var Un=Object.defineProperty;var Vp=Object.getOwnPropertyDescriptor;var Wp=Objec
           ${T[3]}
           ${T.length===4?S.setByOffset("global_idx","value"):T.slice(4).join(`
 `)}
-        }`},getRunData:()=>({outputs:[{dims:l,dataType:i}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},...N(p,l)]})}},lo=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),re({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},ot=(e,t,r,n)=>{let o=e.inputs,i=o.length===1?r:lo(o,r);e.compute(Hr(t,{hint:i.cacheKey,inputDependencies:["rank"]},[o[0]],i.noopWithEmptyAxes&&i.axes.length===0?Sm:n,i.axes,o[0].dataType,i.keepDims,i.noopWithEmptyAxes),{inputs:[0]})},Tm=(e,t)=>{nt(e.inputs),ot(e,"ReduceLogSum",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,"value = log(value);"])},Im=(e,t)=>{nt(e.inputs),ot(e,"ReduceL1",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += abs(${n.getByIndices("input_indices")});`,""])},Cm=(e,t)=>{nt(e.inputs),ot(e,"ReduceL2",t,(n,o)=>[`var t = ${o.type.value}(0); var value = ${o.type.value}(0);`,"",`t = ${n.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Am=(e,t)=>{nt(e.inputs),ot(e,"ReduceLogSumExp",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += exp(${n.getByIndices("input_indices")});`,"value = log(value);"])},km=(e,t)=>{nt(e.inputs),ot(e,"ReduceMax",t,(n,o,i)=>{let a=[];for(let d=0;d<n.rank;d++)(i.indexOf(d)>=0||i.length===0)&&a.push(n.indicesSet("input_indices",d,0));return[`${a.join(`
+        }`},getRunData:()=>({outputs:[{dims:l,dataType:i}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},...N(p,l)]})}},uo=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),ee({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},ot=(e,t,r,n)=>{let o=e.inputs,i=o.length===1?r:uo(o,r);e.compute(Gr(t,{hint:i.cacheKey,inputDependencies:["rank"]},[o[0]],i.noopWithEmptyAxes&&i.axes.length===0?Sm:n,i.axes,o[0].dataType,i.keepDims,i.noopWithEmptyAxes),{inputs:[0]})},Tm=(e,t)=>{nt(e.inputs),ot(e,"ReduceLogSum",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,"value = log(value);"])},Im=(e,t)=>{nt(e.inputs),ot(e,"ReduceL1",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += abs(${n.getByIndices("input_indices")});`,""])},Cm=(e,t)=>{nt(e.inputs),ot(e,"ReduceL2",t,(n,o)=>[`var t = ${o.type.value}(0); var value = ${o.type.value}(0);`,"",`t = ${n.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Am=(e,t)=>{nt(e.inputs),ot(e,"ReduceLogSumExp",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += exp(${n.getByIndices("input_indices")});`,"value = log(value);"])},km=(e,t)=>{nt(e.inputs),ot(e,"ReduceMax",t,(n,o,i)=>{let a=[];for(let d=0;d<n.rank;d++)(i.indexOf(d)>=0||i.length===0)&&a.push(n.indicesSet("input_indices",d,0));return[`${a.join(`
 `)}`,`var value = ${n.getByIndices("input_indices")};`,`value = max(value, ${n.getByIndices("input_indices")});`,""]})},Em=(e,t)=>{nt(e.inputs),ot(e,"ReduceMean",t,(n,o,i)=>{let a=1;for(let d=0;d<n.rank;d++)(i.indexOf(d)>=0||i.length===0)&&(a*=e.inputs[0].dims[d]);return["var sum = f32(0);","",`sum += f32(${n.getByIndices("input_indices")});`,`let value = ${o.type.value}(sum / ${a});`]})},Pm=(e,t)=>{nt(e.inputs),ot(e,"ReduceMin",t,(n,o,i)=>{let a=[];for(let d=0;d<n.rank;d++)(i.indexOf(d)>=0||i.length===0)&&a.push(`input_indices[${d}] = 0;`);return[`${a.join(`
-`)}`,`var value = ${n.getByIndices("input_indices")};`,`value = min(value, ${n.getByIndices("input_indices")});`,""]})},zm=(e,t)=>{nt(e.inputs),ot(e,"ReduceProd",t,(n,o)=>[`var value = ${o.type.storage}(1);`,"",`value *= ${n.getByIndices("input_indices")};`,""])},Om=(e,t)=>{nt(e.inputs),ot(e,"ReduceSum",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,""])},Dm=(e,t)=>{nt(e.inputs),ot(e,"ReduceSumSquare",t,(n,o)=>[`var t = ${o.type.value}(0); var value = ${o.type.value}(0);`,"",`t = ${n.getByIndices("input_indices")}; value += t * t;`,""])},it=(e,t,r)=>{if(t.length===0)return r;let n=1,o=1;for(let i=0;i<t.length;i++)t.indexOf(i)===-1?n*=e[i]:o*=e[i];return o<32&&n>1024},ys=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Em(e,t):ss(e,t)},_s=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Im(e,t):us(e,t)},ws=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Cm(e,t):ds(e,t)},vs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Am(e,t):ls(e,t)},$s=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?km(e,t):cs(e,t)},xs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Pm(e,t):ps(e,t)},Ss=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?zm(e,t):ms(e,t)},Ts=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Om(e,t):fs(e,t)},Is=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Dm(e,t):hs(e,t)},Cs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Tm(e,t):gs(e,t)}});var As,ks,Es,co,Ps=U(()=>{"use strict";te();Se();Gr();As=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},ks=(e,t)=>{As(e.inputs);let r=(n,o,i)=>{let a=[];for(let d=0;d<n.rank;d++)(i.indexOf(d)>=0||i.length===0)&&a.push(`input_indices[${d}] = 0;`);return[`${a.join(`
+`)}`,`var value = ${n.getByIndices("input_indices")};`,`value = min(value, ${n.getByIndices("input_indices")});`,""]})},zm=(e,t)=>{nt(e.inputs),ot(e,"ReduceProd",t,(n,o)=>[`var value = ${o.type.storage}(1);`,"",`value *= ${n.getByIndices("input_indices")};`,""])},Om=(e,t)=>{nt(e.inputs),ot(e,"ReduceSum",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,""])},Dm=(e,t)=>{nt(e.inputs),ot(e,"ReduceSumSquare",t,(n,o)=>[`var t = ${o.type.value}(0); var value = ${o.type.value}(0);`,"",`t = ${n.getByIndices("input_indices")}; value += t * t;`,""])},it=(e,t,r)=>{if(t.length===0)return r;let n=1,o=1;for(let i=0;i<t.length;i++)t.indexOf(i)===-1?n*=e[i]:o*=e[i];return o<32&&n>1024},bs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Em(e,t):as(e,t)},ys=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Im(e,t):ss(e,t)},_s=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Cm(e,t):us(e,t)},ws=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Am(e,t):ds(e,t)},vs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?km(e,t):ls(e,t)},$s=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Pm(e,t):cs(e,t)},xs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?zm(e,t):ps(e,t)},Ss=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Om(e,t):ms(e,t)},Ts=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Dm(e,t):fs(e,t)},Is=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Tm(e,t):hs(e,t)}});var Cs,As,ks,lo,Es=U(()=>{"use strict";J();xe();Lr();Cs=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},As=(e,t)=>{Cs(e.inputs);let r=(n,o,i)=>{let a=[];for(let d=0;d<n.rank;d++)(i.indexOf(d)>=0||i.length===0)&&a.push(`input_indices[${d}] = 0;`);return[`${a.join(`
 `)}`,`var value = ${n.getByIndices("input_indices")};
 var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {
          value = ${n.getByIndices("input_indices")};
          best_index = i32(last_index);
-       }`,"",o.setByOffset("global_idx","best_index")]};e.compute(Hr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},Es=(e,t)=>{As(e.inputs);let r=(n,o,i)=>{let a=[];for(let d=0;d<n.rank;d++)(i.indexOf(d)>=0||i.length===0)&&a.push(`input_indices[${d}] = 0;`);return[`${a.join(`
+       }`,"",o.setByOffset("global_idx","best_index")]};e.compute(Gr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ks=(e,t)=>{Cs(e.inputs);let r=(n,o,i)=>{let a=[];for(let d=0;d<n.rank;d++)(i.indexOf(d)>=0||i.length===0)&&a.push(`input_indices[${d}] = 0;`);return[`${a.join(`
 `)}`,`var value = ${n.getByIndices("input_indices")};
 var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {
          value = ${n.getByIndices("input_indices")};
          best_index = i32(last_index);
-       }`,"",o.setByOffset("global_idx","best_index")]};e.compute(Hr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},co=e=>re(e)});var Bm,po,Mm,Rm,Um,Rt,Nm,zs,Fr=U(()=>{"use strict";te();oe();Nr();ae();Bm=(e,t)=>{let r=e[0],n=e[1],o=e[2],i=e[3],a=e[4],d=e[5];if(a&&d)throw new Error("Attention cannot have both past and attention_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let l=r.dims[0],p=r.dims[1],m=r.dims[2];if(o.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(n.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(n.dims[0]!==m)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(o.dims[0]!==n.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let u=o.dims[0]/3,h=u,_=h;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let S of t.qkvHiddenSizes)if(S%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");u=t.qkvHiddenSizes[0],h=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let y=p;if(u!==h)throw new Error("qkv_hidden_sizes first element should be same as the second");if(o.dims[0]!==u+h+_)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let g=0;if(a){if(h!==_)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(a.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(a.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(a.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(a.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(a.dims[4]!==h/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(g=a.dims[3])}let x=y+g,$=-1,v=0;if(i)throw new Error("Mask not supported");if(a)throw new Error("past is not supported");if(d){if(d.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(d.dims[0]!==l||d.dims[1]!==t.numHeads||d.dims[2]!==p||d.dims[3]!==x)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:l,sequenceLength:p,pastSequenceLength:g,kvSequenceLength:y,totalSequenceLength:x,maxSequenceLength:$,inputHiddenSize:m,hiddenSize:u,vHiddenSize:_,headSize:Math.floor(u/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:v,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},po=(e,t,r)=>t&&e?`
+       }`,"",o.setByOffset("global_idx","best_index")]};e.compute(Gr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},lo=e=>ee(e)});var Bm,co,Mm,Rm,Um,Ut,Nm,Ps,Hr=U(()=>{"use strict";J();ne();Ur();ae();Bm=(e,t)=>{let r=e[0],n=e[1],o=e[2],i=e[3],a=e[4],d=e[5];if(a&&d)throw new Error("Attention cannot have both past and attention_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let l=r.dims[0],p=r.dims[1],m=r.dims[2];if(o.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(n.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(n.dims[0]!==m)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(o.dims[0]!==n.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let u=o.dims[0]/3,h=u,_=h;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let S of t.qkvHiddenSizes)if(S%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");u=t.qkvHiddenSizes[0],h=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let y=p;if(u!==h)throw new Error("qkv_hidden_sizes first element should be same as the second");if(o.dims[0]!==u+h+_)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let g=0;if(a){if(h!==_)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(a.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(a.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(a.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(a.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(a.dims[4]!==h/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(g=a.dims[3])}let x=y+g,$=-1,v=0;if(i)throw new Error("Mask not supported");if(a)throw new Error("past is not supported");if(d){if(d.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(d.dims[0]!==l||d.dims[1]!==t.numHeads||d.dims[2]!==p||d.dims[3]!==x)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:l,sequenceLength:p,pastSequenceLength:g,kvSequenceLength:y,totalSequenceLength:x,maxSequenceLength:$,inputHiddenSize:m,hiddenSize:u,vHiddenSize:_,headSize:Math.floor(u/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:v,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},co=(e,t,r)=>t&&e?`
       let total_sequence_length_input = u32(${t.getByOffset("0")});
       let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
       let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
@@ -4090,7 +4145,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
        `:`
     ${r?"let past_sequence_length = uniforms.past_sequence_length":""};
     let present_sequence_length = total_sequence_length;
-    `,Mm=(e,t,r,n,o,i,a,d)=>{let l=me(a?1:i),p=64,m=i/l;m<p&&(p=32);let u=Math.ceil(i/l/p),h=[{type:12,data:t},{type:12,data:r},{type:12,data:n},{type:12,data:o},{type:12,data:m},{type:12,data:u}],_=_e(e.dataType,l),y=Ee(1,l),g=["type"];a&&g.push("type"),d&&g.push("type");let x=$=>{let v=M("x",e.dataType,e.dims,l),S=[v],T=a?E("seq_lens",a.dataType,a.dims):void 0;T&&S.push(T);let A=d?E("total_sequence_length_input",d.dataType,d.dims):void 0;A&&S.push(A);let k=Ee(e.dataType),P=[{name:"batch_size",type:"u32"},{name:"num_heads",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"sequence_length",type:"u32"},{name:"total_sequence_length",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
+    `,Mm=(e,t,r,n,o,i,a,d)=>{let l=me(a?1:i),p=64,m=i/l;m<p&&(p=32);let u=Math.ceil(i/l/p),h=[{type:12,data:t},{type:12,data:r},{type:12,data:n},{type:12,data:o},{type:12,data:m},{type:12,data:u}],_=ye(e.dataType,l),y=Ee(1,l),g=["type"];a&&g.push("type"),d&&g.push("type");let x=$=>{let v=M("x",e.dataType,e.dims,l),S=[v],T=a?E("seq_lens",a.dataType,a.dims):void 0;T&&S.push(T);let A=d?E("total_sequence_length_input",d.dataType,d.dims):void 0;A&&S.push(A);let C=Ee(e.dataType),P=[{name:"batch_size",type:"u32"},{name:"num_heads",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"sequence_length",type:"u32"},{name:"total_sequence_length",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
   var<workgroup> thread_max: array<f32, ${p}>;
   var<workgroup> thread_sum: array<f32, ${p}>;
   ${$.registerUniforms(P).declareVariables(...S)}
@@ -4099,7 +4154,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     let headIdx = workgroup_id.z % uniforms.num_heads;
     let sequence_length = uniforms.sequence_length;
     var total_sequence_length = uniforms.total_sequence_length;
-    ${po(T,A,!1)}
+    ${co(T,A,!1)}
     let local_offset = local_idx * uniforms.elements_per_thread;
     let offset = (global_idx / ${p}) * uniforms.total_sequence_length + local_offset;
     let seq_causal_length = ${a?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
@@ -4129,7 +4184,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
 
     if (sum == 0) {
       for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
-        x[offset + i] = ${v.type.value}(${k}(1.0) / ${k}(seq_causal_length));
+        x[offset + i] = ${v.type.value}(${C}(1.0) / ${C}(seq_causal_length));
       }
     } else {
       for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
@@ -4139,14 +4194,14 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     }
       ${a?`
         for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) {
-          x[offset + total_seq_id] = ${v.type.value}(${k}(0));
+          x[offset + total_seq_id] = ${v.type.value}(${C}(0));
         }`:""};
-  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${p};${_};${l}`,inputDependencies:g},getShaderSource:x,getRunData:()=>({outputs:[],dispatchGroup:{x:Math.ceil(i/p),y:o,z:t*r},programUniforms:h})}},Rm=(e,t,r,n,o,i,a,d,l)=>{let p=a+i.kvSequenceLength,m=[i.batchSize,i.numHeads,i.sequenceLength,p],u=e>1&&n,h=i.kvNumHeads?i.kvNumHeads:i.numHeads,_=u?[i.batchSize,h,p,i.headSize]:void 0,y=i.nReps?i.nReps:1,g=i.scale===0?1/Math.sqrt(i.headSize):i.scale,x=me(i.headSize),$=i.headSize/x,v=12,S={x:Math.ceil(p/v),y:Math.ceil(i.sequenceLength/v),z:i.batchSize*i.numHeads},T=[{type:12,data:i.sequenceLength},{type:12,data:$},{type:12,data:p},{type:12,data:i.numHeads},{type:12,data:i.headSize},{type:1,data:g},{type:12,data:a},{type:12,data:i.kvSequenceLength},{type:12,data:y}],A=u&&n&&C.size(n.dims)>0,k=["type","type"];A&&k.push("type"),o&&k.push("type"),d&&k.push("type"),l&&k.push("type");let P=[{dims:m,dataType:t.dataType,gpuDataType:0}];u&&P.push({dims:_,dataType:t.dataType,gpuDataType:0});let D=R=>{let G=E("q",t.dataType,t.dims,x),K=E("key",r.dataType,r.dims,x),j=[G,K];if(A){let ne=E("past_key",n.dataType,n.dims,x);j.push(ne)}o&&j.push(E("attention_bias",o.dataType,o.dims));let V=d?E("seq_lens",d.dataType,d.dims):void 0;V&&j.push(V);let Q=l?E("total_sequence_length_input",l.dataType,l.dims):void 0;Q&&j.push(Q);let se=M("output",t.dataType,m),Y=[se];u&&Y.push(M("present_key",t.dataType,_,x));let ee=Ee(1,x),J=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
+  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${p};${_};${l}`,inputDependencies:g},getShaderSource:x,getRunData:()=>({outputs:[],dispatchGroup:{x:Math.ceil(i/p),y:o,z:t*r},programUniforms:h})}},Rm=(e,t,r,n,o,i,a,d,l)=>{let p=a+i.kvSequenceLength,m=[i.batchSize,i.numHeads,i.sequenceLength,p],u=e>1&&n,h=i.kvNumHeads?i.kvNumHeads:i.numHeads,_=u?[i.batchSize,h,p,i.headSize]:void 0,y=i.nReps?i.nReps:1,g=i.scale===0?1/Math.sqrt(i.headSize):i.scale,x=me(i.headSize),$=i.headSize/x,v=12,S={x:Math.ceil(p/v),y:Math.ceil(i.sequenceLength/v),z:i.batchSize*i.numHeads},T=[{type:12,data:i.sequenceLength},{type:12,data:$},{type:12,data:p},{type:12,data:i.numHeads},{type:12,data:i.headSize},{type:1,data:g},{type:12,data:a},{type:12,data:i.kvSequenceLength},{type:12,data:y}],A=u&&n&&k.size(n.dims)>0,C=["type","type"];A&&C.push("type"),o&&C.push("type"),d&&C.push("type"),l&&C.push("type");let P=[{dims:m,dataType:t.dataType,gpuDataType:0}];u&&P.push({dims:_,dataType:t.dataType,gpuDataType:0});let D=R=>{let H=E("q",t.dataType,t.dims,x),L=E("key",r.dataType,r.dims,x),re=[H,L];if(A){let te=E("past_key",n.dataType,n.dims,x);re.push(te)}o&&re.push(E("attention_bias",o.dataType,o.dims));let V=d?E("seq_lens",d.dataType,d.dims):void 0;V&&re.push(V);let K=l?E("total_sequence_length_input",l.dataType,l.dims):void 0;K&&re.push(K);let we=M("output",t.dataType,m),j=[we];u&&j.push(M("present_key",t.dataType,_,x));let Q=Ee(1,x),ie=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
   const TILE_SIZE = ${v}u;
 
-  var<workgroup> tileQ: array<${G.type.storage}, ${v*v}>;
-  var<workgroup> tileK: array<${G.type.storage}, ${v*v}>;
-  ${R.registerUniforms(J).declareVariables(...j,...Y)}
+  var<workgroup> tileQ: array<${H.type.storage}, ${v*v}>;
+  var<workgroup> tileK: array<${H.type.storage}, ${v*v}>;
+  ${R.registerUniforms(ie).declareVariables(...re,...j)}
   ${R.mainStart([v,v,1])}
     // x holds the N and y holds the M
     let headIdx = workgroup_id.z % uniforms.num_heads;
@@ -4157,13 +4212,13 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     let n = workgroup_id.x * TILE_SIZE;
     let sequence_length = uniforms.M;
     var total_sequence_length = uniforms.N;
-    ${po(V,Q,!0)}
+    ${co(V,K,!0)}
     let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
     let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
     ${A&&u?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
     let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
     ${u?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
-    var value = ${ee}(0);
+    var value = ${Q}(0);
     for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
       if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {
         tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];
@@ -4186,7 +4241,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
       workgroupBarrier();
 
       for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
-          value += ${ee}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
+          value += ${Q}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
       }
 
       workgroupBarrier();
@@ -4196,14 +4251,14 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
       let headOffset = workgroup_id.z * uniforms.M * uniforms.N;
       let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
       var sum: f32 = ${(()=>{switch(x){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${x}`)}})()};
-        output[outputIdx] = ${se.type.value} (sum * uniforms.alpha) + ${o?"attention_bias[outputIdx]":"0.0"};
+        output[outputIdx] = ${we.type.value} (sum * uniforms.alpha) + ${o?"attention_bias[outputIdx]":"0.0"};
     }
-  }`};return{name:"AttentionProbs",shaderCache:{hint:`${x};${o!==void 0};${n!==void 0};${e}`,inputDependencies:k},getRunData:()=>({outputs:P,dispatchGroup:S,programUniforms:T}),getShaderSource:D}},Um=(e,t,r,n,o,i,a=void 0,d=void 0)=>{let l=i+o.kvSequenceLength,p=o.nReps?o.nReps:1,m=o.vHiddenSize*p,u=e>1&&n,h=o.kvNumHeads?o.kvNumHeads:o.numHeads,_=u?[o.batchSize,h,l,o.headSize]:void 0,y=[o.batchSize,o.sequenceLength,m],g=12,x={x:Math.ceil(o.vHeadSize/g),y:Math.ceil(o.sequenceLength/g),z:o.batchSize*o.numHeads},$=[{type:12,data:o.sequenceLength},{type:12,data:l},{type:12,data:o.vHeadSize},{type:12,data:o.numHeads},{type:12,data:o.headSize},{type:12,data:m},{type:12,data:i},{type:12,data:o.kvSequenceLength},{type:12,data:p}],v=u&&n&&C.size(n.dims)>0,S=["type","type"];v&&S.push("type"),a&&S.push("type"),d&&S.push("type");let T=[{dims:y,dataType:t.dataType,gpuDataType:0}];u&&T.push({dims:_,dataType:t.dataType,gpuDataType:0});let A=k=>{let P=E("probs",t.dataType,t.dims),D=E("v",r.dataType,r.dims),R=[P,D];v&&R.push(E("past_value",n.dataType,n.dims));let G=a?E("seq_lens",a.dataType,a.dims):void 0;a&&R.push(G);let K=d?E("total_sequence_length_input",d.dataType,d.dims):void 0;d&&R.push(K);let V=[M("output",t.dataType,y)];u&&V.push(M("present_value",t.dataType,_));let Q=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
+  }`};return{name:"AttentionProbs",shaderCache:{hint:`${x};${o!==void 0};${n!==void 0};${e}`,inputDependencies:C},getRunData:()=>({outputs:P,dispatchGroup:S,programUniforms:T}),getShaderSource:D}},Um=(e,t,r,n,o,i,a=void 0,d=void 0)=>{let l=i+o.kvSequenceLength,p=o.nReps?o.nReps:1,m=o.vHiddenSize*p,u=e>1&&n,h=o.kvNumHeads?o.kvNumHeads:o.numHeads,_=u?[o.batchSize,h,l,o.headSize]:void 0,y=[o.batchSize,o.sequenceLength,m],g=12,x={x:Math.ceil(o.vHeadSize/g),y:Math.ceil(o.sequenceLength/g),z:o.batchSize*o.numHeads},$=[{type:12,data:o.sequenceLength},{type:12,data:l},{type:12,data:o.vHeadSize},{type:12,data:o.numHeads},{type:12,data:o.headSize},{type:12,data:m},{type:12,data:i},{type:12,data:o.kvSequenceLength},{type:12,data:p}],v=u&&n&&k.size(n.dims)>0,S=["type","type"];v&&S.push("type"),a&&S.push("type"),d&&S.push("type");let T=[{dims:y,dataType:t.dataType,gpuDataType:0}];u&&T.push({dims:_,dataType:t.dataType,gpuDataType:0});let A=C=>{let P=E("probs",t.dataType,t.dims),D=E("v",r.dataType,r.dims),R=[P,D];v&&R.push(E("past_value",n.dataType,n.dims));let H=a?E("seq_lens",a.dataType,a.dims):void 0;a&&R.push(H);let L=d?E("total_sequence_length_input",d.dataType,d.dims):void 0;d&&R.push(L);let V=[M("output",t.dataType,y)];u&&V.push(M("present_value",t.dataType,_));let K=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
   const TILE_SIZE = ${g}u;
   var<workgroup> tileQ: array<${P.type.value}, ${g*g}>;
   var<workgroup> tileV: array<${P.type.value}, ${g*g}>;
-  ${k.registerUniforms(Q).declareVariables(...R,...V)}
-  ${k.mainStart([g,g,1])}
+  ${C.registerUniforms(K).declareVariables(...R,...V)}
+  ${C.mainStart([g,g,1])}
    let headIdx = workgroup_id.z % uniforms.num_heads;
    let batchIdx = workgroup_id.z / uniforms.num_heads;
    let kvHeadIdx = ${p===1?"headIdx":"headIdx / uniforms.n_reps"};
@@ -4212,7 +4267,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
    let n = global_id.x;
    let sequence_length = uniforms.M;
    var total_sequence_length = uniforms.K;
-   ${po(G,K,!0)}
+   ${co(H,L,!0)}
    let offsetA = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; // kvHeadIdx is relative to the batch
    ${v&&u?"let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;":""};
@@ -4253,7 +4308,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
        + headIdx * uniforms.N + n;
      output[outputIdx] = value;
    }
-  }`};return{name:"AttentionScore",shaderCache:{hint:`${n!==void 0};${e}`,inputDependencies:S},getRunData:()=>({outputs:T,dispatchGroup:x,programUniforms:$}),getShaderSource:A}},Rt=(e,t,r,n,o,i,a,d,l,p,m=void 0,u=void 0)=>{let h=Math.min(e.outputCount,1+(a?1:0)+(d?1:0)),_=h>1?p.pastSequenceLength:0,y=_+p.kvSequenceLength,g=l&&C.size(l.dims)>0?l:void 0,x=[t,r];h>1&&a&&C.size(a.dims)>0&&x.push(a),g&&x.push(g),m&&x.push(m),u&&x.push(u);let $=e.compute(Rm(h,t,r,a,g,p,_,m,u),{inputs:x,outputs:h>1?[-1,1]:[-1]})[0];e.compute(Mm($,p.batchSize,p.numHeads,_,p.sequenceLength,y,m,u),{inputs:m&&u?[$,m,u]:[$],outputs:[]});let v=[$,n];h>1&&d&&C.size(d.dims)>0&&v.push(d),m&&v.push(m),u&&v.push(u),e.compute(Um(h,$,n,d,p,_,m,u),{inputs:v,outputs:h>1?[0,2]:[0]})},Nm=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],n=t.sequenceLength,o=t.inputHiddenSize,i=t.headSize,a=12,d={x:Math.ceil(t.headSize/a),y:Math.ceil(t.sequenceLength/a),z:t.batchSize*t.numHeads},l=[e.inputs[0],e.inputs[1],e.inputs[2]],p=[{type:12,data:n},{type:12,data:o},{type:12,data:i},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],m=u=>{let h=M("output_q",l[0].dataType,r),_=M("output_k",l[0].dataType,r),y=M("output_v",l[0].dataType,r),g=E("input",l[0].dataType,l[0].dims),x=E("weight",l[1].dataType,l[1].dims),$=E("bias",l[2].dataType,l[2].dims),v=g.type.storage,S=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`
+  }`};return{name:"AttentionScore",shaderCache:{hint:`${n!==void 0};${e}`,inputDependencies:S},getRunData:()=>({outputs:T,dispatchGroup:x,programUniforms:$}),getShaderSource:A}},Ut=(e,t,r,n,o,i,a,d,l,p,m=void 0,u=void 0)=>{let h=Math.min(e.outputCount,1+(a?1:0)+(d?1:0)),_=h>1?p.pastSequenceLength:0,y=_+p.kvSequenceLength,g=l&&k.size(l.dims)>0?l:void 0,x=[t,r];h>1&&a&&k.size(a.dims)>0&&x.push(a),g&&x.push(g),m&&x.push(m),u&&x.push(u);let $=e.compute(Rm(h,t,r,a,g,p,_,m,u),{inputs:x,outputs:h>1?[-1,1]:[-1]})[0];e.compute(Mm($,p.batchSize,p.numHeads,_,p.sequenceLength,y,m,u),{inputs:m&&u?[$,m,u]:[$],outputs:[]});let v=[$,n];h>1&&d&&k.size(d.dims)>0&&v.push(d),m&&v.push(m),u&&v.push(u),e.compute(Um(h,$,n,d,p,_,m,u),{inputs:v,outputs:h>1?[0,2]:[0]})},Nm=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],n=t.sequenceLength,o=t.inputHiddenSize,i=t.headSize,a=12,d={x:Math.ceil(t.headSize/a),y:Math.ceil(t.sequenceLength/a),z:t.batchSize*t.numHeads},l=[e.inputs[0],e.inputs[1],e.inputs[2]],p=[{type:12,data:n},{type:12,data:o},{type:12,data:i},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],m=u=>{let h=M("output_q",l[0].dataType,r),_=M("output_k",l[0].dataType,r),y=M("output_v",l[0].dataType,r),g=E("input",l[0].dataType,l[0].dims),x=E("weight",l[1].dataType,l[1].dims),$=E("bias",l[2].dataType,l[2].dims),v=g.type.storage,S=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`
   const TILE_SIZE = ${a}u;
   var<workgroup> tileInput: array<${v}, ${a*a}>;
   var<workgroup> tileWeightQ: array<${v}, ${a*a}>;
@@ -4308,7 +4363,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
       output_k[outputIdx] = valueK;
       output_v[outputIdx] = valueV;
     }
-  }`};return e.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:d,programUniforms:p}),getShaderSource:m},{inputs:l,outputs:[-1,-1,-1]})},zs=(e,t)=>{let r=Bm(e.inputs,t),[n,o,i]=Nm(e,r);return Rt(e,n,o,i,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r)}});var Vm,Wm,Lm,Os,Ds=U(()=>{"use strict";We();te();oe();Se();ae();Vm=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(n,o,i)=>{let a=o.length;if(a!==n.length)throw new Error(`${i}: num dimensions != ${a}`);o.forEach((d,l)=>{if(d!==n[l])throw new Error(`${i}: dim[${l}] do not match`)})};if(e[0].dims.length>1){let n=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,n,"Invalid input scale"),r(e[2].dims,n,"Invalid input B"),r(e[3].dims,n,"Invalid input mean"),r(e[4].dims,n,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},Wm=(e,t)=>{let{epsilon:r,spatial:n,format:o}=t,i=e[0].dims,a=n?me(i[i.length-1]):1,d=o==="NHWC"&&i.length>1?a:1,l=C.size(i)/a,p=n,m=p?i.length:i,u=E("x",e[0].dataType,e[0].dims,a),h=E("scale",e[1].dataType,e[1].dims,d),_=E("bias",e[2].dataType,e[2].dims,d),y=E("inputMean",e[3].dataType,e[3].dims,d),g=E("inputVar",e[4].dataType,e[4].dims,d),x=M("y",e[0].dataType,m,a),$=()=>{let S="";if(n)S=`let cOffset = ${i.length===1?"0u":o==="NHWC"?`outputIndices[${i.length-1}] / ${a}`:"outputIndices[1]"};`;else if(o==="NCHW")S=`
+  }`};return e.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:d,programUniforms:p}),getShaderSource:m},{inputs:l,outputs:[-1,-1,-1]})},Ps=(e,t)=>{let r=Bm(e.inputs,t),[n,o,i]=Nm(e,r);return Ut(e,n,o,i,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r)}});var Vm,Wm,Lm,zs,Os=U(()=>{"use strict";Le();J();ne();xe();ae();Vm=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(n,o,i)=>{let a=o.length;if(a!==n.length)throw new Error(`${i}: num dimensions != ${a}`);o.forEach((d,l)=>{if(d!==n[l])throw new Error(`${i}: dim[${l}] do not match`)})};if(e[0].dims.length>1){let n=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,n,"Invalid input scale"),r(e[2].dims,n,"Invalid input B"),r(e[3].dims,n,"Invalid input mean"),r(e[4].dims,n,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},Wm=(e,t)=>{let{epsilon:r,spatial:n,format:o}=t,i=e[0].dims,a=n?me(i[i.length-1]):1,d=o==="NHWC"&&i.length>1?a:1,l=k.size(i)/a,p=n,m=p?i.length:i,u=E("x",e[0].dataType,e[0].dims,a),h=E("scale",e[1].dataType,e[1].dims,d),_=E("bias",e[2].dataType,e[2].dims,d),y=E("inputMean",e[3].dataType,e[3].dims,d),g=E("inputVar",e[4].dataType,e[4].dims,d),x=M("y",e[0].dataType,m,a),$=()=>{let S="";if(n)S=`let cOffset = ${i.length===1?"0u":o==="NHWC"?`outputIndices[${i.length-1}] / ${a}`:"outputIndices[1]"};`;else if(o==="NCHW")S=`
             ${x.indicesSet("outputIndices","0","0")}
             let cOffset = ${x.indicesToOffset("outputIndices")};`;else{S=`var cIndices = ${h.type.indices}(0);
                        cIndices[0] = outputIndices[${i.length-1}];`;for(let T=1;T<h.rank;T++)S+=`cIndices[${T}] = outputIndices[${T}];`;S+=`let cOffset = ${h.indicesToOffset("cIndices")};`}return S},v=S=>`
@@ -4325,7 +4380,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     let x = ${u.getByOffset("global_idx")};
     let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;
     ${x.setByOffset("global_idx","value")}
-  }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${n}_${a}`,inputDependencies:p?["rank","type","type","type","type"]:void 0},getShaderSource:v,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:p?[{type:12,data:l},...N(i)]:[{type:12,data:l}]})}},Lm=e=>re(e),Os=(e,t)=>{let{inputs:r,outputCount:n}=e,o=Lm({...t,outputCount:n});if(ve.webgpu.validateInputContent&&Vm(r,o),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(Wm(r,o))}});var Gm,Hm,Bs,Ms=U(()=>{"use strict";oe();ae();Gm=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(e[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},Hm=e=>{let t=e[0].dims,r=e[0].dims[2],n=C.size(t)/4,o=e[0].dataType,i=E("input",o,t,4),a=E("bias",o,[r],4),d=E("residual",o,t,4),l=M("output",o,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)}}),getShaderSource:m=>`
+  }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${n}_${a}`,inputDependencies:p?["rank","type","type","type","type"]:void 0},getShaderSource:v,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:p?[{type:12,data:l},...N(i)]:[{type:12,data:l}]})}},Lm=e=>ee(e),zs=(e,t)=>{let{inputs:r,outputCount:n}=e,o=Lm({...t,outputCount:n});if(_e.webgpu.validateInputContent&&Vm(r,o),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(Wm(r,o))}});var Gm,Hm,Ds,Bs=U(()=>{"use strict";ne();ae();Gm=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(e[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},Hm=e=>{let t=e[0].dims,r=e[0].dims[2],n=k.size(t)/4,o=e[0].dataType,i=E("input",o,t,4),a=E("bias",o,[r],4),d=E("residual",o,t,4),l=M("output",o,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)}}),getShaderSource:m=>`
   const channels = ${r}u / 4;
   ${m.declareVariables(i,a,d,l)}
 
@@ -4334,7 +4389,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     let value = ${i.getByOffset("global_idx")}
       + ${a.getByOffset("global_idx % channels")} + ${d.getByOffset("global_idx")};
     ${l.setByOffset("global_idx","value")}
-  }`}},Bs=e=>{Gm(e.inputs),e.compute(Hm(e.inputs))}});var Fm,fe,Rs,Us,Ns,Vs,Ws,Ls,Gs,Hs,Fs,qm,qs,Ks,js,Ys,Xt,Zs,qr,Qs,Xs,Js,eu,tu,ru,nu,ou,iu,au,su,uu,du,lu,cu,pu,mu,fu,mo,fo,hu,gu,bu,Km,jm,yu,Kr=U(()=>{"use strict";te();oe();Se();ae();Fm=(e,t,r,n,o,i,a)=>{let d=Math.ceil(t/4),l="";typeof o=="string"?l=`${o}(a)`:l=o("a");let p=E("inputData",r,[d],4),m=M("outputData",n,[d],4),u=[{name:"vec_size",type:"u32"}];return a&&u.push(...a),`
+  }`}},Ds=e=>{Gm(e.inputs),e.compute(Hm(e.inputs))}});var Fm,fe,Ms,Rs,Us,Ns,Vs,Ws,Ls,Gs,Hs,qm,Fs,qs,Ks,js,Qt,Ys,Fr,Zs,Qs,Xs,Js,eu,tu,ru,nu,ou,iu,au,su,uu,du,lu,cu,pu,mu,po,mo,fu,hu,gu,Km,jm,bu,qr=U(()=>{"use strict";J();ne();xe();ae();Fm=(e,t,r,n,o,i,a)=>{let d=Math.ceil(t/4),l="";typeof o=="string"?l=`${o}(a)`:l=o("a");let p=E("inputData",r,[d],4),m=M("outputData",n,[d],4),u=[{name:"vec_size",type:"u32"}];return a&&u.push(...a),`
       ${e.registerUniforms(u).declareVariables(p,m)}
 
   ${i??""}
@@ -4344,7 +4399,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
 
     let a = ${p.getByOffset("global_idx")};
     ${m.setByOffset("global_idx",l)}
-  }`},fe=(e,t,r,n,o,i=e.dataType,a,d)=>{let l=[{type:12,data:Math.ceil(C.size(e.dims)/4)}];return a&&l.push(...a),{name:t,shaderCache:{hint:o,inputDependencies:["type"]},getShaderSource:p=>Fm(p,C.size(e.dims),e.dataType,i,r,n,d),getRunData:p=>({outputs:[{dims:e.dims,dataType:i}],dispatchGroup:{x:Math.ceil(C.size(p[0].dims)/64/4)},programUniforms:l})}},Rs=e=>{e.compute(fe(e.inputs[0],"Abs","abs"))},Us=e=>{e.compute(fe(e.inputs[0],"Acos","acos"))},Ns=e=>{e.compute(fe(e.inputs[0],"Acosh","acosh"))},Vs=e=>{e.compute(fe(e.inputs[0],"Asin","asin"))},Ws=e=>{e.compute(fe(e.inputs[0],"Asinh","asinh"))},Ls=e=>{e.compute(fe(e.inputs[0],"Atan","atan"))},Gs=e=>{e.compute(fe(e.inputs[0],"Atanh","atanh"))},Hs=e=>re(e),Fs=(e,t)=>{let r;switch(t.to){case 10:r="vec4<f16>";break;case 1:r="vec4<f32>";break;case 12:r="vec4<u32>";break;case 6:r="vec4<i32>";break;case 9:r="vec4<bool>";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${t.to}`)}e.compute(fe(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},qm=e=>{let t,r,n=e.length>=2&&e[1].data!==0,o=e.length>=3&&e[2].data!==0;switch(e[0].dataType){case 1:t=n?e[1].getFloat32Array()[0]:-34028234663852886e22,r=o?e[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:t=n?e[1].getUint16Array()[0]:64511,r=o?e[2].getUint16Array()[0]:31743;break;default:throw new Error("Unsupport data type")}return re({min:t,max:r})},qs=(e,t)=>{let r=t||qm(e.inputs),n=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Clip",o=>`clamp(${o}, vec4<${n}>(uniforms.min), vec4<${n}>(uniforms.max))`,void 0,r.cacheKey,void 0,[{type:e.inputs[0].dataType,data:r.min},{type:e.inputs[0].dataType,data:r.max}],[{name:"min",type:n},{name:"max",type:n}]),{inputs:[0]})},Ks=e=>{e.compute(fe(e.inputs[0],"Ceil","ceil"))},js=e=>{e.compute(fe(e.inputs[0],"Cos","cos"))},Ys=e=>{e.compute(fe(e.inputs[0],"Cosh","cosh"))},Xt=e=>re(e),Zs=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Elu",n=>`elu_vf32(${n})`,`
+  }`},fe=(e,t,r,n,o,i=e.dataType,a,d)=>{let l=[{type:12,data:Math.ceil(k.size(e.dims)/4)}];return a&&l.push(...a),{name:t,shaderCache:{hint:o,inputDependencies:["type"]},getShaderSource:p=>Fm(p,k.size(e.dims),e.dataType,i,r,n,d),getRunData:p=>({outputs:[{dims:e.dims,dataType:i}],dispatchGroup:{x:Math.ceil(k.size(p[0].dims)/64/4)},programUniforms:l})}},Ms=e=>{e.compute(fe(e.inputs[0],"Abs","abs"))},Rs=e=>{e.compute(fe(e.inputs[0],"Acos","acos"))},Us=e=>{e.compute(fe(e.inputs[0],"Acosh","acosh"))},Ns=e=>{e.compute(fe(e.inputs[0],"Asin","asin"))},Vs=e=>{e.compute(fe(e.inputs[0],"Asinh","asinh"))},Ws=e=>{e.compute(fe(e.inputs[0],"Atan","atan"))},Ls=e=>{e.compute(fe(e.inputs[0],"Atanh","atanh"))},Gs=e=>ee(e),Hs=(e,t)=>{let r;switch(t.to){case 10:r="vec4<f16>";break;case 1:r="vec4<f32>";break;case 12:r="vec4<u32>";break;case 6:r="vec4<i32>";break;case 9:r="vec4<bool>";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${t.to}`)}e.compute(fe(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},qm=e=>{let t,r,n=e.length>=2&&e[1].data!==0,o=e.length>=3&&e[2].data!==0;switch(e[0].dataType){case 1:t=n?e[1].getFloat32Array()[0]:-34028234663852886e22,r=o?e[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:t=n?e[1].getUint16Array()[0]:64511,r=o?e[2].getUint16Array()[0]:31743;break;default:throw new Error("Unsupport data type")}return ee({min:t,max:r})},Fs=(e,t)=>{let r=t||qm(e.inputs),n=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Clip",o=>`clamp(${o}, vec4<${n}>(uniforms.min), vec4<${n}>(uniforms.max))`,void 0,r.cacheKey,void 0,[{type:e.inputs[0].dataType,data:r.min},{type:e.inputs[0].dataType,data:r.max}],[{name:"min",type:n},{name:"max",type:n}]),{inputs:[0]})},qs=e=>{e.compute(fe(e.inputs[0],"Ceil","ceil"))},Ks=e=>{e.compute(fe(e.inputs[0],"Cos","cos"))},js=e=>{e.compute(fe(e.inputs[0],"Cosh","cosh"))},Qt=e=>ee(e),Ys=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Elu",n=>`elu_vf32(${n})`,`
   const elu_alpha_ = ${r}(${t.alpha});
 
   fn elu_f32(a: ${r}) -> ${r} {
@@ -4353,7 +4408,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
 
   fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {
   return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
-  }`,t.cacheKey))},qr=(e="f32")=>`
+  }`,t.cacheKey))},Fr=(e="f32")=>`
 const r0: ${e} = 0.3275911;
 const r1: ${e} = 0.254829592;
 const r2: ${e} = -0.284496736;
@@ -4365,15 +4420,15 @@ fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {
   let absv = abs(v);
   let x = 1.0 / (1.0 + r0 * absv);
   return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
-}`,Qs=e=>{let t=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,qr(t)))},Xs=e=>{e.compute(fe(e.inputs[0],"Exp","exp"))},Js=e=>{e.compute(fe(e.inputs[0],"Floor","floor"))},eu=e=>{let t=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,qr(t)))},tu=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"LeakyRelu",n=>`select(leaky_relu_alpha_ * ${n}, ${n}, ${n} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},ru=e=>{e.compute(fe(e.inputs[0],"Not",t=>`!${t}`))},nu=e=>{e.compute(fe(e.inputs[0],"Neg",t=>`-${t}`))},ou=e=>{e.compute(fe(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},iu=e=>{let t=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},au=e=>{e.compute(fe(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},su=e=>re(e),uu=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${n} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},du=e=>{e.compute(fe(e.inputs[0],"Sin","sin"))},lu=e=>{e.compute(fe(e.inputs[0],"Sinh","sinh"))},cu=e=>{e.compute(fe(e.inputs[0],"Sqrt","sqrt"))},pu=e=>{e.compute(fe(e.inputs[0],"Tan","tan"))},mu=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,fu=e=>{e.compute(fe(e.inputs[0],"Tanh",mu))},mo=(e="f32")=>`
+}`,Zs=e=>{let t=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,Fr(t)))},Qs=e=>{e.compute(fe(e.inputs[0],"Exp","exp"))},Xs=e=>{e.compute(fe(e.inputs[0],"Floor","floor"))},Js=e=>{let t=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Fr(t)))},eu=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"LeakyRelu",n=>`select(leaky_relu_alpha_ * ${n}, ${n}, ${n} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},tu=e=>{e.compute(fe(e.inputs[0],"Not",t=>`!${t}`))},ru=e=>{e.compute(fe(e.inputs[0],"Neg",t=>`-${t}`))},nu=e=>{e.compute(fe(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},ou=e=>{let t=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},iu=e=>{e.compute(fe(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},au=e=>ee(e),su=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${n} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},uu=e=>{e.compute(fe(e.inputs[0],"Sin","sin"))},du=e=>{e.compute(fe(e.inputs[0],"Sinh","sinh"))},lu=e=>{e.compute(fe(e.inputs[0],"Sqrt","sqrt"))},cu=e=>{e.compute(fe(e.inputs[0],"Tan","tan"))},pu=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,mu=e=>{e.compute(fe(e.inputs[0],"Tanh",pu))},po=(e="f32")=>`
 const fast_gelu_a: ${e} = 0.5;
 const fast_gelu_b: ${e} = 0.7978845608028654;
 const fast_gelu_c: ${e} = 0.035677408136300125;
 
 fn tanh_v(v: vec4<${e}>) -> vec4<${e}> {
-  return ${mu("v")};
+  return ${pu("v")};
 }
-`,fo=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,hu=e=>{let t=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"FastGelu",fo,mo(t),void 0,e.inputs[0].dataType))},gu=(e,t)=>{let r=Ee(e.inputs[0].dataType);return e.compute(fe(e.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},bu=e=>{e.compute(fe(e.inputs[0],"Log","log"))},Km=(e,t)=>`
+`,mo=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,fu=e=>{let t=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"FastGelu",mo,po(t),void 0,e.inputs[0].dataType))},hu=(e,t)=>{let r=Ee(e.inputs[0].dataType);return e.compute(fe(e.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},gu=e=>{e.compute(fe(e.inputs[0],"Log","log"))},Km=(e,t)=>`
 const alpha = vec4<${e}>(${t});
 const one = ${e}(1.0);
 const zero = ${e}(0.0);
@@ -4390,13 +4445,13 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
   }
   return x * x1;
 }
-`,jm=e=>`quick_gelu_impl(${e})`,yu=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"QuickGelu",jm,Km(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}});var Ym,Zm,wu,vu=U(()=>{"use strict";oe();ae();Kr();Ym=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(e[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},Zm=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=E("input",e[0].dataType,e[0].dims,4),n=E("bias",e[0].dataType,[e[0].dims[2]],4),o=M("output",e[0].dataType,t,4),i=C.size(t)/4,a=_e(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:l=>`
+`,jm=e=>`quick_gelu_impl(${e})`,bu=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"QuickGelu",jm,Km(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}});var Ym,Zm,_u,wu=U(()=>{"use strict";ne();ae();qr();Ym=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(e[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},Zm=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=E("input",e[0].dataType,e[0].dims,4),n=E("bias",e[0].dataType,[e[0].dims[2]],4),o=M("output",e[0].dataType,t,4),i=k.size(t)/4,a=ye(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:l=>`
   const M_SQRT2 = sqrt(2.0);
   const halfChannels = ${e[0].dims[2]/4/2}u;
 
   ${l.declareVariables(r,n,o)}
 
-  ${qr(a)}
+  ${Fr(a)}
 
   ${l.mainStart()}
     ${l.guardAgainstOutOfBoundsWorkgroupSizes(i)}
@@ -4408,12 +4463,12 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
     let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);
 
     ${o.setByOffset("global_idx","valueLeft * geluRight")}
-  }`}},wu=e=>{Ym(e.inputs),e.compute(Zm(e.inputs))}});var Qm,Xm,at,$u,xu,Su,Tu,Iu,Cu,Au,ku,Eu,Pu,zu=U(()=>{"use strict";te();oe();ae();Qm=(e,t,r,n,o,i,a,d,l,p,m,u)=>{let h,_;typeof d=="string"?h=_=(v,S)=>`${d}((${v}),(${S}))`:typeof d=="function"?h=_=d:(h=d.scalar,_=d.vector);let y=M("outputData",m,n.length,4),g=E("aData",l,t.length,4),x=E("bData",p,r.length,4),$;if(o)if(i){let v=C.size(t)===1,S=C.size(r)===1,T=t.length>0&&t[t.length-1]%4===0,A=r.length>0&&r[r.length-1]%4===0;v||S?$=y.setByOffset("global_idx",_(v?`${g.type.value}(${g.getByOffset("0")}.x)`:g.getByOffset("global_idx"),S?`${x.type.value}(${x.getByOffset("0")}.x)`:x.getByOffset("global_idx"))):$=`
+  }`}},_u=e=>{Ym(e.inputs),e.compute(Zm(e.inputs))}});var Qm,Xm,at,vu,$u,xu,Su,Tu,Iu,Cu,Au,ku,Eu,Pu=U(()=>{"use strict";J();ne();ae();Qm=(e,t,r,n,o,i,a,d,l,p,m,u)=>{let h,_;typeof d=="string"?h=_=(v,S)=>`${d}((${v}),(${S}))`:typeof d=="function"?h=_=d:(h=d.scalar,_=d.vector);let y=M("outputData",m,n.length,4),g=E("aData",l,t.length,4),x=E("bData",p,r.length,4),$;if(o)if(i){let v=k.size(t)===1,S=k.size(r)===1,T=t.length>0&&t[t.length-1]%4===0,A=r.length>0&&r[r.length-1]%4===0;v||S?$=y.setByOffset("global_idx",_(v?`${g.type.value}(${g.getByOffset("0")}.x)`:g.getByOffset("global_idx"),S?`${x.type.value}(${x.getByOffset("0")}.x)`:x.getByOffset("global_idx"))):$=`
             let outputIndices = ${y.offsetToIndices("global_idx * 4u")};
             let offsetA = ${g.broadcastedIndicesToOffset("outputIndices",y)};
             let offsetB = ${x.broadcastedIndicesToOffset("outputIndices",y)};
             ${y.setByOffset("global_idx",_(a||T?g.getByOffset("offsetA / 4u"):`${g.type.value}(${g.getByOffset("offsetA / 4u")}[offsetA % 4u])`,a||A?x.getByOffset("offsetB / 4u"):`${x.type.value}(${x.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
-          `}else $=y.setByOffset("global_idx",_(g.getByOffset("global_idx"),x.getByOffset("global_idx")));else{if(!i)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let v=(S,T,A="")=>{let k=`aData[indexA${T}][componentA${T}]`,P=`bData[indexB${T}][componentB${T}]`;return`
+          `}else $=y.setByOffset("global_idx",_(g.getByOffset("global_idx"),x.getByOffset("global_idx")));else{if(!i)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let v=(S,T,A="")=>{let C=`aData[indexA${T}][componentA${T}]`,P=`bData[indexB${T}][componentB${T}]`;return`
             let outputIndices${T} = ${y.offsetToIndices(`global_idx * 4u + ${T}u`)};
             let offsetA${T} = ${g.broadcastedIndicesToOffset(`outputIndices${T}`,y)};
             let offsetB${T} = ${x.broadcastedIndicesToOffset(`outputIndices${T}`,y)};
@@ -4421,7 +4476,7 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
             let indexB${T} = offsetB${T} / 4u;
             let componentA${T} = offsetA${T} % 4u;
             let componentB${T} = offsetB${T} % 4u;
-            ${S}[${T}] = ${A}(${h(k,P)});
+            ${S}[${T}] = ${A}(${h(C,P)});
           `};m===9?$=`
             var data = vec4<u32>(0);
             ${v("data",0,"u32")}
@@ -4441,7 +4496,7 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
         ${e.mainStart()}
         ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
         ${$}
-      }`},Xm=(e,t,r,n,o,i,a=r.dataType)=>{let d=r.dims.map(g=>Number(g)??1),l=n.dims.map(g=>Number(g)??1),p=!C.areEqual(d,l),m=d,u=C.size(d),h=!1,_=!1,y=[p];if(p){let g=tt.calcShape(d,l,!1);if(!g)throw new Error("Can't perform binary op on the given tensors");m=g.slice(),u=C.size(m);let x=C.size(d)===1,$=C.size(l)===1,v=d.length>0&&d[d.length-1]%4===0,S=l.length>0&&l[l.length-1]%4===0;y.push(x),y.push($),y.push(v),y.push(S);let T=1;for(let A=1;A<m.length;A++){let k=d[d.length-A],P=l[l.length-A];if(k===P)T*=k;else break}T%4===0?(_=!0,h=!0):(x||$||v||S)&&(h=!0)}else h=!0;return y.push(h),{name:e,shaderCache:{hint:t+y.map(g=>g.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:g=>Qm(g,d,l,m,h,p,_,o,r.dataType,n.dataType,a,i),getRunData:()=>({outputs:[{dims:m,dataType:a}],dispatchGroup:{x:Math.ceil(u/64/4)},programUniforms:[{type:12,data:Math.ceil(C.size(m)/4)},...N(d,l,m)]})}},at=(e,t,r,n,o,i)=>{e.compute(Xm(t,o??"",e.inputs[0],e.inputs[1],r,n,i))},$u=e=>{at(e,"Add",(t,r)=>`${t}+${r}`)},xu=e=>{at(e,"Div",(t,r)=>`${t}/${r}`)},Su=e=>{at(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},Tu=e=>{at(e,"Mul",(t,r)=>`${t}*${r}`)},Iu=e=>{let t=E("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;at(e,"Pow",{scalar:(n,o)=>`pow_custom(${n},${o})`,vector:(n,o)=>`pow_vector_custom(${n},${o})`},`
+      }`},Xm=(e,t,r,n,o,i,a=r.dataType)=>{let d=r.dims.map(g=>Number(g)??1),l=n.dims.map(g=>Number(g)??1),p=!k.areEqual(d,l),m=d,u=k.size(d),h=!1,_=!1,y=[p];if(p){let g=tt.calcShape(d,l,!1);if(!g)throw new Error("Can't perform binary op on the given tensors");m=g.slice(),u=k.size(m);let x=k.size(d)===1,$=k.size(l)===1,v=d.length>0&&d[d.length-1]%4===0,S=l.length>0&&l[l.length-1]%4===0;y.push(x),y.push($),y.push(v),y.push(S);let T=1;for(let A=1;A<m.length;A++){let C=d[d.length-A],P=l[l.length-A];if(C===P)T*=C;else break}T%4===0?(_=!0,h=!0):(x||$||v||S)&&(h=!0)}else h=!0;return y.push(h),{name:e,shaderCache:{hint:t+y.map(g=>g.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:g=>Qm(g,d,l,m,h,p,_,o,r.dataType,n.dataType,a,i),getRunData:()=>({outputs:[{dims:m,dataType:a}],dispatchGroup:{x:Math.ceil(u/64/4)},programUniforms:[{type:12,data:Math.ceil(k.size(m)/4)},...N(d,l,m)]})}},at=(e,t,r,n,o,i)=>{e.compute(Xm(t,o??"",e.inputs[0],e.inputs[1],r,n,i))},vu=e=>{at(e,"Add",(t,r)=>`${t}+${r}`)},$u=e=>{at(e,"Div",(t,r)=>`${t}/${r}`)},xu=e=>{at(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},Su=e=>{at(e,"Mul",(t,r)=>`${t}*${r}`)},Tu=e=>{let t=E("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;at(e,"Pow",{scalar:(n,o)=>`pow_custom(${n},${o})`,vector:(n,o)=>`pow_vector_custom(${n},${o})`},`
     fn pow_custom(a : ${t}, b : ${t}) -> ${t} {
       if (b == ${t}(0.0)) {
         return ${t}(1.0);
@@ -4454,7 +4509,7 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
       // TODO: implement vectorized pow
       return vec4<${t}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w));
     }
-      `)},Cu=e=>{at(e,"Sub",(t,r)=>`${t}-${r}`)},Au=e=>{at(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},ku=e=>{at(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},Eu=e=>{at(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},Pu=e=>{at(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}});var ef,tf,rf,nf,Ou,Du,Bu=U(()=>{"use strict";te();oe();Se();ae();ef=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,n=e[r],o=n.dataType,i=n.dims.length;e.forEach((a,d)=>{if(d!==r){if(a.dataType!==o)throw new Error("input tensors should be one type");if(a.dims.length!==i)throw new Error("input tensors should have the same shape");a.dims.forEach((l,p)=>{if(p!==t&&l!==n.dims[p])throw new Error("non concat dimensions must match")})}})},tf=(e,t)=>`
+      `)},Iu=e=>{at(e,"Sub",(t,r)=>`${t}-${r}`)},Cu=e=>{at(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},Au=e=>{at(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},ku=e=>{at(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},Eu=e=>{at(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}});var ef,tf,rf,nf,zu,Ou,Du=U(()=>{"use strict";J();ne();xe();ae();ef=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,n=e[r],o=n.dataType,i=n.dims.length;e.forEach((a,d)=>{if(d!==r){if(a.dataType!==o)throw new Error("input tensors should be one type");if(a.dims.length!==i)throw new Error("input tensors should have the same shape");a.dims.forEach((l,p)=>{if(p!==t&&l!==n.dims[p])throw new Error("non concat dimensions must match")})}})},tf=(e,t)=>`
   fn calculateInputIndex(index: u32) -> u32 {
     let sizeInConcatAxis = array<u32, ${e}u>(${t});
     for (var i: u32 = 0u; i < ${e}; i += 1u ) {
@@ -4464,7 +4519,7 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
     }
     return ${e}u;
   }`,rf=(e,t)=>{let r=e.length,n=[];for(let o=0;o<r;++o){let i=t.setByOffset("global_idx",e[o].getByIndices("indices"));r===1?n.push(i):o===0?n.push(`if (inputIndex == ${o}u) { ${i} }`):o===r-1?n.push(`else { ${i} }`):n.push(`else if (inputIndex == ${o}) { ${i} }`)}return n.join(`
-`)},nf=(e,t,r,n)=>{let o=C.size(r),i=new Array(e.length),a=new Array(e.length),d=0,l=[],p=[],m=[{type:12,data:o}];for(let g=0;g<e.length;++g)d+=e[g].dims[t],i[g]=d,p.push(e[g].dims.length),a[g]=E(`input${g}`,n,p[g]),l.push("rank"),m.push({type:12,data:i[g]});for(let g=0;g<e.length;++g)m.push(...N(e[g].dims));m.push(...N(r));let u=M("output",n,r.length),h=u.indicesGet("indices",t),_=Array.from(Array(i.length).keys()).map(g=>`uniforms.sizeInConcatAxis${g}`).join(","),y=g=>`
+`)},nf=(e,t,r,n)=>{let o=k.size(r),i=new Array(e.length),a=new Array(e.length),d=0,l=[],p=[],m=[{type:12,data:o}];for(let g=0;g<e.length;++g)d+=e[g].dims[t],i[g]=d,p.push(e[g].dims.length),a[g]=E(`input${g}`,n,p[g]),l.push("rank"),m.push({type:12,data:i[g]});for(let g=0;g<e.length;++g)m.push(...N(e[g].dims));m.push(...N(r));let u=M("output",n,r.length),h=u.indicesGet("indices",t),_=Array.from(Array(i.length).keys()).map(g=>`uniforms.sizeInConcatAxis${g}`).join(","),y=g=>`
 
   ${(()=>{g.registerUniform("outputSize","u32");for(let x=0;x<e.length;x++)g.registerUniform(`sizeInConcatAxis${x}`,"u32");return g.declareVariables(...a,u)})()}
 
@@ -4482,11 +4537,11 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
     }
 
     ${rf(a,u)}
-  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:m}),getShaderSource:y}},Ou=(e,t)=>{let r=e.inputs,n=r[0].dims,o=C.normalizeAxis(t.axis,n.length);ef(r,o);let i=n.slice();i[o]=r.reduce((d,l)=>d+(l.dims.length>o?l.dims[o]:0),0);let a=r.filter(d=>C.size(d.dims)>0);e.compute(nf(a,o,i,r[0].dataType),{inputs:a})},Du=e=>re({axis:e.axis})});var qe,Ke,je,jr,yt=U(()=>{"use strict";te();oe();qe=(e,t,r="f32")=>{switch(e.activation){case"Relu":return`value = max(value, ${t}(0.0));`;case"Sigmoid":return`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${t}(${r}(uniforms.clip_min)), ${t}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${t}(0.0), min(${t}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case"Tanh":return`let e2x = exp(-2.0 * abs(value));
+  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:m}),getShaderSource:y}},zu=(e,t)=>{let r=e.inputs,n=r[0].dims,o=k.normalizeAxis(t.axis,n.length);ef(r,o);let i=n.slice();i[o]=r.reduce((d,l)=>d+(l.dims.length>o?l.dims[o]:0),0);let a=r.filter(d=>k.size(d.dims)>0);e.compute(nf(a,o,i,r[0].dataType),{inputs:a})},Ou=e=>ee({axis:e.axis})});var Ke,je,Ye,Kr,yt=U(()=>{"use strict";J();ne();Ke=(e,t,r="f32")=>{switch(e.activation){case"Relu":return`value = max(value, ${t}(0.0));`;case"Sigmoid":return`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${t}(${r}(uniforms.clip_min)), ${t}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${t}(0.0), min(${t}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case"Tanh":return`let e2x = exp(-2.0 * abs(value));
               value = sign(value) * (1.0 - e2x) / (1.0 + e2x);
-        `;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},Ke=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},je=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},jr=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,n]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:n}}else if(t==="Clip"){let[r,n]=e?.activation_params||[es,ts];return{activation:t,clipMax:n,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}});var Ae,Mu,Yr=U(()=>{"use strict";Ae=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Mu=e=>`
+        `;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},je=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},Ye=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},Kr=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,n]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:n}}else if(t==="Clip"){let[r,n]=e?.activation_params||[Ja,es];return{activation:t,clipMax:n,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}});var Ae,Bu,jr=U(()=>{"use strict";Ae=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Bu=e=>`
       ${e?"value = value + getBiasByOutputCoords(coords);":""}
-      `});var Ru,Uu=U(()=>{"use strict";Ru=e=>`
+      `});var Mu,Ru=U(()=>{"use strict";Mu=e=>`
 fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
   return dot(coords, vec4<i32>(
       shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
@@ -4495,18 +4550,18 @@ fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
   return dot(coords, vec4<i32>(
     i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));
 }
-`});var Jt,Zr,Qr=U(()=>{"use strict";te();oe();ae();yt();Jt=(e,t,r,n,o)=>{let i=n-r;return`
+`});var Xt,Yr,Zr=U(()=>{"use strict";J();ne();ae();yt();Xt=(e,t,r,n,o)=>{let i=n-r;return`
       ${Array.from({length:r}).map((a,d)=>`
-      if (${F(t.shape,d,t.rank)} != 1) {
-        ${t.indicesSet(e,d,F(o,d+i,n))}
+      if (${q(t.shape,d,t.rank)} != 1) {
+        ${t.indicesSet(e,d,q(o,d+i,n))}
       } else {
         ${t.indicesSet(e,d,0)}
       }`).join("")}
-`},Zr=(e,t,r,n,o=!1,i)=>{let a=e[0].dims,d=e[1].dims,l=a[a.length-2],p=d[d.length-1],m=a[a.length-1],u=me(p),h=me(m),_=me(l),y=C.size(r)/u/_,g=e.length>2,x=n?n.slice(0,-2):r.slice(0,-2),v=[C.size(x),l,p],S=[{type:12,data:y},{type:12,data:l},{type:12,data:p},{type:12,data:m}];Ke(t,S),S.push(...N(x,a,d)),g&&S.push(...N(e[2].dims)),S.push(...N(v));let T=A=>{let k=Lr("batch_dims",e[0].dataType,x.length),P=E("a",e[0].dataType,a.length,h),D=E("b",e[1].dataType,d.length,u),R=M("output",e[0].dataType,v.length,u),G=_e(R.type.tensor),K=qe(t,R.type.value,G),j=[P,D],V="";if(g){let Y=o?u:1;j.push(E("bias",e[2].dataType,e[2].dims.length,Y)),V=`${o?`value += bias[col / ${Y}];`:`value += ${R.type.value}(bias[row + i]);`}`}let Q=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];je(t,Q);let se=()=>{let Y=`var a_data: ${P.type.value};`;for(let ee=0;ee<h;ee++)Y+=`
-              let b_data${ee} = b[(b_offset + (k + ${ee}) * uniforms.N + col) / ${u}];`;for(let ee=0;ee<_;ee++){Y+=`a_data = a[(a_offset + (row + ${ee}) * uniforms.K + k) / ${h}];`;for(let J=0;J<h;J++)Y+=`
-            values[${ee}] = fma(${D.type.value}(a_data${h===1?"":`[${J}]`}), b_data${J}, values[${ee}]);
-`}return Y};return`
-  ${A.registerUniforms(Q).registerInternalVariables(k).declareVariables(...j,R)}
+`},Yr=(e,t,r,n,o=!1,i)=>{let a=e[0].dims,d=e[1].dims,l=a[a.length-2],p=d[d.length-1],m=a[a.length-1],u=me(p),h=me(m),_=me(l),y=k.size(r)/u/_,g=e.length>2,x=n?n.slice(0,-2):r.slice(0,-2),v=[k.size(x),l,p],S=[{type:12,data:y},{type:12,data:l},{type:12,data:p},{type:12,data:m}];je(t,S),S.push(...N(x,a,d)),g&&S.push(...N(e[2].dims)),S.push(...N(v));let T=A=>{let C=Wr("batch_dims",e[0].dataType,x.length),P=E("a",e[0].dataType,a.length,h),D=E("b",e[1].dataType,d.length,u),R=M("output",e[0].dataType,v.length,u),H=ye(R.type.tensor),L=Ke(t,R.type.value,H),re=[P,D],V="";if(g){let j=o?u:1;re.push(E("bias",e[2].dataType,e[2].dims.length,j)),V=`${o?`value += bias[col / ${j}];`:`value += ${R.type.value}(bias[row + i]);`}`}let K=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Ye(t,K);let we=()=>{let j=`var a_data: ${P.type.value};`;for(let Q=0;Q<h;Q++)j+=`
+              let b_data${Q} = b[(b_offset + (k + ${Q}) * uniforms.N + col) / ${u}];`;for(let Q=0;Q<_;Q++){j+=`a_data = a[(a_offset + (row + ${Q}) * uniforms.K + k) / ${h}];`;for(let ie=0;ie<h;ie++)j+=`
+            values[${Q}] = fma(${D.type.value}(a_data${h===1?"":`[${ie}]`}), b_data${ie}, values[${Q}]);
+`}return j};return`
+  ${A.registerUniforms(K).registerInternalVariables(C).declareVariables(...re,R)}
   ${A.mainStart()}
     ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
     let col = (global_idx % (uniforms.N / ${u})) * ${u};
@@ -4515,33 +4570,33 @@ fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
     let row = (index1 % stride1) * ${_};
     let batch = index1 / stride1;
 
-    ${r.length===2?"":`let batch_indices = ${k.offsetToIndices("batch")};`}
+    ${r.length===2?"":`let batch_indices = ${C.offsetToIndices("batch")};`}
 
     var a_indices: ${P.type.indices};
-    ${Jt("a_indices",P,P.rank-2,k.rank,"batch_indices")}
+    ${Xt("a_indices",P,P.rank-2,C.rank,"batch_indices")}
     ${P.indicesSet("a_indices",P.rank-2,0)}
     ${P.indicesSet("a_indices",P.rank-1,0)}
     let a_offset = ${P.indicesToOffset("a_indices")};
 
     var b_indices: ${D.type.indices};
-    ${Jt("b_indices",D,D.rank-2,k.rank,"batch_indices")}
+    ${Xt("b_indices",D,D.rank-2,C.rank,"batch_indices")}
     ${D.indicesSet("b_indices",D.rank-2,0)}
     ${D.indicesSet("b_indices",D.rank-1,0)}
     let b_offset = ${D.indicesToOffset("b_indices")};
     var values: array<${R.type.value}, ${_}>;
     for (var k: u32 = 0u; k < uniforms.K; k = k + ${h}) {
-      ${se()}
+      ${we()}
     }
     for (var i = 0u; i < ${_}u; i++) {
       var value = values[i];
       ${V}
-      ${K}
+      ${L}
       let cur_indices = ${R.type.indices}(batch, row + i, col);
       let offset = ${R.indicesToOffset("cur_indices")};
       ${R.setByOffset(`offset / ${u}`,"value")};
     }
   }
-  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${u};${h};${_};${o}`,inputDependencies:g?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:S}),getShaderSource:T}}});var of,af,ho,Nu,sf,go,uf,er,Xr=U(()=>{"use strict";te();oe();ae();yt();Qr();Yr();of=(e,t)=>e?`
+  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${u};${h};${_};${o}`,inputDependencies:g?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:S}),getShaderSource:T}}});var of,af,fo,Uu,sf,ho,uf,Jt,Qr=U(()=>{"use strict";J();ne();ae();yt();Zr();jr();of=(e,t)=>e?`
         mm_Asub[inputRow][inputCol] = mm_readA(batch,
           kStart + inputRow,
           globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});
@@ -4566,7 +4621,7 @@ fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
           acc[i] = BCached1 * ACached.y + acc[i];
           acc[i] = BCached2 * ACached.z + acc[i];
           ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}
-        }`,ho=(e,t,r="f32",n,o=!1,i=32,a=!1,d=32)=>{let l=t[1]*e[1],p=t[0]*e[0],m=o?l:i,u=o?i:l,h=m/t[0],_=i/t[1];if(!((o&&h===4&&e[1]===4||!o&&(h===3||h===4))&&m%t[0]===0&&i%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${o} is true, innerElementSize ${h} and workPerThread[1] ${e[1]} must be 4.
+        }`,fo=(e,t,r="f32",n,o=!1,i=32,a=!1,d=32)=>{let l=t[1]*e[1],p=t[0]*e[0],m=o?l:i,u=o?i:l,h=m/t[0],_=i/t[1];if(!((o&&h===4&&e[1]===4||!o&&(h===3||h===4))&&m%t[0]===0&&i%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${o} is true, innerElementSize ${h} and workPerThread[1] ${e[1]} must be 4.
       Otherwise, innerElementSize ${h} must be 3 or 4.
   tileAWidth ${m} must be divisible by workgroupSize[0]${t[0]}. tileInner ${i} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`
 var<workgroup> mm_Asub: array<array<vec${h}<${r}>, ${m/h}>, ${u}>;
@@ -4631,7 +4686,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
   for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
       mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
   }
-}`},Nu=(e,t)=>e?`
+}`},Uu=(e,t)=>e?`
             mm_Asub[inputRow][inputCol] = mm_readA(batch,
               kStart + inputRow,
               globalRowStart + inputCol${t?", batchIndices":""});
@@ -4639,7 +4694,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
             mm_Asub[inputRow][inputCol] = mm_readA(batch,
               globalRowStart + inputRow,
               kStart + inputCol${t?", batchIndices":""});
-            `,sf=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",go=(e,t,r="f32",n,o=!1,i=32,a=!1,d=32,l=!1)=>{let p=e[1]*t[1],m=e[0]*t[0],u=o?p:i,h=o?i:p;if(!(h%t[1]===0&&u%t[0]===0&&i%t[1]===0))throw new Error(`tileAHight ${h} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${u} must be divisible by workgroupSize[0]${t[0]}, tileInner ${i} must be divisible by workgroupSize[1]${t[1]}`);let _=h/t[1],y=u/t[0],g=i/t[1],x=l?`
+            `,sf=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",ho=(e,t,r="f32",n,o=!1,i=32,a=!1,d=32,l=!1)=>{let p=e[1]*t[1],m=e[0]*t[0],u=o?p:i,h=o?i:p;if(!(h%t[1]===0&&u%t[0]===0&&i%t[1]===0))throw new Error(`tileAHight ${h} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${u} must be divisible by workgroupSize[0]${t[0]}, tileInner ${i} must be divisible by workgroupSize[1]${t[1]}`);let _=h/t[1],y=u/t[0],g=i/t[1],x=l?`
     let localRow = i32(localId.y);
     let localCol = i32(localId.x);
     let globalRowStart = i32(workgroupId.y) * ${p};
@@ -4650,7 +4705,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       // Load one tile of A into local memory.
       for (var inputRow = localRow; inputRow < ${h}; inputRow = inputRow + ${t[1]}) {
         for (var inputCol = localCol; inputCol < ${u}; inputCol = inputCol + ${t[0]}) {
-          ${Nu(o,n)}
+          ${Uu(o,n)}
         }
       }
       // Load one tile of B into local memory.
@@ -4705,7 +4760,7 @@ for (var t = 0; t < num_tiles; t = t + 1) {
     for (var innerCol = 0; innerCol < ${y}; innerCol = innerCol + 1) {
       let inputRow = tileRowA + innerRow;
       let inputCol = tileColA + innerCol;
-      ${Nu(o,n)}
+      ${Uu(o,n)}
     }
   }
 
@@ -4765,14 +4820,14 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     var acc : array<array<${r}, colPerThread>, rowPerThread>;
     ${x}
   }
-`},uf=(e,t,r,n,o=!1)=>{let[i,a,d,l]=n,p=_e(n[0].type.tensor);return`
+`},uf=(e,t,r,n,o=!1)=>{let[i,a,d,l]=n,p=ye(n[0].type.tensor);return`
     fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${i.type.indices}) -> ${Ae(e,p)} {
       var value = ${Ae(e,p)}(0.0);
       let col = colIn * ${e};
       if(row < uniforms.dim_a_outer && col < uniforms.dim_inner)
       {
         var aIndices: ${a.type.indices};
-        ${Jt("aIndices",a,a.rank-2,i.rank,"batchIndices")}
+        ${Xt("aIndices",a,a.rank-2,i.rank,"batchIndices")}
         ${a.indicesSet("aIndices",a.rank-2,"u32(row)")}
         ${a.indicesSet("aIndices",a.rank-1,"u32(colIn)")}
         value = ${a.getByIndices("aIndices")};
@@ -4786,7 +4841,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       if(row < uniforms.dim_inner && col < uniforms.dim_b_outer)
       {
         var bIndices: ${d.type.indices};
-        ${Jt("bIndices",d,d.rank-2,i.rank,"batchIndices")}
+        ${Xt("bIndices",d,d.rank-2,i.rank,"batchIndices")}
         ${d.indicesSet("bIndices",d.rank-2,"u32(row)")}
         ${d.indicesSet("bIndices",d.rank-1,"u32(colIn)")}
         value = ${d.getByIndices("bIndices")};
@@ -4804,11 +4859,11 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
         ${l.setByIndices("vec3<u32>(coords)","value")}
       }
     }
-    `},er=(e,t,r,n,o=!1,i)=>{let a=e[0].dims,d=e[1].dims,l=a.slice(0,-2),p=d.slice(0,-2),m=n?n.slice(0,-2):r.slice(0,-2),u=C.size(m),h=a[a.length-2],_=a[a.length-1],y=d[d.length-1],g=_%4===0&&y%4===0,x=h<=8?[4,1,1]:[4,4,1],$=[8,8,1],v=[Math.ceil(y/$[0]/x[0]),Math.ceil(h/$[1]/x[1]),Math.ceil(u/$[2]/x[2])],S=g?4:1,T=[...l,h,_/S],A=T.length,k=[...p,_,y/S],P=k.length,D=[u,h,y/S],R=[{type:6,data:h},{type:6,data:y},{type:6,data:_}];Ke(t,R),R.push(...N(m,T,k));let G=["rank","rank"],K=e.length>2;K&&(R.push(...N(e[2].dims)),G.push("rank")),R.push(...N(D));let j=V=>{let Q=m.length,se=Lr("batchDims",e[0].dataType,Q,1),Y=_e(e[0].dataType),ee=E("a",e[0].dataType,A,S),J=E("b",e[1].dataType,P,S),ne=M("result",e[0].dataType,D.length,S),be=[ee,J];if(K){let q=o?S:1;be.push(E("bias",e[2].dataType,e[2].dims.length,q))}let Oe=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];je(t,Oe);let $e=_e(ne.type.tensor),le=qe(t,ne.type.value,$e),W=uf(S,K,le,[se,ee,J,ne],o);return`
-  ${V.registerUniforms(Oe).registerInternalVariables(se).declareVariables(...be,ne)}
+    `},Jt=(e,t,r,n,o=!1,i)=>{let a=e[0].dims,d=e[1].dims,l=a.slice(0,-2),p=d.slice(0,-2),m=n?n.slice(0,-2):r.slice(0,-2),u=k.size(m),h=a[a.length-2],_=a[a.length-1],y=d[d.length-1],g=_%4===0&&y%4===0,x=h<=8?[4,1,1]:[4,4,1],$=[8,8,1],v=[Math.ceil(y/$[0]/x[0]),Math.ceil(h/$[1]/x[1]),Math.ceil(u/$[2]/x[2])],S=g?4:1,T=[...l,h,_/S],A=T.length,C=[...p,_,y/S],P=C.length,D=[u,h,y/S],R=[{type:6,data:h},{type:6,data:y},{type:6,data:_}];je(t,R),R.push(...N(m,T,C));let H=["rank","rank"],L=e.length>2;L&&(R.push(...N(e[2].dims)),H.push("rank")),R.push(...N(D));let re=V=>{let K=m.length,we=Wr("batchDims",e[0].dataType,K,1),j=ye(e[0].dataType),Q=E("a",e[0].dataType,A,S),ie=E("b",e[1].dataType,P,S),te=M("result",e[0].dataType,D.length,S),be=[Q,ie];if(L){let Y=o?S:1;be.push(E("bias",e[2].dataType,e[2].dims.length,Y))}let Oe=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];Ye(t,Oe);let ve=ye(te.type.tensor),de=Ke(t,te.type.value,ve),W=uf(S,L,de,[we,Q,ie,te],o);return`
+  ${V.registerUniforms(Oe).registerInternalVariables(we).declareVariables(...be,te)}
   ${W}
-  ${g?ho(x,$,Y,se):go(x,$,Y,se)}
-                   `};return{name:"MatMul",shaderCache:{hint:`${x};${t.activation};${g};${o}`,inputDependencies:G},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:v[0],y:v[1],z:v[2]},programUniforms:R}),getShaderSource:j}}});var df,Vu,Wu=U(()=>{"use strict";te();et();ae();yt();Yr();Uu();Xr();df=(e,t,r,n,o=!1,i,a=4,d=4,l=4,p="f32")=>{let m=G=>{switch(G){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${p}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${G} is not supported.`)}},u=G=>{switch(G){case 1:return"return w[row * i32(uniforms.w_shape[3]) + colIn];";case 4:return"return w[row * i32(uniforms.w_shape[3]) / 4 + colIn];";default:throw new Error(`innerElementSize ${G} is not supported.`)}},h=e?`
+  ${g?fo(x,$,j,we):ho(x,$,j,we)}
+                   `};return{name:"MatMul",shaderCache:{hint:`${x};${t.activation};${g};${o}`,inputDependencies:H},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:v[0],y:v[1],z:v[2]},programUniforms:R}),getShaderSource:re}}});var df,Nu,Vu=U(()=>{"use strict";J();et();ae();yt();jr();Ru();Qr();df=(e,t,r,n,o=!1,i,a=4,d=4,l=4,p="f32")=>{let m=H=>{switch(H){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${p}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${H} is not supported.`)}},u=H=>{switch(H){case 1:return"return w[row * i32(uniforms.w_shape[3]) + colIn];";case 4:return"return w[row * i32(uniforms.w_shape[3]) / 4 + colIn];";default:throw new Error(`innerElementSize ${H} is not supported.`)}},h=e?`
     let coord = vec4<i32>(batch, xRow, xCol, xCh);
     `:`
     let coord = vec4<i32>(batch, xCh, xRow, xCol);
@@ -4867,8 +4922,8 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     if (row < uniforms.dim_inner && col < uniforms.dim_a_outer) {
       ${u(d)}
     }
-    return ${Ae(d,p)}(0.0);`,A=Ae(l,p),k=e?Ae(a,p):Ae(d,p),P=e?Ae(d,p):Ae(a,p),D=qe(i,A,p);return`
-    fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${k} {
+    return ${Ae(d,p)}(0.0);`,A=Ae(l,p),C=e?Ae(a,p):Ae(d,p),P=e?Ae(d,p):Ae(a,p),D=Ke(i,A,p);return`
+    fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${C} {
       ${e?S:T}
     }
 
@@ -4883,36 +4938,36 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       var value = valueIn;
       let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
       ${_}
-      ${Mu(o)}
+      ${Bu(o)}
       ${D}
       setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
       }
-    }`},Vu=(e,t,r,n,o,i,a,d,l)=>{let p=t.format==="NHWC",m=p?e[0].dims[3]:e[0].dims[1],u=r[0],h=p?r[2]:r[3],_=p?r[1]:r[2],y=p?r[3]:r[1],g=p&&(m%4===0||m%3===0)&&y%4===0,x=p?y:h*_,$=p?h*_:y,v=[8,8,1],S=n<=8?[4,1,1]:[4,4,1],T=[Math.ceil(x/v[0]/S[0]),Math.ceil($/v[1]/S[1]),Math.ceil(u/v[2]/S[2])];ue("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${T}`);let A=g?p&&m%4!==0?3:4:1,k=v[1]*S[1],P=v[0]*S[0],D=Math.max(v[0]*A,v[1]),R=n%k===0,G=o%P===0,K=i%D===0,j=g?[A,4,4]:[1,1,1],V=[{type:6,data:n},{type:6,data:o},{type:6,data:i},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];Ke(t,V),V.push(...N(e[0].dims,e[1].dims));let Q=["rank","rank"];a&&(V.push(...N(e[2].dims)),Q.push("rank")),V.push(...N(r));let se=Y=>{let ee=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"pad",type:"i32",length:2},{name:"stride",type:"i32",length:2},{name:"dilation",type:"i32",length:2}];je(t,ee);let J=g?4:1,ne=_e(e[0].dataType),be=`
-      fn setOutputAtIndex(flatIndex : i32, value : ${g?`vec4<${ne}>`:ne}) {
-        result[flatIndex] = ${g?`vec4<${ne}>`:ne}(value);
+    }`},Nu=(e,t,r,n,o,i,a,d,l)=>{let p=t.format==="NHWC",m=p?e[0].dims[3]:e[0].dims[1],u=r[0],h=p?r[2]:r[3],_=p?r[1]:r[2],y=p?r[3]:r[1],g=p&&(m%4===0||m%3===0)&&y%4===0,x=p?y:h*_,$=p?h*_:y,v=[8,8,1],S=n<=8?[4,1,1]:[4,4,1],T=[Math.ceil(x/v[0]/S[0]),Math.ceil($/v[1]/S[1]),Math.ceil(u/v[2]/S[2])];se("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${T}`);let A=g?p&&m%4!==0?3:4:1,C=v[1]*S[1],P=v[0]*S[0],D=Math.max(v[0]*A,v[1]),R=n%C===0,H=o%P===0,L=i%D===0,re=g?[A,4,4]:[1,1,1],V=[{type:6,data:n},{type:6,data:o},{type:6,data:i},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];je(t,V),V.push(...N(e[0].dims,e[1].dims));let K=["rank","rank"];a&&(V.push(...N(e[2].dims)),K.push("rank")),V.push(...N(r));let we=j=>{let Q=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"pad",type:"i32",length:2},{name:"stride",type:"i32",length:2},{name:"dilation",type:"i32",length:2}];Ye(t,Q);let ie=g?4:1,te=ye(e[0].dataType),be=`
+      fn setOutputAtIndex(flatIndex : i32, value : ${g?`vec4<${te}>`:te}) {
+        result[flatIndex] = ${g?`vec4<${te}>`:te}(value);
       }
-      fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${g?`vec4<${ne}>`:ne}) {
+      fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${g?`vec4<${te}>`:te}) {
         let flatIndex = getOutputIndexFromCoords(vec4<i32>(d0, d1, d2, d3));
         setOutputAtIndex(flatIndex ${g?"/ 4":""}, value);
-      }`,Oe=E("x",e[0].dataType,e[0].dims.length,A===3?1:A),$e=E("w",e[1].dataType,e[1].dims.length,J),le=[Oe,$e],W=M("result",e[0].dataType,r.length,J);if(a){let q=E("bias",e[2].dataType,e[2].dims.length,J);le.push(q),be+=`
-        fn getBiasByOutputCoords(coords : vec4<i32>) -> ${g?`vec4<${ne}>`:ne} {
+      }`,Oe=E("x",e[0].dataType,e[0].dims.length,A===3?1:A),ve=E("w",e[1].dataType,e[1].dims.length,ie),de=[Oe,ve],W=M("result",e[0].dataType,r.length,ie);if(a){let Y=E("bias",e[2].dataType,e[2].dims.length,ie);de.push(Y),be+=`
+        fn getBiasByOutputCoords(coords : vec4<i32>) -> ${g?`vec4<${te}>`:te} {
           return bias[coords.${p?"w":"y"}${g?"/ 4":""}];
         }`}return`
-        ${Ru("uniforms.result_strides")}
+        ${Mu("uniforms.result_strides")}
         //struct Uniforms { xShape : vec4<i32>, wShape : vec4<i32>, outShape : vec4<i32>,
         //  outShapeStrides: vec3<i32>, filterDims : vec2<i32>, pad : vec2<i32>, stride : vec2<i32>,
         //  dilation : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32 };
-        ${Y.registerUniforms(ee).declareVariables(...le,W)}
+        ${j.registerUniforms(Q).declareVariables(...de,W)}
         ${be}
-        ${df(p,R,G,K,a,t,j[0],j[1],j[2],ne)}
-        ${g?ho(S,v,ne,void 0,!p,D):go(S,v,ne,void 0,!p,D,!1,void 0,d)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${A};${g};${R};${G};${K};${k};${P};${D}`,inputDependencies:Q},getRunData:()=>({outputs:[{dims:l?l(r):r,dataType:e[0].dataType}],dispatchGroup:{x:T[0],y:T[1],z:T[2]},programUniforms:V}),getShaderSource:se}}});var lf,Lu,Jr,cf,Gu,pf,Hu,Fu,qu=U(()=>{"use strict";te();et();oe();ae();yt();Yr();lf=e=>{let t=1;for(let r=0;r<e.length;r++)t*=e[r];return t},Lu=e=>typeof e=="number"?[e,e,e]:e,Jr=(e,t)=>t<=1?e:e+(e-1)*(t-1),cf=(e,t,r,n=1)=>{let o=Jr(t,n);return Math.floor((e[0]*(r-1)-r+o)/2)},Gu=(e,t,r,n,o)=>{o==null&&(o=cf(e,t[0],n[0]));let i=[0,0,0,r];for(let a=0;a<3;a++)e[a]+2*o>=t[a]&&(i[a]=Math.trunc((e[a]-t[a]+2*o)/n[a]+1));return i},pf=(e,t,r,n,o,i,a,d,l,p)=>{let m,u,h,_;if(e==="VALID"&&(e=0),typeof e=="number"){m={top:e,bottom:e,left:e,right:e,front:e,back:e};let y=Gu([t,r,n,1],[d,l,p],1,[o,i,a],e);u=y[0],h=y[1],_=y[2]}else if(Array.isArray(e)){if(!e.every((g,x,$)=>g===$[0]))throw Error(`Unsupported padding parameter: ${e}`);m={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let y=Gu([t,r,n,1],[d,l,p],1,[o,i,a],e[0]);u=y[0],h=y[1],_=y[2]}else if(e==="SAME_UPPER"){u=Math.ceil(t/o),h=Math.ceil(r/i),_=Math.ceil(n/a);let y=(u-1)*o+d-t,g=(h-1)*i+l-r,x=(_-1)*a+p-n,$=Math.floor(y/2),v=y-$,S=Math.floor(g/2),T=g-S,A=Math.floor(x/2),k=x-A;m={top:S,bottom:T,left:A,right:k,front:$,back:v}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:m,outDepth:u,outHeight:h,outWidth:_}},Hu=(e,t,r,n,o,i=!1,a="channelsLast")=>{let d,l,p,m,u;if(a==="channelsLast")[d,l,p,m,u]=e;else if(a==="channelsFirst")[d,u,l,p,m]=e;else throw new Error(`Unknown dataFormat ${a}`);let[h,,_,y,g]=t,[x,$,v]=Lu(r),[S,T,A]=Lu(n),k=Jr(_,S),P=Jr(y,T),D=Jr(g,A),{padInfo:R,outDepth:G,outHeight:K,outWidth:j}=pf(o,l,p,m,x,$,v,k,P,D),V=i?h*u:h,Q=[0,0,0,0,0];return a==="channelsFirst"?Q=[d,V,G,K,j]:a==="channelsLast"&&(Q=[d,G,K,j,V]),{batchSize:d,dataFormat:a,inDepth:l,inHeight:p,inWidth:m,inChannels:u,outDepth:G,outHeight:K,outWidth:j,outChannels:V,padInfo:R,strideDepth:x,strideHeight:$,strideWidth:v,filterDepth:_,filterHeight:y,filterWidth:g,effectiveFilterDepth:k,effectiveFilterHeight:P,effectiveFilterWidth:D,dilationDepth:S,dilationHeight:T,dilationWidth:A,inShape:e,outShape:Q,filterShape:t}},Fu=(e,t,r,n,o,i)=>{let a=i==="channelsLast",d=a?e[0].dims[3]:e[0].dims[1],l=!1,p=[64,1,1],m={x:r.map((v,S)=>S)},u=[Math.ceil(lf(m.x.map(v=>r[v]))/p[0]),1,1];ue("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${u}`);let h=l?a&&d%4!==0?3:4:1,_=C.size(r),y=[{type:12,data:_},{type:12,data:n},{type:12,data:o},{type:12,data:t.strides},{type:12,data:t.dilations}];Ke(t,y),y.push(...N(e[0].dims,e[1].dims));let g=["rank","rank"],x=e.length===3;x&&(y.push(...N(e[2].dims)),g.push("rank")),y.push(...N(r));let $=v=>{let S=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:n.length},{name:"pads",type:"u32",length:o.length},{name:"strides",type:"u32",length:t.strides.length},{name:"dilations",type:"u32",length:t.dilations.length}];je(t,S);let T=l?4:1,A=_e(e[0].dataType),k=E("x",e[0].dataType,e[0].dims.length,h===3?1:h),P=E("W",e[1].dataType,e[1].dims.length,T),D=[k,P],R=M("result",e[0].dataType,r.length,T),G="";if(x){let V=E("bias",e[2].dataType,e[2].dims.length,T);D.push(V),G+=`
+        ${df(p,R,H,L,a,t,re[0],re[1],re[2],te)}
+        ${g?fo(S,v,te,void 0,!p,D):ho(S,v,te,void 0,!p,D,!1,void 0,d)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${A};${g};${R};${H};${L};${C};${P};${D}`,inputDependencies:K},getRunData:()=>({outputs:[{dims:l?l(r):r,dataType:e[0].dataType}],dispatchGroup:{x:T[0],y:T[1],z:T[2]},programUniforms:V}),getShaderSource:we}}});var lf,Wu,Xr,cf,Lu,pf,Gu,Hu,Fu=U(()=>{"use strict";J();et();ne();ae();yt();jr();lf=e=>{let t=1;for(let r=0;r<e.length;r++)t*=e[r];return t},Wu=e=>typeof e=="number"?[e,e,e]:e,Xr=(e,t)=>t<=1?e:e+(e-1)*(t-1),cf=(e,t,r,n=1)=>{let o=Xr(t,n);return Math.floor((e[0]*(r-1)-r+o)/2)},Lu=(e,t,r,n,o)=>{o==null&&(o=cf(e,t[0],n[0]));let i=[0,0,0,r];for(let a=0;a<3;a++)e[a]+2*o>=t[a]&&(i[a]=Math.trunc((e[a]-t[a]+2*o)/n[a]+1));return i},pf=(e,t,r,n,o,i,a,d,l,p)=>{let m,u,h,_;if(e==="VALID"&&(e=0),typeof e=="number"){m={top:e,bottom:e,left:e,right:e,front:e,back:e};let y=Lu([t,r,n,1],[d,l,p],1,[o,i,a],e);u=y[0],h=y[1],_=y[2]}else if(Array.isArray(e)){if(!e.every((g,x,$)=>g===$[0]))throw Error(`Unsupported padding parameter: ${e}`);m={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let y=Lu([t,r,n,1],[d,l,p],1,[o,i,a],e[0]);u=y[0],h=y[1],_=y[2]}else if(e==="SAME_UPPER"){u=Math.ceil(t/o),h=Math.ceil(r/i),_=Math.ceil(n/a);let y=(u-1)*o+d-t,g=(h-1)*i+l-r,x=(_-1)*a+p-n,$=Math.floor(y/2),v=y-$,S=Math.floor(g/2),T=g-S,A=Math.floor(x/2),C=x-A;m={top:S,bottom:T,left:A,right:C,front:$,back:v}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:m,outDepth:u,outHeight:h,outWidth:_}},Gu=(e,t,r,n,o,i=!1,a="channelsLast")=>{let d,l,p,m,u;if(a==="channelsLast")[d,l,p,m,u]=e;else if(a==="channelsFirst")[d,u,l,p,m]=e;else throw new Error(`Unknown dataFormat ${a}`);let[h,,_,y,g]=t,[x,$,v]=Wu(r),[S,T,A]=Wu(n),C=Xr(_,S),P=Xr(y,T),D=Xr(g,A),{padInfo:R,outDepth:H,outHeight:L,outWidth:re}=pf(o,l,p,m,x,$,v,C,P,D),V=i?h*u:h,K=[0,0,0,0,0];return a==="channelsFirst"?K=[d,V,H,L,re]:a==="channelsLast"&&(K=[d,H,L,re,V]),{batchSize:d,dataFormat:a,inDepth:l,inHeight:p,inWidth:m,inChannels:u,outDepth:H,outHeight:L,outWidth:re,outChannels:V,padInfo:R,strideDepth:x,strideHeight:$,strideWidth:v,filterDepth:_,filterHeight:y,filterWidth:g,effectiveFilterDepth:C,effectiveFilterHeight:P,effectiveFilterWidth:D,dilationDepth:S,dilationHeight:T,dilationWidth:A,inShape:e,outShape:K,filterShape:t}},Hu=(e,t,r,n,o,i)=>{let a=i==="channelsLast",d=a?e[0].dims[3]:e[0].dims[1],l=!1,p=[64,1,1],m={x:r.map((v,S)=>S)},u=[Math.ceil(lf(m.x.map(v=>r[v]))/p[0]),1,1];se("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${u}`);let h=l?a&&d%4!==0?3:4:1,_=k.size(r),y=[{type:12,data:_},{type:12,data:n},{type:12,data:o},{type:12,data:t.strides},{type:12,data:t.dilations}];je(t,y),y.push(...N(e[0].dims,e[1].dims));let g=["rank","rank"],x=e.length===3;x&&(y.push(...N(e[2].dims)),g.push("rank")),y.push(...N(r));let $=v=>{let S=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:n.length},{name:"pads",type:"u32",length:o.length},{name:"strides",type:"u32",length:t.strides.length},{name:"dilations",type:"u32",length:t.dilations.length}];Ye(t,S);let T=l?4:1,A=ye(e[0].dataType),C=E("x",e[0].dataType,e[0].dims.length,h===3?1:h),P=E("W",e[1].dataType,e[1].dims.length,T),D=[C,P],R=M("result",e[0].dataType,r.length,T),H="";if(x){let V=E("bias",e[2].dataType,e[2].dims.length,T);D.push(V),H+=`
         fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${l?`vec4<${A}>`:A} {
-          return bias[${a?F("coords",4,5):F("coords",1,5)}${l?"/ 4":""}];
-        }`}let K=Ae(h,A),j=qe(t,K,A);return`
-            ${G}
+          return bias[${a?q("coords",4,5):q("coords",1,5)}${l?"/ 4":""}];
+        }`}let L=Ae(h,A),re=Ke(t,L,A);return`
+            ${H}
             fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
               let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
-              return ${k.getByIndices("aIndices")};
+              return ${C.getByIndices("aIndices")};
             }
             fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
               let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
@@ -4922,18 +4977,18 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
           ${v.mainStart()}
           ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
               let coords = ${R.offsetToIndices("global_idx")};
-              let batch = ${F("coords",0,k.rank)};
-              let d2 = ${a?F("coords",k.rank-1,k.rank):F("coords",1,k.rank)};
-              let xFRCCorner = vec3<u32>(${a?F("coords",1,k.rank):F("coords",2,k.rank)},
-              ${a?F("coords",2,k.rank):F("coords",3,k.rank)},
-              ${a?F("coords",3,k.rank):F("coords",4,k.rank)}) * uniforms.strides - uniforms.pads;
+              let batch = ${q("coords",0,C.rank)};
+              let d2 = ${a?q("coords",C.rank-1,C.rank):q("coords",1,C.rank)};
+              let xFRCCorner = vec3<u32>(${a?q("coords",1,C.rank):q("coords",2,C.rank)},
+              ${a?q("coords",2,C.rank):q("coords",3,C.rank)},
+              ${a?q("coords",3,C.rank):q("coords",4,C.rank)}) * uniforms.strides - uniforms.pads;
               let xFCorner = xFRCCorner.x;
               let xRCorner = xFRCCorner.y;
               let xCCorner = xFRCCorner.z;
-              let xShapeY = ${a?F("uniforms.x_shape",1,k.rank):F("uniforms.x_shape",2,k.rank)};
-              let xShapeZ = ${a?F("uniforms.x_shape",2,k.rank):F("uniforms.x_shape",3,k.rank)};
-              let xShapeW = ${a?F("uniforms.x_shape",3,k.rank):F("uniforms.x_shape",4,k.rank)};
-              let xShapeU = ${a?F("uniforms.x_shape",4,k.rank):F("uniforms.x_shape",1,k.rank)};
+              let xShapeY = ${a?q("uniforms.x_shape",1,C.rank):q("uniforms.x_shape",2,C.rank)};
+              let xShapeZ = ${a?q("uniforms.x_shape",2,C.rank):q("uniforms.x_shape",3,C.rank)};
+              let xShapeW = ${a?q("uniforms.x_shape",3,C.rank):q("uniforms.x_shape",4,C.rank)};
+              let xShapeU = ${a?q("uniforms.x_shape",4,C.rank):q("uniforms.x_shape",1,C.rank)};
               let inputDepthNearestVec4 = (xShapeU / 4) * 4;
               let inputDepthVec4Remainder = xShapeU % 4;
 
@@ -5011,9 +5066,9 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
                 }
               }
               ${x?"value = value + getBiasByOutputCoords(coords)":""};
-              ${j}
+              ${re}
               result[global_idx] = f32(value);
-          }`};return{name:"Conv3DNaive",shaderCache:{hint:`${t.cacheKey};${a};${h};${x}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:u[0],y:u[1],z:u[2]},programUniforms:y}),getShaderSource:$}}});var Ku,ju,Yu=U(()=>{"use strict";te();oe();ae();yt();Ku=(e,t,r,n)=>{let o=e.length>2,i=o?"value += b[output_channel];":"",a=e[0].dims,d=e[1].dims,l=t.format==="NHWC",p=l?r[3]:r[1],m=p/t.group,u=l&&m>=4?me(p):1,h=C.size(r)/u,_=[{type:12,data:h},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:m}];Ke(t,_),_.push(...N(a,[d[0],d[1],d[2],d[3]/u]));let y=o?["rank","rank","rank"]:["rank","rank"];_.push(...N([r[0],r[1],r[2],r[3]/u]));let g=x=>{let $=M("output",e[0].dataType,r.length,u),v=_e($.type.tensor),S=qe(t,$.type.value,v),T=E("x",e[0].dataType,a.length),A=E("w",e[1].dataType,d.length,u),k=[T,A];o&&k.push(E("b",e[2].dataType,e[2].dims,u));let P=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:t.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];je(t,P);let D=l?`
+          }`};return{name:"Conv3DNaive",shaderCache:{hint:`${t.cacheKey};${a};${h};${x}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:u[0],y:u[1],z:u[2]},programUniforms:y}),getShaderSource:$}}});var qu,Ku,ju=U(()=>{"use strict";J();ne();ae();yt();qu=(e,t,r,n)=>{let o=e.length>2,i=o?"value += b[output_channel];":"",a=e[0].dims,d=e[1].dims,l=t.format==="NHWC",p=l?r[3]:r[1],m=p/t.group,u=l&&m>=4?me(p):1,h=k.size(r)/u,_=[{type:12,data:h},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:m}];je(t,_),_.push(...N(a,[d[0],d[1],d[2],d[3]/u]));let y=o?["rank","rank","rank"]:["rank","rank"];_.push(...N([r[0],r[1],r[2],r[3]/u]));let g=x=>{let $=M("output",e[0].dataType,r.length,u),v=ye($.type.tensor),S=Ke(t,$.type.value,v),T=E("x",e[0].dataType,a.length),A=E("w",e[1].dataType,d.length,u),C=[T,A];o&&C.push(E("b",e[2].dataType,e[2].dims,u));let P=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:t.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];Ye(t,P);let D=l?`
       for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[0]; wHeight++) {
         let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];
 
@@ -5058,7 +5113,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
         }
       }
       `;return`
-  ${x.registerUniforms(P).declareVariables(...k,$)}
+  ${x.registerUniforms(P).declareVariables(...C,$)}
 
   ${x.mainStart()}
     ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -5075,8 +5130,8 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     ${i}
     ${S}
     ${$.setByOffset("global_idx","value")}
-  }`};return{name:"GroupedConv",shaderCache:{hint:`${t.cacheKey}_${u}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:n?n(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:_}),getShaderSource:g}},ju=(e,t,r,n)=>{let o=e.length>2,i=me(r[3]),a=me(r[2]),d=C.size(r)/i/a,l=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/i],p=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/i],m=[r[0],r[1],r[2],r[3]/i],u=[{type:12,data:d},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];Ke(t,u),u.push(...N(l,p,m));let h=(a-1)*t.strides[1]+p[1],_=y=>{let g=M("output",e[0].dataType,m.length,i),x=_e(g.type.tensor),$=qe(t,g.type.value,x),v=E("x",e[0].dataType,l.length,i),S=E("w",e[1].dataType,p.length,i),T=[v,S];o&&T.push(E("b",e[2].dataType,e[2].dims,i));let A=o?"value += b[output_channel];":"",k=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return je(t,k),`
-  ${y.registerUniforms(k).declareVariables(...T,g)}
+  }`};return{name:"GroupedConv",shaderCache:{hint:`${t.cacheKey}_${u}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:n?n(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:_}),getShaderSource:g}},Ku=(e,t,r,n)=>{let o=e.length>2,i=me(r[3]),a=me(r[2]),d=k.size(r)/i/a,l=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/i],p=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/i],m=[r[0],r[1],r[2],r[3]/i],u=[{type:12,data:d},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];je(t,u),u.push(...N(l,p,m));let h=(a-1)*t.strides[1]+p[1],_=y=>{let g=M("output",e[0].dataType,m.length,i),x=ye(g.type.tensor),$=Ke(t,g.type.value,x),v=E("x",e[0].dataType,l.length,i),S=E("w",e[1].dataType,p.length,i),T=[v,S];o&&T.push(E("b",e[2].dataType,e[2].dims,i));let A=o?"value += b[output_channel];":"",C=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Ye(t,C),`
+  ${y.registerUniforms(C).declareVariables(...T,g)}
   ${y.mainStart()}
     ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
     let width0 = uniforms.output_shape[3];
@@ -5120,19 +5175,19 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       ${$}
       ${g.set("batch","row","col + i","output_channel","value")};
     }
-  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${i};${a};${h};${p[0]};${p[1]}`,inputDependencies:o?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:n?n(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:u}),getShaderSource:_}}});var mf,bo,ff,yo,_o,Zu,hf,gf,wo,Qu=U(()=>{"use strict";oe();Wu();qu();Xr();Yu();yt();Qr();dt();mf=(e,t,r,n,o,i)=>{let a=e[0],d=e.slice(i?1:2,i?3:4),l=d.length,p=t[0],u=t.slice(2).map((y,g)=>y+(y-1)*(r[g]-1)),_=d.map((y,g)=>y+n[g]+n[g+l]).map((y,g)=>Math.floor((y-u[g]+o[g])/o[g]));return _.splice(0,0,a),_.splice(i?3:1,0,p),_},bo=[2,3,1,0],ff=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length>5)throw new Error("greater than 5D is not supported");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],n=e[1].dims[1]*t.group;if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let o=e[0].dims.length-2;if(t.dilations.length!==o)throw new Error(`dilations should be ${o}D`);if(t.strides.length!==o)throw new Error(`strides should be ${o}D`);if(t.pads.length!==o*2)throw new Error(`pads should be ${o*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},yo=(e,t)=>{let r=e.kernelShape.slice();r.length<t[1].dims.length-2&&r.push(...Array(t[1].dims.length-2-r.length).fill(0));for(let i=2;i<t[1].dims.length;++i)r[i-2]===0&&(r[i-2]=t[1].dims[i]);let n=e.pads.slice();At.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,n,e.format==="NHWC",e.autoPad);let o=Object.assign({},e);return Object.assign(o,{kernelShape:r,pads:n}),o},_o=e=>{let t=jr(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],o=e.dilations,i=e.group,a=e.kernel_shape,d=e.pads,l=e.strides,p=e.w_is_const();return{autoPad:n,format:r,dilations:o,group:i,kernelShape:a,pads:d,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Zu=(e,t,r,n)=>{let o=r.format==="NHWC",i=mf(t[0].dims,t[1].dims,r.dilations,r.pads,r.strides,o);if(r.group!==1){let k=[t[0]];if(o){let D=e.kernelCustomData.wT??e.compute(Pe(t[1],bo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=D),k.push(D)}else k.push(t[1]);t.length===3&&k.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&o&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1?e.compute(ju(k,r,i,n),{inputs:k}):e.compute(Ku(k,r,i,n),{inputs:k});return}let a=t.length===3,d=t[0].dims[o?1:2],l=t[0].dims[o?2:3],p=t[0].dims[o?3:1],m=t[1].dims[2],u=t[1].dims[3],h=i[o?1:2],_=i[o?2:3],y=i[o?3:1],g=o&&m===d&&u===l&&r.pads[0]===0&&r.pads[1]===0;if(g||m===1&&u===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let k=i[0],P,D,R,G=[];if(o){let V=e.kernelCustomData.wT??e.compute(Pe(t[1],bo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=V),g){let Q=d*l*p;P=t[0].reshape([1,k,Q]),D=V.reshape([1,Q,y]),R=[1,k,y]}else P=t[0].reshape([k,d*l,p]),D=V.reshape([1,p,y]),R=[k,h*_,y];G.push(P),G.push(D)}else P=t[0].reshape([k,p,d*l]),D=t[1].reshape([1,y,p]),R=[k,y,h*_],G.push(D),G.push(P);a&&G.push(t[2]);let K=R[2],j=G[0].dims[G[0].dims.length-1];K<8&&j<8?e.compute(Zr(G,r,i,R,o,n),{inputs:G}):e.compute(er(G,r,i,R,o,n),{inputs:G});return}let x=!0,$=e.kernelCustomData.wT??e.compute(Pe(t[1],bo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=$);let v=[t[0],$];a&&v.push(t[2]);let S=o?h*_:y,T=o?y:h*_,A=m*u*p;e.compute(Vu(v,r,i,S,T,A,a,x,n),{inputs:v})},hf=(e,t)=>{let r=t.format==="NHWC",n=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&n.push(e.inputs[2]);let o=[0,t.pads[0],0,t.pads[1]],i=[1].concat(t.strides),a=[1].concat(t.dilations),d=[1].concat(t.kernelShape),l=yo({...t,pads:o,strides:i,dilations:a,kernelShape:d},n);Zu(e,n,l,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]])},gf=(e,t,r)=>{let n=r.format==="NHWC"?"channelsLast":"channelsFirst",o=yo(r,t),i=r.autoPad==="NOTSET"?r.pads:r.autoPad,a=Hu(t[0].dims,t[1].dims,r.strides,r.dilations,i,!1,n);e.compute(Fu(t,o,a.outShape,[a.filterDepth,a.filterHeight,a.filterWidth],[a.padInfo.front,a.padInfo.top,a.padInfo.left],n))},wo=(e,t)=>{if(ff(e.inputs,t),e.inputs[0].dims.length===3)hf(e,t);else if(e.inputs[0].dims.length===5)gf(e,e.inputs,t);else{let r=yo(t,e.inputs);Zu(e,e.inputs,r)}}});var Xu,Ju=U(()=>{"use strict";te();et();oe();ae();Xu=(e,t,r)=>{let n=e.length>2,o=t.outputShape,i=t.format==="NHWC",a=t.group,d=e[1].dims,l=d[2]/a,p=d[3],m=i?me(l):1,u=i?me(p):1,h=i?p===1?m:u:1,_=C.size(o)/u,y=[Math.ceil(_/64),1,1];ue("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${y}`);let g=["rank","rank"],x=[t.strides[0],t.strides[1]],$=[t.kernelShape[i?1:2],t.kernelShape[i?2:3]],v=[t.dilations[0],t.dilations[1]],S=[$[0]+(t.dilations[0]<=1?0:(t.kernelShape[i?1:2]-1)*(t.dilations[0]-1)),$[1]+(t.dilations[1]<=1?0:(t.kernelShape[i?2:3]-1)*(t.dilations[1]-1))],T=[S[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),S[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],A=[{type:12,data:_},{type:12,data:x},{type:12,data:$},{type:12,data:v},{type:12,data:S},{type:6,data:T},{type:12,data:l},{type:12,data:p},...N(e[0].dims,e[1].dims)];n&&(A.push(...N(e[2].dims)),g.push("rank")),A.push(...N(o));let k=P=>{let D=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:x.length},{name:"filter_dims",type:"u32",length:$.length},{name:"dilations",type:"u32",length:$.length},{name:"effective_filter_dims",type:"u32",length:S.length},{name:"pads",type:"i32",length:T.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],R=_e(e[0].dataType),G=i?1:2,K=i?2:3,j=i?3:1,V=E("W",e[1].dataType,e[1].dims.length,h),Q=E("Dy",e[0].dataType,e[0].dims.length,m),se=[Q,V];n&&se.push(E("bias",e[2].dataType,[o[j]].length,u));let Y=M("result",e[0].dataType,o.length,u),ee=()=>{let ne="";if(m===1)ne+=`
+  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${i};${a};${h};${p[0]};${p[1]}`,inputDependencies:o?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:n?n(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:u}),getShaderSource:_}}});var mf,go,ff,bo,yo,Yu,hf,gf,_o,Zu=U(()=>{"use strict";ne();Vu();Fu();Qr();ju();yt();Zr();dt();mf=(e,t,r,n,o,i)=>{let a=e[0],d=e.slice(i?1:2,i?3:4),l=d.length,p=t[0],u=t.slice(2).map((y,g)=>y+(y-1)*(r[g]-1)),_=d.map((y,g)=>y+n[g]+n[g+l]).map((y,g)=>Math.floor((y-u[g]+o[g])/o[g]));return _.splice(0,0,a),_.splice(i?3:1,0,p),_},go=[2,3,1,0],ff=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length>5)throw new Error("greater than 5D is not supported");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],n=e[1].dims[1]*t.group;if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let o=e[0].dims.length-2;if(t.dilations.length!==o)throw new Error(`dilations should be ${o}D`);if(t.strides.length!==o)throw new Error(`strides should be ${o}D`);if(t.pads.length!==o*2)throw new Error(`pads should be ${o*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},bo=(e,t)=>{let r=e.kernelShape.slice();r.length<t[1].dims.length-2&&r.push(...Array(t[1].dims.length-2-r.length).fill(0));for(let i=2;i<t[1].dims.length;++i)r[i-2]===0&&(r[i-2]=t[1].dims[i]);let n=e.pads.slice();kt.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,n,e.format==="NHWC",e.autoPad);let o=Object.assign({},e);return Object.assign(o,{kernelShape:r,pads:n}),o},yo=e=>{let t=Kr(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],o=e.dilations,i=e.group,a=e.kernel_shape,d=e.pads,l=e.strides,p=e.w_is_const();return{autoPad:n,format:r,dilations:o,group:i,kernelShape:a,pads:d,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Yu=(e,t,r,n)=>{let o=r.format==="NHWC",i=mf(t[0].dims,t[1].dims,r.dilations,r.pads,r.strides,o);if(r.group!==1){let C=[t[0]];if(o){let D=e.kernelCustomData.wT??e.compute(Pe(t[1],go),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=D),C.push(D)}else C.push(t[1]);t.length===3&&C.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&o&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1?e.compute(Ku(C,r,i,n),{inputs:C}):e.compute(qu(C,r,i,n),{inputs:C});return}let a=t.length===3,d=t[0].dims[o?1:2],l=t[0].dims[o?2:3],p=t[0].dims[o?3:1],m=t[1].dims[2],u=t[1].dims[3],h=i[o?1:2],_=i[o?2:3],y=i[o?3:1],g=o&&m===d&&u===l&&r.pads[0]===0&&r.pads[1]===0;if(g||m===1&&u===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let C=i[0],P,D,R,H=[];if(o){let V=e.kernelCustomData.wT??e.compute(Pe(t[1],go),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=V),g){let K=d*l*p;P=t[0].reshape([1,C,K]),D=V.reshape([1,K,y]),R=[1,C,y]}else P=t[0].reshape([C,d*l,p]),D=V.reshape([1,p,y]),R=[C,h*_,y];H.push(P),H.push(D)}else P=t[0].reshape([C,p,d*l]),D=t[1].reshape([1,y,p]),R=[C,y,h*_],H.push(D),H.push(P);a&&H.push(t[2]);let L=R[2],re=H[0].dims[H[0].dims.length-1];L<8&&re<8?e.compute(Yr(H,r,i,R,o,n),{inputs:H}):e.compute(Jt(H,r,i,R,o,n),{inputs:H});return}let x=!0,$=e.kernelCustomData.wT??e.compute(Pe(t[1],go),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=$);let v=[t[0],$];a&&v.push(t[2]);let S=o?h*_:y,T=o?y:h*_,A=m*u*p;e.compute(Nu(v,r,i,S,T,A,a,x,n),{inputs:v})},hf=(e,t)=>{let r=t.format==="NHWC",n=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&n.push(e.inputs[2]);let o=[0,t.pads[0],0,t.pads[1]],i=[1].concat(t.strides),a=[1].concat(t.dilations),d=[1].concat(t.kernelShape),l=bo({...t,pads:o,strides:i,dilations:a,kernelShape:d},n);Yu(e,n,l,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]])},gf=(e,t,r)=>{let n=r.format==="NHWC"?"channelsLast":"channelsFirst",o=bo(r,t),i=r.autoPad==="NOTSET"?r.pads:r.autoPad,a=Gu(t[0].dims,t[1].dims,r.strides,r.dilations,i,!1,n);e.compute(Hu(t,o,a.outShape,[a.filterDepth,a.filterHeight,a.filterWidth],[a.padInfo.front,a.padInfo.top,a.padInfo.left],n))},_o=(e,t)=>{if(ff(e.inputs,t),e.inputs[0].dims.length===3)hf(e,t);else if(e.inputs[0].dims.length===5)gf(e,e.inputs,t);else{let r=bo(t,e.inputs);Yu(e,e.inputs,r)}}});var Qu,Xu=U(()=>{"use strict";J();et();ne();ae();Qu=(e,t,r)=>{let n=e.length>2,o=t.outputShape,i=t.format==="NHWC",a=t.group,d=e[1].dims,l=d[2]/a,p=d[3],m=i?me(l):1,u=i?me(p):1,h=i?p===1?m:u:1,_=k.size(o)/u,y=[Math.ceil(_/64),1,1];se("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${y}`);let g=["rank","rank"],x=[t.strides[0],t.strides[1]],$=[t.kernelShape[i?1:2],t.kernelShape[i?2:3]],v=[t.dilations[0],t.dilations[1]],S=[$[0]+(t.dilations[0]<=1?0:(t.kernelShape[i?1:2]-1)*(t.dilations[0]-1)),$[1]+(t.dilations[1]<=1?0:(t.kernelShape[i?2:3]-1)*(t.dilations[1]-1))],T=[S[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),S[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],A=[{type:12,data:_},{type:12,data:x},{type:12,data:$},{type:12,data:v},{type:12,data:S},{type:6,data:T},{type:12,data:l},{type:12,data:p},...N(e[0].dims,e[1].dims)];n&&(A.push(...N(e[2].dims)),g.push("rank")),A.push(...N(o));let C=P=>{let D=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:x.length},{name:"filter_dims",type:"u32",length:$.length},{name:"dilations",type:"u32",length:$.length},{name:"effective_filter_dims",type:"u32",length:S.length},{name:"pads",type:"i32",length:T.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],R=ye(e[0].dataType),H=i?1:2,L=i?2:3,re=i?3:1,V=E("W",e[1].dataType,e[1].dims.length,h),K=E("Dy",e[0].dataType,e[0].dims.length,m),we=[K,V];n&&we.push(E("bias",e[2].dataType,[o[re]].length,u));let j=M("result",e[0].dataType,o.length,u),Q=()=>{let te="";if(m===1)te+=`
         let w_offset = ${V.indicesToOffset(`${V.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
         let wValue = ${V.getByOffset(`w_offset / ${h}`)};
-        dotProd = dotProd + xValue * wValue;`;else if(p===1)ne+=`
+        dotProd = dotProd + xValue * wValue;`;else if(p===1)te+=`
           let wValue = ${V.getByOffset(`${V.indicesToOffset(`${V.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)} / ${h}`)};
-          dotProd = dotProd + dot(xValue, wValue);`;else for(let be=0;be<m;be++)ne+=`
+          dotProd = dotProd + dot(xValue, wValue);`;else for(let be=0;be<m;be++)te+=`
             let wValue${be} = ${V.getByOffset(`${V.indicesToOffset(`${V.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${be}, wOutChannel)`)} / ${h}`)};
-            dotProd = dotProd + xValue[${be}] * wValue${be};`;return ne},J=`
-            let outputIndices = ${Y.offsetToIndices(`global_idx * ${u}`)};
-            let batch = ${Y.indicesGet("outputIndices",0)};
-            let d1 = ${Y.indicesGet("outputIndices",j)};
-            let r = ${Y.indicesGet("outputIndices",G)};
-            let c = ${Y.indicesGet("outputIndices",K)};
+            dotProd = dotProd + xValue[${be}] * wValue${be};`;return te},ie=`
+            let outputIndices = ${j.offsetToIndices(`global_idx * ${u}`)};
+            let batch = ${j.indicesGet("outputIndices",0)};
+            let d1 = ${j.indicesGet("outputIndices",re)};
+            let r = ${j.indicesGet("outputIndices",H)};
+            let c = ${j.indicesGet("outputIndices",L)};
             let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
             let dyRCorner = dyCorner.x;
             let dyCCorner = dyCorner.y;
@@ -5140,47 +5195,57 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
             let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
             // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
             // ? = to be determined. : = across all values in that axis.
-            var dotProd = ${Y.type.value}(0.0);
-            for (var wR: u32 = 0; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
+            var dotProd = ${j.type.value}(0.0);
+            var wR: u32 = 0;
+            if (uniforms.dilations.x == 1) {
+              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
+              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
+            }
+            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
               if (wR % uniforms.dilations.x != 0) {
                 continue;
               }
               let dyR = (${R}(dyRCorner) + ${R}(wR)) / ${R}(uniforms.strides[0]);
               let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
-              if (dyR < 0.0 || dyR >= ${R}(uniforms.Dy_shape[${G}]) || fract(dyR) > 0.0 ||
+              if (dyR < 0.0 || dyR >= ${R}(uniforms.Dy_shape[${H}]) || fract(dyR) > 0.0 ||
                   wRPerm < 0) {
                 continue;
               }
-              wR = wR + uniforms.strides[0] - 1;
               let idyR: u32 = u32(dyR);
+              var wC: u32 = 0;
+              if (uniforms.dilations.y == 1) {
+                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
+                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
+              }
 
-              for (var wC: u32 = 0; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
+              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
                 if (wC % uniforms.dilations.y != 0) {
                   continue;
                 }
                 let dyC = (${R}(dyCCorner) + ${R}(wC)) / ${R}(uniforms.strides.y);
                 let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
-                if (dyC < 0.0 || dyC >= ${R}(uniforms.Dy_shape[${K}]) ||
+                if (dyC < 0.0 || dyC >= ${R}(uniforms.Dy_shape[${L}]) ||
                     fract(dyC) > 0.0 || wCPerm < 0) {
                   continue;
                 }
-                wC = wC + uniforms.strides.y - 1;
                 let idyC: u32 = u32(dyC);
                 var inputChannel = groupId * uniforms.input_channels_per_group;
                 for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + ${m}) {
-                  let xValue = ${i?Q.getByOffset(`${Q.indicesToOffset(`${Q.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${m}`):Q.get("batch","inputChannel","idyR","idyC")};
-                  ${ee()}
+                  let xValue = ${i?K.getByOffset(`${K.indicesToOffset(`${K.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${m}`):K.get("batch","inputChannel","idyR","idyC")};
+                  ${Q()}
                   inputChannel = inputChannel + ${m};
                 }
+                wC = wC + uniforms.strides.y - 1;
               }
+              wR = wR + uniforms.strides[0] - 1;
             }
             let value = dotProd${n?` + bias[d1 / ${u}]`:""};
-            ${Y.setByOffset("global_idx","value")};
+            ${j.setByOffset("global_idx","value")};
           `;return`
-    ${P.registerUniforms(D).declareVariables(...se,Y)}
+    ${P.registerUniforms(D).declareVariables(...we,j)}
       ${P.mainStart()}
       ${P.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
-    ${J}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${m}${h}${u}${p===1}`,inputDependencies:g},getRunData:()=>({dispatchGroup:{x:y[0],y:y[1],z:y[2]},outputs:[{dims:r?r(o):o,dataType:e[0].dataType}],programUniforms:A}),getShaderSource:k}}});var bf,yf,_f,ed,td,wf,rd,vf,nd,od=U(()=>{"use strict";Ju();yt();dt();bf=(e,t,r,n,o,i)=>(e-1)*t+r+(n-1)*o+1-i,yf=(e,t,r,n,o)=>{let i=Math.floor(e/2);t==="SAME_UPPER"?(r[n]=i,r[o]=e-i):t==="SAME_LOWER"&&(r[n]=e-i,r[o]=i)},_f=(e,t,r,n,o,i,a,d,l,p)=>{let m=e.length-2,u=p.length===0;l.length<m&&l.push(...Array(m-l.length).fill(0));let h=e[0],_=t[d?3:1]*o;for(let y=0,g=e.length-m-(d?1:0);y<m;++y,++g){let x=e[g],$=u?x*a[y]:p[y],v=bf(x,a[y],i[y],t[g],r[y],$);yf(v,n,i,y,y+m),u&&p.push(a[y]*(x-1)+l[y]+(t[g]-1)*r[y]+1-i[y]-i[y+m])}p.splice(0,0,h),p.splice(d?3:1,0,_)},ed=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((u,h)=>u*h,1)===0){r.length=0;for(let u=2;u<t[1].dims.length;++u)r.push(t[1].dims[u])}let n=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(n?3:1,0,t[1].dims[1]);let o=e.pads.slice(),i=e.outputShape.slice(),a=e.outputPadding.slice(),d=t[0].dims,l=e.dilations.slice();if(l.reduce((u,h)=>u+h,0)===0){let u=t[0].dims.length-2;l=new Array(u).fill(1)}let p=e.strides.slice();if(p.reduce((u,h)=>u+h,0)===0){let u=t[0].dims.length-2;p=new Array(u).fill(1)}_f(d,r,l,e.autoPad,e.group,o,p,n,a,i);let m=Object.assign({},e);return Object.assign(m,{kernelShape:r,pads:o,outputPadding:a,outputShape:i,dilations:l,strides:p}),m},td=e=>{let t=jr(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],o=e.dilations,i=e.group,a=e.kernelShape,d=e.pads,l=e.strides,p=e.wIsConst(),m=e.outputPadding,u=e.outputShape;return{autoPad:n,format:r,dilations:o,group:i,kernelShape:a,outputPadding:m,outputShape:u,pads:d,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},wf=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],n=e[1].dims[0];if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let o=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==o))throw new Error("invalid bias");let i=e[0].dims.length-2;if(t.dilations.reduce((m,u)=>m+u,0)>0&&t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.reduce((m,u)=>m+u,0)>0&&t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.reduce((m,u)=>m+u,0)>0&&t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.outputPadding.length!==i&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${i}D`);if(t.kernelShape.reduce((m,u)=>m+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},rd=(e,t,r,n)=>{let o=e.kernelCustomData.wT??e.compute(Pe(t[1],[2,3,0,1]),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=o);let i=[t[0],o];t.length===3&&i.push(t[2]),e.compute(Xu(i,r,n),{inputs:i})},vf=(e,t)=>{let r=t.format==="NHWC",n=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&n.push(e.inputs[2]);let o=t.kernelShape;(o.length===0||o[0]===0)&&(o=[e.inputs[1].dims[2]]);let i=t.dilations;(i.length===0||i[0]===0)&&(i=[1]);let a=t.strides;(a.length===0||a[0]===0)&&(a=[1]);let d=t.pads;d.length===0&&(d=[0,0]),d=[0,d[0],0,d[1]],a=[1].concat(a),i=[1].concat(i),o=[1].concat(o);let l=ed({...t,pads:d,strides:a,dilations:i,kernelShape:o},n);rd(e,n,l,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]])},nd=(e,t)=>{if(wf(e.inputs,t),e.inputs[0].dims.length===3)vf(e,t);else{let r=ed(t,e.inputs);rd(e,e.inputs,r)}}});var $f,id,ad,sd=U(()=>{"use strict";te();oe();Se();ae();$f=(e,t,r,n)=>{let o=C.size(t),i=t.length,a=E("input",e,i),d=M("output",e,i),l=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),p=C.normalizeAxis(l,i),m=u=>{let h=` i32(${a.indicesGet("inputIndices","uniforms.axis")}) `,_=F("uniforms.input_shape","uniforms.axis",i),y=n.reverse?h+(n.exclusive?" + 1":""):"0",g=n.reverse?_:h+(n.exclusive?"":" + 1");return`
+    ${ie}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${m}${h}${u}${p===1}`,inputDependencies:g},getRunData:()=>({dispatchGroup:{x:y[0],y:y[1],z:y[2]},outputs:[{dims:r?r(o):o,dataType:e[0].dataType}],programUniforms:A}),getShaderSource:C}}});var bf,yf,_f,Ju,ed,wf,td,vf,rd,nd=U(()=>{"use strict";Xu();yt();dt();bf=(e,t,r,n,o,i)=>(e-1)*t+r+(n-1)*o+1-i,yf=(e,t,r,n,o)=>{let i=Math.floor(e/2);t==="SAME_UPPER"?(r[n]=i,r[o]=e-i):t==="SAME_LOWER"&&(r[n]=e-i,r[o]=i)},_f=(e,t,r,n,o,i,a,d,l,p)=>{let m=e.length-2,u=p.length===0;l.length<m&&l.push(...Array(m-l.length).fill(0));let h=e[0],_=t[d?3:1]*o;for(let y=0,g=e.length-m-(d?1:0);y<m;++y,++g){let x=e[g],$=u?x*a[y]:p[y],v=bf(x,a[y],i[y],t[g],r[y],$);yf(v,n,i,y,y+m),u&&p.push(a[y]*(x-1)+l[y]+(t[g]-1)*r[y]+1-i[y]-i[y+m])}p.splice(0,0,h),p.splice(d?3:1,0,_)},Ju=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((u,h)=>u*h,1)===0){r.length=0;for(let u=2;u<t[1].dims.length;++u)r.push(t[1].dims[u])}let n=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(n?3:1,0,t[1].dims[1]);let o=e.pads.slice(),i=e.outputShape.slice(),a=e.outputPadding.slice(),d=t[0].dims,l=e.dilations.slice();if(l.reduce((u,h)=>u+h,0)===0){let u=t[0].dims.length-2;l=new Array(u).fill(1)}let p=e.strides.slice();if(p.reduce((u,h)=>u+h,0)===0){let u=t[0].dims.length-2;p=new Array(u).fill(1)}_f(d,r,l,e.autoPad,e.group,o,p,n,a,i);let m=Object.assign({},e);return Object.assign(m,{kernelShape:r,pads:o,outputPadding:a,outputShape:i,dilations:l,strides:p}),m},ed=e=>{let t=Kr(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],o=e.dilations,i=e.group,a=e.kernelShape,d=e.pads,l=e.strides,p=e.wIsConst(),m=e.outputPadding,u=e.outputShape;return{autoPad:n,format:r,dilations:o,group:i,kernelShape:a,outputPadding:m,outputShape:u,pads:d,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},wf=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],n=e[1].dims[0];if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let o=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==o))throw new Error("invalid bias");let i=e[0].dims.length-2;if(t.dilations.reduce((m,u)=>m+u,0)>0&&t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.reduce((m,u)=>m+u,0)>0&&t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.reduce((m,u)=>m+u,0)>0&&t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.outputPadding.length!==i&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${i}D`);if(t.kernelShape.reduce((m,u)=>m+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},td=(e,t,r,n)=>{let o=e.kernelCustomData.wT??e.compute(Pe(t[1],[2,3,0,1]),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=o);let i=[t[0],o];t.length===3&&i.push(t[2]),e.compute(Qu(i,r,n),{inputs:i})},vf=(e,t)=>{let r=t.format==="NHWC",n=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&n.push(e.inputs[2]);let o=t.kernelShape;(o.length===0||o[0]===0)&&(o=[e.inputs[1].dims[2]]);let i=t.dilations;(i.length===0||i[0]===0)&&(i=[1]);let a=t.strides;(a.length===0||a[0]===0)&&(a=[1]);let d=t.pads;d.length===0&&(d=[0,0]),d=[0,d[0],0,d[1]],a=[1].concat(a),i=[1].concat(i),o=[1].concat(o);let l=t.outputPadding;l=[0].concat(l);let p=Ju({...t,pads:d,strides:a,dilations:i,kernelShape:o,outputPadding:l},n);td(e,n,p,m=>r?[m[0],m[2],m[3]]:[m[0],m[1],m[3]])},rd=(e,t)=>{if(wf(e.inputs,t),e.inputs[0].dims.length===3)vf(e,t);else{let r=Ju(t,e.inputs);td(e,e.inputs,r)}}});var $f,od,id,ad=U(()=>{"use strict";J();ne();xe();ae();$f=(e,t,r,n)=>{let o=k.size(t),i=t.length,a=E("input",e,i),d=M("output",e,i),l=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),p=k.normalizeAxis(l,i),m=u=>{let h=` i32(${a.indicesGet("inputIndices","uniforms.axis")}) `,_=q("uniforms.input_shape","uniforms.axis",i),y=n.reverse?h+(n.exclusive?" + 1":""):"0",g=n.reverse?_:h+(n.exclusive?"":" + 1");return`
                 ${u.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(a,d)}
                 ${u.mainStart()}
                   ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
@@ -5193,7 +5258,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
                     sum = sum + ${a.getByIndices("inputIndices")};
                   }
                   ${d.setByOffset("global_idx","sum")};
-                }`};return{name:"CumSum",shaderCache:{hint:n.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:[{type:12,data:o},{type:12,data:p},...N(t,t)]}),getShaderSource:m}},id=(e,t)=>{let r=e.inputs[0].dims,n=e.inputs[0].dataType,o=e.inputs[1];e.compute($f(n,r,o,t),{inputs:[0]})},ad=e=>{let t=e.exclusive===1,r=e.reverse===1;return re({exclusive:t,reverse:r})}});var xf,Sf,Tf,ud,dd,ld=U(()=>{"use strict";te();oe();Se();ae();xf=e=>{if(!e||e.length!==1)throw new Error("DepthToSpace requires 1 input.");if(e[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},Sf=(e,t,r,n)=>{let o=[];o.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
+                }`};return{name:"CumSum",shaderCache:{hint:n.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:[{type:12,data:o},{type:12,data:p},...N(t,t)]}),getShaderSource:m}},od=(e,t)=>{let r=e.inputs[0].dims,n=e.inputs[0].dataType,o=e.inputs[1];e.compute($f(n,r,o,t),{inputs:[0]})},id=e=>{let t=e.exclusive===1,r=e.reverse===1;return ee({exclusive:t,reverse:r})}});var xf,Sf,Tf,sd,ud,dd=U(()=>{"use strict";J();ne();xe();ae();xf=e=>{if(!e||e.length!==1)throw new Error("DepthToSpace requires 1 input.");if(e[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},Sf=(e,t,r,n)=>{let o=[];o.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
     var a: ${r.type.indices};`);for(let i=0;i<t;++i)o.push(r.indicesSet("a",e[i],`i[${i}]`));return o.push("return a;}"),o.join(`
 `)},Tf=(e,t)=>{let r,n,o,i,a,d,l=t.format==="NHWC",p=t.blocksize,m=t.mode==="DCR";l?([r,n,o,i]=e.dims,a=m?[r,n,o,p,p,i/p**2]:[r,n,o,i/p**2,p,p],d=m?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,n,o,i]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],a=m?[r,p,p,i/p**2,n,o]:[r,i/p**2,p,p,n,o],d=m?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let u=e.reshape(a),h=u.dims.length,_=e.dataType,y=E("a",_,h),g=M("output",_,h),x=$=>`
   ${$.registerUniform("output_size","u32").declareVariables(y,g)}
@@ -5207,18 +5272,18 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     let aIndices = perm(indices);
 
     ${g.setByOffset("global_idx",y.getByIndices("aIndices"))}
-  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:$=>{let v=l?[r,n*p,o*p,i/p**2]:[r,i/p**2,n*p,o*p],S=C.size(v),T=u.dims,A=C.sortBasedOnPerm(T,d);return{outputs:[{dims:v,dataType:$[0].dataType}],dispatchGroup:{x:Math.ceil(S/64)},programUniforms:[{type:12,data:S},...N(T,A)]}},getShaderSource:x}},ud=(e,t)=>{xf(e.inputs),e.compute(Tf(e.inputs[0],t))},dd=e=>re({blocksize:e.blocksize,mode:e.mode,format:e.format})});var vo,en,cd,If,Cf,$o,xo,pd,Af,md,fd,hd=U(()=>{"use strict";te();oe();Se();ae();vo="[a-zA-Z]|\\.\\.\\.",en="("+vo+")+",cd="^"+en+"$",If="("+en+",)*"+en,Cf="^"+If+"$",$o=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,r){let n=this.symbolToIndices.get(t);n===void 0?n=[r]:n.push(r),this.symbolToIndices.set(t,n)}},xo=class{constructor(t,r){this.equation=r;this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[n,o]=r.includes("->")?r.split("->",2):[r,""];if(!n.match(RegExp(Cf)))throw new Error("Invalid LHS term");if(n.split(",").forEach((d,l)=>{let p=t[l].dims.slice();if(!d.match(RegExp(cd)))throw new Error("Invalid LHS term");let m=this.processTerm(d,!0,p,l);this.lhs.push(m)}),o==="")o+=[...this.symbolToInfo.entries()].filter(([d,l])=>l.count===1||d==="...").map(([d])=>d).join("");else if(!o.match(RegExp(en)))throw new Error("Invalid RHS");o.match(RegExp(vo,"g"))?.forEach(d=>{if(d==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let l=this.symbolToInfo.get(d);if(l===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(l.dimValue)}}),this.rhs=this.processTerm(o,!1,this.outputDims)}addSymbol(t,r,n){let o=this.symbolToInfo.get(t);if(o!==void 0){if(o.dimValue!==r&&o.count!==1)throw new Error("Dimension mismatch");o.count++,o.inputIndices.push(n)}else o={count:1,dimValue:r,inputIndices:[n]};this.symbolToInfo.set(t,o)}processTerm(t,r,n,o=-1){let i=n.length,a=!1,d=[],l=0;if(!t.match(RegExp(cd))&&!r&&t!=="")throw new Error("Invalid LHS term");let p=t.match(RegExp(vo,"g")),m=new $o(o);return p?.forEach((u,h)=>{if(u==="..."){if(a)throw new Error("Only one ellipsis is allowed per input term");a=!0;let _=i-p.length+1;if(_<0)throw new Error("Ellipsis out of bounds");if(d=n.slice(l,l+_),this.hasEllipsis){if(this.ellipsisDims.length!==d.length||this.ellipsisDims.toString()!==d.toString())throw new Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=d;else throw new Error("Ellipsis must be specified in the LHS");for(let y=0;y<d.length;y++){let g=String.fromCharCode("0".charCodeAt(0)+y);m.addSymbol(g,h+y),this.addSymbol(g,n[l++],o)}}else m.addSymbol(u,h+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(u,n[l++],o)}),m}},pd=e=>e+"_max",Af=(e,t,r,n)=>{let i=e.map(m=>m.length).map((m,u)=>E(`input${u}`,t,m)),a=C.size(n),d=M("output",t,n.length),l=[...r.symbolToInfo.keys()].filter(m=>!r.rhs.symbolToIndices.has(m)),p=m=>{let u=[],h="var prod = 1.0;",_="var sum = 0.0;",y="sum += prod;",g=[],x=[],$=[],v=[],S=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((A,k)=>{if(r.rhs.symbolToIndices.has(k)){let P=r.rhs.symbolToIndices.get(k)?.[0];P!==void 0&&r.lhs.forEach((D,R)=>{if(A.inputIndices.includes(R)){let G=D.symbolToIndices.get(k);if(G===void 0)throw new Error("Invalid symbol error");G.forEach(K=>{u.push(`${i[R].indicesSet(`input${R}Indices`,K,d.indicesGet("outputIndices",P))}`)})}})}else r.lhs.forEach((P,D)=>{if(A.inputIndices.includes(D)){let R=P.symbolToIndices.get(k);if(R===void 0)throw new Error("Invalid symbol error");R.forEach(G=>{g.push(`${i[D].indicesSet(`input${D}Indices`,G,`${k}`)}`)}),v.push(`prod *= ${i[D].getByIndices(`input${D}Indices`)};`)}}),x.push(`for(var ${k}: u32 = 0; ${k} < uniforms.${pd(k)}; ${k}++) {`),$.push("}")});let T=S?[...u,`let sum = ${i.map((A,k)=>A.getByIndices(`input${k}Indices`)).join(" * ")};`]:[...u,_,...x,...g,h,...v,y,...$];return`
-            ${m.registerUniforms(l.map(A=>({name:`${pd(A)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...i,d)}
+  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:$=>{let v=l?[r,n*p,o*p,i/p**2]:[r,i/p**2,n*p,o*p],S=k.size(v),T=u.dims,A=k.sortBasedOnPerm(T,d);return{outputs:[{dims:v,dataType:$[0].dataType}],dispatchGroup:{x:Math.ceil(S/64)},programUniforms:[{type:12,data:S},...N(T,A)]}},getShaderSource:x}},sd=(e,t)=>{xf(e.inputs),e.compute(Tf(e.inputs[0],t))},ud=e=>ee({blocksize:e.blocksize,mode:e.mode,format:e.format})});var wo,Jr,ld,If,Cf,vo,$o,cd,Af,pd,md,fd=U(()=>{"use strict";J();ne();xe();ae();wo="[a-zA-Z]|\\.\\.\\.",Jr="("+wo+")+",ld="^"+Jr+"$",If="("+Jr+",)*"+Jr,Cf="^"+If+"$",vo=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,r){let n=this.symbolToIndices.get(t);n===void 0?n=[r]:n.push(r),this.symbolToIndices.set(t,n)}},$o=class{constructor(t,r){this.equation=r;this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[n,o]=r.includes("->")?r.split("->",2):[r,""];if(!n.match(RegExp(Cf)))throw new Error("Invalid LHS term");if(n.split(",").forEach((d,l)=>{let p=t[l].dims.slice();if(!d.match(RegExp(ld)))throw new Error("Invalid LHS term");let m=this.processTerm(d,!0,p,l);this.lhs.push(m)}),o==="")o+=[...this.symbolToInfo.entries()].filter(([d,l])=>l.count===1||d==="...").map(([d])=>d).join("");else if(!o.match(RegExp(Jr)))throw new Error("Invalid RHS");o.match(RegExp(wo,"g"))?.forEach(d=>{if(d==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let l=this.symbolToInfo.get(d);if(l===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(l.dimValue)}}),this.rhs=this.processTerm(o,!1,this.outputDims)}addSymbol(t,r,n){let o=this.symbolToInfo.get(t);if(o!==void 0){if(o.dimValue!==r&&o.count!==1)throw new Error("Dimension mismatch");o.count++,o.inputIndices.push(n)}else o={count:1,dimValue:r,inputIndices:[n]};this.symbolToInfo.set(t,o)}processTerm(t,r,n,o=-1){let i=n.length,a=!1,d=[],l=0;if(!t.match(RegExp(ld))&&!r&&t!=="")throw new Error("Invalid LHS term");let p=t.match(RegExp(wo,"g")),m=new vo(o);return p?.forEach((u,h)=>{if(u==="..."){if(a)throw new Error("Only one ellipsis is allowed per input term");a=!0;let _=i-p.length+1;if(_<0)throw new Error("Ellipsis out of bounds");if(d=n.slice(l,l+_),this.hasEllipsis){if(this.ellipsisDims.length!==d.length||this.ellipsisDims.toString()!==d.toString())throw new Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=d;else throw new Error("Ellipsis must be specified in the LHS");for(let y=0;y<d.length;y++){let g=String.fromCharCode("0".charCodeAt(0)+y);m.addSymbol(g,h+y),this.addSymbol(g,n[l++],o)}}else m.addSymbol(u,h+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(u,n[l++],o)}),m}},cd=e=>e+"_max",Af=(e,t,r,n)=>{let i=e.map(m=>m.length).map((m,u)=>E(`input${u}`,t,m)),a=k.size(n),d=M("output",t,n.length),l=[...r.symbolToInfo.keys()].filter(m=>!r.rhs.symbolToIndices.has(m)),p=m=>{let u=[],h="var prod = 1.0;",_="var sum = 0.0;",y="sum += prod;",g=[],x=[],$=[],v=[],S=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((A,C)=>{if(r.rhs.symbolToIndices.has(C)){let P=r.rhs.symbolToIndices.get(C)?.[0];P!==void 0&&r.lhs.forEach((D,R)=>{if(A.inputIndices.includes(R)){let H=D.symbolToIndices.get(C);if(H===void 0)throw new Error("Invalid symbol error");H.forEach(L=>{u.push(`${i[R].indicesSet(`input${R}Indices`,L,d.indicesGet("outputIndices",P))}`)})}})}else r.lhs.forEach((P,D)=>{if(A.inputIndices.includes(D)){let R=P.symbolToIndices.get(C);if(R===void 0)throw new Error("Invalid symbol error");R.forEach(H=>{g.push(`${i[D].indicesSet(`input${D}Indices`,H,`${C}`)}`)}),v.push(`prod *= ${i[D].getByIndices(`input${D}Indices`)};`)}}),x.push(`for(var ${C}: u32 = 0; ${C} < uniforms.${cd(C)}; ${C}++) {`),$.push("}")});let T=S?[...u,`let sum = ${i.map((A,C)=>A.getByIndices(`input${C}Indices`)).join(" * ")};`]:[...u,_,...x,...g,h,...v,y,...$];return`
+            ${m.registerUniforms(l.map(A=>({name:`${cd(A)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...i,d)}
 
             ${m.mainStart()}
             ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
             var outputIndices = ${d.offsetToIndices("global_idx")};
-            ${i.map((A,k)=>`var input${k}Indices: ${i[k].type.indices};`).join(`
+            ${i.map((A,C)=>`var input${C}Indices: ${i[C].type.indices};`).join(`
 `)}
             ${T.join(`
 `)};
             ${d.setByOffset("global_idx","sum")};
-          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let m=l.filter(h=>r.symbolToInfo.has(h)).map(h=>({type:12,data:r.symbolToInfo.get(h)?.dimValue||0}));m.push({type:12,data:a});let u=e.map((h,_)=>[...N(h)]).reduce((h,_)=>h.concat(_),m);return u.push(...N(n)),{outputs:[{dims:n,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:u}},getShaderSource:p}},md=(e,t)=>{let r=new xo(e.inputs,t.equation),n=r.outputDims,o=e.inputs.map((i,a)=>i.dims);e.compute(Af(o,e.inputs[0].dataType,r,n))},fd=e=>{let t=e.equation.replace(/\s+/g,"");return re({equation:t})}});var kf,gd,Ef,Pf,bd,yd=U(()=>{"use strict";te();oe();ae();kf=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=r.length<t.length?0:r.length-t.length,o=t.length<r.length?0:t.length-r.length;for(;n<r.length&&o<t.length;++n,++o)if(r[n]!==t[o]&&r[n]!==1&&t[o]!==1)throw new Error("Expand requires shape to be broadcastable to input")},gd=(e,t)=>{let r=e.length-t.length,n=[];for(let o=0;o<r;++o)n.push(e[o]);for(let o=0;o<t.length;++o)n.push(t[o]===1?e[o+r]:t[o]);return n},Ef=(e,t)=>e.length>t.length?gd(e,t):gd(t,e),Pf=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=Ef(t,r),o=e[0].dataType,i=o===9||C.size(t)===1,a=o===9||t.length>0&&t[t.length-1]%4===0?4:1,d=i||n.length>0&&n[n.length-1]%4===0?4:1,l=Math.ceil(C.size(n)/d),p=u=>{let h=E("input",o,t.length,a),_=M("output",o,n.length,d),y;if(o===9){let g=(x,$,v="")=>`
+          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let m=l.filter(h=>r.symbolToInfo.has(h)).map(h=>({type:12,data:r.symbolToInfo.get(h)?.dimValue||0}));m.push({type:12,data:a});let u=e.map((h,_)=>[...N(h)]).reduce((h,_)=>h.concat(_),m);return u.push(...N(n)),{outputs:[{dims:n,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:u}},getShaderSource:p}},pd=(e,t)=>{let r=new $o(e.inputs,t.equation),n=r.outputDims,o=e.inputs.map((i,a)=>i.dims);e.compute(Af(o,e.inputs[0].dataType,r,n))},md=e=>{let t=e.equation.replace(/\s+/g,"");return ee({equation:t})}});var kf,hd,Ef,Pf,gd,bd=U(()=>{"use strict";J();ne();ae();kf=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=r.length<t.length?0:r.length-t.length,o=t.length<r.length?0:t.length-r.length;for(;n<r.length&&o<t.length;++n,++o)if(r[n]!==t[o]&&r[n]!==1&&t[o]!==1)throw new Error("Expand requires shape to be broadcastable to input")},hd=(e,t)=>{let r=e.length-t.length,n=[];for(let o=0;o<r;++o)n.push(e[o]);for(let o=0;o<t.length;++o)n.push(t[o]===1?e[o+r]:t[o]);return n},Ef=(e,t)=>e.length>t.length?hd(e,t):hd(t,e),Pf=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=Ef(t,r),o=e[0].dataType,i=o===9||k.size(t)===1,a=o===9||t.length>0&&t[t.length-1]%4===0?4:1,d=i||n.length>0&&n[n.length-1]%4===0?4:1,l=Math.ceil(k.size(n)/d),p=u=>{let h=E("input",o,t.length,a),_=M("output",o,n.length,d),y;if(o===9){let g=(x,$,v="")=>`
           let outputIndices${$} = ${_.offsetToIndices(`outputOffset + ${$}u`)};
           let offset${$} = ${h.broadcastedIndicesToOffset(`outputIndices${$}`,_)};
           let index${$} = offset${$} / 4u;
@@ -5241,28 +5306,28 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     ${u.registerUniform("vec_size","u32").declareVariables(h,_)}
     ${u.mainStart()}
     ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
-    ${y}`},m=[{type:12,data:l},...N(t,n)];return{name:"Expand",shaderCache:{hint:`${n.length};${a}${d}`,inputDependencies:["rank"]},getShaderSource:p,getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:m})}},bd=e=>{kf(e.inputs),e.compute(Pf(e.inputs),{inputs:[0]})}});var zf,_d,wd=U(()=>{"use strict";te();oe();ae();Kr();zf=e=>{let t=e[0].dataType,r=C.size(e[0].dims),n=C.size(e[1].dims),o=n%4===0,i=a=>{let d=E("x",t,[1],4),l=E("bias",t,[1],4),p=M("y",t,[1],4),m=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],u=_=>`
+    ${y}`},m=[{type:12,data:l},...N(t,n)];return{name:"Expand",shaderCache:{hint:`${n.length};${a}${d}`,inputDependencies:["rank"]},getShaderSource:p,getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:m})}},gd=e=>{kf(e.inputs),e.compute(Pf(e.inputs),{inputs:[0]})}});var zf,yd,_d=U(()=>{"use strict";J();ne();ae();qr();zf=e=>{let t=e[0].dataType,r=k.size(e[0].dims),n=k.size(e[1].dims),o=n%4===0,i=a=>{let d=E("x",t,[1],4),l=E("bias",t,[1],4),p=M("y",t,[1],4),m=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],u=_=>`
       let bias${_}_offset: u32 = (global_idx * 4 + ${_}) % uniforms.bias_size;
       let bias${_} = ${l.getByOffset(`bias${_}_offset / 4`)}[bias${_}_offset % 4];`,h=o?`
       let bias = ${l.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${u(0)}${u(1)}${u(2)}${u(3)}
       let bias = ${d.type.value}(bias0, bias1, bias2, bias3);`;return`${a.registerUniforms(m).declareVariables(d,l,p)}
 
-    ${mo(Ee(t))}
+    ${po(Ee(t))}
 
-    ${a.mainStart(kt)}
+    ${a.mainStart(Et)}
       ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}
 
       let x = ${d.getByOffset("global_idx")};
       ${h}
       let x_in = x + bias;
-      ${p.setByOffset("global_idx",fo("x_in"))}
-    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${o}`,inputDependencies:["type","type"]},getShaderSource:i,getRunData:a=>({outputs:[{dims:a[0].dims,dataType:a[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:n}],dispatchGroup:{x:Math.ceil(r/kt/4)}})}},_d=e=>{e.inputs.length<2||C.size(e.inputs[1].dims)===0?hu(e):e.compute(zf(e.inputs))}});var Of,Df,vd,$d,xd=U(()=>{"use strict";te();oe();Se();ae();Of=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},Df=(e,t)=>{let r=e[0].dims,n=e[1].dims,o=r.length,i=C.normalizeAxis(t.axis,o),a=r.slice(0);a.splice(i,1,...n);let d=r[i],l=e[0].dataType===9?4:1,p=Math.ceil(C.size(a)/l),m=[{type:12,data:p},{type:6,data:d},{type:12,data:i},...N(e[0].dims,e[1].dims,a)],u=h=>{let _=E("data",e[0].dataType,e[0].dims.length,l),y=E("inputIndices",e[1].dataType,e[1].dims.length),g=M("output",e[0].dataType,a.length,l),x=v=>{let S=n.length,T=`var indicesIndices${v}  = ${y.type.indices}(0);`;for(let A=0;A<S;A++)T+=`${S>1?`indicesIndices${v}[${A}]`:`indicesIndices${v}`} = ${a.length>1?`outputIndices${v}[uniforms.axis + ${A}]`:`outputIndices${v}`};`;T+=`
+      ${p.setByOffset("global_idx",mo("x_in"))}
+    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${o}`,inputDependencies:["type","type"]},getShaderSource:i,getRunData:a=>({outputs:[{dims:a[0].dims,dataType:a[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:n}],dispatchGroup:{x:Math.ceil(r/Et/4)}})}},yd=e=>{e.inputs.length<2||k.size(e.inputs[1].dims)===0?fu(e):e.compute(zf(e.inputs))}});var Of,Df,wd,vd,$d=U(()=>{"use strict";J();ne();xe();ae();Of=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},Df=(e,t)=>{let r=e[0].dims,n=e[1].dims,o=r.length,i=k.normalizeAxis(t.axis,o),a=r.slice(0);a.splice(i,1,...n);let d=r[i],l=e[0].dataType===9?4:1,p=Math.ceil(k.size(a)/l),m=[{type:12,data:p},{type:6,data:d},{type:12,data:i},...N(e[0].dims,e[1].dims,a)],u=h=>{let _=E("data",e[0].dataType,e[0].dims.length,l),y=E("inputIndices",e[1].dataType,e[1].dims.length),g=M("output",e[0].dataType,a.length,l),x=v=>{let S=n.length,T=`var indicesIndices${v}  = ${y.type.indices}(0);`;for(let A=0;A<S;A++)T+=`${S>1?`indicesIndices${v}[${A}]`:`indicesIndices${v}`} = ${a.length>1?`outputIndices${v}[uniforms.axis + ${A}]`:`outputIndices${v}`};`;T+=`
           var idx${v} = ${y.getByIndices(`indicesIndices${v}`)};
           if (idx${v} < 0) {
             idx${v} = idx${v} + uniforms.axisDimLimit;
           }
           var dataIndices${v} : ${_.type.indices};
-        `;for(let A=0,k=0;A<o;A++)A===i?(T+=`${o>1?`dataIndices${v}[${A}]`:`dataIndices${v}`} = u32(idx${v});`,k+=S):(T+=`${o>1?`dataIndices${v}[${A}]`:`dataIndices${v}`} = ${a.length>1?`outputIndices${v}[${k}]`:`outputIndices${v}`};`,k++);return T},$;if(e[0].dataType===9){let v=(S,T,A="")=>`
+        `;for(let A=0,C=0;A<o;A++)A===i?(T+=`${o>1?`dataIndices${v}[${A}]`:`dataIndices${v}`} = u32(idx${v});`,C+=S):(T+=`${o>1?`dataIndices${v}[${A}]`:`dataIndices${v}`} = ${a.length>1?`outputIndices${v}[${C}]`:`outputIndices${v}`};`,C++);return T},$;if(e[0].dataType===9){let v=(S,T,A="")=>`
           let outputIndices${T} = ${g.offsetToIndices(`outputOffset + ${T}u`)};
           ${x(T)};
           let offset${T} = ${_.indicesToOffset(`dataIndices${T}`)};
@@ -5287,7 +5352,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       ${h.mainStart()}
         ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
         ${$}
-      }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:u}},vd=e=>re({axis:e.axis}),$d=(e,t)=>{let r=e.inputs;Of(r),e.compute(Df(e.inputs,t))}});var Bf,Sd,Td,Id=U(()=>{"use strict";te();oe();ae();Bf=(e,t,r,n,o,i,a,d,l)=>{let p=[{type:12,data:i},{type:12,data:n},{type:12,data:o},{type:12,data:r},{type:12,data:a},{type:12,data:d},{type:12,data:l}],m=[i];p.push(...N(t.dims,m));let u=h=>{let _=E("indices_data",t.dataType,t.dims.length),y=M("input_slice_offsets_data",12,1,1),g=[_,y],x=[{name:"output_size",type:"u32"},{name:"batch_dims",type:"u32"},{name:"input_dims",type:"u32",length:o.length},{name:"sizes_from_slice_dims_data",type:"u32",length:r.length},{name:"num_slices_per_batch",type:"u32"},{name:"input_batch_stride",type:"u32"},{name:"num_slice_dims",type:"u32"}];return`
+      }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:u}},wd=e=>ee({axis:e.axis}),vd=(e,t)=>{let r=e.inputs;Of(r),e.compute(Df(e.inputs,t))}});var Bf,xd,Sd,Td=U(()=>{"use strict";J();ne();ae();Bf=(e,t,r,n,o,i,a,d,l)=>{let p=[{type:12,data:i},{type:12,data:n},{type:12,data:o},{type:12,data:r},{type:12,data:a},{type:12,data:d},{type:12,data:l}],m=[i];p.push(...N(t.dims,m));let u=h=>{let _=E("indices_data",t.dataType,t.dims.length),y=M("input_slice_offsets_data",12,1,1),g=[_,y],x=[{name:"output_size",type:"u32"},{name:"batch_dims",type:"u32"},{name:"input_dims",type:"u32",length:o.length},{name:"sizes_from_slice_dims_data",type:"u32",length:r.length},{name:"num_slices_per_batch",type:"u32"},{name:"input_batch_stride",type:"u32"},{name:"num_slice_dims",type:"u32"}];return`
   ${h.registerUniforms(x).declareVariables(...g)}
   ${h.mainStart()}
     ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -5306,13 +5371,13 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     }
 
     input_slice_offsets_data[global_idx] =  base_offset + u32(relative_slice_offset);
-  }`};return e.compute({name:"computeSliceOffsets",shaderCache:{hint:`${o.length}_${r.length}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:m,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:p}),getShaderSource:u},{inputs:[t],outputs:[-1]})[0]},Sd=(e,t)=>{let r=e.inputs,n=r[0].dims,o=r[0].dataType,i=r[1].dims,a=i[i.length-1],d=C.sizeToDimension(i,i.length-1),l=C.sizeFromDimension(n,t.batchDims+a),p=C.sizeToDimension(n,t.batchDims),m=C.sizeFromDimension(n,t.batchDims),u=d/p,h=new Array(a),_=l;for(let T=0;T<a;++T)h[a-1-T]=_,_*=n[t.batchDims+a-1-T];let y=Bf(e,r[1],h,t.batchDims,n,d,u,m,a),g=t.batchDims+a;if(g>n.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let x=i.slice(0,-1).concat(n.slice(g)),$=C.size(x),v=[{type:12,data:$},{type:12,data:l},...N(r[0].dims,y.dims,x)],S=T=>{let A=E("data",r[0].dataType,r[0].dims.length),k=E("slice_offsets",12,y.dims.length),P=M("output",r[0].dataType,x.length);return`
-          ${T.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(A,k,P)}
+  }`};return e.compute({name:"computeSliceOffsets",shaderCache:{hint:`${o.length}_${r.length}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:m,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:p}),getShaderSource:u},{inputs:[t],outputs:[-1]})[0]},xd=(e,t)=>{let r=e.inputs,n=r[0].dims,o=r[0].dataType,i=r[1].dims,a=i[i.length-1],d=k.sizeToDimension(i,i.length-1),l=k.sizeFromDimension(n,t.batchDims+a),p=k.sizeToDimension(n,t.batchDims),m=k.sizeFromDimension(n,t.batchDims),u=d/p,h=new Array(a),_=l;for(let T=0;T<a;++T)h[a-1-T]=_,_*=n[t.batchDims+a-1-T];let y=Bf(e,r[1],h,t.batchDims,n,d,u,m,a),g=t.batchDims+a;if(g>n.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let x=i.slice(0,-1).concat(n.slice(g)),$=k.size(x),v=[{type:12,data:$},{type:12,data:l},...N(r[0].dims,y.dims,x)],S=T=>{let A=E("data",r[0].dataType,r[0].dims.length),C=E("slice_offsets",12,y.dims.length),P=M("output",r[0].dataType,x.length);return`
+          ${T.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(A,C,P)}
             ${T.mainStart()}
             ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
           let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
           output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
-        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:x,dataType:o}],dispatchGroup:{x:Math.ceil($/64)},programUniforms:v}),getShaderSource:S},{inputs:[r[0],y]})},Td=e=>({batchDims:e.batch_dims,cacheKey:""})});var Mf,Rf,Cd,Ad,kd=U(()=>{"use strict";te();oe();Se();ae();Mf=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let r=C.normalizeAxis(t.quantizeAxis,e[0].dims.length),n=t.blockSize,o=e[0],i=e[2],a=e.length===4?e[3]:void 0;if(i.dims.length!==o.dims.length||!o.dims.map((d,l)=>l===r?Math.ceil(d/n)===i.dims[l]:d===i.dims[l]).reduce((d,l)=>d&&l,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(a){if(a.dataType!==o.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(a.dims.length!==i.dims.length||!a.dims.map((d,l)=>d===i.dims[l]).reduce((d,l)=>d&&l,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},Rf=(e,t)=>{let r=e[0].dims,n=e[1].dims,o=r.length,i=C.normalizeAxis(t.gatherAxis,o),a=C.normalizeAxis(t.quantizeAxis,o),d=r.slice(0);d.splice(i,1,...n);let l=C.size(d),p=e[2].dataType,u=e[0].dataType===22,h=[{type:12,data:l},{type:12,data:a},{type:12,data:i},{type:12,data:t.blockSize},...N(...e.map((y,g)=>y.dims),d)],_=y=>{let g=E("data",e[0].dataType,e[0].dims.length),x=E("inputIndices",e[1].dataType,e[1].dims.length),$=E("scales",e[2].dataType,e[2].dims.length),v=e.length>3?E("zeroPoint",e[3].dataType,e[3].dims.length):void 0,S=M("output",p,d.length),T=[g,x,$];v&&T.push(v);let A=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
+        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:x,dataType:o}],dispatchGroup:{x:Math.ceil($/64)},programUniforms:v}),getShaderSource:S},{inputs:[r[0],y]})},Sd=e=>({batchDims:e.batch_dims,cacheKey:""})});var Mf,Rf,Id,Cd,Ad=U(()=>{"use strict";J();ne();xe();ae();Mf=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let r=k.normalizeAxis(t.quantizeAxis,e[0].dims.length),n=t.blockSize,o=e[0],i=e[2],a=e.length===4?e[3]:void 0;if(i.dims.length!==o.dims.length||!o.dims.map((d,l)=>l===r?Math.ceil(d/n)===i.dims[l]:d===i.dims[l]).reduce((d,l)=>d&&l,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(a){if(a.dataType!==o.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(a.dims.length!==i.dims.length||!a.dims.map((d,l)=>d===i.dims[l]).reduce((d,l)=>d&&l,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},Rf=(e,t)=>{let r=e[0].dims,n=e[1].dims,o=r.length,i=k.normalizeAxis(t.gatherAxis,o),a=k.normalizeAxis(t.quantizeAxis,o),d=r.slice(0);d.splice(i,1,...n);let l=k.size(d),p=e[2].dataType,u=e[0].dataType===22,h=[{type:12,data:l},{type:12,data:a},{type:12,data:i},{type:12,data:t.blockSize},...N(...e.map((y,g)=>y.dims),d)],_=y=>{let g=E("data",e[0].dataType,e[0].dims.length),x=E("inputIndices",e[1].dataType,e[1].dims.length),$=E("scales",e[2].dataType,e[2].dims.length),v=e.length>3?E("zeroPoint",e[3].dataType,e[3].dims.length):void 0,S=M("output",p,d.length),T=[g,x,$];v&&T.push(v);let A=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
         ${y.registerUniforms(A).declareVariables(...T,S)}
         ${y.mainStart()}
         let output_indices = ${S.offsetToIndices("global_idx")};
@@ -5357,8 +5422,8 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
               let zero_point = zero_point_vec[zero_point_index / 2];`:"var zero_point = 0")()};
         let dequantized_data = ${Ee(p)}(quantized_data - zero_point) * scale;
         ${S.setByOffset("global_idx","dequantized_data")};
-    }`};return{name:"GatherBlockQuantized",shaderCache:{hint:`${t.cacheKey};${e.filter((y,g)=>g!==1).map(y=>y.dims.join("_")).join(";")}`,inputDependencies:Array.from({length:e.length},(y,g)=>"rank")},getRunData:()=>({outputs:[{dims:d,dataType:p}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:h}),getShaderSource:_}},Cd=(e,t)=>{let r=e.inputs;Mf(r,t),e.compute(Rf(e.inputs,t))},Ad=e=>re({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})});var Uf,Nf,Ed,Pd,zd=U(()=>{"use strict";te();oe();Se();ae();Uf=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and
-                     indices input tensors be of same rank.`)},Nf=(e,t)=>{let r=e[0].dims,n=e[0].dataType,o=r.length,i=e[1].dims,a=e[1].dataType,d=C.normalizeAxis(t.axis,o),l=r[d],p=i.slice(0),m=C.size(p),u=E("input",n,o),h=E("indicesInput",a,i.length),_=M("output",n,p.length),y=[{type:12,data:m},{type:6,data:l},{type:12,data:d}];return y.push(...N(r,i,p)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:y}),getShaderSource:$=>`
+    }`};return{name:"GatherBlockQuantized",shaderCache:{hint:`${t.cacheKey};${e.filter((y,g)=>g!==1).map(y=>y.dims.join("_")).join(";")}`,inputDependencies:Array.from({length:e.length},(y,g)=>"rank")},getRunData:()=>({outputs:[{dims:d,dataType:p}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:h}),getShaderSource:_}},Id=(e,t)=>{let r=e.inputs;Mf(r,t),e.compute(Rf(e.inputs,t))},Cd=e=>ee({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})});var Uf,Nf,kd,Ed,Pd=U(()=>{"use strict";J();ne();xe();ae();Uf=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and
+                     indices input tensors be of same rank.`)},Nf=(e,t)=>{let r=e[0].dims,n=e[0].dataType,o=r.length,i=e[1].dims,a=e[1].dataType,d=k.normalizeAxis(t.axis,o),l=r[d],p=i.slice(0),m=k.size(p),u=E("input",n,o),h=E("indicesInput",a,i.length),_=M("output",n,p.length),y=[{type:12,data:m},{type:6,data:l},{type:12,data:d}];return y.push(...N(r,i,p)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:y}),getShaderSource:$=>`
       ${$.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(u,h,_)}
       ${$.mainStart()}
       ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
@@ -5374,8 +5439,8 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       let value = ${u.getByIndices("inputIndices")};
 
       ${_.setByOffset("global_idx","value")};
-  }`}},Ed=e=>re({axis:e.axis}),Pd=(e,t)=>{let r=e.inputs;Uf(r),e.compute(Nf(e.inputs,t))}});var Vf,Wf,Od,Dd,Bd=U(()=>{"use strict";te();oe();ae();Vf=e=>{if(!e)throw new Error("Input is missing");if(e.length<2||e.length>3)throw new Error("Invaid input number.");if(e.length===3&&e[2].dims.length>2)throw new Error("Invalid input shape of C");if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("Input types are mismatched")},Wf=(e,t)=>{let r=e[0].dims.slice(),n=e[1].dims.slice(),[o,i,a]=Vr.getShapeOfGemmResult(r,t.transA,n,t.transB,e.length===3?e[2].dims:void 0),d=[o,i];if(!d)throw new Error("Can't use gemm on the given tensors");let l=16,p=Math.ceil(i/l),m=Math.ceil(o/l),u=!0,h=C.size(d),_=[{type:12,data:u?p:h},{type:12,data:o},{type:12,data:i},{type:12,data:a},{type:1,data:t.alpha},{type:1,data:t.beta}],y=["type","type"];e.length===3&&(_.push(...N(e[2].dims)),y.push("rank")),_.push(...N(d));let g=$=>{let v="";t.transA&&t.transB?v="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?v="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?v="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(v="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let S=t.alpha===1?"":"value *= uniforms.alpha;",T=E("a",e[0].dataType,e[0].dims),A=E("b",e[1].dataType,e[1].dims),k=T.type.value,P=null,D=[T,A];e.length===3&&(P=E("c",e[2].dataType,e[2].dims.length),D.push(P));let R=M("output",e[0].dataType,d.length);D.push(R);let G=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`
-  ${$.registerUniforms(G).declareVariables(...D)}
+  }`}},kd=e=>ee({axis:e.axis}),Ed=(e,t)=>{let r=e.inputs;Uf(r),e.compute(Nf(e.inputs,t))}});var Vf,Wf,zd,Od,Dd=U(()=>{"use strict";J();ne();ae();Vf=e=>{if(!e)throw new Error("Input is missing");if(e.length<2||e.length>3)throw new Error("Invaid input number.");if(e.length===3&&e[2].dims.length>2)throw new Error("Invalid input shape of C");if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("Input types are mismatched")},Wf=(e,t)=>{let r=e[0].dims.slice(),n=e[1].dims.slice(),[o,i,a]=Nr.getShapeOfGemmResult(r,t.transA,n,t.transB,e.length===3?e[2].dims:void 0),d=[o,i];if(!d)throw new Error("Can't use gemm on the given tensors");let l=16,p=Math.ceil(i/l),m=Math.ceil(o/l),u=!0,h=k.size(d),_=[{type:12,data:u?p:h},{type:12,data:o},{type:12,data:i},{type:12,data:a},{type:1,data:t.alpha},{type:1,data:t.beta}],y=["type","type"];e.length===3&&(_.push(...N(e[2].dims)),y.push("rank")),_.push(...N(d));let g=$=>{let v="";t.transA&&t.transB?v="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?v="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?v="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(v="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let S=t.alpha===1?"":"value *= uniforms.alpha;",T=E("a",e[0].dataType,e[0].dims),A=E("b",e[1].dataType,e[1].dims),C=T.type.value,P=null,D=[T,A];e.length===3&&(P=E("c",e[2].dataType,e[2].dims.length),D.push(P));let R=M("output",e[0].dataType,d.length);D.push(R);let H=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`
+  ${$.registerUniforms(H).declareVariables(...D)}
 
   ${$.mainStart()}
     ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -5383,15 +5448,15 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     let m = global_idx / uniforms.N;
     let n = global_idx % uniforms.N;
 
-    var value = ${k}(0);
+    var value = ${C}(0);
     for (var k: u32 = 0u; k < uniforms.K; k++) {
       ${v}
     }
 
     ${S}
-    ${(()=>P!=null?`let cOffset = ${P.broadcastedIndicesToOffset("vec2(m, n)",R)}; value += ${k}(uniforms.beta) * ${P.getByOffset("cOffset")};`:"")()}
+    ${(()=>P!=null?`let cOffset = ${P.broadcastedIndicesToOffset("vec2(m, n)",R)}; value += ${C}(uniforms.beta) * ${P.getByOffset("cOffset")};`:"")()}
     output[global_idx] = value;
-  }`},x=$=>{let v=E("a",e[0].dataType,e[0].dims),S=E("b",e[1].dataType,e[1].dims),T=null,A=[v,S];e.length===3&&(T=E("c",e[2].dataType,e[2].dims.length),A.push(T));let k=M("output",e[0].dataType,d.length);A.push(k);let P=[{name:"num_tile_n",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}],D="",R="";t.transA&&t.transB?(R=`
+  }`},x=$=>{let v=E("a",e[0].dataType,e[0].dims),S=E("b",e[1].dataType,e[1].dims),T=null,A=[v,S];e.length===3&&(T=E("c",e[2].dataType,e[2].dims.length),A.push(T));let C=M("output",e[0].dataType,d.length);A.push(C);let P=[{name:"num_tile_n",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}],D="",R="";t.transA&&t.transB?(R=`
       var col = tile_row_start + local_id.x;
       var row = k_start + local_id.y;
       if (col < uniforms.M && row < uniforms.K) {
@@ -5455,7 +5520,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       } else {
         tile_b[local_id.y][local_id.x] = ${S.type.value}(0);
       }
-      `,D="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let G=t.alpha===1?"":"value *= uniforms.alpha;";return`
+      `,D="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let H=t.alpha===1?"":"value *= uniforms.alpha;";return`
   ${$.registerUniforms(P).declareVariables(...A)}
   var<workgroup> tile_a: array<array<${v.type.storage}, ${l}>, ${l}>;
   var<workgroup> tile_b: array<array<${S.type.storage}, ${l}>, ${l}>;
@@ -5464,7 +5529,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     let tile_row_start = (workgroup_index / uniforms.num_tile_n) * ${l};
     let num_tiles = (uniforms.K - 1) / ${l} + 1;
     var k_start = 0u;
-    var value = ${k.type.value}(0);
+    var value = ${C.type.value}(0);
     for (var t: u32 = 0u; t < num_tiles; t++) {
       ${R}
       k_start = k_start + ${l};
@@ -5476,14 +5541,14 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       workgroupBarrier();
     }
 
-    ${G}
+    ${H}
     let m = tile_row_start + local_id.y;
     let n = tile_col_start + local_id.x;
-    ${(()=>T!=null?`let cOffset = ${T.broadcastedIndicesToOffset("vec2(m, n)",k)}; value += ${k.type.value}(uniforms.beta) * ${T.getByOffset("cOffset")};`:"")()}
+    ${(()=>T!=null?`let cOffset = ${T.broadcastedIndicesToOffset("vec2(m, n)",C)}; value += ${C.type.value}(uniforms.beta) * ${T.getByOffset("cOffset")};`:"")()}
     if (m < uniforms.M && n < uniforms.N) {
       output[m * uniforms.N + n] = value;
     }
-  }`};return u?{name:"GemmShared",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:d,dataType:e[0].dataType}],dispatchGroup:{x:p*m},programUniforms:_}),getShaderSource:x}:{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:d,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:_}),getShaderSource:g}},Od=e=>{let t=e.transA,r=e.transB,n=e.alpha,o=e.beta;return{transA:t,transB:r,alpha:n,beta:o,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},Dd=(e,t)=>{Vf(e.inputs),e.compute(Wf(e.inputs,t))}});var lt,_t,Ut,Nt,Lf,Gf,Hf,Ff,qf,Kf,jf,Yf,Md,Rd,Ud=U(()=>{"use strict";te();oe();Se();ae();[lt,_t,Ut,Nt]=[0,1,2,3],Lf=e=>{if(e[0].dims.length!==4)throw new Error("only 4-D tensor is supported.");if(e[0].dims.length!==e[1].dims.length)throw new Error("input dimensions must be equal to grid dimensions");if(e[0].dims.length-2!==e[1].dims[e[1].dims.length-1])throw new Error(`last dimension of grid must be equal to ${e[0].dims.length-2}`);if(e[0].dims[0]!==e[1].dims[0])throw new Error("grid batch size must match input batch size")},Gf=`
+  }`};return u?{name:"GemmShared",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:d,dataType:e[0].dataType}],dispatchGroup:{x:p*m},programUniforms:_}),getShaderSource:x}:{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:d,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:_}),getShaderSource:g}},zd=e=>{let t=e.transA,r=e.transB,n=e.alpha,o=e.beta;return{transA:t,transB:r,alpha:n,beta:o,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},Od=(e,t)=>{Vf(e.inputs),e.compute(Wf(e.inputs,t))}});var lt,_t,Nt,Vt,Lf,Gf,Hf,Ff,qf,Kf,jf,Yf,Bd,Md,Rd=U(()=>{"use strict";J();ne();xe();ae();[lt,_t,Nt,Vt]=[0,1,2,3],Lf=e=>{if(e[0].dims.length!==4)throw new Error("only 4-D tensor is supported.");if(e[0].dims.length!==e[1].dims.length)throw new Error("input dimensions must be equal to grid dimensions");if(e[0].dims.length-2!==e[1].dims[e[1].dims.length-1])throw new Error(`last dimension of grid must be equal to ${e[0].dims.length-2}`);if(e[0].dims[0]!==e[1].dims[0])throw new Error("grid batch size must match input batch size")},Gf=`
   fn gs_get_cubic_coeffs(x: f32) -> vec4<f32> {
     let cubic_alpha = -0.75f;
     let x_abs = abs(x);
@@ -5549,15 +5614,15 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
      indices[${lt}] = batch;
      indices[${_t}] = channel;`+(()=>{switch(r.paddingMode){case"zeros":return`
           if (r >= 0 && r < H && c >=0 && c < W) {
-            indices[${Ut}] = u32(r);
-            indices[${Nt}] = u32(c);
+            indices[${Nt}] = u32(r);
+            indices[${Vt}] = u32(c);
           }
         `;case"border":return`
-          indices[${Ut}] = u32(clamp(r, 0, H - 1));
-          indices[${Nt}] = u32(clamp(c, 0, W - 1));
+          indices[${Nt}] = u32(clamp(r, 0, H - 1));
+          indices[${Vt}] = u32(clamp(c, 0, W - 1));
         `;case"reflection":return`
-          indices[${Ut}] = gs_reflect(r, border[1], border[3]);
-          indices[${Nt}] = gs_reflect(c, border[0], border[2]);
+          indices[${Nt}] = gs_reflect(r, border[1], border[3]);
+          indices[${Vt}] = gs_reflect(c, border[0], border[2]);
         `;default:throw new Error(`padding mode ${r.paddingMode} is not supported`)}})()+`
     return ${e.getByIndices("indices")};
   }
@@ -5592,7 +5657,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
           let dx = x - f32(x0 + 1);
           let dy = y - f32(y0 + 1);
           let result = gs_bicubic_interpolate(p, dx, dy);
-        `;default:throw new Error(`mode ${r.mode} is not supported`)}})()+`${e.setByOffset("global_idx","result")}`,Yf=(e,t)=>{let r=E("x",e[0].dataType,e[0].dims.length),n=[e[1].dims[0],e[1].dims[1],e[1].dims[2]],o=E("grid",e[1].dataType,n.length,2),i=[e[0].dims[0],e[0].dims[1],e[1].dims[1],e[1].dims[2]];t.format==="NHWC"&&(i=[e[0].dims[0],e[1].dims[1],e[1].dims[2],e[0].dims[3]],[lt,_t,Ut,Nt]=[0,3,1,2]);let a=M("output",e[0].dataType,i.length),d=r.type.value,l=C.size(i),p=[{type:12,data:l},...N(e[0].dims,n,i)],m=u=>`
+        `;default:throw new Error(`mode ${r.mode} is not supported`)}})()+`${e.setByOffset("global_idx","result")}`,Yf=(e,t)=>{let r=E("x",e[0].dataType,e[0].dims.length),n=[e[1].dims[0],e[1].dims[1],e[1].dims[2]],o=E("grid",e[1].dataType,n.length,2),i=[e[0].dims[0],e[0].dims[1],e[1].dims[1],e[1].dims[2]];t.format==="NHWC"&&(i=[e[0].dims[0],e[1].dims[1],e[1].dims[2],e[0].dims[3]],[lt,_t,Nt,Vt]=[0,3,1,2]);let a=M("output",e[0].dataType,i.length),d=r.type.value,l=k.size(i),p=[{type:12,data:l},...N(e[0].dims,n,i)],m=u=>`
   ${u.registerUniform("output_size","u32").declareVariables(r,o,a)}
   ${Gf}
   ${Hf(d)}
@@ -5602,8 +5667,8 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
 
   ${u.mainStart()}
     ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
-      let H_in = i32(uniforms.x_shape[${Ut}]);
-      let W_in = i32(uniforms.x_shape[${Nt}]);
+      let H_in = i32(uniforms.x_shape[${Nt}]);
+      let W_in = i32(uniforms.x_shape[${Vt}]);
 
       ${t.alignCorners===0?`
       let x_min = -0.5;
@@ -5619,23 +5684,23 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       let border = vec4<f32>(x_min, y_min, x_max, y_max);
 
       let indices = ${a.offsetToIndices("global_idx")};
-      var grid_indices = vec3<u32>(indices[${lt}], indices[${Ut}], indices[${Nt}]);
+      var grid_indices = vec3<u32>(indices[${lt}], indices[${Nt}], indices[${Vt}]);
       let nxy = ${o.getByIndices("grid_indices")};
       var x = gs_denormalize(f32(nxy[0]), W_in);
       var y = gs_denormalize(f32(nxy[1]), H_in);
 
       ${jf(a,d,t)}
-  }`;return{name:"GridSample",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:["type","type"]},getRunData:u=>{let h=C.size(i);return{outputs:[{dims:i,dataType:u[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:p}},getShaderSource:m}},Md=(e,t)=>{Lf(e.inputs),e.compute(Yf(e.inputs,t))},Rd=e=>re({alignCorners:e.align_corners,mode:e.mode,paddingMode:e.padding_mode,format:e.format})});var Be,Xf,Vd,Nd,Jf,tr,Wd,So=U(()=>{"use strict";te();oe();Se();Nr();Fr();ae();dt();Be=(e,t)=>e.length>t&&e[t].dims.length>0?e[t]:void 0,Xf=(e,t)=>{let r=e[0],n=Be(e,1),o=Be(e,2),i=Be(e,3),a=Be(e,4),d=Be(e,5),l=Be(e,6),p=Be(e,7);if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let m=r.dims[0],u=r.dims[1],h=r.dims.length===3?r.dims[2]:t.numHeads*r.dims[4],_=u,y=0,g=0,x=Math.floor(h/t.numHeads);if(l&&p&&C.size(l.dims)&&C.size(p.dims)){if(l.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(l.dims[0]!==m||l.dims[1]!==t.numHeads||l.dims[3]!==x)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(p.dims[0]!==m||p.dims[1]!==t.numHeads||p.dims[3]!==x)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(l.dims[2]!==p.dims[2])throw new Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(p.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');y=l.dims[2],g=l.dims[2]}else if(l&&C.size(l.dims)||p&&C.size(p.dims))throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $;if(n&&C.size(n.dims)>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(n.dims[2]!==r.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');$=2,_=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==x)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(o)throw new Error('Expect "value" be none when "key" has packed kv format.');$=5,_=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==x)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,_=n.dims[2]}}else{if(r.dims.length!==5)throw new Error('Input "query" is expected to have 5 dimensions when key is empty');if(r.dims[2]!==t.numHeads||r.dims[3]!==3)throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');$=3}if(i&&C.size(i.dims)>0){if(i.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(n&&n.dims.length===5&&n.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let v=y+_,S=0;if(a&&C.size(a.dims)>0){S=8;let P=a.dims;throw P.length===1?P[0]===m?S=1:P[0]===3*m+2&&(S=3):P.length===2&&P[0]===m&&P[1]===v&&(S=5),S===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let T=!1,A=h;if(o&&C.size(o.dims)>0){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(_!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');A=o.dims[2]}else{if(_!==o.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');A=o.dims[1]*o.dims[3],T=!0}}let k=!1;if(a&&C.size(a.dims)>0)throw new Error("Key padding mask is not supported");if(d&&C.size(d.dims)>0){if(d.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(d.dims[0]!==m||d.dims[1]!==t.numHeads||d.dims[2]!==u||d.dims[3]!==v)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:m,sequenceLength:u,pastSequenceLength:y,kvSequenceLength:_,totalSequenceLength:v,maxSequenceLength:g,inputHiddenSize:0,hiddenSize:h,vHiddenSize:A,headSize:x,vHeadSize:Math.floor(A/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:S,scale:t.scale,broadcastResPosBias:k,passPastInKv:T,qkvFormat:$}},Vd=e=>re({...e}),Nd=re({perm:[0,2,1,3]}),Jf=(e,t,r,n,o,i,a)=>{let d=[n,o,i],l=C.size(d),p=[{type:12,data:l},{type:12,data:a},{type:12,data:i}],m=u=>{let h=M("qkv_with_bias",t.dataType,d),_=E("qkv",t.dataType,d),y=E("bias",r.dataType,d),g=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
+  }`;return{name:"GridSample",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:["type","type"]},getRunData:u=>{let h=k.size(i);return{outputs:[{dims:i,dataType:u[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:p}},getShaderSource:m}},Bd=(e,t)=>{Lf(e.inputs),e.compute(Yf(e.inputs,t))},Md=e=>ee({alignCorners:e.align_corners,mode:e.mode,paddingMode:e.padding_mode,format:e.format})});var Me,Xf,Nd,Ud,Jf,er,Vd,xo=U(()=>{"use strict";J();ne();xe();Ur();Hr();ae();dt();Me=(e,t)=>e.length>t&&e[t].dims.length>0?e[t]:void 0,Xf=(e,t)=>{let r=e[0],n=Me(e,1),o=Me(e,2),i=Me(e,3),a=Me(e,4),d=Me(e,5),l=Me(e,6),p=Me(e,7);if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let m=r.dims[0],u=r.dims[1],h=r.dims.length===3?r.dims[2]:t.numHeads*r.dims[4],_=u,y=0,g=0,x=Math.floor(h/t.numHeads);if(l&&p&&k.size(l.dims)&&k.size(p.dims)){if(l.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(l.dims[0]!==m||l.dims[1]!==t.numHeads||l.dims[3]!==x)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(p.dims[0]!==m||p.dims[1]!==t.numHeads||p.dims[3]!==x)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(l.dims[2]!==p.dims[2])throw new Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(p.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');y=l.dims[2],g=l.dims[2]}else if(l&&k.size(l.dims)||p&&k.size(p.dims))throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $;if(n&&k.size(n.dims)>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(n.dims[2]!==r.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');$=2,_=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==x)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(o)throw new Error('Expect "value" be none when "key" has packed kv format.');$=5,_=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==x)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,_=n.dims[2]}}else{if(r.dims.length!==5)throw new Error('Input "query" is expected to have 5 dimensions when key is empty');if(r.dims[2]!==t.numHeads||r.dims[3]!==3)throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');$=3}if(i&&k.size(i.dims)>0){if(i.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(n&&n.dims.length===5&&n.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let v=y+_,S=0;if(a&&k.size(a.dims)>0){S=8;let P=a.dims;throw P.length===1?P[0]===m?S=1:P[0]===3*m+2&&(S=3):P.length===2&&P[0]===m&&P[1]===v&&(S=5),S===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let T=!1,A=h;if(o&&k.size(o.dims)>0){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(_!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');A=o.dims[2]}else{if(_!==o.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');A=o.dims[1]*o.dims[3],T=!0}}let C=!1;if(a&&k.size(a.dims)>0)throw new Error("Key padding mask is not supported");if(d&&k.size(d.dims)>0){if(d.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(d.dims[0]!==m||d.dims[1]!==t.numHeads||d.dims[2]!==u||d.dims[3]!==v)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:m,sequenceLength:u,pastSequenceLength:y,kvSequenceLength:_,totalSequenceLength:v,maxSequenceLength:g,inputHiddenSize:0,hiddenSize:h,vHiddenSize:A,headSize:x,vHeadSize:Math.floor(A/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:S,scale:t.scale,broadcastResPosBias:C,passPastInKv:T,qkvFormat:$}},Nd=e=>ee({...e}),Ud=ee({perm:[0,2,1,3]}),Jf=(e,t,r,n,o,i,a)=>{let d=[n,o,i],l=k.size(d),p=[{type:12,data:l},{type:12,data:a},{type:12,data:i}],m=u=>{let h=M("qkv_with_bias",t.dataType,d),_=E("qkv",t.dataType,d),y=E("bias",r.dataType,d),g=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
   ${u.registerUniforms(g).declareVariables(_,y,h)}
   ${u.mainStart()}
     ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
     let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;
 
     qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
-  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:d,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:p}),getShaderSource:m},{inputs:[t,r],outputs:[-1]})[0]},tr=(e,t,r,n,o,i,a,d)=>{let l=i;if(a&&C.size(a.dims)>0){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=Jf(e,i,a,t,n,r*o,d),l=l.reshape([t,n,r,o]),r===1||n===1?l:e.compute(Pe(l,Nd.perm),{inputs:[l],outputs:[-1]})[0]}else return i.dims.length===3&&(l=i.reshape([t,n,r,o])),r===1||n===1?l:e.compute(Pe(l,Nd.perm),{inputs:[l],outputs:[-1]})[0]},Wd=(e,t)=>{let r=Xf(e.inputs,t),n=e.inputs[0],o=Be(e.inputs,1),i=Be(e.inputs,2),a=Be(e.inputs,3),d=Be(e.inputs,4),l=Be(e.inputs,5),p=Be(e.inputs,6),m=Be(e.inputs,7);if(n.dims.length===5)throw new Error("Packed QKV is not implemented");if(o?.dims.length===5)throw new Error("Packed KV is not implemented");let u=o&&i&&o.dims.length===4&&i.dims.length===4,h=tr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,n,a,0);if(u)return Rt(e,h,o,i,d,void 0,p,m,l,r);if(!o||!i)throw new Error("key and value must be provided");let _=tr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,o,a,r.hiddenSize),y=tr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,a,2*r.hiddenSize);Rt(e,h,_,y,d,void 0,p,m,l,r)}});var eh,th,rh,nh,To,Ld,Gd,Io=U(()=>{"use strict";te();oe();Se();ae();eh=e=>{if(!e||e.length<1)throw new Error("too few inputs")},th=(e,t)=>{let r=[],n=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(o=>r.push(Number(o))),n=r.length),re({numOutputs:n,axis:t.axis,splitSizes:r})},rh=e=>`
+  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:d,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:p}),getShaderSource:m},{inputs:[t,r],outputs:[-1]})[0]},er=(e,t,r,n,o,i,a,d)=>{let l=i;if(a&&k.size(a.dims)>0){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=Jf(e,i,a,t,n,r*o,d),l=l.reshape([t,n,r,o]),r===1||n===1?l:e.compute(Pe(l,Ud.perm),{inputs:[l],outputs:[-1]})[0]}else return i.dims.length===3&&(l=i.reshape([t,n,r,o])),r===1||n===1?l:e.compute(Pe(l,Ud.perm),{inputs:[l],outputs:[-1]})[0]},Vd=(e,t)=>{let r=Xf(e.inputs,t),n=e.inputs[0],o=Me(e.inputs,1),i=Me(e.inputs,2),a=Me(e.inputs,3),d=Me(e.inputs,4),l=Me(e.inputs,5),p=Me(e.inputs,6),m=Me(e.inputs,7);if(n.dims.length===5)throw new Error("Packed QKV is not implemented");if(o?.dims.length===5)throw new Error("Packed KV is not implemented");let u=o&&i&&o.dims.length===4&&i.dims.length===4,h=er(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,n,a,0);if(u)return Ut(e,h,o,i,d,void 0,p,m,l,r);if(!o||!i)throw new Error("key and value must be provided");let _=er(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,o,a,r.hiddenSize),y=er(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,a,2*r.hiddenSize);Ut(e,h,_,y,d,void 0,p,m,l,r)}});var eh,th,rh,nh,So,Wd,Ld,To=U(()=>{"use strict";J();ne();xe();ae();eh=e=>{if(!e||e.length<1)throw new Error("too few inputs")},th=(e,t)=>{let r=[],n=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(o=>r.push(Number(o))),n=r.length),ee({numOutputs:n,axis:t.axis,splitSizes:r})},rh=e=>`
 fn calculateOutputIndex(index: u32) -> u32 {
     for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
-    if (index < ${F("uniforms.size_in_split_axis","i",e)}) {
+    if (index < ${q("uniforms.size_in_split_axis","i",e)}) {
         return i;
     }
     }
@@ -5644,7 +5709,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
       fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
         ${r.join(`
 `)}
-      }`},To=(e,t)=>{let r=e[0].dims,n=C.size(r),o=e[0].dataType,i=C.normalizeAxis(t.axis,r.length),a=new Array(t.numOutputs),d=E("input",o,r.length),l=new Array(t.numOutputs),p=[],m=[],u=0,h=[{type:12,data:n}];for(let y=0;y<t.numOutputs;y++){u+=t.splitSizes[y],l[y]=u;let g=r.slice();g[i]=t.splitSizes[y],m.push(g),a[y]=M(`output${y}`,o,g.length),p.push({dims:m[y],dataType:e[0].dataType})}h.push({type:12,data:l},...N(r,...m));let _=y=>`
+      }`},So=(e,t)=>{let r=e[0].dims,n=k.size(r),o=e[0].dataType,i=k.normalizeAxis(t.axis,r.length),a=new Array(t.numOutputs),d=E("input",o,r.length),l=new Array(t.numOutputs),p=[],m=[],u=0,h=[{type:12,data:n}];for(let y=0;y<t.numOutputs;y++){u+=t.splitSizes[y],l[y]=u;let g=r.slice();g[i]=t.splitSizes[y],m.push(g),a[y]=M(`output${y}`,o,g.length),p.push({dims:m[y],dataType:e[0].dataType})}h.push({type:12,data:l},...N(r,...m));let _=y=>`
   ${y.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",l.length).declareVariables(d,...a)}
   ${rh(l.length)}
   ${nh(a)}
@@ -5656,11 +5721,11 @@ fn calculateOutputIndex(index: u32) -> u32 {
     var index = ${d.indicesGet("indices",i)};
     let output_number = calculateOutputIndex(index);
     if (output_number != 0) {
-      index -= ${F("uniforms.size_in_split_axis","output_number - 1u",l.length)};
+      index -= ${q("uniforms.size_in_split_axis","output_number - 1u",l.length)};
       ${d.indicesSet("indices",i,"index")};
     }
     writeBufferData(output_number, indices, global_idx);
-  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:p,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:h})}},Ld=(e,t)=>{eh(e.inputs);let r=e.inputs.length===1?t:th(e.inputs,t);e.compute(To(e.inputs,r),{inputs:[0]})},Gd=e=>{let t=e.axis,r=e.splitSizes,n=e.numOutputs<0?r.length:e.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return re({axis:t,numOutputs:n,splitSizes:r})}});var oh,ih,Hd,Fd,qd=U(()=>{"use strict";Se();Fr();So();Io();dt();oh=(e,t)=>{if(t.doRotary)throw new Error("GroupQuerryAttention do_rotary attribute is not supported");if(t.doRotary&&e.length<=7)throw new Error("cos_cache and sin_cache inputs are required if do_rotary is specified");let r=e[0],n=e[1],o=e[2],i=e[3],a=e[4];if(t.localWindowSize!==-1)throw new Error("Local attention is not supported");if(t.softcap!==0)throw new Error("Softcap is not supported");if(t.rotaryInterleaved!==0)throw new Error("Rotary interleaved is not supported");if(t.smoothSoftmax)throw new Error("Smooth softmax is not supported");if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let d=!1,l=r.dims[0],p=r.dims[1],m=r.dims.length===3?d?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],u=p,h=0,_=!n||n.dims.length===0,y=Math.floor(_?m/(t.numHeads+2*t.kvNumHeads):m/t.numHeads);_&&(m=y*t.numHeads);let g=i&&i.dims.length!==0,x=a&&a.dims.length!==0;if(g&&i.dims.length===4&&i.dims[0]===l&&i.dims[1]!==t.kvNumHeads&&i.dims[2]===t.kvNumHeads&&i.dims[3]===y)throw new Error("BSNH pastKey/pastValue is not supported");if(g&&x){if(i.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(a.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');h=i.dims[2]}else if(g||x)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let v=1;if(n&&n.dims.length>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(r.dims[2]%n.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');u=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==y)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(o)throw new Error('Expect "value" be none when "key" has packed kv format.');u=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==y)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');u=n.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');v=3}let S=0,T=!1,A=t.kvNumHeads?y*t.kvNumHeads:m;if(o&&o.dims.length>0){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(u!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');A=o.dims[2]}else{if(u!==o.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');A=o.dims[1]*o.dims[3],T=!0}}let k=e.length>4?e[5]:void 0;if(k&&k.dims.length!==1&&k.dims[0]!==l)throw new Error('Input "seqlens" is expected to have 1 dimension and the same dim 0 as batch_size');let P=-1,D=-1,R=!1;return{batchSize:l,sequenceLength:p,pastSequenceLength:h,kvSequenceLength:u,totalSequenceLength:P,maxSequenceLength:D,inputHiddenSize:0,hiddenSize:m,vHiddenSize:A,headSize:y,vHeadSize:Math.floor(A/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:S,scale:t.scale,broadcastResPosBias:R,passPastInKv:T,qkvFormat:v}},ih=re({perm:[0,2,1,3]}),Hd=(e,t,r)=>{let n=t,o=r.kvNumHeads;return t.dims.length===3&&r.kvSequenceLength!==0&&(n=t.reshape([r.batchSize,r.kvSequenceLength,o,r.headSize]),n=e.compute(Pe(n,ih.perm),{inputs:[n],outputs:[-1]})[0]),n},Fd=(e,t)=>{let r=oh(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(e.inputs[1]?.dims.length===5)throw new Error("Packed KV is not implemented");let n=e.inputs[0],o=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,i=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,a=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,d=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,l=e.inputs.length>4?e.inputs[5]:void 0,p=e.inputs.length>5?e.inputs[6]:void 0,m=r.kvNumHeads?r.kvNumHeads:r.numHeads,u=re({axis:2,numOutputs:3,splitSizes:[r.numHeads*r.headSize,m*r.headSize,m*r.headSize]}),[h,_,y]=!o&&!i?e.compute(To([n],u),{inputs:[n],outputs:[-1,-1,-1]}):[n,o,i],g=tr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,h,void 0,0);Rt(e,g,Hd(e,_,r),Hd(e,y,r),void 0,void 0,a,d,void 0,r,l,p)}});var Kd,ah,sh,jd,Yd=U(()=>{"use strict";te();oe();dt();ae();Kd=(e,t,r,n,o,i,a,d)=>{let l=me(i),p=l===1?"f32":`vec${l}f`,m=l===1?"vec2f":`mat2x${l}f`,u=o*a,h=64;u===1&&(h=256);let _=[o,a,i/l],y=[o,a,2],g=["rank","type","type"],x=[];x.push(...N(_,y));let $=v=>{let S=E("x",t.dataType,3,l),T=E("scale",r.dataType,r.dims),A=E("bias",n.dataType,n.dims),k=M("output",1,3,2),P=[S,T,A,k];return`
+  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:p,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:h})}},Wd=(e,t)=>{eh(e.inputs);let r=e.inputs.length===1?t:th(e.inputs,t);e.compute(So(e.inputs,r),{inputs:[0]})},Ld=e=>{let t=e.axis,r=e.splitSizes,n=e.numOutputs<0?r.length:e.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return ee({axis:t,numOutputs:n,splitSizes:r})}});var oh,ih,Gd,Hd,Fd=U(()=>{"use strict";xe();Hr();xo();To();dt();oh=(e,t)=>{if(t.doRotary)throw new Error("GroupQuerryAttention do_rotary attribute is not supported");if(t.doRotary&&e.length<=7)throw new Error("cos_cache and sin_cache inputs are required if do_rotary is specified");let r=e[0],n=e[1],o=e[2],i=e[3],a=e[4];if(t.localWindowSize!==-1)throw new Error("Local attention is not supported");if(t.softcap!==0)throw new Error("Softcap is not supported");if(t.rotaryInterleaved!==0)throw new Error("Rotary interleaved is not supported");if(t.smoothSoftmax)throw new Error("Smooth softmax is not supported");if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let d=!1,l=r.dims[0],p=r.dims[1],m=r.dims.length===3?d?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],u=p,h=0,_=!n||n.dims.length===0,y=Math.floor(_?m/(t.numHeads+2*t.kvNumHeads):m/t.numHeads);_&&(m=y*t.numHeads);let g=i&&i.dims.length!==0,x=a&&a.dims.length!==0;if(g&&i.dims.length===4&&i.dims[0]===l&&i.dims[1]!==t.kvNumHeads&&i.dims[2]===t.kvNumHeads&&i.dims[3]===y)throw new Error("BSNH pastKey/pastValue is not supported");if(g&&x){if(i.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(a.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');h=i.dims[2]}else if(g||x)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let v=1;if(n&&n.dims.length>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(r.dims[2]%n.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');u=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==y)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(o)throw new Error('Expect "value" be none when "key" has packed kv format.');u=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==y)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');u=n.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');v=3}let S=0,T=!1,A=t.kvNumHeads?y*t.kvNumHeads:m;if(o&&o.dims.length>0){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(u!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');A=o.dims[2]}else{if(u!==o.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');A=o.dims[1]*o.dims[3],T=!0}}let C=e.length>4?e[5]:void 0;if(C&&C.dims.length!==1&&C.dims[0]!==l)throw new Error('Input "seqlens" is expected to have 1 dimension and the same dim 0 as batch_size');let P=-1,D=-1,R=!1;return{batchSize:l,sequenceLength:p,pastSequenceLength:h,kvSequenceLength:u,totalSequenceLength:P,maxSequenceLength:D,inputHiddenSize:0,hiddenSize:m,vHiddenSize:A,headSize:y,vHeadSize:Math.floor(A/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:S,scale:t.scale,broadcastResPosBias:R,passPastInKv:T,qkvFormat:v}},ih=ee({perm:[0,2,1,3]}),Gd=(e,t,r)=>{let n=t,o=r.kvNumHeads;return t.dims.length===3&&r.kvSequenceLength!==0&&(n=t.reshape([r.batchSize,r.kvSequenceLength,o,r.headSize]),n=e.compute(Pe(n,ih.perm),{inputs:[n],outputs:[-1]})[0]),n},Hd=(e,t)=>{let r=oh(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(e.inputs[1]?.dims.length===5)throw new Error("Packed KV is not implemented");let n=e.inputs[0],o=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,i=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,a=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,d=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,l=e.inputs.length>4?e.inputs[5]:void 0,p=e.inputs.length>5?e.inputs[6]:void 0,m=r.kvNumHeads?r.kvNumHeads:r.numHeads,u=ee({axis:2,numOutputs:3,splitSizes:[r.numHeads*r.headSize,m*r.headSize,m*r.headSize]}),[h,_,y]=!o&&!i?e.compute(So([n],u),{inputs:[n],outputs:[-1,-1,-1]}):[n,o,i],g=er(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,h,void 0,0);Ut(e,g,Gd(e,_,r),Gd(e,y,r),void 0,void 0,a,d,void 0,r,l,p)}});var qd,ah,sh,Kd,jd=U(()=>{"use strict";J();ne();dt();ae();qd=(e,t,r,n,o,i,a,d)=>{let l=me(i),p=l===1?"f32":`vec${l}f`,m=l===1?"vec2f":`mat2x${l}f`,u=o*a,h=64;u===1&&(h=256);let _=[o,a,i/l],y=[o,a,2],g=["rank","type","type"],x=[];x.push(...N(_,y));let $=v=>{let S=E("x",t.dataType,3,l),T=E("scale",r.dataType,r.dims),A=E("bias",n.dataType,n.dims),C=M("output",1,3,2),P=[S,T,A,C];return`
   var<workgroup> workgroup_shared : array<${m}, ${h}>;
   const workgroup_size = ${h}u;
   ${v.declareVariables(...P)}
@@ -5686,15 +5751,15 @@ fn calculateOutputIndex(index: u32) -> u32 {
       workgroupBarrier();
     }
     if (local_idx == 0) {
-      let sum_final = ${Fe("workgroup_shared[0][0]",l)} / f32(hight * ${l});
-      let squared_sum_final = ${Fe("workgroup_shared[0][1]",l)} / f32(hight * ${l});
+      let sum_final = ${qe("workgroup_shared[0][0]",l)} / f32(hight * ${l});
+      let squared_sum_final = ${qe("workgroup_shared[0][1]",l)} / f32(hight * ${l});
 
       let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${d}));
       let channel_scale = inv_std_dev * f32(scale[channel]);
       let channel_shift = f32(bias[channel]) - sum_final * channel_scale;
       output[workgroup_index] = vec2f(channel_scale, channel_shift);
     }
-  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${d};${h}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:y,dataType:1}],dispatchGroup:{x:u},programUniforms:x}),getShaderSource:$},{inputs:[t,r,n],outputs:[-1]})[0]},ah=(e,t,r)=>{let n=t[0].dims,o=n,i=2,a=n[0],d=n[1],l=C.sizeFromDimension(n,i),p=me(l),m=C.size(o)/p,u=Kd(e,t[0],t[1],t[2],a,l,d,r.epsilon),h=[a,d,l/p],_=[a,d],y=["type","none"],g=x=>{let $=E("x",t[0].dataType,h.length,p),v=E("scale_shift",1,_.length,2),S=M("output",t[0].dataType,h.length,p),T=[$,v,S];return`
+  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${d};${h}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:y,dataType:1}],dispatchGroup:{x:u},programUniforms:x}),getShaderSource:$},{inputs:[t,r,n],outputs:[-1]})[0]},ah=(e,t,r)=>{let n=t[0].dims,o=n,i=2,a=n[0],d=n[1],l=k.sizeFromDimension(n,i),p=me(l),m=k.size(o)/p,u=qd(e,t[0],t[1],t[2],a,l,d,r.epsilon),h=[a,d,l/p],_=[a,d],y=["type","none"],g=x=>{let $=E("x",t[0].dataType,h.length,p),v=E("scale_shift",1,_.length,2),S=M("output",t[0].dataType,h.length,p),T=[$,v,S];return`
   ${x.registerUniform("output_size","u32").declareVariables(...T)}
   ${x.mainStart()}
   ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -5704,10 +5769,10 @@ fn calculateOutputIndex(index: u32) -> u32 {
       let scale_shift = ${v.getByIndices("vec2<u32>(batch, channel)")};
       let value = ${$.getByOffset("global_idx")} * ${S.type.value}(scale_shift.x) + ${S.type.value}(scale_shift.y);
       ${S.setByOffset("global_idx","value")};
-  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${p}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},...N(h,_,h)]}),getShaderSource:g},{inputs:[t[0],u]})},sh=(e,t,r)=>{let n=t[0].dims,o=n,i=n[0],a=n[n.length-1],d=C.sizeFromDimension(n,1)/a,l=me(a),p=C.size(o)/l,m=[{type:12,data:d},{type:12,data:Math.floor(a/l)}],u=["type","type"],h=!1,_=[0,n.length-1];for(let $=0;$<n.length-2;$++)h=h||n[$+1]!==1,_.push($+1);h=h&&n[n.length-1]!==1;let y=h?e.compute(Pe(e.inputs[0],_),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:n.length},($,v)=>n[_[v]])),g=Kd(e,y,t[1],t[2],i,d,a,r.epsilon),x=$=>{let v=_e(t[0].dataType),S=l===1?"vec2f":`mat${l}x2f`,T=P=>{let D=P===0?"x":"y",R=l===1?"f32":`vec${l}f`;switch(l){case 1:return`${v}(${R}(scale.${D}))`;case 2:return`vec2<${v}>(${R}(scale[0].${D}, scale[1].${D}))`;case 4:return`vec4<${v}>(${R}(scale[0].${D}, scale[1].${D}, scale[2].${D}, scale[3].${D}))`;default:throw new Error(`Not supported compoents ${l}`)}},A=E("input",t[0].dataType,t[0].dims,l),k=M("output",t[0].dataType,o,l);return`
+  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${p}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},...N(h,_,h)]}),getShaderSource:g},{inputs:[t[0],u]})},sh=(e,t,r)=>{let n=t[0].dims,o=n,i=n[0],a=n[n.length-1],d=k.sizeFromDimension(n,1)/a,l=me(a),p=k.size(o)/l,m=[{type:12,data:d},{type:12,data:Math.floor(a/l)}],u=["type","type"],h=!1,_=[0,n.length-1];for(let $=0;$<n.length-2;$++)h=h||n[$+1]!==1,_.push($+1);h=h&&n[n.length-1]!==1;let y=h?e.compute(Pe(e.inputs[0],_),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:n.length},($,v)=>n[_[v]])),g=qd(e,y,t[1],t[2],i,d,a,r.epsilon),x=$=>{let v=ye(t[0].dataType),S=l===1?"vec2f":`mat${l}x2f`,T=P=>{let D=P===0?"x":"y",R=l===1?"f32":`vec${l}f`;switch(l){case 1:return`${v}(${R}(scale.${D}))`;case 2:return`vec2<${v}>(${R}(scale[0].${D}, scale[1].${D}))`;case 4:return`vec4<${v}>(${R}(scale[0].${D}, scale[1].${D}, scale[2].${D}, scale[3].${D}))`;default:throw new Error(`Not supported compoents ${l}`)}},A=E("input",t[0].dataType,t[0].dims,l),C=M("output",t[0].dataType,o,l);return`
   @group(0) @binding(0) var<storage, read> input : array<${A.type.storage}>;
   @group(0) @binding(1) var<storage, read> scale_input : array<${S}>;
-  @group(0) @binding(2) var<storage, read_write> output : array<${k.type.storage}>;
+  @group(0) @binding(2) var<storage, read_write> output : array<${C.type.storage}>;
   struct Uniforms {H: u32, C : u32};
   @group(0) @binding(3) var<uniform> uniforms: Uniforms;
 
@@ -5718,52 +5783,52 @@ fn calculateOutputIndex(index: u32) -> u32 {
     let scale_offset = current_image_number * uniforms.C + current_channel_number;
     let scale = scale_input[scale_offset];
     output[global_idx] = fma(input[global_idx], ${T(0)}, ${T(1)});
-  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${l}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:x},{inputs:[t[0],g]})},jd=(e,t)=>{t.format==="NHWC"?sh(e,e.inputs,t):ah(e,e.inputs,t)}});var uh,dh,Zd,Qd=U(()=>{"use strict";te();oe();ae();uh=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},dh=(e,t,r)=>{let n=t.simplified,o=e[0].dims,i=e[1],a=!n&&e[2],d=o,l=C.normalizeAxis(t.axis,o.length),p=C.sizeToDimension(o,l),m=C.sizeFromDimension(o,l),u=C.size(i.dims),h=a?C.size(a.dims):0;if(u!==m||a&&h!==m)throw new Error(`Size of X.shape()[axis:] == ${m}.
+  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${l}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:x},{inputs:[t[0],g]})},Kd=(e,t)=>{t.format==="NHWC"?sh(e,e.inputs,t):ah(e,e.inputs,t)}});var uh,dh,Yd,Zd=U(()=>{"use strict";J();ne();ae();uh=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},dh=(e,t,r)=>{let n=t.simplified,o=e[0].dims,i=e[1],a=!n&&e[2],d=o,l=k.normalizeAxis(t.axis,o.length),p=k.sizeToDimension(o,l),m=k.sizeFromDimension(o,l),u=k.size(i.dims),h=a?k.size(a.dims):0;if(u!==m||a&&h!==m)throw new Error(`Size of X.shape()[axis:] == ${m}.
        Size of scale and bias (if provided) must match this.
-       Got scale size of ${u} and bias size of ${h}`);let _=[];for(let A=0;A<o.length;++A)A<l?_.push(o[A]):_.push(1);let y=me(m),g=["type","type"],x=[{type:12,data:p},{type:1,data:m},{type:12,data:Math.floor(m/y)},{type:1,data:t.epsilon}];a&&g.push("type");let $=r>1,v=r>2,S=A=>{let k=_e(e[0].dataType),P=[E("x",e[0].dataType,e[0].dims,y),E("scale",i.dataType,i.dims,y)];a&&P.push(E("bias",a.dataType,a.dims,y)),P.push(M("output",e[0].dataType,d,y)),$&&P.push(M("mean_data_output",1,_)),v&&P.push(M("inv_std_output",1,_));let D=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
+       Got scale size of ${u} and bias size of ${h}`);let _=[];for(let A=0;A<o.length;++A)A<l?_.push(o[A]):_.push(1);let y=me(m),g=["type","type"],x=[{type:12,data:p},{type:1,data:m},{type:12,data:Math.floor(m/y)},{type:1,data:t.epsilon}];a&&g.push("type");let $=r>1,v=r>2,S=A=>{let C=ye(e[0].dataType),P=[E("x",e[0].dataType,e[0].dims,y),E("scale",i.dataType,i.dims,y)];a&&P.push(E("bias",a.dataType,a.dims,y)),P.push(M("output",e[0].dataType,d,y)),$&&P.push(M("mean_data_output",1,_)),v&&P.push(M("inv_std_output",1,_));let D=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
   ${A.registerUniforms(D).declareVariables(...P)}
   ${A.mainStart()}
     ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
     let offset = global_idx * uniforms.norm_size_vectorized;
-    var mean_vector = ${uo("f32",y)};
-    var mean_square_vector = ${uo("f32",y)};
+    var mean_vector = ${so("f32",y)};
+    var mean_square_vector = ${so("f32",y)};
 
     for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
-      let value = ${Et(k,y,"x[h + offset]")};
+      let value = ${Pt(C,y,"x[h + offset]")};
       mean_vector += value;
       mean_square_vector += value * value;
     }
-    let mean = ${Fe("mean_vector",y)} / uniforms.norm_size;
-    let inv_std_dev = inverseSqrt(${Fe("mean_square_vector",y)} / uniforms.norm_size ${n?"":"- mean * mean"} + uniforms.epsilon);
+    let mean = ${qe("mean_vector",y)} / uniforms.norm_size;
+    let inv_std_dev = inverseSqrt(${qe("mean_square_vector",y)} / uniforms.norm_size ${n?"":"- mean * mean"} + uniforms.epsilon);
 
     for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
-      let f32input = ${Et(k,y,"x[j + offset]")};
-      let f32scale = ${Et(k,y,"scale[j]")};
+      let f32input = ${Pt(C,y,"x[j + offset]")};
+      let f32scale = ${Pt(C,y,"scale[j]")};
       output[j + offset] = ${P[0].type.value}((f32input ${n?"":"- mean"}) * inv_std_dev * f32scale
-        ${a?`+ ${Et(k,y,"bias[j]")}`:""}
+        ${a?`+ ${Pt(C,y,"bias[j]")}`:""}
       );
     }
 
     ${$?"mean_data_output[global_idx] = mean":""};
     ${v?"inv_std_output[global_idx] = inv_std_dev":""};
-  }`},T=[{dims:d,dataType:e[0].dataType}];return $&&T.push({dims:_,dataType:1}),v&&T.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${y};${r};${n}`,inputDependencies:g},getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(p/64)},programUniforms:x}),getShaderSource:S}},Zd=(e,t)=>{uh(e.inputs),e.compute(dh(e.inputs,t,e.outputCount))}});var lh,Xd,Jd=U(()=>{"use strict";oe();Qr();Xr();lh=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},Xd=e=>{lh(e.inputs);let t=tt.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can't use matmul on the given tensors");let r=t[t.length-1],n=e.inputs[0].dims[e.inputs[0].dims.length-1];if(r<8&&n<8)e.compute(Zr(e.inputs,{activation:""},t));else{let o=t[t.length-2],i=C.size(e.inputs[0].dims.slice(0,-2)),a=C.size(e.inputs[1].dims.slice(0,-2));if(i!==1&&o===1&&a===1){let d=e.inputs[0].reshape([1,i,n]),l=e.inputs[1].reshape([1,n,r]),p=[1,i,r],m=[d,l];e.compute(er(m,{activation:""},t,p),{inputs:m})}else e.compute(er(e.inputs,{activation:""},t))}}});var ch,ph,mh,el,tl,rl=U(()=>{"use strict";te();oe();Se();ae();ch=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],n=r.dims.length;if(r.dims[n-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let o=Math.floor((t.k+t.blockSize-1)/t.blockSize),i=t.blockSize/8*t.bits,a=e[1];if(!C.areEqual(a.dims,[t.n,o,i]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let l=e[2].dims;if(C.size(l)!==t.n*o)throw new Error("scales input size error.");if(e.length===4){let m=e[3].dims,u=t.bits>4?t.n*o:t.n*Math.floor((o+1)/2);if(C.size(m)!==u)throw new Error("zeroPoints input size error.")}},ph=(e,t)=>{let r=e[0].dims,n=r.length,o=r[n-2],i=t.k,a=t.n,d=r.slice(0,n-2),l=C.size(d),m=e[1].dims[2]/4,u=e[0].dataType,h=me(t.k),_=me(m),y=me(a),g=d.concat([o,a]),x=o>1&&a/y%2===0?2:1,$=C.size(g)/y/x,v=64,S=[],T=[l,o,i/h],A=C.convertShape(e[1].dims).slice();A.splice(-1,1,m/_),S.push(...N(T)),S.push(...N(A)),S.push(...N(e[2].dims)),e.length===4&&S.push(...N(C.convertShape(e[3].dims)));let k=[l,o,a/y];S.push(...N(k));let P=D=>{let R=T.length,G=E("a",e[0].dataType,R,h),K=E("b",12,A.length,_),j=E("scales",e[2].dataType,e[2].dims.length),V=[G,K,j],Q=e.length===4?E("zero_points",12,e[3].dims.length):void 0;Q&&V.push(Q);let se=k.length,Y=M("output",e[0].dataType,se,y),ee=_e(e[0].dataType),J=(()=>{switch(h){case 1:return`array<${ee}, 8>`;case 2:return`mat4x2<${ee}>`;case 4:return`mat2x4<${ee}>`;default:throw new Error(`${h}-component is not supported.`)}})(),ne=()=>{let $e=`
+  }`},T=[{dims:d,dataType:e[0].dataType}];return $&&T.push({dims:_,dataType:1}),v&&T.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${y};${r};${n}`,inputDependencies:g},getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(p/64)},programUniforms:x}),getShaderSource:S}},Yd=(e,t)=>{uh(e.inputs),e.compute(dh(e.inputs,t,e.outputCount))}});var lh,Qd,Xd=U(()=>{"use strict";ne();Zr();Qr();lh=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},Qd=e=>{lh(e.inputs);let t=tt.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can't use matmul on the given tensors");let r=t[t.length-1],n=e.inputs[0].dims[e.inputs[0].dims.length-1];if(r<8&&n<8)e.compute(Yr(e.inputs,{activation:""},t));else{let o=t[t.length-2],i=k.size(e.inputs[0].dims.slice(0,-2)),a=k.size(e.inputs[1].dims.slice(0,-2));if(i!==1&&o===1&&a===1){let d=e.inputs[0].reshape([1,i,n]),l=e.inputs[1].reshape([1,n,r]),p=[1,i,r],m=[d,l];e.compute(Jt(m,{activation:""},t,p),{inputs:m})}else e.compute(Jt(e.inputs,{activation:""},t))}}});var ch,ph,mh,Jd,el,tl=U(()=>{"use strict";J();ne();xe();ae();ch=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],n=r.dims.length;if(r.dims[n-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let o=Math.floor((t.k+t.blockSize-1)/t.blockSize),i=t.blockSize/8*t.bits,a=e[1];if(!k.areEqual(a.dims,[t.n,o,i]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let l=e[2].dims;if(k.size(l)!==t.n*o)throw new Error("scales input size error.");if(e.length===4){let m=e[3].dims,u=t.bits>4?t.n*o:t.n*Math.floor((o+1)/2);if(k.size(m)!==u)throw new Error("zeroPoints input size error.")}},ph=(e,t)=>{let r=e[0].dims,n=r.length,o=r[n-2],i=t.k,a=t.n,d=r.slice(0,n-2),l=k.size(d),m=e[1].dims[2]/4,u=e[0].dataType,h=me(t.k),_=me(m),y=me(a),g=d.concat([o,a]),x=o>1&&a/y%2===0?2:1,$=k.size(g)/y/x,v=64,S=[],T=[l,o,i/h],A=k.convertShape(e[1].dims).slice();A.splice(-1,1,m/_),S.push(...N(T)),S.push(...N(A)),S.push(...N(e[2].dims)),e.length===4&&S.push(...N(k.convertShape(e[3].dims)));let C=[l,o,a/y];S.push(...N(C));let P=D=>{let R=T.length,H=E("a",e[0].dataType,R,h),L=E("b",12,A.length,_),re=E("scales",e[2].dataType,e[2].dims.length),V=[H,L,re],K=e.length===4?E("zero_points",12,e[3].dims.length):void 0;K&&V.push(K);let we=C.length,j=M("output",e[0].dataType,we,y),Q=ye(e[0].dataType),ie=(()=>{switch(h){case 1:return`array<${Q}, 8>`;case 2:return`mat4x2<${Q}>`;case 4:return`mat2x4<${Q}>`;default:throw new Error(`${h}-component is not supported.`)}})(),te=()=>{let ve=`
           // reuse a data
-            var input_offset = ${G.indicesToOffset(`${G.type.indices}(batch, row, word_offset)`)};
-            var a_data: ${J};
+            var input_offset = ${H.indicesToOffset(`${H.type.indices}(batch, row, word_offset)`)};
+            var a_data: ${ie};
             for (var j: u32 = 0; j < ${8/h}; j++) {
-              a_data[j] = ${G.getByOffset("input_offset")};
+              a_data[j] = ${H.getByOffset("input_offset")};
               input_offset++;
             }
-          `;for(let le=0;le<y*x;le++)$e+=`
-            b_value = ${_===1?`b${le}_data`:`b${le}_data[i]`};
+          `;for(let de=0;de<y*x;de++)ve+=`
+            b_value = ${_===1?`b${de}_data`:`b${de}_data[i]`};
             b_value_lower = unpack4xU8(b_value & b_mask);
             b_value_upper = unpack4xU8((b_value >> 4) & b_mask);
-            b_quantized_values = ${J}(${Array.from({length:4},(W,q)=>`${ee}(b_value_lower[${q}]), ${ee}(b_value_upper[${q}])`).join(", ")});
-            b_dequantized_values = ${(()=>h===1?`${J}(${Array.from({length:8},(W,q)=>`(b_quantized_values[${q}] - ${Q?`zero_point${le}`:"zero_point"}) * scale${le}`).join(", ")});`:`(b_quantized_values - ${J}(${Array(8).fill(`${Q?`zero_point${le}`:"zero_point"}`).join(",")})) * scale${le};`)()};
-            workgroup_shared[local_id.x * ${x} + ${Math.floor(le/y)}]${y>1?`[${le%y}]`:""} += ${Array.from({length:8/h},(W,q)=>`${h===1?`a_data[${q}] * b_dequantized_values[${q}]`:`dot(a_data[${q}], b_dequantized_values[${q}])`}`).join(" + ")};
-          `;return $e},be=()=>{let $e=`
+            b_quantized_values = ${ie}(${Array.from({length:4},(W,Y)=>`${Q}(b_value_lower[${Y}]), ${Q}(b_value_upper[${Y}])`).join(", ")});
+            b_dequantized_values = ${(()=>h===1?`${ie}(${Array.from({length:8},(W,Y)=>`(b_quantized_values[${Y}] - ${K?`zero_point${de}`:"zero_point"}) * scale${de}`).join(", ")});`:`(b_quantized_values - ${ie}(${Array(8).fill(`${K?`zero_point${de}`:"zero_point"}`).join(",")})) * scale${de};`)()};
+            workgroup_shared[local_id.x * ${x} + ${Math.floor(de/y)}]${y>1?`[${de%y}]`:""} += ${Array.from({length:8/h},(W,Y)=>`${h===1?`a_data[${Y}] * b_dequantized_values[${Y}]`:`dot(a_data[${Y}], b_dequantized_values[${Y}])`}`).join(" + ")};
+          `;return ve},be=()=>{let ve=`
             var col_index = col * ${y};
-            ${Q?`
+            ${K?`
             let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2;
             var zero_point_byte_count: u32;
             var zero_point_word_index: u32;
@@ -5772,29 +5837,29 @@ fn calculateOutputIndex(index: u32) -> u32 {
             var zero_point_bits_offset: u32;
             var zero_point_word: u32;`:`
             // The default zero point is 8 for unsigned 4-bit quantization.
-            let zero_point = ${ee}(8);`}
-            `;for(let le=0;le<y*x;le++)$e+=`
-            let scale${le} = ${j.getByOffset("col_index * nBlocksPerCol + block")};
-            ${Q?`
+            let zero_point = ${Q}(8);`}
+            `;for(let de=0;de<y*x;de++)ve+=`
+            let scale${de} = ${re.getByOffset("col_index * nBlocksPerCol + block")};
+            ${K?`
             zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u);
             zero_point_word_index = zero_point_byte_count >> 0x2u;
             zero_point_byte_offset = zero_point_byte_count & 0x3u;
             zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
-            zero_point_word = ${Q.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
-            let zero_point${le} = ${ee}((zero_point_word) & 0xFu);`:""}
-            col_index += 1;`;return $e},Oe=()=>{let $e=`col_index = col * ${y};`;for(let le=0;le<y*x;le++)$e+=`
-            let b${le}_data = ${K.getByIndices(`${K.type.indices}(col_index, block, word)`)};
-            col_index += 1;`;return $e+=`
+            zero_point_word = ${K.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
+            let zero_point${de} = ${Q}((zero_point_word) & 0xFu);`:""}
+            col_index += 1;`;return ve},Oe=()=>{let ve=`col_index = col * ${y};`;for(let de=0;de<y*x;de++)ve+=`
+            let b${de}_data = ${L.getByIndices(`${L.type.indices}(col_index, block, word)`)};
+            col_index += 1;`;return ve+=`
             var b_value: u32;
             let b_mask: u32 = 0x0F0F0F0Fu;
             var b_value_lower: vec4<u32>;
             var b_value_upper: vec4<u32>;
-            var b_quantized_values: ${J};
-            var b_dequantized_values: ${J};`,$e};return`
-        var<workgroup> workgroup_shared: array<${Y.type.value}, ${x*v}>;
-        ${D.declareVariables(...V,Y)}
+            var b_quantized_values: ${ie};
+            var b_dequantized_values: ${ie};`,ve};return`
+        var<workgroup> workgroup_shared: array<${j.type.value}, ${x*v}>;
+        ${D.declareVariables(...V,j)}
         ${D.mainStart([v,1,1])}
-          let output_indices = ${Y.offsetToIndices(`(global_idx / ${v}) * ${x}`)};
+          let output_indices = ${j.offsetToIndices(`(global_idx / ${v}) * ${x}`)};
           let col = output_indices[2];
           let row = output_indices[1];
           let batch = output_indices[0];
@@ -5807,7 +5872,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
             for (var word: u32 = 0; word < ${m}; word += ${_}) {
               ${Oe()}
               for (var i: u32 = 0; i < ${_}; i++) {
-                ${ne()}
+                ${te()}
                 word_offset += ${8/h};
               }
             }
@@ -5815,15 +5880,15 @@ fn calculateOutputIndex(index: u32) -> u32 {
           workgroupBarrier();
 
           if (local_id.x < ${x}) {
-            var output_value: ${Y.type.value} = ${Y.type.value}(0);
+            var output_value: ${j.type.value} = ${j.type.value}(0);
             var workgroup_shared_offset: u32 = local_id.x;
             for (var b: u32 = 0u; b < ${v}u; b++) {
               output_value += workgroup_shared[workgroup_shared_offset];
               workgroup_shared_offset += ${x};
             }
-            ${Y.setByIndices(`${Y.type.indices}(batch, row, col + local_id.x)`,"output_value")};
+            ${j.setByIndices(`${j.type.indices}(batch, row, col + local_id.x)`,"output_value")};
           }
-        }`};return{name:"MatMulNBits",shaderCache:{hint:`${t.blockSize};${t.bits};${h};${_};${y};${x};${v}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:g,dataType:u}],dispatchGroup:{x:$},programUniforms:S}),getShaderSource:P}},mh=(e,t)=>{let r=e[0].dims,n=r.length,o=r[n-2],i=t.k,a=t.n,d=r.slice(0,n-2),l=C.size(d),m=e[1].dims[2]/4,u=e[0].dataType,h=me(t.k),_=me(m),y=d.concat([o,a]),g=128,x=a%8===0?8:a%4===0?4:1,$=g/x,v=$*_*8,S=v/h,T=v/t.blockSize,A=C.size(y)/x,k=[],P=[l,o,i/h],D=C.convertShape(e[1].dims).slice();D.splice(-1,1,m/_),k.push(...N(P)),k.push(...N(D)),k.push(...N(e[2].dims)),e.length===4&&k.push(...N(C.convertShape(e[3].dims)));let R=[l,o,a];k.push(...N(R));let G=K=>{let j=P.length,V=E("a",e[0].dataType,j,h),Q=E("b",12,D.length,_),se=E("scales",e[2].dataType,e[2].dims.length),Y=[V,Q,se],ee=e.length===4?E("zero_points",12,e[3].dims.length):void 0;ee&&Y.push(ee);let J=R.length,ne=M("output",e[0].dataType,J),be=_e(e[0].dataType),Oe=()=>{switch(h){case 1:return`
+        }`};return{name:"MatMulNBits",shaderCache:{hint:`${t.blockSize};${t.bits};${h};${_};${y};${x};${v}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:g,dataType:u}],dispatchGroup:{x:$},programUniforms:S}),getShaderSource:P}},mh=(e,t)=>{let r=e[0].dims,n=r.length,o=r[n-2],i=t.k,a=t.n,d=r.slice(0,n-2),l=k.size(d),m=e[1].dims[2]/4,u=e[0].dataType,h=me(t.k),_=me(m),y=d.concat([o,a]),g=128,x=a%8===0?8:a%4===0?4:1,$=g/x,v=$*_*8,S=v/h,T=v/t.blockSize,A=k.size(y)/x,C=[],P=[l,o,i/h],D=k.convertShape(e[1].dims).slice();D.splice(-1,1,m/_),C.push(...N(P)),C.push(...N(D)),C.push(...N(e[2].dims)),e.length===4&&C.push(...N(k.convertShape(e[3].dims)));let R=[l,o,a];C.push(...N(R));let H=L=>{let re=P.length,V=E("a",e[0].dataType,re,h),K=E("b",12,D.length,_),we=E("scales",e[2].dataType,e[2].dims.length),j=[V,K,we],Q=e.length===4?E("zero_points",12,e[3].dims.length):void 0;Q&&j.push(Q);let ie=R.length,te=M("output",e[0].dataType,ie),be=ye(e[0].dataType),Oe=()=>{switch(h){case 1:return`
           let a_data0 = vec4<${be}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]);
           let a_data1 = vec4<${be}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return`
           let a_data0 = vec4<${be}>(sub_a[word_offset], sub_a[word_offset + 1]);
@@ -5831,10 +5896,10 @@ fn calculateOutputIndex(index: u32) -> u32 {
           let a_data0 = sub_a[word_offset];
           let a_data1 = sub_a[word_offset + 1];`;default:throw new Error(`${h}-component is not supported.`)}};return`
         var<workgroup> sub_a: array<${V.type.value}, ${S}>;
-        var<workgroup> inter_results: array<array<${ne.type.value}, ${$}>, ${x}>;
-        ${K.declareVariables(...Y,ne)}
-        ${K.mainStart([$,x,1])}
-          let output_indices = ${ne.offsetToIndices(`workgroup_index * ${x}`)};
+        var<workgroup> inter_results: array<array<${te.type.value}, ${$}>, ${x}>;
+        ${L.declareVariables(...j,te)}
+        ${L.mainStart([$,x,1])}
+          let output_indices = ${te.offsetToIndices(`workgroup_index * ${x}`)};
           let col = output_indices[2];
           let row = output_indices[1];
           let batch = output_indices[0];
@@ -5860,52 +5925,52 @@ fn calculateOutputIndex(index: u32) -> u32 {
             // each thread process one block
             let b_row = col + local_id.y;
             let block = tile * ${T} + local_id.x;
-            ${ee?`
+            ${Q?`
             let zero_point_bytes_per_col = (n_blocks_per_col + 1) / 2;
             let zero_point_byte_count = b_row * zero_point_bytes_per_col + (block >> 0x1u);
             let zero_point_word_index = zero_point_byte_count >> 0x2u;
             let zero_point_byte_offset = zero_point_byte_count & 0x3u;
             let zero_point_nibble_offset: u32 = block & 0x1u;
             let zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
-            let zero_point_word = ${ee.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
+            let zero_point_word = ${Q.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
             let zero_point = ${be}((zero_point_word) & 0xFu);`:`
             // The default zero point is 8 for unsigned 4-bit quantization.
             let zero_point = ${be}(8);`}
-            let scale = ${se.getByOffset("b_row * n_blocks_per_col + block")};
-            let b_data = ${Q.getByIndices(`${Q.type.indices}(b_row, block, 0)`)};
+            let scale = ${we.getByOffset("b_row * n_blocks_per_col + block")};
+            let b_data = ${K.getByIndices(`${K.type.indices}(b_row, block, 0)`)};
             var word_offset = local_id.x * ${t.blockSize/h};
             for (var i: u32 = 0; i < ${_}; i++) {
               ${Oe()}
               let b_value = ${_===1?"b_data":"b_data[i]"};
               let b_value_lower = unpack4xU8(b_value & 0x0F0F0F0Fu);
               let b_value_upper = unpack4xU8((b_value >> 4) & 0x0F0F0F0Fu);
-              let b_quantized_values = mat2x4<${be}>(${Array.from({length:4},($e,le)=>`${be}(b_value_lower[${le}]), ${be}(b_value_upper[${le}])`).join(", ")});
+              let b_quantized_values = mat2x4<${be}>(${Array.from({length:4},(ve,de)=>`${be}(b_value_lower[${de}]), ${be}(b_value_upper[${de}])`).join(", ")});
               let b_dequantized_values = (b_quantized_values - mat2x4<${be}>(${Array(8).fill("zero_point").join(",")})) * scale;
-              inter_results[local_id.y][local_id.x] += ${Array.from({length:2},($e,le)=>`${`dot(a_data${le}, b_dequantized_values[${le}])`}`).join(" + ")};
+              inter_results[local_id.y][local_id.x] += ${Array.from({length:2},(ve,de)=>`${`dot(a_data${de}, b_dequantized_values[${de}])`}`).join(" + ")};
               word_offset += ${8/h};
             }
             workgroupBarrier();
           }
 
           if (local_idx < ${x}) {
-            var output_value: ${ne.type.value} = ${ne.type.value}(0);
+            var output_value: ${te.type.value} = ${te.type.value}(0);
             for (var b = 0u; b < ${$}; b++) {
               output_value += inter_results[local_idx][b];
             }
             if (col + local_idx < uniforms.output_shape[2])
             {
-              ${ne.setByIndices(`${ne.type.indices}(batch, row, col + local_idx)`,"output_value")}
+              ${te.setByIndices(`${te.type.indices}(batch, row, col + local_idx)`,"output_value")}
             }
           }
-        }`};return{name:"BlockwiseMatMulNBits32",shaderCache:{hint:`${t.blockSize};${h};${_};${$};${x}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:y,dataType:u}],dispatchGroup:{x:A},programUniforms:k}),getShaderSource:G}},el=(e,t)=>{ch(e.inputs,t),t.blockSize===32&&e.adapterInfo.isVendor("intel")&&e.adapterInfo.isArchitecture("gen-12lp")?e.compute(mh(e.inputs,t)):e.compute(ph(e.inputs,t))},tl=e=>re(e)});var fh,hh,gh,bh,yh,_h,wh,vh,nl,ol=U(()=>{"use strict";te();oe();ae();fh=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1&&e[0].dataType!==10)throw new Error("Input type must be float or float16.");if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},hh=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
-            k = i32(${e.indicesGet("indices",o)}) - ${F("uniforms.pads",o,r)};
+        }`};return{name:"BlockwiseMatMulNBits32",shaderCache:{hint:`${t.blockSize};${h};${_};${$};${x}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:y,dataType:u}],dispatchGroup:{x:A},programUniforms:C}),getShaderSource:H}},Jd=(e,t)=>{ch(e.inputs,t),t.blockSize===32&&e.adapterInfo.isVendor("intel")&&e.adapterInfo.isArchitecture("gen-12lp")?e.compute(mh(e.inputs,t)):e.compute(ph(e.inputs,t))},el=e=>ee(e)});var fh,hh,gh,bh,yh,_h,wh,vh,rl,nl=U(()=>{"use strict";J();ne();ae();fh=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1&&e[0].dataType!==10)throw new Error("Input type must be float or float16.");if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},hh=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
+            k = i32(${e.indicesGet("indices",o)}) - ${q("uniforms.pads",o,r)};
             if (k < 0) {
               break;
             }
-            if (k >= i32(${F("uniforms.x_shape",o,t)})) {
+            if (k >= i32(${q("uniforms.x_shape",o,t)})) {
               break;
             }
-            offset += k * i32(${F("uniforms.x_strides",o,t)});
+            offset += k * i32(${q("uniforms.x_strides",o,t)});
         `;return`
           value = ${e.type.value}(uniforms.constant_value);
           for (var i = 0; i < 1; i++) {
@@ -5915,52 +5980,52 @@ fn calculateOutputIndex(index: u32) -> u32 {
             value = x[offset];
           }
       `},gh=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
-                k = i32(${e.indicesGet("indices",o)}) - ${F("uniforms.pads",o,r)};
+                k = i32(${e.indicesGet("indices",o)}) - ${q("uniforms.pads",o,r)};
                 if (k < 0) {
                   k = -k;
                 }
                 {
-                  let _2n_1 = 2 * (i32(${F("uniforms.x_shape",o,t)}) - 1);
+                  let _2n_1 = 2 * (i32(${q("uniforms.x_shape",o,t)}) - 1);
                   k = k % _2n_1;
-                  if(k >= i32(${F("uniforms.x_shape",o,t)})) {
+                  if(k >= i32(${q("uniforms.x_shape",o,t)})) {
                     k = _2n_1 - k;
                   }
                 }
-                offset += k * i32(${F("uniforms.x_strides",o,t)});
+                offset += k * i32(${q("uniforms.x_strides",o,t)});
             `;return`
               var offset = 0;
               var k = 0;
               ${n}
               value = x[offset];
           `},bh=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
-                k = i32(${e.indicesGet("indices",o)}) - ${F("uniforms.pads",o,r)};
+                k = i32(${e.indicesGet("indices",o)}) - ${q("uniforms.pads",o,r)};
                 if (k < 0) {
                   k = 0;
                 }
-                if (k >= i32(${F("uniforms.x_shape",o,t)})) {
-                  k = i32(${F("uniforms.x_shape",o,t)}) - 1;
+                if (k >= i32(${q("uniforms.x_shape",o,t)})) {
+                  k = i32(${q("uniforms.x_shape",o,t)}) - 1;
                 }
-                offset += k * i32(${F("uniforms.x_strides",o,t)});
+                offset += k * i32(${q("uniforms.x_strides",o,t)});
             `;return`
               var offset = 0;
               var k = 0;
               ${n}
               value = x[offset];
           `},yh=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
-                k = i32(${e.indicesGet("indices",o)}) - ${F("uniforms.pads",o,r)};
+                k = i32(${e.indicesGet("indices",o)}) - ${q("uniforms.pads",o,r)};
                 if (k < 0)  {
-                  k += i32(${F("uniforms.x_shape",o,t)}]);
+                  k += i32(${q("uniforms.x_shape",o,t)}]);
                 }
-                if (k >= i32(${F("uniforms.x_shape",o,t)})) {
-                  k -= i32(${F("uniforms.x_shape",o,t)});
+                if (k >= i32(${q("uniforms.x_shape",o,t)})) {
+                  k -= i32(${q("uniforms.x_shape",o,t)});
                 }
-                offset += k * i32(${F("uniforms.x_strides",o,t)});
+                offset += k * i32(${q("uniforms.x_strides",o,t)});
             `;return`
               var offset = 0;
               var k = 0;
               ${n}
               value = x[offset];
-          `},_h=(e,t,r)=>{switch(r.mode){case 0:return hh(e,t,r.pads.length);case 1:return gh(e,t,r.pads.length);case 2:return bh(e,t,r.pads.length);case 3:return yh(e,t,r.pads.length);default:throw new Error("Invalid mode")}},wh=(e,t)=>{let r=C.padShape(e[0].dims.slice(),t.pads),n=e[0].dims,o=C.size(r),i=[{type:12,data:o},{type:6,data:t.pads}],a=e.length>=3&&e[2].data;t.mode===0&&i.push({type:a?e[2].dataType:1,data:t.value}),i.push(...N(e[0].dims,r));let d=["rank"],l=p=>{let m=M("output",e[0].dataType,r.length),u=E("x",e[0].dataType,n.length),h=u.type.value,_=_h(m,n.length,t),y=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&y.push({name:"constant_value",type:a?h:"f32"}),`
+          `},_h=(e,t,r)=>{switch(r.mode){case 0:return hh(e,t,r.pads.length);case 1:return gh(e,t,r.pads.length);case 2:return bh(e,t,r.pads.length);case 3:return yh(e,t,r.pads.length);default:throw new Error("Invalid mode")}},wh=(e,t)=>{let r=k.padShape(e[0].dims.slice(),t.pads),n=e[0].dims,o=k.size(r),i=[{type:12,data:o},{type:6,data:t.pads}],a=e.length>=3&&e[2].data;t.mode===0&&i.push({type:a?e[2].dataType:1,data:t.value}),i.push(...N(e[0].dims,r));let d=["rank"],l=p=>{let m=M("output",e[0].dataType,r.length),u=E("x",e[0].dataType,n.length),h=u.type.value,_=_h(m,n.length,t),y=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&y.push({name:"constant_value",type:a?h:"f32"}),`
             ${p.registerUniforms(y).declareVariables(u,m)}
             ${p.mainStart()}
             ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -5970,7 +6035,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
             var value = ${h}(0);
             ${_}
             output[global_idx] = value;
-        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}${a}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(C.size(r)/64)},programUniforms:i}),getShaderSource:l}},vh=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),n=e.length>=3&&e[2].data?e[2].dataType===10?e[2].getUint16Array()[0]:e[2].getFloat32Array()[0]:0,o=e[0].dims.length,i=new Int32Array(2*o).fill(0);if(e.length>=4){let d=e[3].getBigInt64Array();for(let l=0;l<d.length;l++)i[Number(d[l])]=Number(r[l]),i[Number(d[l])+o]=Number(r[l+d.length])}else r.forEach((d,l)=>i[Number(l)]=Number(d));let a=[];return i.forEach(d=>a.push(d)),{mode:t.mode,value:n,pads:a}}else return t},nl=(e,t)=>{fh(e.inputs);let r=vh(e.inputs,t);e.compute(wh(e.inputs,r),{inputs:[0]})}});var tn,il,al,sl,ul,$h,xh,dl,ll,cl,pl,ml,fl,hl,gl,bl,yl,_l,wl,vl=U(()=>{"use strict";We();te();oe();ae();tn=e=>{if(ve.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},il=(e,t,r)=>{let n=t.format==="NHWC",o=e.dims.slice();n&&o.splice(1,0,o.pop());let i=Object.hasOwnProperty.call(t,"dilations"),a=t.kernelShape.slice(),d=t.strides.slice(),l=i?t.dilations.slice():[],p=t.pads.slice();At.adjustPoolAttributes(r,o,a,d,l,p);let m=At.computePoolOutputShape(r,o,d,l,a,p,t.autoPad),u=Object.assign({},t);i?Object.assign(u,{kernelShape:a,strides:d,pads:p,dilations:l,cacheKey:t.cacheKey}):Object.assign(u,{kernelShape:a,strides:d,pads:p,cacheKey:t.cacheKey});let h=m.slice();return h.push(h.splice(1,1)[0]),[u,n?h:m]},al=(e,t)=>{let r=t.format==="NHWC",n=C.size(e),o=C.size(t.kernelShape),i=[{type:12,data:n},{type:12,data:o}],a=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let d=t.kernelShape[t.kernelShape.length-1],l=t.strides[t.strides.length-1],p=t.pads[t.pads.length/2-1],m=t.pads[t.pads.length-1],u=!!(p+m);i.push({type:12,data:d},{type:12,data:l},{type:12,data:p},{type:12,data:m}),a.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let h=!1;if(t.kernelShape.length===2){let _=t.kernelShape[t.kernelShape.length-2],y=t.strides[t.strides.length-2],g=t.pads[t.pads.length/2-2],x=t.pads[t.pads.length-2];h=!!(g+x),i.push({type:12,data:_},{type:12,data:y},{type:12,data:g},{type:12,data:x}),a.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[i,a,!0,u,h]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let d=C.computeStrides(t.kernelShape);i.push({type:12,data:d},{type:12,data:t.pads},{type:12,data:t.strides}),a.push({name:"kernelStrides",type:"u32",length:d.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let l=t.pads.reduce((p,m)=>p+m);return[i,a,!!l,!1,!1]}},sl=(e,t,r,n,o,i,a,d,l,p,m,u)=>{let h=o.format==="NHWC",_=t.type.value,y=M("output",t.type.tensor,n);if(o.kernelShape.length<=2){let g="",x="",$="",v=r-(h?2:1);if(m?g=`
+        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}${a}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(k.size(r)/64)},programUniforms:i}),getShaderSource:l}},vh=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),n=e.length>=3&&e[2].data?e[2].dataType===10?e[2].getUint16Array()[0]:e[2].getFloat32Array()[0]:0,o=e[0].dims.length,i=new Int32Array(2*o).fill(0);if(e.length>=4){let d=e[3].getBigInt64Array();for(let l=0;l<d.length;l++)i[Number(d[l])]=Number(r[l]),i[Number(d[l])+o]=Number(r[l+d.length])}else r.forEach((d,l)=>i[Number(l)]=Number(d));let a=[];return i.forEach(d=>a.push(d)),{mode:t.mode,value:n,pads:a}}else return t},rl=(e,t)=>{fh(e.inputs);let r=vh(e.inputs,t);e.compute(wh(e.inputs,r),{inputs:[0]})}});var en,ol,il,al,sl,$h,xh,ul,dl,ll,cl,pl,ml,fl,hl,gl,bl,yl,_l,wl=U(()=>{"use strict";Le();J();ne();ae();en=e=>{if(_e.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},ol=(e,t,r)=>{let n=t.format==="NHWC",o=e.dims.slice();n&&o.splice(1,0,o.pop());let i=Object.hasOwnProperty.call(t,"dilations"),a=t.kernelShape.slice(),d=t.strides.slice(),l=i?t.dilations.slice():[],p=t.pads.slice();kt.adjustPoolAttributes(r,o,a,d,l,p);let m=kt.computePoolOutputShape(r,o,d,l,a,p,t.autoPad),u=Object.assign({},t);i?Object.assign(u,{kernelShape:a,strides:d,pads:p,dilations:l,cacheKey:t.cacheKey}):Object.assign(u,{kernelShape:a,strides:d,pads:p,cacheKey:t.cacheKey});let h=m.slice();return h.push(h.splice(1,1)[0]),[u,n?h:m]},il=(e,t)=>{let r=t.format==="NHWC",n=k.size(e),o=k.size(t.kernelShape),i=[{type:12,data:n},{type:12,data:o}],a=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let d=t.kernelShape[t.kernelShape.length-1],l=t.strides[t.strides.length-1],p=t.pads[t.pads.length/2-1],m=t.pads[t.pads.length-1],u=!!(p+m);i.push({type:12,data:d},{type:12,data:l},{type:12,data:p},{type:12,data:m}),a.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let h=!1;if(t.kernelShape.length===2){let _=t.kernelShape[t.kernelShape.length-2],y=t.strides[t.strides.length-2],g=t.pads[t.pads.length/2-2],x=t.pads[t.pads.length-2];h=!!(g+x),i.push({type:12,data:_},{type:12,data:y},{type:12,data:g},{type:12,data:x}),a.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[i,a,!0,u,h]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let d=k.computeStrides(t.kernelShape);i.push({type:12,data:d},{type:12,data:t.pads},{type:12,data:t.strides}),a.push({name:"kernelStrides",type:"u32",length:d.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let l=t.pads.reduce((p,m)=>p+m);return[i,a,!!l,!1,!1]}},al=(e,t,r,n,o,i,a,d,l,p,m,u)=>{let h=o.format==="NHWC",_=t.type.value,y=M("output",t.type.tensor,n);if(o.kernelShape.length<=2){let g="",x="",$="",v=r-(h?2:1);if(m?g=`
                 for (var i: u32 = 0u; i < uniforms.kw; i++) {
                   xIndices[${v}] = indices[${v}] * uniforms.sw - uniforms.pwStart + i;
                   if (xIndices[${v}] < 0 || xIndices[${v}]
@@ -6045,27 +6110,27 @@ fn calculateOutputIndex(index: u32) -> u32 {
               for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {
                 var offset = i;
                 for (var j = 0u; j < ${g-1}u; j++) {
-                  offsets[j] = offset / ${F("uniforms.kernelStrides","j",g)};
-                  offset -= offsets[j] * ${F("uniforms.kernelStrides","j",g)};
+                  offsets[j] = offset / ${q("uniforms.kernelStrides","j",g)};
+                  offset -= offsets[j] * ${q("uniforms.kernelStrides","j",g)};
                 }
                 offsets[${g-1}] = offset;
 
                 isPad = false;
                 for (var j = ${r-g}u; j < ${r}u; j++) {
-                  xIndices[j] = indices[j] * ${F("uniforms.strides",`j - ${r-g}u`,g)}
-                    + offsets[j - ${r-g}u] - ${F("uniforms.pads","j - 2u",x)};
+                  xIndices[j] = indices[j] * ${q("uniforms.strides",`j - ${r-g}u`,g)}
+                    + offsets[j - ${r-g}u] - ${q("uniforms.pads","j - 2u",x)};
                   ${$}
               }
               ${a}
 
               output[global_idx] = value;
-            }`}},ul=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,$h=e=>`${ul(e)};${e.countIncludePad}`,xh=e=>`${ul(e)};${e.storageOrder};${e.dilations}`,dl=e=>({format:e.format,autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],ceilMode:e.ceil_mode,kernelShape:e.kernel_shape,strides:e.strides,pads:e.pads}),ll=(e,t,r,n)=>{let[o,i]=il(t,n,r),a=E("x",t.dataType,t.dims.length),d=a.type.value,l="value += x_val;",p="";o.countIncludePad?p+=`value /= ${d}(uniforms.kernelSize);`:p+=`value /= ${d}(i32(uniforms.kernelSize) - pad);`;let[m,u,h,_,y]=al(i,o);m.push(...N(t.dims,i));let g=["rank"];return{name:e,shaderCache:{hint:`${n.cacheKey};${h};${_};${y}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(C.size(i)/64)},programUniforms:m}),getShaderSource:x=>sl(x,a,t.dims.length,i.length,o,l,p,0,u,h,_,y)}},cl=e=>{let t=e.count_include_pad!==0,r=dl(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let n={countIncludePad:t,...r,cacheKey:""};return{...n,cacheKey:$h(n)}},pl=(e,t)=>{tn(e.inputs),e.compute(ll("AveragePool",e.inputs[0],!1,t))},ml={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},fl=e=>{let t=e.format;return{format:t,...ml,cacheKey:t}},hl=(e,t)=>{tn(e.inputs),e.compute(ll("GlobalAveragePool",e.inputs[0],!0,t))},gl=(e,t,r,n)=>{let[o,i]=il(t,n,r),a=`
+            }`}},sl=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,$h=e=>`${sl(e)};${e.countIncludePad}`,xh=e=>`${sl(e)};${e.storageOrder};${e.dilations}`,ul=e=>({format:e.format,autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],ceilMode:e.ceil_mode,kernelShape:e.kernel_shape,strides:e.strides,pads:e.pads}),dl=(e,t,r,n)=>{let[o,i]=ol(t,n,r),a=E("x",t.dataType,t.dims.length),d=a.type.value,l="value += x_val;",p="";o.countIncludePad?p+=`value /= ${d}(uniforms.kernelSize);`:p+=`value /= ${d}(i32(uniforms.kernelSize) - pad);`;let[m,u,h,_,y]=il(i,o);m.push(...N(t.dims,i));let g=["rank"];return{name:e,shaderCache:{hint:`${n.cacheKey};${h};${_};${y}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(k.size(i)/64)},programUniforms:m}),getShaderSource:x=>al(x,a,t.dims.length,i.length,o,l,p,0,u,h,_,y)}},ll=e=>{let t=e.count_include_pad!==0,r=ul(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let n={countIncludePad:t,...r,cacheKey:""};return{...n,cacheKey:$h(n)}},cl=(e,t)=>{en(e.inputs),e.compute(dl("AveragePool",e.inputs[0],!1,t))},pl={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},ml=e=>{let t=e.format;return{format:t,...pl,cacheKey:t}},fl=(e,t)=>{en(e.inputs),e.compute(dl("GlobalAveragePool",e.inputs[0],!0,t))},hl=(e,t,r,n)=>{let[o,i]=ol(t,n,r),a=`
       value = max(x_val, value);
-    `,d="",l=E("x",t.dataType,t.dims.length),p=["rank"],[m,u,h,_,y]=al(i,o);return m.push(...N(t.dims,i)),{name:e,shaderCache:{hint:`${n.cacheKey};${h};${_};${y}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(C.size(i)/64)},programUniforms:m}),getShaderSource:g=>sl(g,l,t.dims.length,i.length,o,a,d,t.dataType===10?-65504:-1e5,u,h,_,y)}},bl=(e,t)=>{tn(e.inputs),e.compute(gl("MaxPool",e.inputs[0],!1,t))},yl=e=>{let t=e.storage_order,r=e.dilations,n=dl(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(n.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let o={storageOrder:t,dilations:r,...n,cacheKey:""};return{...o,cacheKey:xh(o)}},_l=e=>{let t=e.format;return{format:t,...ml,cacheKey:t}},wl=(e,t)=>{tn(e.inputs),e.compute(gl("GlobalMaxPool",e.inputs[0],!0,t))}});var Th,Ih,$l,xl,Sl=U(()=>{"use strict";te();oe();Se();ae();Th=(e,t)=>{if(e.length<2||e.length>3)throw new Error("DequantizeLinear requires 2 or 3 inputs.");if(e.length===3&&e[1].dims===e[2].dims)throw new Error("x-scale and x-zero-point must have the same shape.");if(e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("x and x-zero-point must have the same data type.");if(e[0].dataType===6&&e.length>2)throw new Error("In the case of dequantizing int32 there is no zero point.");if(e[1].dims.length!==0&&e[1].dims.length!==1&&e[1].dims.length!==e[0].dims.length)throw new Error("scale input must be a scalar, a 1D tensor, or have the same rank as the input tensor.");if(e.length>2){if(e[0].dataType!==e[2].dataType)throw new Error("x and x-zero-point must have the same data type.");if(e[1].dims.length!==e[2].dims.length)throw new Error("scale and zero-point inputs must have the same rank.");if(!e[1].dims.map((r,n)=>r===e[2].dims[n]).reduce((r,n)=>r&&n,!0))throw new Error("scale and zero-point inputs must have the same shape.")}if(t.blockSize>0){if(e[1].dims.length===0||e[1].dims.length===1&&e[1].dims[0]===1)throw new Error("blockSize must be set only for block quantization.");if(!e[1].dims.map((o,i)=>i===t.axis||o===e[0].dims[i]).reduce((o,i)=>o&&i,!0))throw new Error("For block qunatization, scale input shape to match the input shape except for the axis");if(e[1].dims.length!==e[0].dims.length)throw new Error("For block qunatization the scale input rank must be the same as the x rank.");let r=e[0].dims[t.axis],n=e[1].dims[t.axis];if(t.blockSize<Math.ceil(r/n)||t.blockSize>Math.ceil(r/(n-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},Ih=(e,t)=>{let r=C.normalizeAxis(t.axis,e[0].dims.length),n=e[0].dataType,o=n===3,i=e[0].dims,a=e[1].dataType,d=C.size(i),l=n===3||n===2,p=l?[Math.ceil(C.size(e[0].dims)/4)]:e[0].dims,m=e[1].dims,u=e.length>2?e[2]:void 0,h=u?l?[Math.ceil(C.size(u.dims)/4)]:u.dims:void 0,_=m.length===0||m.length===1&&m[0]===1,y=_===!1&&m.length===1,g=me(d),x=_&&(!l||g===4),$=x?g:1,v=x&&!l?g:1,S=E("input",l?12:n,p.length,v),T=E("scale",a,m.length),A=u?E("zero_point",l?12:n,h.length):void 0,k=M("output",a,i.length,$),P=[S,T];A&&P.push(A);let D=[p,m];u&&D.push(h);let R=[{type:12,data:d/$},{type:12,data:r},{type:12,data:t.blockSize},...N(...D,i)],G=K=>{let j=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
-      ${K.registerUniforms(j).declareVariables(...P,k)}
-      ${K.mainStart()}
-          ${K.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
-          let output_indices = ${k.offsetToIndices("global_idx")};
+    `,d="",l=E("x",t.dataType,t.dims.length),p=["rank"],[m,u,h,_,y]=il(i,o);return m.push(...N(t.dims,i)),{name:e,shaderCache:{hint:`${n.cacheKey};${h};${_};${y}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(k.size(i)/64)},programUniforms:m}),getShaderSource:g=>al(g,l,t.dims.length,i.length,o,a,d,t.dataType===10?-65504:-1e5,u,h,_,y)}},gl=(e,t)=>{en(e.inputs),e.compute(hl("MaxPool",e.inputs[0],!1,t))},bl=e=>{let t=e.storage_order,r=e.dilations,n=ul(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(n.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let o={storageOrder:t,dilations:r,...n,cacheKey:""};return{...o,cacheKey:xh(o)}},yl=e=>{let t=e.format;return{format:t,...pl,cacheKey:t}},_l=(e,t)=>{en(e.inputs),e.compute(hl("GlobalMaxPool",e.inputs[0],!0,t))}});var Th,Ih,vl,$l,xl=U(()=>{"use strict";J();ne();xe();ae();Th=(e,t)=>{if(e.length<2||e.length>3)throw new Error("DequantizeLinear requires 2 or 3 inputs.");if(e.length===3&&e[1].dims===e[2].dims)throw new Error("x-scale and x-zero-point must have the same shape.");if(e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("x and x-zero-point must have the same data type.");if(e[0].dataType===6&&e.length>2)throw new Error("In the case of dequantizing int32 there is no zero point.");if(e[1].dims.length!==0&&e[1].dims.length!==1&&e[1].dims.length!==e[0].dims.length)throw new Error("scale input must be a scalar, a 1D tensor, or have the same rank as the input tensor.");if(e.length>2){if(e[0].dataType!==e[2].dataType)throw new Error("x and x-zero-point must have the same data type.");if(e[1].dims.length!==e[2].dims.length)throw new Error("scale and zero-point inputs must have the same rank.");if(!e[1].dims.map((r,n)=>r===e[2].dims[n]).reduce((r,n)=>r&&n,!0))throw new Error("scale and zero-point inputs must have the same shape.")}if(t.blockSize>0){if(e[1].dims.length===0||e[1].dims.length===1&&e[1].dims[0]===1)throw new Error("blockSize must be set only for block quantization.");if(!e[1].dims.map((o,i)=>i===t.axis||o===e[0].dims[i]).reduce((o,i)=>o&&i,!0))throw new Error("For block qunatization, scale input shape to match the input shape except for the axis");if(e[1].dims.length!==e[0].dims.length)throw new Error("For block qunatization the scale input rank must be the same as the x rank.");let r=e[0].dims[t.axis],n=e[1].dims[t.axis];if(t.blockSize<Math.ceil(r/n)||t.blockSize>Math.ceil(r/(n-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},Ih=(e,t)=>{let r=k.normalizeAxis(t.axis,e[0].dims.length),n=e[0].dataType,o=n===3,i=e[0].dims,a=e[1].dataType,d=k.size(i),l=n===3||n===2,p=l?[Math.ceil(k.size(e[0].dims)/4)]:e[0].dims,m=e[1].dims,u=e.length>2?e[2]:void 0,h=u?l?[Math.ceil(k.size(u.dims)/4)]:u.dims:void 0,_=m.length===0||m.length===1&&m[0]===1,y=_===!1&&m.length===1,g=me(d),x=_&&(!l||g===4),$=x?g:1,v=x&&!l?g:1,S=E("input",l?12:n,p.length,v),T=E("scale",a,m.length),A=u?E("zero_point",l?12:n,h.length):void 0,C=M("output",a,i.length,$),P=[S,T];A&&P.push(A);let D=[p,m];u&&D.push(h);let R=[{type:12,data:d/$},{type:12,data:r},{type:12,data:t.blockSize},...N(...D,i)],H=L=>{let re=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
+      ${L.registerUniforms(re).declareVariables(...P,C)}
+      ${L.mainStart()}
+          ${L.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+          let output_indices = ${C.offsetToIndices("global_idx")};
 
           // Set input x
           ${(()=>l?`
@@ -6075,7 +6140,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
 
           // Set scale input
           ${(()=>_?`let scale_value= ${T.getByOffset("0")}`:y?`
-            let scale_index = ${k.indicesGet("output_indices","uniforms.axis")};
+            let scale_index = ${C.indicesGet("output_indices","uniforms.axis")};
             let scale_value= ${T.getByOffset("scale_index")};`:`
             var scale_indices: ${T.type.indices} = output_indices;
             let index = ${T.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
@@ -6087,24 +6152,24 @@ fn calculateOutputIndex(index: u32) -> u32 {
                 let zero_point_input = ${A.getByOffset("0")};
                 let zero_point_vec =  ${o?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
                 let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${A.getByOffset("0")}`:y?l?`
-                let zero_point_index = ${k.indicesGet("output_indices","uniforms.axis")};
+                let zero_point_index = ${C.indicesGet("output_indices","uniforms.axis")};
                 let zero_point_input = ${A.getByOffset("zero_point_index / 4")};
                 let zero_point_vec =  ${o?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
                 let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
-                let zero_point_index = ${k.indicesGet("output_indices","uniforms.axis")};
+                let zero_point_index = ${C.indicesGet("output_indices","uniforms.axis")};
                 let zero_point_value = ${A.getByOffset("zero_point_index")};`:l?`
                 let zero_point_offset = ${T.indicesToOffset("scale_indices")};
                 let zero_point_input = ${A.getByOffset("zero_point_offset / 4")};
                 let zero_point_vec = ${o?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
                 let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${A.getByIndices("scale_indices")};`:`let zero_point_value = ${l?o?"i32":"u32":S.type.value}(0);`)()};
       // Compute and write output
-      ${k.setByOffset("global_idx",`${k.type.value}(x_value - zero_point_value) * scale_value`)};
-      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:A?["rank","rank","rank"]:["rank","rank"]},getShaderSource:G,getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:Math.ceil(d/$/64),y:1,z:1},programUniforms:R})}},$l=(e,t)=>{Th(e.inputs,t),e.compute(Ih(e.inputs,t))},xl=e=>re({axis:e.axis,blockSize:e.blockSize})});var Ch,Ah,Tl,Il=U(()=>{"use strict";We();te();ae();Ch=(e,t,r)=>{let n=e===t,o=e<t&&r<0,i=e>t&&r>0;if(n||o||i)throw new Error("Range these inputs' contents are invalid.")},Ah=(e,t,r,n)=>{let o=Math.abs(Math.ceil((t-e)/r)),i=[o],a=o,d=[{type:12,data:a},{type:n,data:e},{type:n,data:r},...N(i)],l=p=>{let m=M("output",n,i.length),u=m.type.value,h=[{name:"outputSize",type:"u32"},{name:"start",type:u},{name:"delta",type:u}];return`
+      ${C.setByOffset("global_idx",`${C.type.value}(x_value - zero_point_value) * scale_value`)};
+      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:A?["rank","rank","rank"]:["rank","rank"]},getShaderSource:H,getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:Math.ceil(d/$/64),y:1,z:1},programUniforms:R})}},vl=(e,t)=>{Th(e.inputs,t),e.compute(Ih(e.inputs,t))},$l=e=>ee({axis:e.axis,blockSize:e.blockSize})});var Ch,Ah,Sl,Tl=U(()=>{"use strict";Le();J();ae();Ch=(e,t,r)=>{let n=e===t,o=e<t&&r<0,i=e>t&&r>0;if(n||o||i)throw new Error("Range these inputs' contents are invalid.")},Ah=(e,t,r,n)=>{let o=Math.abs(Math.ceil((t-e)/r)),i=[o],a=o,d=[{type:12,data:a},{type:n,data:e},{type:n,data:r},...N(i)],l=p=>{let m=M("output",n,i.length),u=m.type.value,h=[{name:"outputSize",type:"u32"},{name:"start",type:u},{name:"delta",type:u}];return`
         ${p.registerUniforms(h).declareVariables(m)}
         ${p.mainStart()}
         ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
         output[global_idx] = uniforms.start + ${u}(global_idx) * uniforms.delta;
-      }`};return{name:"Range",shaderCache:{hint:`${n}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:i,dataType:n}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d})}},Tl=e=>{let t=0,r=0,n=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],n=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],n=e.inputs[2].getFloat32Array()[0]),ve.webgpu.validateInputContent&&Ch(t,r,n),e.compute(Ah(t,r,n,e.inputs[0].dataType),{inputs:[]})}});var kh,Eh,Cl,Al,kl=U(()=>{"use strict";te();oe();Se();ae();kh=(e,t,r,n)=>{if(e!=="none"&&n!=="i32"&&n!=="u32"&&n!=="f32")throw new Error(`Input ${n} is not supported with reduction ${e}.`);let o=`{
+      }`};return{name:"Range",shaderCache:{hint:`${n}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:i,dataType:n}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d})}},Sl=e=>{let t=0,r=0,n=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],n=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],n=e.inputs[2].getFloat32Array()[0]),_e.webgpu.validateInputContent&&Ch(t,r,n),e.compute(Ah(t,r,n,e.inputs[0].dataType),{inputs:[]})}});var kh,Eh,Il,Cl,Al=U(()=>{"use strict";J();ne();xe();ae();kh=(e,t,r,n)=>{if(e!=="none"&&n!=="i32"&&n!=="u32"&&n!=="f32")throw new Error(`Input ${n} is not supported with reduction ${e}.`);let o=`{
                 var oldValue = 0;
                 loop {
                   let newValueF32 =`,i=`;
@@ -6117,7 +6182,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
                 }
               }`;switch(e){case"none":return`${t}=${r};`;case"add":return n==="i32"||n==="u32"?`atomicAdd(&${t}, bitcast<${n}>(${r}));`:`
               ${o}bitcast<${n}>(oldValue) + (${r})${i}`;case"max":return n==="i32"||n==="u32"?`atomicMax(&${t}, bitcast<${n}>(${r}));`:`
-                ${o}max(bitcast<f32>(oldValue), (${r}))${i}`;case"min":return n==="i32"||n==="u32"?`atomicMin(&${t}, bitcast<${n}>(${r}));`:`${o}min(bitcast<${n}>(oldValue), (${r}))${i}`;case"mul":return`${o}(bitcast<${n}>(oldValue) * (${r}))${i}`;default:throw new Error(`Reduction ${e} is not supported.`)}},Eh=(e,t)=>{let r=e[0].dims,n=e[1].dims,o=r,i=1,a=Math.ceil(C.size(n)/i),d=n[n.length-1],l=C.sizeFromDimension(r,d),p=[{type:12,data:a},{type:12,data:d},{type:12,data:l},...N(e[1].dims,e[2].dims,o)],m=u=>{let h=E("indices",e[1].dataType,e[1].dims.length),_=E("updates",e[2].dataType,e[2].dims.length,i),y=t.reduction!=="none"&&t.reduction!==""?rs("output",e[0].dataType,o.length):M("output",e[0].dataType,o.length,i);return`
+                ${o}max(bitcast<f32>(oldValue), (${r}))${i}`;case"min":return n==="i32"||n==="u32"?`atomicMin(&${t}, bitcast<${n}>(${r}));`:`${o}min(bitcast<${n}>(oldValue), (${r}))${i}`;case"mul":return`${o}(bitcast<${n}>(oldValue) * (${r}))${i}`;default:throw new Error(`Reduction ${e} is not supported.`)}},Eh=(e,t)=>{let r=e[0].dims,n=e[1].dims,o=r,i=1,a=Math.ceil(k.size(n)/i),d=n[n.length-1],l=k.sizeFromDimension(r,d),p=[{type:12,data:a},{type:12,data:d},{type:12,data:l},...N(e[1].dims,e[2].dims,o)],m=u=>{let h=E("indices",e[1].dataType,e[1].dims.length),_=E("updates",e[2].dataType,e[2].dims.length,i),y=t.reduction!=="none"&&t.reduction!==""?ts("output",e[0].dataType,o.length):M("output",e[0].dataType,o.length,i);return`
       ${u.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(h,_,y)}
       ${u.mainStart()}
         ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -6150,22 +6215,30 @@ fn calculateOutputIndex(index: u32) -> u32 {
     ${kh(t.reduction,"output[data_offset + i]","value",y.type.value)}
   }
 
-      }`};return{name:"ScatterND",shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p}),getShaderSource:m}},Cl=e=>re({reduction:e.reduction}),Al=(e,t)=>{e.compute(Eh(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}});var Ph,zh,Oh,Dh,Bh,Mh,Rh,Uh,Nh,Vh,Wh,El,Lh,Gh,Hh,Fh,qh,Pl,zl,Ol=U(()=>{"use strict";te();oe();Se();ae();Ph=(e,t)=>{if(e.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),e.length>0){if(t.mode==="linear"){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and
-            one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(t.mode==="cubic"&&!(e.length===2||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},zh=(e,t,r)=>{t.every(o=>o>=0&&o<r||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let n=new Array(r).fill(1);return t.forEach((o,i)=>n[o]=e[i]),n},Oh=(e,t,r,n,o,i)=>{let[a,d,l]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],p=e[0].dims.length;if(a>0&&e.length>a&&e[a].dims.length>0)e[a].getFloat32Array().forEach(m=>i.push(m));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(d>0&&e.length>d&&e[d].dims.length===1&&e[d].dims[0]>0){if(e[d].getFloat32Array().forEach(m=>n.push(m)),n.length!==0&&n.length!==p&&r>=18&&n.length!==t.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");Ph(n,t),t.axes.length>0&&zh(n,t.axes,p).forEach((m,u)=>n[u]=m)}if(l>0&&e.length>l&&e[l].dims.length===1&&e[l].dims[0]>0&&(e[l].getBigInt64Array().forEach(m=>o.push(Number(m))),o.length!==0&&o.length!==p&&r>=18&&o.length!==t.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(t.axes.length>0){if(n.length!==0&&n.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(o.length!==0&&o.length!==t.axes.length)throw new Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if(typeof n<"u"&&typeof o<"u"&&n.length>0&&o.length>p)throw new Error("Resize requires only of scales or sizes to be specified")},Dh=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
-     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`return ${t}(xResized) / ${t}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) {
+      }`};return{name:"ScatterND",shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p}),getShaderSource:m}},Il=e=>ee({reduction:e.reduction}),Cl=(e,t)=>{e.compute(Eh(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}});var Ph,zh,Oh,kl,Dh,Bh,Mh,Rh,Uh,Nh,Vh,Wh,El,Lh,Gh,Hh,Fh,qh,Pl,zl,Ol=U(()=>{"use strict";J();ne();xe();ae();Ph=(e,t)=>{if(e.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),e.length>0){if(t.mode==="linear"){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and
+            one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(t.mode==="cubic"&&!(e.length===2||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},zh=(e,t,r)=>{t.every(o=>o>=0&&o<r||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let n=new Array(r).fill(1);return t.forEach((o,i)=>n[o]=e[i]),n},Oh=(e,t,r,n,o,i)=>{let[a,d,l]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],p=e[0].dims.length;if(a>0&&e.length>a&&e[a].dims.length>0)e[a].getFloat32Array().forEach(m=>i.push(m));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(d>0&&e.length>d&&e[d].dims.length===1&&e[d].dims[0]>0){if(e[d].getFloat32Array().forEach(m=>n.push(m)),n.length!==0&&n.length!==p&&r>=18&&n.length!==t.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");Ph(n,t),t.axes.length>0&&zh(n,t.axes,p).forEach((m,u)=>n[u]=m)}if(l>0&&e.length>l&&e[l].dims.length===1&&e[l].dims[0]>0&&(e[l].getBigInt64Array().forEach(m=>o.push(Number(m))),o.length!==0&&o.length!==p&&r>=18&&o.length!==t.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(t.axes.length>0){if(n.length!==0&&n.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(o.length!==0&&o.length!==t.axes.length)throw new Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if(typeof n<"u"&&typeof o<"u"&&n.length>0&&o.length>p)throw new Error("Resize requires only of scales or sizes to be specified")},kl=(e,t,r,n)=>`
+  // The whole part and the fractional part are calculated separately due to inaccuracy of floating
+  // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
+  // offset-by-one error later in floor().
+  let big = (${e}) * (${t});
+  let whole = ${n}(big / (${r}));
+  let fract = ${n}(big % (${r})) / ${n}(${r});
+  return whole + fract;
+`,Dh=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
+     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`
+          if (xScale < 1.0 || floor(xScale) != xScale) {
+            return ${t}(xResized) / ${t}(xScale);
+          } else {
+            ${kl("xResized","lengthOriginal","lengthResized",t)}
+          }
+        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
                     return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
                   } else {
                     return 0.0;
                   }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
                     return 0.0;
                   } else {
-                    // The whole part and the fractional part are calculated separately due to inaccuracy of floating
-                    // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
-                    // offset-by-one error later in floor().
-                    let whole = ${t}(xResized * (lengthOriginal - 1) / (lengthResized - 1));
-                    let fract =
-                        ${t}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${t}(lengthResized - 1);
-                    return whole + fract;
+                    ${kl("xResized","lengthOriginal - 1","lengthResized - 1",t)}
                   }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
                     return ${t}(roiStart) * ${t}(lengthOriginal - 1) +
                         (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /
@@ -6181,14 +6254,14 @@ fn calculateOutputIndex(index: u32) -> u32 {
       var original_indices: array<${e.type.value}, ${r.length}>;
       for (var i:u32 = 0; i < ${r.length}; i++) {
         var output_index = ${e.indicesGet("output_indices","i")};
-        var scale = ${F("uniforms.scales","i",n)};
-        var roi_low = ${F("uniforms.roi","i",o)};
-        var roi_hi = ${F("uniforms.roi",`i + ${t.length}`,o)};
+        var scale = ${q("uniforms.scales","i",n)};
+        var roi_low = ${q("uniforms.roi","i",o)};
+        var roi_hi = ${q("uniforms.roi",`i + ${t.length}`,o)};
         if (scale == 1.0) {
           original_indices[i] = ${e.type.value}(output_index);
         } else {
-          var input_shape_i = ${F("uniforms.input_shape","i",t.length)};
-          var output_shape_i = ${F("uniforms.output_shape","i",r.length)};
+          var input_shape_i = ${q("uniforms.input_shape","i",t.length)};
+          var output_shape_i = ${q("uniforms.output_shape","i",r.length)};
           original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
                                                                            input_shape_i, roi_low, roi_hi);
         }
@@ -6200,14 +6273,14 @@ fn calculateOutputIndex(index: u32) -> u32 {
       for (var i:u32 = 0; i < ${n.length}; i++) {
         var output_index = ${t.indicesGet("output_indices","i")};
         var input_index: u32;
-        var scale = ${F("uniforms.scales","i",o)};
+        var scale = ${q("uniforms.scales","i",o)};
         if (scale == 1.0) {
           input_index = output_index;
         } else {
-          var roi_low = ${F("uniforms.roi","i",i)};
-          var roi_hi = ${F("uniforms.roi",`i + ${r.length}`,i)};
-          var input_shape_i = ${F("uniforms.input_shape","i",r.length)};
-          var output_shape_i = ${F("uniforms.output_shape","i",n.length)};
+          var roi_low = ${q("uniforms.roi","i",i)};
+          var roi_hi = ${q("uniforms.roi",`i + ${r.length}`,i)};
+          var input_shape_i = ${q("uniforms.input_shape","i",r.length)};
+          var output_shape_i = ${q("uniforms.output_shape","i",n.length)};
           var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
                                                                         input_shape_i, roi_low, roi_hi);
           if (!${a} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {
@@ -6222,14 +6295,14 @@ fn calculateOutputIndex(index: u32) -> u32 {
             input_index = u32(original_idx);
           }
         }
-        ${e.indicesSet("input_indices","i"," input_index")}
+        ${e.indicesSet("input_indices","i","input_index")}
       }
       return input_indices;
     }`,Wh=(e,t)=>`
     fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {
       for (var i:u32 = 0; i < ${t.length}; i++) {
         var input_index = ${e.indicesGet("input_indices","i")};
-        if (input_index < 0 || input_index >= ${F("uniforms.input_shape","i",t.length)}) {
+        if (input_index < 0 || input_index >= ${q("uniforms.input_shape","i",t.length)}) {
           return false;
         }
       }
@@ -6385,7 +6458,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
       }
       return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +
               x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);
-    }`},Fh=(e,t,r,n,o,i)=>{let a=e.dims,d=Mh(i,t.axes,a.length),l=Rh(a,n,o,t.axes),p=n.slice();n.length===0&&(p=a.map((v,S)=>v===0?1:l[S]/v),t.keepAspectRatioPolicy!=="stretch"&&(l=Uh(a,p,t)));let m=M("output",e.dataType,l.length),u=E("input",e.dataType,a.length),h=C.size(l),_=a.length===l.length&&a.every((v,S)=>v===l[S]),y=t.coordinateTransformMode==="tf_crop_and_resize",g=t.extrapolationValue,x=u.type.value,$=v=>`
+    }`},Fh=(e,t,r,n,o,i)=>{let a=e.dims,d=Mh(i,t.axes,a.length),l=Rh(a,n,o,t.axes),p=n.slice();n.length===0&&(p=a.map((v,S)=>v===0?1:l[S]/v),t.keepAspectRatioPolicy!=="stretch"&&(l=Uh(a,p,t)));let m=M("output",e.dataType,l.length),u=E("input",e.dataType,a.length),h=k.size(l),_=a.length===l.length&&a.every((v,S)=>v===l[S]),y=t.coordinateTransformMode==="tf_crop_and_resize",g=t.extrapolationValue,x=u.type.value,$=v=>`
       ${_?"":`
       ${Dh(t.coordinateTransformMode,x)};
       ${(()=>{switch(t.mode){case"nearest":return`
@@ -6412,10 +6485,10 @@ fn calculateOutputIndex(index: u32) -> u32 {
                   output[global_idx] = ${t.extrapolationValue};
                 }`;case"linear":return`output[global_idx] = ${a.length===2||a.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};
 `}
-      }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${p.length>0?p:""}|${o.length>0?o:""}|${d.length>0?d:""}|${_}|${a}`,inputDependencies:["rank"]},getShaderSource:$,getRunData:()=>({outputs:[{dims:l,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:[{type:12,data:h},{type:1,data:p},{type:1,data:d},...N(a,l)]})}},qh=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},Pl=(e,t)=>{let r=[],n=[],o=[],i=qh(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");Oh(e.inputs,t,i,r,n,o),e.compute(Fh(e.inputs[0],t,i,r,n,o),{inputs:[0]})},zl=e=>{let t=e.antialias,r=e.axes,n=e.coordinateTransformMode,o=e.cubicCoeffA,i=e.excludeOutside!==0,a=e.extrapolationValue,d=e.keepAspectRatioPolicy,l=e.mode,p=e.nearestMode===""?"simple":e.nearestMode;return re({antialias:t,axes:r,coordinateTransformMode:n,cubicCoeffA:o,excludeOutside:i,extrapolationValue:a,keepAspectRatioPolicy:d,mode:l,nearestMode:p})}});var Kh,jh,Dl,Bl=U(()=>{"use strict";te();oe();Se();ae();Kh=(e,t)=>{let[r,n,o,i]=e,{numHeads:a,rotaryEmbeddingDim:d}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!C.areEqual(n.dims,[])&&!C.areEqual(n.dims,[1])&&n.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${n.dims.length}`);if(o.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${o.dims.length}`);if(i.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${i.dims.length}`);if(!C.areEqual(o.dims,i.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(d>0&&a===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let l=r.dims[0],p=r.dims[r.dims.length-2],m=o.dims[0],u=C.sizeFromDimension(r.dims,1)/p,h=d===0?o.dims[1]*2:u/a;if(d>h)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(n.dims.length===2){if(l!==n.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${n.dims[0]}`);if(p!==n.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${n.dims[1]}`)}if(h/2!==o.dims[1]&&d/2!==o.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${o.dims[1]}`);if(p>m)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},jh=(e,t)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:o,scale:i}=t,a=e[0].dims[0],d=C.sizeFromDimension(e[0].dims,1),l=e[0].dims[e[0].dims.length-2],p=d/l,m=e[2].dims[1],u=o===0?m*2:p/n,h=new Array(a,l,p/u,u-m),_=C.computeStrides(h),y=[{type:1,data:i},{type:12,data:h},{type:12,data:_},...e[0].dims.length===3?new Array({type:12,data:[d,p,u,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[d,u,l*u,1]}):[],...N(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],g=x=>{let $=E("input",e[0].dataType,e[0].dims.length),v=E("position_ids",e[1].dataType,e[1].dims.length),S=E("cos_cache",e[2].dataType,e[2].dims.length),T=E("sin_cache",e[3].dataType,e[3].dims.length),A=M("output",e[0].dataType,e[0].dims.length);return x.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:h.length},{name:"global_strides",type:"u32",length:_.length},{name:"input_output_strides",type:"u32",length:_.length}]),`
+      }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${p.length>0?t.mode==="cubic"?p:p.length:""}|${o.length>0?o:""}|${d.length>0?d:""}|${_}|${t.mode==="nearest"?a.length:a}`,inputDependencies:["rank"]},getShaderSource:$,getRunData:()=>({outputs:[{dims:l,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:[{type:12,data:h},{type:1,data:p},{type:1,data:d},...N(a,l)]})}},qh=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},Pl=(e,t)=>{let r=[],n=[],o=[],i=qh(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");Oh(e.inputs,t,i,r,n,o),e.compute(Fh(e.inputs[0],t,i,r,n,o),{inputs:[0]})},zl=e=>{let t=e.antialias,r=e.axes,n=e.coordinateTransformMode,o=e.cubicCoeffA,i=e.excludeOutside!==0,a=e.extrapolationValue,d=e.keepAspectRatioPolicy,l=e.mode,p=e.nearestMode===""?"simple":e.nearestMode;return ee({antialias:t,axes:r,coordinateTransformMode:n,cubicCoeffA:o,excludeOutside:i,extrapolationValue:a,keepAspectRatioPolicy:d,mode:l,nearestMode:p})}});var Kh,jh,Dl,Bl=U(()=>{"use strict";J();ne();xe();ae();Kh=(e,t)=>{let[r,n,o,i]=e,{numHeads:a,rotaryEmbeddingDim:d}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!k.areEqual(n.dims,[])&&!k.areEqual(n.dims,[1])&&n.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${n.dims.length}`);if(o.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${o.dims.length}`);if(i.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${i.dims.length}`);if(!k.areEqual(o.dims,i.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(d>0&&a===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let l=r.dims[0],p=r.dims[r.dims.length-2],m=o.dims[0],u=k.sizeFromDimension(r.dims,1)/p,h=d===0?o.dims[1]*2:u/a;if(d>h)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(n.dims.length===2){if(l!==n.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${n.dims[0]}`);if(p!==n.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${n.dims[1]}`)}if(h/2!==o.dims[1]&&d/2!==o.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${o.dims[1]}`);if(p>m)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},jh=(e,t)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:o,scale:i}=t,a=e[0].dims[0],d=k.sizeFromDimension(e[0].dims,1),l=e[0].dims[e[0].dims.length-2],p=d/l,m=e[2].dims[1],u=o===0?m*2:p/n,h=new Array(a,l,p/u,u-m),_=k.computeStrides(h),y=[{type:1,data:i},{type:12,data:h},{type:12,data:_},...e[0].dims.length===3?new Array({type:12,data:[d,p,u,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[d,u,l*u,1]}):[],...N(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],g=x=>{let $=E("input",e[0].dataType,e[0].dims.length),v=E("position_ids",e[1].dataType,e[1].dims.length),S=E("cos_cache",e[2].dataType,e[2].dims.length),T=E("sin_cache",e[3].dataType,e[3].dims.length),A=M("output",e[0].dataType,e[0].dims.length);return x.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:h.length},{name:"global_strides",type:"u32",length:_.length},{name:"input_output_strides",type:"u32",length:_.length}]),`
         ${x.declareVariables($,v,S,T,A)}
 
-        ${x.mainStart(kt)}
+        ${x.mainStart(Et)}
           let half_rotary_emb_dim = uniforms.${S.name}_shape[1];
           let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
           let size = uniforms.global_shape[0] * uniforms.global_strides[0];
@@ -6438,9 +6511,9 @@ fn calculateOutputIndex(index: u32) -> u32 {
             let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;
             ${A.setByOffset("k",$.getByOffset("k"))}
           }
-        }`};return{name:"RotaryEmbedding",shaderCache:{hint:re({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:g,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(C.size(h)/kt)},programUniforms:y})}},Dl=(e,t)=>{Kh(e.inputs,t),e.compute(jh(e.inputs,t))}});var Yh,Zh,Ml,Rl=U(()=>{"use strict";te();oe();ae();Yh=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],n=e[2];if(t.dataType!==r.dataType||t.dataType!==n.dataType)throw new Error("All inputs must have the same data type");if(t.dims.length!==3&&t.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let o=t.dims[t.dims.length-1],i=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==o)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==i)throw new Error("Skip must have the same sequence length as input");if(n.dims.length!==1)throw new Error("Gamma must be 1D");if(n.dims[n.dims.length-1]!==o)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let a=e[3];if(a.dims.length!==1)throw new Error("Beta must be 1D");if(a.dims[a.dims.length-1]!==o)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let a=e[4];if(a.dims.length!==1)throw new Error("Bias must be 1D");if(a.dims[a.dims.length-1]!==o)throw new Error("Bias must have the same hidden size as input")}},Zh=(e,t,r,n)=>{let o=t.simplified,i=e[0].dims,a=C.size(i),d=i,l=a,p=i.slice(-1)[0],m=n?i.slice(0,-1).concat(1):[],u=!o&&e.length>3,h=e.length>4,_=n&&r>1,y=n&&r>2,g=r>3,x=64,$=me(p),v=[{type:12,data:l},{type:12,data:$},{type:12,data:p},{type:1,data:t.epsilon}],S=A=>{let k=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],P=[E("x",e[0].dataType,e[0].dims,$),E("skip",e[1].dataType,e[1].dims,$),E("gamma",e[2].dataType,e[2].dims,$)];u&&P.push(E("beta",e[3].dataType,e[3].dims,$)),h&&P.push(E("bias",e[4].dataType,e[4].dims,$)),P.push(M("output",e[0].dataType,d,$)),_&&P.push(M("mean_output",1,m)),y&&P.push(M("inv_std_output",1,m)),g&&P.push(M("input_skip_bias_sum",e[0].dataType,d,$));let D=_e(e[0].dataType),R=_e(1,$);return`
+        }`};return{name:"RotaryEmbedding",shaderCache:{hint:ee({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:g,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(k.size(h)/Et)},programUniforms:y})}},Dl=(e,t)=>{Kh(e.inputs,t),e.compute(jh(e.inputs,t))}});var Yh,Zh,Ml,Rl=U(()=>{"use strict";J();ne();ae();Yh=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],n=e[2];if(t.dataType!==r.dataType||t.dataType!==n.dataType)throw new Error("All inputs must have the same data type");if(t.dims.length!==3&&t.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let o=t.dims[t.dims.length-1],i=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==o)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==i)throw new Error("Skip must have the same sequence length as input");if(n.dims.length!==1)throw new Error("Gamma must be 1D");if(n.dims[n.dims.length-1]!==o)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let a=e[3];if(a.dims.length!==1)throw new Error("Beta must be 1D");if(a.dims[a.dims.length-1]!==o)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let a=e[4];if(a.dims.length!==1)throw new Error("Bias must be 1D");if(a.dims[a.dims.length-1]!==o)throw new Error("Bias must have the same hidden size as input")}},Zh=(e,t,r,n)=>{let o=t.simplified,i=e[0].dims,a=k.size(i),d=i,l=a,p=i.slice(-1)[0],m=n?i.slice(0,-1).concat(1):[],u=!o&&e.length>3,h=e.length>4,_=n&&r>1,y=n&&r>2,g=r>3,x=64,$=me(p),v=[{type:12,data:l},{type:12,data:$},{type:12,data:p},{type:1,data:t.epsilon}],S=A=>{let C=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],P=[E("x",e[0].dataType,e[0].dims,$),E("skip",e[1].dataType,e[1].dims,$),E("gamma",e[2].dataType,e[2].dims,$)];u&&P.push(E("beta",e[3].dataType,e[3].dims,$)),h&&P.push(E("bias",e[4].dataType,e[4].dims,$)),P.push(M("output",e[0].dataType,d,$)),_&&P.push(M("mean_output",1,m)),y&&P.push(M("inv_std_output",1,m)),g&&P.push(M("input_skip_bias_sum",e[0].dataType,d,$));let D=ye(e[0].dataType),R=ye(1,$);return`
 
-      ${A.registerUniforms(k).declareVariables(...P)}
+      ${A.registerUniforms(C).declareVariables(...P)}
       var<workgroup> sum_shared : array<${R}, ${x}>;
       var<workgroup> sum_squared_shared : array<${R}, ${x}>;
 
@@ -6462,7 +6535,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
           let value = input_value + skip_value + bias_value;
           ${g?"input_skip_bias_sum[offset + i] = value;":""}
           output[offset + i] = value;
-          let f32_value = ${Et(D,$,"value")};
+          let f32_value = ${Pt(D,$,"value")};
           sum_shared[ix] += f32_value;
           sum_squared_shared[ix] += f32_value * f32_value;
         }
@@ -6480,8 +6553,8 @@ fn calculateOutputIndex(index: u32) -> u32 {
 
         let sum = sum_shared[0];
         let square_sum = sum_squared_shared[0];
-        let mean = ${Fe("sum",$)} / f32(uniforms.hidden_size);
-        let inv_std_dev = inverseSqrt(${Fe("square_sum",$)} / f32(uniforms.hidden_size) ${o?"":"- mean * mean"} + uniforms.epsilon);
+        let mean = ${qe("sum",$)} / f32(uniforms.hidden_size);
+        let inv_std_dev = inverseSqrt(${qe("square_sum",$)} / f32(uniforms.hidden_size) ${o?"":"- mean * mean"} + uniforms.epsilon);
         ${_?"mean_output[global_idx] = mean;":""}
         ${y?"inv_std_output[global_idx] = inv_std_dev;":""}
 
@@ -6490,14 +6563,14 @@ fn calculateOutputIndex(index: u32) -> u32 {
             ${D}(inv_std_dev) * gamma[offset1d + i]
             ${u?"+ beta[offset1d + i]":""};
         }
-      }`},T=[{dims:d,dataType:e[0].dataType}];return r>1&&T.push({dims:m,dataType:1}),r>2&&T.push({dims:m,dataType:1}),r>3&&T.push({dims:i,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${$};${_};${y};${g}`,inputDependencies:e.map((A,k)=>"type")},getShaderSource:S,getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(l/p)},programUniforms:v})}},Ml=(e,t)=>{Yh(e.inputs);let n=[0];e.outputCount>1&&n.push(-3),e.outputCount>2&&n.push(-3),e.outputCount>3&&n.push(3),e.compute(Zh(e.inputs,t,e.outputCount,!1),{outputs:n})}});var Qh,rn,Xh,Ul,Jh,eg,Nl,Vl,Wl=U(()=>{"use strict";te();oe();Se();ae();Qh=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(t.starts.length!==t.ends.length)throw new Error("starts and ends must have the same length");e.slice(1).forEach((r,n)=>{if(e[n+1].dataType!==6&&e[n+1].dataType!==7)throw new Error(`Input ${n} must be an array of int32 or int64`)})},rn=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(n=>r.push(Number(n)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(n=>r.push(Number(n)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Xh=(e,t)=>{if(e.length>1){let r=rn(e,1),n=rn(e,2),o=rn(e,3);return o.length===0&&(o=[...Array(e[0].dims.length).keys()]),re({starts:r,ends:n,axes:o})}else return t},Ul=(e,t,r,n,o)=>{let i=e;return e<0&&(i+=r[n[t]]),o[t]<0?Math.max(0,Math.min(i,r[n[t]]-1)):Math.max(0,Math.min(i,r[n[t]]))},Jh=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
+      }`},T=[{dims:d,dataType:e[0].dataType}];return r>1&&T.push({dims:m,dataType:1}),r>2&&T.push({dims:m,dataType:1}),r>3&&T.push({dims:i,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${$};${_};${y};${g}`,inputDependencies:e.map((A,C)=>"type")},getShaderSource:S,getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(l/p)},programUniforms:v})}},Ml=(e,t)=>{Yh(e.inputs);let n=[0];e.outputCount>1&&n.push(-3),e.outputCount>2&&n.push(-3),e.outputCount>3&&n.push(3),e.compute(Zh(e.inputs,t,e.outputCount,!1),{outputs:n})}});var Qh,tn,Xh,Ul,Jh,eg,Nl,Vl,Wl=U(()=>{"use strict";J();ne();xe();ae();Qh=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(t.starts.length!==t.ends.length)throw new Error("starts and ends must have the same length");e.slice(1).forEach((r,n)=>{if(e[n+1].dataType!==6&&e[n+1].dataType!==7)throw new Error(`Input ${n} must be an array of int32 or int64`)})},tn=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(n=>r.push(Number(n)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(n=>r.push(Number(n)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Xh=(e,t)=>{if(e.length>1){let r=tn(e,1),n=tn(e,2),o=tn(e,3);return o.length===0&&(o=[...Array(e[0].dims.length).keys()]),ee({starts:r,ends:n,axes:o})}else return t},Ul=(e,t,r,n,o)=>{let i=e;return e<0&&(i+=r[n[t]]),o[t]<0?Math.max(0,Math.min(i,r[n[t]]-1)):Math.max(0,Math.min(i,r[n[t]]))},Jh=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
           var input_indices: ${e.type.indices};
           var carry = 0u;
           for (var i = ${r.length}; i >= 0; i--) {
-            let input_shape_i = ${F("uniforms.input_shape","i",r.length)};
-            let steps_i = ${F("uniforms.steps","i",r.length)};
-            let signs_i = ${F("uniforms.signs","i",r.length)};
-            let starts_i = ${F("uniforms.starts","i",r.length)};
+            let input_shape_i = ${q("uniforms.input_shape","i",r.length)};
+            let steps_i = ${q("uniforms.steps","i",r.length)};
+            let signs_i = ${q("uniforms.signs","i",r.length)};
+            let starts_i = ${q("uniforms.starts","i",r.length)};
             var output_index = ${t.indicesGet("output_indices","i")};
             var input_index = output_index * steps_i + starts_i + carry;
             carry = input_index / input_shape_i;
@@ -6508,7 +6581,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
             ${e.indicesSet("input_indices","i","input_index")};
           }
           return input_indices;
-      }`,eg=(e,t)=>{let r=e[0].dims,n=C.size(r),o=t.axes.length>0?C.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],i=rn(e,4);i.forEach($=>$!==0||(()=>{throw new Error("step cannot be 0")})),i.length===0&&(i=Array(o.length).fill(1));let a=t.starts.map(($,v)=>Ul($,v,r,o,i)),d=t.ends.map(($,v)=>Ul($,v,r,o,i));if(o.length!==a.length||o.length!==d.length)throw new Error("start, ends and axes should have the same number of elements");if(o.length!==r.length)for(let $=0;$<r.length;++$)o.includes($)||(a.splice($,0,0),d.splice($,0,r[$]),i.splice($,0,1));let l=i.map($=>Math.sign($));i.forEach(($,v,S)=>{if($<0){let T=(d[v]-a[v])/$,A=a[v],k=A+T*i[v];a[v]=k,d[v]=A,S[v]=-$}});let p=r.slice(0);o.forEach(($,v)=>{p[$]=Math.ceil((d[$]-a[$])/i[$])});let m={dims:p,dataType:e[0].dataType},u=M("output",e[0].dataType,p.length),h=E("input",e[0].dataType,e[0].dims.length),_=C.size(p),y=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:a.length},{name:"signs",type:"i32",length:l.length},{name:"steps",type:"u32",length:i.length}],g=[{type:12,data:_},{type:12,data:a},{type:6,data:l},{type:12,data:i},...N(e[0].dims,p)],x=$=>`
+      }`,eg=(e,t)=>{let r=e[0].dims,n=k.size(r),o=t.axes.length>0?k.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],i=tn(e,4);i.forEach($=>$!==0||(()=>{throw new Error("step cannot be 0")})),i.length===0&&(i=Array(o.length).fill(1));let a=t.starts.map(($,v)=>Ul($,v,r,o,i)),d=t.ends.map(($,v)=>Ul($,v,r,o,i));if(o.length!==a.length||o.length!==d.length)throw new Error("start, ends and axes should have the same number of elements");if(o.length!==r.length)for(let $=0;$<r.length;++$)o.includes($)||(a.splice($,0,0),d.splice($,0,r[$]),i.splice($,0,1));let l=i.map($=>Math.sign($));i.forEach(($,v,S)=>{if($<0){let T=(d[v]-a[v])/$,A=a[v],C=A+T*i[v];a[v]=C,d[v]=A,S[v]=-$}});let p=r.slice(0);o.forEach(($,v)=>{p[$]=Math.ceil((d[$]-a[$])/i[$])});let m={dims:p,dataType:e[0].dataType},u=M("output",e[0].dataType,p.length),h=E("input",e[0].dataType,e[0].dims.length),_=k.size(p),y=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:a.length},{name:"signs",type:"i32",length:l.length},{name:"steps",type:"u32",length:i.length}],g=[{type:12,data:_},{type:12,data:a},{type:6,data:l},{type:12,data:i},...N(e[0].dims,p)],x=$=>`
       ${$.registerUniforms(y).declareVariables(h,u)}
         ${Jh(h,u,r)}
         ${$.mainStart()}
@@ -6516,7 +6589,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
           let output_indices = ${u.offsetToIndices("global_idx")};
           let input_indices = calculateInputIndices(output_indices);
           ${u.setByOffset("global_idx",h.getByIndices("input_indices"))}
-      }`;return{name:"Slice",shaderCache:{hint:`${l.length}_${a.length}_${i.length}`,inputDependencies:["rank"]},getShaderSource:x,getRunData:()=>({outputs:[m],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:g})}},Nl=(e,t)=>{Qh(e.inputs,t);let r=Xh(e.inputs,t);e.compute(eg(e.inputs,r),{inputs:[0]})},Vl=e=>{let t=e.starts,r=e.ends,n=e.axes;return re({starts:t,ends:r,axes:n})}});var tg,rg,Ll,Gl,Hl=U(()=>{"use strict";te();oe();Se();dt();ae();tg=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},rg=(e,t)=>{let r=e.inputs[0],n=r.dims,o=C.size(n),i=n.length,a=C.normalizeAxis(t.axis,i),d=a<n.length-1,l,p=[];d?(p=Array.from({length:i},(P,D)=>D),p[a]=i-1,p[i-1]=a,l=e.compute(Pe(r,p),{inputs:[r],outputs:[-1]})[0]):l=r;let m=l.dims,u=m[i-1],h=o/u,_=me(u),y=u/_,g=64;h===1&&(g=256);let x=(P,D)=>D===4?`max(max(${P}.x, ${P}.y), max(${P}.z, ${P}.w))`:D===2?`max(${P}.x, ${P}.y)`:D===3?`max(max(${P}.x, ${P}.y), ${P}.z)`:P,$=E("x",l.dataType,l.dims,_),v=M("result",l.dataType,l.dims,_),S=$.type.value,T=_e(l.dataType)==="f32"?`var threadMax = ${S}(-3.402823e+38f);`:`var threadMax = ${S}(-65504.0h);`,A=P=>`
+      }`;return{name:"Slice",shaderCache:{hint:`${l.length}_${a.length}_${i.length}`,inputDependencies:["rank"]},getShaderSource:x,getRunData:()=>({outputs:[m],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:g})}},Nl=(e,t)=>{Qh(e.inputs,t);let r=Xh(e.inputs,t);e.compute(eg(e.inputs,r),{inputs:[0]})},Vl=e=>{let t=e.starts,r=e.ends,n=e.axes;return ee({starts:t,ends:r,axes:n})}});var tg,rg,Ll,Gl,Hl=U(()=>{"use strict";J();ne();xe();dt();ae();tg=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},rg=(e,t)=>{let r=e.inputs[0],n=r.dims,o=k.size(n),i=n.length,a=k.normalizeAxis(t.axis,i),d=a<n.length-1,l,p=[];d?(p=Array.from({length:i},(P,D)=>D),p[a]=i-1,p[i-1]=a,l=e.compute(Pe(r,p),{inputs:[r],outputs:[-1]})[0]):l=r;let m=l.dims,u=m[i-1],h=o/u,_=me(u),y=u/_,g=64;h===1&&(g=256);let x=(P,D)=>D===4?`max(max(${P}.x, ${P}.y), max(${P}.z, ${P}.w))`:D===2?`max(${P}.x, ${P}.y)`:D===3?`max(max(${P}.x, ${P}.y), ${P}.z)`:P,$=E("x",l.dataType,l.dims,_),v=M("result",l.dataType,l.dims,_),S=$.type.value,T=ye(l.dataType)==="f32"?`var threadMax = ${S}(-3.402823e+38f);`:`var threadMax = ${S}(-65504.0h);`,A=P=>`
       var<workgroup> rowMaxShared : ${S};
       var<workgroup> rowSumShared : ${S};
       var<workgroup> threadShared : array<${S}, ${g}>;
@@ -6579,7 +6652,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
           workgroupBarrier();
         }
         if (lindex == 0) {
-          rowSumShared = ${S}(${Fe("threadShared[0]",_)});
+          rowSumShared = ${S}(${qe("threadShared[0]",_)});
         }
         workgroupBarrier();
 
@@ -6588,7 +6661,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
           let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;
           setValue(row, col, row_stride, value);
         }
-      }`,k=e.compute({name:"Softmax",shaderCache:{hint:`${_};${g}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:m,dataType:l.dataType}],dispatchGroup:{x:h},programUniforms:[{type:6,data:y}]}),getShaderSource:A},{inputs:[l],outputs:[d?-1:0]})[0];d&&e.compute(Pe(k,p),{inputs:[k]})},Ll=(e,t)=>{tg(e.inputs),rg(e,t)},Gl=e=>re({axis:e.axis})});var Fl,ng,og,ig,ql,Kl=U(()=>{"use strict";te();oe();ae();Fl=e=>Array.from(e.getBigInt64Array(),Number),ng=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==10&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, float16, int32, and uint32 data types");if(e[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(Fl(e[1]).length!==e[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},og=(e,t)=>{let r=[];for(let n=0;n<e.length;++n)r.push(e[n]*t[n]);return r},ig=(e,t)=>{let r=e[0].dims,n=t??Fl(e[1]),o=og(r,n),i=C.size(o),a=e[0].dataType,d=E("input",a,r.length),l=M("output",a,o.length),p=m=>`
+      }`,C=e.compute({name:"Softmax",shaderCache:{hint:`${_};${g}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:m,dataType:l.dataType}],dispatchGroup:{x:h},programUniforms:[{type:6,data:y}]}),getShaderSource:A},{inputs:[l],outputs:[d?-1:0]})[0];d&&e.compute(Pe(C,p),{inputs:[C]})},Ll=(e,t)=>{tg(e.inputs),rg(e,t)},Gl=e=>ee({axis:e.axis})});var Fl,ng,og,ig,ql,Kl=U(()=>{"use strict";J();ne();ae();Fl=e=>Array.from(e.getBigInt64Array(),Number),ng=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==10&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, float16, int32, and uint32 data types");if(e[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(Fl(e[1]).length!==e[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},og=(e,t)=>{let r=[];for(let n=0;n<e.length;++n)r.push(e[n]*t[n]);return r},ig=(e,t)=>{let r=e[0].dims,n=t??Fl(e[1]),o=og(r,n),i=k.size(o),a=e[0].dataType,d=E("input",a,r.length),l=M("output",a,o.length),p=m=>`
       const inputShape = ${d.indices(...r)};
       ${m.registerUniform("output_size","u32").declareVariables(d,l)}
       ${m.mainStart()}
@@ -6602,7 +6675,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
         ${d.indicesSet("input_indices","i","input_dim_value")}
       }
       ${l.setByOffset("global_idx",d.getByIndices("input_indices"))}
-    }`;return{name:"Tile",shaderCache:{hint:`${n}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},...N(e[0].dims,o)]}),getShaderSource:p}},ql=e=>{ng(e.inputs),e.compute(ig(e.inputs),{inputs:[0]})}});var ag,sg,jl,Yl=U(()=>{"use strict";te();oe();ae();ag=(e,t,r,n,o)=>{let i=M("output_data",o,r.length,4),a=E("a_data",t[1].dataType,t[1].dims.length,4),d=E("b_data",t[2].dataType,t[2].dims.length,4),l=E("c_data",t[0].dataType,t[0].dims.length,4),p,m=(u,h,_)=>`select(${h}, ${u}, ${_})`;if(!n)p=i.setByOffset("global_idx",m(a.getByOffset("global_idx"),d.getByOffset("global_idx"),l.getByOffset("global_idx")));else{let u=(h,_,y="")=>{let g=`a_data[index_a${_}][component_a${_}]`,x=`b_data[index_b${_}][component_b${_}]`,$=`bool(c_data[index_c${_}] & (0xffu << (component_c${_} * 8)))`;return`
+    }`;return{name:"Tile",shaderCache:{hint:`${n}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},...N(e[0].dims,o)]}),getShaderSource:p}},ql=e=>{ng(e.inputs),e.compute(ig(e.inputs),{inputs:[0]})}});var ag,sg,jl,Yl=U(()=>{"use strict";J();ne();ae();ag=(e,t,r,n,o)=>{let i=M("output_data",o,r.length,4),a=E("a_data",t[1].dataType,t[1].dims.length,4),d=E("b_data",t[2].dataType,t[2].dims.length,4),l=E("c_data",t[0].dataType,t[0].dims.length,4),p,m=(u,h,_)=>`select(${h}, ${u}, ${_})`;if(!n)p=i.setByOffset("global_idx",m(a.getByOffset("global_idx"),d.getByOffset("global_idx"),l.getByOffset("global_idx")));else{let u=(h,_,y="")=>{let g=`a_data[index_a${_}][component_a${_}]`,x=`b_data[index_b${_}][component_b${_}]`,$=`bool(c_data[index_c${_}] & (0xffu << (component_c${_} * 8)))`;return`
             let output_indices${_} = ${i.offsetToIndices(`global_idx * 4u + ${_}u`)};
             let offset_a${_} = ${a.broadcastedIndicesToOffset(`output_indices${_}`,i)};
             let offset_b${_} = ${d.broadcastedIndicesToOffset(`output_indices${_}`,i)};
@@ -6630,10 +6703,10 @@ fn calculateOutputIndex(index: u32) -> u32 {
         ${e.mainStart()}
         ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
         ${p}
-      }`},sg=e=>{let t=e[1].dims,r=e[2].dims,n=e[0].dims,o=e[1].dataType,i=!(C.areEqual(t,r)&&C.areEqual(r,n)),a=t,d=C.size(t);if(i){let p=tt.calcShape(tt.calcShape(t,r,!1),n,!1);if(!p)throw new Error("Can't perform where op on the given tensors");a=p,d=C.size(a)}let l=Math.ceil(d/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:p=>ag(p,e,a,i,o),getRunData:()=>({outputs:[{dims:a,dataType:o}],dispatchGroup:{x:Math.ceil(d/64/4)},programUniforms:[{type:12,data:l},...N(n,t,r,a)]})}},jl=e=>{e.compute(sg(e.inputs))}});var Zl,Ql=U(()=>{"use strict";Ps();Fr();Ds();Ms();vu();zu();Bu();Qu();od();sd();ld();hd();yd();wd();xd();Id();kd();zd();Bd();Ud();qd();Yd();Qd();Jd();rl();So();ol();vl();Sl();Il();kl();Gr();Ol();Bl();Rl();Wl();Hl();Io();Kl();dt();Kr();Yl();Zl=new Map([["Abs",[Rs]],["Acos",[Us]],["Acosh",[Ns]],["Add",[$u]],["ArgMax",[Es,co]],["ArgMin",[ks,co]],["Asin",[Vs]],["Asinh",[Ws]],["Atan",[Ls]],["Atanh",[Gs]],["Attention",[zs]],["AveragePool",[pl,cl]],["BatchNormalization",[Os]],["BiasAdd",[Bs]],["BiasSplitGelu",[wu]],["Cast",[Fs,Hs]],["Ceil",[Ks]],["Clip",[qs]],["Concat",[Ou,Du]],["Conv",[wo,_o]],["ConvTranspose",[nd,td]],["Cos",[js]],["Cosh",[Ys]],["CumSum",[id,ad]],["DepthToSpace",[ud,dd]],["DequantizeLinear",[$l,xl]],["Div",[xu]],["Einsum",[md,fd]],["Elu",[Zs,Xt]],["Equal",[Su]],["Erf",[Qs]],["Exp",[Xs]],["Expand",[bd]],["FastGelu",[_d]],["Floor",[Js]],["FusedConv",[wo,_o]],["Gather",[$d,vd]],["GatherElements",[Pd,Ed]],["GatherBlockQuantized",[Cd,Ad]],["GatherND",[Sd,Td]],["Gelu",[eu]],["Gemm",[Dd,Od]],["GlobalAveragePool",[hl,fl]],["GlobalMaxPool",[wl,_l]],["Greater",[Au]],["GreaterOrEqual",[Eu]],["GridSample",[Md,Rd]],["GroupQueryAttention",[Fd]],["HardSigmoid",[uu,su]],["InstanceNormalization",[jd]],["LayerNormalization",[Zd]],["LeakyRelu",[tu,Xt]],["Less",[ku]],["LessOrEqual",[Pu]],["Log",[bu]],["MatMul",[Xd]],["MatMulNBits",[el,tl]],["MaxPool",[bl,yl]],["Mul",[Tu]],["MultiHeadAttention",[Wd,Vd]],["Neg",[nu]],["Not",[ru]],["Pad",[nl]],["Pow",[Iu]],["QuickGelu",[yu,Xt]],["Range",[Tl]],["Reciprocal",[ou]],["ReduceMin",[xs]],["ReduceMean",[ys]],["ReduceMax",[$s]],["ReduceSum",[Ts]],["ReduceProd",[Ss]],["ReduceL1",[_s]],["ReduceL2",[ws]],["ReduceLogSum",[Cs]],["ReduceLogSumExp",[vs]],["ReduceSumSquare",[Is]],["Relu",[iu]],["Resize",[Pl,zl]],["RotaryEmbedding",[Dl]],["ScatterND",[Al,Cl]],["Sigmoid",[au]],["Sin",[du]],["Sinh",[lu]],["Slice",[Nl,Vl]],["SkipLayerNormalization",[Ml]],["Split",[Ld,Gd]],["Sqrt",[cu]],["Softmax",[Ll,Gl]],["Sub",[Cu]],["Tan",[pu]],["Tanh",[fu]],["ThresholdedRelu",[gu,Xt]],["Tile",[ql]],["Transpose",[is,as]],["Where",[jl]]])});var nn,Xl=U(()=>{"use strict";We();et();ae();nn=class{constructor(t){this.backend=t;this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,r){this.repo.set(t,r)}run(t,r,n,o,i){Ue(t.programInfo.name);let a=this.backend.device,d=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let l=[];for(let m of r)l.push({binding:l.length,resource:{buffer:m.buffer}});for(let m of n)l.push({binding:l.length,resource:{buffer:m.buffer}});i&&l.push({binding:l.length,resource:i});let p=a.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:l,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let m={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:p,dispatchGroup:o};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(m)}d.setPipeline(t.computePipeline),d.setBindGroup(0,p),d.dispatchWorkgroups(...o),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),De(t.programInfo.name)}dispose(){}build(t,r){Ue(t.name);let n=this.backend.device,o=[];[{feature:"shader-f16",extension:"f16"},{feature:"subgroups",extension:"subgroups"},{feature:"subgroups-f16",extension:"subgroups_f16"}].forEach(u=>{n.features.has(u.feature)&&o.push(`enable ${u.extension};`)});let a=ns(r,this.backend.device.limits),d=t.getShaderSource(a),l=`${o.join(`
+      }`},sg=e=>{let t=e[1].dims,r=e[2].dims,n=e[0].dims,o=e[1].dataType,i=!(k.areEqual(t,r)&&k.areEqual(r,n)),a=t,d=k.size(t);if(i){let p=tt.calcShape(tt.calcShape(t,r,!1),n,!1);if(!p)throw new Error("Can't perform where op on the given tensors");a=p,d=k.size(a)}let l=Math.ceil(d/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:p=>ag(p,e,a,i,o),getRunData:()=>({outputs:[{dims:a,dataType:o}],dispatchGroup:{x:Math.ceil(d/64/4)},programUniforms:[{type:12,data:l},...N(n,t,r,a)]})}},jl=e=>{e.compute(sg(e.inputs))}});var Zl,Ql=U(()=>{"use strict";Es();Hr();Os();Bs();wu();Pu();Du();Zu();nd();ad();dd();fd();bd();_d();$d();Td();Ad();Pd();Dd();Rd();Fd();jd();Zd();Xd();tl();xo();nl();wl();xl();Tl();Al();Lr();Ol();Bl();Rl();Wl();Hl();To();Kl();dt();qr();Yl();Zl=new Map([["Abs",[Ms]],["Acos",[Rs]],["Acosh",[Us]],["Add",[vu]],["ArgMax",[ks,lo]],["ArgMin",[As,lo]],["Asin",[Ns]],["Asinh",[Vs]],["Atan",[Ws]],["Atanh",[Ls]],["Attention",[Ps]],["AveragePool",[cl,ll]],["BatchNormalization",[zs]],["BiasAdd",[Ds]],["BiasSplitGelu",[_u]],["Cast",[Hs,Gs]],["Ceil",[qs]],["Clip",[Fs]],["Concat",[zu,Ou]],["Conv",[_o,yo]],["ConvTranspose",[rd,ed]],["Cos",[Ks]],["Cosh",[js]],["CumSum",[od,id]],["DepthToSpace",[sd,ud]],["DequantizeLinear",[vl,$l]],["Div",[$u]],["Einsum",[pd,md]],["Elu",[Ys,Qt]],["Equal",[xu]],["Erf",[Zs]],["Exp",[Qs]],["Expand",[gd]],["FastGelu",[yd]],["Floor",[Xs]],["FusedConv",[_o,yo]],["Gather",[vd,wd]],["GatherElements",[Ed,kd]],["GatherBlockQuantized",[Id,Cd]],["GatherND",[xd,Sd]],["Gelu",[Js]],["Gemm",[Od,zd]],["GlobalAveragePool",[fl,ml]],["GlobalMaxPool",[_l,yl]],["Greater",[Cu]],["GreaterOrEqual",[ku]],["GridSample",[Bd,Md]],["GroupQueryAttention",[Hd]],["HardSigmoid",[su,au]],["InstanceNormalization",[Kd]],["LayerNormalization",[Yd]],["LeakyRelu",[eu,Qt]],["Less",[Au]],["LessOrEqual",[Eu]],["Log",[gu]],["MatMul",[Qd]],["MatMulNBits",[Jd,el]],["MaxPool",[gl,bl]],["Mul",[Su]],["MultiHeadAttention",[Vd,Nd]],["Neg",[ru]],["Not",[tu]],["Pad",[rl]],["Pow",[Tu]],["QuickGelu",[bu,Qt]],["Range",[Sl]],["Reciprocal",[nu]],["ReduceMin",[$s]],["ReduceMean",[bs]],["ReduceMax",[vs]],["ReduceSum",[Ss]],["ReduceProd",[xs]],["ReduceL1",[ys]],["ReduceL2",[_s]],["ReduceLogSum",[Is]],["ReduceLogSumExp",[ws]],["ReduceSumSquare",[Ts]],["Relu",[ou]],["Resize",[Pl,zl]],["RotaryEmbedding",[Dl]],["ScatterND",[Cl,Il]],["Sigmoid",[iu]],["Sin",[uu]],["Sinh",[du]],["Slice",[Nl,Vl]],["SkipLayerNormalization",[Ml]],["Split",[Wd,Ld]],["Sqrt",[lu]],["Softmax",[Ll,Gl]],["Sub",[Iu]],["Tan",[cu]],["Tanh",[mu]],["ThresholdedRelu",[hu,Qt]],["Tile",[ql]],["Transpose",[os,is]],["Where",[jl]]])});var rn,Xl=U(()=>{"use strict";Le();et();ae();rn=class{constructor(t){this.backend=t;this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,r){this.repo.set(t,r)}run(t,r,n,o,i){Ne(t.programInfo.name);let a=this.backend.device,d=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let l=[];for(let m of r)l.push({binding:l.length,resource:{buffer:m.buffer}});for(let m of n)l.push({binding:l.length,resource:{buffer:m.buffer}});i&&l.push({binding:l.length,resource:i});let p=a.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:l,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let m={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:p,dispatchGroup:o};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(m)}d.setPipeline(t.computePipeline),d.setBindGroup(0,p),d.dispatchWorkgroups(...o),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),Be(t.programInfo.name)}dispose(){}build(t,r){Ne(t.name);let n=this.backend.device,o=[];[{feature:"shader-f16",extension:"f16"},{feature:"subgroups",extension:"subgroups"},{feature:"subgroups-f16",extension:"subgroups_f16"}].forEach(u=>{n.features.has(u.feature)&&o.push(`enable ${u.extension};`)});let a=rs(r,this.backend.device.limits),d=t.getShaderSource(a),l=`${o.join(`
 `)}
 ${a.additionalImplementations}
-${d}`,p=n.createShaderModule({code:l,label:t.name});ue("verbose",()=>`[WebGPU] ${t.name} shader code: ${l}`);let m=n.createComputePipeline({compute:{module:p,entryPoint:"main"},layout:"auto",label:t.name});return De(t.name),{programInfo:t,computePipeline:m,uniformVariablesInfo:a.variablesInfo}}normalizeDispatchGroupSize(t){let r=typeof t=="number"?t:t.x,n=typeof t=="number"?1:t.y||1,o=typeof t=="number"?1:t.z||1,i=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(r<=i&&n<=i&&o<=i)return[r,n,o];let a=r*n*o,d=Math.ceil(Math.sqrt(a));if(d>i){if(d=Math.ceil(Math.cbrt(a)),d>i)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[d,d,d]}else return[d,d,1]}}});var ug,dg,Co,Ao,on,Jl=U(()=>{"use strict";We();te();et();Jn();Ja();Ql();Xl();ug=(e,t)=>{if(t.length!==e.length)throw new Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let r=[];for(let n=0;n<e.length;++n){let o=e[n].dataType;switch(t[n]){case"none":{r.push("");break}case"type":{r.push(`${o}`);break}case"rank":{let i=e[n].dims.length;r.push(`${o};${i}`);break}case"dims":{let i=e[n].dims.join(",");r.push(`${o};${i}`);break}default:throw new Error(`unsupported input dependency: ${t[n]}`)}}return r.join("|")},dg=(e,t,r)=>{let n=e.name;return e.shaderCache?.hint&&(n+="["+e.shaderCache.hint+"]"),n+=":"+r+`:${ug(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,n},Co=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},Ao=class{constructor(t){this.subgroupsSupported=t.features.has("subgroups"),this.subgroupsF16Supported=t.features.has("subgroups");let r=t.limits;!this.subgroupsSupported||!r.minSubgroupSize||!r.maxSubgroupSize?this.subgroupSizeRange=void 0:this.subgroupSizeRange=[r.minSubgroupSize,r.maxSubgroupSize]}},on=class{constructor(){this.currentSessionId=null;this.currentKernelId=null;this.commandEncoder=null;this.computePassEncoder=null;this.maxDispatchNumber=16;this.pendingDispatchNumber=0;this.pendingKernels=[];this.pendingQueries=new Map;this.sessionStatus="default";this.capturedCommandList=new Map;this.capturedPendingKernels=new Map;this.sessionExternalDataMapping=new Map}get currentKernelCustomData(){if(this.currentKernelId===null)throw new Error("currentKernelCustomData(): currentKernelId is null. (should not happen)");let t=this.kernelCustomData.get(this.currentKernelId);return t||(t={},this.kernelCustomData.set(this.currentKernelId,t)),t}async initialize(t,r){this.env=t;let n=[],o={requiredLimits:{maxComputeWorkgroupStorageSize:r.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.limits.maxStorageBufferBindingSize,maxBufferSize:r.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:r.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:r.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:r.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:r.limits.maxComputeWorkgroupSizeZ},requiredFeatures:n},i=a=>r.features.has(a)&&n.push(a)&&!0;i("chromium-experimental-timestamp-query-inside-passes")||i("timestamp-query"),i("shader-f16"),i("subgroups")&&i("subgroups-f16"),this.device=await r.requestDevice(o),this.deviceInfo=new Ao(this.device),this.adapterInfo=new Co(r.info||await r.requestAdapterInfo()),this.gpuDataManager=Xa(this),this.programManager=new nn(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,Rr(t.logLevel,!!t.debug),this.device.onuncapturederror=a=>{a.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${a.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value:r,writable:!1,enumerable:!0,configurable:!1}),this.setQueryType()}dispose(){typeof this.querySet<"u"&&this.querySet.destroy(),this.gpuDataManager.dispose()}getCommandEncoder(){return this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder()),this.commandEncoder}getComputePassEncoder(){if(!this.computePassEncoder){let t=this.getCommandEncoder(),r={};this.queryType==="at-passes"&&(r.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=t.beginComputePass(r)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;Ue(),this.endComputePass();let t;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),t=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(t,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&t.mapAsync(GPUMapMode.READ).then(()=>{let r=new BigUint64Array(t.getMappedRange()),n=this.pendingQueries.get(t);for(let o=0;o<r.length/2;o++){let i=n[o],a=i.kernelId,d=this.kernels.get(a),l=d.kernelType,p=d.kernelName,m=i.programName,u=i.inputTensorViews,h=i.outputTensorViews,_=r[o*2],y=r[o*2+1];typeof this.queryTimeBase>"u"&&(this.queryTimeBase=_);let g=Number(_-this.queryTimeBase),x=Number(y-this.queryTimeBase);if(!Number.isSafeInteger(g)||!Number.isSafeInteger(x))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:u.map($=>({dims:$.dims,dataType:bt($.dataType)})),outputsMetadata:h.map($=>({dims:$.dims,dataType:bt($.dataType)})),kernelId:a,kernelType:l,kernelName:p,programName:m,startTime:g,endTime:x});else{let $="";u.forEach((S,T)=>{$+=`input[${T}]: [${S.dims}] | ${bt(S.dataType)}, `});let v="";h.forEach((S,T)=>{v+=`output[${T}]: [${S.dims}] | ${bt(S.dataType)}, `}),console.log(`[profiling] kernel "${a}|${l}|${p}|${m}" ${$}${v}execution time: ${x-g} ns`)}wr("GPU",`${m}::${_}::${y}`)}t.unmap(),this.pendingQueries.delete(t)}),De()}run(t,r,n,o,i,a){Ue(t.name);let d=[];for(let S=0;S<r.length;++S){let T=r[S].data;if(T===0)continue;let A=this.gpuDataManager.get(T);if(!A)throw new Error(`no GPU data for input: ${T}`);d.push(A)}let{outputs:l,dispatchGroup:p,programUniforms:m}=t.getRunData(r),u=n.length===0?l.map((S,T)=>T):n;if(u.length!==l.length)throw new Error(`Output size ${u.length} must be equal to ${l.length}.`);let h=[],_=[];for(let S=0;S<l.length;++S){if(!Number.isInteger(u[S])||u[S]<-3||u[S]>=a)throw new Error(`Invalid output index: ${u[S]}`);if(u[S]===-3)continue;let T=u[S]===-1,A=u[S]===-2,k=T||A?i(l[S].dataType,l[S].dims):o(u[S],l[S].dataType,l[S].dims);if(h.push(k),k.data===0)continue;let P=this.gpuDataManager.get(k.data);if(!P)throw new Error(`no GPU data for output: ${k.data}`);if(T&&this.temporaryData.push(P),A){let D=this.kernelPersistentData.get(this.currentKernelId);D||(D=[],this.kernelPersistentData.set(this.currentKernelId,D)),D.push(P)}_.push(P)}if(d.length!==r.length||_.length!==h.length){if(_.length===0)return De(t.name),h;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let y;if(m){let S=0,T=[];m.forEach(D=>{let R=typeof D.data=="number"?[D.data]:D.data;if(R.length===0)return;let G=D.type===10?2:4,K,j;D.type===10?(j=R.length>4?16:R.length>2?8:R.length*G,K=R.length>4?16:G*R.length):(j=R.length<=2?R.length*G:16,K=16),S=Math.ceil(S/j)*j,T.push(S);let V=D.type===10?8:4;S+=R.length>4?Math.ceil(R.length/V)*K:R.length*G});let A=16;S=Math.ceil(S/A)*A;let k=new ArrayBuffer(S);m.forEach((D,R)=>{let G=T[R],K=typeof D.data=="number"?[D.data]:D.data;if(D.type===6)new Int32Array(k,G,K.length).set(K);else if(D.type===12)new Uint32Array(k,G,K.length).set(K);else if(D.type===10)new Uint16Array(k,G,K.length).set(K);else if(D.type===1)new Float32Array(k,G,K.length).set(K);else throw new Error(`Unsupported uniform type: ${bt(D.type)}`)});let P=this.gpuDataManager.create(S,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(P.buffer,0,k,0,S),this.gpuDataManager.release(P.id),y={offset:0,size:S,buffer:P.buffer}}let g=this.programManager.normalizeDispatchGroupSize(p),x=g[1]===1&&g[2]===1,$=dg(t,r,x),v=this.programManager.getArtifact($);if(v||(v=this.programManager.build(t,g),this.programManager.setArtifact($,v),ue("info",()=>`[artifact] key: ${$}, programName: ${t.name}`)),m&&v.uniformVariablesInfo){if(m.length!==v.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${v.uniformVariablesInfo.length}, got ${m.length} in program "${v.programInfo.name}".`);for(let S=0;S<m.length;S++){let T=m[S],A=T.type,k=typeof T.data=="number"?1:T.data.length,[P,D]=v.uniformVariablesInfo[S];if(A!==P||k!==D)throw new Error(`Uniform variable ${S} mismatch: expect type ${P} with size ${D}, got type ${A} with size ${k} in program "${v.programInfo.name}".`)}}if(ue("info",()=>`[ProgramManager] run "${t.name}" (key=${$}) with ${g[0]}x${g[1]}x${g[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let S={kernelId:this.currentKernelId,programName:v.programInfo.name,inputTensorViews:r,outputTensorViews:h};this.pendingKernels.push(S),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(S)}return this.programManager.run(v,d,_,g,y),De(t.name),h}upload(t,r){this.gpuDataManager.upload(t,r)}memcpy(t,r){this.gpuDataManager.memcpy(t,r)}async download(t,r){await this.gpuDataManager.download(t,r)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,r,n,o){let i=Zl.get(t);if(!i)throw new Error(`kernel not implemented: ${t}`);let a={kernelType:t,kernelName:o,kernelEntry:i[0],attributes:[i[1],n]};this.kernels.set(r,a)}releaseKernel(t){let r=this.kernelPersistentData.get(t);if(r){for(let n of r)this.gpuDataManager.release(n.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,r,n){let o=this.kernels.get(t);if(!o)throw new Error(`kernel not created: ${t}`);let i=o.kernelType,a=o.kernelName,d=o.kernelEntry,l=o.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${i}] ${a}" is not allowed to be called recursively`);this.currentKernelId=t,l[0]&&(l[1]=l[0](l[1]),l[0]=void 0),ue("info",()=>`[WebGPU] Start to run kernel "[${i}] ${a}"...`);let p=this.env.debug;this.temporaryData=[];try{return p&&this.device.pushErrorScope("validation"),d(r,l[1]),0}catch(m){return n.push(Promise.resolve(`[WebGPU] Kernel "[${i}] ${a}" failed. ${m}`)),1}finally{p&&n.push(this.device.popErrorScope().then(m=>m?`GPU validation error for kernel "[${i}] ${a}": ${m.message}`:null));for(let m of this.temporaryData)this.gpuDataManager.release(m.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,r,n,o){let i=this.sessionExternalDataMapping.get(t);i||(i=new Map,this.sessionExternalDataMapping.set(t,i));let a=i.get(r),d=this.gpuDataManager.registerExternalBuffer(n,o,a);return i.set(r,[d,n]),d}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(n=>this.gpuDataManager.unregisterExternalBuffer(n[0])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let r=this.gpuDataManager.get(t);if(!r)throw new Error(`no GPU data for buffer: ${t}`);return r.buffer}createDownloader(t,r,n){return async()=>{let o=await no(this,t,r);return Ur(o.buffer,n)}}writeTimestamp(t){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,t)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||(typeof this.env.trace>"u"?this.env.wasm.trace:this.env.trace))&&(this.device.features.has("chromium-experimental-timestamp-query-inside-passes")?this.queryType="inside-passes":this.device.features.has("timestamp-query")&&(this.queryType="at-passes"),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){ue("info","captureBegin"),this.capturedCommandList.get(this.currentSessionId)||this.capturedCommandList.set(this.currentSessionId,[]),this.capturedPendingKernels.get(this.currentSessionId)||this.capturedPendingKernels.set(this.currentSessionId,[]),this.flush(),this.sessionStatus="capturing"}captureEnd(){ue("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){ue("info","replay"),this.sessionStatus="replaying";let t=this.capturedCommandList.get(this.currentSessionId),r=this.capturedPendingKernels.get(this.currentSessionId),n=t.length;this.pendingKernels=[];for(let o=0;o<n;o++){let i=this.getComputePassEncoder(),a=t[o];this.writeTimestamp(this.pendingDispatchNumber*2),i.setPipeline(a.computePipeline),i.setBindGroup(0,a.bindGroup),i.dispatchWorkgroups(...a.dispatchGroup),this.writeTimestamp(this.pendingDispatchNumber*2+1),this.pendingDispatchNumber++,this.queryType!=="none"&&this.pendingKernels.push(r[o]),(this.pendingDispatchNumber>=this.maxDispatchNumber||this.queryType==="at-passes")&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}onCreateSession(){this.gpuDataManager.onCreateSession()}onReleaseSession(t){this.unregisterBuffers(t),this.capturedCommandList.has(t)&&this.capturedCommandList.delete(t),this.capturedPendingKernels.has(t)&&this.capturedPendingKernels.delete(t),this.gpuDataManager.onReleaseSession(t)}onRunStart(t){this.currentSessionId=t,this.setQueryType()}}});var lg,ec,cg,tc,an,sn,ko,rc,nc=U(()=>{"use strict";et();lg=1,ec=()=>lg++,cg=new Map([["float32",32],["float16",16],["int32",32],["uint32",32],["int64",64],["uint64",64],["int8",8],["uint8",8],["int4",4],["uint4",4]]),tc=(e,t)=>{let r=cg.get(e);if(!r)throw new Error("Unsupported data type.");return t.length>0?Math.ceil(t.reduce((n,o)=>n*o)*r/8):0},an=class{constructor(t){this.sessionId=t.sessionId,this.mlContext=t.context,this.mlTensor=t.tensor,this.dataType=t.dataType,this.tensorShape=t.shape}get tensor(){return this.mlTensor}get type(){return this.dataType}get shape(){return this.tensorShape}get byteLength(){return tc(this.dataType,this.tensorShape)}destroy(){ue("verbose",()=>"[WebNN] TensorWrapper.destroy"),this.mlTensor.destroy()}write(t){this.mlContext.writeTensor(this.mlTensor,t)}async read(t){return t?this.mlContext.readTensor(this.mlTensor,t):this.mlContext.readTensor(this.mlTensor)}canReuseTensor(t,r,n){return this.mlContext===t&&this.dataType===r&&this.tensorShape.length===n.length&&this.tensorShape.every((o,i)=>o===n[i])}},sn=class{constructor(t,r){this.tensorManager=t;this.wrapper=r}get tensorWrapper(){return this.wrapper}releaseTensor(){this.tensorWrapper&&(this.tensorManager.releaseTensor(this.tensorWrapper),this.wrapper=void 0)}async ensureTensor(t,r,n,o){if(this.wrapper){if(this.wrapper.canReuseTensor(t,r,n))return this.wrapper.tensor;if(o){if(this.wrapper.byteLength!==tc(r,n))throw new Error("Unable to copy data to tensor with different size.");this.activeUpload=new Uint8Array(await this.wrapper.read())}this.tensorManager.releaseTensor(this.wrapper)}let i=typeof MLTensorUsage>"u"?void 0:MLTensorUsage.READ|MLTensorUsage.WRITE;return this.wrapper=await this.tensorManager.getCachedTensor(r,n,i,!0,!0),o&&this.activeUpload&&(this.wrapper.write(this.activeUpload),this.activeUpload=void 0),this.wrapper.tensor}upload(t){if(this.wrapper)if(t.byteLength===this.wrapper.byteLength){this.wrapper.write(t);return}else ue("verbose",()=>"Data size does not match tensor size. Releasing tensor."),this.releaseTensor();this.activeUpload?this.activeUpload.set(t):this.activeUpload=new Uint8Array(t)}async download(t){if(this.activeUpload)if(t){t instanceof ArrayBuffer?new Uint8Array(t).set(this.activeUpload):new Uint8Array(t.buffer,t.byteOffset,t.byteLength).set(this.activeUpload);return}else return this.activeUpload.buffer;if(!this.wrapper)throw new Error("Tensor has not been created.");return t?this.wrapper.read(t):this.wrapper.read()}},ko=class{constructor(t){this.backend=t;this.tensorTrackersById=new Map;this.freeTensors=[];this.externalTensors=new Set}reserveTensorId(){let t=ec();return this.tensorTrackersById.set(t,new sn(this)),t}releaseTensorId(t){let r=this.tensorTrackersById.get(t);r&&(this.tensorTrackersById.delete(t),r.tensorWrapper&&this.releaseTensor(r.tensorWrapper))}async ensureTensor(t,r,n,o){ue("verbose",()=>`[WebNN] TensorManager.ensureTensor {tensorId: ${t}, dataType: ${r}, shape: ${n}, copyOld: ${o}}`);let i=this.tensorTrackersById.get(t);if(!i)throw new Error("Tensor not found.");return i.ensureTensor(this.backend.currentContext,r,n,o)}upload(t,r){let n=this.tensorTrackersById.get(t);if(!n)throw new Error("Tensor not found.");n.upload(r)}async download(t,r){ue("verbose",()=>`[WebNN] TensorManager.download {tensorId: ${t}, dstBuffer: ${r?.byteLength}}`);let n=this.tensorTrackersById.get(t);if(!n)throw new Error("Tensor not found.");return n.download(r)}releaseTensorsForSession(t){for(let r of this.freeTensors)r.sessionId===t&&r.destroy();this.freeTensors=this.freeTensors.filter(r=>r.sessionId!==t)}registerTensor(t,r,n,o){let i=ec(),a=new an({sessionId:this.backend.currentSessionId,context:t,tensor:r,dataType:n,shape:o});return this.tensorTrackersById.set(i,new sn(this,a)),this.externalTensors.add(a),i}async getCachedTensor(t,r,n,o,i){let a=this.backend.currentSessionId,d=this.backend.currentContext;for(let[p,m]of this.freeTensors.entries())if(m.canReuseTensor(d,t,r)){ue("verbose",()=>`[WebNN] Reusing tensor {dataType: ${t}, shape: ${r}}`);let u=this.freeTensors.splice(p,1)[0];return u.sessionId=a,u}ue("verbose",()=>`[WebNN] MLContext.createTensor {dataType: ${t}, shape: ${r}}`);let l=await d.createTensor({dataType:t,shape:r,dimensions:r,usage:n,writable:o,readable:i});return new an({sessionId:a,context:d,tensor:l,dataType:t,shape:r})}releaseTensor(t){this.externalTensors.has(t)&&this.externalTensors.delete(t),this.freeTensors.push(t)}},rc=(...e)=>new ko(...e)});var oc,pg,un,ic=U(()=>{"use strict";te();gt();Jn();nc();et();oc=new Map([[1,"float32"],[10,"float16"],[6,"int32"],[12,"uint32"],[7,"int64"],[13,"uint64"],[22,"int4"],[21,"uint4"],[3,"int8"],[2,"uint8"],[9,"uint8"]]),pg=(e,t)=>{if(e===t)return!0;if(e===void 0||t===void 0)return!1;let r=Object.keys(e).sort(),n=Object.keys(t).sort();return r.length===n.length&&r.every((o,i)=>o===n[i]&&e[o]===t[o])},un=class{constructor(t){this.tensorManager=rc(this);this.mlContextBySessionId=new Map;this.sessionIdsByMLContext=new Map;this.mlContextCache=[];Rr(t.logLevel,!!t.debug)}get currentSessionId(){if(this.activeSessionId===void 0)throw new Error("No active session");return this.activeSessionId}onRunStart(t){this.activeSessionId=t}async createMLContext(t){if(t instanceof GPUDevice){let n=this.mlContextCache.findIndex(o=>o.gpuDevice===t);if(n!==-1)return this.mlContextCache[n].mlContext;{let o=await navigator.ml.createContext(t);return this.mlContextCache.push({gpuDevice:t,mlContext:o}),o}}else if(t===void 0){let n=this.mlContextCache.findIndex(o=>o.options===void 0&&o.gpuDevice===void 0);if(n!==-1)return this.mlContextCache[n].mlContext;{let o=await navigator.ml.createContext();return this.mlContextCache.push({mlContext:o}),o}}let r=this.mlContextCache.findIndex(n=>pg(n.options,t));if(r!==-1)return this.mlContextCache[r].mlContext;{let n=await navigator.ml.createContext(t);return this.mlContextCache.push({options:t,mlContext:n}),n}}get currentContext(){let t=this.getMLContext(this.currentSessionId);if(!t)throw new Error(`No MLContext found for session ${this.currentSessionId}`);return t}registerMLContext(t,r){this.mlContextBySessionId.set(t,r);let n=this.sessionIdsByMLContext.get(r);n||(n=new Set,this.sessionIdsByMLContext.set(r,n)),n.add(t)}onReleaseSession(t){let r=this.mlContextBySessionId.get(t);if(!r)return;this.tensorManager.releaseTensorsForSession(t),this.mlContextBySessionId.delete(t);let n=this.sessionIdsByMLContext.get(r);if(n.delete(t),n.size===0){this.sessionIdsByMLContext.delete(r);let o=this.mlContextCache.findIndex(i=>i.mlContext===r);o!==-1&&this.mlContextCache.splice(o,1)}}getMLContext(t){return this.mlContextBySessionId.get(t)}reserveTensorId(){return this.tensorManager.reserveTensorId()}releaseTensorId(t){ue("verbose",()=>`[WebNN] releaseTensorId {tensorId: ${t}}`),this.tensorManager.releaseTensorId(t)}async ensureTensor(t,r,n,o){let i=oc.get(r);if(!i)throw new Error(`Unsupported ONNX data type: ${r}`);return this.tensorManager.ensureTensor(t,i,n,o)}uploadTensor(t,r){if(!Ie().shouldTransferToMLTensor)throw new Error("Trying to upload to a MLTensor while shouldTransferToMLTensor is false");ue("verbose",()=>`[WebNN] uploadTensor {tensorId: ${t}, data: ${r.byteLength}}`),this.tensorManager.upload(t,r)}async downloadTensor(t,r){return this.tensorManager.download(t,r)}createMLTensorDownloader(t,r){return async()=>{let n=await this.tensorManager.download(t);return Ur(n,r)}}registerMLTensor(t,r,n){let o=oc.get(r);if(!o)throw new Error(`Unsupported ONNX data type: ${r}`);let i=this.tensorManager.registerTensor(this.currentContext,t,o,n);return ue("verbose",()=>`[WebNN] registerMLTensor {tensor: ${t}, dataType: ${o}, dimensions: ${n}} -> {tensorId: ${i}}`),i}registerMLConstant(t,r,n,o,i,a){if(!a)throw new Error("External mounted files are not available.");let d=t;t.startsWith("./")&&(d=t.substring(2));let l=a.get(d);if(!l)throw new Error(`File with name ${d} not found in preloaded files.`);if(r+n>l.byteLength)throw new Error("Out of bounds: data offset and length exceed the external file data size.");let p=l.slice(r,r+n).buffer,m;switch(i.dataType){case"float32":m=new Float32Array(p);break;case"float16":m=new Uint16Array(p);break;case"int32":m=new Int32Array(p);break;case"uint32":m=new Uint32Array(p);break;case"int64":m=new BigInt64Array(p);break;case"uint64":m=new BigUint64Array(p);break;case"int8":m=new Int8Array(p);break;case"int4":case"uint4":case"uint8":m=new Uint8Array(p);break;default:throw new Error(`Unsupported data type: ${i.dataType} in creating WebNN Constant from external data.`)}return ue("verbose",()=>`[WebNN] registerMLConstant {dataType: ${i.dataType}, shape: ${i.shape}}}`),o.constant(i,m)}flush(){}}});var ac={};Ft(ac,{init:()=>mg});var rr,Eo,mg,sc=U(()=>{"use strict";te();Jl();et();oe();ic();rr=class e{constructor(t,r,n,o){this.module=t;this.dataType=r;this.data=n;this.dims=o}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=C.size(this.dims);return t===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,t)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let t=C.size(this.dims);return t===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,t)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let t=C.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}getUint16Array(){if(this.dataType!==10&&this.dataType!==4)throw new Error("Invalid data type");let t=C.size(this.dims);return t===0?new Uint16Array:new Uint16Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(C.size(t)!==C.size(this.dims))throw new Error("Invalid new shape");return new e(this.module,this.dataType,this.data,t)}},Eo=class{constructor(t,r,n){this.module=t;this.backend=r;this.customDataOffset=0;this.customDataSize=0;this.adapterInfo=r.adapterInfo,this.deviceInfo=r.deviceInfo;let o=t.PTR_SIZE,i=n/t.PTR_SIZE,a=o===4?"i32":"i64";this.opKernelContext=Number(t.getValue(o*i++,a));let d=Number(t.getValue(o*i++,a));this.outputCount=Number(t.getValue(o*i++,a)),this.customDataOffset=Number(t.getValue(o*i++,"*")),this.customDataSize=Number(t.getValue(o*i++,a));let l=[];for(let p=0;p<d;p++){let m=Number(t.getValue(o*i++,a)),u=Number(t.getValue(o*i++,"*")),h=Number(t.getValue(o*i++,a)),_=[];for(let y=0;y<h;y++)_.push(Number(t.getValue(o*i++,a)));l.push(new rr(t,m,u,_))}this.inputs=l}get kernelCustomData(){return this.backend.currentKernelCustomData}get customDataBuffer(){return this.module.HEAPU8.subarray(this.customDataOffset,this.customDataOffset+this.customDataSize)}compute(t,r){let n=r?.inputs?.map(d=>typeof d=="number"?this.inputs[d]:d)??this.inputs,o=r?.outputs??[],i=(d,l,p)=>new rr(this.module,l,this.output(d,p),p),a=(d,l)=>{let p=Ct(d,l);if(!p)throw new Error(`Unsupported data type: ${d}`);let m=p>0?this.backend.gpuDataManager.create(p).id:0;return new rr(this.module,d,m,l)};return this.backend.run(t,n,o,i,a,this.outputCount)}output(t,r){let n=this.module.stackSave();try{let o=this.module.PTR_SIZE,i=o===4?"i32":"i64",a=this.module.stackAlloc((1+r.length)*o);this.module.setValue(a,r.length,i);for(let d=0;d<r.length;d++)this.module.setValue(a+o*(d+1),r[d],i);return this.module._JsepOutput(this.opKernelContext,t,a)}catch(o){throw new Error(`Failed to generate kernel's output[${t}] with dims [${r}]. If you are running with pre-allocated output, please make sure the output type/dims are correct. Error: ${o}`)}finally{this.module.stackRestore(n)}}},mg=async(e,t,r,n)=>{let o=t.jsepInit;if(!o)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let i=new on;await i.initialize(r,n),o("webgpu",[i,a=>i.alloc(Number(a)),a=>i.free(a),(a,d,l,p=!1)=>{if(p)ue("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${Number(a)}, dst=${Number(d)}, size=${Number(l)}`),i.memcpy(Number(a),Number(d));else{ue("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${Number(a)}, gpuDataId=${Number(d)}, size=${Number(l)}`);let m=t.HEAPU8.subarray(Number(a>>>0),Number(a>>>0)+Number(l));i.upload(Number(d),m)}},async(a,d,l)=>{ue("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${a}, dataOffset=${d}, size=${l}`),await i.download(Number(a),()=>t.HEAPU8.subarray(Number(d)>>>0,Number(d+l)>>>0))},(a,d,l)=>i.createKernel(a,Number(d),l,t.UTF8ToString(t._JsepGetNodeName(Number(d)))),a=>i.releaseKernel(a),(a,d,l,p)=>{ue("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${l}, kernel=${a}, contextDataOffset=${d}`);let m=new Eo(t,i,Number(d));return i.computeKernel(Number(a),m,p)},()=>i.captureBegin(),()=>i.captureEnd(),()=>i.replay()])}else{let i=new un(r);o("webnn",[i,()=>i.reserveTensorId(),a=>i.releaseTensorId(a),async(a,d,l,p)=>i.ensureTensor(a,d,l,p),(a,d)=>{i.uploadTensor(a,d)},async(a,d)=>i.downloadTensor(a,d)])}}});var fg,Tr,Ir,Pt,hg,Kt,Cr,Ar,uc,kr,Er,Pr,qn=U(()=>{"use strict";Ha();qa();te();gt();Or();Xn();fg=(e,t)=>{Ie()._OrtInit(e,t)!==0&&pe("Can't initialize onnxruntime.")},Tr=async e=>{fg(e.wasm.numThreads,Zt(e.logLevel))},Ir=async(e,t)=>{{let r=(sc(),br(ac)).init;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let n=e.webgpu.adapter;if(n){if(typeof n.limits!="object"||typeof n.features!="object"||typeof n.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let o=e.webgpu.powerPreference;if(o!==void 0&&o!=="low-power"&&o!=="high-performance")throw new Error(`Invalid powerPreference setting: "${o}"`);let i=e.webgpu.forceFallbackAdapter;if(i!==void 0&&typeof i!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${i}"`);if(n=await navigator.gpu.requestAdapter({powerPreference:o,forceFallbackAdapter:i}),!n)throw new Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.')}await r("webgpu",Ie(),e,n)}if(t==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",Ie(),e)}}},Pt=new Map,hg=e=>{let t=Ie(),r=t.stackSave();try{let n=t.PTR_SIZE,o=t.stackAlloc(2*n);t._OrtGetInputOutputCount(e,o,o+n)!==0&&pe("Can't get session input/output count.");let a=n===4?"i32":"i64";return[Number(t.getValue(o,a)),Number(t.getValue(o+n,a))]}finally{t.stackRestore(r)}},Kt=e=>{let t=Ie(),r=t._malloc(e.byteLength);if(r===0)throw new Error(`Can't create a session. failed to allocate a buffer of size ${e.byteLength}.`);return t.HEAPU8.set(e,r),[r,e.byteLength]},Cr=async(e,t)=>{let r,n,o=Ie();Array.isArray(e)?[r,n]=e:e.buffer===o.HEAPU8.buffer?[r,n]=[e.byteOffset,e.byteLength]:[r,n]=Kt(e);let i=0,a=0,d=0,l=[],p=[],m=[];try{if([a,l]=Fa(t),t?.externalData&&o.mountExternalData){let v=[];for(let S of t.externalData){let T=typeof S=="string"?S:S.path;v.push(Qt(typeof S=="string"?S:S.data).then(A=>{o.mountExternalData(T,A)}))}await Promise.all(v)}for(let v of t?.executionProviders??[])if((typeof v=="string"?v:v.name)==="webnn"){if(o.shouldTransferToMLTensor=!1,typeof v!="string"){let T=v,A=T?.context,k=T?.gpuDevice,P=T?.deviceType,D=T?.powerPreference;A?o.currentContext=A:k?o.currentContext=await o.jsepCreateMLContext(k):o.currentContext=await o.jsepCreateMLContext({deviceType:P,powerPreference:D})}else o.currentContext=await o.jsepCreateMLContext();break}i=await o._OrtCreateSession(r,n,a),i===0&&pe("Can't create a session."),o.jsepOnCreateSession?.(),o.currentContext&&(o.jsepRegisterMLContext(i,o.currentContext),o.currentContext=void 0,o.shouldTransferToMLTensor=!0);let[u,h]=hg(i),_=!!t?.enableGraphCapture,y=[],g=[],x=[];for(let v=0;v<u;v++){let S=o._OrtGetInputName(i,v);S===0&&pe("Can't get an input name."),p.push(S),y.push(o.UTF8ToString(S))}for(let v=0;v<h;v++){let S=o._OrtGetOutputName(i,v);S===0&&pe("Can't get an output name."),m.push(S);let T=o.UTF8ToString(S);g.push(T);{if(_&&t?.preferredOutputLocation===void 0){x.push("gpu-buffer");continue}let A=typeof t?.preferredOutputLocation=="string"?t.preferredOutputLocation:t?.preferredOutputLocation?.[T]??"cpu";if(A!=="cpu"&&A!=="cpu-pinned"&&A!=="gpu-buffer"&&A!=="ml-tensor")throw new Error(`Not supported preferred output location: ${A}.`);if(_&&A!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${A}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);x.push(A)}}let $=null;return x.some(v=>v==="gpu-buffer"||v==="ml-tensor")&&(d=o._OrtCreateBinding(i),d===0&&pe("Can't create IO binding."),$={handle:d,outputPreferredLocations:x,outputPreferredLocationsEncoded:x.map(v=>Qn(v))}),Pt.set(i,[i,p,m,$,_,!1]),[i,y,g]}catch(u){throw p.forEach(h=>o._OrtFree(h)),m.forEach(h=>o._OrtFree(h)),d!==0&&o._OrtReleaseBinding(d)!==0&&pe("Can't release IO binding."),i!==0&&o._OrtReleaseSession(i)!==0&&pe("Can't release session."),u}finally{o._free(r),a!==0&&o._OrtReleaseSessionOptions(a)!==0&&pe("Can't release session options."),l.forEach(u=>o._free(u)),o.unmountExternalData?.()}},Ar=e=>{let t=Ie(),r=Pt.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[n,o,i,a,d]=r;a&&(d&&t._OrtClearBoundOutputs(a.handle)!==0&&pe("Can't clear bound outputs."),t._OrtReleaseBinding(a.handle)!==0&&pe("Can't release IO binding.")),t.jsepOnReleaseSession?.(e),o.forEach(l=>t._OrtFree(l)),i.forEach(l=>t._OrtFree(l)),t._OrtReleaseSession(n)!==0&&pe("Can't release session."),Pt.delete(e)},uc=(e,t,r,n,o,i=!1)=>{if(!e){t.push(0);return}let a=Ie(),d=a.PTR_SIZE,l=e[0],p=e[1],m=e[3],u,h;if(l==="string"&&(m==="gpu-buffer"||m==="ml-tensor"))throw new Error("String tensor is not supported on GPU.");if(i&&m!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${o} when enableGraphCapture is true.`);if(m==="gpu-buffer"){let g=e[2].gpuBuffer;h=Ct(Yt(l),p);let x=a.jsepRegisterBuffer;if(!x)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');u=x(n,o,g,h)}else if(m==="ml-tensor"){let g=e[2].mlTensor;h=Ct(Yt(l),p);let x=a.jsepRegisterMLTensor;if(!x)throw new Error('Tensor location "ml-tensor" is not supported without using WebNN.');u=x(g,Yt(l),p)}else{let g=e[2];if(Array.isArray(g)){h=d*g.length,u=a._malloc(h),r.push(u);for(let x=0;x<g.length;x++){if(typeof g[x]!="string")throw new TypeError(`tensor data at index ${x} is not a string`);a.setValue(u+x*d,ke(g[x],r),"*")}}else h=g.byteLength,u=a._malloc(h),r.push(u),a.HEAPU8.set(new Uint8Array(g.buffer,g.byteOffset,h),u)}let _=a.stackSave(),y=a.stackAlloc(4*p.length);try{p.forEach((x,$)=>a.setValue(y+$*d,x,d===4?"i32":"i64"));let g=a._OrtCreateTensor(Yt(l),u,h,y,p.length,Qn(m));g===0&&pe(`Can't create tensor for input/output. session=${n}, index=${o}.`),t.push(g)}finally{a.stackRestore(_)}},kr=async(e,t,r,n,o,i)=>{let a=Ie(),d=a.PTR_SIZE,l=Pt.get(e);if(!l)throw new Error(`cannot run inference. invalid session id: ${e}`);let p=l[0],m=l[1],u=l[2],h=l[3],_=l[4],y=l[5],g=t.length,x=n.length,$=0,v=[],S=[],T=[],A=[],k=a.stackSave(),P=a.stackAlloc(g*d),D=a.stackAlloc(g*d),R=a.stackAlloc(x*d),G=a.stackAlloc(x*d);try{a.jsepOnRunStart?.(p),[$,v]=Ga(i);for(let V=0;V<g;V++)uc(r[V],S,A,e,t[V],_);for(let V=0;V<x;V++)uc(o[V],T,A,e,g+n[V],_);for(let V=0;V<g;V++)a.setValue(P+V*d,S[V],"*"),a.setValue(D+V*d,m[t[V]],"*");for(let V=0;V<x;V++)a.setValue(R+V*d,T[V],"*"),a.setValue(G+V*d,u[n[V]],"*");if(h&&!y){let{handle:V,outputPreferredLocations:Q,outputPreferredLocationsEncoded:se}=h;if(m.length!==g)throw new Error(`input count from feeds (${g}) is expected to be always equal to model's input count (${m.length}).`);for(let Y=0;Y<g;Y++){let ee=t[Y];await a._OrtBindInput(V,m[ee],S[Y])!==0&&pe(`Can't bind input[${Y}] for session=${e}.`)}for(let Y=0;Y<x;Y++){let ee=n[Y];o[Y]?.[3]?a._OrtBindOutput(V,u[ee],T[Y],0)!==0&&pe(`Can't bind pre-allocated output[${Y}] for session=${e}.`):a._OrtBindOutput(V,u[ee],0,se[ee])!==0&&pe(`Can't bind output[${Y}] to ${Q[Y]} for session=${e}.`)}Pt.set(e,[p,m,u,h,_,!0])}let K;h?K=await a._OrtRunWithBinding(p,h.handle,x,R,$):K=await a._OrtRun(p,D,P,g,G,x,R,$),K!==0&&pe("failed to call OrtRun().");let j=[];for(let V=0;V<x;V++){let Q=Number(a.getValue(R+V*d,"*"));if(Q===T[V]){j.push(o[V]);continue}let se=a.stackSave(),Y=a.stackAlloc(4*d),ee=!1,J,ne=0;try{a._OrtGetTensorData(Q,Y,Y+d,Y+2*d,Y+3*d)!==0&&pe(`Can't access output tensor data on index ${V}.`);let Oe=d===4?"i32":"i64",$e=Number(a.getValue(Y,Oe));ne=a.getValue(Y+d,"*");let le=a.getValue(Y+d*2,"*"),W=Number(a.getValue(Y+d*3,Oe)),q=[];for(let we=0;we<W;we++)q.push(Number(a.getValue(le+we*d,Oe)));a._OrtFree(le)!==0&&pe("Can't free memory for tensor dims.");let he=q.reduce((we,ye)=>we*ye,1);J=bt($e);let Ge=h?.outputPreferredLocations[n[V]];if(J==="string"){if(Ge==="gpu-buffer"||Ge==="ml-tensor")throw new Error("String tensor is not supported on GPU.");let we=[];for(let ye=0;ye<he;ye++){let Ye=a.getValue(ne+ye*d,"*"),Lt=a.getValue(ne+(ye+1)*d,"*"),fn=ye===he-1?void 0:Lt-Ye;we.push(a.UTF8ToString(Ye,fn))}j.push([J,q,we,"cpu"])}else if(Ge==="gpu-buffer"&&he>0){let we=a.jsepGetBuffer;if(!we)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let ye=we(ne),Ye=Ct($e,he);if(Ye===void 0||!Br(J))throw new Error(`Unsupported data type: ${J}`);ee=!0,j.push([J,q,{gpuBuffer:ye,download:a.jsepCreateDownloader(ye,Ye,J),dispose:()=>{a._OrtReleaseTensor(Q)!==0&&pe("Can't release tensor.")}},"gpu-buffer"])}else if(Ge==="ml-tensor"&&he>0){let we=a.jsepEnsureTensor;if(!we)throw new Error('preferredLocation "ml-tensor" is not supported without using WebNN.');if(Ct($e,he)===void 0||!Mr(J))throw new Error(`Unsupported data type: ${J}`);let Ye=await we(ne,$e,q,!1);ee=!0,j.push([J,q,{mlTensor:Ye,download:a.jsepCreateMLTensorDownloader(ne,J),dispose:()=>{a.jsepReleaseTensorId(ne),a._OrtReleaseTensor(Q)}},"ml-tensor"])}else{let we=Dr(J),ye=new we(he);new Uint8Array(ye.buffer,ye.byteOffset,ye.byteLength).set(a.HEAPU8.subarray(ne,ne+ye.byteLength)),j.push([J,q,ye,"cpu"])}}finally{a.stackRestore(se),J==="string"&&ne&&a._free(ne),ee||a._OrtReleaseTensor(Q)}}return h&&!_&&(a._OrtClearBoundOutputs(h.handle)!==0&&pe("Can't clear bound outputs."),Pt.set(e,[p,m,u,h,_,!1])),j}finally{a.stackRestore(k),S.forEach(K=>a._OrtReleaseTensor(K)),T.forEach(K=>a._OrtReleaseTensor(K)),A.forEach(K=>a._free(K)),$!==0&&a._OrtReleaseRunOptions($),v.forEach(K=>a._free(K))}},Er=e=>{let t=Ie(),r=Pt.get(e);if(!r)throw new Error("invalid session id");let n=r[0],o=t._OrtEndProfiling(n);o===0&&pe("Can't get an profile file name."),t._OrtFree(o)},Pr=e=>{let t=[];for(let r of e){let n=r[2];!Array.isArray(n)&&"buffer"in n&&t.push(n.buffer)}return t}});var zt,Le,nr,ln,cn,dn,Po,zo,Vt,Wt,bg,dc,lc,cc,pc,mc,fc,hc,Oo=U(()=>{"use strict";We();qn();gt();xr();zt=()=>!!ve.wasm.proxy&&typeof document<"u",nr=!1,ln=!1,cn=!1,zo=new Map,Vt=(e,t)=>{let r=zo.get(e);r?r.push(t):zo.set(e,[t])},Wt=()=>{if(nr||!ln||cn||!Le)throw new Error("worker not ready")},bg=e=>{switch(e.data.type){case"init-wasm":nr=!1,e.data.err?(cn=!0,Po[1](e.data.err)):(ln=!0,Po[0]()),dn&&(URL.revokeObjectURL(dn),dn=void 0);break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let t=zo.get(e.data.type);e.data.err?t.shift()[1](e.data.err):t.shift()[0](e.data.out);break}default:}},dc=async()=>{if(!ln){if(nr)throw new Error("multiple calls to 'initWasm()' detected.");if(cn)throw new Error("previous call to 'initWasm()' failed.");if(nr=!0,zt())return new Promise((e,t)=>{Le?.terminate(),Va().then(([r,n])=>{try{Le=n,Le.onerror=i=>t(i),Le.onmessage=bg,Po=[e,t];let o={type:"init-wasm",in:ve};!o.in.wasm.wasmPaths&&(r||import.meta.url?.startsWith("file:"))&&(o.in.wasm.wasmPaths={wasm:new URL(/* asset import */ __webpack_require__(/*! ort-wasm-simd-threaded.jsep.wasm */ "./node_modules/onnxruntime-web/dist/ort-wasm-simd-threaded.jsep.wasm"), __webpack_require__.b).href}),Le.postMessage(o),dn=r}catch(o){t(o)}},t)});try{await Sr(ve.wasm),await Tr(ve),ln=!0}catch(e){throw cn=!0,e}finally{nr=!1}}},lc=async e=>{if(zt())return Wt(),new Promise((t,r)=>{Vt("init-ep",[t,r]);let n={type:"init-ep",in:{epName:e,env:ve}};Le.postMessage(n)});await Ir(ve,e)},cc=async e=>zt()?(Wt(),new Promise((t,r)=>{Vt("copy-from",[t,r]);let n={type:"copy-from",in:{buffer:e}};Le.postMessage(n,[e.buffer])})):Kt(e),pc=async(e,t)=>{if(zt()){if(t?.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return Wt(),new Promise((r,n)=>{Vt("create",[r,n]);let o={type:"create",in:{model:e,options:{...t}}},i=[];e instanceof Uint8Array&&i.push(e.buffer),Le.postMessage(o,i)})}else return Cr(e,t)},mc=async e=>{if(zt())return Wt(),new Promise((t,r)=>{Vt("release",[t,r]);let n={type:"release",in:e};Le.postMessage(n)});Ar(e)},fc=async(e,t,r,n,o,i)=>{if(zt()){if(r.some(a=>a[3]!=="cpu"))throw new Error("input tensor on GPU is not supported for proxy.");if(o.some(a=>a))throw new Error("pre-allocated output tensor is not supported for proxy.");return Wt(),new Promise((a,d)=>{Vt("run",[a,d]);let l=r,p={type:"run",in:{sessionId:e,inputIndices:t,inputs:l,outputIndices:n,options:i}};Le.postMessage(p,Pr(l))})}else return kr(e,t,r,n,o,i)},hc=async e=>{if(zt())return Wt(),new Promise((t,r)=>{Vt("end-profiling",[t,r]);let n={type:"end-profiling",in:e};Le.postMessage(n)});Er(e)}});var gc,yg,pn,bc=U(()=>{"use strict";We();Oo();te();$r();Xn();gc=(e,t)=>{switch(e.location){case"cpu":return[e.type,e.dims,e.data,"cpu"];case"gpu-buffer":return[e.type,e.dims,{gpuBuffer:e.gpuBuffer},"gpu-buffer"];case"ml-tensor":return[e.type,e.dims,{mlTensor:e.mlTensor},"ml-tensor"];default:throw new Error(`invalid data location: ${e.location} for ${t()}`)}},yg=e=>{switch(e[3]){case"cpu":return new He(e[0],e[2],e[1]);case"gpu-buffer":{let t=e[0];if(!Br(t))throw new Error(`not supported data type: ${t} for deserializing GPU tensor`);let{gpuBuffer:r,download:n,dispose:o}=e[2];return He.fromGpuBuffer(r,{dataType:t,dims:e[1],download:n,dispose:o})}case"ml-tensor":{let t=e[0];if(!Mr(t))throw new Error(`not supported data type: ${t} for deserializing MLTensor tensor`);let{mlTensor:r,download:n,dispose:o}=e[2];return He.fromMLTensor(r,{dataType:t,dims:e[1],download:n,dispose:o})}default:throw new Error(`invalid data location: ${e[3]}`)}},pn=class{async fetchModelAndCopyToWasmMemory(t){return cc(await Qt(t))}async loadModel(t,r){Ue();let n;typeof t=="string"? false?0:n=await this.fetchModelAndCopyToWasmMemory(t):n=t,[this.sessionId,this.inputNames,this.outputNames]=await pc(n,r),De()}async dispose(){return mc(this.sessionId)}async run(t,r,n){Ue();let o=[],i=[];Object.entries(t).forEach(h=>{let _=h[0],y=h[1],g=this.inputNames.indexOf(_);if(g===-1)throw new Error(`invalid input '${_}'`);o.push(y),i.push(g)});let a=[],d=[];Object.entries(r).forEach(h=>{let _=h[0],y=h[1],g=this.outputNames.indexOf(_);if(g===-1)throw new Error(`invalid output '${_}'`);a.push(y),d.push(g)});let l=o.map((h,_)=>gc(h,()=>`input "${this.inputNames[i[_]]}"`)),p=a.map((h,_)=>h?gc(h,()=>`output "${this.outputNames[d[_]]}"`):null),m=await fc(this.sessionId,i,l,d,p,n),u={};for(let h=0;h<m.length;h++)u[this.outputNames[d[h]]]=a[h]??yg(m[h]);return De(),u}startProfiling(){}endProfiling(){hc(this.sessionId)}}});var _c={};Ft(_c,{OnnxruntimeWebAssemblyBackend:()=>mn,initializeFlags:()=>yc,wasmBackend:()=>_g});var yc,mn,_g,wc=U(()=>{"use strict";We();Oo();bc();yc=()=>{if((typeof ve.wasm.initTimeout!="number"||ve.wasm.initTimeout<0)&&(ve.wasm.initTimeout=0),ve.wasm.simd===!1&&console.warn('Deprecated property "env.wasm.simd" is set to false. non-SIMD build is no longer provided, and this setting will be ignored.'),typeof ve.wasm.proxy!="boolean"&&(ve.wasm.proxy=!1),typeof ve.wasm.trace!="boolean"&&(ve.wasm.trace=!1),typeof ve.wasm.numThreads!="number"||!Number.isInteger(ve.wasm.numThreads)||ve.wasm.numThreads<=0)if(typeof self<"u"&&!self.crossOriginIsolated)ve.wasm.numThreads=1;else{let e=typeof navigator>"u"?Nn("node:os").cpus().length:navigator.hardwareConcurrency;ve.wasm.numThreads=Math.min(4,Math.ceil((e||1)/2))}},mn=class{async init(t){yc(),await dc(),await lc(t)}async createInferenceSessionHandler(t,r){let n=new pn;return await n.loadModel(t,r),Promise.resolve(n)}},_g=new mn});We();We();We();var Ca="1.21.0-dev.20250114-228dd16893";var $1=Fn;{let e=(wc(),br(_c)).wasmBackend;St("webgpu",e,5),St("webnn",e,5),St("cpu",e,10),St("wasm",e,10)}Object.defineProperty(ve.versions,"web",{value:Ca,enumerable:!0});
+${d}`,p=n.createShaderModule({code:l,label:t.name});se("verbose",()=>`[WebGPU] ${t.name} shader code: ${l}`);let m=n.createComputePipeline({compute:{module:p,entryPoint:"main"},layout:"auto",label:t.name});return Be(t.name),{programInfo:t,computePipeline:m,uniformVariablesInfo:a.variablesInfo}}normalizeDispatchGroupSize(t){let r=typeof t=="number"?t:t.x,n=typeof t=="number"?1:t.y||1,o=typeof t=="number"?1:t.z||1,i=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(r<=i&&n<=i&&o<=i)return[r,n,o];let a=r*n*o,d=Math.ceil(Math.sqrt(a));if(d>i){if(d=Math.ceil(Math.cbrt(a)),d>i)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[d,d,d]}else return[d,d,1]}}});var ug,dg,Io,Co,nn,Jl=U(()=>{"use strict";Le();J();et();Xn();Xa();Ql();Xl();ug=(e,t)=>{if(t.length!==e.length)throw new Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let r=[];for(let n=0;n<e.length;++n){let o=e[n].dataType;switch(t[n]){case"none":{r.push("");break}case"type":{r.push(`${o}`);break}case"rank":{let i=e[n].dims.length;r.push(`${o};${i}`);break}case"dims":{let i=e[n].dims.join(",");r.push(`${o};${i}`);break}default:throw new Error(`unsupported input dependency: ${t[n]}`)}}return r.join("|")},dg=(e,t,r)=>{let n=e.name;return e.shaderCache?.hint&&(n+="["+e.shaderCache.hint+"]"),n+=":"+r+`:${ug(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,n},Io=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},Co=class{constructor(t){this.subgroupsSupported=t.features.has("subgroups"),this.subgroupsF16Supported=t.features.has("subgroups");let r=t.limits;!this.subgroupsSupported||!r.minSubgroupSize||!r.maxSubgroupSize?this.subgroupSizeRange=void 0:this.subgroupSizeRange=[r.minSubgroupSize,r.maxSubgroupSize]}},nn=class{constructor(){this.currentSessionId=null;this.currentKernelId=null;this.commandEncoder=null;this.computePassEncoder=null;this.maxDispatchNumber=16;this.pendingDispatchNumber=0;this.pendingKernels=[];this.pendingQueries=new Map;this.sessionStatus="default";this.capturedCommandList=new Map;this.capturedPendingKernels=new Map;this.sessionExternalDataMapping=new Map}get currentKernelCustomData(){if(this.currentKernelId===null)throw new Error("currentKernelCustomData(): currentKernelId is null. (should not happen)");let t=this.kernelCustomData.get(this.currentKernelId);return t||(t={},this.kernelCustomData.set(this.currentKernelId,t)),t}async initialize(t,r){this.env=t;let n=[],o={requiredLimits:{maxComputeWorkgroupStorageSize:r.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.limits.maxStorageBufferBindingSize,maxBufferSize:r.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:r.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:r.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:r.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:r.limits.maxComputeWorkgroupSizeZ},requiredFeatures:n},i=a=>r.features.has(a)&&n.push(a)&&!0;i("chromium-experimental-timestamp-query-inside-passes")||i("timestamp-query"),i("shader-f16"),i("subgroups")&&i("subgroups-f16"),this.device=await r.requestDevice(o),this.deviceInfo=new Co(this.device),this.adapterInfo=new Io(r.info||await r.requestAdapterInfo()),this.gpuDataManager=Qa(this),this.programManager=new rn(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,Mr(t.logLevel,!!t.debug),this.device.onuncapturederror=a=>{a.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${a.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value:r,writable:!1,enumerable:!0,configurable:!1}),this.setQueryType()}dispose(){typeof this.querySet<"u"&&this.querySet.destroy(),this.gpuDataManager.dispose()}getCommandEncoder(){return this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder()),this.commandEncoder}getComputePassEncoder(){if(!this.computePassEncoder){let t=this.getCommandEncoder(),r={};this.queryType==="at-passes"&&(r.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=t.beginComputePass(r)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;Ne(),this.endComputePass();let t;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),t=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(t,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&t.mapAsync(GPUMapMode.READ).then(()=>{let r=new BigUint64Array(t.getMappedRange()),n=this.pendingQueries.get(t);for(let o=0;o<r.length/2;o++){let i=n[o],a=i.kernelId,d=this.kernels.get(a),l=d.kernelType,p=d.kernelName,m=i.programName,u=i.inputTensorViews,h=i.outputTensorViews,_=r[o*2],y=r[o*2+1];typeof this.queryTimeBase>"u"&&(this.queryTimeBase=_);let g=Number(_-this.queryTimeBase),x=Number(y-this.queryTimeBase);if(!Number.isSafeInteger(g)||!Number.isSafeInteger(x))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:u.map($=>({dims:$.dims,dataType:bt($.dataType)})),outputsMetadata:h.map($=>({dims:$.dims,dataType:bt($.dataType)})),kernelId:a,kernelType:l,kernelName:p,programName:m,startTime:g,endTime:x});else{let $="";u.forEach((S,T)=>{$+=`input[${T}]: [${S.dims}] | ${bt(S.dataType)}, `});let v="";h.forEach((S,T)=>{v+=`output[${T}]: [${S.dims}] | ${bt(S.dataType)}, `}),console.log(`[profiling] kernel "${a}|${l}|${p}|${m}" ${$}${v}execution time: ${x-g} ns`)}_r("GPU",`${m}::${_}::${y}`)}t.unmap(),this.pendingQueries.delete(t)}),Be()}run(t,r,n,o,i,a){Ne(t.name);let d=[];for(let S=0;S<r.length;++S){let T=r[S].data;if(T===0)continue;let A=this.gpuDataManager.get(T);if(!A)throw new Error(`no GPU data for input: ${T}`);d.push(A)}let{outputs:l,dispatchGroup:p,programUniforms:m}=t.getRunData(r),u=n.length===0?l.map((S,T)=>T):n;if(u.length!==l.length)throw new Error(`Output size ${u.length} must be equal to ${l.length}.`);let h=[],_=[];for(let S=0;S<l.length;++S){if(!Number.isInteger(u[S])||u[S]<-3||u[S]>=a)throw new Error(`Invalid output index: ${u[S]}`);if(u[S]===-3)continue;let T=u[S]===-1,A=u[S]===-2,C=T||A?i(l[S].dataType,l[S].dims):o(u[S],l[S].dataType,l[S].dims);if(h.push(C),C.data===0)continue;let P=this.gpuDataManager.get(C.data);if(!P)throw new Error(`no GPU data for output: ${C.data}`);if(T&&this.temporaryData.push(P),A){let D=this.kernelPersistentData.get(this.currentKernelId);D||(D=[],this.kernelPersistentData.set(this.currentKernelId,D)),D.push(P)}_.push(P)}if(d.length!==r.length||_.length!==h.length){if(_.length===0)return Be(t.name),h;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let y;if(m){let S=0,T=[];m.forEach(D=>{let R=typeof D.data=="number"?[D.data]:D.data;if(R.length===0)return;let H=D.type===10?2:4,L,re;D.type===10?(re=R.length>4?16:R.length>2?8:R.length*H,L=R.length>4?16:H*R.length):(re=R.length<=2?R.length*H:16,L=16),S=Math.ceil(S/re)*re,T.push(S);let V=D.type===10?8:4;S+=R.length>4?Math.ceil(R.length/V)*L:R.length*H});let A=16;S=Math.ceil(S/A)*A;let C=new ArrayBuffer(S);m.forEach((D,R)=>{let H=T[R],L=typeof D.data=="number"?[D.data]:D.data;if(D.type===6)new Int32Array(C,H,L.length).set(L);else if(D.type===12)new Uint32Array(C,H,L.length).set(L);else if(D.type===10)new Uint16Array(C,H,L.length).set(L);else if(D.type===1)new Float32Array(C,H,L.length).set(L);else throw new Error(`Unsupported uniform type: ${bt(D.type)}`)});let P=this.gpuDataManager.create(S,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(P.buffer,0,C,0,S),this.gpuDataManager.release(P.id),y={offset:0,size:S,buffer:P.buffer}}let g=this.programManager.normalizeDispatchGroupSize(p),x=g[1]===1&&g[2]===1,$=dg(t,r,x),v=this.programManager.getArtifact($);if(v||(v=this.programManager.build(t,g),this.programManager.setArtifact($,v),se("info",()=>`[artifact] key: ${$}, programName: ${t.name}`)),m&&v.uniformVariablesInfo){if(m.length!==v.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${v.uniformVariablesInfo.length}, got ${m.length} in program "${v.programInfo.name}".`);for(let S=0;S<m.length;S++){let T=m[S],A=T.type,C=typeof T.data=="number"?1:T.data.length,[P,D]=v.uniformVariablesInfo[S];if(A!==P||C!==D)throw new Error(`Uniform variable ${S} mismatch: expect type ${P} with size ${D}, got type ${A} with size ${C} in program "${v.programInfo.name}".`)}}if(se("info",()=>`[ProgramManager] run "${t.name}" (key=${$}) with ${g[0]}x${g[1]}x${g[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let S={kernelId:this.currentKernelId,programName:v.programInfo.name,inputTensorViews:r,outputTensorViews:h};this.pendingKernels.push(S),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(S)}return this.programManager.run(v,d,_,g,y),Be(t.name),h}upload(t,r){this.gpuDataManager.upload(t,r)}memcpy(t,r){this.gpuDataManager.memcpy(t,r)}async download(t,r){await this.gpuDataManager.download(t,r)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,r,n,o){let i=Zl.get(t);if(!i)throw new Error(`kernel not implemented: ${t}`);let a={kernelType:t,kernelName:o,kernelEntry:i[0],attributes:[i[1],n]};this.kernels.set(r,a)}releaseKernel(t){let r=this.kernelPersistentData.get(t);if(r){for(let n of r)this.gpuDataManager.release(n.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,r,n){let o=this.kernels.get(t);if(!o)throw new Error(`kernel not created: ${t}`);let i=o.kernelType,a=o.kernelName,d=o.kernelEntry,l=o.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${i}] ${a}" is not allowed to be called recursively`);this.currentKernelId=t,l[0]&&(l[1]=l[0](l[1]),l[0]=void 0),se("info",()=>`[WebGPU] Start to run kernel "[${i}] ${a}"...`);let p=this.env.debug;this.temporaryData=[];try{return p&&this.device.pushErrorScope("validation"),d(r,l[1]),0}catch(m){return n.push(Promise.resolve(`[WebGPU] Kernel "[${i}] ${a}" failed. ${m}`)),1}finally{p&&n.push(this.device.popErrorScope().then(m=>m?`GPU validation error for kernel "[${i}] ${a}": ${m.message}`:null));for(let m of this.temporaryData)this.gpuDataManager.release(m.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,r,n,o){let i=this.sessionExternalDataMapping.get(t);i||(i=new Map,this.sessionExternalDataMapping.set(t,i));let a=i.get(r),d=this.gpuDataManager.registerExternalBuffer(n,o,a);return i.set(r,[d,n]),d}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(n=>this.gpuDataManager.unregisterExternalBuffer(n[0])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let r=this.gpuDataManager.get(t);if(!r)throw new Error(`no GPU data for buffer: ${t}`);return r.buffer}createDownloader(t,r,n){return async()=>{let o=await ro(this,t,r);return Rr(o.buffer,n)}}writeTimestamp(t){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,t)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||(typeof this.env.trace>"u"?this.env.wasm.trace:this.env.trace))&&(this.device.features.has("chromium-experimental-timestamp-query-inside-passes")?this.queryType="inside-passes":this.device.features.has("timestamp-query")&&(this.queryType="at-passes"),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){se("info","captureBegin"),this.capturedCommandList.get(this.currentSessionId)||this.capturedCommandList.set(this.currentSessionId,[]),this.capturedPendingKernels.get(this.currentSessionId)||this.capturedPendingKernels.set(this.currentSessionId,[]),this.flush(),this.sessionStatus="capturing"}captureEnd(){se("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){se("info","replay"),this.sessionStatus="replaying";let t=this.capturedCommandList.get(this.currentSessionId),r=this.capturedPendingKernels.get(this.currentSessionId),n=t.length;this.pendingKernels=[];for(let o=0;o<n;o++){let i=this.getComputePassEncoder(),a=t[o];this.writeTimestamp(this.pendingDispatchNumber*2),i.setPipeline(a.computePipeline),i.setBindGroup(0,a.bindGroup),i.dispatchWorkgroups(...a.dispatchGroup),this.writeTimestamp(this.pendingDispatchNumber*2+1),this.pendingDispatchNumber++,this.queryType!=="none"&&this.pendingKernels.push(r[o]),(this.pendingDispatchNumber>=this.maxDispatchNumber||this.queryType==="at-passes")&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}onCreateSession(){this.gpuDataManager.onCreateSession()}onReleaseSession(t){this.unregisterBuffers(t),this.capturedCommandList.has(t)&&this.capturedCommandList.delete(t),this.capturedPendingKernels.has(t)&&this.capturedPendingKernels.delete(t),this.gpuDataManager.onReleaseSession(t)}onRunStart(t){this.currentSessionId=t,this.setQueryType()}}});var lg,ec,cg,tc,on,an,Ao,rc,nc=U(()=>{"use strict";et();lg=1,ec=()=>lg++,cg=new Map([["float32",32],["float16",16],["int32",32],["uint32",32],["int64",64],["uint64",64],["int8",8],["uint8",8],["int4",4],["uint4",4]]),tc=(e,t)=>{let r=cg.get(e);if(!r)throw new Error("Unsupported data type.");return t.length>0?Math.ceil(t.reduce((n,o)=>n*o)*r/8):0},on=class{constructor(t){this.sessionId=t.sessionId,this.mlContext=t.context,this.mlTensor=t.tensor,this.dataType=t.dataType,this.tensorShape=t.shape}get tensor(){return this.mlTensor}get type(){return this.dataType}get shape(){return this.tensorShape}get byteLength(){return tc(this.dataType,this.tensorShape)}destroy(){se("verbose",()=>"[WebNN] TensorWrapper.destroy"),this.mlTensor.destroy()}write(t){this.mlContext.writeTensor(this.mlTensor,t)}async read(t){return t?this.mlContext.readTensor(this.mlTensor,t):this.mlContext.readTensor(this.mlTensor)}canReuseTensor(t,r,n){return this.mlContext===t&&this.dataType===r&&this.tensorShape.length===n.length&&this.tensorShape.every((o,i)=>o===n[i])}},an=class{constructor(t,r){this.tensorManager=t;this.wrapper=r}get tensorWrapper(){return this.wrapper}releaseTensor(){this.tensorWrapper&&(this.tensorManager.releaseTensor(this.tensorWrapper),this.wrapper=void 0)}async ensureTensor(t,r,n,o){if(this.wrapper){if(this.wrapper.canReuseTensor(t,r,n))return this.wrapper.tensor;if(o){if(this.wrapper.byteLength!==tc(r,n))throw new Error("Unable to copy data to tensor with different size.");this.activeUpload=new Uint8Array(await this.wrapper.read())}this.tensorManager.releaseTensor(this.wrapper)}let i=typeof MLTensorUsage>"u"?void 0:MLTensorUsage.READ|MLTensorUsage.WRITE;return this.wrapper=await this.tensorManager.getCachedTensor(r,n,i,!0,!0),o&&this.activeUpload&&(this.wrapper.write(this.activeUpload),this.activeUpload=void 0),this.wrapper.tensor}upload(t){if(this.wrapper)if(t.byteLength===this.wrapper.byteLength){this.wrapper.write(t);return}else se("verbose",()=>"Data size does not match tensor size. Releasing tensor."),this.releaseTensor();this.activeUpload?this.activeUpload.set(t):this.activeUpload=new Uint8Array(t)}async download(t){if(this.activeUpload)if(t){t instanceof ArrayBuffer?new Uint8Array(t).set(this.activeUpload):new Uint8Array(t.buffer,t.byteOffset,t.byteLength).set(this.activeUpload);return}else return this.activeUpload.buffer;if(!this.wrapper)throw new Error("Tensor has not been created.");return t?this.wrapper.read(t):this.wrapper.read()}},Ao=class{constructor(t){this.backend=t;this.tensorTrackersById=new Map;this.freeTensors=[];this.externalTensors=new Set}reserveTensorId(){let t=ec();return this.tensorTrackersById.set(t,new an(this)),t}releaseTensorId(t){let r=this.tensorTrackersById.get(t);r&&(this.tensorTrackersById.delete(t),r.tensorWrapper&&this.releaseTensor(r.tensorWrapper))}async ensureTensor(t,r,n,o){se("verbose",()=>`[WebNN] TensorManager.ensureTensor {tensorId: ${t}, dataType: ${r}, shape: ${n}, copyOld: ${o}}`);let i=this.tensorTrackersById.get(t);if(!i)throw new Error("Tensor not found.");return i.ensureTensor(this.backend.currentContext,r,n,o)}upload(t,r){let n=this.tensorTrackersById.get(t);if(!n)throw new Error("Tensor not found.");n.upload(r)}async download(t,r){se("verbose",()=>`[WebNN] TensorManager.download {tensorId: ${t}, dstBuffer: ${r?.byteLength}}`);let n=this.tensorTrackersById.get(t);if(!n)throw new Error("Tensor not found.");return n.download(r)}releaseTensorsForSession(t){for(let r of this.freeTensors)r.sessionId===t&&r.destroy();this.freeTensors=this.freeTensors.filter(r=>r.sessionId!==t)}registerTensor(t,r,n,o){let i=ec(),a=new on({sessionId:this.backend.currentSessionId,context:t,tensor:r,dataType:n,shape:o});return this.tensorTrackersById.set(i,new an(this,a)),this.externalTensors.add(a),i}async getCachedTensor(t,r,n,o,i){let a=this.backend.currentSessionId,d=this.backend.currentContext;for(let[p,m]of this.freeTensors.entries())if(m.canReuseTensor(d,t,r)){se("verbose",()=>`[WebNN] Reusing tensor {dataType: ${t}, shape: ${r}}`);let u=this.freeTensors.splice(p,1)[0];return u.sessionId=a,u}se("verbose",()=>`[WebNN] MLContext.createTensor {dataType: ${t}, shape: ${r}}`);let l=await d.createTensor({dataType:t,shape:r,dimensions:r,usage:n,writable:o,readable:i});return new on({sessionId:a,context:d,tensor:l,dataType:t,shape:r})}releaseTensor(t){this.externalTensors.has(t)&&this.externalTensors.delete(t),this.freeTensors.push(t)}},rc=(...e)=>new Ao(...e)});var oc,pg,sn,ic=U(()=>{"use strict";J();gt();Xn();nc();et();oc=new Map([[1,"float32"],[10,"float16"],[6,"int32"],[12,"uint32"],[7,"int64"],[13,"uint64"],[22,"int4"],[21,"uint4"],[3,"int8"],[2,"uint8"],[9,"uint8"]]),pg=(e,t)=>{if(e===t)return!0;if(e===void 0||t===void 0)return!1;let r=Object.keys(e).sort(),n=Object.keys(t).sort();return r.length===n.length&&r.every((o,i)=>o===n[i]&&e[o]===t[o])},sn=class{constructor(t){this.tensorManager=rc(this);this.mlContextBySessionId=new Map;this.sessionIdsByMLContext=new Map;this.mlContextCache=[];Mr(t.logLevel,!!t.debug)}get currentSessionId(){if(this.activeSessionId===void 0)throw new Error("No active session");return this.activeSessionId}onRunStart(t){this.activeSessionId=t}async createMLContext(t){if(t instanceof GPUDevice){let n=this.mlContextCache.findIndex(o=>o.gpuDevice===t);if(n!==-1)return this.mlContextCache[n].mlContext;{let o=await navigator.ml.createContext(t);return this.mlContextCache.push({gpuDevice:t,mlContext:o}),o}}else if(t===void 0){let n=this.mlContextCache.findIndex(o=>o.options===void 0&&o.gpuDevice===void 0);if(n!==-1)return this.mlContextCache[n].mlContext;{let o=await navigator.ml.createContext();return this.mlContextCache.push({mlContext:o}),o}}let r=this.mlContextCache.findIndex(n=>pg(n.options,t));if(r!==-1)return this.mlContextCache[r].mlContext;{let n=await navigator.ml.createContext(t);return this.mlContextCache.push({options:t,mlContext:n}),n}}get currentContext(){let t=this.getMLContext(this.currentSessionId);if(!t)throw new Error(`No MLContext found for session ${this.currentSessionId}`);return t}registerMLContext(t,r){this.mlContextBySessionId.set(t,r);let n=this.sessionIdsByMLContext.get(r);n||(n=new Set,this.sessionIdsByMLContext.set(r,n)),n.add(t)}onReleaseSession(t){let r=this.mlContextBySessionId.get(t);if(!r)return;this.tensorManager.releaseTensorsForSession(t),this.mlContextBySessionId.delete(t);let n=this.sessionIdsByMLContext.get(r);if(n.delete(t),n.size===0){this.sessionIdsByMLContext.delete(r);let o=this.mlContextCache.findIndex(i=>i.mlContext===r);o!==-1&&this.mlContextCache.splice(o,1)}}getMLContext(t){return this.mlContextBySessionId.get(t)}reserveTensorId(){return this.tensorManager.reserveTensorId()}releaseTensorId(t){se("verbose",()=>`[WebNN] releaseTensorId {tensorId: ${t}}`),this.tensorManager.releaseTensorId(t)}async ensureTensor(t,r,n,o){let i=oc.get(r);if(!i)throw new Error(`Unsupported ONNX data type: ${r}`);return this.tensorManager.ensureTensor(t,i,n,o)}uploadTensor(t,r){if(!Te().shouldTransferToMLTensor)throw new Error("Trying to upload to a MLTensor while shouldTransferToMLTensor is false");se("verbose",()=>`[WebNN] uploadTensor {tensorId: ${t}, data: ${r.byteLength}}`),this.tensorManager.upload(t,r)}async downloadTensor(t,r){return this.tensorManager.download(t,r)}createMLTensorDownloader(t,r){return async()=>{let n=await this.tensorManager.download(t);return Rr(n,r)}}registerMLTensor(t,r,n){let o=oc.get(r);if(!o)throw new Error(`Unsupported ONNX data type: ${r}`);let i=this.tensorManager.registerTensor(this.currentContext,t,o,n);return se("verbose",()=>`[WebNN] registerMLTensor {tensor: ${t}, dataType: ${o}, dimensions: ${n}} -> {tensorId: ${i}}`),i}registerMLConstant(t,r,n,o,i,a){if(!a)throw new Error("External mounted files are not available.");let d=t;t.startsWith("./")&&(d=t.substring(2));let l=a.get(d);if(!l)throw new Error(`File with name ${d} not found in preloaded files.`);if(r+n>l.byteLength)throw new Error("Out of bounds: data offset and length exceed the external file data size.");let p=l.slice(r,r+n).buffer,m;switch(i.dataType){case"float32":m=new Float32Array(p);break;case"float16":m=new Uint16Array(p);break;case"int32":m=new Int32Array(p);break;case"uint32":m=new Uint32Array(p);break;case"int64":m=new BigInt64Array(p);break;case"uint64":m=new BigUint64Array(p);break;case"int8":m=new Int8Array(p);break;case"int4":case"uint4":case"uint8":m=new Uint8Array(p);break;default:throw new Error(`Unsupported data type: ${i.dataType} in creating WebNN Constant from external data.`)}return se("verbose",()=>`[WebNN] registerMLConstant {dataType: ${i.dataType}, shape: ${i.shape}}}`),o.constant(i,m)}flush(){}}});var ac={};Ht(ac,{init:()=>mg});var tr,ko,mg,sc=U(()=>{"use strict";J();Jl();et();ne();ic();tr=class e{constructor(t,r,n,o){this.module=t;this.dataType=r;this.data=n;this.dims=o}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=k.size(this.dims);return t===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,t)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let t=k.size(this.dims);return t===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,t)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let t=k.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}getUint16Array(){if(this.dataType!==10&&this.dataType!==4)throw new Error("Invalid data type");let t=k.size(this.dims);return t===0?new Uint16Array:new Uint16Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(k.size(t)!==k.size(this.dims))throw new Error("Invalid new shape");return new e(this.module,this.dataType,this.data,t)}},ko=class{constructor(t,r,n){this.module=t;this.backend=r;this.customDataOffset=0;this.customDataSize=0;this.adapterInfo=r.adapterInfo,this.deviceInfo=r.deviceInfo;let o=t.PTR_SIZE,i=n/t.PTR_SIZE,a=o===4?"i32":"i64";this.opKernelContext=Number(t.getValue(o*i++,a));let d=Number(t.getValue(o*i++,a));this.outputCount=Number(t.getValue(o*i++,a)),this.customDataOffset=Number(t.getValue(o*i++,"*")),this.customDataSize=Number(t.getValue(o*i++,a));let l=[];for(let p=0;p<d;p++){let m=Number(t.getValue(o*i++,a)),u=Number(t.getValue(o*i++,"*")),h=Number(t.getValue(o*i++,a)),_=[];for(let y=0;y<h;y++)_.push(Number(t.getValue(o*i++,a)));l.push(new tr(t,m,u,_))}this.inputs=l}get kernelCustomData(){return this.backend.currentKernelCustomData}get customDataBuffer(){return this.module.HEAPU8.subarray(this.customDataOffset,this.customDataOffset+this.customDataSize)}compute(t,r){let n=r?.inputs?.map(d=>typeof d=="number"?this.inputs[d]:d)??this.inputs,o=r?.outputs??[],i=(d,l,p)=>new tr(this.module,l,this.output(d,p),p),a=(d,l)=>{let p=At(d,l);if(!p)throw new Error(`Unsupported data type: ${d}`);let m=p>0?this.backend.gpuDataManager.create(p).id:0;return new tr(this.module,d,m,l)};return this.backend.run(t,n,o,i,a,this.outputCount)}output(t,r){let n=this.module.stackSave();try{let o=this.module.PTR_SIZE,i=o===4?"i32":"i64",a=this.module.stackAlloc((1+r.length)*o);this.module.setValue(a,r.length,i);for(let d=0;d<r.length;d++)this.module.setValue(a+o*(d+1),r[d],i);return this.module._JsepOutput(this.opKernelContext,t,a)}catch(o){throw new Error(`Failed to generate kernel's output[${t}] with dims [${r}]. If you are running with pre-allocated output, please make sure the output type/dims are correct. Error: ${o}`)}finally{this.module.stackRestore(n)}}},mg=async(e,t,r,n)=>{let o=t.jsepInit;if(!o)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let i=new nn;await i.initialize(r,n),o("webgpu",[i,a=>i.alloc(Number(a)),a=>i.free(a),(a,d,l,p=!1)=>{if(p)se("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${Number(a)}, dst=${Number(d)}, size=${Number(l)}`),i.memcpy(Number(a),Number(d));else{se("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${Number(a)}, gpuDataId=${Number(d)}, size=${Number(l)}`);let m=t.HEAPU8.subarray(Number(a>>>0),Number(a>>>0)+Number(l));i.upload(Number(d),m)}},async(a,d,l)=>{se("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${a}, dataOffset=${d}, size=${l}`),await i.download(Number(a),()=>t.HEAPU8.subarray(Number(d)>>>0,Number(d+l)>>>0))},(a,d,l)=>i.createKernel(a,Number(d),l,t.UTF8ToString(t._JsepGetNodeName(Number(d)))),a=>i.releaseKernel(a),(a,d,l,p)=>{se("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${l}, kernel=${a}, contextDataOffset=${d}`);let m=new ko(t,i,Number(d));return i.computeKernel(Number(a),m,p)},()=>i.captureBegin(),()=>i.captureEnd(),()=>i.replay()])}else{let i=new sn(r);o("webnn",[i,()=>i.reserveTensorId(),a=>i.releaseTensorId(a),async(a,d,l,p)=>i.ensureTensor(a,d,l,p),(a,d)=>{i.uploadTensor(a,d)},async(a,d)=>i.downloadTensor(a,d)])}}});var fg,Sr,Tr,zt,hg,qt,Ir,Cr,uc,Ar,kr,Er,Fn=U(()=>{"use strict";Ga();Fa();J();gt();zr();Qn();fg=(e,t)=>{Te()._OrtInit(e,t)!==0&&ce("Can't initialize onnxruntime.")},Sr=async e=>{fg(e.wasm.numThreads,Yt(e.logLevel))},Tr=async(e,t)=>{{let r=(sc(),gr(ac)).init;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let n=e.webgpu.adapter;if(n){if(typeof n.limits!="object"||typeof n.features!="object"||typeof n.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let o=e.webgpu.powerPreference;if(o!==void 0&&o!=="low-power"&&o!=="high-performance")throw new Error(`Invalid powerPreference setting: "${o}"`);let i=e.webgpu.forceFallbackAdapter;if(i!==void 0&&typeof i!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${i}"`);if(n=await navigator.gpu.requestAdapter({powerPreference:o,forceFallbackAdapter:i}),!n)throw new Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.')}await r("webgpu",Te(),e,n)}if(t==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",Te(),e)}}},zt=new Map,hg=e=>{let t=Te(),r=t.stackSave();try{let n=t.PTR_SIZE,o=t.stackAlloc(2*n);t._OrtGetInputOutputCount(e,o,o+n)!==0&&ce("Can't get session input/output count.");let a=n===4?"i32":"i64";return[Number(t.getValue(o,a)),Number(t.getValue(o+n,a))]}finally{t.stackRestore(r)}},qt=e=>{let t=Te(),r=t._malloc(e.byteLength);if(r===0)throw new Error(`Can't create a session. failed to allocate a buffer of size ${e.byteLength}.`);return t.HEAPU8.set(e,r),[r,e.byteLength]},Ir=async(e,t)=>{let r,n,o=Te();Array.isArray(e)?[r,n]=e:e.buffer===o.HEAPU8.buffer?[r,n]=[e.byteOffset,e.byteLength]:[r,n]=qt(e);let i=0,a=0,d=0,l=[],p=[],m=[];try{if([a,l]=Ha(t),t?.externalData&&o.mountExternalData){let v=[];for(let S of t.externalData){let T=typeof S=="string"?S:S.path;v.push(Zt(typeof S=="string"?S:S.data).then(A=>{o.mountExternalData(T,A)}))}await Promise.all(v)}for(let v of t?.executionProviders??[])if((typeof v=="string"?v:v.name)==="webnn"){if(o.shouldTransferToMLTensor=!1,typeof v!="string"){let T=v,A=T?.context,C=T?.gpuDevice,P=T?.deviceType,D=T?.powerPreference;A?o.currentContext=A:C?o.currentContext=await o.jsepCreateMLContext(C):o.currentContext=await o.jsepCreateMLContext({deviceType:P,powerPreference:D})}else o.currentContext=await o.jsepCreateMLContext();break}i=await o._OrtCreateSession(r,n,a),i===0&&ce("Can't create a session."),o.jsepOnCreateSession?.(),o.currentContext&&(o.jsepRegisterMLContext(i,o.currentContext),o.currentContext=void 0,o.shouldTransferToMLTensor=!0);let[u,h]=hg(i),_=!!t?.enableGraphCapture,y=[],g=[],x=[];for(let v=0;v<u;v++){let S=o._OrtGetInputName(i,v);S===0&&ce("Can't get an input name."),p.push(S),y.push(o.UTF8ToString(S))}for(let v=0;v<h;v++){let S=o._OrtGetOutputName(i,v);S===0&&ce("Can't get an output name."),m.push(S);let T=o.UTF8ToString(S);g.push(T);{if(_&&t?.preferredOutputLocation===void 0){x.push("gpu-buffer");continue}let A=typeof t?.preferredOutputLocation=="string"?t.preferredOutputLocation:t?.preferredOutputLocation?.[T]??"cpu";if(A!=="cpu"&&A!=="cpu-pinned"&&A!=="gpu-buffer"&&A!=="ml-tensor")throw new Error(`Not supported preferred output location: ${A}.`);if(_&&A!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${A}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);x.push(A)}}let $=null;return x.some(v=>v==="gpu-buffer"||v==="ml-tensor")&&(d=o._OrtCreateBinding(i),d===0&&ce("Can't create IO binding."),$={handle:d,outputPreferredLocations:x,outputPreferredLocationsEncoded:x.map(v=>Zn(v))}),zt.set(i,[i,p,m,$,_,!1]),[i,y,g]}catch(u){throw p.forEach(h=>o._OrtFree(h)),m.forEach(h=>o._OrtFree(h)),d!==0&&o._OrtReleaseBinding(d)!==0&&ce("Can't release IO binding."),i!==0&&o._OrtReleaseSession(i)!==0&&ce("Can't release session."),u}finally{o._free(r),a!==0&&o._OrtReleaseSessionOptions(a)!==0&&ce("Can't release session options."),l.forEach(u=>o._free(u)),o.unmountExternalData?.()}},Cr=e=>{let t=Te(),r=zt.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[n,o,i,a,d]=r;a&&(d&&t._OrtClearBoundOutputs(a.handle)!==0&&ce("Can't clear bound outputs."),t._OrtReleaseBinding(a.handle)!==0&&ce("Can't release IO binding.")),t.jsepOnReleaseSession?.(e),o.forEach(l=>t._OrtFree(l)),i.forEach(l=>t._OrtFree(l)),t._OrtReleaseSession(n)!==0&&ce("Can't release session."),zt.delete(e)},uc=(e,t,r,n,o,i=!1)=>{if(!e){t.push(0);return}let a=Te(),d=a.PTR_SIZE,l=e[0],p=e[1],m=e[3],u,h;if(l==="string"&&(m==="gpu-buffer"||m==="ml-tensor"))throw new Error("String tensor is not supported on GPU.");if(i&&m!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${o} when enableGraphCapture is true.`);if(m==="gpu-buffer"){let g=e[2].gpuBuffer;h=At(jt(l),p);let x=a.jsepRegisterBuffer;if(!x)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');u=x(n,o,g,h)}else if(m==="ml-tensor"){let g=e[2].mlTensor;h=At(jt(l),p);let x=a.jsepRegisterMLTensor;if(!x)throw new Error('Tensor location "ml-tensor" is not supported without using WebNN.');u=x(g,jt(l),p)}else{let g=e[2];if(Array.isArray(g)){h=d*g.length,u=a._malloc(h),r.push(u);for(let x=0;x<g.length;x++){if(typeof g[x]!="string")throw new TypeError(`tensor data at index ${x} is not a string`);a.setValue(u+x*d,ke(g[x],r),"*")}}else h=g.byteLength,u=a._malloc(h),r.push(u),a.HEAPU8.set(new Uint8Array(g.buffer,g.byteOffset,h),u)}let _=a.stackSave(),y=a.stackAlloc(4*p.length);try{p.forEach((x,$)=>a.setValue(y+$*d,x,d===4?"i32":"i64"));let g=a._OrtCreateTensor(jt(l),u,h,y,p.length,Zn(m));g===0&&ce(`Can't create tensor for input/output. session=${n}, index=${o}.`),t.push(g)}finally{a.stackRestore(_)}},Ar=async(e,t,r,n,o,i)=>{let a=Te(),d=a.PTR_SIZE,l=zt.get(e);if(!l)throw new Error(`cannot run inference. invalid session id: ${e}`);let p=l[0],m=l[1],u=l[2],h=l[3],_=l[4],y=l[5],g=t.length,x=n.length,$=0,v=[],S=[],T=[],A=[],C=a.stackSave(),P=a.stackAlloc(g*d),D=a.stackAlloc(g*d),R=a.stackAlloc(x*d),H=a.stackAlloc(x*d);try{a.jsepOnRunStart?.(p),[$,v]=La(i);for(let V=0;V<g;V++)uc(r[V],S,A,e,t[V],_);for(let V=0;V<x;V++)uc(o[V],T,A,e,g+n[V],_);for(let V=0;V<g;V++)a.setValue(P+V*d,S[V],"*"),a.setValue(D+V*d,m[t[V]],"*");for(let V=0;V<x;V++)a.setValue(R+V*d,T[V],"*"),a.setValue(H+V*d,u[n[V]],"*");if(h&&!y){let{handle:V,outputPreferredLocations:K,outputPreferredLocationsEncoded:we}=h;if(m.length!==g)throw new Error(`input count from feeds (${g}) is expected to be always equal to model's input count (${m.length}).`);for(let j=0;j<g;j++){let Q=t[j];await a._OrtBindInput(V,m[Q],S[j])!==0&&ce(`Can't bind input[${j}] for session=${e}.`)}for(let j=0;j<x;j++){let Q=n[j];o[j]?.[3]?a._OrtBindOutput(V,u[Q],T[j],0)!==0&&ce(`Can't bind pre-allocated output[${j}] for session=${e}.`):a._OrtBindOutput(V,u[Q],0,we[Q])!==0&&ce(`Can't bind output[${j}] to ${K[j]} for session=${e}.`)}zt.set(e,[p,m,u,h,_,!0])}let L;h?L=await a._OrtRunWithBinding(p,h.handle,x,R,$):L=await a._OrtRun(p,D,P,g,H,x,R,$),L!==0&&ce("failed to call OrtRun().");let re=[];for(let V=0;V<x;V++){let K=Number(a.getValue(R+V*d,"*"));if(K===T[V]){re.push(o[V]);continue}let we=a.stackSave(),j=a.stackAlloc(4*d),Q=!1,ie,te=0;try{a._OrtGetTensorData(K,j,j+d,j+2*d,j+3*d)!==0&&ce(`Can't access output tensor data on index ${V}.`);let Oe=d===4?"i32":"i64",ve=Number(a.getValue(j,Oe));te=a.getValue(j+d,"*");let de=a.getValue(j+d*2,"*"),W=Number(a.getValue(j+d*3,Oe)),Y=[];for(let pe=0;pe<W;pe++)Y.push(Number(a.getValue(de+pe*d,Oe)));a._OrtFree(de)!==0&&ce("Can't free memory for tensor dims.");let he=Y.reduce((pe,Ie)=>pe*Ie,1);ie=bt(ve);let De=h?.outputPreferredLocations[n[V]];if(ie==="string"){if(De==="gpu-buffer"||De==="ml-tensor")throw new Error("String tensor is not supported on GPU.");let pe=[];for(let Ie=0;Ie<he;Ie++){let He=a.getValue(te+Ie*d,"*"),mn=a.getValue(te+(Ie+1)*d,"*"),wt=Ie===he-1?void 0:mn-He;pe.push(a.UTF8ToString(He,wt))}re.push([ie,Y,pe,"cpu"])}else if(De==="gpu-buffer"&&he>0){let pe=a.jsepGetBuffer;if(!pe)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let Ie=pe(te),He=At(ve,he);if(He===void 0||!Dr(ie))throw new Error(`Unsupported data type: ${ie}`);Q=!0,re.push([ie,Y,{gpuBuffer:Ie,download:a.jsepCreateDownloader(Ie,He,ie),dispose:()=>{a._OrtReleaseTensor(K)!==0&&ce("Can't release tensor.")}},"gpu-buffer"])}else if(De==="ml-tensor"&&he>0){let pe=a.jsepEnsureTensor;if(!pe)throw new Error('preferredLocation "ml-tensor" is not supported without using WebNN.');if(At(ve,he)===void 0||!Br(ie))throw new Error(`Unsupported data type: ${ie}`);let He=await pe(te,ve,Y,!1);Q=!0,re.push([ie,Y,{mlTensor:He,download:a.jsepCreateMLTensorDownloader(te,ie),dispose:()=>{a.jsepReleaseTensorId(te),a._OrtReleaseTensor(K)}},"ml-tensor"])}else{let pe=Or(ie),Ie=new pe(he);new Uint8Array(Ie.buffer,Ie.byteOffset,Ie.byteLength).set(a.HEAPU8.subarray(te,te+Ie.byteLength)),re.push([ie,Y,Ie,"cpu"])}}finally{a.stackRestore(we),ie==="string"&&te&&a._free(te),Q||a._OrtReleaseTensor(K)}}return h&&!_&&(a._OrtClearBoundOutputs(h.handle)!==0&&ce("Can't clear bound outputs."),zt.set(e,[p,m,u,h,_,!1])),re}finally{a.stackRestore(C),S.forEach(L=>a._OrtReleaseTensor(L)),T.forEach(L=>a._OrtReleaseTensor(L)),A.forEach(L=>a._free(L)),$!==0&&a._OrtReleaseRunOptions($),v.forEach(L=>a._free(L))}},kr=e=>{let t=Te(),r=zt.get(e);if(!r)throw new Error("invalid session id");let n=r[0],o=t._OrtEndProfiling(n);o===0&&ce("Can't get an profile file name."),t._OrtFree(o)},Er=e=>{let t=[];for(let r of e){let n=r[2];!Array.isArray(n)&&"buffer"in n&&t.push(n.buffer)}return t}});var Ot,Ge,rr,dn,ln,un,Eo,Po,Wt,Lt,bg,dc,lc,cc,pc,mc,fc,hc,zo=U(()=>{"use strict";Le();Fn();gt();$r();Ot=()=>!!_e.wasm.proxy&&typeof document<"u",rr=!1,dn=!1,ln=!1,Po=new Map,Wt=(e,t)=>{let r=Po.get(e);r?r.push(t):Po.set(e,[t])},Lt=()=>{if(rr||!dn||ln||!Ge)throw new Error("worker not ready")},bg=e=>{switch(e.data.type){case"init-wasm":rr=!1,e.data.err?(ln=!0,Eo[1](e.data.err)):(dn=!0,Eo[0]()),un&&(URL.revokeObjectURL(un),un=void 0);break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let t=Po.get(e.data.type);e.data.err?t.shift()[1](e.data.err):t.shift()[0](e.data.out);break}default:}},dc=async()=>{if(!dn){if(rr)throw new Error("multiple calls to 'initWasm()' detected.");if(ln)throw new Error("previous call to 'initWasm()' failed.");if(rr=!0,Ot())return new Promise((e,t)=>{Ge?.terminate(),Na().then(([r,n])=>{try{Ge=n,Ge.onerror=i=>t(i),Ge.onmessage=bg,Eo=[e,t];let o={type:"init-wasm",in:_e};!o.in.wasm.wasmPaths&&(r||import.meta.url?.startsWith("file:"))&&(o.in.wasm.wasmPaths={wasm:new URL(/* asset import */ __webpack_require__(/*! ort-wasm-simd-threaded.jsep.wasm */ "./node_modules/onnxruntime-web/dist/ort-wasm-simd-threaded.jsep.wasm"), __webpack_require__.b).href}),Ge.postMessage(o),un=r}catch(o){t(o)}},t)});try{await xr(_e.wasm),await Sr(_e),dn=!0}catch(e){throw ln=!0,e}finally{rr=!1}}},lc=async e=>{if(Ot())return Lt(),new Promise((t,r)=>{Wt("init-ep",[t,r]);let n={type:"init-ep",in:{epName:e,env:_e}};Ge.postMessage(n)});await Tr(_e,e)},cc=async e=>Ot()?(Lt(),new Promise((t,r)=>{Wt("copy-from",[t,r]);let n={type:"copy-from",in:{buffer:e}};Ge.postMessage(n,[e.buffer])})):qt(e),pc=async(e,t)=>{if(Ot()){if(t?.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return Lt(),new Promise((r,n)=>{Wt("create",[r,n]);let o={type:"create",in:{model:e,options:{...t}}},i=[];e instanceof Uint8Array&&i.push(e.buffer),Ge.postMessage(o,i)})}else return Ir(e,t)},mc=async e=>{if(Ot())return Lt(),new Promise((t,r)=>{Wt("release",[t,r]);let n={type:"release",in:e};Ge.postMessage(n)});Cr(e)},fc=async(e,t,r,n,o,i)=>{if(Ot()){if(r.some(a=>a[3]!=="cpu"))throw new Error("input tensor on GPU is not supported for proxy.");if(o.some(a=>a))throw new Error("pre-allocated output tensor is not supported for proxy.");return Lt(),new Promise((a,d)=>{Wt("run",[a,d]);let l=r,p={type:"run",in:{sessionId:e,inputIndices:t,inputs:l,outputIndices:n,options:i}};Ge.postMessage(p,Er(l))})}else return Ar(e,t,r,n,o,i)},hc=async e=>{if(Ot())return Lt(),new Promise((t,r)=>{Wt("end-profiling",[t,r]);let n={type:"end-profiling",in:e};Ge.postMessage(n)});kr(e)}});var gc,yg,cn,bc=U(()=>{"use strict";Le();zo();J();vr();Qn();gc=(e,t)=>{switch(e.location){case"cpu":return[e.type,e.dims,e.data,"cpu"];case"gpu-buffer":return[e.type,e.dims,{gpuBuffer:e.gpuBuffer},"gpu-buffer"];case"ml-tensor":return[e.type,e.dims,{mlTensor:e.mlTensor},"ml-tensor"];default:throw new Error(`invalid data location: ${e.location} for ${t()}`)}},yg=e=>{switch(e[3]){case"cpu":return new Fe(e[0],e[2],e[1]);case"gpu-buffer":{let t=e[0];if(!Dr(t))throw new Error(`not supported data type: ${t} for deserializing GPU tensor`);let{gpuBuffer:r,download:n,dispose:o}=e[2];return Fe.fromGpuBuffer(r,{dataType:t,dims:e[1],download:n,dispose:o})}case"ml-tensor":{let t=e[0];if(!Br(t))throw new Error(`not supported data type: ${t} for deserializing MLTensor tensor`);let{mlTensor:r,download:n,dispose:o}=e[2];return Fe.fromMLTensor(r,{dataType:t,dims:e[1],download:n,dispose:o})}default:throw new Error(`invalid data location: ${e[3]}`)}},cn=class{async fetchModelAndCopyToWasmMemory(t){return cc(await Zt(t))}async loadModel(t,r){Ne();let n;typeof t=="string"? false?0:n=await this.fetchModelAndCopyToWasmMemory(t):n=t,[this.sessionId,this.inputNames,this.outputNames]=await pc(n,r),Be()}async dispose(){return mc(this.sessionId)}async run(t,r,n){Ne();let o=[],i=[];Object.entries(t).forEach(h=>{let _=h[0],y=h[1],g=this.inputNames.indexOf(_);if(g===-1)throw new Error(`invalid input '${_}'`);o.push(y),i.push(g)});let a=[],d=[];Object.entries(r).forEach(h=>{let _=h[0],y=h[1],g=this.outputNames.indexOf(_);if(g===-1)throw new Error(`invalid output '${_}'`);a.push(y),d.push(g)});let l=o.map((h,_)=>gc(h,()=>`input "${this.inputNames[i[_]]}"`)),p=a.map((h,_)=>h?gc(h,()=>`output "${this.outputNames[d[_]]}"`):null),m=await fc(this.sessionId,i,l,d,p,n),u={};for(let h=0;h<m.length;h++)u[this.outputNames[d[h]]]=a[h]??yg(m[h]);return Be(),u}startProfiling(){}endProfiling(){hc(this.sessionId)}}});var _c={};Ht(_c,{OnnxruntimeWebAssemblyBackend:()=>pn,initializeFlags:()=>yc,wasmBackend:()=>_g});var yc,pn,_g,wc=U(()=>{"use strict";Le();zo();bc();yc=()=>{if((typeof _e.wasm.initTimeout!="number"||_e.wasm.initTimeout<0)&&(_e.wasm.initTimeout=0),_e.wasm.simd===!1&&console.warn('Deprecated property "env.wasm.simd" is set to false. non-SIMD build is no longer provided, and this setting will be ignored.'),typeof _e.wasm.proxy!="boolean"&&(_e.wasm.proxy=!1),typeof _e.wasm.trace!="boolean"&&(_e.wasm.trace=!1),typeof _e.wasm.numThreads!="number"||!Number.isInteger(_e.wasm.numThreads)||_e.wasm.numThreads<=0)if(typeof self<"u"&&!self.crossOriginIsolated)_e.wasm.numThreads=1;else{let e=typeof navigator>"u"?Un("node:os").cpus().length:navigator.hardwareConcurrency;_e.wasm.numThreads=Math.min(4,Math.ceil((e||1)/2))}},pn=class{async init(t){yc(),await dc(),await lc(t)}async createInferenceSessionHandler(t,r){let n=new cn;return await n.loadModel(t,r),Promise.resolve(n)}},_g=new pn});Le();Le();Le();var Ia="1.21.0-dev.20250206-d981b153d3";var $1=Hn;{let e=(wc(),gr(_c)).wasmBackend;Tt("webgpu",e,5),Tt("webnn",e,5),Tt("cpu",e,10),Tt("wasm",e,10)}Object.defineProperty(_e.versions,"web",{value:Ia,enumerable:!0});
 /**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
@@ -8440,6 +8513,8 @@ function getNormalizedConfig(config) {
             break;
         case 'gemma':
         case 'gemma2':
+        case 'glm':
+        case 'helium':
             mapping['num_heads'] = 'num_key_value_heads';
             mapping['num_layers'] = 'num_hidden_layers';
             mapping['dim_kv'] = 'head_dim';
@@ -8512,12 +8587,17 @@ function getNormalizedConfig(config) {
             mapping['encoder_hidden_size'] = mapping['decoder_hidden_size'] = 'd_model';
             break;
         case 'musicgen_decoder':
-        case 'moonshine':
             mapping['num_encoder_layers'] = mapping['num_decoder_layers'] = 'num_hidden_layers';
             mapping['num_encoder_heads'] = mapping['num_decoder_heads'] = 'num_attention_heads';
             mapping['encoder_hidden_size'] = mapping['decoder_hidden_size'] = 'hidden_size';
             break;
-
+        case 'moonshine':
+            mapping['num_decoder_layers'] = 'decoder_num_hidden_layers';
+            mapping['num_decoder_heads'] = 'decoder_num_key_value_heads';
+            mapping['num_encoder_layers'] = 'encoder_num_hidden_layers';
+            mapping['num_encoder_heads'] = 'encoder_num_key_value_heads';
+            mapping['encoder_hidden_size'] = mapping['decoder_hidden_size'] = 'hidden_size';
+            break;
         case 'vision-encoder-decoder':
             // @ts-expect-error TS2339
             const decoderConfig = getNormalizedConfig(config.decoder);
@@ -8762,7 +8842,7 @@ __webpack_require__.r(__webpack_exports__);
 
 
 
-const VERSION = '3.3.1';
+const VERSION = '3.3.3';
 
 // Check if various APIs are available (depends on environment)
 const IS_BROWSER_ENV = typeof window !== "undefined" && typeof window.document !== "undefined";
@@ -10894,6 +10974,9 @@ __webpack_require__.r(__webpack_exports__);
 /* harmony export */   GemmaForCausalLM: () => (/* binding */ GemmaForCausalLM),
 /* harmony export */   GemmaModel: () => (/* binding */ GemmaModel),
 /* harmony export */   GemmaPreTrainedModel: () => (/* binding */ GemmaPreTrainedModel),
+/* harmony export */   GlmForCausalLM: () => (/* binding */ GlmForCausalLM),
+/* harmony export */   GlmModel: () => (/* binding */ GlmModel),
+/* harmony export */   GlmPreTrainedModel: () => (/* binding */ GlmPreTrainedModel),
 /* harmony export */   GraniteForCausalLM: () => (/* binding */ GraniteForCausalLM),
 /* harmony export */   GraniteModel: () => (/* binding */ GraniteModel),
 /* harmony export */   GranitePreTrainedModel: () => (/* binding */ GranitePreTrainedModel),
@@ -10901,6 +10984,9 @@ __webpack_require__.r(__webpack_exports__);
 /* harmony export */   GroundingDinoPreTrainedModel: () => (/* binding */ GroundingDinoPreTrainedModel),
 /* harmony export */   GroupViTModel: () => (/* binding */ GroupViTModel),
 /* harmony export */   GroupViTPreTrainedModel: () => (/* binding */ GroupViTPreTrainedModel),
+/* harmony export */   HeliumForCausalLM: () => (/* binding */ HeliumForCausalLM),
+/* harmony export */   HeliumModel: () => (/* binding */ HeliumModel),
+/* harmony export */   HeliumPreTrainedModel: () => (/* binding */ HeliumPreTrainedModel),
 /* harmony export */   HieraForImageClassification: () => (/* binding */ HieraForImageClassification),
 /* harmony export */   HieraModel: () => (/* binding */ HieraModel),
 /* harmony export */   HieraPreTrainedModel: () => (/* binding */ HieraPreTrainedModel),
@@ -15439,6 +15525,19 @@ class LlamaModel extends LlamaPreTrainedModel { }
 class LlamaForCausalLM extends LlamaPreTrainedModel { }
 //////////////////////////////////////////////////
 
+//////////////////////////////////////////////////
+// Helium models
+class HeliumPreTrainedModel extends PreTrainedModel { }
+class HeliumModel extends HeliumPreTrainedModel { }
+class HeliumForCausalLM extends HeliumPreTrainedModel { }
+//////////////////////////////////////////////////
+
+//////////////////////////////////////////////////
+// Glm models
+class GlmPreTrainedModel extends PreTrainedModel { }
+class GlmModel extends GlmPreTrainedModel { }
+class GlmForCausalLM extends GlmPreTrainedModel { }
+//////////////////////////////////////////////////
 
 //////////////////////////////////////////////////
 // EXAONE models
@@ -18293,6 +18392,8 @@ const MODEL_MAPPING_NAMES_DECODER_ONLY = new Map([
     ['cohere', ['CohereModel', CohereModel]],
     ['gemma', ['GemmaModel', GemmaModel]],
     ['gemma2', ['Gemma2Model', Gemma2Model]],
+    ['helium', ['HeliumModel', HeliumModel]],
+    ['glm', ['GlmModel', GlmModel]],
     ['openelm', ['OpenELMModel', OpenELMModel]],
     ['qwen2', ['Qwen2Model', Qwen2Model]],
     ['phi', ['PhiModel', PhiModel]],
@@ -18389,6 +18490,8 @@ const MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = new Map([
     ['cohere', ['CohereForCausalLM', CohereForCausalLM]],
     ['gemma', ['GemmaForCausalLM', GemmaForCausalLM]],
     ['gemma2', ['Gemma2ForCausalLM', Gemma2ForCausalLM]],
+    ['helium', ['HeliumForCausalLM', HeliumForCausalLM]],
+    ['glm', ['GlmForCausalLM', GlmForCausalLM]],
     ['openelm', ['OpenELMForCausalLM', OpenELMForCausalLM]],
     ['qwen2', ['Qwen2ForCausalLM', Qwen2ForCausalLM]],
     ['phi', ['PhiForCausalLM', PhiForCausalLM]],
@@ -21932,8 +22035,9 @@ __webpack_require__.r(__webpack_exports__);
 /* harmony export */   SamProcessor: () => (/* reexport safe */ _sam_processing_sam_js__WEBPACK_IMPORTED_MODULE_12__.SamProcessor),
 /* harmony export */   SpeechT5Processor: () => (/* reexport safe */ _speecht5_processing_speecht5_js__WEBPACK_IMPORTED_MODULE_13__.SpeechT5Processor),
 /* harmony export */   VLChatProcessor: () => (/* reexport safe */ _janus_processing_janus_js__WEBPACK_IMPORTED_MODULE_3__.VLChatProcessor),
-/* harmony export */   Wav2Vec2ProcessorWithLM: () => (/* reexport safe */ _wav2vec2_processing_wav2vec2_js__WEBPACK_IMPORTED_MODULE_14__.Wav2Vec2ProcessorWithLM),
-/* harmony export */   WhisperProcessor: () => (/* reexport safe */ _whisper_processing_whisper_js__WEBPACK_IMPORTED_MODULE_15__.WhisperProcessor)
+/* harmony export */   Wav2Vec2Processor: () => (/* reexport safe */ _wav2vec2_processing_wav2vec2_js__WEBPACK_IMPORTED_MODULE_14__.Wav2Vec2Processor),
+/* harmony export */   Wav2Vec2ProcessorWithLM: () => (/* reexport safe */ _wav2vec2_with_lm_processing_wav2vec2_with_lm_js__WEBPACK_IMPORTED_MODULE_15__.Wav2Vec2ProcessorWithLM),
+/* harmony export */   WhisperProcessor: () => (/* reexport safe */ _whisper_processing_whisper_js__WEBPACK_IMPORTED_MODULE_16__.WhisperProcessor)
 /* harmony export */ });
 /* harmony import */ var _florence2_processing_florence2_js__WEBPACK_IMPORTED_MODULE_0__ = __webpack_require__(/*! ./florence2/processing_florence2.js */ "./src/models/florence2/processing_florence2.js");
 /* harmony import */ var _grounding_dino_processing_grounding_dino_js__WEBPACK_IMPORTED_MODULE_1__ = __webpack_require__(/*! ./grounding_dino/processing_grounding_dino.js */ "./src/models/grounding_dino/processing_grounding_dino.js");
@@ -21950,7 +22054,9 @@ __webpack_require__.r(__webpack_exports__);
 /* harmony import */ var _sam_processing_sam_js__WEBPACK_IMPORTED_MODULE_12__ = __webpack_require__(/*! ./sam/processing_sam.js */ "./src/models/sam/processing_sam.js");
 /* harmony import */ var _speecht5_processing_speecht5_js__WEBPACK_IMPORTED_MODULE_13__ = __webpack_require__(/*! ./speecht5/processing_speecht5.js */ "./src/models/speecht5/processing_speecht5.js");
 /* harmony import */ var _wav2vec2_processing_wav2vec2_js__WEBPACK_IMPORTED_MODULE_14__ = __webpack_require__(/*! ./wav2vec2/processing_wav2vec2.js */ "./src/models/wav2vec2/processing_wav2vec2.js");
-/* harmony import */ var _whisper_processing_whisper_js__WEBPACK_IMPORTED_MODULE_15__ = __webpack_require__(/*! ./whisper/processing_whisper.js */ "./src/models/whisper/processing_whisper.js");
+/* harmony import */ var _wav2vec2_with_lm_processing_wav2vec2_with_lm_js__WEBPACK_IMPORTED_MODULE_15__ = __webpack_require__(/*! ./wav2vec2_with_lm/processing_wav2vec2_with_lm.js */ "./src/models/wav2vec2_with_lm/processing_wav2vec2_with_lm.js");
+/* harmony import */ var _whisper_processing_whisper_js__WEBPACK_IMPORTED_MODULE_16__ = __webpack_require__(/*! ./whisper/processing_whisper.js */ "./src/models/whisper/processing_whisper.js");
+
 
 
 
@@ -23161,16 +23267,53 @@ class Wav2Vec2FeatureExtractor extends _base_feature_extraction_utils_js__WEBPAC
   \****************************************************/
 /***/ ((__unused_webpack___webpack_module__, __webpack_exports__, __webpack_require__) => {
 
+__webpack_require__.r(__webpack_exports__);
+/* harmony export */ __webpack_require__.d(__webpack_exports__, {
+/* harmony export */   Wav2Vec2Processor: () => (/* binding */ Wav2Vec2Processor)
+/* harmony export */ });
+/* harmony import */ var _tokenizers_js__WEBPACK_IMPORTED_MODULE_0__ = __webpack_require__(/*! ../../tokenizers.js */ "./src/tokenizers.js");
+/* harmony import */ var _auto_feature_extraction_auto_js__WEBPACK_IMPORTED_MODULE_1__ = __webpack_require__(/*! ../auto/feature_extraction_auto.js */ "./src/models/auto/feature_extraction_auto.js");
+/* harmony import */ var _base_processing_utils_js__WEBPACK_IMPORTED_MODULE_2__ = __webpack_require__(/*! ../../base/processing_utils.js */ "./src/base/processing_utils.js");
+
+
+
+
+class Wav2Vec2Processor extends _base_processing_utils_js__WEBPACK_IMPORTED_MODULE_2__.Processor {
+    static tokenizer_class = _tokenizers_js__WEBPACK_IMPORTED_MODULE_0__.AutoTokenizer
+    static feature_extractor_class = _auto_feature_extraction_auto_js__WEBPACK_IMPORTED_MODULE_1__.AutoFeatureExtractor
+
+    /**
+     * Calls the feature_extractor function with the given audio input.
+     * @param {any} audio The audio input to extract features from.
+     * @returns {Promise<any>} A Promise that resolves with the extracted features.
+     */
+    async _call(audio) {
+        return await this.feature_extractor(audio)
+    }
+}
+
+
+/***/ }),
+
+/***/ "./src/models/wav2vec2_with_lm/processing_wav2vec2_with_lm.js":
+/*!********************************************************************!*\
+  !*** ./src/models/wav2vec2_with_lm/processing_wav2vec2_with_lm.js ***!
+  \********************************************************************/
+/***/ ((__unused_webpack___webpack_module__, __webpack_exports__, __webpack_require__) => {
+
 __webpack_require__.r(__webpack_exports__);
 /* harmony export */ __webpack_require__.d(__webpack_exports__, {
 /* harmony export */   Wav2Vec2ProcessorWithLM: () => (/* binding */ Wav2Vec2ProcessorWithLM)
 /* harmony export */ });
-/* harmony import */ var _base_processing_utils_js__WEBPACK_IMPORTED_MODULE_0__ = __webpack_require__(/*! ../../base/processing_utils.js */ "./src/base/processing_utils.js");
+/* harmony import */ var _tokenizers_js__WEBPACK_IMPORTED_MODULE_0__ = __webpack_require__(/*! ../../tokenizers.js */ "./src/tokenizers.js");
 /* harmony import */ var _auto_feature_extraction_auto_js__WEBPACK_IMPORTED_MODULE_1__ = __webpack_require__(/*! ../auto/feature_extraction_auto.js */ "./src/models/auto/feature_extraction_auto.js");
+/* harmony import */ var _base_processing_utils_js__WEBPACK_IMPORTED_MODULE_2__ = __webpack_require__(/*! ../../base/processing_utils.js */ "./src/base/processing_utils.js");
 
 
 
-class Wav2Vec2ProcessorWithLM extends _base_processing_utils_js__WEBPACK_IMPORTED_MODULE_0__.Processor {
+
+class Wav2Vec2ProcessorWithLM extends _base_processing_utils_js__WEBPACK_IMPORTED_MODULE_2__.Processor {
+    static tokenizer_class = _tokenizers_js__WEBPACK_IMPORTED_MODULE_0__.AutoTokenizer
     static feature_extractor_class = _auto_feature_extraction_auto_js__WEBPACK_IMPORTED_MODULE_1__.AutoFeatureExtractor
 
     /**
@@ -23751,9 +23894,12 @@ __webpack_require__.r(__webpack_exports__);
 /* harmony export */ });
 /* harmony import */ var _backends_onnx_js__WEBPACK_IMPORTED_MODULE_0__ = __webpack_require__(/*! ../backends/onnx.js */ "./src/backends/onnx.js");
 /* harmony import */ var _utils_tensor_js__WEBPACK_IMPORTED_MODULE_1__ = __webpack_require__(/*! ../utils/tensor.js */ "./src/utils/tensor.js");
+/* harmony import */ var _env_js__WEBPACK_IMPORTED_MODULE_2__ = __webpack_require__(/*! ../env.js */ "./src/env.js");
 
 
 
+
+const IS_WEB_ENV = _env_js__WEBPACK_IMPORTED_MODULE_2__.apis.IS_BROWSER_ENV || _env_js__WEBPACK_IMPORTED_MODULE_2__.apis.IS_WEBWORKER_ENV;
 /**
  * Asynchronously creates a wrapper function for running an ONNX inference session.
  *
@@ -23769,10 +23915,16 @@ const wrap = async (session_bytes, session_options, names) => {
     const session = await (0,_backends_onnx_js__WEBPACK_IMPORTED_MODULE_0__.createInferenceSession)(
         new Uint8Array(session_bytes), session_options,
     );
+
+    /** @type {Promise<any>} */
+    let chain = Promise.resolve();
+
     return /** @type {any} */(async (/** @type {Record<string, Tensor>} */ inputs) => {
         const proxied = (0,_backends_onnx_js__WEBPACK_IMPORTED_MODULE_0__.isONNXProxy)();
         const ortFeed = Object.fromEntries(Object.entries(inputs).map(([k, v]) => [k, (proxied ? v.clone() : v).ort_tensor]));
-        const outputs = await session.run(ortFeed);
+
+        // When running in-browser via WASM, we need to chain calls to session.run to avoid "Error: Session already started"
+        const outputs = await (chain = IS_WEB_ENV ? chain.then(() => session.run(ortFeed)) : session.run(ortFeed));
 
         if (Array.isArray(names)) {
             return names.map((n) => new _utils_tensor_js__WEBPACK_IMPORTED_MODULE_1__.Tensor(outputs[n]));
@@ -27716,13 +27868,15 @@ class TokenizerModel extends _utils_generic_js__WEBPACK_IMPORTED_MODULE_0__.Call
                 return new BPE(config);
 
             default:
-                // Some tokenizers, like for google-t5/t5-small, do not have a `type` field.
-                // In this case, we can infer the tokenizer type based on the structure of the `vocab` field.
+                // Some older tokenizers, like `google-t5/t5-small` and `distilbert/distilbert-base-uncased`, do not have a `type` field.
+                // In this case, we can infer the tokenizer type based on the structure of the `vocab` field and other properties.
                 if (config.vocab) {
                     if (Array.isArray(config.vocab)) {
                         // config.vocab is of type `[string, number][]`
                         // @ts-ignore
                         return new Unigram(config, ...args);
+                    } else if (typeof config.vocab === 'object' && config.continuing_subword_prefix && config.unk_token) {
+                        return new WordPieceTokenizer(config);
                     } else {
                         // @ts-ignore
                         return new LegacyTokenizerModel(config, ...args);
@@ -37798,7 +37952,7 @@ function quantize_embeddings(tensor, precision) {
 /******/ 
 /******/ /* webpack/runtime/import chunk loading */
 /******/ (() => {
-/******/ 	__webpack_require__.b = new URL("./", import.meta.url);
+/******/ 	__webpack_require__.b = new URL(import.meta.url);
 /******/ 	
 /******/ 	// object to store loaded and loading chunks
 /******/ 	// undefined = chunk not loaded, null = chunk preloaded/prefetched
@@ -38082,6 +38236,9 @@ __webpack_require__.r(__webpack_exports__);
 /* harmony export */   GemmaModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GemmaModel),
 /* harmony export */   GemmaPreTrainedModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GemmaPreTrainedModel),
 /* harmony export */   GemmaTokenizer: () => (/* reexport safe */ _tokenizers_js__WEBPACK_IMPORTED_MODULE_3__.GemmaTokenizer),
+/* harmony export */   GlmForCausalLM: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GlmForCausalLM),
+/* harmony export */   GlmModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GlmModel),
+/* harmony export */   GlmPreTrainedModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GlmPreTrainedModel),
 /* harmony export */   GraniteForCausalLM: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GraniteForCausalLM),
 /* harmony export */   GraniteModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GraniteModel),
 /* harmony export */   GranitePreTrainedModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GranitePreTrainedModel),
@@ -38092,6 +38249,9 @@ __webpack_require__.r(__webpack_exports__);
 /* harmony export */   GroundingDinoProcessor: () => (/* reexport safe */ _models_processors_js__WEBPACK_IMPORTED_MODULE_16__.GroundingDinoProcessor),
 /* harmony export */   GroupViTModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GroupViTModel),
 /* harmony export */   GroupViTPreTrainedModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.GroupViTPreTrainedModel),
+/* harmony export */   HeliumForCausalLM: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.HeliumForCausalLM),
+/* harmony export */   HeliumModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.HeliumModel),
+/* harmony export */   HeliumPreTrainedModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.HeliumPreTrainedModel),
 /* harmony export */   HerbertTokenizer: () => (/* reexport safe */ _tokenizers_js__WEBPACK_IMPORTED_MODULE_3__.HerbertTokenizer),
 /* harmony export */   HieraForImageClassification: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.HieraForImageClassification),
 /* harmony export */   HieraModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.HieraModel),
@@ -38462,6 +38622,7 @@ __webpack_require__.r(__webpack_exports__);
 /* harmony export */   Wav2Vec2ForSequenceClassification: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.Wav2Vec2ForSequenceClassification),
 /* harmony export */   Wav2Vec2Model: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.Wav2Vec2Model),
 /* harmony export */   Wav2Vec2PreTrainedModel: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.Wav2Vec2PreTrainedModel),
+/* harmony export */   Wav2Vec2Processor: () => (/* reexport safe */ _models_processors_js__WEBPACK_IMPORTED_MODULE_16__.Wav2Vec2Processor),
 /* harmony export */   Wav2Vec2ProcessorWithLM: () => (/* reexport safe */ _models_processors_js__WEBPACK_IMPORTED_MODULE_16__.Wav2Vec2ProcessorWithLM),
 /* harmony export */   WavLMForAudioFrameClassification: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.WavLMForAudioFrameClassification),
 /* harmony export */   WavLMForCTC: () => (/* reexport safe */ _models_js__WEBPACK_IMPORTED_MODULE_2__.WavLMForCTC),
@@ -38869,6 +39030,9 @@ var __webpack_exports__GemmaForCausalLM = __webpack_exports__.GemmaForCausalLM;
 var __webpack_exports__GemmaModel = __webpack_exports__.GemmaModel;
 var __webpack_exports__GemmaPreTrainedModel = __webpack_exports__.GemmaPreTrainedModel;
 var __webpack_exports__GemmaTokenizer = __webpack_exports__.GemmaTokenizer;
+var __webpack_exports__GlmForCausalLM = __webpack_exports__.GlmForCausalLM;
+var __webpack_exports__GlmModel = __webpack_exports__.GlmModel;
+var __webpack_exports__GlmPreTrainedModel = __webpack_exports__.GlmPreTrainedModel;
 var __webpack_exports__GraniteForCausalLM = __webpack_exports__.GraniteForCausalLM;
 var __webpack_exports__GraniteModel = __webpack_exports__.GraniteModel;
 var __webpack_exports__GranitePreTrainedModel = __webpack_exports__.GranitePreTrainedModel;
@@ -38879,6 +39043,9 @@ var __webpack_exports__GroundingDinoPreTrainedModel = __webpack_exports__.Ground
 var __webpack_exports__GroundingDinoProcessor = __webpack_exports__.GroundingDinoProcessor;
 var __webpack_exports__GroupViTModel = __webpack_exports__.GroupViTModel;
 var __webpack_exports__GroupViTPreTrainedModel = __webpack_exports__.GroupViTPreTrainedModel;
+var __webpack_exports__HeliumForCausalLM = __webpack_exports__.HeliumForCausalLM;
+var __webpack_exports__HeliumModel = __webpack_exports__.HeliumModel;
+var __webpack_exports__HeliumPreTrainedModel = __webpack_exports__.HeliumPreTrainedModel;
 var __webpack_exports__HerbertTokenizer = __webpack_exports__.HerbertTokenizer;
 var __webpack_exports__HieraForImageClassification = __webpack_exports__.HieraForImageClassification;
 var __webpack_exports__HieraModel = __webpack_exports__.HieraModel;
@@ -39249,6 +39416,7 @@ var __webpack_exports__Wav2Vec2ForCTC = __webpack_exports__.Wav2Vec2ForCTC;
 var __webpack_exports__Wav2Vec2ForSequenceClassification = __webpack_exports__.Wav2Vec2ForSequenceClassification;
 var __webpack_exports__Wav2Vec2Model = __webpack_exports__.Wav2Vec2Model;
 var __webpack_exports__Wav2Vec2PreTrainedModel = __webpack_exports__.Wav2Vec2PreTrainedModel;
+var __webpack_exports__Wav2Vec2Processor = __webpack_exports__.Wav2Vec2Processor;
 var __webpack_exports__Wav2Vec2ProcessorWithLM = __webpack_exports__.Wav2Vec2ProcessorWithLM;
 var __webpack_exports__WavLMForAudioFrameClassification = __webpack_exports__.WavLMForAudioFrameClassification;
 var __webpack_exports__WavLMForCTC = __webpack_exports__.WavLMForCTC;
@@ -39337,6 +39505,6 @@ var __webpack_exports__topk = __webpack_exports__.topk;
 var __webpack_exports__window_function = __webpack_exports__.window_function;
 var __webpack_exports__zeros = __webpack_exports__.zeros;
 var __webpack_exports__zeros_like = __webpack_exports__.zeros_like;
-export { __webpack_exports__ASTFeatureExtractor as ASTFeatureExtractor, __webpack_exports__ASTForAudioClassification as ASTForAudioClassification, __webpack_exports__ASTModel as ASTModel, __webpack_exports__ASTPreTrainedModel as ASTPreTrainedModel, __webpack_exports__AlbertForMaskedLM as AlbertForMaskedLM, __webpack_exports__AlbertForQuestionAnswering as AlbertForQuestionAnswering, __webpack_exports__AlbertForSequenceClassification as AlbertForSequenceClassification, __webpack_exports__AlbertModel as AlbertModel, __webpack_exports__AlbertPreTrainedModel as AlbertPreTrainedModel, __webpack_exports__AlbertTokenizer as AlbertTokenizer, __webpack_exports__AudioClassificationPipeline as AudioClassificationPipeline, __webpack_exports__AutoConfig as AutoConfig, __webpack_exports__AutoFeatureExtractor as AutoFeatureExtractor, __webpack_exports__AutoImageProcessor as AutoImageProcessor, __webpack_exports__AutoModel as AutoModel, __webpack_exports__AutoModelForAudioClassification as AutoModelForAudioClassification, __webpack_exports__AutoModelForAudioFrameClassification as AutoModelForAudioFrameClassification, __webpack_exports__AutoModelForCTC as AutoModelForCTC, __webpack_exports__AutoModelForCausalLM as AutoModelForCausalLM, __webpack_exports__AutoModelForDepthEstimation as AutoModelForDepthEstimation, __webpack_exports__AutoModelForDocumentQuestionAnswering as AutoModelForDocumentQuestionAnswering, __webpack_exports__AutoModelForImageClassification as AutoModelForImageClassification, __webpack_exports__AutoModelForImageFeatureExtraction as AutoModelForImageFeatureExtraction, __webpack_exports__AutoModelForImageMatting as AutoModelForImageMatting, __webpack_exports__AutoModelForImageSegmentation as AutoModelForImageSegmentation, __webpack_exports__AutoModelForImageToImage as AutoModelForImageToImage, __webpack_exports__AutoModelForMaskGeneration as AutoModelForMaskGeneration, __webpack_exports__AutoModelForMaskedLM as AutoModelForMaskedLM, __webpack_exports__AutoModelForNormalEstimation as AutoModelForNormalEstimation, __webpack_exports__AutoModelForObjectDetection as AutoModelForObjectDetection, __webpack_exports__AutoModelForPoseEstimation as AutoModelForPoseEstimation, __webpack_exports__AutoModelForQuestionAnswering as AutoModelForQuestionAnswering, __webpack_exports__AutoModelForSemanticSegmentation as AutoModelForSemanticSegmentation, __webpack_exports__AutoModelForSeq2SeqLM as AutoModelForSeq2SeqLM, __webpack_exports__AutoModelForSequenceClassification as AutoModelForSequenceClassification, __webpack_exports__AutoModelForSpeechSeq2Seq as AutoModelForSpeechSeq2Seq, __webpack_exports__AutoModelForTextToSpectrogram as AutoModelForTextToSpectrogram, __webpack_exports__AutoModelForTextToWaveform as AutoModelForTextToWaveform, __webpack_exports__AutoModelForTokenClassification as AutoModelForTokenClassification, __webpack_exports__AutoModelForUniversalSegmentation as AutoModelForUniversalSegmentation, __webpack_exports__AutoModelForVision2Seq as AutoModelForVision2Seq, __webpack_exports__AutoModelForXVector as AutoModelForXVector, __webpack_exports__AutoModelForZeroShotObjectDetection as AutoModelForZeroShotObjectDetection, __webpack_exports__AutoProcessor as AutoProcessor, __webpack_exports__AutoTokenizer as AutoTokenizer, __webpack_exports__AutomaticSpeechRecognitionPipeline as AutomaticSpeechRecognitionPipeline, __webpack_exports__BartForConditionalGeneration as BartForConditionalGeneration, __webpack_exports__BartForSequenceClassification as BartForSequenceClassification, __webpack_exports__BartModel as BartModel, __webpack_exports__BartPretrainedModel as BartPretrainedModel, __webpack_exports__BartTokenizer as BartTokenizer, __webpack_exports__BaseModelOutput as BaseModelOutput, __webpack_exports__BaseStreamer as BaseStreamer, __webpack_exports__BeitFeatureExtractor as BeitFeatureExtractor, __webpack_exports__BeitForImageClassification as BeitForImageClassification, __webpack_exports__BeitModel as BeitModel, __webpack_exports__BeitPreTrainedModel as BeitPreTrainedModel, __webpack_exports__BertForMaskedLM as BertForMaskedLM, __webpack_exports__BertForQuestionAnswering as BertForQuestionAnswering, __webpack_exports__BertForSequenceClassification as BertForSequenceClassification, __webpack_exports__BertForTokenClassification as BertForTokenClassification, __webpack_exports__BertModel as BertModel, __webpack_exports__BertPreTrainedModel as BertPreTrainedModel, __webpack_exports__BertTokenizer as BertTokenizer, __webpack_exports__BitImageProcessor as BitImageProcessor, __webpack_exports__BlenderbotForConditionalGeneration as BlenderbotForConditionalGeneration, __webpack_exports__BlenderbotModel as BlenderbotModel, __webpack_exports__BlenderbotPreTrainedModel as BlenderbotPreTrainedModel, __webpack_exports__BlenderbotSmallForConditionalGeneration as BlenderbotSmallForConditionalGeneration, __webpack_exports__BlenderbotSmallModel as BlenderbotSmallModel, __webpack_exports__BlenderbotSmallPreTrainedModel as BlenderbotSmallPreTrainedModel, __webpack_exports__BlenderbotSmallTokenizer as BlenderbotSmallTokenizer, __webpack_exports__BlenderbotTokenizer as BlenderbotTokenizer, __webpack_exports__BloomForCausalLM as BloomForCausalLM, __webpack_exports__BloomModel as BloomModel, __webpack_exports__BloomPreTrainedModel as BloomPreTrainedModel, __webpack_exports__BloomTokenizer as BloomTokenizer, __webpack_exports__CLIPFeatureExtractor as CLIPFeatureExtractor, __webpack_exports__CLIPImageProcessor as CLIPImageProcessor, __webpack_exports__CLIPModel as CLIPModel, __webpack_exports__CLIPPreTrainedModel as CLIPPreTrainedModel, __webpack_exports__CLIPSegForImageSegmentation as CLIPSegForImageSegmentation, __webpack_exports__CLIPSegModel as CLIPSegModel, __webpack_exports__CLIPSegPreTrainedModel as CLIPSegPreTrainedModel, __webpack_exports__CLIPTextModel as CLIPTextModel, __webpack_exports__CLIPTextModelWithProjection as CLIPTextModelWithProjection, __webpack_exports__CLIPTokenizer as CLIPTokenizer, __webpack_exports__CLIPVisionModel as CLIPVisionModel, __webpack_exports__CLIPVisionModelWithProjection as CLIPVisionModelWithProjection, __webpack_exports__CamembertForMaskedLM as CamembertForMaskedLM, __webpack_exports__CamembertForQuestionAnswering as CamembertForQuestionAnswering, __webpack_exports__CamembertForSequenceClassification as CamembertForSequenceClassification, __webpack_exports__CamembertForTokenClassification as CamembertForTokenClassification, __webpack_exports__CamembertModel as CamembertModel, __webpack_exports__CamembertPreTrainedModel as CamembertPreTrainedModel, __webpack_exports__CamembertTokenizer as CamembertTokenizer, __webpack_exports__CausalLMOutput as CausalLMOutput, __webpack_exports__CausalLMOutputWithPast as CausalLMOutputWithPast, __webpack_exports__ChineseCLIPFeatureExtractor as ChineseCLIPFeatureExtractor, __webpack_exports__ChineseCLIPModel as ChineseCLIPModel, __webpack_exports__ChineseCLIPPreTrainedModel as ChineseCLIPPreTrainedModel, __webpack_exports__ClapAudioModelWithProjection as ClapAudioModelWithProjection, __webpack_exports__ClapFeatureExtractor as ClapFeatureExtractor, __webpack_exports__ClapModel as ClapModel, __webpack_exports__ClapPreTrainedModel as ClapPreTrainedModel, __webpack_exports__ClapTextModelWithProjection as ClapTextModelWithProjection, __webpack_exports__ClassifierFreeGuidanceLogitsProcessor as ClassifierFreeGuidanceLogitsProcessor, __webpack_exports__CodeGenForCausalLM as CodeGenForCausalLM, __webpack_exports__CodeGenModel as CodeGenModel, __webpack_exports__CodeGenPreTrainedModel as CodeGenPreTrainedModel, __webpack_exports__CodeGenTokenizer as CodeGenTokenizer, __webpack_exports__CodeLlamaTokenizer as CodeLlamaTokenizer, __webpack_exports__CohereForCausalLM as CohereForCausalLM, __webpack_exports__CohereModel as CohereModel, __webpack_exports__CoherePreTrainedModel as CoherePreTrainedModel, __webpack_exports__CohereTokenizer as CohereTokenizer, __webpack_exports__ConvBertForMaskedLM as ConvBertForMaskedLM, __webpack_exports__ConvBertForQuestionAnswering as ConvBertForQuestionAnswering, __webpack_exports__ConvBertForSequenceClassification as ConvBertForSequenceClassification, __webpack_exports__ConvBertForTokenClassification as ConvBertForTokenClassification, __webpack_exports__ConvBertModel as ConvBertModel, __webpack_exports__ConvBertPreTrainedModel as ConvBertPreTrainedModel, __webpack_exports__ConvBertTokenizer as ConvBertTokenizer, __webpack_exports__ConvNextFeatureExtractor as ConvNextFeatureExtractor, __webpack_exports__ConvNextForImageClassification as ConvNextForImageClassification, __webpack_exports__ConvNextImageProcessor as ConvNextImageProcessor, __webpack_exports__ConvNextModel as ConvNextModel, __webpack_exports__ConvNextPreTrainedModel as ConvNextPreTrainedModel, __webpack_exports__ConvNextV2ForImageClassification as ConvNextV2ForImageClassification, __webpack_exports__ConvNextV2Model as ConvNextV2Model, __webpack_exports__ConvNextV2PreTrainedModel as ConvNextV2PreTrainedModel, __webpack_exports__DPTFeatureExtractor as DPTFeatureExtractor, __webpack_exports__DPTForDepthEstimation as DPTForDepthEstimation, __webpack_exports__DPTImageProcessor as DPTImageProcessor, __webpack_exports__DPTModel as DPTModel, __webpack_exports__DPTPreTrainedModel as DPTPreTrainedModel, __webpack_exports__DebertaForMaskedLM as DebertaForMaskedLM, __webpack_exports__DebertaForQuestionAnswering as DebertaForQuestionAnswering, __webpack_exports__DebertaForSequenceClassification as DebertaForSequenceClassification, __webpack_exports__DebertaForTokenClassification as DebertaForTokenClassification, __webpack_exports__DebertaModel as DebertaModel, __webpack_exports__DebertaPreTrainedModel as DebertaPreTrainedModel, __webpack_exports__DebertaTokenizer as DebertaTokenizer, __webpack_exports__DebertaV2ForMaskedLM as DebertaV2ForMaskedLM, __webpack_exports__DebertaV2ForQuestionAnswering as DebertaV2ForQuestionAnswering, __webpack_exports__DebertaV2ForSequenceClassification as DebertaV2ForSequenceClassification, __webpack_exports__DebertaV2ForTokenClassification as DebertaV2ForTokenClassification, __webpack_exports__DebertaV2Model as DebertaV2Model, __webpack_exports__DebertaV2PreTrainedModel as DebertaV2PreTrainedModel, __webpack_exports__DebertaV2Tokenizer as DebertaV2Tokenizer, __webpack_exports__DecisionTransformerModel as DecisionTransformerModel, __webpack_exports__DecisionTransformerPreTrainedModel as DecisionTransformerPreTrainedModel, __webpack_exports__DeiTFeatureExtractor as DeiTFeatureExtractor, __webpack_exports__DeiTForImageClassification as DeiTForImageClassification, __webpack_exports__DeiTImageProcessor as DeiTImageProcessor, __webpack_exports__DeiTModel as DeiTModel, __webpack_exports__DeiTPreTrainedModel as DeiTPreTrainedModel, __webpack_exports__DepthAnythingForDepthEstimation as DepthAnythingForDepthEstimation, __webpack_exports__DepthAnythingPreTrainedModel as DepthAnythingPreTrainedModel, __webpack_exports__DepthEstimationPipeline as DepthEstimationPipeline, __webpack_exports__DepthProForDepthEstimation as DepthProForDepthEstimation, __webpack_exports__DepthProPreTrainedModel as DepthProPreTrainedModel, __webpack_exports__DetrFeatureExtractor as DetrFeatureExtractor, __webpack_exports__DetrForObjectDetection as DetrForObjectDetection, __webpack_exports__DetrForSegmentation as DetrForSegmentation, __webpack_exports__DetrImageProcessor as DetrImageProcessor, __webpack_exports__DetrModel as DetrModel, __webpack_exports__DetrObjectDetectionOutput as DetrObjectDetectionOutput, __webpack_exports__DetrPreTrainedModel as DetrPreTrainedModel, __webpack_exports__DetrSegmentationOutput as DetrSegmentationOutput, __webpack_exports__Dinov2ForImageClassification as Dinov2ForImageClassification, __webpack_exports__Dinov2Model as Dinov2Model, __webpack_exports__Dinov2PreTrainedModel as Dinov2PreTrainedModel, __webpack_exports__Dinov2WithRegistersForImageClassification as Dinov2WithRegistersForImageClassification, __webpack_exports__Dinov2WithRegistersModel as Dinov2WithRegistersModel, __webpack_exports__Dinov2WithRegistersPreTrainedModel as Dinov2WithRegistersPreTrainedModel, __webpack_exports__DistilBertForMaskedLM as DistilBertForMaskedLM, __webpack_exports__DistilBertForQuestionAnswering as DistilBertForQuestionAnswering, __webpack_exports__DistilBertForSequenceClassification as DistilBertForSequenceClassification, __webpack_exports__DistilBertForTokenClassification as DistilBertForTokenClassification, __webpack_exports__DistilBertModel as DistilBertModel, __webpack_exports__DistilBertPreTrainedModel as DistilBertPreTrainedModel, __webpack_exports__DistilBertTokenizer as DistilBertTokenizer, __webpack_exports__DocumentQuestionAnsweringPipeline as DocumentQuestionAnsweringPipeline, __webpack_exports__DonutFeatureExtractor as DonutFeatureExtractor, __webpack_exports__DonutImageProcessor as DonutImageProcessor, __webpack_exports__DonutSwinModel as DonutSwinModel, __webpack_exports__DonutSwinPreTrainedModel as DonutSwinPreTrainedModel, __webpack_exports__EfficientNetForImageClassification as EfficientNetForImageClassification, __webpack_exports__EfficientNetImageProcessor as EfficientNetImageProcessor, __webpack_exports__EfficientNetModel as EfficientNetModel, __webpack_exports__EfficientNetPreTrainedModel as EfficientNetPreTrainedModel, __webpack_exports__ElectraForMaskedLM as ElectraForMaskedLM, __webpack_exports__ElectraForQuestionAnswering as ElectraForQuestionAnswering, __webpack_exports__ElectraForSequenceClassification as ElectraForSequenceClassification, __webpack_exports__ElectraForTokenClassification as ElectraForTokenClassification, __webpack_exports__ElectraModel as ElectraModel, __webpack_exports__ElectraPreTrainedModel as ElectraPreTrainedModel, __webpack_exports__ElectraTokenizer as ElectraTokenizer, __webpack_exports__EosTokenCriteria as EosTokenCriteria, __webpack_exports__EsmForMaskedLM as EsmForMaskedLM, __webpack_exports__EsmForSequenceClassification as EsmForSequenceClassification, __webpack_exports__EsmForTokenClassification as EsmForTokenClassification, __webpack_exports__EsmModel as EsmModel, __webpack_exports__EsmPreTrainedModel as EsmPreTrainedModel, __webpack_exports__EsmTokenizer as EsmTokenizer, __webpack_exports__ExaoneForCausalLM as ExaoneForCausalLM, __webpack_exports__ExaoneModel as ExaoneModel, __webpack_exports__ExaonePreTrainedModel as ExaonePreTrainedModel, __webpack_exports__FFT as FFT, __webpack_exports__FalconForCausalLM as FalconForCausalLM, __webpack_exports__FalconModel as FalconModel, __webpack_exports__FalconPreTrainedModel as FalconPreTrainedModel, __webpack_exports__FalconTokenizer as FalconTokenizer, __webpack_exports__FastViTForImageClassification as FastViTForImageClassification, __webpack_exports__FastViTModel as FastViTModel, __webpack_exports__FastViTPreTrainedModel as FastViTPreTrainedModel, __webpack_exports__FeatureExtractionPipeline as FeatureExtractionPipeline, __webpack_exports__FeatureExtractor as FeatureExtractor, __webpack_exports__FillMaskPipeline as FillMaskPipeline, __webpack_exports__Florence2ForConditionalGeneration as Florence2ForConditionalGeneration, __webpack_exports__Florence2PreTrainedModel as Florence2PreTrainedModel, __webpack_exports__Florence2Processor as Florence2Processor, __webpack_exports__ForcedBOSTokenLogitsProcessor as ForcedBOSTokenLogitsProcessor, __webpack_exports__ForcedEOSTokenLogitsProcessor as ForcedEOSTokenLogitsProcessor, __webpack_exports__GLPNFeatureExtractor as GLPNFeatureExtractor, __webpack_exports__GLPNForDepthEstimation as GLPNForDepthEstimation, __webpack_exports__GLPNModel as GLPNModel, __webpack_exports__GLPNPreTrainedModel as GLPNPreTrainedModel, __webpack_exports__GPT2LMHeadModel as GPT2LMHeadModel, __webpack_exports__GPT2Model as GPT2Model, __webpack_exports__GPT2PreTrainedModel as GPT2PreTrainedModel, __webpack_exports__GPT2Tokenizer as GPT2Tokenizer, __webpack_exports__GPTBigCodeForCausalLM as GPTBigCodeForCausalLM, __webpack_exports__GPTBigCodeModel as GPTBigCodeModel, __webpack_exports__GPTBigCodePreTrainedModel as GPTBigCodePreTrainedModel, __webpack_exports__GPTJForCausalLM as GPTJForCausalLM, __webpack_exports__GPTJModel as GPTJModel, __webpack_exports__GPTJPreTrainedModel as GPTJPreTrainedModel, __webpack_exports__GPTNeoForCausalLM as GPTNeoForCausalLM, __webpack_exports__GPTNeoModel as GPTNeoModel, __webpack_exports__GPTNeoPreTrainedModel as GPTNeoPreTrainedModel, __webpack_exports__GPTNeoXForCausalLM as GPTNeoXForCausalLM, __webpack_exports__GPTNeoXModel as GPTNeoXModel, __webpack_exports__GPTNeoXPreTrainedModel as GPTNeoXPreTrainedModel, __webpack_exports__GPTNeoXTokenizer as GPTNeoXTokenizer, __webpack_exports__Gemma2ForCausalLM as Gemma2ForCausalLM, __webpack_exports__Gemma2Model as Gemma2Model, __webpack_exports__Gemma2PreTrainedModel as Gemma2PreTrainedModel, __webpack_exports__GemmaForCausalLM as GemmaForCausalLM, __webpack_exports__GemmaModel as GemmaModel, __webpack_exports__GemmaPreTrainedModel as GemmaPreTrainedModel, __webpack_exports__GemmaTokenizer as GemmaTokenizer, __webpack_exports__GraniteForCausalLM as GraniteForCausalLM, __webpack_exports__GraniteModel as GraniteModel, __webpack_exports__GranitePreTrainedModel as GranitePreTrainedModel, __webpack_exports__Grok1Tokenizer as Grok1Tokenizer, __webpack_exports__GroundingDinoForObjectDetection as GroundingDinoForObjectDetection, __webpack_exports__GroundingDinoImageProcessor as GroundingDinoImageProcessor, __webpack_exports__GroundingDinoPreTrainedModel as GroundingDinoPreTrainedModel, __webpack_exports__GroundingDinoProcessor as GroundingDinoProcessor, __webpack_exports__GroupViTModel as GroupViTModel, __webpack_exports__GroupViTPreTrainedModel as GroupViTPreTrainedModel, __webpack_exports__HerbertTokenizer as HerbertTokenizer, __webpack_exports__HieraForImageClassification as HieraForImageClassification, __webpack_exports__HieraModel as HieraModel, __webpack_exports__HieraPreTrainedModel as HieraPreTrainedModel, __webpack_exports__HubertForCTC as HubertForCTC, __webpack_exports__HubertForSequenceClassification as HubertForSequenceClassification, __webpack_exports__HubertModel as HubertModel, __webpack_exports__HubertPreTrainedModel as HubertPreTrainedModel, __webpack_exports__IJepaForImageClassification as IJepaForImageClassification, __webpack_exports__IJepaModel as IJepaModel, __webpack_exports__IJepaPreTrainedModel as IJepaPreTrainedModel, __webpack_exports__Idefics3ForConditionalGeneration as Idefics3ForConditionalGeneration, __webpack_exports__Idefics3ImageProcessor as Idefics3ImageProcessor, __webpack_exports__Idefics3PreTrainedModel as Idefics3PreTrainedModel, __webpack_exports__Idefics3Processor as Idefics3Processor, __webpack_exports__ImageClassificationPipeline as ImageClassificationPipeline, __webpack_exports__ImageFeatureExtractionPipeline as ImageFeatureExtractionPipeline, __webpack_exports__ImageFeatureExtractor as ImageFeatureExtractor, __webpack_exports__ImageMattingOutput as ImageMattingOutput, __webpack_exports__ImageProcessor as ImageProcessor, __webpack_exports__ImageSegmentationPipeline as ImageSegmentationPipeline, __webpack_exports__ImageToImagePipeline as ImageToImagePipeline, __webpack_exports__ImageToTextPipeline as ImageToTextPipeline, __webpack_exports__InterruptableStoppingCriteria as InterruptableStoppingCriteria, __webpack_exports__JAISLMHeadModel as JAISLMHeadModel, __webpack_exports__JAISModel as JAISModel, __webpack_exports__JAISPreTrainedModel as JAISPreTrainedModel, __webpack_exports__JinaCLIPImageProcessor as JinaCLIPImageProcessor, __webpack_exports__JinaCLIPModel as JinaCLIPModel, __webpack_exports__JinaCLIPPreTrainedModel as JinaCLIPPreTrainedModel, __webpack_exports__JinaCLIPProcessor as JinaCLIPProcessor, __webpack_exports__JinaCLIPTextModel as JinaCLIPTextModel, __webpack_exports__JinaCLIPVisionModel as JinaCLIPVisionModel, __webpack_exports__LlamaForCausalLM as LlamaForCausalLM, __webpack_exports__LlamaModel as LlamaModel, __webpack_exports__LlamaPreTrainedModel as LlamaPreTrainedModel, __webpack_exports__LlamaTokenizer as LlamaTokenizer, __webpack_exports__LlavaForConditionalGeneration as LlavaForConditionalGeneration, __webpack_exports__LlavaOnevisionForConditionalGeneration as LlavaOnevisionForConditionalGeneration, __webpack_exports__LlavaOnevisionImageProcessor as LlavaOnevisionImageProcessor, __webpack_exports__LlavaPreTrainedModel as LlavaPreTrainedModel, __webpack_exports__LogitsProcessor as LogitsProcessor, __webpack_exports__LogitsProcessorList as LogitsProcessorList, __webpack_exports__LogitsWarper as LogitsWarper, __webpack_exports__LongT5ForConditionalGeneration as LongT5ForConditionalGeneration, __webpack_exports__LongT5Model as LongT5Model, __webpack_exports__LongT5PreTrainedModel as LongT5PreTrainedModel, __webpack_exports__M2M100ForConditionalGeneration as M2M100ForConditionalGeneration, __webpack_exports__M2M100Model as M2M100Model, __webpack_exports__M2M100PreTrainedModel as M2M100PreTrainedModel, __webpack_exports__M2M100Tokenizer as M2M100Tokenizer, __webpack_exports__MBart50Tokenizer as MBart50Tokenizer, __webpack_exports__MBartForCausalLM as MBartForCausalLM, __webpack_exports__MBartForConditionalGeneration as MBartForConditionalGeneration, __webpack_exports__MBartForSequenceClassification as MBartForSequenceClassification, __webpack_exports__MBartModel as MBartModel, __webpack_exports__MBartPreTrainedModel as MBartPreTrainedModel, __webpack_exports__MBartTokenizer as MBartTokenizer, __webpack_exports__MPNetForMaskedLM as MPNetForMaskedLM, __webpack_exports__MPNetForQuestionAnswering as MPNetForQuestionAnswering, __webpack_exports__MPNetForSequenceClassification as MPNetForSequenceClassification, __webpack_exports__MPNetForTokenClassification as MPNetForTokenClassification, __webpack_exports__MPNetModel as MPNetModel, __webpack_exports__MPNetPreTrainedModel as MPNetPreTrainedModel, __webpack_exports__MPNetTokenizer as MPNetTokenizer, __webpack_exports__MT5ForConditionalGeneration as MT5ForConditionalGeneration, __webpack_exports__MT5Model as MT5Model, __webpack_exports__MT5PreTrainedModel as MT5PreTrainedModel, __webpack_exports__MarianMTModel as MarianMTModel, __webpack_exports__MarianModel as MarianModel, __webpack_exports__MarianPreTrainedModel as MarianPreTrainedModel, __webpack_exports__MarianTokenizer as MarianTokenizer, __webpack_exports__Mask2FormerImageProcessor as Mask2FormerImageProcessor, __webpack_exports__MaskFormerFeatureExtractor as MaskFormerFeatureExtractor, __webpack_exports__MaskFormerForInstanceSegmentation as MaskFormerForInstanceSegmentation, __webpack_exports__MaskFormerImageProcessor as MaskFormerImageProcessor, __webpack_exports__MaskFormerModel as MaskFormerModel, __webpack_exports__MaskFormerPreTrainedModel as MaskFormerPreTrainedModel, __webpack_exports__MaskedLMOutput as MaskedLMOutput, __webpack_exports__MaxLengthCriteria as MaxLengthCriteria, __webpack_exports__MgpstrForSceneTextRecognition as MgpstrForSceneTextRecognition, __webpack_exports__MgpstrModelOutput as MgpstrModelOutput, __webpack_exports__MgpstrPreTrainedModel as MgpstrPreTrainedModel, __webpack_exports__MgpstrProcessor as MgpstrProcessor, __webpack_exports__MgpstrTokenizer as MgpstrTokenizer, __webpack_exports__MinLengthLogitsProcessor as MinLengthLogitsProcessor, __webpack_exports__MinNewTokensLengthLogitsProcessor as MinNewTokensLengthLogitsProcessor, __webpack_exports__MistralForCausalLM as MistralForCausalLM, __webpack_exports__MistralModel as MistralModel, __webpack_exports__MistralPreTrainedModel as MistralPreTrainedModel, __webpack_exports__MobileBertForMaskedLM as MobileBertForMaskedLM, __webpack_exports__MobileBertForQuestionAnswering as MobileBertForQuestionAnswering, __webpack_exports__MobileBertForSequenceClassification as MobileBertForSequenceClassification, __webpack_exports__MobileBertModel as MobileBertModel, __webpack_exports__MobileBertPreTrainedModel as MobileBertPreTrainedModel, __webpack_exports__MobileBertTokenizer as MobileBertTokenizer, __webpack_exports__MobileLLMForCausalLM as MobileLLMForCausalLM, __webpack_exports__MobileLLMModel as MobileLLMModel, __webpack_exports__MobileLLMPreTrainedModel as MobileLLMPreTrainedModel, __webpack_exports__MobileNetV1FeatureExtractor as MobileNetV1FeatureExtractor, __webpack_exports__MobileNetV1ForImageClassification as MobileNetV1ForImageClassification, __webpack_exports__MobileNetV1ImageProcessor as MobileNetV1ImageProcessor, __webpack_exports__MobileNetV1Model as MobileNetV1Model, __webpack_exports__MobileNetV1PreTrainedModel as MobileNetV1PreTrainedModel, __webpack_exports__MobileNetV2FeatureExtractor as MobileNetV2FeatureExtractor, __webpack_exports__MobileNetV2ForImageClassification as MobileNetV2ForImageClassification, __webpack_exports__MobileNetV2ImageProcessor as MobileNetV2ImageProcessor, __webpack_exports__MobileNetV2Model as MobileNetV2Model, __webpack_exports__MobileNetV2PreTrainedModel as MobileNetV2PreTrainedModel, __webpack_exports__MobileNetV3FeatureExtractor as MobileNetV3FeatureExtractor, __webpack_exports__MobileNetV3ForImageClassification as MobileNetV3ForImageClassification, __webpack_exports__MobileNetV3ImageProcessor as MobileNetV3ImageProcessor, __webpack_exports__MobileNetV3Model as MobileNetV3Model, __webpack_exports__MobileNetV3PreTrainedModel as MobileNetV3PreTrainedModel, __webpack_exports__MobileNetV4FeatureExtractor as MobileNetV4FeatureExtractor, __webpack_exports__MobileNetV4ForImageClassification as MobileNetV4ForImageClassification, __webpack_exports__MobileNetV4ImageProcessor as MobileNetV4ImageProcessor, __webpack_exports__MobileNetV4Model as MobileNetV4Model, __webpack_exports__MobileNetV4PreTrainedModel as MobileNetV4PreTrainedModel, __webpack_exports__MobileViTFeatureExtractor as MobileViTFeatureExtractor, __webpack_exports__MobileViTForImageClassification as MobileViTForImageClassification, __webpack_exports__MobileViTImageProcessor as MobileViTImageProcessor, __webpack_exports__MobileViTModel as MobileViTModel, __webpack_exports__MobileViTPreTrainedModel as MobileViTPreTrainedModel, __webpack_exports__MobileViTV2ForImageClassification as MobileViTV2ForImageClassification, __webpack_exports__MobileViTV2Model as MobileViTV2Model, __webpack_exports__MobileViTV2PreTrainedModel as MobileViTV2PreTrainedModel, __webpack_exports__ModelOutput as ModelOutput, __webpack_exports__ModernBertForMaskedLM as ModernBertForMaskedLM, __webpack_exports__ModernBertForSequenceClassification as ModernBertForSequenceClassification, __webpack_exports__ModernBertForTokenClassification as ModernBertForTokenClassification, __webpack_exports__ModernBertModel as ModernBertModel, __webpack_exports__ModernBertPreTrainedModel as ModernBertPreTrainedModel, __webpack_exports__Moondream1ForConditionalGeneration as Moondream1ForConditionalGeneration, __webpack_exports__MoonshineFeatureExtractor as MoonshineFeatureExtractor, __webpack_exports__MoonshineForConditionalGeneration as MoonshineForConditionalGeneration, __webpack_exports__MoonshineModel as MoonshineModel, __webpack_exports__MoonshinePreTrainedModel as MoonshinePreTrainedModel, __webpack_exports__MoonshineProcessor as MoonshineProcessor, __webpack_exports__MptForCausalLM as MptForCausalLM, __webpack_exports__MptModel as MptModel, __webpack_exports__MptPreTrainedModel as MptPreTrainedModel, __webpack_exports__MultiModalityCausalLM as MultiModalityCausalLM, __webpack_exports__MultiModalityPreTrainedModel as MultiModalityPreTrainedModel, __webpack_exports__MusicgenForCausalLM as MusicgenForCausalLM, __webpack_exports__MusicgenForConditionalGeneration as MusicgenForConditionalGeneration, __webpack_exports__MusicgenModel as MusicgenModel, __webpack_exports__MusicgenPreTrainedModel as MusicgenPreTrainedModel, __webpack_exports__NllbTokenizer as NllbTokenizer, __webpack_exports__NoBadWordsLogitsProcessor as NoBadWordsLogitsProcessor, __webpack_exports__NoRepeatNGramLogitsProcessor as NoRepeatNGramLogitsProcessor, __webpack_exports__NomicBertModel as NomicBertModel, __webpack_exports__NomicBertPreTrainedModel as NomicBertPreTrainedModel, __webpack_exports__NougatImageProcessor as NougatImageProcessor, __webpack_exports__NougatTokenizer as NougatTokenizer, __webpack_exports__OPTForCausalLM as OPTForCausalLM, __webpack_exports__OPTModel as OPTModel, __webpack_exports__OPTPreTrainedModel as OPTPreTrainedModel, __webpack_exports__ObjectDetectionPipeline as ObjectDetectionPipeline, __webpack_exports__Olmo2ForCausalLM as Olmo2ForCausalLM, __webpack_exports__Olmo2Model as Olmo2Model, __webpack_exports__Olmo2PreTrainedModel as Olmo2PreTrainedModel, __webpack_exports__OlmoForCausalLM as OlmoForCausalLM, __webpack_exports__OlmoModel as OlmoModel, __webpack_exports__OlmoPreTrainedModel as OlmoPreTrainedModel, __webpack_exports__OpenELMForCausalLM as OpenELMForCausalLM, __webpack_exports__OpenELMModel as OpenELMModel, __webpack_exports__OpenELMPreTrainedModel as OpenELMPreTrainedModel, __webpack_exports__OwlViTFeatureExtractor as OwlViTFeatureExtractor, __webpack_exports__OwlViTForObjectDetection as OwlViTForObjectDetection, __webpack_exports__OwlViTImageProcessor as OwlViTImageProcessor, __webpack_exports__OwlViTModel as OwlViTModel, __webpack_exports__OwlViTPreTrainedModel as OwlViTPreTrainedModel, __webpack_exports__OwlViTProcessor as OwlViTProcessor, __webpack_exports__Owlv2ForObjectDetection as Owlv2ForObjectDetection, __webpack_exports__Owlv2ImageProcessor as Owlv2ImageProcessor, __webpack_exports__Owlv2Model as Owlv2Model, __webpack_exports__Owlv2PreTrainedModel as Owlv2PreTrainedModel, __webpack_exports__PaliGemmaForConditionalGeneration as PaliGemmaForConditionalGeneration, __webpack_exports__PaliGemmaPreTrainedModel as PaliGemmaPreTrainedModel, __webpack_exports__PaliGemmaProcessor as PaliGemmaProcessor, __webpack_exports__PatchTSMixerForPrediction as PatchTSMixerForPrediction, __webpack_exports__PatchTSMixerModel as PatchTSMixerModel, __webpack_exports__PatchTSMixerPreTrainedModel as PatchTSMixerPreTrainedModel, __webpack_exports__PatchTSTForPrediction as PatchTSTForPrediction, __webpack_exports__PatchTSTModel as PatchTSTModel, __webpack_exports__PatchTSTPreTrainedModel as PatchTSTPreTrainedModel, __webpack_exports__Phi3ForCausalLM as Phi3ForCausalLM, __webpack_exports__Phi3Model as Phi3Model, __webpack_exports__Phi3PreTrainedModel as Phi3PreTrainedModel, __webpack_exports__Phi3VForCausalLM as Phi3VForCausalLM, __webpack_exports__Phi3VImageProcessor as Phi3VImageProcessor, __webpack_exports__Phi3VPreTrainedModel as Phi3VPreTrainedModel, __webpack_exports__Phi3VProcessor as Phi3VProcessor, __webpack_exports__PhiForCausalLM as PhiForCausalLM, __webpack_exports__PhiModel as PhiModel, __webpack_exports__PhiPreTrainedModel as PhiPreTrainedModel, __webpack_exports__Pipeline as Pipeline, __webpack_exports__PreTrainedModel as PreTrainedModel, __webpack_exports__PreTrainedTokenizer as PreTrainedTokenizer, __webpack_exports__PretrainedConfig as PretrainedConfig, __webpack_exports__PretrainedMixin as PretrainedMixin, __webpack_exports__Processor as Processor, __webpack_exports__PvtForImageClassification as PvtForImageClassification, __webpack_exports__PvtImageProcessor as PvtImageProcessor, __webpack_exports__PvtModel as PvtModel, __webpack_exports__PvtPreTrainedModel as PvtPreTrainedModel, __webpack_exports__PyAnnoteFeatureExtractor as PyAnnoteFeatureExtractor, __webpack_exports__PyAnnoteForAudioFrameClassification as PyAnnoteForAudioFrameClassification, __webpack_exports__PyAnnoteModel as PyAnnoteModel, __webpack_exports__PyAnnotePreTrainedModel as PyAnnotePreTrainedModel, __webpack_exports__PyAnnoteProcessor as PyAnnoteProcessor, __webpack_exports__QuestionAnsweringModelOutput as QuestionAnsweringModelOutput, __webpack_exports__QuestionAnsweringPipeline as QuestionAnsweringPipeline, __webpack_exports__Qwen2ForCausalLM as Qwen2ForCausalLM, __webpack_exports__Qwen2Model as Qwen2Model, __webpack_exports__Qwen2PreTrainedModel as Qwen2PreTrainedModel, __webpack_exports__Qwen2Tokenizer as Qwen2Tokenizer, __webpack_exports__Qwen2VLForConditionalGeneration as Qwen2VLForConditionalGeneration, __webpack_exports__Qwen2VLImageProcessor as Qwen2VLImageProcessor, __webpack_exports__Qwen2VLPreTrainedModel as Qwen2VLPreTrainedModel, __webpack_exports__Qwen2VLProcessor as Qwen2VLProcessor, __webpack_exports__RTDetrForObjectDetection as RTDetrForObjectDetection, __webpack_exports__RTDetrImageProcessor as RTDetrImageProcessor, __webpack_exports__RTDetrModel as RTDetrModel, __webpack_exports__RTDetrObjectDetectionOutput as RTDetrObjectDetectionOutput, __webpack_exports__RTDetrPreTrainedModel as RTDetrPreTrainedModel, __webpack_exports__RawAudio as RawAudio, __webpack_exports__RawImage as RawImage, __webpack_exports__RepetitionPenaltyLogitsProcessor as RepetitionPenaltyLogitsProcessor, __webpack_exports__ResNetForImageClassification as ResNetForImageClassification, __webpack_exports__ResNetModel as ResNetModel, __webpack_exports__ResNetPreTrainedModel as ResNetPreTrainedModel, __webpack_exports__RoFormerForMaskedLM as RoFormerForMaskedLM, __webpack_exports__RoFormerForQuestionAnswering as RoFormerForQuestionAnswering, __webpack_exports__RoFormerForSequenceClassification as RoFormerForSequenceClassification, __webpack_exports__RoFormerForTokenClassification as RoFormerForTokenClassification, __webpack_exports__RoFormerModel as RoFormerModel, __webpack_exports__RoFormerPreTrainedModel as RoFormerPreTrainedModel, __webpack_exports__RoFormerTokenizer as RoFormerTokenizer, __webpack_exports__RobertaForMaskedLM as RobertaForMaskedLM, __webpack_exports__RobertaForQuestionAnswering as RobertaForQuestionAnswering, __webpack_exports__RobertaForSequenceClassification as RobertaForSequenceClassification, __webpack_exports__RobertaForTokenClassification as RobertaForTokenClassification, __webpack_exports__RobertaModel as RobertaModel, __webpack_exports__RobertaPreTrainedModel as RobertaPreTrainedModel, __webpack_exports__RobertaTokenizer as RobertaTokenizer, __webpack_exports__SamImageProcessor as SamImageProcessor, __webpack_exports__SamImageSegmentationOutput as SamImageSegmentationOutput, __webpack_exports__SamModel as SamModel, __webpack_exports__SamPreTrainedModel as SamPreTrainedModel, __webpack_exports__SamProcessor as SamProcessor, __webpack_exports__SapiensForDepthEstimation as SapiensForDepthEstimation, __webpack_exports__SapiensForNormalEstimation as SapiensForNormalEstimation, __webpack_exports__SapiensForSemanticSegmentation as SapiensForSemanticSegmentation, __webpack_exports__SapiensPreTrainedModel as SapiensPreTrainedModel, __webpack_exports__SeamlessM4TFeatureExtractor as SeamlessM4TFeatureExtractor, __webpack_exports__SegformerFeatureExtractor as SegformerFeatureExtractor, __webpack_exports__SegformerForImageClassification as SegformerForImageClassification, __webpack_exports__SegformerForSemanticSegmentation as SegformerForSemanticSegmentation, __webpack_exports__SegformerImageProcessor as SegformerImageProcessor, __webpack_exports__SegformerModel as SegformerModel, __webpack_exports__SegformerPreTrainedModel as SegformerPreTrainedModel, __webpack_exports__Seq2SeqLMOutput as Seq2SeqLMOutput, __webpack_exports__SequenceClassifierOutput as SequenceClassifierOutput, __webpack_exports__SiglipImageProcessor as SiglipImageProcessor, __webpack_exports__SiglipModel as SiglipModel, __webpack_exports__SiglipPreTrainedModel as SiglipPreTrainedModel, __webpack_exports__SiglipTextModel as SiglipTextModel, __webpack_exports__SiglipTokenizer as SiglipTokenizer, __webpack_exports__SiglipVisionModel as SiglipVisionModel, __webpack_exports__SpeechT5FeatureExtractor as SpeechT5FeatureExtractor, __webpack_exports__SpeechT5ForSpeechToText as SpeechT5ForSpeechToText, __webpack_exports__SpeechT5ForTextToSpeech as SpeechT5ForTextToSpeech, __webpack_exports__SpeechT5HifiGan as SpeechT5HifiGan, __webpack_exports__SpeechT5Model as SpeechT5Model, __webpack_exports__SpeechT5PreTrainedModel as SpeechT5PreTrainedModel, __webpack_exports__SpeechT5Processor as SpeechT5Processor, __webpack_exports__SpeechT5Tokenizer as SpeechT5Tokenizer, __webpack_exports__SqueezeBertForMaskedLM as SqueezeBertForMaskedLM, __webpack_exports__SqueezeBertForQuestionAnswering as SqueezeBertForQuestionAnswering, __webpack_exports__SqueezeBertForSequenceClassification as SqueezeBertForSequenceClassification, __webpack_exports__SqueezeBertModel as SqueezeBertModel, __webpack_exports__SqueezeBertPreTrainedModel as SqueezeBertPreTrainedModel, __webpack_exports__SqueezeBertTokenizer as SqueezeBertTokenizer, __webpack_exports__StableLmForCausalLM as StableLmForCausalLM, __webpack_exports__StableLmModel as StableLmModel, __webpack_exports__StableLmPreTrainedModel as StableLmPreTrainedModel, __webpack_exports__Starcoder2ForCausalLM as Starcoder2ForCausalLM, __webpack_exports__Starcoder2Model as Starcoder2Model, __webpack_exports__Starcoder2PreTrainedModel as Starcoder2PreTrainedModel, __webpack_exports__StoppingCriteria as StoppingCriteria, __webpack_exports__StoppingCriteriaList as StoppingCriteriaList, __webpack_exports__StyleTextToSpeech2Model as StyleTextToSpeech2Model, __webpack_exports__StyleTextToSpeech2PreTrainedModel as StyleTextToSpeech2PreTrainedModel, __webpack_exports__SummarizationPipeline as SummarizationPipeline, __webpack_exports__SuppressTokensAtBeginLogitsProcessor as SuppressTokensAtBeginLogitsProcessor, __webpack_exports__Swin2SRForImageSuperResolution as Swin2SRForImageSuperResolution, __webpack_exports__Swin2SRImageProcessor as Swin2SRImageProcessor, __webpack_exports__Swin2SRModel as Swin2SRModel, __webpack_exports__Swin2SRPreTrainedModel as Swin2SRPreTrainedModel, __webpack_exports__SwinForImageClassification as SwinForImageClassification, __webpack_exports__SwinModel as SwinModel, __webpack_exports__SwinPreTrainedModel as SwinPreTrainedModel, __webpack_exports__T5ForConditionalGeneration as T5ForConditionalGeneration, __webpack_exports__T5Model as T5Model, __webpack_exports__T5PreTrainedModel as T5PreTrainedModel, __webpack_exports__T5Tokenizer as T5Tokenizer, __webpack_exports__TableTransformerForObjectDetection as TableTransformerForObjectDetection, __webpack_exports__TableTransformerModel as TableTransformerModel, __webpack_exports__TableTransformerObjectDetectionOutput as TableTransformerObjectDetectionOutput, __webpack_exports__TableTransformerPreTrainedModel as TableTransformerPreTrainedModel, __webpack_exports__TemperatureLogitsWarper as TemperatureLogitsWarper, __webpack_exports__Tensor as Tensor, __webpack_exports__Text2TextGenerationPipeline as Text2TextGenerationPipeline, __webpack_exports__TextClassificationPipeline as TextClassificationPipeline, __webpack_exports__TextGenerationPipeline as TextGenerationPipeline, __webpack_exports__TextStreamer as TextStreamer, __webpack_exports__TextToAudioPipeline as TextToAudioPipeline, __webpack_exports__TokenClassificationPipeline as TokenClassificationPipeline, __webpack_exports__TokenClassifierOutput as TokenClassifierOutput, __webpack_exports__TokenizerModel as TokenizerModel, __webpack_exports__TopKLogitsWarper as TopKLogitsWarper, __webpack_exports__TopPLogitsWarper as TopPLogitsWarper, __webpack_exports__TrOCRForCausalLM as TrOCRForCausalLM, __webpack_exports__TrOCRPreTrainedModel as TrOCRPreTrainedModel, __webpack_exports__TranslationPipeline as TranslationPipeline, __webpack_exports__UniSpeechForCTC as UniSpeechForCTC, __webpack_exports__UniSpeechForSequenceClassification as UniSpeechForSequenceClassification, __webpack_exports__UniSpeechModel as UniSpeechModel, __webpack_exports__UniSpeechPreTrainedModel as UniSpeechPreTrainedModel, __webpack_exports__UniSpeechSatForAudioFrameClassification as UniSpeechSatForAudioFrameClassification, __webpack_exports__UniSpeechSatForCTC as UniSpeechSatForCTC, __webpack_exports__UniSpeechSatForSequenceClassification as UniSpeechSatForSequenceClassification, __webpack_exports__UniSpeechSatModel as UniSpeechSatModel, __webpack_exports__UniSpeechSatPreTrainedModel as UniSpeechSatPreTrainedModel, __webpack_exports__VLChatProcessor as VLChatProcessor, __webpack_exports__VLMImageProcessor as VLMImageProcessor, __webpack_exports__ViTFeatureExtractor as ViTFeatureExtractor, __webpack_exports__ViTForImageClassification as ViTForImageClassification, __webpack_exports__ViTImageProcessor as ViTImageProcessor, __webpack_exports__ViTMAEModel as ViTMAEModel, __webpack_exports__ViTMAEPreTrainedModel as ViTMAEPreTrainedModel, __webpack_exports__ViTMSNForImageClassification as ViTMSNForImageClassification, __webpack_exports__ViTMSNModel as ViTMSNModel, __webpack_exports__ViTMSNPreTrainedModel as ViTMSNPreTrainedModel, __webpack_exports__ViTModel as ViTModel, __webpack_exports__ViTPreTrainedModel as ViTPreTrainedModel, __webpack_exports__VisionEncoderDecoderModel as VisionEncoderDecoderModel, __webpack_exports__VitMatteForImageMatting as VitMatteForImageMatting, __webpack_exports__VitMatteImageProcessor as VitMatteImageProcessor, __webpack_exports__VitMattePreTrainedModel as VitMattePreTrainedModel, __webpack_exports__VitPoseForPoseEstimation as VitPoseForPoseEstimation, __webpack_exports__VitPoseImageProcessor as VitPoseImageProcessor, __webpack_exports__VitPosePreTrainedModel as VitPosePreTrainedModel, __webpack_exports__VitsModel as VitsModel, __webpack_exports__VitsModelOutput as VitsModelOutput, __webpack_exports__VitsPreTrainedModel as VitsPreTrainedModel, __webpack_exports__VitsTokenizer as VitsTokenizer, __webpack_exports__Wav2Vec2BertForCTC as Wav2Vec2BertForCTC, __webpack_exports__Wav2Vec2BertForSequenceClassification as Wav2Vec2BertForSequenceClassification, __webpack_exports__Wav2Vec2BertModel as Wav2Vec2BertModel, __webpack_exports__Wav2Vec2BertPreTrainedModel as Wav2Vec2BertPreTrainedModel, __webpack_exports__Wav2Vec2CTCTokenizer as Wav2Vec2CTCTokenizer, __webpack_exports__Wav2Vec2FeatureExtractor as Wav2Vec2FeatureExtractor, __webpack_exports__Wav2Vec2ForAudioFrameClassification as Wav2Vec2ForAudioFrameClassification, __webpack_exports__Wav2Vec2ForCTC as Wav2Vec2ForCTC, __webpack_exports__Wav2Vec2ForSequenceClassification as Wav2Vec2ForSequenceClassification, __webpack_exports__Wav2Vec2Model as Wav2Vec2Model, __webpack_exports__Wav2Vec2PreTrainedModel as Wav2Vec2PreTrainedModel, __webpack_exports__Wav2Vec2ProcessorWithLM as Wav2Vec2ProcessorWithLM, __webpack_exports__WavLMForAudioFrameClassification as WavLMForAudioFrameClassification, __webpack_exports__WavLMForCTC as WavLMForCTC, __webpack_exports__WavLMForSequenceClassification as WavLMForSequenceClassification, __webpack_exports__WavLMForXVector as WavLMForXVector, __webpack_exports__WavLMModel as WavLMModel, __webpack_exports__WavLMPreTrainedModel as WavLMPreTrainedModel, __webpack_exports__WeSpeakerFeatureExtractor as WeSpeakerFeatureExtractor, __webpack_exports__WeSpeakerResNetModel as WeSpeakerResNetModel, __webpack_exports__WeSpeakerResNetPreTrainedModel as WeSpeakerResNetPreTrainedModel, __webpack_exports__WhisperFeatureExtractor as WhisperFeatureExtractor, __webpack_exports__WhisperForConditionalGeneration as WhisperForConditionalGeneration, __webpack_exports__WhisperModel as WhisperModel, __webpack_exports__WhisperPreTrainedModel as WhisperPreTrainedModel, __webpack_exports__WhisperProcessor as WhisperProcessor, __webpack_exports__WhisperTextStreamer as WhisperTextStreamer, __webpack_exports__WhisperTimeStampLogitsProcessor as WhisperTimeStampLogitsProcessor, __webpack_exports__WhisperTokenizer as WhisperTokenizer, __webpack_exports__XLMForQuestionAnswering as XLMForQuestionAnswering, __webpack_exports__XLMForSequenceClassification as XLMForSequenceClassification, __webpack_exports__XLMForTokenClassification as XLMForTokenClassification, __webpack_exports__XLMModel as XLMModel, __webpack_exports__XLMPreTrainedModel as XLMPreTrainedModel, __webpack_exports__XLMRobertaForMaskedLM as XLMRobertaForMaskedLM, __webpack_exports__XLMRobertaForQuestionAnswering as XLMRobertaForQuestionAnswering, __webpack_exports__XLMRobertaForSequenceClassification as XLMRobertaForSequenceClassification, __webpack_exports__XLMRobertaForTokenClassification as XLMRobertaForTokenClassification, __webpack_exports__XLMRobertaModel as XLMRobertaModel, __webpack_exports__XLMRobertaPreTrainedModel as XLMRobertaPreTrainedModel, __webpack_exports__XLMRobertaTokenizer as XLMRobertaTokenizer, __webpack_exports__XLMTokenizer as XLMTokenizer, __webpack_exports__XLMWithLMHeadModel as XLMWithLMHeadModel, __webpack_exports__XVectorOutput as XVectorOutput, __webpack_exports__YolosFeatureExtractor as YolosFeatureExtractor, __webpack_exports__YolosForObjectDetection as YolosForObjectDetection, __webpack_exports__YolosImageProcessor as YolosImageProcessor, __webpack_exports__YolosModel as YolosModel, __webpack_exports__YolosObjectDetectionOutput as YolosObjectDetectionOutput, __webpack_exports__YolosPreTrainedModel as YolosPreTrainedModel, __webpack_exports__ZeroShotAudioClassificationPipeline as ZeroShotAudioClassificationPipeline, __webpack_exports__ZeroShotClassificationPipeline as ZeroShotClassificationPipeline, __webpack_exports__ZeroShotImageClassificationPipeline as ZeroShotImageClassificationPipeline, __webpack_exports__ZeroShotObjectDetectionPipeline as ZeroShotObjectDetectionPipeline, __webpack_exports__bankers_round as bankers_round, __webpack_exports__cat as cat, __webpack_exports__cos_sim as cos_sim, __webpack_exports__dot as dot, __webpack_exports__dynamic_time_warping as dynamic_time_warping, __webpack_exports__env as env, __webpack_exports__full as full, __webpack_exports__full_like as full_like, __webpack_exports__getKeyValueShapes as getKeyValueShapes, __webpack_exports__hamming as hamming, __webpack_exports__hanning as hanning, __webpack_exports__interpolate as interpolate, __webpack_exports__interpolate_4d as interpolate_4d, __webpack_exports__interpolate_data as interpolate_data, __webpack_exports__is_chinese_char as is_chinese_char, __webpack_exports__layer_norm as layer_norm, __webpack_exports__load_image as load_image, __webpack_exports__log_softmax as log_softmax, __webpack_exports__magnitude as magnitude, __webpack_exports__matmul as matmul, __webpack_exports__max as max, __webpack_exports__mean as mean, __webpack_exports__mean_pooling as mean_pooling, __webpack_exports__medianFilter as medianFilter, __webpack_exports__mel_filter_bank as mel_filter_bank, __webpack_exports__min as min, __webpack_exports__ones as ones, __webpack_exports__ones_like as ones_like, __webpack_exports__permute as permute, __webpack_exports__permute_data as permute_data, __webpack_exports__pipeline as pipeline, __webpack_exports__quantize_embeddings as quantize_embeddings, __webpack_exports__rand as rand, __webpack_exports__read_audio as read_audio, __webpack_exports__rfft as rfft, __webpack_exports__round as round, __webpack_exports__slice as slice, __webpack_exports__softmax as softmax, __webpack_exports__spectrogram as spectrogram, __webpack_exports__stack as stack, __webpack_exports__std_mean as std_mean, __webpack_exports__topk as topk, __webpack_exports__window_function as window_function, __webpack_exports__zeros as zeros, __webpack_exports__zeros_like as zeros_like };
+export { __webpack_exports__ASTFeatureExtractor as ASTFeatureExtractor, __webpack_exports__ASTForAudioClassification as ASTForAudioClassification, __webpack_exports__ASTModel as ASTModel, __webpack_exports__ASTPreTrainedModel as ASTPreTrainedModel, __webpack_exports__AlbertForMaskedLM as AlbertForMaskedLM, __webpack_exports__AlbertForQuestionAnswering as AlbertForQuestionAnswering, __webpack_exports__AlbertForSequenceClassification as AlbertForSequenceClassification, __webpack_exports__AlbertModel as AlbertModel, __webpack_exports__AlbertPreTrainedModel as AlbertPreTrainedModel, __webpack_exports__AlbertTokenizer as AlbertTokenizer, __webpack_exports__AudioClassificationPipeline as AudioClassificationPipeline, __webpack_exports__AutoConfig as AutoConfig, __webpack_exports__AutoFeatureExtractor as AutoFeatureExtractor, __webpack_exports__AutoImageProcessor as AutoImageProcessor, __webpack_exports__AutoModel as AutoModel, __webpack_exports__AutoModelForAudioClassification as AutoModelForAudioClassification, __webpack_exports__AutoModelForAudioFrameClassification as AutoModelForAudioFrameClassification, __webpack_exports__AutoModelForCTC as AutoModelForCTC, __webpack_exports__AutoModelForCausalLM as AutoModelForCausalLM, __webpack_exports__AutoModelForDepthEstimation as AutoModelForDepthEstimation, __webpack_exports__AutoModelForDocumentQuestionAnswering as AutoModelForDocumentQuestionAnswering, __webpack_exports__AutoModelForImageClassification as AutoModelForImageClassification, __webpack_exports__AutoModelForImageFeatureExtraction as AutoModelForImageFeatureExtraction, __webpack_exports__AutoModelForImageMatting as AutoModelForImageMatting, __webpack_exports__AutoModelForImageSegmentation as AutoModelForImageSegmentation, __webpack_exports__AutoModelForImageToImage as AutoModelForImageToImage, __webpack_exports__AutoModelForMaskGeneration as AutoModelForMaskGeneration, __webpack_exports__AutoModelForMaskedLM as AutoModelForMaskedLM, __webpack_exports__AutoModelForNormalEstimation as AutoModelForNormalEstimation, __webpack_exports__AutoModelForObjectDetection as AutoModelForObjectDetection, __webpack_exports__AutoModelForPoseEstimation as AutoModelForPoseEstimation, __webpack_exports__AutoModelForQuestionAnswering as AutoModelForQuestionAnswering, __webpack_exports__AutoModelForSemanticSegmentation as AutoModelForSemanticSegmentation, __webpack_exports__AutoModelForSeq2SeqLM as AutoModelForSeq2SeqLM, __webpack_exports__AutoModelForSequenceClassification as AutoModelForSequenceClassification, __webpack_exports__AutoModelForSpeechSeq2Seq as AutoModelForSpeechSeq2Seq, __webpack_exports__AutoModelForTextToSpectrogram as AutoModelForTextToSpectrogram, __webpack_exports__AutoModelForTextToWaveform as AutoModelForTextToWaveform, __webpack_exports__AutoModelForTokenClassification as AutoModelForTokenClassification, __webpack_exports__AutoModelForUniversalSegmentation as AutoModelForUniversalSegmentation, __webpack_exports__AutoModelForVision2Seq as AutoModelForVision2Seq, __webpack_exports__AutoModelForXVector as AutoModelForXVector, __webpack_exports__AutoModelForZeroShotObjectDetection as AutoModelForZeroShotObjectDetection, __webpack_exports__AutoProcessor as AutoProcessor, __webpack_exports__AutoTokenizer as AutoTokenizer, __webpack_exports__AutomaticSpeechRecognitionPipeline as AutomaticSpeechRecognitionPipeline, __webpack_exports__BartForConditionalGeneration as BartForConditionalGeneration, __webpack_exports__BartForSequenceClassification as BartForSequenceClassification, __webpack_exports__BartModel as BartModel, __webpack_exports__BartPretrainedModel as BartPretrainedModel, __webpack_exports__BartTokenizer as BartTokenizer, __webpack_exports__BaseModelOutput as BaseModelOutput, __webpack_exports__BaseStreamer as BaseStreamer, __webpack_exports__BeitFeatureExtractor as BeitFeatureExtractor, __webpack_exports__BeitForImageClassification as BeitForImageClassification, __webpack_exports__BeitModel as BeitModel, __webpack_exports__BeitPreTrainedModel as BeitPreTrainedModel, __webpack_exports__BertForMaskedLM as BertForMaskedLM, __webpack_exports__BertForQuestionAnswering as BertForQuestionAnswering, __webpack_exports__BertForSequenceClassification as BertForSequenceClassification, __webpack_exports__BertForTokenClassification as BertForTokenClassification, __webpack_exports__BertModel as BertModel, __webpack_exports__BertPreTrainedModel as BertPreTrainedModel, __webpack_exports__BertTokenizer as BertTokenizer, __webpack_exports__BitImageProcessor as BitImageProcessor, __webpack_exports__BlenderbotForConditionalGeneration as BlenderbotForConditionalGeneration, __webpack_exports__BlenderbotModel as BlenderbotModel, __webpack_exports__BlenderbotPreTrainedModel as BlenderbotPreTrainedModel, __webpack_exports__BlenderbotSmallForConditionalGeneration as BlenderbotSmallForConditionalGeneration, __webpack_exports__BlenderbotSmallModel as BlenderbotSmallModel, __webpack_exports__BlenderbotSmallPreTrainedModel as BlenderbotSmallPreTrainedModel, __webpack_exports__BlenderbotSmallTokenizer as BlenderbotSmallTokenizer, __webpack_exports__BlenderbotTokenizer as BlenderbotTokenizer, __webpack_exports__BloomForCausalLM as BloomForCausalLM, __webpack_exports__BloomModel as BloomModel, __webpack_exports__BloomPreTrainedModel as BloomPreTrainedModel, __webpack_exports__BloomTokenizer as BloomTokenizer, __webpack_exports__CLIPFeatureExtractor as CLIPFeatureExtractor, __webpack_exports__CLIPImageProcessor as CLIPImageProcessor, __webpack_exports__CLIPModel as CLIPModel, __webpack_exports__CLIPPreTrainedModel as CLIPPreTrainedModel, __webpack_exports__CLIPSegForImageSegmentation as CLIPSegForImageSegmentation, __webpack_exports__CLIPSegModel as CLIPSegModel, __webpack_exports__CLIPSegPreTrainedModel as CLIPSegPreTrainedModel, __webpack_exports__CLIPTextModel as CLIPTextModel, __webpack_exports__CLIPTextModelWithProjection as CLIPTextModelWithProjection, __webpack_exports__CLIPTokenizer as CLIPTokenizer, __webpack_exports__CLIPVisionModel as CLIPVisionModel, __webpack_exports__CLIPVisionModelWithProjection as CLIPVisionModelWithProjection, __webpack_exports__CamembertForMaskedLM as CamembertForMaskedLM, __webpack_exports__CamembertForQuestionAnswering as CamembertForQuestionAnswering, __webpack_exports__CamembertForSequenceClassification as CamembertForSequenceClassification, __webpack_exports__CamembertForTokenClassification as CamembertForTokenClassification, __webpack_exports__CamembertModel as CamembertModel, __webpack_exports__CamembertPreTrainedModel as CamembertPreTrainedModel, __webpack_exports__CamembertTokenizer as CamembertTokenizer, __webpack_exports__CausalLMOutput as CausalLMOutput, __webpack_exports__CausalLMOutputWithPast as CausalLMOutputWithPast, __webpack_exports__ChineseCLIPFeatureExtractor as ChineseCLIPFeatureExtractor, __webpack_exports__ChineseCLIPModel as ChineseCLIPModel, __webpack_exports__ChineseCLIPPreTrainedModel as ChineseCLIPPreTrainedModel, __webpack_exports__ClapAudioModelWithProjection as ClapAudioModelWithProjection, __webpack_exports__ClapFeatureExtractor as ClapFeatureExtractor, __webpack_exports__ClapModel as ClapModel, __webpack_exports__ClapPreTrainedModel as ClapPreTrainedModel, __webpack_exports__ClapTextModelWithProjection as ClapTextModelWithProjection, __webpack_exports__ClassifierFreeGuidanceLogitsProcessor as ClassifierFreeGuidanceLogitsProcessor, __webpack_exports__CodeGenForCausalLM as CodeGenForCausalLM, __webpack_exports__CodeGenModel as CodeGenModel, __webpack_exports__CodeGenPreTrainedModel as CodeGenPreTrainedModel, __webpack_exports__CodeGenTokenizer as CodeGenTokenizer, __webpack_exports__CodeLlamaTokenizer as CodeLlamaTokenizer, __webpack_exports__CohereForCausalLM as CohereForCausalLM, __webpack_exports__CohereModel as CohereModel, __webpack_exports__CoherePreTrainedModel as CoherePreTrainedModel, __webpack_exports__CohereTokenizer as CohereTokenizer, __webpack_exports__ConvBertForMaskedLM as ConvBertForMaskedLM, __webpack_exports__ConvBertForQuestionAnswering as ConvBertForQuestionAnswering, __webpack_exports__ConvBertForSequenceClassification as ConvBertForSequenceClassification, __webpack_exports__ConvBertForTokenClassification as ConvBertForTokenClassification, __webpack_exports__ConvBertModel as ConvBertModel, __webpack_exports__ConvBertPreTrainedModel as ConvBertPreTrainedModel, __webpack_exports__ConvBertTokenizer as ConvBertTokenizer, __webpack_exports__ConvNextFeatureExtractor as ConvNextFeatureExtractor, __webpack_exports__ConvNextForImageClassification as ConvNextForImageClassification, __webpack_exports__ConvNextImageProcessor as ConvNextImageProcessor, __webpack_exports__ConvNextModel as ConvNextModel, __webpack_exports__ConvNextPreTrainedModel as ConvNextPreTrainedModel, __webpack_exports__ConvNextV2ForImageClassification as ConvNextV2ForImageClassification, __webpack_exports__ConvNextV2Model as ConvNextV2Model, __webpack_exports__ConvNextV2PreTrainedModel as ConvNextV2PreTrainedModel, __webpack_exports__DPTFeatureExtractor as DPTFeatureExtractor, __webpack_exports__DPTForDepthEstimation as DPTForDepthEstimation, __webpack_exports__DPTImageProcessor as DPTImageProcessor, __webpack_exports__DPTModel as DPTModel, __webpack_exports__DPTPreTrainedModel as DPTPreTrainedModel, __webpack_exports__DebertaForMaskedLM as DebertaForMaskedLM, __webpack_exports__DebertaForQuestionAnswering as DebertaForQuestionAnswering, __webpack_exports__DebertaForSequenceClassification as DebertaForSequenceClassification, __webpack_exports__DebertaForTokenClassification as DebertaForTokenClassification, __webpack_exports__DebertaModel as DebertaModel, __webpack_exports__DebertaPreTrainedModel as DebertaPreTrainedModel, __webpack_exports__DebertaTokenizer as DebertaTokenizer, __webpack_exports__DebertaV2ForMaskedLM as DebertaV2ForMaskedLM, __webpack_exports__DebertaV2ForQuestionAnswering as DebertaV2ForQuestionAnswering, __webpack_exports__DebertaV2ForSequenceClassification as DebertaV2ForSequenceClassification, __webpack_exports__DebertaV2ForTokenClassification as DebertaV2ForTokenClassification, __webpack_exports__DebertaV2Model as DebertaV2Model, __webpack_exports__DebertaV2PreTrainedModel as DebertaV2PreTrainedModel, __webpack_exports__DebertaV2Tokenizer as DebertaV2Tokenizer, __webpack_exports__DecisionTransformerModel as DecisionTransformerModel, __webpack_exports__DecisionTransformerPreTrainedModel as DecisionTransformerPreTrainedModel, __webpack_exports__DeiTFeatureExtractor as DeiTFeatureExtractor, __webpack_exports__DeiTForImageClassification as DeiTForImageClassification, __webpack_exports__DeiTImageProcessor as DeiTImageProcessor, __webpack_exports__DeiTModel as DeiTModel, __webpack_exports__DeiTPreTrainedModel as DeiTPreTrainedModel, __webpack_exports__DepthAnythingForDepthEstimation as DepthAnythingForDepthEstimation, __webpack_exports__DepthAnythingPreTrainedModel as DepthAnythingPreTrainedModel, __webpack_exports__DepthEstimationPipeline as DepthEstimationPipeline, __webpack_exports__DepthProForDepthEstimation as DepthProForDepthEstimation, __webpack_exports__DepthProPreTrainedModel as DepthProPreTrainedModel, __webpack_exports__DetrFeatureExtractor as DetrFeatureExtractor, __webpack_exports__DetrForObjectDetection as DetrForObjectDetection, __webpack_exports__DetrForSegmentation as DetrForSegmentation, __webpack_exports__DetrImageProcessor as DetrImageProcessor, __webpack_exports__DetrModel as DetrModel, __webpack_exports__DetrObjectDetectionOutput as DetrObjectDetectionOutput, __webpack_exports__DetrPreTrainedModel as DetrPreTrainedModel, __webpack_exports__DetrSegmentationOutput as DetrSegmentationOutput, __webpack_exports__Dinov2ForImageClassification as Dinov2ForImageClassification, __webpack_exports__Dinov2Model as Dinov2Model, __webpack_exports__Dinov2PreTrainedModel as Dinov2PreTrainedModel, __webpack_exports__Dinov2WithRegistersForImageClassification as Dinov2WithRegistersForImageClassification, __webpack_exports__Dinov2WithRegistersModel as Dinov2WithRegistersModel, __webpack_exports__Dinov2WithRegistersPreTrainedModel as Dinov2WithRegistersPreTrainedModel, __webpack_exports__DistilBertForMaskedLM as DistilBertForMaskedLM, __webpack_exports__DistilBertForQuestionAnswering as DistilBertForQuestionAnswering, __webpack_exports__DistilBertForSequenceClassification as DistilBertForSequenceClassification, __webpack_exports__DistilBertForTokenClassification as DistilBertForTokenClassification, __webpack_exports__DistilBertModel as DistilBertModel, __webpack_exports__DistilBertPreTrainedModel as DistilBertPreTrainedModel, __webpack_exports__DistilBertTokenizer as DistilBertTokenizer, __webpack_exports__DocumentQuestionAnsweringPipeline as DocumentQuestionAnsweringPipeline, __webpack_exports__DonutFeatureExtractor as DonutFeatureExtractor, __webpack_exports__DonutImageProcessor as DonutImageProcessor, __webpack_exports__DonutSwinModel as DonutSwinModel, __webpack_exports__DonutSwinPreTrainedModel as DonutSwinPreTrainedModel, __webpack_exports__EfficientNetForImageClassification as EfficientNetForImageClassification, __webpack_exports__EfficientNetImageProcessor as EfficientNetImageProcessor, __webpack_exports__EfficientNetModel as EfficientNetModel, __webpack_exports__EfficientNetPreTrainedModel as EfficientNetPreTrainedModel, __webpack_exports__ElectraForMaskedLM as ElectraForMaskedLM, __webpack_exports__ElectraForQuestionAnswering as ElectraForQuestionAnswering, __webpack_exports__ElectraForSequenceClassification as ElectraForSequenceClassification, __webpack_exports__ElectraForTokenClassification as ElectraForTokenClassification, __webpack_exports__ElectraModel as ElectraModel, __webpack_exports__ElectraPreTrainedModel as ElectraPreTrainedModel, __webpack_exports__ElectraTokenizer as ElectraTokenizer, __webpack_exports__EosTokenCriteria as EosTokenCriteria, __webpack_exports__EsmForMaskedLM as EsmForMaskedLM, __webpack_exports__EsmForSequenceClassification as EsmForSequenceClassification, __webpack_exports__EsmForTokenClassification as EsmForTokenClassification, __webpack_exports__EsmModel as EsmModel, __webpack_exports__EsmPreTrainedModel as EsmPreTrainedModel, __webpack_exports__EsmTokenizer as EsmTokenizer, __webpack_exports__ExaoneForCausalLM as ExaoneForCausalLM, __webpack_exports__ExaoneModel as ExaoneModel, __webpack_exports__ExaonePreTrainedModel as ExaonePreTrainedModel, __webpack_exports__FFT as FFT, __webpack_exports__FalconForCausalLM as FalconForCausalLM, __webpack_exports__FalconModel as FalconModel, __webpack_exports__FalconPreTrainedModel as FalconPreTrainedModel, __webpack_exports__FalconTokenizer as FalconTokenizer, __webpack_exports__FastViTForImageClassification as FastViTForImageClassification, __webpack_exports__FastViTModel as FastViTModel, __webpack_exports__FastViTPreTrainedModel as FastViTPreTrainedModel, __webpack_exports__FeatureExtractionPipeline as FeatureExtractionPipeline, __webpack_exports__FeatureExtractor as FeatureExtractor, __webpack_exports__FillMaskPipeline as FillMaskPipeline, __webpack_exports__Florence2ForConditionalGeneration as Florence2ForConditionalGeneration, __webpack_exports__Florence2PreTrainedModel as Florence2PreTrainedModel, __webpack_exports__Florence2Processor as Florence2Processor, __webpack_exports__ForcedBOSTokenLogitsProcessor as ForcedBOSTokenLogitsProcessor, __webpack_exports__ForcedEOSTokenLogitsProcessor as ForcedEOSTokenLogitsProcessor, __webpack_exports__GLPNFeatureExtractor as GLPNFeatureExtractor, __webpack_exports__GLPNForDepthEstimation as GLPNForDepthEstimation, __webpack_exports__GLPNModel as GLPNModel, __webpack_exports__GLPNPreTrainedModel as GLPNPreTrainedModel, __webpack_exports__GPT2LMHeadModel as GPT2LMHeadModel, __webpack_exports__GPT2Model as GPT2Model, __webpack_exports__GPT2PreTrainedModel as GPT2PreTrainedModel, __webpack_exports__GPT2Tokenizer as GPT2Tokenizer, __webpack_exports__GPTBigCodeForCausalLM as GPTBigCodeForCausalLM, __webpack_exports__GPTBigCodeModel as GPTBigCodeModel, __webpack_exports__GPTBigCodePreTrainedModel as GPTBigCodePreTrainedModel, __webpack_exports__GPTJForCausalLM as GPTJForCausalLM, __webpack_exports__GPTJModel as GPTJModel, __webpack_exports__GPTJPreTrainedModel as GPTJPreTrainedModel, __webpack_exports__GPTNeoForCausalLM as GPTNeoForCausalLM, __webpack_exports__GPTNeoModel as GPTNeoModel, __webpack_exports__GPTNeoPreTrainedModel as GPTNeoPreTrainedModel, __webpack_exports__GPTNeoXForCausalLM as GPTNeoXForCausalLM, __webpack_exports__GPTNeoXModel as GPTNeoXModel, __webpack_exports__GPTNeoXPreTrainedModel as GPTNeoXPreTrainedModel, __webpack_exports__GPTNeoXTokenizer as GPTNeoXTokenizer, __webpack_exports__Gemma2ForCausalLM as Gemma2ForCausalLM, __webpack_exports__Gemma2Model as Gemma2Model, __webpack_exports__Gemma2PreTrainedModel as Gemma2PreTrainedModel, __webpack_exports__GemmaForCausalLM as GemmaForCausalLM, __webpack_exports__GemmaModel as GemmaModel, __webpack_exports__GemmaPreTrainedModel as GemmaPreTrainedModel, __webpack_exports__GemmaTokenizer as GemmaTokenizer, __webpack_exports__GlmForCausalLM as GlmForCausalLM, __webpack_exports__GlmModel as GlmModel, __webpack_exports__GlmPreTrainedModel as GlmPreTrainedModel, __webpack_exports__GraniteForCausalLM as GraniteForCausalLM, __webpack_exports__GraniteModel as GraniteModel, __webpack_exports__GranitePreTrainedModel as GranitePreTrainedModel, __webpack_exports__Grok1Tokenizer as Grok1Tokenizer, __webpack_exports__GroundingDinoForObjectDetection as GroundingDinoForObjectDetection, __webpack_exports__GroundingDinoImageProcessor as GroundingDinoImageProcessor, __webpack_exports__GroundingDinoPreTrainedModel as GroundingDinoPreTrainedModel, __webpack_exports__GroundingDinoProcessor as GroundingDinoProcessor, __webpack_exports__GroupViTModel as GroupViTModel, __webpack_exports__GroupViTPreTrainedModel as GroupViTPreTrainedModel, __webpack_exports__HeliumForCausalLM as HeliumForCausalLM, __webpack_exports__HeliumModel as HeliumModel, __webpack_exports__HeliumPreTrainedModel as HeliumPreTrainedModel, __webpack_exports__HerbertTokenizer as HerbertTokenizer, __webpack_exports__HieraForImageClassification as HieraForImageClassification, __webpack_exports__HieraModel as HieraModel, __webpack_exports__HieraPreTrainedModel as HieraPreTrainedModel, __webpack_exports__HubertForCTC as HubertForCTC, __webpack_exports__HubertForSequenceClassification as HubertForSequenceClassification, __webpack_exports__HubertModel as HubertModel, __webpack_exports__HubertPreTrainedModel as HubertPreTrainedModel, __webpack_exports__IJepaForImageClassification as IJepaForImageClassification, __webpack_exports__IJepaModel as IJepaModel, __webpack_exports__IJepaPreTrainedModel as IJepaPreTrainedModel, __webpack_exports__Idefics3ForConditionalGeneration as Idefics3ForConditionalGeneration, __webpack_exports__Idefics3ImageProcessor as Idefics3ImageProcessor, __webpack_exports__Idefics3PreTrainedModel as Idefics3PreTrainedModel, __webpack_exports__Idefics3Processor as Idefics3Processor, __webpack_exports__ImageClassificationPipeline as ImageClassificationPipeline, __webpack_exports__ImageFeatureExtractionPipeline as ImageFeatureExtractionPipeline, __webpack_exports__ImageFeatureExtractor as ImageFeatureExtractor, __webpack_exports__ImageMattingOutput as ImageMattingOutput, __webpack_exports__ImageProcessor as ImageProcessor, __webpack_exports__ImageSegmentationPipeline as ImageSegmentationPipeline, __webpack_exports__ImageToImagePipeline as ImageToImagePipeline, __webpack_exports__ImageToTextPipeline as ImageToTextPipeline, __webpack_exports__InterruptableStoppingCriteria as InterruptableStoppingCriteria, __webpack_exports__JAISLMHeadModel as JAISLMHeadModel, __webpack_exports__JAISModel as JAISModel, __webpack_exports__JAISPreTrainedModel as JAISPreTrainedModel, __webpack_exports__JinaCLIPImageProcessor as JinaCLIPImageProcessor, __webpack_exports__JinaCLIPModel as JinaCLIPModel, __webpack_exports__JinaCLIPPreTrainedModel as JinaCLIPPreTrainedModel, __webpack_exports__JinaCLIPProcessor as JinaCLIPProcessor, __webpack_exports__JinaCLIPTextModel as JinaCLIPTextModel, __webpack_exports__JinaCLIPVisionModel as JinaCLIPVisionModel, __webpack_exports__LlamaForCausalLM as LlamaForCausalLM, __webpack_exports__LlamaModel as LlamaModel, __webpack_exports__LlamaPreTrainedModel as LlamaPreTrainedModel, __webpack_exports__LlamaTokenizer as LlamaTokenizer, __webpack_exports__LlavaForConditionalGeneration as LlavaForConditionalGeneration, __webpack_exports__LlavaOnevisionForConditionalGeneration as LlavaOnevisionForConditionalGeneration, __webpack_exports__LlavaOnevisionImageProcessor as LlavaOnevisionImageProcessor, __webpack_exports__LlavaPreTrainedModel as LlavaPreTrainedModel, __webpack_exports__LogitsProcessor as LogitsProcessor, __webpack_exports__LogitsProcessorList as LogitsProcessorList, __webpack_exports__LogitsWarper as LogitsWarper, __webpack_exports__LongT5ForConditionalGeneration as LongT5ForConditionalGeneration, __webpack_exports__LongT5Model as LongT5Model, __webpack_exports__LongT5PreTrainedModel as LongT5PreTrainedModel, __webpack_exports__M2M100ForConditionalGeneration as M2M100ForConditionalGeneration, __webpack_exports__M2M100Model as M2M100Model, __webpack_exports__M2M100PreTrainedModel as M2M100PreTrainedModel, __webpack_exports__M2M100Tokenizer as M2M100Tokenizer, __webpack_exports__MBart50Tokenizer as MBart50Tokenizer, __webpack_exports__MBartForCausalLM as MBartForCausalLM, __webpack_exports__MBartForConditionalGeneration as MBartForConditionalGeneration, __webpack_exports__MBartForSequenceClassification as MBartForSequenceClassification, __webpack_exports__MBartModel as MBartModel, __webpack_exports__MBartPreTrainedModel as MBartPreTrainedModel, __webpack_exports__MBartTokenizer as MBartTokenizer, __webpack_exports__MPNetForMaskedLM as MPNetForMaskedLM, __webpack_exports__MPNetForQuestionAnswering as MPNetForQuestionAnswering, __webpack_exports__MPNetForSequenceClassification as MPNetForSequenceClassification, __webpack_exports__MPNetForTokenClassification as MPNetForTokenClassification, __webpack_exports__MPNetModel as MPNetModel, __webpack_exports__MPNetPreTrainedModel as MPNetPreTrainedModel, __webpack_exports__MPNetTokenizer as MPNetTokenizer, __webpack_exports__MT5ForConditionalGeneration as MT5ForConditionalGeneration, __webpack_exports__MT5Model as MT5Model, __webpack_exports__MT5PreTrainedModel as MT5PreTrainedModel, __webpack_exports__MarianMTModel as MarianMTModel, __webpack_exports__MarianModel as MarianModel, __webpack_exports__MarianPreTrainedModel as MarianPreTrainedModel, __webpack_exports__MarianTokenizer as MarianTokenizer, __webpack_exports__Mask2FormerImageProcessor as Mask2FormerImageProcessor, __webpack_exports__MaskFormerFeatureExtractor as MaskFormerFeatureExtractor, __webpack_exports__MaskFormerForInstanceSegmentation as MaskFormerForInstanceSegmentation, __webpack_exports__MaskFormerImageProcessor as MaskFormerImageProcessor, __webpack_exports__MaskFormerModel as MaskFormerModel, __webpack_exports__MaskFormerPreTrainedModel as MaskFormerPreTrainedModel, __webpack_exports__MaskedLMOutput as MaskedLMOutput, __webpack_exports__MaxLengthCriteria as MaxLengthCriteria, __webpack_exports__MgpstrForSceneTextRecognition as MgpstrForSceneTextRecognition, __webpack_exports__MgpstrModelOutput as MgpstrModelOutput, __webpack_exports__MgpstrPreTrainedModel as MgpstrPreTrainedModel, __webpack_exports__MgpstrProcessor as MgpstrProcessor, __webpack_exports__MgpstrTokenizer as MgpstrTokenizer, __webpack_exports__MinLengthLogitsProcessor as MinLengthLogitsProcessor, __webpack_exports__MinNewTokensLengthLogitsProcessor as MinNewTokensLengthLogitsProcessor, __webpack_exports__MistralForCausalLM as MistralForCausalLM, __webpack_exports__MistralModel as MistralModel, __webpack_exports__MistralPreTrainedModel as MistralPreTrainedModel, __webpack_exports__MobileBertForMaskedLM as MobileBertForMaskedLM, __webpack_exports__MobileBertForQuestionAnswering as MobileBertForQuestionAnswering, __webpack_exports__MobileBertForSequenceClassification as MobileBertForSequenceClassification, __webpack_exports__MobileBertModel as MobileBertModel, __webpack_exports__MobileBertPreTrainedModel as MobileBertPreTrainedModel, __webpack_exports__MobileBertTokenizer as MobileBertTokenizer, __webpack_exports__MobileLLMForCausalLM as MobileLLMForCausalLM, __webpack_exports__MobileLLMModel as MobileLLMModel, __webpack_exports__MobileLLMPreTrainedModel as MobileLLMPreTrainedModel, __webpack_exports__MobileNetV1FeatureExtractor as MobileNetV1FeatureExtractor, __webpack_exports__MobileNetV1ForImageClassification as MobileNetV1ForImageClassification, __webpack_exports__MobileNetV1ImageProcessor as MobileNetV1ImageProcessor, __webpack_exports__MobileNetV1Model as MobileNetV1Model, __webpack_exports__MobileNetV1PreTrainedModel as MobileNetV1PreTrainedModel, __webpack_exports__MobileNetV2FeatureExtractor as MobileNetV2FeatureExtractor, __webpack_exports__MobileNetV2ForImageClassification as MobileNetV2ForImageClassification, __webpack_exports__MobileNetV2ImageProcessor as MobileNetV2ImageProcessor, __webpack_exports__MobileNetV2Model as MobileNetV2Model, __webpack_exports__MobileNetV2PreTrainedModel as MobileNetV2PreTrainedModel, __webpack_exports__MobileNetV3FeatureExtractor as MobileNetV3FeatureExtractor, __webpack_exports__MobileNetV3ForImageClassification as MobileNetV3ForImageClassification, __webpack_exports__MobileNetV3ImageProcessor as MobileNetV3ImageProcessor, __webpack_exports__MobileNetV3Model as MobileNetV3Model, __webpack_exports__MobileNetV3PreTrainedModel as MobileNetV3PreTrainedModel, __webpack_exports__MobileNetV4FeatureExtractor as MobileNetV4FeatureExtractor, __webpack_exports__MobileNetV4ForImageClassification as MobileNetV4ForImageClassification, __webpack_exports__MobileNetV4ImageProcessor as MobileNetV4ImageProcessor, __webpack_exports__MobileNetV4Model as MobileNetV4Model, __webpack_exports__MobileNetV4PreTrainedModel as MobileNetV4PreTrainedModel, __webpack_exports__MobileViTFeatureExtractor as MobileViTFeatureExtractor, __webpack_exports__MobileViTForImageClassification as MobileViTForImageClassification, __webpack_exports__MobileViTImageProcessor as MobileViTImageProcessor, __webpack_exports__MobileViTModel as MobileViTModel, __webpack_exports__MobileViTPreTrainedModel as MobileViTPreTrainedModel, __webpack_exports__MobileViTV2ForImageClassification as MobileViTV2ForImageClassification, __webpack_exports__MobileViTV2Model as MobileViTV2Model, __webpack_exports__MobileViTV2PreTrainedModel as MobileViTV2PreTrainedModel, __webpack_exports__ModelOutput as ModelOutput, __webpack_exports__ModernBertForMaskedLM as ModernBertForMaskedLM, __webpack_exports__ModernBertForSequenceClassification as ModernBertForSequenceClassification, __webpack_exports__ModernBertForTokenClassification as ModernBertForTokenClassification, __webpack_exports__ModernBertModel as ModernBertModel, __webpack_exports__ModernBertPreTrainedModel as ModernBertPreTrainedModel, __webpack_exports__Moondream1ForConditionalGeneration as Moondream1ForConditionalGeneration, __webpack_exports__MoonshineFeatureExtractor as MoonshineFeatureExtractor, __webpack_exports__MoonshineForConditionalGeneration as MoonshineForConditionalGeneration, __webpack_exports__MoonshineModel as MoonshineModel, __webpack_exports__MoonshinePreTrainedModel as MoonshinePreTrainedModel, __webpack_exports__MoonshineProcessor as MoonshineProcessor, __webpack_exports__MptForCausalLM as MptForCausalLM, __webpack_exports__MptModel as MptModel, __webpack_exports__MptPreTrainedModel as MptPreTrainedModel, __webpack_exports__MultiModalityCausalLM as MultiModalityCausalLM, __webpack_exports__MultiModalityPreTrainedModel as MultiModalityPreTrainedModel, __webpack_exports__MusicgenForCausalLM as MusicgenForCausalLM, __webpack_exports__MusicgenForConditionalGeneration as MusicgenForConditionalGeneration, __webpack_exports__MusicgenModel as MusicgenModel, __webpack_exports__MusicgenPreTrainedModel as MusicgenPreTrainedModel, __webpack_exports__NllbTokenizer as NllbTokenizer, __webpack_exports__NoBadWordsLogitsProcessor as NoBadWordsLogitsProcessor, __webpack_exports__NoRepeatNGramLogitsProcessor as NoRepeatNGramLogitsProcessor, __webpack_exports__NomicBertModel as NomicBertModel, __webpack_exports__NomicBertPreTrainedModel as NomicBertPreTrainedModel, __webpack_exports__NougatImageProcessor as NougatImageProcessor, __webpack_exports__NougatTokenizer as NougatTokenizer, __webpack_exports__OPTForCausalLM as OPTForCausalLM, __webpack_exports__OPTModel as OPTModel, __webpack_exports__OPTPreTrainedModel as OPTPreTrainedModel, __webpack_exports__ObjectDetectionPipeline as ObjectDetectionPipeline, __webpack_exports__Olmo2ForCausalLM as Olmo2ForCausalLM, __webpack_exports__Olmo2Model as Olmo2Model, __webpack_exports__Olmo2PreTrainedModel as Olmo2PreTrainedModel, __webpack_exports__OlmoForCausalLM as OlmoForCausalLM, __webpack_exports__OlmoModel as OlmoModel, __webpack_exports__OlmoPreTrainedModel as OlmoPreTrainedModel, __webpack_exports__OpenELMForCausalLM as OpenELMForCausalLM, __webpack_exports__OpenELMModel as OpenELMModel, __webpack_exports__OpenELMPreTrainedModel as OpenELMPreTrainedModel, __webpack_exports__OwlViTFeatureExtractor as OwlViTFeatureExtractor, __webpack_exports__OwlViTForObjectDetection as OwlViTForObjectDetection, __webpack_exports__OwlViTImageProcessor as OwlViTImageProcessor, __webpack_exports__OwlViTModel as OwlViTModel, __webpack_exports__OwlViTPreTrainedModel as OwlViTPreTrainedModel, __webpack_exports__OwlViTProcessor as OwlViTProcessor, __webpack_exports__Owlv2ForObjectDetection as Owlv2ForObjectDetection, __webpack_exports__Owlv2ImageProcessor as Owlv2ImageProcessor, __webpack_exports__Owlv2Model as Owlv2Model, __webpack_exports__Owlv2PreTrainedModel as Owlv2PreTrainedModel, __webpack_exports__PaliGemmaForConditionalGeneration as PaliGemmaForConditionalGeneration, __webpack_exports__PaliGemmaPreTrainedModel as PaliGemmaPreTrainedModel, __webpack_exports__PaliGemmaProcessor as PaliGemmaProcessor, __webpack_exports__PatchTSMixerForPrediction as PatchTSMixerForPrediction, __webpack_exports__PatchTSMixerModel as PatchTSMixerModel, __webpack_exports__PatchTSMixerPreTrainedModel as PatchTSMixerPreTrainedModel, __webpack_exports__PatchTSTForPrediction as PatchTSTForPrediction, __webpack_exports__PatchTSTModel as PatchTSTModel, __webpack_exports__PatchTSTPreTrainedModel as PatchTSTPreTrainedModel, __webpack_exports__Phi3ForCausalLM as Phi3ForCausalLM, __webpack_exports__Phi3Model as Phi3Model, __webpack_exports__Phi3PreTrainedModel as Phi3PreTrainedModel, __webpack_exports__Phi3VForCausalLM as Phi3VForCausalLM, __webpack_exports__Phi3VImageProcessor as Phi3VImageProcessor, __webpack_exports__Phi3VPreTrainedModel as Phi3VPreTrainedModel, __webpack_exports__Phi3VProcessor as Phi3VProcessor, __webpack_exports__PhiForCausalLM as PhiForCausalLM, __webpack_exports__PhiModel as PhiModel, __webpack_exports__PhiPreTrainedModel as PhiPreTrainedModel, __webpack_exports__Pipeline as Pipeline, __webpack_exports__PreTrainedModel as PreTrainedModel, __webpack_exports__PreTrainedTokenizer as PreTrainedTokenizer, __webpack_exports__PretrainedConfig as PretrainedConfig, __webpack_exports__PretrainedMixin as PretrainedMixin, __webpack_exports__Processor as Processor, __webpack_exports__PvtForImageClassification as PvtForImageClassification, __webpack_exports__PvtImageProcessor as PvtImageProcessor, __webpack_exports__PvtModel as PvtModel, __webpack_exports__PvtPreTrainedModel as PvtPreTrainedModel, __webpack_exports__PyAnnoteFeatureExtractor as PyAnnoteFeatureExtractor, __webpack_exports__PyAnnoteForAudioFrameClassification as PyAnnoteForAudioFrameClassification, __webpack_exports__PyAnnoteModel as PyAnnoteModel, __webpack_exports__PyAnnotePreTrainedModel as PyAnnotePreTrainedModel, __webpack_exports__PyAnnoteProcessor as PyAnnoteProcessor, __webpack_exports__QuestionAnsweringModelOutput as QuestionAnsweringModelOutput, __webpack_exports__QuestionAnsweringPipeline as QuestionAnsweringPipeline, __webpack_exports__Qwen2ForCausalLM as Qwen2ForCausalLM, __webpack_exports__Qwen2Model as Qwen2Model, __webpack_exports__Qwen2PreTrainedModel as Qwen2PreTrainedModel, __webpack_exports__Qwen2Tokenizer as Qwen2Tokenizer, __webpack_exports__Qwen2VLForConditionalGeneration as Qwen2VLForConditionalGeneration, __webpack_exports__Qwen2VLImageProcessor as Qwen2VLImageProcessor, __webpack_exports__Qwen2VLPreTrainedModel as Qwen2VLPreTrainedModel, __webpack_exports__Qwen2VLProcessor as Qwen2VLProcessor, __webpack_exports__RTDetrForObjectDetection as RTDetrForObjectDetection, __webpack_exports__RTDetrImageProcessor as RTDetrImageProcessor, __webpack_exports__RTDetrModel as RTDetrModel, __webpack_exports__RTDetrObjectDetectionOutput as RTDetrObjectDetectionOutput, __webpack_exports__RTDetrPreTrainedModel as RTDetrPreTrainedModel, __webpack_exports__RawAudio as RawAudio, __webpack_exports__RawImage as RawImage, __webpack_exports__RepetitionPenaltyLogitsProcessor as RepetitionPenaltyLogitsProcessor, __webpack_exports__ResNetForImageClassification as ResNetForImageClassification, __webpack_exports__ResNetModel as ResNetModel, __webpack_exports__ResNetPreTrainedModel as ResNetPreTrainedModel, __webpack_exports__RoFormerForMaskedLM as RoFormerForMaskedLM, __webpack_exports__RoFormerForQuestionAnswering as RoFormerForQuestionAnswering, __webpack_exports__RoFormerForSequenceClassification as RoFormerForSequenceClassification, __webpack_exports__RoFormerForTokenClassification as RoFormerForTokenClassification, __webpack_exports__RoFormerModel as RoFormerModel, __webpack_exports__RoFormerPreTrainedModel as RoFormerPreTrainedModel, __webpack_exports__RoFormerTokenizer as RoFormerTokenizer, __webpack_exports__RobertaForMaskedLM as RobertaForMaskedLM, __webpack_exports__RobertaForQuestionAnswering as RobertaForQuestionAnswering, __webpack_exports__RobertaForSequenceClassification as RobertaForSequenceClassification, __webpack_exports__RobertaForTokenClassification as RobertaForTokenClassification, __webpack_exports__RobertaModel as RobertaModel, __webpack_exports__RobertaPreTrainedModel as RobertaPreTrainedModel, __webpack_exports__RobertaTokenizer as RobertaTokenizer, __webpack_exports__SamImageProcessor as SamImageProcessor, __webpack_exports__SamImageSegmentationOutput as SamImageSegmentationOutput, __webpack_exports__SamModel as SamModel, __webpack_exports__SamPreTrainedModel as SamPreTrainedModel, __webpack_exports__SamProcessor as SamProcessor, __webpack_exports__SapiensForDepthEstimation as SapiensForDepthEstimation, __webpack_exports__SapiensForNormalEstimation as SapiensForNormalEstimation, __webpack_exports__SapiensForSemanticSegmentation as SapiensForSemanticSegmentation, __webpack_exports__SapiensPreTrainedModel as SapiensPreTrainedModel, __webpack_exports__SeamlessM4TFeatureExtractor as SeamlessM4TFeatureExtractor, __webpack_exports__SegformerFeatureExtractor as SegformerFeatureExtractor, __webpack_exports__SegformerForImageClassification as SegformerForImageClassification, __webpack_exports__SegformerForSemanticSegmentation as SegformerForSemanticSegmentation, __webpack_exports__SegformerImageProcessor as SegformerImageProcessor, __webpack_exports__SegformerModel as SegformerModel, __webpack_exports__SegformerPreTrainedModel as SegformerPreTrainedModel, __webpack_exports__Seq2SeqLMOutput as Seq2SeqLMOutput, __webpack_exports__SequenceClassifierOutput as SequenceClassifierOutput, __webpack_exports__SiglipImageProcessor as SiglipImageProcessor, __webpack_exports__SiglipModel as SiglipModel, __webpack_exports__SiglipPreTrainedModel as SiglipPreTrainedModel, __webpack_exports__SiglipTextModel as SiglipTextModel, __webpack_exports__SiglipTokenizer as SiglipTokenizer, __webpack_exports__SiglipVisionModel as SiglipVisionModel, __webpack_exports__SpeechT5FeatureExtractor as SpeechT5FeatureExtractor, __webpack_exports__SpeechT5ForSpeechToText as SpeechT5ForSpeechToText, __webpack_exports__SpeechT5ForTextToSpeech as SpeechT5ForTextToSpeech, __webpack_exports__SpeechT5HifiGan as SpeechT5HifiGan, __webpack_exports__SpeechT5Model as SpeechT5Model, __webpack_exports__SpeechT5PreTrainedModel as SpeechT5PreTrainedModel, __webpack_exports__SpeechT5Processor as SpeechT5Processor, __webpack_exports__SpeechT5Tokenizer as SpeechT5Tokenizer, __webpack_exports__SqueezeBertForMaskedLM as SqueezeBertForMaskedLM, __webpack_exports__SqueezeBertForQuestionAnswering as SqueezeBertForQuestionAnswering, __webpack_exports__SqueezeBertForSequenceClassification as SqueezeBertForSequenceClassification, __webpack_exports__SqueezeBertModel as SqueezeBertModel, __webpack_exports__SqueezeBertPreTrainedModel as SqueezeBertPreTrainedModel, __webpack_exports__SqueezeBertTokenizer as SqueezeBertTokenizer, __webpack_exports__StableLmForCausalLM as StableLmForCausalLM, __webpack_exports__StableLmModel as StableLmModel, __webpack_exports__StableLmPreTrainedModel as StableLmPreTrainedModel, __webpack_exports__Starcoder2ForCausalLM as Starcoder2ForCausalLM, __webpack_exports__Starcoder2Model as Starcoder2Model, __webpack_exports__Starcoder2PreTrainedModel as Starcoder2PreTrainedModel, __webpack_exports__StoppingCriteria as StoppingCriteria, __webpack_exports__StoppingCriteriaList as StoppingCriteriaList, __webpack_exports__StyleTextToSpeech2Model as StyleTextToSpeech2Model, __webpack_exports__StyleTextToSpeech2PreTrainedModel as StyleTextToSpeech2PreTrainedModel, __webpack_exports__SummarizationPipeline as SummarizationPipeline, __webpack_exports__SuppressTokensAtBeginLogitsProcessor as SuppressTokensAtBeginLogitsProcessor, __webpack_exports__Swin2SRForImageSuperResolution as Swin2SRForImageSuperResolution, __webpack_exports__Swin2SRImageProcessor as Swin2SRImageProcessor, __webpack_exports__Swin2SRModel as Swin2SRModel, __webpack_exports__Swin2SRPreTrainedModel as Swin2SRPreTrainedModel, __webpack_exports__SwinForImageClassification as SwinForImageClassification, __webpack_exports__SwinModel as SwinModel, __webpack_exports__SwinPreTrainedModel as SwinPreTrainedModel, __webpack_exports__T5ForConditionalGeneration as T5ForConditionalGeneration, __webpack_exports__T5Model as T5Model, __webpack_exports__T5PreTrainedModel as T5PreTrainedModel, __webpack_exports__T5Tokenizer as T5Tokenizer, __webpack_exports__TableTransformerForObjectDetection as TableTransformerForObjectDetection, __webpack_exports__TableTransformerModel as TableTransformerModel, __webpack_exports__TableTransformerObjectDetectionOutput as TableTransformerObjectDetectionOutput, __webpack_exports__TableTransformerPreTrainedModel as TableTransformerPreTrainedModel, __webpack_exports__TemperatureLogitsWarper as TemperatureLogitsWarper, __webpack_exports__Tensor as Tensor, __webpack_exports__Text2TextGenerationPipeline as Text2TextGenerationPipeline, __webpack_exports__TextClassificationPipeline as TextClassificationPipeline, __webpack_exports__TextGenerationPipeline as TextGenerationPipeline, __webpack_exports__TextStreamer as TextStreamer, __webpack_exports__TextToAudioPipeline as TextToAudioPipeline, __webpack_exports__TokenClassificationPipeline as TokenClassificationPipeline, __webpack_exports__TokenClassifierOutput as TokenClassifierOutput, __webpack_exports__TokenizerModel as TokenizerModel, __webpack_exports__TopKLogitsWarper as TopKLogitsWarper, __webpack_exports__TopPLogitsWarper as TopPLogitsWarper, __webpack_exports__TrOCRForCausalLM as TrOCRForCausalLM, __webpack_exports__TrOCRPreTrainedModel as TrOCRPreTrainedModel, __webpack_exports__TranslationPipeline as TranslationPipeline, __webpack_exports__UniSpeechForCTC as UniSpeechForCTC, __webpack_exports__UniSpeechForSequenceClassification as UniSpeechForSequenceClassification, __webpack_exports__UniSpeechModel as UniSpeechModel, __webpack_exports__UniSpeechPreTrainedModel as UniSpeechPreTrainedModel, __webpack_exports__UniSpeechSatForAudioFrameClassification as UniSpeechSatForAudioFrameClassification, __webpack_exports__UniSpeechSatForCTC as UniSpeechSatForCTC, __webpack_exports__UniSpeechSatForSequenceClassification as UniSpeechSatForSequenceClassification, __webpack_exports__UniSpeechSatModel as UniSpeechSatModel, __webpack_exports__UniSpeechSatPreTrainedModel as UniSpeechSatPreTrainedModel, __webpack_exports__VLChatProcessor as VLChatProcessor, __webpack_exports__VLMImageProcessor as VLMImageProcessor, __webpack_exports__ViTFeatureExtractor as ViTFeatureExtractor, __webpack_exports__ViTForImageClassification as ViTForImageClassification, __webpack_exports__ViTImageProcessor as ViTImageProcessor, __webpack_exports__ViTMAEModel as ViTMAEModel, __webpack_exports__ViTMAEPreTrainedModel as ViTMAEPreTrainedModel, __webpack_exports__ViTMSNForImageClassification as ViTMSNForImageClassification, __webpack_exports__ViTMSNModel as ViTMSNModel, __webpack_exports__ViTMSNPreTrainedModel as ViTMSNPreTrainedModel, __webpack_exports__ViTModel as ViTModel, __webpack_exports__ViTPreTrainedModel as ViTPreTrainedModel, __webpack_exports__VisionEncoderDecoderModel as VisionEncoderDecoderModel, __webpack_exports__VitMatteForImageMatting as VitMatteForImageMatting, __webpack_exports__VitMatteImageProcessor as VitMatteImageProcessor, __webpack_exports__VitMattePreTrainedModel as VitMattePreTrainedModel, __webpack_exports__VitPoseForPoseEstimation as VitPoseForPoseEstimation, __webpack_exports__VitPoseImageProcessor as VitPoseImageProcessor, __webpack_exports__VitPosePreTrainedModel as VitPosePreTrainedModel, __webpack_exports__VitsModel as VitsModel, __webpack_exports__VitsModelOutput as VitsModelOutput, __webpack_exports__VitsPreTrainedModel as VitsPreTrainedModel, __webpack_exports__VitsTokenizer as VitsTokenizer, __webpack_exports__Wav2Vec2BertForCTC as Wav2Vec2BertForCTC, __webpack_exports__Wav2Vec2BertForSequenceClassification as Wav2Vec2BertForSequenceClassification, __webpack_exports__Wav2Vec2BertModel as Wav2Vec2BertModel, __webpack_exports__Wav2Vec2BertPreTrainedModel as Wav2Vec2BertPreTrainedModel, __webpack_exports__Wav2Vec2CTCTokenizer as Wav2Vec2CTCTokenizer, __webpack_exports__Wav2Vec2FeatureExtractor as Wav2Vec2FeatureExtractor, __webpack_exports__Wav2Vec2ForAudioFrameClassification as Wav2Vec2ForAudioFrameClassification, __webpack_exports__Wav2Vec2ForCTC as Wav2Vec2ForCTC, __webpack_exports__Wav2Vec2ForSequenceClassification as Wav2Vec2ForSequenceClassification, __webpack_exports__Wav2Vec2Model as Wav2Vec2Model, __webpack_exports__Wav2Vec2PreTrainedModel as Wav2Vec2PreTrainedModel, __webpack_exports__Wav2Vec2Processor as Wav2Vec2Processor, __webpack_exports__Wav2Vec2ProcessorWithLM as Wav2Vec2ProcessorWithLM, __webpack_exports__WavLMForAudioFrameClassification as WavLMForAudioFrameClassification, __webpack_exports__WavLMForCTC as WavLMForCTC, __webpack_exports__WavLMForSequenceClassification as WavLMForSequenceClassification, __webpack_exports__WavLMForXVector as WavLMForXVector, __webpack_exports__WavLMModel as WavLMModel, __webpack_exports__WavLMPreTrainedModel as WavLMPreTrainedModel, __webpack_exports__WeSpeakerFeatureExtractor as WeSpeakerFeatureExtractor, __webpack_exports__WeSpeakerResNetModel as WeSpeakerResNetModel, __webpack_exports__WeSpeakerResNetPreTrainedModel as WeSpeakerResNetPreTrainedModel, __webpack_exports__WhisperFeatureExtractor as WhisperFeatureExtractor, __webpack_exports__WhisperForConditionalGeneration as WhisperForConditionalGeneration, __webpack_exports__WhisperModel as WhisperModel, __webpack_exports__WhisperPreTrainedModel as WhisperPreTrainedModel, __webpack_exports__WhisperProcessor as WhisperProcessor, __webpack_exports__WhisperTextStreamer as WhisperTextStreamer, __webpack_exports__WhisperTimeStampLogitsProcessor as WhisperTimeStampLogitsProcessor, __webpack_exports__WhisperTokenizer as WhisperTokenizer, __webpack_exports__XLMForQuestionAnswering as XLMForQuestionAnswering, __webpack_exports__XLMForSequenceClassification as XLMForSequenceClassification, __webpack_exports__XLMForTokenClassification as XLMForTokenClassification, __webpack_exports__XLMModel as XLMModel, __webpack_exports__XLMPreTrainedModel as XLMPreTrainedModel, __webpack_exports__XLMRobertaForMaskedLM as XLMRobertaForMaskedLM, __webpack_exports__XLMRobertaForQuestionAnswering as XLMRobertaForQuestionAnswering, __webpack_exports__XLMRobertaForSequenceClassification as XLMRobertaForSequenceClassification, __webpack_exports__XLMRobertaForTokenClassification as XLMRobertaForTokenClassification, __webpack_exports__XLMRobertaModel as XLMRobertaModel, __webpack_exports__XLMRobertaPreTrainedModel as XLMRobertaPreTrainedModel, __webpack_exports__XLMRobertaTokenizer as XLMRobertaTokenizer, __webpack_exports__XLMTokenizer as XLMTokenizer, __webpack_exports__XLMWithLMHeadModel as XLMWithLMHeadModel, __webpack_exports__XVectorOutput as XVectorOutput, __webpack_exports__YolosFeatureExtractor as YolosFeatureExtractor, __webpack_exports__YolosForObjectDetection as YolosForObjectDetection, __webpack_exports__YolosImageProcessor as YolosImageProcessor, __webpack_exports__YolosModel as YolosModel, __webpack_exports__YolosObjectDetectionOutput as YolosObjectDetectionOutput, __webpack_exports__YolosPreTrainedModel as YolosPreTrainedModel, __webpack_exports__ZeroShotAudioClassificationPipeline as ZeroShotAudioClassificationPipeline, __webpack_exports__ZeroShotClassificationPipeline as ZeroShotClassificationPipeline, __webpack_exports__ZeroShotImageClassificationPipeline as ZeroShotImageClassificationPipeline, __webpack_exports__ZeroShotObjectDetectionPipeline as ZeroShotObjectDetectionPipeline, __webpack_exports__bankers_round as bankers_round, __webpack_exports__cat as cat, __webpack_exports__cos_sim as cos_sim, __webpack_exports__dot as dot, __webpack_exports__dynamic_time_warping as dynamic_time_warping, __webpack_exports__env as env, __webpack_exports__full as full, __webpack_exports__full_like as full_like, __webpack_exports__getKeyValueShapes as getKeyValueShapes, __webpack_exports__hamming as hamming, __webpack_exports__hanning as hanning, __webpack_exports__interpolate as interpolate, __webpack_exports__interpolate_4d as interpolate_4d, __webpack_exports__interpolate_data as interpolate_data, __webpack_exports__is_chinese_char as is_chinese_char, __webpack_exports__layer_norm as layer_norm, __webpack_exports__load_image as load_image, __webpack_exports__log_softmax as log_softmax, __webpack_exports__magnitude as magnitude, __webpack_exports__matmul as matmul, __webpack_exports__max as max, __webpack_exports__mean as mean, __webpack_exports__mean_pooling as mean_pooling, __webpack_exports__medianFilter as medianFilter, __webpack_exports__mel_filter_bank as mel_filter_bank, __webpack_exports__min as min, __webpack_exports__ones as ones, __webpack_exports__ones_like as ones_like, __webpack_exports__permute as permute, __webpack_exports__permute_data as permute_data, __webpack_exports__pipeline as pipeline, __webpack_exports__quantize_embeddings as quantize_embeddings, __webpack_exports__rand as rand, __webpack_exports__read_audio as read_audio, __webpack_exports__rfft as rfft, __webpack_exports__round as round, __webpack_exports__slice as slice, __webpack_exports__softmax as softmax, __webpack_exports__spectrogram as spectrogram, __webpack_exports__stack as stack, __webpack_exports__std_mean as std_mean, __webpack_exports__topk as topk, __webpack_exports__window_function as window_function, __webpack_exports__zeros as zeros, __webpack_exports__zeros_like as zeros_like };
 
 //# sourceMappingURL=transformers.js.map