@huggingface/transformers 3.0.0-alpha.7 → 3.0.0-alpha.8

package/dist/transformers.js

@@ -3771,7 +3771,7 @@ __webpack_require__.r(__webpack_exports__);
 // Licensed under the MIT License.
 // This file is generated by /js/scripts/update-version.ts
 // Do not modify file content manually.
-const version = '1.18.0';
+const version = '1.19.0';
 //# sourceMappingURL=version.js.map
 
 /***/ }),
@@ -3784,34 +3784,33 @@ const version = '1.18.0';
 
 __webpack_require__.r(__webpack_exports__);
 /* harmony export */ __webpack_require__.d(__webpack_exports__, {
-/* harmony export */   "InferenceSession": () => (/* binding */ dp),
-/* harmony export */   "TRACE": () => (/* binding */ Sr),
-/* harmony export */   "TRACE_FUNC_BEGIN": () => (/* binding */ We),
-/* harmony export */   "TRACE_FUNC_END": () => (/* binding */ Me),
-/* harmony export */   "Tensor": () => (/* binding */ De),
-/* harmony export */   "TrainingSession": () => (/* binding */ cp),
-/* harmony export */   "default": () => (/* binding */ K_),
-/* harmony export */   "env": () => (/* binding */ be),
-/* harmony export */   "registerBackend": () => (/* binding */ _t)
+/* harmony export */   "InferenceSession": () => (/* binding */ op),
+/* harmony export */   "TRACE": () => (/* binding */ $r),
+/* harmony export */   "TRACE_FUNC_BEGIN": () => (/* binding */ Ve),
+/* harmony export */   "TRACE_FUNC_END": () => (/* binding */ De),
+/* harmony export */   "Tensor": () => (/* binding */ Pe),
+/* harmony export */   "TrainingSession": () => (/* binding */ ap),
+/* harmony export */   "default": () => (/* binding */ ex),
+/* harmony export */   "env": () => (/* binding */ fe),
+/* harmony export */   "registerBackend": () => (/* binding */ wt)
 /* harmony export */ });
 /*!
- * ONNX Runtime Web v1.19.0
+ * ONNX Runtime Web v1.20.0-dev.20240818-417aa00406
  * Copyright (c) Microsoft Corporation. All rights reserved.
  * Licensed under the MIT License.
  */
-var Rn=Object.defineProperty;var op=Object.getOwnPropertyDescriptor;var ip=Object.getOwnPropertyNames;var ap=Object.prototype.hasOwnProperty;var Mn=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,r)=>(typeof require<"u"?require:t)[r]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')});var U=(e,t)=>()=>(e&&(t=e(e=0)),t);var Ft=(e,t)=>{for(var r in t)Rn(e,r,{get:t[r],enumerable:!0})},sp=(e,t,r,n)=>{if(t&&typeof t=="object"||typeof t=="function")for(let o of ip(t))!ap.call(e,o)&&o!==r&&Rn(e,o,{get:()=>t[o],enumerable:!(n=op(t,o))||n.enumerable});return e};var br=e=>sp(Rn({},"__esModule",{value:!0}),e);var wr,$t,_t,up,vr,$r=U(()=>{"use strict";wr=new Map,$t=[],_t=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let n=wr.get(e);if(n===void 0)wr.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 ${r}`)}if(r>=0){let o=$t.indexOf(e);o!==-1&&$t.splice(o,1);for(let i=0;i<$t.length;i++)if(wr.get($t[i]).priority<=r){$t.splice(i,0,e);return}$t.push(e)}return}throw new TypeError("not a valid backend")},up=async e=>{let t=wr.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(n){return r||(t.error=`${n}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},vr=async e=>{let t=e.executionProviders||[],r=t.map(d=>typeof d=="string"?d:d.name),n=r.length===0?$t:r,o,i=[],a=new Set;for(let d of n){let p=await up(d);typeof p=="string"?i.push({name:d,err:p}):(o||(o=p),o===p&&a.add(d))}if(!o)throw new Error(`no available backend found. ERR: ${i.map(d=>`[${d.name}] ${d.err}`).join(", ")}`);for(let{name:d,err:p}of i)r.includes(d)&&console.warn(`removing requested execution provider "${d}" from session options because it is not available: ${p}`);let l=t.filter(d=>a.has(typeof d=="string"?d:d.name));return[o,new Proxy(e,{get:(d,p)=>p==="executionProviders"?l:Reflect.get(d,p)})]}});var Yi=U(()=>{"use strict";$r()});var Xi,Zi=U(()=>{"use strict";Xi="1.19.0"});var Qi,Ne,Un=U(()=>{"use strict";Zi();Qi="warning",Ne={wasm:{},webgl:{},webgpu:{},versions:{common:Xi},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: ${e}`);Qi=e}},get logLevel(){return Qi}};Object.defineProperty(Ne,"logLevel",{enumerable:!0})});var be,Ji=U(()=>{"use strict";Un();be=Ne});var ea,ta,ra=U(()=>{"use strict";ea=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let n=r.getContext("2d");if(n!=null){let o,i;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],i=e.dims[3]):(o=e.dims[3],i=e.dims[2]);let a=t?.format!==void 0?t.format:"RGB",l=t?.norm,d,p;l===void 0||l.mean===void 0?d=[255,255,255,255]:typeof l.mean=="number"?d=[l.mean,l.mean,l.mean,l.mean]:(d=[l.mean[0],l.mean[1],l.mean[2],0],l.mean[3]!==void 0&&(d[3]=l.mean[3])),l===void 0||l.bias===void 0?p=[0,0,0,0]:typeof l.bias=="number"?p=[l.bias,l.bias,l.bias,l.bias]:(p=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(p[3]=l.bias[3]));let m=i*o,u=0,h=m,w=m*2,g=-1;a==="RGBA"?(u=0,h=m,w=m*2,g=m*3):a==="RGB"?(u=0,h=m,w=m*2):a==="RBG"&&(u=0,w=m,h=m*2);for(let b=0;b<i;b++)for(let x=0;x<o;x++){let _=(e.data[u++]-p[0])*d[0],$=(e.data[h++]-p[1])*d[1],S=(e.data[w++]-p[2])*d[2],I=g===-1?255:(e.data[g++]-p[3])*d[3];n.fillStyle="rgba("+_+","+$+","+S+","+I+")",n.fillRect(x,b,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},ta=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let o,i,a;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],i=e.dims[1],a=e.dims[3]):(o=e.dims[3],i=e.dims[2],a=e.dims[1]);let l=t!==void 0&&t.format!==void 0?t.format:"RGB",d=t?.norm,p,m;d===void 0||d.mean===void 0?p=[255,255,255,255]:typeof d.mean=="number"?p=[d.mean,d.mean,d.mean,d.mean]:(p=[d.mean[0],d.mean[1],d.mean[2],255],d.mean[3]!==void 0&&(p[3]=d.mean[3])),d===void 0||d.bias===void 0?m=[0,0,0,0]:typeof d.bias=="number"?m=[d.bias,d.bias,d.bias,d.bias]:(m=[d.bias[0],d.bias[1],d.bias[2],0],d.bias[3]!==void 0&&(m[3]=d.bias[3]));let u=i*o;if(t!==void 0&&(t.format!==void 0&&a===4&&t.format!=="RGBA"||a===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let h=4,w=0,g=1,b=2,x=3,_=0,$=u,S=u*2,I=-1;l==="RGBA"?(_=0,$=u,S=u*2,I=u*3):l==="RGB"?(_=0,$=u,S=u*2):l==="RBG"&&(_=0,S=u,$=u*2),n=r.createImageData(o,i);for(let T=0;T<i*o;w+=h,g+=h,b+=h,x+=h,T++)n.data[w]=(e.data[_++]-m[0])*p[0],n.data[g]=(e.data[$++]-m[1])*p[1],n.data[b]=(e.data[S++]-m[2])*p[2],n.data[x]=I===-1?255:(e.data[I++]-m[3])*p[3]}else throw new Error("Can not access image data");return n}});var Vn,na,oa,ia,aa,sa=U(()=>{"use strict";_r();Vn=(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 l=t.format!==void 0?t.format:"RGBA",d=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",p=r*n,m=d==="RGBA"?new Float32Array(p*4):new Float32Array(p*3),u=4,h=0,w=1,g=2,b=3,x=0,_=p,$=p*2,S=-1;l==="RGB"&&(u=3,h=0,w=1,g=2,b=-1),d==="RGBA"?S=p*3:d==="RBG"?(x=0,$=p,_=p*2):d==="BGR"&&($=0,_=p,x=p*2);for(let T=0;T<p;T++,h+=u,g+=u,w+=u,b+=u)m[x++]=(e[h]+a[0])/i[0],m[_++]=(e[w]+a[1])/i[1],m[$++]=(e[g]+a[2])/i[2],S!==-1&&b!==-1&&(m[S++]=(e[b]+a[3])/i[3]);return d==="RGBA"?new Be("float32",m,[1,4,r,n]):new Be("float32",m,[1,3,r,n])},na=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,l=t??{},d=()=>{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=>m instanceof HTMLCanvasElement||m instanceof OffscreenCanvas?m.getContext("2d"):null;if(r){let m=d();m.width=e.width,m.height=e.height;let u=p(m);if(u!=null){let h=e.height,w=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(h=t.resizedHeight,w=t.resizedWidth),t!==void 0){if(l=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");l.tensorFormat="RGBA",l.height=h,l.width=w}else l.tensorFormat="RGBA",l.height=h,l.width=w;u.drawImage(e,0,0),a=u.getImageData(0,0,w,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&&(l=t),l.format="RGBA",l.height=m,l.width=u,t!==void 0){let h=d();h.width=u,h.height=m;let w=p(h);if(w!=null)w.putImageData(e,0,0),a=w.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=d();m.width=e.width,m.height=e.height;let u=p(m);if(u!=null){let h=e.height,w=e.width;return u.drawImage(e,0,0,w,h),a=u.getImageData(0,0,w,h).data,l.height=h,l.width=w,Vn(a,l)}else throw new Error("Can not access image data")}else{if(i)return new Promise((m,u)=>{let h=d(),w=p(h);if(!e||!w)return u();let g=new Image;g.crossOrigin="Anonymous",g.src=e,g.onload=()=>{h.width=g.width,h.height=g.height,w.drawImage(g,0,0,h.width,h.height);let b=w.getImageData(0,0,h.width,h.height);l.height=h.height,l.width=h.width,m(Vn(b.data,l))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(a!==void 0)return Vn(a,l);throw new Error("Input data provided is not supported - aborted tensor creation")},oa=(e,t)=>{let{width:r,height:n,download:o,dispose:i}=t,a=[1,n,r,4];return new Be({location:"texture",type:"float32",texture:e,dims:a,download:o,dispose:i})},ia=(e,t)=>{let{dataType:r,dims:n,download:o,dispose:i}=t;return new Be({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:n,download:o,dispose:i})},aa=(e,t,r)=>new Be({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})});var xt,qt,ua,da,la=U(()=>{"use strict";xt=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),qt=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),ua=!1,da=()=>{if(!ua){ua=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;e&&(xt.set("int64",BigInt64Array),qt.set(BigInt64Array,"int64")),t&&(xt.set("uint64",BigUint64Array),qt.set(BigUint64Array,"uint64")),r?(xt.set("float16",Float16Array),qt.set(Float16Array,"float16")):xt.set("float16",Uint16Array)}}});var ca,pa,ma=U(()=>{"use strict";_r();ca=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},pa=(e,t)=>{switch(e.location){case"cpu":return new Be(e.type,e.data,t);case"cpu-pinned":return new Be({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new Be({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new Be({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}});var Be,_r=U(()=>{"use strict";ra();sa();la();ma();Be=class{constructor(t,r,n){da();let o,i;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,o=t.type,i=t.dims,t.location){case"cpu-pinned":{let l=xt.get(o);if(!l)throw new TypeError(`unsupported type "${o}" to create tensor from pinned buffer`);if(!(t.data instanceof l))throw new TypeError(`buffer should be of type ${l.name}`);this.cpuData=t.data;break}case"texture":{if(o!=="float32")throw new TypeError(`unsupported type "${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")throw new TypeError(`unsupported type "${o}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let l,d;if(typeof t=="string")if(o=t,d=n,t==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");l=r}else{let p=xt.get(t);if(p===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(r)){if(t==="float16"&&p===Uint16Array)throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");t==="uint64"||t==="int64"?l=p.from(r,BigInt):l=p.from(r)}else if(r instanceof p)l=r;else throw new TypeError(`A ${o} tensor's data must be type of ${p}`)}else if(d=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",l=t;else if(p==="boolean")o="bool",l=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${p}.`)}else{let p=qt.get(t.constructor);if(p===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);o=p,l=t}if(d===void 0)d=[l.length];else if(!Array.isArray(d))throw new TypeError("A tensor's dims must be a number array");i=d,this.cpuData=l,this.dataLocation="cpu"}let a=ca(i);if(this.cpuData&&a!==this.cpuData.length)throw new Error(`Tensor's size(${a}) does not match data length(${this.cpuData.length}).`);this.type=o,this.dims=i,this.size=a}static async fromImage(t,r){return na(t,r)}static fromTexture(t,r){return oa(t,r)}static fromGpuBuffer(t,r){return ia(t,r)}static fromPinnedBuffer(t,r,n){return aa(t,r,n)}toDataURL(t){return ea(this,t)}toImageData(t){return ta(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}async getData(t){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":{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: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return pa(this,t)}}});var De,xr=U(()=>{"use strict";_r();De=Be});var Sr,fa,We,Me,Nn=U(()=>{"use strict";Un();Sr=(e,t)=>{(typeof Ne.trace>"u"?!Ne.wasm.trace:!Ne.trace)||console.timeStamp(`${e}::ORT::${t}`)},fa=(e,t)=>{let r=new Error().stack?.split(/\r\n|\r|\n/g)||[],n=!1;for(let o=0;o<r.length;o++){if(n&&!r[o].includes("TRACE_FUNC")){let i=`FUNC_${e}::${r[o].trim().split(" ")[1]}`;t&&(i+=`::${t}`),Sr("CPU",i);return}r[o].includes("TRACE_FUNC")&&(n=!0)}},We=e=>{(typeof Ne.trace>"u"?!Ne.wasm.trace:!Ne.trace)||fa("BEGIN",e)},Me=e=>{(typeof Ne.trace>"u"?!Ne.wasm.trace:!Ne.trace)||fa("END",e)}});var Ir,ha=U(()=>{"use strict";$r();xr();Nn();Ir=class e{constructor(t){this.handler=t}async run(t,r,n){We();let o={},i={};if(typeof t!="object"||t===null||t instanceof De||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 De)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: ${p}.`);o[p]=null}if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let p=!1,m=Object.getOwnPropertyNames(r);for(let u of this.outputNames)if(m.indexOf(u)!==-1){let h=r[u];(h===null||h instanceof De)&&(p=!0,a=!1,o[u]=h)}if(p){if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else i=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let p of this.inputNames)if(typeof t[p]>"u")throw new Error(`input '${p}' is missing in 'feeds'.`);if(a)for(let p of this.outputNames)o[p]=null;let l=await this.handler.run(t,o,i),d={};for(let p in l)if(Object.hasOwnProperty.call(l,p)){let m=l[p];m instanceof De?d[p]=m:d[p]=new De(m.type,m.data,m.dims)}return Me(),d}async release(){return this.handler.dispose()}static async create(t,r,n,o){We();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, ${m.byteLength-u}].`);if(typeof o=="object"&&o!==null)a=o;else if(typeof o<"u")throw new TypeError("'options' must be an object.")}else if(typeof n<"u")throw new TypeError("'byteLength' must be a number.")}else if(typeof r<"u")throw new TypeError("'options' must be an object.");i=new Uint8Array(m,u,h)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[l,d]=await vr(a),p=await l.createInferenceSessionHandler(i,d);return Me(),new e(p)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}});var dp,ga=U(()=>{"use strict";ha();dp=Ir});var ya=U(()=>{"use strict"});var ba=U(()=>{"use strict"});var wa=U(()=>{"use strict"});var va=U(()=>{"use strict"});var lp,Cr,$a=U(()=>{"use strict";$r();xr();lp="Training backend could not be resolved. Make sure you're using the correct configuration & WebAssembly files.",Cr=class e{constructor(t,r,n){this.handler=t,this.hasOptimizerModel=r,this.hasEvalModel=n}get trainingInputNames(){return this.handler.inputNames}get trainingOutputNames(){return this.handler.outputNames}get evalInputNames(){if(this.hasEvalModel)return this.handler.evalInputNames;throw new Error("This training session has no evalModel loaded.")}get evalOutputNames(){if(this.hasEvalModel)return this.handler.evalOutputNames;throw new Error("This training session has no evalModel loaded.")}static async create(t,r){let n=t.evalModel||"",o=t.optimizerModel||"",i=r||{},[a,l]=await vr(i);if(a.createTrainingSessionHandler){let d=await a.createTrainingSessionHandler(t.checkpointState,t.trainModel,n,o,l);return new e(d,!!t.optimizerModel,!!t.evalModel)}else throw new Error(lp)}typeNarrowingForRunStep(t,r,n,o,i){let a={},l={};if(typeof n!="object"||n===null||n instanceof De||Array.isArray(n))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let d=!0;if(typeof o=="object"){if(o===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(o instanceof De)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(o)){if(o.length===0)throw new TypeError("'fetches' cannot be an empty array.");d=!1;for(let p of o){if(typeof p!="string")throw new TypeError("'fetches' must be a string array or an object.");if(r.indexOf(p)===-1)throw new RangeError(`'fetches' contains invalid output name: ${p}.`);a[p]=null}if(typeof i=="object"&&i!==null)l=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else{let p=!1,m=Object.getOwnPropertyNames(o);for(let u of r)if(m.indexOf(u)!==-1){let h=o[u];(h===null||h instanceof De)&&(p=!0,d=!1,a[u]=h)}if(p){if(typeof i=="object"&&i!==null)l=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else l=o}}else if(typeof o<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let p of t)if(typeof n[p]>"u")throw new Error(`input '${p}' is missing in 'feeds'.`);if(d)for(let p of r)a[p]=null;return[a,l]}convertHandlerReturnTypeToMapOfTensors(t){let r={};for(let n in t)if(Object.hasOwnProperty.call(t,n)){let o=t[n];o instanceof De?r[n]=o:r[n]=new De(o.type,o.data,o.dims)}return r}async lazyResetGrad(){await this.handler.lazyResetGrad()}async runTrainStep(t,r,n){let[o,i]=this.typeNarrowingForRunStep(this.trainingInputNames,this.trainingOutputNames,t,r,n),a=await this.handler.runTrainStep(t,o,i);return this.convertHandlerReturnTypeToMapOfTensors(a)}async runOptimizerStep(t){if(this.hasOptimizerModel)await this.handler.runOptimizerStep(t||{});else throw new Error("This TrainingSession has no OptimizerModel loaded.")}async runEvalStep(t,r,n){if(this.hasEvalModel){let[o,i]=this.typeNarrowingForRunStep(this.evalInputNames,this.evalOutputNames,t,r,n),a=await this.handler.runEvalStep(t,o,i);return this.convertHandlerReturnTypeToMapOfTensors(a)}else throw new Error("This TrainingSession has no EvalModel loaded.")}async getParametersSize(t=!0){return this.handler.getParametersSize(t)}async loadParametersBuffer(t,r=!0){let n=await this.getParametersSize(r);if(t.length!==4*n)throw new Error("Size of the buffer passed into loadParametersBuffer must match the number of parameters in the model. Please use getParametersSize method to check.");return this.handler.loadParametersBuffer(t,r)}async getContiguousParameters(t=!0){return this.handler.getContiguousParameters(t)}async release(){return this.handler.dispose()}}});var cp,_a=U(()=>{"use strict";$a();cp=Cr});var Wn={};Ft(Wn,{InferenceSession:()=>dp,TRACE:()=>Sr,TRACE_FUNC_BEGIN:()=>We,TRACE_FUNC_END:()=>Me,Tensor:()=>De,TrainingSession:()=>cp,env:()=>be,registerBackend:()=>_t});var Ke=U(()=>{"use strict";Yi();Ji();ga();xr();ya();ba();Nn();wa();va();_a()});var Tr=U(()=>{"use strict"});var Ca={};Ft(Ca,{default:()=>pp});var Sa,Ia,pp,Ta=U(()=>{"use strict";Hn();St();jt();Sa="ort-wasm-proxy-worker",Ia=globalThis.self?.name===Sa;Ia&&(self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":Ar(r.wasm).then(()=>{Er(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;kr(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;Pr(n,o).then(i=>{postMessage({type:t,out:i})},i=>{postMessage({type:t,err:i})});break}case"release":Or(r),postMessage({type:t});break;case"run":{let{sessionId:n,inputIndices:o,inputs:i,outputIndices:a,options:l}=r;Dr(n,o,i,a,new Array(a.length).fill(null),l).then(d=>{d.some(p=>p[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:d},Br([...i,...d]))},d=>{postMessage({type:t,err:d})});break}case"end-profiling":zr(r),postMessage({type:t});break;default:}}catch(n){postMessage({type:t,err:n})}});pp=Ia?null:e=>new Worker(e??Ut,{type:"module",name:Sa})});var Ea={};Ft(Ea,{default:()=>mp});var Gn,Aa,mp,ka=U(()=>{"use strict";Aa=(Gn=import.meta.url,async function(e={}){function t(){return le.buffer!=re.buffer&&we(),re}function r(){return le.buffer!=re.buffer&&we(),se}function n(){return le.buffer!=re.buffer&&we(),Z}function o(){return le.buffer!=re.buffer&&we(),ne}function i(){return le.buffer!=re.buffer&&we(),J}function a(){return le.buffer!=re.buffer&&we(),ke}function l(){return le.buffer!=re.buffer&&we(),R}function d(){return le.buffer!=re.buffer&&we(),Re}var p,m,u=Object.assign({},e),h=new Promise((s,c)=>{p=s,m=c}),w=typeof window=="object",g=typeof importScripts=="function",b=g&&self.name=="em-pthread";u.mountExternalData=(s,c)=>{(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,y,v)=>(...C)=>{let O=Je,B=y?.();C=f(...C);let N=y?.();return B!==N&&(f=N,v(B),y=v=null),Je!=O?new Promise((L,j)=>{An={resolve:L,reject:j}}):C},c=f=>async(...y)=>{try{if(u.Eb)throw Error("Session already started");let v=u.Eb={bc:y[0],errors:[]},C=await f(...y);if(u.Eb!==v)throw Error("Session mismatch");u.Mb?.flush();let O=v.errors;if(0<O.length){let B=await Promise.all(O);if(B=B.filter(N=>N),0<B.length)throw Error(B.join(`
-`))}return C}finally{u.Eb=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.Mb,u.Tb,u.Xb,u.Nb,u.Wb,u.jb,u.Yb,u.$b,u.Ub,u.Vb,u.Zb]=c;let f=u.Mb;u.jsepRegisterBuffer=(y,v,C,O)=>f.registerBuffer(y,v,C,O),u.jsepGetBuffer=y=>f.getBuffer(y),u.jsepCreateDownloader=(y,v,C)=>f.createDownloader(y,v,C),u.jsepOnReleaseSession=y=>{f.onReleaseSession(y)},u.jsepOnRunStart=y=>f.onRunStart(y)}};var $,S,I=Object.assign({},u),T="./this.program",A=(s,c)=>{throw c},D="";(w||g)&&(g?D=self.location.href:typeof document<"u"&&document.currentScript&&(D=document.currentScript.src),Gn&&(D=Gn),D=D.startsWith("blob:")?"":D.substr(0,D.replace(/[?#].*/,"").lastIndexOf("/")+1),g&&(S=s=>{var c=new XMLHttpRequest;return c.open("GET",s,!1),c.responseType="arraybuffer",c.send(null),new Uint8Array(c.response)}),$=(s,c,f)=>{var y=new XMLHttpRequest;y.open("GET",s,!0),y.responseType="arraybuffer",y.onload=()=>{y.status==200||y.status==0&&y.response?c(y.response):f()},y.onerror=f,y.send(null)});var z,H=console.log.bind(console),W=console.error.bind(console),X=H,Y=W;if(Object.assign(u,I),I=null,b){let s=function(c){try{var f=c.data,y=f.cmd;if(y==="load"){let v=[];self.onmessage=C=>v.push(C),self.startWorker=()=>{postMessage({cmd:"loaded"});for(let C of v)s(C);self.onmessage=s};for(let C of f.handlers)u[C]&&!u[C].proxy||(u[C]=(...O)=>{postMessage({Lb:"callHandler",kc:C,args:O})},C=="print"&&(X=u[C]),C=="printErr"&&(Y=u[C]));le=f.wasmMemory,we(),pe(f.wasmModule)}else if(y==="run"){On(f.pthread_ptr,0,0,1,0,0),In(f.pthread_ptr),Hl(),No(),ue||(Mi(),ue=!0);try{Gl(f.start_routine,f.arg)}catch(v){if(v!="unwind")throw v}}else y==="cancel"?Mt()&&gr(-1):f.target!=="setimmediate"&&(y==="checkMailbox"?ue&&sr():y&&(Y(`worker: received unknown command ${y}`),Y(f)))}catch(v){throw Ui(),v}};var Ih=s,pe,ue=!1;Y=function(...c){c=c.join(" "),console.error(c)},self.alert=function(...c){postMessage({Lb:"alert",text:c.join(" "),mc:Mt()})},u.instantiateWasm=(c,f)=>new Promise(y=>{pe=v=>{v=new WebAssembly.Instance(v,Bo()),f(v),y()}}),self.onunhandledrejection=c=>{throw c.reason||c},self.onmessage=s}u.wasmBinary&&(z=u.wasmBinary);var le,K,de,re,se,Z,ne,J,ke,R,F,ce,Re,Se=!1;function we(){var s=le.buffer;u.HEAP8=re=new Int8Array(s),u.HEAP16=Z=new Int16Array(s),u.HEAPU8=se=new Uint8Array(s),u.HEAPU16=ne=new Uint16Array(s),u.HEAP32=J=new Int32Array(s),u.HEAPU32=ke=new Uint32Array(s),u.HEAPF32=R=new Float32Array(s),u.HEAPF64=Re=new Float64Array(s),u.HEAP64=F=new BigInt64Array(s),u.HEAPU64=ce=new BigUint64Array(s)}if(!b){if(!((le=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0})).buffer instanceof x))throw Y("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),Error("bad memory");we()}var Lt=[],Pe=[],ze=[],Fe=0,Ot=null,yt=null;function ko(){if(--Fe==0&&(Ot!==null&&(clearInterval(Ot),Ot=null),yt)){var s=yt;yt=null,s()}}function Dt(s){throw Y(s="Aborted("+s+")"),Se=!0,de=1,s=new WebAssembly.RuntimeError(s+". Build with -sASSERTIONS for more info."),m(s),s}var mn,Po=s=>s.startsWith("data:application/octet-stream;base64,"),Oo=s=>s.startsWith("file://");function Do(s){if(s==mn&&z)return new Uint8Array(z);if(S)return S(s);throw"both async and sync fetching of the wasm failed"}function zo(s,c,f){return function(y){if(!z&&(w||g)){if(typeof fetch=="function"&&!Oo(y))return fetch(y,{credentials:"same-origin"}).then(v=>{if(!v.ok)throw`failed to load wasm binary file at '${y}'`;return v.arrayBuffer()}).catch(()=>Do(y));if($)return new Promise((v,C)=>{$(y,O=>v(new Uint8Array(O)),C)})}return Promise.resolve().then(()=>Do(y))}(s).then(y=>WebAssembly.instantiate(y,c)).then(f,y=>{Y(`failed to asynchronously prepare wasm: ${y}`),Dt(y)})}function Bo(){return{a:{M:Wl,za:Nl,b:Fl,$:Lo,z:jo,pa:Ko,X:Xo,Z:Zo,qa:Qo,na:Jo,ga:ei,ma:ti,J:ri,Y:ni,V:oi,oa:ii,W:ai,va:ql,D:Kl,P:Yl,O:Zl,C:Jl,s:ec,p:tc,E:rc,y:dc,Q:lc,ta:cc,ja:pc,T:mc,aa:fc,F:hc,ia:In,sa:gc,u:yc,B:vc,o:$c,m:xc,c:xn,n:Sc,k:Tc,Aa:Ac,r:Ec,f:kc,v:Pc,l:Oc,g:Dc,i:zc,j:Bc,h:Rc,e:Mc,da:Uc,ea:Vc,fa:Nc,ba:vi,ca:$i,S:Wc,d:Hc,N:Gc,G:Lc,K:Fc,w:qc,ra:jc,U:Kc,t:xi,x:Yc,L:Xc,R:Zc,ya:Qc,xa:Jc,ka:Ci,la:Ti,_:bn,A:Ai,I:Ei,ha:ki,H:Pi,a:le,wa:yn,ua:zi,q:rp}}}var fn={849460:(s,c,f,y)=>{if(u===void 0||!u.Fb)return 1;if((s=Ae(s>>>0)).startsWith("./")&&(s=s.substring(2)),!(s=u.Fb.get(s)))return 2;if(y>>>=0,(c>>>=0)+(f>>>=0)>s.byteLength)return 3;try{return r().set(s.subarray(c,c+f),y>>>0),0}catch{return 4}},849961:()=>{u.Ub()},849992:()=>{u.Vb()},850021:()=>{u.Zb()},850046:s=>u.Tb(s),850079:s=>u.Xb(s),850111:(s,c,f)=>{u.Nb(s,c,f,!0)},850150:(s,c,f)=>{u.Nb(s,c,f)},850183:()=>typeof wasmOffsetConverter<"u",850240:s=>{u.jb("Abs",s,void 0)},850291:s=>{u.jb("Neg",s,void 0)},850342:s=>{u.jb("Floor",s,void 0)},850395:s=>{u.jb("Ceil",s,void 0)},850447:s=>{u.jb("Reciprocal",s,void 0)},850505:s=>{u.jb("Sqrt",s,void 0)},850557:s=>{u.jb("Exp",s,void 0)},850608:s=>{u.jb("Erf",s,void 0)},850659:s=>{u.jb("Sigmoid",s,void 0)},850714:(s,c,f)=>{u.jb("HardSigmoid",s,{alpha:c,beta:f})},850793:s=>{u.jb("Log",s,void 0)},850844:s=>{u.jb("Sin",s,void 0)},850895:s=>{u.jb("Cos",s,void 0)},850946:s=>{u.jb("Tan",s,void 0)},850997:s=>{u.jb("Asin",s,void 0)},851049:s=>{u.jb("Acos",s,void 0)},851101:s=>{u.jb("Atan",s,void 0)},851153:s=>{u.jb("Sinh",s,void 0)},851205:s=>{u.jb("Cosh",s,void 0)},851257:s=>{u.jb("Asinh",s,void 0)},851310:s=>{u.jb("Acosh",s,void 0)},851363:s=>{u.jb("Atanh",s,void 0)},851416:s=>{u.jb("Tanh",s,void 0)},851468:s=>{u.jb("Not",s,void 0)},851519:(s,c,f)=>{u.jb("Clip",s,{min:c,max:f})},851588:s=>{u.jb("Clip",s,void 0)},851640:(s,c)=>{u.jb("Elu",s,{alpha:c})},851698:s=>{u.jb("Relu",s,void 0)},851750:(s,c)=>{u.jb("LeakyRelu",s,{alpha:c})},851814:(s,c)=>{u.jb("ThresholdedRelu",s,{alpha:c})},851884:(s,c)=>{u.jb("Cast",s,{to:c})},851942:s=>{u.jb("Add",s,void 0)},851993:s=>{u.jb("Sub",s,void 0)},852044:s=>{u.jb("Mul",s,void 0)},852095:s=>{u.jb("Div",s,void 0)},852146:s=>{u.jb("Pow",s,void 0)},852197:s=>{u.jb("Equal",s,void 0)},852250:s=>{u.jb("Greater",s,void 0)},852305:s=>{u.jb("GreaterOrEqual",s,void 0)},852367:s=>{u.jb("Less",s,void 0)},852419:s=>{u.jb("LessOrEqual",s,void 0)},852478:(s,c,f,y,v)=>{u.jb("ReduceMean",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},852637:(s,c,f,y,v)=>{u.jb("ReduceMax",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},852795:(s,c,f,y,v)=>{u.jb("ReduceMin",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},852953:(s,c,f,y,v)=>{u.jb("ReduceProd",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},853112:(s,c,f,y,v)=>{u.jb("ReduceSum",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},853270:(s,c,f,y,v)=>{u.jb("ReduceL1",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},853427:(s,c,f,y,v)=>{u.jb("ReduceL2",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},853584:(s,c,f,y,v)=>{u.jb("ReduceLogSum",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},853745:(s,c,f,y,v)=>{u.jb("ReduceSumSquare",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},853909:(s,c,f,y,v)=>{u.jb("ReduceLogSumExp",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},854073:s=>{u.jb("Where",s,void 0)},854126:(s,c,f)=>{u.jb("Transpose",s,{perm:c?Array.from(i().subarray(c>>>0,f>>>0)):[]})},854234:(s,c,f,y)=>{u.jb("DepthToSpace",s,{blocksize:c,mode:Ae(f),format:y?"NHWC":"NCHW"})},854367:(s,c,f,y)=>{u.jb("DepthToSpace",s,{blocksize:c,mode:Ae(f),format:y?"NHWC":"NCHW"})},854500:(s,c,f,y,v,C,O,B,N,L,j,me,he,P,ee)=>{u.jb("ConvTranspose",s,{format:N?"NHWC":"NCHW",autoPad:c,dilations:[f],group:y,kernelShape:[v],pads:[C,O],strides:[B],wIsConst:()=>!!t()[L>>>0],outputPadding:j?Array.from(i().subarray(j>>>0,me>>>0)):[],outputShape:he?Array.from(i().subarray(he>>>0,P>>>0)):[],activation:Ae(ee)})},854901:(s,c,f,y,v,C,O,B,N,L,j,me,he,P)=>{u.jb("ConvTranspose",s,{format:B?"NHWC":"NCHW",autoPad:c,dilations:Array.from(i().subarray(f>>>0,2+(f>>>0)>>>0)),group:y,kernelShape:Array.from(i().subarray(v>>>0,2+(v>>>0)>>>0)),pads:Array.from(i().subarray(C>>>0,4+(C>>>0)>>>0)),strides:Array.from(i().subarray(O>>>0,2+(O>>>0)>>>0)),wIsConst:()=>!!t()[N>>>0],outputPadding:L?Array.from(i().subarray(L>>>0,j>>>0)):[],outputShape:me?Array.from(i().subarray(me>>>0,he>>>0)):[],activation:Ae(P)})},855466:(s,c,f,y,v,C,O,B,N,L,j,me,he,P,ee)=>{u.jb("ConvTranspose",s,{format:N?"NHWC":"NCHW",autoPad:c,dilations:[f],group:y,kernelShape:[v],pads:[C,O],strides:[B],wIsConst:()=>!!t()[L>>>0],outputPadding:j?Array.from(i().subarray(j>>>0,me>>>0)):[],outputShape:he?Array.from(i().subarray(he>>>0,P>>>0)):[],activation:Ae(ee)})},855867:(s,c,f,y,v,C,O,B,N,L,j,me,he,P)=>{u.jb("ConvTranspose",s,{format:B?"NHWC":"NCHW",autoPad:c,dilations:Array.from(i().subarray(f>>>0,2+(f>>>0)>>>0)),group:y,kernelShape:Array.from(i().subarray(v>>>0,2+(v>>>0)>>>0)),pads:Array.from(i().subarray(C>>>0,4+(C>>>0)>>>0)),strides:Array.from(i().subarray(O>>>0,2+(O>>>0)>>>0)),wIsConst:()=>!!t()[N>>>0],outputPadding:L?Array.from(i().subarray(L>>>0,j>>>0)):[],outputShape:me?Array.from(i().subarray(me>>>0,he>>>0)):[],activation:Ae(P)})},856432:(s,c)=>{u.jb("GlobalAveragePool",s,{format:c?"NHWC":"NCHW"})},856523:(s,c,f,y,v,C,O,B,N,L,j,me,he,P,ee,ye)=>{u.jb("AveragePool",s,{format:ye?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:y,storage_order:v,dilations:[C,O],kernel_shape:[B,N],pads:[L,j,me,he],strides:[P,ee]})},856807:(s,c)=>{u.jb("GlobalAveragePool",s,{format:c?"NHWC":"NCHW"})},856898:(s,c,f,y,v,C,O,B,N,L,j,me,he,P,ee,ye)=>{u.jb("AveragePool",s,{format:ye?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:y,storage_order:v,dilations:[C,O],kernel_shape:[B,N],pads:[L,j,me,he],strides:[P,ee]})},857182:(s,c)=>{u.jb("GlobalMaxPool",s,{format:c?"NHWC":"NCHW"})},857269:(s,c,f,y,v,C,O,B,N,L,j,me,he,P,ee,ye)=>{u.jb("MaxPool",s,{format:ye?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:y,storage_order:v,dilations:[C,O],kernel_shape:[B,N],pads:[L,j,me,he],strides:[P,ee]})},857549:(s,c)=>{u.jb("GlobalMaxPool",s,{format:c?"NHWC":"NCHW"})},857636:(s,c,f,y,v,C,O,B,N,L,j,me,he,P,ee,ye)=>{u.jb("MaxPool",s,{format:ye?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:y,storage_order:v,dilations:[C,O],kernel_shape:[B,N],pads:[L,j,me,he],strides:[P,ee]})},857916:(s,c,f,y,v)=>{u.jb("Gemm",s,{alpha:c,beta:f,transA:y,transB:v})},858020:s=>{u.jb("MatMul",s,void 0)},858074:(s,c,f,y)=>{u.jb("ArgMax",s,{keepDims:!!c,selectLastIndex:!!f,axis:y})},858182:(s,c,f,y)=>{u.jb("ArgMin",s,{keepDims:!!c,selectLastIndex:!!f,axis:y})},858290:(s,c)=>{u.jb("Softmax",s,{axis:c})},858353:(s,c)=>{u.jb("Concat",s,{axis:c})},858413:(s,c,f,y,v)=>{u.jb("Split",s,{axis:c,numOutputs:f,splitSizes:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},858553:s=>{u.jb("Expand",s,void 0)},858607:(s,c)=>{u.jb("Gather",s,{axis:Number(c)})},858678:(s,c)=>{u.jb("GatherElements",s,{axis:Number(c)})},858757:(s,c,f,y,v,C,O,B,N,L,j)=>{u.jb("Resize",s,{antialias:c,axes:f?Array.from(i().subarray(f>>>0,y>>>0)):[],coordinateTransformMode:Ae(v),cubicCoeffA:C,excludeOutside:O,extrapolationValue:B,keepAspectRatioPolicy:Ae(N),mode:Ae(L),nearestMode:Ae(j)})},859103:(s,c,f,y,v,C,O)=>{u.jb("Slice",s,{starts:c?Array.from(i().subarray(c>>>0,f>>>0)):[],ends:y?Array.from(i().subarray(y>>>0,v>>>0)):[],axes:C?Array.from(i().subarray(C>>>0,O>>>0)):[]})},859319:s=>{u.jb("Tile",s,void 0)},859371:(s,c,f)=>{u.jb("InstanceNormalization",s,{epsilon:c,format:f?"NHWC":"NCHW"})},859485:(s,c,f)=>{u.jb("InstanceNormalization",s,{epsilon:c,format:f?"NHWC":"NCHW"})},859599:s=>{u.jb("Range",s,void 0)},859652:(s,c)=>{u.jb("Einsum",s,{equation:Ae(c)})},859733:(s,c,f,y,v)=>{u.jb("Pad",s,{mode:c,value:f,pads:y?Array.from(i().subarray(y>>>0,v>>>0)):[]})},859860:(s,c,f,y,v,C)=>{u.jb("BatchNormalization",s,{epsilon:c,momentum:f,spatial:!!v,trainingMode:!!y,format:C?"NHWC":"NCHW"})},860029:(s,c,f,y,v,C)=>{u.jb("BatchNormalization",s,{epsilon:c,momentum:f,spatial:!!v,trainingMode:!!y,format:C?"NHWC":"NCHW"})},860198:(s,c,f)=>{u.jb("CumSum",s,{exclusive:Number(c),reverse:Number(f)})},860295:(s,c,f,y,v,C,O,B,N)=>{u.jb("Attention",s,{numHeads:c,isUnidirectional:f,maskFilterValue:y,scale:v,doRotary:C,qkvHiddenSizes:O?Array.from(i().subarray(Number(B)>>>0,Number(B)+O>>>0)):[],pastPresentShareBuffer:!!N})},860567:s=>{u.jb("BiasAdd",s,void 0)},860622:s=>{u.jb("BiasSplitGelu",s,void 0)},860683:s=>{u.jb("FastGelu",s,void 0)},860739:(s,c,f,y,v,C,O,B,N,L,j,me,he,P,ee,ye)=>{u.jb("Conv",s,{format:me?"NHWC":"NCHW",auto_pad:c,dilations:f?Array.from(i().subarray(f>>>0,y>>>0)):[],group:v,kernel_shape:C?Array.from(i().subarray(C>>>0,O>>>0)):[],pads:B?Array.from(i().subarray(B>>>0,N>>>0)):[],strides:L?Array.from(i().subarray(L>>>0,j>>>0)):[],w_is_const:()=>!!t()[he>>>0],activation:Ae(P),activation_params:ee?Array.from(l().subarray(ee>>>0,ye>>>0)):[]})},861235:s=>{u.jb("Gelu",s,void 0)},861287:(s,c,f,y)=>{u.jb("GroupQueryAttention",s,{numHeads:c,kvNumHeads:f,scale:y})},861400:(s,c,f,y)=>{u.jb("LayerNormalization",s,{axis:c,epsilon:f,simplified:!!y})},861511:(s,c,f,y)=>{u.jb("LayerNormalization",s,{axis:c,epsilon:f,simplified:!!y})},861622:(s,c,f,y,v,C)=>{u.jb("MatMulNBits",s,{k:c,n:f,accuracyLevel:y,bits:v,blockSize:C})},861749:(s,c,f,y,v,C)=>{u.jb("MultiHeadAttention",s,{numHeads:c,isUnidirectional:f,maskFilterValue:y,scale:v,doRotary:C})},861908:(s,c)=>{u.jb("QuickGelu",s,{alpha:c})},861972:(s,c,f,y,v)=>{u.jb("RotaryEmbedding",s,{interleaved:!!c,numHeads:f,rotaryEmbeddingDim:y,scale:v})},862111:(s,c,f)=>{u.jb("SkipLayerNormalization",s,{epsilon:c,simplified:!!f})},862213:s=>{u.Yb(s)},862247:(s,c)=>u.$b(s,c,u.Eb.bc,u.Eb.errors),862359:(s,c,f)=>{u.jb("SkipLayerNormalization",s,{epsilon:c,simplified:!!f})}};function Nl(s,c,f){return hi(async()=>{await u.Wb(s,c,f)})}function Wl(){return typeof wasmOffsetConverter<"u"}function hn(s){this.name="ExitStatus",this.message=`Program terminated with exit(${s})`,this.status=s}var gn=s=>{s.terminate(),s.onmessage=()=>{}},Ro=s=>{ct.length==0&&(Ho(),Wo(ct[0]));var c=ct.pop();if(!c)return 6;wt.push(c),Ze[s.Ab]=c,c.Ab=s.Ab;var f={cmd:"run",start_routine:s.cc,arg:s.Pb,pthread_ptr:s.Ab};return c.postMessage(f,s.ic),0},bt=0,xe=(s,c,...f)=>{for(var y=2*f.length,v=Bn(),C=zn(8*y),O=C>>>3,B=0;B<f.length;B++){var N=f[B];typeof N=="bigint"?(F[O+2*B]=1n,F[O+2*B+1]=N):(F[O+2*B]=0n,d()[O+2*B+1>>>0]=N)}return s=Vi(s,0,y,C,c),yr(v),s};function yn(s){if(b)return xe(0,1,s);if(de=s,!(0<bt)){for(var c of wt)gn(c);for(c of ct)gn(c);ct=[],wt=[],Ze=[],Se=!0}A(s,new hn(s))}function Mo(s){if(b)return xe(1,0,s);bn(s)}var bn=s=>{if(de=s,b)throw Mo(s),"unwind";yn(s)},ct=[],wt=[],Uo=[],Ze={},Vo=s=>{var c=s.Ab;delete Ze[c],ct.push(s),wt.splice(wt.indexOf(s),1),s.Ab=0,Dn(c)};function No(){Uo.forEach(s=>s())}var Wo=s=>new Promise(c=>{s.onmessage=v=>{var C=(v=v.data).cmd;if(v.targetThread&&v.targetThread!=Mt()){var O=Ze[v.targetThread];O?O.postMessage(v,v.transferList):Y(`Internal error! Worker sent a message "${C}" to target pthread ${v.targetThread}, but that thread no longer exists!`)}else C==="checkMailbox"?sr():C==="spawnThread"?Ro(v):C==="cleanupThread"?Vo(Ze[v.thread]):C==="killThread"?(v=v.thread,C=Ze[v],delete Ze[v],gn(C),Dn(v),wt.splice(wt.indexOf(C),1),C.Ab=0):C==="cancelThread"?Ze[v.thread].postMessage({cmd:"cancel"}):C==="loaded"?(s.loaded=!0,c(s)):C==="alert"?alert(`Thread ${v.threadId}: ${v.text}`):v.target==="setimmediate"?s.postMessage(v):C==="callHandler"?u[v.handler](...v.args):C&&Y(`worker sent an unknown command ${C}`)},s.onerror=v=>{throw Y(`worker sent an error! ${v.filename}:${v.lineno}: ${v.message}`),v};var f,y=[];for(f of[])u.hasOwnProperty(f)&&y.push(f);s.postMessage({cmd:"load",handlers:y,wasmMemory:le,wasmModule:K})});function Ho(){var s=new Worker(new URL(import.meta.url),{type:"module",workerData:"em-pthread",name:"em-pthread"});ct.push(s)}var ar=s=>{for(;0<s.length;)s.shift()(u)},Hl=()=>{var s=Mt(),c=a()[s+52>>>2>>>0];s=a()[s+56>>>2>>>0],Wi(c,c-s),yr(c)},Gl=(s,c)=>{bt=0,s=Hi(s,c),0<bt?de=s:gr(s)};class Ll{constructor(c){this.Ib=c-24}}function Fl(s,c,f){var y=new Ll(s>>>=0);throw c>>>=0,f>>>=0,a()[y.Ib+16>>>2>>>0]=0,a()[y.Ib+4>>>2>>>0]=c,a()[y.Ib+8>>>2>>>0]=f,s}function Go(s,c,f,y){return b?xe(2,1,s,c,f,y):Lo(s,c,f,y)}function Lo(s,c,f,y){if(s>>>=0,c>>>=0,f>>>=0,y>>>=0,x===void 0)return Y("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var v=[];return b&&v.length===0?Go(s,c,f,y):(s={cc:f,Ab:s,Pb:y,ic:v},b?(s.Lb="spawnThread",postMessage(s,v),0):Ro(s))}var Fo=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,qo=(s,c,f)=>{var y=(c>>>=0)+f;for(f=c;s[f]&&!(f>=y);)++f;if(16<f-c&&s.buffer&&Fo)return Fo.decode(s.buffer instanceof x?s.slice(c,f):s.subarray(c,f));for(y="";c<f;){var v=s[c++];if(128&v){var C=63&s[c++];if((224&v)==192)y+=String.fromCharCode((31&v)<<6|C);else{var O=63&s[c++];65536>(v=(240&v)==224?(15&v)<<12|C<<6|O:(7&v)<<18|C<<12|O<<6|63&s[c++])?y+=String.fromCharCode(v):(v-=65536,y+=String.fromCharCode(55296|v>>10,56320|1023&v))}}else y+=String.fromCharCode(v)}return y},Ae=(s,c)=>(s>>>=0)?qo(r(),s,c):"";function jo(s,c,f){return b?xe(3,1,s,c,f):0}function Ko(s,c){if(b)return xe(4,1,s,c)}var wn=s=>{for(var c=0,f=0;f<s.length;++f){var y=s.charCodeAt(f);127>=y?c++:2047>=y?c+=2:55296<=y&&57343>=y?(c+=4,++f):c+=3}return c},Yo=(s,c,f,y)=>{if(!(0<y))return 0;var v=f>>>=0;y=f+y-1;for(var C=0;C<s.length;++C){var O=s.charCodeAt(C);if(55296<=O&&57343>=O&&(O=65536+((1023&O)<<10)|1023&s.charCodeAt(++C)),127>=O){if(f>=y)break;c[f++>>>0]=O}else{if(2047>=O){if(f+1>=y)break;c[f++>>>0]=192|O>>6}else{if(65535>=O){if(f+2>=y)break;c[f++>>>0]=224|O>>12}else{if(f+3>=y)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-v},zt=(s,c,f)=>Yo(s,r(),c,f);function Xo(s,c){if(b)return xe(5,1,s,c)}function Zo(s,c,f){if(b)return xe(6,1,s,c,f)}function Qo(s,c,f){return b?xe(7,1,s,c,f):0}function Jo(s,c){if(b)return xe(8,1,s,c)}function ei(s,c,f){if(b)return xe(9,1,s,c,f)}function ti(s,c,f,y){if(b)return xe(10,1,s,c,f,y)}function ri(s,c,f,y){if(b)return xe(11,1,s,c,f,y)}function ni(s,c,f,y){if(b)return xe(12,1,s,c,f,y)}function oi(s){if(b)return xe(13,1,s)}function ii(s,c){if(b)return xe(14,1,s,c)}function ai(s,c,f){if(b)return xe(15,1,s,c,f)}var si,pt,ql=()=>{Dt("")},Qe=s=>{for(var c="";r()[s>>>0];)c+=si[r()[s++>>>0]];return c},vn={},$n={},jl={};function st(s,c,f={}){if(!("argPackAdvance"in c))throw new TypeError("registerType registeredInstance requires argPackAdvance");return function(y,v,C={}){var O=v.name;if(!y)throw new pt(`type "${O}" must have a positive integer typeid pointer`);if($n.hasOwnProperty(y)){if(C.Rb)return;throw new pt(`Cannot register type '${O}' twice`)}$n[y]=v,delete jl[y],vn.hasOwnProperty(y)&&(v=vn[y],delete vn[y],v.forEach(B=>B()))}(s,c,f)}var ui=(s,c,f)=>{switch(c){case 1:return f?y=>t()[y>>>0]:y=>r()[y>>>0];case 2:return f?y=>n()[y>>>1>>>0]:y=>o()[y>>>1>>>0];case 4:return f?y=>i()[y>>>2>>>0]:y=>a()[y>>>2>>>0];case 8:return f?y=>F[y>>>3]:y=>ce[y>>>3];default:throw new TypeError(`invalid integer width (${c}): ${s}`)}};function Kl(s,c,f){f>>>=0,st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:y=>y,toWireType:function(y,v){if(typeof v!="bigint"&&typeof v!="number")throw v=v===null?"null":(y=typeof v)=="object"||y==="array"||y==="function"?v.toString():""+v,new TypeError(`Cannot convert "${v}" to ${this.name}`);return typeof v=="number"&&(v=BigInt(v)),v},argPackAdvance:mt,readValueFromPointer:ui(c,f,c.indexOf("u")==-1),Db:null})}var mt=8;function Yl(s,c,f,y){st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:function(v){return!!v},toWireType:function(v,C){return C?f:y},argPackAdvance:mt,readValueFromPointer:function(v){return this.fromWireType(r()[v>>>0])},Db:null})}var _n=[],ut=[];function xn(s){9<(s>>>=0)&&--ut[s+1]==0&&(ut[s]=void 0,_n.push(s))}var qe=s=>{if(!s)throw new pt("Cannot use deleted val. handle = "+s);return ut[s]},je=s=>{switch(s){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let c=_n.pop()||ut.length;return ut[c]=s,ut[c+1]=1,c}};function Sn(s){return this.fromWireType(a()[s>>>2>>>0])}var Xl={name:"emscripten::val",fromWireType:s=>{var c=qe(s);return xn(s),c},toWireType:(s,c)=>je(c),argPackAdvance:mt,readValueFromPointer:Sn,Db:null};function Zl(s){return st(s>>>0,Xl)}var Ql=(s,c)=>{switch(c){case 4:return function(f){return this.fromWireType(l()[f>>>2>>>0])};case 8:return function(f){return this.fromWireType(d()[f>>>3>>>0])};default:throw new TypeError(`invalid float width (${c}): ${s}`)}};function Jl(s,c,f){f>>>=0,st(s>>>=0,{name:c=Qe(c>>>0),fromWireType:y=>y,toWireType:(y,v)=>v,argPackAdvance:mt,readValueFromPointer:Ql(c,f),Db:null})}function ec(s,c,f,y,v){if(s>>>=0,f>>>=0,c=Qe(c>>>0),v===-1&&(v=4294967295),v=B=>B,y===0){var C=32-8*f;v=B=>B<<C>>>C}var O=c.includes("unsigned")?function(B,N){return N>>>0}:function(B,N){return N};st(s,{name:c,fromWireType:v,toWireType:O,argPackAdvance:mt,readValueFromPointer:ui(c,f,y!==0),Db:null})}function tc(s,c,f){function y(C){var O=a()[C>>>2>>>0];return C=a()[C+4>>>2>>>0],new v(t().buffer,C,O)}var v=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][c];st(s>>>=0,{name:f=Qe(f>>>0),fromWireType:y,argPackAdvance:mt,readValueFromPointer:y},{Rb:!0})}function rc(s,c){s>>>=0;var f=(c=Qe(c>>>0))==="std::string";st(s,{name:c,fromWireType:function(y){var v=a()[y>>>2>>>0],C=y+4;if(f)for(var O=C,B=0;B<=v;++B){var N=C+B;if(B==v||r()[N>>>0]==0){if(O=Ae(O,N-O),L===void 0)var L=O;else L+=String.fromCharCode(0),L+=O;O=N+1}}else{for(L=Array(v),B=0;B<v;++B)L[B]=String.fromCharCode(r()[C+B>>>0]);L=L.join("")}return et(y),L},toWireType:function(y,v){v instanceof ArrayBuffer&&(v=new Uint8Array(v));var C=typeof v=="string";if(!(C||v instanceof Uint8Array||v instanceof Uint8ClampedArray||v instanceof Int8Array))throw new pt("Cannot pass non-string to std::string");var O=f&&C?wn(v):v.length,B=hr(4+O+1),N=B+4;if(a()[B>>>2>>>0]=O,f&&C)zt(v,N,O+1);else if(C)for(C=0;C<O;++C){var L=v.charCodeAt(C);if(255<L)throw et(N),new pt("String has UTF-16 code units that do not fit in 8 bits");r()[N+C>>>0]=L}else for(C=0;C<O;++C)r()[N+C>>>0]=v[C];return y!==null&&y.push(et,B),B},argPackAdvance:mt,readValueFromPointer:Sn,Db(y){et(y)}})}var di=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,nc=(s,c)=>{for(var f=s>>1,y=f+c/2;!(f>=y)&&o()[f>>>0];)++f;if(32<(f<<=1)-s&&di)return di.decode(r().slice(s,f));for(f="",y=0;!(y>=c/2);++y){var v=n()[s+2*y>>>1>>>0];if(v==0)break;f+=String.fromCharCode(v)}return f},oc=(s,c,f)=>{if(f??=2147483647,2>f)return 0;var y=c;f=(f-=2)<2*s.length?f/2:s.length;for(var v=0;v<f;++v){var C=s.charCodeAt(v);n()[c>>>1>>>0]=C,c+=2}return n()[c>>>1>>>0]=0,c-y},ic=s=>2*s.length,ac=(s,c)=>{for(var f=0,y="";!(f>=c/4);){var v=i()[s+4*f>>>2>>>0];if(v==0)break;++f,65536<=v?(v-=65536,y+=String.fromCharCode(55296|v>>10,56320|1023&v)):y+=String.fromCharCode(v)}return y},sc=(s,c,f)=>{if(c>>>=0,f??=2147483647,4>f)return 0;var y=c;f=y+f-4;for(var v=0;v<s.length;++v){var C=s.charCodeAt(v);if(55296<=C&&57343>=C&&(C=65536+((1023&C)<<10)|1023&s.charCodeAt(++v)),i()[c>>>2>>>0]=C,(c+=4)+4>f)break}return i()[c>>>2>>>0]=0,c-y},uc=s=>{for(var c=0,f=0;f<s.length;++f){var y=s.charCodeAt(f);55296<=y&&57343>=y&&++f,c+=4}return c};function dc(s,c,f){if(s>>>=0,c>>>=0,f=Qe(f>>>=0),c===2)var y=nc,v=oc,C=ic,O=B=>o()[B>>>1>>>0];else c===4&&(y=ac,v=sc,C=uc,O=B=>a()[B>>>2>>>0]);st(s,{name:f,fromWireType:B=>{for(var N,L=a()[B>>>2>>>0],j=B+4,me=0;me<=L;++me){var he=B+4+me*c;me!=L&&O(he)!=0||(j=y(j,he-j),N===void 0?N=j:(N+=String.fromCharCode(0),N+=j),j=he+c)}return et(B),N},toWireType:(B,N)=>{if(typeof N!="string")throw new pt(`Cannot pass non-string to C++ string type ${f}`);var L=C(N),j=hr(4+L+c);return a()[j>>>2>>>0]=L/c,v(N,j+4,L+c),B!==null&&B.push(et,j),j},argPackAdvance:mt,readValueFromPointer:Sn,Db(B){et(B)}})}function lc(s,c){st(s>>>=0,{Sb:!0,name:c=Qe(c>>>0),argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})}var cc=()=>1;function pc(s){On(s>>>0,!g,1,!w,131072,!1),No()}var li=s=>{if(!Se)try{if(s(),!(0<bt))try{b?gr(de):bn(de)}catch(c){c instanceof hn||c=="unwind"||A(1,c)}}catch(c){c instanceof hn||c=="unwind"||A(1,c)}};function In(s){s>>>=0,typeof Atomics.jc=="function"&&(Atomics.jc(i(),s>>>2,s).value.then(sr),s+=128,Atomics.store(i(),s>>>2,1))}var sr=()=>{var s=Mt();s&&(In(s),li(Ni))};function mc(s,c){(s>>>=0)==c>>>0?setTimeout(sr):b?postMessage({targetThread:s,cmd:"checkMailbox"}):(s=Ze[s])&&s.postMessage({cmd:"checkMailbox"})}var Cn=[];function fc(s,c,f,y,v){for(c>>>=0,y/=2,Cn.length=y,f=v>>>0>>>3,v=0;v<y;v++)Cn[v]=F[f+2*v]?F[f+2*v+1]:d()[f+2*v+1>>>0];return(c?fn[c]:np[s])(...Cn)}function hc(s){s>>>=0,b?postMessage({cmd:"cleanupThread",thread:s}):Vo(Ze[s])}function gc(s){}var Tn=(s,c)=>{var f=$n[s];if(f===void 0)throw s=Ri(s),f=Qe(s),et(s),new pt(`${c} has unknown type ${f}`);return f},ci=(s,c,f)=>{var y=[];return s=s.toWireType(y,f),y.length&&(a()[c>>>2>>>0]=je(y)),s};function yc(s,c,f){return c>>>=0,f>>>=0,s=qe(s>>>0),c=Tn(c,"emval::as"),ci(c,f,s)}var ur=s=>{try{s()}catch(c){Dt(c)}},ft=0,Je=null,pi=0,dr=[],mi={},fi={},bc=0,An=null,wc=[];function hi(s){return function(c){if(!Se){if(ft===0){var f=!1,y=!1;c((v=0)=>{if(!Se&&(pi=v,f=!0,y)){ft=2,ur(()=>Fi(Je)),typeof Browser<"u"&&Browser.Jb.Qb&&Browser.Jb.resume(),v=!1;try{var C=function(){var N=i()[Je+8>>>2>>>0];return N=q[fi[N]],--bt,N()}()}catch(N){C=N,v=!0}var O=!1;if(!Je){var B=An;B&&(An=null,(v?B.reject:B.resolve)(C),O=!0)}if(v&&!O)throw C}}),y=!0,f||(ft=1,Je=function(){var v=hr(65548),C=v+12;a()[v>>>2>>>0]=C,a()[v+4>>>2>>>0]=C+65536,C=dr[0];var O=mi[C];return O===void 0&&(O=bc++,mi[C]=O,fi[O]=C),C=O,i()[v+8>>>2>>>0]=C,v}(),typeof Browser<"u"&&Browser.Jb.Qb&&Browser.Jb.pause(),ur(()=>Gi(Je)))}else ft===2?(ft=0,ur(qi),et(Je),Je=null,wc.forEach(li)):Dt(`invalid state: ${ft}`);return pi}}(c=>{s().then(c)})}function vc(s){return s>>>=0,hi(()=>(s=qe(s)).then(je))}var lr=[];function $c(s,c,f,y){return f>>>=0,y>>>=0,(s=lr[s>>>0])(null,c=qe(c>>>0),f,y)}var _c={},cr=s=>{var c=_c[s];return c===void 0?Qe(s):c};function xc(s,c,f,y,v){return f>>>=0,y>>>=0,v>>>=0,(s=lr[s>>>0])(c=qe(c>>>0),c[f=cr(f)],y,v)}var gi=()=>typeof globalThis=="object"?globalThis:Function("return this")();function Sc(s){return(s>>>=0)==0?je(gi()):(s=cr(s),je(gi()[s]))}var Ic=s=>{var c=lr.length;return lr.push(s),c},Cc=(s,c)=>{for(var f=Array(s),y=0;y<s;++y)f[y]=Tn(a()[c+4*y>>>2>>>0],"parameter "+y);return f},yi=(s,c)=>Object.defineProperty(c,"name",{value:s});function Tc(s,c,f){var y=(c=Cc(s,c>>>0)).shift();s--;var v=`return function (obj, func, destructorsRef, args) {
-`,C=0,O=[];f===0&&O.push("obj");for(var B=["retType"],N=[y],L=0;L<s;++L)O.push("arg"+L),B.push("argType"+L),N.push(c[L]),v+=`  var arg${L} = argType${L}.readValueFromPointer(args${C?"+"+C:""});
-`,C+=c[L].argPackAdvance;return v+=`  var rv = ${f===1?"new func":"func.call"}(${O.join(", ")});
-`,y.Sb||(B.push("emval_returnValue"),N.push(ci),v+=`  return emval_returnValue(retType, destructorsRef, rv);
+var On=Object.defineProperty;var Jc=Object.getOwnPropertyDescriptor;var ep=Object.getOwnPropertyNames;var tp=Object.prototype.hasOwnProperty;var zn=(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 M=(e,t)=>()=>(e&&(t=e(e=0)),t);var Lt=(e,t)=>{for(var r in t)On(e,r,{get:t[r],enumerable:!0})},rp=(e,t,r,n)=>{if(t&&typeof t=="object"||typeof t=="function")for(let o of ep(t))!tp.call(e,o)&&o!==r&&On(e,o,{get:()=>t[o],enumerable:!(n=Jc(t,o))||n.enumerable});return e};var hr=e=>rp(On({},"__esModule",{value:!0}),e);var gr,bt,wt,np,yr,br=M(()=>{"use strict";gr=new Map,bt=[],wt=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let n=gr.get(e);if(n===void 0)gr.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 ${r}`)}if(r>=0){let o=bt.indexOf(e);o!==-1&&bt.splice(o,1);for(let i=0;i<bt.length;i++)if(gr.get(bt[i]).priority<=r){bt.splice(i,0,e);return}bt.push(e)}return}throw new TypeError("not a valid backend")},np=async e=>{let t=gr.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(n){return r||(t.error=`${n}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},yr=async e=>{let t=e.executionProviders||[],r=t.map(d=>typeof d=="string"?d:d.name),n=r.length===0?bt:r,o,i=[],a=new Set;for(let d of n){let p=await np(d);typeof p=="string"?i.push({name:d,err:p}):(o||(o=p),o===p&&a.add(d))}if(!o)throw new Error(`no available backend found. ERR: ${i.map(d=>`[${d.name}] ${d.err}`).join(", ")}`);for(let{name:d,err:p}of i)r.includes(d)&&console.warn(`removing requested execution provider "${d}" from session options because it is not available: ${p}`);let l=t.filter(d=>a.has(typeof d=="string"?d:d.name));return[o,new Proxy(e,{get:(d,p)=>p==="executionProviders"?l:Reflect.get(d,p)})]}});var Li=M(()=>{"use strict";br()});var Hi,Gi=M(()=>{"use strict";Hi="1.20.0-dev.20240816-b2d603abda"});var Fi,Ue,Bn=M(()=>{"use strict";Gi();Fi="warning",Ue={wasm:{},webgl:{},webgpu:{},versions:{common:Hi},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: ${e}`);Fi=e}},get logLevel(){return Fi}};Object.defineProperty(Ue,"logLevel",{enumerable:!0})});var fe,qi=M(()=>{"use strict";Bn();fe=Ue});var Ki,ji,Yi=M(()=>{"use strict";Ki=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let n=r.getContext("2d");if(n!=null){let o,i;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],i=e.dims[3]):(o=e.dims[3],i=e.dims[2]);let a=t?.format!==void 0?t.format:"RGB",l=t?.norm,d,p;l===void 0||l.mean===void 0?d=[255,255,255,255]:typeof l.mean=="number"?d=[l.mean,l.mean,l.mean,l.mean]:(d=[l.mean[0],l.mean[1],l.mean[2],0],l.mean[3]!==void 0&&(d[3]=l.mean[3])),l===void 0||l.bias===void 0?p=[0,0,0,0]:typeof l.bias=="number"?p=[l.bias,l.bias,l.bias,l.bias]:(p=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(p[3]=l.bias[3]));let m=i*o,u=0,h=m,w=m*2,g=-1;a==="RGBA"?(u=0,h=m,w=m*2,g=m*3):a==="RGB"?(u=0,h=m,w=m*2):a==="RBG"&&(u=0,w=m,h=m*2);for(let y=0;y<i;y++)for(let x=0;x<o;x++){let _=(e.data[u++]-p[0])*d[0],$=(e.data[h++]-p[1])*d[1],S=(e.data[w++]-p[2])*d[2],I=g===-1?255:(e.data[g++]-p[3])*d[3];n.fillStyle="rgba("+_+","+$+","+S+","+I+")",n.fillRect(x,y,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},ji=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let o,i,a;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],i=e.dims[1],a=e.dims[3]):(o=e.dims[3],i=e.dims[2],a=e.dims[1]);let l=t!==void 0&&t.format!==void 0?t.format:"RGB",d=t?.norm,p,m;d===void 0||d.mean===void 0?p=[255,255,255,255]:typeof d.mean=="number"?p=[d.mean,d.mean,d.mean,d.mean]:(p=[d.mean[0],d.mean[1],d.mean[2],255],d.mean[3]!==void 0&&(p[3]=d.mean[3])),d===void 0||d.bias===void 0?m=[0,0,0,0]:typeof d.bias=="number"?m=[d.bias,d.bias,d.bias,d.bias]:(m=[d.bias[0],d.bias[1],d.bias[2],0],d.bias[3]!==void 0&&(m[3]=d.bias[3]));let u=i*o;if(t!==void 0&&(t.format!==void 0&&a===4&&t.format!=="RGBA"||a===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let h=4,w=0,g=1,y=2,x=3,_=0,$=u,S=u*2,I=-1;l==="RGBA"?(_=0,$=u,S=u*2,I=u*3):l==="RGB"?(_=0,$=u,S=u*2):l==="RBG"&&(_=0,S=u,$=u*2),n=r.createImageData(o,i);for(let T=0;T<i*o;w+=h,g+=h,y+=h,x+=h,T++)n.data[w]=(e.data[_++]-m[0])*p[0],n.data[g]=(e.data[$++]-m[1])*p[1],n.data[y]=(e.data[S++]-m[2])*p[2],n.data[x]=I===-1?255:(e.data[I++]-m[3])*p[3]}else throw new Error("Can not access image data");return n}});var Dn,Xi,Zi,Qi,Ji,ea=M(()=>{"use strict";wr();Dn=(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 l=t.format!==void 0?t.format:"RGBA",d=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",p=r*n,m=d==="RGBA"?new Float32Array(p*4):new Float32Array(p*3),u=4,h=0,w=1,g=2,y=3,x=0,_=p,$=p*2,S=-1;l==="RGB"&&(u=3,h=0,w=1,g=2,y=-1),d==="RGBA"?S=p*3:d==="RBG"?(x=0,$=p,_=p*2):d==="BGR"&&($=0,_=p,x=p*2);for(let T=0;T<p;T++,h+=u,g+=u,w+=u,y+=u)m[x++]=(e[h]+a[0])/i[0],m[_++]=(e[w]+a[1])/i[1],m[$++]=(e[g]+a[2])/i[2],S!==-1&&y!==-1&&(m[S++]=(e[y]+a[3])/i[3]);return d==="RGBA"?new ze("float32",m,[1,4,r,n]):new ze("float32",m,[1,3,r,n])},Xi=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,l=t??{},d=()=>{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=>m instanceof HTMLCanvasElement||m instanceof OffscreenCanvas?m.getContext("2d"):null;if(r){let m=d();m.width=e.width,m.height=e.height;let u=p(m);if(u!=null){let h=e.height,w=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(h=t.resizedHeight,w=t.resizedWidth),t!==void 0){if(l=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");l.tensorFormat="RGBA",l.height=h,l.width=w}else l.tensorFormat="RGBA",l.height=h,l.width=w;u.drawImage(e,0,0),a=u.getImageData(0,0,w,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&&(l=t),l.format="RGBA",l.height=m,l.width=u,t!==void 0){let h=d();h.width=u,h.height=m;let w=p(h);if(w!=null)w.putImageData(e,0,0),a=w.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=d();m.width=e.width,m.height=e.height;let u=p(m);if(u!=null){let h=e.height,w=e.width;return u.drawImage(e,0,0,w,h),a=u.getImageData(0,0,w,h).data,l.height=h,l.width=w,Dn(a,l)}else throw new Error("Can not access image data")}else{if(i)return new Promise((m,u)=>{let h=d(),w=p(h);if(!e||!w)return u();let g=new Image;g.crossOrigin="Anonymous",g.src=e,g.onload=()=>{h.width=g.width,h.height=g.height,w.drawImage(g,0,0,h.width,h.height);let y=w.getImageData(0,0,h.width,h.height);l.height=h.height,l.width=h.width,m(Dn(y.data,l))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(a!==void 0)return Dn(a,l);throw new Error("Input data provided is not supported - aborted tensor creation")},Zi=(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})},Qi=(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})},Ji=(e,t,r)=>new ze({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})});var vt,Ht,ta,ra,na=M(()=>{"use strict";vt=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),Ht=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),ta=!1,ra=()=>{if(!ta){ta=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;e&&(vt.set("int64",BigInt64Array),Ht.set(BigInt64Array,"int64")),t&&(vt.set("uint64",BigUint64Array),Ht.set(BigUint64Array,"uint64")),r?(vt.set("float16",Float16Array),Ht.set(Float16Array,"float16")):vt.set("float16",Uint16Array)}}});var oa,ia,aa=M(()=>{"use strict";wr();oa=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},ia=(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});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}});var ze,wr=M(()=>{"use strict";Yi();ea();na();aa();ze=class{constructor(t,r,n){ra();let o,i;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,o=t.type,i=t.dims,t.location){case"cpu-pinned":{let l=vt.get(o);if(!l)throw new TypeError(`unsupported type "${o}" to create tensor from pinned buffer`);if(!(t.data instanceof l))throw new TypeError(`buffer should be of type ${l.name}`);this.cpuData=t.data;break}case"texture":{if(o!=="float32")throw new TypeError(`unsupported type "${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")throw new TypeError(`unsupported type "${o}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let l,d;if(typeof t=="string")if(o=t,d=n,t==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");l=r}else{let p=vt.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"?l=p.from(r,BigInt):l=p.from(r)}else if(r instanceof p)l=r;else throw new TypeError(`A ${o} tensor's data must be type of ${p}`)}else if(d=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",l=t;else if(p==="boolean")o="bool",l=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${p}.`)}else{let p=Ht.get(t.constructor);if(p===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);o=p,l=t}if(d===void 0)d=[l.length];else if(!Array.isArray(d))throw new TypeError("A tensor's dims must be a number array");i=d,this.cpuData=l,this.dataLocation="cpu"}let a=oa(i);if(this.cpuData&&a!==this.cpuData.length&&!((o==="uint4"||o==="int4")&&Math.ceil(a/2)===this.cpuData.length))throw new Error(`Tensor's size(${a}) does not match data length(${this.cpuData.length}).`);this.type=o,this.dims=i,this.size=a}static async fromImage(t,r){return Xi(t,r)}static fromTexture(t,r){return Zi(t,r)}static fromGpuBuffer(t,r){return Qi(t,r)}static fromPinnedBuffer(t,r,n){return Ji(t,r,n)}toDataURL(t){return Ki(this,t)}toImageData(t){return ji(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}async getData(t){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":{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.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 ia(this,t)}}});var Pe,vr=M(()=>{"use strict";wr();Pe=ze});var $r,sa,Ve,De,Rn=M(()=>{"use strict";Bn();$r=(e,t)=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||console.timeStamp(`${e}::ORT::${t}`)},sa=(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)}},Ve=e=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||sa("BEGIN",e)},De=e=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||sa("END",e)}});var _r,ua=M(()=>{"use strict";br();vr();Rn();_r=class e{constructor(t){this.handler=t}async run(t,r,n){Ve();let o={},i={};if(typeof t!="object"||t===null||t instanceof Pe||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 Pe)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: ${p}.`);o[p]=null}if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let p=!1,m=Object.getOwnPropertyNames(r);for(let u of this.outputNames)if(m.indexOf(u)!==-1){let h=r[u];(h===null||h instanceof Pe)&&(p=!0,a=!1,o[u]=h)}if(p){if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else i=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let p of this.inputNames)if(typeof t[p]>"u")throw new Error(`input '${p}' is missing in 'feeds'.`);if(a)for(let p of this.outputNames)o[p]=null;let l=await this.handler.run(t,o,i),d={};for(let p in l)if(Object.hasOwnProperty.call(l,p)){let m=l[p];m instanceof Pe?d[p]=m:d[p]=new Pe(m.type,m.data,m.dims)}return De(),d}async release(){return this.handler.dispose()}static async create(t,r,n,o){Ve();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[l,d]=await yr(a),p=await l.createInferenceSessionHandler(i,d);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 op,da=M(()=>{"use strict";ua();op=_r});var la=M(()=>{"use strict"});var ca=M(()=>{"use strict"});var pa=M(()=>{"use strict"});var ma=M(()=>{"use strict"});var ip,xr,fa=M(()=>{"use strict";br();vr();ip="Training backend could not be resolved. Make sure you're using the correct configuration & WebAssembly files.",xr=class e{constructor(t,r,n){this.handler=t,this.hasOptimizerModel=r,this.hasEvalModel=n}get trainingInputNames(){return this.handler.inputNames}get trainingOutputNames(){return this.handler.outputNames}get evalInputNames(){if(this.hasEvalModel)return this.handler.evalInputNames;throw new Error("This training session has no evalModel loaded.")}get evalOutputNames(){if(this.hasEvalModel)return this.handler.evalOutputNames;throw new Error("This training session has no evalModel loaded.")}static async create(t,r){let n=t.evalModel||"",o=t.optimizerModel||"",i=r||{},[a,l]=await yr(i);if(a.createTrainingSessionHandler){let d=await a.createTrainingSessionHandler(t.checkpointState,t.trainModel,n,o,l);return new e(d,!!t.optimizerModel,!!t.evalModel)}else throw new Error(ip)}typeNarrowingForRunStep(t,r,n,o,i){let a={},l={};if(typeof n!="object"||n===null||n instanceof Pe||Array.isArray(n))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let d=!0;if(typeof o=="object"){if(o===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(o instanceof Pe)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(o)){if(o.length===0)throw new TypeError("'fetches' cannot be an empty array.");d=!1;for(let p of o){if(typeof p!="string")throw new TypeError("'fetches' must be a string array or an object.");if(r.indexOf(p)===-1)throw new RangeError(`'fetches' contains invalid output name: ${p}.`);a[p]=null}if(typeof i=="object"&&i!==null)l=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else{let p=!1,m=Object.getOwnPropertyNames(o);for(let u of r)if(m.indexOf(u)!==-1){let h=o[u];(h===null||h instanceof Pe)&&(p=!0,d=!1,a[u]=h)}if(p){if(typeof i=="object"&&i!==null)l=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else l=o}}else if(typeof o<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let p of t)if(typeof n[p]>"u")throw new Error(`input '${p}' is missing in 'feeds'.`);if(d)for(let p of r)a[p]=null;return[a,l]}convertHandlerReturnTypeToMapOfTensors(t){let r={};for(let n in t)if(Object.hasOwnProperty.call(t,n)){let o=t[n];o instanceof Pe?r[n]=o:r[n]=new Pe(o.type,o.data,o.dims)}return r}async lazyResetGrad(){await this.handler.lazyResetGrad()}async runTrainStep(t,r,n){let[o,i]=this.typeNarrowingForRunStep(this.trainingInputNames,this.trainingOutputNames,t,r,n),a=await this.handler.runTrainStep(t,o,i);return this.convertHandlerReturnTypeToMapOfTensors(a)}async runOptimizerStep(t){if(this.hasOptimizerModel)await this.handler.runOptimizerStep(t||{});else throw new Error("This TrainingSession has no OptimizerModel loaded.")}async runEvalStep(t,r,n){if(this.hasEvalModel){let[o,i]=this.typeNarrowingForRunStep(this.evalInputNames,this.evalOutputNames,t,r,n),a=await this.handler.runEvalStep(t,o,i);return this.convertHandlerReturnTypeToMapOfTensors(a)}else throw new Error("This TrainingSession has no EvalModel loaded.")}async getParametersSize(t=!0){return this.handler.getParametersSize(t)}async loadParametersBuffer(t,r=!0){let n=await this.getParametersSize(r);if(t.length!==4*n)throw new Error("Size of the buffer passed into loadParametersBuffer must match the number of parameters in the model. Please use getParametersSize method to check.");return this.handler.loadParametersBuffer(t,r)}async getContiguousParameters(t=!0){return this.handler.getContiguousParameters(t)}async release(){return this.handler.dispose()}}});var ap,ha=M(()=>{"use strict";fa();ap=xr});var Mn={};Lt(Mn,{InferenceSession:()=>op,TRACE:()=>$r,TRACE_FUNC_BEGIN:()=>Ve,TRACE_FUNC_END:()=>De,Tensor:()=>Pe,TrainingSession:()=>ap,env:()=>fe,registerBackend:()=>wt});var qe=M(()=>{"use strict";Li();qi();da();vr();la();ca();Rn();pa();ma();ha()});var Sr=M(()=>{"use strict"});var wa={};Lt(wa,{default:()=>sp});var ya,ba,sp,va=M(()=>{"use strict";Un();$t();Gt();ya="ort-wasm-proxy-worker",ba=globalThis.self?.name===ya;ba&&(self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":Ir(r.wasm).then(()=>{Cr(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=Ft(n);postMessage({type:t,out:o});break}case"create":{let{model:n,options:o}=r;Ar(n,o).then(i=>{postMessage({type:t,out:i})},i=>{postMessage({type:t,err:i})});break}case"release":kr(r),postMessage({type:t});break;case"run":{let{sessionId:n,inputIndices:o,inputs:i,outputIndices:a,options:l}=r;Er(n,o,i,a,new Array(a.length).fill(null),l).then(d=>{d.some(p=>p[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:d},Or([...i,...d]))},d=>{postMessage({type:t,err:d})});break}case"end-profiling":Pr(r),postMessage({type:t});break;default:}}catch(n){postMessage({type:t,err:n})}});sp=ba?null:e=>new Worker(e??Dt,{type:"module",name:ya})});var _a={};Lt(_a,{default:()=>up});var Vn,$a,up,xa=M(()=>{"use strict";$a=(Vn=import.meta.url,async function(e={}){function t(){return ie.buffer!=ne.buffer&&Se(),ne}function r(){return ie.buffer!=ne.buffer&&Se(),Z}function n(){return ie.buffer!=ne.buffer&&Se(),ee}function o(){return ie.buffer!=ne.buffer&&Se(),K}function i(){return ie.buffer!=ne.buffer&&Se(),me}function a(){return ie.buffer!=ne.buffer&&Se(),$e}function l(){return ie.buffer!=ne.buffer&&Se(),R}function d(){return ie.buffer!=ne.buffer&&Se(),Be}var p,m,u=e,h=new Promise((s,c)=>{p=s,m=c}),w=typeof window=="object",g=typeof importScripts=="function",y=g&&self.name=="em-pthread";u.mountExternalData=(s,c)=>{s.startsWith("./")&&(s=s.substring(2)),(u.Ab||(u.Ab=new Map)).set(s,c)},u.unmountExternalData=()=>{delete u.Ab};var x=globalThis.SharedArrayBuffer??new WebAssembly.Memory({initial:0,maximum:0,shared:!0}).buffer.constructor;let _=()=>{let s=(f,b,v)=>(...C)=>{let P=Ze,B=b?.();C=f(...C);let N=b?.();return B!==N&&(f=N,v(B),b=v=null),Ze!=P?new Promise((H,J)=>{Sn={resolve:H,reject:J}}):C},c=f=>async(...b)=>{try{if(u.zb)throw Error("Session already started");let v=u.zb={Ub:b[0],errors:[]},C=await f(...b);if(u.zb!==v)throw Error("Session mismatch");u.Eb?.flush();let P=v.errors;if(0<P.length){let B=await Promise.all(P);if(B=B.filter(N=>N),0<B.length)throw Error(B.join(`
+`))}return C}finally{u.zb=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.Eb,u.Lb,u.Pb,u.Fb,u.Ob,u.hb,u.Qb,u.Sb,u.Mb,u.Nb,u.Rb]=c;let f=u.Eb;u.jsepRegisterBuffer=(b,v,C,P)=>f.registerBuffer(b,v,C,P),u.jsepGetBuffer=b=>f.getBuffer(b),u.jsepCreateDownloader=(b,v,C)=>f.createDownloader(b,v,C),u.jsepOnReleaseSession=b=>{f.onReleaseSession(b)},u.jsepOnRunStart=b=>f.onRunStart(b)}};var $,S,I=Object.assign({},u),T="./this.program",A=(s,c)=>{throw c},O="";(w||g)&&(g?O=self.location.href:typeof document<"u"&&document.currentScript&&(O=document.currentScript.src),Vn&&(O=Vn),O=O.startsWith("blob:")?"":O.substr(0,O.replace(/[?#].*/,"").lastIndexOf("/")+1),g&&(S=s=>{var c=new XMLHttpRequest;return c.open("GET",s,!1),c.responseType="arraybuffer",c.send(null),new Uint8Array(c.response)}),$=s=>Ao(s)?new Promise((c,f)=>{var b=new XMLHttpRequest;b.open("GET",s,!0),b.responseType="arraybuffer",b.onload=()=>{(b.status==200||b.status==0&&b.response)&&f(b.response),c(b.status)},b.onerror=c,b.send(null)}):fetch(s,{credentials:"same-origin"}).then(c=>c.ok?c.arrayBuffer():Promise.reject(Error(c.status+" : "+c.url))));var z,W=console.log.bind(console),V=console.error.bind(console),j=W,Y=V;if(Object.assign(u,I),I=null,y){let s=function(c){try{var f=c.data,b=f.cmd;if(b==="load"){let v=[];self.onmessage=C=>v.push(C),self.startWorker=()=>{postMessage({cmd:"loaded"});for(let C of v)s(C);self.onmessage=s};for(let C of f.handlers)u[C]&&!u[C].proxy||(u[C]=(...P)=>{postMessage({Db:"callHandler",Yb:C,args:P})},C=="print"&&(j=u[C]),C=="printErr"&&(Y=u[C]));ie=f.wasmMemory,Se(),ae(f.wasmModule)}else if(b==="run"){An(f.pthread_ptr,0,0,1,0,0),$n(f.pthread_ptr),Ul(),Ro(),se||(Pi(),se=!0);try{Vl(f.start_routine,f.arg)}catch(v){if(v!="unwind")throw v}}else b==="cancel"?zt()&&mr(-1):f.target!=="setimmediate"&&(b==="checkMailbox"?se&&or():b&&(Y(`worker: received unknown command ${b}`),Y(f)))}catch(v){throw Oi(),v}};var _h=s,ae,se=!1;Y=function(...c){c=c.join(" "),console.error(c)},self.alert=function(...c){postMessage({Db:"alert",text:c.join(" "),$b:zt()})},u.instantiateWasm=(c,f)=>new Promise(b=>{ae=v=>{v=new WebAssembly.Instance(v,Po()),f(v),b()}}),self.onunhandledrejection=c=>{throw c.reason||c},self.onmessage=s}u.wasmBinary&&(z=u.wasmBinary);var ie,q,ce,ne,Z,ee,K,me,$e,R,G,ue,Be,ge=!1;function Se(){var s=ie.buffer;u.HEAP8=ne=new Int8Array(s),u.HEAP16=ee=new Int16Array(s),u.HEAPU8=Z=new Uint8Array(s),u.HEAPU16=K=new Uint16Array(s),u.HEAP32=me=new Int32Array(s),u.HEAPU32=$e=new Uint32Array(s),u.HEAPF32=R=new Float32Array(s),u.HEAPF64=Be=new Float64Array(s),u.HEAP64=G=new BigInt64Array(s),u.HEAPU64=ue=new BigUint64Array(s)}if(!y){if(!((ie=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0})).buffer instanceof x))throw Y("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");Se()}var Wt=[],Te=[],Oe=[],He=0,kt=null,ht=null;function Co(){if(--He==0&&(kt!==null&&(clearInterval(kt),kt=null),ht)){var s=ht;ht=null,s()}}function Et(s){throw Y(s="Aborted("+s+")"),ge=!0,ce=1,s=new WebAssembly.RuntimeError(s+". Build with -sASSERTIONS for more info."),m(s),s}var ln,To=s=>s.startsWith("data:application/octet-stream;base64,"),Ao=s=>s.startsWith("file://");function ko(s){if(s==ln&&z)return new Uint8Array(z);if(S)return S(s);throw"both async and sync fetching of the wasm failed"}function Eo(s,c,f){return function(b){return z?Promise.resolve().then(()=>ko(b)):$(b).then(v=>new Uint8Array(v),()=>ko(b))}(s).then(b=>WebAssembly.instantiate(b,c)).then(f,b=>{Y(`failed to asynchronously prepare wasm: ${b}`),Et(b)})}function Po(){return{a:{L:Ml,xa:Rl,b:Wl,Z:No,y:Ho,na:Go,V:qo,X:Ko,oa:jo,la:Yo,ea:Xo,ka:Zo,I:Qo,W:Jo,T:ei,ma:ti,U:ri,sa:Ll,B:Gl,O:Fl,N:Kl,A:Yl,r:Xl,p:Zl,C:Ql,x:ic,P:ac,ra:sc,ha:uc,R:dc,_:lc,E:cc,ga:$n,qa:pc,t:mc,D:gc,o:yc,m:wc,c:wn,n:vc,k:xc,ya:Sc,q:Ic,g:Cc,u:Tc,l:Ac,f:kc,i:Ec,j:Pc,h:Oc,e:zc,ba:Bc,ca:Dc,da:Rc,$:yi,aa:bi,ua:Mc,d:Uc,M:Vc,F:Nc,J:Wc,v:Lc,pa:Hc,S:Gc,s:vi,w:Fc,K:qc,Q:Kc,wa:jc,va:Yc,ia:xi,ja:Si,Y:hn,z:Ii,H:Ci,fa:Ti,G:Ai,a:ie,ta:fn}}}var cn={858772:(s,c,f,b)=>{if(u===void 0||!u.Ab)return 1;if((s=Ee(s>>>0)).startsWith("./")&&(s=s.substring(2)),!(s=u.Ab.get(s)))return 2;if(b>>>=0,(c>>>=0)+(f>>>=0)>s.byteLength)return 3;try{return r().set(s.subarray(c,c+f),b>>>0),0}catch{return 4}},859273:()=>{u.Mb()},859304:()=>{u.Nb()},859333:()=>{u.Rb()},859358:s=>u.Lb(s),859391:s=>u.Pb(s),859423:(s,c,f)=>{u.Fb(s,c,f,!0)},859462:(s,c,f)=>{u.Fb(s,c,f)},859495:()=>typeof wasmOffsetConverter<"u",859552:s=>{u.hb("Abs",s,void 0)},859603:s=>{u.hb("Neg",s,void 0)},859654:s=>{u.hb("Floor",s,void 0)},859707:s=>{u.hb("Ceil",s,void 0)},859759:s=>{u.hb("Reciprocal",s,void 0)},859817:s=>{u.hb("Sqrt",s,void 0)},859869:s=>{u.hb("Exp",s,void 0)},859920:s=>{u.hb("Erf",s,void 0)},859971:s=>{u.hb("Sigmoid",s,void 0)},860026:(s,c,f)=>{u.hb("HardSigmoid",s,{alpha:c,beta:f})},860105:s=>{u.hb("Log",s,void 0)},860156:s=>{u.hb("Sin",s,void 0)},860207:s=>{u.hb("Cos",s,void 0)},860258:s=>{u.hb("Tan",s,void 0)},860309:s=>{u.hb("Asin",s,void 0)},860361:s=>{u.hb("Acos",s,void 0)},860413:s=>{u.hb("Atan",s,void 0)},860465:s=>{u.hb("Sinh",s,void 0)},860517:s=>{u.hb("Cosh",s,void 0)},860569:s=>{u.hb("Asinh",s,void 0)},860622:s=>{u.hb("Acosh",s,void 0)},860675:s=>{u.hb("Atanh",s,void 0)},860728:s=>{u.hb("Tanh",s,void 0)},860780:s=>{u.hb("Not",s,void 0)},860831:(s,c,f)=>{u.hb("Clip",s,{min:c,max:f})},860900:s=>{u.hb("Clip",s,void 0)},860952:(s,c)=>{u.hb("Elu",s,{alpha:c})},861010:s=>{u.hb("Gelu",s,void 0)},861062:s=>{u.hb("Relu",s,void 0)},861114:(s,c)=>{u.hb("LeakyRelu",s,{alpha:c})},861178:(s,c)=>{u.hb("ThresholdedRelu",s,{alpha:c})},861248:(s,c)=>{u.hb("Cast",s,{to:c})},861306:s=>{u.hb("Add",s,void 0)},861357:s=>{u.hb("Sub",s,void 0)},861408:s=>{u.hb("Mul",s,void 0)},861459:s=>{u.hb("Div",s,void 0)},861510:s=>{u.hb("Pow",s,void 0)},861561:s=>{u.hb("Equal",s,void 0)},861614:s=>{u.hb("Greater",s,void 0)},861669:s=>{u.hb("GreaterOrEqual",s,void 0)},861731:s=>{u.hb("Less",s,void 0)},861783:s=>{u.hb("LessOrEqual",s,void 0)},861842:(s,c,f,b,v)=>{u.hb("ReduceMean",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},862001:(s,c,f,b,v)=>{u.hb("ReduceMax",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},862159:(s,c,f,b,v)=>{u.hb("ReduceMin",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},862317:(s,c,f,b,v)=>{u.hb("ReduceProd",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},862476:(s,c,f,b,v)=>{u.hb("ReduceSum",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},862634:(s,c,f,b,v)=>{u.hb("ReduceL1",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},862791:(s,c,f,b,v)=>{u.hb("ReduceL2",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},862948:(s,c,f,b,v)=>{u.hb("ReduceLogSum",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},863109:(s,c,f,b,v)=>{u.hb("ReduceSumSquare",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},863273:(s,c,f,b,v)=>{u.hb("ReduceLogSumExp",s,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},863437:s=>{u.hb("Where",s,void 0)},863490:(s,c,f)=>{u.hb("Transpose",s,{perm:c?Array.from(i().subarray(c>>>0,f>>>0)):[]})},863598:(s,c,f,b)=>{u.hb("DepthToSpace",s,{blocksize:c,mode:Ee(f),format:b?"NHWC":"NCHW"})},863731:(s,c,f,b)=>{u.hb("DepthToSpace",s,{blocksize:c,mode:Ee(f),format:b?"NHWC":"NCHW"})},863864:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae,Bt)=>{u.hb("ConvTranspose",s,{format:N?"NHWC":"NCHW",autoPad:c,dilations:[f],group:b,kernelShape:[v],pads:[C,P],strides:[B],wIsConst:()=>!!t()[H>>>0],outputPadding:J?Array.from(i().subarray(J>>>0,de>>>0)):[],outputShape:ye?Array.from(i().subarray(ye>>>0,Ae>>>0)):[],activation:Ee(Bt)})},864265:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae)=>{u.hb("ConvTranspose",s,{format:B?"NHWC":"NCHW",autoPad:c,dilations:Array.from(i().subarray(f>>>0,2+(f>>>0)>>>0)),group:b,kernelShape:Array.from(i().subarray(v>>>0,2+(v>>>0)>>>0)),pads:Array.from(i().subarray(C>>>0,4+(C>>>0)>>>0)),strides:Array.from(i().subarray(P>>>0,2+(P>>>0)>>>0)),wIsConst:()=>!!t()[N>>>0],outputPadding:H?Array.from(i().subarray(H>>>0,J>>>0)):[],outputShape:de?Array.from(i().subarray(de>>>0,ye>>>0)):[],activation:Ee(Ae)})},864830:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae,Bt)=>{u.hb("ConvTranspose",s,{format:N?"NHWC":"NCHW",autoPad:c,dilations:[f],group:b,kernelShape:[v],pads:[C,P],strides:[B],wIsConst:()=>!!t()[H>>>0],outputPadding:J?Array.from(i().subarray(J>>>0,de>>>0)):[],outputShape:ye?Array.from(i().subarray(ye>>>0,Ae>>>0)):[],activation:Ee(Bt)})},865231:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae)=>{u.hb("ConvTranspose",s,{format:B?"NHWC":"NCHW",autoPad:c,dilations:Array.from(i().subarray(f>>>0,2+(f>>>0)>>>0)),group:b,kernelShape:Array.from(i().subarray(v>>>0,2+(v>>>0)>>>0)),pads:Array.from(i().subarray(C>>>0,4+(C>>>0)>>>0)),strides:Array.from(i().subarray(P>>>0,2+(P>>>0)>>>0)),wIsConst:()=>!!t()[N>>>0],outputPadding:H?Array.from(i().subarray(H>>>0,J>>>0)):[],outputShape:de?Array.from(i().subarray(de>>>0,ye>>>0)):[],activation:Ee(Ae)})},865796:(s,c)=>{u.hb("GlobalAveragePool",s,{format:c?"NHWC":"NCHW"})},865887:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae)=>{u.hb("AveragePool",s,{format:Ae?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:v,dilations:C?Array.from(i().subarray(C>>>0,P>>>0)):[],kernel_shape:B?Array.from(i().subarray(B>>>0,N>>>0)):[],pads:H?Array.from(i().subarray(H>>>0,J>>>0)):[],strides:de?Array.from(i().subarray(de>>>0,ye>>>0)):[]})},866302:(s,c)=>{u.hb("GlobalAveragePool",s,{format:c?"NHWC":"NCHW"})},866393:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae)=>{u.hb("AveragePool",s,{format:Ae?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:v,dilations:C?Array.from(i().subarray(C>>>0,P>>>0)):[],kernel_shape:B?Array.from(i().subarray(B>>>0,N>>>0)):[],pads:H?Array.from(i().subarray(H>>>0,J>>>0)):[],strides:de?Array.from(i().subarray(de>>>0,ye>>>0)):[]})},866808:(s,c)=>{u.hb("GlobalMaxPool",s,{format:c?"NHWC":"NCHW"})},866895:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae)=>{u.hb("MaxPool",s,{format:Ae?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:v,dilations:C?Array.from(i().subarray(C>>>0,P>>>0)):[],kernel_shape:B?Array.from(i().subarray(B>>>0,N>>>0)):[],pads:H?Array.from(i().subarray(H>>>0,J>>>0)):[],strides:de?Array.from(i().subarray(de>>>0,ye>>>0)):[]})},867306:(s,c)=>{u.hb("GlobalMaxPool",s,{format:c?"NHWC":"NCHW"})},867393:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae)=>{u.hb("MaxPool",s,{format:Ae?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:b,storage_order:v,dilations:C?Array.from(i().subarray(C>>>0,P>>>0)):[],kernel_shape:B?Array.from(i().subarray(B>>>0,N>>>0)):[],pads:H?Array.from(i().subarray(H>>>0,J>>>0)):[],strides:de?Array.from(i().subarray(de>>>0,ye>>>0)):[]})},867804:(s,c,f,b,v)=>{u.hb("Gemm",s,{alpha:c,beta:f,transA:b,transB:v})},867908:s=>{u.hb("MatMul",s,void 0)},867962:(s,c,f,b)=>{u.hb("ArgMax",s,{keepDims:!!c,selectLastIndex:!!f,axis:b})},868070:(s,c,f,b)=>{u.hb("ArgMin",s,{keepDims:!!c,selectLastIndex:!!f,axis:b})},868178:(s,c)=>{u.hb("Softmax",s,{axis:c})},868241:(s,c)=>{u.hb("Concat",s,{axis:c})},868301:(s,c,f,b,v)=>{u.hb("Split",s,{axis:c,numOutputs:f,splitSizes:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},868441:s=>{u.hb("Expand",s,void 0)},868495:(s,c)=>{u.hb("Gather",s,{axis:Number(c)})},868566:(s,c)=>{u.hb("GatherElements",s,{axis:Number(c)})},868645:(s,c,f,b,v,C,P,B,N,H,J)=>{u.hb("Resize",s,{antialias:c,axes:f?Array.from(i().subarray(f>>>0,b>>>0)):[],coordinateTransformMode:Ee(v),cubicCoeffA:C,excludeOutside:P,extrapolationValue:B,keepAspectRatioPolicy:Ee(N),mode:Ee(H),nearestMode:Ee(J)})},868991:(s,c,f,b,v,C,P)=>{u.hb("Slice",s,{starts:c?Array.from(i().subarray(c>>>0,f>>>0)):[],ends:b?Array.from(i().subarray(b>>>0,v>>>0)):[],axes:C?Array.from(i().subarray(C>>>0,P>>>0)):[]})},869207:s=>{u.hb("Tile",s,void 0)},869259:(s,c,f)=>{u.hb("InstanceNormalization",s,{epsilon:c,format:f?"NHWC":"NCHW"})},869373:(s,c,f)=>{u.hb("InstanceNormalization",s,{epsilon:c,format:f?"NHWC":"NCHW"})},869487:s=>{u.hb("Range",s,void 0)},869540:(s,c)=>{u.hb("Einsum",s,{equation:Ee(c)})},869621:(s,c,f,b,v)=>{u.hb("Pad",s,{mode:c,value:f,pads:b?Array.from(i().subarray(b>>>0,v>>>0)):[]})},869748:(s,c,f,b,v,C)=>{u.hb("BatchNormalization",s,{epsilon:c,momentum:f,spatial:!!v,trainingMode:!!b,format:C?"NHWC":"NCHW"})},869917:(s,c,f,b,v,C)=>{u.hb("BatchNormalization",s,{epsilon:c,momentum:f,spatial:!!v,trainingMode:!!b,format:C?"NHWC":"NCHW"})},870086:(s,c,f)=>{u.hb("CumSum",s,{exclusive:Number(c),reverse:Number(f)})},870183:(s,c,f)=>{u.hb("DequantizeLinear",s,{axis:c,blockSize:f})},870273:(s,c,f,b,v,C,P,B,N)=>{u.hb("Attention",s,{numHeads:c,isUnidirectional:f,maskFilterValue:b,scale:v,doRotary:C,qkvHiddenSizes:P?Array.from(i().subarray(Number(B)>>>0,Number(B)+P>>>0)):[],pastPresentShareBuffer:!!N})},870545:s=>{u.hb("BiasAdd",s,void 0)},870600:s=>{u.hb("BiasSplitGelu",s,void 0)},870661:s=>{u.hb("FastGelu",s,void 0)},870717:(s,c,f,b,v,C,P,B,N,H,J,de,ye,Ae,Bt,Qc)=>{u.hb("Conv",s,{format:de?"NHWC":"NCHW",auto_pad:c,dilations:f?Array.from(i().subarray(f>>>0,b>>>0)):[],group:v,kernel_shape:C?Array.from(i().subarray(C>>>0,P>>>0)):[],pads:B?Array.from(i().subarray(B>>>0,N>>>0)):[],strides:H?Array.from(i().subarray(H>>>0,J>>>0)):[],w_is_const:()=>!!t()[ye>>>0],activation:Ee(Ae),activation_params:Bt?Array.from(l().subarray(Bt>>>0,Qc>>>0)):[]})},871213:s=>{u.hb("Gelu",s,void 0)},871265:(s,c,f,b)=>{u.hb("GroupQueryAttention",s,{numHeads:c,kvNumHeads:f,scale:b})},871378:(s,c,f,b)=>{u.hb("LayerNormalization",s,{axis:c,epsilon:f,simplified:!!b})},871489:(s,c,f,b)=>{u.hb("LayerNormalization",s,{axis:c,epsilon:f,simplified:!!b})},871600:(s,c,f,b,v,C)=>{u.hb("MatMulNBits",s,{k:c,n:f,accuracyLevel:b,bits:v,blockSize:C})},871727:(s,c,f,b,v,C)=>{u.hb("MultiHeadAttention",s,{numHeads:c,isUnidirectional:f,maskFilterValue:b,scale:v,doRotary:C})},871886:(s,c)=>{u.hb("QuickGelu",s,{alpha:c})},871950:(s,c,f,b,v)=>{u.hb("RotaryEmbedding",s,{interleaved:!!c,numHeads:f,rotaryEmbeddingDim:b,scale:v})},872089:(s,c,f)=>{u.hb("SkipLayerNormalization",s,{epsilon:c,simplified:!!f})},872191:s=>{u.Qb(s)},872225:(s,c)=>u.Sb(s,c,u.zb.Ub,u.zb.errors),872337:(s,c,f)=>{u.hb("SkipLayerNormalization",s,{epsilon:c,simplified:!!f})}};function Rl(s,c,f){return ci(async()=>{await u.Ob(s,c,f)})}function Ml(){return typeof wasmOffsetConverter<"u"}function pn(s){this.name="ExitStatus",this.message=`Program terminated with exit(${s})`,this.status=s}var mn=s=>{s.terminate(),s.onmessage=()=>{}},Oo=s=>{dt.length==0&&(Uo(),Mo(dt[0]));var c=dt.pop();if(!c)return 6;yt.push(c),Ye[s.sb]=c,c.sb=s.sb;var f={cmd:"run",start_routine:s.Vb,arg:s.Hb,pthread_ptr:s.sb};return c.postMessage(f,s.Wb),0},gt=0,ve=(s,c,...f)=>{for(var b=2*f.length,v=Pn(),C=En(8*b),P=C>>>3,B=0;B<f.length;B++){var N=f[B];typeof N=="bigint"?(G[P+2*B]=1n,G[P+2*B+1]=N):(G[P+2*B]=0n,d()[P+2*B+1>>>0]=N)}return s=zi(s,0,b,C,c),fr(v),s};function fn(s){if(y)return ve(0,1,s);if(ce=s,!(0<gt)){for(var c of yt)mn(c);for(c of dt)mn(c);dt=[],yt=[],Ye=[],ge=!0}A(s,new pn(s))}function zo(s){if(y)return ve(1,0,s);hn(s)}var hn=s=>{if(ce=s,y)throw zo(s),"unwind";fn(s)},dt=[],yt=[],Bo=[],Ye={},Do=s=>{var c=s.sb;delete Ye[c],dt.push(s),yt.splice(yt.indexOf(s),1),s.sb=0,kn(c)};function Ro(){Bo.forEach(s=>s())}var Mo=s=>new Promise(c=>{s.onmessage=v=>{var C=(v=v.data).cmd;if(v.targetThread&&v.targetThread!=zt()){var P=Ye[v.targetThread];P?P.postMessage(v,v.transferList):Y(`Internal error! Worker sent a message "${C}" to target pthread ${v.targetThread}, but that thread no longer exists!`)}else C==="checkMailbox"?or():C==="spawnThread"?Oo(v):C==="cleanupThread"?Do(Ye[v.thread]):C==="killThread"?(v=v.thread,C=Ye[v],delete Ye[v],mn(C),kn(v),yt.splice(yt.indexOf(C),1),C.sb=0):C==="cancelThread"?Ye[v.thread].postMessage({cmd:"cancel"}):C==="loaded"?(s.loaded=!0,c(s)):C==="alert"?alert(`Thread ${v.threadId}: ${v.text}`):v.target==="setimmediate"?s.postMessage(v):C==="callHandler"?u[v.handler](...v.args):C&&Y(`worker sent an unknown command ${C}`)},s.onerror=v=>{throw Y(`worker sent an error! ${v.filename}:${v.lineno}: ${v.message}`),v};var f,b=[];for(f of[])u.propertyIsEnumerable(f)&&b.push(f);s.postMessage({cmd:"load",handlers:b,wasmMemory:ie,wasmModule:q})});function Uo(){var s=new Worker(new URL(import.meta.url),{type:"module",workerData:"em-pthread",name:"em-pthread"});dt.push(s)}var nr=s=>{for(;0<s.length;)s.shift()(u)},Ul=()=>{var s=zt(),c=a()[s+52>>>2>>>0];s=a()[s+56>>>2>>>0],Di(c,c-s),fr(c)},Vl=(s,c)=>{gt=0,s=Ri(s,c),0<gt?ce=s:mr(s)};class Nl{constructor(c){this.Bb=c-24}}function Wl(s,c,f){var b=new Nl(s>>>=0);throw c>>>=0,f>>>=0,a()[b.Bb+16>>>2>>>0]=0,a()[b.Bb+4>>>2>>>0]=c,a()[b.Bb+8>>>2>>>0]=f,s}function Vo(s,c,f,b){return y?ve(2,1,s,c,f,b):No(s,c,f,b)}function No(s,c,f,b){if(s>>>=0,c>>>=0,f>>>=0,b>>>=0,x===void 0)return Y("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var v=[];return y&&v.length===0?Vo(s,c,f,b):(s={Vb:f,sb:s,Hb:b,Wb:v},y?(s.Db="spawnThread",postMessage(s,v),0):Oo(s))}var Wo=typeof TextDecoder<"u"?new TextDecoder:void 0,Lo=(s,c,f)=>{var b=(c>>>=0)+f;for(f=c;s[f]&&!(f>=b);)++f;if(16<f-c&&s.buffer&&Wo)return Wo.decode(s.buffer instanceof x?s.slice(c,f):s.subarray(c,f));for(b="";c<f;){var v=s[c++];if(128&v){var C=63&s[c++];if((224&v)==192)b+=String.fromCharCode((31&v)<<6|C);else{var P=63&s[c++];65536>(v=(240&v)==224?(15&v)<<12|C<<6|P:(7&v)<<18|C<<12|P<<6|63&s[c++])?b+=String.fromCharCode(v):(v-=65536,b+=String.fromCharCode(55296|v>>10,56320|1023&v))}}else b+=String.fromCharCode(v)}return b},Ee=(s,c)=>(s>>>=0)?Lo(r(),s,c):"";function Ho(s,c,f){return y?ve(3,1,s,c,f):0}function Go(s,c){if(y)return ve(4,1,s,c)}var Fo=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},Pt=(s,c,f)=>{var b=r();if(c>>>=0,0<f){var v=c;f=c+f-1;for(var C=0;C<s.length;++C){var P=s.charCodeAt(C);if(55296<=P&&57343>=P&&(P=65536+((1023&P)<<10)|1023&s.charCodeAt(++C)),127>=P){if(c>=f)break;b[c++>>>0]=P}else{if(2047>=P){if(c+1>=f)break;b[c++>>>0]=192|P>>6}else{if(65535>=P){if(c+2>=f)break;b[c++>>>0]=224|P>>12}else{if(c+3>=f)break;b[c++>>>0]=240|P>>18,b[c++>>>0]=128|P>>12&63}b[c++>>>0]=128|P>>6&63}b[c++>>>0]=128|63&P}}b[c>>>0]=0,s=c-v}else s=0;return s};function qo(s,c){if(y)return ve(5,1,s,c)}function Ko(s,c,f){if(y)return ve(6,1,s,c,f)}function jo(s,c,f){return y?ve(7,1,s,c,f):0}function Yo(s,c){if(y)return ve(8,1,s,c)}function Xo(s,c,f){if(y)return ve(9,1,s,c,f)}function Zo(s,c,f,b){if(y)return ve(10,1,s,c,f,b)}function Qo(s,c,f,b){if(y)return ve(11,1,s,c,f,b)}function Jo(s,c,f,b){if(y)return ve(12,1,s,c,f,b)}function ei(s){if(y)return ve(13,1,s)}function ti(s,c){if(y)return ve(14,1,s,c)}function ri(s,c,f){if(y)return ve(15,1,s,c,f)}var ni,lt,Ll=()=>{Et("")},Xe=s=>{for(var c="";r()[s>>>0];)c+=ni[r()[s++>>>0]];return c},gn={},yn={},Hl={};function it(s,c,f={}){if(!("argPackAdvance"in c))throw new TypeError("registerType registeredInstance requires argPackAdvance");return function(b,v,C={}){var P=v.name;if(!b)throw new lt(`type "${P}" must have a positive integer typeid pointer`);if(yn.hasOwnProperty(b)){if(C.Jb)return;throw new lt(`Cannot register type '${P}' twice`)}yn[b]=v,delete Hl[b],gn.hasOwnProperty(b)&&(v=gn[b],delete gn[b],v.forEach(B=>B()))}(s,c,f)}var oi=(s,c,f)=>{switch(c){case 1:return f?b=>t()[b>>>0]:b=>r()[b>>>0];case 2:return f?b=>n()[b>>>1>>>0]:b=>o()[b>>>1>>>0];case 4:return f?b=>i()[b>>>2>>>0]:b=>a()[b>>>2>>>0];case 8:return f?b=>G[b>>>3]:b=>ue[b>>>3];default:throw new TypeError(`invalid integer width (${c}): ${s}`)}};function Gl(s,c,f){f>>>=0,it(s>>>=0,{name:c=Xe(c>>>0),fromWireType:b=>b,toWireType:function(b,v){if(typeof v!="bigint"&&typeof v!="number")throw v=v===null?"null":(b=typeof v)=="object"||b==="array"||b==="function"?v.toString():""+v,new TypeError(`Cannot convert "${v}" to ${this.name}`);return typeof v=="number"&&(v=BigInt(v)),v},argPackAdvance:ct,readValueFromPointer:oi(c,f,c.indexOf("u")==-1),yb:null})}var ct=8;function Fl(s,c,f,b){it(s>>>=0,{name:c=Xe(c>>>0),fromWireType:function(v){return!!v},toWireType:function(v,C){return C?f:b},argPackAdvance:ct,readValueFromPointer:function(v){return this.fromWireType(r()[v>>>0])},yb:null})}var bn=[],at=[];function wn(s){9<(s>>>=0)&&--at[s+1]==0&&(at[s]=void 0,bn.push(s))}var Ge=s=>{if(!s)throw new lt("Cannot use deleted val. handle = "+s);return at[s]},Fe=s=>{switch(s){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let c=bn.pop()||at.length;return at[c]=s,at[c+1]=1,c}};function vn(s){return this.fromWireType(a()[s>>>2>>>0])}var ql={name:"emscripten::val",fromWireType:s=>{var c=Ge(s);return wn(s),c},toWireType:(s,c)=>Fe(c),argPackAdvance:ct,readValueFromPointer:vn,yb:null};function Kl(s){return it(s>>>0,ql)}var jl=(s,c)=>{switch(c){case 4:return function(f){return this.fromWireType(l()[f>>>2>>>0])};case 8:return function(f){return this.fromWireType(d()[f>>>3>>>0])};default:throw new TypeError(`invalid float width (${c}): ${s}`)}};function Yl(s,c,f){f>>>=0,it(s>>>=0,{name:c=Xe(c>>>0),fromWireType:b=>b,toWireType:(b,v)=>v,argPackAdvance:ct,readValueFromPointer:jl(c,f),yb:null})}function Xl(s,c,f,b,v){if(s>>>=0,f>>>=0,c=Xe(c>>>0),v===-1&&(v=4294967295),v=B=>B,b===0){var C=32-8*f;v=B=>B<<C>>>C}var P=c.includes("unsigned")?function(B,N){return N>>>0}:function(B,N){return N};it(s,{name:c,fromWireType:v,toWireType:P,argPackAdvance:ct,readValueFromPointer:oi(c,f,b!==0),yb:null})}function Zl(s,c,f){function b(C){var P=a()[C>>>2>>>0];return C=a()[C+4>>>2>>>0],new v(t().buffer,C,P)}var v=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][c];it(s>>>=0,{name:f=Xe(f>>>0),fromWireType:b,argPackAdvance:ct,readValueFromPointer:b},{Jb:!0})}function Ql(s,c){s>>>=0;var f=(c=Xe(c>>>0))==="std::string";it(s,{name:c,fromWireType:function(b){var v=a()[b>>>2>>>0],C=b+4;if(f)for(var P=C,B=0;B<=v;++B){var N=C+B;if(B==v||r()[N>>>0]==0){if(P=Ee(P,N-P),H===void 0)var H=P;else H+=String.fromCharCode(0),H+=P;P=N+1}}else{for(H=Array(v),B=0;B<v;++B)H[B]=String.fromCharCode(r()[C+B>>>0]);H=H.join("")}return Qe(b),H},toWireType:function(b,v){v instanceof ArrayBuffer&&(v=new Uint8Array(v));var C=typeof v=="string";if(!(C||v instanceof Uint8Array||v instanceof Uint8ClampedArray||v instanceof Int8Array))throw new lt("Cannot pass non-string to std::string");var P=f&&C?Fo(v):v.length,B=pr(4+P+1),N=B+4;if(a()[B>>>2>>>0]=P,f&&C)Pt(v,N,P+1);else if(C)for(C=0;C<P;++C){var H=v.charCodeAt(C);if(255<H)throw Qe(N),new lt("String has UTF-16 code units that do not fit in 8 bits");r()[N+C>>>0]=H}else for(C=0;C<P;++C)r()[N+C>>>0]=v[C];return b!==null&&b.push(Qe,B),B},argPackAdvance:ct,readValueFromPointer:vn,yb(b){Qe(b)}})}var ii=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,Jl=(s,c)=>{for(var f=s>>1,b=f+c/2;!(f>=b)&&o()[f>>>0];)++f;if(32<(f<<=1)-s&&ii)return ii.decode(r().slice(s,f));for(f="",b=0;!(b>=c/2);++b){var v=n()[s+2*b>>>1>>>0];if(v==0)break;f+=String.fromCharCode(v)}return f},ec=(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 v=0;v<f;++v){var C=s.charCodeAt(v);n()[c>>>1>>>0]=C,c+=2}return n()[c>>>1>>>0]=0,c-b},tc=s=>2*s.length,rc=(s,c)=>{for(var f=0,b="";!(f>=c/4);){var v=i()[s+4*f>>>2>>>0];if(v==0)break;++f,65536<=v?(v-=65536,b+=String.fromCharCode(55296|v>>10,56320|1023&v)):b+=String.fromCharCode(v)}return b},nc=(s,c,f)=>{if(c>>>=0,f??=2147483647,4>f)return 0;var b=c;f=b+f-4;for(var v=0;v<s.length;++v){var C=s.charCodeAt(v);if(55296<=C&&57343>=C&&(C=65536+((1023&C)<<10)|1023&s.charCodeAt(++v)),i()[c>>>2>>>0]=C,(c+=4)+4>f)break}return i()[c>>>2>>>0]=0,c-b},oc=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 ic(s,c,f){if(s>>>=0,c>>>=0,f=Xe(f>>>=0),c===2)var b=Jl,v=ec,C=tc,P=B=>o()[B>>>1>>>0];else c===4&&(b=rc,v=nc,C=oc,P=B=>a()[B>>>2>>>0]);it(s,{name:f,fromWireType:B=>{for(var N,H=a()[B>>>2>>>0],J=B+4,de=0;de<=H;++de){var ye=B+4+de*c;de!=H&&P(ye)!=0||(J=b(J,ye-J),N===void 0?N=J:(N+=String.fromCharCode(0),N+=J),J=ye+c)}return Qe(B),N},toWireType:(B,N)=>{if(typeof N!="string")throw new lt(`Cannot pass non-string to C++ string type ${f}`);var H=C(N),J=pr(4+H+c);return a()[J>>>2>>>0]=H/c,v(N,J+4,H+c),B!==null&&B.push(Qe,J),J},argPackAdvance:ct,readValueFromPointer:vn,yb(B){Qe(B)}})}function ac(s,c){it(s>>>=0,{Kb:!0,name:c=Xe(c>>>0),argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})}var sc=()=>1;function uc(s){An(s>>>0,!g,1,!w,131072,!1),Ro()}var ai=s=>{if(!ge)try{if(s(),!(0<gt))try{y?mr(ce):hn(ce)}catch(c){c instanceof pn||c=="unwind"||A(1,c)}}catch(c){c instanceof pn||c=="unwind"||A(1,c)}};function $n(s){s>>>=0,typeof Atomics.Xb=="function"&&(Atomics.Xb(i(),s>>>2,s).value.then(or),s+=128,Atomics.store(i(),s>>>2,1))}var or=()=>{var s=zt();s&&($n(s),ai(Bi))};function dc(s,c){(s>>>=0)==c>>>0?setTimeout(or):y?postMessage({targetThread:s,cmd:"checkMailbox"}):(s=Ye[s])&&s.postMessage({cmd:"checkMailbox"})}var _n=[];function lc(s,c,f,b,v){for(c>>>=0,b/=2,_n.length=b,f=v>>>0>>>3,v=0;v<b;v++)_n[v]=G[f+2*v]?G[f+2*v+1]:d()[f+2*v+1>>>0];return(c?cn[c]:Zc[s])(..._n)}function cc(s){s>>>=0,y?postMessage({cmd:"cleanupThread",thread:s}):Do(Ye[s])}function pc(s){}var xn=(s,c)=>{var f=yn[s];if(f===void 0)throw s=Ei(s),f=Xe(s),Qe(s),new lt(`${c} has unknown type ${f}`);return f},si=(s,c,f)=>{var b=[];return s=s.toWireType(b,f),b.length&&(a()[c>>>2>>>0]=Fe(b)),s};function mc(s,c,f){return c>>>=0,f>>>=0,s=Ge(s>>>0),c=xn(c,"emval::as"),si(c,f,s)}var ir=s=>{try{s()}catch(c){Et(c)}},pt=0,Ze=null,ui=0,ar=[],di={},li={},fc=0,Sn=null,hc=[];function ci(s){return function(c){if(!ge){if(pt===0){var f=!1,b=!1;c((v=0)=>{if(!ge&&(ui=v,f=!0,b)){pt=2,ir(()=>Vi(Ze)),typeof Browser<"u"&&Browser.Cb.Ib&&Browser.Cb.resume(),v=!1;try{var C=function(){var N=i()[Ze+8>>>2>>>0];return N=F[li[N]],--gt,N()}()}catch(N){C=N,v=!0}var P=!1;if(!Ze){var B=Sn;B&&(Sn=null,(v?B.reject:B.resolve)(C),P=!0)}if(v&&!P)throw C}}),b=!0,f||(pt=1,Ze=function(){var v=pr(65548),C=v+12;a()[v>>>2>>>0]=C,a()[v+4>>>2>>>0]=C+65536,C=ar[0];var P=di[C];return P===void 0&&(P=fc++,di[C]=P,li[P]=C),C=P,i()[v+8>>>2>>>0]=C,v}(),typeof Browser<"u"&&Browser.Cb.Ib&&Browser.Cb.pause(),ir(()=>Mi(Ze)))}else pt===2?(pt=0,ir(Ni),Qe(Ze),Ze=null,hc.forEach(ai)):Et(`invalid state: ${pt}`);return ui}}(c=>{s().then(c)})}function gc(s){return s>>>=0,ci(()=>(s=Ge(s)).then(Fe))}var sr=[];function yc(s,c,f,b){return f>>>=0,b>>>=0,(s=sr[s>>>0])(null,c=Ge(c>>>0),f,b)}var bc={},ur=s=>{var c=bc[s];return c===void 0?Xe(s):c};function wc(s,c,f,b,v){return f>>>=0,b>>>=0,v>>>=0,(s=sr[s>>>0])(c=Ge(c>>>0),c[f=ur(f)],b,v)}var pi=()=>typeof globalThis=="object"?globalThis:Function("return this")();function vc(s){return(s>>>=0)==0?Fe(pi()):(s=ur(s),Fe(pi()[s]))}var $c=s=>{var c=sr.length;return sr.push(s),c},_c=(s,c)=>{for(var f=Array(s),b=0;b<s;++b)f[b]=xn(a()[c+4*b>>>2>>>0],"parameter "+b);return f},mi=(s,c)=>Object.defineProperty(c,"name",{value:s});function xc(s,c,f){var b=(c=_c(s,c>>>0)).shift();s--;var v=`return function (obj, func, destructorsRef, args) {
+`,C=0,P=[];f===0&&P.push("obj");for(var B=["retType"],N=[b],H=0;H<s;++H)P.push("arg"+H),B.push("argType"+H),N.push(c[H]),v+=`  var arg${H} = argType${H}.readValueFromPointer(args${C?"+"+C:""});
+`,C+=c[H].argPackAdvance;return v+=`  var rv = ${f===1?"new func":"func.call"}(${P.join(", ")});
+`,b.Kb||(B.push("emval_returnValue"),N.push(si),v+=`  return emval_returnValue(retType, destructorsRef, rv);
 `),B.push(v+`};
-`),s=function(j){var me=Function;if(!(me instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof me} which is not a function`);var he=yi(me.name||"unknownFunctionName",function(){});return he.prototype=me.prototype,he=new he,(j=me.apply(he,j))instanceof Object?j:he}(B)(...N),f=`methodCaller<(${c.map(j=>j.name).join(", ")}) => ${y.name}>`,Ic(yi(f,s))}function Ac(s){return s=cr(s>>>0),je(u[s])}function Ec(s,c){return c>>>=0,s=qe(s>>>0),c=qe(c),je(s[c])}function kc(s){9<(s>>>=0)&&(ut[s+1]+=1)}function Pc(){return je([])}function Oc(s){s=qe(s>>>0);for(var c=Array(s.length),f=0;f<s.length;f++)c[f]=s[f];return je(c)}function Dc(s){return je(cr(s>>>0))}function zc(){return je({})}function Bc(s){for(var c=qe(s>>>=0);c.length;){var f=c.pop();c.pop()(f)}xn(s)}function Rc(s,c,f){c>>>=0,f>>>=0,s=qe(s>>>0),c=qe(c),f=qe(f),s[c]=f}function Mc(s,c){return c>>>=0,s=(s=Tn(s>>>0,"_emval_take_value")).readValueFromPointer(c),je(s)}function Uc(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),bi=[0,31,60,91,121,152,182,213,244,274,305,335],wi=[0,31,59,90,120,151,181,212,243,273,304,334];function Vc(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())?bi:wi)[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 y=new Date(s.getFullYear(),0,1).getTimezoneOffset();s=0|(f!=y&&s.getTimezoneOffset()==Math.min(y,f)),i()[c+32>>>2>>>0]=s}function Nc(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],y=c.getTimezoneOffset(),v=new Date(c.getFullYear(),6,1).getTimezoneOffset(),C=new Date(c.getFullYear(),0,1).getTimezoneOffset(),O=Math.min(C,v);return 0>f?i()[s+32>>>2>>>0]=+(v!=C&&O==y):0<f!=(O==y)&&(v=Math.max(C,v),c.setTime(c.getTime()+6e4*((0<f?O:v)-y))),i()[s+24>>>2>>>0]=c.getDay(),f=(Bt(c.getFullYear())?bi:wi)[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 vi(s,c,f,y,v,C,O){return b?xe(16,1,s,c,f,y,v,C,O):-52}function $i(s,c,f,y,v,C){if(b)return xe(17,1,s,c,f,y,v,C)}function Wc(s,c,f,y){s>>>=0,c>>>=0,f>>>=0,y>>>=0;var v=new Date().getFullYear(),C=new Date(v,0,1),O=new Date(v,6,1);v=C.getTimezoneOffset();var B=O.getTimezoneOffset(),N=Math.max(v,B);a()[s>>>2>>>0]=60*N,i()[c>>>2>>>0]=+(v!=B),C=(s=L=>L.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1])(C),O=s(O),B<v?(zt(C,f,17),zt(O,y,17)):(zt(C,y,17),zt(O,f,17))}var En=[],_i=(s,c)=>{En.length=0;for(var f;f=r()[s++>>>0];){var y=f!=105;c+=(y&=f!=112)&&c%8?4:0,En.push(f==112?a()[c>>>2>>>0]:f==106?F[c>>>3]:f==105?i()[c>>>2>>>0]:d()[c>>>3>>>0]),c+=y?8:4}return En};function Hc(s,c,f){return s>>>=0,c=_i(c>>>0,f>>>0),fn[s](...c)}function Gc(s,c,f){return s>>>=0,c=_i(c>>>0,f>>>0),fn[s](...c)}var Lc=()=>{},Fc=()=>Date.now();function qc(s,c){return Y(Ae(s>>>0,c>>>0))}var xi,jc=()=>{throw bt+=1,"unwind"};function Kc(){return 4294901760}xi=()=>performance.timeOrigin+performance.now();var Yc=()=>navigator.hardwareConcurrency;function Xc(){return Dt("Cannot use emscripten_pc_get_function without -sUSE_OFFSET_CONVERTER"),0}function Zc(s){s>>>=0;var c=r().length;if(s<=c||4294901760<s)return!1;for(var f=1;4>=f;f*=2){var y=c*(1+.2/f);y=Math.min(y,s+100663296);var v=Math;y=Math.max(s,y);e:{v=(v.min.call(v,4294901760,y+(65536-y%65536)%65536)-le.buffer.byteLength+65535)/65536;try{le.grow(v),we();var C=1;break e}catch{}C=void 0}if(C)return!0}return!1}var pr=()=>(Dt("Cannot use convertFrameToPC (needed by __builtin_return_address) without -sUSE_OFFSET_CONVERTER"),0),Rt={},Si=s=>{s.forEach(c=>{var f=pr();f&&(Rt[f]=c)})};function Qc(){var s=Error().stack.toString().split(`
-`);return s[0]=="Error"&&s.shift(),Si(s),Rt.Ob=pr(),Rt.ac=s,Rt.Ob}function Jc(s,c,f){if(s>>>=0,c>>>=0,Rt.Ob==s)var y=Rt.ac;else(y=Error().stack.toString().split(`
-`))[0]=="Error"&&y.shift(),Si(y);for(var v=3;y[v]&&pr()!=s;)++v;for(s=0;s<f&&y[s+v];++s)i()[c+4*s>>>2>>>0]=pr();return s}var kn,Pn={},Ii=()=>{if(!kn){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",_:T||"./this.program"};for(s in Pn)Pn[s]===void 0?delete c[s]:c[s]=Pn[s];var f=[];for(s in c)f.push(`${s}=${c[s]}`);kn=f}return kn};function Ci(s,c){if(b)return xe(18,1,s,c);s>>>=0,c>>>=0;var f=0;return Ii().forEach((y,v)=>{var C=c+f;for(v=a()[s+4*v>>>2>>>0]=C,C=0;C<y.length;++C)t()[v++>>>0]=y.charCodeAt(C);t()[v>>>0]=0,f+=y.length+1}),0}function Ti(s,c){if(b)return xe(19,1,s,c);s>>>=0,c>>>=0;var f=Ii();a()[s>>>2>>>0]=f.length;var y=0;return f.forEach(v=>y+=v.length+1),a()[c>>>2>>>0]=y,0}function Ai(s){return b?xe(20,1,s):52}function Ei(s,c,f,y){return b?xe(21,1,s,c,f,y):52}function ki(s,c,f,y){return b?xe(22,1,s,c,f,y):70}var ep=[null,[],[]];function Pi(s,c,f,y){if(b)return xe(23,1,s,c,f,y);c>>>=0,f>>>=0,y>>>=0;for(var v=0,C=0;C<f;C++){var O=a()[c>>>2>>>0],B=a()[c+4>>>2>>>0];c+=8;for(var N=0;N<B;N++){var L=r()[O+N>>>0],j=ep[s];L===0||L===10?((s===1?X:Y)(qo(j,0)),j.length=0):j.push(L)}v+=B}return a()[y>>>2>>>0]=v,0}var Oi=[31,29,31,30,31,30,31,31,30,31,30,31],Di=[31,28,31,30,31,30,31,31,30,31,30,31],tp=(s,c)=>{t().set(s,c>>>0)};function zi(s,c,f,y){function v(P,ee,ye){for(P=typeof P=="number"?P.toString():P||"";P.length<ee;)P=ye[0]+P;return P}function C(P,ee){return v(P,ee,"0")}function O(P,ee){function ye(Ki){return 0>Ki?-1:0<Ki?1:0}var vt;return(vt=ye(P.getFullYear()-ee.getFullYear()))===0&&(vt=ye(P.getMonth()-ee.getMonth()))===0&&(vt=ye(P.getDate()-ee.getDate())),vt}function B(P){switch(P.getDay()){case 0:return new Date(P.getFullYear()-1,11,29);case 1:return P;case 2:return new Date(P.getFullYear(),0,3);case 3:return new Date(P.getFullYear(),0,2);case 4:return new Date(P.getFullYear(),0,1);case 5:return new Date(P.getFullYear()-1,11,31);case 6:return new Date(P.getFullYear()-1,11,30)}}function N(P){var ee=P.Bb;for(P=new Date(new Date(P.Cb+1900,0,1).getTime());0<ee;){var ye=P.getMonth(),vt=(Bt(P.getFullYear())?Oi:Di)[ye];if(!(ee>vt-P.getDate())){P.setDate(P.getDate()+ee);break}ee-=vt-P.getDate()+1,P.setDate(1),11>ye?P.setMonth(ye+1):(P.setMonth(0),P.setFullYear(P.getFullYear()+1))}return ye=new Date(P.getFullYear()+1,0,4),ee=B(new Date(P.getFullYear(),0,4)),ye=B(ye),0>=O(ee,P)?0>=O(ye,P)?P.getFullYear()+1:P.getFullYear():P.getFullYear()-1}s>>>=0,c>>>=0,f>>>=0,y>>>=0;var L=a()[y+40>>>2>>>0];for(var j in y={fc:i()[y>>>2>>>0],ec:i()[y+4>>>2>>>0],Gb:i()[y+8>>>2>>>0],Kb:i()[y+12>>>2>>>0],Hb:i()[y+16>>>2>>>0],Cb:i()[y+20>>>2>>>0],ub:i()[y+24>>>2>>>0],Bb:i()[y+28>>>2>>>0],nc:i()[y+32>>>2>>>0],dc:i()[y+36>>>2>>>0],hc:L?Ae(L):""},f=Ae(f),L={"%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(j,"g"),L[j]);var me="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),he="January February March April May June July August September October November December".split(" ");for(j in L={"%a":P=>me[P.ub].substring(0,3),"%A":P=>me[P.ub],"%b":P=>he[P.Hb].substring(0,3),"%B":P=>he[P.Hb],"%C":P=>C((P.Cb+1900)/100|0,2),"%d":P=>C(P.Kb,2),"%e":P=>v(P.Kb,2," "),"%g":P=>N(P).toString().substring(2),"%G":N,"%H":P=>C(P.Gb,2),"%I":P=>((P=P.Gb)==0?P=12:12<P&&(P-=12),C(P,2)),"%j":P=>{for(var ee=0,ye=0;ye<=P.Hb-1;ee+=(Bt(P.Cb+1900)?Oi:Di)[ye++]);return C(P.Kb+ee,3)},"%m":P=>C(P.Hb+1,2),"%M":P=>C(P.ec,2),"%n":()=>`
-`,"%p":P=>0<=P.Gb&&12>P.Gb?"AM":"PM","%S":P=>C(P.fc,2),"%t":()=>"	","%u":P=>P.ub||7,"%U":P=>C(Math.floor((P.Bb+7-P.ub)/7),2),"%V":P=>{var ee=Math.floor((P.Bb+7-(P.ub+6)%7)/7);if(2>=(P.ub+371-P.Bb-2)%7&&ee++,ee)ee==53&&((ye=(P.ub+371-P.Bb)%7)==4||ye==3&&Bt(P.Cb)||(ee=1));else{ee=52;var ye=(P.ub+7-P.Bb-1)%7;(ye==4||ye==5&&Bt(P.Cb%400-1))&&ee++}return C(ee,2)},"%w":P=>P.ub,"%W":P=>C(Math.floor((P.Bb+7-(P.ub+6)%7)/7),2),"%y":P=>(P.Cb+1900).toString().substring(2),"%Y":P=>P.Cb+1900,"%z":P=>{var ee=0<=(P=P.dc);return P=Math.abs(P)/60,(ee?"+":"-")+("0000"+(P/60*100+P%60)).slice(-4)},"%Z":P=>P.hc,"%%":()=>"%"},f=f.replace(/%%/g,"\0\0"),L)f.includes(j)&&(f=f.replace(new RegExp(j,"g"),L[j](y)));return j=function(P){var ee=Array(wn(P)+1);return Yo(P,ee,0,ee.length),ee}(f=f.replace(/\0\0/g,"%")),j.length>c?0:(tp(j,s),j.length-1)}function rp(s,c,f,y){return zi(s>>>0,c>>>0,f>>>0,y>>>0)}b||function(){for(var s=u.numThreads-1;s--;)Ho();Lt.unshift(()=>{Fe++,function(c){b?c():Promise.all(ct.map(Wo)).then(c)}(()=>ko())})}();for(var Bi=Array(256),mr=0;256>mr;++mr)Bi[mr]=String.fromCharCode(mr);si=Bi,pt=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-_n.length;var np=[yn,Mo,Go,jo,Ko,Xo,Zo,Qo,Jo,ei,ti,ri,ni,oi,ii,ai,vi,$i,Ci,Ti,Ai,Ei,ki,Pi],q=function(){function s(f,y){return q=f.exports,q=function(){var v=q,C={};for(let[O,B]of Object.entries(v))C[O]=typeof B=="function"?(...N)=>{dr.push(O);try{return B(...N)}finally{Se||(dr.pop(),Je&&ft===1&&dr.length===0&&(ft=0,bt+=1,ur(Li),typeof Fibers<"u"&&Fibers.oc()))}}:B;return C}(),q=function(){var v=q,C=B=>N=>B(N)>>>0,O=B=>()=>B()>>>0;return(v=Object.assign({},v)).Ca=C(v.Ca),v.fb=O(v.fb),v.gb=C(v.gb),v.emscripten_main_runtime_thread_id=O(v.emscripten_main_runtime_thread_id),v.sb=C(v.sb),v.tb=O(v.tb),v}(),Uo.push(q.ib),Pe.unshift(q.Ba),K=y,ko(),q}var c=Bo();if(Fe++,u.instantiateWasm)try{return u.instantiateWasm(c,s)}catch(f){Y(`Module.instantiateWasm callback failed with error: ${f}`),m(f)}return mn||=u.locateFile?Po("ort-wasm-simd-threaded.jsep.wasm")?"ort-wasm-simd-threaded.jsep.wasm":u.locateFile?u.locateFile("ort-wasm-simd-threaded.jsep.wasm",D):D+"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,y){var v=mn;return z||typeof WebAssembly.instantiateStreaming!="function"||Po(v)||Oo(v)||typeof fetch!="function"?zo(v,f,y):fetch(v,{credentials:"same-origin"}).then(C=>WebAssembly.instantiateStreaming(C,f).then(y,function(O){return Y(`wasm streaming compile failed: ${O}`),Y("falling back to ArrayBuffer instantiation"),zo(v,f,y)}))}(c,function(f){s(f.instance,f.module)}).catch(m),{}}(),Ri=s=>(Ri=q.Ca)(s),Mi=()=>(Mi=q.Da)();u._OrtInit=(s,c)=>(u._OrtInit=q.Ea)(s,c),u._OrtGetLastError=(s,c)=>(u._OrtGetLastError=q.Fa)(s,c),u._OrtCreateSessionOptions=(s,c,f,y,v,C,O,B,N,L)=>(u._OrtCreateSessionOptions=q.Ga)(s,c,f,y,v,C,O,B,N,L),u._OrtAppendExecutionProvider=(s,c)=>(u._OrtAppendExecutionProvider=q.Ha)(s,c),u._OrtAddFreeDimensionOverride=(s,c,f)=>(u._OrtAddFreeDimensionOverride=q.Ia)(s,c,f),u._OrtAddSessionConfigEntry=(s,c,f)=>(u._OrtAddSessionConfigEntry=q.Ja)(s,c,f),u._OrtReleaseSessionOptions=s=>(u._OrtReleaseSessionOptions=q.Ka)(s),u._OrtCreateSession=(s,c,f)=>(u._OrtCreateSession=q.La)(s,c,f),u._OrtReleaseSession=s=>(u._OrtReleaseSession=q.Ma)(s),u._OrtGetInputOutputCount=(s,c,f)=>(u._OrtGetInputOutputCount=q.Na)(s,c,f),u._OrtGetInputName=(s,c)=>(u._OrtGetInputName=q.Oa)(s,c),u._OrtGetOutputName=(s,c)=>(u._OrtGetOutputName=q.Pa)(s,c),u._OrtFree=s=>(u._OrtFree=q.Qa)(s),u._OrtCreateTensor=(s,c,f,y,v,C)=>(u._OrtCreateTensor=q.Ra)(s,c,f,y,v,C),u._OrtGetTensorData=(s,c,f,y,v)=>(u._OrtGetTensorData=q.Sa)(s,c,f,y,v),u._OrtReleaseTensor=s=>(u._OrtReleaseTensor=q.Ta)(s),u._OrtCreateRunOptions=(s,c,f,y)=>(u._OrtCreateRunOptions=q.Ua)(s,c,f,y),u._OrtAddRunConfigEntry=(s,c,f)=>(u._OrtAddRunConfigEntry=q.Va)(s,c,f),u._OrtReleaseRunOptions=s=>(u._OrtReleaseRunOptions=q.Wa)(s),u._OrtCreateBinding=s=>(u._OrtCreateBinding=q.Xa)(s),u._OrtBindInput=(s,c,f)=>(u._OrtBindInput=q.Ya)(s,c,f),u._OrtBindOutput=(s,c,f,y)=>(u._OrtBindOutput=q.Za)(s,c,f,y),u._OrtClearBoundOutputs=s=>(u._OrtClearBoundOutputs=q._a)(s),u._OrtReleaseBinding=s=>(u._OrtReleaseBinding=q.$a)(s),u._OrtRunWithBinding=(s,c,f,y,v)=>(u._OrtRunWithBinding=q.ab)(s,c,f,y,v),u._OrtRun=(s,c,f,y,v,C,O,B)=>(u._OrtRun=q.bb)(s,c,f,y,v,C,O,B),u._OrtEndProfiling=s=>(u._OrtEndProfiling=q.cb)(s),u._JsepOutput=(s,c,f)=>(u._JsepOutput=q.db)(s,c,f),u._JsepGetNodeName=s=>(u._JsepGetNodeName=q.eb)(s);var fr,Mt=()=>(Mt=q.fb)(),hr=u._malloc=s=>(hr=u._malloc=q.gb)(s),et=u._free=s=>(et=u._free=q.hb)(s),On=(s,c,f,y,v,C)=>(On=q.kb)(s,c,f,y,v,C),Ui=()=>(Ui=q.lb)(),Vi=(s,c,f,y,v)=>(Vi=q.mb)(s,c,f,y,v),Dn=s=>(Dn=q.nb)(s),gr=s=>(gr=q.ob)(s),Ni=()=>(Ni=q.pb)(),Wi=(s,c)=>(Wi=q.qb)(s,c),yr=s=>(yr=q.rb)(s),zn=s=>(zn=q.sb)(s),Bn=()=>(Bn=q.tb)(),Hi=u.dynCall_ii=(s,c)=>(Hi=u.dynCall_ii=q.vb)(s,c),Gi=s=>(Gi=q.wb)(s),Li=()=>(Li=q.xb)(),Fi=s=>(Fi=q.yb)(s),qi=()=>(qi=q.zb)();function ji(){0<Fe||(b?(p(u),b||ar(Pe),startWorker(u)):(ar(Lt),0<Fe||fr||(fr=!0,u.calledRun=!0,Se||(b||ar(Pe),p(u),b||ar(ze)))))}return u.___start_em_js=862461,u.___stop_em_js=862683,u.stackSave=()=>Bn(),u.stackRestore=s=>yr(s),u.stackAlloc=s=>zn(s),u.UTF8ToString=Ae,u.stringToUTF8=zt,u.lengthBytesUTF8=wn,yt=function s(){fr||ji(),fr||(yt=s)},ji(),h}),mp=Aa;globalThis.self?.name==="em-pthread"&&Aa()});var Ut,fp,hp,gp,Pa,Oa,yp,Da,jt=U(()=>{"use strict";Tr();Ut= false?0:import.meta.url??(typeof document<"u"?document.currentScript?.src:typeof self<"u"?self.location?.href:void 0),fp= false||typeof location>"u"?void 0:location.origin,hp=(e,t)=>{try{let r=t??Ut;return(r?new URL(e,r):new URL(e)).origin===fp}catch{return!1}},gp=async e=>{let r=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(r)},Pa=(Ta(),br(Ca)).default,Oa=async()=>{if(!Ut)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(hp(Ut))return[void 0,Pa()];let e=await gp(Ut);return[e,Pa(e)]},yp=(ka(),br(Ea)).default,Da=async(e,t,r)=>[void 0,yp]});var Ln,Fn,Rr,za,bp,wp,Ar,Ie,St=U(()=>{"use strict";jt();Fn=!1,Rr=!1,za=!1,bp=()=>{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}},wp=()=>{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}},Ar=async e=>{if(Fn)return Promise.resolve();if(Rr)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(za)throw new Error("previous call to 'initializeWebAssembly()' failed.");Rr=!0;let t=e.initTimeout,r=e.numThreads;if(!wp())throw new Error("WebAssembly SIMD is not supported in the current environment.");let n=bp();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,l=a?.href??a,d=o?.wasm,p=d?.href??d,m=e.wasmBinary,[u,h]=await Da(l,i,r>1),w=!1,g=[];if(t>0&&g.push(new Promise(b=>{setTimeout(()=>{w=!0,b()},t)})),g.push(new Promise((b,x)=>{let _={numThreads:r};m?_.wasmBinary=m:(p||i)&&(_.locateFile=($,S)=>p??(i??S)+$),h(_).then($=>{Rr=!1,Fn=!0,Ln=$,b(),u&&URL.revokeObjectURL(u)},$=>{Rr=!1,za=!0,x($)})})),await Promise.race(g),w)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Ie=()=>{if(Fn&&Ln)return Ln;throw new Error("WebAssembly is not initialized yet.")}});var Te,Yt,_e,Mr=U(()=>{"use strict";St();Te=(e,t)=>{let r=Ie(),n=r.lengthBytesUTF8(e)+1,o=r._malloc(n);return r.stringToUTF8(e,o,n),t.push(o),o},Yt=(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")Yt(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: ${typeof i}`)})},_e=e=>{let t=Ie(),r=t.stackSave();try{let n=t.stackAlloc(8);t._OrtGetLastError(n,n+4);let o=t.HEAP32[n/4],i=t.HEAPU32[n/4+1],a=i?t.UTF8ToString(i):"";throw new Error(`${e} ERROR_CODE: ${o}, ERROR_MESSAGE: ${a}`)}finally{t.stackRestore(r)}}});var Ba,Ra=U(()=>{"use strict";St();Mr();Ba=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=Te(e.tag,n)),r=t._OrtCreateRunOptions(o.logSeverityLevel,o.logVerbosityLevel,!!o.terminate,i),r===0&&_e("Can't create run options."),e?.extra!==void 0&&Yt(e.extra,"",new WeakSet,(a,l)=>{let d=Te(a,n),p=Te(l,n);t._OrtAddRunConfigEntry(r,d,p)!==0&&_e(`Can't set a run config entry: ${a} - ${l}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseRunOptions(r),n.forEach(a=>t._free(a)),i}}});var vp,$p,_p,xp,Ma,Ua=U(()=>{"use strict";St();Mr();vp=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: ${e}`)}},$p=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},_p=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)},xp=(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 l=n?.deviceType;if(l){let d=Te("deviceType",r),p=Te(l,r);Ie()._OrtAddSessionConfigEntry(e,d,p)!==0&&_e(`Can't set a session config entry: 'deviceType' - ${l}.`)}}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 l=Te("preferredLayout",r),d=Te(a.preferredLayout,r);Ie()._OrtAddSessionConfigEntry(e,l,d)!==0&&_e(`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=Te(o,r);Ie()._OrtAppendExecutionProvider(e,i)!==0&&_e(`Can't append execution provider: ${o}.`)}},Ma=e=>{let t=Ie(),r=0,n=[],o=e||{};_p(o);try{let i=vp(o.graphOptimizationLevel??"all"),a=$p(o.executionMode??"sequential"),l=typeof o.logId=="string"?Te(o.logId,n):0,d=o.logSeverityLevel??2;if(!Number.isInteger(d)||d<0||d>4)throw new Error(`log serverity level is not valid: ${d}`);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"?Te(o.optimizedModelFilePath,n):0;if(r=t._OrtCreateSessionOptions(i,!!o.enableCpuMemArena,!!o.enableMemPattern,a,!!o.enableProfiling,0,l,d,p,m),r===0&&_e("Can't create session options."),o.executionProviders&&xp(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=Te("enableGraphCapture",n),h=Te(o.enableGraphCapture.toString(),n);t._OrtAddSessionConfigEntry(r,u,h)!==0&&_e(`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 w=Te(u,n);t._OrtAddFreeDimensionOverride(r,w,h)!==0&&_e(`Can't set a free dimension override: ${u} - ${h}.`)}return o.extra!==void 0&&Yt(o.extra,"",new WeakSet,(u,h)=>{let w=Te(u,n),g=Te(h,n);t._OrtAddSessionConfigEntry(r,w,g)!==0&&_e(`Can't set a session config entry: ${u} - ${h}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseSessionOptions(r),n.forEach(a=>t._free(a)),i}}});var qn,ht,It,Ur,Xt,Vr,jn,Q=U(()=>{"use strict";qn=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;default:throw new Error(`unsupported data type: ${e}`)}},ht=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";default:throw new Error(`unsupported data type: ${e}`)}},It=e=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][e],Ur=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}`)}},Xt=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}`)}},Vr=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool",jn=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;default:throw new Error(`unsupported data location: ${e}`)}}});var Zt,Kn=U(()=>{"use strict";Tr();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: ${e}, no response body.`);let o=t.body.getReader(),i;try{i=new ArrayBuffer(n)}catch(l){if(l instanceof RangeError){let d=Math.ceil(n/65536);i=new WebAssembly.Memory({initial:d,maximum:d}).buffer}else throw l}let a=0;for(;;){let{done:l,value:d}=await o.read();if(l)break;let p=d.byteLength;new Uint8Array(i,a,p).set(d),a+=p}return new Uint8Array(i,0,n)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}});var Sp,Ip,Va,Na,Wa,Cp,ve,dt=U(()=>{"use strict";Q();Sp=["V","I","W","E","F"],Ip=(e,t)=>{console.log(`[${Sp[e]},${new Date().toISOString()}]${t}`)},Wa=(e,t)=>{Va=e,Na=t},Cp=(e,t)=>{let r=Xt(e),n=Xt(Va);r>=n&&Ip(r,typeof t=="function"?t():t)},ve=(...e)=>{Na&&Cp(...e)}});var Ha,Ga=U(()=>{"use strict";Q();Ha=(e,t)=>new(Ur(t))(e)});var Nr=U(()=>{"use strict"});var La,Yn,Xn,Tp,Ap,Fa,Qn,Zn,ja,Ka=U(()=>{"use strict";dt();Nr();La=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]]),Yn=[],Xn=e=>Math.ceil(e/16)*16,Tp=e=>{for(let t=0;t<Yn.length;t++){let r=Yn[t];if(e<=r)return r}return Math.ceil(e/16)*16},Ap=1,Fa=()=>Ap++,Qn=async(e,t,r,n)=>{let o=Xn(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 l=i.getMappedRange();if(n){let d=n();return d.set(new Uint8Array(l,0,r)),d}else return new Uint8Array(l.slice(0,r))}finally{i.destroy()}},Zn=class{constructor(t){this.backend=t;this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map,this.capturedPendingBuffers=new Map;for(let[r]of La)Yn.push(r),this.freeBuffers.set(r,[]),this.freeUniformBuffers.set(r,[])}upload(t,r){let n=r.buffer,o=r.byteOffset,i=r.byteLength,a=Xn(i),l=this.storageCache.get(t);if(!l)throw new Error("gpu data for uploading does not exist");if(l.originalSize!==i)throw new Error(`inconsistent data size. gpu data size=${l.originalSize}, data size=${i}`);let d=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),p=d.getMappedRange();new Uint8Array(p).set(new Uint8Array(n,o,i)),d.unmap();let m=this.backend.getCommandEncoder();this.backend.endComputePass(),m.copyBufferToBuffer(d,0,l.gpuData.buffer,0,a),ve("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(d)}memcpy(t,r){let n=this.storageCache.get(t);if(!n)throw new Error("source gpu data for memcpy does not exist");let o=this.storageCache.get(r);if(!o)throw new Error("destination gpu data for memcpy does not exist");if(n.originalSize!==o.originalSize)throw new Error("inconsistent source and destination gpu data size");let i=Xn(n.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(n.gpuData.buffer,0,o.gpuData.buffer,0,i)}registerExternalBuffer(t,r,n){let o;if(n){if(o=this.externalBuffers.get(n),o===void 0)throw new Error("previous buffer is not registered");if(t===n)return ve("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${o}, buffer is the same, skip.`),o;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
-             Please use the previous external buffer!`);this.externalBuffers.delete(n)}else o=Fa();return this.storageCache.set(o,{gpuData:{id:o,type:0,buffer:t},originalSize:r}),this.externalBuffers.set(t,o),ve("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${o}, registered.`),o}unregisterExternalBuffer(t){let r=this.externalBuffers.get(t);r!==void 0&&(this.storageCache.delete(r),this.externalBuffers.delete(t),ve("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${r}`))}create(t,r=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let n=Tp(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 l={id:Fa(),type:0,buffer:o};return this.storageCache.set(l.id,{gpuData:l,originalSize:t}),ve("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${l.id}`),l}get(t){return this.storageCache.get(t)?.gpuData}release(t){let r=this.storageCache.get(t);if(!r)throw new Error("releasing data does not exist");return ve("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(t,r){let n=this.storageCache.get(t);if(!n)throw new Error("data does not exist");await Qn(this.backend,n.gpuData.buffer,n.originalSize,r)}refreshPendingBuffers(){for(let t of this.buffersForUploadingPending)t.destroy();if(this.buffersForUploadingPending=[],this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let r=La.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}onReleaseSession(t){let r=this.capturedPendingBuffers.get(t);r&&(r.forEach(n=>{n.destroy()}),this.capturedPendingBuffers.delete(t))}},ja=(...e)=>new Zn(...e)});var Jn,te,Ce=U(()=>{"use strict";Jn=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}},te=e=>new Jn(e)});var eo,tt,k,Ct,Wr,Hr,Gr,ae=U(()=>{"use strict";eo=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),l=new Array(a);if(n){if(o<2||i<2)return;let d=eo.calcMatMulShape([t[o-2],t[o-1]],[r[i-2],r[i-1]]);if(d===void 0)return;[l[a-2],l[a-1]]=d}for(let d=n?3:1;d<=a;d++){let p=o-d<0?1:t[o-d],m=i-d<0?1:r[i-d];if(p!==m&&p>1&&m>1)return;let u=Math.max(p,m);if(p&&m)l[a-d]=Math.max(p,m);else{if(u>1)return;l[a-d]=0}}return l}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*=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])}},Ct=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 l=0;l<r.length-2;l++)l>=n.length?n.push(r[l+2]):n[l]=r[l+2];for(let l=0;l<n.length;l++)if(l<o.length){if(o[l]<0)throw new Error("strides should be greater than or equal to 1")}else o.push(1);for(let l=0;l<n.length;l++)if(l<i.length){if(i[l]<0)throw new Error("dilations should be greater than or equal to 1")}else i.push(1);for(let l=0;l<n.length*2;l++)if(l<a.length){if(a[l]<0)throw new Error("pad should be greater than or equal to 1")}else a.push(0);for(let l=0;l<n.length;l++){if(n[l]<=0)throw new Error("kernel shapes need to be greater than 0");if(a[l]>=n[l]||a[l+n.length]>=n[l])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,n,o,i,a,l){if(l){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 d=0;d<t.length-2;d++)e.adjustPadAndReturnShape(t[d+(a?1:2)],r[d],n[d],o[d],i,d,d+t.length-2,l)}}static computePoolOutputShape(t,r,n,o,i,a,l){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let d=[r[0],r[1]];return e.computeShapeHelper(t,r,d,n,o,i,a,l),d}static computeConvOutputShape(t,r,n,o,i,a,l){if(t.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let d=[t[0],r[0]];return e.computeShapeHelper(!1,t,d,n,o,i,a,l),d}static computeShapeHelper(t,r,n,o,i,a,l,d){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],l,p,p+r.length-2,d))}static adjustPadAndReturnShape(t,r,n,o,i,a,l,d){let p=n*(o-1)+1;if(d&&d!=="NOTSET")switch(d){case"VALID":return i[a]=0,i[l]=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(d==="SAME_LOWER"?(u+1)/2:u/2),i[l]=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[l]-p)/r+1)}},Wr=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,l,d;r?(a=t[1],l=t[0]):(a=t[0],l=t[1]);let p=-1;if(o?(d=n[0],p=1):(d=n[1],p=0),n[p]!==l)throw new Error("dimension mismatch");if(a<=0||d<=0||l<=0)throw new Error("invalid shape specified");if(i&&!tt.isValidBroadcast(i,[a,d]))throw new Error("gemm: invalid bias shape for broadcast");return[a,d,l]}},Hr=-34028234663852886e22,Gr=34028234663852886e22});var Tt,ro,ge,Oe,V,$e,gt,At,Xe,G,no,E,M,Lr,to,Ya,Nt,ie=U(()=>{"use strict";Q();ae();Tt=64,ro=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(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>"];default:throw new Error(`Unknown data type: ${e}`)}},ge=(e,t=1)=>{let r=ro(e,t);return typeof r=="string"?r:r[0]},Oe=(e,t=1)=>{let r=ro(e,t);return typeof r=="string"?r:r[1]},V=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:k.computeStrides(r)})}),t},$e=e=>e%4===0?4:e%2===0?2:1,gt=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,At=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,Xe=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,G=(e,t,r,n)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?n==="f16"?`${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}[${t}]`:e,no=(e,t,r,n,o)=>{let i=typeof r=="number",a=i?r:r.length,l=[...new Array(a).keys()],d=a<2?"u32":a<=4?`vec${a}<u32>`:`array<u32, ${a}>`,p=ro(t,o),m=typeof p=="string"?p:p[1],u=typeof p=="string"?p:p[0],h={indices:d,value:m,storage:u,tensor:t},w=R=>typeof R=="string"?R:`${R}u`,g={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},b=i?"uniforms.":"",x=`${b}${e}_shape`,_=`${b}${e}_strides`,$="";for(let R=0;R<a-1;R++)$+=`
-    let dim${R} = current / ${G(_,R,a)};
-    let rest${R} = current % ${G(_,R,a)};
+`),s=function(J){var de=Function;if(!(de instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof de} which is not a function`);var ye=mi(de.name||"unknownFunctionName",function(){});return ye.prototype=de.prototype,ye=new ye,(J=de.apply(ye,J))instanceof Object?J:ye}(B)(...N),f=`methodCaller<(${c.map(J=>J.name).join(", ")}) => ${b.name}>`,$c(mi(f,s))}function Sc(s){return s=ur(s>>>0),Fe(u[s])}function Ic(s,c){return c>>>=0,s=Ge(s>>>0),c=Ge(c),Fe(s[c])}function Cc(s){9<(s>>>=0)&&(at[s+1]+=1)}function Tc(){return Fe([])}function Ac(s){s=Ge(s>>>0);for(var c=Array(s.length),f=0;f<s.length;f++)c[f]=s[f];return Fe(c)}function kc(s){return Fe(ur(s>>>0))}function Ec(){return Fe({})}function Pc(s){for(var c=Ge(s>>>=0);c.length;){var f=c.pop();c.pop()(f)}wn(s)}function Oc(s,c,f){c>>>=0,f>>>=0,s=Ge(s>>>0),c=Ge(c),f=Ge(f),s[c]=f}function zc(s,c){return c>>>=0,s=(s=xn(s>>>0,"_emval_take_value")).readValueFromPointer(c),Fe(s)}function Bc(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 fi=s=>s%4==0&&(s%100!=0||s%400==0),hi=[0,31,60,91,121,152,182,213,244,274,305,335],gi=[0,31,59,90,120,151,181,212,243,273,304,334];function Dc(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=(fi(s.getFullYear())?hi:gi)[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 Rc(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(),v=new Date(c.getFullYear(),6,1).getTimezoneOffset(),C=new Date(c.getFullYear(),0,1).getTimezoneOffset(),P=Math.min(C,v);return 0>f?i()[s+32>>>2>>>0]=+(v!=C&&P==b):0<f!=(P==b)&&(v=Math.max(C,v),c.setTime(c.getTime()+6e4*((0<f?P:v)-b))),i()[s+24>>>2>>>0]=c.getDay(),f=(fi(c.getFullYear())?hi:gi)[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 yi(s,c,f,b,v,C,P){return y?ve(16,1,s,c,f,b,v,C,P):-52}function bi(s,c,f,b,v,C){if(y)return ve(17,1,s,c,f,b,v,C)}function Mc(s,c,f,b){s>>>=0,c>>>=0,f>>>=0,b>>>=0;var v=new Date().getFullYear(),C=new Date(v,0,1),P=new Date(v,6,1);v=C.getTimezoneOffset();var B=P.getTimezoneOffset(),N=Math.max(v,B);a()[s>>>2>>>0]=60*N,i()[c>>>2>>>0]=+(v!=B),C=(s=H=>H.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1])(C),P=s(P),B<v?(Pt(C,f,17),Pt(P,b,17)):(Pt(C,b,17),Pt(P,f,17))}var In=[],wi=(s,c)=>{In.length=0;for(var f;f=r()[s++>>>0];){var b=f!=105;c+=(b&=f!=112)&&c%8?4:0,In.push(f==112?a()[c>>>2>>>0]:f==106?G[c>>>3]:f==105?i()[c>>>2>>>0]:d()[c>>>3>>>0]),c+=b?8:4}return In};function Uc(s,c,f){return s>>>=0,c=wi(c>>>0,f>>>0),cn[s](...c)}function Vc(s,c,f){return s>>>=0,c=wi(c>>>0,f>>>0),cn[s](...c)}var Nc=()=>{},Wc=()=>Date.now();function Lc(s,c){return Y(Ee(s>>>0,c>>>0))}var vi,Hc=()=>{throw gt+=1,"unwind"};function Gc(){return 4294901760}vi=()=>performance.timeOrigin+performance.now();var Fc=()=>navigator.hardwareConcurrency;function qc(){return Et("Cannot use emscripten_pc_get_function without -sUSE_OFFSET_CONVERTER"),0}function Kc(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 v=Math;b=Math.max(s,b);e:{v=(v.min.call(v,4294901760,b+(65536-b%65536)%65536)-ie.buffer.byteLength+65535)/65536;try{ie.grow(v),Se();var C=1;break e}catch{}C=void 0}if(C)return!0}return!1}var dr=()=>(Et("Cannot use convertFrameToPC (needed by __builtin_return_address) without -sUSE_OFFSET_CONVERTER"),0),Ot={},$i=s=>{s.forEach(c=>{var f=dr();f&&(Ot[f]=c)})};function jc(){var s=Error().stack.toString().split(`
+`);return s[0]=="Error"&&s.shift(),$i(s),Ot.Gb=dr(),Ot.Tb=s,Ot.Gb}function Yc(s,c,f){if(s>>>=0,c>>>=0,Ot.Gb==s)var b=Ot.Tb;else(b=Error().stack.toString().split(`
+`))[0]=="Error"&&b.shift(),$i(b);for(var v=3;b[v]&&dr()!=s;)++v;for(s=0;s<f&&b[s+v];++s)i()[c+4*s>>>2>>>0]=dr();return s}var Cn,Tn={},_i=()=>{if(!Cn){var s,c={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:T||"./this.program"};for(s in Tn)Tn[s]===void 0?delete c[s]:c[s]=Tn[s];var f=[];for(s in c)f.push(`${s}=${c[s]}`);Cn=f}return Cn};function xi(s,c){if(y)return ve(18,1,s,c);s>>>=0,c>>>=0;var f=0;return _i().forEach((b,v)=>{var C=c+f;for(v=a()[s+4*v>>>2>>>0]=C,C=0;C<b.length;++C)t()[v++>>>0]=b.charCodeAt(C);t()[v>>>0]=0,f+=b.length+1}),0}function Si(s,c){if(y)return ve(19,1,s,c);s>>>=0,c>>>=0;var f=_i();a()[s>>>2>>>0]=f.length;var b=0;return f.forEach(v=>b+=v.length+1),a()[c>>>2>>>0]=b,0}function Ii(s){return y?ve(20,1,s):52}function Ci(s,c,f,b){return y?ve(21,1,s,c,f,b):52}function Ti(s,c,f,b){return y?ve(22,1,s,c,f,b):70}var Xc=[null,[],[]];function Ai(s,c,f,b){if(y)return ve(23,1,s,c,f,b);c>>>=0,f>>>=0,b>>>=0;for(var v=0,C=0;C<f;C++){var P=a()[c>>>2>>>0],B=a()[c+4>>>2>>>0];c+=8;for(var N=0;N<B;N++){var H=r()[P+N>>>0],J=Xc[s];H===0||H===10?((s===1?j:Y)(Lo(J,0)),J.length=0):J.push(H)}v+=B}return a()[b>>>2>>>0]=v,0}y||function(){for(var s=u.numThreads-1;s--;)Uo();Wt.unshift(()=>{He++,function(c){y?c():Promise.all(dt.map(Mo)).then(c)}(()=>Co())})}();for(var ki=Array(256),lr=0;256>lr;++lr)ki[lr]=String.fromCharCode(lr);ni=ki,lt=u.BindingError=class extends Error{constructor(s){super(s),this.name="BindingError"}},u.InternalError=class extends Error{constructor(s){super(s),this.name="InternalError"}},at.push(0,1,void 0,1,null,1,!0,1,!1,1),u.count_emval_handles=()=>at.length/2-5-bn.length;var Zc=[fn,zo,Vo,Ho,Go,qo,Ko,jo,Yo,Xo,Zo,Qo,Jo,ei,ti,ri,yi,bi,xi,Si,Ii,Ci,Ti,Ai],F=function(){function s(f,b){return F=f.exports,F=function(){var v=F,C={};for(let[P,B]of Object.entries(v))C[P]=typeof B=="function"?(...N)=>{ar.push(P);try{return B(...N)}finally{ge||(ar.pop(),Ze&&pt===1&&ar.length===0&&(pt=0,gt+=1,ir(Ui),typeof Fibers<"u"&&Fibers.ac()))}}:B;return C}(),F=function(){var v=F,C=B=>N=>B(N)>>>0,P=B=>()=>B()>>>0;return(v=Object.assign({},v)).Aa=C(v.Aa),v.db=P(v.db),v.eb=C(v.eb),v.emscripten_main_runtime_thread_id=P(v.emscripten_main_runtime_thread_id),v.qb=C(v.qb),v.rb=P(v.rb),v}(),Bo.push(F.gb),Te.unshift(F.za),q=b,Co(),F}var c=Po();if(He++,u.instantiateWasm)try{return u.instantiateWasm(c,s)}catch(f){Y(`Module.instantiateWasm callback failed with error: ${f}`),m(f)}return ln||=u.locateFile?To("ort-wasm-simd-threaded.jsep.wasm")?"ort-wasm-simd-threaded.jsep.wasm":u.locateFile?u.locateFile("ort-wasm-simd-threaded.jsep.wasm",O):O+"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 v=ln;return z||typeof WebAssembly.instantiateStreaming!="function"||To(v)||Ao(v)||typeof fetch!="function"?Eo(v,f,b):fetch(v,{credentials:"same-origin"}).then(C=>WebAssembly.instantiateStreaming(C,f).then(b,function(P){return Y(`wasm streaming compile failed: ${P}`),Y("falling back to ArrayBuffer instantiation"),Eo(v,f,b)}))}(c,function(f){s(f.instance,f.module)}).catch(m),{}}(),Ei=s=>(Ei=F.Aa)(s),Pi=()=>(Pi=F.Ba)();u._OrtInit=(s,c)=>(u._OrtInit=F.Ca)(s,c),u._OrtGetLastError=(s,c)=>(u._OrtGetLastError=F.Da)(s,c),u._OrtCreateSessionOptions=(s,c,f,b,v,C,P,B,N,H)=>(u._OrtCreateSessionOptions=F.Ea)(s,c,f,b,v,C,P,B,N,H),u._OrtAppendExecutionProvider=(s,c)=>(u._OrtAppendExecutionProvider=F.Fa)(s,c),u._OrtAddFreeDimensionOverride=(s,c,f)=>(u._OrtAddFreeDimensionOverride=F.Ga)(s,c,f),u._OrtAddSessionConfigEntry=(s,c,f)=>(u._OrtAddSessionConfigEntry=F.Ha)(s,c,f),u._OrtReleaseSessionOptions=s=>(u._OrtReleaseSessionOptions=F.Ia)(s),u._OrtCreateSession=(s,c,f)=>(u._OrtCreateSession=F.Ja)(s,c,f),u._OrtReleaseSession=s=>(u._OrtReleaseSession=F.Ka)(s),u._OrtGetInputOutputCount=(s,c,f)=>(u._OrtGetInputOutputCount=F.La)(s,c,f),u._OrtGetInputName=(s,c)=>(u._OrtGetInputName=F.Ma)(s,c),u._OrtGetOutputName=(s,c)=>(u._OrtGetOutputName=F.Na)(s,c),u._OrtFree=s=>(u._OrtFree=F.Oa)(s),u._OrtCreateTensor=(s,c,f,b,v,C)=>(u._OrtCreateTensor=F.Pa)(s,c,f,b,v,C),u._OrtGetTensorData=(s,c,f,b,v)=>(u._OrtGetTensorData=F.Qa)(s,c,f,b,v),u._OrtReleaseTensor=s=>(u._OrtReleaseTensor=F.Ra)(s),u._OrtCreateRunOptions=(s,c,f,b)=>(u._OrtCreateRunOptions=F.Sa)(s,c,f,b),u._OrtAddRunConfigEntry=(s,c,f)=>(u._OrtAddRunConfigEntry=F.Ta)(s,c,f),u._OrtReleaseRunOptions=s=>(u._OrtReleaseRunOptions=F.Ua)(s),u._OrtCreateBinding=s=>(u._OrtCreateBinding=F.Va)(s),u._OrtBindInput=(s,c,f)=>(u._OrtBindInput=F.Wa)(s,c,f),u._OrtBindOutput=(s,c,f,b)=>(u._OrtBindOutput=F.Xa)(s,c,f,b),u._OrtClearBoundOutputs=s=>(u._OrtClearBoundOutputs=F.Ya)(s),u._OrtReleaseBinding=s=>(u._OrtReleaseBinding=F.Za)(s),u._OrtRunWithBinding=(s,c,f,b,v)=>(u._OrtRunWithBinding=F._a)(s,c,f,b,v),u._OrtRun=(s,c,f,b,v,C,P,B)=>(u._OrtRun=F.$a)(s,c,f,b,v,C,P,B),u._OrtEndProfiling=s=>(u._OrtEndProfiling=F.ab)(s),u._JsepOutput=(s,c,f)=>(u._JsepOutput=F.bb)(s,c,f),u._JsepGetNodeName=s=>(u._JsepGetNodeName=F.cb)(s);var cr,zt=()=>(zt=F.db)(),pr=u._malloc=s=>(pr=u._malloc=F.eb)(s),Qe=u._free=s=>(Qe=u._free=F.fb)(s),An=(s,c,f,b,v,C)=>(An=F.ib)(s,c,f,b,v,C),Oi=()=>(Oi=F.jb)(),zi=(s,c,f,b,v)=>(zi=F.kb)(s,c,f,b,v),kn=s=>(kn=F.lb)(s),mr=s=>(mr=F.mb)(s),Bi=()=>(Bi=F.nb)(),Di=(s,c)=>(Di=F.ob)(s,c),fr=s=>(fr=F.pb)(s),En=s=>(En=F.qb)(s),Pn=()=>(Pn=F.rb)(),Ri=u.dynCall_ii=(s,c)=>(Ri=u.dynCall_ii=F.tb)(s,c),Mi=s=>(Mi=F.ub)(s),Ui=()=>(Ui=F.vb)(),Vi=s=>(Vi=F.wb)(s),Ni=()=>(Ni=F.xb)();function Wi(){0<He||(y?(p(u),y||nr(Te),startWorker(u)):(nr(Wt),0<He||cr||(cr=!0,u.calledRun=!0,ge||(y||nr(Te),p(u),y||nr(Oe)))))}return u.stackSave=()=>Pn(),u.stackRestore=s=>fr(s),u.stackAlloc=s=>En(s),u.UTF8ToString=Ee,u.stringToUTF8=Pt,u.lengthBytesUTF8=Fo,ht=function s(){cr||Wi(),cr||(ht=s)},Wi(),h}),up=$a;globalThis.self?.name==="em-pthread"&&$a()});var Dt,dp,lp,cp,Sa,Ia,pp,Ca,Gt=M(()=>{"use strict";Sr();Dt= false?0:import.meta.url??(typeof document<"u"?document.currentScript?.src:typeof self<"u"?self.location?.href:void 0),dp= false||typeof location>"u"?void 0:location.origin,lp=(e,t)=>{try{let r=t??Dt;return(r?new URL(e,r):new URL(e)).origin===dp}catch{return!1}},cp=async e=>{let r=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(r)},Sa=(va(),hr(wa)).default,Ia=async()=>{if(!Dt)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(lp(Dt))return[void 0,Sa()];let e=await cp(Dt);return[e,Sa(e)]},pp=(xa(),hr(_a)).default,Ca=async(e,t,r)=>[void 0,pp]});var Nn,Wn,zr,Ta,mp,fp,Ir,_e,$t=M(()=>{"use strict";Gt();Wn=!1,zr=!1,Ta=!1,mp=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},fp=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},Ir=async e=>{if(Wn)return Promise.resolve();if(zr)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Ta)throw new Error("previous call to 'initializeWebAssembly()' failed.");zr=!0;let t=e.initTimeout,r=e.numThreads;if(!fp())throw new Error("WebAssembly SIMD is not supported in the current environment.");let n=mp();r>1&&!n&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+r+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=r=1);let o=e.wasmPaths,i=typeof o=="string"?o:void 0,a=o?.mjs,l=a?.href??a,d=o?.wasm,p=d?.href??d,m=e.wasmBinary,[u,h]=await Ca(l,i,r>1),w=!1,g=[];if(t>0&&g.push(new Promise(y=>{setTimeout(()=>{w=!0,y()},t)})),g.push(new Promise((y,x)=>{let _={numThreads:r};m?_.wasmBinary=m:(p||i)&&(_.locateFile=($,S)=>p??(i??S)+$),h(_).then($=>{zr=!1,Wn=!0,Nn=$,y(),u&&URL.revokeObjectURL(u)},$=>{zr=!1,Ta=!0,x($)})})),await Promise.race(g),w)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},_e=()=>{if(Wn&&Nn)return Nn;throw new Error("WebAssembly is not initialized yet.")}});var Ie,qt,we,Br=M(()=>{"use strict";$t();Ie=(e,t)=>{let r=_e(),n=r.lengthBytesUTF8(e)+1,o=r._malloc(n);return r.stringToUTF8(e,o,n),t.push(o),o},qt=(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")qt(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: ${typeof i}`)})},we=e=>{let t=_e(),r=t.stackSave();try{let n=t.stackAlloc(8);t._OrtGetLastError(n,n+4);let o=t.HEAP32[n/4],i=t.HEAPU32[n/4+1],a=i?t.UTF8ToString(i):"";throw new Error(`${e} ERROR_CODE: ${o}, ERROR_MESSAGE: ${a}`)}finally{t.stackRestore(r)}}});var Aa,ka=M(()=>{"use strict";$t();Br();Aa=e=>{let t=_e(),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=Ie(e.tag,n)),r=t._OrtCreateRunOptions(o.logSeverityLevel,o.logVerbosityLevel,!!o.terminate,i),r===0&&we("Can't create run options."),e?.extra!==void 0&&qt(e.extra,"",new WeakSet,(a,l)=>{let d=Ie(a,n),p=Ie(l,n);t._OrtAddRunConfigEntry(r,d,p)!==0&&we(`Can't set a run config entry: ${a} - ${l}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseRunOptions(r),n.forEach(a=>t._free(a)),i}}});var hp,gp,yp,bp,Ea,Pa=M(()=>{"use strict";$t();Br();hp=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}`)}},gp=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},yp=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)},bp=(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 l=n?.deviceType;if(l){let d=Ie("deviceType",r),p=Ie(l,r);_e()._OrtAddSessionConfigEntry(e,d,p)!==0&&we(`Can't set a session config entry: 'deviceType' - ${l}.`)}}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 l=Ie("preferredLayout",r),d=Ie(a.preferredLayout,r);_e()._OrtAddSessionConfigEntry(e,l,d)!==0&&we(`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=Ie(o,r);_e()._OrtAppendExecutionProvider(e,i)!==0&&we(`Can't append execution provider: ${o}.`)}},Ea=e=>{let t=_e(),r=0,n=[],o=e||{};yp(o);try{let i=hp(o.graphOptimizationLevel??"all"),a=gp(o.executionMode??"sequential"),l=typeof o.logId=="string"?Ie(o.logId,n):0,d=o.logSeverityLevel??2;if(!Number.isInteger(d)||d<0||d>4)throw new Error(`log serverity level is not valid: ${d}`);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"?Ie(o.optimizedModelFilePath,n):0;if(r=t._OrtCreateSessionOptions(i,!!o.enableCpuMemArena,!!o.enableMemPattern,a,!!o.enableProfiling,0,l,d,p,m),r===0&&we("Can't create session options."),o.executionProviders&&bp(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=Ie("enableGraphCapture",n),h=Ie(o.enableGraphCapture.toString(),n);t._OrtAddSessionConfigEntry(r,u,h)!==0&&we(`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 w=Ie(u,n);t._OrtAddFreeDimensionOverride(r,w,h)!==0&&we(`Can't set a free dimension override: ${u} - ${h}.`)}return o.extra!==void 0&&qt(o.extra,"",new WeakSet,(u,h)=>{let w=Ie(u,n),g=Ie(h,n);t._OrtAddSessionConfigEntry(r,w,g)!==0&&we(`Can't set a session config entry: ${u} - ${h}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseSessionOptions(r),n.forEach(a=>t._free(a)),i}}});var Ln,mt,_t,Dr,Kt,Rr,Hn,X=M(()=>{"use strict";Ln=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}`)}},mt=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}`)}},_t=(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}`)}},Kt=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}`)}},Rr=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool",Hn=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;default:throw new Error(`unsupported data location: ${e}`)}}});var jt,Gn=M(()=>{"use strict";Sr();jt=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: ${e}, no response body.`);let o=t.body.getReader(),i;try{i=new ArrayBuffer(n)}catch(l){if(l instanceof RangeError){let d=Math.ceil(n/65536);i=new WebAssembly.Memory({initial:d,maximum:d}).buffer}else throw l}let a=0;for(;;){let{done:l,value:d}=await o.read();if(l)break;let p=d.byteLength;new Uint8Array(i,a,p).set(d),a+=p}return new Uint8Array(i,0,n)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}});var wp,vp,Oa,za,Ba,$p,be,st=M(()=>{"use strict";X();wp=["V","I","W","E","F"],vp=(e,t)=>{console.log(`[${wp[e]},${new Date().toISOString()}]${t}`)},Ba=(e,t)=>{Oa=e,za=t},$p=(e,t)=>{let r=Kt(e),n=Kt(Oa);r>=n&&vp(r,typeof t=="function"?t():t)},be=(...e)=>{za&&$p(...e)}});var Da,Ra=M(()=>{"use strict";X();Da=(e,t)=>new(Dr(t))(e)});var Mr=M(()=>{"use strict"});var Ma,Fn,qn,_p,xp,Ua,jn,Kn,Na,Wa=M(()=>{"use strict";st();Mr();Ma=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),Fn=[],qn=e=>Math.ceil(e/16)*16,_p=e=>{for(let t=0;t<Fn.length;t++){let r=Fn[t];if(e<=r)return r}return Math.ceil(e/16)*16},xp=1,Ua=()=>xp++,jn=async(e,t,r,n)=>{let o=qn(r),i=e.device.createBuffer({size:o,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let a=e.getCommandEncoder();e.endComputePass(),a.copyBufferToBuffer(t,0,i,0,o),e.flush(),await i.mapAsync(GPUMapMode.READ);let l=i.getMappedRange();if(n){let d=n();return d.set(new Uint8Array(l,0,r)),d}else return new Uint8Array(l.slice(0,r))}finally{i.destroy()}},Kn=class{constructor(t){this.backend=t;this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map,this.capturedPendingBuffers=new Map;for(let[r]of Ma)Fn.push(r),this.freeBuffers.set(r,[]),this.freeUniformBuffers.set(r,[])}upload(t,r){let n=r.buffer,o=r.byteOffset,i=r.byteLength,a=qn(i),l=this.storageCache.get(t);if(!l)throw new Error("gpu data for uploading does not exist");if(l.originalSize!==i)throw new Error(`inconsistent data size. gpu data size=${l.originalSize}, data size=${i}`);let d=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),p=d.getMappedRange();new Uint8Array(p).set(new Uint8Array(n,o,i)),d.unmap();let m=this.backend.getCommandEncoder();this.backend.endComputePass(),m.copyBufferToBuffer(d,0,l.gpuData.buffer,0,a),be("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(d)}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=qn(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=this.externalBuffers.get(n),o===void 0)throw new Error("previous buffer is not registered");if(t===n)return be("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!`);this.externalBuffers.delete(n)}else o=Ua();return this.storageCache.set(o,{gpuData:{id:o,type:0,buffer:t},originalSize:r}),this.externalBuffers.set(t,o),be("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${o}, registered.`),o}unregisterExternalBuffer(t){let r=this.externalBuffers.get(t);r!==void 0&&(this.storageCache.delete(r),this.externalBuffers.delete(t),be("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${r}`))}create(t,r=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let n=_p(t),o,i=(r&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(r&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(i||a){let p=(i?this.freeBuffers:this.freeUniformBuffers).get(n);p?p.length>0?o=p.pop():o=this.backend.device.createBuffer({size:n,usage:r}):o=this.backend.device.createBuffer({size:n,usage:r})}else o=this.backend.device.createBuffer({size:n,usage:r});let l={id:Ua(),type:0,buffer:o};return this.storageCache.set(l.id,{gpuData:l,originalSize:t}),be("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${l.id}`),l}get(t){return this.storageCache.get(t)?.gpuData}release(t){let r=this.storageCache.get(t);if(!r)throw new Error("releasing data does not exist");return be("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(t,r){let n=this.storageCache.get(t);if(!n)throw new Error("data does not exist");await jn(this.backend,n.gpuData.buffer,n.originalSize,r)}refreshPendingBuffers(){for(let t of this.buffersForUploadingPending)t.destroy();if(this.buffersForUploadingPending=[],this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let r=Ma.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let n=this.freeBuffers.get(t.size)||[];r===void 0||n.length>=r?t.destroy():n.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let n=this.freeUniformBuffers.get(t.size)||[];r===void 0||n.length>=r?t.destroy():n.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let r of this.buffersPending)t.push(r);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onReleaseSession(t){let r=this.capturedPendingBuffers.get(t);r&&(r.forEach(n=>{n.destroy()}),this.capturedPendingBuffers.delete(t))}},Na=(...e)=>new Kn(...e)});var Yn,Q,xe=M(()=>{"use strict";Yn=class{constructor(t){Object.assign(this,t)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(t=>`${this[t]}`).join(";")),this.key}},Q=e=>new Yn(e)});var Xn,Je,E,xt,Ur,Vr,Nr,oe=M(()=>{"use strict";Xn=class{static calcMatMulShape(t,r){return t[1]!==r[0]?void 0:[t[0],r[1]]}},Je=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),l=new Array(a);if(n){if(o<2||i<2)return;let d=Xn.calcMatMulShape([t[o-2],t[o-1]],[r[i-2],r[i-1]]);if(d===void 0)return;[l[a-2],l[a-1]]=d}for(let d=n?3:1;d<=a;d++){let p=o-d<0?1:t[o-d],m=i-d<0?1:r[i-d];if(p!==m&&p>1&&m>1)return;let u=Math.max(p,m);if(p&&m)l[a-d]=Math.max(p,m);else{if(u>1)return;l[a-d]=0}}return l}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}},E=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*=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])}},xt=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 l=0;l<r.length-2;l++)l>=n.length?n.push(r[l+2]):n[l]=r[l+2];for(let l=0;l<n.length;l++)if(l<o.length){if(o[l]<0)throw new Error("strides should be greater than or equal to 1")}else o.push(1);for(let l=0;l<n.length;l++)if(l<i.length){if(i[l]<0)throw new Error("dilations should be greater than or equal to 1")}else i.push(1);for(let l=0;l<n.length*2;l++)if(l<a.length){if(a[l]<0)throw new Error("pad should be greater than or equal to 1")}else a.push(0);for(let l=0;l<n.length;l++){if(n[l]<=0)throw new Error("kernel shapes need to be greater than 0");if(a[l]>=n[l]||a[l+n.length]>=n[l])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,n,o,i,a,l){if(l){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 d=0;d<t.length-2;d++)e.adjustPadAndReturnShape(t[d+(a?1:2)],r[d],n[d],o[d],i,d,d+t.length-2,l)}}static computePoolOutputShape(t,r,n,o,i,a,l){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let d=[r[0],r[1]];return e.computeShapeHelper(t,r,d,n,o,i,a,l),d}static computeConvOutputShape(t,r,n,o,i,a,l){if(t.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let d=[t[0],r[0]];return e.computeShapeHelper(!1,t,d,n,o,i,a,l),d}static computeShapeHelper(t,r,n,o,i,a,l,d){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],l,p,p+r.length-2,d))}static adjustPadAndReturnShape(t,r,n,o,i,a,l,d){let p=n*(o-1)+1;if(d&&d!=="NOTSET")switch(d){case"VALID":return i[a]=0,i[l]=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(d==="SAME_LOWER"?(u+1)/2:u/2),i[l]=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[l]-p)/r+1)}},Ur=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,l,d;r?(a=t[1],l=t[0]):(a=t[0],l=t[1]);let p=-1;if(o?(d=n[0],p=1):(d=n[1],p=0),n[p]!==l)throw new Error("dimension mismatch");if(a<=0||d<=0||l<=0)throw new Error("invalid shape specified");if(i&&!Je.isValidBroadcast(i,[a,d]))throw new Error("gemm: invalid bias shape for broadcast");return[a,d,l]}},Vr=-34028234663852886e22,Nr=34028234663852886e22});var St,Qn,pe,ke,U,he,ft,It,je,L,Jn,k,D,Wr,Zn,La,Mt,re=M(()=>{"use strict";X();oe();St=64,Qn=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(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>"];default:throw new Error(`Unknown data type: ${e}`)}},pe=(e,t=1)=>{let r=Qn(e,t);return typeof r=="string"?r:r[0]},ke=(e,t=1)=>{let r=Qn(e,t);return typeof r=="string"?r:r[1]},U=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:E.computeStrides(r)})}),t},he=e=>e%4===0?4:e%2===0?2:1,ft=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,It=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,je=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,L=(e,t,r,n)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?n==="f16"?`${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}[${t}]`:e,Jn=(e,t,r,n,o)=>{let i=typeof r=="number",a=i?r:r.length,l=[...new Array(a).keys()],d=a<2?"u32":a<=4?`vec${a}<u32>`:`array<u32, ${a}>`,p=Qn(t,o),m=typeof p=="string"?p:p[1],u=typeof p=="string"?p:p[0],h={indices:d,value:m,storage:u,tensor:t},w=R=>typeof R=="string"?R:`${R}u`,g={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},y=i?"uniforms.":"",x=`${y}${e}_shape`,_=`${y}${e}_strides`,$="";for(let R=0;R<a-1;R++)$+=`
+    let dim${R} = current / ${L(_,R,a)};
+    let rest${R} = current % ${L(_,R,a)};
     indices[${R}] = dim${R};
     current = rest${R};
     `;$+=`indices[${a-1}] = current;`;let S=a<2?"":`
@@ -3820,25 +3819,25 @@ var Rn=Object.defineProperty;var op=Object.getOwnPropertyDescriptor;var ip=Objec
     var current = offset;
     ${$}
     return indices;
-  }`,I=R=>(g.offsetToIndices=!0,a<2?R:`o2i_${e}(${R})`),T=[];if(a>=2)for(let R=a-1;R>=0;R--)T.push(`${G(_,R,a)} * (indices[${R}])`);let A=a<2?"":`
+  }`,I=R=>(g.offsetToIndices=!0,a<2?R:`o2i_${e}(${R})`),T=[];if(a>=2)for(let R=a-1;R>=0;R--)T.push(`${L(_,R,a)} * (indices[${R}])`);let A=a<2?"":`
   fn i2o_${e}(indices: ${h.indices}) -> u32 {
     return ${T.join("+")};
-  }`,D=R=>(g.indicesToOffset=!0,a<2?R:`i2o_${e}(${R})`),z=(...R)=>a===0?"0u":`${h.indices}(${R.map(w).join(",")})`,H=(R,F)=>a<2?`${R}`:`${G(R,F,a)}`,W=(R,F,ce)=>a<2?`${R}=${ce};`:`${G(R,F,a)}=${ce};`,X={},Y=(R,F)=>{g.broadcastedIndicesToOffset=!0;let ce=`${F.name}broadcastedIndicesTo${e}Offset`;if(ce in X)return`${ce}(${R})`;let Re=[];for(let Se=a-1;Se>=0;Se--){let we=F.indicesGet("outputIndices",Se+F.rank-a);Re.push(`${H(_,Se)} * (${we} % ${H(x,Se)})`)}return X[ce]=`fn ${ce}(outputIndices: ${F.type.indices}) -> u32 {
-             return ${Re.length>0?Re.join("+"):"0u"};
-           }`,`${ce}(${R})`},pe=(R,F)=>(()=>{if(h.storage===h.value)return`${e}[${R}]=${F};`;if(h.storage==="vec2<u32>"&&h.value==="i32")return`${e}[${R}]=vec2<u32>(u32(${F}), select(0u, 0xFFFFFFFFu, ${F} < 0));`;if(h.storage==="vec2<u32>"&&h.value==="u32")return`${e}[${R}]=vec2<u32>(u32(${F}), 0u);`;if(h.storage==="u32"&&h.value==="vec4<bool>")return`${e}[${R}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${F}));`;throw new Error(`not supported combination of storage type ${h.storage} and value type ${h.value} yet`)})(),ue=R=>(()=>{if(h.storage===h.value)return`${e}[${R}]`;if(h.storage==="vec2<u32>"&&h.value==="i32")return`i32(${e}[${R}].x)`;if(h.storage==="vec2<u32>"&&h.value==="u32")return`u32(${e}[${R}].x)`;if(h.storage==="u32"&&h.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${R}] & 0xFFu), bool(${e}[${R}] & 0xFF00u), bool(${e}[${R}] & 0xFF0000u), bool(${e}[${R}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${h.storage} and value type ${h.value} yet`)})(),le=a<2?"":`
+  }`,O=R=>(g.indicesToOffset=!0,a<2?R:`i2o_${e}(${R})`),z=(...R)=>a===0?"0u":`${h.indices}(${R.map(w).join(",")})`,W=(R,G)=>a<2?`${R}`:`${L(R,G,a)}`,V=(R,G,ue)=>a<2?`${R}=${ue};`:`${L(R,G,a)}=${ue};`,j={},Y=(R,G)=>{g.broadcastedIndicesToOffset=!0;let ue=`${G.name}broadcastedIndicesTo${e}Offset`;if(ue in j)return`${ue}(${R})`;let Be=[];for(let ge=a-1;ge>=0;ge--){let Se=G.indicesGet("outputIndices",ge+G.rank-a);Be.push(`${W(_,ge)} * (${Se} % ${W(x,ge)})`)}return j[ue]=`fn ${ue}(outputIndices: ${G.type.indices}) -> u32 {
+             return ${Be.length>0?Be.join("+"):"0u"};
+           }`,`${ue}(${R})`},ae=(R,G)=>(()=>{if(h.storage===h.value)return`${e}[${R}]=${G};`;if(h.storage==="vec2<u32>"&&h.value==="i32")return`${e}[${R}]=vec2<u32>(u32(${G}), select(0u, 0xFFFFFFFFu, ${G} < 0));`;if(h.storage==="vec2<u32>"&&h.value==="u32")return`${e}[${R}]=vec2<u32>(u32(${G}), 0u);`;if(h.storage==="u32"&&h.value==="vec4<bool>")return`${e}[${R}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${G}));`;throw new Error(`not supported combination of storage type ${h.storage} and value type ${h.value} yet`)})(),se=R=>(()=>{if(h.storage===h.value)return`${e}[${R}]`;if(h.storage==="vec2<u32>"&&h.value==="i32")return`i32(${e}[${R}].x)`;if(h.storage==="vec2<u32>"&&h.value==="u32")return`u32(${e}[${R}].x)`;if(h.storage==="u32"&&h.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${R}] & 0xFFu), bool(${e}[${R}] & 0xFF00u), bool(${e}[${R}] & 0xFF0000u), bool(${e}[${R}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${h.storage} and value type ${h.value} yet`)})(),ie=a<2?"":`
   fn get_${e}ByIndices(indices: ${h.indices}) -> ${m} {
-    return ${ue(`i2o_${e}(indices)`)};
-  }`,K=a<2?"":(()=>{let R=l.map(ce=>`d${ce}: u32`).join(", "),F=l.map(ce=>`d${ce}`).join(", ");return`
+    return ${se(`i2o_${e}(indices)`)};
+  }`,q=a<2?"":(()=>{let R=l.map(ue=>`d${ue}: u32`).join(", "),G=l.map(ue=>`d${ue}`).join(", ");return`
   fn get_${e}(${R}) -> ${m} {
-    return get_${e}ByIndices(${z(F)});
-  }`})(),de=(...R)=>{if(R.length!==a)throw new Error(`indices length must be ${a}`);let F=R.map(w).join(",");return a===0?ue("0u"):a===1?ue(F[0]):(g.get=!0,g.getByIndices=!0,g.indicesToOffset=!0,`get_${e}(${F})`)},re=R=>a<2?ue(R):(g.getByIndices=!0,g.indicesToOffset=!0,`get_${e}ByIndices(${R})`),se=a<2?"":`
+    return get_${e}ByIndices(${z(G)});
+  }`})(),ce=(...R)=>{if(R.length!==a)throw new Error(`indices length must be ${a}`);let G=R.map(w).join(",");return a===0?se("0u"):a===1?se(G[0]):(g.get=!0,g.getByIndices=!0,g.indicesToOffset=!0,`get_${e}(${G})`)},ne=R=>a<2?se(R):(g.getByIndices=!0,g.indicesToOffset=!0,`get_${e}ByIndices(${R})`),Z=a<2?"":`
   fn set_${e}ByIndices(indices: ${h.indices}, value: ${m}) {
-    ${pe(`i2o_${e}(indices)`,"value")}
-  }`,Z=a<2?"":(()=>{let R=l.map(ce=>`d${ce}: u32`).join(", "),F=l.map(ce=>`d${ce}`).join(", ");return`
+    ${ae(`i2o_${e}(indices)`,"value")}
+  }`,ee=a<2?"":(()=>{let R=l.map(ue=>`d${ue}: u32`).join(", "),G=l.map(ue=>`d${ue}`).join(", ");return`
   fn set_${e}(${R}, value: ${m}) {
-    set_${e}ByIndices(${z(F)}, value);
-  }`})();return{impl:()=>{let R=[],F=!1;return g.offsetToIndices&&(R.push(S),F=!0),g.indicesToOffset&&(R.push(A),F=!0),g.broadcastedIndicesToOffset&&(Object.values(X).forEach(ce=>R.push(ce)),F=!0),g.set&&(R.push(Z),F=!0),g.setByIndices&&(R.push(se),F=!0),g.get&&(R.push(K),F=!0),g.getByIndices&&(R.push(le),F=!0),!i&&F&&R.unshift(`const ${x} = ${h.indices}(${r.join(",")});`,`const ${_} = ${h.indices}(${k.computeStrides(r).join(",")});`),R.join(`
-`)},type:h,offsetToIndices:I,indicesToOffset:D,broadcastedIndicesToOffset:Y,indices:z,indicesGet:H,indicesSet:W,set:(...R)=>{if(R.length!==a+1)throw new Error(`indices length must be ${a}`);let F=R[a];if(typeof F!="string")throw new Error("value must be string");let ce=R.slice(0,a).map(w).join(",");return a===0?pe("0u",F):a===1?pe(ce[0],F):(g.set=!0,g.setByIndices=!0,g.indicesToOffset=!0,`set_${e}(${ce}, ${F})`)},setByOffset:pe,setByIndices:(R,F)=>a<2?pe(R,F):(g.setByIndices=!0,g.indicesToOffset=!0,`set_${e}ByIndices(${R}, ${F});`),get:de,getByOffset:ue,getByIndices:re,usage:n,name:e,strides:_,shape:x,rank:a}},E=(e,t,r,n=1)=>no(e,t,r,"input",n),M=(e,t,r,n=1)=>no(e,t,r,"output",n),Lr=(e,t,r,n=1)=>no(e,t,r,"internal",n),to=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=Tt){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(${z(G)}, value);
+  }`})();return{impl:()=>{let R=[],G=!1;return g.offsetToIndices&&(R.push(S),G=!0),g.indicesToOffset&&(R.push(A),G=!0),g.broadcastedIndicesToOffset&&(Object.values(j).forEach(ue=>R.push(ue)),G=!0),g.set&&(R.push(ee),G=!0),g.setByIndices&&(R.push(Z),G=!0),g.get&&(R.push(q),G=!0),g.getByIndices&&(R.push(ie),G=!0),!i&&G&&R.unshift(`const ${x} = ${h.indices}(${r.join(",")});`,`const ${_} = ${h.indices}(${E.computeStrides(r).join(",")});`),R.join(`
+`)},type:h,offsetToIndices:I,indicesToOffset:O,broadcastedIndicesToOffset:Y,indices:z,indicesGet:W,indicesSet:V,set:(...R)=>{if(R.length!==a+1)throw new Error(`indices length must be ${a}`);let G=R[a];if(typeof G!="string")throw new Error("value must be string");let ue=R.slice(0,a).map(w).join(",");return a===0?ae("0u",G):a===1?ae(ue[0],G):(g.set=!0,g.setByIndices=!0,g.indicesToOffset=!0,`set_${e}(${ue}, ${G})`)},setByOffset:ae,setByIndices:(R,G)=>a<2?ae(R,G):(g.setByIndices=!0,g.indicesToOffset=!0,`set_${e}ByIndices(${R}, ${G});`),get:ce,getByOffset:se,getByIndices:ne,usage:n,name:e,strides:_,shape:x,rank:a}},k=(e,t,r,n=1)=>Jn(e,t,r,"input",n),D=(e,t,r,n=1)=>Jn(e,t,r,"output",n),Wr=(e,t,r,n=1)=>Jn(e,t,r,"internal",n),Zn=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=St){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_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
                                              @builtin(local_invocation_id) local_id : vec3<u32>,
@@ -3853,9 +3852,9 @@ var Rn=Object.defineProperty;var op=Object.getOwnPropertyDescriptor;var ip=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])}},Ya=(e,t)=>new to(e,t),Nt=(e,t)=>{let r=e.length,n=[];for(let o=0;o<r;o++){let i=r-1-o,a=e[i]||1;(t[t.length-1-o]||1)>1&&a===1&&n.unshift(i)}return n}});var Ep,Xa,kp,Pp,Ue,Za,Qa,Et=U(()=>{"use strict";Q();ae();Ce();ie();Ep=e=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.")},Xa=(e,t)=>t&&t.length!==e?[...new Array(e).keys()].reverse():t,kp=(e,t)=>k.sortBasedOnPerm(e,Xa(e.length,t)),Pp=(e,t,r,n)=>{let o=[];o.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
+`)}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])}},La=(e,t)=>new Zn(e,t),Mt=(e,t)=>{let r=e.length,n=[];for(let o=0;o<r;o++){let i=r-1-o,a=e[i]||1;(t[t.length-1-o]||1)>1&&a===1&&n.unshift(i)}return n}});var Sp,Ha,Ip,Cp,Re,Ga,Fa,Ct=M(()=>{"use strict";X();oe();xe();re();Sp=e=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.")},Ha=(e,t)=>t&&t.length!==e?[...new Array(e).keys()].reverse():t,Ip=(e,t)=>E.sortBasedOnPerm(e,Ha(e.length,t)),Cp=(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(`
-`)},Ue=(e,t)=>{let r=e.dataType,n=e.dims.length,o=Xa(n,t),i=kp(e.dims,o),a=M("output",r,i.length),l=E("a",r,n),d;if(o.length===2&&o[0]===1&&o[1]===0){let p=a.type.value,m=[16,16,1];d=u=>`
+`)},Re=(e,t)=>{let r=e.dataType,n=e.dims.length,o=Ha(n,t),i=Ip(e.dims,o),a=D("output",r,i.length),l=k("a",r,n),d;if(o.length===2&&o[0]===1&&o[1]===0){let p=a.type.value,m=[16,16,1];d=u=>`
   ${u.registerUniform("output_size","u32").declareVariables(l,a)}
   var<workgroup> tile : array<array<${p}, ${m[0]+1}>, ${m[0]}>;
   ${u.mainStart(m)}
@@ -3875,7 +3874,7 @@ var Rn=Object.defineProperty;var op=Object.getOwnPropertyDescriptor;var ip=Objec
   }`}else d=p=>`
   ${p.registerUniform("output_size","u32").declareVariables(l,a)}
 
-  ${Pp(o,n,l,a)}
+  ${Cp(o,n,l,a)}
 
   ${p.mainStart()}
     ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -3884,24 +3883,24 @@ var Rn=Object.defineProperty;var op=Object.getOwnPropertyDescriptor;var ip=Objec
     let aIndices = perm(indices);
 
     ${a.setByOffset("global_idx",l.getByIndices("aIndices"))}
-  }`;return{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:p=>{let m=k.size(i);return{outputs:[{dims:i,dataType:p[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},...V(p[0].dims,i)]}},getShaderSource:d}},Za=(e,t)=>{Ep(e.inputs),e.compute(Ue(e.inputs[0],t.perm))},Qa=e=>te({perm:e.perm})});var Op,Dp,zp,Bp,Rp,Mp,Up,Vp,Np,Wp,rt,Ja,es,ts,rs,ns,os,is,as,ss,us,ds=U(()=>{"use strict";Q();ae();ie();Fr();Et();Op={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"},Dp={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"},zp={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Bp={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Rp=(e,t)=>{let r=[];for(let n=t-e;n<t;++n)r.push(n);return r},Mp=(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]},Up=(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},Vp=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},Np=(e,t)=>{let r=[];if(!Vp(e,t)){for(let n=0;n<t;++n)e.indexOf(n)===-1&&r.push(n);e.forEach(n=>r.push(n))}return r},Wp=(e,t,r,n,o,i,a)=>{let l=r[0].dims,d=k.size(i),p=k.size(a),m=E("_A",r[0].dataType,l),u=M("output",o,i),h=32,w=`
+  }`;return{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:p=>{let m=E.size(i);return{outputs:[{dims:i,dataType:p[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},...U(p[0].dims,i)]}},getShaderSource:d}},Ga=(e,t)=>{Sp(e.inputs),e.compute(Re(e.inputs[0],t.perm))},Fa=e=>Q({perm:e.perm})});var Tp,Ap,kp,Ep,Pp,Op,zp,Bp,Dp,Rp,et,qa,Ka,ja,Ya,Xa,Za,Qa,Ja,es,ts,rs=M(()=>{"use strict";X();oe();re();Lr();Ct();Tp={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"},Ap={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"},kp={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Ep={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Pp=(e,t)=>{let r=[];for(let n=t-e;n<t;++n)r.push(n);return r},Op=(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]},zp=(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},Bp=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},Dp=(e,t)=>{let r=[];if(!Bp(e,t)){for(let n=0;n<t;++n)e.indexOf(n)===-1&&r.push(n);e.forEach(n=>r.push(n))}return r},Rp=(e,t,r,n,o,i,a)=>{let l=r[0].dims,d=E.size(i),p=E.size(a),m=k("_A",r[0].dataType,l),u=D("output",o,i),h=32,w=`
           var<workgroup> aBestValues : array<f32, ${h}>;
-       `;return{name:e,shaderCache:t,getShaderSource:b=>`
-        ${b.registerUniform("reduceSize","u32").declareVariables(m,u)}
+       `;return{name:e,shaderCache:t,getShaderSource:y=>`
+        ${y.registerUniform("reduceSize","u32").declareVariables(m,u)}
         ${w}
         fn DIV_CEIL(a : u32, b : u32) -> u32 {
           return ((a - 1u) / b + 1u);
          }
-         ${b.mainStart(h)}
+         ${y.mainStart(h)}
 
           let outputIndex = global_idx / ${h};
           let offset = outputIndex * uniforms.reduceSize;
 
-          var bestValue = f32(${zp[n]});
+          var bestValue = f32(${kp[n]});
           let Length = uniforms.reduceSize;
           for (var k = local_idx; k < Length; k = k + ${h}) {
            let candidate = f32(${m.getByOffset("offset + k")});
-           bestValue = ${Op[n]};
+           bestValue = ${Tp[n]};
           }
           aBestValues[local_idx] = bestValue;
           workgroupBarrier();
@@ -3912,7 +3911,7 @@ var Rn=Object.defineProperty;var op=Object.getOwnPropertyDescriptor;var ip=Objec
            let interval = DIV_CEIL(reduceSize, 2u);
            if (local_idx < currentSize) {
             let candidate = aBestValues[local_idx + interval];
-            bestValue = ${Dp[n]};
+            bestValue = ${Ap[n]};
             aBestValues[local_idx] = bestValue;
            }
            reduceSize = interval;
@@ -3920,13 +3919,13 @@ var Rn=Object.defineProperty;var op=Object.getOwnPropertyDescriptor;var ip=Objec
          }
 
          if (local_idx == 0u) {
-          ${u.setByOffset("outputIndex",`${n==="mean"?`${u.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${u.type.storage}(${Bp[n]})`}`)};
+          ${u.setByOffset("outputIndex",`${n==="mean"?`${u.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${u.type.storage}(${Ep[n]})`}`)};
          }
-        }`,getRunData:()=>({outputs:[{dims:i,dataType:o}],dispatchGroup:{x:d},programUniforms:[{type:12,data:p}]})}},rt=(e,t,r,n)=>{let o=e.inputs.length===1?r:oo(e.inputs,r),i=o.axes;i.length===0&&!o.noopWithEmptyAxes&&(i=e.inputs[0].dims.map((w,g)=>g));let a=k.normalizeAxes(i,e.inputs[0].dims.length),l=a,d=e.inputs[0],p=Np(l,e.inputs[0].dims.length);p.length>0&&(d=e.compute(Ue(e.inputs[0],p),{inputs:[0],outputs:[-1]})[0],l=Rp(l.length,d.dims.length));let[m,u]=Mp(d.dims,l),h=m;o.keepDims&&(h=Up(m,a)),e.compute(Wp(t,{hint:o.cacheKey,inputDependencies:["type"]},[d],n,e.inputs[0].dataType,h,u),{inputs:[d]})},Ja=(e,t)=>{rt(e,"ReduceMeanShared",t,"mean")},es=(e,t)=>{rt(e,"ReduceL1Shared",t,"l1")},ts=(e,t)=>{rt(e,"ReduceL2Shared",t,"l2")},rs=(e,t)=>{rt(e,"ReduceLogSumExpShared",t,"logSumExp")},ns=(e,t)=>{rt(e,"ReduceMaxShared",t,"max")},os=(e,t)=>{rt(e,"ReduceMinShared",t,"min")},is=(e,t)=>{rt(e,"ReduceProdShared",t,"prod")},as=(e,t)=>{rt(e,"ReduceSumShared",t,"sum")},ss=(e,t)=>{rt(e,"ReduceSumSquareShared",t,"sumSquare")},us=(e,t)=>{rt(e,"ReduceLogSumShared",t,"logSum")}});var nt,Hp,qr,oo,ot,Gp,Lp,Fp,qp,jp,Kp,Yp,Xp,Zp,Qp,it,ls,cs,ps,ms,fs,hs,gs,ys,bs,ws,Fr=U(()=>{"use strict";Q();ae();Ce();ie();ds();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.")},Hp=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],qr=(e,t,r,n,o,i,a=!1,l=!1)=>{let d=[],p=r[0].dims,m=p.length,u=k.normalizeAxes(o,m),h=!l&&u.length===0;p.forEach((x,_)=>{h||u.indexOf(_)>=0?a&&d.push(1):d.push(x)});let w=d.length,g=k.size(d);return{name:e,shaderCache:t,getShaderSource:x=>{let _=[],$=E("_A",r[0].dataType,m),S=M("output",i,w),I=n($,S,u),T=I[2];for(let A=0,D=0;A<m;A++)h||u.indexOf(A)>=0?(a&&D++,T=`for(var j${A}: u32 = 0; j${A} < ${p[A]}; j${A}++) {
+        }`,getRunData:()=>({outputs:[{dims:i,dataType:o}],dispatchGroup:{x:d},programUniforms:[{type:12,data:p}]})}},et=(e,t,r,n)=>{let o=e.inputs.length===1?r:eo(e.inputs,r),i=o.axes;i.length===0&&!o.noopWithEmptyAxes&&(i=e.inputs[0].dims.map((w,g)=>g));let a=E.normalizeAxes(i,e.inputs[0].dims.length),l=a,d=e.inputs[0],p=Dp(l,e.inputs[0].dims.length);p.length>0&&(d=e.compute(Re(e.inputs[0],p),{inputs:[0],outputs:[-1]})[0],l=Pp(l.length,d.dims.length));let[m,u]=Op(d.dims,l),h=m;o.keepDims&&(h=zp(m,a)),e.compute(Rp(t,{hint:o.cacheKey,inputDependencies:["type"]},[d],n,e.inputs[0].dataType,h,u),{inputs:[d]})},qa=(e,t)=>{et(e,"ReduceMeanShared",t,"mean")},Ka=(e,t)=>{et(e,"ReduceL1Shared",t,"l1")},ja=(e,t)=>{et(e,"ReduceL2Shared",t,"l2")},Ya=(e,t)=>{et(e,"ReduceLogSumExpShared",t,"logSumExp")},Xa=(e,t)=>{et(e,"ReduceMaxShared",t,"max")},Za=(e,t)=>{et(e,"ReduceMinShared",t,"min")},Qa=(e,t)=>{et(e,"ReduceProdShared",t,"prod")},Ja=(e,t)=>{et(e,"ReduceSumShared",t,"sum")},es=(e,t)=>{et(e,"ReduceSumSquareShared",t,"sumSquare")},ts=(e,t)=>{et(e,"ReduceLogSumShared",t,"logSum")}});var tt,Mp,Hr,eo,rt,Up,Vp,Np,Wp,Lp,Hp,Gp,Fp,qp,Kp,nt,ns,os,is,as,ss,us,ds,ls,cs,ps,Lr=M(()=>{"use strict";X();oe();xe();re();rs();tt=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.")},Mp=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Hr=(e,t,r,n,o,i,a=!1,l=!1)=>{let d=[],p=r[0].dims,m=p.length,u=E.normalizeAxes(o,m),h=!l&&u.length===0;p.forEach((x,_)=>{h||u.indexOf(_)>=0?a&&d.push(1):d.push(x)});let w=d.length,g=E.size(d);return{name:e,shaderCache:t,getShaderSource:x=>{let _=[],$=k("_A",r[0].dataType,m),S=D("output",i,w),I=n($,S,u),T=I[2];for(let A=0,O=0;A<m;A++)h||u.indexOf(A)>=0?(a&&O++,T=`for(var j${A}: u32 = 0; j${A} < ${p[A]}; j${A}++) {
                   ${I[2].includes("last_index")?`let last_index = j${A};`:""}
                   ${$.indicesSet("input_indices",A,`j${A}`)}
                   ${T}
-                }`):(_.push(`${$.indicesSet("input_indices",A,S.indicesGet("output_indices",D))};`),D++);return`
+                }`):(_.push(`${$.indicesSet("input_indices",A,S.indicesGet("output_indices",O))};`),O++);return`
 
         ${x.registerUniform("output_size","u32").declareVariables($,S)}
 
@@ -3943,25 +3942,25 @@ var Rn=Object.defineProperty;var op=Object.getOwnPropertyDescriptor;var ip=Objec
           ${I[3]}
           ${I.length===4?S.setByOffset("global_idx","value"):I.slice(4).join(`
 `)}
-        }`},getRunData:()=>({outputs:[{dims:d,dataType:i}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:[{type:12,data:g},...V(p,d)]})}},oo=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),te({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},ot=(e,t,r,n)=>{let o=e.inputs,i=o.length===1?r:oo(o,r);e.compute(qr(t,{hint:i.cacheKey,inputDependencies:["rank"]},[o[0]],i.noopWithEmptyAxes&&i.axes.length===0?Hp:n,i.axes,o[0].dataType,i.keepDims,i.noopWithEmptyAxes),{inputs:[0]})},Gp=(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);"])},Lp=(e,t)=>{nt(e.inputs),ot(e,"ReduceL1",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += abs(${n.getByIndices("input_indices")});`,""])},Fp=(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);"])},qp=(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);"])},jp=(e,t)=>{nt(e.inputs),ot(e,"ReduceMax",t,(n,o,i)=>{let a=[];for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&a.push(n.indicesSet("input_indices",l,0));return[`${a.join(`
-`)}`,`var value = ${n.getByIndices("input_indices")};`,`value = max(value, ${n.getByIndices("input_indices")});`,""]})},Kp=(e,t)=>{nt(e.inputs),ot(e,"ReduceMean",t,(n,o,i)=>{let a=1;for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&(a*=e.inputs[0].dims[l]);return["var sum = f32(0);","",`sum += f32(${n.getByIndices("input_indices")});`,`let value = ${o.type.value}(sum / ${a});`]})},Yp=(e,t)=>{nt(e.inputs),ot(e,"ReduceMin",t,(n,o,i)=>{let a=[];for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&a.push(`input_indices[${l}] = 0;`);return[`${a.join(`
-`)}`,`var value = ${n.getByIndices("input_indices")};`,`value = min(value, ${n.getByIndices("input_indices")});`,""]})},Xp=(e,t)=>{nt(e.inputs),ot(e,"ReduceProd",t,(n,o)=>[`var value = ${o.type.storage}(1);`,"",`value *= ${n.getByIndices("input_indices")};`,""])},Zp=(e,t)=>{nt(e.inputs),ot(e,"ReduceSum",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,""])},Qp=(e,t)=>{nt(e.inputs),ot(e,"ReduceSumSquare",t,(n,o)=>[`var t = ${o.type.value}(0); var value = ${o.type.value}(0);`,"",`t = ${n.getByIndices("input_indices")}; value += t * t;`,""])},it=(e,t,r)=>{if(t.length===0)return r;let n=1,o=1;for(let i=0;i<t.length;i++)t.indexOf(i)===-1?n*=e[i]:o*=e[i];return o<32&&n>1024},ls=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Kp(e,t):Ja(e,t)},cs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Lp(e,t):es(e,t)},ps=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Fp(e,t):ts(e,t)},ms=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?qp(e,t):rs(e,t)},fs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?jp(e,t):ns(e,t)},hs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Yp(e,t):os(e,t)},gs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Xp(e,t):is(e,t)},ys=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Zp(e,t):as(e,t)},bs=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Qp(e,t):ss(e,t)},ws=(e,t)=>{it(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Gp(e,t):us(e,t)}});var vs,$s,_s,io,xs=U(()=>{"use strict";Q();Ce();Fr();vs=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.")},$s=(e,t)=>{vs(e.inputs);let r=(n,o,i)=>{let a=[];for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&a.push(`input_indices[${l}] = 0;`);return[`${a.join(`
+        }`},getRunData:()=>({outputs:[{dims:d,dataType:i}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:[{type:12,data:g},...U(p,d)]})}},eo=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),Q({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},rt=(e,t,r,n)=>{let o=e.inputs,i=o.length===1?r:eo(o,r);e.compute(Hr(t,{hint:i.cacheKey,inputDependencies:["rank"]},[o[0]],i.noopWithEmptyAxes&&i.axes.length===0?Mp:n,i.axes,o[0].dataType,i.keepDims,i.noopWithEmptyAxes),{inputs:[0]})},Up=(e,t)=>{tt(e.inputs),rt(e,"ReduceLogSum",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,"value = log(value);"])},Vp=(e,t)=>{tt(e.inputs),rt(e,"ReduceL1",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += abs(${n.getByIndices("input_indices")});`,""])},Np=(e,t)=>{tt(e.inputs),rt(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);"])},Wp=(e,t)=>{tt(e.inputs),rt(e,"ReduceLogSumExp",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += exp(${n.getByIndices("input_indices")});`,"value = log(value);"])},Lp=(e,t)=>{tt(e.inputs),rt(e,"ReduceMax",t,(n,o,i)=>{let a=[];for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&a.push(n.indicesSet("input_indices",l,0));return[`${a.join(`
+`)}`,`var value = ${n.getByIndices("input_indices")};`,`value = max(value, ${n.getByIndices("input_indices")});`,""]})},Hp=(e,t)=>{tt(e.inputs),rt(e,"ReduceMean",t,(n,o,i)=>{let a=1;for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&(a*=e.inputs[0].dims[l]);return["var sum = f32(0);","",`sum += f32(${n.getByIndices("input_indices")});`,`let value = ${o.type.value}(sum / ${a});`]})},Gp=(e,t)=>{tt(e.inputs),rt(e,"ReduceMin",t,(n,o,i)=>{let a=[];for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&a.push(`input_indices[${l}] = 0;`);return[`${a.join(`
+`)}`,`var value = ${n.getByIndices("input_indices")};`,`value = min(value, ${n.getByIndices("input_indices")});`,""]})},Fp=(e,t)=>{tt(e.inputs),rt(e,"ReduceProd",t,(n,o)=>[`var value = ${o.type.storage}(1);`,"",`value *= ${n.getByIndices("input_indices")};`,""])},qp=(e,t)=>{tt(e.inputs),rt(e,"ReduceSum",t,(n,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,""])},Kp=(e,t)=>{tt(e.inputs),rt(e,"ReduceSumSquare",t,(n,o)=>[`var t = ${o.type.value}(0); var value = ${o.type.value}(0);`,"",`t = ${n.getByIndices("input_indices")}; value += t * t;`,""])},nt=(e,t,r)=>{if(t.length===0)return r;let n=1,o=1;for(let i=0;i<t.length;i++)t.indexOf(i)===-1?n*=e[i]:o*=e[i];return o<32&&n>1024},ns=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Hp(e,t):qa(e,t)},os=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Vp(e,t):Ka(e,t)},is=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Np(e,t):ja(e,t)},as=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Wp(e,t):Ya(e,t)},ss=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Lp(e,t):Xa(e,t)},us=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Gp(e,t):Za(e,t)},ds=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Fp(e,t):Qa(e,t)},ls=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?qp(e,t):Ja(e,t)},cs=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Kp(e,t):es(e,t)},ps=(e,t)=>{nt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Up(e,t):ts(e,t)}});var ms,fs,hs,to,gs=M(()=>{"use strict";X();xe();Lr();ms=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},fs=(e,t)=>{ms(e.inputs);let r=(n,o,i)=>{let a=[];for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&a.push(`input_indices[${l}] = 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 = ${n.getByIndices("input_indices")};
          best_index = i32(last_index);
-       }`,"",o.setByOffset("global_idx","best_index")]};e.compute(qr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},_s=(e,t)=>{vs(e.inputs);let r=(n,o,i)=>{let a=[];for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&a.push(`input_indices[${l}] = 0;`);return[`${a.join(`
+       }`,"",o.setByOffset("global_idx","best_index")]};e.compute(Hr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},hs=(e,t)=>{ms(e.inputs);let r=(n,o,i)=>{let a=[];for(let l=0;l<n.rank;l++)(i.indexOf(l)>=0||i.length===0)&&a.push(`input_indices[${l}] = 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(qr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},io=e=>te(e)});var Jp,em,tm,rm,Wt,nm,Ss,jr=U(()=>{"use strict";Q();Nr();ie();Jp=(e,t)=>{let r=e[0],n=e[1],o=e[2],i=e[3],a=e[4],l=e[5];if(a&&l)throw new Error("Attention cannot have both past and relative_position_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let d=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,w=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],w=t.qkvHiddenSizes[2]}let g=p;if(u!==h)throw new Error("qkv_hidden_sizes first element should be same as the second");if(o.dims[0]!==u+h+w)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let b=0;if(a){if(h!==w)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]!==d)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||(b=a.dims[3])}let x=g+b,_=-1,$=0;if(i)throw new Error("Mask not supported");if(a)throw new Error("past is not supported");return{batchSize:d,sequenceLength:p,pastSequenceLength:b,kvSequenceLength:g,totalSequenceLength:x,maxSequenceLength:_,inputHiddenSize:m,hiddenSize:u,vHiddenSize:w,headSize:Math.floor(u/t.numHeads),vHeadSize:Math.floor(w/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:$,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},em=(e,t,r,n)=>{let o=$e(n),i=64,a=n/o;a<i?i=1:a/8<64&&(i=Math.ceil(a/8));let l=Math.ceil(n/o/i),d=[{type:t.dataType,data:1/n},{type:12,data:a},{type:12,data:l}],p=ge(t.dataType,o),m=Oe(1,o),u=h=>{let w=M("x",t.dataType,t.dims,o),b=[{name:"d_inv",type:Oe(t.dataType)},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
+       }`,"",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]})},to=e=>Q(e)});var jp,Yp,Xp,Zp,Ut,Qp,ys,Gr=M(()=>{"use strict";X();Mr();re();jp=(e,t)=>{let r=e[0],n=e[1],o=e[2],i=e[3],a=e[4],l=e[5];if(a&&l)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 d=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,w=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],w=t.qkvHiddenSizes[2]}let g=p;if(u!==h)throw new Error("qkv_hidden_sizes first element should be same as the second");if(o.dims[0]!==u+h+w)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let y=0;if(a){if(h!==w)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]!==d)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||(y=a.dims[3])}let x=g+y,_=-1,$=0;if(i)throw new Error("Mask not supported");if(a)throw new Error("past is not supported");if(l){if(l.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(l.dims[0]!==d||l.dims[1]!==t.numHeads||l.dims[2]!==p||l.dims[3]!==x)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:d,sequenceLength:p,pastSequenceLength:y,kvSequenceLength:g,totalSequenceLength:x,maxSequenceLength:_,inputHiddenSize:m,hiddenSize:u,vHiddenSize:w,headSize:Math.floor(u/t.numHeads),vHeadSize:Math.floor(w/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:$,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Yp=(e,t,r,n)=>{let o=he(n),i=64,a=n/o;a<i&&(i=32);let l=Math.ceil(n/o/i),d=[{type:1,data:1/n},{type:12,data:a},{type:12,data:l}],p=pe(t.dataType,o),m=ke(1,o),u=["type"],h=w=>{let g=D("x",t.dataType,t.dims,o),y=ke(t.dataType),x=[{name:"d_inv",type:"f32"},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
   var<workgroup> thread_max: array<f32, ${i}>;
   var<workgroup> thread_sum: array<f32, ${i}>;
-  ${h.registerUniforms(b).declareVariables(w)}
-  ${h.mainStart([i,1,1])}
+  ${w.registerUniforms(x).declareVariables(g)}
+  ${w.mainStart([i,1,1])}
     let local_offset = local_idx * uniforms.elements_per_thread;
-    let offset = workgroup_id.x * uniforms.d_comp + local_offset;
+    let offset = (global_idx / ${i}) * uniforms.d_comp + local_offset;
 
     var thread_max_vector = ${m}(-3.402823e+38f);
     for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {
@@ -3989,21 +3988,21 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
 
     if (sum == 0) {
       for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {
-        x[offset + i] = ${w.type.value}(uniforms.d_inv);
+        x[offset + i] = ${g.type.value}(${y}(uniforms.d_inv));
       }
     } else {
       for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {
         var f32input = ${m}(x[offset + i]);
-        x[offset + i] = ${w.type.value}(exp(f32input - max_value) / sum);
+        x[offset + i] = ${g.type.value}(exp(f32input - max_value) / sum);
       }
     }
-  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${i};${p};${o}`},getShaderSource:u,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:d})}},tm=(e,t,r,n,o,i,a,l)=>{let d=l+i.kvSequenceLength,p=[i.batchSize,i.numHeads,i.sequenceLength,d],m=i.kvNumHeads===void 0&&e.outputCount>1,u=m?[i.batchSize,i.numHeads,d,i.headSize]:void 0,h=a.scale===0?1/Math.sqrt(i.headSize):a.scale,w=$e(i.headSize),g=i.headSize/w,b=12,x={x:Math.ceil(d/b),y:Math.ceil(i.sequenceLength/b),z:i.batchSize*i.numHeads},_=[{type:12,data:i.sequenceLength},{type:12,data:g},{type:12,data:d},{type:12,data:i.numHeads},{type:1,data:h},{type:12,data:l},{type:12,data:i.kvSequenceLength}],$=["type","type"];n&&$.push("type"),o&&$.push("type");let S=[{dims:p,dataType:t.dataType,gpuDataType:0}];m&&S.push({dims:u,dataType:t.dataType,gpuDataType:0});let I=T=>{let A=E("q",t.dataType,t.dims,w),D=E("key",r.dataType,r.dims,w),z=[A,D];if(n){let pe=E("past_key",n.dataType,n.dims,w);z.push(pe)}o&&z.push(E("relative_position_bias",o.dataType,o.dims));let H=M("output",t.dataType,p),W=[H];m&&W.push(M("present_key",t.dataType,u,w));let X=Oe(1,w),Y=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"}];return`
-  const TILE_SIZE = ${b}u;
+  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${i};${p};${o}`,inputDependencies:u},getShaderSource:h,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:d})}},Xp=(e,t,r,n,o,i,a,l)=>{let d=l+i.kvSequenceLength,p=[i.batchSize,i.numHeads,i.sequenceLength,d],m=i.kvNumHeads===void 0&&e.outputCount>1,u=m?[i.batchSize,i.numHeads,d,i.headSize]:void 0,h=a.scale===0?1/Math.sqrt(i.headSize):a.scale,w=he(i.headSize),g=i.headSize/w,y=12,x={x:Math.ceil(d/y),y:Math.ceil(i.sequenceLength/y),z:i.batchSize*i.numHeads},_=[{type:12,data:i.sequenceLength},{type:12,data:g},{type:12,data:d},{type:12,data:i.numHeads},{type:1,data:h},{type:12,data:l},{type:12,data:i.kvSequenceLength}],$=["type","type"];n&&$.push("type"),o&&$.push("type");let S=[{dims:p,dataType:t.dataType,gpuDataType:0}];m&&S.push({dims:u,dataType:t.dataType,gpuDataType:0});let I=T=>{let A=k("q",t.dataType,t.dims,w),O=k("key",r.dataType,r.dims,w),z=[A,O];if(n){let ae=k("past_key",n.dataType,n.dims,w);z.push(ae)}o&&z.push(k("attention_bias",o.dataType,o.dims));let W=D("output",t.dataType,p),V=[W];m&&V.push(D("present_key",t.dataType,u,w));let j=ke(1,w),Y=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"}];return`
+  const TILE_SIZE = ${y}u;
 
-  var<workgroup> tileQ: array<${A.type.storage}, ${b*b}>;
-  var<workgroup> tileK: array<${A.type.storage}, ${b*b}>;
-  ${T.registerUniforms(Y).declareVariables(...z,...W)}
-  ${T.mainStart([b,b,1])}
+  var<workgroup> tileQ: array<${A.type.storage}, ${y*y}>;
+  var<workgroup> tileK: array<${A.type.storage}, ${y*y}>;
+  ${T.registerUniforms(Y).declareVariables(...z,...V)}
+  ${T.mainStart([y,y,1])}
     // x holds the N and y holds the M
     let headIdx = workgroup_id.z;
     let m = workgroup_id.y * TILE_SIZE;
@@ -4014,7 +4013,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     let pastKeyOffset = uniforms.past_sequence_length * uniforms.K * headIdx;`:`
     let kOffset = uniforms.N * uniforms.K * headIdx + n * uniforms.K;`)()}
     ${m?"let presentKeyOffset = headIdx * uniforms.N * uniforms.K;":""}
-    var value = ${X}(0);
+    var value = ${j}(0);
     for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
       if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {
         tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];
@@ -4033,7 +4032,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
       workgroupBarrier();
 
       for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
-        value += ${X}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
+        value += ${j}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
       }
 
       workgroupBarrier();
@@ -4043,13 +4042,13 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     if (global_id.y < uniforms.M && global_id.x < uniforms.N) {
       let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
       var sum: f32 = ${(()=>{switch(w){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: ${w}`)}})()};
-        output[outputIdx] = ${H.type.value} (sum * uniforms.alpha) + ${o?"relative_position_bias[outputIdx]":"0.0"};
+        output[outputIdx] = ${W.type.value} (sum * uniforms.alpha) + ${o?"attention_bias[outputIdx]":"0.0"};
     }
-  }`};return{name:"AttentionProbs",shaderCache:{hint:`${w};${o!==void 0};${n!==void 0};${e.outputCount}`,inputDependencies:$},getRunData:()=>({outputs:S,dispatchGroup:x,programUniforms:_}),getShaderSource:I}},rm=(e,t,r,n,o,i)=>{let a=i+o.kvSequenceLength,l=o.nReps?o.nReps:1,d=o.vHiddenSize*l,p=o.kvNumHeads==null&&e.outputCount>1,m=p?[o.batchSize,o.numHeads,a,o.headSize]:void 0,u=[o.batchSize,o.sequenceLength,d],h=12,w={x:Math.ceil(o.vHeadSize/h),y:Math.ceil(o.sequenceLength/h),z:o.batchSize*o.numHeads},g=[{type:12,data:o.sequenceLength},{type:12,data:a},{type:12,data:o.vHeadSize},{type:12,data:o.numHeads},{type:12,data:d},{type:12,data:i},{type:12,data:o.kvSequenceLength}],b=n?["type","type","type"]:["type","type"],x=[{dims:u,dataType:t.dataType,gpuDataType:0}];p&&x.push({dims:m,dataType:t.dataType,gpuDataType:0});let _=$=>{let S=E("probs",t.dataType,t.dims),I=E("v",r.dataType,r.dims),T=[S,I];n&&T.push(E("past_value",n.dataType,n.dims));let D=[M("output",t.dataType,u)];p&&D.push(M("present_value",t.dataType,m));let z=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"}];return`
+  }`};return{name:"AttentionProbs",shaderCache:{hint:`${w};${o!==void 0};${n!==void 0};${e.outputCount}`,inputDependencies:$},getRunData:()=>({outputs:S,dispatchGroup:x,programUniforms:_}),getShaderSource:I}},Zp=(e,t,r,n,o,i)=>{let a=i+o.kvSequenceLength,l=o.nReps?o.nReps:1,d=o.vHiddenSize*l,p=o.kvNumHeads==null&&e.outputCount>1,m=p?[o.batchSize,o.numHeads,a,o.headSize]:void 0,u=[o.batchSize,o.sequenceLength,d],h=12,w={x:Math.ceil(o.vHeadSize/h),y:Math.ceil(o.sequenceLength/h),z:o.batchSize*o.numHeads},g=[{type:12,data:o.sequenceLength},{type:12,data:a},{type:12,data:o.vHeadSize},{type:12,data:o.numHeads},{type:12,data:d},{type:12,data:i},{type:12,data:o.kvSequenceLength}],y=n?["type","type","type"]:["type","type"],x=[{dims:u,dataType:t.dataType,gpuDataType:0}];p&&x.push({dims:m,dataType:t.dataType,gpuDataType:0});let _=$=>{let S=k("probs",t.dataType,t.dims),I=k("v",r.dataType,r.dims),T=[S,I];n&&T.push(k("past_value",n.dataType,n.dims));let O=[D("output",t.dataType,u)];p&&O.push(D("present_value",t.dataType,m));let z=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"}];return`
   const TILE_SIZE = ${h}u;
   var<workgroup> tileQ: array<${S.type.value}, ${h*h}>;
   var<workgroup> tileK: array<${S.type.value}, ${h*h}>;
-  ${$.registerUniforms(z).declareVariables(...T,...D)}
+  ${$.registerUniforms(z).declareVariables(...T,...O)}
   ${$.mainStart([h,h,1])}
    let headIdx = workgroup_id.z;
    let m = global_id.y;
@@ -4096,13 +4095,13 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
        + currentBatchHeadNumber * uniforms.N + n;
      output[outputIdx] = value;
    }
-  }`};return{name:"AttentionScore",shaderCache:{hint:`${n!==void 0};${e.outputCount}`,inputDependencies:b},getRunData:()=>({outputs:x,dispatchGroup:w,programUniforms:g}),getShaderSource:_}},Wt=(e,t,r,n,o,i,a,l,d,p,m)=>{let u=e.outputCount,h=p.kvNumHeads!==void 0||u>1?p.pastSequenceLength:0,w=h+p.kvSequenceLength,g=p.kvNumHeads===void 0&&u>1&&a?[t,r,a]:[t,r];d&&g.push(d);let b=e.compute(tm(e,t,r,u>1?a:void 0,d,p,m,h),{inputs:g,outputs:p.kvNumHeads===void 0&&u>1?[-1,1]:[-1]})[0];e.compute(em(e,b,p.batchSize*p.numHeads*p.sequenceLength,w),{inputs:[b],outputs:[]});let x=p.kvNumHeads===void 0&&u>1&&l?[b,n,l]:[b,n];e.compute(rm(e,b,n,u>1&&l?l:void 0,p,h),{inputs:x,outputs:p.kvNumHeads===void 0&&u>1?[0,2]:[0]})},nm=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],n=t.sequenceLength,o=t.inputHiddenSize,i=t.headSize,a=12,l={x:Math.ceil(t.headSize/a),y:Math.ceil(t.sequenceLength/a),z:t.batchSize*t.numHeads},d=[e.inputs[0],e.inputs[1],e.inputs[2]],p=[{type:12,data:n},{type:12,data:o},{type:12,data:i},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],m=u=>{let h=M("output_q",d[0].dataType,r),w=M("output_k",d[0].dataType,r),g=M("output_v",d[0].dataType,r),b=E("input",d[0].dataType,d[0].dims),x=E("weight",d[1].dataType,d[1].dims),_=E("bias",d[2].dataType,d[2].dims),$=b.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.outputCount}`,inputDependencies:y},getRunData:()=>({outputs:x,dispatchGroup:w,programUniforms:g}),getShaderSource:_}},Ut=(e,t,r,n,o,i,a,l,d,p,m)=>{let u=e.outputCount,h=p.kvNumHeads!==void 0||u>1?p.pastSequenceLength:0,w=h+p.kvSequenceLength,g=p.kvNumHeads===void 0&&u>1&&a?[t,r,a]:[t,r];d&&g.push(d);let y=e.compute(Xp(e,t,r,u>1?a:void 0,d,p,m,h),{inputs:g,outputs:p.kvNumHeads===void 0&&u>1?[-1,1]:[-1]})[0];e.compute(Yp(e,y,p.batchSize*p.numHeads*p.sequenceLength,w),{inputs:[y],outputs:[]});let x=p.kvNumHeads===void 0&&u>1&&l?[y,n,l]:[y,n];e.compute(Zp(e,y,n,u>1&&l?l:void 0,p,h),{inputs:x,outputs:p.kvNumHeads===void 0&&u>1?[0,2]:[0]})},Qp=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],n=t.sequenceLength,o=t.inputHiddenSize,i=t.headSize,a=12,l={x:Math.ceil(t.headSize/a),y:Math.ceil(t.sequenceLength/a),z:t.batchSize*t.numHeads},d=[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=D("output_q",d[0].dataType,r),w=D("output_k",d[0].dataType,r),g=D("output_v",d[0].dataType,r),y=k("input",d[0].dataType,d[0].dims),x=k("weight",d[1].dataType,d[1].dims),_=k("bias",d[2].dataType,d[2].dims),$=y.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 = ${a}u;
   var<workgroup> tileInput: array<${$}, ${a*a}>;
   var<workgroup> tileWeightQ: array<${$}, ${a*a}>;
   var<workgroup> tileWeightK: array<${$}, ${a*a}>;
   var<workgroup> tileWeightV: array<${$}, ${a*a}>;
-  ${u.registerUniforms(S).declareVariables(b,x,_,h,w,g)}
+  ${u.registerUniforms(S).declareVariables(y,x,_,h,w,g)}
   ${u.mainStart([a,a,1])}
     let batchIndex = workgroup_id.z / uniforms.num_heads;
     let headNumber = workgroup_id.z % uniforms.num_heads;
@@ -4151,12 +4150,12 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
       output_k[outputIdx] = valueK;
       output_v[outputIdx] = valueV;
     }
-  }`};return e.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:l,programUniforms:p}),getShaderSource:m},{inputs:d,outputs:[-1,-1,-1]})},Ss=(e,t)=>{let r=Jp(e.inputs,t),[n,o,i]=nm(e,r);return Wt(e,n,o,i,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t)}});var om,im,am,Is,Cs=U(()=>{"use strict";Ke();Q();ae();Ce();ie();om=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(n,o,i)=>{let a=o.length;if(a!==n.length)throw new Error(`${i}: num dimensions != ${a}`);o.forEach((l,d)=>{if(l!==n[d])throw new Error(`${i}: dim[${d}] do not match`)})};if(e[0].dims.length>1){let n=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,n,"Invalid input scale"),r(e[2].dims,n,"Invalid input B"),r(e[3].dims,n,"Invalid input mean"),r(e[4].dims,n,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},im=(e,t)=>{let{epsilon:r,spatial:n,format:o}=t,i=e[0].dims,a=n?$e(i[i.length-1]):1,l=o==="NHWC"&&i.length>1?a:1,d=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,l),w=E("bias",e[2].dataType,e[2].dims,l),g=E("inputMean",e[3].dataType,e[3].dims,l),b=E("inputVar",e[4].dataType,e[4].dims,l),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}] / ${a}`:"outputIndices[1]"};`;else if(o==="NCHW")S=`
+  }`};return e.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:l,programUniforms:p}),getShaderSource:m},{inputs:d,outputs:[-1,-1,-1]})},ys=(e,t)=>{let r=jp(e.inputs,t),[n,o,i]=Qp(e,r);return Ut(e,n,o,i,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t)}});var Jp,em,tm,bs,ws=M(()=>{"use strict";qe();X();oe();xe();re();Jp=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(n,o,i)=>{let a=o.length;if(a!==n.length)throw new Error(`${i}: num dimensions != ${a}`);o.forEach((l,d)=>{if(l!==n[d])throw new Error(`${i}: dim[${d}] 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")},em=(e,t)=>{let{epsilon:r,spatial:n,format:o}=t,i=e[0].dims,a=n?he(i[i.length-1]):1,l=o==="NHWC"&&i.length>1?a:1,d=E.size(i)/a,p=n,m=p?i.length:i,u=k("x",e[0].dataType,e[0].dims,a),h=k("scale",e[1].dataType,e[1].dims,l),w=k("bias",e[2].dataType,e[2].dims,l),g=k("inputMean",e[3].dataType,e[3].dims,l),y=k("inputVar",e[4].dataType,e[4].dims,l),x=D("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 I=1;I<h.rank;I++)S+=`cIndices[${I}] = outputIndices[${I}];`;S+=`let cOffset = ${h.indicesToOffset("cIndices")};`}return S},$=S=>`
   const epsilon = ${r};
-  ${S.registerUniform("outputSize","u32").declareVariables(u,h,w,g,b,x)}
+  ${S.registerUniform("outputSize","u32").declareVariables(u,h,w,g,y,x)}
   ${S.mainStart()}
   ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
     var outputIndices = ${x.offsetToIndices(`global_idx * ${a}`)};
@@ -4164,11 +4163,11 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     let scale = ${h.getByOffset("cOffset")};
     let bias = ${w.getByOffset("cOffset")};
     let inputMean = ${g.getByOffset("cOffset")};
-    let inputVar = ${b.getByOffset("cOffset")};
+    let inputVar = ${y.getByOffset("cOffset")};
     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}_${a}`,inputDependencies:p?["rank","type","type","type","type"]:void 0},getShaderSource:$,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p?[{type:12,data:d},...V(i)]:[{type:12,data:d}]})}},am=e=>te(e),Is=(e,t)=>{let{inputs:r,outputCount:n}=e,o=am({...t,outputCount:n});if(be.webgpu.validateInputContent&&om(r,o),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(im(r,o))}});var sm,um,Ts,As=U(()=>{"use strict";ae();ie();sm=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")},um=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),l=E("residual",o,t,4),d=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:$,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p?[{type:12,data:d},...U(i)]:[{type:12,data:d}]})}},tm=e=>Q(e),bs=(e,t)=>{let{inputs:r,outputCount:n}=e,o=tm({...t,outputCount:n});if(fe.webgpu.validateInputContent&&Jp(r,o),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(em(r,o))}});var rm,nm,vs,$s=M(()=>{"use strict";oe();re();rm=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")},nm=e=>{let t=e[0].dims,r=e[0].dims[2],n=E.size(t)/4,o=e[0].dataType,i=k("input",o,t,4),a=k("bias",o,[r],4),l=k("residual",o,t,4),d=D("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,l,d)}
 
@@ -4177,7 +4176,7 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
     let value = ${i.getByOffset("global_idx")}
       + ${a.getByOffset("global_idx % channels")} + ${l.getByOffset("global_idx")};
     ${d.setByOffset("global_idx","value")}
-  }`}},Ts=e=>{sm(e.inputs),e.compute(um(e.inputs))}});var dm,fe,Es,ks,Ps,Os,Ds,zs,Bs,Rs,Ms,lm,Us,Vs,Ns,Ws,Qt,Hs,Kr,Gs,Ls,Fs,qs,js,Ks,Ys,Xs,Zs,Qs,Js,eu,tu,ru,nu,ou,iu,au,ao,so,su,uu,du,cm,pm,lu,Yr=U(()=>{"use strict";Q();ae();Ce();ie();dm=(e,t,r,n,o,i)=>{let a=Math.ceil(t/4),l="";typeof o=="string"?l=`${o}(a)`:l=o("a");let d=E("inputData",r,[a],4),p=M("outputData",n,[a],4);return`
+  }`}},vs=e=>{rm(e.inputs),e.compute(nm(e.inputs))}});var om,le,_s,xs,Ss,Is,Cs,Ts,As,ks,Es,im,Ps,Os,zs,Bs,Yt,Ds,Fr,Rs,Ms,Us,Vs,Ns,Ws,Ls,Hs,Gs,Fs,qs,Ks,js,Ys,Xs,Zs,Qs,Js,ro,no,eu,tu,ru,am,sm,nu,qr=M(()=>{"use strict";X();oe();xe();re();om=(e,t,r,n,o,i)=>{let a=Math.ceil(t/4),l="";typeof o=="string"?l=`${o}(a)`:l=o("a");let d=k("inputData",r,[a],4),p=D("outputData",n,[a],4);return`
       ${e.registerUniform("vec_size","u32").declareVariables(d,p)}
 
   ${i??""}
@@ -4187,10 +4186,10 @@ var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLas
 
     let a = ${d.getByOffset("global_idx")};
     ${p.setByOffset("global_idx",l)}
-  }`},fe=(e,t,r,n,o,i=e.dataType)=>({name:t,shaderCache:{hint:o,inputDependencies:["type"]},getShaderSource:a=>dm(a,k.size(e.dims),e.dataType,i,r,n),getRunData:a=>({outputs:[{dims:e.dims,dataType:i}],dispatchGroup:{x:Math.ceil(k.size(a[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(k.size(e.dims)/4)}]})}),Es=e=>{e.compute(fe(e.inputs[0],"Abs","abs"))},ks=e=>{e.compute(fe(e.inputs[0],"Acos","acos"))},Ps=e=>{e.compute(fe(e.inputs[0],"Acosh","acosh"))},Os=e=>{e.compute(fe(e.inputs[0],"Asin","asin"))},Ds=e=>{e.compute(fe(e.inputs[0],"Asinh","asinh"))},zs=e=>{e.compute(fe(e.inputs[0],"Atan","atan"))},Bs=e=>{e.compute(fe(e.inputs[0],"Atanh","atanh"))},Rs=e=>te(e),Ms=(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): ${t.to}`)}e.compute(fe(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},lm=e=>{let t=e.length>=2&&e[1].data!==0?e[1].getFloat32Array()[0]:Hr,r=e.length>=3&&e[2].data!==0?e[2].getFloat32Array()[0]:Gr;return te({min:t,max:r})},Us=(e,t)=>{let r=e.inputs.length===1?t:lm(e.inputs),n=Oe(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Clip",o=>`clamp(${o}, clip_min_, clip_max_)`,`
+  }`},le=(e,t,r,n,o,i=e.dataType)=>({name:t,shaderCache:{hint:o,inputDependencies:["type"]},getShaderSource:a=>om(a,E.size(e.dims),e.dataType,i,r,n),getRunData:a=>({outputs:[{dims:e.dims,dataType:i}],dispatchGroup:{x:Math.ceil(E.size(a[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(E.size(e.dims)/4)}]})}),_s=e=>{e.compute(le(e.inputs[0],"Abs","abs"))},xs=e=>{e.compute(le(e.inputs[0],"Acos","acos"))},Ss=e=>{e.compute(le(e.inputs[0],"Acosh","acosh"))},Is=e=>{e.compute(le(e.inputs[0],"Asin","asin"))},Cs=e=>{e.compute(le(e.inputs[0],"Asinh","asinh"))},Ts=e=>{e.compute(le(e.inputs[0],"Atan","atan"))},As=e=>{e.compute(le(e.inputs[0],"Atanh","atanh"))},ks=e=>Q(e),Es=(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): ${t.to}`)}e.compute(le(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},im=e=>{let t=e.length>=2&&e[1].data!==0?e[1].getFloat32Array()[0]:Vr,r=e.length>=3&&e[2].data!==0?e[2].getFloat32Array()[0]:Nr;return Q({min:t,max:r})},Ps=(e,t)=>{let r=e.inputs.length===1?t:im(e.inputs),n=ke(e.inputs[0].dataType);e.compute(le(e.inputs[0],"Clip",o=>`clamp(${o}, clip_min_, clip_max_)`,`
     const clip_min_: vec4<${n}> = vec4(${n}(${r.min}));
     const clip_max_: vec4<${n}> = vec4(${n}(${r.max}));
-`,r.cacheKey),{inputs:[0]})},Vs=e=>{e.compute(fe(e.inputs[0],"Ceil","ceil"))},Ns=e=>{e.compute(fe(e.inputs[0],"Cos","cos"))},Ws=e=>{e.compute(fe(e.inputs[0],"Cosh","cosh"))},Qt=e=>te(e),Hs=(e,t)=>{let r=Oe(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Elu",n=>`elu_vf32(${n})`,`
+`,r.cacheKey),{inputs:[0]})},Os=e=>{e.compute(le(e.inputs[0],"Ceil","ceil"))},zs=e=>{e.compute(le(e.inputs[0],"Cos","cos"))},Bs=e=>{e.compute(le(e.inputs[0],"Cosh","cosh"))},Yt=e=>Q(e),Ds=(e,t)=>{let r=ke(e.inputs[0].dataType);e.compute(le(e.inputs[0],"Elu",n=>`elu_vf32(${n})`,`
   const elu_alpha_ = ${r}(${t.alpha});
 
   fn elu_f32(a: ${r}) -> ${r} {
@@ -4199,7 +4198,7 @@ var best_index : i32 = 0;`,`if (${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))},Kr=(e="f32")=>`
+  }`,t.cacheKey))},Fr=(e="f32")=>`
 const r0: ${e} = 0.3275911;
 const r1: ${e} = 0.254829592;
 const r2: ${e} = -0.284496736;
@@ -4211,15 +4210,15 @@ fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {
   let absv = abs(v);
   let x = 1.0 / (1.0 + r0 * absv);
   return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
-}`,Gs=e=>{let t=Oe(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,Kr(t)))},Ls=e=>{e.compute(fe(e.inputs[0],"Exp","exp"))},Fs=e=>{e.compute(fe(e.inputs[0],"Floor","floor"))},qs=e=>{let t=Oe(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Kr(t)))},js=(e,t)=>{let r=Oe(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))},Ks=e=>{e.compute(fe(e.inputs[0],"Not",t=>`!${t}`))},Ys=e=>{e.compute(fe(e.inputs[0],"Neg",t=>`-${t}`))},Xs=e=>{e.compute(fe(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},Zs=e=>{let t=Oe(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},Qs=e=>{e.compute(fe(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},Js=e=>te(e),eu=(e,t)=>{let r=Oe(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${n} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},tu=e=>{e.compute(fe(e.inputs[0],"Sin","sin"))},ru=e=>{e.compute(fe(e.inputs[0],"Sinh","sinh"))},nu=e=>{e.compute(fe(e.inputs[0],"Sqrt","sqrt"))},ou=e=>{e.compute(fe(e.inputs[0],"Tan","tan"))},iu=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,au=e=>{e.compute(fe(e.inputs[0],"Tanh",iu))},ao=(e="f32")=>`
+}`,Rs=e=>{let t=ke(e.inputs[0].dataType);e.compute(le(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,Fr(t)))},Ms=e=>{e.compute(le(e.inputs[0],"Exp","exp"))},Us=e=>{e.compute(le(e.inputs[0],"Floor","floor"))},Vs=e=>{let t=ke(e.inputs[0].dataType);e.compute(le(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Fr(t)))},Ns=(e,t)=>{let r=ke(e.inputs[0].dataType);e.compute(le(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))},Ws=e=>{e.compute(le(e.inputs[0],"Not",t=>`!${t}`))},Ls=e=>{e.compute(le(e.inputs[0],"Neg",t=>`-${t}`))},Hs=e=>{e.compute(le(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},Gs=e=>{let t=ke(e.inputs[0].dataType);e.compute(le(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},Fs=e=>{e.compute(le(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},qs=e=>Q(e),Ks=(e,t)=>{let r=ke(e.inputs[0].dataType);e.compute(le(e.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${n} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},js=e=>{e.compute(le(e.inputs[0],"Sin","sin"))},Ys=e=>{e.compute(le(e.inputs[0],"Sinh","sinh"))},Xs=e=>{e.compute(le(e.inputs[0],"Sqrt","sqrt"))},Zs=e=>{e.compute(le(e.inputs[0],"Tan","tan"))},Qs=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Js=e=>{e.compute(le(e.inputs[0],"Tanh",Qs))},ro=(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 ${iu("v")};
+  return ${Qs("v")};
 }
-`,so=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,su=e=>{let t=Oe(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"FastGelu",so,ao(t),void 0,e.inputs[0].dataType))},uu=(e,t)=>{let r=Oe(e.inputs[0].dataType);return e.compute(fe(e.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},du=e=>{e.compute(fe(e.inputs[0],"Log","log"))},cm=(e,t)=>`
+`,no=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,eu=e=>{let t=ke(e.inputs[0].dataType);e.compute(le(e.inputs[0],"FastGelu",no,ro(t),void 0,e.inputs[0].dataType))},tu=(e,t)=>{let r=ke(e.inputs[0].dataType);return e.compute(le(e.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},ru=e=>{e.compute(le(e.inputs[0],"Log","log"))},am=(e,t)=>`
 const alpha = vec4<${e}>(${t});
 const one = ${e}(1.0);
 const zero = ${e}(0.0);
@@ -4236,13 +4235,13 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
   }
   return x * x1;
 }
-`,pm=e=>`quick_gelu_impl(${e})`,lu=(e,t)=>{let r=Oe(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"QuickGelu",pm,cm(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}});var mm,fm,pu,mu=U(()=>{"use strict";ae();ie();Yr();mm=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")},fm=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=ge(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:d=>`
+`,sm=e=>`quick_gelu_impl(${e})`,nu=(e,t)=>{let r=ke(e.inputs[0].dataType);e.compute(le(e.inputs[0],"QuickGelu",sm,am(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}});var um,dm,iu,au=M(()=>{"use strict";oe();re();qr();um=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")},dm=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=k("input",e[0].dataType,e[0].dims,4),n=k("bias",e[0].dataType,[e[0].dims[2]],4),o=D("output",e[0].dataType,t,4),i=E.size(t)/4,a=pe(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:d=>`
   const M_SQRT2 = sqrt(2.0);
   const halfChannels = ${e[0].dims[2]/4/2}u;
 
   ${d.declareVariables(r,n,o)}
 
-  ${Kr(a)}
+  ${Fr(a)}
 
   ${d.mainStart()}
     ${d.guardAgainstOutOfBoundsWorkgroupSizes(i)}
@@ -4254,20 +4253,20 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
     let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);
 
     ${o.setByOffset("global_idx","valueLeft * geluRight")}
-  }`}},pu=e=>{mm(e.inputs),e.compute(fm(e.inputs))}});var hm,gm,at,fu,hu,gu,yu,bu,wu,vu,$u,_u,xu,Su=U(()=>{"use strict";Q();ae();ie();hm=(e,t,r,n,o,i,a,l,d,p,m,u)=>{let h,w;typeof l=="string"?h=w=($,S)=>`${l}((${$}),(${S}))`:typeof l=="function"?h=w=l:(h=l.scalar,w=l.vector);let g=M("outputData",m,n.length,4),b=E("aData",d,t.length,4),x=E("bData",p,r.length,4),_;if(o)if(i){let $=k.size(t)===1,S=k.size(r)===1,I=t.length>0&&t[t.length-1]%4===0,T=r.length>0&&r[r.length-1]%4===0;$||S?_=g.setByOffset("global_idx",w($?`${b.type.value}(${b.getByOffset("0")}.x)`:b.getByOffset("global_idx"),S?`${x.type.value}(${x.getByOffset("0")}.x)`:x.getByOffset("global_idx"))):_=`
+  }`}},iu=e=>{um(e.inputs),e.compute(dm(e.inputs))}});var lm,cm,ot,su,uu,du,lu,cu,pu,mu,fu,hu,gu,yu=M(()=>{"use strict";X();oe();re();lm=(e,t,r,n,o,i,a,l,d,p,m,u)=>{let h,w;typeof l=="string"?h=w=($,S)=>`${l}((${$}),(${S}))`:typeof l=="function"?h=w=l:(h=l.scalar,w=l.vector);let g=D("outputData",m,n.length,4),y=k("aData",d,t.length,4),x=k("bData",p,r.length,4),_;if(o)if(i){let $=E.size(t)===1,S=E.size(r)===1,I=t.length>0&&t[t.length-1]%4===0,T=r.length>0&&r[r.length-1]%4===0;$||S?_=g.setByOffset("global_idx",w($?`${y.type.value}(${y.getByOffset("0")}.x)`:y.getByOffset("global_idx"),S?`${x.type.value}(${x.getByOffset("0")}.x)`:x.getByOffset("global_idx"))):_=`
             let outputIndices = ${g.offsetToIndices("global_idx * 4u")};
-            let offsetA = ${b.broadcastedIndicesToOffset("outputIndices",g)};
+            let offsetA = ${y.broadcastedIndicesToOffset("outputIndices",g)};
             let offsetB = ${x.broadcastedIndicesToOffset("outputIndices",g)};
-            ${g.setByOffset("global_idx",w(a||I?b.getByOffset("offsetA / 4u"):`${b.type.value}(${b.getByOffset("offsetA / 4u")}[offsetA % 4u])`,a||T?x.getByOffset("offsetB / 4u"):`${x.type.value}(${x.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
-          `}else _=g.setByOffset("global_idx",w(b.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 $=(S,I,T="")=>{let A=`aData[indexA${I}][componentA${I}]`,D=`bData[indexB${I}][componentB${I}]`;return`
+            ${g.setByOffset("global_idx",w(a||I?y.getByOffset("offsetA / 4u"):`${y.type.value}(${y.getByOffset("offsetA / 4u")}[offsetA % 4u])`,a||T?x.getByOffset("offsetB / 4u"):`${x.type.value}(${x.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
+          `}else _=g.setByOffset("global_idx",w(y.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 $=(S,I,T="")=>{let A=`aData[indexA${I}][componentA${I}]`,O=`bData[indexB${I}][componentB${I}]`;return`
             let outputIndices${I} = ${g.offsetToIndices(`global_idx * 4u + ${I}u`)};
-            let offsetA${I} = ${b.broadcastedIndicesToOffset(`outputIndices${I}`,g)};
+            let offsetA${I} = ${y.broadcastedIndicesToOffset(`outputIndices${I}`,g)};
             let offsetB${I} = ${x.broadcastedIndicesToOffset(`outputIndices${I}`,g)};
             let indexA${I} = offsetA${I} / 4u;
             let indexB${I} = offsetB${I} / 4u;
             let componentA${I} = offsetA${I} % 4u;
             let componentB${I} = offsetB${I} % 4u;
-            ${S}[${I}] = ${T}(${h(A,D)});
+            ${S}[${I}] = ${T}(${h(A,O)});
           `};m===9?_=`
             var data = vec4<u32>(0);
             ${$("data",0,"u32")}
@@ -4280,14 +4279,14 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
             ${$("outputData[global_idx]",2)}
             ${$("outputData[global_idx]",3)}
           `}return`
-        ${e.registerUniform("vec_size","u32").declareVariables(b,x,g)}
+        ${e.registerUniform("vec_size","u32").declareVariables(y,x,g)}
 
         ${u??""}
 
         ${e.mainStart()}
         ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
         ${_}
-      }`},gm=(e,t,r,n,o,i,a=r.dataType)=>{let l=!k.areEqual(r.dims,n.dims),d=r.dims,p=k.size(r.dims),m=!1,u=!1,h=[l];if(l){let w=tt.calcShape(r.dims,n.dims,!1);if(!w)throw new Error("Can't perform binary op on the given tensors");d=w,p=k.size(d);let g=k.size(r.dims)===1,b=k.size(n.dims)===1,x=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,_=n.dims.length>0&&n.dims[n.dims.length-1]%4===0;h.push(g),h.push(b),h.push(x),h.push(_);let $=1;for(let S=1;S<d.length;S++){let I=r.dims[r.dims.length-S]??1,T=n.dims[n.dims.length-S]??1;if(I===T)$*=I;else break}$%4===0?(u=!0,m=!0):(g||b||x||_)&&(m=!0)}else m=!0;return h.push(m),{name:e,shaderCache:{hint:t+h.map(w=>w.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:w=>hm(w,r.dims,n.dims,d,m,l,u,o,r.dataType,n.dataType,a,i),getRunData:()=>({outputs:[{dims:d,dataType:a}],dispatchGroup:{x:Math.ceil(p/64/4)},programUniforms:[{type:12,data:Math.ceil(k.size(d)/4)},...V(r.dims,n.dims,d)]})}},at=(e,t,r,n,o,i)=>{e.compute(gm(t,o??"",e.inputs[0],e.inputs[1],r,n,i))},fu=e=>{at(e,"Add",(t,r)=>`${t}+${r}`)},hu=e=>{at(e,"Div",(t,r)=>`${t}/${r}`)},gu=e=>{at(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},yu=e=>{at(e,"Mul",(t,r)=>`${t}*${r}`)},bu=e=>{let t=E("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;at(e,"Pow",{scalar:(n,o)=>`pow_custom(${n},${o})`,vector:(n,o)=>`pow_vector_custom(${n},${o})`},`
+      }`},cm=(e,t,r,n,o,i,a=r.dataType)=>{let l=!E.areEqual(r.dims,n.dims),d=r.dims,p=E.size(r.dims),m=!1,u=!1,h=[l];if(l){let w=Je.calcShape(r.dims,n.dims,!1);if(!w)throw new Error("Can't perform binary op on the given tensors");d=w,p=E.size(d);let g=E.size(r.dims)===1,y=E.size(n.dims)===1,x=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,_=n.dims.length>0&&n.dims[n.dims.length-1]%4===0;h.push(g),h.push(y),h.push(x),h.push(_);let $=1;for(let S=1;S<d.length;S++){let I=r.dims[r.dims.length-S]??1,T=n.dims[n.dims.length-S]??1;if(I===T)$*=I;else break}$%4===0?(u=!0,m=!0):(g||y||x||_)&&(m=!0)}else m=!0;return h.push(m),{name:e,shaderCache:{hint:t+h.map(w=>w.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:w=>lm(w,r.dims,n.dims,d,m,l,u,o,r.dataType,n.dataType,a,i),getRunData:()=>({outputs:[{dims:d,dataType:a}],dispatchGroup:{x:Math.ceil(p/64/4)},programUniforms:[{type:12,data:Math.ceil(E.size(d)/4)},...U(r.dims,n.dims,d)]})}},ot=(e,t,r,n,o,i)=>{e.compute(cm(t,o??"",e.inputs[0],e.inputs[1],r,n,i))},su=e=>{ot(e,"Add",(t,r)=>`${t}+${r}`)},uu=e=>{ot(e,"Div",(t,r)=>`${t}/${r}`)},du=e=>{ot(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},lu=e=>{ot(e,"Mul",(t,r)=>`${t}*${r}`)},cu=e=>{let t=k("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;ot(e,"Pow",{scalar:(n,o)=>`pow_custom(${n},${o})`,vector:(n,o)=>`pow_vector_custom(${n},${o})`},`
     fn pow_custom(a : ${t}, b : ${t}) -> ${t} {
       if (b == ${t}(0.0)) {
         return ${t}(1.0);
@@ -4300,7 +4299,7 @@ fn quick_gelu_impl(x: vec4<${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));
     }
-      `)},wu=e=>{at(e,"Sub",(t,r)=>`${t}-${r}`)},vu=e=>{at(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},$u=e=>{at(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},_u=e=>{at(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},xu=e=>{at(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}});var bm,wm,vm,$m,Iu,Cu,Tu=U(()=>{"use strict";Q();ae();Ce();ie();bm=(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,l)=>{if(l!==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((d,p)=>{if(p!==t&&d!==n.dims[p])throw new Error("non concat dimensions must match")})}})},wm=(e,t)=>`
+      `)},pu=e=>{ot(e,"Sub",(t,r)=>`${t}-${r}`)},mu=e=>{ot(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},fu=e=>{ot(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},hu=e=>{ot(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},gu=e=>{ot(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}});var mm,fm,hm,gm,bu,wu,vu=M(()=>{"use strict";X();oe();xe();re();mm=(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,l)=>{if(l!==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((d,p)=>{if(p!==t&&d!==n.dims[p])throw new Error("non concat dimensions must match")})}})},fm=(e,t)=>`
   fn calculateInputIndex(index: u32) -> u32 {
     let sizeInConcatAxis = array<u32, ${e}u>(${t});
     for (var i: u32 = 0u; i < ${e}; i += 1u ) {
@@ -4309,15 +4308,15 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
       }
     }
     return ${e}u;
-  }`,vm=(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}) { ${i} }`)}return n.join(`
-`)},$m=(e,t,r,n)=>{let o=k.size(r),i=new Array(e.length),a=new Array(e.length),l=0,d=[],p=[],m=[{type:12,data:o}];for(let b=0;b<e.length;++b)l+=e[b].dims[t],i[b]=l,p.push(e[b].dims.length),a[b]=E(`input${b}`,n,p[b]),d.push("rank"),m.push({type:12,data:i[b]});for(let b=0;b<e.length;++b)m.push(...V(e[b].dims));m.push(...V(r));let u=M("output",n,r.length),h=u.indicesGet("indices",t),w=Array.from(Array(i.length).keys()).map(b=>`uniforms.sizeInConcatAxis${b}`).join(","),g=b=>`
+  }`,hm=(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}) { ${i} }`)}return n.join(`
+`)},gm=(e,t,r,n)=>{let o=E.size(r),i=new Array(e.length),a=new Array(e.length),l=0,d=[],p=[],m=[{type:12,data:o}];for(let y=0;y<e.length;++y)l+=e[y].dims[t],i[y]=l,p.push(e[y].dims.length),a[y]=k(`input${y}`,n,p[y]),d.push("rank"),m.push({type:12,data:i[y]});for(let y=0;y<e.length;++y)m.push(...U(e[y].dims));m.push(...U(r));let u=D("output",n,r.length),h=u.indicesGet("indices",t),w=Array.from(Array(i.length).keys()).map(y=>`uniforms.sizeInConcatAxis${y}`).join(","),g=y=>`
 
-  ${(()=>{b.registerUniform("outputSize","u32");for(let x=0;x<e.length;x++)b.registerUniform(`sizeInConcatAxis${x}`,"u32");return b.declareVariables(...a,u)})()}
+  ${(()=>{y.registerUniform("outputSize","u32");for(let x=0;x<e.length;x++)y.registerUniform(`sizeInConcatAxis${x}`,"u32");return y.declareVariables(...a,u)})()}
 
-  ${wm(i.length,w)}
+  ${fm(i.length,w)}
 
-  ${b.mainStart()}
-    ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+  ${y.mainStart()}
+    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
 
     var indices = ${u.offsetToIndices("global_idx")};
 
@@ -4327,12 +4326,12 @@ fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
       ${h} -= sizeInConcatAxis[inputIndex - 1u];
     }
 
-    ${vm(a,u)}
-  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:m}),getShaderSource:g}},Iu=(e,t)=>{let r=e.inputs,n=r[0].dims,o=k.normalizeAxis(t.axis,n.length);bm(r,o);let i=n.slice();i[o]=r.reduce((l,d)=>l+(d.dims.length>o?d.dims[o]:0),0);let a=r.filter(l=>k.size(l.dims)>0);e.compute($m(a,o,i,r[0].dataType),{inputs:a})},Cu=e=>te({axis:e.axis})});var He,Ge,Le,Xr,lt=U(()=>{"use strict";Q();ae();He=(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));
+    ${hm(a,u)}
+  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:m}),getShaderSource:g}},bu=(e,t)=>{let r=e.inputs,n=r[0].dims,o=E.normalizeAxis(t.axis,n.length);mm(r,o);let i=n.slice();i[o]=r.reduce((l,d)=>l+(d.dims.length>o?d.dims[o]:0),0);let a=r.filter(l=>E.size(l.dims)>0);e.compute(gm(a,o,i,r[0].dataType),{inputs:a})},wu=e=>Q({axis:e.axis})});var Ne,We,Le,Kr,ut=M(()=>{"use strict";X();oe();Ne=(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 ${e.activation}`)}},Ge=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},Le=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},Xr=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,n]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:n}}else if(t==="Clip"){let[r,n]=e?.activation_params||[Hr,Gr];return{activation:t,clipMax:n,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}});var Ee,Zr,Jt=U(()=>{"use strict";Ee=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Zr=e=>`
+        `;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},We=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},Le=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},Kr=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,n]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:n}}else if(t==="Clip"){let[r,n]=e?.activation_params||[Vr,Nr];return{activation:t,clipMax:n,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}});var Ce,jr,Xt=M(()=>{"use strict";Ce=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},jr=e=>`
       ${e?"value = value + getBiasByOutputCoords(coords);":""}
-      `});var Qr,uo=U(()=>{"use strict";Qr=e=>`
+      `});var Yr,oo=M(()=>{"use strict";Yr=e=>`
 fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
   return dot(coords, vec4<i32>(
       shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
@@ -4341,7 +4340,7 @@ fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
   return dot(coords, vec4<i32>(
     i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));
 }
-`});var _m,xm,er,Au,Sm,tr,Im,Jr,rr=U(()=>{"use strict";Q();ae();ie();lt();Jt();_m=(e,t)=>e?`
+`});var ym,bm,Zt,$u,wm,Qt,vm,Xr,Jt=M(()=>{"use strict";X();oe();re();ut();Xt();ym=(e,t)=>e?`
         mm_Asub[inputRow][inputCol] = mm_readA(batch,
           kStart + inputRow,
           globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});
@@ -4349,7 +4348,7 @@ fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
         mm_Asub[inputRow][inputCol] = mm_readA(batch,
           globalRow + innerRow,
           kStart / innerElementSize + inputCol${t?", batchIndices":""});
-        `,xm=(e,t)=>e?`
+        `,bm=(e,t)=>e?`
         let ACached0 = mm_Asub[k * innerElementSize][localRow];
         let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];
         let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];
@@ -4366,7 +4365,7 @@ 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];"}
-        }`,er=(e,t,r="f32",n,o=!1,i=32,a=!1,l=32)=>{let d=t[1]*e[1],p=t[0]*e[0],m=o?d:i,u=o?i:d,h=m/t[0],w=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.
+        }`,Zt=(e,t,r="f32",n,o=!1,i=32,a=!1,l=32)=>{let d=t[1]*e[1],p=t[0]*e[0],m=o?d:i,u=o?i:d,h=m/t[0],w=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}>;
@@ -4403,7 +4402,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
       for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
           let inputRow = tileRow + innerRow;
           let inputCol = tileCol;
-          ${_m(o,n)}
+          ${ym(o,n)}
       }
 
       // Load one tile of B into local memory.
@@ -4422,7 +4421,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
           let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];
           ${h===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}
 
-          ${xm(o,h)}
+          ${bm(o,h)}
       }
 
       workgroupBarrier();
@@ -4431,7 +4430,7 @@ 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]);
   }
-}`},Au=(e,t)=>e?`
+}`},$u=(e,t)=>e?`
             mm_Asub[inputRow][inputCol] = mm_readA(batch,
               kStart + inputRow,
               globalRowStart + inputCol${t?", batchIndices":""});
@@ -4439,7 +4438,7 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
             mm_Asub[inputRow][inputCol] = mm_readA(batch,
               globalRowStart + inputRow,
               kStart + inputCol${t?", batchIndices":""});
-            `,Sm=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",tr=(e,t,r="f32",n,o=!1,i=32,a=!1,l=32,d=!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 w=h/t[1],g=u/t[0],b=i/t[1],x=d?`
+            `,wm=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",Qt=(e,t,r="f32",n,o=!1,i=32,a=!1,l=32,d=!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 w=h/t[1],g=u/t[0],y=i/t[1],x=d?`
     let localRow = i32(localId.y);
     let localCol = i32(localId.x);
     let globalRowStart = i32(workgroupId.y) * ${p};
@@ -4450,7 +4449,7 @@ 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]}) {
-          ${Au(o,n)}
+          ${$u(o,n)}
         }
       }
       // Load one tile of B into local memory.
@@ -4497,7 +4496,7 @@ let globalRowStart = i32(workgroupId.y) * ${p};
 
 let tileRowA = i32(localId.y) * ${w};
 let tileColA = i32(localId.x) * ${g};
-let tileRowB = i32(localId.y) * ${b};
+let tileRowB = i32(localId.y) * ${y};
 // Loop over shared dimension.
 for (var t = 0; t < num_tiles; t = t + 1) {
   // Load one tile of A into local memory.
@@ -4505,12 +4504,12 @@ for (var t = 0; t < num_tiles; t = t + 1) {
     for (var innerCol = 0; innerCol < ${g}; innerCol = innerCol + 1) {
       let inputRow = tileRowA + innerRow;
       let inputCol = tileColA + innerCol;
-      ${Au(o,n)}
+      ${$u(o,n)}
     }
   }
 
   // Load one tile of B into local memory.
-  for (var innerRow = 0; innerRow < ${b}; innerRow = innerRow + 1) {
+  for (var innerRow = 0; innerRow < ${y}; innerRow = innerRow + 1) {
     for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
       let inputRow = tileRowB + innerRow;
       let inputCol = tileCol + innerCol;
@@ -4530,7 +4529,7 @@ for (var t = 0; t < num_tiles; t = t + 1) {
     }
 
     for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
-      ${Sm(o)}
+      ${wm(o)}
       for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
         acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
       }
@@ -4565,17 +4564,17 @@ fn main(@builtin(local_invocation_id) localId : vec3<u32>,
     var acc : array<array<${r}, colPerThread>, rowPerThread>;
     ${x}
   }
-`},Im=(e,t,r,n,o,i=!1)=>{let[a,l,d]=o,[p,m,u,h]=n,w=Nt(a,d),g=Nt(l,d),b=ge(n[0].type.tensor),x=()=>{let S=m.rank,I=p.rank,T=`var aIndices: ${m.type.indices};`;for(let A=S-2-1,D=I-1;A>=0;A--,D--)T+=`
-aIndices[${A}] = ${I>1?`batchIndices[${D}]`:"batchIndices"};`;return w.forEach(A=>{T+=`
+`},vm=(e,t,r,n,o,i=!1)=>{let[a,l,d]=o,[p,m,u,h]=n,w=Mt(a,d),g=Mt(l,d),y=pe(n[0].type.tensor),x=()=>{let S=m.rank,I=p.rank,T=`var aIndices: ${m.type.indices};`;for(let A=S-2-1,O=I-1;A>=0;A--,O--)T+=`
+aIndices[${A}] = ${I>1?`batchIndices[${O}]`:"batchIndices"};`;return w.forEach(A=>{T+=`
 aIndices[${A}] = 0;`}),T+=`
 aIndices[${S-2}] = u32(row);
-                   aIndices[${S-1}] = u32(colIn);`,T},_=()=>{let S=u.rank,I=p.rank,T=`var bIndices: ${u.type.indices};`;for(let A=S-2-1,D=I-1;A>=0;A--,D--)T+=`
-bIndices[${A}] = ${I>1?`batchIndices[${D}]`:"batchIndices"};`;return g.forEach(A=>{T+=`
+                   aIndices[${S-1}] = u32(colIn);`,T},_=()=>{let S=u.rank,I=p.rank,T=`var bIndices: ${u.type.indices};`;for(let A=S-2-1,O=I-1;A>=0;A--,O--)T+=`
+bIndices[${A}] = ${I>1?`batchIndices[${O}]`:"batchIndices"};`;return g.forEach(A=>{T+=`
 bIndices[${A}] = 0;`}),T+=`
 bIndices[${S-2}] = u32(row);
                    bIndices[${S-1}] = u32(colIn);`,T};return`
-    fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${p.type.indices}) -> ${Ee(e,b)} {
-      var value = ${Ee(e,b)}(0.0);
+    fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${p.type.indices}) -> ${Ce(e,y)} {
+      var value = ${Ce(e,y)}(0.0);
       let col = colIn * ${e};
       if(row < uniforms.dim_a_outer && col < uniforms.dim_inner)
       {
@@ -4585,8 +4584,8 @@ bIndices[${S-2}] = u32(row);
       return value;
     }
 
-    fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${p.type.indices}) -> ${Ee(e,b)} {
-      var value = ${Ee(e,b)}(0.0);
+    fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${p.type.indices}) -> ${Ce(e,y)} {
+      var value = ${Ce(e,y)}(0.0);
       let col = colIn * ${e};
       if(row < uniforms.dim_inner && col < uniforms.dim_b_outer)
       {
@@ -4596,21 +4595,21 @@ bIndices[${S-2}] = u32(row);
       return value;
     }
 
-    fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${Ee(e,b)}) {
+    fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${Ce(e,y)}) {
       let col = colIn * ${e};
       if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {
         var value = valueIn;
         let coords = vec3<i32>(batch, row, colIn);
-        ${t?`value = value + ${i?"bias[colIn]":`${Ee(e,b)}(bias[row])`};`:""}
+        ${t?`value = value + ${i?"bias[colIn]":`${Ce(e,y)}(bias[row])`};`:""}
         ${r}
         ${h.setByIndices("vec3<u32>(coords)","value")}
       }
     }
-    `},Jr=(e,t,r,n,o=!1)=>{let i=e[0].dims,a=e[1].dims,l=i.slice(0,-2),d=a.slice(0,-2),p=n?n.slice(0,-2):r.slice(0,-2),m=k.size(p),u=i[i.length-2],h=i[i.length-1],w=a[a.length-1],g=h%4===0&&w%4===0,b=u<=8?[4,1,1]:[4,4,1],x=[8,8,1],_=[Math.ceil(w/x[0]/b[0]),Math.ceil(u/x[1]/b[1]),Math.ceil(m/x[2]/b[2])],$=g?4:1,S=[...l,u,h/$],I=S.length,T=[...d,h,w/$],A=T.length,D=[m,u,w/$],z=[{type:6,data:u},{type:6,data:w},{type:6,data:h}];Ge(t,z),z.push(...V(p,S,T));let H=["rank","rank"],W=e.length>2;W&&(z.push(...V(e[2].dims)),H.push("rank")),z.push(...V(D));let X=Y=>{let pe=p.length,ue=Lr("batchDims",e[0].dataType,pe,1),le=ge(e[0].dataType),K=E("a",e[0].dataType,I,$),de=E("b",e[1].dataType,A,$),re=M("result",e[0].dataType,D.length,$),se=[K,de];if(W){let R=o?$:1;se.push(E("bias",e[2].dataType,e[2].dims.length,R))}let Z=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];Le(t,Z);let ne=ge(re.type.tensor),J=He(t,re.type.value,ne),ke=Im($,W,J,[ue,K,de,re],[l,d,p],o);return`
-  ${Y.registerUniforms(Z).registerInternalVariables(ue).declareVariables(...se,re)}
-  ${ke}
-  ${g?er(b,x,le,ue):tr(b,x,le,ue)}
-                   `};return{name:"MatMul",shaderCache:{hint:`${b};${t.activation};${g};${o}`,inputDependencies:H},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:_[0],y:_[1],z:_[2]},programUniforms:z}),getShaderSource:X}}});var Cm,Eu,ku=U(()=>{"use strict";Q();dt();ie();lt();Jt();uo();rr();Cm=(e,t,r,n,o=!1,i,a=4,l=4,d=4,p="f32")=>{let m=W=>{switch(W){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${p}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${W} is not supported.`)}},u=W=>{switch(W){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 ${W} is not supported.`)}},h=e?`
+    `},Xr=(e,t,r,n,o=!1)=>{let i=e[0].dims,a=e[1].dims,l=i.slice(0,-2),d=a.slice(0,-2),p=n?n.slice(0,-2):r.slice(0,-2),m=E.size(p),u=i[i.length-2],h=i[i.length-1],w=a[a.length-1],g=h%4===0&&w%4===0,y=u<=8?[4,1,1]:[4,4,1],x=[8,8,1],_=[Math.ceil(w/x[0]/y[0]),Math.ceil(u/x[1]/y[1]),Math.ceil(m/x[2]/y[2])],$=g?4:1,S=[...l,u,h/$],I=S.length,T=[...d,h,w/$],A=T.length,O=[m,u,w/$],z=[{type:6,data:u},{type:6,data:w},{type:6,data:h}];We(t,z),z.push(...U(p,S,T));let W=["rank","rank"],V=e.length>2;V&&(z.push(...U(e[2].dims)),W.push("rank")),z.push(...U(O));let j=Y=>{let ae=p.length,se=Wr("batchDims",e[0].dataType,ae,1),ie=pe(e[0].dataType),q=k("a",e[0].dataType,I,$),ce=k("b",e[1].dataType,A,$),ne=D("result",e[0].dataType,O.length,$),Z=[q,ce];if(V){let R=o?$:1;Z.push(k("bias",e[2].dataType,e[2].dims.length,R))}let ee=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];Le(t,ee);let K=pe(ne.type.tensor),me=Ne(t,ne.type.value,K),$e=vm($,V,me,[se,q,ce,ne],[l,d,p],o);return`
+  ${Y.registerUniforms(ee).registerInternalVariables(se).declareVariables(...Z,ne)}
+  ${$e}
+  ${g?Zt(y,x,ie,se):Qt(y,x,ie,se)}
+                   `};return{name:"MatMul",shaderCache:{hint:`${y};${t.activation};${g};${o}`,inputDependencies:W},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:_[0],y:_[1],z:_[2]},programUniforms:z}),getShaderSource:j}}});var $m,_u,xu=M(()=>{"use strict";X();st();re();ut();Xt();oo();Jt();$m=(e,t,r,n,o=!1,i,a=4,l=4,d=4,p="f32")=>{let m=V=>{switch(V){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 ${V} is not supported.`)}},u=V=>{switch(V){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 ${V} is not supported.`)}},h=e?`
     let coord = vec4<i32>(batch, xRow, xCol, xCh);
     `:`
     let coord = vec4<i32>(batch, xCh, xRow, xCol);
@@ -4626,7 +4625,7 @@ bIndices[${S-2}] = u32(row);
       row,
       col / outWidth,
       col % outWidth);
-    `,g=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",b=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",x=e?"row":"col",_=e?"col":"row",$=`
+    `,g=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",y=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",x=e?"row":"col",_=e?"col":"row",$=`
     let inChannels = i32(uniforms.w_shape[2]);
     let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
     let outRow = ${x} / outWidth;
@@ -4637,10 +4636,10 @@ bIndices[${S-2}] = u32(row);
     let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0];
     let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1];
     let xCh = ${_} % inChannels;
-    var resData = ${Ee(a,p)}(0.0);
+    var resData = ${Ce(a,p)}(0.0);
     // The bounds checking is always needed since we use it to pad zero for
     // the 'same' padding type.
-    if (xRow >= 0 && xRow < ${g} && xCol >= 0 && xCol < ${b}) {
+    if (xRow >= 0 && xRow < ${g} && xCol >= 0 && xCol < ${y}) {
       ${h}
       let xIndex = getIndexFromCoords4D(coord, vec4<i32>(uniforms.x_shape));
       ${m(a)}
@@ -4652,19 +4651,19 @@ bIndices[${S-2}] = u32(row);
     if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {
       ${$}
     }
-    return ${Ee(a,p)}(0.0);`:n&&r?`
+    return ${Ce(a,p)}(0.0);`:n&&r?`
     let col = colIn * ${a};
     ${$}`:`
     let col = colIn * ${a};
     if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {
       ${$}
     }
-    return ${Ee(a,p)}(0.0);`,I=`${u(l)}`,T=Ee(d,p),A=e?Ee(a,p):Ee(l,p),D=e?Ee(l,p):Ee(a,p),z=He(i,T,p);return`
+    return ${Ce(a,p)}(0.0);`,I=`${u(l)}`,T=Ce(d,p),A=e?Ce(a,p):Ce(l,p),O=e?Ce(l,p):Ce(a,p),z=Ne(i,T,p);return`
     fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${A} {
       ${e?S:I}
     }
 
-    fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${D} {
+    fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${O} {
       ${e?I:S}
     }
 
@@ -4675,57 +4674,57 @@ bIndices[${S-2}] = u32(row);
       var value = valueIn;
       let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
       ${w}
-      ${Zr(o)}
+      ${jr(o)}
       ${z}
       setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
       }
-    }`},Eu=(e,t,r,n,o,i,a,l)=>{let d=t.format==="NHWC",p=d?e[0].dims[3]:e[0].dims[1],m=r[0],u=d?r[2]:r[3],h=d?r[1]:r[2],w=d?r[3]:r[1],g=d&&(p%4===0||p%3===0)&&w%4===0,b=d?w:u*h,x=d?u*h:w,_=[8,8,1],$=n<=8?[4,1,1]:[4,4,1],S=[Math.ceil(b/_[0]/$[0]),Math.ceil(x/_[1]/$[1]),Math.ceil(m/_[2]/$[2])];ve("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${S}`);let I=g?d&&p%4!==0?3:4:1,T=_[1]*$[1],A=_[0]*$[0],D=Math.max(_[0]*I,_[1]),z=n%T===0,H=o%A===0,W=i%D===0,X=g?[I,4,4]:[1,1,1],Y=[{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}];Ge(t,Y),Y.push(...V(e[0].dims,e[1].dims));let pe=["rank","rank"];a&&(Y.push(...V(e[2].dims)),pe.push("rank")),Y.push(...V(r));let ue=le=>{let K=[{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}];Le(t,K);let de=g?4:1,re=ge(e[0].dataType),se=`
-      fn setOutputAtIndex(flatIndex : i32, value : ${g?`vec4<${re}>`:re}) {
-        result[flatIndex] = ${g?`vec4<${re}>`:re}(value);
+    }`},_u=(e,t,r,n,o,i,a,l)=>{let d=t.format==="NHWC",p=d?e[0].dims[3]:e[0].dims[1],m=r[0],u=d?r[2]:r[3],h=d?r[1]:r[2],w=d?r[3]:r[1],g=d&&(p%4===0||p%3===0)&&w%4===0,y=d?w:u*h,x=d?u*h:w,_=[8,8,1],$=n<=8?[4,1,1]:[4,4,1],S=[Math.ceil(y/_[0]/$[0]),Math.ceil(x/_[1]/$[1]),Math.ceil(m/_[2]/$[2])];be("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${S}`);let I=g?d&&p%4!==0?3:4:1,T=_[1]*$[1],A=_[0]*$[0],O=Math.max(_[0]*I,_[1]),z=n%T===0,W=o%A===0,V=i%O===0,j=g?[I,4,4]:[1,1,1],Y=[{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}];We(t,Y),Y.push(...U(e[0].dims,e[1].dims));let ae=["rank","rank"];a&&(Y.push(...U(e[2].dims)),ae.push("rank")),Y.push(...U(r));let se=ie=>{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}];Le(t,q);let ce=g?4:1,ne=pe(e[0].dataType),Z=`
+      fn setOutputAtIndex(flatIndex : i32, value : ${g?`vec4<${ne}>`:ne}) {
+        result[flatIndex] = ${g?`vec4<${ne}>`:ne}(value);
       }
-      fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${g?`vec4<${re}>`:re}) {
+      fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${g?`vec4<${ne}>`:ne}) {
         let flatIndex = getOutputIndexFromCoords(vec4<i32>(d0, d1, d2, d3));
         setOutputAtIndex(flatIndex ${g?"/ 4":""}, value);
-      }`,Z=E("x",e[0].dataType,e[0].dims.length,I===3?1:I),ne=E("w",e[1].dataType,e[1].dims.length,de),J=[Z,ne],ke=M("result",e[0].dataType,r.length,de);if(a){let R=E("bias",e[2].dataType,e[2].dims.length,de);J.push(R),se+=`
-        fn getBiasByOutputCoords(coords : vec4<i32>) -> ${g?`vec4<${re}>`:re} {
+      }`,ee=k("x",e[0].dataType,e[0].dims.length,I===3?1:I),K=k("w",e[1].dataType,e[1].dims.length,ce),me=[ee,K],$e=D("result",e[0].dataType,r.length,ce);if(a){let R=k("bias",e[2].dataType,e[2].dims.length,ce);me.push(R),Z+=`
+        fn getBiasByOutputCoords(coords : vec4<i32>) -> ${g?`vec4<${ne}>`:ne} {
           return bias[coords.${d?"w":"y"}${g?"/ 4":""}];
         }`}return`
-        ${Qr("uniforms.result_strides")}
+        ${Yr("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 };
-        ${le.registerUniforms(K).declareVariables(...J,ke)}
-        ${se}
-        ${Cm(d,z,H,W,a,t,X[0],X[1],X[2],re)}
-        ${g?er($,_,re,void 0,!d,D):tr($,_,re,void 0,!d,D,!1,void 0,l)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${I};${g};${z};${H};${W};${T};${A};${D}`,inputDependencies:pe},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:S[0],y:S[1],z:S[2]},programUniforms:Y}),getShaderSource:ue}}});var Tm,Pu,en,Am,Ou,Em,Du,zu,Bu=U(()=>{"use strict";Q();dt();ae();ie();lt();Jt();Tm=e=>{let t=1;for(let r=0;r<e.length;r++)t*=e[r];return t},Pu=e=>typeof e=="number"?[e,e,e]:e,en=(e,t)=>t<=1?e:e+(e-1)*(t-1),Am=(e,t,r,n=1)=>{let o=en(t,n);return Math.floor((e[0]*(r-1)-r+o)/2)},Ou=(e,t,r,n,o)=>{o==null&&(o=Am(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},Em=(e,t,r,n,o,i,a,l,d,p)=>{let m,u,h,w;if(e==="VALID"&&(e=0),typeof e=="number"){m={top:e,bottom:e,left:e,right:e,front:e,back:e};let g=Ou([t,r,n,1],[l,d,p],1,[o,i,a],e);u=g[0],h=g[1],w=g[2]}else if(Array.isArray(e)){if(!e.every((b,x,_)=>b===_[0]))throw Error(`Unsupported padding parameter: ${e}`);m={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let g=Ou([t,r,n,1],[l,d,p],1,[o,i,a],e[0]);u=g[0],h=g[1],w=g[2]}else if(e==="SAME_UPPER"){u=Math.ceil(t/o),h=Math.ceil(r/i),w=Math.ceil(n/a);let g=(u-1)*o+l-t,b=(h-1)*i+d-r,x=(w-1)*a+p-n,_=Math.floor(g/2),$=g-_,S=Math.floor(b/2),I=b-S,T=Math.floor(x/2),A=x-T;m={top:S,bottom:I,left:T,right:A,front:_,back:$}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:m,outDepth:u,outHeight:h,outWidth:w}},Du=(e,t,r,n,o,i=!1,a="channelsLast")=>{let l,d,p,m,u;if(a==="channelsLast")[l,d,p,m,u]=e;else if(a==="channelsFirst")[l,u,d,p,m]=e;else throw new Error(`Unknown dataFormat ${a}`);let[h,,w,g,b]=t,[x,_,$]=Pu(r),[S,I,T]=Pu(n),A=en(w,S),D=en(g,I),z=en(b,T),{padInfo:H,outDepth:W,outHeight:X,outWidth:Y}=Em(o,d,p,m,x,_,$,A,D,z),pe=i?h*u:h,ue=[0,0,0,0,0];return a==="channelsFirst"?ue=[l,pe,W,X,Y]:a==="channelsLast"&&(ue=[l,W,X,Y,pe]),{batchSize:l,dataFormat:a,inDepth:d,inHeight:p,inWidth:m,inChannels:u,outDepth:W,outHeight:X,outWidth:Y,outChannels:pe,padInfo:H,strideDepth:x,strideHeight:_,strideWidth:$,filterDepth:w,filterHeight:g,filterWidth:b,effectiveFilterDepth:A,effectiveFilterHeight:D,effectiveFilterWidth:z,dilationDepth:S,dilationHeight:I,dilationWidth:T,inShape:e,outShape:ue,filterShape:t}},zu=(e,t,r,n,o,i)=>{let a=i==="channelsLast",l=a?e[0].dims[3]:e[0].dims[1],d=!1,p=[64,1,1],m={x:r.map(($,S)=>S)},u=[Math.ceil(Tm(m.x.map($=>r[$]))/p[0]),1,1];ve("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${u}`);let h=d?a&&l%4!==0?3:4:1,w=k.size(r),g=[{type:12,data:w},{type:12,data:n},{type:12,data:o},{type:12,data:t.strides},{type:12,data:t.dilations}];Ge(t,g),g.push(...V(e[0].dims,e[1].dims));let b=["rank","rank"],x=e.length===3;x&&(g.push(...V(e[2].dims)),b.push("rank")),g.push(...V(r));let _=$=>{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}];Le(t,S);let I=d?4:1,T=ge(e[0].dataType),A=E("x",e[0].dataType,e[0].dims.length,h===3?1:h),D=E("W",e[1].dataType,e[1].dims.length,I),z=[A,D],H=M("result",e[0].dataType,r.length,I),W="";if(x){let pe=E("bias",e[2].dataType,e[2].dims.length,I);z.push(pe),W+=`
+        ${ie.registerUniforms(q).declareVariables(...me,$e)}
+        ${Z}
+        ${$m(d,z,W,V,a,t,j[0],j[1],j[2],ne)}
+        ${g?Zt($,_,ne,void 0,!d,O):Qt($,_,ne,void 0,!d,O,!1,void 0,l)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${I};${g};${z};${W};${V};${T};${A};${O}`,inputDependencies:ae},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:S[0],y:S[1],z:S[2]},programUniforms:Y}),getShaderSource:se}}});var _m,Su,Zr,xm,Iu,Sm,Cu,Tu,Au=M(()=>{"use strict";X();st();oe();re();ut();Xt();_m=e=>{let t=1;for(let r=0;r<e.length;r++)t*=e[r];return t},Su=e=>typeof e=="number"?[e,e,e]:e,Zr=(e,t)=>t<=1?e:e+(e-1)*(t-1),xm=(e,t,r,n=1)=>{let o=Zr(t,n);return Math.floor((e[0]*(r-1)-r+o)/2)},Iu=(e,t,r,n,o)=>{o==null&&(o=xm(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},Sm=(e,t,r,n,o,i,a,l,d,p)=>{let m,u,h,w;if(e==="VALID"&&(e=0),typeof e=="number"){m={top:e,bottom:e,left:e,right:e,front:e,back:e};let g=Iu([t,r,n,1],[l,d,p],1,[o,i,a],e);u=g[0],h=g[1],w=g[2]}else if(Array.isArray(e)){if(!e.every((y,x,_)=>y===_[0]))throw Error(`Unsupported padding parameter: ${e}`);m={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let g=Iu([t,r,n,1],[l,d,p],1,[o,i,a],e[0]);u=g[0],h=g[1],w=g[2]}else if(e==="SAME_UPPER"){u=Math.ceil(t/o),h=Math.ceil(r/i),w=Math.ceil(n/a);let g=(u-1)*o+l-t,y=(h-1)*i+d-r,x=(w-1)*a+p-n,_=Math.floor(g/2),$=g-_,S=Math.floor(y/2),I=y-S,T=Math.floor(x/2),A=x-T;m={top:S,bottom:I,left:T,right:A,front:_,back:$}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:m,outDepth:u,outHeight:h,outWidth:w}},Cu=(e,t,r,n,o,i=!1,a="channelsLast")=>{let l,d,p,m,u;if(a==="channelsLast")[l,d,p,m,u]=e;else if(a==="channelsFirst")[l,u,d,p,m]=e;else throw new Error(`Unknown dataFormat ${a}`);let[h,,w,g,y]=t,[x,_,$]=Su(r),[S,I,T]=Su(n),A=Zr(w,S),O=Zr(g,I),z=Zr(y,T),{padInfo:W,outDepth:V,outHeight:j,outWidth:Y}=Sm(o,d,p,m,x,_,$,A,O,z),ae=i?h*u:h,se=[0,0,0,0,0];return a==="channelsFirst"?se=[l,ae,V,j,Y]:a==="channelsLast"&&(se=[l,V,j,Y,ae]),{batchSize:l,dataFormat:a,inDepth:d,inHeight:p,inWidth:m,inChannels:u,outDepth:V,outHeight:j,outWidth:Y,outChannels:ae,padInfo:W,strideDepth:x,strideHeight:_,strideWidth:$,filterDepth:w,filterHeight:g,filterWidth:y,effectiveFilterDepth:A,effectiveFilterHeight:O,effectiveFilterWidth:z,dilationDepth:S,dilationHeight:I,dilationWidth:T,inShape:e,outShape:se,filterShape:t}},Tu=(e,t,r,n,o,i)=>{let a=i==="channelsLast",l=a?e[0].dims[3]:e[0].dims[1],d=!1,p=[64,1,1],m={x:r.map(($,S)=>S)},u=[Math.ceil(_m(m.x.map($=>r[$]))/p[0]),1,1];be("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${u}`);let h=d?a&&l%4!==0?3:4:1,w=E.size(r),g=[{type:12,data:w},{type:12,data:n},{type:12,data:o},{type:12,data:t.strides},{type:12,data:t.dilations}];We(t,g),g.push(...U(e[0].dims,e[1].dims));let y=["rank","rank"],x=e.length===3;x&&(g.push(...U(e[2].dims)),y.push("rank")),g.push(...U(r));let _=$=>{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}];Le(t,S);let I=d?4:1,T=pe(e[0].dataType),A=k("x",e[0].dataType,e[0].dims.length,h===3?1:h),O=k("W",e[1].dataType,e[1].dims.length,I),z=[A,O],W=D("result",e[0].dataType,r.length,I),V="";if(x){let ae=k("bias",e[2].dataType,e[2].dims.length,I);z.push(ae),V+=`
         fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${d?`vec4<${T}>`:T} {
-          return bias[${a?G("coords",4,5):G("coords",1,5)}${d?"/ 4":""}];
-        }`}let X=Ee(h,T),Y=He(t,X,T);return`
-            ${W}
+          return bias[${a?L("coords",4,5):L("coords",1,5)}${d?"/ 4":""}];
+        }`}let j=Ce(h,T),Y=Ne(t,j,T);return`
+            ${V}
             fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
               let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
               return ${A.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);
-              return ${D.getByIndices("aIndices")};
+              return ${O.getByIndices("aIndices")};
             }
-          ${$.registerUniforms(S).declareVariables(...z,H)}
+          ${$.registerUniforms(S).declareVariables(...z,W)}
           ${$.mainStart()}
           ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
-              let coords = ${H.offsetToIndices("global_idx")};
-              let batch = ${G("coords",0,A.rank)};
-              let d2 = ${a?G("coords",A.rank-1,A.rank):G("coords",1,A.rank)};
-              let xFRCCorner = vec3<u32>(${a?G("coords",1,A.rank):G("coords",2,A.rank)},
-              ${a?G("coords",2,A.rank):G("coords",3,A.rank)},
-              ${a?G("coords",3,A.rank):G("coords",4,A.rank)}) * uniforms.strides - uniforms.pads;
+              let coords = ${W.offsetToIndices("global_idx")};
+              let batch = ${L("coords",0,A.rank)};
+              let d2 = ${a?L("coords",A.rank-1,A.rank):L("coords",1,A.rank)};
+              let xFRCCorner = vec3<u32>(${a?L("coords",1,A.rank):L("coords",2,A.rank)},
+              ${a?L("coords",2,A.rank):L("coords",3,A.rank)},
+              ${a?L("coords",3,A.rank):L("coords",4,A.rank)}) * uniforms.strides - uniforms.pads;
               let xFCorner = xFRCCorner.x;
               let xRCorner = xFRCCorner.y;
               let xCCorner = xFRCCorner.z;
-              let xShapeY = ${a?G("uniforms.x_shape",1,A.rank):G("uniforms.x_shape",2,A.rank)};
-              let xShapeZ = ${a?G("uniforms.x_shape",2,A.rank):G("uniforms.x_shape",3,A.rank)};
-              let xShapeW = ${a?G("uniforms.x_shape",3,A.rank):G("uniforms.x_shape",4,A.rank)};
-              let xShapeU = ${a?G("uniforms.x_shape",4,A.rank):G("uniforms.x_shape",1,A.rank)};
+              let xShapeY = ${a?L("uniforms.x_shape",1,A.rank):L("uniforms.x_shape",2,A.rank)};
+              let xShapeZ = ${a?L("uniforms.x_shape",2,A.rank):L("uniforms.x_shape",3,A.rank)};
+              let xShapeW = ${a?L("uniforms.x_shape",3,A.rank):L("uniforms.x_shape",4,A.rank)};
+              let xShapeU = ${a?L("uniforms.x_shape",4,A.rank):L("uniforms.x_shape",1,A.rank)};
               let inputDepthNearestVec4 = (xShapeU / 4) * 4;
               let inputDepthVec4Remainder = xShapeU % 4;
 
@@ -4805,19 +4804,19 @@ bIndices[${S-2}] = u32(row);
               ${x?"value = value + getBiasByOutputCoords(coords)":""};
               ${Y}
               result[global_idx] = f32(value);
-          }`};return{name:"Conv3DNaive",shaderCache:{hint:`${t.cacheKey};${a};${h};${x}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:u[0],y:u[1],z:u[2]},programUniforms:g}),getShaderSource:_}}});var lo,Ru,Mu=U(()=>{"use strict";Q();ae();ie();co();lt();lo=(e,t,r)=>{let n=e.length>2,o=n?"value += b[output_channel];":"",i=e[0].dims,a=e[1].dims,l=a[0]/t.group,d=t.format==="NHWC",p=tn(i,a,t.dilations,t.pads,t.strides,d),m=k.size(p),u=[{type:12,data:m},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:l}];Ge(t,u),u.push(...V(i,a));let h=["rank","rank"];n&&(u.push(...V(e[2].dims)),h.push("rank")),u.push(...V(p));let w=g=>{let b=M("output",e[0].dataType,p.length),x=ge(b.type.tensor),_=He(t,b.type.value,x),$=E("x",e[0].dataType,i.length),S=E("w",e[1].dataType,a.length),I=[$,S];n&&I.push(E("b",e[2].dataType,e[2].dims.length));let T=[{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"}];return Le(t,T),`
-  ${g.registerUniforms(T).declareVariables(...I,b)}
+          }`};return{name:"Conv3DNaive",shaderCache:{hint:`${t.cacheKey};${a};${h};${x}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:u[0],y:u[1],z:u[2]},programUniforms:g}),getShaderSource:_}}});var io,ku,Eu=M(()=>{"use strict";X();oe();re();ao();ut();io=(e,t,r)=>{let n=e.length>2,o=n?"value += b[output_channel];":"",i=e[0].dims,a=e[1].dims,l=a[0]/t.group,d=t.format==="NHWC",p=Qr(i,a,t.dilations,t.pads,t.strides,d),m=E.size(p),u=[{type:12,data:m},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:l}];We(t,u),u.push(...U(i,a));let h=["rank","rank"];n&&(u.push(...U(e[2].dims)),h.push("rank")),u.push(...U(p));let w=g=>{let y=D("output",e[0].dataType,p.length),x=pe(y.type.tensor),_=Ne(t,y.type.value,x),$=k("x",e[0].dataType,i.length),S=k("w",e[1].dataType,a.length),I=[$,S];n&&I.push(k("b",e[2].dataType,e[2].dims.length));let T=[{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"}];return Le(t,T),`
+  ${g.registerUniforms(T).declareVariables(...I,y)}
 
   ${g.mainStart()}
     ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
 
-    let outputIndices = ${b.offsetToIndices("global_idx")};
+    let outputIndices = ${y.offsetToIndices("global_idx")};
     let batch: u32 = outputIndices[0];
     let output_channel: u32 = outputIndices[${d?3:1}];
     let xRCCorner: vec2<u32> = vec2<u32>(outputIndices[${d?1:2}], outputIndices[${d?2:3}]) * uniforms.strides - uniforms.pads;
     let group_id: u32 = output_channel / uniforms.output_channels_per_group;
 
-    var value: ${b.type.value} = ${b.type.value}(0);
+    var value: ${y.type.value} = ${y.type.value}(0);
     for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {
       let input_channel = group_id * uniforms.w_shape[1] + wInChannel;
       for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) {
@@ -4841,8 +4840,8 @@ bIndices[${S-2}] = u32(row);
     }
     ${o}
     ${_}
-    ${b.setByOffset("global_idx","value")}
-  }`};return{name:"GroupedConv",shaderCache:{hint:t.cacheKey,inputDependencies:h},getRunData:()=>({outputs:[{dims:r?r(p):p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:u}),getShaderSource:w}},Ru=(e,t,r)=>{let n=e.length>2,o=$e(r[3]),i=$e(r[2]),a=k.size(r)/o/i,l=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/o],d=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/o],p=[r[0],r[1],r[2],r[3]/o],m=[{type:12,data:a},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];Ge(t,m),m.push(...V(l,d,p));let u=(i-1)*t.strides[1]+d[1],h=w=>{let g=M("output",e[0].dataType,p.length,o),b=ge(g.type.tensor),x=He(t,g.type.value,b),_=E("x",e[0].dataType,l.length,o),$=E("w",e[1].dataType,d.length,o),S=[_,$];n&&S.push(E("b",e[2].dataType,e[2].dims,o));let I=n?"value += b[output_channel];":"",T=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Le(t,T),`
+    ${y.setByOffset("global_idx","value")}
+  }`};return{name:"GroupedConv",shaderCache:{hint:t.cacheKey,inputDependencies:h},getRunData:()=>({outputs:[{dims:r?r(p):p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:u}),getShaderSource:w}},ku=(e,t,r)=>{let n=e.length>2,o=he(r[3]),i=he(r[2]),a=E.size(r)/o/i,l=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/o],d=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/o],p=[r[0],r[1],r[2],r[3]/o],m=[{type:12,data:a},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];We(t,m),m.push(...U(l,d,p));let u=(i-1)*t.strides[1]+d[1],h=w=>{let g=D("output",e[0].dataType,p.length,o),y=pe(g.type.tensor),x=Ne(t,g.type.value,y),_=k("x",e[0].dataType,l.length,o),$=k("w",e[1].dataType,d.length,o),S=[_,$];n&&S.push(k("b",e[2].dataType,e[2].dims,o));let I=n?"value += b[output_channel];":"",T=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Le(t,T),`
   ${w.registerUniforms(T).declareVariables(...S,g)}
   ${w.mainStart()}
     ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -4887,14 +4886,14 @@ bIndices[${S-2}] = u32(row);
       ${x}
       ${g.set("batch","row","col + i","output_channel","value")};
     }
-  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${o};${i};${u};${d[0]};${d[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:m}),getShaderSource:h}}});var po,km,Uu,mo=U(()=>{"use strict";Q();ae();rr();ie();lt();po=(e,t,r,n,o=!1)=>{let i=e[0].dims,a=e[1].dims,l=i[i.length-2],d=a[a.length-1],p=i[i.length-1],m=$e(d),u=$e(p),h=$e(l),w=k.size(r)/m/h,g=e.length>2,b=n?n.slice(0,-2):r.slice(0,-2),_=[k.size(b),l,d],$=[{type:12,data:w},{type:12,data:l},{type:12,data:d},{type:12,data:p}];Ge(t,$),$.push(...V(b,i,a)),g&&$.push(...V(e[2].dims)),$.push(...V(_));let S=I=>{let T=Lr("batch_dims",e[0].dataType,b.length),A=E("a",e[0].dataType,i.length,u),D=E("b",e[1].dataType,a.length,m),z=M("output",e[0].dataType,_.length,m),H=ge(z.type.tensor),W=He(t,z.type.value,H),X=[A,D],Y="";if(g){let Z=o?m:1;X.push(E("bias",e[2].dataType,e[2].dims.length,Z)),Y=`${o?`value += bias[col / ${Z}];`:`value += ${z.type.value}(bias[row + i]);`}`}let pe=i.slice(0,-2),ue=a.slice(0,-2),le=Nt(pe,b),K=Nt(ue,b),de=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Le(t,de);let re=(Z,ne)=>{let J=Z.rank,ke=Z.name;if(J===2)return`var ${ke}_indices = ${Z.type.indices}(0u, 0u);`;let R=T.rank,F=`var ${ke}_indices: ${Z.type.indices};`;for(let ce=J-2-1,Re=R-1;ce>=0;ce--,Re--)F+=`
-${ke}_indices[${ce}] = ${R>1?`batch_indices[${Re}]`:"batch_indices"};`;return ne.forEach(ce=>{F+=`
-${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
-                     ${ke}_indices[${J-1}] = 0u;`,F},se=()=>{let Z=`var a_data: ${A.type.value};`;for(let ne=0;ne<u;ne++)Z+=`
-              let b_data${ne} = b[(b_offset + (k + ${ne}) * uniforms.N + col) / ${m}];`;for(let ne=0;ne<h;ne++){Z+=`a_data = a[(a_offset + (row + ${ne}) * uniforms.K + k) / ${u}];`;for(let J=0;J<u;J++)Z+=`
-            values[${ne}] = fma(${D.type.value}(a_data${u===1?"":`[${J}]`}), b_data${J}, values[${ne}]);
-`}return Z};return`
-  ${I.registerUniforms(de).registerInternalVariables(T).declareVariables(...X,z)}
+  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${o};${i};${u};${d[0]};${d[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:m}),getShaderSource:h}}});var so,Im,Pu,uo=M(()=>{"use strict";X();oe();Jt();re();ut();so=(e,t,r,n,o=!1)=>{let i=e[0].dims,a=e[1].dims,l=i[i.length-2],d=a[a.length-1],p=i[i.length-1],m=he(d),u=he(p),h=he(l),w=E.size(r)/m/h,g=e.length>2,y=n?n.slice(0,-2):r.slice(0,-2),_=[E.size(y),l,d],$=[{type:12,data:w},{type:12,data:l},{type:12,data:d},{type:12,data:p}];We(t,$),$.push(...U(y,i,a)),g&&$.push(...U(e[2].dims)),$.push(...U(_));let S=I=>{let T=Wr("batch_dims",e[0].dataType,y.length),A=k("a",e[0].dataType,i.length,u),O=k("b",e[1].dataType,a.length,m),z=D("output",e[0].dataType,_.length,m),W=pe(z.type.tensor),V=Ne(t,z.type.value,W),j=[A,O],Y="";if(g){let ee=o?m:1;j.push(k("bias",e[2].dataType,e[2].dims.length,ee)),Y=`${o?`value += bias[col / ${ee}];`:`value += ${z.type.value}(bias[row + i]);`}`}let ae=i.slice(0,-2),se=a.slice(0,-2),ie=Mt(ae,y),q=Mt(se,y),ce=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Le(t,ce);let ne=(ee,K)=>{let me=ee.rank,$e=ee.name;if(me===2)return`var ${$e}_indices = ${ee.type.indices}(0u, 0u);`;let R=T.rank,G=`var ${$e}_indices: ${ee.type.indices};`;for(let ue=me-2-1,Be=R-1;ue>=0;ue--,Be--)G+=`
+${$e}_indices[${ue}] = ${R>1?`batch_indices[${Be}]`:"batch_indices"};`;return K.forEach(ue=>{G+=`
+${$e}_indices[${ue}] = 0;`}),G+=`${$e}_indices[${me-2}] = 0u;
+                     ${$e}_indices[${me-1}] = 0u;`,G},Z=()=>{let ee=`var a_data: ${A.type.value};`;for(let K=0;K<u;K++)ee+=`
+              let b_data${K} = b[(b_offset + (k + ${K}) * uniforms.N + col) / ${m}];`;for(let K=0;K<h;K++){ee+=`a_data = a[(a_offset + (row + ${K}) * uniforms.K + k) / ${u}];`;for(let me=0;me<u;me++)ee+=`
+            values[${K}] = fma(${O.type.value}(a_data${u===1?"":`[${me}]`}), b_data${me}, values[${K}]);
+`}return ee};return`
+  ${I.registerUniforms(ce).registerInternalVariables(T).declareVariables(...j,z)}
   ${I.mainStart()}
     ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
     let col = (global_idx % (uniforms.N / ${m})) * ${m};
@@ -4904,24 +4903,24 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
     let batch = index1 / stride1;
 
     ${r.length===2?"":`let batch_indices = ${T.offsetToIndices("batch")};`}
-    ${re(A,le)}
+    ${ne(A,ie)}
     let a_offset = ${A.indicesToOffset("a_indices")};
-    ${re(D,K)}
-    let b_offset = ${D.indicesToOffset("b_indices")};
+    ${ne(O,q)}
+    let b_offset = ${O.indicesToOffset("b_indices")};
     var values: array<${z.type.value}, ${h}>;
     for (var k: u32 = 0u; k < uniforms.K; k = k + ${u}) {
-      ${se()}
+      ${Z()}
     }
     for (var i = 0u; i < ${h}u; i++) {
       var value = values[i];
       ${Y}
-      ${W}
+      ${V}
       let cur_indices = ${z.type.indices}(batch, row + i, col);
       let offset = ${z.indicesToOffset("cur_indices")};
       ${z.setByOffset(`offset / ${m}`,"value")};
     }
   }
-  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${m};${u};${h};${o}`,inputDependencies:g?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:$}),getShaderSource:S}},km=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.")},Uu=e=>{km(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];r<8&&n<8?e.compute(po(e.inputs,{activation:""},t)):e.compute(Jr(e.inputs,{activation:""},t))}});var tn,fo,Pm,ho,go,Om,Dm,zm,yo,co=U(()=>{"use strict";ae();ku();Bu();rr();Mu();lt();mo();Et();tn=(e,t,r,n,o,i)=>{let a=e[0],l=e.slice(i?1:2,i?3:4),d=l.length,p=t[0],u=t.slice(2).map((g,b)=>g+(g-1)*(r[b]-1)),w=l.map((g,b)=>g+n[b]+n[b+d]).map((g,b)=>Math.floor((g-u[b]+o[b])/o[b]));return w.splice(0,0,a),w.splice(i?3:1,0,p),w},fo=[2,3,1,0],Pm=(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 ${o*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},ho=(e,t)=>{let r=e.kernelShape.slice();for(let i=2;i<t[1].dims.length;++i)r[i-2]===0&&(r[i-2]=t[1].dims[i]);let n=e.pads.slice();Ct.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,n,e.format==="NHWC",e.autoPad);let o=Object.assign({},e);return Object.assign(o,{kernelShape:r,pads:n}),o},go=e=>{let t=Xr(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],o=e.dilations,i=e.group,a=e.kernel_shape,l=e.pads,d=e.strides,p=e.w_is_const();return{autoPad:n,format:r,dilations:o,group:i,kernelShape:a,pads:l,strides:d,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Om=(e,t,r)=>{let n=ho(r,t),o=r.format==="NHWC";if(r.group!==1){if(!e.adapterInfo.isArchitecture("ampere")&&o&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let D=tn(t[0].dims,t[1].dims,r.dilations,n.pads,r.strides,o),z=e.kernelCustomData.wT??e.compute(Ue(t[1],fo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=z);let H=[t[0],z];t.length===3&&H.push(t[2]),e.compute(Ru(H,n,D),{inputs:H})}else e.compute(lo(t,n));return}let i=t.length===3,a=t[0].dims[o?1:2],l=t[0].dims[o?2:3],d=t[0].dims[o?3:1],p=t[1].dims[2],m=t[1].dims[3],u=tn(t[0].dims,t[1].dims,r.dilations,n.pads,r.strides,o),h=u[o?1:2],w=u[o?2:3],g=u[o?3:1],b=o&&p===a&&m===l&&r.pads[0]===0&&r.pads[1]===0;if(b||p===1&&m===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 A=u[0],D,z,H,W=[];if(o){let pe=e.kernelCustomData.wT??e.compute(Ue(t[1],fo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=pe),b){let ue=a*l*d;D=t[0].reshape([1,A,ue]),z=pe.reshape([1,ue,g]),H=[1,A,g]}else D=t[0].reshape([A,a*l,d]),z=pe.reshape([1,d,g]),H=[A,h*w,g];W.push(D),W.push(z)}else D=t[0].reshape([A,d,a*l]),z=t[1].reshape([1,g,d]),H=[A,g,h*w],W.push(z),W.push(D);i&&W.push(t[2]);let X=H[2],Y=W[0].dims[W[0].dims.length-1];X<8&&Y<8?e.compute(po(W,n,u,H,o),{inputs:W}):e.compute(Jr(W,n,u,H,o),{inputs:W});return}let x=!0,_=e.kernelCustomData.wT??e.compute(Ue(t[1],fo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=_);let $=[t[0],_];i&&$.push(t[2]);let S=o?h*w:g,I=o?g:h*w,T=p*m*d;e.compute(Eu($,n,u,S,I,T,i,x),{inputs:$})},Dm=(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),l=[1].concat(t.kernelShape),d=ho({...t,pads:o,strides:i,dilations:a,kernelShape:l},n);e.compute(lo(n,d,p=>r?[p[0],p[2],p[3]]:[]))},zm=(e,t,r)=>{let n=r.format==="NHWC"?"channelsLast":"channelsFirst",o=ho(r,t),i=r.autoPad==="NOTSET"?r.pads:r.autoPad,a=Du(t[0].dims,t[1].dims,r.strides,r.dilations,i,!1,n);e.compute(zu(t,o,a.outShape,[a.filterDepth,a.filterHeight,a.filterWidth],[a.padInfo.front,a.padInfo.top,a.padInfo.left],n))},yo=(e,t)=>{Pm(e.inputs,t),e.inputs[0].dims.length===3?Dm(e,t):e.inputs[0].dims.length===5?zm(e,e.inputs,t):Om(e,e.inputs,t)}});var Bm,Vu,Nu=U(()=>{"use strict";Q();dt();ie();lt();Jt();uo();rr();Bm=(e,t=!1,r,n,o=4)=>{let i=_=>{switch(_){case 1:return"return w[getIndexFromCoords4D(coord, vec4<i32>(uniforms.w_shape))];";case 4:return`
+  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${m};${u};${h};${o}`,inputDependencies:g?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:$}),getShaderSource:S}},Im=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.")},Pu=e=>{Im(e.inputs);let t=Je.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];r<8&&n<8?e.compute(so(e.inputs,{activation:""},t)):e.compute(Xr(e.inputs,{activation:""},t))}});var Qr,lo,Cm,co,po,Tm,Am,km,mo,ao=M(()=>{"use strict";oe();xu();Au();Jt();Eu();ut();uo();Ct();Qr=(e,t,r,n,o,i)=>{let a=e[0],l=e.slice(i?1:2,i?3:4),d=l.length,p=t[0],u=t.slice(2).map((g,y)=>g+(g-1)*(r[y]-1)),w=l.map((g,y)=>g+n[y]+n[y+d]).map((g,y)=>Math.floor((g-u[y]+o[y])/o[y]));return w.splice(0,0,a),w.splice(i?3:1,0,p),w},lo=[2,3,1,0],Cm=(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 ${o*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},co=(e,t)=>{let r=e.kernelShape.slice();for(let i=2;i<t[1].dims.length;++i)r[i-2]===0&&(r[i-2]=t[1].dims[i]);let n=e.pads.slice();xt.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,n,e.format==="NHWC",e.autoPad);let o=Object.assign({},e);return Object.assign(o,{kernelShape:r,pads:n}),o},po=e=>{let t=Kr(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],o=e.dilations,i=e.group,a=e.kernel_shape,l=e.pads,d=e.strides,p=e.w_is_const();return{autoPad:n,format:r,dilations:o,group:i,kernelShape:a,pads:l,strides:d,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Tm=(e,t,r)=>{let n=co(r,t),o=r.format==="NHWC";if(r.group!==1){if(!e.adapterInfo.isArchitecture("ampere")&&o&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let O=Qr(t[0].dims,t[1].dims,r.dilations,n.pads,r.strides,o),z=e.kernelCustomData.wT??e.compute(Re(t[1],lo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=z);let W=[t[0],z];t.length===3&&W.push(t[2]),e.compute(ku(W,n,O),{inputs:W})}else e.compute(io(t,n));return}let i=t.length===3,a=t[0].dims[o?1:2],l=t[0].dims[o?2:3],d=t[0].dims[o?3:1],p=t[1].dims[2],m=t[1].dims[3],u=Qr(t[0].dims,t[1].dims,r.dilations,n.pads,r.strides,o),h=u[o?1:2],w=u[o?2:3],g=u[o?3:1],y=o&&p===a&&m===l&&r.pads[0]===0&&r.pads[1]===0;if(y||p===1&&m===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 A=u[0],O,z,W,V=[];if(o){let ae=e.kernelCustomData.wT??e.compute(Re(t[1],lo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=ae),y){let se=a*l*d;O=t[0].reshape([1,A,se]),z=ae.reshape([1,se,g]),W=[1,A,g]}else O=t[0].reshape([A,a*l,d]),z=ae.reshape([1,d,g]),W=[A,h*w,g];V.push(O),V.push(z)}else O=t[0].reshape([A,d,a*l]),z=t[1].reshape([1,g,d]),W=[A,g,h*w],V.push(z),V.push(O);i&&V.push(t[2]);let j=W[2],Y=V[0].dims[V[0].dims.length-1];j<8&&Y<8?e.compute(so(V,n,u,W,o),{inputs:V}):e.compute(Xr(V,n,u,W,o),{inputs:V});return}let x=!0,_=e.kernelCustomData.wT??e.compute(Re(t[1],lo),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=_);let $=[t[0],_];i&&$.push(t[2]);let S=o?h*w:g,I=o?g:h*w,T=p*m*d;e.compute(_u($,n,u,S,I,T,i,x),{inputs:$})},Am=(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),l=[1].concat(t.kernelShape),d=co({...t,pads:o,strides:i,dilations:a,kernelShape:l},n);e.compute(io(n,d,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]]))},km=(e,t,r)=>{let n=r.format==="NHWC"?"channelsLast":"channelsFirst",o=co(r,t),i=r.autoPad==="NOTSET"?r.pads:r.autoPad,a=Cu(t[0].dims,t[1].dims,r.strides,r.dilations,i,!1,n);e.compute(Tu(t,o,a.outShape,[a.filterDepth,a.filterHeight,a.filterWidth],[a.padInfo.front,a.padInfo.top,a.padInfo.left],n))},mo=(e,t)=>{Cm(e.inputs,t),e.inputs[0].dims.length===3?Am(e,t):e.inputs[0].dims.length===5?km(e,e.inputs,t):Tm(e,e.inputs,t)}});var Em,Ou,zu=M(()=>{"use strict";X();st();re();ut();Xt();oo();Jt();Em=(e,t=!1,r,n,o=4)=>{let i=_=>{switch(_){case 1:return"return w[getIndexFromCoords4D(coord, vec4<i32>(uniforms.w_shape))];";case 4:return`
             let coord1 = vec4<i32>(coordX, coordY, col + 1, rowInner);
             let coord2 = vec4<i32>(coordX, coordY, col + 2, rowInner);
             let coord3 = vec4<i32>(coordX, coordY, col + 3, rowInner);
@@ -4987,7 +4986,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
         ${i(o)}
       }
       return ${n}(0.0);
-      `,b=He(r,n);return`
+      `,y=Ne(r,n);return`
   fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${n} {
     ${e?w:g}
   }
@@ -5002,25 +5001,25 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
       var value = valueInput;
       let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
       ${l}
-      ${Zr(t)}
-      ${b}
+      ${jr(t)}
+      ${y}
       result[getIndexFromCoords4D(coords, vec4<i32>(uniforms.result_shape))/${o}] = value;
     }
-  }`},Vu=(e,t,r,n,o,i,a,l)=>{let d=t.format==="NHWC",p=d?e[0].dims[3]:e[0].dims[1],m=r[0],u=d?r[2]:r[3],h=d?r[1]:r[2],w=d?r[3]:r[1],g=d&&p%4===0&&p%3&&w%4===0,b=d?w:u*h,x=d?u*h:w,_=[8,8,1],$=n<=8?[4,1,1]:[4,4,1],S=[Math.ceil(b/_[0]/$[0]),Math.ceil(x/_[1]/$[1]),Math.ceil(m/_[2]/$[2])];ve("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${S}`);let I=g?4:1,T=Math.max(_[0]*I,_[1]),A=g?4:1,D=[t.kernelShape[d?1:2],t.kernelShape[d?2:3]],z=[D[0]+(t.dilations[0]<=1?0:(D[0]-1)*(t.dilations[0]-1)),D[1]+(t.dilations[1]<=1?0:(D[1]-1)*(t.dilations[1]-1))],H=[z[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),z[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],W=[{type:6,data:n},{type:6,data:o},{type:6,data:i},{type:6,data:t.strides},{type:6,data:t.dilations},{type:6,data:D},{type:6,data:H}];Ge(t,W),W.push(...V(e[0].dims,e[1].dims));let X=["rank","rank"];a&&(W.push(...V(e[2].dims)),X.push("rank")),W.push(...V(r));let Y=pe=>{let ue=E("x",e[0].dataType,e[0].dims.length,A),le=E("w",e[1].dataType,e[1].dims.length,1),K=M("result",e[0].dataType,r.length,A),de=[ue,le],re="";if(a){let ne=E("bias",e[2].dataType,e[2].dims.length,A);de.push(ne),re+=`
-          fn getBiasByOutputCoords(coords : vec4<i32>) -> ${ne.type.value} {
+  }`},Ou=(e,t,r,n,o,i,a,l)=>{let d=t.format==="NHWC",p=d?e[0].dims[3]:e[0].dims[1],m=r[0],u=d?r[2]:r[3],h=d?r[1]:r[2],w=d?r[3]:r[1],g=d&&p%4===0&&p%3&&w%4===0,y=d?w:u*h,x=d?u*h:w,_=[8,8,1],$=n<=8?[4,1,1]:[4,4,1],S=[Math.ceil(y/_[0]/$[0]),Math.ceil(x/_[1]/$[1]),Math.ceil(m/_[2]/$[2])];be("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${S}`);let I=g?4:1,T=Math.max(_[0]*I,_[1]),A=g?4:1,O=[t.kernelShape[d?1:2],t.kernelShape[d?2:3]],z=[O[0]+(t.dilations[0]<=1?0:(O[0]-1)*(t.dilations[0]-1)),O[1]+(t.dilations[1]<=1?0:(O[1]-1)*(t.dilations[1]-1))],W=[z[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),z[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],V=[{type:6,data:n},{type:6,data:o},{type:6,data:i},{type:6,data:t.strides},{type:6,data:t.dilations},{type:6,data:O},{type:6,data:W}];We(t,V),V.push(...U(e[0].dims,e[1].dims));let j=["rank","rank"];a&&(V.push(...U(e[2].dims)),j.push("rank")),V.push(...U(r));let Y=ae=>{let se=k("x",e[0].dataType,e[0].dims.length,A),ie=k("w",e[1].dataType,e[1].dims.length,1),q=D("result",e[0].dataType,r.length,A),ce=[se,ie],ne="";if(a){let K=k("bias",e[2].dataType,e[2].dims.length,A);ce.push(K),ne+=`
+          fn getBiasByOutputCoords(coords : vec4<i32>) -> ${K.type.value} {
             return bias[coords.${d?"w":"y"}${g?"/ 4":""}];
-          }`}let se=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"strides",type:"i32",length:2},{name:"dilations",type:"i32",length:2},{name:"filter_dims",type:"i32",length:D.length},{name:"pads",type:"i32",length:H.length}];Le(t,se);let Z=ge(e[0].dataType,1);if(Z!=="f16"&&Z!=="f32")throw new Error(`elemType ${Z} is not supported.`);return`
-        ${Qr("uniforms.result_strides")}
-        ${pe.registerUniforms(se).declareVariables(...de,K)};
-        ${re}
-        ${Bm(d,a,t,ue.type.value,I)}
-        ${g?er($,_,Z,void 0,!d,T):tr($,_,Z,void 0,!d,T,!1,void 0,l)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${t.cacheKey};${$};${_};${g}`,inputDependencies:X},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:S[0],y:S[1],z:S[2]},programUniforms:W}),getShaderSource:Y}}});var Rm,bo,Wu=U(()=>{"use strict";Q();dt();ae();ie();Rm=(e,t,r,n,o,i=!1,a,l,d=!1)=>{let p=d?1:2,m=d?2:3,u=d?3:1,h=i?2:1,w=`
+          }`}let Z=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"strides",type:"i32",length:2},{name:"dilations",type:"i32",length:2},{name:"filter_dims",type:"i32",length:O.length},{name:"pads",type:"i32",length:W.length}];Le(t,Z);let ee=pe(e[0].dataType,1);if(ee!=="f16"&&ee!=="f32")throw new Error(`elemType ${ee} is not supported.`);return`
+        ${Yr("uniforms.result_strides")}
+        ${ae.registerUniforms(Z).declareVariables(...ce,q)};
+        ${ne}
+        ${Em(d,a,t,se.type.value,I)}
+        ${g?Zt($,_,ee,void 0,!d,T):Qt($,_,ee,void 0,!d,T,!1,void 0,l)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${t.cacheKey};${$};${_};${g}`,inputDependencies:j},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:S[0],y:S[1],z:S[2]},programUniforms:V}),getShaderSource:Y}}});var Pm,fo,Bu=M(()=>{"use strict";X();st();oe();re();Pm=(e,t,r,n,o,i=!1,a,l,d=!1)=>{let p=d?1:2,m=d?2:3,u=d?3:1,h=i?2:1,w=`
   fn setOutputAtIndex(flatIndex : u32, value : ${i?`vec4<${a}>`:a}) {
     result[flatIndex] = ${i?`vec4<${a}>`:a}(value);
   }`;n&&(w+=`
     fn getBiasByOutputCoords(coords : vec4<u32>) -> ${i?`vec4<${a}>`:a} {
       return bias[coords.${d?"w":"y"}${i?"/ 4":""}];
-    }`);let g=i?4:1,b=E("W",t[1].dataType,t[1].dims.length,g),x=E("Dy",t[0].dataType,t[0].dims.length,g),_=[x,b];n&&_.push(E("bias",t[2].dataType,[r[u]].length,g));let $=M("result",t[0].dataType,r.length,g),S=`{
+    }`);let g=i?4:1,y=k("W",t[1].dataType,t[1].dims.length,g),x=k("Dy",t[0].dataType,t[0].dims.length,g),_=[x,y];n&&_.push(k("bias",t[2].dataType,[r[u]].length,g));let $=D("result",t[0].dataType,r.length,g),S=`{
         let batch: u32 = ${o?"global_id.z":"workgroup_id.z"} / uniforms.result_shape[1];
         let r = ${o?"global_id.z":"workgroup_id.z"} % uniforms.result_shape[1];
         let c = ${o?"global_id.y":"workgroup_id.y"} * ${h};
@@ -5066,10 +5065,10 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             if (bDyCVal && bDyCVal2) {
               let d2Length = uniforms.Dy_shape[3];
               for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) {
-                let wValue0 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
-                let wValue1 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
-                let wValue2 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
-                let wValue3 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};
+                let wValue0 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
+                let wValue1 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
+                let wValue2 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
+                let wValue3 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};
 
                 var xValue = ${x.get("batch","idyR","idyC","d2")};
                 let tmpval = vec4<${a}>(dot(xValue, wValue0),
@@ -5088,10 +5087,10 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             } else if (bDyCVal) {
               let d2Length = uniforms.Dy_shape[${u}];
               for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {
-                let wValue0 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
-                let wValue1 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
-                let wValue2 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
-                let wValue3 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};
+                let wValue0 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
+                let wValue1 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
+                let wValue2 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
+                let wValue3 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};
 
                 var xValue = ${x.get("batch","idyR","idyC","d2")};
                 let tmpval = vec4<${a}>(dot(xValue, wValue0),
@@ -5103,10 +5102,10 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             } else if (bDyCVal2) {
               let d2Length = uniforms.Dy_shape[3];
               for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {
-                let wValue0 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
-                let wValue1 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
-                let wValue2 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
-                let wValue3 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};
+                let wValue0 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};
+                let wValue1 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};
+                let wValue2 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};
+                let wValue3 = ${y.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};
 
                 var xValue = ${x.get("batch","idyR","idyC2","d2")};
                 let tmpval = vec4<${a}>(dot(xValue, wValue0),
@@ -5163,7 +5162,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
               var inputChannel = groupId * uniforms.input_channels_per_group;
               for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + 1) {
                 let xValue = ${d?x.get("batch","idyR","idyC","inputChannel"):x.get("batch","inputChannel","idyR","idyC")};
-                let wValue = ${b.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")};
+                let wValue = ${y.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")};
                 dotProd = dotProd + xValue * wValue;
                 inputChannel = inputChannel + 1;
               }
@@ -5177,35 +5176,35 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
 
     ${e.mainStart()}
     ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
-  ${i?S:I}}`},bo=(e,t,r)=>{let n=e.length>2,o=t.outputShape,i=k.size(o),a=[Math.ceil(i/64),1,1];ve("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${a}`);let l=t.format==="NHWC",d=["rank","rank"],p=[t.strides[0],t.strides[1]],m=[t.kernelShape[l?1:2],t.kernelShape[l?2:3]],u=[t.dilations[0],t.dilations[1]],h=[m[0]+(t.dilations[0]<=1?0:(t.kernelShape[l?1:2]-1)*(t.dilations[0]-1)),m[1]+(t.dilations[1]<=1?0:(t.kernelShape[l?2:3]-1)*(t.dilations[1]-1))],w=[h[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),h[1]-1-Math.floor(t.pads[1]+t.pads[3])/2],g=!1,b=t.group,x=e[1].dims,_=x[0]/b,$=x[1],S=[{type:12,data:i},{type:12,data:p},{type:12,data:m},{type:12,data:u},{type:12,data:h},{type:6,data:w},{type:12,data:_},{type:12,data:$},...V(e[0].dims,e[1].dims)];n&&(S.push(...V(e[2].dims)),d.push("rank")),S.push(...V(o));let I=a[1]===1&&a[2]===1,T=A=>{let D=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:p.length},{name:"filter_dims",type:"u32",length:m.length},{name:"dilations",type:"u32",length:m.length},{name:"effective_filter_dims",type:"u32",length:h.length},{name:"pads",type:"i32",length:w.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],z=ge(e[0].dataType);return`${Rm(A,e,o,n,I,g,z,D,l)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};`,inputDependencies:d},getRunData:()=>({dispatchGroup:{x:a[0],y:a[1],z:a[2]},outputs:[{dims:r?r(o):o,dataType:e[0].dataType}],programUniforms:S}),getShaderSource:T}}});var Mm,Um,Vm,Hu,Gu,Nm,Wm,Hm,Gm,Lu,Fu=U(()=>{"use strict";Nu();Wu();lt();Et();Mm=(e,t,r,n,o,i)=>(e-1)*t+r+(n-1)*o+1-i,Um=(e,t,r,n,o)=>{let i=Math.floor(e/2);t==="SAME_UPPER"?(r[n]=i,r[o]=e-i):t==="SAME_LOWER"&&(r[n]=e-i,r[o]=i)},Vm=(e,t,r,n,o,i,a,l,d,p)=>{let m=e.length-2,u=p.length===0;if(d.length===0)for(let g=0;g<m;++g)d.push(0);let h=e[0],w=t[l?3:1]*o;for(let g=0,b=e.length-m-(l?1:0);g<m;++g,++b){let x=e[b],_=u?x*a[g]:p[g],$=Mm(x,a[g],i[g],t[b],r[g],_);Um($,n,i,g,g+m),u&&p.push(a[g]*(x-1)+d[g]+(t[b]-1)*r[g]+1-i[g]-i[g+m])}p.splice(0,0,h),p.splice(l?3:1,0,w)},Hu=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((u,h)=>u*h,1)===0){r.length=0;for(let u=2;u<t[1].dims.length;++u)r.push(t[1].dims[u])}let n=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(n?3:1,0,t[1].dims[1]);let o=e.pads.slice(),i=e.outputShape.slice(),a=e.outputPadding.slice(),l=t[0].dims,d=e.dilations.slice();if(d.reduce((u,h)=>u+h,0)===0){let u=t[0].dims.length-2;d=new Array(u).fill(1)}let p=e.strides.slice();if(p.reduce((u,h)=>u+h,0)===0){let u=t[0].dims.length-2;p=new Array(u).fill(1)}Vm(l,r,d,e.autoPad,e.group,o,p,n,a,i);let m=Object.assign({},e);return Object.assign(m,{kernelShape:r,pads:o,outputPadding:a,outputShape:i,dilations:d,strides:p}),m},Gu=e=>{let t=Xr(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],o=e.dilations,i=e.group,a=e.kernelShape,l=e.pads,d=e.strides,p=e.wIsConst(),m=e.outputPadding,u=e.outputShape;return{autoPad:n,format:r,dilations:o,group:i,kernelShape:a,outputPadding:m,outputShape:u,pads:l,strides:d,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Nm=(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 ${i}D`);if(t.kernelShape.reduce((m,u)=>m+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Wm=[2,3,1,0],Hm=(e,t,r)=>{let n=Hu(r,t),o=r.format==="NHWC",i=n.outputShape,a=i[o?3:1],l=t[0].dims[o?3:1];if(n.group!==1||a===1&&l===1){e.compute(bo(t,n));return}let d=i[o?1:2],p=i[o?2:3],m=t[1].dims[2],u=t[1].dims[3],h=o?d*p:a,w=o?a:d*p,g=m*u*l,b=!0,x=e.kernelCustomData.wT??e.compute(Ue(t[1],Wm),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=x);let _=[t[0],x],$=t.length===3;$&&(!o&&t[2].dims.length===1?_.push(t[2].reshape([t[2].dims[0],1,1])):_.push(t[2])),e.compute(Vu(_,n,i,h,w,g,$,b),{inputs:_})},Gm=(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 l=t.pads;l.length===0&&(l=[0,0]),l=[0,l[0],0,l[1]],a=[1].concat(a),i=[1].concat(i),o=[1].concat(o);let d=Hu({...t,pads:l,strides:a,dilations:i,kernelShape:o},n);e.compute(bo(n,d,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]]))},Lu=(e,t)=>{Nm(e.inputs,t),e.inputs[0].dims.length===3?Gm(e,t):Hm(e,e.inputs,t)}});var Lm,qu,ju,Ku=U(()=>{"use strict";Q();ae();Ce();ie();Lm=(e,t,r,n)=>{let o=k.size(t),i=t.length,a=E("input",e,i),l=M("output",e,i),d=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),p=k.normalizeAxis(d,i),m=u=>{let h=` i32(${a.indicesGet("inputIndices","uniforms.axis")}) `,w=G("uniforms.input_shape","uniforms.axis",i),g=n.reverse?h+(n.exclusive?" + 1":""):"0",b=n.reverse?w:h+(n.exclusive?"":" + 1");return`
+  ${i?S:I}}`},fo=(e,t,r)=>{let n=e.length>2,o=t.outputShape,i=E.size(o),a=[Math.ceil(i/64),1,1];be("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${a}`);let l=t.format==="NHWC",d=["rank","rank"],p=[t.strides[0],t.strides[1]],m=[t.kernelShape[l?1:2],t.kernelShape[l?2:3]],u=[t.dilations[0],t.dilations[1]],h=[m[0]+(t.dilations[0]<=1?0:(t.kernelShape[l?1:2]-1)*(t.dilations[0]-1)),m[1]+(t.dilations[1]<=1?0:(t.kernelShape[l?2:3]-1)*(t.dilations[1]-1))],w=[h[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),h[1]-1-Math.floor(t.pads[1]+t.pads[3])/2],g=!1,y=t.group,x=e[1].dims,_=x[0]/y,$=x[1],S=[{type:12,data:i},{type:12,data:p},{type:12,data:m},{type:12,data:u},{type:12,data:h},{type:6,data:w},{type:12,data:_},{type:12,data:$},...U(e[0].dims,e[1].dims)];n&&(S.push(...U(e[2].dims)),d.push("rank")),S.push(...U(o));let I=a[1]===1&&a[2]===1,T=A=>{let O=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:p.length},{name:"filter_dims",type:"u32",length:m.length},{name:"dilations",type:"u32",length:m.length},{name:"effective_filter_dims",type:"u32",length:h.length},{name:"pads",type:"i32",length:w.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],z=pe(e[0].dataType);return`${Pm(A,e,o,n,I,g,z,O,l)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};`,inputDependencies:d},getRunData:()=>({dispatchGroup:{x:a[0],y:a[1],z:a[2]},outputs:[{dims:r?r(o):o,dataType:e[0].dataType}],programUniforms:S}),getShaderSource:T}}});var Om,zm,Bm,Du,Ru,Dm,Rm,Mm,Um,Mu,Uu=M(()=>{"use strict";zu();Bu();ut();Ct();Om=(e,t,r,n,o,i)=>(e-1)*t+r+(n-1)*o+1-i,zm=(e,t,r,n,o)=>{let i=Math.floor(e/2);t==="SAME_UPPER"?(r[n]=i,r[o]=e-i):t==="SAME_LOWER"&&(r[n]=e-i,r[o]=i)},Bm=(e,t,r,n,o,i,a,l,d,p)=>{let m=e.length-2,u=p.length===0;if(d.length===0)for(let g=0;g<m;++g)d.push(0);let h=e[0],w=t[l?3:1]*o;for(let g=0,y=e.length-m-(l?1:0);g<m;++g,++y){let x=e[y],_=u?x*a[g]:p[g],$=Om(x,a[g],i[g],t[y],r[g],_);zm($,n,i,g,g+m),u&&p.push(a[g]*(x-1)+d[g]+(t[y]-1)*r[g]+1-i[g]-i[g+m])}p.splice(0,0,h),p.splice(l?3:1,0,w)},Du=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((u,h)=>u*h,1)===0){r.length=0;for(let u=2;u<t[1].dims.length;++u)r.push(t[1].dims[u])}let n=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(n?3:1,0,t[1].dims[1]);let o=e.pads.slice(),i=e.outputShape.slice(),a=e.outputPadding.slice(),l=t[0].dims,d=e.dilations.slice();if(d.reduce((u,h)=>u+h,0)===0){let u=t[0].dims.length-2;d=new Array(u).fill(1)}let p=e.strides.slice();if(p.reduce((u,h)=>u+h,0)===0){let u=t[0].dims.length-2;p=new Array(u).fill(1)}Bm(l,r,d,e.autoPad,e.group,o,p,n,a,i);let m=Object.assign({},e);return Object.assign(m,{kernelShape:r,pads:o,outputPadding:a,outputShape:i,dilations:d,strides:p}),m},Ru=e=>{let t=Kr(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],o=e.dilations,i=e.group,a=e.kernelShape,l=e.pads,d=e.strides,p=e.wIsConst(),m=e.outputPadding,u=e.outputShape;return{autoPad:n,format:r,dilations:o,group:i,kernelShape:a,outputPadding:m,outputShape:u,pads:l,strides:d,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Dm=(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 ${i}D`);if(t.kernelShape.reduce((m,u)=>m+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Rm=[2,3,1,0],Mm=(e,t,r)=>{let n=Du(r,t),o=r.format==="NHWC",i=n.outputShape,a=i[o?3:1],l=t[0].dims[o?3:1];if(n.group!==1||a===1&&l===1){e.compute(fo(t,n));return}let d=i[o?1:2],p=i[o?2:3],m=t[1].dims[2],u=t[1].dims[3],h=o?d*p:a,w=o?a:d*p,g=m*u*l,y=!0,x=e.kernelCustomData.wT??e.compute(Re(t[1],Rm),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=x);let _=[t[0],x],$=t.length===3;$&&(!o&&t[2].dims.length===1?_.push(t[2].reshape([t[2].dims[0],1,1])):_.push(t[2])),e.compute(Ou(_,n,i,h,w,g,$,y),{inputs:_})},Um=(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 l=t.pads;l.length===0&&(l=[0,0]),l=[0,l[0],0,l[1]],a=[1].concat(a),i=[1].concat(i),o=[1].concat(o);let d=Du({...t,pads:l,strides:a,dilations:i,kernelShape:o},n);e.compute(fo(n,d,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]]))},Mu=(e,t)=>{Dm(e.inputs,t),e.inputs[0].dims.length===3?Um(e,t):Mm(e,e.inputs,t)}});var Vm,Vu,Nu,Wu=M(()=>{"use strict";X();oe();xe();re();Vm=(e,t,r,n)=>{let o=E.size(t),i=t.length,a=k("input",e,i),l=D("output",e,i),d=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),p=E.normalizeAxis(d,i),m=u=>{let h=` i32(${a.indicesGet("inputIndices","uniforms.axis")}) `,w=L("uniforms.input_shape","uniforms.axis",i),g=n.reverse?h+(n.exclusive?" + 1":""):"0",y=n.reverse?w:h+(n.exclusive?"":" + 1");return`
                 ${u.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(a,l)}
                 ${u.mainStart()}
                   ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
                   var inputIndices = ${l.offsetToIndices("global_idx")};
                   var sum = ${l.type.value}(0);
                   let first : i32 = ${g};
-                  let last : i32 = ${b};
+                  let last : i32 = ${y};
                   for (var i : i32 = first; i < last; i++) {
                     ${a.indicesSet("inputIndices","uniforms.axis","u32(i)")};
                     sum = sum + ${a.getByIndices("inputIndices")};
                   }
                   ${l.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},...V(t,t)]}),getShaderSource:m}},qu=(e,t)=>{let r=e.inputs[0].dims,n=e.inputs[0].dataType,o=e.inputs[1];e.compute(Lm(n,r,o,t),{inputs:[0]})},ju=e=>{let t=e.exclusive===1,r=e.reverse===1;return te({exclusive:t,reverse:r})}});var Fm,qm,jm,Yu,Xu,Zu=U(()=>{"use strict";Q();ae();Ce();ie();Fm=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.")},qm=(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},...U(t,t)]}),getShaderSource:m}},Vu=(e,t)=>{let r=e.inputs[0].dims,n=e.inputs[0].dataType,o=e.inputs[1];e.compute(Vm(n,r,o,t),{inputs:[0]})},Nu=e=>{let t=e.exclusive===1,r=e.reverse===1;return Q({exclusive:t,reverse:r})}});var Nm,Wm,Lm,Lu,Hu,Gu=M(()=>{"use strict";X();oe();xe();re();Nm=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.")},Wm=(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(`
-`)},jm=(e,t)=>{let r,n,o,i,a,l,d=t.format==="NHWC",p=t.blocksize,m=t.mode==="DCR";d?([r,n,o,i]=e.dims,a=m?[r,n,o,p,p,i/p**2]:[r,n,o,i/p**2,p,p],l=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],l=m?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let u=e.reshape(a),h=u.dims.length,w=e.dataType,g=E("a",w,h),b=M("output",w,h),x=_=>`
-  ${_.registerUniform("output_size","u32").declareVariables(g,b)}
+`)},Lm=(e,t)=>{let r,n,o,i,a,l,d=t.format==="NHWC",p=t.blocksize,m=t.mode==="DCR";d?([r,n,o,i]=e.dims,a=m?[r,n,o,p,p,i/p**2]:[r,n,o,i/p**2,p,p],l=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],l=m?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let u=e.reshape(a),h=u.dims.length,w=e.dataType,g=k("a",w,h),y=D("output",w,h),x=_=>`
+  ${_.registerUniform("output_size","u32").declareVariables(g,y)}
 
-  ${qm(l,h,g,b)}
+  ${Wm(l,h,g,y)}
 
   ${_.mainStart()}
     ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
 
-    let indices = ${b.offsetToIndices("global_idx")};
+    let indices = ${y.offsetToIndices("global_idx")};
     let aIndices = perm(indices);
 
-    ${b.setByOffset("global_idx",g.getByIndices("aIndices"))}
-  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:_=>{let $=d?[r,n*p,o*p,i/p**2]:[r,i/p**2,n*p,o*p],S=k.size($),I=u.dims,T=k.sortBasedOnPerm(I,l);return{outputs:[{dims:$,dataType:_[0].dataType}],dispatchGroup:{x:Math.ceil(S/64)},programUniforms:[{type:12,data:S},...V(I,T)]}},getShaderSource:x}},Yu=(e,t)=>{Fm(e.inputs),e.compute(jm(e.inputs[0],t))},Xu=e=>te({blocksize:e.blocksize,mode:e.mode,format:e.format})});var wo,rn,Qu,Km,Ym,vo,$o,Ju,Xm,ed,td,rd=U(()=>{"use strict";Q();ae();Ce();ie();wo="[a-zA-Z]|\\.\\.\\.",rn="("+wo+")+",Qu="^"+rn+"$",Km="("+rn+",)*"+rn,Ym="^"+Km+"$",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(Ym)))throw new Error("Invalid LHS term");if(n.split(",").forEach((l,d)=>{let p=t[d].dims.slice();if(!l.match(RegExp(Qu)))throw new Error("Invalid LHS term");let m=this.processTerm(l,!0,p,d);this.lhs.push(m)}),o==="")o+=[...this.symbolToInfo.entries()].filter(([l,d])=>d.count===1||l==="...").map(([l])=>l).join("");else if(!o.match(RegExp(rn)))throw new Error("Invalid RHS");o.match(RegExp(wo,"g"))?.forEach(l=>{if(l==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let d=this.symbolToInfo.get(l);if(d===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(d.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,l=[],d=0;if(!t.match(RegExp(Qu))&&!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 w=i-p.length+1;if(w<0)throw new Error("Ellipsis out of bounds");if(l=n.slice(d,d+w),this.hasEllipsis){if(this.ellipsisDims.length!==l.length||this.ellipsisDims.toString()!==l.toString())throw new Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=l;else throw new Error("Ellipsis must be specified in the LHS");for(let g=0;g<l.length;g++){let b=String.fromCharCode("0".charCodeAt(0)+g);m.addSymbol(b,h+g),this.addSymbol(b,n[d++],o)}}else m.addSymbol(u,h+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(u,n[d++],o)}),m}},Ju=e=>e+"_max",Xm=(e,t,r,n)=>{let i=e.map(m=>m.length).map((m,u)=>E(`input${u}`,t,m)),a=k.size(n),l=M("output",t,n.length),d=[...r.symbolToInfo.keys()].filter(m=>!r.rhs.symbolToIndices.has(m)),p=m=>{let u=[],h="var prod = 1.0;",w="var sum = 0.0;",g="sum += prod;",b=[],x=[],_=[],$=[],S=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((T,A)=>{if(r.rhs.symbolToIndices.has(A)){let D=r.rhs.symbolToIndices.get(A)?.[0];D!==void 0&&r.lhs.forEach((z,H)=>{if(T.inputIndices.includes(H)){let W=z.symbolToIndices.get(A);if(W===void 0)throw new Error("Invalid symbol error");W.forEach(X=>{u.push(`${i[H].indicesSet(`input${H}Indices`,X,l.indicesGet("outputIndices",D))}`)})}})}else r.lhs.forEach((D,z)=>{if(T.inputIndices.includes(z)){let H=D.symbolToIndices.get(A);if(H===void 0)throw new Error("Invalid symbol error");H.forEach(W=>{b.push(`${i[z].indicesSet(`input${z}Indices`,W,`${A}`)}`)}),$.push(`prod *= ${i[z].getByIndices(`input${z}Indices`)};`)}}),x.push(`for(var ${A}: u32 = 0; ${A} < uniforms.${Ju(A)}; ${A}++) {`),_.push("}")});let I=S?[...u,`let sum = ${i.map((T,A)=>T.getByIndices(`input${A}Indices`)).join(" * ")};`]:[...u,w,...x,...b,h,...$,g,..._];return`
-            ${m.registerUniforms(d.map(T=>({name:`${Ju(T)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...i,l)}
+    ${y.setByOffset("global_idx",g.getByIndices("aIndices"))}
+  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:_=>{let $=d?[r,n*p,o*p,i/p**2]:[r,i/p**2,n*p,o*p],S=E.size($),I=u.dims,T=E.sortBasedOnPerm(I,l);return{outputs:[{dims:$,dataType:_[0].dataType}],dispatchGroup:{x:Math.ceil(S/64)},programUniforms:[{type:12,data:S},...U(I,T)]}},getShaderSource:x}},Lu=(e,t)=>{Nm(e.inputs),e.compute(Lm(e.inputs[0],t))},Hu=e=>Q({blocksize:e.blocksize,mode:e.mode,format:e.format})});var ho,Jr,Fu,Hm,Gm,go,yo,qu,Fm,Ku,ju,Yu=M(()=>{"use strict";X();oe();xe();re();ho="[a-zA-Z]|\\.\\.\\.",Jr="("+ho+")+",Fu="^"+Jr+"$",Hm="("+Jr+",)*"+Jr,Gm="^"+Hm+"$",go=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)}},yo=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(Gm)))throw new Error("Invalid LHS term");if(n.split(",").forEach((l,d)=>{let p=t[d].dims.slice();if(!l.match(RegExp(Fu)))throw new Error("Invalid LHS term");let m=this.processTerm(l,!0,p,d);this.lhs.push(m)}),o==="")o+=[...this.symbolToInfo.entries()].filter(([l,d])=>d.count===1||l==="...").map(([l])=>l).join("");else if(!o.match(RegExp(Jr)))throw new Error("Invalid RHS");o.match(RegExp(ho,"g"))?.forEach(l=>{if(l==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let d=this.symbolToInfo.get(l);if(d===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(d.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,l=[],d=0;if(!t.match(RegExp(Fu))&&!r&&t!=="")throw new Error("Invalid LHS term");let p=t.match(RegExp(ho,"g")),m=new go(o);return p?.forEach((u,h)=>{if(u==="..."){if(a)throw new Error("Only one ellipsis is allowed per input term");a=!0;let w=i-p.length+1;if(w<0)throw new Error("Ellipsis out of bounds");if(l=n.slice(d,d+w),this.hasEllipsis){if(this.ellipsisDims.length!==l.length||this.ellipsisDims.toString()!==l.toString())throw new Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=l;else throw new Error("Ellipsis must be specified in the LHS");for(let g=0;g<l.length;g++){let y=String.fromCharCode("0".charCodeAt(0)+g);m.addSymbol(y,h+g),this.addSymbol(y,n[d++],o)}}else m.addSymbol(u,h+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(u,n[d++],o)}),m}},qu=e=>e+"_max",Fm=(e,t,r,n)=>{let i=e.map(m=>m.length).map((m,u)=>k(`input${u}`,t,m)),a=E.size(n),l=D("output",t,n.length),d=[...r.symbolToInfo.keys()].filter(m=>!r.rhs.symbolToIndices.has(m)),p=m=>{let u=[],h="var prod = 1.0;",w="var sum = 0.0;",g="sum += prod;",y=[],x=[],_=[],$=[],S=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((T,A)=>{if(r.rhs.symbolToIndices.has(A)){let O=r.rhs.symbolToIndices.get(A)?.[0];O!==void 0&&r.lhs.forEach((z,W)=>{if(T.inputIndices.includes(W)){let V=z.symbolToIndices.get(A);if(V===void 0)throw new Error("Invalid symbol error");V.forEach(j=>{u.push(`${i[W].indicesSet(`input${W}Indices`,j,l.indicesGet("outputIndices",O))}`)})}})}else r.lhs.forEach((O,z)=>{if(T.inputIndices.includes(z)){let W=O.symbolToIndices.get(A);if(W===void 0)throw new Error("Invalid symbol error");W.forEach(V=>{y.push(`${i[z].indicesSet(`input${z}Indices`,V,`${A}`)}`)}),$.push(`prod *= ${i[z].getByIndices(`input${z}Indices`)};`)}}),x.push(`for(var ${A}: u32 = 0; ${A} < uniforms.${qu(A)}; ${A}++) {`),_.push("}")});let I=S?[...u,`let sum = ${i.map((T,A)=>T.getByIndices(`input${A}Indices`)).join(" * ")};`]:[...u,w,...x,...y,h,...$,g,..._];return`
+            ${m.registerUniforms(d.map(T=>({name:`${qu(T)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...i,l)}
 
             ${m.mainStart()}
             ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
@@ -5215,12 +5214,12 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             ${I.join(`
 `)};
             ${l.setByOffset("global_idx","sum")};
-          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let m=d.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,w)=>[...V(h)]).reduce((h,w)=>h.concat(w),m);return u.push(...V(n)),{outputs:[{dims:n,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:u}},getShaderSource:p}},ed=(e,t)=>{let r=new $o(e.inputs,t.equation),n=r.outputDims,o=e.inputs.map((i,a)=>i.dims);e.compute(Xm(o,e.inputs[0].dataType,r,n))},td=e=>{let t=e.equation.replace(/\s+/g,"");return te({equation:t})}});var Zm,nd,Qm,Jm,od,id=U(()=>{"use strict";Q();ae();ie();Zm=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=r.length<t.length?0:r.length-t.length,o=t.length<r.length?0:t.length-r.length;for(;n<r.length&&o<t.length;++n,++o)if(r[n]!==t[o]&&r[n]!==1&&t[o]!==1)throw new Error("Expand requires shape to be broadcastable to input")},nd=(e,t)=>{let r=e.length-t.length,n=[];for(let o=0;o<r;++o)n.push(e[o]);for(let o=0;o<t.length;++o)n.push(t[o]===1?e[o+r]:t[o]);return n},Qm=(e,t)=>e.length>t.length?nd(e,t):nd(t,e),Jm=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=Qm(t,r),o=e[0].dataType,i=o===9?4:1,a=Math.ceil(k.size(n)/i),l=p=>{let m=E("input",o,t.length,i),u=M("output",o,n.length,i),h;if(o===9){let w=(g,b,x="")=>`
-          let outputIndices${b} = ${u.offsetToIndices(`outputOffset + ${b}u`)};
-          let offset${b} = ${m.broadcastedIndicesToOffset(`outputIndices${b}`,u)};
-          let index${b} = offset${b} / 4u;
-          let component${b} = offset${b} % 4u;
-          ${g}[${b}] = ${x}(${m.getByOffset(`index${b}`)}[component${b}]);
+          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let m=d.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,w)=>[...U(h)]).reduce((h,w)=>h.concat(w),m);return u.push(...U(n)),{outputs:[{dims:n,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:u}},getShaderSource:p}},Ku=(e,t)=>{let r=new yo(e.inputs,t.equation),n=r.outputDims,o=e.inputs.map((i,a)=>i.dims);e.compute(Fm(o,e.inputs[0].dataType,r,n))},ju=e=>{let t=e.equation.replace(/\s+/g,"");return Q({equation:t})}});var qm,Xu,Km,jm,Zu,Qu=M(()=>{"use strict";X();oe();re();qm=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")},Xu=(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},Km=(e,t)=>e.length>t.length?Xu(e,t):Xu(t,e),jm=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=Km(t,r),o=e[0].dataType,i=o===9?4:1,a=Math.ceil(E.size(n)/i),l=p=>{let m=k("input",o,t.length,i),u=D("output",o,n.length,i),h;if(o===9){let w=(g,y,x="")=>`
+          let outputIndices${y} = ${u.offsetToIndices(`outputOffset + ${y}u`)};
+          let offset${y} = ${m.broadcastedIndicesToOffset(`outputIndices${y}`,u)};
+          let index${y} = offset${y} / 4u;
+          let component${y} = offset${y} % 4u;
+          ${g}[${y}] = ${x}(${m.getByOffset(`index${y}`)}[component${y}]);
         `;h=`
         let outputOffset = global_idx * ${i};
         var data = vec4<u32>(0);
@@ -5237,29 +5236,29 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
     ${p.registerUniform("vec_size","u32").declareVariables(m,u)}
     ${p.mainStart()}
     ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
-    ${h}`},d=[{type:12,data:a},...V(t,n)];return{name:"Expand",shaderCache:{hint:`${n.length}`,inputDependencies:["rank"]},getShaderSource:l,getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d})}},od=e=>{Zm(e.inputs),e.compute(Jm(e.inputs),{inputs:[0]})}});var ef,ad,sd=U(()=>{"use strict";Q();ae();ie();Yr();ef=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 l=E("x",t,[1],4),d=E("bias",t,[1],4),p=M("y",t,[1],4),m=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],u=w=>`
+    ${h}`},d=[{type:12,data:a},...U(t,n)];return{name:"Expand",shaderCache:{hint:`${n.length}`,inputDependencies:["rank"]},getShaderSource:l,getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d})}},Zu=e=>{qm(e.inputs),e.compute(jm(e.inputs),{inputs:[0]})}});var Ym,Ju,ed=M(()=>{"use strict";X();oe();re();qr();Ym=e=>{let t=e[0].dataType,r=E.size(e[0].dims),n=E.size(e[1].dims),o=n%4===0,i=a=>{let l=k("x",t,[1],4),d=k("bias",t,[1],4),p=D("y",t,[1],4),m=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],u=w=>`
       let bias${w}_offset: u32 = (global_idx * 4 + ${w}) % uniforms.bias_size;
       let bias${w} = ${d.getByOffset(`bias${w}_offset / 4`)}[bias${w}_offset % 4];`,h=o?`
       let bias = ${d.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${u(0)}${u(1)}${u(2)}${u(3)}
       let bias = ${l.type.value}(bias0, bias1, bias2, bias3);`;return`${a.registerUniforms(m).declareVariables(l,d,p)}
 
-    ${ao(Oe(t))}
+    ${ro(ke(t))}
 
-    ${a.mainStart(Tt)}
+    ${a.mainStart(St)}
       ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}
 
       let x = ${l.getByOffset("global_idx")};
       ${h}
       let x_in = x + bias;
-      ${p.setByOffset("global_idx",so("x_in"))}
-    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${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/Tt/4)}})}},ad=e=>{e.inputs.length<2||k.size(e.inputs[1].dims)===0?su(e):e.compute(ef(e.inputs))}});var tf,rf,ud,dd,ld=U(()=>{"use strict";Q();ae();Ce();ie();tf=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},rf=(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 l=r[i],d=e[0].dataType===9?4:1,p=Math.ceil(k.size(a)/d),m=[{type:12,data:p},{type:6,data:l},{type:12,data:i},...V(e[0].dims,e[1].dims,a)],u=h=>{let w=E("data",e[0].dataType,e[0].dims.length,d),g=E("inputIndices",e[1].dataType,e[1].dims.length),b=M("output",e[0].dataType,a.length,d),x=$=>{let S=n.length,I=`var indicesIndices${$}  = ${g.type.indices}(0);`;for(let T=0;T<S;T++)I+=`${S>1?`indicesIndices${$}[${T}]`:`indicesIndices${$}`} = ${a.length>1?`outputIndices${$}[uniforms.axis + ${T}]`:`outputIndices${$}`};`;I+=`
+      ${p.setByOffset("global_idx",no("x_in"))}
+    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${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/St/4)}})}},Ju=e=>{e.inputs.length<2||E.size(e.inputs[1].dims)===0?eu(e):e.compute(Ym(e.inputs))}});var Xm,Zm,td,rd,nd=M(()=>{"use strict";X();oe();xe();re();Xm=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},Zm=(e,t)=>{let r=e[0].dims,n=e[1].dims,o=r.length,i=E.normalizeAxis(t.axis,o),a=r.slice(0);a.splice(i,1,...n);let l=r[i],d=e[0].dataType===9?4:1,p=Math.ceil(E.size(a)/d),m=[{type:12,data:p},{type:6,data:l},{type:12,data:i},...U(e[0].dims,e[1].dims,a)],u=h=>{let w=k("data",e[0].dataType,e[0].dims.length,d),g=k("inputIndices",e[1].dataType,e[1].dims.length),y=D("output",e[0].dataType,a.length,d),x=$=>{let S=n.length,I=`var indicesIndices${$}  = ${g.type.indices}(0);`;for(let T=0;T<S;T++)I+=`${S>1?`indicesIndices${$}[${T}]`:`indicesIndices${$}`} = ${a.length>1?`outputIndices${$}[uniforms.axis + ${T}]`:`outputIndices${$}`};`;I+=`
           var idx${$} = ${g.getByIndices(`indicesIndices${$}`)};
           if (idx${$} < 0) {
             idx${$} = idx${$} + uniforms.axisDimLimit;
           }
           var dataIndices${$} : ${w.type.indices};
         `;for(let T=0,A=0;T<o;T++)T===i?(I+=`${o>1?`dataIndices${$}[${T}]`:`dataIndices${$}`} = u32(idx${$});`,A+=S):(I+=`${o>1?`dataIndices${$}[${T}]`:`dataIndices${$}`} = ${a.length>1?`outputIndices${$}[${A}]`:`outputIndices${$}`};`,A++);return I},_;if(e[0].dataType===9){let $=(S,I,T="")=>`
-          let outputIndices${I} = ${b.offsetToIndices(`outputOffset + ${I}u`)};
+          let outputIndices${I} = ${y.offsetToIndices(`outputOffset + ${I}u`)};
           ${x(I)};
           let offset${I} = ${w.indicesToOffset(`dataIndices${I}`)};
           let index${I} = offset${I} / 4u;
@@ -5272,19 +5271,19 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
         ${$("value",1,"u32")}
         ${$("value",2,"u32")}
         ${$("value",3,"u32")}
-        ${b.setByOffset("global_idx","value")}
+        ${y.setByOffset("global_idx","value")}
       `}else _=`
-      let outputIndices = ${b.offsetToIndices("global_idx")};
+      let outputIndices = ${y.offsetToIndices("global_idx")};
       ${x("")};
       let value = ${w.getByIndices("dataIndices")};
-      ${b.setByOffset("global_idx","value")};
+      ${y.setByOffset("global_idx","value")};
       `;return`
-      ${h.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(w,g,b)}
+      ${h.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(w,g,y)}
       ${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}},ud=e=>te({axis:e.axis}),dd=(e,t)=>{let r=e.inputs;tf(r),e.compute(rf(e.inputs,t))}});var nf,of,cd,pd,md=U(()=>{"use strict";Q();ae();Ce();ie();nf=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.`)},of=(e,t)=>{let r=e[0].dims,n=e[0].dataType,o=r.length,i=e[1].dims,a=e[1].dataType,l=k.normalizeAxis(t.axis,o),d=r[l],p=i.slice(0),m=k.size(p),u=E("input",n,o),h=E("indicesInput",a,i.length),w=M("output",n,p.length),g=[{type:12,data:m},{type:6,data:d},{type:12,data:l}];return g.push(...V(r,i,p)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:g}),getShaderSource:_=>`
+      }`};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}},td=e=>Q({axis:e.axis}),rd=(e,t)=>{let r=e.inputs;Xm(r),e.compute(Zm(e.inputs,t))}});var Qm,Jm,od,id,ad=M(()=>{"use strict";X();oe();xe();re();Qm=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.`)},Jm=(e,t)=>{let r=e[0].dims,n=e[0].dataType,o=r.length,i=e[1].dims,a=e[1].dataType,l=E.normalizeAxis(t.axis,o),d=r[l],p=i.slice(0),m=E.size(p),u=k("input",n,o),h=k("indicesInput",a,i.length),w=D("output",n,p.length),g=[{type:12,data:m},{type:6,data:d},{type:12,data:l}];return g.push(...U(r,i,p)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:g}),getShaderSource:_=>`
       ${_.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(u,h,w)}
       ${_.mainStart()}
       ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
@@ -5300,7 +5299,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
       let value = ${u.getByIndices("inputIndices")};
 
       ${w.setByOffset("global_idx","value")};
-  }`}},cd=e=>te({axis:e.axis}),pd=(e,t)=>{let r=e.inputs;nf(r),e.compute(of(e.inputs,t))}});var af,sf,fd,hd,gd=U(()=>{"use strict";Q();ae();ie();af=e=>{if(!e)throw new Error("Input is missing");if(e.length<2||e.length>3)throw new Error("Invaid input number.");if(e.length===3&&e[2].dims.length>2)throw new Error("Invalid input shape of C");if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("Input types are mismatched")},sf=(e,t)=>{let r=e[0].dims.slice(),n=e[1].dims.slice(),[o,i,a]=Wr.getShapeOfGemmResult(r,t.transA,n,t.transB,e.length===3?e[2].dims:void 0),l=[o,i];if(!l)throw new Error("Can't use gemm on the given tensors");let d=k.size(l),p=[{type:12,data:d},{type:12,data:o},{type:12,data:i},{type:12,data:a},{type:1,data:t.alpha},{type:1,data:t.beta}],m=["type","type"];e.length===3&&(p.push(...V(e[2].dims)),m.push("rank")),p.push(...V(l));let u=h=>{let w="";t.transA&&t.transB?w="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?w="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?w="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(w="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let g=t.alpha===1?"":"value *= uniforms.alpha;",b=E("a",e[0].dataType,e[0].dims),x=E("b",e[1].dataType,e[1].dims),_=b.type.value,$=null,S=[b,x];e.length===3&&($=E("c",e[2].dataType,e[2].dims.length),S.push($));let I=M("output",e[0].dataType,l.length);S.push(I);let T=[{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`
+  }`}},od=e=>Q({axis:e.axis}),id=(e,t)=>{let r=e.inputs;Qm(r),e.compute(Jm(e.inputs,t))}});var ef,tf,sd,ud,dd=M(()=>{"use strict";X();oe();re();ef=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")},tf=(e,t)=>{let r=e[0].dims.slice(),n=e[1].dims.slice(),[o,i,a]=Ur.getShapeOfGemmResult(r,t.transA,n,t.transB,e.length===3?e[2].dims:void 0),l=[o,i];if(!l)throw new Error("Can't use gemm on the given tensors");let d=E.size(l),p=[{type:12,data:d},{type:12,data:o},{type:12,data:i},{type:12,data:a},{type:1,data:t.alpha},{type:1,data:t.beta}],m=["type","type"];e.length===3&&(p.push(...U(e[2].dims)),m.push("rank")),p.push(...U(l));let u=h=>{let w="";t.transA&&t.transB?w="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?w="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?w="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(w="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let g=t.alpha===1?"":"value *= uniforms.alpha;",y=k("a",e[0].dataType,e[0].dims),x=k("b",e[1].dataType,e[1].dims),_=y.type.value,$=null,S=[y,x];e.length===3&&($=k("c",e[2].dataType,e[2].dims.length),S.push($));let I=D("output",e[0].dataType,l.length);S.push(I);let T=[{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`
   ${h.registerUniforms(T).declareVariables(...S)}
 
   ${h.mainStart()}
@@ -5317,14 +5316,14 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
     ${g}
     ${(()=>$!=null?`let cOffset = ${$.broadcastedIndicesToOffset("vec2(m, n)",I)}; value += ${_}(uniforms.beta) * ${$.getByOffset("cOffset")};`:"")()}
     output[global_idx] = value;
-  }`};return{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:l,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p}),getShaderSource:u}},fd=e=>{let t=e.transA,r=e.transB,n=e.alpha,o=e.beta;return{transA:t,transB:r,alpha:n,beta:o,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},hd=(e,t)=>{af(e.inputs),e.compute(sf(e.inputs,t))}});var Ve,lf,bd,yd,cf,nr,wd,_o=U(()=>{"use strict";Q();ae();Ce();Nr();jr();ie();Et();Ve=(e,t)=>e.length>t&&e[t].dims.length>0&&k.size(e[t].dims)>0?e[t]:void 0,lf=(e,t)=>{let r=e[0],n=Ve(e,1),o=Ve(e,2),i=Ve(e,3),a=Ve(e,4),l=Ve(e,5),d=Ve(e,6),p=Ve(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=!1,u=r.dims[0],h=r.dims[1],w=r.dims.length===3?m?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],g=h,b=0,x=0,_=Math.floor(w/t.numHeads);if(d&&p){if(d.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(d.dims[0]!==u||d.dims[1]!==t.numHeads||d.dims[3]!==_)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(p.dims[0]!==u||p.dims[1]!==t.numHeads||p.dims[3]!==_)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(d.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');b=d.dims[2],x=d.dims[2]}else if(d||p)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $;if(n){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(n.dims[2]!==r.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');$=2,g=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==_)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.');$=5,g=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==_)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,g=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');$=3}if(i){if(i.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(o&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let S=0;if(a){S=8;let z=a.dims;throw z.length===1?z[0]===u?S=1:z[0]===3*u+2&&(S=3):z.length===2&&z[0]===u&&z[1]===g&&(S=5),S===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)'):new Error("Mask not supported")}let I=!1,T=w;if(o){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(g!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');T=o.dims[2]}else{if(g!==o.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');T=o.dims[1]*o.dims[3],I=!0}}let A=b+g,D=!1;if(a)throw new Error("Key padding mask is not supported");if(l){if(l.dims.length!==4)throw new Error('Input "relative_position_bias" is expected to have 4 dimensions');if(l.dims[0]!==u&&l.dims[0]!==1||l.dims[1]!==t.numHeads||l.dims[2]!==h||l.dims[3]!==A)throw new Error('Input "relative_position_bias" shape (batch_size, 1, sequence_length, kv_sequence_length)')}return{batchSize:u,sequenceLength:h,pastSequenceLength:b,kvSequenceLength:g,totalSequenceLength:A,maxSequenceLength:x,inputHiddenSize:0,hiddenSize:w,vHiddenSize:T,headSize:_,vHeadSize:Math.floor(T/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:S,scale:t.scale,broadcastResPosBias:D,passPastInKv:I,qkvFormat:$}},bd=e=>te({...e}),yd=te({perm:[0,2,1,3]}),cf=(e,t,r,n,o,i,a)=>{let l=[n,o,i],d=k.size(l),p=[{type:12,data:d},{type:12,data:a},{type:12,data:i}],m=u=>{let h=M("qkv_with_bias",t.dataType,l),w=E("qkv",t.dataType,l),g=E("bias",r.dataType,l),b=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
-  ${u.registerUniforms(b).declareVariables(w,g,h)}
+  }`};return{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:l,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p}),getShaderSource:u}},sd=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}`}},ud=(e,t)=>{ef(e.inputs),e.compute(tf(e.inputs,t))}});var Me,of,cd,ld,af,er,pd,bo=M(()=>{"use strict";X();oe();xe();Mr();Gr();re();Ct();Me=(e,t)=>e.length>t&&e[t].dims.length>0&&E.size(e[t].dims)>0?e[t]:void 0,of=(e,t)=>{let r=e[0],n=Me(e,1),o=Me(e,2),i=Me(e,3),a=Me(e,4),l=Me(e,5),d=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],w=u,g=0,y=0,x=Math.floor(h/t.numHeads);if(d&&p){if(d.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(d.dims[0]!==m||d.dims[1]!==t.numHeads||d.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(d.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');g=d.dims[2],y=d.dims[2]}else if(d||p)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let _;if(n){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,w=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,w=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,w=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){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 $=g+w,S=0;if(a){S=8;let O=a.dims;throw O.length===1?O[0]===m?S=1:O[0]===3*m+2&&(S=3):O.length===2&&O[0]===m&&O[1]===$&&(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 I=!1,T=h;if(o){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(w!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');T=o.dims[2]}else{if(w!==o.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');T=o.dims[1]*o.dims[3],I=!0}}let A=!1;if(a)throw new Error("Key padding mask is not supported");if(l){if(l.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(l.dims[0]!==m||l.dims[1]!==t.numHeads||l.dims[2]!==u||l.dims[3]!==$)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:m,sequenceLength:u,pastSequenceLength:g,kvSequenceLength:w,totalSequenceLength:$,maxSequenceLength:y,inputHiddenSize:0,hiddenSize:h,vHiddenSize:T,headSize:x,vHeadSize:Math.floor(T/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:S,scale:t.scale,broadcastResPosBias:A,passPastInKv:I,qkvFormat:_}},cd=e=>Q({...e}),ld=Q({perm:[0,2,1,3]}),af=(e,t,r,n,o,i,a)=>{let l=[n,o,i],d=E.size(l),p=[{type:12,data:d},{type:12,data:a},{type:12,data:i}],m=u=>{let h=D("qkv_with_bias",t.dataType,l),w=k("qkv",t.dataType,l),g=k("bias",r.dataType,l),y=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
+  ${u.registerUniforms(y).declareVariables(w,g,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:l,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p}),getShaderSource:m},{inputs:[t,r],outputs:[-1]})[0]},nr=(e,t,r,n,o,i,a,l)=>{let d=i;if(a){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return d=cf(e,i,a,t,n,r*o,l),d=d.reshape([t,n,r,o]),e.compute(Ue(d,yd.perm),{inputs:[d],outputs:[-1]})[0]}else return i.dims.length===3&&(d=i.reshape([t,n,r,o])),e.compute(Ue(d,yd.perm),{inputs:[d],outputs:[-1]})[0]},wd=(e,t)=>{let r=lf(e.inputs,t),n=e.inputs[0],o=Ve(e.inputs,1),i=Ve(e.inputs,2),a=Ve(e.inputs,3),l=Ve(e.inputs,4),d=Ve(e.inputs,5),p=Ve(e.inputs,6),m=Ve(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=nr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,n,a,0);if(u)return Wt(e,h,o,i,l,void 0,p,m,d,r,t);if(!o||!i)throw new Error("key and value must be provided");let w=nr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,o,a,r.hiddenSize),g=nr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,a,2*r.hiddenSize);Wt(e,h,w,g,l,void 0,p,m,d,r,t)}});var vd,pf,mf,xo,$d,So=U(()=>{"use strict";Q();ae();ie();vd=e=>Array.from(e.getBigInt64Array(),Number),pf=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(vd(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")},mf=(e,t)=>{let r=[];for(let n=0;n<e.length;++n)r.push(e[n]*t[n]);return r},xo=(e,t)=>{let r=e[0].dims,n=t??vd(e[1]),o=mf(r,n),i=k.size(o),a=e[0].dataType,l=E("input",a,r.length),d=M("output",a,o.length),p=m=>`
+  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:l,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p}),getShaderSource:m},{inputs:[t,r],outputs:[-1]})[0]},er=(e,t,r,n,o,i,a,l)=>{let d=i;if(a){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return d=af(e,i,a,t,n,r*o,l),d=d.reshape([t,n,r,o]),e.compute(Re(d,ld.perm),{inputs:[d],outputs:[-1]})[0]}else return i.dims.length===3&&(d=i.reshape([t,n,r,o])),e.compute(Re(d,ld.perm),{inputs:[d],outputs:[-1]})[0]},pd=(e,t)=>{let r=of(e.inputs,t),n=e.inputs[0],o=Me(e.inputs,1),i=Me(e.inputs,2),a=Me(e.inputs,3),l=Me(e.inputs,4),d=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,l,void 0,p,m,d,r,t);if(!o||!i)throw new Error("key and value must be provided");let w=er(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,o,a,r.hiddenSize),g=er(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,a,2*r.hiddenSize);Ut(e,h,w,g,l,void 0,p,m,d,r,t)}});var md,sf,uf,wo,fd,vo=M(()=>{"use strict";X();oe();re();md=e=>Array.from(e.getBigInt64Array(),Number),sf=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(md(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")},uf=(e,t)=>{let r=[];for(let n=0;n<e.length;++n)r.push(e[n]*t[n]);return r},wo=(e,t)=>{let r=e[0].dims,n=t??md(e[1]),o=uf(r,n),i=E.size(o),a=e[0].dataType,l=k("input",a,r.length),d=D("output",a,o.length),p=m=>`
       const inputShape = ${l.indices(...r)};
       ${m.registerUniform("output_size","u32").declareVariables(l,d)}
       ${m.mainStart()}
@@ -5338,7 +5337,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
         ${l.indicesSet("input_indices","i","input_dim_value")}
       }
       ${d.setByOffset("global_idx",l.getByIndices("input_indices"))}
-    }`;return{name:"Tile",shaderCache:{hint:`${n}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},...V(e[0].dims,o)]}),getShaderSource:p}},$d=e=>{pf(e.inputs),e.compute(xo(e.inputs),{inputs:[0]})}});var ff,_d,Sd,hf,xd,Id,Cd=U(()=>{"use strict";Q();ae();Ce();jr();ie();_o();So();Et();ff=(e,t)=>{let r=e[0],n=e[1],o=e[2],i=e[3],a=e[4];if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let l=!1,d=r.dims[0],p=r.dims[1],m=r.dims.length===3?l?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],u=p,h=0,w=0,g=Math.floor(m/t.numHeads),b=i&&i.dims.length!==0,x=a&&a.dims.length!==0,_=!0;if(b&&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[1],w=i.dims[1]):(h=i.dims[2],w=i.dims[2])}else if(b||x)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $;if(n){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(r.dims[2]%n.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');$=2,u=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==g)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.');$=5,u=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==g)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,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');$=3}let S=0,I=!1,T=m;if(o){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(u!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');T=o.dims[2]}else{if(u!==o.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');T=o.dims[1]*o.dims[3],I=!0}}let A=h+u,D=!1;return{batchSize:d,sequenceLength:p,pastSequenceLength:h,kvSequenceLength:u,totalSequenceLength:A,maxSequenceLength:w,inputHiddenSize:0,hiddenSize:m,vHiddenSize:T,headSize:g,vHeadSize:Math.floor(T/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:S,scale:t.scale,broadcastResPosBias:D,passPastInKv:I,qkvFormat:$,isPastkvBSNH:_}},_d=(e,t,r,n)=>{let o=[n.batchSize,n.totalSequenceLength,n.kvNumHeads,n.headSize],i=4,a=k.size(o)/i,l=n.totalSequenceLength,d=M("present_kv",r,o.length,i),p=E("new_kv",e.dataType,e.dims.length,i),m=t?E("past_kv",t.dataType,t.dims.length,i):void 0,u=Math.ceil(n.headSize/i),h={x:l,y:e.dims[0],z:1},w=t?["rank","rank"]:["rank"],g=[{type:12,data:a},{type:12,data:n.pastSequenceLength},{type:12,data:n.kvSequenceLength},{type:12,data:n.totalSequenceLength}],b=[p];m?(g.push(...V(e.dims),...V(t.dims),...V(o)),b.push(m)):g.push(...V(e.dims),...V(o));let x=[{name:"output_size",type:"u32"},{name:"past_seqlen",type:"u32"},{name:"new_seqlen",type:"u32"},{name:"present_seqlen",type:"u32"}],_=`      let past_batch_stride = uniforms.past_seqlen * num_heads * H;
+    }`;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},...U(e[0].dims,o)]}),getShaderSource:p}},fd=e=>{sf(e.inputs),e.compute(wo(e.inputs),{inputs:[0]})}});var df,hd,yd,lf,gd,bd,wd=M(()=>{"use strict";X();oe();xe();Gr();re();bo();vo();Ct();df=(e,t)=>{let r=e[0],n=e[1],o=e[2],i=e[3],a=e[4];if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let l=!1,d=r.dims[0],p=r.dims[1],m=r.dims.length===3?l?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],u=p,h=0,w=0,g=Math.floor(m/t.numHeads),y=i&&i.dims.length!==0,x=a&&a.dims.length!==0,_=!0;if(y&&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[1],w=i.dims[1]):(h=i.dims[2],w=i.dims[2])}else if(y||x)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $;if(n){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(r.dims[2]%n.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');$=2,u=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==g)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.');$=5,u=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==g)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,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');$=3}let S=0,I=!1,T=m;if(o){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(u!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');T=o.dims[2]}else{if(u!==o.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');T=o.dims[1]*o.dims[3],I=!0}}let A=h+u,O=!1;return{batchSize:d,sequenceLength:p,pastSequenceLength:h,kvSequenceLength:u,totalSequenceLength:A,maxSequenceLength:w,inputHiddenSize:0,hiddenSize:m,vHiddenSize:T,headSize:g,vHeadSize:Math.floor(T/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:S,scale:t.scale,broadcastResPosBias:O,passPastInKv:I,qkvFormat:$,isPastkvBSNH:_}},hd=(e,t,r,n)=>{let o=[n.batchSize,n.totalSequenceLength,n.kvNumHeads,n.headSize],i=4,a=E.size(o)/i,l=n.totalSequenceLength,d=D("present_kv",r,o.length,i),p=k("new_kv",e.dataType,e.dims.length,i),m=t?k("past_kv",t.dataType,t.dims.length,i):void 0,u=Math.ceil(n.headSize/i),h={x:l,y:e.dims[0],z:1},w=t?["rank","rank"]:["rank"],g=[{type:12,data:a},{type:12,data:n.pastSequenceLength},{type:12,data:n.kvSequenceLength},{type:12,data:n.totalSequenceLength}],y=[p];m?(g.push(...U(e.dims),...U(t.dims),...U(o)),y.push(m)):g.push(...U(e.dims),...U(o));let x=[{name:"output_size",type:"u32"},{name:"past_seqlen",type:"u32"},{name:"new_seqlen",type:"u32"},{name:"present_seqlen",type:"u32"}],_=`      let past_batch_stride = uniforms.past_seqlen * num_heads * H;
         var past_head_stride = uniforms.past_seqlen * H;
         if (is_bsnh) {
           past_head_stride = H;
@@ -5356,7 +5355,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
           ${$}
         }`,I=T=>`
 
-  ${T.registerUniforms(x).declareVariables(...b,d)}
+  ${T.registerUniforms(x).declareVariables(...y,d)}
   ${T.mainStart([u,n.kvNumHeads,1])}
     ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
     var indices = ${d.offsetToIndices("global_idx")};
@@ -5384,7 +5383,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
 
     let out_offset = b * present_batch_stride + s * row_stride + n * present_head_stride + h;
     ${S}
-  }`;return{name:"ConcatPastNew",shaderCache:{hint:`${n.kvNumHeads}${u}${!!t}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:o,dataType:r}],dispatchGroup:h,programUniforms:g}),getShaderSource:I}},Sd=e=>te({...e}),hf=te({perm:[0,2,1,3]}),xd=(e,t,r,n,o)=>{let i=t,a=n.kvNumHeads,l=n.nReps;return t.dims.length===3&&n.kvSequenceLength!==0&&(i=t.reshape([n.batchSize,n.kvSequenceLength,a,n.headSize])),r?i=e.compute(_d(i,r,i.dataType,n),{inputs:[i,r],outputs:[n.isPastkvBSNH?o:-1]})[0]:i=e.compute(_d(i,void 0,i.dataType,n),{inputs:[i],outputs:[n.isPastkvBSNH?o:-1]})[0],l!==1&&(i=e.compute(xo([i],[1,1,1,l]),{inputs:[i],outputs:[-1]})[0],i=i.reshape([n.batchSize,n.totalSequenceLength,a*l,n.headSize])),e.compute(Ue(i,hf.perm),{inputs:[i],outputs:[-1]})[0]},Id=(e,t)=>{let r=ff(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=nr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,e.inputs[0],void 0,0),o=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,i=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,a=xd(e,e.inputs[1],o,r,1),l=xd(e,e.inputs[2],i,r,2);Wt(e,n,a,l,void 0,void 0,void 0,void 0,void 0,r,t)}});var gf,yf,bf,Td,Ad=U(()=>{"use strict";Q();ae();ie();gf=(e,t)=>{let r=e[0].dims,n=r,o=2,i=k.sizeToDimension(r,o),a=k.sizeFromDimension(r,o),l=$e(a),d=a/l,p=[r[0],r[1],d],m=["rank","type","type"],u=[{type:12,data:a},{type:12,data:d}];u.push(...V(p,p));let h=w=>{let g=E("x",e[0].dataType,p.length,l),b=E("scale",e[1].dataType,e[1].dims),x=E("bias",e[2].dataType,e[2].dims),_=M("output",e[0].dataType,p.length,l),$=[g,b,x,_],S=g.type.value,I=l===1?"f32":`vec${l}<f32>`,T=64,A=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return`
+  }`;return{name:"ConcatPastNew",shaderCache:{hint:`${n.kvNumHeads}${u}${!!t}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:o,dataType:r}],dispatchGroup:h,programUniforms:g}),getShaderSource:I}},yd=e=>Q({...e}),lf=Q({perm:[0,2,1,3]}),gd=(e,t,r,n,o)=>{let i=t,a=n.kvNumHeads,l=n.nReps;return t.dims.length===3&&n.kvSequenceLength!==0&&(i=t.reshape([n.batchSize,n.kvSequenceLength,a,n.headSize])),r?i=e.compute(hd(i,r,i.dataType,n),{inputs:[i,r],outputs:[n.isPastkvBSNH?o:-1]})[0]:i=e.compute(hd(i,void 0,i.dataType,n),{inputs:[i],outputs:[n.isPastkvBSNH?o:-1]})[0],l!==1&&(i=e.compute(wo([i],[1,1,1,l]),{inputs:[i],outputs:[-1]})[0],i=i.reshape([n.batchSize,n.totalSequenceLength,a*l,n.headSize])),e.compute(Re(i,lf.perm),{inputs:[i],outputs:[-1]})[0]},bd=(e,t)=>{let r=df(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=er(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,e.inputs[0],void 0,0),o=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,i=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,a=gd(e,e.inputs[1],o,r,1),l=gd(e,e.inputs[2],i,r,2);Ut(e,n,a,l,void 0,void 0,void 0,void 0,void 0,r,t)}});var cf,pf,mf,vd,$d=M(()=>{"use strict";X();oe();re();cf=(e,t)=>{let r=e[0].dims,n=r,o=2,i=E.sizeToDimension(r,o),a=E.sizeFromDimension(r,o),l=he(a),d=a/l,p=[r[0],r[1],d],m=["rank","type","type"],u=[{type:12,data:a},{type:12,data:d}];u.push(...U(p,p));let h=w=>{let g=k("x",e[0].dataType,p.length,l),y=k("scale",e[1].dataType,e[1].dims),x=k("bias",e[2].dataType,e[2].dims),_=D("output",e[0].dataType,p.length,l),$=[g,y,x,_],S=g.type.value,I=l===1?"f32":`vec${l}<f32>`,T=64,A=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return`
   var<workgroup> meanShared : f32;
   var<workgroup> squaredNormShared : f32;
   var<workgroup> workgroupShared : array<${I}, ${T}>;
@@ -5412,7 +5411,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
       workgroupBarrier();
     }
     if (localIndex == 0) {
-      meanShared = ${Xe("workgroupShared[0]",l)} / f32(uniforms.normSize);
+      meanShared = ${je("workgroupShared[0]",l)} / f32(uniforms.normSize);
     }
     workgroupBarrier();
 
@@ -5433,19 +5432,19 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
       workgroupBarrier();
     }
     if (localIndex == 0) {
-      squaredNormShared = ${Xe("workgroupShared[0]",l)};
+      squaredNormShared = ${je("workgroupShared[0]",l)};
     }
     workgroupBarrier();
 
     let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${t.epsilon}));
-    let channelScale = invStdDev * f32(${b.getByOffset("channel")});
+    let channelScale = invStdDev * f32(${y.getByOffset("channel")});
     let channelShift = f32(${x.getByOffset("channel")}) - meanShared * channelScale;
     for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {
       let value = ${g.get("batch","channel","h")} * ${S}(${I}(channelScale)) + ${S}(${I}(channelShift));
       ${_.set("batch","channel","h","value")};
     }
-  }`};return{name:"InstanceNormalization",shaderCache:{hint:`${t.epsilon};${l}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:i},programUniforms:u}),getShaderSource:h}},yf=(e,t,r,n,o,i,a,l)=>{let d=$e(a),p=64,m=d===1?"vec2f":`mat2x${d}f`,u=d===1?"f32":`vec${d}f`,h=(A,D)=>`${m}(${A}, ${D})`,w=o*a/d,g=Math.ceil(i/p),b=["type"],x=[{type:12,data:g},{type:12,data:i},{type:12,data:Math.floor(a/d)},{type:12,data:Math.floor(i*a/d)}],_=A=>{let D=E("input",t.dataType,t.dims,d);return`
-  ${A.declareVariables(D)}
+  }`};return{name:"InstanceNormalization",shaderCache:{hint:`${t.epsilon};${l}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:i},programUniforms:u}),getShaderSource:h}},pf=(e,t,r,n,o,i,a,l)=>{let d=he(a),p=64,m=d===1?"vec2f":`mat2x${d}f`,u=d===1?"f32":`vec${d}f`,h=(A,O)=>`${m}(${A}, ${O})`,w=o*a/d,g=Math.ceil(i/p),y=["type"],x=[{type:12,data:g},{type:12,data:i},{type:12,data:Math.floor(a/d)},{type:12,data:Math.floor(i*a/d)}],_=A=>{let O=k("input",t.dataType,t.dims,d);return`
+  ${A.declareVariables(O)}
   @group(0) @binding(1) var<storage, read_write> output : array<${m}>;
   struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32};
   @group(0) @binding(2) var<uniform> uniforms: Uniforms;
@@ -5460,17 +5459,17 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
     let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H);
 
     let offset = currentImageNumber * uniforms.image_size + currentChannelNumber;
-    var sum = ${gt("f32",d)};
-    var squaredSum = ${gt("f32",d)};
+    var sum = ${ft("f32",d)};
+    var squaredSum = ${ft("f32",d)};
     for (var i: u32 = wgOffset; i < wgMax; i++) {
         let value = ${u}(input[offset + i * uniforms.C]);
         sum += value;
         squaredSum += value * value;
     }
     output[global_idx] = ${h("sum","squaredSum")};
-  }`},$=e.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${d}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:[o,a,p,2],dataType:1}],dispatchGroup:{x:o*a/d},programUniforms:x}),getShaderSource:_},{inputs:[t],outputs:[-1]})[0],S=[{type:12,data:w},{type:12,data:i},{type:12,data:Math.floor(a/d)},{type:12,data:Math.floor(p*a/d)}],I=["type","type","type"],T=A=>{let D=E("scale",r.dataType,r.dims,d),z=E("bias",n.dataType,n.dims,d);return`
+  }`},$=e.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${d}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:[o,a,p,2],dataType:1}],dispatchGroup:{x:o*a/d},programUniforms:x}),getShaderSource:_},{inputs:[t],outputs:[-1]})[0],S=[{type:12,data:w},{type:12,data:i},{type:12,data:Math.floor(a/d)},{type:12,data:Math.floor(p*a/d)}],I=["type","type","type"],T=A=>{let O=k("scale",r.dataType,r.dims,d),z=k("bias",n.dataType,n.dims,d);return`
   @group(0) @binding(0) var<storage, read> input : array<${m}>;
-  @group(0) @binding(1) var<storage, read> scale : array<${D.type.storage}>;
+  @group(0) @binding(1) var<storage, read> scale : array<${O.type.storage}>;
   @group(0) @binding(2) var<storage, read> bias : array<${z.type.storage}>;
   @group(0) @binding(3) var<storage, read_write> output : array<${m}>;
   struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32};
@@ -5482,8 +5481,8 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
     let currentChannelNumber = global_idx % uniforms.C;
 
     let offset = currentImageNumber * uniforms.image_size;
-    var sum = ${gt("f32",d)};
-    var squaredSum = ${gt("f32",d)};
+    var sum = ${ft("f32",d)};
+    var squaredSum = ${ft("f32",d)};
     for (var i: u32 = 0; i < min(${p}, uniforms.H); i++) {
         let value = input[offset + i + currentChannelNumber * ${p}];
         sum += value[0];
@@ -5496,7 +5495,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
     let channelShift = ${u}(bias[currentChannelNumber]) - sum * channelScale;
 
     output[global_idx] = ${h("channelScale","channelShift")};
-  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${d};${l}`,inputDependencies:I},getRunData:()=>({outputs:[{dims:[o,a,2],dataType:1}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:S}),getShaderSource:T},{inputs:[$,r,n],outputs:[-1]})[0]},bf=(e,t,r)=>{let n=t[0].dims,o=n,i=n[0],a=n[n.length-1],l=k.sizeFromDimension(n,1)/a,d=$e(a),p=k.size(o)/d,m=[{type:12,data:l},{type:12,data:Math.floor(a/d)}],u=["type","type"],h=yf(e,t[0],t[1],t[2],i,l,a,r.epsilon),w=g=>{let b=ge(t[0].dataType),x=d===1?"vec2f":`mat2x${d}f`,_=d===1?b:`vec${d}<${b}>`,$=E("input",t[0].dataType,t[0].dims,d),S=M("output",t[0].dataType,o,d);return`
+  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${d};${l}`,inputDependencies:I},getRunData:()=>({outputs:[{dims:[o,a,2],dataType:1}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:S}),getShaderSource:T},{inputs:[$,r,n],outputs:[-1]})[0]},mf=(e,t,r)=>{let n=t[0].dims,o=n,i=n[0],a=n[n.length-1],l=E.sizeFromDimension(n,1)/a,d=he(a),p=E.size(o)/d,m=[{type:12,data:l},{type:12,data:Math.floor(a/d)}],u=["type","type"],h=pf(e,t[0],t[1],t[2],i,l,a,r.epsilon),w=g=>{let y=pe(t[0].dataType),x=d===1?"vec2f":`mat2x${d}f`,_=d===1?y:`vec${d}<${y}>`,$=k("input",t[0].dataType,t[0].dims,d),S=D("output",t[0].dataType,o,d);return`
   @group(0) @binding(0) var<storage, read> input : array<${$.type.storage}>;
   @group(0) @binding(1) var<storage, read> scaleInput : array<${x}>;
   @group(0) @binding(2) var<storage, read_write> output : array<${S.type.storage}>;
@@ -5510,77 +5509,77 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
     let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber;
     let scale = scaleInput[scaleOffset];
     output[global_idx] = fma(input[global_idx], ${_}(scale[0]), ${_}(scale[1]));
-  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${d}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:w},{inputs:[t[0],h]})},Td=(e,t)=>{t.format==="NHWC"?bf(e,e.inputs,t):e.compute(gf(e.inputs,t))}});var wf,vf,Ed,kd=U(()=>{"use strict";Q();ae();ie();wf=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},vf=(e,t,r)=>{let n=t.simplified,o=e[0].dims,i=e[1],a=!n&&e[2],l=o,d=k.normalizeAxis(t.axis,o.length),p=k.sizeToDimension(o,d),m=k.sizeFromDimension(o,d),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:] == ${m}.
+  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${d}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:w},{inputs:[t[0],h]})},vd=(e,t)=>{t.format==="NHWC"?mf(e,e.inputs,t):e.compute(cf(e.inputs,t))}});var ff,hf,_d,xd=M(()=>{"use strict";X();oe();re();ff=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},hf=(e,t,r)=>{let n=t.simplified,o=e[0].dims,i=e[1],a=!n&&e[2],l=o,d=E.normalizeAxis(t.axis,o.length),p=E.sizeToDimension(o,d),m=E.sizeFromDimension(o,d),u=E.size(i.dims),h=a?E.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 ${h}`);let w=[];for(let T=0;T<o.length;++T)T<d?w.push(o[T]):w.push(1);let g=$e(m),b=["type","type"],x=[{type:12,data:p},{type:1,data:m},{type:12,data:Math.floor(m/g)},{type:1,data:t.epsilon}];a&&b.push("type");let _=r>1,$=r>2,S=T=>{let A=ge(e[0].dataType),D=[E("x",e[0].dataType,e[0].dims,g),E("scale",i.dataType,i.dims,g)];a&&D.push(E("bias",a.dataType,a.dims,g)),D.push(M("output",e[0].dataType,l,g)),_&&D.push(M("mean_data_output",1,w)),$&&D.push(M("inv_std_output",1,w));let z=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
-  ${T.registerUniforms(z).declareVariables(...D)}
+       Got scale size of ${u} and bias size of ${h}`);let w=[];for(let T=0;T<o.length;++T)T<d?w.push(o[T]):w.push(1);let g=he(m),y=["type","type"],x=[{type:12,data:p},{type:1,data:m},{type:12,data:Math.floor(m/g)},{type:1,data:t.epsilon}];a&&y.push("type");let _=r>1,$=r>2,S=T=>{let A=pe(e[0].dataType),O=[k("x",e[0].dataType,e[0].dims,g),k("scale",i.dataType,i.dims,g)];a&&O.push(k("bias",a.dataType,a.dims,g)),O.push(D("output",e[0].dataType,l,g)),_&&O.push(D("mean_data_output",1,w)),$&&O.push(D("inv_std_output",1,w));let z=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
+  ${T.registerUniforms(z).declareVariables(...O)}
   ${T.mainStart()}
     ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
     let offset = global_idx * uniforms.norm_size_vectorized;
-    var mean_vector = ${gt("f32",g)};
-    var mean_square_vector = ${gt("f32",g)};
+    var mean_vector = ${ft("f32",g)};
+    var mean_square_vector = ${ft("f32",g)};
 
     for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
-      let value = ${At(A,g,"x[h + offset]")};
+      let value = ${It(A,g,"x[h + offset]")};
       mean_vector += value;
       mean_square_vector += value * value;
     }
-    let mean = ${Xe("mean_vector",g)} / uniforms.norm_size;
-    let inv_std_dev = inverseSqrt(${Xe("mean_square_vector",g)} / uniforms.norm_size ${n?"":"- mean * mean"} + uniforms.epsilon);
+    let mean = ${je("mean_vector",g)} / uniforms.norm_size;
+    let inv_std_dev = inverseSqrt(${je("mean_square_vector",g)} / uniforms.norm_size ${n?"":"- mean * mean"} + uniforms.epsilon);
 
     for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
-      let f32input = ${At(A,g,"x[j + offset]")};
-      let f32scale = ${At(A,g,"scale[j]")};
-      output[j + offset] = ${D[0].type.value}((f32input ${n?"":"- mean"}) * inv_std_dev * f32scale
-        ${a?`+ ${At(A,g,"bias[j]")}`:""}
+      let f32input = ${It(A,g,"x[j + offset]")};
+      let f32scale = ${It(A,g,"scale[j]")};
+      output[j + offset] = ${O[0].type.value}((f32input ${n?"":"- mean"}) * inv_std_dev * f32scale
+        ${a?`+ ${It(A,g,"bias[j]")}`:""}
       );
     }
 
     ${_?"mean_data_output[global_idx] = mean":""};
     ${$?"inv_std_output[global_idx] = inv_std_dev":""};
-  }`},I=[{dims:l,dataType:e[0].dataType}];return _&&I.push({dims:w,dataType:1}),$&&I.push({dims:w,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${g};${r};${n}`,inputDependencies:b},getRunData:()=>({outputs:I,dispatchGroup:{x:Math.ceil(p/64)},programUniforms:x}),getShaderSource:S}},Ed=(e,t)=>{wf(e.inputs),e.compute(vf(e.inputs,t,e.outputCount))}});var $f,_f,Pd,Od,Dd=U(()=>{"use strict";Q();ae();Ce();ie();$f=(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 d=e[2].dims;if(k.size(d)!==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.")}},_f=(e,t,r,n)=>{let o=e[0].dims,i=o.length,a=Math.floor((t.k+t.blockSize-1)/t.blockSize),l=o[i-2],d=t.k,p=t.n,m=o.slice(0,i-2),u=k.size(m),w=t.blockSize/8*t.bits/4,g=e[0].dataType,b=$e(l),x=$e(t.k),_=$e(w),$=It(g),S=l*a*$,I=Math.floor(n/S),T=a<=r[0]&&I>0,A=!T||I>=4?$e(p):I>=2&&$e(p)>=2?2:1,D=m.concat([l,p]),z=k.size(D)/A/b,H=T?[]:[{type:12,data:z},{type:12,data:t.blockSize}],W=[u,l,d/x],X=k.convertShape(e[1].dims).slice();X.splice(-1,1,w/_),H.push(...V(W)),H.push(...V(X)),H.push(...V(e[2].dims)),e.length===4&&H.push(...V(k.convertShape(e[3].dims)));let Y=[u,l,p/A];H.push(...V(Y));let pe=ue=>{let le=W.length,K=E("a",e[0].dataType,le,x),de=E("b",12,X.length,_),re=E("scales",e[2].dataType,e[2].dims.length),se=[K,de,re],Z=e.length===4?E("zero_points",12,e[3].dims.length):void 0;Z&&se.push(Z);let ne=Y.length,J=M("output",e[0].dataType,ne,A),ke=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],R=ge(e[0].dataType),F=(()=>{switch(x){case 1:return`array<${R}, 8>`;case 2:return`mat4x2<${R}>`;case 4:return`mat2x4<${R}>`;default:throw new Error(`${x}-component is not supported.`)}})(),ce=`
+  }`},I=[{dims:l,dataType:e[0].dataType}];return _&&I.push({dims:w,dataType:1}),$&&I.push({dims:w,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${g};${r};${n}`,inputDependencies:y},getRunData:()=>({outputs:I,dispatchGroup:{x:Math.ceil(p/64)},programUniforms:x}),getShaderSource:S}},_d=(e,t)=>{ff(e.inputs),e.compute(hf(e.inputs,t,e.outputCount))}});var gf,yf,Sd,Id,Cd=M(()=>{"use strict";X();oe();xe();re();gf=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],n=r.dims.length;if(r.dims[n-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let o=Math.floor((t.k+t.blockSize-1)/t.blockSize),i=t.blockSize/8*t.bits,a=e[1];if(!E.areEqual(a.dims,[t.n,o,i]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let d=e[2].dims;if(E.size(d)!==t.n*o)throw new Error("scales input size error.");if(e.length===4){let m=e[3].dims,u=t.bits>4?t.n*o:t.n*Math.floor((o+1)/2);if(E.size(m)!==u)throw new Error("zeroPoints input size error.")}},yf=(e,t,r,n)=>{let o=e[0].dims,i=o.length,a=Math.floor((t.k+t.blockSize-1)/t.blockSize),l=o[i-2],d=t.k,p=t.n,m=o.slice(0,i-2),u=E.size(m),w=t.blockSize/8*t.bits/4,g=e[0].dataType,y=he(l),x=he(t.k),_=he(w),$=_t(g,l*a),S=Math.floor(n/$),I=a<=r[0]&&S>0,T=!I||S>=4?he(p):S>=2&&he(p)>=2?2:1,A=m.concat([l,p]),O=E.size(A)/T/y,z=I?[]:[{type:12,data:O},{type:12,data:t.blockSize}],W=[u,l,d/x],V=E.convertShape(e[1].dims).slice();V.splice(-1,1,w/_),z.push(...U(W)),z.push(...U(V)),z.push(...U(e[2].dims)),e.length===4&&z.push(...U(E.convertShape(e[3].dims)));let j=[u,l,p/T];z.push(...U(j));let Y=ae=>{let se=W.length,ie=k("a",e[0].dataType,se,x),q=k("b",12,V.length,_),ce=k("scales",e[2].dataType,e[2].dims.length),ne=[ie,q,ce],Z=e.length===4?k("zero_points",12,e[3].dims.length):void 0;Z&&ne.push(Z);let ee=j.length,K=D("output",e[0].dataType,ee,T),me=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],$e=pe(e[0].dataType),R=(()=>{switch(x){case 1:return`array<${$e}, 8>`;case 2:return`mat4x2<${$e}>`;case 4:return`mat2x4<${$e}>`;default:throw new Error(`${x}-component is not supported.`)}})(),G=`
         for (var word: u32 = 0; word < ${w}; word += ${_}) {
-          ${de.indicesSet("b_indices","2","word")};
-          let b_data = ${de.getByIndices("b_indices")};
+          ${q.indicesSet("b_indices","2","word")};
+          let b_data = ${q.getByIndices("b_indices")};
           for (var i: u32 = 0; i < ${_}; i++) {
             let b_value: u32 = ${_===1?"b_data":"b_data[word + i]"};
             let b_mask: u32 = 0x0F0F0F0Fu;
             let b_value_lower: vec4<u32> = unpack4xU8(b_value & b_mask);
             let b_value_upper: vec4<u32> = unpack4xU8((b_value >> 4) & b_mask);
-            let b_quantized_values = ${F}(${Array.from({length:4},(Se,we)=>`${R}(b_value_lower[${we}]), ${R}(b_value_upper[${we}])`).join(", ")});
-            let b_dequantized_values = ${(()=>x===1?`${F}(${Array.from({length:8},(Se,we)=>`(b_quantized_values[${we}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${F}(${Array(8).fill("zero_point").join(",")})) * scale;`)()};
+            let b_quantized_values = ${R}(${Array.from({length:4},(Be,ge)=>`${$e}(b_value_lower[${ge}]), ${$e}(b_value_upper[${ge}])`).join(", ")});
+            let b_dequantized_values = ${(()=>x===1?`${R}(${Array.from({length:8},(Be,ge)=>`(b_quantized_values[${ge}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${R}(${Array(8).fill("zero_point").join(",")})) * scale;`)()};
             // Number of B elements per 32-bit word is 32/bits = 32/4 = 8
-            for (var m: u32 = 0; m < ${T?l:b}u; m++) {
-              ${K.indicesSet("a_indices",le-2,T?"m":`row * ${b} + m`)};
-              ${K.indicesSet("a_indices",le-1,"word_offset")};
-              var input_offset = ${K.indicesToOffset("a_indices")};
-              var a_data: ${F};
+            for (var m: u32 = 0; m < ${I?l:y}u; m++) {
+              ${ie.indicesSet("a_indices",se-2,I?"m":`row * ${y} + m`)};
+              ${ie.indicesSet("a_indices",se-1,"word_offset")};
+              var input_offset = ${ie.indicesToOffset("a_indices")};
+              var a_data: ${R};
               for (var j: u32 = 0; j < ${8/x}; j++) {
-                a_data[j] = ${K.getByOffset("input_offset")};
+                a_data[j] = ${ie.getByOffset("input_offset")};
                 input_offset++;
               }
-              ${T?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${A>1?"[c]":""} += ${Array.from({length:8/x},(Se,we)=>`${x===1?`a_data[${we}] * b_dequantized_values[${we}]`:`dot(a_data[${we}], b_dequantized_values[${we}])`}`).join(" + ")};
+              ${I?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${T>1?"[c]":""} += ${Array.from({length:8/x},(Be,ge)=>`${x===1?`a_data[${ge}] * b_dequantized_values[${ge}]`:`dot(a_data[${ge}], b_dequantized_values[${ge}])`}`).join(" + ")};
             }
             word_offset += ${8/x};
           }
-        }`,Re=Z?`
+        }`,ue=Z?`
           zero_point_offset += 4;
           if (zero_point_offset == 32) {
             zero_point_offset = 0;
             zero_point_index++;
             zero_point_word = ${Z.getByOffset("zero_point_index")};
-          }`:"";return T?`
-        var<workgroup> workgroup_shared: array<${J.type.value}, ${l*a}>;
-        ${ue.declareVariables(...se,J)}
-        ${ue.mainStart([a,1,1])}
-          var a_indices: ${K.type.indices};
+          }`:"";return I?`
+        var<workgroup> workgroup_shared: array<${K.type.value}, ${l*a}>;
+        ${ae.declareVariables(...ne,K)}
+        ${ae.mainStart([a,1,1])}
+          var a_indices: ${ie.type.indices};
           var block = local_id.x;
           var col = workgroup_id.y;
           var batch = workgroup_id.z;
-          ${K.indicesSet("a_indices","0","batch")};
+          ${ie.indicesSet("a_indices","0","batch")};
           // Two zero points are packed into one byte when uniforms.bits is 4.
-          for (var c: u32 = 0; c < ${A}; c++) {
-            let col_times_components_plus_c = col * ${A} + c;
+          for (var c: u32 = 0; c < ${T}; c++) {
+            let col_times_components_plus_c = col * ${T} + c;
               ${Z?`
             var zero_point_bytes_per_col: u32 = (${a} + 1) / 2;
             var zero_point_byte_count: u32 = col_times_components_plus_c * zero_point_bytes_per_col + (block >> 0x1u);
@@ -5589,90 +5588,90 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             var zero_point_nibble_offset: u32 = block & 0x1u;
             var zero_point_bits_offset: u32 = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
             var zero_point_word: u32 = ${Z.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;`:""}
-            var b_indices: ${de.type.indices};
-            ${de.indicesSet("b_indices","0","col_times_components_plus_c")};
+            var b_indices: ${q.type.indices};
+            ${q.indicesSet("b_indices","0","col_times_components_plus_c")};
             // The scale and zero points are computed per block.
             var scales_index = col_times_components_plus_c * ${a} + block;
-            let scale = ${re.getByOffset("scales_index")};
+            let scale = ${ce.getByOffset("scales_index")};
             // The default zero point is 8 for unsigned 4-bit quantization.
-            let zero_point = ${R}(${Z?"(zero_point_word) & 0xFu":8});
-            ${de.indicesSet("b_indices","1","block")};
+            let zero_point = ${$e}(${Z?"(zero_point_word) & 0xFu":8});
+            ${q.indicesSet("b_indices","1","block")};
             var word_offset: u32 = block * ${t.blockSize/x};
             var workgroup_shared_offset: u32 = block * ${l};
-            ${ce}
+            ${G}
           }
           workgroupBarrier();
-          var output_indices: ${J.type.indices};
+          var output_indices: ${K.type.indices};
           var elements_per_thread: u32 = ${Math.ceil(l/a)};
-          ${J.indicesSet("output_indices","0","batch")};
-          ${J.indicesSet("output_indices",ne-1,"col")};
-          ${J.indicesSet("output_indices",ne-2,"local_id.x * elements_per_thread")};
-          var output_offset = ${J.indicesToOffset("output_indices")};
+          ${K.indicesSet("output_indices","0","batch")};
+          ${K.indicesSet("output_indices",ee-1,"col")};
+          ${K.indicesSet("output_indices",ee-2,"local_id.x * elements_per_thread")};
+          var output_offset = ${K.indicesToOffset("output_indices")};
           for (var m: u32 = 0u; m < elements_per_thread; m++) {
             var row = m + local_id.x * elements_per_thread;
             if (row < ${l}) {
-              var output_value: ${J.type.value} = ${J.type.value}(0);
+              var output_value: ${K.type.value} = ${K.type.value}(0);
               var workgroup_shared_offset: u32 = row;
               for (var b: u32 = 0u; b < ${a}u; b++) {
                 output_value += workgroup_shared[workgroup_shared_offset];
                 workgroup_shared_offset += ${l};
               }
-              ${J.setByOffset("output_offset","output_value")};
-              output_offset += ${p/A};
+              ${K.setByOffset("output_offset","output_value")};
+              output_offset += ${p/T};
             }
           }
         }`:`
-        ${ue.registerUniforms(ke).declareVariables(...se,J)}
-        ${ue.mainStart()}
-          ${ue.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
-          var output_values: array<${J.type.value}, ${b}>;
-          var output_indices = ${J.offsetToIndices("global_idx")};
-          var col = ${J.indicesGet("output_indices",ne-1)};
-          var row = ${J.indicesGet("output_indices",ne-2)};
-          var a_indices: ${K.type.indices} = output_indices;
+        ${ae.registerUniforms(me).declareVariables(...ne,K)}
+        ${ae.mainStart()}
+          ${ae.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+          var output_values: array<${K.type.value}, ${y}>;
+          var output_indices = ${K.offsetToIndices("global_idx")};
+          var col = ${K.indicesGet("output_indices",ee-1)};
+          var row = ${K.indicesGet("output_indices",ee-2)};
+          var a_indices: ${ie.type.indices} = output_indices;
           // Two zero points are packed into one byte because uniforms.bits <= 4.
           // zero_point_offset is either 0 or 4. It is bit offset within one byte.
           // TODO support zero_point_offset for bits > 4
           ${Z?`
-          var zero_point_abs_offset = col * ${A} * ((${a} + 1) / 2);
+          var zero_point_abs_offset = col * ${T} * ((${a} + 1) / 2);
           var zero_point_index: u32 = zero_point_abs_offset / 4;
           var zero_point_word: u32 = ${Z.getByOffset("zero_point_index")};
           var zero_point_offset: u32 = (zero_point_abs_offset % 4) * 8;`:""}
-          var scale_index = col * ${a*A};
-          var b_indices: ${de.type.indices};
-          for (var c: u32 = 0; c < ${A}; c++) {
-            ${de.indicesSet("b_indices","0",`col * ${A} + c`)};
+          var scale_index = col * ${a*T};
+          var b_indices: ${q.type.indices};
+          for (var c: u32 = 0; c < ${T}; c++) {
+            ${q.indicesSet("b_indices","0",`col * ${T} + c`)};
             var block_offset: u32 = 0;
             for (var block: u32 = 0; block < ${a}; block++) {
               // The scale and zero points are computed per block.
-              let scale = ${re.getByOffset("scale_index")};
+              let scale = ${ce.getByOffset("scale_index")};
               // The default zero point is 8 for unsigned 4-bit quantization.
-              let zero_point = ${R}(${Z?"extractBits(zero_point_word, zero_point_offset, 4)":8});
-              ${de.indicesSet("b_indices","1","block")};
+              let zero_point = ${$e}(${Z?"extractBits(zero_point_word, zero_point_offset, 4)":8});
+              ${q.indicesSet("b_indices","1","block")};
               var word_offset: u32 = block_offset;
-              ${ce}
+              ${G}
               scale_index++;
-              ${Re}
+              ${ue}
               block_offset += uniforms.block_size / ${x};
             }
             // Drop the trailing 4 bits if the zero_poit_offset is not a byte boundary to align with the next byte.
             ${Z?`if (zero_point_offset % 8 > 0) {
-                ${Re}
+                ${ue}
               }`:""}
             }
-            for (var k: u32 = 0u; k < ${b}u; k++) {
-              ${J.indicesSet("output_indices",ne-2,`${b} * row + k`)};
-              ${J.setByIndices("output_indices","output_values[k]")}
+            for (var k: u32 = 0u; k < ${y}u; k++) {
+              ${K.indicesSet("output_indices",ee-2,`${y} * row + k`)};
+              ${K.setByIndices("output_indices","output_values[k]")}
             }
-        }`};return{name:T?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${t.cacheKey};${l};${g};${e.length}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:D,dataType:g}],name:T?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:T?{x:1,y:Math.ceil(p/A),z:u}:{x:Math.ceil(z/64)},programUniforms:H}),getShaderSource:pe}},Pd=(e,t)=>{$f(e.inputs,t);let r=e.getMaxComputeWorkgroupSizes(),n=e.getMaxComputeWorkgroupStoragesize();e.compute(_f(e.inputs,t,r,n))},Od=e=>te(e)});var xf,Sf,If,Cf,Tf,Af,Ef,kf,zd,Bd=U(()=>{"use strict";Q();ae();ie();xf=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].")}},Sf=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
-            k = i32(${e.indicesGet("indices",o)}) - ${G("uniforms.pads",o,r)};
+        }`};return{name:I?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${t.cacheKey};${l};${g};${e.length}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:A,dataType:g}],name:I?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:I?{x:1,y:Math.ceil(p/T),z:u}:{x:Math.ceil(O/64)},programUniforms:z}),getShaderSource:Y}},Sd=(e,t)=>{gf(e.inputs,t);let r=e.getMaxComputeWorkgroupSizes(),n=e.getMaxComputeWorkgroupStoragesize();e.compute(yf(e.inputs,t,r,n))},Id=e=>Q(e)});var bf,wf,vf,$f,_f,xf,Sf,If,Td,Ad=M(()=>{"use strict";X();oe();re();bf=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].")}},wf=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
+            k = i32(${e.indicesGet("indices",o)}) - ${L("uniforms.pads",o,r)};
             if (k < 0) {
               break;
             }
-            if (k >= i32(${G("uniforms.x_shape",o,t)})) {
+            if (k >= i32(${L("uniforms.x_shape",o,t)})) {
               break;
             }
-            offset += k * i32(${G("uniforms.x_strides",o,t)});
+            offset += k * i32(${L("uniforms.x_strides",o,t)});
         `;return`
           value = ${e.type.value}(uniforms.constant_value);
           for (var i = 0; i < 1; i++) {
@@ -5681,53 +5680,53 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             ${n}
             value = x[offset];
           }
-      `},If=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
-                k = i32(${e.indicesGet("indices",o)}) - ${G("uniforms.pads",o,r)};
+      `},vf=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
+                k = i32(${e.indicesGet("indices",o)}) - ${L("uniforms.pads",o,r)};
                 if (k < 0) {
                   k = -k;
                 }
                 {
-                  let _2n_1 = 2 * (i32(${G("uniforms.x_shape",o,t)}) - 1);
+                  let _2n_1 = 2 * (i32(${L("uniforms.x_shape",o,t)}) - 1);
                   k = k % _2n_1;
-                  if(k >= i32(${G("uniforms.x_shape",o,t)})) {
+                  if(k >= i32(${L("uniforms.x_shape",o,t)})) {
                     k = _2n_1 - k;
                   }
                 }
-                offset += k * i32(${G("uniforms.x_strides",o,t)});
+                offset += k * i32(${L("uniforms.x_strides",o,t)});
             `;return`
               var offset = 0;
               var k = 0;
               ${n}
               value = x[offset];
-          `},Cf=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
-                k = i32(${e.indicesGet("indices",o)}) - ${G("uniforms.pads",o,r)};
+          `},$f=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
+                k = i32(${e.indicesGet("indices",o)}) - ${L("uniforms.pads",o,r)};
                 if (k < 0) {
                   k = 0;
                 }
-                if (k >= i32(${G("uniforms.x_shape",o,t)})) {
-                  k = i32(${G("uniforms.x_shape",o,t)}) - 1;
+                if (k >= i32(${L("uniforms.x_shape",o,t)})) {
+                  k = i32(${L("uniforms.x_shape",o,t)}) - 1;
                 }
-                offset += k * i32(${G("uniforms.x_strides",o,t)});
+                offset += k * i32(${L("uniforms.x_strides",o,t)});
             `;return`
               var offset = 0;
               var k = 0;
               ${n}
               value = x[offset];
-          `},Tf=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
-                k = i32(${e.indicesGet("indices",o)}) - ${G("uniforms.pads",o,r)};
+          `},_f=(e,t,r)=>{let n="";for(let o=t-1;o>=0;--o)n+=`
+                k = i32(${e.indicesGet("indices",o)}) - ${L("uniforms.pads",o,r)};
                 if (k < 0)  {
-                  k += i32(${G("uniforms.x_shape",o,t)}]);
+                  k += i32(${L("uniforms.x_shape",o,t)}]);
                 }
-                if (k >= i32(${G("uniforms.x_shape",o,t)})) {
-                  k -= i32(${G("uniforms.x_shape",o,t)});
+                if (k >= i32(${L("uniforms.x_shape",o,t)})) {
+                  k -= i32(${L("uniforms.x_shape",o,t)});
                 }
-                offset += k * i32(${G("uniforms.x_strides",o,t)});
+                offset += k * i32(${L("uniforms.x_strides",o,t)});
             `;return`
               var offset = 0;
               var k = 0;
               ${n}
               value = x[offset];
-          `},Af=(e,t,r)=>{switch(r.mode){case 0:return Sf(e,t,r.pads.length);case 1:return If(e,t,r.pads.length);case 2:return Cf(e,t,r.pads.length);case 3:return Tf(e,t,r.pads.length);default:throw new Error("Invalid mode")}},Ef=(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}];t.mode===0&&i.push({type:e[0].dataType,data:t.value}),i.push(...V(e[0].dims,r));let a=["rank"],l=d=>{let p=M("output",e[0].dataType,r.length),m=E("x",e[0].dataType,n.length),u=m.type.value,h=Af(p,n.length,t),w=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&w.push({name:"constant_value",type:u}),`
+          `},xf=(e,t,r)=>{switch(r.mode){case 0:return wf(e,t,r.pads.length);case 1:return vf(e,t,r.pads.length);case 2:return $f(e,t,r.pads.length);case 3:return _f(e,t,r.pads.length);default:throw new Error("Invalid mode")}},Sf=(e,t)=>{let r=E.padShape(e[0].dims.slice(),t.pads),n=e[0].dims,o=E.size(r),i=[{type:12,data:o},{type:6,data:t.pads}];t.mode===0&&i.push({type:e[0].dataType,data:t.value}),i.push(...U(e[0].dims,r));let a=["rank"],l=d=>{let p=D("output",e[0].dataType,r.length),m=k("x",e[0].dataType,n.length),u=m.type.value,h=xf(p,n.length,t),w=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&w.push({name:"constant_value",type:u}),`
             ${d.registerUniforms(w).declareVariables(m,p)}
             ${d.mainStart()}
             ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
@@ -5737,7 +5736,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             var value = ${u}(0);
             ${h}
             output[global_idx] = value;
-        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}`,inputDependencies:a},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(k.size(r)/64)},programUniforms:i}),getShaderSource:l}},kf=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),n=e.length>=3&&e[2].data?e[2].getFloat32Array()[0]:0,o=e[0].dims.length,i=new Int32Array(2*o).fill(0);if(e.length>=4){let l=e[3].getBigInt64Array();for(let d=0;d<l.length;d++)i[Number(l[d])]=Number(r[d]),i[Number(l[d])+o]=Number(r[d+l.length])}else r.forEach((l,d)=>i[Number(d)]=Number(l));let a=[];return i.forEach(l=>a.push(l)),{mode:t.mode,value:n,pads:a}}else return t},zd=(e,t)=>{xf(e.inputs);let r=kf(e.inputs,t);e.compute(Ef(e.inputs,r),{inputs:[0]})}});var nn,Rd,Md,Ud,Vd,Pf,Of,Nd,Wd,Hd,Gd,Ld,Fd,qd,jd,Kd,Yd,Xd,Zd,Qd=U(()=>{"use strict";Ke();Q();ae();ie();nn=e=>{if(be.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Rd=(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(),l=t.strides.slice(),d=i?t.dilations.slice():[],p=t.pads.slice();Ct.adjustPoolAttributes(r,o,a,l,d,p);let m=Ct.computePoolOutputShape(r,o,l,d,a,p,t.autoPad),u=Object.assign({},t);i?Object.assign(u,{kernelShape:a,strides:l,pads:p,dilations:d,cacheKey:t.cacheKey}):Object.assign(u,{kernelShape:a,strides:l,pads:p,cacheKey:t.cacheKey});let h=m.slice();return h.push(h.splice(1,1)[0]),[u,n?h:m]},Md=(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 l=t.kernelShape[t.kernelShape.length-1],d=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:l},{type:12,data:d},{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 w=t.kernelShape[t.kernelShape.length-2],g=t.strides[t.strides.length-2],b=t.pads[t.pads.length/2-2],x=t.pads[t.pads.length-2];h=!!(b+x),i.push({type:12,data:w},{type:12,data:g},{type:12,data:b},{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 l=k.computeStrides(t.kernelShape);i.push({type:12,data:l},{type:12,data:t.pads},{type:12,data:t.strides}),a.push({name:"kernelStrides",type:"u32",length:l.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let d=t.pads.reduce((p,m)=>p+m);return[i,a,!!d,!1,!1]}},Ud=(e,t,r,n,o,i,a,l,d,p,m,u)=>{let h=o.format==="NHWC",w=t.type.value,g=M("output",t.type.tensor,n);if(o.kernelShape.length<=2){let b="",x="",_="",$=r-(h?2:1);if(m?b=`
+        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}`,inputDependencies:a},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(E.size(r)/64)},programUniforms:i}),getShaderSource:l}},If=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),n=e.length>=3&&e[2].data?e[2].getFloat32Array()[0]:0,o=e[0].dims.length,i=new Int32Array(2*o).fill(0);if(e.length>=4){let l=e[3].getBigInt64Array();for(let d=0;d<l.length;d++)i[Number(l[d])]=Number(r[d]),i[Number(l[d])+o]=Number(r[d+l.length])}else r.forEach((l,d)=>i[Number(d)]=Number(l));let a=[];return i.forEach(l=>a.push(l)),{mode:t.mode,value:n,pads:a}}else return t},Td=(e,t)=>{bf(e.inputs);let r=If(e.inputs,t);e.compute(Sf(e.inputs,r),{inputs:[0]})}});var en,kd,Ed,Pd,Od,Cf,Tf,zd,Bd,Dd,Rd,Md,Ud,Vd,Nd,Wd,Ld,Hd,Gd,Fd=M(()=>{"use strict";qe();X();oe();re();en=e=>{if(fe.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},kd=(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(),l=t.strides.slice(),d=i?t.dilations.slice():[],p=t.pads.slice();xt.adjustPoolAttributes(r,o,a,l,d,p);let m=xt.computePoolOutputShape(r,o,l,d,a,p,t.autoPad),u=Object.assign({},t);i?Object.assign(u,{kernelShape:a,strides:l,pads:p,dilations:d,cacheKey:t.cacheKey}):Object.assign(u,{kernelShape:a,strides:l,pads:p,cacheKey:t.cacheKey});let h=m.slice();return h.push(h.splice(1,1)[0]),[u,n?h:m]},Ed=(e,t)=>{let r=t.format==="NHWC",n=E.size(e),o=E.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 l=t.kernelShape[t.kernelShape.length-1],d=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:l},{type:12,data:d},{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 w=t.kernelShape[t.kernelShape.length-2],g=t.strides[t.strides.length-2],y=t.pads[t.pads.length/2-2],x=t.pads[t.pads.length-2];h=!!(y+x),i.push({type:12,data:w},{type:12,data:g},{type:12,data:y},{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 l=E.computeStrides(t.kernelShape);i.push({type:12,data:l},{type:12,data:t.pads},{type:12,data:t.strides}),a.push({name:"kernelStrides",type:"u32",length:l.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let d=t.pads.reduce((p,m)=>p+m);return[i,a,!!d,!1,!1]}},Pd=(e,t,r,n,o,i,a,l,d,p,m,u)=>{let h=o.format==="NHWC",w=t.type.value,g=D("output",t.type.tensor,n);if(o.kernelShape.length<=2){let y="",x="",_="",$=r-(h?2:1);if(m?y=`
                 for (var i: u32 = 0u; i < uniforms.kw; i++) {
                   xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
                   if (xIndices[${$}] < 0 || xIndices[${$}]
@@ -5747,7 +5746,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
                   }
                   let x_val = x[${t.indicesToOffset("xIndices")}];
                   ${i}
-                }`:b=`
+                }`:y=`
                 for (var i: u32 = 0u; i < uniforms.kw; i++) {
                   xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
                   let x_val = x[${t.indicesToOffset("xIndices")}];
@@ -5776,12 +5775,12 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
               var value = ${w}(${l});
               var pad = 0;
               ${x}
-              ${b}
+              ${y}
               ${_}
               ${a}
 
               output[global_idx] = value;
-            }`}else{if(h)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let b=o.kernelShape.length,x=o.pads.length,_="";return p?_=`
+            }`}else{if(h)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let y=o.kernelShape.length,x=o.pads.length,_="";return p?_=`
                 if (xIndices[j] >= uniforms.x_shape[j]) {
                   pad++;
                   isPad = true;
@@ -5803,7 +5802,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
               let indices = ${g.offsetToIndices("global_idx")};
               var xIndices = ${g.offsetToIndices("global_idx")};
 
-              var offsets: array<u32, ${b}>;
+              var offsets: array<u32, ${y}>;
 
               var value = ${w}(${l});
               var pad = 0;
@@ -5811,30 +5810,68 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
 
               for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {
                 var offset = i;
-                for (var j = 0u; j < ${b-1}u; j++) {
-                  offsets[j] = offset / ${G("uniforms.kernelStrides","j",b)};
-                  offset -= offsets[j] * ${G("uniforms.kernelStrides","j",b)};
+                for (var j = 0u; j < ${y-1}u; j++) {
+                  offsets[j] = offset / ${L("uniforms.kernelStrides","j",y)};
+                  offset -= offsets[j] * ${L("uniforms.kernelStrides","j",y)};
                 }
-                offsets[${b-1}] = offset;
+                offsets[${y-1}] = offset;
 
                 isPad = false;
-                for (var j = ${r-b}u; j < ${r}u; j++) {
-                  xIndices[j] = indices[j] * ${G("uniforms.strides",`j - ${r-b}u`,b)}
-                    + offsets[j - ${r-b}u] - ${G("uniforms.pads","j - 2u",x)};
+                for (var j = ${r-y}u; j < ${r}u; j++) {
+                  xIndices[j] = indices[j] * ${L("uniforms.strides",`j - ${r-y}u`,y)}
+                    + offsets[j - ${r-y}u] - ${L("uniforms.pads","j - 2u",x)};
                   ${_}
               }
               ${a}
 
               output[global_idx] = value;
-            }`}},Vd=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,Pf=e=>`${Vd(e)};${e.countIncludePad}`,Of=e=>`${Vd(e)};${e.storageOrder};${e.dilations}`,Nd=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}),Wd=(e,t,r,n)=>{let[o,i]=Rd(t,n,r),a=E("x",t.dataType,t.dims.length),l=a.type.value,d="value += x_val;",p="";o.countIncludePad?p+=`value /= ${l}(uniforms.kernelSize);`:p+=`value /= ${l}(i32(uniforms.kernelSize) - pad);`;let[m,u,h,w,g]=Md(i,o);m.push(...V(t.dims,i));let b=["rank"];return{name:e,shaderCache:{hint:`${n.cacheKey};${h};${w};${g}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(k.size(i)/64)},programUniforms:m}),getShaderSource:x=>Ud(x,a,t.dims.length,i.length,o,d,p,0,u,h,w,g)}},Hd=e=>{let t=e.count_include_pad!==0,r=Nd(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let n={countIncludePad:t,...r,cacheKey:""};return{...n,cacheKey:Pf(n)}},Gd=(e,t)=>{nn(e.inputs),e.compute(Wd("AveragePool",e.inputs[0],!1,t))},Ld={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},Fd=e=>{let t=e.format;return{format:t,...Ld,cacheKey:t}},qd=(e,t)=>{nn(e.inputs),e.compute(Wd("GlobalAveragePool",e.inputs[0],!0,t))},jd=(e,t,r,n)=>{let[o,i]=Rd(t,n,r),a=`
+            }`}},Od=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,Cf=e=>`${Od(e)};${e.countIncludePad}`,Tf=e=>`${Od(e)};${e.storageOrder};${e.dilations}`,zd=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}),Bd=(e,t,r,n)=>{let[o,i]=kd(t,n,r),a=k("x",t.dataType,t.dims.length),l=a.type.value,d="value += x_val;",p="";o.countIncludePad?p+=`value /= ${l}(uniforms.kernelSize);`:p+=`value /= ${l}(i32(uniforms.kernelSize) - pad);`;let[m,u,h,w,g]=Ed(i,o);m.push(...U(t.dims,i));let y=["rank"];return{name:e,shaderCache:{hint:`${n.cacheKey};${h};${w};${g}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(E.size(i)/64)},programUniforms:m}),getShaderSource:x=>Pd(x,a,t.dims.length,i.length,o,d,p,0,u,h,w,g)}},Dd=e=>{let t=e.count_include_pad!==0,r=zd(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let n={countIncludePad:t,...r,cacheKey:""};return{...n,cacheKey:Cf(n)}},Rd=(e,t)=>{en(e.inputs),e.compute(Bd("AveragePool",e.inputs[0],!1,t))},Md={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},Ud=e=>{let t=e.format;return{format:t,...Md,cacheKey:t}},Vd=(e,t)=>{en(e.inputs),e.compute(Bd("GlobalAveragePool",e.inputs[0],!0,t))},Nd=(e,t,r,n)=>{let[o,i]=kd(t,n,r),a=`
       value = max(x_val, value);
-    `,l="",d=E("x",t.dataType,t.dims.length),p=["rank"],[m,u,h,w,g]=Md(i,o);return m.push(...V(t.dims,i)),{name:e,shaderCache:{hint:`${n.cacheKey};${h};${w};${g}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(k.size(i)/64)},programUniforms:m}),getShaderSource:b=>Ud(b,d,t.dims.length,i.length,o,a,l,t.dataType===10?-65504:-1e5,u,h,w,g)}},Kd=(e,t)=>{nn(e.inputs),e.compute(jd("MaxPool",e.inputs[0],!1,t))},Yd=e=>{let t=e.storage_order,r=e.dilations,n=Nd(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(n.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let o={storageOrder:t,dilations:r,...n,cacheKey:""};return{...o,cacheKey:Of(o)}},Xd=e=>{let t=e.format;return{format:t,...Ld,cacheKey:t}},Zd=(e,t)=>{nn(e.inputs),e.compute(jd("GlobalMaxPool",e.inputs[0],!0,t))}});var zf,Bf,Jd,el=U(()=>{"use strict";Ke();Q();ie();zf=(e,t,r)=>{let n=e===t,o=e<t&&r<0,i=e>t&&r>0;if(n||o||i)throw new Error("Range these inputs' contents are invalid.")},Bf=(e,t,r,n)=>{let o=Math.abs(Math.ceil((t-e)/r)),i=[o],a=o,l=[{type:12,data:a},{type:n,data:e},{type:n,data:r},...V(i)],d=p=>{let m=M("output",n,i.length),u=m.type.value,h=[{name:"outputSize",type:"u32"},{name:"start",type:u},{name:"delta",type:u}];return`
+    `,l="",d=k("x",t.dataType,t.dims.length),p=["rank"],[m,u,h,w,g]=Ed(i,o);return m.push(...U(t.dims,i)),{name:e,shaderCache:{hint:`${n.cacheKey};${h};${w};${g}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(E.size(i)/64)},programUniforms:m}),getShaderSource:y=>Pd(y,d,t.dims.length,i.length,o,a,l,t.dataType===10?-65504:-1e5,u,h,w,g)}},Wd=(e,t)=>{en(e.inputs),e.compute(Nd("MaxPool",e.inputs[0],!1,t))},Ld=e=>{let t=e.storage_order,r=e.dilations,n=zd(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:Tf(o)}},Hd=e=>{let t=e.format;return{format:t,...Md,cacheKey:t}},Gd=(e,t)=>{en(e.inputs),e.compute(Nd("GlobalMaxPool",e.inputs[0],!0,t))}});var kf,Ef,qd,Kd,jd=M(()=>{"use strict";X();oe();xe();re();kf=(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)].")}},Ef=(e,t)=>{let r=E.normalizeAxis(t.axis,e[0].dims.length),n=e[0].dataType,o=n===3,i=e[0].dims,a=e[1].dataType,l=E.size(i),d=n===3||n===2,p=d?[Math.ceil(E.size(e[0].dims)/4)]:e[0].dims,m=e[1].dims,u=e.length>2?e[2]:void 0,h=u?d?[Math.ceil(E.size(u.dims)/4)]:u.dims:void 0,w=m.length===0||m.length===1&&m[0]===1,g=w===!1&&m.length===1,y=he(l),x=w&&(!d||y===4),_=x?y:1,$=x&&!d?y:1,S=k("input",d?12:n,p.length,$),I=k("scale",a,m.length),T=u?k("zero_point",d?12:n,h.length):void 0,A=D("output",a,i.length,_),O=[S,I];T&&O.push(T);let z=[p,m];u&&z.push(h);let W=[{type:12,data:l/_},{type:12,data:r},{type:12,data:t.blockSize},...U(...z,i)],V=j=>{let Y=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
+      ${j.registerUniforms(Y).declareVariables(...O,A)}
+      ${j.mainStart()}
+          ${j.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+          let output_indices = ${A.offsetToIndices("global_idx")};
+
+          // Set input x
+          ${(()=>d?`
+            let input = ${S.getByOffset("global_idx / 4")};
+            let x_vec = ${o?"unpack4xI8(input)":"unpack4xU8(input)"};
+            let x_value = ${_===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${S.getByOffset("global_idx")};`)()};
+
+          // Set scale input
+          ${(()=>w?`let scale_value= ${I.getByOffset("0")}`:g?`
+            let scale_index = ${A.indicesGet("output_indices","uniforms.axis")};
+            let scale_value= ${I.getByOffset("scale_index")};`:`
+            var scale_indices: ${I.type.indices} = output_indices;
+            let index = ${I.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
+            ${I.indicesSet("scale_indices","uniforms.axis","index")};
+            let scale_value= ${I.getByIndices("scale_indices")};`)()};
+
+          // Set zero-point input
+          ${(()=>T?w?d?`
+                let zero_point_input = ${T.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 = ${T.getByOffset("0")}`:g?d?`
+                let zero_point_index = ${A.indicesGet("output_indices","uniforms.axis")};
+                let zero_point_input = ${T.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 = ${A.indicesGet("output_indices","uniforms.axis")};
+                let zero_point_value = ${T.getByOffset("zero_point_index")};`:d?`
+                let zero_point_offset = ${I.indicesToOffset("scale_indices")};
+                let zero_point_input = ${T.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 = ${T.getByIndices("scale_indices")};`:`let zero_point_value = ${d?o?"i32":"u32":S.type.value}(0);`)()};
+      // Compute and write output
+      ${A.setByOffset("global_idx",`${A.type.value}(x_value - zero_point_value) * scale_value`)};
+      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:T?["rank","rank","rank"]:["rank","rank"]},getShaderSource:V,getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:Math.ceil(l/_/64),y:1,z:1},programUniforms:W})}},qd=(e,t)=>{kf(e.inputs,t),e.compute(Ef(e.inputs,t))},Kd=e=>Q({axis:e.axis,blockSize:e.blockSize})});var Pf,Of,Yd,Xd=M(()=>{"use strict";qe();X();re();Pf=(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.")},Of=(e,t,r,n)=>{let o=Math.abs(Math.ceil((t-e)/r)),i=[o],a=o,l=[{type:12,data:a},{type:n,data:e},{type:n,data:r},...U(i)],d=p=>{let m=D("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:d,getRunData:()=>({outputs:[{dims:i,dataType:n}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:l})}},Jd=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]),be.webgpu.validateInputContent&&zf(t,r,n),e.compute(Bf(t,r,n,e.inputs[0].dataType),{inputs:[]})}});var Rf,Mf,Uf,Vf,Nf,Wf,Hf,Gf,Lf,Ff,qf,tl,jf,Kf,Yf,Xf,Zf,rl,nl,ol=U(()=>{"use strict";Q();ae();Ce();ie();Rf=(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")}},Mf=(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},Uf=(e,t,r,n,o,i)=>{let[a,l,d]=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(l>0&&e.length>l&&e[l].dims.length>0){if(e[l].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");Rf(n,t),t.axes.length>0&&Mf(n,t.axes,p).forEach((m,u)=>n[u]=m)}if(d>0&&e.length>d&&(e[d].getBigInt64Array().forEach(m=>o.push(Number(m))),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!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(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")},Vf=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
+      }`};return{name:"Range",shaderCache:{hint:`${n}`},getShaderSource:d,getRunData:()=>({outputs:[{dims:i,dataType:n}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:l})}},Yd=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]),fe.webgpu.validateInputContent&&Pf(t,r,n),e.compute(Of(t,r,n,e.inputs[0].dataType),{inputs:[]})}});var zf,Bf,Df,Rf,Mf,Uf,Vf,Nf,Wf,Lf,Hf,Zd,Gf,Ff,qf,Kf,jf,Qd,Jd,el=M(()=>{"use strict";X();oe();xe();re();zf=(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")}},Bf=(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},Df=(e,t,r,n,o,i)=>{let[a,l,d]=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(l>0&&e.length>l&&e[l].dims.length>0){if(e[l].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");zf(n,t),t.axes.length>0&&Bf(n,t.axes,p).forEach((m,u)=>n[u]=m)}if(d>0&&e.length>d&&(e[d].getBigInt64Array().forEach(m=>o.push(Number(m))),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!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(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")},Rf=(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) / ${t}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) {
                     return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
                   } else {
@@ -5859,38 +5896,38 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
                   const adjustment = ${t}(lengthResized) / outputWidth;
                   const center = ${t}(lengthOriginal) / 2;
                   const offset = center * (1 - adjustment);
-                  return offset + ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;case"half_pixel":return`return ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${e} is not supported`)}})()+"}",Nf=(e,t,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(e){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) {             return ceil(xOriginal);           } else {             return round(xOriginal);           }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) {                     return floor(xOriginal);                   } else {                     return round(xOriginal);                   }";case"simple":default:if(t<11)return"if (isDownSample)                     {                       return ceil(xOriginal);                     } else {                       return xOriginal;                     }";throw new Error(`Nearest mode ${e} is not supported`)}})()+"}",Wf=(e,t,r)=>{let n=new Array(r).fill(0).concat(new Array(r).fill(1)),o=e.length===0?n:e.slice();return t.length>0?(t.forEach((i,a)=>{n[i]=o[a],n[a+r]=o[t.length+a]}),n):o},Hf=(e,t,r,n)=>{let o=[];if(r.length>0)if(n.length>0){if(e.forEach(i=>o.push(i)),Math.max(...n)>e.length)throw new Error("axes is out of bound");n.forEach((i,a)=>o[i]=r[a])}else r.forEach(i=>o.push(i));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");o=e.map((i,a)=>Math.round(i*t[a]))}return o},Gf=(e,t,r)=>{let n=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(i=>t[i]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(i=>t[i]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let o=e.slice();return r.axes.length>0?(r.axes.forEach(i=>t[i]=n),r.axes.forEach(i=>o[i]=Math.round(e[i]*t[i]))):(t.fill(n,0,t.length),o.forEach((i,a)=>o[a]=Math.round(i*t[a]))),o},Lf=(e,t,r,n,o)=>`
+                  return offset + ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;case"half_pixel":return`return ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${e} is not supported`)}})()+"}",Mf=(e,t,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(e){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) {             return ceil(xOriginal);           } else {             return round(xOriginal);           }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) {                     return floor(xOriginal);                   } else {                     return round(xOriginal);                   }";case"simple":default:if(t<11)return"if (isDownSample)                     {                       return ceil(xOriginal);                     } else {                       return xOriginal;                     }";throw new Error(`Nearest mode ${e} is not supported`)}})()+"}",Uf=(e,t,r)=>{let n=new Array(r).fill(0).concat(new Array(r).fill(1)),o=e.length===0?n:e.slice();return t.length>0?(t.forEach((i,a)=>{n[i]=o[a],n[a+r]=o[t.length+a]}),n):o},Vf=(e,t,r,n)=>{let o=[];if(r.length>0)if(n.length>0){if(e.forEach(i=>o.push(i)),Math.max(...n)>e.length)throw new Error("axes is out of bound");n.forEach((i,a)=>o[i]=r[a])}else r.forEach(i=>o.push(i));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");o=e.map((i,a)=>Math.round(i*t[a]))}return o},Nf=(e,t,r)=>{let n=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(i=>t[i]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(i=>t[i]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let o=e.slice();return r.axes.length>0?(r.axes.forEach(i=>t[i]=n),r.axes.forEach(i=>o[i]=Math.round(e[i]*t[i]))):(t.fill(n,0,t.length),o.forEach((i,a)=>o[a]=Math.round(i*t[a]))),o},Wf=(e,t,r,n,o)=>`
     fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {
       var original_indices: array<${e.type.value}, ${r.length}>;
       for (var i:u32 = 0; i < ${r.length}; i++) {
         var output_index = ${e.indicesGet("output_indices","i")};
-        var scale = ${G("uniforms.scales","i",n)};
-        var roi_low = ${G("uniforms.roi","i",o)};
-        var roi_hi = ${G("uniforms.roi",`i + ${t.length}`,o)};
+        var scale = ${L("uniforms.scales","i",n)};
+        var roi_low = ${L("uniforms.roi","i",o)};
+        var roi_hi = ${L("uniforms.roi",`i + ${t.length}`,o)};
         if (scale == 1.0) {
           original_indices[i] = ${e.type.value}(output_index);
         } else {
-          var input_shape_i = ${G("uniforms.input_shape","i",t.length)};
-          var output_shape_i = ${G("uniforms.output_shape","i",r.length)};
+          var input_shape_i = ${L("uniforms.input_shape","i",t.length)};
+          var output_shape_i = ${L("uniforms.output_shape","i",r.length)};
           original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
                                                                            input_shape_i, roi_low, roi_hi);
         }
       }
       return original_indices;
-    }`,Ff=(e,t,r,n,o,i,a)=>`
+    }`,Lf=(e,t,r,n,o,i,a)=>`
     fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
       var input_indices: ${e.type.indices};
       for (var i:u32 = 0; i < ${n.length}; i++) {
         var output_index = ${t.indicesGet("output_indices","i")};
         var input_index: u32;
-        var scale = ${G("uniforms.scales","i",o)};
+        var scale = ${L("uniforms.scales","i",o)};
         if (scale == 1.0) {
           input_index = output_index;
         } else {
-          var roi_low = ${G("uniforms.roi","i",i)};
-          var roi_hi = ${G("uniforms.roi",`i + ${r.length}`,i)};
-          var input_shape_i = ${G("uniforms.input_shape","i",r.length)};
-          var output_shape_i = ${G("uniforms.output_shape","i",n.length)};
+          var roi_low = ${L("uniforms.roi","i",i)};
+          var roi_hi = ${L("uniforms.roi",`i + ${r.length}`,i)};
+          var input_shape_i = ${L("uniforms.input_shape","i",r.length)};
+          var output_shape_i = ${L("uniforms.output_shape","i",n.length)};
           var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
                                                                         input_shape_i, roi_low, roi_hi);
           if (!${a} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {
@@ -5908,24 +5945,24 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
         ${e.indicesSet("input_indices","i"," input_index")}
       }
       return input_indices;
-    }`,qf=(e,t)=>`
+    }`,Hf=(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 >= ${G("uniforms.input_shape","i",t.length)}) {
+        if (input_index < 0 || input_index >= ${L("uniforms.input_shape","i",t.length)}) {
           return false;
         }
       }
       return true;
-    }`,tl=(e,t,r,n)=>e.rank>n?`
+    }`,Zd=(e,t,r,n)=>e.rank>n?`
     ${e.indicesSet("input_indices",t,"channel")};
     ${e.indicesSet("input_indices",r,"batch")};
-`:"",jf=(e,t,r,n,o)=>{let[a,l,d,p]=r.length===2?[-1,0,1,-1]:[0,2,3,1],m=e.type.value;return`
+`:"",Gf=(e,t,r,n,o)=>{let[a,l,d,p]=r.length===2?[-1,0,1,-1]:[0,2,3,1],m=e.type.value;return`
     fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${m} {
       var input_indices: ${e.type.indices};
       ${e.indicesSet("input_indices",l,`max(0, min(row, ${r[l]} - 1))`)};
       ${e.indicesSet("input_indices",d,`max(0, min(col, ${r[d]} - 1))`)};
-      ${tl(e,p,a,2)}
+      ${Zd(e,p,a,2)}
       return ${e.getByIndices("input_indices")};
     }
 
@@ -5961,7 +5998,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
         dy2 = 0.5;
       }
       return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);
-    }`},Kf=(e,t,r,n,o,i,a,l,d,p)=>{let m=r.length===2,u=!0,[h,w]=m?[0,1]:u?[2,3]:[1,2],g=e.type.value,b=x=>{let _=x===h?"row":"col";return`
+    }`},Ff=(e,t,r,n,o,i,a,l,d,p)=>{let m=r.length===2,u=!0,[h,w]=m?[0,1]:u?[2,3]:[1,2],g=e.type.value,y=x=>{let _=x===h?"row":"col";return`
       fn ${_}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${g} {
         var output_index = ${t.indicesGet("output_indices",x)};
         var originalIdx: ${g} = getOriginalCoordinateFromResizedCoordinate(output_index, ${o[x]},
@@ -5985,8 +6022,8 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
         }
         return cubicInterpolation1D(data, coefs);
       }`};return`
-    ${b(h)};
-    ${b(w)};
+    ${y(h)};
+    ${y(w)};
   fn getCubicInterpolationCoefs(s: ${g}) -> array<${g}, 4> {
     var absS = abs(s);
     var coeffs: array<${g}, 4> = array<${g}, 4>(0.0, 0.0, 0.0, 0.0);
@@ -6009,13 +6046,13 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
     var input_indices: ${e.type.indices} = output_indices;
     return colCubicInterpolation(input_indices, output_indices);
   }
-    `},Yf=(e,t,r,n,o)=>{let[a,l,d,p,m]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],u=e.type.value;return`
+    `},qf=(e,t,r,n,o)=>{let[a,l,d,p,m]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],u=e.type.value;return`
     fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${u} {
       var input_indices: ${e.type.indices};
       ${e.indicesSet("input_indices",l,`max(0, min(depth, ${r[l]} - 1))`)};
       ${e.indicesSet("input_indices",d,`max(0, min(height, ${r[d]} - 1))`)};
       ${e.indicesSet("input_indices",p,`max(0, min(width, ${r[p]} - 1))`)};
-      ${tl(e,m,a,3)}
+      ${Zd(e,m,a,3)}
       return ${e.getByIndices("input_indices")};
     }
 
@@ -6068,18 +6105,18 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
       }
       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);
-    }`},Xf=(e,t,r,n,o,i)=>{let a=e.dims,l=Wf(i,t.axes,a.length),d=Hf(a,n,o,t.axes),p=n.slice();n.length===0&&(p=a.map(($,S)=>$===0?1:d[S]/$),t.keepAspectRatioPolicy!=="stretch"&&(d=Gf(a,p,t)));let m=M("output",e.dataType,d.length),u=E("input",e.dataType,a.length),h=k.size(d),w=a.length===d.length&&a.every(($,S)=>$===d[S]),g=t.coordinateTransformMode==="tf_crop_and_resize",b=t.extrapolationValue,x=u.type.value,_=$=>`
+    }`},Kf=(e,t,r,n,o,i)=>{let a=e.dims,l=Uf(i,t.axes,a.length),d=Vf(a,n,o,t.axes),p=n.slice();n.length===0&&(p=a.map(($,S)=>$===0?1:d[S]/$),t.keepAspectRatioPolicy!=="stretch"&&(d=Nf(a,p,t)));let m=D("output",e.dataType,d.length),u=k("input",e.dataType,a.length),h=E.size(d),w=a.length===d.length&&a.every(($,S)=>$===d[S]),g=t.coordinateTransformMode==="tf_crop_and_resize",y=t.extrapolationValue,x=u.type.value,_=$=>`
       ${w?"":`
-      ${Vf(t.coordinateTransformMode,x)};
+      ${Rf(t.coordinateTransformMode,x)};
       ${(()=>{switch(t.mode){case"nearest":return`
-              ${qf(u,a)};
-              ${Nf(t.nearestMode,r,x)};
-              ${Ff(u,m,a,d,p.length,l.length,g)};
+              ${Hf(u,a)};
+              ${Mf(t.nearestMode,r,x)};
+              ${Lf(u,m,a,d,p.length,l.length,g)};
               `;case"linear":return`
-              ${Lf(m,a,d,p.length,l.length)};
-              ${(()=>{if(a.length===2||a.length===4)return`${jf(u,m,a,g,b)}`;if(a.length===3||a.length===5)return`${Yf(u,m,a,g,b)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
+              ${Wf(m,a,d,p.length,l.length)};
+              ${(()=>{if(a.length===2||a.length===4)return`${Gf(u,m,a,g,y)}`;if(a.length===3||a.length===5)return`${qf(u,m,a,g,y)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
             `;case"cubic":return`
-            ${(()=>{if(a.length===2||a.length===4)return`${Kf(u,m,a,d,p,l,t.cubicCoeffA,g,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
+            ${(()=>{if(a.length===2||a.length===4)return`${Ff(u,m,a,d,p,l,t.cubicCoeffA,g,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
             `;default:throw Error("Invalid resize mode")}})()};
       `}
       ${$.registerUniform("output_size","u32").registerUniform("scales","f32",p.length).registerUniform("roi","f32",l.length).declareVariables(u,m)}
@@ -6095,10 +6132,10 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
                   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:""}|${o.length>0?o:""}|${l.length>0?l:""}|${w}|${a}`,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:[{dims:d,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:[{type:12,data:h},{type:1,data:p},{type:1,data:l},...V(a,d)]})}},Zf=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},rl=(e,t)=>{let r=[],n=[],o=[],i=Zf(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");Uf(e.inputs,t,i,r,n,o),e.compute(Xf(e.inputs[0],t,i,r,n,o),{inputs:[0]})},nl=e=>{let t=e.antialias,r=e.axes,n=e.coordinateTransformMode,o=e.cubicCoeffA,i=e.excludeOutside!==0,a=e.extrapolationValue,l=e.keepAspectRatioPolicy,d=e.mode,p=e.nearestMode===""?"simple":e.nearestMode;return te({antialias:t,axes:r,coordinateTransformMode:n,cubicCoeffA:o,excludeOutside:i,extrapolationValue:a,keepAspectRatioPolicy:l,mode:d,nearestMode:p})}});var Qf,Jf,il,al=U(()=>{"use strict";Q();ae();Ce();ie();Qf=(e,t)=>{let[r,n,o,i]=e,{numHeads:a,rotaryEmbeddingDim:l}=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(l>0&&a===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let d=r.dims[0],p=r.dims[r.dims.length-2],m=o.dims[0],u=k.sizeFromDimension(r.dims,1)/p,h=l===0?o.dims[1]*2:u/a;if(l>h)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(n.dims.length===2){if(d!==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]&&l/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 ${o.dims[1]}`);if(p>m)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Jf=(e,t)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:o,scale:i}=t,a=e[0].dims[0],l=k.sizeFromDimension(e[0].dims,1),d=e[0].dims[e[0].dims.length-2],p=l/d,m=e[2].dims[1],u=o===0?m*2:p/n,h=new Array(a,d,p/u,u-m),w=k.computeStrides(h),g=[{type:1,data:i},{type:12,data:h},{type:12,data:w},...e[0].dims.length===3?new Array({type:12,data:[l,p,u,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[l,u,d*u,1]}):[],...V(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],b=x=>{let _=E("input",e[0].dataType,e[0].dims.length),$=E("position_ids",e[1].dataType,e[1].dims.length),S=E("cos_cache",e[2].dataType,e[2].dims.length),I=E("sin_cache",e[3].dataType,e[3].dims.length),T=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:w.length},{name:"input_output_strides",type:"u32",length:w.length}]),`
+      }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${p.length>0?p:""}|${o.length>0?o:""}|${l.length>0?l:""}|${w}|${a}`,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:[{dims:d,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:[{type:12,data:h},{type:1,data:p},{type:1,data:l},...U(a,d)]})}},jf=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},Qd=(e,t)=>{let r=[],n=[],o=[],i=jf(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");Df(e.inputs,t,i,r,n,o),e.compute(Kf(e.inputs[0],t,i,r,n,o),{inputs:[0]})},Jd=e=>{let t=e.antialias,r=e.axes,n=e.coordinateTransformMode,o=e.cubicCoeffA,i=e.excludeOutside!==0,a=e.extrapolationValue,l=e.keepAspectRatioPolicy,d=e.mode,p=e.nearestMode===""?"simple":e.nearestMode;return Q({antialias:t,axes:r,coordinateTransformMode:n,cubicCoeffA:o,excludeOutside:i,extrapolationValue:a,keepAspectRatioPolicy:l,mode:d,nearestMode:p})}});var Yf,Xf,tl,rl=M(()=>{"use strict";X();oe();xe();re();Yf=(e,t)=>{let[r,n,o,i]=e,{numHeads:a,rotaryEmbeddingDim:l}=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(!E.areEqual(n.dims,[])&&!E.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(!E.areEqual(o.dims,i.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(l>0&&a===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let d=r.dims[0],p=r.dims[r.dims.length-2],m=o.dims[0],u=E.sizeFromDimension(r.dims,1)/p,h=l===0?o.dims[1]*2:u/a;if(l>h)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(n.dims.length===2){if(d!==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]&&l/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 ${o.dims[1]}`);if(p>m)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Xf=(e,t)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:o,scale:i}=t,a=e[0].dims[0],l=E.sizeFromDimension(e[0].dims,1),d=e[0].dims[e[0].dims.length-2],p=l/d,m=e[2].dims[1],u=o===0?m*2:p/n,h=new Array(a,d,p/u,u-m),w=E.computeStrides(h),g=[{type:1,data:i},{type:12,data:h},{type:12,data:w},...e[0].dims.length===3?new Array({type:12,data:[l,p,u,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[l,u,d*u,1]}):[],...U(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],y=x=>{let _=k("input",e[0].dataType,e[0].dims.length),$=k("position_ids",e[1].dataType,e[1].dims.length),S=k("cos_cache",e[2].dataType,e[2].dims.length),I=k("sin_cache",e[3].dataType,e[3].dims.length),T=D("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:w.length},{name:"input_output_strides",type:"u32",length:w.length}]),`
         ${x.declareVariables(_,$,S,I,T)}
 
-        ${x.mainStart(Tt)}
+        ${x.mainStart(St)}
           let half_rotary_emb_dim = uniforms.${S.name}_shape[1];
           let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
           let size = uniforms.global_shape[0] * uniforms.global_strides[0];
@@ -6106,7 +6143,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
 
           if (bsnh[3] < half_rotary_emb_dim) {
             let position_ids_idx =
-                ${$.broadcastedIndicesToOffset("bsnh.xy",M("",$.type.tensor,2))};
+                ${$.broadcastedIndicesToOffset("bsnh.xy",D("",$.type.tensor,2))};
             let position_id =
                 u32(${$.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);
             let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${r});
@@ -6121,11 +6158,11 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;
             ${T.setByOffset("k",_.getByOffset("k"))}
           }
-        }`};return{name:"RotaryEmbedding",shaderCache:{hint:te({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:b,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(k.size(h)/Tt)},programUniforms:g})}},il=(e,t)=>{Qf(e.inputs,t),e.compute(Jf(e.inputs,t))}});var eh,th,sl,ul=U(()=>{"use strict";Q();ae();ie();eh=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],n=e[2];if(t.dataType!==r.dataType||t.dataType!==n.dataType)throw new Error("All inputs must have the same data type");if(t.dims.length!==3&&t.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let o=t.dims[t.dims.length-1],i=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==o)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==i)throw new Error("Skip must have the same sequence length as input");if(n.dims.length!==1)throw new Error("Gamma must be 1D");if(n.dims[n.dims.length-1]!==o)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let a=e[3];if(a.dims.length!==1)throw new Error("Beta must be 1D");if(a.dims[a.dims.length-1]!==o)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let a=e[4];if(a.dims.length!==1)throw new Error("Bias must be 1D");if(a.dims[a.dims.length-1]!==o)throw new Error("Bias must have the same hidden size as input")}},th=(e,t,r,n)=>{let o=t.simplified,i=e[0].dims,a=k.size(i),l=i,d=a,p=i.slice(-1)[0],m=n?i.slice(0,-1).concat(1):[],u=!o&&e.length>3,h=e.length>4,w=n&&r>1,g=n&&r>2,b=r>3,x=64,_=$e(p),$=[{type:12,data:d},{type:12,data:_},{type:12,data:p},{type:1,data:t.epsilon}],S=T=>{let A=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],D=[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&&D.push(E("beta",e[3].dataType,e[3].dims,_)),h&&D.push(E("bias",e[4].dataType,e[4].dims,_)),D.push(M("output",e[0].dataType,l,_)),w&&D.push(M("mean_output",1,m)),g&&D.push(M("inv_std_output",1,m)),b&&D.push(M("input_skip_bias_sum",e[0].dataType,l,_));let z=ge(e[0].dataType),H=ge(1,_);return`
+        }`};return{name:"RotaryEmbedding",shaderCache:{hint:Q({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:y,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(E.size(h)/St)},programUniforms:g})}},tl=(e,t)=>{Yf(e.inputs,t),e.compute(Xf(e.inputs,t))}});var Zf,Qf,nl,ol=M(()=>{"use strict";X();oe();re();Zf=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")}},Qf=(e,t,r,n)=>{let o=t.simplified,i=e[0].dims,a=E.size(i),l=i,d=a,p=i.slice(-1)[0],m=n?i.slice(0,-1).concat(1):[],u=!o&&e.length>3,h=e.length>4,w=n&&r>1,g=n&&r>2,y=r>3,x=64,_=he(p),$=[{type:12,data:d},{type:12,data:_},{type:12,data:p},{type:1,data:t.epsilon}],S=T=>{let A=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],O=[k("x",e[0].dataType,e[0].dims,_),k("skip",e[1].dataType,e[1].dims,_),k("gamma",e[2].dataType,e[2].dims,_)];u&&O.push(k("beta",e[3].dataType,e[3].dims,_)),h&&O.push(k("bias",e[4].dataType,e[4].dims,_)),O.push(D("output",e[0].dataType,l,_)),w&&O.push(D("mean_output",1,m)),g&&O.push(D("inv_std_output",1,m)),y&&O.push(D("input_skip_bias_sum",e[0].dataType,l,_));let z=pe(e[0].dataType),W=pe(1,_);return`
 
-      ${T.registerUniforms(A).declareVariables(...D)}
-      var<workgroup> sum_shared : array<${H}, ${x}>;
-      var<workgroup> sum_squared_shared : array<${H}, ${x}>;
+      ${T.registerUniforms(A).declareVariables(...O)}
+      var<workgroup> sum_shared : array<${W}, ${x}>;
+      var<workgroup> sum_squared_shared : array<${W}, ${x}>;
 
       ${T.mainStart([x,1,1])}
         let ix = local_id.x;
@@ -6143,9 +6180,9 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
           let bias_value = ${h?"bias[offset1d + i]":z+"(0.0)"};
           let input_value = x[offset + i];
           let value = input_value + skip_value + bias_value;
-          ${b?"input_skip_bias_sum[offset + i] = value;":""}
+          ${y?"input_skip_bias_sum[offset + i] = value;":""}
           output[offset + i] = value;
-          let f32_value = ${At(z,_,"value")};
+          let f32_value = ${It(z,_,"value")};
           sum_shared[ix] += f32_value;
           sum_squared_shared[ix] += f32_value * f32_value;
         }
@@ -6163,8 +6200,8 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
 
         let sum = sum_shared[0];
         let square_sum = sum_squared_shared[0];
-        let mean = ${Xe("sum",_)} / f32(uniforms.hidden_size);
-        let inv_std_dev = inverseSqrt(${Xe("square_sum",_)} / f32(uniforms.hidden_size) ${o?"":"- mean * mean"} + uniforms.epsilon);
+        let mean = ${je("sum",_)} / f32(uniforms.hidden_size);
+        let inv_std_dev = inverseSqrt(${je("square_sum",_)} / f32(uniforms.hidden_size) ${o?"":"- mean * mean"} + uniforms.epsilon);
         ${w?"mean_output[global_idx] = mean;":""}
         ${g?"inv_std_output[global_idx] = inv_std_dev;":""}
 
@@ -6173,14 +6210,14 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             ${z}(inv_std_dev) * gamma[offset1d + i]
             ${u?"+ beta[offset1d + i]":""};
         }
-      }`},I=[{dims:l,dataType:e[0].dataType}];return r>1&&I.push({dims:m,dataType:1}),r>2&&I.push({dims:m,dataType:1}),r>3&&I.push({dims:i,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${_};${w};${g};${b}`,inputDependencies:e.map((T,A)=>"type")},getShaderSource:S,getRunData:()=>({outputs:I,dispatchGroup:{x:Math.ceil(d/p)},programUniforms:$})}},sl=(e,t)=>{eh(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(th(e.inputs,t,e.outputCount,!1),{outputs:n})}});var rh,on,nh,dl,oh,ih,ll,cl,pl=U(()=>{"use strict";Q();ae();Ce();ie();rh=(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`)})},on=(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},nh=(e,t)=>{if(e.length>1){let r=on(e,1),n=on(e,2),o=on(e,3);return o.length===0&&(o=[...Array(e[0].dims.length).keys()]),te({starts:r,ends:n,axes:o})}else return t},dl=(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]]))},oh=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
+      }`},I=[{dims:l,dataType:e[0].dataType}];return r>1&&I.push({dims:m,dataType:1}),r>2&&I.push({dims:m,dataType:1}),r>3&&I.push({dims:i,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${_};${w};${g};${y}`,inputDependencies:e.map((T,A)=>"type")},getShaderSource:S,getRunData:()=>({outputs:I,dispatchGroup:{x:Math.ceil(d/p)},programUniforms:$})}},nl=(e,t)=>{Zf(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(Qf(e.inputs,t,e.outputCount,!1),{outputs:n})}});var Jf,tn,eh,il,th,rh,al,sl,ul=M(()=>{"use strict";X();oe();xe();re();Jf=(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},eh=(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()]),Q({starts:r,ends:n,axes:o})}else return t},il=(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]]))},th=(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 = ${G("uniforms.input_shape","i",r.length)};
-            let steps_i = ${G("uniforms.steps","i",r.length)};
-            let signs_i = ${G("uniforms.signs","i",r.length)};
-            let starts_i = ${G("uniforms.starts","i",r.length)};
+            let input_shape_i = ${L("uniforms.input_shape","i",r.length)};
+            let steps_i = ${L("uniforms.steps","i",r.length)};
+            let signs_i = ${L("uniforms.signs","i",r.length)};
+            let starts_i = ${L("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;
@@ -6191,15 +6228,15 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
             ${e.indicesSet("input_indices","i","input_index")};
           }
           return input_indices;
-      }`,ih=(e,t)=>{let r=e[0].dims,n=k.size(r),o=t.axes.length>0?k.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],i=on(e,4);i.forEach(_=>_!==0||(()=>{throw new Error("step cannot be 0")})),i.length===0&&(i=Array(o.length).fill(1));let a=t.starts.map((_,$)=>dl(_,$,r,o,i)),l=t.ends.map((_,$)=>dl(_,$,r,o,i));if(o.length!==a.length||o.length!==l.length)throw new Error("start, ends and axes should have the same number of elements");if(o.length!==r.length)for(let _=0;_<r.length;++_)o.includes(_)||(a.splice(_,0,0),l.splice(_,0,r[_]),i.splice(_,0,1));let d=i.map(_=>Math.sign(_));i.forEach((_,$,S)=>{if(_<0){let I=(l[$]-a[$])/_,T=a[$],A=T+I*i[$];a[$]=A,l[$]=T,S[$]=-_}});let p=r.slice(0);o.forEach((_,$)=>{p[_]=Math.ceil((l[_]-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),w=k.size(p),g=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:a.length},{name:"signs",type:"i32",length:d.length},{name:"steps",type:"u32",length:i.length}],b=[{type:12,data:w},{type:12,data:a},{type:6,data:d},{type:12,data:i},...V(e[0].dims,p)],x=_=>`
+      }`,rh=(e,t)=>{let r=e[0].dims,n=E.size(r),o=t.axes.length>0?E.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((_,$)=>il(_,$,r,o,i)),l=t.ends.map((_,$)=>il(_,$,r,o,i));if(o.length!==a.length||o.length!==l.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),l.splice(_,0,r[_]),i.splice(_,0,1));let d=i.map(_=>Math.sign(_));i.forEach((_,$,S)=>{if(_<0){let I=(l[$]-a[$])/_,T=a[$],A=T+I*i[$];a[$]=A,l[$]=T,S[$]=-_}});let p=r.slice(0);o.forEach((_,$)=>{p[_]=Math.ceil((l[_]-a[_])/i[_])});let m={dims:p,dataType:e[0].dataType},u=D("output",e[0].dataType,p.length),h=k("input",e[0].dataType,e[0].dims.length),w=E.size(p),g=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:a.length},{name:"signs",type:"i32",length:d.length},{name:"steps",type:"u32",length:i.length}],y=[{type:12,data:w},{type:12,data:a},{type:6,data:d},{type:12,data:i},...U(e[0].dims,p)],x=_=>`
       ${_.registerUniforms(g).declareVariables(h,u)}
-        ${oh(h,u,r)}
+        ${th(h,u,r)}
         ${_.mainStart()}
           ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
           let output_indices = ${u.offsetToIndices("global_idx")};
           let input_indices = calculateInputIndices(output_indices);
           ${u.setByOffset("global_idx",h.getByIndices("input_indices"))}
-      }`;return{name:"Slice",shaderCache:{hint:`${d.length}_${a.length}_${i.length}`,inputDependencies:["rank"]},getShaderSource:x,getRunData:()=>({outputs:[m],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:b})}},ll=(e,t)=>{rh(e.inputs,t);let r=nh(e.inputs,t);e.compute(ih(e.inputs,r),{inputs:[0]})},cl=e=>{let t=e.starts,r=e.ends,n=e.axes;return te({starts:t,ends:r,axes:n})}});var ah,sh,ml,fl,hl=U(()=>{"use strict";Q();ae();Ce();ie();ah=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},sh=(e,t)=>{let r=e.dims,n=k.size(r),o=64,i=t.axis;if(i<0&&(i=r.length+i),i<r.length-1)throw new Error("softmax only supports last axis for now.");let a=r[i],l=n/a,d=$e(a),p=a/d,m=(x,_)=>_===4?`max(max(${x}.x, ${x}.y), max(${x}.z, ${x}.w))`:_===2?`max(${x}.x, ${x}.y)`:_===3?`max(max(${x}.x, ${x}.y), ${x}.z)`:x,u=E("x",e.dataType,e.dims,d),h=M("result",e.dataType,e.dims,d),w=u.type.value,g=ge(e.dataType)==="f32"?`var threadMax = ${w}(-3.402823e+38f);`:`var threadMax = ${w}(-65504.0h);`,b=x=>`
+      }`;return{name:"Slice",shaderCache:{hint:`${d.length}_${a.length}_${i.length}`,inputDependencies:["rank"]},getShaderSource:x,getRunData:()=>({outputs:[m],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:y})}},al=(e,t)=>{Jf(e.inputs,t);let r=eh(e.inputs,t);e.compute(rh(e.inputs,r),{inputs:[0]})},sl=e=>{let t=e.starts,r=e.ends,n=e.axes;return Q({starts:t,ends:r,axes:n})}});var nh,oh,dl,ll,cl=M(()=>{"use strict";X();oe();xe();re();nh=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},oh=(e,t)=>{let r=e.dims,n=E.size(r),o=64,i=t.axis;if(i<0&&(i=r.length+i),i<r.length-1)throw new Error("softmax only supports last axis for now.");let a=r[i],l=n/a,d=he(a),p=a/d,m=(x,_)=>_===4?`max(max(${x}.x, ${x}.y), max(${x}.z, ${x}.w))`:_===2?`max(${x}.x, ${x}.y)`:_===3?`max(max(${x}.x, ${x}.y), ${x}.z)`:x,u=k("x",e.dataType,e.dims,d),h=D("result",e.dataType,e.dims,d),w=u.type.value,g=pe(e.dataType)==="f32"?`var threadMax = ${w}(-3.402823e+38f);`:`var threadMax = ${w}(-65504.0h);`,y=x=>`
       var<workgroup> rowMaxShared : ${w};
       var<workgroup> rowSumShared : ${w};
       var<workgroup> threadShared : array<${w}, ${o}>;
@@ -6262,7 +6299,7 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
           workgroupBarrier();
         }
         if (lindex == 0) {
-          rowSumShared = ${w}(${Xe("threadShared[0]",d)});
+          rowSumShared = ${w}(${je("threadShared[0]",d)});
         }
         workgroupBarrier();
 
@@ -6271,22 +6308,22 @@ ${ke}_indices[${ce}] = 0;`}),F+=`${ke}_indices[${J-2}] = 0u;
           let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;
           setValue(row, col, row_stride, value);
         }
-      }`;return{name:"Softmax",shaderCache:{hint:`${d}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.dataType}],dispatchGroup:{x:l},programUniforms:[{type:6,data:p}]}),getShaderSource:b}},ml=(e,t)=>{ah(e.inputs),e.compute(sh(e.inputs[0],t))},fl=e=>te({axis:e.axis})});var uh,dh,lh,ch,ph,gl,yl,bl=U(()=>{"use strict";Q();ae();Ce();ie();uh=e=>{if(!e||e.length<1)throw new Error("too few inputs")},dh=(e,t)=>{let r=[],n=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(o=>r.push(Number(o))),n=r.length),te({numOutputs:n,axis:t.axis,splitSizes:r})},lh=e=>`
+      }`;return{name:"Softmax",shaderCache:{hint:`${d}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.dataType}],dispatchGroup:{x:l},programUniforms:[{type:6,data:p}]}),getShaderSource:y}},dl=(e,t)=>{nh(e.inputs),e.compute(oh(e.inputs[0],t))},ll=e=>Q({axis:e.axis})});var ih,ah,sh,uh,dh,pl,ml,fl=M(()=>{"use strict";X();oe();xe();re();ih=e=>{if(!e||e.length<1)throw new Error("too few inputs")},ah=(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),Q({numOutputs:n,axis:t.axis,splitSizes:r})},sh=e=>`
 fn calculateOutputIndex(index: u32) -> u32 {
     for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
-    if (index < ${G("uniforms.size_in_split_axis","i",e)}) {
+    if (index < ${L("uniforms.size_in_split_axis","i",e)}) {
         return i;
     }
     }
     return ${e}u;
-}`,ch=e=>{let t=e.length,r=[];for(let n=0;n<t;++n){let o=e[n].setByIndices("indices","input[global_idx]");t===1?r.push(o):n===0?r.push(`if (output_number == ${n}u) { ${o} }`):n===t-1?r.push(`else { ${o} }`):r.push(`else if (output_number == ${n}) { ${o} }`)}return`
+}`,uh=e=>{let t=e.length,r=[];for(let n=0;n<t;++n){let o=e[n].setByIndices("indices","input[global_idx]");t===1?r.push(o):n===0?r.push(`if (output_number == ${n}u) { ${o} }`):n===t-1?r.push(`else { ${o} }`):r.push(`else if (output_number == ${n}) { ${o} }`)}return`
       fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
         ${r.join(`
 `)}
-      }`},ph=(e,t)=>{let r=e[0].dims,n=k.size(r),o=e[0].dataType,i=k.normalizeAxis(t.axis,r.length),a=new Array(t.numOutputs),l=E("input",o,r.length),d=new Array(t.numOutputs),p=[],m=[],u=0,h=[{type:12,data:n}];for(let g=0;g<t.numOutputs;g++){u+=t.splitSizes[g],d[g]=u;let b=r.slice();b[t.axis]=t.splitSizes[g],m.push(b),a[g]=M(`output${g}`,o,b.length),p.push({dims:m[g],dataType:e[0].dataType})}h.push({type:12,data:d},...V(r,...m));let w=g=>`
+      }`},dh=(e,t)=>{let r=e[0].dims,n=E.size(r),o=e[0].dataType,i=E.normalizeAxis(t.axis,r.length),a=new Array(t.numOutputs),l=k("input",o,r.length),d=new Array(t.numOutputs),p=[],m=[],u=0,h=[{type:12,data:n}];for(let g=0;g<t.numOutputs;g++){u+=t.splitSizes[g],d[g]=u;let y=r.slice();y[i]=t.splitSizes[g],m.push(y),a[g]=D(`output${g}`,o,y.length),p.push({dims:m[g],dataType:e[0].dataType})}h.push({type:12,data:d},...U(r,...m));let w=g=>`
   ${g.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",d.length).declareVariables(l,...a)}
-  ${lh(d.length)}
-  ${ch(a)}
+  ${sh(d.length)}
+  ${uh(a)}
 
   ${g.mainStart()}
     ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}
@@ -6295,11 +6332,11 @@ fn calculateOutputIndex(index: u32) -> u32 {
     var index = ${l.indicesGet("indices",i)};
     let output_number = calculateOutputIndex(index);
     if (output_number != 0) {
-      index -= ${G("uniforms.size_in_split_axis","output_number - 1u",d.length)};
+      index -= ${L("uniforms.size_in_split_axis","output_number - 1u",d.length)};
       ${l.indicesSet("indices",i,"index")};
     }
     writeBufferData(output_number, indices, global_idx);
-  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:w,getRunData:()=>({outputs:p,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:h})}},gl=(e,t)=>{uh(e.inputs);let r=e.inputs.length===1?t:dh(e.inputs,t);e.compute(ph(e.inputs,r),{inputs:[0]})},yl=e=>{let t=e.axis,r=e.splitSizes,n=e.numOutputs<0?r.length:e.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return te({axis:t,numOutputs:n,splitSizes:r})}});var mh,fh,wl,vl=U(()=>{"use strict";Q();ae();ie();mh=(e,t,r,n,o)=>{let i=M("output_data",o,r.length,4),a=E("a_data",t[1].dataType,t[1].dims.length,4),l=E("b_data",t[2].dataType,t[2].dims.length,4),d=E("c_data",t[0].dataType,t[0].dims.length,4),p,m=(u,h,w)=>`select(${h}, ${u}, ${w})`;if(!n)p=i.setByOffset("global_idx",m(a.getByOffset("global_idx"),l.getByOffset("global_idx"),d.getByOffset("global_idx")));else{let u=(h,w,g="")=>{let b=`a_data[index_a${w}][component_a${w}]`,x=`b_data[index_b${w}][component_b${w}]`,_=`bool(c_data[index_c${w}] & (0xffu << (component_c${w} * 8)))`;return`
+  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:w,getRunData:()=>({outputs:p,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:h})}},pl=(e,t)=>{ih(e.inputs);let r=e.inputs.length===1?t:ah(e.inputs,t);e.compute(dh(e.inputs,r),{inputs:[0]})},ml=e=>{let t=e.axis,r=e.splitSizes,n=e.numOutputs<0?r.length:e.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return Q({axis:t,numOutputs:n,splitSizes:r})}});var lh,ch,hl,gl=M(()=>{"use strict";X();oe();re();lh=(e,t,r,n,o)=>{let i=D("output_data",o,r.length,4),a=k("a_data",t[1].dataType,t[1].dims.length,4),l=k("b_data",t[2].dataType,t[2].dims.length,4),d=k("c_data",t[0].dataType,t[0].dims.length,4),p,m=(u,h,w)=>`select(${h}, ${u}, ${w})`;if(!n)p=i.setByOffset("global_idx",m(a.getByOffset("global_idx"),l.getByOffset("global_idx"),d.getByOffset("global_idx")));else{let u=(h,w,g="")=>{let y=`a_data[index_a${w}][component_a${w}]`,x=`b_data[index_b${w}][component_b${w}]`,_=`bool(c_data[index_c${w}] & (0xffu << (component_c${w} * 8)))`;return`
             let output_indices${w} = ${i.offsetToIndices(`global_idx * 4u + ${w}u`)};
             let offset_a${w} = ${a.broadcastedIndicesToOffset(`output_indices${w}`,i)};
             let offset_b${w} = ${l.broadcastedIndicesToOffset(`output_indices${w}`,i)};
@@ -6310,7 +6347,7 @@ fn calculateOutputIndex(index: u32) -> u32 {
             let component_a${w} = offset_a${w} % 4u;
             let component_b${w} = offset_b${w} % 4u;
             let component_c${w} = offset_c${w} % 4u;
-            ${h}[${w}] = ${g}(${m(b,x,_)});
+            ${h}[${w}] = ${g}(${m(y,x,_)});
           `};o===9?p=`
             var data = vec4<u32>(0);
             ${u("data",0,"u32")}
@@ -6327,10 +6364,10 @@ fn calculateOutputIndex(index: u32) -> u32 {
         ${e.mainStart()}
         ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
         ${p}
-      }`},fh=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,l=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,l=k.size(a)}let d=Math.ceil(l/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:p=>mh(p,e,a,i,o),getRunData:()=>({outputs:[{dims:a,dataType:o}],dispatchGroup:{x:Math.ceil(l/64/4)},programUniforms:[{type:12,data:d},...V(n,t,r,a)]})}},wl=e=>{e.compute(fh(e.inputs))}});var $l,_l=U(()=>{"use strict";xs();jr();Cs();As();mu();Su();Tu();co();Fu();Ku();Zu();rd();id();sd();ld();md();gd();Cd();Ad();kd();mo();Dd();_o();Bd();Qd();el();Fr();ol();al();ul();pl();hl();bl();So();Et();Yr();vl();$l=new Map([["Abs",[Es]],["Acos",[ks]],["Acosh",[Ps]],["Add",[fu]],["ArgMax",[_s,io]],["ArgMin",[$s,io]],["Asin",[Os]],["Asinh",[Ds]],["Atan",[zs]],["Atanh",[Bs]],["Attention",[Ss]],["AveragePool",[Gd,Hd]],["BatchNormalization",[Is]],["BiasAdd",[Ts]],["BiasSplitGelu",[pu]],["Cast",[Ms,Rs]],["Ceil",[Vs]],["Clip",[Us]],["Concat",[Iu,Cu]],["Conv",[yo,go]],["ConvTranspose",[Lu,Gu]],["Cos",[Ns]],["Cosh",[Ws]],["CumSum",[qu,ju]],["DepthToSpace",[Yu,Xu]],["Div",[hu]],["Einsum",[ed,td]],["Elu",[Hs,Qt]],["Equal",[gu]],["Erf",[Gs]],["Exp",[Ls]],["Expand",[od]],["FastGelu",[ad]],["Floor",[Fs]],["FusedConv",[yo,go]],["Gather",[dd,ud]],["GatherElements",[pd,cd]],["Gelu",[qs]],["Gemm",[hd,fd]],["GlobalAveragePool",[qd,Fd]],["GlobalMaxPool",[Zd,Xd]],["Greater",[vu]],["GreaterOrEqual",[_u]],["GroupQueryAttention",[Id,Sd]],["HardSigmoid",[eu,Js]],["InstanceNormalization",[Td]],["LayerNormalization",[Ed]],["LeakyRelu",[js,Qt]],["Less",[$u]],["LessOrEqual",[xu]],["Log",[du]],["MatMul",[Uu]],["MatMulNBits",[Pd,Od]],["MaxPool",[Kd,Yd]],["Mul",[yu]],["MultiHeadAttention",[wd,bd]],["Neg",[Ys]],["Not",[Ks]],["Pad",[zd]],["Pow",[bu]],["QuickGelu",[lu,Qt]],["Range",[Jd]],["Reciprocal",[Xs]],["ReduceMin",[hs]],["ReduceMean",[ls]],["ReduceMax",[fs]],["ReduceSum",[ys]],["ReduceProd",[gs]],["ReduceL1",[cs]],["ReduceL2",[ps]],["ReduceLogSum",[ws]],["ReduceLogSumExp",[ms]],["ReduceSumSquare",[bs]],["Relu",[Zs]],["Resize",[rl,nl]],["RotaryEmbedding",[il]],["Sigmoid",[Qs]],["Sin",[tu]],["Sinh",[ru]],["Slice",[ll,cl]],["SkipLayerNormalization",[sl]],["Split",[gl,yl]],["Sqrt",[nu]],["Softmax",[ml,fl]],["Sub",[wu]],["Tan",[ou]],["Tanh",[au]],["ThresholdedRelu",[uu,Qt]],["Tile",[$d]],["Transpose",[Za,Qa]],["Where",[wl]]])});var an,xl=U(()=>{"use strict";Ke();dt();ie();an=class{constructor(t){this.backend=t;this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,r){this.repo.set(t,r)}run(t,r,n,o,i){We(t.programInfo.name);let a=this.backend.device,l=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let d=[];for(let m of r)d.push({binding:d.length,resource:{buffer:m.buffer}});for(let m of n)d.push({binding:d.length,resource:{buffer:m.buffer}});i&&d.push({binding:d.length,resource:i});let p=a.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:d,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let m={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:p,dispatchGroup:o};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(m)}l.setPipeline(t.computePipeline),l.setBindGroup(0,p),l.dispatchWorkgroups(...o),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),Me(t.programInfo.name)}dispose(){}build(t,r){We(t.name);let n=this.backend.device,o=[];n.features.has("shader-f16")&&o.push("enable f16;");let i=Ya(r,this.backend.device.limits),a=t.getShaderSource(i),l=`${o.join(`
+      }`},ch=e=>{let t=e[1].dims,r=e[2].dims,n=e[0].dims,o=e[1].dataType,i=!(E.areEqual(t,r)&&E.areEqual(r,n)),a=t,l=E.size(t);if(i){let p=Je.calcShape(Je.calcShape(t,r,!1),n,!1);if(!p)throw new Error("Can't perform where op on the given tensors");a=p,l=E.size(a)}let d=Math.ceil(l/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:p=>lh(p,e,a,i,o),getRunData:()=>({outputs:[{dims:a,dataType:o}],dispatchGroup:{x:Math.ceil(l/64/4)},programUniforms:[{type:12,data:d},...U(n,t,r,a)]})}},hl=e=>{e.compute(ch(e.inputs))}});var yl,bl=M(()=>{"use strict";gs();Gr();ws();$s();au();yu();vu();ao();Uu();Wu();Gu();Yu();Qu();ed();nd();ad();dd();wd();$d();xd();uo();Cd();bo();Ad();Fd();jd();Xd();Lr();el();rl();ol();ul();cl();fl();vo();Ct();qr();gl();yl=new Map([["Abs",[_s]],["Acos",[xs]],["Acosh",[Ss]],["Add",[su]],["ArgMax",[hs,to]],["ArgMin",[fs,to]],["Asin",[Is]],["Asinh",[Cs]],["Atan",[Ts]],["Atanh",[As]],["Attention",[ys]],["AveragePool",[Rd,Dd]],["BatchNormalization",[bs]],["BiasAdd",[vs]],["BiasSplitGelu",[iu]],["Cast",[Es,ks]],["Ceil",[Os]],["Clip",[Ps]],["Concat",[bu,wu]],["Conv",[mo,po]],["ConvTranspose",[Mu,Ru]],["Cos",[zs]],["Cosh",[Bs]],["CumSum",[Vu,Nu]],["DepthToSpace",[Lu,Hu]],["DequantizeLinear",[qd,Kd]],["Div",[uu]],["Einsum",[Ku,ju]],["Elu",[Ds,Yt]],["Equal",[du]],["Erf",[Rs]],["Exp",[Ms]],["Expand",[Zu]],["FastGelu",[Ju]],["Floor",[Us]],["FusedConv",[mo,po]],["Gather",[rd,td]],["GatherElements",[id,od]],["Gelu",[Vs]],["Gemm",[ud,sd]],["GlobalAveragePool",[Vd,Ud]],["GlobalMaxPool",[Gd,Hd]],["Greater",[mu]],["GreaterOrEqual",[hu]],["GroupQueryAttention",[bd,yd]],["HardSigmoid",[Ks,qs]],["InstanceNormalization",[vd]],["LayerNormalization",[_d]],["LeakyRelu",[Ns,Yt]],["Less",[fu]],["LessOrEqual",[gu]],["Log",[ru]],["MatMul",[Pu]],["MatMulNBits",[Sd,Id]],["MaxPool",[Wd,Ld]],["Mul",[lu]],["MultiHeadAttention",[pd,cd]],["Neg",[Ls]],["Not",[Ws]],["Pad",[Td]],["Pow",[cu]],["QuickGelu",[nu,Yt]],["Range",[Yd]],["Reciprocal",[Hs]],["ReduceMin",[us]],["ReduceMean",[ns]],["ReduceMax",[ss]],["ReduceSum",[ls]],["ReduceProd",[ds]],["ReduceL1",[os]],["ReduceL2",[is]],["ReduceLogSum",[ps]],["ReduceLogSumExp",[as]],["ReduceSumSquare",[cs]],["Relu",[Gs]],["Resize",[Qd,Jd]],["RotaryEmbedding",[tl]],["Sigmoid",[Fs]],["Sin",[js]],["Sinh",[Ys]],["Slice",[al,sl]],["SkipLayerNormalization",[nl]],["Split",[pl,ml]],["Sqrt",[Xs]],["Softmax",[dl,ll]],["Sub",[pu]],["Tan",[Zs]],["Tanh",[Js]],["ThresholdedRelu",[tu,Yt]],["Tile",[fd]],["Transpose",[Ga,Fa]],["Where",[hl]]])});var rn,wl=M(()=>{"use strict";qe();st();re();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){Ve(t.programInfo.name);let a=this.backend.device,l=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let d=[];for(let m of r)d.push({binding:d.length,resource:{buffer:m.buffer}});for(let m of n)d.push({binding:d.length,resource:{buffer:m.buffer}});i&&d.push({binding:d.length,resource:i});let p=a.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:d,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)}l.setPipeline(t.computePipeline),l.setBindGroup(0,p),l.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){Ve(t.name);let n=this.backend.device,o=[];n.features.has("shader-f16")&&o.push("enable f16;");let i=La(r,this.backend.device.limits),a=t.getShaderSource(i),l=`${o.join(`
 `)}
 ${i.additionalImplementations}
-${a}`,d=n.createShaderModule({code:l,label:t.name});ve("verbose",()=>`[WebGPU] ${t.name} shader code: ${l}`);let p=n.createComputePipeline({compute:{module:d,entryPoint:"main"},layout:"auto",label:t.name});return Me(t.name),{programInfo:t,computePipeline:p,uniformVariablesInfo:i.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,l=Math.ceil(Math.sqrt(a));if(l>i){if(l=Math.ceil(Math.cbrt(a)),l>i)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[l,l,l]}else return[l,l,1]}}});var hh,gh,Io,sn,Sl=U(()=>{"use strict";Ke();Q();dt();Ga();Ka();_l();xl();hh=(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: ${t[n]}`)}}return r.join("|")},gh=(e,t,r)=>{let n=e.name;return e.shaderCache?.hint&&(n+="["+e.shaderCache.hint+"]"),n+=":"+r+`:${hh(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}},sn=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};r.features.has("chromium-experimental-timestamp-query-inside-passes")?n.push("chromium-experimental-timestamp-query-inside-passes"):r.features.has("timestamp-query")&&n.push("timestamp-query"),r.features.has("shader-f16")&&n.push("shader-f16"),this.device=await r.requestDevice(o),this.adapterInfo=new Io(r.info||await r.requestAdapterInfo()),this.gpuDataManager=ja(this),this.programManager=new an(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,Wa(t.logLevel,!!t.debug),this.device.onuncapturederror=i=>{i.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${i.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;We(),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,l=this.kernels.get(a),d=l.kernelType,p=l.kernelName,m=i.programName,u=i.inputTensorViews,h=i.outputTensorViews,w=r[o*2],g=r[o*2+1];typeof this.queryTimeBase>"u"&&(this.queryTimeBase=w);let b=Number(w-this.queryTimeBase),x=Number(g-this.queryTimeBase);if(!Number.isSafeInteger(b)||!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:ht(_.dataType)})),outputsMetadata:h.map(_=>({dims:_.dims,dataType:ht(_.dataType)})),kernelId:a,kernelType:d,kernelName:p,programName:m,startTime:b,endTime:x});else{let _="";u.forEach((S,I)=>{_+=`input[${I}]: [${S.dims}] | ${ht(S.dataType)}, `});let $="";h.forEach((S,I)=>{$+=`output[${I}]: [${S.dims}] | ${ht(S.dataType)}, `}),console.log(`[profiling] kernel "${a}|${d}|${p}|${m}" ${_}${$}execution time: ${x-b} ns`)}Sr("GPU",`${m}::${w}::${g}`)}t.unmap(),this.pendingQueries.delete(t)}),Me()}run(t,r,n,o,i,a){We(t.name);let l=[];for(let S=0;S<r.length;++S){let I=r[S].data;if(I===0)continue;let T=this.gpuDataManager.get(I);if(!T)throw new Error(`no GPU data for input: ${I}`);l.push(T)}let{outputs:d,dispatchGroup:p,programUniforms:m}=t.getRunData(r),u=n.length===0?d.map((S,I)=>I):n;if(u.length!==d.length)throw new Error(`Output size ${u.length} must be equal to ${d.length}.`);let h=[],w=[];for(let S=0;S<d.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 I=u[S]===-1,T=u[S]===-2,A=I||T?i(d[S].dataType,d[S].dims):o(u[S],d[S].dataType,d[S].dims);if(h.push(A),A.data===0)continue;let D=this.gpuDataManager.get(A.data);if(!D)throw new Error(`no GPU data for output: ${A.data}`);if(I&&this.temporaryData.push(D),T){let z=this.kernelPersistentData.get(this.currentKernelId);z||(z=[],this.kernelPersistentData.set(this.currentKernelId,z)),z.push(D)}w.push(D)}if(l.length!==r.length||w.length!==h.length){if(w.length===0)return Me(t.name),h;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let g;if(m){let S=0,I=[];m.forEach(z=>{let H=typeof z.data=="number"?[z.data]:z.data;if(H.length===0)return;let W=z.type===10?2:4,X,Y;z.type===10?(Y=H.length>4?16:H.length>2?8:H.length*W,X=H.length>4?16:W*H.length):(Y=H.length<=2?H.length*W:16,X=16),S=Math.ceil(S/Y)*Y,I.push(S);let pe=z.type===10?8:4;S+=H.length>4?Math.ceil(H.length/pe)*X:H.length*W});let T=16;S=Math.ceil(S/T)*T;let A=new ArrayBuffer(S);m.forEach((z,H)=>{let W=I[H],X=typeof z.data=="number"?[z.data]:z.data;if(z.type===6)new Int32Array(A,W,X.length).set(X);else if(z.type===12)new Uint32Array(A,W,X.length).set(X);else if(z.type===10)new Uint16Array(A,W,X.length).set(X);else if(z.type===1)new Float32Array(A,W,X.length).set(X);else throw new Error(`Unsupported uniform type: ${ht(z.type)}`)});let D=this.gpuDataManager.create(S,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(D.buffer,0,A,0,S),this.gpuDataManager.release(D.id),g={offset:0,size:S,buffer:D.buffer}}let b=this.programManager.normalizeDispatchGroupSize(p),x=b[1]===1&&b[2]===1,_=gh(t,r,x),$=this.programManager.getArtifact(_);if($||($=this.programManager.build(t,b),this.programManager.setArtifact(_,$),ve("info",()=>`[artifact] key: ${_}, programName: ${t.name}`)),m&&$.uniformVariablesInfo){if(m.length!==$.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${$.uniformVariablesInfo.length}, got ${m.length} in program "${$.programInfo.name}".`);for(let S=0;S<m.length;S++){let I=m[S],T=I.type,A=typeof I.data=="number"?1:I.data.length,[D,z]=$.uniformVariablesInfo[S];if(T!==D||A!==z)throw new Error(`Uniform variable ${S} mismatch: expect type ${D} with size ${z}, got type ${T} with size ${A} in program "${$.programInfo.name}".`)}}if(ve("info",()=>`[ProgramManager] run "${t.name}" (key=${_}) with ${b[0]}x${b[1]}x${b[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let S={kernelId:this.currentKernelId,programName:$.programInfo.name,inputTensorViews:r,outputTensorViews:h};this.pendingKernels.push(S),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(S)}return this.programManager.run($,l,w,b,g),Me(t.name),h}upload(t,r){this.gpuDataManager.upload(t,r)}memcpy(t,r){this.gpuDataManager.memcpy(t,r)}async download(t,r){await this.gpuDataManager.download(t,r)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,r,n,o){let i=$l.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: ${t}`);let i=o.kernelType,a=o.kernelName,l=o.kernelEntry,d=o.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${i}] ${a}" is not allowed to be called recursively`);this.currentKernelId=t,d[0]&&(d[1]=d[0](d[1]),d[0]=void 0),ve("info",()=>`[WebGPU] Start to run kernel "[${i}] ${a}"...`);let p=this.env.debug;this.temporaryData=[];try{return p&&this.device.pushErrorScope("validation"),l(r,d[1]),0}catch(m){return n.push(Promise.resolve(`[WebGPU] Kernel "[${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),l=this.gpuDataManager.registerExternalBuffer(n,o,a?.[1]);return i.set(r,[l,n]),l}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(n=>this.gpuDataManager.unregisterExternalBuffer(n[1])),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 Qn(this,t,r);return Ha(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(){ve("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(){ve("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){ve("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"}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 Il={};Ft(Il,{init:()=>yh});var or,Co,yh,Cl=U(()=>{"use strict";Q();Sl();dt();ae();or=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)}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)}},Co=class{constructor(t,r,n){this.module=t;this.backend=r;this.customDataOffset=0;this.customDataSize=0;this.adapterInfo=r.adapterInfo;let o=t.HEAPU32,i=n>>>2;this.opKernelContext=o[i++];let a=o[i++];this.outputCount=o[i++],this.customDataOffset=o[i++],this.customDataSize=o[i++];let l=[];for(let d=0;d<a;d++){let p=o[i++],m=o[i++],u=o[i++],h=[];for(let w=0;w<u;w++)h.push(o[i++]);l.push(new or(t,p,m,h))}this.inputs=l}get kernelCustomData(){return this.backend.currentKernelCustomData}get customDataBuffer(){return this.module.HEAPU8.subarray(this.customDataOffset,this.customDataOffset+this.customDataSize)}getMaxComputeWorkgroupSizes(){return[this.backend.device.limits.maxComputeWorkgroupSizeX,this.backend.device.limits.maxComputeWorkgroupSizeY,this.backend.device.limits.maxComputeWorkgroupSizeZ]}getMaxComputeWorkgroupStoragesize(){return this.backend.device.limits.maxComputeWorkgroupStorageSize}compute(t,r){let n=r?.inputs?.map(l=>typeof l=="number"?this.inputs[l]:l)??this.inputs,o=r?.outputs??[],i=(l,d,p)=>new or(this.module,d,this.output(l,p),p),a=(l,d)=>{let p=It(l);if(!p)throw new Error(`Unsupported data type: ${l}`);let m=p*k.size(d),u=m>0?this.backend.gpuDataManager.create(m).id:0;return new or(this.module,l,u,d)};return this.backend.run(t,n,o,i,a,this.outputCount)}output(t,r){let n=this.module.stackSave();try{let o=this.module.stackAlloc((1+r.length)*4),i=o>>2;this.module.HEAPU32[i++]=r.length;for(let a=0;a<r.length;a++)this.module.HEAPU32[i++]=r[a];return this.module._JsepOutput(this.opKernelContext,t,o)}catch(o){throw new Error(`Failed to generate kernel's output[${t}] with dims [${r}]. If you are running with pre-allocated output, please make sure the output type/dims are correct. Error: ${o}`)}finally{this.module.stackRestore(n)}}},yh=async(e,t,r,n)=>{let o=t.jsepInit;if(!o)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let i=new sn;await i.initialize(r,n),o("webgpu",[i,a=>i.alloc(a),a=>i.free(a),(a,l,d,p=!1)=>{if(p)ve("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${a}, dst=${l}, size=${d}`),i.memcpy(a,l);else{ve("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${a}, gpuDataId=${l}, size=${d}`);let m=t.HEAPU8.subarray(a>>>0,(a>>>0)+d);i.upload(l,m)}},async(a,l,d)=>{ve("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${a}, dataOffset=${l}, size=${d}`),await i.download(a,()=>t.HEAPU8.subarray(l>>>0,(l>>>0)+d))},(a,l,d)=>i.createKernel(a,l,d,t.UTF8ToString(t._JsepGetNodeName(l))),a=>i.releaseKernel(a),(a,l,d,p)=>{ve("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${d}, kernel=${a}, contextDataOffset=${l}`);let m=new Co(t,i,l);return i.computeKernel(a,m,p)},()=>i.captureBegin(),()=>i.captureEnd(),()=>i.replay()])}else o("webnn")}});var bh,Er,kr,kt,wh,Kt,Pr,Or,Tl,Dr,zr,Br,Hn=U(()=>{"use strict";Ra();Ua();Q();St();Mr();Kn();bh=(e,t)=>{Ie()._OrtInit(e,t)!==0&&_e("Can't initialize onnxruntime.")},Er=async e=>{bh(e.wasm.numThreads,Xt(e.logLevel))},kr=async(e,t)=>{{let r=(Cl(),br(Il)).init;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let n=e.webgpu.adapter;if(n){if(typeof n.limits!="object"||typeof n.features!="object"||typeof n.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let o=e.webgpu.powerPreference;if(o!==void 0&&o!=="low-power"&&o!=="high-performance")throw new Error(`Invalid powerPreference setting: "${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)}}},kt=new Map,wh=e=>{let t=Ie(),r=t.stackSave();try{let n=t.stackAlloc(8);return t._OrtGetInputOutputCount(e,n,n+4)!==0&&_e("Can't get session input/output count."),[t.HEAP32[n/4],t.HEAP32[n/4+1]]}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]},Pr=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,l=0,d=[],p=[],m=[];try{if([a,d]=Ma(t),t?.externalData&&o.mountExternalData){let $=[];for(let S of t.externalData){let I=typeof S=="string"?S:S.path;$.push(Zt(typeof S=="string"?S:S.data).then(T=>{o.mountExternalData(I,T)}))}await Promise.all($)}for(let $ of t?.executionProviders??[])if((typeof $=="string"?$:$.name)==="webnn"){if(o.currentContext)throw new Error("WebNN execution provider is already set.");if(typeof $!="string"){let I=$,T=I?.context,A=I?.gpuDevice,D=I?.deviceType,z=I?.numThreads,H=I?.powerPreference;T?o.currentContext=T:A?o.currentContext=await navigator.ml.createContext(A):o.currentContext=await navigator.ml.createContext({deviceType:D,numThreads:z,powerPreference:H})}else o.currentContext=await navigator.ml.createContext();break}i=await o._OrtCreateSession(r,n,a),i===0&&_e("Can't create a session."),o.currentContext&&(o.currentContext=void 0);let[u,h]=wh(i),w=!!t?.enableGraphCapture,g=[],b=[],x=[];for(let $=0;$<u;$++){let S=o._OrtGetInputName(i,$);S===0&&_e("Can't get an input name."),p.push(S),g.push(o.UTF8ToString(S))}for(let $=0;$<h;$++){let S=o._OrtGetOutputName(i,$);S===0&&_e("Can't get an output name."),m.push(S);let I=o.UTF8ToString(S);b.push(I);{if(w&&t?.preferredOutputLocation===void 0){x.push("gpu-buffer");continue}let T=typeof t?.preferredOutputLocation=="string"?t.preferredOutputLocation:t?.preferredOutputLocation?.[I]??"cpu";if(T!=="cpu"&&T!=="cpu-pinned"&&T!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${T}.`);if(w&&T!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${T}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);x.push(T)}}let _=null;return x.some($=>$==="gpu-buffer")&&(l=o._OrtCreateBinding(i),l===0&&_e("Can't create IO binding."),_={handle:l,outputPreferredLocations:x,outputPreferredLocationsEncoded:x.map($=>jn($))}),kt.set(i,[i,p,m,_,w,!1]),[i,g,b]}catch(u){throw p.forEach(h=>o._OrtFree(h)),m.forEach(h=>o._OrtFree(h)),l!==0&&o._OrtReleaseBinding(l),i!==0&&o._OrtReleaseSession(i),u}finally{o._free(r),a!==0&&o._OrtReleaseSessionOptions(a),d.forEach(u=>o._free(u)),o.unmountExternalData?.()}},Or=e=>{let t=Ie(),r=kt.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[n,o,i,a,l]=r;a&&(l&&t._OrtClearBoundOutputs(a.handle),t._OrtReleaseBinding(a.handle)),t.jsepOnReleaseSession?.(e),o.forEach(d=>t._OrtFree(d)),i.forEach(d=>t._OrtFree(d)),t._OrtReleaseSession(n),kt.delete(e)},Tl=(e,t,r,n,o,i=!1)=>{if(!e){t.push(0);return}let a=Ie(),l=e[0],d=e[1],p=e[3],m,u;if(l==="string"&&p==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(i&&p!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${o} when enableGraphCapture is true.`);if(p==="gpu-buffer"){let g=e[2].gpuBuffer,b=It(qn(l));u=d.reduce((_,$)=>_*$,1)*b;let x=a.jsepRegisterBuffer;if(!x)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');m=x(n,o,g,u)}else{let g=e[2];if(Array.isArray(g)){u=4*g.length,m=a._malloc(u),r.push(m);let b=m/4;for(let x=0;x<g.length;x++){if(typeof g[x]!="string")throw new TypeError(`tensor data at index ${x} is not a string`);a.HEAPU32[b++]=Te(g[x],r)}}else u=g.byteLength,m=a._malloc(u),r.push(m),a.HEAPU8.set(new Uint8Array(g.buffer,g.byteOffset,u),m)}let h=a.stackSave(),w=a.stackAlloc(4*d.length);try{let g=w/4;d.forEach(x=>a.HEAP32[g++]=x);let b=a._OrtCreateTensor(qn(l),m,u,w,d.length,jn(p));b===0&&_e(`Can't create tensor for input/output. session=${n}, index=${o}.`),t.push(b)}finally{a.stackRestore(h)}},Dr=async(e,t,r,n,o,i)=>{let a=Ie(),l=kt.get(e);if(!l)throw new Error(`cannot run inference. invalid session id: ${e}`);let d=l[0],p=l[1],m=l[2],u=l[3],h=l[4],w=l[5],g=t.length,b=n.length,x=0,_=[],$=[],S=[],I=[],T=a.stackSave(),A=a.stackAlloc(g*4),D=a.stackAlloc(g*4),z=a.stackAlloc(b*4),H=a.stackAlloc(b*4);try{[x,_]=Ba(i);for(let K=0;K<g;K++)Tl(r[K],$,I,e,t[K],h);for(let K=0;K<b;K++)Tl(o[K],S,I,e,g+n[K],h);let W=A/4,X=D/4,Y=z/4,pe=H/4;for(let K=0;K<g;K++)a.HEAPU32[W++]=$[K],a.HEAPU32[X++]=p[t[K]];for(let K=0;K<b;K++)a.HEAPU32[Y++]=S[K],a.HEAPU32[pe++]=m[n[K]];if(u&&!w){let{handle:K,outputPreferredLocations:de,outputPreferredLocationsEncoded:re}=u;if(p.length!==g)throw new Error(`input count from feeds (${g}) is expected to be always equal to model's input count (${p.length}).`);for(let se=0;se<g;se++){let Z=t[se];await a._OrtBindInput(K,p[Z],$[se])!==0&&_e(`Can't bind input[${se}] for session=${e}.`)}for(let se=0;se<b;se++){let Z=n[se];o[se]?.[3]?a._OrtBindOutput(K,m[Z],S[se],0)!==0&&_e(`Can't bind pre-allocated output[${se}] for session=${e}.`):a._OrtBindOutput(K,m[Z],0,re[Z])!==0&&_e(`Can't bind output[${se}] to ${de[se]} for session=${e}.`)}kt.set(e,[d,p,m,u,h,!0])}a.jsepOnRunStart?.(d);let ue;u?ue=await a._OrtRunWithBinding(d,u.handle,b,z,x):ue=await a._OrtRun(d,D,A,g,H,b,z,x),ue!==0&&_e("failed to call OrtRun().");let le=[];for(let K=0;K<b;K++){let de=a.HEAPU32[z/4+K];if(de===S[K]){le.push(o[K]);continue}let re=a.stackSave(),se=a.stackAlloc(4*4),Z=!1,ne,J=0;try{a._OrtGetTensorData(de,se,se+4,se+8,se+12)!==0&&_e(`Can't access output tensor data on index ${K}.`);let R=se/4,F=a.HEAPU32[R++];J=a.HEAPU32[R++];let ce=a.HEAPU32[R++],Re=a.HEAPU32[R++],Se=[];for(let Pe=0;Pe<Re;Pe++)Se.push(a.HEAPU32[ce/4+Pe]);a._OrtFree(ce);let we=Se.reduce((Pe,ze)=>Pe*ze,1);ne=ht(F);let Lt=u?.outputPreferredLocations[n[K]];if(ne==="string"){if(Lt==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Pe=[],ze=J/4;for(let Fe=0;Fe<we;Fe++){let Ot=a.HEAPU32[ze++],yt=Fe===we-1?void 0:a.HEAPU32[ze]-Ot;Pe.push(a.UTF8ToString(Ot,yt))}le.push([ne,Se,Pe,"cpu"])}else if(Lt==="gpu-buffer"&&we>0){let Pe=a.jsepGetBuffer;if(!Pe)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let ze=Pe(J),Fe=It(F);if(Fe===void 0||!Vr(ne))throw new Error(`Unsupported data type: ${ne}`);Z=!0,le.push([ne,Se,{gpuBuffer:ze,download:a.jsepCreateDownloader(ze,we*Fe,ne),dispose:()=>{a._OrtReleaseTensor(de)}},"gpu-buffer"])}else{let Pe=Ur(ne),ze=new Pe(we);new Uint8Array(ze.buffer,ze.byteOffset,ze.byteLength).set(a.HEAPU8.subarray(J,J+ze.byteLength)),le.push([ne,Se,ze,"cpu"])}}finally{a.stackRestore(re),ne==="string"&&J&&a._free(J),Z||a._OrtReleaseTensor(de)}}return u&&!h&&(a._OrtClearBoundOutputs(u.handle),kt.set(e,[d,p,m,u,h,!1])),le}finally{a.stackRestore(T),$.forEach(W=>a._OrtReleaseTensor(W)),S.forEach(W=>a._OrtReleaseTensor(W)),I.forEach(W=>a._free(W)),x!==0&&a._OrtReleaseRunOptions(x),_.forEach(W=>a._free(W))}},zr=e=>{let t=Ie(),r=kt.get(e);if(!r)throw new Error("invalid session id");let n=r[0],o=t._OrtEndProfiling(n);o===0&&_e("Can't get an profile file name."),t._OrtFree(o)},Br=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 Pt,Ye,ir,dn,ln,un,To,Ao,Ht,Gt,$h,Al,El,kl,Pl,Ol,Dl,zl,Eo=U(()=>{"use strict";Ke();Hn();St();jt();Pt=()=>!!be.wasm.proxy&&typeof document<"u",ir=!1,dn=!1,ln=!1,Ao=new Map,Ht=(e,t)=>{let r=Ao.get(e);r?r.push(t):Ao.set(e,[t])},Gt=()=>{if(ir||!dn||ln||!Ye)throw new Error("worker not ready")},$h=e=>{switch(e.data.type){case"init-wasm":ir=!1,e.data.err?(ln=!0,To[1](e.data.err)):(dn=!0,To[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=Ao.get(e.data.type);e.data.err?t.shift()[1](e.data.err):t.shift()[0](e.data.out);break}default:}},Al=async()=>{if(!dn){if(ir)throw new Error("multiple calls to 'initWasm()' detected.");if(ln)throw new Error("previous call to 'initWasm()' failed.");if(ir=!0,Pt())return new Promise((e,t)=>{Ye?.terminate(),Oa().then(([r,n])=>{try{Ye=n,Ye.onerror=i=>t(i),Ye.onmessage=$h,To=[e,t];let o={type:"init-wasm",in:be};Ye.postMessage(o),un=r}catch(o){t(o)}},t)});try{await Ar(be.wasm),await Er(be),dn=!0}catch(e){throw ln=!0,e}finally{ir=!1}}},El=async e=>{if(Pt())return Gt(),new Promise((t,r)=>{Ht("init-ep",[t,r]);let n={type:"init-ep",in:{epName:e,env:be}};Ye.postMessage(n)});await kr(be,e)},kl=async e=>Pt()?(Gt(),new Promise((t,r)=>{Ht("copy-from",[t,r]);let n={type:"copy-from",in:{buffer:e}};Ye.postMessage(n,[e.buffer])})):Kt(e),Pl=async(e,t)=>{if(Pt()){if(t?.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return Gt(),new Promise((r,n)=>{Ht("create",[r,n]);let o={type:"create",in:{model:e,options:{...t}}},i=[];e instanceof Uint8Array&&i.push(e.buffer),Ye.postMessage(o,i)})}else return Pr(e,t)},Ol=async e=>{if(Pt())return Gt(),new Promise((t,r)=>{Ht("release",[t,r]);let n={type:"release",in:e};Ye.postMessage(n)});Or(e)},Dl=async(e,t,r,n,o,i)=>{if(Pt()){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 Gt(),new Promise((a,l)=>{Ht("run",[a,l]);let d=r,p={type:"run",in:{sessionId:e,inputIndices:t,inputs:d,outputIndices:n,options:i}};Ye.postMessage(p,Br(d))})}else return Dr(e,t,r,n,o,i)},zl=async e=>{if(Pt())return Gt(),new Promise((t,r)=>{Ht("end-profiling",[t,r]);let n={type:"end-profiling",in:e};Ye.postMessage(n)});zr(e)}});var Bl,_h,cn,Rl=U(()=>{"use strict";Ke();Eo();Q();Tr();Kn();Bl=(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"];default:throw new Error(`invalid data location: ${e.location} for ${t()}`)}},_h=e=>{switch(e[3]){case"cpu":return new De(e[0],e[2],e[1]);case"gpu-buffer":{let t=e[0];if(!Vr(t))throw new Error(`not supported data type: ${t} for deserializing GPU tensor`);let{gpuBuffer:r,download:n,dispose:o}=e[2];return De.fromGpuBuffer(r,{dataType:t,dims:e[1],download:n,dispose:o})}default:throw new Error(`invalid data location: ${e[3]}`)}},cn=class{async fetchModelAndCopyToWasmMemory(t){return kl(await Zt(t))}async loadModel(t,r){We();let n;typeof t=="string"? false?0:n=await this.fetchModelAndCopyToWasmMemory(t):n=t,[this.sessionId,this.inputNames,this.outputNames]=await Pl(n,r),Me()}async dispose(){return Ol(this.sessionId)}async run(t,r,n){We();let o=[],i=[];Object.entries(t).forEach(h=>{let w=h[0],g=h[1],b=this.inputNames.indexOf(w);if(b===-1)throw new Error(`invalid input '${w}'`);o.push(g),i.push(b)});let a=[],l=[];Object.entries(r).forEach(h=>{let w=h[0],g=h[1],b=this.outputNames.indexOf(w);if(b===-1)throw new Error(`invalid output '${w}'`);a.push(g),l.push(b)});let d=o.map((h,w)=>Bl(h,()=>`input "${this.inputNames[i[w]]}"`)),p=a.map((h,w)=>h?Bl(h,()=>`output "${this.outputNames[l[w]]}"`):null),m=await Dl(this.sessionId,i,d,l,p,n),u={};for(let h=0;h<m.length;h++)u[this.outputNames[l[h]]]=a[h]??_h(m[h]);return Me(),u}startProfiling(){}endProfiling(){zl(this.sessionId)}}});var xh,pn,Ml=U(()=>{"use strict";Ke();Eo();Rl();jt();xh=()=>{if((typeof be.wasm.initTimeout!="number"||be.wasm.initTimeout<0)&&(be.wasm.initTimeout=0),be.wasm.simd===!1&&console.warn('Deprecated property "env.wasm.simd" is set to false. non-SIMD build is no longer provided, and this setting will be ignored.'),typeof be.wasm.proxy!="boolean"&&(be.wasm.proxy=!1),typeof be.wasm.trace!="boolean"&&(be.wasm.trace=!1),typeof be.wasm.numThreads!="number"||!Number.isInteger(be.wasm.numThreads)||be.wasm.numThreads<=0)if(typeof self<"u"&&!self.crossOriginIsolated)be.wasm.numThreads=1;else{let e=typeof navigator>"u"?Mn("node:os").cpus().length:navigator.hardwareConcurrency;be.wasm.numThreads=Math.min(4,Math.ceil((e||1)/2))}},pn=class{async init(t){xh(),await Al(),await El(t)}async createInferenceSessionHandler(t,r){let n=new cn;return await n.loadModel(t,r),Promise.resolve(n)}}});var Ul={};Ft(Ul,{wasmBackend:()=>Sh});var Sh,Vl=U(()=>{"use strict";Ml();Sh=new pn});Ke();Ke();Ke();var xa="1.19.0";var K_=Wn;{let e=(Vl(),br(Ul)).wasmBackend;_t("webgpu",e,5),_t("webnn",e,5),_t("cpu",e,10),_t("wasm",e,10)}Object.defineProperty(be.versions,"web",{value:xa,enumerable:!0});
+${a}`,d=n.createShaderModule({code:l,label:t.name});be("verbose",()=>`[WebGPU] ${t.name} shader code: ${l}`);let p=n.createComputePipeline({compute:{module:d,entryPoint:"main"},layout:"auto",label:t.name});return De(t.name),{programInfo:t,computePipeline:p,uniformVariablesInfo:i.variablesInfo}}normalizeDispatchGroupSize(t){let r=typeof t=="number"?t:t.x,n=typeof t=="number"?1:t.y||1,o=typeof t=="number"?1:t.z||1,i=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(r<=i&&n<=i&&o<=i)return[r,n,o];let a=r*n*o,l=Math.ceil(Math.sqrt(a));if(l>i){if(l=Math.ceil(Math.cbrt(a)),l>i)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[l,l,l]}else return[l,l,1]}}});var ph,mh,$o,nn,vl=M(()=>{"use strict";qe();X();st();Ra();Wa();bl();wl();ph=(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: ${t[n]}`)}}return r.join("|")},mh=(e,t,r)=>{let n=e.name;return e.shaderCache?.hint&&(n+="["+e.shaderCache.hint+"]"),n+=":"+r+`:${ph(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,n},$o=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},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};r.features.has("chromium-experimental-timestamp-query-inside-passes")?n.push("chromium-experimental-timestamp-query-inside-passes"):r.features.has("timestamp-query")&&n.push("timestamp-query"),r.features.has("shader-f16")&&n.push("shader-f16"),this.device=await r.requestDevice(o),this.adapterInfo=new $o(r.info||await r.requestAdapterInfo()),this.gpuDataManager=Na(this),this.programManager=new rn(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,Ba(t.logLevel,!!t.debug),this.device.onuncapturederror=i=>{i.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${i.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;Ve(),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,l=this.kernels.get(a),d=l.kernelType,p=l.kernelName,m=i.programName,u=i.inputTensorViews,h=i.outputTensorViews,w=r[o*2],g=r[o*2+1];typeof this.queryTimeBase>"u"&&(this.queryTimeBase=w);let y=Number(w-this.queryTimeBase),x=Number(g-this.queryTimeBase);if(!Number.isSafeInteger(y)||!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:mt(_.dataType)})),outputsMetadata:h.map(_=>({dims:_.dims,dataType:mt(_.dataType)})),kernelId:a,kernelType:d,kernelName:p,programName:m,startTime:y,endTime:x});else{let _="";u.forEach((S,I)=>{_+=`input[${I}]: [${S.dims}] | ${mt(S.dataType)}, `});let $="";h.forEach((S,I)=>{$+=`output[${I}]: [${S.dims}] | ${mt(S.dataType)}, `}),console.log(`[profiling] kernel "${a}|${d}|${p}|${m}" ${_}${$}execution time: ${x-y} ns`)}$r("GPU",`${m}::${w}::${g}`)}t.unmap(),this.pendingQueries.delete(t)}),De()}run(t,r,n,o,i,a){Ve(t.name);let l=[];for(let S=0;S<r.length;++S){let I=r[S].data;if(I===0)continue;let T=this.gpuDataManager.get(I);if(!T)throw new Error(`no GPU data for input: ${I}`);l.push(T)}let{outputs:d,dispatchGroup:p,programUniforms:m}=t.getRunData(r),u=n.length===0?d.map((S,I)=>I):n;if(u.length!==d.length)throw new Error(`Output size ${u.length} must be equal to ${d.length}.`);let h=[],w=[];for(let S=0;S<d.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 I=u[S]===-1,T=u[S]===-2,A=I||T?i(d[S].dataType,d[S].dims):o(u[S],d[S].dataType,d[S].dims);if(h.push(A),A.data===0)continue;let O=this.gpuDataManager.get(A.data);if(!O)throw new Error(`no GPU data for output: ${A.data}`);if(I&&this.temporaryData.push(O),T){let z=this.kernelPersistentData.get(this.currentKernelId);z||(z=[],this.kernelPersistentData.set(this.currentKernelId,z)),z.push(O)}w.push(O)}if(l.length!==r.length||w.length!==h.length){if(w.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 g;if(m){let S=0,I=[];m.forEach(z=>{let W=typeof z.data=="number"?[z.data]:z.data;if(W.length===0)return;let V=z.type===10?2:4,j,Y;z.type===10?(Y=W.length>4?16:W.length>2?8:W.length*V,j=W.length>4?16:V*W.length):(Y=W.length<=2?W.length*V:16,j=16),S=Math.ceil(S/Y)*Y,I.push(S);let ae=z.type===10?8:4;S+=W.length>4?Math.ceil(W.length/ae)*j:W.length*V});let T=16;S=Math.ceil(S/T)*T;let A=new ArrayBuffer(S);m.forEach((z,W)=>{let V=I[W],j=typeof z.data=="number"?[z.data]:z.data;if(z.type===6)new Int32Array(A,V,j.length).set(j);else if(z.type===12)new Uint32Array(A,V,j.length).set(j);else if(z.type===10)new Uint16Array(A,V,j.length).set(j);else if(z.type===1)new Float32Array(A,V,j.length).set(j);else throw new Error(`Unsupported uniform type: ${mt(z.type)}`)});let O=this.gpuDataManager.create(S,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(O.buffer,0,A,0,S),this.gpuDataManager.release(O.id),g={offset:0,size:S,buffer:O.buffer}}let y=this.programManager.normalizeDispatchGroupSize(p),x=y[1]===1&&y[2]===1,_=mh(t,r,x),$=this.programManager.getArtifact(_);if($||($=this.programManager.build(t,y),this.programManager.setArtifact(_,$),be("info",()=>`[artifact] key: ${_}, programName: ${t.name}`)),m&&$.uniformVariablesInfo){if(m.length!==$.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${$.uniformVariablesInfo.length}, got ${m.length} in program "${$.programInfo.name}".`);for(let S=0;S<m.length;S++){let I=m[S],T=I.type,A=typeof I.data=="number"?1:I.data.length,[O,z]=$.uniformVariablesInfo[S];if(T!==O||A!==z)throw new Error(`Uniform variable ${S} mismatch: expect type ${O} with size ${z}, got type ${T} with size ${A} in program "${$.programInfo.name}".`)}}if(be("info",()=>`[ProgramManager] run "${t.name}" (key=${_}) with ${y[0]}x${y[1]}x${y[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let S={kernelId:this.currentKernelId,programName:$.programInfo.name,inputTensorViews:r,outputTensorViews:h};this.pendingKernels.push(S),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(S)}return this.programManager.run($,l,w,y,g),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=yl.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: ${t}`);let i=o.kernelType,a=o.kernelName,l=o.kernelEntry,d=o.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${i}] ${a}" is not allowed to be called recursively`);this.currentKernelId=t,d[0]&&(d[1]=d[0](d[1]),d[0]=void 0),be("info",()=>`[WebGPU] Start to run kernel "[${i}] ${a}"...`);let p=this.env.debug;this.temporaryData=[];try{return p&&this.device.pushErrorScope("validation"),l(r,d[1]),0}catch(m){return n.push(Promise.resolve(`[WebGPU] Kernel "[${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),l=this.gpuDataManager.registerExternalBuffer(n,o,a?.[1]);return i.set(r,[l,n]),l}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(n=>this.gpuDataManager.unregisterExternalBuffer(n[1])),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 jn(this,t,r);return Da(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(){be("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(){be("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){be("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"}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 $l={};Lt($l,{init:()=>fh});var tr,_o,fh,_l=M(()=>{"use strict";X();vl();st();oe();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=E.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=E.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=E.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(E.size(t)!==E.size(this.dims))throw new Error("Invalid new shape");return new e(this.module,this.dataType,this.data,t)}},_o=class{constructor(t,r,n){this.module=t;this.backend=r;this.customDataOffset=0;this.customDataSize=0;this.adapterInfo=r.adapterInfo;let o=t.HEAPU32,i=n>>>2;this.opKernelContext=o[i++];let a=o[i++];this.outputCount=o[i++],this.customDataOffset=o[i++],this.customDataSize=o[i++];let l=[];for(let d=0;d<a;d++){let p=o[i++],m=o[i++],u=o[i++],h=[];for(let w=0;w<u;w++)h.push(o[i++]);l.push(new tr(t,p,m,h))}this.inputs=l}get kernelCustomData(){return this.backend.currentKernelCustomData}get customDataBuffer(){return this.module.HEAPU8.subarray(this.customDataOffset,this.customDataOffset+this.customDataSize)}getMaxComputeWorkgroupSizes(){return[this.backend.device.limits.maxComputeWorkgroupSizeX,this.backend.device.limits.maxComputeWorkgroupSizeY,this.backend.device.limits.maxComputeWorkgroupSizeZ]}getMaxComputeWorkgroupStoragesize(){return this.backend.device.limits.maxComputeWorkgroupStorageSize}compute(t,r){let n=r?.inputs?.map(l=>typeof l=="number"?this.inputs[l]:l)??this.inputs,o=r?.outputs??[],i=(l,d,p)=>new tr(this.module,d,this.output(l,p),p),a=(l,d)=>{let p=_t(l,d);if(!p)throw new Error(`Unsupported data type: ${l}`);let m=p>0?this.backend.gpuDataManager.create(p).id:0;return new tr(this.module,l,m,d)};return this.backend.run(t,n,o,i,a,this.outputCount)}output(t,r){let n=this.module.stackSave();try{let o=this.module.stackAlloc((1+r.length)*4),i=o>>2;this.module.HEAPU32[i++]=r.length;for(let a=0;a<r.length;a++)this.module.HEAPU32[i++]=r[a];return this.module._JsepOutput(this.opKernelContext,t,o)}catch(o){throw new Error(`Failed to generate kernel's output[${t}] with dims [${r}]. If you are running with pre-allocated output, please make sure the output type/dims are correct. Error: ${o}`)}finally{this.module.stackRestore(n)}}},fh=async(e,t,r,n)=>{let o=t.jsepInit;if(!o)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let i=new nn;await i.initialize(r,n),o("webgpu",[i,a=>i.alloc(a),a=>i.free(a),(a,l,d,p=!1)=>{if(p)be("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${a}, dst=${l}, size=${d}`),i.memcpy(a,l);else{be("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${a}, gpuDataId=${l}, size=${d}`);let m=t.HEAPU8.subarray(a>>>0,(a>>>0)+d);i.upload(l,m)}},async(a,l,d)=>{be("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${a}, dataOffset=${l}, size=${d}`),await i.download(a,()=>t.HEAPU8.subarray(l>>>0,(l>>>0)+d))},(a,l,d)=>i.createKernel(a,l,d,t.UTF8ToString(t._JsepGetNodeName(l))),a=>i.releaseKernel(a),(a,l,d,p)=>{be("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${d}, kernel=${a}, contextDataOffset=${l}`);let m=new _o(t,i,l);return i.computeKernel(a,m,p)},()=>i.captureBegin(),()=>i.captureEnd(),()=>i.replay()])}else o("webnn")}});var hh,Cr,Tr,Tt,gh,Ft,Ar,kr,xl,Er,Pr,Or,Un=M(()=>{"use strict";ka();Pa();X();$t();Br();Gn();hh=(e,t)=>{_e()._OrtInit(e,t)!==0&&we("Can't initialize onnxruntime.")},Cr=async e=>{hh(e.wasm.numThreads,Kt(e.logLevel))},Tr=async(e,t)=>{{let r=(_l(),hr($l)).init;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let n=e.webgpu.adapter;if(n){if(typeof n.limits!="object"||typeof n.features!="object"||typeof n.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let o=e.webgpu.powerPreference;if(o!==void 0&&o!=="low-power"&&o!=="high-performance")throw new Error(`Invalid powerPreference setting: "${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",_e(),e,n)}if(t==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",_e(),e)}}},Tt=new Map,gh=e=>{let t=_e(),r=t.stackSave();try{let n=t.stackAlloc(8);return t._OrtGetInputOutputCount(e,n,n+4)!==0&&we("Can't get session input/output count."),[t.HEAP32[n/4],t.HEAP32[n/4+1]]}finally{t.stackRestore(r)}},Ft=e=>{let t=_e(),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]},Ar=async(e,t)=>{let r,n,o=_e();Array.isArray(e)?[r,n]=e:e.buffer===o.HEAPU8.buffer?[r,n]=[e.byteOffset,e.byteLength]:[r,n]=Ft(e);let i=0,a=0,l=0,d=[],p=[],m=[];try{if([a,d]=Ea(t),t?.externalData&&o.mountExternalData){let $=[];for(let S of t.externalData){let I=typeof S=="string"?S:S.path;$.push(jt(typeof S=="string"?S:S.data).then(T=>{o.mountExternalData(I,T)}))}await Promise.all($)}for(let $ of t?.executionProviders??[])if((typeof $=="string"?$:$.name)==="webnn"){if(o.currentContext)throw new Error("WebNN execution provider is already set.");if(typeof $!="string"){let I=$,T=I?.context,A=I?.gpuDevice,O=I?.deviceType,z=I?.numThreads,W=I?.powerPreference;T?o.currentContext=T:A?o.currentContext=await navigator.ml.createContext(A):o.currentContext=await navigator.ml.createContext({deviceType:O,numThreads:z,powerPreference:W})}else o.currentContext=await navigator.ml.createContext();break}i=await o._OrtCreateSession(r,n,a),i===0&&we("Can't create a session."),o.currentContext&&(o.currentContext=void 0);let[u,h]=gh(i),w=!!t?.enableGraphCapture,g=[],y=[],x=[];for(let $=0;$<u;$++){let S=o._OrtGetInputName(i,$);S===0&&we("Can't get an input name."),p.push(S),g.push(o.UTF8ToString(S))}for(let $=0;$<h;$++){let S=o._OrtGetOutputName(i,$);S===0&&we("Can't get an output name."),m.push(S);let I=o.UTF8ToString(S);y.push(I);{if(w&&t?.preferredOutputLocation===void 0){x.push("gpu-buffer");continue}let T=typeof t?.preferredOutputLocation=="string"?t.preferredOutputLocation:t?.preferredOutputLocation?.[I]??"cpu";if(T!=="cpu"&&T!=="cpu-pinned"&&T!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${T}.`);if(w&&T!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${T}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);x.push(T)}}let _=null;return x.some($=>$==="gpu-buffer")&&(l=o._OrtCreateBinding(i),l===0&&we("Can't create IO binding."),_={handle:l,outputPreferredLocations:x,outputPreferredLocationsEncoded:x.map($=>Hn($))}),Tt.set(i,[i,p,m,_,w,!1]),[i,g,y]}catch(u){throw p.forEach(h=>o._OrtFree(h)),m.forEach(h=>o._OrtFree(h)),l!==0&&o._OrtReleaseBinding(l),i!==0&&o._OrtReleaseSession(i),u}finally{o._free(r),a!==0&&o._OrtReleaseSessionOptions(a),d.forEach(u=>o._free(u)),o.unmountExternalData?.()}},kr=e=>{let t=_e(),r=Tt.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[n,o,i,a,l]=r;a&&(l&&t._OrtClearBoundOutputs(a.handle),t._OrtReleaseBinding(a.handle)),t.jsepOnReleaseSession?.(e),o.forEach(d=>t._OrtFree(d)),i.forEach(d=>t._OrtFree(d)),t._OrtReleaseSession(n),Tt.delete(e)},xl=(e,t,r,n,o,i=!1)=>{if(!e){t.push(0);return}let a=_e(),l=e[0],d=e[1],p=e[3],m,u;if(l==="string"&&p==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(i&&p!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${o} when enableGraphCapture is true.`);if(p==="gpu-buffer"){let g=e[2].gpuBuffer;u=_t(Ln(l),d);let y=a.jsepRegisterBuffer;if(!y)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');m=y(n,o,g,u)}else{let g=e[2];if(Array.isArray(g)){u=4*g.length,m=a._malloc(u),r.push(m);let y=m/4;for(let x=0;x<g.length;x++){if(typeof g[x]!="string")throw new TypeError(`tensor data at index ${x} is not a string`);a.HEAPU32[y++]=Ie(g[x],r)}}else u=g.byteLength,m=a._malloc(u),r.push(m),a.HEAPU8.set(new Uint8Array(g.buffer,g.byteOffset,u),m)}let h=a.stackSave(),w=a.stackAlloc(4*d.length);try{let g=w/4;d.forEach(x=>a.HEAP32[g++]=x);let y=a._OrtCreateTensor(Ln(l),m,u,w,d.length,Hn(p));y===0&&we(`Can't create tensor for input/output. session=${n}, index=${o}.`),t.push(y)}finally{a.stackRestore(h)}},Er=async(e,t,r,n,o,i)=>{let a=_e(),l=Tt.get(e);if(!l)throw new Error(`cannot run inference. invalid session id: ${e}`);let d=l[0],p=l[1],m=l[2],u=l[3],h=l[4],w=l[5],g=t.length,y=n.length,x=0,_=[],$=[],S=[],I=[],T=a.stackSave(),A=a.stackAlloc(g*4),O=a.stackAlloc(g*4),z=a.stackAlloc(y*4),W=a.stackAlloc(y*4);try{[x,_]=Aa(i);for(let q=0;q<g;q++)xl(r[q],$,I,e,t[q],h);for(let q=0;q<y;q++)xl(o[q],S,I,e,g+n[q],h);let V=A/4,j=O/4,Y=z/4,ae=W/4;for(let q=0;q<g;q++)a.HEAPU32[V++]=$[q],a.HEAPU32[j++]=p[t[q]];for(let q=0;q<y;q++)a.HEAPU32[Y++]=S[q],a.HEAPU32[ae++]=m[n[q]];if(u&&!w){let{handle:q,outputPreferredLocations:ce,outputPreferredLocationsEncoded:ne}=u;if(p.length!==g)throw new Error(`input count from feeds (${g}) is expected to be always equal to model's input count (${p.length}).`);for(let Z=0;Z<g;Z++){let ee=t[Z];await a._OrtBindInput(q,p[ee],$[Z])!==0&&we(`Can't bind input[${Z}] for session=${e}.`)}for(let Z=0;Z<y;Z++){let ee=n[Z];o[Z]?.[3]?a._OrtBindOutput(q,m[ee],S[Z],0)!==0&&we(`Can't bind pre-allocated output[${Z}] for session=${e}.`):a._OrtBindOutput(q,m[ee],0,ne[ee])!==0&&we(`Can't bind output[${Z}] to ${ce[Z]} for session=${e}.`)}Tt.set(e,[d,p,m,u,h,!0])}a.jsepOnRunStart?.(d);let se;u?se=await a._OrtRunWithBinding(d,u.handle,y,z,x):se=await a._OrtRun(d,O,A,g,W,y,z,x),se!==0&&we("failed to call OrtRun().");let ie=[];for(let q=0;q<y;q++){let ce=a.HEAPU32[z/4+q];if(ce===S[q]){ie.push(o[q]);continue}let ne=a.stackSave(),Z=a.stackAlloc(4*4),ee=!1,K,me=0;try{a._OrtGetTensorData(ce,Z,Z+4,Z+8,Z+12)!==0&&we(`Can't access output tensor data on index ${q}.`);let R=Z/4,G=a.HEAPU32[R++];me=a.HEAPU32[R++];let ue=a.HEAPU32[R++],Be=a.HEAPU32[R++],ge=[];for(let Te=0;Te<Be;Te++)ge.push(a.HEAPU32[ue/4+Te]);a._OrtFree(ue);let Se=ge.reduce((Te,Oe)=>Te*Oe,1);K=mt(G);let Wt=u?.outputPreferredLocations[n[q]];if(K==="string"){if(Wt==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Te=[],Oe=me/4;for(let He=0;He<Se;He++){let kt=a.HEAPU32[Oe++],ht=He===Se-1?void 0:a.HEAPU32[Oe]-kt;Te.push(a.UTF8ToString(kt,ht))}ie.push([K,ge,Te,"cpu"])}else if(Wt==="gpu-buffer"&&Se>0){let Te=a.jsepGetBuffer;if(!Te)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let Oe=Te(me),He=_t(G,Se);if(He===void 0||!Rr(K))throw new Error(`Unsupported data type: ${K}`);ee=!0,ie.push([K,ge,{gpuBuffer:Oe,download:a.jsepCreateDownloader(Oe,He,K),dispose:()=>{a._OrtReleaseTensor(ce)}},"gpu-buffer"])}else{let Te=Dr(K),Oe=new Te(Se);new Uint8Array(Oe.buffer,Oe.byteOffset,Oe.byteLength).set(a.HEAPU8.subarray(me,me+Oe.byteLength)),ie.push([K,ge,Oe,"cpu"])}}finally{a.stackRestore(ne),K==="string"&&me&&a._free(me),ee||a._OrtReleaseTensor(ce)}}return u&&!h&&(a._OrtClearBoundOutputs(u.handle),Tt.set(e,[d,p,m,u,h,!1])),ie}finally{a.stackRestore(T),$.forEach(V=>a._OrtReleaseTensor(V)),S.forEach(V=>a._OrtReleaseTensor(V)),I.forEach(V=>a._free(V)),x!==0&&a._OrtReleaseRunOptions(x),_.forEach(V=>a._free(V))}},Pr=e=>{let t=_e(),r=Tt.get(e);if(!r)throw new Error("invalid session id");let n=r[0],o=t._OrtEndProfiling(n);o===0&&we("Can't get an profile file name."),t._OrtFree(o)},Or=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 At,Ke,rr,an,sn,on,xo,So,Vt,Nt,bh,Sl,Il,Cl,Tl,Al,kl,El,Io=M(()=>{"use strict";qe();Un();$t();Gt();At=()=>!!fe.wasm.proxy&&typeof document<"u",rr=!1,an=!1,sn=!1,So=new Map,Vt=(e,t)=>{let r=So.get(e);r?r.push(t):So.set(e,[t])},Nt=()=>{if(rr||!an||sn||!Ke)throw new Error("worker not ready")},bh=e=>{switch(e.data.type){case"init-wasm":rr=!1,e.data.err?(sn=!0,xo[1](e.data.err)):(an=!0,xo[0]()),on&&(URL.revokeObjectURL(on),on=void 0);break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let t=So.get(e.data.type);e.data.err?t.shift()[1](e.data.err):t.shift()[0](e.data.out);break}default:}},Sl=async()=>{if(!an){if(rr)throw new Error("multiple calls to 'initWasm()' detected.");if(sn)throw new Error("previous call to 'initWasm()' failed.");if(rr=!0,At())return new Promise((e,t)=>{Ke?.terminate(),Ia().then(([r,n])=>{try{Ke=n,Ke.onerror=i=>t(i),Ke.onmessage=bh,xo=[e,t];let o={type:"init-wasm",in:fe};Ke.postMessage(o),on=r}catch(o){t(o)}},t)});try{await Ir(fe.wasm),await Cr(fe),an=!0}catch(e){throw sn=!0,e}finally{rr=!1}}},Il=async e=>{if(At())return Nt(),new Promise((t,r)=>{Vt("init-ep",[t,r]);let n={type:"init-ep",in:{epName:e,env:fe}};Ke.postMessage(n)});await Tr(fe,e)},Cl=async e=>At()?(Nt(),new Promise((t,r)=>{Vt("copy-from",[t,r]);let n={type:"copy-from",in:{buffer:e}};Ke.postMessage(n,[e.buffer])})):Ft(e),Tl=async(e,t)=>{if(At()){if(t?.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return Nt(),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),Ke.postMessage(o,i)})}else return Ar(e,t)},Al=async e=>{if(At())return Nt(),new Promise((t,r)=>{Vt("release",[t,r]);let n={type:"release",in:e};Ke.postMessage(n)});kr(e)},kl=async(e,t,r,n,o,i)=>{if(At()){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 Nt(),new Promise((a,l)=>{Vt("run",[a,l]);let d=r,p={type:"run",in:{sessionId:e,inputIndices:t,inputs:d,outputIndices:n,options:i}};Ke.postMessage(p,Or(d))})}else return Er(e,t,r,n,o,i)},El=async e=>{if(At())return Nt(),new Promise((t,r)=>{Vt("end-profiling",[t,r]);let n={type:"end-profiling",in:e};Ke.postMessage(n)});Pr(e)}});var Pl,wh,un,Ol=M(()=>{"use strict";qe();Io();X();Sr();Gn();Pl=(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"];default:throw new Error(`invalid data location: ${e.location} for ${t()}`)}},wh=e=>{switch(e[3]){case"cpu":return new Pe(e[0],e[2],e[1]);case"gpu-buffer":{let t=e[0];if(!Rr(t))throw new Error(`not supported data type: ${t} for deserializing GPU tensor`);let{gpuBuffer:r,download:n,dispose:o}=e[2];return Pe.fromGpuBuffer(r,{dataType:t,dims:e[1],download:n,dispose:o})}default:throw new Error(`invalid data location: ${e[3]}`)}},un=class{async fetchModelAndCopyToWasmMemory(t){return Cl(await jt(t))}async loadModel(t,r){Ve();let n;typeof t=="string"? false?0:n=await this.fetchModelAndCopyToWasmMemory(t):n=t,[this.sessionId,this.inputNames,this.outputNames]=await Tl(n,r),De()}async dispose(){return Al(this.sessionId)}async run(t,r,n){Ve();let o=[],i=[];Object.entries(t).forEach(h=>{let w=h[0],g=h[1],y=this.inputNames.indexOf(w);if(y===-1)throw new Error(`invalid input '${w}'`);o.push(g),i.push(y)});let a=[],l=[];Object.entries(r).forEach(h=>{let w=h[0],g=h[1],y=this.outputNames.indexOf(w);if(y===-1)throw new Error(`invalid output '${w}'`);a.push(g),l.push(y)});let d=o.map((h,w)=>Pl(h,()=>`input "${this.inputNames[i[w]]}"`)),p=a.map((h,w)=>h?Pl(h,()=>`output "${this.outputNames[l[w]]}"`):null),m=await kl(this.sessionId,i,d,l,p,n),u={};for(let h=0;h<m.length;h++)u[this.outputNames[l[h]]]=a[h]??wh(m[h]);return De(),u}startProfiling(){}endProfiling(){El(this.sessionId)}}});var vh,dn,zl=M(()=>{"use strict";qe();Io();Ol();Gt();vh=()=>{if((typeof fe.wasm.initTimeout!="number"||fe.wasm.initTimeout<0)&&(fe.wasm.initTimeout=0),fe.wasm.simd===!1&&console.warn('Deprecated property "env.wasm.simd" is set to false. non-SIMD build is no longer provided, and this setting will be ignored.'),typeof fe.wasm.proxy!="boolean"&&(fe.wasm.proxy=!1),typeof fe.wasm.trace!="boolean"&&(fe.wasm.trace=!1),typeof fe.wasm.numThreads!="number"||!Number.isInteger(fe.wasm.numThreads)||fe.wasm.numThreads<=0)if(typeof self<"u"&&!self.crossOriginIsolated)fe.wasm.numThreads=1;else{let e=typeof navigator>"u"?zn("node:os").cpus().length:navigator.hardwareConcurrency;fe.wasm.numThreads=Math.min(4,Math.ceil((e||1)/2))}},dn=class{async init(t){vh(),await Sl(),await Il(t)}async createInferenceSessionHandler(t,r){let n=new un;return await n.loadModel(t,r),Promise.resolve(n)}}});var Bl={};Lt(Bl,{wasmBackend:()=>$h});var $h,Dl=M(()=>{"use strict";zl();$h=new dn});qe();qe();qe();var ga="1.20.0-dev.20240818-417aa00406";var ex=Mn;{let e=(Dl(),hr(Bl)).wasmBackend;wt("webgpu",e,5),wt("webnn",e,5),wt("cpu",e,10),wt("wasm",e,10)}Object.defineProperty(fe.versions,"web",{value:ga,enumerable:!0});
 /**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
@@ -7036,7 +7073,7 @@ __webpack_require__.r(__webpack_exports__);
 
 
 
-const VERSION = '3.0.0-alpha.7';
+const VERSION = '3.0.0-alpha.8';
 
 // Check if various APIs are available (depends on environment)
 const IS_BROWSER_ENV = typeof self !== 'undefined';