diff --git "a/assets/index-8bAK_Ga0.js" "b/assets/index-8bAK_Ga0.js" new file mode 100644--- /dev/null +++ "b/assets/index-8bAK_Ga0.js" @@ -0,0 +1,2373 @@ +(function(){const e=document.createElement("link").relList;if(e&&e.supports&&e.supports("modulepreload"))return;for(const a of document.querySelectorAll('link[rel="modulepreload"]'))n(a);new MutationObserver(a=>{for(const i of a)if(i.type==="childList")for(const s of i.addedNodes)s.tagName==="LINK"&&s.rel==="modulepreload"&&n(s)}).observe(document,{childList:!0,subtree:!0});function r(a){const i={};return a.integrity&&(i.integrity=a.integrity),a.referrerPolicy&&(i.referrerPolicy=a.referrerPolicy),a.crossOrigin==="use-credentials"?i.credentials="include":a.crossOrigin==="anonymous"?i.credentials="omit":i.credentials="same-origin",i}function n(a){if(a.ep)return;a.ep=!0;const i=r(a);fetch(a.href,i)}})();const Je={},k0=Object.freeze(Object.defineProperty({__proto__:null,default:Je},Symbol.toStringTag,{value:"Module"})),E0="3.0.0-alpha.0",rs=typeof self<"u",C0=rs&&self.constructor.name==="DedicatedWorkerGlobalScope",zp=rs&&"caches"in self,T0=typeof navigator<"u"&&"gpu"in navigator,Pp=typeof process<"u",A0=Pp&&process?.release?.name==="node",Fi=!Dp(Je),Rp=!Dp(Je),an=Object.freeze({IS_BROWSER_ENV:rs,IS_WEBWORKER_ENV:C0,IS_WEB_CACHE_AVAILABLE:zp,IS_WEBGPU_AVAILABLE:T0,IS_PROCESS_AVAILABLE:Pp,IS_NODE_ENV:A0,IS_FS_AVAILABLE:Fi,IS_PATH_AVAILABLE:Rp}),Do=Fi&&Rp,Bp=Do?Je.dirname(Je.dirname(Je.fileURLToPath(import.meta.url))):"./",I0=Do?Je.join(Bp,"/.cache/"):null,ru="/models/",M0=Do?Je.join(Bp,ru):ru,zt={version:E0,backends:{onnx:{},tfjs:{}},allowRemoteModels:!0,remoteHost:"https://huggingface.co/",remotePathTemplate:"{model}/resolve/{revision}/",allowLocalModels:!rs,localModelPath:M0,useFS:Fi,useBrowserCache:zp,useFSCache:Fi,cacheDir:I0,useCustomCache:!1,customCache:null};function Dp(t){return Object.keys(t).length===0}const bt=class{constructor(){let t=function(...e){return t._call(...e)};return Object.setPrototypeOf(t,new.target.prototype)}_call(...t){throw Error("Must implement _call method in subclass")}};function An(t,e){t&&t(e)}function O0(t){return Object.fromEntries(Object.entries(t).map(([e,r])=>[r,e]))}function Np(t){return t.replace(/[.*+?^${}()|[\]\\]/g,"\\$&")}function z0(t){return Number.isInteger(t)||typeof t=="bigint"}function nu(t){const e=[];let r=t;for(;Array.isArray(r);)e.push(r.length),r=r[0];return e}function ft(...t){return Array.prototype.concat.apply([],t)}function P0(...t){return t.reduce((e,r)=>e.flatMap(n=>r.map(a=>[n,a])))}function Li(t,e){return Math.abs((t+e)%(2*e)-e)}function rn(t,e){return Object.assign({},...e.map(r=>{if(t[r]!==void 0)return{[r]:t[r]}}))}var Ms={};class ns{_CONTENT_TYPE_MAP={txt:"text/plain",html:"text/html",css:"text/css",js:"text/javascript",json:"application/json",png:"image/png",jpg:"image/jpeg",jpeg:"image/jpeg",gif:"image/gif"};constructor(e){if(this.filePath=e,this.headers=new Headers,this.exists=Je.existsSync(e),this.exists){this.status=200,this.statusText="OK";let r=Je.statSync(e);this.headers.set("content-length",r.size.toString()),this.updateContentType();let n=this;this.body=new ReadableStream({start(a){n.arrayBuffer().then(i=>{a.enqueue(new Uint8Array(i)),a.close()})}})}else this.status=404,this.statusText="Not Found",this.body=null}updateContentType(){const e=this.filePath.toString().split(".").pop().toLowerCase();this.headers.set("content-type",this._CONTENT_TYPE_MAP[e]??"application/octet-stream")}clone(){let e=new ns(this.filePath);return e.exists=this.exists,e.status=this.status,e.statusText=this.statusText,e.headers=new Headers(this.headers),e}async arrayBuffer(){return(await Je.promises.readFile(this.filePath)).buffer}async blob(){const e=await Je.promises.readFile(this.filePath);return new Blob([e],{type:this.headers.get("content-type")})}async text(){return await Je.promises.readFile(this.filePath,"utf8")}async json(){return JSON.parse(await this.text())}}function wo(t,e=null,r=null){let n;try{n=new URL(t)}catch{return!1}return!(e&&!e.includes(n.protocol)||r&&!r.includes(n.hostname))}async function Ui(t){if(zt.useFS&&!wo(t,["http:","https:","blob:"]))return new ns(t);if(typeof process<"u"&&process?.release?.name==="node"){const e=!!Ms?.TESTING_REMOTELY,r=zt.version,n=new Headers;if(n.set("User-Agent",`transformers.js/${r}; is_ci/${e};`),wo(t,["http:","https:"],["huggingface.co","hf.co"])){const i=Ms?.HF_TOKEN??Ms?.HF_ACCESS_TOKEN;i&&n.set("Authorization",`Bearer ${i}`)}return fetch(t,{headers:n})}else return fetch(t)}const R0={400:"Bad request error occurred while trying to load file",401:"Unauthorized access to file",403:"Forbidden access to file",404:"Could not locate file",408:"Request timeout error occurred while trying to load file",500:"Internal server error error occurred while trying to load file",502:"Bad gateway error occurred while trying to load file",503:"Service unavailable error occurred while trying to load file",504:"Gateway timeout error occurred while trying to load file"};function B0(t,e,r){if(!r)return null;const n=R0[t]??`Error (${t}) occurred while trying to load file`;throw Error(`${n}: "${e}".`)}class au{constructor(e){this.path=e}async match(e){let r=Je.join(this.path,e),n=new ns(r);if(n.exists)return n}async put(e,r){const n=Buffer.from(await r.arrayBuffer());let a=Je.join(this.path,e);try{await Je.promises.mkdir(Je.dirname(a),{recursive:!0}),await Je.promises.writeFile(a,n)}catch(i){console.warn("An error occurred while writing the file to cache:",i)}}}async function D0(t,...e){for(let r of e)try{let n=await t.match(r);if(n)return n}catch{continue}}async function Ii(t,e,r=!0,n={}){if(!zt.allowLocalModels&&n.local_files_only)throw Error("Invalid configuration detected: local models are disabled (`env.allowLocalModels=false`) but you have requested to only use local models (`local_files_only=true`).");An(n.progress_callback,{status:"initiate",name:t,file:e});let a;if(!a&&zt.useBrowserCache){if(typeof caches>"u")throw Error("Browser cache is not available in this environment.");try{a=await caches.open("transformers-cache")}catch(S){console.warn("An error occurred while opening the browser cache:",S)}}if(!a&&zt.useFSCache&&(a=new au(n.cache_dir??zt.cacheDir)),!a&&zt.useCustomCache)throw Error("`env.useCustomCache=true`, but `env.customCache` is not defined.");const i=n.revision??"main";let s=gi(t,e),o=gi(zt.localModelPath,s),l=gi(zt.remoteHost,zt.remotePathTemplate.replaceAll("{model}",t).replaceAll("{revision}",encodeURIComponent(i)),e),u=i==="main"?s:gi(t,i,e),p,h=a instanceof au?u:l,m=!1,d;a&&(d=await D0(a,o,h));const _=d!==void 0;if(d===void 0){if(zt.allowLocalModels)if(wo(s,["http:","https:"])){if(n.local_files_only)throw new Error(`\`local_files_only=true\`, but attempted to load a remote file from: ${s}.`)}else try{d=await Ui(o),p=o}catch($){console.warn(`Unable to load from local path "${o}": "${$}"`)}if(d===void 0||d.status===404){if(n.local_files_only||!zt.allowRemoteModels){if(r)throw Error(`\`local_files_only=true\` or \`env.allowRemoteModels=false\` and file was not found locally at "${o}".`);return null}if(d=await Ui(l),d.status!==200)return B0(d.status,l,r);p=h}m=a&&typeof Response<"u"&&d instanceof Response&&d.status===200}An(n.progress_callback,{status:"download",name:t,file:e});const w={status:"progress",name:t,file:e};let v;return n.progress_callback?_&&typeof navigator<"u"&&/firefox/i.test(navigator.userAgent)?(v=new Uint8Array(await d.arrayBuffer()),An(n.progress_callback,{...w,progress:100,loaded:v.length,total:v.length})):v=await N0(d,S=>{An(n.progress_callback,{...w,...S})}):v=new Uint8Array(await d.arrayBuffer()),m&&p&&await a.match(p)===void 0&&await a.put(p,new Response(v,{headers:d.headers})).catch(S=>{console.warn(`Unable to add response to browser cache: ${S}.`)}),An(n.progress_callback,{status:"done",name:t,file:e}),v}async function Br(t,e,r=!0,n={}){let a=await Ii(t,e,r,n);if(a===null)return{};let s=new TextDecoder("utf-8").decode(a);return JSON.parse(s)}async function N0(t,e){const r=t.headers.get("Content-Length");r===null&&console.warn("Unable to determine content-length from response headers. Will expand buffer when needed.");let n=parseInt(r??"0"),a=new Uint8Array(n),i=0;const s=t.body.getReader();async function o(){const{done:l,value:u}=await s.read();if(l)return;let p=i+u.length;if(p>n){n=p;let m=new Uint8Array(n);m.set(a),a=m}a.set(u,i),i=p;const h=i/n*100;return e({progress:h,loaded:i,total:n}),o()}return await o(),a}function gi(...t){return t=t.map((e,r)=>(r&&(e=e.replace(new RegExp("^/"),"")),r!==t.length-1&&(e=e.replace(new RegExp("/$"),"")),e)),t.join("/")}function F0(t,[e,r,n],[a,i],s="bilinear",o=!1){const l=i/n,u=a/r,p=new t.constructor(a*i*e),h=r*n,m=a*i;for(let d=0;d=0;--o)a[o]=l,n[o]=e[r[o]],l*=n[o];const i=r.map((o,l)=>a[r.indexOf(l)]),s=new t.constructor(t.length);for(let o=0;o=0;--u)l+=p%e[u]*i[u],p=Math.floor(p/e[u]);s[l]=t[o]}return[s,n]}function wt(t){const e=Kt(t)[0],r=t.map(i=>Math.exp(i-e)),n=r.reduce((i,s)=>i+s,0);return r.map(i=>i/n)}function U0(t){return wt(t).map(n=>Math.log(n))}function Fp(t){if(t.length===0)throw Error("Array must not be empty");let e=t[0],r=0;for(let n=1;ne&&(e=t[n],r=n);return[Number(e),r]}function Lp(t){return t>0&&(t&t-1)===0}class Up{constructor(e){if(this.size=e|0,this.size<=1||!Lp(this.size))throw new Error("FFT size must be a power of two larger than 1");this._csize=e<<1,this.table=new Float64Array(this.size*2);for(let n=0;nn;n<<=1)++r;this._width=r%2===0?r-1:r,this._bitrev=new Int32Array(1<>>a&3)<>>1);for(let a=0;a>>1]=e[a];return n}toComplexArray(e,r){const n=r||this.createComplexArray();for(let a=0;a>>1],n[a+1]=0;return n}transform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._transform4(e,r,1)}realTransform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._realTransform4(e,r,1)}inverseTransform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._transform4(e,r,-1);for(let n=0;n>=2;s>=2;s>>=2){o=a/s<<1;const m=o>>>2;for(l=0;l>>1,s>>>1)}else for(l=0,u=0;l>>1,s>>>1,n)}const h=this.table;for(s>>=2;s>=2;s>>=2){o=a/s<<1;const d=o>>>1,_=d>>>1,w=_>>>1;for(l=0;l>>1;for(let d=2;d>1;++p){const h=(p+1-e)**2/2,m=Math.sqrt(l**2+u**2)**h,d=h*Math.atan2(u,l),_=2*p;i[_]=m*Math.cos(d),i[_+1]=m*Math.sin(d),s[_]=i[_],s[_+1]=-i[_+1]}this._slicedChirpBuffer=i.subarray(r,n),this._f=new Up(a>>1),this._f.transform(this._chirpBuffer,s)}_transform(e,r,n){const a=this._buffer1,i=this._buffer2,s=this._outBuffer1,o=this._outBuffer2,l=this._chirpBuffer,u=this._slicedChirpBuffer,p=this._a;if(n)for(let h=0;h>1,_=r[d];a[h]=_*u[h],a[m]=_*u[m]}else for(let h=0;h=t.length&&(l=2*(t.length-1)-l),n[s++]=t[l]}n.sort(),r[i]=n[a]}return r}function Mi(t,e){const r=Math.pow(10,e);return Math.round(t*r)/r}function H0(t){const e=Math.round(t);return Math.abs(t)%1===.5?e%2===0?e:e-1:e}/*! + * ONNX Runtime Web v1.18.0 + * Copyright (c) Microsoft Corporation. All rights reserved. + * Licensed under the MIT License. + */var No=Object.defineProperty,j0=Object.getOwnPropertyDescriptor,q0=Object.getOwnPropertyNames,K0=Object.prototype.hasOwnProperty,J=(t,e)=>()=>(t&&(e=t(t=0)),e),Bn=(t,e)=>()=>(e||t((e={exports:{}}).exports,e),e.exports),Dn=(t,e)=>{for(var r in e)No(t,r,{get:e[r],enumerable:!0})},Y0=(t,e,r,n)=>{if(e&&typeof e=="object"||typeof e=="function")for(let a of q0(e))!K0.call(t,a)&&a!==r&&No(t,a,{get:()=>e[a],enumerable:!(n=j0(e,a))||n.enumerable});return t},nn=t=>Y0(No({},"__esModule",{value:!0}),t),aa,Or,tn,iu,Fo,Lo=J(()=>{aa=new Map,Or=[],tn=(t,e,r)=>{if(e&&typeof e.init=="function"&&typeof e.createInferenceSessionHandler=="function"){let n=aa.get(t);if(n===void 0)aa.set(t,{backend:e,priority:r});else{if(n.priority>r)return;if(n.priority===r&&n.backend!==e)throw new Error(`cannot register backend "${t}" using priority ${r}`)}if(r>=0){let a=Or.indexOf(t);a!==-1&&Or.splice(a,1);for(let i=0;i{let e=aa.get(t);if(!e)return"backend not found.";if(e.initialized)return e.backend;if(e.aborted)return e.error;{let r=!!e.initPromise;try{return r||(e.initPromise=e.backend.init(t)),await e.initPromise,e.initialized=!0,e.backend}catch(n){return r||(e.error=`${n}`,e.aborted=!0),e.error}finally{delete e.initPromise}}},Fo=async t=>{let e=t.executionProviders||[],r=e.map(l=>typeof l=="string"?l:l.name),n=r.length===0?Or:r,a,i=[],s=new Set;for(let l of n){let u=await iu(l);typeof u=="string"?i.push({name:l,err:u}):(a||(a=u),a===u&&s.add(l))}if(!a)throw new Error(`no available backend found. ERR: ${i.map(l=>`[${l.name}] ${l.err}`).join(", ")}`);for(let{name:l,err:u}of i)r.includes(l)&&console.warn(`removing requested execution provider "${l}" from session options because it is not available: ${u}`);let o=e.filter(l=>s.has(typeof l=="string"?l:l.name));return[a,new Proxy(t,{get:(l,u)=>u==="executionProviders"?o:Reflect.get(l,u)})]}}),X0=J(()=>{Lo()}),Wp,Q0=J(()=>{Wp="1.18.0"}),Os,Ht,Vp=J(()=>{Q0(),Os="warning",Ht={wasm:{},webgl:{},webgpu:{},versions:{common:Wp},set logLevel(t){if(t!==void 0){if(typeof t!="string"||["verbose","info","warning","error","fatal"].indexOf(t)===-1)throw new Error(`Unsupported logging level: ${t}`);Os=t}},get logLevel(){return Os}},Object.defineProperty(Ht,"logLevel",{enumerable:!0})}),Ue,Z0=J(()=>{Vp(),Ue=Ht}),Gp,Hp,J0=J(()=>{Gp=(t,e)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=t.dims[3],r.height=t.dims[2];let n=r.getContext("2d");if(n!=null){let a,i;e?.tensorLayout!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],i=t.dims[3]):(a=t.dims[3],i=t.dims[2]);let s=e?.format!==void 0?e.format:"RGB",o=e?.norm,l,u;o===void 0||o.mean===void 0?l=[255,255,255,255]:typeof o.mean=="number"?l=[o.mean,o.mean,o.mean,o.mean]:(l=[o.mean[0],o.mean[1],o.mean[2],0],o.mean[3]!==void 0&&(l[3]=o.mean[3])),o===void 0||o.bias===void 0?u=[0,0,0,0]:typeof o.bias=="number"?u=[o.bias,o.bias,o.bias,o.bias]:(u=[o.bias[0],o.bias[1],o.bias[2],0],o.bias[3]!==void 0&&(u[3]=o.bias[3]));let p=i*a,h=0,m=p,d=p*2,_=-1;s==="RGBA"?(h=0,m=p,d=p*2,_=p*3):s==="RGB"?(h=0,m=p,d=p*2):s==="RBG"&&(h=0,d=p,m=p*2);for(let w=0;w{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let a,i,s;e?.tensorLayout!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],i=t.dims[1],s=t.dims[3]):(a=t.dims[3],i=t.dims[2],s=t.dims[1]);let o=e!==void 0&&e.format!==void 0?e.format:"RGB",l=e?.norm,u,p;l===void 0||l.mean===void 0?u=[255,255,255,255]:typeof l.mean=="number"?u=[l.mean,l.mean,l.mean,l.mean]:(u=[l.mean[0],l.mean[1],l.mean[2],255],l.mean[3]!==void 0&&(u[3]=l.mean[3])),l===void 0||l.bias===void 0?p=[0,0,0,0]:typeof l.bias=="number"?p=[l.bias,l.bias,l.bias,l.bias]:(p=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(p[3]=l.bias[3]));let h=i*a;if(e!==void 0&&(e.format!==void 0&&s===4&&e.format!=="RGBA"||s===3&&e.format!=="RGB"&&e.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let m=4,d=0,_=1,w=2,v=3,S=0,$=h,E=h*2,T=-1;o==="RGBA"?(S=0,$=h,E=h*2,T=h*3):o==="RGB"?(S=0,$=h,E=h*2):o==="RBG"&&(S=0,E=h,$=h*2),n=r.createImageData(a,i);for(let A=0;A{Uo(),_i=(t,e)=>{if(t===void 0)throw new Error("Image buffer must be defined");if(e.height===void 0||e.width===void 0)throw new Error("Image height and width must be defined");if(e.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:n}=e,a=e.norm??{mean:255,bias:0},i,s;typeof a.mean=="number"?i=[a.mean,a.mean,a.mean,a.mean]:i=[a.mean[0],a.mean[1],a.mean[2],a.mean[3]??255],typeof a.bias=="number"?s=[a.bias,a.bias,a.bias,a.bias]:s=[a.bias[0],a.bias[1],a.bias[2],a.bias[3]??0];let o=e.format!==void 0?e.format:"RGBA",l=e.tensorFormat!==void 0&&e.tensorFormat!==void 0?e.tensorFormat:"RGB",u=r*n,p=l==="RGBA"?new Float32Array(u*4):new Float32Array(u*3),h=4,m=0,d=1,_=2,w=3,v=0,S=u,$=u*2,E=-1;o==="RGB"&&(h=3,m=0,d=1,_=2,w=-1),l==="RGBA"?E=u*3:l==="RBG"?(v=0,$=u,S=u*2):l==="BGR"&&($=0,S=u,v=u*2);for(let T=0;T{let r=typeof HTMLImageElement<"u"&&t instanceof HTMLImageElement,n=typeof ImageData<"u"&&t instanceof ImageData,a=typeof ImageBitmap<"u"&&t instanceof ImageBitmap,i=typeof t=="string",s,o=e??{},l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},u=p=>p instanceof HTMLCanvasElement||p instanceof OffscreenCanvas?p.getContext("2d"):null;if(r){let p=l();p.width=t.width,p.height=t.height;let h=u(p);if(h!=null){let m=t.height,d=t.width;if(e!==void 0&&e.resizedHeight!==void 0&&e.resizedWidth!==void 0&&(m=e.resizedHeight,d=e.resizedWidth),e!==void 0){if(o=e,e.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=m,o.width=d}else o.tensorFormat="RGBA",o.height=m,o.width=d;h.drawImage(t,0,0),s=h.getImageData(0,0,d,m).data}else throw new Error("Can not access image data")}else if(n){let p,h;if(e!==void 0&&e.resizedWidth!==void 0&&e.resizedHeight!==void 0?(p=e.resizedHeight,h=e.resizedWidth):(p=t.height,h=t.width),e!==void 0&&(o=e),o.format="RGBA",o.height=p,o.width=h,e!==void 0){let m=l();m.width=h,m.height=p;let d=u(m);if(d!=null)d.putImageData(t,0,0),s=d.getImageData(0,0,h,p).data;else throw new Error("Can not access image data")}else s=t.data}else if(a){if(e===void 0)throw new Error("Please provide image config with format for Imagebitmap");let p=l();p.width=t.width,p.height=t.height;let h=u(p);if(h!=null){let m=t.height,d=t.width;return h.drawImage(t,0,0,d,m),s=h.getImageData(0,0,d,m).data,o.height=m,o.width=d,_i(s,o)}else throw new Error("Can not access image data")}else{if(i)return new Promise((p,h)=>{let m=l(),d=u(m);if(!t||!d)return h();let _=new Image;_.crossOrigin="Anonymous",_.src=t,_.onload=()=>{m.width=_.width,m.height=_.height,d.drawImage(_,0,0,m.width,m.height);let w=d.getImageData(0,0,m.width,m.height);o.height=m.height,o.width=m.width,p(_i(w.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return _i(s,o);throw new Error("Input data provided is not supported - aborted tensor creation")},qp=(t,e)=>{let{width:r,height:n,download:a,dispose:i}=e,s=[1,n,r,4];return new lr({location:"texture",type:"float32",texture:t,dims:s,download:a,dispose:i})},Kp=(t,e)=>{let{dataType:r,dims:n,download:a,dispose:i}=e;return new lr({location:"gpu-buffer",type:r??"float32",gpuBuffer:t,dims:n,download:a,dispose:i})},Yp=(t,e,r)=>new lr({location:"cpu-pinned",type:t,data:e,dims:r??[e.length]})}),Zr,ha,zs,Xp,ty=J(()=>{Zr=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),ha=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),zs=!1,Xp=()=>{if(!zs){zs=!0;let t=typeof BigInt64Array<"u"&&BigInt64Array.from,e=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;t&&(Zr.set("int64",BigInt64Array),ha.set(BigInt64Array,"int64")),e&&(Zr.set("uint64",BigUint64Array),ha.set(BigUint64Array,"uint64")),r?(Zr.set("float16",Float16Array),ha.set(Float16Array,"float16")):Zr.set("float16",Uint16Array)}}}),Qp,Zp,ry=J(()=>{Uo(),Qp=t=>{let e=1;for(let r=0;r{switch(t.location){case"cpu":return new lr(t.type,t.data,e);case"cpu-pinned":return new lr({location:"cpu-pinned",data:t.data,type:t.type,dims:e});case"texture":return new lr({location:"texture",texture:t.texture,type:t.type,dims:e});case"gpu-buffer":return new lr({location:"gpu-buffer",gpuBuffer:t.gpuBuffer,type:t.type,dims:e});default:throw new Error(`tensorReshape: tensor location ${t.location} is not supported`)}}}),lr,Uo=J(()=>{J0(),ey(),ty(),ry(),lr=class{constructor(t,e,r){Xp();let n,a;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,n=t.type,a=t.dims,t.location){case"cpu-pinned":{let s=Zr.get(n);if(!s)throw new TypeError(`unsupported type "${n}" to create tensor from pinned buffer`);if(!(t.data instanceof s))throw new TypeError(`buffer should be of type ${s.name}`);this.cpuData=t.data;break}case"texture":{if(n!=="float32")throw new TypeError(`unsupported type "${n}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break}case"gpu-buffer":{if(n!=="float32"&&n!=="float16"&&n!=="int32"&&n!=="int64"&&n!=="uint32"&&n!=="uint8"&&n!=="bool")throw new TypeError(`unsupported type "${n}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let s,o;if(typeof t=="string")if(n=t,o=r,t==="string"){if(!Array.isArray(e))throw new TypeError("A string tensor's data must be a string array.");s=e}else{let l=Zr.get(t);if(l===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(e)){if(t==="float16"&&l===Uint16Array)throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");t==="uint64"||t==="int64"?s=l.from(e,BigInt):s=l.from(e)}else if(e instanceof l)s=e;else throw new TypeError(`A ${n} tensor's data must be type of ${l}`)}else if(o=e,Array.isArray(t)){if(t.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let l=typeof t[0];if(l==="string")n="string",s=t;else if(l==="boolean")n="bool",s=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${l}.`)}else{let l=ha.get(t.constructor);if(l===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);n=l,s=t}if(o===void 0)o=[s.length];else if(!Array.isArray(o))throw new TypeError("A tensor's dims must be a number array");a=o,this.cpuData=s,this.dataLocation="cpu"}let i=Qp(a);if(this.cpuData&&i!==this.cpuData.length)throw new Error(`Tensor's size(${i}) does not match data length(${this.cpuData.length}).`);this.type=n,this.dims=a,this.size=i}static async fromImage(t,e){return jp(t,e)}static fromTexture(t,e){return qp(t,e)}static fromGpuBuffer(t,e){return Kp(t,e)}static fromPinnedBuffer(t,e,r){return Yp(t,e,r)}toDataURL(t){return Gp(this,t)}toImageData(t){return Hp(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 e=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=e,t&&this.disposer&&(this.disposer(),this.disposer=void 0),e}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Zp(this,t)}}}),Et,Wo=J(()=>{Uo(),Et=lr}),va,Ps,ur,jt,Jp=J(()=>{Vp(),va=(t,e)=>{(typeof Ht.trace>"u"?!Ht.wasm.trace:!Ht.trace)||console.timeStamp(`${t}::ORT::${e}`)},Ps=(t,e)=>{let r=new Error().stack?.split(/\r\n|\r|\n/g)||[],n=!1;for(let a=0;a{(typeof Ht.trace>"u"?!Ht.wasm.trace:!Ht.trace)||Ps("BEGIN",t)},jt=t=>{(typeof Ht.trace>"u"?!Ht.wasm.trace:!Ht.trace)||Ps("END",t)}}),eh,ny=J(()=>{Lo(),Wo(),Jp(),eh=class th{constructor(e){this.handler=e}async run(e,r,n){ur();let a={},i={};if(typeof e!="object"||e===null||e instanceof Et||Array.isArray(e))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let s=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof Et)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.");s=!1;for(let u of r){if(typeof u!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(u)===-1)throw new RangeError(`'fetches' contains invalid output name: ${u}.`);a[u]=null}if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let u=!1,p=Object.getOwnPropertyNames(r);for(let h of this.outputNames)if(p.indexOf(h)!==-1){let m=r[h];(m===null||m instanceof Et)&&(u=!0,s=!1,a[h]=m)}if(u){if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else i=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let u of this.inputNames)if(typeof e[u]>"u")throw new Error(`input '${u}' is missing in 'feeds'.`);if(s)for(let u of this.outputNames)a[u]=null;let o=await this.handler.run(e,a,i),l={};for(let u in o)if(Object.hasOwnProperty.call(o,u)){let p=o[u];p instanceof Et?l[u]=p:l[u]=new Et(p.type,p.data,p.dims)}return jt(),l}async release(){return this.handler.dispose()}static async create(e,r,n,a){ur();let i,s={};if(typeof e=="string"){if(i=e,typeof r=="object"&&r!==null)s=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(e instanceof Uint8Array){if(i=e,typeof r=="object"&&r!==null)s=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(e instanceof ArrayBuffer||typeof SharedArrayBuffer<"u"&&e instanceof SharedArrayBuffer){let p=e,h=0,m=e.byteLength;if(typeof r=="object"&&r!==null)s=r;else if(typeof r=="number"){if(h=r,!Number.isSafeInteger(h))throw new RangeError("'byteOffset' must be an integer.");if(h<0||h>=p.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${p.byteLength}).`);if(m=e.byteLength-h,typeof n=="number"){if(m=n,!Number.isSafeInteger(m))throw new RangeError("'byteLength' must be an integer.");if(m<=0||h+m>p.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${p.byteLength-h}].`);if(typeof a=="object"&&a!==null)s=a;else if(typeof a<"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(p,h,m)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[o,l]=await Fo(s),u=await o.createInferenceSessionHandler(i,l);return jt(),new th(u)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}}),Vo,ay=J(()=>{ny(),Vo=eh}),iy=J(()=>{}),sy=J(()=>{}),oy=J(()=>{}),ly=J(()=>{}),su,rh,uy=J(()=>{Lo(),Wo(),su="Training backend could not be resolved. Make sure you're using the correct configuration & WebAssembly files.",rh=class nh{constructor(e,r,n){this.handler=e,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(e,r){let n=e.evalModel||"",a=e.optimizerModel||"",i=r||{},[s,o]=await Fo(i);if(s.createTrainingSessionHandler){let l=await s.createTrainingSessionHandler(e.checkpointState,e.trainModel,n,a,o);return new nh(l,!!e.optimizerModel,!!e.evalModel)}else throw new Error(su)}typeNarrowingForRunStep(e,r,n,a,i){let s={},o={};if(typeof n!="object"||n===null||n instanceof Et||Array.isArray(n))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let l=!0;if(typeof a=="object"){if(a===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(a instanceof Et)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(a)){if(a.length===0)throw new TypeError("'fetches' cannot be an empty array.");l=!1;for(let u of a){if(typeof u!="string")throw new TypeError("'fetches' must be a string array or an object.");if(r.indexOf(u)===-1)throw new RangeError(`'fetches' contains invalid output name: ${u}.`);s[u]=null}if(typeof i=="object"&&i!==null)o=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else{let u=!1,p=Object.getOwnPropertyNames(a);for(let h of r)if(p.indexOf(h)!==-1){let m=a[h];(m===null||m instanceof Et)&&(u=!0,l=!1,s[h]=m)}if(u){if(typeof i=="object"&&i!==null)o=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else o=a}}else if(typeof a<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let u of e)if(typeof n[u]>"u")throw new Error(`input '${u}' is missing in 'feeds'.`);if(l)for(let u of r)s[u]=null;return[s,o]}convertHandlerReturnTypeToMapOfTensors(e){let r={};for(let n in e)if(Object.hasOwnProperty.call(e,n)){let a=e[n];a instanceof Et?r[n]=a:r[n]=new Et(a.type,a.data,a.dims)}return r}async lazyResetGrad(){await this.handler.lazyResetGrad()}async runTrainStep(e,r,n){let[a,i]=this.typeNarrowingForRunStep(this.trainingInputNames,this.trainingOutputNames,e,r,n),s=await this.handler.runTrainStep(e,a,i);return this.convertHandlerReturnTypeToMapOfTensors(s)}async runOptimizerStep(e){if(this.hasOptimizerModel)await this.handler.runOptimizerStep(e||{});else throw new Error("This TrainingSession has no OptimizerModel loaded.")}async runEvalStep(e,r,n){if(this.hasEvalModel){let[a,i]=this.typeNarrowingForRunStep(this.evalInputNames,this.evalOutputNames,e,r,n),s=await this.handler.runEvalStep(e,a,i);return this.convertHandlerReturnTypeToMapOfTensors(s)}else throw new Error("This TrainingSession has no EvalModel loaded.")}async getParametersSize(e=!0){return this.handler.getParametersSize(e)}async loadParametersBuffer(e,r=!0){let n=await this.getParametersSize(r);if(e.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(e,r)}async getContiguousParameters(e=!0){return this.handler.getContiguousParameters(e)}async release(){return this.handler.dispose()}}}),Go,dy=J(()=>{uy(),Go=rh}),ah={};Dn(ah,{InferenceSession:()=>Vo,TRACE:()=>va,TRACE_FUNC_BEGIN:()=>ur,TRACE_FUNC_END:()=>jt,Tensor:()=>Et,TrainingSession:()=>Go,env:()=>Ue,registerBackend:()=>tn});var dr=J(()=>{X0(),Z0(),ay(),Wo(),iy(),sy(),Jp(),oy(),ly(),dy()}),Ho={};Dn(Ho,{createReadStream:()=>oh,readFile:()=>ih,readFileSync:()=>sh});var ih,sh,oh,lh=J(()=>{ih=void 0,sh=void 0,oh=void 0}),jo={};Dn(jo,{join:()=>uh});var uh,dh=J(()=>{uh=void 0}),cy=Bn((t,e)=>{var r=(()=>{var n=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(n||=__filename),function(a={}){var i=a,s,o,l=new Promise((f,y)=>{s=f,o=y});i.mountExternalData=(f,y)=>{(i.eb||(i.eb=new Map)).set(f,y)},i.unmountExternalData=()=>{delete i.eb};let u=()=>{let f=(x,C,O)=>(...N)=>{let j=It,Z=C?.();N=x(...N);let ce=C?.();return Z!==ce&&(x=ce,O(Z),C=O=null),It!=j?ri():N},y=x=>async(...C)=>{try{if(i.cb)throw Error("Session already started");let O=i.cb={xb:C[0],errors:[]},N=await x(...C);if(i.cb!==O)throw Error("Session mismatch");i.kb?.flush();let j=O.errors;if(0ce),0i._OrtCreateSession,x=>i._OrtCreateSession=x),i._OrtRun=y(f(i._OrtRun,()=>i._OrtRun,x=>i._OrtRun=x)),i._OrtRunWithBinding=y(f(i._OrtRunWithBinding,()=>i._OrtRunWithBinding,x=>i._OrtRunWithBinding=x)),i._OrtBindInput=f(i._OrtBindInput,()=>i._OrtBindInput,x=>i._OrtBindInput=x),u=void 0};i.jsepInit=(f,y)=>{if(u?.(),f==="webgpu"){[i.kb,i.pb,i.tb,i.lb,i.sb,i.Ra,i.ub,i.wb,i.qb,i.rb,i.vb]=y;let x=i.kb;i.jsepRegisterBuffer=(C,O,N,j)=>x.registerBuffer(C,O,N,j),i.jsepGetBuffer=C=>x.getBuffer(C),i.jsepCreateDownloader=(C,O,N)=>x.createDownloader(C,O,N),i.jsepOnReleaseSession=C=>{x.onReleaseSession(C)},i.jsepOnRunStart=C=>x.onRunStart(C)}};var p=Object.assign({},i),h="./this.program",m=(f,y)=>{throw y},d=typeof window=="object",_=typeof importScripts=="function",w=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",v="",S,$,E;if(w){var T=(lh(),nn(Ho)),A=(dh(),nn(jo));v=_?A.dirname(v)+"/":__dirname+"/",S=(f,y)=>(f=pt(f)?new URL(f):A.normalize(f),T.readFileSync(f,y?void 0:"utf8")),E=f=>(f=S(f,!0),f.buffer||(f=new Uint8Array(f)),f),$=(f,y,x,C=!0)=>{f=pt(f)?new URL(f):A.normalize(f),T.readFile(f,C?void 0:"utf8",(O,N)=>{O?x(O):y(C?N.buffer:N)})},!i.thisProgram&&1{throw process.exitCode=f,y}}else(d||_)&&(_?v=self.location.href:typeof document<"u"&&document.currentScript&&(v=document.currentScript.src),n&&(v=n),v.startsWith("blob:")?v="":v=v.substr(0,v.replace(/[?#].*/,"").lastIndexOf("/")+1),S=f=>{var y=new XMLHttpRequest;return y.open("GET",f,!1),y.send(null),y.responseText},_&&(E=f=>{var y=new XMLHttpRequest;return y.open("GET",f,!1),y.responseType="arraybuffer",y.send(null),new Uint8Array(y.response)}),$=(f,y,x)=>{var C=new XMLHttpRequest;C.open("GET",f,!0),C.responseType="arraybuffer",C.onload=()=>{C.status==200||C.status==0&&C.response?y(C.response):x()},C.onerror=x,C.send(null)});var P=console.log.bind(console),B=console.error.bind(console);Object.assign(i,p),p=null;var D,q=!1,H,ie,te,de,se,M,R,G,re,ee,ne;function W(){var f=D.buffer;i.HEAP8=ie=new Int8Array(f),i.HEAP16=de=new Int16Array(f),i.HEAPU8=te=new Uint8Array(f),i.HEAPU16=se=new Uint16Array(f),i.HEAP32=M=new Int32Array(f),i.HEAPU32=R=new Uint32Array(f),i.HEAPF32=G=new Float32Array(f),i.HEAPF64=ne=new Float64Array(f),i.HEAP64=re=new BigInt64Array(f),i.HEAPU64=ee=new BigUint64Array(f)}var ae=[],fe=[],Ce=[],Be=0,Ye=null;function Qe(f){throw f="Aborted("+f+")",B(f),q=!0,H=1,f=new WebAssembly.RuntimeError(f+". Build with -sASSERTIONS for more info."),o(f),f}var We=f=>f.startsWith("data:application/octet-stream;base64,"),pt=f=>f.startsWith("file://"),Ve;if(Ve="ort-wasm-simd.wasm",!We(Ve)){var Ne=Ve;Ve=i.locateFile?i.locateFile(Ne,v):v+Ne}function gt(f){if(E)return E(f);throw"both async and sync fetching of the wasm failed"}function Xt(f){if(d||_){if(typeof fetch=="function"&&!pt(f))return fetch(f,{credentials:"same-origin"}).then(y=>{if(!y.ok)throw`failed to load wasm binary file at '${f}'`;return y.arrayBuffer()}).catch(()=>gt(f));if($)return new Promise((y,x)=>{$(f,C=>y(new Uint8Array(C)),x)})}return Promise.resolve().then(()=>gt(f))}function lt(f,y,x){return Xt(f).then(C=>WebAssembly.instantiate(C,y)).then(x,C=>{B(`failed to asynchronously prepare wasm: ${C}`),Qe(C)})}function Fr(f,y){var x=Ve;return typeof WebAssembly.instantiateStreaming!="function"||We(x)||pt(x)||w||typeof fetch!="function"?lt(x,f,y):fetch(x,{credentials:"same-origin"}).then(C=>WebAssembly.instantiateStreaming(C,f).then(y,function(O){return B(`wasm streaming compile failed: ${O}`),B("falling back to ArrayBuffer instantiation"),lt(x,f,y)}))}var wr={824920:(f,y,x,C)=>{if(typeof i>"u"||!i.eb)return 1;if(f=Xe(f>>>0),f.startsWith("./")&&(f=f.substring(2)),f=i.eb.get(f),!f)return 2;if(y>>>=0,x>>>=0,y+x>f.byteLength)return 3;try{return te.set(f.subarray(y,y+x),C>>>0>>>0),0}catch{return 4}},825421:()=>{i.qb()},825452:()=>{i.rb()},825481:()=>{i.vb()},825506:f=>i.pb(f),825539:f=>i.tb(f),825571:(f,y,x)=>{i.lb(f,y,x,!0)},825610:(f,y,x)=>{i.lb(f,y,x)},825643:f=>{i.Ra("Abs",f,void 0)},825694:f=>{i.Ra("Neg",f,void 0)},825745:f=>{i.Ra("Floor",f,void 0)},825798:f=>{i.Ra("Ceil",f,void 0)},825850:f=>{i.Ra("Reciprocal",f,void 0)},825908:f=>{i.Ra("Sqrt",f,void 0)},825960:f=>{i.Ra("Exp",f,void 0)},826011:f=>{i.Ra("Erf",f,void 0)},826062:f=>{i.Ra("Sigmoid",f,void 0)},826117:(f,y,x)=>{i.Ra("HardSigmoid",f,{alpha:y,beta:x})},826196:f=>{i.Ra("Log",f,void 0)},826247:f=>{i.Ra("Sin",f,void 0)},826298:f=>{i.Ra("Cos",f,void 0)},826349:f=>{i.Ra("Tan",f,void 0)},826400:f=>{i.Ra("Asin",f,void 0)},826452:f=>{i.Ra("Acos",f,void 0)},826504:f=>{i.Ra("Atan",f,void 0)},826556:f=>{i.Ra("Sinh",f,void 0)},826608:f=>{i.Ra("Cosh",f,void 0)},826660:f=>{i.Ra("Asinh",f,void 0)},826713:f=>{i.Ra("Acosh",f,void 0)},826766:f=>{i.Ra("Atanh",f,void 0)},826819:f=>{i.Ra("Tanh",f,void 0)},826871:f=>{i.Ra("Not",f,void 0)},826922:(f,y,x)=>{i.Ra("Clip",f,{min:y,max:x})},826991:f=>{i.Ra("Clip",f,void 0)},827043:(f,y)=>{i.Ra("Elu",f,{alpha:y})},827101:f=>{i.Ra("Relu",f,void 0)},827153:(f,y)=>{i.Ra("LeakyRelu",f,{alpha:y})},827217:(f,y)=>{i.Ra("ThresholdedRelu",f,{alpha:y})},827287:(f,y)=>{i.Ra("Cast",f,{to:y})},827345:f=>{i.Ra("Add",f,void 0)},827396:f=>{i.Ra("Sub",f,void 0)},827447:f=>{i.Ra("Mul",f,void 0)},827498:f=>{i.Ra("Div",f,void 0)},827549:f=>{i.Ra("Pow",f,void 0)},827600:f=>{i.Ra("Equal",f,void 0)},827653:f=>{i.Ra("Greater",f,void 0)},827708:f=>{i.Ra("GreaterOrEqual",f,void 0)},827770:f=>{i.Ra("Less",f,void 0)},827822:f=>{i.Ra("LessOrEqual",f,void 0)},827881:(f,y,x,C,O)=>{i.Ra("ReduceMean",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},828040:(f,y,x,C,O)=>{i.Ra("ReduceMax",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},828198:(f,y,x,C,O)=>{i.Ra("ReduceMin",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},828356:(f,y,x,C,O)=>{i.Ra("ReduceProd",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},828515:(f,y,x,C,O)=>{i.Ra("ReduceSum",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},828673:(f,y,x,C,O)=>{i.Ra("ReduceL1",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},828830:(f,y,x,C,O)=>{i.Ra("ReduceL2",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},828987:(f,y,x,C,O)=>{i.Ra("ReduceLogSum",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},829148:(f,y,x,C,O)=>{i.Ra("ReduceSumSquare",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},829312:(f,y,x,C,O)=>{i.Ra("ReduceLogSumExp",f,{keepDims:!!y,noopWithEmptyAxes:!!x,axes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},829476:f=>{i.Ra("Where",f,void 0)},829529:(f,y,x)=>{i.Ra("Transpose",f,{perm:y?Array.from(M.subarray(y>>>0,x>>>0)):[]})},829637:(f,y,x,C)=>{i.Ra("DepthToSpace",f,{blocksize:y,mode:Xe(x),format:C?"NHWC":"NCHW"})},829770:(f,y,x,C)=>{i.Ra("DepthToSpace",f,{blocksize:y,mode:Xe(x),format:C?"NHWC":"NCHW"})},829903:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F,xe)=>{i.Ra("ConvTranspose",f,{format:ce?"NHWC":"NCHW",autoPad:y,dilations:[x],group:C,kernelShape:[O],pads:[N,j],strides:[Z],wIsConst:()=>!!ie[oe>>>0],outputPadding:me?Array.from(M.subarray(me>>>0,ze>>>0)):[],outputShape:Fe?Array.from(M.subarray(Fe>>>0,F>>>0)):[],activation:Xe(xe)})},830304:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F)=>{i.Ra("ConvTranspose",f,{format:Z?"NHWC":"NCHW",autoPad:y,dilations:Array.from(M.subarray(x>>>0,(x>>>0)+2>>>0)),group:C,kernelShape:Array.from(M.subarray(O>>>0,(O>>>0)+2>>>0)),pads:Array.from(M.subarray(N>>>0,(N>>>0)+4>>>0)),strides:Array.from(M.subarray(j>>>0,(j>>>0)+2>>>0)),wIsConst:()=>!!ie[ce>>>0],outputPadding:oe?Array.from(M.subarray(oe>>>0,me>>>0)):[],outputShape:ze?Array.from(M.subarray(ze>>>0,Fe>>>0)):[],activation:Xe(F)})},830869:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F,xe)=>{i.Ra("ConvTranspose",f,{format:ce?"NHWC":"NCHW",autoPad:y,dilations:[x],group:C,kernelShape:[O],pads:[N,j],strides:[Z],wIsConst:()=>!!ie[oe>>>0],outputPadding:me?Array.from(M.subarray(me>>>0,ze>>>0)):[],outputShape:Fe?Array.from(M.subarray(Fe>>>0,F>>>0)):[],activation:Xe(xe)})},831270:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F)=>{i.Ra("ConvTranspose",f,{format:Z?"NHWC":"NCHW",autoPad:y,dilations:Array.from(M.subarray(x>>>0,(x>>>0)+2>>>0)),group:C,kernelShape:Array.from(M.subarray(O>>>0,(O>>>0)+2>>>0)),pads:Array.from(M.subarray(N>>>0,(N>>>0)+4>>>0)),strides:Array.from(M.subarray(j>>>0,(j>>>0)+2>>>0)),wIsConst:()=>!!ie[ce>>>0],outputPadding:oe?Array.from(M.subarray(oe>>>0,me>>>0)):[],outputShape:ze?Array.from(M.subarray(ze>>>0,Fe>>>0)):[],activation:Xe(F)})},831835:(f,y)=>{i.Ra("GlobalAveragePool",f,{format:y?"NHWC":"NCHW"})},831926:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F,xe,Oe)=>{i.Ra("AveragePool",f,{format:Oe?"NHWC":"NCHW",auto_pad:y,ceil_mode:x,count_include_pad:C,storage_order:O,dilations:[N,j],kernel_shape:[Z,ce],pads:[oe,me,ze,Fe],strides:[F,xe]})},832210:(f,y)=>{i.Ra("GlobalAveragePool",f,{format:y?"NHWC":"NCHW"})},832301:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F,xe,Oe)=>{i.Ra("AveragePool",f,{format:Oe?"NHWC":"NCHW",auto_pad:y,ceil_mode:x,count_include_pad:C,storage_order:O,dilations:[N,j],kernel_shape:[Z,ce],pads:[oe,me,ze,Fe],strides:[F,xe]})},832585:(f,y)=>{i.Ra("GlobalMaxPool",f,{format:y?"NHWC":"NCHW"})},832672:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F,xe,Oe)=>{i.Ra("MaxPool",f,{format:Oe?"NHWC":"NCHW",auto_pad:y,ceil_mode:x,count_include_pad:C,storage_order:O,dilations:[N,j],kernel_shape:[Z,ce],pads:[oe,me,ze,Fe],strides:[F,xe]})},832952:(f,y)=>{i.Ra("GlobalMaxPool",f,{format:y?"NHWC":"NCHW"})},833039:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F,xe,Oe)=>{i.Ra("MaxPool",f,{format:Oe?"NHWC":"NCHW",auto_pad:y,ceil_mode:x,count_include_pad:C,storage_order:O,dilations:[N,j],kernel_shape:[Z,ce],pads:[oe,me,ze,Fe],strides:[F,xe]})},833319:(f,y,x,C,O)=>{i.Ra("Gemm",f,{alpha:y,beta:x,transA:C,transB:O})},833423:f=>{i.Ra("MatMul",f,void 0)},833477:(f,y,x,C)=>{i.Ra("ArgMax",f,{keepDims:!!y,selectLastIndex:!!x,axis:C})},833585:(f,y,x,C)=>{i.Ra("ArgMin",f,{keepDims:!!y,selectLastIndex:!!x,axis:C})},833693:(f,y)=>{i.Ra("Softmax",f,{axis:y})},833756:(f,y)=>{i.Ra("Concat",f,{axis:y})},833816:(f,y,x,C,O)=>{i.Ra("Split",f,{axis:y,numOutputs:x,splitSizes:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},833956:f=>{i.Ra("Expand",f,void 0)},834010:(f,y)=>{i.Ra("Gather",f,{axis:Number(y)})},834081:(f,y)=>{i.Ra("GatherElements",f,{axis:Number(y)})},834160:(f,y,x,C,O,N,j,Z,ce,oe,me)=>{i.Ra("Resize",f,{antialias:y,axes:x?Array.from(M.subarray(x>>>0,C>>>0)):[],coordinateTransformMode:Xe(O),cubicCoeffA:N,excludeOutside:j,extrapolationValue:Z,keepAspectRatioPolicy:Xe(ce),mode:Xe(oe),nearestMode:Xe(me)})},834506:(f,y,x,C,O,N,j)=>{i.Ra("Slice",f,{starts:y?Array.from(M.subarray(y>>>0,x>>>0)):[],ends:C?Array.from(M.subarray(C>>>0,O>>>0)):[],axes:N?Array.from(M.subarray(N>>>0,j>>>0)):[]})},834722:f=>{i.Ra("Tile",f,void 0)},834774:(f,y,x,C)=>{i.Ra("LayerNormalization",f,{axis:y,epsilon:x,simplified:!!C})},834885:(f,y,x)=>{i.Ra("InstanceNormalization",f,{epsilon:y,format:x?"NHWC":"NCHW"})},834999:(f,y,x)=>{i.Ra("InstanceNormalization",f,{epsilon:y,format:x?"NHWC":"NCHW"})},835113:f=>{i.Ra("Range",f,void 0)},835166:(f,y)=>{i.Ra("Einsum",f,{equation:Xe(y)})},835247:(f,y,x,C,O)=>{i.Ra("Pad",f,{mode:y,value:x,pads:C?Array.from(M.subarray(C>>>0,O>>>0)):[]})},835374:(f,y,x,C,O,N)=>{i.Ra("BatchNormalization",f,{epsilon:y,momentum:x,spatial:!!O,trainingMode:!!C,format:N?"NHWC":"NCHW"})},835543:(f,y,x,C,O,N)=>{i.Ra("BatchNormalization",f,{epsilon:y,momentum:x,spatial:!!O,trainingMode:!!C,format:N?"NHWC":"NCHW"})},835712:(f,y,x)=>{i.Ra("CumSum",f,{exclusive:Number(y),reverse:Number(x)})},835809:(f,y,x,C,O,N,j,Z,ce)=>{i.Ra("Attention",f,{numHeads:y,isUnidirectional:x,maskFilterValue:C,scale:O,doRotary:N,qkvHiddenSizes:j?Array.from(M.subarray(Number(Z)>>>0,Number(Z)+j>>>0)):[],pastPresentShareBuffer:!!ce})},836081:f=>{i.Ra("BiasAdd",f,void 0)},836136:f=>{i.Ra("BiasSplitGelu",f,void 0)},836197:f=>{i.Ra("FastGelu",f,void 0)},836253:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe)=>{i.Ra("Conv",f,{format:ce?"NHWC":"NCHW",auto_pad:y,dilations:[x],group:C,kernel_shape:[O],pads:N?Array.from(M.subarray(N>>>0,j>>>0)):[],strides:[Z],w_is_const:()=>!!ie[oe>>>0],activation:Xe(me),activation_params:ze?Array.from(G.subarray(ze>>>0,Fe>>>0)):[]})},836623:(f,y,x,C,O,N,j,Z,ce,oe,me,ze,Fe,F,xe,Oe)=>{i.Ra("Conv",f,{format:ze?"NHWC":"NCHW",auto_pad:y,dilations:[x,C],group:O,kernel_shape:[N,j],pads:Z?Array.from(M.subarray(Z>>>0,ce>>>0)):[],strides:[oe,me],w_is_const:()=>!!ie[Fe>>>0],activation:Xe(F),activation_params:xe?Array.from(G.subarray(xe>>>0,Oe>>>0)):[]})},837014:f=>{i.Ra("Gelu",f,void 0)},837066:(f,y,x,C,O,N)=>{i.Ra("MatMulNBits",f,{k:y,n:x,accuracyLevel:C,bits:O,blockSize:N})},837193:(f,y,x,C,O,N)=>{i.Ra("MultiHeadAttention",f,{numHeads:y,isUnidirectional:x,maskFilterValue:C,scale:O,doRotary:N})},837352:(f,y,x,C,O)=>{i.Ra("RotaryEmbedding",f,{interleaved:!!y,numHeads:x,rotaryEmbeddingDim:C,scale:O})},837491:(f,y,x)=>{i.Ra("SkipLayerNormalization",f,{epsilon:y,simplified:!!x})},837593:(f,y,x)=>{i.Ra("SkipLayerNormalization",f,{epsilon:y,simplified:!!x})},837695:(f,y,x,C)=>{i.Ra("LayerNormalization",f,{axis:y,epsilon:x,simplified:!!C})},837806:f=>{i.ub(f)},837840:(f,y)=>i.wb(f,y,i.cb.xb,i.cb.errors)};function cr(f){this.name="ExitStatus",this.message=`Program terminated with exit(${f})`,this.status=f}class _s{constructor(y){this.hb=y-24}}var br=0,Ar=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Wn=(f,y,x)=>{y>>>=0;var C=y+x;for(x=y;f[x]&&!(x>=C);)++x;if(16O?C+=String.fromCharCode(O):(O-=65536,C+=String.fromCharCode(55296|O>>10,56320|O&1023))}}else C+=String.fromCharCode(O)}return C},Xe=(f,y)=>(f>>>=0)?Wn(te,f,y):"",hn=f=>{for(var y=0,x=0;x=C?y++:2047>=C?y+=2:55296<=C&&57343>=C?(y+=4,++x):y+=3}return y},Wt=(f,y,x,C)=>{if(x>>>=0,!(0=j){var Z=f.charCodeAt(++N);j=65536+((j&1023)<<10)|Z&1023}if(127>=j){if(x>=C)break;y[x++>>>0]=j}else{if(2047>=j){if(x+1>=C)break;y[x++>>>0]=192|j>>6}else{if(65535>=j){if(x+2>=C)break;y[x++>>>0]=224|j>>12}else{if(x+3>=C)break;y[x++>>>0]=240|j>>18,y[x++>>>0]=128|j>>12&63}y[x++>>>0]=128|j>>6&63}y[x++>>>0]=128|j&63}}return y[x>>>0]=0,x-O},Ir,Bt=f=>{for(var y="";te[f>>>0];)y+=Ir[te[f++>>>0]];return y},Vn={},fn={},pr;function Gn(f,y,x={}){var C=y.name;if(!f)throw new pr(`type "${C}" must have a positive integer typeid pointer`);if(fn.hasOwnProperty(f)){if(x.nb)return;throw new pr(`Cannot register type '${C}' twice`)}fn[f]=y,Vn.hasOwnProperty(f)&&(y=Vn[f],delete Vn[f],y.forEach(O=>O()))}function Tt(f,y,x={}){if(!("argPackAdvance"in y))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Gn(f,y,x)}var mn=(f,y,x)=>{switch(y){case 1:return x?C=>ie[C>>>0]:C=>te[C>>>0];case 2:return x?C=>de[C>>>1>>>0]:C=>se[C>>>1>>>0];case 4:return x?C=>M[C>>>2>>>0]:C=>R[C>>>2>>>0];case 8:return x?C=>re[C>>>3]:C=>ee[C>>>3];default:throw new TypeError(`invalid integer width (${y}): ${f}`)}},gn=[],ut=[];function Hn(f){f>>>=0,9{if(!f)throw new pr("Cannot use deleted val. handle = "+f);return ut[f]},At=f=>{switch(f){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let y=gn.pop()||ut.length;return ut[y]=f,ut[y+1]=1,y}};function Lr(f){return this.fromWireType(R[f>>>2>>>0])}var ys={name:"emscripten::val",fromWireType:f=>{var y=Ie(f);return Hn(f),y},toWireType:(f,y)=>At(y),argPackAdvance:8,readValueFromPointer:Lr,bb:null},Ka=(f,y)=>{switch(y){case 4:return function(x){return this.fromWireType(G[x>>>2>>>0])};case 8:return function(x){return this.fromWireType(ne[x>>>3>>>0])};default:throw new TypeError(`invalid float width (${y}): ${f}`)}},Ya=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,be=(f,y)=>{for(var x=f>>1,C=x+y/2;!(x>=C)&&se[x>>>0];)++x;if(x<<=1,32>>0,x>>>0));for(x="",C=0;!(C>=y/2);++C){var O=de[f+2*C>>>1>>>0];if(O==0)break;x+=String.fromCharCode(O)}return x},_n=(f,y,x)=>{if(x??=2147483647,2>x)return 0;x-=2;var C=y;x=x<2*f.length?x/2:f.length;for(var O=0;O>>1>>>0]=f.charCodeAt(O),y+=2;return de[y>>>1>>>0]=0,y-C},Xa=f=>2*f.length,ws=(f,y)=>{for(var x=0,C="";!(x>=y/4);){var O=M[f+4*x>>>2>>>0];if(O==0)break;++x,65536<=O?(O-=65536,C+=String.fromCharCode(55296|O>>10,56320|O&1023)):C+=String.fromCharCode(O)}return C},Qa=(f,y,x)=>{if(y>>>=0,x??=2147483647,4>x)return 0;var C=y;x=C+x-4;for(var O=0;O=N){var j=f.charCodeAt(++O);N=65536+((N&1023)<<10)|j&1023}if(M[y>>>2>>>0]=N,y+=4,y+4>x)break}return M[y>>>2>>>0]=0,y-C},Za=f=>{for(var y=0,x=0;x=C&&++x,y+=4}return y},yn=(f,y)=>{var x=fn[f];if(x===void 0)throw f=ta(f),x=Bt(f),xt(f),new pr(`${y} has unknown type ${x}`);return x},jn=(f,y,x)=>{var C=[];return f=f.toWireType(C,x),C.length&&(R[y>>>2>>>0]=At(C)),f},Ur=f=>{try{f()}catch(y){Qe(y)}},it=f=>{if(!q)try{f();try{H=H=f=H,i.onExit?.(f),q=!0,m(f,new cr(f))}catch(y){y instanceof cr||y=="unwind"||m(1,y)}}catch(y){y instanceof cr||y=="unwind"||m(1,y)}};function Ja(){var f=we,y={};for(let[x,C]of Object.entries(f))y[x]=typeof C=="function"?(...O)=>{Zt.push(x);try{return C(...O)}finally{q||(Zt.pop(),It&&Qt===1&&Zt.length===0&&(Qt=0,Ur(ui),typeof Fibers<"u"&&Fibers.Db()))}}:C;return y}var Qt=0,It=null,qn=0,Zt=[],Kn={},Yn={},ei=0,wn=null,ti=[];function ri(){return new Promise((f,y)=>{wn={resolve:f,reject:y}})}function ni(){var f=Mr(65548),y=f+12;R[f>>>2>>>0]=y,R[f+4>>>2>>>0]=y+65536,y=Zt[0];var x=Kn[y];return x===void 0&&(x=ei++,Kn[y]=x,Yn[x]=y),M[f+8>>>2>>>0]=x,f}function ai(f){if(!q){if(Qt===0){var y=!1,x=!1;f((C=0)=>{if(!q&&(qn=C,y=!0,x)){Qt=2,Ur(()=>er(It)),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.resume(),C=!1;try{var O=(0,we[Yn[M[It+8>>>2>>>0]]])()}catch(Z){O=Z,C=!0}var N=!1;if(!It){var j=wn;j&&(wn=null,(C?j.reject:j.resolve)(O),N=!0)}if(C&&!N)throw O}}),x=!0,y||(Qt=1,It=ni(),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.pause(),Ur(()=>Gr(It)))}else Qt===2?(Qt=0,Ur(Ft),xt(It),It=null,ti.forEach(it)):Qe(`invalid state: ${Qt}`);return qn}}function Xn(f){return ai(y=>{f().then(y)})}var Wr=[],ii={},bn=f=>{var y=ii[f];return y===void 0?Bt(f):y},Dt=()=>typeof globalThis=="object"?globalThis:Function("return this")(),Qn=f=>{var y=Wr.length;return Wr.push(f),y},Zn=(f,y)=>{for(var x=Array(f),C=0;C>>2>>>0],"parameter "+C);return x},Jt=(f,y)=>Object.defineProperty(y,"name",{value:f});function bs(f){var y=Function;if(!(y instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof y} which is not a function`);var x=Jt(y.name||"unknownFunctionName",function(){});return x.prototype=y.prototype,x=new x,f=y.apply(x,f),f instanceof Object?f:x}var _t=f=>f%4===0&&(f%100!==0||f%400===0),Jn=[0,31,60,91,121,152,182,213,244,274,305,335],vn=[0,31,59,90,120,151,181,212,243,273,304,334],Nt=[],$n=(f,y)=>{Nt.length=0;for(var x;x=te[f++>>>0];){var C=x!=105;C&=x!=112,y+=C&&y%8?4:0,Nt.push(x==112?R[y>>>2>>>0]:x==106?re[y>>>3]:x==105?M[y>>>2>>>0]:ne[y>>>3>>>0]),y+=C?8:4}return Nt},dt={},Mt=()=>{if(!Vr){var f={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",_:h||"./this.program"},y;for(y in dt)dt[y]===void 0?delete f[y]:f[y]=dt[y];var x=[];for(y in f)x.push(`${y}=${f[y]}`);Vr=x}return Vr},Vr,vs=[null,[],[]],si=[31,29,31,30,31,30,31,31,30,31,30,31],ea=[31,28,31,30,31,30,31,31,30,31,30,31];function $s(f){var y=Array(hn(f)+1);return Wt(f,y,0,y.length),y}function oi(f,y,x,C){function O(F,xe,Oe){for(F=typeof F=="number"?F.toString():F||"";F.lengthxr?-1:0$r-F.getDate())xe-=$r-F.getDate()+1,F.setDate(1),11>Oe?F.setMonth(Oe+1):(F.setMonth(0),F.setFullYear(F.getFullYear()+1));else{F.setDate(F.getDate()+xe);break}}return Oe=new Date(F.getFullYear()+1,0,4),xe=Z(new Date(F.getFullYear(),0,4)),Oe=Z(Oe),0>=j(xe,F)?0>=j(Oe,F)?F.getFullYear()+1:F.getFullYear():F.getFullYear()-1}f>>>=0,y>>>=0,x>>>=0,C>>>=0;var oe=R[C+40>>>2>>>0];C={Ab:M[C>>>2>>>0],zb:M[C+4>>>2>>>0],fb:M[C+8>>>2>>>0],jb:M[C+12>>>2>>>0],gb:M[C+16>>>2>>>0],ab:M[C+20>>>2>>>0],Va:M[C+24>>>2>>>0],$a:M[C+28>>>2>>>0],Cb:M[C+32>>>2>>>0],yb:M[C+36>>>2>>>0],Bb:oe?Xe(oe):""},x=Xe(x),oe={"%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"};for(var me in oe)x=x.replace(new RegExp(me,"g"),oe[me]);var ze="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Fe="January February March April May June July August September October November December".split(" ");oe={"%a":F=>ze[F.Va].substring(0,3),"%A":F=>ze[F.Va],"%b":F=>Fe[F.gb].substring(0,3),"%B":F=>Fe[F.gb],"%C":F=>N((F.ab+1900)/100|0,2),"%d":F=>N(F.jb,2),"%e":F=>O(F.jb,2," "),"%g":F=>ce(F).toString().substring(2),"%G":ce,"%H":F=>N(F.fb,2),"%I":F=>(F=F.fb,F==0?F=12:12{for(var xe=0,Oe=0;Oe<=F.gb-1;xe+=(_t(F.ab+1900)?si:ea)[Oe++]);return N(F.jb+xe,3)},"%m":F=>N(F.gb+1,2),"%M":F=>N(F.zb,2),"%n":()=>` +`,"%p":F=>0<=F.fb&&12>F.fb?"AM":"PM","%S":F=>N(F.Ab,2),"%t":()=>" ","%u":F=>F.Va||7,"%U":F=>N(Math.floor((F.$a+7-F.Va)/7),2),"%V":F=>{var xe=Math.floor((F.$a+7-(F.Va+6)%7)/7);if(2>=(F.Va+371-F.$a-2)%7&&xe++,xe)xe==53&&(Oe=(F.Va+371-F.$a)%7,Oe==4||Oe==3&&_t(F.ab)||(xe=1));else{xe=52;var Oe=(F.Va+7-F.$a-1)%7;(Oe==4||Oe==5&&_t(F.ab%400-1))&&xe++}return N(xe,2)},"%w":F=>F.Va,"%W":F=>N(Math.floor((F.$a+7-(F.Va+6)%7)/7),2),"%y":F=>(F.ab+1900).toString().substring(2),"%Y":F=>F.ab+1900,"%z":F=>{F=F.yb;var xe=0<=F;return F=Math.abs(F)/60,(xe?"+":"-")+("0000"+(F/60*100+F%60)).slice(-4)},"%Z":F=>F.Bb,"%%":()=>"%"},x=x.replace(/%%/g,"\0\0");for(me in oe)x.includes(me)&&(x=x.replace(new RegExp(me,"g"),oe[me](C)));return x=x.replace(/\0\0/g,"%"),me=$s(x),me.length>y?0:(ie.set(me,f>>>0),me.length-1)}for(var li=Array(256),xn=0;256>xn;++xn)li[xn]=String.fromCharCode(xn);Ir=li,pr=i.BindingError=class extends Error{constructor(f){super(f),this.name="BindingError"}},i.InternalError=class extends Error{constructor(f){super(f),this.name="InternalError"}},ut.push(0,1,void 0,1,null,1,!0,1,!1,1),i.count_emval_handles=()=>ut.length/2-5-gn.length;var xs={ia:function(f,y,x){return Xn(async()=>{await i.sb(f,y,x)})},a:function(f,y,x){f>>>=0;var C=new _s(f);throw R[C.hb+16>>>2>>>0]=0,R[C.hb+4>>>2>>>0]=y>>>0,R[C.hb+8>>>2>>>0]=x>>>0,br=f,br},y:function(){return 0},ea:function(){},R:function(){},T:function(){},ga:function(){return 0},ca:function(){},Z:function(){},ba:function(){},G:function(){},S:function(){},P:function(){},da:function(){},Q:function(){},C:function(f,y,x){y=Bt(y>>>0),Tt(f>>>0,{name:y,fromWireType:C=>C,toWireType:function(C,O){if(typeof O!="bigint"&&typeof O!="number")throw O===null?O="null":(C=typeof O,O=C==="object"||C==="array"||C==="function"?O.toString():""+O),new TypeError(`Cannot convert "${O}" to ${this.name}`);return typeof O=="number"&&(O=BigInt(O)),O},argPackAdvance:8,readValueFromPointer:mn(y,x>>>0,y.indexOf("u")==-1),bb:null})},K:function(f,y,x,C){y=Bt(y>>>0),Tt(f>>>0,{name:y,fromWireType:function(O){return!!O},toWireType:function(O,N){return N?x:C},argPackAdvance:8,readValueFromPointer:function(O){return this.fromWireType(te[O>>>0])},bb:null})},J:function(f){return Tt(f>>>0,ys)},B:function(f,y,x){y=Bt(y>>>0),Tt(f>>>0,{name:y,fromWireType:C=>C,toWireType:(C,O)=>O,argPackAdvance:8,readValueFromPointer:Ka(y,x>>>0),bb:null})},s:function(f,y,x,C,O){if(f>>>=0,x>>>=0,y=Bt(y>>>0),O===-1&&(O=4294967295),O=Z=>Z,C===0){var N=32-8*x;O=Z=>Z<>>N}var j=y.includes("unsigned")?function(Z,ce){return ce>>>0}:function(Z,ce){return ce};Tt(f,{name:y,fromWireType:O,toWireType:j,argPackAdvance:8,readValueFromPointer:mn(y,x,C!==0),bb:null})},o:function(f,y,x){function C(N){return new O(ie.buffer,R[N+4>>>2>>>0],R[N>>>2>>>0])}var O=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][y];x=Bt(x>>>0),Tt(f>>>0,{name:x,fromWireType:C,argPackAdvance:8,readValueFromPointer:C},{nb:!0})},D:function(f,y){y=Bt(y>>>0);var x=y==="std::string";Tt(f>>>0,{name:y,fromWireType:function(C){var O=R[C>>>2>>>0],N=C+4;if(x)for(var j=N,Z=0;Z<=O;++Z){var ce=N+Z;if(Z==O||te[ce>>>0]==0){if(j=Xe(j,ce-j),oe===void 0)var oe=j;else oe+="\0",oe+=j;j=ce+1}}else{for(oe=Array(O),Z=0;Z>>0]);oe=oe.join("")}return xt(C),oe},toWireType:function(C,O){O instanceof ArrayBuffer&&(O=new Uint8Array(O));var N=typeof O=="string";if(!(N||O instanceof Uint8Array||O instanceof Uint8ClampedArray||O instanceof Int8Array))throw new pr("Cannot pass non-string to std::string");var j=x&&N?hn(O):O.length,Z=Mr(4+j+1),ce=Z+4;if(R[Z>>>2>>>0]=j,x&&N)Wt(O,te,ce,j+1);else if(N)for(N=0;N>>0]=oe}else for(N=0;N>>0]=O[N];return C!==null&&C.push(xt,Z),Z},argPackAdvance:8,readValueFromPointer:Lr,bb(C){xt(C)}})},x:function(f,y,x){if(y>>>=0,x>>>=0,x=Bt(x),y===2)var C=be,O=_n,N=Xa,j=Z=>se[Z>>>1>>>0];else y===4&&(C=ws,O=Qa,N=Za,j=Z=>R[Z>>>2>>>0]);Tt(f>>>0,{name:x,fromWireType:Z=>{for(var ce=R[Z>>>2>>>0],oe,me=Z+4,ze=0;ze<=ce;++ze){var Fe=Z+4+ze*y;(ze==ce||j(Fe)==0)&&(me=C(me,Fe-me),oe===void 0?oe=me:(oe+="\0",oe+=me),me=Fe+y)}return xt(Z),oe},toWireType:(Z,ce)=>{if(typeof ce!="string")throw new pr(`Cannot pass non-string to C++ string type ${x}`);var oe=N(ce),me=Mr(4+oe+y);return R[me>>>2>>>0]=oe/y,O(ce,me+4,oe+y),Z!==null&&Z.push(xt,me),me},argPackAdvance:8,readValueFromPointer:Lr,bb(Z){xt(Z)}})},L:function(f,y){y=Bt(y>>>0),Tt(f>>>0,{ob:!0,name:y,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ha:()=>1,u:function(f,y,x){return y>>>=0,x>>>=0,f=Ie(f>>>0),y=yn(y,"emval::as"),jn(y,x,f)},w:function(f){return f>>>=0,Xn(()=>(f=Ie(f),f.then(At)))},n:function(f,y,x,C){return x>>>=0,C>>>=0,f=Wr[f>>>0],y=Ie(y>>>0),f(null,y,x,C)},j:function(f,y,x,C,O){return x>>>=0,C>>>=0,O>>>=0,f=Wr[f>>>0],y=Ie(y>>>0),x=bn(x),f(y,y[x],C,O)},b:Hn,A:function(f,y){return y>>>=0,f=Ie(f>>>0),y=Ie(y),f==y},m:function(f){return f>>>=0,f===0?At(Dt()):(f=bn(f),At(Dt()[f]))},i:function(f,y,x){y=Zn(f,y>>>0);var C=y.shift();f--;var O=`return function (obj, func, destructorsRef, args) { +`,N=0,j=[];x===0&&j.push("obj");for(var Z=["retType"],ce=[C],oe=0;oeme.name).join(", ")}) => ${C.name}>`,Qn(Jt(x,f))},r:function(f,y){return y>>>=0,f=Ie(f>>>0),y=Ie(y),At(f[y])},e:function(f){f>>>=0,9>>0);for(var y=Array(f.length),x=0;x>>0))},k:function(){return At({})},h:function(f){f>>>=0;for(var y=Ie(f);y.length;){var x=y.pop();y.pop()(x)}Hn(f)},g:function(f,y,x){y>>>=0,x>>>=0,f=Ie(f>>>0),y=Ie(y),x=Ie(x),f[y]=x},c:function(f,y){return y>>>=0,f=yn(f>>>0,"_emval_take_value"),f=f.readValueFromPointer(y),At(f)},W:function(f,y){f=-9007199254740992>f||9007199254740992>>=0,f=new Date(1e3*f),M[y>>>2>>>0]=f.getUTCSeconds(),M[y+4>>>2>>>0]=f.getUTCMinutes(),M[y+8>>>2>>>0]=f.getUTCHours(),M[y+12>>>2>>>0]=f.getUTCDate(),M[y+16>>>2>>>0]=f.getUTCMonth(),M[y+20>>>2>>>0]=f.getUTCFullYear()-1900,M[y+24>>>2>>>0]=f.getUTCDay(),M[y+28>>>2>>>0]=(f.getTime()-Date.UTC(f.getUTCFullYear(),0,1,0,0,0,0))/864e5|0},X:function(f,y){f=-9007199254740992>f||9007199254740992>>=0,f=new Date(1e3*f),M[y>>>2>>>0]=f.getSeconds(),M[y+4>>>2>>>0]=f.getMinutes(),M[y+8>>>2>>>0]=f.getHours(),M[y+12>>>2>>>0]=f.getDate(),M[y+16>>>2>>>0]=f.getMonth(),M[y+20>>>2>>>0]=f.getFullYear()-1900,M[y+24>>>2>>>0]=f.getDay(),M[y+28>>>2>>>0]=(_t(f.getFullYear())?Jn:vn)[f.getMonth()]+f.getDate()-1|0,M[y+36>>>2>>>0]=-(60*f.getTimezoneOffset());var x=new Date(f.getFullYear(),6,1).getTimezoneOffset(),C=new Date(f.getFullYear(),0,1).getTimezoneOffset();M[y+32>>>2>>>0]=(x!=C&&f.getTimezoneOffset()==Math.min(C,x))|0},Y:function(f){f>>>=0;var y=new Date(M[f+20>>>2>>>0]+1900,M[f+16>>>2>>>0],M[f+12>>>2>>>0],M[f+8>>>2>>>0],M[f+4>>>2>>>0],M[f>>>2>>>0],0),x=M[f+32>>>2>>>0],C=y.getTimezoneOffset(),O=new Date(y.getFullYear(),6,1).getTimezoneOffset(),N=new Date(y.getFullYear(),0,1).getTimezoneOffset(),j=Math.min(N,O);return 0>x?M[f+32>>>2>>>0]=+(O!=N&&j==C):0>>2>>>0]=y.getDay(),M[f+28>>>2>>>0]=(_t(y.getFullYear())?Jn:vn)[y.getMonth()]+y.getDate()-1|0,M[f>>>2>>>0]=y.getSeconds(),M[f+4>>>2>>>0]=y.getMinutes(),M[f+8>>>2>>>0]=y.getHours(),M[f+12>>>2>>>0]=y.getDate(),M[f+16>>>2>>>0]=y.getMonth(),M[f+20>>>2>>>0]=y.getYear(),f=y.getTime(),BigInt(isNaN(f)?-1:f/1e3)},U:function(){return-52},V:function(){},N:function(f,y,x,C){x>>>=0,C>>>=0;var O=new Date().getFullYear(),N=new Date(O,0,1),j=new Date(O,6,1);O=N.getTimezoneOffset();var Z=j.getTimezoneOffset();R[f>>>0>>>2>>>0]=60*Math.max(O,Z),M[y>>>0>>>2>>>0]=+(O!=Z),f=ce=>ce.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],N=f(N),j=f(j),Z{Qe("")},d:function(f,y,x){return f>>>=0,y=$n(y>>>0,x>>>0),wr[f](...y)},I:function(f,y,x){return f>>>=0,y=$n(y>>>0,x>>>0),wr[f](...y)},H:()=>Date.now(),O:function(){return 4294901760},q:()=>performance.now(),M:function(f){f>>>=0;var y=te.length;if(4294901760=x;x*=2){var C=y*(1+.2/x);C=Math.min(C,f+100663296);var O=Math;C=Math.max(f,C);e:{O=(O.min.call(O,4294901760,C+(65536-C%65536)%65536)-D.buffer.byteLength+65535)/65536;try{D.grow(O),W();var N=1;break e}catch{}N=void 0}if(N)return!0}return!1},$:function(f,y){f>>>=0,y>>>=0;var x=0;return Mt().forEach((C,O)=>{var N=y+x;for(O=R[f+4*O>>>2>>>0]=N,N=0;N>>0]=C.charCodeAt(N);ie[O>>>0]=0,x+=C.length+1}),0},aa:function(f,y){f>>>=0,y>>>=0;var x=Mt();R[f>>>2>>>0]=x.length;var C=0;return x.forEach(O=>C+=O.length+1),R[y>>>2>>>0]=C,0},z:()=>52,F:function(){return 52},_:function(){return 70},E:function(f,y,x,C){y>>>=0,x>>>=0,C>>>=0;for(var O=0,N=0;N>>2>>>0],Z=R[y+4>>>2>>>0];y+=8;for(var ce=0;ce>>0],me=vs[f];oe===0||oe===10?((f===1?P:B)(Wn(me,0)),me.length=0):me.push(oe)}O+=Z}return R[C>>>2>>>0]=O,0},fa:oi,p:function(f,y,x,C){return oi(f>>>0,y>>>0,x>>>0,C>>>0)}},we=function(){function f(x){return we=x.exports,we=Ja(),we=di(),D=we.ja,W(),fe.unshift(we.ka),Be--,Be==0&&Ye&&(x=Ye,Ye=null,x()),we}var y={a:xs};if(Be++,i.instantiateWasm)try{return i.instantiateWasm(y,f)}catch(x){B(`Module.instantiateWasm callback failed with error: ${x}`),o(x)}return Fr(y,function(x){f(x.instance)}).catch(o),{}}(),ta=f=>(ta=we.la)(f);i._OrtInit=(f,y)=>(i._OrtInit=we.ma)(f,y),i._OrtGetLastError=(f,y)=>(i._OrtGetLastError=we.na)(f,y),i._OrtCreateSessionOptions=(f,y,x,C,O,N,j,Z,ce,oe)=>(i._OrtCreateSessionOptions=we.oa)(f,y,x,C,O,N,j,Z,ce,oe),i._OrtAppendExecutionProvider=(f,y)=>(i._OrtAppendExecutionProvider=we.pa)(f,y),i._OrtAddFreeDimensionOverride=(f,y,x)=>(i._OrtAddFreeDimensionOverride=we.qa)(f,y,x),i._OrtAddSessionConfigEntry=(f,y,x)=>(i._OrtAddSessionConfigEntry=we.ra)(f,y,x),i._OrtReleaseSessionOptions=f=>(i._OrtReleaseSessionOptions=we.sa)(f),i._OrtCreateSession=(f,y,x)=>(i._OrtCreateSession=we.ta)(f,y,x),i._OrtReleaseSession=f=>(i._OrtReleaseSession=we.ua)(f),i._OrtGetInputOutputCount=(f,y,x)=>(i._OrtGetInputOutputCount=we.va)(f,y,x),i._OrtGetInputName=(f,y)=>(i._OrtGetInputName=we.wa)(f,y),i._OrtGetOutputName=(f,y)=>(i._OrtGetOutputName=we.xa)(f,y),i._OrtFree=f=>(i._OrtFree=we.ya)(f),i._OrtCreateTensor=(f,y,x,C,O,N)=>(i._OrtCreateTensor=we.za)(f,y,x,C,O,N),i._OrtGetTensorData=(f,y,x,C,O)=>(i._OrtGetTensorData=we.Aa)(f,y,x,C,O),i._OrtReleaseTensor=f=>(i._OrtReleaseTensor=we.Ba)(f),i._OrtCreateRunOptions=(f,y,x,C)=>(i._OrtCreateRunOptions=we.Ca)(f,y,x,C),i._OrtAddRunConfigEntry=(f,y,x)=>(i._OrtAddRunConfigEntry=we.Da)(f,y,x),i._OrtReleaseRunOptions=f=>(i._OrtReleaseRunOptions=we.Ea)(f),i._OrtCreateBinding=f=>(i._OrtCreateBinding=we.Fa)(f),i._OrtBindInput=(f,y,x)=>(i._OrtBindInput=we.Ga)(f,y,x),i._OrtBindOutput=(f,y,x,C)=>(i._OrtBindOutput=we.Ha)(f,y,x,C),i._OrtClearBoundOutputs=f=>(i._OrtClearBoundOutputs=we.Ia)(f),i._OrtReleaseBinding=f=>(i._OrtReleaseBinding=we.Ja)(f),i._OrtRunWithBinding=(f,y,x,C,O)=>(i._OrtRunWithBinding=we.Ka)(f,y,x,C,O),i._OrtRun=(f,y,x,C,O,N,j,Z)=>(i._OrtRun=we.La)(f,y,x,C,O,N,j,Z),i._OrtEndProfiling=f=>(i._OrtEndProfiling=we.Ma)(f),i._JsepOutput=(f,y,x)=>(i._JsepOutput=we.Na)(f,y,x),i._JsepGetNodeName=f=>(i._JsepGetNodeName=we.Oa)(f);var Mr=i._malloc=f=>(Mr=i._malloc=we.Pa)(f),xt=i._free=f=>(xt=i._free=we.Qa)(f),Sn=f=>(Sn=we.Sa)(f),kn=f=>(kn=we.Ta)(f),ra=()=>(ra=we.Ua)(),Gr=f=>(Gr=we.Wa)(f),ui=()=>(ui=we.Xa)(),er=f=>(er=we.Ya)(f),Ft=()=>(Ft=we.Za)();i.___start_em_js=837952,i.___stop_em_js=838113;function di(){var f=we;f=Object.assign({},f);var y=x=>C=>x(C)>>>0;return f.la=y(f.la),f.Pa=y(f.Pa),f.Ta=y(f.Ta),f.Ua=(x=>()=>x()>>>0)(f.Ua),f}i.stackSave=()=>ra(),i.stackRestore=f=>Sn(f),i.stackAlloc=f=>kn(f),i.UTF8ToString=Xe,i.stringToUTF8=(f,y,x)=>Wt(f,te,y,x),i.lengthBytesUTF8=hn;var vr;Ye=function f(){vr||na(),vr||(Ye=f)};function na(){if(!(0r)}),py=Bn(()=>{}),hy=Bn(()=>{}),ch={};Dn(ch,{cpus:()=>ph});var ph,fy=J(()=>{ph=void 0}),my=Bn((t,e)=>{var r=(()=>{var n=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(n||=__filename),function(a={}){function i(){return ne.buffer!=Ce.buffer&<(),Ce}function s(){return ne.buffer!=Ce.buffer&<(),Be}function o(){return ne.buffer!=Ce.buffer&<(),Ye}function l(){return ne.buffer!=Ce.buffer&<(),Qe}function u(){return ne.buffer!=Ce.buffer&<(),We}function p(){return ne.buffer!=Ce.buffer&<(),pt}function h(){return ne.buffer!=Ce.buffer&<(),Ve}function m(){return ne.buffer!=Ce.buffer&<(),Xt}var d=a,_,w,v=new Promise((c,g)=>{_=c,w=g});d.mountExternalData=(c,g)=>{(d.Db||(d.Db=new Map)).set(c,g)},d.unmountExternalData=()=>{delete d.Db};let S=()=>{let c=(b,k,I)=>(...z)=>{let V=Ft,Y=k?.();z=b(...z);let pe=k?.();return Y!==pe&&(b=pe,I(Y),k=I=null),Ft!=V?O():z},g=b=>async(...k)=>{try{if(d.Cb)throw Error("Session already started");let I=d.Cb={ec:k[0],errors:[]},z=await b(...k);if(d.Cb!==I)throw Error("Session mismatch");d.Kb?.flush();let V=I.errors;if(0pe),0d._OrtCreateSession,b=>d._OrtCreateSession=b),d._OrtRun=g(c(d._OrtRun,()=>d._OrtRun,b=>d._OrtRun=b)),d._OrtRunWithBinding=g(c(d._OrtRunWithBinding,()=>d._OrtRunWithBinding,b=>d._OrtRunWithBinding=b)),d._OrtBindInput=c(d._OrtBindInput,()=>d._OrtBindInput,b=>d._OrtBindInput=b),S=void 0};d.jsepInit=(c,g)=>{if(S?.(),c==="webgpu"){[d.Kb,d.Wb,d.$b,d.Lb,d.Zb,d.ob,d.ac,d.cc,d.Xb,d.Yb,d.bc]=g;let b=d.Kb;d.jsepRegisterBuffer=(k,I,z,V)=>b.registerBuffer(k,I,z,V),d.jsepGetBuffer=k=>b.getBuffer(k),d.jsepCreateDownloader=(k,I,z)=>b.createDownloader(k,I,z),d.jsepOnReleaseSession=k=>{b.onReleaseSession(k)},d.jsepOnRunStart=k=>b.onRunStart(k)}};var $=Object.assign({},d),E="./this.program",T=(c,g)=>{throw g},A=typeof window=="object",P=typeof importScripts=="function",B=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",D=d.ENVIRONMENT_IS_PTHREAD||!1,q="";function H(c){return d.locateFile?d.locateFile(c,q):q+c}var ie,te,de;if(B){var se=(lh(),nn(Ho)),M=(dh(),nn(jo));q=P?M.dirname(q)+"/":__dirname+"/",ie=(c,g)=>(c=Wt(c)?new URL(c):M.normalize(c),se.readFileSync(c,g?void 0:"utf8")),de=c=>(c=ie(c,!0),c.buffer||(c=new Uint8Array(c)),c),te=(c,g,b,k=!0)=>{c=Wt(c)?new URL(c):M.normalize(c),se.readFile(c,k?void 0:"utf8",(I,z)=>{I?b(I):g(k?z.buffer:z)})},!d.thisProgram&&1{throw process.exitCode=c,g},global.Worker=py().Worker}else(A||P)&&(P?q=self.location.href:typeof document<"u"&&document.currentScript&&(q=document.currentScript.src),typeof n<"u"&&n&&(q=n),q.startsWith("blob:")?q="":q=q.substr(0,q.replace(/[?#].*/,"").lastIndexOf("/")+1),B||(ie=c=>{var g=new XMLHttpRequest;return g.open("GET",c,!1),g.send(null),g.responseText},P&&(de=c=>{var g=new XMLHttpRequest;return g.open("GET",c,!1),g.responseType="arraybuffer",g.send(null),new Uint8Array(g.response)}),te=(c,g,b)=>{var k=new XMLHttpRequest;k.open("GET",c,!0),k.responseType="arraybuffer",k.onload=()=>{k.status==200||k.status==0&&k.response?g(k.response):b()},k.onerror=b,k.send(null)}));B&&typeof performance>"u"&&(global.performance=hy().performance);var R=console.log.bind(console),G=console.error.bind(console);B&&(R=(...c)=>se.writeSync(1,c.join(" ")+` +`),G=(...c)=>se.writeSync(2,c.join(" ")+` +`));var re=R,ee=G;Object.assign(d,$),$=null;var ne,W,ae=!1,fe,Ce,Be,Ye,Qe,We,pt,Ve,Ne,gt,Xt;function lt(){var c=ne.buffer;d.HEAP8=Ce=new Int8Array(c),d.HEAP16=Ye=new Int16Array(c),d.HEAPU8=Be=new Uint8Array(c),d.HEAPU16=Qe=new Uint16Array(c),d.HEAP32=We=new Int32Array(c),d.HEAPU32=pt=new Uint32Array(c),d.HEAPF32=Ve=new Float32Array(c),d.HEAPF64=Xt=new Float64Array(c),d.HEAP64=Ne=new BigInt64Array(c),d.HEAPU64=gt=new BigUint64Array(c)}var Fr=16777216;if(D)ne=d.wasmMemory;else if(d.wasmMemory)ne=d.wasmMemory;else if(ne=new WebAssembly.Memory({initial:Fr/65536,maximum:65536,shared:!0}),!(ne.buffer instanceof SharedArrayBuffer))throw ee("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),B&&ee("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");lt(),Fr=ne.buffer.byteLength;var wr=[],cr=[],_s=[],br=0,Ar=null;function Wn(){if(br--,br==0&&Ar){var c=Ar;Ar=null,c()}}function Xe(c){throw c="Aborted("+c+")",ee(c),ae=!0,fe=1,c=new WebAssembly.RuntimeError(c+". Build with -sASSERTIONS for more info."),w(c),c}var hn=c=>c.startsWith("data:application/octet-stream;base64,"),Wt=c=>c.startsWith("file://"),Ir;Ir="ort-wasm-simd-threaded.wasm",hn(Ir)||(Ir=H(Ir));function Bt(c){if(de)return de(c);throw"both async and sync fetching of the wasm failed"}function Vn(c){if(A||P){if(typeof fetch=="function"&&!Wt(c))return fetch(c,{credentials:"same-origin"}).then(g=>{if(!g.ok)throw`failed to load wasm binary file at '${c}'`;return g.arrayBuffer()}).catch(()=>Bt(c));if(te)return new Promise((g,b)=>{te(c,k=>g(new Uint8Array(k)),b)})}return Promise.resolve().then(()=>Bt(c))}function fn(c,g,b){return Vn(c).then(k=>WebAssembly.instantiate(k,g)).then(b,k=>{ee(`failed to asynchronously prepare wasm: ${k}`),Xe(k)})}function pr(c,g){var b=Ir;return typeof WebAssembly.instantiateStreaming!="function"||hn(b)||Wt(b)||B||typeof fetch!="function"?fn(b,c,g):fetch(b,{credentials:"same-origin"}).then(k=>WebAssembly.instantiateStreaming(k,c).then(g,function(I){return ee(`wasm streaming compile failed: ${I}`),ee("falling back to ArrayBuffer instantiation"),fn(b,c,g)}))}var Gn={826468:(c,g,b,k)=>{if(typeof d>"u"||!d.Db)return 1;if(c=it(c>>>0),c.startsWith("./")&&(c=c.substring(2)),c=d.Db.get(c),!c)return 2;if(g>>>=0,b>>>=0,k>>>=0,g+b>c.byteLength)return 3;try{return s().set(c.subarray(g,g+b),k>>>0),0}catch{return 4}},826969:()=>{d.Xb()},827e3:()=>{d.Yb()},827029:()=>{d.bc()},827054:c=>d.Wb(c),827087:c=>d.$b(c),827119:(c,g,b)=>{d.Lb(c,g,b,!0)},827158:(c,g,b)=>{d.Lb(c,g,b)},827191:c=>{d.ob("Abs",c,void 0)},827242:c=>{d.ob("Neg",c,void 0)},827293:c=>{d.ob("Floor",c,void 0)},827346:c=>{d.ob("Ceil",c,void 0)},827398:c=>{d.ob("Reciprocal",c,void 0)},827456:c=>{d.ob("Sqrt",c,void 0)},827508:c=>{d.ob("Exp",c,void 0)},827559:c=>{d.ob("Erf",c,void 0)},827610:c=>{d.ob("Sigmoid",c,void 0)},827665:(c,g,b)=>{d.ob("HardSigmoid",c,{alpha:g,beta:b})},827744:c=>{d.ob("Log",c,void 0)},827795:c=>{d.ob("Sin",c,void 0)},827846:c=>{d.ob("Cos",c,void 0)},827897:c=>{d.ob("Tan",c,void 0)},827948:c=>{d.ob("Asin",c,void 0)},828e3:c=>{d.ob("Acos",c,void 0)},828052:c=>{d.ob("Atan",c,void 0)},828104:c=>{d.ob("Sinh",c,void 0)},828156:c=>{d.ob("Cosh",c,void 0)},828208:c=>{d.ob("Asinh",c,void 0)},828261:c=>{d.ob("Acosh",c,void 0)},828314:c=>{d.ob("Atanh",c,void 0)},828367:c=>{d.ob("Tanh",c,void 0)},828419:c=>{d.ob("Not",c,void 0)},828470:(c,g,b)=>{d.ob("Clip",c,{min:g,max:b})},828539:c=>{d.ob("Clip",c,void 0)},828591:(c,g)=>{d.ob("Elu",c,{alpha:g})},828649:c=>{d.ob("Relu",c,void 0)},828701:(c,g)=>{d.ob("LeakyRelu",c,{alpha:g})},828765:(c,g)=>{d.ob("ThresholdedRelu",c,{alpha:g})},828835:(c,g)=>{d.ob("Cast",c,{to:g})},828893:c=>{d.ob("Add",c,void 0)},828944:c=>{d.ob("Sub",c,void 0)},828995:c=>{d.ob("Mul",c,void 0)},829046:c=>{d.ob("Div",c,void 0)},829097:c=>{d.ob("Pow",c,void 0)},829148:c=>{d.ob("Equal",c,void 0)},829201:c=>{d.ob("Greater",c,void 0)},829256:c=>{d.ob("GreaterOrEqual",c,void 0)},829318:c=>{d.ob("Less",c,void 0)},829370:c=>{d.ob("LessOrEqual",c,void 0)},829429:(c,g,b,k,I)=>{d.ob("ReduceMean",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},829588:(c,g,b,k,I)=>{d.ob("ReduceMax",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},829746:(c,g,b,k,I)=>{d.ob("ReduceMin",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},829904:(c,g,b,k,I)=>{d.ob("ReduceProd",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},830063:(c,g,b,k,I)=>{d.ob("ReduceSum",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},830221:(c,g,b,k,I)=>{d.ob("ReduceL1",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},830378:(c,g,b,k,I)=>{d.ob("ReduceL2",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},830535:(c,g,b,k,I)=>{d.ob("ReduceLogSum",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},830696:(c,g,b,k,I)=>{d.ob("ReduceSumSquare",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},830860:(c,g,b,k,I)=>{d.ob("ReduceLogSumExp",c,{keepDims:!!g,noopWithEmptyAxes:!!b,axes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},831024:c=>{d.ob("Where",c,void 0)},831077:(c,g,b)=>{d.ob("Transpose",c,{perm:g?Array.from(u().subarray(g>>>0,b>>>0)):[]})},831185:(c,g,b,k)=>{d.ob("DepthToSpace",c,{blocksize:g,mode:it(b),format:k?"NHWC":"NCHW"})},831318:(c,g,b,k)=>{d.ob("DepthToSpace",c,{blocksize:g,mode:it(b),format:k?"NHWC":"NCHW"})},831451:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L,ke)=>{d.ob("ConvTranspose",c,{format:pe?"NHWC":"NCHW",autoPad:g,dilations:[b],group:k,kernelShape:[I],pads:[z,V],strides:[Y],wIsConst:()=>!!i()[le>>>0],outputPadding:_e?Array.from(u().subarray(_e>>>0,Le>>>0)):[],outputShape:Ge?Array.from(u().subarray(Ge>>>0,L>>>0)):[],activation:it(ke)})},831852:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L)=>{d.ob("ConvTranspose",c,{format:Y?"NHWC":"NCHW",autoPad:g,dilations:Array.from(u().subarray(b>>>0,(b>>>0)+2>>>0)),group:k,kernelShape:Array.from(u().subarray(I>>>0,(I>>>0)+2>>>0)),pads:Array.from(u().subarray(z>>>0,(z>>>0)+4>>>0)),strides:Array.from(u().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!i()[pe>>>0],outputPadding:le?Array.from(u().subarray(le>>>0,_e>>>0)):[],outputShape:Le?Array.from(u().subarray(Le>>>0,Ge>>>0)):[],activation:it(L)})},832417:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L,ke)=>{d.ob("ConvTranspose",c,{format:pe?"NHWC":"NCHW",autoPad:g,dilations:[b],group:k,kernelShape:[I],pads:[z,V],strides:[Y],wIsConst:()=>!!i()[le>>>0],outputPadding:_e?Array.from(u().subarray(_e>>>0,Le>>>0)):[],outputShape:Ge?Array.from(u().subarray(Ge>>>0,L>>>0)):[],activation:it(ke)})},832818:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L)=>{d.ob("ConvTranspose",c,{format:Y?"NHWC":"NCHW",autoPad:g,dilations:Array.from(u().subarray(b>>>0,(b>>>0)+2>>>0)),group:k,kernelShape:Array.from(u().subarray(I>>>0,(I>>>0)+2>>>0)),pads:Array.from(u().subarray(z>>>0,(z>>>0)+4>>>0)),strides:Array.from(u().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!i()[pe>>>0],outputPadding:le?Array.from(u().subarray(le>>>0,_e>>>0)):[],outputShape:Le?Array.from(u().subarray(Le>>>0,Ge>>>0)):[],activation:it(L)})},833383:(c,g)=>{d.ob("GlobalAveragePool",c,{format:g?"NHWC":"NCHW"})},833474:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L,ke,De)=>{d.ob("AveragePool",c,{format:De?"NHWC":"NCHW",auto_pad:g,ceil_mode:b,count_include_pad:k,storage_order:I,dilations:[z,V],kernel_shape:[Y,pe],pads:[le,_e,Le,Ge],strides:[L,ke]})},833758:(c,g)=>{d.ob("GlobalAveragePool",c,{format:g?"NHWC":"NCHW"})},833849:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L,ke,De)=>{d.ob("AveragePool",c,{format:De?"NHWC":"NCHW",auto_pad:g,ceil_mode:b,count_include_pad:k,storage_order:I,dilations:[z,V],kernel_shape:[Y,pe],pads:[le,_e,Le,Ge],strides:[L,ke]})},834133:(c,g)=>{d.ob("GlobalMaxPool",c,{format:g?"NHWC":"NCHW"})},834220:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L,ke,De)=>{d.ob("MaxPool",c,{format:De?"NHWC":"NCHW",auto_pad:g,ceil_mode:b,count_include_pad:k,storage_order:I,dilations:[z,V],kernel_shape:[Y,pe],pads:[le,_e,Le,Ge],strides:[L,ke]})},834500:(c,g)=>{d.ob("GlobalMaxPool",c,{format:g?"NHWC":"NCHW"})},834587:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L,ke,De)=>{d.ob("MaxPool",c,{format:De?"NHWC":"NCHW",auto_pad:g,ceil_mode:b,count_include_pad:k,storage_order:I,dilations:[z,V],kernel_shape:[Y,pe],pads:[le,_e,Le,Ge],strides:[L,ke]})},834867:(c,g,b,k,I)=>{d.ob("Gemm",c,{alpha:g,beta:b,transA:k,transB:I})},834971:c=>{d.ob("MatMul",c,void 0)},835025:(c,g,b,k)=>{d.ob("ArgMax",c,{keepDims:!!g,selectLastIndex:!!b,axis:k})},835133:(c,g,b,k)=>{d.ob("ArgMin",c,{keepDims:!!g,selectLastIndex:!!b,axis:k})},835241:(c,g)=>{d.ob("Softmax",c,{axis:g})},835304:(c,g)=>{d.ob("Concat",c,{axis:g})},835364:(c,g,b,k,I)=>{d.ob("Split",c,{axis:g,numOutputs:b,splitSizes:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},835504:c=>{d.ob("Expand",c,void 0)},835558:(c,g)=>{d.ob("Gather",c,{axis:Number(g)})},835629:(c,g)=>{d.ob("GatherElements",c,{axis:Number(g)})},835708:(c,g,b,k,I,z,V,Y,pe,le,_e)=>{d.ob("Resize",c,{antialias:g,axes:b?Array.from(u().subarray(b>>>0,k>>>0)):[],coordinateTransformMode:it(I),cubicCoeffA:z,excludeOutside:V,extrapolationValue:Y,keepAspectRatioPolicy:it(pe),mode:it(le),nearestMode:it(_e)})},836054:(c,g,b,k,I,z,V)=>{d.ob("Slice",c,{starts:g?Array.from(u().subarray(g>>>0,b>>>0)):[],ends:k?Array.from(u().subarray(k>>>0,I>>>0)):[],axes:z?Array.from(u().subarray(z>>>0,V>>>0)):[]})},836270:c=>{d.ob("Tile",c,void 0)},836322:(c,g,b,k)=>{d.ob("LayerNormalization",c,{axis:g,epsilon:b,simplified:!!k})},836433:(c,g,b)=>{d.ob("InstanceNormalization",c,{epsilon:g,format:b?"NHWC":"NCHW"})},836547:(c,g,b)=>{d.ob("InstanceNormalization",c,{epsilon:g,format:b?"NHWC":"NCHW"})},836661:c=>{d.ob("Range",c,void 0)},836714:(c,g)=>{d.ob("Einsum",c,{equation:it(g)})},836795:(c,g,b,k,I)=>{d.ob("Pad",c,{mode:g,value:b,pads:k?Array.from(u().subarray(k>>>0,I>>>0)):[]})},836922:(c,g,b,k,I,z)=>{d.ob("BatchNormalization",c,{epsilon:g,momentum:b,spatial:!!I,trainingMode:!!k,format:z?"NHWC":"NCHW"})},837091:(c,g,b,k,I,z)=>{d.ob("BatchNormalization",c,{epsilon:g,momentum:b,spatial:!!I,trainingMode:!!k,format:z?"NHWC":"NCHW"})},837260:(c,g,b)=>{d.ob("CumSum",c,{exclusive:Number(g),reverse:Number(b)})},837357:(c,g,b,k,I,z,V,Y,pe)=>{d.ob("Attention",c,{numHeads:g,isUnidirectional:b,maskFilterValue:k,scale:I,doRotary:z,qkvHiddenSizes:V?Array.from(u().subarray(Number(Y)>>>0,Number(Y)+V>>>0)):[],pastPresentShareBuffer:!!pe})},837629:c=>{d.ob("BiasAdd",c,void 0)},837684:c=>{d.ob("BiasSplitGelu",c,void 0)},837745:c=>{d.ob("FastGelu",c,void 0)},837801:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge)=>{d.ob("Conv",c,{format:pe?"NHWC":"NCHW",auto_pad:g,dilations:[b],group:k,kernel_shape:[I],pads:z?Array.from(u().subarray(z>>>0,V>>>0)):[],strides:[Y],w_is_const:()=>!!i()[le>>>0],activation:it(_e),activation_params:Le?Array.from(h().subarray(Le>>>0,Ge>>>0)):[]})},838171:(c,g,b,k,I,z,V,Y,pe,le,_e,Le,Ge,L,ke,De)=>{d.ob("Conv",c,{format:Le?"NHWC":"NCHW",auto_pad:g,dilations:[b,k],group:I,kernel_shape:[z,V],pads:Y?Array.from(u().subarray(Y>>>0,pe>>>0)):[],strides:[le,_e],w_is_const:()=>!!i()[Ge>>>0],activation:it(L),activation_params:ke?Array.from(h().subarray(ke>>>0,De>>>0)):[]})},838562:c=>{d.ob("Gelu",c,void 0)},838614:(c,g,b,k,I,z)=>{d.ob("MatMulNBits",c,{k:g,n:b,accuracyLevel:k,bits:I,blockSize:z})},838741:(c,g,b,k,I,z)=>{d.ob("MultiHeadAttention",c,{numHeads:g,isUnidirectional:b,maskFilterValue:k,scale:I,doRotary:z})},838900:(c,g,b,k,I)=>{d.ob("RotaryEmbedding",c,{interleaved:!!g,numHeads:b,rotaryEmbeddingDim:k,scale:I})},839039:(c,g,b)=>{d.ob("SkipLayerNormalization",c,{epsilon:g,simplified:!!b})},839141:(c,g,b)=>{d.ob("SkipLayerNormalization",c,{epsilon:g,simplified:!!b})},839243:(c,g,b,k)=>{d.ob("LayerNormalization",c,{axis:g,epsilon:b,simplified:!!k})},839354:c=>{d.ac(c)},839388:(c,g)=>d.cc(c,g,d.Cb.ec,d.Cb.errors)};function Tt(c){this.name="ExitStatus",this.message=`Program terminated with exit(${c})`,this.status=c}var mn=c=>{c.terminate(),c.onmessage=()=>{}},gn=c=>{be.xb.length==0&&(Ka(),be.Mb(be.xb[0]));var g=be.xb.pop();if(!g)return 6;be.yb.push(g),be.ub[c.wb]=g,g.wb=c.wb;var b={cmd:"run",start_routine:c.fc,arg:c.Rb,pthread_ptr:c.wb};return B&&g.unref(),g.postMessage(b,c.mc),0},ut=0,Hn=c=>{var g=Is();return c=c(),fi(g),c},Ie=(c,g,...b)=>Hn(()=>{for(var k=2*b.length,I=As(8*k),z=I>>>3,V=0;V>>0]=Y)}return jl(c,0,k,I,g)});function At(c){if(D)return Ie(0,1,c);fe=c,0{if(fe=c,D)throw Xa(c),"unwind";At(c)};function ys(){for(var c=d.numThreads;c--;)Ka();wr.unshift(()=>{br++,Ya(()=>Wn())})}function Ka(){var c=H("ort-wasm-simd-threaded.worker.js");c=new Worker(c),be.xb.push(c)}function Ya(c){D?c():Promise.all(be.xb.map(be.Mb)).then(c)}var be={xb:[],yb:[],Qb:[],ub:{},Gb(){D?(be.receiveObjectTransfer=be.dc,be.threadInitTLS=be.Pb,be.setExitStatus=be.Ob):ys()},Ob:c=>fe=c,pc:["$terminateWorker"],hc:()=>{for(var c of be.yb)mn(c);for(c of be.xb)mn(c);be.xb=[],be.yb=[],be.ub=[]},Nb:c=>{var g=c.wb;delete be.ub[g],be.xb.push(c),be.yb.splice(be.yb.indexOf(c),1),c.wb=0,Cs(g)},dc(){},Pb(){be.Qb.forEach(c=>c())},Mb:c=>new Promise(g=>{c.onmessage=z=>{z=z.data;var V=z.cmd;if(z.targetThread&&z.targetThread!=pi()){var Y=be.ub[z.targetThread];Y?Y.postMessage(z,z.transferList):ee(`Internal error! Worker sent a message "${V}" to target pthread ${z.targetThread}, but that thread no longer exists!`)}else V==="checkMailbox"?xt():V==="spawnThread"?gn(z):V==="cleanupThread"?be.Nb(be.ub[z.thread]):V==="killThread"?(z=z.thread,V=be.ub[z],delete be.ub[z],mn(V),Cs(z),be.yb.splice(be.yb.indexOf(V),1),V.wb=0):V==="cancelThread"?be.ub[z.thread].postMessage({cmd:"cancel"}):V==="loaded"?(c.loaded=!0,B&&!c.wb&&c.unref(),g(c)):V==="alert"?alert(`Thread ${z.threadId}: ${z.text}`):z.target==="setimmediate"?c.postMessage(z):V==="callHandler"?d[z.handler](...z.args):V&&ee(`worker sent an unknown command ${V}`)},c.onerror=z=>{throw ee(`worker sent an error! ${z.filename}:${z.lineno}: ${z.message}`),z},B&&(c.on("message",z=>c.onmessage({data:z})),c.on("error",z=>c.onerror(z)));var b=[],k=["onExit"],I;for(I of k)d.hasOwnProperty(I)&&b.push(I);c.postMessage({cmd:"load",handlers:b,urlOrBlob:d.mainScriptUrlOrBlob||n,wasmMemory:ne,wasmModule:W})})};d.PThread=be;var _n=c=>{for(;0{var c=pi(),g=p()[c+52>>>2>>>0];c=p()[c+56>>>2>>>0],Kl(g,g-c),fi(g)};function Xa(c){if(D)return Ie(1,0,c);Lr(c)}d.invokeEntryPoint=(c,g)=>{ut=0,c=Yl(c,g),0>>2>>>0]=0,p()[this.Ib+4>>>2>>>0]=g,p()[this.Ib+8>>>2>>>0]=b}}var Qa=0;function Za(c,g,b,k){return D?Ie(2,1,c,g,b,k):yn(c,g,b,k)}function yn(c,g,b,k){if(c>>>=0,g>>>=0,b>>>=0,k>>>=0,typeof SharedArrayBuffer>"u")return ee("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var I=[];return D&&I.length===0?Za(c,g,b,k):(c={fc:b,wb:c,Rb:k,mc:I},D?(c.oc="spawnThread",postMessage(c,I),0):gn(c))}var jn=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Ur=(c,g,b)=>{g>>>=0;var k=g+b;for(b=g;c[b]&&!(b>=k);)++b;if(16I?k+=String.fromCharCode(I):(I-=65536,k+=String.fromCharCode(55296|I>>10,56320|I&1023))}}else k+=String.fromCharCode(I)}return k},it=(c,g)=>(c>>>=0)?Ur(s(),c,g):"";function Ja(c,g,b){return D?Ie(3,1,c,g,b):0}function Qt(c,g){if(D)return Ie(4,1,c,g)}var It=c=>{for(var g=0,b=0;b=k?g++:2047>=k?g+=2:55296<=k&&57343>=k?(g+=4,++b):g+=3}return g},qn=(c,g,b,k)=>{if(b>>>=0,!(0=V){var Y=c.charCodeAt(++z);V=65536+((V&1023)<<10)|Y&1023}if(127>=V){if(b>=k)break;g[b++>>>0]=V}else{if(2047>=V){if(b+1>=k)break;g[b++>>>0]=192|V>>6}else{if(65535>=V){if(b+2>=k)break;g[b++>>>0]=224|V>>12}else{if(b+3>=k)break;g[b++>>>0]=240|V>>18,g[b++>>>0]=128|V>>12&63}g[b++>>>0]=128|V>>6&63}g[b++>>>0]=128|V&63}}return g[b>>>0]=0,b-I},Zt=(c,g,b)=>qn(c,s(),g,b);function Kn(c,g){if(D)return Ie(5,1,c,g)}function Yn(c,g,b){if(D)return Ie(6,1,c,g,b)}function ei(c,g,b){return D?Ie(7,1,c,g,b):0}function wn(c,g){if(D)return Ie(8,1,c,g)}function ti(c,g,b){if(D)return Ie(9,1,c,g,b)}function ri(c,g,b,k){if(D)return Ie(10,1,c,g,b,k)}function ni(c,g,b,k){if(D)return Ie(11,1,c,g,b,k)}function ai(c,g,b,k){if(D)return Ie(12,1,c,g,b,k)}function Xn(c){if(D)return Ie(13,1,c)}function Wr(c,g){if(D)return Ie(14,1,c,g)}function ii(c,g,b){if(D)return Ie(15,1,c,g,b)}var bn,Dt=c=>{for(var g="";s()[c>>>0];)g+=bn[s()[c++>>>0]];return g},Qn={},Zn={},Jt;function bs(c,g,b={}){var k=g.name;if(!c)throw new Jt(`type "${k}" must have a positive integer typeid pointer`);if(Zn.hasOwnProperty(c)){if(b.Ub)return;throw new Jt(`Cannot register type '${k}' twice`)}Zn[c]=g,Qn.hasOwnProperty(c)&&(g=Qn[c],delete Qn[c],g.forEach(I=>I()))}function _t(c,g,b={}){if(!("argPackAdvance"in g))throw new TypeError("registerType registeredInstance requires argPackAdvance");return bs(c,g,b)}var Jn=(c,g,b)=>{switch(g){case 1:return b?k=>i()[k>>>0]:k=>s()[k>>>0];case 2:return b?k=>o()[k>>>1>>>0]:k=>l()[k>>>1>>>0];case 4:return b?k=>u()[k>>>2>>>0]:k=>p()[k>>>2>>>0];case 8:return b?k=>Ne[k>>>3]:k=>gt[k>>>3];default:throw new TypeError(`invalid integer width (${g}): ${c}`)}},vn=[],Nt=[];function $n(c){c>>>=0,9{if(!c)throw new Jt("Cannot use deleted val. handle = "+c);return Nt[c]},Mt=c=>{switch(c){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let g=vn.pop()||Nt.length;return Nt[g]=c,Nt[g+1]=1,g}};function Vr(c){return this.fromWireType(p()[c>>>2>>>0])}var vs={name:"emscripten::val",fromWireType:c=>{var g=dt(c);return $n(c),g},toWireType:(c,g)=>Mt(g),argPackAdvance:8,readValueFromPointer:Vr,Bb:null},si=(c,g)=>{switch(g){case 4:return function(b){return this.fromWireType(h()[b>>>2>>>0])};case 8:return function(b){return this.fromWireType(m()[b>>>3>>>0])};default:throw new TypeError(`invalid float width (${g}): ${c}`)}},ea=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,$s=(c,g)=>{for(var b=c>>1,k=b+g/2;!(b>=k)&&l()[b>>>0];)++b;if(b<<=1,32=g/2);++k){var I=o()[c+2*k>>>1>>>0];if(I==0)break;b+=String.fromCharCode(I)}return b},oi=(c,g,b)=>{if(b??=2147483647,2>b)return 0;b-=2;var k=g;b=b<2*c.length?b/2:c.length;for(var I=0;I>>1>>>0]=z,g+=2}return o()[g>>>1>>>0]=0,g-k},li=c=>2*c.length,xn=(c,g)=>{for(var b=0,k="";!(b>=g/4);){var I=u()[c+4*b>>>2>>>0];if(I==0)break;++b,65536<=I?(I-=65536,k+=String.fromCharCode(55296|I>>10,56320|I&1023)):k+=String.fromCharCode(I)}return k},xs=(c,g,b)=>{if(g>>>=0,b??=2147483647,4>b)return 0;var k=g;b=k+b-4;for(var I=0;I=z){var V=c.charCodeAt(++I);z=65536+((z&1023)<<10)|V&1023}if(u()[g>>>2>>>0]=z,g+=4,g+4>b)break}return u()[g>>>2>>>0]=0,g-k},we=c=>{for(var g=0,b=0;b=k&&++b,g+=4}return g},ta=c=>{if(!ae)try{if(c(),!(0>>=0,typeof Atomics.nc=="function"&&(Atomics.nc(u(),c>>>2,c).value.then(xt),c+=128,Atomics.store(u(),c>>>2,1))}d.__emscripten_thread_mailbox_await=Mr;var xt=()=>{var c=pi();c&&(Mr(c),ta(ql))};d.checkMailbox=xt;var Sn=[],kn=(c,g)=>{var b=Zn[c];if(b===void 0)throw c=Gl(c),b=Dt(c),hr(c),new Jt(`${g} has unknown type ${b}`);return b},ra=(c,g,b)=>{var k=[];return c=c.toWireType(k,b),k.length&&(p()[g>>>2>>>0]=Mt(k)),c},Gr=c=>{try{c()}catch(g){Xe(g)}};function ui(){var c=he,g={};for(let[b,k]of Object.entries(c))g[b]=typeof k=="function"?(...I)=>{vr.push(b);try{return k(...I)}finally{ae||(vr.pop(),Ft&&er===1&&vr.length===0&&(er=0,ut+=1,Gr(Ql),typeof Fibers<"u"&&Fibers.rc()))}}:k;return g}var er=0,Ft=null,di=0,vr=[],na={},f={},y=0,x=null,C=[];function O(){return new Promise((c,g)=>{x={resolve:c,reject:g}})}function N(){var c=hi(65548),g=c+12;p()[c>>>2>>>0]=g,p()[c+4>>>2>>>0]=g+65536,g=vr[0];var b=na[g];return b===void 0&&(b=y++,na[g]=b,f[b]=g),g=b,u()[c+8>>>2>>>0]=g,c}function j(){var c=u()[Ft+8>>>2>>>0];return c=he[f[c]],--ut,c()}function Z(c){if(!ae){if(er===0){var g=!1,b=!1;c((k=0)=>{if(!ae&&(di=k,g=!0,b)){er=2,Gr(()=>Zl(Ft)),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.resume(),k=!1;try{var I=j()}catch(Y){I=Y,k=!0}var z=!1;if(!Ft){var V=x;V&&(x=null,(k?V.reject:V.resolve)(I),z=!0)}if(k&&!z)throw I}}),b=!0,g||(er=1,Ft=N(),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.pause(),Gr(()=>Xl(Ft)))}else er===2?(er=0,Gr(Jl),hr(Ft),Ft=null,C.forEach(ta)):Xe(`invalid state: ${er}`);return di}}function ce(c){return Z(g=>{c().then(g)})}var oe=[],me={},ze=c=>{var g=me[c];return g===void 0?Dt(c):g},Fe=()=>typeof globalThis=="object"?globalThis:Function("return this")(),F=c=>{var g=oe.length;return oe.push(c),g},xe=(c,g)=>{for(var b=Array(c),k=0;k>>2>>>0],"parameter "+k);return b},Oe=(c,g)=>Object.defineProperty(g,"name",{value:c});function $r(c){var g=Function;if(!(g instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof g} which is not a function`);var b=Oe(g.name||"unknownFunctionName",function(){});return b.prototype=g.prototype,b=new b,c=g.apply(b,c),c instanceof Object?c:b}var xr=c=>c%4===0&&(c%100!==0||c%400===0),Tl=[0,31,60,91,121,152,182,213,244,274,305,335],Al=[0,31,59,90,120,151,181,212,243,273,304,334];function Il(c,g,b,k,I,z,V){return D?Ie(16,1,c,g,b,k,I,z,V):-52}function Ml(c,g,b,k,I,z){if(D)return Ie(17,1,c,g,b,k,I,z)}var Ss=[],Ol=(c,g)=>{Ss.length=0;for(var b;b=s()[c++>>>0];){var k=b!=105;k&=b!=112,g+=k&&g%8?4:0,Ss.push(b==112?p()[g>>>2>>>0]:b==106?Ne[g>>>3]:b==105?u()[g>>>2>>>0]:m()[g>>>3>>>0]),g+=k?8:4}return Ss},ks={},zl=()=>{if(!Es){var c={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:E||"./this.program"},g;for(g in ks)ks[g]===void 0?delete c[g]:c[g]=ks[g];var b=[];for(g in c)b.push(`${g}=${c[g]}`);Es=b}return Es},Es;function Pl(c,g){if(D)return Ie(18,1,c,g);c>>>=0,g>>>=0;var b=0;return zl().forEach((k,I)=>{var z=g+b;for(I=p()[c+4*I>>>2>>>0]=z,z=0;z>>0]=k.charCodeAt(z);i()[I>>>0]=0,b+=k.length+1}),0}function Rl(c,g){if(D)return Ie(19,1,c,g);c>>>=0,g>>>=0;var b=zl();p()[c>>>2>>>0]=b.length;var k=0;return b.forEach(I=>k+=I.length+1),p()[g>>>2>>>0]=k,0}function Bl(c){return D?Ie(20,1,c):52}function Dl(c,g,b,k){return D?Ie(21,1,c,g,b,k):52}function Nl(c,g,b,k){return D?Ie(22,1,c,g,b,k):70}var w0=[null,[],[]];function Fl(c,g,b,k){if(D)return Ie(23,1,c,g,b,k);g>>>=0,b>>>=0,k>>>=0;for(var I=0,z=0;z>>2>>>0],Y=p()[g+4>>>2>>>0];g+=8;for(var pe=0;pe>>0],_e=w0[c];le===0||le===10?((c===1?re:ee)(Ur(_e,0)),_e.length=0):_e.push(le)}I+=Y}return p()[k>>>2>>>0]=I,0}var Ll=[31,29,31,30,31,30,31,31,30,31,30,31],Ul=[31,28,31,30,31,30,31,31,30,31,30,31];function b0(c){var g=Array(It(c)+1);return qn(c,g,0,g.length),g}var v0=(c,g)=>{i().set(c,g>>>0)};function Wl(c,g,b,k){function I(L,ke,De){for(L=typeof L=="number"?L.toString():L||"";L.lengthtu?-1:0Hr-L.getDate())ke-=Hr-L.getDate()+1,L.setDate(1),11>De?L.setMonth(De+1):(L.setMonth(0),L.setFullYear(L.getFullYear()+1));else{L.setDate(L.getDate()+ke);break}}return De=new Date(L.getFullYear()+1,0,4),ke=Y(new Date(L.getFullYear(),0,4)),De=Y(De),0>=V(ke,L)?0>=V(De,L)?L.getFullYear()+1:L.getFullYear():L.getFullYear()-1}c>>>=0,g>>>=0,b>>>=0,k>>>=0;var le=p()[k+40>>>2>>>0];k={kc:u()[k>>>2>>>0],jc:u()[k+4>>>2>>>0],Eb:u()[k+8>>>2>>>0],Jb:u()[k+12>>>2>>>0],Fb:u()[k+16>>>2>>>0],Ab:u()[k+20>>>2>>>0],vb:u()[k+24>>>2>>>0],zb:u()[k+28>>>2>>>0],qc:u()[k+32>>>2>>>0],ic:u()[k+36>>>2>>>0],lc:le?it(le):""},b=it(b),le={"%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"};for(var _e in le)b=b.replace(new RegExp(_e,"g"),le[_e]);var Le="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Ge="January February March April May June July August September October November December".split(" ");le={"%a":L=>Le[L.vb].substring(0,3),"%A":L=>Le[L.vb],"%b":L=>Ge[L.Fb].substring(0,3),"%B":L=>Ge[L.Fb],"%C":L=>z((L.Ab+1900)/100|0,2),"%d":L=>z(L.Jb,2),"%e":L=>I(L.Jb,2," "),"%g":L=>pe(L).toString().substring(2),"%G":pe,"%H":L=>z(L.Eb,2),"%I":L=>(L=L.Eb,L==0?L=12:12{for(var ke=0,De=0;De<=L.Fb-1;ke+=(xr(L.Ab+1900)?Ll:Ul)[De++]);return z(L.Jb+ke,3)},"%m":L=>z(L.Fb+1,2),"%M":L=>z(L.jc,2),"%n":()=>` +`,"%p":L=>0<=L.Eb&&12>L.Eb?"AM":"PM","%S":L=>z(L.kc,2),"%t":()=>" ","%u":L=>L.vb||7,"%U":L=>z(Math.floor((L.zb+7-L.vb)/7),2),"%V":L=>{var ke=Math.floor((L.zb+7-(L.vb+6)%7)/7);if(2>=(L.vb+371-L.zb-2)%7&&ke++,ke)ke==53&&(De=(L.vb+371-L.zb)%7,De==4||De==3&&xr(L.Ab)||(ke=1));else{ke=52;var De=(L.vb+7-L.zb-1)%7;(De==4||De==5&&xr(L.Ab%400-1))&&ke++}return z(ke,2)},"%w":L=>L.vb,"%W":L=>z(Math.floor((L.zb+7-(L.vb+6)%7)/7),2),"%y":L=>(L.Ab+1900).toString().substring(2),"%Y":L=>L.Ab+1900,"%z":L=>{L=L.ic;var ke=0<=L;return L=Math.abs(L)/60,(ke?"+":"-")+("0000"+(L/60*100+L%60)).slice(-4)},"%Z":L=>L.lc,"%%":()=>"%"},b=b.replace(/%%/g,"\0\0");for(_e in le)b.includes(_e)&&(b=b.replace(new RegExp(_e,"g"),le[_e](k)));return b=b.replace(/\0\0/g,"%"),_e=b0(b),_e.length>g?0:(v0(_e,c),_e.length-1)}be.Gb();for(var Vl=Array(256),ci=0;256>ci;++ci)Vl[ci]=String.fromCharCode(ci);bn=Vl,Jt=d.BindingError=class extends Error{constructor(c){super(c),this.name="BindingError"}},d.InternalError=class extends Error{constructor(c){super(c),this.name="InternalError"}},Nt.push(0,1,void 0,1,null,1,!0,1,!1,1),d.count_emval_handles=()=>Nt.length/2-5-vn.length;var $0=[At,Xa,Za,Ja,Qt,Kn,Yn,ei,wn,ti,ri,ni,ai,Xn,Wr,ii,Il,Ml,Pl,Rl,Bl,Dl,Nl,Fl],x0={ua:function(c,g,b){return ce(async()=>{await d.Zb(c,g,b)})},b:function(c,g,b){throw c>>>=0,new ws(c).Gb(g>>>0,b>>>0),Qa=c,Qa},ia:function(c){Hl(c>>>0,!P,1,!A,131072,!1),be.Pb()},G:function(c){c>>>=0,D?postMessage({cmd:"cleanupThread",thread:c}):be.Nb(be.ub[c])},_:yn,A:Ja,pa:Qt,W:Kn,Y:Yn,qa:ei,na:wn,fa:ti,ma:ri,K:ni,X:ai,U:Xn,oa:Wr,V:ii,E:function(c,g,b){c>>>=0,b>>>=0,g=Dt(g>>>0),_t(c,{name:g,fromWireType:k=>k,toWireType:function(k,I){if(typeof I!="bigint"&&typeof I!="number")throw I===null?I="null":(k=typeof I,I=k==="object"||k==="array"||k==="function"?I.toString():""+I),new TypeError(`Cannot convert "${I}" to ${this.name}`);return typeof I=="number"&&(I=BigInt(I)),I},argPackAdvance:8,readValueFromPointer:Jn(g,b,g.indexOf("u")==-1),Bb:null})},O:function(c,g,b,k){c>>>=0,g=Dt(g>>>0),_t(c,{name:g,fromWireType:function(I){return!!I},toWireType:function(I,z){return z?b:k},argPackAdvance:8,readValueFromPointer:function(I){return this.fromWireType(s()[I>>>0])},Bb:null})},N:function(c){return _t(c>>>0,vs)},D:function(c,g,b){c>>>=0,b>>>=0,g=Dt(g>>>0),_t(c,{name:g,fromWireType:k=>k,toWireType:(k,I)=>I,argPackAdvance:8,readValueFromPointer:si(g,b),Bb:null})},t:function(c,g,b,k,I){if(c>>>=0,b>>>=0,g=Dt(g>>>0),I===-1&&(I=4294967295),I=Y=>Y,k===0){var z=32-8*b;I=Y=>Y<>>z}var V=g.includes("unsigned")?function(Y,pe){return pe>>>0}:function(Y,pe){return pe};_t(c,{name:g,fromWireType:I,toWireType:V,argPackAdvance:8,readValueFromPointer:Jn(g,b,k!==0),Bb:null})},p:function(c,g,b){function k(z){var V=p()[z>>>2>>>0];return z=p()[z+4>>>2>>>0],new I(i().buffer,z,V)}c>>>=0;var I=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][g];b=Dt(b>>>0),_t(c,{name:b,fromWireType:k,argPackAdvance:8,readValueFromPointer:k},{Ub:!0})},F:function(c,g){c>>>=0,g=Dt(g>>>0);var b=g==="std::string";_t(c,{name:g,fromWireType:function(k){var I=p()[k>>>2>>>0],z=k+4;if(b)for(var V=z,Y=0;Y<=I;++Y){var pe=z+Y;if(Y==I||s()[pe>>>0]==0){if(V=it(V,pe-V),le===void 0)var le=V;else le+="\0",le+=V;V=pe+1}}else{for(le=Array(I),Y=0;Y>>0]);le=le.join("")}return hr(k),le},toWireType:function(k,I){I instanceof ArrayBuffer&&(I=new Uint8Array(I));var z=typeof I=="string";if(!(z||I instanceof Uint8Array||I instanceof Uint8ClampedArray||I instanceof Int8Array))throw new Jt("Cannot pass non-string to std::string");var V=b&&z?It(I):I.length,Y=hi(4+V+1),pe=Y+4;if(p()[Y>>>2>>>0]=V,b&&z)Zt(I,pe,V+1);else if(z)for(z=0;z>>0]=le}else for(z=0;z>>0]=I[z];return k!==null&&k.push(hr,Y),Y},argPackAdvance:8,readValueFromPointer:Vr,Bb(k){hr(k)}})},z:function(c,g,b){if(c>>>=0,g>>>=0,b>>>=0,b=Dt(b),g===2)var k=$s,I=oi,z=li,V=Y=>l()[Y>>>1>>>0];else g===4&&(k=xn,I=xs,z=we,V=Y=>p()[Y>>>2>>>0]);_t(c,{name:b,fromWireType:Y=>{for(var pe=p()[Y>>>2>>>0],le,_e=Y+4,Le=0;Le<=pe;++Le){var Ge=Y+4+Le*g;(Le==pe||V(Ge)==0)&&(_e=k(_e,Ge-_e),le===void 0?le=_e:(le+="\0",le+=_e),_e=Ge+g)}return hr(Y),le},toWireType:(Y,pe)=>{if(typeof pe!="string")throw new Jt(`Cannot pass non-string to C++ string type ${b}`);var le=z(pe),_e=hi(4+le+g);return p()[_e>>>2>>>0]=le/g,I(pe,_e+4,le+g),Y!==null&&Y.push(hr,_e),_e},argPackAdvance:8,readValueFromPointer:Vr,Bb(Y){hr(Y)}})},P:function(c,g){c>>>=0,g=Dt(g>>>0),_t(c,{Vb:!0,name:g,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ta:()=>1,S:function(c,g){c>>>=0,c==g>>>0?setTimeout(xt):D?postMessage({targetThread:c,cmd:"checkMailbox"}):(c=be.ub[c])&&c.postMessage({cmd:"checkMailbox"})},$:function(c,g,b,k,I){g>>>=0,b>>>=0,k/=2,Sn.length=k,I=I>>>0>>>3;for(var z=0;z>>0];return c=g?Gn[g]:$0[c],be.Sb=b,b=c(...Sn),be.Sb=0,b},ha:Mr,sa:function(c){B&&be.ub[c>>>0].ref()},w:function(c,g,b){return g>>>=0,b>>>=0,c=dt(c>>>0),g=kn(g,"emval::as"),ra(g,b,c)},y:function(c){return c>>>=0,ce(()=>(c=dt(c),c.then(Mt)))},o:function(c,g,b,k){return b>>>=0,k>>>=0,c=oe[c>>>0],g=dt(g>>>0),c(null,g,b,k)},k:function(c,g,b,k,I){return b>>>=0,k>>>=0,I>>>=0,c=oe[c>>>0],g=dt(g>>>0),b=ze(b),c(g,g[b],k,I)},c:$n,C:function(c,g){return g>>>=0,c=dt(c>>>0),g=dt(g),c==g},n:function(c){return c>>>=0,c===0?Mt(Fe()):(c=ze(c),Mt(Fe()[c]))},j:function(c,g,b){g=xe(c,g>>>0);var k=g.shift();c--;var I=`return function (obj, func, destructorsRef, args) { +`,z=0,V=[];b===0&&V.push("obj");for(var Y=["retType"],pe=[k],le=0;le_e.name).join(", ")}) => ${k.name}>`,F(Oe(b,c))},s:function(c,g){return g>>>=0,c=dt(c>>>0),g=dt(g),Mt(c[g])},d:function(c){c>>>=0,9>>0);for(var g=Array(c.length),b=0;b>>0))},l:function(){return Mt({})},i:function(c){c>>>=0;for(var g=dt(c);g.length;){var b=g.pop();g.pop()(b)}$n(c)},h:function(c,g,b){g>>>=0,b>>>=0,c=dt(c>>>0),g=dt(g),b=dt(b),c[g]=b},e:function(c,g){return g>>>=0,c=kn(c>>>0,"_emval_take_value"),c=c.readValueFromPointer(g),Mt(c)},ca:function(c,g){c=-9007199254740992>c||9007199254740992>>=0,c=new Date(1e3*c),u()[g>>>2>>>0]=c.getUTCSeconds(),u()[g+4>>>2>>>0]=c.getUTCMinutes(),u()[g+8>>>2>>>0]=c.getUTCHours(),u()[g+12>>>2>>>0]=c.getUTCDate(),u()[g+16>>>2>>>0]=c.getUTCMonth(),u()[g+20>>>2>>>0]=c.getUTCFullYear()-1900,u()[g+24>>>2>>>0]=c.getUTCDay(),c=(c.getTime()-Date.UTC(c.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,u()[g+28>>>2>>>0]=c},da:function(c,g){c=-9007199254740992>c||9007199254740992>>=0,c=new Date(1e3*c),u()[g>>>2>>>0]=c.getSeconds(),u()[g+4>>>2>>>0]=c.getMinutes(),u()[g+8>>>2>>>0]=c.getHours(),u()[g+12>>>2>>>0]=c.getDate(),u()[g+16>>>2>>>0]=c.getMonth(),u()[g+20>>>2>>>0]=c.getFullYear()-1900,u()[g+24>>>2>>>0]=c.getDay();var b=(xr(c.getFullYear())?Tl:Al)[c.getMonth()]+c.getDate()-1|0;u()[g+28>>>2>>>0]=b,u()[g+36>>>2>>>0]=-(60*c.getTimezoneOffset()),b=new Date(c.getFullYear(),6,1).getTimezoneOffset();var k=new Date(c.getFullYear(),0,1).getTimezoneOffset();c=(b!=k&&c.getTimezoneOffset()==Math.min(k,b))|0,u()[g+32>>>2>>>0]=c},ea:function(c){c>>>=0;var g=new Date(u()[c+20>>>2>>>0]+1900,u()[c+16>>>2>>>0],u()[c+12>>>2>>>0],u()[c+8>>>2>>>0],u()[c+4>>>2>>>0],u()[c>>>2>>>0],0),b=u()[c+32>>>2>>>0],k=g.getTimezoneOffset(),I=new Date(g.getFullYear(),6,1).getTimezoneOffset(),z=new Date(g.getFullYear(),0,1).getTimezoneOffset(),V=Math.min(z,I);return 0>b?u()[c+32>>>2>>>0]=+(I!=z&&V==k):0>>2>>>0]=g.getDay(),b=(xr(g.getFullYear())?Tl:Al)[g.getMonth()]+g.getDate()-1|0,u()[c+28>>>2>>>0]=b,u()[c>>>2>>>0]=g.getSeconds(),u()[c+4>>>2>>>0]=g.getMinutes(),u()[c+8>>>2>>>0]=g.getHours(),u()[c+12>>>2>>>0]=g.getDate(),u()[c+16>>>2>>>0]=g.getMonth(),u()[c+20>>>2>>>0]=g.getYear(),c=g.getTime(),BigInt(isNaN(c)?-1:c/1e3)},aa:Il,ba:Ml,R:function(c,g,b,k){c>>>=0,g>>>=0,b>>>=0,k>>>=0;var I=new Date().getFullYear(),z=new Date(I,0,1),V=new Date(I,6,1);I=z.getTimezoneOffset();var Y=V.getTimezoneOffset(),pe=Math.max(I,Y);p()[c>>>2>>>0]=60*pe,u()[g>>>2>>>0]=+(I!=Y),c=le=>le.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],z=c(z),V=c(V),Y{Xe("")},f:function(c,g,b){return c>>>=0,g=Ol(g>>>0,b>>>0),Gn[c](...g)},M:function(c,g,b){return c>>>=0,g=Ol(g>>>0,b>>>0),Gn[c](...g)},H:()=>{},L:()=>Date.now(),ra:()=>{throw ut+=1,"unwind"},T:function(){return 4294901760},u:()=>performance.timeOrigin+performance.now(),x:()=>B?(fy(),nn(ch)).cpus().length:navigator.hardwareConcurrency,Q:function(c){c>>>=0;var g=s().length;if(c<=g||4294901760=b;b*=2){var k=g*(1+.2/b);k=Math.min(k,c+100663296);var I=Math;k=Math.max(c,k);e:{I=(I.min.call(I,4294901760,k+(65536-k%65536)%65536)-ne.buffer.byteLength+65535)/65536;try{ne.grow(I),lt();var z=1;break e}catch{}z=void 0}if(z)return!0}return!1},ja:Pl,la:Rl,Z:Lr,B:Bl,J:Dl,ga:Nl,I:Fl,a:ne||d.wasmMemory,ka:Wl,r:function(c,g,b,k){return Wl(c>>>0,g>>>0,b>>>0,k>>>0)}},he=function(){function c(b,k){return he=b.exports,he=ui(),he=S0(),be.Qb.push(he.cb),cr.unshift(he.va),W=k,Wn(),he}var g={a:x0};if(br++,d.instantiateWasm)try{return d.instantiateWasm(g,c)}catch(b){ee(`Module.instantiateWasm callback failed with error: ${b}`),w(b)}return pr(g,function(b){c(b.instance,b.module)}).catch(w),{}}(),Gl=c=>(Gl=he.wa)(c);d.__embind_initialize_bindings=()=>(d.__embind_initialize_bindings=he.xa)(),d._OrtInit=(c,g)=>(d._OrtInit=he.ya)(c,g),d._OrtGetLastError=(c,g)=>(d._OrtGetLastError=he.za)(c,g),d._OrtCreateSessionOptions=(c,g,b,k,I,z,V,Y,pe,le)=>(d._OrtCreateSessionOptions=he.Aa)(c,g,b,k,I,z,V,Y,pe,le),d._OrtAppendExecutionProvider=(c,g)=>(d._OrtAppendExecutionProvider=he.Ba)(c,g),d._OrtAddFreeDimensionOverride=(c,g,b)=>(d._OrtAddFreeDimensionOverride=he.Ca)(c,g,b),d._OrtAddSessionConfigEntry=(c,g,b)=>(d._OrtAddSessionConfigEntry=he.Da)(c,g,b),d._OrtReleaseSessionOptions=c=>(d._OrtReleaseSessionOptions=he.Ea)(c),d._OrtCreateSession=(c,g,b)=>(d._OrtCreateSession=he.Fa)(c,g,b),d._OrtReleaseSession=c=>(d._OrtReleaseSession=he.Ga)(c),d._OrtGetInputOutputCount=(c,g,b)=>(d._OrtGetInputOutputCount=he.Ha)(c,g,b),d._OrtGetInputName=(c,g)=>(d._OrtGetInputName=he.Ia)(c,g),d._OrtGetOutputName=(c,g)=>(d._OrtGetOutputName=he.Ja)(c,g),d._OrtFree=c=>(d._OrtFree=he.Ka)(c),d._OrtCreateTensor=(c,g,b,k,I,z)=>(d._OrtCreateTensor=he.La)(c,g,b,k,I,z),d._OrtGetTensorData=(c,g,b,k,I)=>(d._OrtGetTensorData=he.Ma)(c,g,b,k,I),d._OrtReleaseTensor=c=>(d._OrtReleaseTensor=he.Na)(c),d._OrtCreateRunOptions=(c,g,b,k)=>(d._OrtCreateRunOptions=he.Oa)(c,g,b,k),d._OrtAddRunConfigEntry=(c,g,b)=>(d._OrtAddRunConfigEntry=he.Pa)(c,g,b),d._OrtReleaseRunOptions=c=>(d._OrtReleaseRunOptions=he.Qa)(c),d._OrtCreateBinding=c=>(d._OrtCreateBinding=he.Ra)(c),d._OrtBindInput=(c,g,b)=>(d._OrtBindInput=he.Sa)(c,g,b),d._OrtBindOutput=(c,g,b,k)=>(d._OrtBindOutput=he.Ta)(c,g,b,k),d._OrtClearBoundOutputs=c=>(d._OrtClearBoundOutputs=he.Ua)(c),d._OrtReleaseBinding=c=>(d._OrtReleaseBinding=he.Va)(c),d._OrtRunWithBinding=(c,g,b,k,I)=>(d._OrtRunWithBinding=he.Wa)(c,g,b,k,I),d._OrtRun=(c,g,b,k,I,z,V,Y)=>(d._OrtRun=he.Xa)(c,g,b,k,I,z,V,Y),d._OrtEndProfiling=c=>(d._OrtEndProfiling=he.Ya)(c),d._JsepOutput=(c,g,b)=>(d._JsepOutput=he.Za)(c,g,b),d._JsepGetNodeName=c=>(d._JsepGetNodeName=he._a)(c);var pi=d._pthread_self=()=>(pi=d._pthread_self=he.$a)(),hi=d._malloc=c=>(hi=d._malloc=he.ab)(c),hr=d._free=c=>(hr=d._free=he.bb)(c);d.__emscripten_tls_init=()=>(d.__emscripten_tls_init=he.cb)();var Hl=d.__emscripten_thread_init=(c,g,b,k,I,z)=>(Hl=d.__emscripten_thread_init=he.eb)(c,g,b,k,I,z);d.__emscripten_thread_crashed=()=>(d.__emscripten_thread_crashed=he.fb)();var jl=(c,g,b,k,I)=>(jl=he.gb)(c,g,b,k,I),Cs=c=>(Cs=he.hb)(c),Ts=d.__emscripten_thread_exit=c=>(Ts=d.__emscripten_thread_exit=he.ib)(c),ql=()=>(ql=he.jb)(),Kl=(c,g)=>(Kl=he.kb)(c,g),fi=c=>(fi=he.lb)(c),As=c=>(As=he.mb)(c),Is=()=>(Is=he.nb)(),Yl=d.dynCall_ii=(c,g)=>(Yl=d.dynCall_ii=he.pb)(c,g),Xl=c=>(Xl=he.qb)(c),Ql=()=>(Ql=he.rb)(),Zl=c=>(Zl=he.sb)(c),Jl=()=>(Jl=he.tb)();d.___start_em_js=839500,d.___stop_em_js=839661;function S0(){var c=he;c=Object.assign({},c);var g=k=>I=>k(I)>>>0,b=k=>()=>k()>>>0;return c.wa=g(c.wa),c.$a=b(c.$a),c.ab=g(c.ab),c.emscripten_main_runtime_thread_id=b(c.emscripten_main_runtime_thread_id),c.mb=g(c.mb),c.nb=b(c.nb),c}d.wasmMemory=ne,d.stackSave=()=>Is(),d.stackRestore=c=>fi(c),d.stackAlloc=c=>As(c),d.keepRuntimeAlive=()=>0r)}),gy=Bn((t,e)=>{e.exports='"use strict";var Module={},ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads"),parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",e=>onmessage({data:e}));var fs=require("fs"),vm=require("vm");Object.assign(global,{self:global,require,Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:e=>vm.runInThisContext(fs.readFileSync(e,"utf8"),{filename:e}),postMessage:e=>parentPort.postMessage(e),performance:global.performance||{now:Date.now}})}var initializedJS=!1;function threadPrintErr(...e){var a=e.join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,a+`\n`);return}console.error(a)}function threadAlert(...e){var a=e.join(" ");postMessage({cmd:"alert",text:a,threadId:Module._pthread_self()})}var err=threadPrintErr;self.alert=threadAlert,Module.instantiateWasm=(e,a)=>{var r=Module.wasmModule;Module.wasmModule=null;var t=new WebAssembly.Instance(r,e);return a(t)},self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let r=[];self.onmessage=t=>r.push(t),self.startWorker=t=>{Module=t,postMessage({cmd:"loaded"});for(let s of r)handleMessage(s);self.onmessage=handleMessage},Module.wasmModule=e.data.wasmModule;for(const t of e.data.handlers)Module[t]=(...s)=>{postMessage({cmd:"callHandler",handler:t,args:s})};if(Module.wasmMemory=e.data.wasmMemory,Module.buffer=Module.wasmMemory.buffer,Module.ENVIRONMENT_IS_PTHREAD=!0,typeof e.data.urlOrBlob=="string")importScripts(e.data.urlOrBlob);else{var a=URL.createObjectURL(e.data.urlOrBlob);importScripts(a),URL.revokeObjectURL(a)}ortWasmThreaded(Module)}else if(e.data.cmd==="run"){Module.__emscripten_thread_init(e.data.pthread_ptr,0,0,1),Module.__emscripten_thread_mailbox_await(e.data.pthread_ptr),Module.establishStackSpace(),Module.PThread.receiveObjectTransfer(e.data),Module.PThread.threadInitTLS(),initializedJS||(initializedJS=!0);try{Module.invokeEntryPoint(e.data.start_routine,e.data.arg)}catch(r){if(r!="unwind")throw r}}else e.data.cmd==="cancel"?Module._pthread_self()&&Module.__emscripten_thread_exit(-1):e.data.target==="setimmediate"||(e.data.cmd==="checkMailbox"?initializedJS&&Module.checkMailbox():e.data.cmd&&(err(`worker.js received unknown command ${e.data.cmd}`),err(e.data)))}catch(r){throw Module.__emscripten_thread_crashed?.(),r}}self.onmessage=handleMessage;\n'}),ou,lu,Rs,yi,ia,Bs,uu,du,cu,hh,st,Aa=J(()=>{ou=cy(),lu=my(),yi=!1,ia=!1,Bs=!1,uu=t=>{if(t===1)return!1;if(typeof SharedArrayBuffer>"u")return typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+t+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),!1;typeof process<"u"&&process.versions&&process.versions.node&&console.warn("env.wasm.numThreads is set to "+t+", however, currently onnxruntime-web does not support multi-threads in Node.js. Please consider using onnxruntime-node for performance critical scenarios.");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}},du=()=>{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}},cu=(t,e)=>t?e?"ort-wasm-simd-threaded.wasm":"ort-wasm-simd.wasm":e?"ort-wasm-threaded.wasm":"ort-wasm.wasm",hh=async t=>{if(yi)return Promise.resolve();if(ia)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Bs)throw new Error("previous call to 'initializeWebAssembly()' failed.");ia=!0;let e=t.initTimeout,r=t.numThreads,n=t.simd,a=uu(r),i=n&&du(),s=t.wasmPaths,o=typeof s=="string"?s:void 0,l=cu(i,a),u=typeof s=="object"?s[l]:void 0,p=!1,h=[];if(e>0&&h.push(new Promise(m=>{setTimeout(()=>{p=!0,m()},e)})),h.push(new Promise((m,d)=>{let _=a?lu:ou,w={locateFile:(v,S)=>{if(a&&v.endsWith(".worker.js")&&typeof Blob<"u")return URL.createObjectURL(new Blob([gy()],{type:"text/javascript"}));if(v.endsWith(".wasm")){if(u)return u;let $=o??S;return l==="ort-wasm-simd.wasm"?$+"ort-wasm-simd.jsep.wasm":l==="ort-wasm-simd-threaded.wasm"?$+"ort-wasm-simd-threaded.jsep.wasm":$+l}return S+v}};if(a)if(w.numThreads=r,typeof Blob>"u")w.mainScriptUrlOrBlob=(void 0)(__dirname,"ort-wasm-threaded.js");else{let v=`var ortWasmThreaded=${_.toString()};`;w.mainScriptUrlOrBlob=new Blob([v],{type:"text/javascript"})}_(w).then(v=>{ia=!1,yi=!0,Rs=v,m()},v=>{ia=!1,Bs=!0,d(v)})})),await Promise.race(h),p)throw new Error(`WebAssembly backend initializing failed due to timeout: ${e}ms`)},st=()=>{if(yi&&Rs)return Rs;throw new Error("WebAssembly is not initialized yet.")}}),nt,Wi,qe,qo=J(()=>{Aa(),nt=(t,e)=>{let r=st(),n=r.lengthBytesUTF8(t)+1,a=r._malloc(n);return r.stringToUTF8(t,a,n),e.push(a),a},Wi=(t,e,r,n)=>{if(typeof t=="object"&&t!==null){if(r.has(t))throw new Error("Circular reference in options");r.add(t)}Object.entries(t).forEach(([a,i])=>{let s=e?e+a:a;if(typeof i=="object")Wi(i,s+".",r,n);else if(typeof i=="string"||typeof i=="number")n(s,i.toString());else if(typeof i=="boolean")n(s,i?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof i}`)})},qe=t=>{let e=st(),r=e.stackSave();try{let n=e.stackAlloc(8);e._OrtGetLastError(n,n+4);let a=e.HEAP32[n/4],i=e.HEAPU32[n/4+1],s=i?e.UTF8ToString(i):"";throw new Error(`${t} ERROR_CODE: ${a}, ERROR_MESSAGE: ${s}`)}finally{e.stackRestore(r)}}}),fh,_y=J(()=>{Aa(),qo(),fh=t=>{let e=st(),r=0,n=[],a=t||{};try{if(t?.logSeverityLevel===void 0)a.logSeverityLevel=2;else if(typeof t.logSeverityLevel!="number"||!Number.isInteger(t.logSeverityLevel)||t.logSeverityLevel<0||t.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${t.logSeverityLevel}`);if(t?.logVerbosityLevel===void 0)a.logVerbosityLevel=0;else if(typeof t.logVerbosityLevel!="number"||!Number.isInteger(t.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${t.logVerbosityLevel}`);t?.terminate===void 0&&(a.terminate=!1);let i=0;return t?.tag!==void 0&&(i=nt(t.tag,n)),r=e._OrtCreateRunOptions(a.logSeverityLevel,a.logVerbosityLevel,!!a.terminate,i),r===0&&qe("Can't create run options."),t?.extra!==void 0&&Wi(t.extra,"",new WeakSet,(s,o)=>{let l=nt(s,n),u=nt(o,n);e._OrtAddRunConfigEntry(r,l,u)!==0&&qe(`Can't set a run config entry: ${s} - ${o}.`)}),[r,n]}catch(i){throw r!==0&&e._OrtReleaseRunOptions(r),n.forEach(s=>e._free(s)),i}}}),pu,hu,fu,mu,mh,yy=J(()=>{Aa(),qo(),pu=t=>{switch(t){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${t}`)}},hu=t=>{switch(t){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${t}`)}},fu=t=>{t.extra||(t.extra={}),t.extra.session||(t.extra.session={});let e=t.extra.session;e.use_ort_model_bytes_directly||(e.use_ort_model_bytes_directly="1"),t.executionProviders&&t.executionProviders.some(r=>(typeof r=="string"?r:r.name)==="webgpu")&&(t.enableMemPattern=!1)},mu=(t,e,r)=>{for(let n of e){let a=typeof n=="string"?n:n.name;switch(a){case"webnn":if(a="WEBNN",typeof n!="string"){let s=n;if(s?.deviceType){let o=nt("deviceType",r),l=nt(s.deviceType,r);st()._OrtAddSessionConfigEntry(t,o,l)!==0&&qe(`Can't set a session config entry: 'deviceType' - ${s.deviceType}.`)}if(s?.numThreads){let o=s.numThreads;(typeof o!="number"||!Number.isInteger(o)||o<0)&&(o=0);let l=nt("numThreads",r),u=nt(o.toString(),r);st()._OrtAddSessionConfigEntry(t,l,u)!==0&&qe(`Can't set a session config entry: 'numThreads' - ${s.numThreads}.`)}if(s?.powerPreference){let o=nt("powerPreference",r),l=nt(s.powerPreference,r);st()._OrtAddSessionConfigEntry(t,o,l)!==0&&qe(`Can't set a session config entry: 'powerPreference' - ${s.powerPreference}.`)}}break;case"webgpu":if(a="JS",typeof n!="string"){let s=n;if(s?.preferredLayout){if(s.preferredLayout!=="NCHW"&&s.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${s.preferredLayout}`);let o=nt("preferredLayout",r),l=nt(s.preferredLayout,r);st()._OrtAddSessionConfigEntry(t,o,l)!==0&&qe(`Can't set a session config entry: 'preferredLayout' - ${s.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${a}`)}let i=nt(a,r);st()._OrtAppendExecutionProvider(t,i)!==0&&qe(`Can't append execution provider: ${a}.`)}},mh=t=>{let e=st(),r=0,n=[],a=t||{};fu(a);try{let i=pu(a.graphOptimizationLevel??"all"),s=hu(a.executionMode??"sequential"),o=typeof a.logId=="string"?nt(a.logId,n):0,l=a.logSeverityLevel??2;if(!Number.isInteger(l)||l<0||l>4)throw new Error(`log serverity level is not valid: ${l}`);let u=a.logVerbosityLevel??0;if(!Number.isInteger(u)||u<0||u>4)throw new Error(`log verbosity level is not valid: ${u}`);let p=typeof a.optimizedModelFilePath=="string"?nt(a.optimizedModelFilePath,n):0;if(r=e._OrtCreateSessionOptions(i,!!a.enableCpuMemArena,!!a.enableMemPattern,s,!!a.enableProfiling,0,o,l,u,p),r===0&&qe("Can't create session options."),a.executionProviders&&mu(r,a.executionProviders,n),a.enableGraphCapture!==void 0){if(typeof a.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${a.enableGraphCapture}`);let h=nt("enableGraphCapture",n),m=nt(a.enableGraphCapture.toString(),n);e._OrtAddSessionConfigEntry(r,h,m)!==0&&qe(`Can't set a session config entry: 'enableGraphCapture' - ${a.enableGraphCapture}.`)}if(a.freeDimensionOverrides)for(let[h,m]of Object.entries(a.freeDimensionOverrides)){if(typeof h!="string")throw new Error(`free dimension override name must be a string: ${h}`);if(typeof m!="number"||!Number.isInteger(m)||m<0)throw new Error(`free dimension override value must be a non-negative integer: ${m}`);let d=nt(h,n);e._OrtAddFreeDimensionOverride(r,d,m)!==0&&qe(`Can't set a free dimension override: ${h} - ${m}.`)}return a.extra!==void 0&&Wi(a.extra,"",new WeakSet,(h,m)=>{let d=nt(h,n),_=nt(m,n);e._OrtAddSessionConfigEntry(r,d,_)!==0&&qe(`Can't set a session config entry: ${h} - ${m}.`)}),[r,n]}catch(i){throw r!==0&&e._OrtReleaseSessionOptions(r),n.forEach(s=>e._free(s)),i}}}),bo,Jr,$a,Ko,Vi,Yo,vo,$e=J(()=>{bo=t=>{switch(t){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: ${t}`)}},Jr=t=>{switch(t){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: ${t}`)}},$a=t=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][t],Ko=t=>{switch(t){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: ${t}`)}},Vi=t=>{switch(t){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: ${t}`)}},Yo=t=>t==="float32"||t==="float16"||t==="int32"||t==="int64"||t==="uint32"||t==="uint8"||t==="bool",vo=t=>{switch(t){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: ${t}`)}}}),Gi,gh=J(()=>{Gi=async t=>{if(typeof t=="string")if(typeof process<"u"&&process.versions&&process.versions.node)try{return new Uint8Array(await(void 0)(t))}catch(e){if(e.code==="ERR_FS_FILE_TOO_LARGE"){let r=(void 0)(t),n=[];for await(let a of r)n.push(a);return new Uint8Array(Buffer.concat(n))}throw e}else{let e=await fetch(t);if(!e.ok)throw new Error(`failed to load external data file: ${t}`);let r=e.headers.get("Content-Length"),n=r?parseInt(r,10):0;if(n<1073741824)return new Uint8Array(await e.arrayBuffer());{if(!e.body)throw new Error(`failed to load external data file: ${t}, no response body.`);let a=e.body.getReader(),i;try{i=new ArrayBuffer(n)}catch(o){if(o instanceof RangeError){let l=Math.ceil(n/65536);i=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw o}let s=0;for(;;){let{done:o,value:l}=await a.read();if(o)break;let u=l.byteLength;new Uint8Array(i,s,u).set(l),s+=u}return new Uint8Array(i,0,n)}}else return t instanceof Blob?new Uint8Array(await t.arrayBuffer()):t instanceof Uint8Array?t:new Uint8Array(t)}}),gu,_u,yu,wu,_h,bu,at,un=J(()=>{$e(),gu=["V","I","W","E","F"],_u=(t,e)=>{console.log(`[${gu[t]},${new Date().toISOString()}]${e}`)},_h=(t,e)=>{yu=t,wu=e},bu=(t,e)=>{let r=Vi(t),n=Vi(yu);r>=n&&_u(r,typeof e=="function"?e():e)},at=(...t)=>{wu&&bu(...t)}}),yh,wy=J(()=>{$e(),yh=(t,e)=>new(Ko(e))(t)}),Xo=J(()=>{}),Ds,wi,bi,vu,$u,Ns,$o,xu,wh,by=J(()=>{un(),Xo(),Ds=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]]),wi=[],bi=t=>Math.ceil(t/16)*16,vu=t=>{for(let e=0;e$u++,$o=async(t,e,r,n)=>{let a=bi(r),i=t.device.createBuffer({size:a,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let s=t.getCommandEncoder();t.endComputePass(),s.copyBufferToBuffer(e,0,i,0,a),t.flush(),await i.mapAsync(GPUMapMode.READ);let o=i.getMappedRange();if(n){let l=n();return l.set(new Uint8Array(o,0,r)),l}else return new Uint8Array(o.slice(0,r))}finally{i.destroy()}},xu=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[e]of Ds)wi.push(e),this.freeBuffers.set(e,[]),this.freeUniformBuffers.set(e,[])}upload(t,e){let r=e.buffer,n=e.byteOffset,a=e.byteLength,i=bi(a),s=this.storageCache.get(t);if(!s)throw new Error("gpu data for uploading does not exist");if(s.originalSize!==a)throw new Error(`inconsistent data size. gpu data size=${s.originalSize}, data size=${a}`);let o=this.backend.device.createBuffer({mappedAtCreation:!0,size:i,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),l=o.getMappedRange();new Uint8Array(l).set(new Uint8Array(r,n,a)),o.unmap();let u=this.backend.getCommandEncoder();this.backend.endComputePass(),u.copyBufferToBuffer(o,0,s.gpuData.buffer,0,i),at("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(o)}memcpy(t,e){let r=this.storageCache.get(t);if(!r)throw new Error("source gpu data for memcpy does not exist");let n=this.storageCache.get(e);if(!n)throw new Error("destination gpu data for memcpy does not exist");if(r.originalSize!==n.originalSize)throw new Error("inconsistent source and destination gpu data size");let a=bi(r.originalSize),i=this.backend.getCommandEncoder();this.backend.endComputePass(),i.copyBufferToBuffer(r.gpuData.buffer,0,n.gpuData.buffer,0,a)}registerExternalBuffer(t,e,r){let n;if(r){if(n=this.externalBuffers.get(r),n===void 0)throw new Error("previous buffer is not registered");if(t===r)return at("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${e}) => id=${n}, buffer is the same, skip.`),n;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet. + Please use the previous external buffer!`);this.externalBuffers.delete(r)}else n=Ns();return this.storageCache.set(n,{gpuData:{id:n,type:0,buffer:t},originalSize:e}),this.externalBuffers.set(t,n),at("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${e}) => id=${n}, registered.`),n}unregisterExternalBuffer(t){let e=this.externalBuffers.get(t);e!==void 0&&(this.storageCache.delete(e),this.externalBuffers.delete(t),at("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${e}`))}create(t,e=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=vu(t),n,a=(e&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,i=(e&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(a||i){let o=(a?this.freeBuffers:this.freeUniformBuffers).get(r);o?o.length>0?n=o.pop():n=this.backend.device.createBuffer({size:r,usage:e}):n=this.backend.device.createBuffer({size:r,usage:e})}else n=this.backend.device.createBuffer({size:r,usage:e});let s={id:Ns(),type:0,buffer:n};return this.storageCache.set(s.id,{gpuData:s,originalSize:t}),at("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${s.id}`),s}get(t){return this.storageCache.get(t)?.gpuData}release(t){let e=this.storageCache.get(t);if(!e)throw new Error("releasing data does not exist");return at("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${e.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(e.gpuData.buffer),e.originalSize}async download(t,e){let r=this.storageCache.get(t);if(!r)throw new Error("data does not exist");await $o(this.backend,r.gpuData.buffer,r.originalSize,e)}refreshPendingBuffers(){for(let t of this.buffersForUploadingPending)t.destroy();if(this.buffersForUploadingPending=[],this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let e=Ds.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(t.size)||[];e===void 0||r.length>=e?t.destroy():r.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(t.size)||[];e===void 0||r.length>=e?t.destroy():r.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let e of this.buffersPending)t.push(e);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onReleaseSession(t){let e=this.capturedPendingBuffers.get(t);e&&(e.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(t))}},wh=(...t)=>new xu(...t)}),Su,Ke,ct=J(()=>{Su=class{constructor(t){Object.assign(this,t)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(t=>`${this[t]}`).join(";")),this.key}},Ke=t=>new Su(t)}),ku,zn,K,Hi,bh,Qo,Zo,Me=J(()=>{ku=class{static calcMatMulShape(t,e){return t[1]!==e[0]?void 0:[t[0],e[1]]}},zn=class{static calcShape(t,e,r=!1){let n=t.length,a=e.length;if(n===0)return e;if(a===0)return t;let i=Math.max(t.length,e.length),s=new Array(i);if(r){if(n<2||a<2)return;let o=ku.calcMatMulShape([t[n-2],t[n-1]],[e[a-2],e[a-1]]);if(o===void 0)return;[s[i-2],s[i-1]]=o}for(let o=r?3:1;o<=i;o++){let l=n-o<0?1:t[n-o],u=a-o<0?1:e[a-o];if(l!==u&&l>1&&u>1)return;let p=Math.max(l,u);if(l&&u)s[i-o]=Math.max(l,u);else{if(p>1)return;s[i-o]=0}}return s}static isValidBroadcast(t,e){let r=t.length,n=e.length;if(r>n)return!1;for(let a=1;a<=r;a++)if(t[r-a]!==1&&t[r-a]!==e[n-a])return!1;return!0}},K=class Oi{static size(e){return Oi.getSizeFromDimensionRange(e,0,e.length)}static convertShape(e,r=4){let n=e.length;if(n===0)return[];let a=new Array(n),i=n-1;for(;i>=0;){if(e[i]%r===0){a[i]=e[i]/r;break}if(r%e[i]!==0)throw new Error("cannot convert shape");a[i]=1,r/=e[i],i--}for(i--;i>=0;i--)a[i]=e[i];return a}static sizeFromDimension(e,r){if(r<0||r>e.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${e.length} dimensions.`);return Oi.getSizeFromDimensionRange(e,r,e.length)}static sizeToDimension(e,r){if(r<0||r>e.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${e.length} dimensions.`);return Oi.getSizeFromDimensionRange(e,0,r)}static getSizeFromDimensionRange(e,r,n){let a=1;for(let i=r;i=0;--a)n[a]=n[a+1]*e[a+1];return n}static normalizeAxis(e,r){if(e<-r&&e>=r)throw new Error("unsupported axis for this operation.");return e<0?e+r:e}static normalizeAxes(e,r){return e.map(n=>this.normalizeAxis(n,r??e.length))}static sortBasedOnPerm(e,r){return r?r.map(n=>e[n]):e.slice().reverse()}static padShape(e,r){let n=e.length;return e.map((a,i)=>a+r[i]+r[i+n])}static areEqual(e,r){return e.length!==r.length?!1:e.every((n,a)=>n===r[a])}},Hi=class fa{static adjustPoolAttributes(e,r,n,a,i,s){if(!e&&n.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(e)for(let o=0;o=n.length?n.push(r[o+2]):n[o]=r[o+2];for(let o=0;o=n[o]||s[o+n.length]>=n[o])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(e,r,n,a,i,s,o){if(o){if(i.length!==2*(e.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==e.length-2)throw new Error("length of strides should be the length of data dimensions");if(a.length!==e.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let l=0;l{$e(),Me(),Pn=64,vi=(t,e)=>{if(e===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(t){case 10:return e>1?`vec${e}`:"f16";case 1:return e>1?`vec${e}`:"f32";case 6:return e>1?`vec${e}`:"i32";case 12:return e>1?`vec${e}`:"u32";case 7:if(e>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","i32"];case 13:if(e>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","u32"];case 9:if(e!==4)throw new Error("bool must be vec4");return["u32","vec4"];default:throw new Error(`Unknown data type: ${t}`)}},mt=(t,e=1)=>{let r=vi(t,e);return typeof r=="string"?r:r[0]},Pt=(t,e=1)=>{let r=vi(t,e);return typeof r=="string"?r:r[1]},ye=(...t)=>{let e=[];return t.forEach(r=>{r.length!==0&&e.push({type:12,data:r},{type:12,data:K.computeStrides(r)})}),e},ot=t=>t%4===0?4:t%2===0?2:1,Sr=(t="f32",e,r="0")=>!e||e===1?`${t}(${r})`:`vec${e}<${t}>(${r})`,Mn=(t,e,r)=>t==="f32"?r:e===1?`f32(${r})`:`vec${e}(${r})`,Dr=(t,e)=>e===4?`(${t}.x + ${t}.y + ${t}.z + ${t}.w)`:e===2?`(${t}.x + ${t}.y)`:e===3?`(${t}.x + ${t}.y + ${t}.z)`:t,Se=(t,e,r,n)=>t.startsWith("uniforms.")&&r>4?typeof e=="string"?n==="f16"?`${t}[(${e}) / 8][(${e}) % 8 / 4][(${e}) % 8 % 4]`:`${t}[(${e}) / 4][(${e}) % 4]`:n==="f16"?`${t}[${Math.floor(e/8)}][${Math.floor(e%8/4)}][${e%8%4}]`:`${t}[${Math.floor(e/4)}][${e%4}]`:r>1?`${t}[${e}]`:t,$i=(t,e,r,n,a)=>{let i=typeof r=="number",s=i?r:r.length,o=[...new Array(s).keys()],l=s<2?"u32":s<=4?`vec${s}`:`array`,u=vi(e,a),p=typeof u=="string"?u:u[1],h=typeof u=="string"?u:u[0],m={indices:l,value:p,storage:h,tensor:e},d=W=>typeof W=="string"?W:`${W}u`,_={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},w=i?"uniforms.":"",v=`${w}${t}_shape`,S=`${w}${t}_strides`,$="";for(let W=0;W ${m.indices} { + var indices: ${m.indices}; + var current = offset; + ${$} + return indices; + }`,T=W=>(_.offsetToIndices=!0,s<2?W:`o2i_${t}(${W})`),A=[];if(s>=2)for(let W=s-1;W>=0;W--)A.push(`${Se(S,W,s)} * (indices[${W}])`);let P=s<2?"":` + fn i2o_${t}(indices: ${m.indices}) -> u32 { + return ${A.join("+")}; + }`,B=W=>(_.indicesToOffset=!0,s<2?W:`i2o_${t}(${W})`),D=(...W)=>s===0?"0u":`${m.indices}(${W.map(d).join(",")})`,q=(W,ae)=>s<2?`${W}`:`${Se(W,ae,s)}`,H=(W,ae,fe)=>s<2?`${W}=${fe};`:`${Se(W,ae,s)}=${fe};`,ie={},te=(W,ae)=>{_.broadcastedIndicesToOffset=!0;let fe=`${ae.name}broadcastedIndicesTo${t}Offset`;if(fe in ie)return`${fe}(${W})`;let Ce=[];for(let Be=s-1;Be>=0;Be--){let Ye=ae.indicesGet("outputIndices",Be+ae.rank-s);Ce.push(`${q(S,Be)} * (${Ye} % ${q(v,Be)})`)}return ie[fe]=`fn ${fe}(outputIndices: ${ae.type.indices}) -> u32 { + return ${Ce.length>0?Ce.join("+"):"0u"}; + }`,`${fe}(${W})`},de=(W,ae)=>(()=>{if(m.storage===m.value)return`${t}[${W}]=${ae};`;if(m.storage==="vec2"&&m.value==="i32")return`${t}[${W}]=vec2(u32(${ae}), select(0u, 0xFFFFFFFFu, ${ae} < 0));`;if(m.storage==="vec2"&&m.value==="u32")return`${t}[${W}]=vec2(u32(${ae}), 0u);`;if(m.storage==="u32"&&m.value==="vec4")return`${t}[${W}]=dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(${ae}));`;throw new Error(`not supported combination of storage type ${m.storage} and value type ${m.value} yet`)})(),se=W=>(()=>{if(m.storage===m.value)return`${t}[${W}]`;if(m.storage==="vec2"&&m.value==="i32")return`i32(${t}[${W}].x)`;if(m.storage==="vec2"&&m.value==="u32")return`u32(${t}[${W}].x)`;if(m.storage==="u32"&&m.value==="vec4")return`vec4(bool(${t}[${W}] & 0xFFu), bool(${t}[${W}] & 0xFF00u), bool(${t}[${W}] & 0xFF0000u), bool(${t}[${W}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${m.storage} and value type ${m.value} yet`)})(),M=s<2?"":` + fn get_${t}ByIndices(indices: ${m.indices}) -> ${p} { + return ${se(`i2o_${t}(indices)`)}; + }`,R=s<2?"":(()=>{let W=o.map(fe=>`d${fe}: u32`).join(", "),ae=o.map(fe=>`d${fe}`).join(", ");return` + fn get_${t}(${W}) -> ${p} { + return get_${t}ByIndices(${D(ae)}); + }`})(),G=(...W)=>{if(W.length!==s)throw new Error(`indices length must be ${s}`);let ae=W.map(d).join(",");return s===0?se("0u"):s===1?se(ae[0]):(_.get=!0,_.getByIndices=!0,_.indicesToOffset=!0,`get_${t}(${ae})`)},re=W=>s<2?se(W):(_.getByIndices=!0,_.indicesToOffset=!0,`get_${t}ByIndices(${W})`),ee=s<2?"":` + fn set_${t}ByIndices(indices: ${m.indices}, value: ${p}) { + ${de(`i2o_${t}(indices)`,"value")} + }`,ne=s<2?"":(()=>{let W=o.map(fe=>`d${fe}: u32`).join(", "),ae=o.map(fe=>`d${fe}`).join(", ");return` + fn set_${t}(${W}, value: ${p}) { + set_${t}ByIndices(${D(ae)}, value); + }`})();return{impl:()=>{let W=[],ae=!1;return _.offsetToIndices&&(W.push(E),ae=!0),_.indicesToOffset&&(W.push(P),ae=!0),_.broadcastedIndicesToOffset&&(Object.values(ie).forEach(fe=>W.push(fe)),ae=!0),_.set&&(W.push(ne),ae=!0),_.setByIndices&&(W.push(ee),ae=!0),_.get&&(W.push(R),ae=!0),_.getByIndices&&(W.push(M),ae=!0),!i&&ae&&W.unshift(`const ${v} = ${m.indices}(${r.join(",")});`,`const ${S} = ${m.indices}(${K.computeStrides(r).join(",")});`),W.join(` +`)},type:m,offsetToIndices:T,indicesToOffset:B,broadcastedIndicesToOffset:te,indices:D,indicesGet:q,indicesSet:H,set:(...W)=>{if(W.length!==s+1)throw new Error(`indices length must be ${s}`);let ae=W[s];if(typeof ae!="string")throw new Error("value must be string");let fe=W.slice(0,s).map(d).join(",");return s===0?de("0u",ae):s===1?de(fe[0],ae):(_.set=!0,_.setByIndices=!0,_.indicesToOffset=!0,`set_${t}(${fe}, ${ae})`)},setByOffset:de,setByIndices:(W,ae)=>s<2?de(W,ae):(_.setByIndices=!0,_.indicesToOffset=!0,`set_${t}ByIndices(${W}, ${ae});`),get:G,getByOffset:se,getByIndices:re,usage:n,name:t,strides:S,shape:v,rank:s}},X=(t,e,r,n=1)=>$i(t,e,r,"input",n),ge=(t,e,r,n=1)=>$i(t,e,r,"output",n),Jo=(t,e,r,n=1)=>$i(t,e,r,"internal",n),Eu=class{constructor(t,e){this.normalizedDispatchGroup=t,this.limits=e,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=Pn){let e=typeof t=="number"?t:t[0],r=typeof t=="number"?1:t[1],n=typeof t=="number"?1:t[2];if(e>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||n>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${e}, ${r}, ${n}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(e*r*n>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${e}, ${r}, ${n}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let a=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,i=a?`@builtin(global_invocation_id) global_id : vec3, + @builtin(workgroup_id) workgroup_id : vec3, + @builtin(local_invocation_id) local_id : vec3`:`@builtin(global_invocation_id) global_id : vec3, + @builtin(local_invocation_id) local_id : vec3, + @builtin(local_invocation_index) local_idx : u32, + @builtin(workgroup_id) workgroup_id : vec3, + @builtin(num_workgroups) num_workgroups : vec3`,s=a?"let global_idx = global_id.x; let local_idx = local_id.x;":`let global_idx = (workgroup_id.z * num_workgroups[0] * num_workgroups[1] + + workgroup_id.y * num_workgroups[0] + workgroup_id.x) * ${e*r*n}u + local_idx;`;return`@compute @workgroup_size(${e}, ${r}, ${n}) + fn main(${i}) { + ${s} + `}appendVariableUniforms(t){t.rank!==0&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,e){if(t.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let r=t.usage==="input"?"read":"read_write",n=t.type.storage;return`@group(0) @binding(${e}) var ${t.name}: array<${n}>;`}declareVariables(...t){return t.map(e=>this.declareVariable(e,this.variableIndex++)).join(` +`)}registerInternalVariable(t){if(t.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(e=>this.registerInternalVariable(e)),this}registerUniform(t,e,r=1){return this.uniforms.push({name:t,type:e,length:r}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let t=[];for(let{name:e,type:r,length:n}of this.uniforms)if(n&&n>4)r==="f16"?t.push(`@align(16) ${e}:array, ${Math.ceil(n/8)}>`):t.push(`${e}:array, ${Math.ceil(n/4)}>`);else{let a=n==null||n===1?r:`vec${n}<${r}>`;t.push(`${e}:${a}`)}return` + struct Uniforms { ${t.join(", ")} }; + @group(0) @binding(${this.variableIndex}) var 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=e=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(e)];return this.uniforms.map(e=>[t(e.type),e.length??1])}},vh=(t,e)=>new Eu(t,e),xa=(t,e)=>{let r=t.length,n=[];for(let a=0;a1&&s===1&&n.unshift(i)}return n}}),Cu,Fs,Tu,Au,Er,$h,xh,Ia=J(()=>{$e(),Me(),ct(),Te(),Cu=t=>{if(!t||t.length!==1)throw new Error("Transpose requires 1 input.")},Fs=(t,e)=>e&&e.length!==t?[...new Array(t).keys()].reverse():e,Tu=(t,e)=>K.sortBasedOnPerm(t,Fs(t.length,e)),Au=(t,e,r,n)=>{let a=[];a.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} { + var a: ${r.type.indices};`);for(let i=0;i{let r=t.dataType,n=t.dims.length,a=Fs(n,e),i=Tu(t.dims,a),s=ge("output",r,i.length),o=X("a",r,n),l=u=>` + ${u.registerUniform("output_size","u32").declareVariables(o,s)} + + ${Au(a,n,o,s)} + + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${s.offsetToIndices("global_idx")}; + let aIndices = perm(indices); + + ${s.setByOffset("global_idx",o.getByIndices("aIndices"))} + }`;return{name:"Transpose",shaderCache:{hint:`${e}`,inputDependencies:["rank"]},getRunData:u=>{let p=K.size(i);return{outputs:[{dims:i,dataType:u[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:[{type:12,data:p},...ye(u[0].dims,i)]}},getShaderSource:l}},$h=(t,e)=>{Cu(t.inputs),t.compute(Er(t.inputs[0],e.perm))},xh=t=>Ke({perm:t.perm})}),Iu,Mu,Ou,zu,Pu,Ru,Bu,Du,Nu,Fu,tr,Sh,kh,Eh,Ch,Th,Ah,Ih,Mh,Oh,zh,vy=J(()=>{$e(),Me(),Te(),el(),Ia(),Iu={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"},Mu={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"},Ou={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},zu={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Pu=(t,e)=>{let r=[];for(let n=e-t;n{let r=[],n=t.length;for(let i=0;it[i]);return[r,a]},Bu=(t,e)=>{let r=t.length+e.length,n=[],a=0;for(let i=0;i{for(let r=0;r{let r=[];if(!Du(t,e)){for(let n=0;nr.push(n))}return r},Fu=(t,e,r,n,a,i,s)=>{let o=r[0].dims,l=K.size(i),u=K.size(s),p=X("_A",r[0].dataType,o),h=ge("output",a,i),m=32,d=` + var aBestValues : array; + `;return{name:t,shaderCache:e,getShaderSource:_=>` + ${_.registerUniform("reduceSize","u32").declareVariables(p,h)} + ${d} + fn DIV_CEIL(a : u32, b : u32) -> u32 { + return ((a - 1u) / b + 1u); + } + ${_.mainStart(m)} + + let outputIndex = global_idx / ${m}; + let offset = outputIndex * uniforms.reduceSize; + + var bestValue = f32(${Ou[n]}); + let Length = uniforms.reduceSize; + for (var k = local_idx; k < Length; k = k + ${m}) { + let candidate = f32(${p.getByOffset("offset + k")}); + bestValue = ${Iu[n]}; + } + aBestValues[local_idx] = bestValue; + workgroupBarrier(); + + var reduceSize = min(Length, ${m}u); + for (var currentSize = reduceSize / 2u; reduceSize > 1u; + currentSize = reduceSize / 2u) { + let interval = DIV_CEIL(reduceSize, 2u); + if (local_idx < currentSize) { + let candidate = aBestValues[local_idx + interval]; + bestValue = ${Mu[n]}; + aBestValues[local_idx] = bestValue; + } + reduceSize = interval; + workgroupBarrier(); + } + + if (local_idx == 0u) { + ${h.setByOffset("outputIndex",`${n==="mean"?`${h.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${h.type.storage}(${zu[n]})`}`)}; + } + }`,getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:l},programUniforms:[{type:12,data:u}]})}},tr=(t,e,r,n)=>{let a=t.inputs.length===1?r:xo(t.inputs,r),i=a.axes;i.length===0&&!a.noopWithEmptyAxes&&(i=t.inputs[0].dims.map((d,_)=>_));let s=K.normalizeAxes(i,t.inputs[0].dims.length),o=s,l=t.inputs[0],u=Nu(o,t.inputs[0].dims.length);u.length>0&&(l=t.compute(Er(t.inputs[0],u),{inputs:[0],outputs:[-1]})[0],o=Pu(o.length,l.dims.length));let[p,h]=Ru(l.dims,o),m=p;a.keepDims&&(m=Bu(p,s)),t.compute(Fu(e,{hint:a.cacheKey,inputDependencies:["type"]},[l],n,t.inputs[0].dataType,m,h),{inputs:[l]})},Sh=(t,e)=>{tr(t,"ReduceMeanShared",e,"mean")},kh=(t,e)=>{tr(t,"ReduceL1Shared",e,"l1")},Eh=(t,e)=>{tr(t,"ReduceL2Shared",e,"l2")},Ch=(t,e)=>{tr(t,"ReduceLogSumExpShared",e,"logSumExp")},Th=(t,e)=>{tr(t,"ReduceMaxShared",e,"max")},Ah=(t,e)=>{tr(t,"ReduceMinShared",e,"min")},Ih=(t,e)=>{tr(t,"ReduceProdShared",e,"prod")},Mh=(t,e)=>{tr(t,"ReduceSumShared",e,"sum")},Oh=(t,e)=>{tr(t,"ReduceSumSquareShared",e,"sumSquare")},zh=(t,e)=>{tr(t,"ReduceLogSumShared",e,"logSum")}}),rr,Lu,ji,xo,nr,Uu,Wu,Vu,Gu,Hu,ju,qu,Ku,Yu,Xu,ar,Ph,Rh,Bh,Dh,Nh,Fh,Lh,Uh,Wh,Vh,el=J(()=>{$e(),Me(),ct(),Te(),vy(),rr=t=>{if(!t||t.length===0||t.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(t.length===2&&t[1].dims.length!==1)throw new Error("Invalid axes input dims.")},Lu=t=>["","",`var value = ${t.getByIndices("input_indices")};`,""],ji=(t,e,r,n,a,i,s=!1,o=!1)=>{let l=[],u=r[0].dims,p=u.length,h=K.normalizeAxes(a,p),m=!o&&h.length===0;u.forEach((w,v)=>{m||h.indexOf(v)>=0?s&&l.push(1):l.push(w)});let d=l.length,_=K.size(l);return{name:t,shaderCache:e,getShaderSource:w=>{let v=[],S=X("_A",r[0].dataType,p),$=ge("output",i,d),E=n(S,$,h),T=E[2];for(let A=0,P=0;A=0?(s&&P++,T=`for(var j${A}: u32 = 0; j${A} < ${u[A]}; j${A}++) { + ${E[2].includes("last_index")?`let last_index = j${A};`:""} + ${S.indicesSet("input_indices",A,`j${A}`)} + ${T} + }`):(v.push(`${S.indicesSet("input_indices",A,$.indicesGet("output_indices",P))};`),P++);return` + + ${w.registerUniform("output_size","u32").declareVariables(S,$)} + + ${w.mainStart()} + ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + var input_indices: ${S.type.indices}; + let output_indices = ${$.offsetToIndices("global_idx")}; + + ${v.join(` +`)} + ${E[0]} // init ops for reduce max/min + ${E[1]} + ${T} + ${E[3]} + ${E.length===4?$.setByOffset("global_idx","value"):E.slice(4).join(` +`)} + }`},getRunData:()=>({outputs:[{dims:l,dataType:i}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:[{type:12,data:_},...ye(u,l)]})}},xo=(t,e)=>{let r=[];return t[1].dims[0]>0&&t[1].getBigInt64Array().forEach(n=>r.push(Number(n))),Ke({axes:r,keepDims:e.keepDims,noopWithEmptyAxes:e.noopWithEmptyAxes})},nr=(t,e,r,n)=>{let a=t.inputs,i=a.length===1?r:xo(a,r);t.compute(ji(e,{hint:i.cacheKey,inputDependencies:["rank"]},[a[0]],i.noopWithEmptyAxes&&i.axes.length===0?Lu:n,i.axes,a[0].dataType,i.keepDims,i.noopWithEmptyAxes),{inputs:[0]})},Uu=(t,e)=>{rr(t.inputs),nr(t,"ReduceLogSum",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},Wu=(t,e)=>{rr(t.inputs),nr(t,"ReduceL1",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},Vu=(t,e)=>{rr(t.inputs),nr(t,"ReduceL2",e,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Gu=(t,e)=>{rr(t.inputs),nr(t,"ReduceLogSumExp",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},Hu=(t,e)=>{rr(t.inputs),nr(t,"ReduceMax",e,(r,n,a)=>{let i=[];for(let s=0;s=0||a.length===0)&&i.push(r.indicesSet("input_indices",s,0));return[`${i.join(` +`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},ju=(t,e)=>{rr(t.inputs),nr(t,"ReduceMean",e,(r,n,a)=>{let i=1;for(let s=0;s=0||a.length===0)&&(i*=t.inputs[0].dims[s]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${n.type.value}(sum / ${i});`]})},qu=(t,e)=>{rr(t.inputs),nr(t,"ReduceMin",e,(r,n,a)=>{let i=[];for(let s=0;s=0||a.length===0)&&i.push(`input_indices[${s}] = 0;`);return[`${i.join(` +`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},Ku=(t,e)=>{rr(t.inputs),nr(t,"ReduceProd",e,(r,n)=>[`var value = ${n.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},Yu=(t,e)=>{rr(t.inputs),nr(t,"ReduceSum",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},Xu=(t,e)=>{rr(t.inputs),nr(t,"ReduceSumSquare",e,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},ar=(t,e,r)=>{if(e.length===0)return r;let n=1,a=1;for(let i=0;i1024},Ph=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?ju(t,e):Sh(t,e)},Rh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Wu(t,e):kh(t,e)},Bh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Vu(t,e):Eh(t,e)},Dh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Gu(t,e):Ch(t,e)},Nh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Hu(t,e):Th(t,e)},Fh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?qu(t,e):Ah(t,e)},Lh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Ku(t,e):Ih(t,e)},Uh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Yu(t,e):Mh(t,e)},Wh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Xu(t,e):Oh(t,e)},Vh=(t,e)=>{ar(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Uu(t,e):zh(t,e)}}),Ls,Gh,Hh,So,$y=J(()=>{$e(),ct(),el(),Ls=t=>{if(!t||t.length===0||t.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(t[0].dataType!==1)throw new Error("Invalid input type.")},Gh=(t,e)=>{Ls(t.inputs);let r=(n,a,i)=>{let s=[];for(let o=0;o=0||i.length===0)&&s.push(`input_indices[${o}] = 0;`);return[`${s.join(` +`)}`,`var value = ${n.getByIndices("input_indices")}; +var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${e.selectLastIndex>0?"<=":"<"} value) { + value = ${n.getByIndices("input_indices")}; + best_index = i32(last_index); + }`,"",a.setByOffset("global_idx","best_index")]};t.compute(ji("ArgMin",{hint:e.cacheKey,inputDependencies:["rank"]},[t.inputs[0]],r,[e.axis],7,e.keepDims),{inputs:[0]})},Hh=(t,e)=>{Ls(t.inputs);let r=(n,a,i)=>{let s=[];for(let o=0;o=0||i.length===0)&&s.push(`input_indices[${o}] = 0;`);return[`${s.join(` +`)}`,`var value = ${n.getByIndices("input_indices")}; +var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${e.selectLastIndex>0?">=":">"} value) { + value = ${n.getByIndices("input_indices")}; + best_index = i32(last_index); + }`,"",a.setByOffset("global_idx","best_index")]};t.compute(ji("argMax",{hint:e.cacheKey,inputDependencies:["rank"]},[t.inputs[0]],r,[e.axis],7,e.keepDims),{inputs:[0]})},So=t=>Ke(t)}),Qu,Zu,Ju,qi,jh,qh,Kh=J(()=>{$e(),Me(),ct(),Te(),Qu=(t,e)=>{if(!t||t.length<1)throw new Error("too few inputs");let r=0,n=t[r],a=n.dataType,i=n.dims.length;t.forEach((s,o)=>{if(o!==r){if(s.dataType!==a)throw new Error("input tensors should be one type");if(s.dims.length!==i)throw new Error("input tensors should have the same shape");s.dims.forEach((l,u)=>{if(u!==e&&l!==n.dims[u])throw new Error("non concat dimensions must match")})}})},Zu=(t,e)=>` + fn calculateInputIndex(index: u32) -> u32 { + let sizeInConcatAxis = array(${e}); + for (var i: u32 = 0u; i < ${t}; i += 1u ) { + if (index < sizeInConcatAxis[i]) { + return i; + } + } + return ${t}u; + }`,Ju=(t,e)=>{let r=t.length,n=[];for(let a=0;a{let a=K.size(r),i=new Array(t.length),s=new Array(t.length),o=0,l=[],u=[],p=[{type:12,data:a}];for(let w=0;w`uniforms.sizeInConcatAxis${w}`).join(","),_=w=>` + + ${(()=>{w.registerUniform("outputSize","u32");for(let v=0;v(${d}); + ${m} -= sizeInConcatAxis[inputIndex - 1u]; + } + + ${Ju(s,h)} + }`;return{name:"Concat",shaderCache:{hint:`${e}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p}),getShaderSource:_}},jh=(t,e)=>{let r=t.inputs,n=r[0].dims,a=K.normalizeAxis(e.axis,n.length);Qu(r,a);let i=n.slice();i[a]=r.reduce((o,l)=>o+(l.dims.length>a?l.dims[a]:0),0);let s=r.filter(o=>K.size(o.dims)>0);t.compute(qi(s,a,i,r[0].dataType),{inputs:s})},qh=t=>Ke({axis:t.axis})}),ed,td,rd,nd,Ki,ad,Yh,Xh=J(()=>{$e(),Xo(),Te(),Kh(),ed=(t,e)=>{let r=t[0],n=t[1],a=t[2],i=t[3],s=t[4],o=t[5];if(s&&o)throw new Error("Attention cannot have both past and relative_position_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let l=r.dims[0],u=r.dims[1],p=r.dims[2];if(a.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(n.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(n.dims[0]!==p)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(a.dims[0]!==n.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let h=a.dims[0]/3,m=h,d=m;if(e.qkvHiddenSizes.length>0){if(e.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let E of e.qkvHiddenSizes)if(E%e.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");h=e.qkvHiddenSizes[0],m=e.qkvHiddenSizes[1],d=e.qkvHiddenSizes[2]}let _=u;if(h!==m)throw new Error("qkv_hidden_sizes first element should be same as the second");if(a.dims[0]!==h+m+d)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let w=0;if(s){if(m!==d)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(s.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(s.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(s.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(s.dims[2]!==e.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(s.dims[4]!==m/e.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');e.pastPresentShareBuffer||(w=s.dims[3])}let v=_+w,S=-1,$=0;if(i)throw new Error("Mask not supported");if(s)throw new Error("past is not supported");return{batchSize:l,sequenceLength:u,pastSequenceLength:w,kvSequenceLength:_,totalSequenceLength:v,maxSequenceLength:S,inputHiddenSize:p,hiddenSize:h,vHiddenSize:d,headSize:Math.floor(h/e.numHeads),vHeadSize:Math.floor(d/e.numHeads),numHeads:e.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:e.maskFilterValue,maskType:$,scale:e.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},td=(t,e,r,n)=>{let a=ot(n),i=64,s=n/a;s{let d=ge("x",e.dataType,e.dims,a),_=[{name:"d_inv",type:Pt(e.dataType)},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return` + var thread_max: array; + var thread_sum: array; + ${m.registerUniforms(_).declareVariables(d)} + ${m.mainStart([i,1,1])} + let local_offset = local_idx * uniforms.elements_per_thread; + let offset = workgroup_id.x * uniforms.d_comp + local_offset; + + var thread_max_vector = ${p}(-3.402823e+38f); + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + thread_max_vector = max(${p}(x[offset + i]), thread_max_vector); + } + thread_max[local_idx] = ${(()=>{switch(a){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${a}`)}})()}; + workgroupBarrier(); + + var max_value = f32(-3.402823e+38f); + for (var i = 0u; i < ${i}; i++) { + max_value = max(thread_max[i], max_value); + } + + var sum_vector = ${p}(0); + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + sum_vector += exp(${p}(x[offset + i]) - max_value); + } + thread_sum[local_idx] = ${(()=>{switch(a){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${a}`)}})()}; + workgroupBarrier(); + + var sum: f32 = 0; + for (var i = 0u; i < ${i}; i++) { + sum += thread_sum[i]; + } + + if (sum == 0) { + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + x[offset + i] = ${d.type.value}(uniforms.d_inv); + } + } else { + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + var f32input = ${p}(x[offset + i]); + x[offset + i] = ${d.type.value}(exp(f32input - max_value) / sum); + } + } + }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${i};${u};${a}`},getShaderSource:h,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:l})}},rd=(t,e,r,n,a,i,s)=>{let o=s+a.kvSequenceLength,l=[a.batchSize,a.numHeads,a.sequenceLength,o],u=i.scale===0?1/Math.sqrt(a.headSize):i.scale,p=ot(a.headSize),h=a.headSize/p,m=12,d={x:Math.ceil(o/m),y:Math.ceil(a.sequenceLength/m),z:a.batchSize*a.numHeads},_=[{type:12,data:a.sequenceLength},{type:12,data:h},{type:12,data:o},{type:12,data:a.numHeads},{type:1,data:u}],w=n?["type","type","type"]:["type","type"],v=S=>{let $=X("q",e.dataType,e.dims,p),E=X("key",r.dataType,r.dims,p),T=[$,E];n&&T.push(X("relative_position_bias",n.dataType,n.dims));let A=ge("output",e.dataType,l),P=Pt(1,p),B=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"alpha",type:"f32"}];return` + const TILE_SIZE = ${m}u; + + var tileQ: array<${$.type.storage}, ${m*m}>; + var tileK: array<${$.type.storage}, ${m*m}>; + ${S.registerUniforms(B).declareVariables(...T,A)} + ${S.mainStart([m,m,1])} + // x holds the N and y holds the M + let headIdx = workgroup_id.z; + let m = workgroup_id.y * TILE_SIZE; + let n = workgroup_id.x * TILE_SIZE; + let qOffset = uniforms.M * uniforms.K * headIdx + m * uniforms.K; + let kOffset = uniforms.N * uniforms.K * headIdx + n * uniforms.K; + + var value = ${P}(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]; + } + if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) { + tileK[TILE_SIZE * local_id.y + local_id.x] = key[kOffset + local_id.y * uniforms.K + w + local_id.x]; + } + workgroupBarrier(); + + for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) { + value += ${P}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]); + } + + workgroupBarrier(); + } + + let headOffset = headIdx * uniforms.M * uniforms.N; + 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(p){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: ${p}`)}})()}; + output[outputIdx] = ${A.type.value} (sum * uniforms.alpha) + ${n?"relative_position_bias[outputIdx]":"0.0"}; + } + }`};return{name:"AttentionProbs",shaderCache:{hint:`${p}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:l,dataType:e.dataType,gpuDataType:0}],dispatchGroup:d,programUniforms:_}),getShaderSource:v}},nd=(t,e,r,n,a)=>{let i=a+n.kvSequenceLength,s=[n.batchSize,n.sequenceLength,n.vHiddenSize],o=12,l={x:Math.ceil(n.vHeadSize/o),y:Math.ceil(n.sequenceLength/o),z:n.batchSize*n.numHeads},u=[{type:12,data:n.sequenceLength},{type:12,data:i},{type:12,data:n.vHeadSize},{type:12,data:n.numHeads},{type:12,data:n.vHiddenSize}];return{name:"AttentionScore",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:s,dataType:e.dataType,gpuDataType:0}],dispatchGroup:l,programUniforms:u}),getShaderSource:p=>{let h=X("probs",e.dataType,e.dims),m=X("v",r.dataType,r.dims),d=ge("output",e.dataType,s),_=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"}];return` + const TILE_SIZE = ${o}u; + var tileQ: array<${h.type.value}, ${o*o}>; + var tileK: array<${h.type.value}, ${o*o}>; + ${p.registerUniforms(_).declareVariables(h,m,d)} + ${p.mainStart([o,o,1])} + let headIdx = workgroup_id.z; + let m = global_id.y; + let n = global_id.x; + + let offsetA = headIdx * (uniforms.M * uniforms.K) + m * uniforms.K; + let offsetB = headIdx * (uniforms.N * uniforms.K) + n; + + var value = ${h.type.storage}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + tileK[TILE_SIZE * local_id.y + local_id.x] = v[offsetB + (w + local_id.y) * uniforms.N]; + } + workgroupBarrier(); + for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) { + value += tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * k + local_id.x]; + } + workgroupBarrier(); + } + + // we need to transpose output from BNSH_v to BSND_v + let batchIdx = workgroup_id.z / uniforms.num_heads; + let currentBatchHeadNumber = workgroup_id.z % uniforms.num_heads; + if (m < uniforms.M && n < uniforms.N) { + let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size + + currentBatchHeadNumber * uniforms.N + n; + output[outputIdx] = value; + } + }`}}},Ki=(t,e,r,n,a,i,s,o,l,u,p)=>{let h=t.outputCount>1,m=t.outputCount>2,d=h&&m?u.pastSequenceLength:0,_=d+u.kvSequenceLength,w=[u.batchSize,u.numHeads,_,u.headSize],v=s?[s,r]:[r],S=h?t.compute(qi(v,2,w,r.dataType),{inputs:v,outputs:[1]})[0]:r,$=[u.batchSize,u.numHeads,_,u.headSize],E=o?[o,n]:[n],T=m?t.compute(qi(E,2,$,n.dataType),{inputs:E,outputs:[2]})[0]:n,A=[e,S];l&&A.push(l);let P=t.compute(rd(t,e,S,l,u,p,d),{inputs:A,outputs:[-1]})[0];t.compute(td(t,P,u.batchSize*u.numHeads*u.sequenceLength,_),{inputs:[P],outputs:[]});let B=[P,T];t.compute(nd(t,P,T,u,d),{inputs:B,outputs:[0]})},ad=(t,e)=>{let r=[e.batchSize,e.numHeads,e.sequenceLength,e.headSize],n=e.sequenceLength,a=e.inputHiddenSize,i=e.headSize,s=12,o={x:Math.ceil(e.headSize/s),y:Math.ceil(e.sequenceLength/s),z:e.batchSize*e.numHeads},l=[t.inputs[0],t.inputs[1],t.inputs[2]],u=[{type:12,data:n},{type:12,data:a},{type:12,data:i},{type:12,data:e.numHeads},{type:12,data:e.headSize},{type:12,data:e.hiddenSize},{type:12,data:e.hiddenSize+e.hiddenSize+e.vHiddenSize}],p=h=>{let m=ge("output_q",l[0].dataType,r),d=ge("output_k",l[0].dataType,r),_=ge("output_v",l[0].dataType,r),w=X("input",l[0].dataType,l[0].dims),v=X("weight",l[1].dataType,l[1].dims),S=X("bias",l[2].dataType,l[2].dims),$=w.type.storage,E=[{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 = ${s}u; + var tileInput: array<${$}, ${s*s}>; + var tileWeightQ: array<${$}, ${s*s}>; + var tileWeightK: array<${$}, ${s*s}>; + var tileWeightV: array<${$}, ${s*s}>; + ${h.registerUniforms(E).declareVariables(w,v,S,m,d,_)} + ${h.mainStart([s,s,1])} + let batchIndex = workgroup_id.z / uniforms.num_heads; + let headNumber = workgroup_id.z % uniforms.num_heads; + let m = global_id.y; + let n = global_id.x; + + let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K; + let biasOffsetQ = headNumber * uniforms.head_size; + let biasOffsetK = uniforms.hidden_size + biasOffsetQ; + let biasOffsetV = uniforms.hidden_size + biasOffsetK; + + var valueQ = ${$}(0); + var valueK = ${$}(0); + var valueV = ${$}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + let offset = n + (w + local_id.y) * uniforms.ldb; + tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset]; + tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset]; + tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset]; + } + workgroupBarrier(); + for (var k: u32 = 0u; k({outputs:[{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0}],dispatchGroup:o,programUniforms:u}),getShaderSource:p},{inputs:l,outputs:[-1,-1,-1]})},Yh=(t,e)=>{let r=ed(t.inputs,e),[n,a,i]=ad(t,r);return Ki(t,n,a,i,t.inputs[4],void 0,void 0,void 0,t.inputs[5],r,e)}}),id,sd,od,Qh,xy=J(()=>{dr(),$e(),Me(),ct(),Te(),id=(t,e)=>{if(!t||t.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(n,a,i)=>{let s=a.length;if(s!==n.length)throw new Error(`${i}: num dimensions != ${s}`);a.forEach((o,l)=>{if(o!==n[l])throw new Error(`${i}: dim[${l}] do not match`)})};if(t[0].dims.length>1){let n=e.format==="NHWC"?e.spatial?t[0].dims.slice(-1):t[0].dims.slice(-1).concat(t[0].dims.slice(1,t[0].dims.length-1)):t[0].dims.slice(1,e.spatial?2:void 0);r(t[1].dims,n,"Invalid input scale"),r(t[2].dims,n,"Invalid input B"),r(t[3].dims,n,"Invalid input mean"),r(t[4].dims,n,"Invalid input var")}else r(t[1].dims,[1],"Invalid input scale"),r(t[2].dims,[1],"Invalid input B"),r(t[3].dims,[1],"Invalid input mean"),r(t[4].dims,[1],"Invalid input var")},sd=(t,e)=>{let{epsilon:r,spatial:n,format:a}=e,i=t[0].dims,s=n?ot(i[i.length-1]):1,o=a==="NHWC"&&i.length>1?s:1,l=K.size(i)/s,u=n,p=u?i.length:i,h=X("x",t[0].dataType,t[0].dims,s),m=X("scale",t[1].dataType,t[1].dims,o),d=X("bias",t[2].dataType,t[2].dims,o),_=X("inputMean",t[3].dataType,t[3].dims,o),w=X("inputVar",t[4].dataType,t[4].dims,o),v=ge("y",t[0].dataType,p,s),S=()=>{let E="";if(n)E=`let cOffset = ${i.length===1?"0u":a==="NHWC"?`outputIndices[${i.length-1}] / ${s}`:"outputIndices[1]"};`;else if(a==="NCHW")E=` + ${v.indicesSet("outputIndices","0","0")} + let cOffset = ${v.indicesToOffset("outputIndices")};`;else{E=`var cIndices = ${m.type.indices}(0); + cIndices[0] = outputIndices[${i.length-1}];`;for(let T=1;T` + const epsilon = ${r}; + ${E.registerUniform("outputSize","u32").declareVariables(h,m,d,_,w,v)} + ${E.mainStart()} + ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var outputIndices = ${v.offsetToIndices(`global_idx * ${s}`)}; + ${S()} + let scale = ${m.getByOffset("cOffset")}; + let bias = ${d.getByOffset("cOffset")}; + let inputMean = ${_.getByOffset("cOffset")}; + let inputVar = ${w.getByOffset("cOffset")}; + let x = ${h.getByOffset("global_idx")}; + let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias; + ${v.setByOffset("global_idx","value")} + }`;return{name:"BatchNormalization",shaderCache:{hint:`${e.epsilon}_${e.format}_${n}_${s}`,inputDependencies:u?["rank","type","type","type","type"]:void 0},getShaderSource:$,getRunData:()=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u?[{type:12,data:l},...ye(i)]:[{type:12,data:l}]})}},od=t=>Ke(t),Qh=(t,e)=>{let{inputs:r,outputCount:n}=t,a=od({...e,outputCount:n});if(Ue.webgpu.validateInputContent&&id(r,a),e.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");t.compute(sd(r,a))}}),ld,ud,Zh,Sy=J(()=>{Me(),Te(),ld=t=>{if(t[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(t[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(t[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(t[0].dims[2]!==t[1].dims[0])throw new Error("last dimension of input and bias are not the same")},ud=t=>{let e=t[0].dims,r=t[0].dims[2],n=K.size(e)/4,a=t[0].dataType,i=X("input",a,e,4),s=X("bias",a,[r],4),o=X("residual",a,e,4),l=ge("output",a,e,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:e,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)}}),getShaderSource:u=>` + const channels = ${r}u / 4; + ${u.declareVariables(i,s,o,l)} + + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes(n)} + let value = ${i.getByOffset("global_idx")} + + ${s.getByOffset("global_idx % channels")} + ${o.getByOffset("global_idx")}; + ${l.setByOffset("global_idx","value")} + }`}},Zh=t=>{ld(t.inputs),t.compute(ud(t.inputs))}}),dd,Re,Jh,ef,tf,rf,nf,af,sf,of,lf,cd,uf,df,cf,pf,zi,hf,Pi,ff,mf,gf,_f,yf,wf,bf,vf,$f,xf,Sf,kf,Ef,Cf,Tf,Af,Us,If,ko,Eo,Mf,Of,zf,tl=J(()=>{$e(),Me(),ct(),Te(),dd=(t,e,r,n,a,i)=>{let s=Math.ceil(e/4),o="";typeof a=="string"?o=`${a}(a)`:o=a("a");let l=X("inputData",r,[s],4),u=ge("outputData",n,[s],4);return` + ${t.registerUniform("vec_size","u32").declareVariables(l,u)} + + ${i??""} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + + let a = ${l.getByOffset("global_idx")}; + ${u.setByOffset("global_idx",o)} + }`},Re=(t,e,r,n,a,i=t.dataType)=>({name:e,shaderCache:{hint:a,inputDependencies:["type"]},getShaderSource:s=>dd(s,K.size(t.dims),t.dataType,i,r,n),getRunData:s=>({outputs:[{dims:t.dims,dataType:i}],dispatchGroup:{x:Math.ceil(K.size(s[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(K.size(t.dims)/4)}]})}),Jh=t=>{t.compute(Re(t.inputs[0],"Abs","abs"))},ef=t=>{t.compute(Re(t.inputs[0],"Acos","acos"))},tf=t=>{t.compute(Re(t.inputs[0],"Acosh","acosh"))},rf=t=>{t.compute(Re(t.inputs[0],"Asin","asin"))},nf=t=>{t.compute(Re(t.inputs[0],"Asinh","asinh"))},af=t=>{t.compute(Re(t.inputs[0],"Atan","atan"))},sf=t=>{t.compute(Re(t.inputs[0],"Atanh","atanh"))},of=t=>Ke(t),lf=(t,e)=>{let r;switch(e.to){case 10:r="vec4";break;case 1:r="vec4";break;case 12:r="vec4";break;case 6:r="vec4";break;case 9:r="vec4";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${e.to}`)}t.compute(Re(t.inputs[0],"Cast",r,void 0,e.cacheKey,e.to))},cd=t=>{let e=t.length>=2&&t[1].data!==0?t[1].getFloat32Array()[0]:Qo,r=t.length>=3&&t[2].data!==0?t[2].getFloat32Array()[0]:Zo;return Ke({min:e,max:r})},uf=(t,e)=>{let r=t.inputs.length===1?e:cd(t.inputs),n=Pt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Clip",a=>`clamp(${a}, 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]})},df=t=>{t.compute(Re(t.inputs[0],"Ceil","ceil"))},cf=t=>{t.compute(Re(t.inputs[0],"Cos","cos"))},pf=t=>{t.compute(Re(t.inputs[0],"Cosh","cosh"))},zi=t=>Ke(t),hf=(t,e)=>{let r=Pt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Elu",n=>`elu_vf32(${n})`,` + const elu_alpha_ = ${r}(${e.alpha}); + + fn elu_f32(a: ${r}) -> ${r} { + return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0); + } + + fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> { + return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w)); + }`,e.cacheKey))},Pi=(t="f32")=>` +const r0: ${t} = 0.3275911; +const r1: ${t} = 0.254829592; +const r2: ${t} = -0.284496736; +const r3: ${t} = 1.421413741; +const r4: ${t} = -1.453152027; +const r5: ${t} = 1.061405429; + +fn erf_vf32(v: vec4<${t}>) -> vec4<${t}> { + let absv = abs(v); + let x = 1.0 / (1.0 + r0 * absv); + return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv)); +}`,ff=t=>{let e=Pt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Erf",r=>`erf_vf32(${r})`,Pi(e)))},mf=t=>{t.compute(Re(t.inputs[0],"Exp","exp"))},gf=t=>{t.compute(Re(t.inputs[0],"Floor","floor"))},_f=t=>{let e=Pt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Pi(e)))},yf=(t,e)=>{let r=Pt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"LeakyRelu",n=>`select(leaky_relu_alpha_ * ${n}, ${n}, ${n} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${e.alpha});`,e.cacheKey))},wf=t=>{t.compute(Re(t.inputs[0],"Not",e=>`!${e}`))},bf=t=>{t.compute(Re(t.inputs[0],"Neg",e=>`-${e}`))},vf=t=>{t.compute(Re(t.inputs[0],"Reciprocal",e=>`1.0/${e}`))},$f=t=>{let e=Pt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"Relu",r=>`select(vec4<${e}>(0.0), ${r}, ${r} > vec4<${e}>(0.0))`))},xf=t=>{t.compute(Re(t.inputs[0],"Sigmoid",e=>`(1.0 / (1.0 + exp(-${e})))`))},Sf=t=>Ke(t),kf=(t,e)=>{let r=Pt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${e.alpha} * ${n} + vec4<${r}>(${e.beta})))`,void 0,e.cacheKey))},Ef=t=>{t.compute(Re(t.inputs[0],"Sin","sin"))},Cf=t=>{t.compute(Re(t.inputs[0],"Sinh","sinh"))},Tf=t=>{t.compute(Re(t.inputs[0],"Sqrt","sqrt"))},Af=t=>{t.compute(Re(t.inputs[0],"Tan","tan"))},Us=t=>`sign(${t}) * (1 - exp(-2 * abs(${t}))) / (1 + exp(-2 * abs(${t})))`,If=t=>{t.compute(Re(t.inputs[0],"Tanh",Us))},ko=(t="f32")=>` +const fast_gelu_a: ${t} = 0.5; +const fast_gelu_b: ${t} = 0.7978845608028654; +const fast_gelu_c: ${t} = 0.035677408136300125; + +fn tanh_v(v: vec4<${t}>) -> vec4<${t}> { + return ${Us("v")}; +} +`,Eo=t=>`(fast_gelu_a + fast_gelu_a * tanh_v(${t} * (fast_gelu_c * ${t} * ${t} + fast_gelu_b))) * ${t}`,Mf=t=>{let e=Pt(t.inputs[0].dataType);t.compute(Re(t.inputs[0],"FastGelu",Eo,ko(e),void 0,t.inputs[0].dataType))},Of=(t,e)=>{let r=Pt(t.inputs[0].dataType);return t.compute(Re(t.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${e.alpha});`,e.cacheKey)),0},zf=t=>{t.compute(Re(t.inputs[0],"Log","log"))}}),pd,hd,Pf,ky=J(()=>{Me(),Te(),tl(),pd=t=>{if(t[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(t[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(t[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(t[0].dims[2]!==t[1].dims[0])throw new Error("last dimension of input and bias are not the same")},hd=t=>{let e=t[0].dims.slice();e[2]=e[2]/2;let r=X("input",t[0].dataType,t[0].dims,4),n=X("bias",t[0].dataType,[t[0].dims[2]],4),a=ge("output",t[0].dataType,e,4),i=K.size(e)/4,s=mt(t[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:e,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:o=>` + const M_SQRT2 = sqrt(2.0); + const halfChannels = ${t[0].dims[2]/4/2}u; + + ${o.declareVariables(r,n,a)} + + ${Pi(s)} + + ${o.mainStart()} + ${o.guardAgainstOutOfBoundsWorkgroupSizes(i)} + let biasIdx = global_idx % halfChannels; + let batchIndex = global_idx / halfChannels; + let inputOffset = biasIdx + batchIndex * halfChannels * 2; + let valueLeft = input[inputOffset] + bias[biasIdx]; + let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels]; + let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1); + + ${a.setByOffset("global_idx","valueLeft * geluRight")} + }`}},Pf=t=>{pd(t.inputs),t.compute(hd(t.inputs))}}),fd,md,ir,Rf,Bf,Df,Nf,Ff,Lf,Uf,Wf,Vf,Gf,Ey=J(()=>{$e(),Me(),Te(),fd=(t,e,r,n,a,i,s,o,l,u,p,h)=>{let m,d;typeof o=="string"?m=d=($,E)=>`${o}((${$}),(${E}))`:typeof o=="function"?m=d=o:(m=o.scalar,d=o.vector);let _=ge("outputData",p,n.length,4),w=X("aData",l,e.length,4),v=X("bData",u,r.length,4),S;if(a)if(i){let $=K.size(e)===1,E=K.size(r)===1,T=e.length>0&&e[e.length-1]%4===0,A=r.length>0&&r[r.length-1]%4===0;$||E?S=_.setByOffset("global_idx",d($?`${w.type.value}(${w.getByOffset("0")}.x)`:w.getByOffset("global_idx"),E?`${v.type.value}(${v.getByOffset("0")}.x)`:v.getByOffset("global_idx"))):S=` + let outputIndices = ${_.offsetToIndices("global_idx * 4u")}; + let offsetA = ${w.broadcastedIndicesToOffset("outputIndices",_)}; + let offsetB = ${v.broadcastedIndicesToOffset("outputIndices",_)}; + ${_.setByOffset("global_idx",d(s||T?w.getByOffset("offsetA / 4u"):`${w.type.value}(${w.getByOffset("offsetA / 4u")}[offsetA % 4u])`,s||A?v.getByOffset("offsetB / 4u"):`${v.type.value}(${v.getByOffset("offsetB / 4u")}[offsetB % 4u])`))} + `}else S=_.setByOffset("global_idx",d(w.getByOffset("global_idx"),v.getByOffset("global_idx")));else{if(!i)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let $=(E,T,A="")=>{let P=`aData[indexA${T}][componentA${T}]`,B=`bData[indexB${T}][componentB${T}]`;return` + let outputIndices${T} = ${_.offsetToIndices(`global_idx * 4u + ${T}u`)}; + let offsetA${T} = ${w.broadcastedIndicesToOffset(`outputIndices${T}`,_)}; + let offsetB${T} = ${v.broadcastedIndicesToOffset(`outputIndices${T}`,_)}; + let indexA${T} = offsetA${T} / 4u; + let indexB${T} = offsetB${T} / 4u; + let componentA${T} = offsetA${T} % 4u; + let componentB${T} = offsetB${T} % 4u; + ${E}[${T}] = ${A}(${m(P,B)}); + `};p===9?S=` + var data = vec4(0); + ${$("data",0,"u32")} + ${$("data",1,"u32")} + ${$("data",2,"u32")} + ${$("data",3,"u32")} + outputData[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:S=` + ${$("outputData[global_idx]",0)} + ${$("outputData[global_idx]",1)} + ${$("outputData[global_idx]",2)} + ${$("outputData[global_idx]",3)} + `}return` + ${t.registerUniform("vec_size","u32").declareVariables(w,v,_)} + + ${h??""} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${S} + }`},md=(t,e,r,n,a,i,s=r.dataType)=>{let o=!K.areEqual(r.dims,n.dims),l=r.dims,u=K.size(r.dims),p=!1,h=!1,m=[o];if(o){let d=zn.calcShape(r.dims,n.dims,!1);if(!d)throw new Error("Can't perform binary op on the given tensors");l=d,u=K.size(l);let _=K.size(r.dims)===1,w=K.size(n.dims)===1,v=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,S=n.dims.length>0&&n.dims[n.dims.length-1]%4===0;m.push(_),m.push(w),m.push(v),m.push(S);let $=1;for(let E=1;Ed.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:d=>fd(d,r.dims,n.dims,l,p,o,h,a,r.dataType,n.dataType,s,i),getRunData:()=>({outputs:[{dims:l,dataType:s}],dispatchGroup:{x:Math.ceil(u/64/4)},programUniforms:[{type:12,data:Math.ceil(K.size(l)/4)},...ye(r.dims,n.dims,l)]})}},ir=(t,e,r,n,a,i)=>{t.compute(md(e,a??"",t.inputs[0],t.inputs[1],r,n,i))},Rf=t=>{ir(t,"Add",(e,r)=>`${e}+${r}`)},Bf=t=>{ir(t,"Div",(e,r)=>`${e}/${r}`)},Df=t=>{ir(t,"Equal",{scalar:(e,r)=>`u32(${e}==${r})`,vector:(e,r)=>`vec4(${e}==${r})`},void 0,void 0,9)},Nf=t=>{ir(t,"Mul",(e,r)=>`${e}*${r}`)},Ff=t=>{let e=X("input",t.inputs[0].dataType,t.inputs[0].dims).type.value;ir(t,"Pow",{scalar:(r,n)=>`pow_custom(${r},${n})`,vector:(r,n)=>`pow_vector_custom(${r},${n})`},` + fn pow_custom(a : ${e}, b : ${e}) -> ${e} { + if (b == ${e}(0.0)) { + return ${e}(1.0); + } else if (a < ${e}(0.0) && f32(b) != floor(f32(b))) { + return ${e}(pow(f32(a), f32(b))); // NaN + } + return select(sign(a), ${e}(1.0), round(f32(abs(b) % ${e}(2.0))) != 1.0) * ${e}(${e==="i32"?"round":""}(pow(f32(abs(a)), f32(b)))); + } + fn pow_vector_custom(a : vec4<${e}>, b : vec4<${e}>) -> vec4<${e}> { + // TODO: implement vectorized pow + return vec4<${e}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w)); + } + `)},Lf=t=>{ir(t,"Sub",(e,r)=>`${e}-${r}`)},Uf=t=>{ir(t,"Greater",{scalar:(e,r)=>`u32(${e}>${r})`,vector:(e,r)=>`vec4(${e}>${r})`},void 0,void 0,9)},Wf=t=>{ir(t,"Less",{scalar:(e,r)=>`u32(${e}<${r})`,vector:(e,r)=>`vec4(${e}<${r})`},void 0,void 0,9)},Vf=t=>{ir(t,"GreaterOrEqual",{scalar:(e,r)=>`u32(${e}>=${r})`,vector:(e,r)=>`vec4(${e}>=${r})`},void 0,void 0,9)},Gf=t=>{ir(t,"LessOrEqual",{scalar:(e,r)=>`u32(${e}<=${r})`,vector:(e,r)=>`vec4(${e}<=${r})`},void 0,void 0,9)}}),sn,on,ln,rl,dn=J(()=>{$e(),Me(),sn=(t,e,r="f32")=>{switch(t.activation){case"Relu":return`value = max(value, ${e}(0.0));`;case"Sigmoid":return`value = (${e}(1.0) / (${e}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${e}(${r}(uniforms.clip_min)), ${e}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${e}(0.0), min(${e}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${e}(0.0));`;case"":return"";default:throw new Error(`Unsupported activation ${t.activation}`)}},on=(t,e)=>{t.activation==="Clip"?e.push({type:1,data:t.clipMax},{type:1,data:t.clipMin}):t.activation==="HardSigmoid"?e.push({type:1,data:t.alpha},{type:1,data:t.beta}):t.activation==="LeakyRelu"&&e.push({type:1,data:t.alpha})},ln=(t,e)=>{t.activation==="Clip"?e.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):t.activation==="HardSigmoid"?e.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):t.activation==="LeakyRelu"&&e.push({name:"alpha",type:"f32"})},rl=t=>{let e=t?.activation||"";if(e==="HardSigmoid"){let[r,n]=t?.activation_params||[.2,.5];return{activation:e,alpha:r,beta:n}}else if(e==="Clip"){let[r,n]=t?.activation_params||[Qo,Zo];return{activation:e,clipMax:n,clipMin:r}}else if(e==="LeakyRelu"){let[r]=t?.activation_params||[.01];return{activation:e,alpha:r}}return{activation:e}}}),kt,nl,al=J(()=>{kt=(t,e)=>{switch(t){case 1:return e;case 2:return`vec2<${e}>`;case 3:return`vec3<${e}>`;case 4:return`vec4<${e}>`;default:throw new Error(`${t}-component is not supported.`)}},nl=t=>` + ${t?"value = value + getBiasByOutputCoords(coords);":""} + `}),il,Hf=J(()=>{il=t=>` +fn getIndexFromCoords4D(coords : vec4, shape : vec4) -> i32 { + return dot(coords, vec4( + shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1)); +} +fn getOutputIndexFromCoords(coords : vec4) -> i32 { + return dot(coords, vec4( + i32(${t}.x), i32(${t}.y), i32(${t}.z), 1)); +} +`}),gd,_d,Yi,Ws,yd,Xi,wd,sl,as=J(()=>{$e(),Me(),Te(),dn(),al(),gd=(t,e)=>t?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart / innerElementSize + inputCol${e?", batchIndices":""}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRow + innerRow, + kStart / innerElementSize + inputCol${e?", batchIndices":""}); + `,_d=(t,e)=>t?` + let ACached0 = mm_Asub[k * innerElementSize][localRow]; + let ACached1 = mm_Asub[k * innerElementSize + 1][localRow]; + let ACached2 = mm_Asub[k * innerElementSize + 2][localRow]; + ${e===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"} + for (var i = 0; i < rowPerThread; i = i + 1) { + acc[i] = BCached0 * ACached0[i] + acc[i]; + acc[i] = BCached1 * ACached1[i] + acc[i]; + acc[i] = BCached2 * ACached2[i] + acc[i]; + ${e===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"} + }`:` + for (var i = 0; i < rowPerThread; i = i + 1) { + let ACached = mm_Asub[tileRow + i][k]; + acc[i] = BCached0 * ACached.x + acc[i]; + acc[i] = BCached1 * ACached.y + acc[i]; + acc[i] = BCached2 * ACached.z + acc[i]; + ${e===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"} + }`,Yi=(t,e,r="f32",n,a=!1,i=32,s=!1,o=32)=>{let l=e[1]*t[1],u=e[0]*t[0],p=a?l:i,h=a?i:l,m=p/e[0],d=i/e[1];if(!((a&&m===4&&t[1]===4||!a&&(m===3||m===4))&&p%e[0]===0&&i%e[1]===0&&t[0]===4))throw new Error(`If transposeA ${a} is true, innerElementSize ${m} and workPerThread[1] ${t[1]} must be 4. + Otherwise, innerElementSize ${m} must be 3 or 4. + tileAWidth ${p} must be divisible by workgroupSize[0]${e[0]}. tileInner ${i} must be divisible by workgroupSize[1] ${e[1]}. colPerThread ${t[0]} must be 4.`);return` +var mm_Asub: array, ${p/m}>, ${h}>; +var mm_Bsub: array, ${u/t[0]}>, ${i}>; + +const rowPerThread = ${t[1]}; +const colPerThread = ${t[0]}; +const innerElementSize = ${m}; +const tileInner = ${i}; + +@compute @workgroup_size(${e[0]}, ${e[1]}, ${e[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let localRow = i32(localId.y); + let tileRow = localRow * rowPerThread; + let tileCol = i32(localId.x); + + let globalRow =i32(globalId.y) * rowPerThread; + let globalCol = i32(globalId.x); + let batch = ${s?"0":"i32(globalId.z)"}; + ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""} + let globalRowStart = i32(workgroupId.y) * ${l}; + + let num_tiles = ${s?`${Math.ceil(o/i)}`:"(uniforms.dim_inner - 1) / tileInner + 1"}; + var kStart = ${s?`i32(globalId.z) * ${o}`:"0"}; + + var acc: array, rowPerThread>; + + // Loop over shared dimension. + let tileRowB = localRow * ${d}; + for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let inputRow = tileRow + innerRow; + let inputCol = tileCol; + ${gd(a,n)} + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${d}; innerRow = innerRow + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${n?", batchIndices":""}); + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + for (var k = 0; k < tileInner / innerElementSize; k = k + 1) { + let BCached0 = mm_Bsub[k * innerElementSize][tileCol]; + let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol]; + let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol]; + ${m===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"} + + ${_d(a,m)} + } + + workgroupBarrier(); + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]); + } +}`},Ws=(t,e)=>t?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart + inputCol${e?", batchIndices":""}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRowStart + inputRow, + kStart + inputCol${e?", batchIndices":""}); + `,yd=t=>t?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",Xi=(t,e,r="f32",n,a=!1,i=32,s=!1,o=32,l=!1)=>{let u=t[1]*e[1],p=t[0]*e[0],h=a?u:i,m=a?i:u;if(!(m%e[1]===0&&h%e[0]===0&&i%e[1]===0))throw new Error(`tileAHight ${m} must be divisible by workgroupSize[1]${e[1]}, tileAWidth ${h} must be divisible by workgroupSize[0]${e[0]}, tileInner ${i} must be divisible by workgroupSize[1]${e[1]}`);let d=m/e[1],_=h/e[0],w=i/e[1],v=l?` + let localRow = i32(localId.y); + let localCol = i32(localId.x); + let globalRowStart = i32(workgroupId.y) * ${u}; + let globalColStart = i32(workgroupId.x) * ${p}; + + // Loop over shared dimension. + for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var inputRow = localRow; inputRow < ${m}; inputRow = inputRow + ${e[1]}) { + for (var inputCol = localCol; inputCol < ${h}; inputCol = inputCol + ${e[0]}) { + ${Ws(a,n)} + } + } + // Load one tile of B into local memory. + for (var inputRow = localRow; inputRow < ${i}; inputRow = inputRow + ${e[1]}) { + for (var inputCol = localCol; inputCol < ${p}; inputCol = inputCol + ${e[0]}) { + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalColStart + inputCol${n?", batchIndices":""}); + } + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + var BCached : array<${r}, colPerThread>; + for (var k = 0; k < tileInner; k = k + 1) { + for (var inner = 0; inner < colPerThread; inner = inner + 1) { + BCached[inner] = mm_Bsub[k][localCol + inner * ${e[0]}]; + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let ACached = ${a?`mm_Asub[k][localRow + innerRow * ${e[1]}];`:`mm_Asub[localRow + innerRow * ${e[1]}][k];`} + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = acc[innerRow][innerCol] + + ACached * BCached[innerCol]; + } + } + } + workgroupBarrier(); + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let gRow = globalRowStart + localRow + innerRow * ${e[1]}; + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let gCol = globalColStart + localCol + innerCol * ${e[0]}; + mm_write(batch, gRow, gCol, acc[innerRow][innerCol]); + } + } + `:` +let tileRow = i32(localId.y) * rowPerThread; +let tileCol = i32(localId.x) * colPerThread; + +let globalRow = i32(globalId.y) * rowPerThread; +let globalCol = i32(globalId.x) * colPerThread; +let globalRowStart = i32(workgroupId.y) * ${u}; + +let tileRowA = i32(localId.y) * ${d}; +let tileColA = i32(localId.x) * ${_}; +let tileRowB = i32(localId.y) * ${w}; +// Loop over shared dimension. +for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var innerRow = 0; innerRow < ${d}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < ${_}; innerCol = innerCol + 1) { + let inputRow = tileRowA + innerRow; + let inputCol = tileColA + innerCol; + ${Ws(a,n)} + } + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${w}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol + innerCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalCol + innerCol${n?", batchIndices":""}); + } + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + var BCached : array<${r}, colPerThread>; + for (var k = 0; k < tileInner; k = k + 1) { + for (var inner = 0; inner < colPerThread; inner = inner + 1) { + BCached[inner] = mm_Bsub[k][tileCol + inner]; + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + ${yd(a)} + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol]; + } + } + } + + workgroupBarrier(); +} + +for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + mm_write(batch, globalRow + innerRow, globalCol + innerCol, + acc[innerRow][innerCol]); + } +} +`;return` + var mm_Asub : array, ${m}>; + var mm_Bsub : array, ${i}>; + const rowPerThread = ${t[1]}; + const colPerThread = ${t[0]}; + const tileInner = ${i}; + +@compute @workgroup_size(${e[0]}, ${e[1]}, ${e[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let batch = ${s?"0":"i32(globalId.z)"}; + ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""} + let num_tiles = ${s?`${Math.ceil(o/i)}`:"(uniforms.dim_inner - 1) / tileInner + 1"}; + var kStart = ${s?`i32(globalId.z) * ${o}`:"0"}; + + var acc : array, rowPerThread>; + + // Without this initialization strange values show up in acc. + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = 0.0; + } + } + ${v} + } +`},wd=(t,e,r,n,a,i=!1)=>{let[s,o,l]=a,[u,p,h,m]=n,d=xa(s,l),_=xa(o,l),w=mt(n[0].type.tensor),v=()=>{let $=p.rank,E=u.rank,T=`var aIndices: ${p.type.indices};`;for(let A=$-2-1,P=E-1;A>=0;A--,P--)T+=` +aIndices[${A}] = ${E>1?`batchIndices[${P}]`:"batchIndices"};`;return d.forEach(A=>{T+=` +aIndices[${A}] = 0;`}),T+=` +aIndices[${$-2}] = u32(row); + aIndices[${$-1}] = u32(colIn);`,T},S=()=>{let $=h.rank,E=u.rank,T=`var bIndices: ${h.type.indices};`;for(let A=$-2-1,P=E-1;A>=0;A--,P--)T+=` +bIndices[${A}] = ${E>1?`batchIndices[${P}]`:"batchIndices"};`;return _.forEach(A=>{T+=` +bIndices[${A}] = 0;`}),T+=` +bIndices[${$-2}] = u32(row); + bIndices[${$-1}] = u32(colIn);`,T};return` + fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${u.type.indices}) -> ${kt(t,w)} { + var value = ${kt(t,w)}(0.0); + let col = colIn * ${t}; + if(row < uniforms.dim_a_outer && col < uniforms.dim_inner) + { + ${v()} + value = ${p.getByIndices("aIndices")}; + } + return value; + } + + fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${u.type.indices}) -> ${kt(t,w)} { + var value = ${kt(t,w)}(0.0); + let col = colIn * ${t}; + if(row < uniforms.dim_inner && col < uniforms.dim_b_outer) + { + ${S()} + value = ${h.getByIndices("bIndices")}; + } + return value; + } + + fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${kt(t,w)}) { + let col = colIn * ${t}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) { + var value = valueIn; + let coords = vec3(batch, row, colIn); + ${e?`value = value + ${i?"bias[colIn]":`${kt(t,w)}(bias[row])`};`:""} + ${r} + ${m.setByIndices("vec3(coords)","value")} + } + } + `},sl=(t,e,r,n,a=!1)=>{let i=t[0].dims,s=t[1].dims,o=i.slice(0,-2),l=s.slice(0,-2),u=n?n.slice(0,-2):r.slice(0,-2),p=K.size(u),h=i[i.length-2],m=i[i.length-1],d=s[s.length-1],_=m%4===0&&d%4===0,w=h<=8?[4,1,1]:[4,4,1],v=[8,8,1],S=[Math.ceil(d/v[0]/w[0]),Math.ceil(h/v[1]/w[1]),Math.ceil(p/v[2]/w[2])],$=_?4:1,E=[...o,h,m/$],T=E.length,A=[...l,m,d/$],P=A.length,B=[p,h,d/$],D=[{type:6,data:h},{type:6,data:d},{type:6,data:m}];on(e,D),D.push(...ye(u,E,A));let q=["rank","rank"],H=t.length>2;H&&(D.push(...ye(t[2].dims)),q.push("rank")),D.push(...ye(B));let ie=te=>{let de=u.length,se=Jo("batchDims",t[0].dataType,de,1),M=mt(t[0].dataType),R=X("a",t[0].dataType,T,$),G=X("b",t[1].dataType,P,$),re=ge("result",t[0].dataType,B.length,$),ee=[R,G];if(H){let Ce=a?$:1;ee.push(X("bias",t[2].dataType,t[2].dims.length,Ce))}let ne=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];ln(e,ne);let W=mt(re.type.tensor),ae=sn(e,re.type.value,W),fe=wd($,H,ae,[se,R,G,re],[o,l,u],a);return` + ${te.registerUniforms(ne).registerInternalVariables(se).declareVariables(...ee,re)} + ${fe} + ${_?Yi(w,v,M,se):Xi(w,v,M,se)} + `};return{name:"MatMul",shaderCache:{hint:`${w};${e.activation};${_};${a}`,inputDependencies:q},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:S[0],y:S[1],z:S[2]},programUniforms:D}),getShaderSource:ie}}}),bd,jf,Cy=J(()=>{$e(),un(),Te(),dn(),al(),Hf(),as(),bd=(t,e,r,n,a=!1,i,s=4,o=4,l=4,u="f32")=>{let p=q=>{switch(q){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${u}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${q} is not supported.`)}},h=q=>{switch(q){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 ${q} is not supported.`)}},m=t?` + let coord = vec4(batch, xRow, xCol, xCh); + `:` + let coord = vec4(batch, xCh, xRow, xCol); + `,d=t?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,_=t?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",w=t?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",v=t?"row":"col",S=t?"col":"row",$=` + let inChannels = i32(uniforms.w_shape[2]); + let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + let outRow = ${v} / outWidth; + let outCol = ${v} % outWidth; + + let WRow = ${S} / (i32(uniforms.w_shape[1]) * inChannels); + let WCol = ${S} / inChannels % i32(uniforms.w_shape[1]); + 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 = ${S} % inChannels; + var resData = ${kt(s,u)}(0.0); + // The bounds checking is always needed since we use it to pad zero for + // the 'same' padding type. + if (xRow >= 0 && xRow < ${_} && xCol >= 0 && xCol < ${w}) { + ${m} + let xIndex = getIndexFromCoords4D(coord, vec4(uniforms.x_shape)); + ${p(s)} + } + return resData;`,E=t?e&&n?` + let col = colIn * ${s}; + ${$}`:` + let col = colIn * ${s}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) { + ${$} + } + return ${kt(s,u)}(0.0);`:n&&r?` + let col = colIn * ${s}; + ${$}`:` + let col = colIn * ${s}; + if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { + ${$} + } + return ${kt(s,u)}(0.0);`,T=`${h(o)}`,A=kt(l,u),P=kt(t?s:o,u),B=kt(t?o:s,u),D=sn(i,A,u);return` + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${P} { + ${t?E:T} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${B} { + ${t?T:E} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${A}) { + let col = colIn * ${l}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) + { + var value = valueIn; + let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + ${d} + ${nl(a)} + ${D} + setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value); + } + }`},jf=(t,e,r,n,a,i,s,o)=>{let l=e.format==="NHWC",u=l?t[0].dims[3]:t[0].dims[1],p=r[0],h=l?r[2]:r[3],m=l?r[1]:r[2],d=l?r[3]:r[1],_=l&&(u%4===0||u%3===0)&&d%4===0,w=l?d:h*m,v=l?h*m:d,S=[8,8,1],$=n<=8?[4,1,1]:[4,4,1],E=[Math.ceil(w/S[0]/$[0]),Math.ceil(v/S[1]/$[1]),Math.ceil(p/S[2]/$[2])];at("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${E}`);let T=_?l&&u%4!==0?3:4:1,A=S[1]*$[1],P=S[0]*$[0],B=Math.max(S[0]*T,S[1]),D=n%A===0,q=a%P===0,H=i%B===0,ie=_?[T,4,4]:[1,1,1],te=[{type:6,data:n},{type:6,data:a},{type:6,data:i},{type:6,data:[e.pads[0],e.pads[1]]},{type:6,data:e.strides},{type:6,data:e.dilations}];on(e,te),te.push(...ye(t[0].dims,t[1].dims));let de=["rank","rank"];s&&(te.push(...ye(t[2].dims)),de.push("rank")),te.push(...ye(r));let se=M=>{let R=[{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}];ln(e,R);let G=_?4:1,re=mt(t[0].dataType),ee=` + fn setOutputAtIndex(flatIndex : i32, value : ${_?`vec4<${re}>`:re}) { + result[flatIndex] = ${_?`vec4<${re}>`:re}(value); + } + fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${_?`vec4<${re}>`:re}) { + let flatIndex = getOutputIndexFromCoords(vec4(d0, d1, d2, d3)); + setOutputAtIndex(flatIndex ${_?"/ 4":""}, value); + }`,ne=X("x",t[0].dataType,t[0].dims.length,T===3?1:T),W=X("w",t[1].dataType,t[1].dims.length,G),ae=[ne,W],fe=ge("result",t[0].dataType,r.length,G);if(s){let Ce=X("bias",t[2].dataType,t[2].dims.length,G);ae.push(Ce),ee+=` + fn getBiasByOutputCoords(coords : vec4) -> ${_?`vec4<${re}>`:re} { + return bias[coords.${l?"w":"y"}${_?"/ 4":""}]; + }`}return` + ${il("uniforms.result_strides")} + //struct Uniforms { xShape : vec4, wShape : vec4, outShape : vec4, + // outShapeStrides: vec3, filterDims : vec2, pad : vec2, stride : vec2, + // dilation : vec2, dimAOuter : i32, dimBOuter : i32, dimInner : i32 }; + ${M.registerUniforms(R).declareVariables(...ae,fe)} + ${ee} + ${bd(l,D,q,H,s,e,ie[0],ie[1],ie[2],re)} + ${_?Yi($,S,re,void 0,!l,B):Xi($,S,re,void 0,!l,B,!1,void 0,o)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${e.cacheKey};${T};${_};${D};${q};${H};${A};${P};${B}`,inputDependencies:de},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:te}),getShaderSource:se}}}),Co,qf,Ty=J(()=>{$e(),Me(),Te(),Xf(),dn(),Co=(t,e,r)=>{let n=t.length>2,a=n?"value += b[output_channel];":"",i=t[0].dims,s=t[1].dims,o=s[0]/e.group,l=e.format==="NHWC",u=Ri(i,s,e.dilations,e.pads,e.strides,l),p=K.size(u),h=[{type:12,data:p},{type:12,data:e.dilations},{type:12,data:[e.strides[0],e.strides[1]]},{type:12,data:[e.pads[0],e.pads[1]]},{type:12,data:o}];on(e,h),h.push(...ye(i,s));let m=["rank","rank"];n&&(h.push(...ye(t[2].dims)),m.push("rank")),h.push(...ye(u));let d=_=>{let w=ge("output",t[0].dataType,u.length),v=mt(w.type.tensor),S=sn(e,w.type.value,v),$=X("x",t[0].dataType,i.length),E=X("w",t[1].dataType,s.length),T=[$,E];n&&T.push(X("b",t[2].dataType,t[2].dims.length));let A=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:e.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];return ln(e,A),` + ${_.registerUniforms(A).declareVariables(...T,w)} + + ${_.mainStart()} + ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let outputIndices = ${w.offsetToIndices("global_idx")}; + let batch: u32 = outputIndices[0]; + let output_channel: u32 = outputIndices[${l?3:1}]; + let xRCCorner: vec2 = vec2(outputIndices[${l?1:2}], outputIndices[${l?2:3}]) * uniforms.strides - uniforms.pads; + let group_id: u32 = output_channel / uniforms.output_channels_per_group; + + var value: ${w.type.value} = ${w.type.value}(0); + for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) { + let input_channel = group_id * uniforms.w_shape[1] + wInChannel; + for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) { + let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0]; + + if (xHeight < 0u || xHeight >= uniforms.x_shape[${l?1:2}]) { + continue; + } + + for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) { + let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1]; + if (xWidth < 0u || xWidth >= uniforms.x_shape[${l?2:3}]) { + continue; + } + + let xVal = ${l?$.get("batch","xHeight","xWidth","input_channel"):$.get("batch","input_channel","xHeight","xWidth")}; + let wVal = ${E.get("output_channel","wInChannel","wHeight","wWidth")}; + value += xVal*wVal; + } + } + } + ${a} + ${S} + ${w.setByOffset("global_idx","value")} + }`};return{name:"GroupedConv",shaderCache:{hint:e.cacheKey,inputDependencies:m},getRunData:()=>({outputs:[{dims:r?r(u):u,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:d}},qf=(t,e,r)=>{let n=t.length>2,a=ot(r[3]),i=ot(r[2]),s=K.size(r)/a/i,o=[t[0].dims[0],t[0].dims[1],t[0].dims[2],t[0].dims[3]/a],l=[t[1].dims[0],t[1].dims[1],t[1].dims[2],t[1].dims[3]/a],u=[r[0],r[1],r[2],r[3]/a],p=[{type:12,data:s},{type:6,data:[e.strides[0],e.strides[1]]},{type:6,data:[e.pads[0],e.pads[1]]}];on(e,p),p.push(...ye(o,l,u));let h=(i-1)*e.strides[1]+l[1],m=d=>{let _=ge("output",t[0].dataType,u.length,a),w=mt(_.type.tensor),v=sn(e,_.type.value,w),S=X("x",t[0].dataType,o.length,a),$=X("w",t[1].dataType,l.length,a),E=[S,$];n&&E.push(X("b",t[2].dataType,t[2].dims,a));let T=n?"value += b[output_channel];":"",A=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return ln(e,A),` + ${d.registerUniforms(A).declareVariables(...E,_)} + ${d.mainStart()} + ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let width0 = uniforms.output_shape[3]; + let output_channel = global_idx % width0; + var index1 = global_idx / width0; + let width1 = uniforms.output_shape[2] / ${i}u; + let col = (index1 % width1) * ${i}u; + index1 = index1 / width1; + let row = index1 % uniforms.output_shape[1]; + let batch = index1 / uniforms.output_shape[1]; + + let x_corner = vec2(i32(row), i32(col)) * uniforms.strides - uniforms.pads; + + var x_vals: array<${S.type.value}, ${h}>; + var values: array<${_.type.value}, ${i}>; + let input_channel = output_channel; + // Use constant instead of uniform can give better performance for w's height/width. + for (var w_height: u32 = 0u; w_height < ${l[0]}; w_height++) { + let x_height = x_corner.x + i32(w_height); + if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) { + for (var i = 0; i < ${h}; i++) { + let x_width = x_corner.y + i; + if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) { + x_vals[i] = ${S.get("batch","u32(x_height)","u32(x_width)","input_channel")}; + } else { + x_vals[i] = ${S.type.value}(0); + } + } + for (var w_width: u32 = 0u; w_width < ${l[1]}; w_width++) { + let w_val = ${$.get("w_height","w_width","0","output_channel")}; + for (var i = 0u; i < ${i}u; i++) { + values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]); + } + } + } + } + + for (var i = 0u; i < ${i}u; i++) { + var value = values[i]; + ${T} + ${v} + ${_.set("batch","row","col + i","output_channel","value")}; + } + }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${e.cacheKey};${a};${i};${h};${l[0]};${l[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:p}),getShaderSource:m}}}),To,vd,Kf,Yf=J(()=>{$e(),Me(),as(),Te(),dn(),To=(t,e,r,n,a=!1)=>{let i=t[0].dims,s=t[1].dims,o=i[i.length-2],l=s[s.length-1],u=i[i.length-1],p=ot(l),h=ot(u),m=ot(o),d=K.size(r)/p/m,_=t.length>2,w=n?n.slice(0,-2):r.slice(0,-2),v=[K.size(w),o,l],S=[{type:12,data:d},{type:12,data:o},{type:12,data:l},{type:12,data:u}];on(e,S),S.push(...ye(w,i,s)),_&&S.push(...ye(t[2].dims)),S.push(...ye(v));let $=E=>{let T=Jo("batch_dims",t[0].dataType,w.length),A=X("a",t[0].dataType,i.length,h),P=X("b",t[1].dataType,s.length,p),B=ge("output",t[0].dataType,v.length,p),D=mt(B.type.tensor),q=sn(e,B.type.value,D),H=[A,P],ie="";if(_){let ee=a?p:1;H.push(X("bias",t[2].dataType,t[2].dims.length,ee)),ie=`${a?`value += bias[col / ${ee}];`:`value += ${B.type.value}(bias[row + i]);`}`}let te=i.slice(0,-2),de=s.slice(0,-2),se=xa(te,w),M=xa(de,w),R=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];ln(e,R);let G=(ee,ne)=>{let W=ee.rank,ae=ee.name;if(W===2)return`var ${ae}_indices = ${ee.type.indices}(0u, 0u);`;let fe=T.rank,Ce=`var ${ae}_indices: ${ee.type.indices};`;for(let Be=W-2-1,Ye=fe-1;Be>=0;Be--,Ye--)Ce+=` +${ae}_indices[${Be}] = ${fe>1?`batch_indices[${Ye}]`:"batch_indices"};`;return ne.forEach(Be=>{Ce+=` +${ae}_indices[${Be}] = 0;`}),Ce+=`${ae}_indices[${W-2}] = 0u; + ${ae}_indices[${W-1}] = 0u;`,Ce},re=()=>{let ee=`var a_data: ${A.type.value};`;for(let ne=0;ne; + for (var k: u32 = 0u; k < uniforms.K; k = k + ${h}) { + ${re()} + } + for (var i = 0u; i < ${m}u; i++) { + var value = values[i]; + ${ie} + ${q} + let cur_indices = ${B.type.indices}(batch, row + i, col); + let offset = ${B.indicesToOffset("cur_indices")}; + ${B.setByOffset(`offset / ${p}`,"value")}; + } + } + `};return{name:"MatMulNaive",shaderCache:{hint:`${e.activation};${p};${h};${m};${a}`,inputDependencies:_?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:S}),getShaderSource:$}},vd=t=>{if(!t||t.length!==2)throw new Error("MatMul requires 2 inputs.");if(t[0].dims[t[0].dims.length-1]!==t[1].dims[t[1].dims.length-2])throw new Error("shared dimension does not match.")},Kf=t=>{vd(t.inputs);let e=zn.calcShape(t.inputs[0].dims,t.inputs[1].dims,!0);if(!e)throw new Error("Can't use matmul on the given tensors");let r=e[e.length-1],n=t.inputs[0].dims[t.inputs[0].dims.length-1];r<8&&n<8?t.compute(To(t.inputs,{activation:""},e)):t.compute(sl(t.inputs,{activation:""},e))}}),Ri,xi,$d,Vs,Ao,xd,Sd,Io,Xf=J(()=>{Me(),Cy(),as(),Ty(),dn(),Yf(),Ia(),Ri=(t,e,r,n,a,i)=>{let s=t[0],o=t.slice(i?1:2,i?3:4),l=o.length,u=e[0],p=e.slice(2).map((m,d)=>m+(m-1)*(r[d]-1)),h=o.map((m,d)=>m+n[d]+n[d+l]).map((m,d)=>Math.floor((m-p[d]+a[d])/a[d]));return h.splice(0,0,s),h.splice(i?3:1,0,u),h},xi=[2,3,1,0],$d=(t,e)=>{if(!t||t.length!==2&&t.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(t[0].dims.length!==4&&t[0].dims.length!==3)throw new Error("currently only support conv 1D and 2D");if(t[0].dims.length!==t[1].dims.length)throw new Error("filter does not have same dimension as input");let r=t[0].dims[e.format==="NHWC"?t[0].dims.length-1:1],n=t[1].dims[1]*e.group;if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(t.length===3&&(t[2].dims.length!==1||t[1].dims[0]!==t[2].dims[0]))throw new Error("invalid bias");let a=t[0].dims.length-2;if(e.dilations.length!==a)throw new Error(`dilations should be ${a}D`);if(e.strides.length!==a)throw new Error(`strides should be ${a}D`);if(e.pads.length!==a*2)throw new Error(`pads should be ${a*2}D`);if(e.kernelShape.length!==0&&e.kernelShape.length!==t[1].dims.length-2)throw new Error("invalid kernel shape")},Vs=(t,e)=>{let r=t.kernelShape.slice();for(let i=2;i{let e=rl(t),r=t.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][t.auto_pad],a=t.dilations,i=t.group,s=t.kernel_shape,o=t.pads,l=t.strides,u=t.w_is_const();return{autoPad:n,format:r,dilations:a,group:i,kernelShape:s,pads:o,strides:l,wIsConst:u,...e,cacheKey:`${t.format};${e.activation};`}},xd=(t,e,r)=>{let n=Vs(r,e),a=r.format==="NHWC";if(r.group!==1){if(!t.adapterInfo.isArchitecture("ampere")&&a&&e[1].dims[0]===r.group&&e[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let P=Ri(e[0].dims,e[1].dims,r.dilations,n.pads,r.strides,a),B=t.kernelCustomData.wT??t.compute(Er(e[1],xi),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=B);let D=[e[0],B];e.length===3&&D.push(e[2]),t.compute(qf(D,n,P),{inputs:D})}else t.compute(Co(e,n));return}let i=e.length===3,s=e[0].dims[a?1:2],o=e[0].dims[a?2:3],l=e[0].dims[a?3:1],u=e[1].dims[2],p=e[1].dims[3],h=Ri(e[0].dims,e[1].dims,r.dilations,n.pads,r.strides,a),m=h[a?1:2],d=h[a?2:3],_=h[a?3:1],w=a&&u===s&&p===o&&r.pads[0]===0&&r.pads[1]===0;if(w||u===1&&p===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 P=h[0],B,D,q,H=[];if(a){let de=t.kernelCustomData.wT??t.compute(Er(e[1],xi),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=de),w){let se=s*o*l;B=e[0].reshape([1,P,se]),D=de.reshape([1,se,_]),q=[1,P,_]}else B=e[0].reshape([P,s*o,l]),D=de.reshape([1,l,_]),q=[P,m*d,_];H.push(B),H.push(D)}else B=e[0].reshape([P,l,s*o]),D=e[1].reshape([1,_,l]),q=[P,_,m*d],H.push(D),H.push(B);i&&H.push(e[2]);let ie=q[2],te=H[0].dims[H[0].dims.length-1];ie<8&&te<8?t.compute(To(H,n,h,q,a),{inputs:H}):t.compute(sl(H,n,h,q,a),{inputs:H});return}let v=!0,S=t.kernelCustomData.wT??t.compute(Er(e[1],xi),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=S);let $=[e[0],S];i&&$.push(e[2]);let E=a?m*d:_,T=a?_:m*d,A=u*p*l;t.compute(jf($,n,h,E,T,A,i,v),{inputs:$})},Sd=(t,e)=>{let r=e.format==="NHWC",n=[t.inputs[0].reshape(r?[t.inputs[0].dims[0],1,t.inputs[0].dims[1],t.inputs[0].dims[2]]:[t.inputs[0].dims[0],t.inputs[0].dims[1],1,t.inputs[0].dims[2]]),t.inputs[1].reshape([t.inputs[1].dims[0],t.inputs[1].dims[1],1,t.inputs[1].dims[2]])];t.inputs.length===3&&n.push(t.inputs[2]);let a=[0,e.pads[0],0,e.pads[1]],i=[1].concat(e.strides),s=[1].concat(e.dilations),o=[1].concat(e.kernelShape),l=Vs({...e,pads:a,strides:i,dilations:s,kernelShape:o},n);t.compute(Co(n,l,u=>r?[u[0],u[2],u[3]]:[]))},Io=(t,e)=>{$d(t.inputs,e),t.inputs[0].dims.length===3?Sd(t,e):xd(t,t.inputs,e)}}),kd,Qf,Ay=J(()=>{$e(),un(),Te(),dn(),al(),Hf(),as(),kd=(t,e=!1,r,n,a=4)=>{let i=v=>{switch(v){case 1:return"return w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];";case 4:return` + let coord1 = vec4(coordX, coordY, col + 1, rowInner); + let coord2 = vec4(coordX, coordY, col + 2, rowInner); + let coord3 = vec4(coordX, coordY, col + 3, rowInner); + let v0 = w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))]; + let v1 = w[getIndexFromCoords4D(coord1, vec4(uniforms.w_shape))]; + let v2 = w[getIndexFromCoords4D(coord2, vec4(uniforms.w_shape))]; + let v3 = w[getIndexFromCoords4D(coord3, vec4(uniforms.w_shape))]; + return ${n}(v0, v1, v2, v3); + `;default:throw new Error(`innerElementSize ${v} is not supported.`)}},s=t?` + let coord = vec4(batch, iXR, iXC, xCh); + `:` + let coord = vec4(batch, xCh, iXR, iXC); + `,o=t?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,l=t?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",u=t?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",p=t?"row":"col",h=t?"col":"row",m=` + let inChannels = ${t?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"}; + let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + let outRow = ${p} / outWidth; + let outCol = ${p} % outWidth; + + let WRow = ${h} / (uniforms.filter_dims[1] * inChannels); + let WCol = ${h} / inChannels % uniforms.filter_dims[1]; + let xR = f32(outRow - uniforms.pads[0] + uniforms.dilations[0] * WRow) / f32(uniforms.strides[0]); + let xC = f32(outCol - uniforms.pads[1] + uniforms.dilations[1] * WCol) / f32(uniforms.strides[1]); + if (xR < 0.0 || xR >= f32(${l}) || fract(xR) > 0.0) { + return ${n}(0.0); + } + if (xC < 0.0 || xC >= f32(${u}) || fract(xC) > 0.0) { + return ${n}(0.0); + } + let iXR = i32(xR); + let iXC = i32(xC); + let xCh = ${h} % inChannels; + ${s} + return x[getIndexFromCoords4D(coord, vec4(uniforms.x_shape))/${a}];`,d=t?` + let col = colIn * ${a}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) { + ${m} + } + return ${n}(0.0);`:` + let col = colIn * ${a}; + if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { + ${m} + } + return ${n}(0.0);`,_=` + let col = colIn * ${a}; + let inChannels = ${t?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"}; + let coordX = uniforms.filter_dims[0] - 1 - row / (uniforms.filter_dims[1] * inChannels); + let coordY = uniforms.filter_dims[1] - 1 - (row / inChannels) % uniforms.filter_dims[1]; + if (${t?"row < uniforms.dim_inner && col < uniforms.dim_b_outer":"row < uniforms.dim_inner && col < uniforms.dim_a_outer"} && coordX >= 0 && coordY >= 0) { + let rowInner = row % inChannels; + let coord = vec4(coordX, coordY, col, rowInner); + ${i(a)} + } + return ${n}(0.0); + `,w=sn(r,n);return` + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${n} { + ${t?d:_} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${n} { + ${t?_:d} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${n}) { + let col = colIn * ${a}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) { + var value = valueInput; + let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + ${o} + ${nl(e)} + ${w} + result[getIndexFromCoords4D(coords, vec4(uniforms.result_shape))/${a}] = value; + } + }`},Qf=(t,e,r,n,a,i,s,o)=>{let l=e.format==="NHWC",u=l?t[0].dims[3]:t[0].dims[1],p=r[0],h=l?r[2]:r[3],m=l?r[1]:r[2],d=l?r[3]:r[1],_=l&&u%4===0&&u%3&&d%4===0,w=l?d:h*m,v=l?h*m:d,S=[8,8,1],$=n<=8?[4,1,1]:[4,4,1],E=[Math.ceil(w/S[0]/$[0]),Math.ceil(v/S[1]/$[1]),Math.ceil(p/S[2]/$[2])];at("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${E}`);let T=_?4:1,A=Math.max(S[0]*T,S[1]),P=_?4:1,B=[e.kernelShape[l?1:2],e.kernelShape[l?2:3]],D=[B[0]+(e.dilations[0]<=1?0:(B[0]-1)*(e.dilations[0]-1)),B[1]+(e.dilations[1]<=1?0:(B[1]-1)*(e.dilations[1]-1))],q=[D[0]-1-Math.floor((e.pads[0]+e.pads[2])/2),D[1]-1-Math.floor((e.pads[1]+e.pads[3])/2)],H=[{type:6,data:n},{type:6,data:a},{type:6,data:i},{type:6,data:e.strides},{type:6,data:e.dilations},{type:6,data:B},{type:6,data:q}];on(e,H),H.push(...ye(t[0].dims,t[1].dims));let ie=["rank","rank"];s&&(H.push(...ye(t[2].dims)),ie.push("rank")),H.push(...ye(r));let te=de=>{let se=X("x",t[0].dataType,t[0].dims.length,P),M=X("w",t[1].dataType,t[1].dims.length,1),R=ge("result",t[0].dataType,r.length,P),G=[se,M],re="";if(s){let W=X("bias",t[2].dataType,t[2].dims.length,P);G.push(W),re+=` + fn getBiasByOutputCoords(coords : vec4) -> ${W.type.value} { + return bias[coords.${l?"w":"y"}${_?"/ 4":""}]; + }`}let ee=[{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:B.length},{name:"pads",type:"i32",length:q.length}];ln(e,ee);let ne=mt(t[0].dataType,1);if(ne!=="f16"&&ne!=="f32")throw new Error(`elemType ${ne} is not supported.`);return` + ${il("uniforms.result_strides")} + ${de.registerUniforms(ee).declareVariables(...G,R)}; + ${re} + ${kd(l,s,e,se.type.value,T)} + ${_?Yi($,S,ne,void 0,!l,A):Xi($,S,ne,void 0,!l,A,!1,void 0,o)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${e.cacheKey};${$};${S};${_}`,inputDependencies:ie},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:H}),getShaderSource:te}}}),Ed,Mo,Iy=J(()=>{$e(),un(),Me(),Te(),Ed=(t,e,r,n,a,i=!1,s,o,l=!1)=>{let u=l?1:2,p=l?2:3,h=l?3:1,m=i?2:1,d=` + fn setOutputAtIndex(flatIndex : u32, value : ${i?`vec4<${s}>`:s}) { + result[flatIndex] = ${i?`vec4<${s}>`:s}(value); + }`;n&&(d+=` + fn getBiasByOutputCoords(coords : vec4) -> ${i?`vec4<${s}>`:s} { + return bias[coords.${l?"w":"y"}${i?"/ 4":""}]; + }`);let _=i?4:1,w=X("W",e[1].dataType,e[1].dims.length,_),v=X("Dy",e[0].dataType,e[0].dims.length,_),S=[v,w];n&&S.push(X("bias",e[2].dataType,[r[h]].length,_));let $=ge("result",e[0].dataType,r.length,_),E=`{ + let batch: u32 = ${a?"global_id.z":"workgroup_id.z"} / uniforms.result_shape[1]; + let r = ${a?"global_id.z":"workgroup_id.z"} % uniforms.result_shape[1]; + let c = ${a?"global_id.y":"workgroup_id.y"} * ${m}; + let d1: u32 = ${a?"global_id.x":"workgroup_id.x"} * 4; + + let dyCorner = vec2(i32(r), i32(c)) - vec2(uniforms.pads); + + // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1). + // ? = to be determined. : = across all values in that axis. + var dotProd: array, ${m}>; + for (var i = 0; i < ${m}; i++) { + dotProd[i] = vec4<${s}>(0.0); + } + for (var wR: u32 = 0; wR < uniforms.filter_dims[0]; wR = wR + 1) { + var dyR = (${s}(dyCorner.x) + ${s}(wR)) / ${s}(uniforms.strides.x); + let wRPerm = uniforms.filter_dims[0] - 1 - wR; + if (dyR < 0.0 || dyR >= ${s}(uniforms.Dy_shape[1]) || + fract(dyR) > 0.0 || wRPerm < 0) { + continue; + } + let idyR: u32 = u32(dyR); + + for (var wC: u32 = 0; wC < uniforms.filter_dims[1]; wC = wC + 1) { + let dyC = (${s}(dyCorner.y) + ${s}(wC)) / ${s}(uniforms.strides.y); + let dyC2 = (${s}(dyCorner.y) + 1.0 + ${s}(wC)) / ${s}(uniforms.strides.y); + let wCPerm = uniforms.filter_dims[1] - 1 - wC; + if (wCPerm < 0) { + continue; + } + var bDyCVal = true; + var bDyCVal2 = true; + if (dyC < 0.0 || dyC >= ${s}(uniforms.Dy_shape[2]) || + fract(dyC) > 0.0) { + bDyCVal = false; + } + if (dyC2 < 0.0 || dyC2 >= ${s}(uniforms.Dy_shape[2]) || + fract(dyC2) > 0.0) { + bDyCVal2 = false; + } + + let idyC: u32 = u32(dyC); + let idyC2: u32 = u32(dyC2); + if (bDyCVal && bDyCVal2) { + let d2Length = uniforms.Dy_shape[3]; + for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${v.get("batch","idyR","idyC","d2")}; + let tmpval = vec4<${s}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[0] = dotProd[0] + tmpval; + + xValue = ${v.get("batch","idyR","idyC2","d2")}; + + dotProd[1] = dotProd[1] + vec4<${s}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + } + } else if (bDyCVal) { + let d2Length = uniforms.Dy_shape[${h}]; + for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${v.get("batch","idyR","idyC","d2")}; + let tmpval = vec4<${s}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[0] = dotProd[0] + tmpval; + } + } else if (bDyCVal2) { + let d2Length = uniforms.Dy_shape[3]; + for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${w.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${v.get("batch","idyR","idyC2","d2")}; + let tmpval = vec4<${s}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[1] = dotProd[1] + tmpval; + } + } + } + } + + for (var i: u32 = 0; i < ${m}; i = i + 1) { + let value = dotProd[i] + ${n?"bias[c+i]":`vec4<${s}>(0.0)`}; + ${$.set("batch","r","c + i","d1","value")}; + } + }`,T=` + let outputIndices = ${$.offsetToIndices("global_idx")}; + let batch = ${$.indicesGet("outputIndices",0)}; + let d1 = ${$.indicesGet("outputIndices",h)}; + let r = ${$.indicesGet("outputIndices",u)}; + let c = ${$.indicesGet("outputIndices",p)}; + let dyCorner = vec2(i32(r), i32(c)) - uniforms.pads; + let dyRCorner = dyCorner.x; + let dyCCorner = dyCorner.y; + let groupId = d1 / uniforms.output_channels_per_group; + let wOutChannel = d1 - groupId * uniforms.output_channels_per_group; + // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1). + // ? = to be determined. : = across all values in that axis. + var dotProd = ${s}(0.0); + for (var wR: u32 = 0; wR < uniforms.effective_filter_dims.x; wR = wR + 1) { + if (wR % uniforms.dilations.x != 0) { + continue; + } + let dyR = (${s}(dyRCorner) + ${s}(wR)) / ${s}(uniforms.strides[0]); + let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x; + if (dyR < 0.0 || dyR >= ${s}(uniforms.Dy_shape[${u}]) || fract(dyR) > 0.0 || + wRPerm < 0) { + continue; + } + let idyR: u32 = u32(dyR); + + for (var wC: u32 = 0; wC < uniforms.effective_filter_dims.y; wC = wC + 1) { + if (wC % uniforms.dilations.y != 0) { + continue; + } + let dyC = (${s}(dyCCorner) + ${s}(wC)) / ${s}(uniforms.strides.y); + let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y; + if (dyC < 0.0 || dyC >= ${s}(uniforms.Dy_shape[${p}]) || + fract(dyC) > 0.0 || wCPerm < 0) { + continue; + } + let idyC: u32 = u32(dyC); + var inputChannel = groupId * uniforms.input_channels_per_group; + for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + 1) { + let xValue = ${l?v.get("batch","idyR","idyC","inputChannel"):v.get("batch","inputChannel","idyR","idyC")}; + let wValue = ${w.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")}; + dotProd = dotProd + xValue * wValue; + inputChannel = inputChannel + 1; + } + } + } + let value = dotProd + ${n?"bias[d1]":`${s}(0.0)`}; + ${$.setByOffset("global_idx","value")}; + `;return` + ${t.registerUniforms(o).declareVariables(...S,$)} + ${d} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}; + ${i?E:T}}`},Mo=(t,e,r)=>{let n=t.length>2,a=e.outputShape,i=K.size(a),s=[Math.ceil(i/64),1,1];at("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${s}`);let o=e.format==="NHWC",l=["rank","rank"],u=[e.strides[0],e.strides[1]],p=[e.kernelShape[o?1:2],e.kernelShape[o?2:3]],h=[e.dilations[0],e.dilations[1]],m=[p[0]+(e.dilations[0]<=1?0:(e.kernelShape[o?1:2]-1)*(e.dilations[0]-1)),p[1]+(e.dilations[1]<=1?0:(e.kernelShape[o?2:3]-1)*(e.dilations[1]-1))],d=[m[0]-1-Math.floor((e.pads[0]+e.pads[2])/2),m[1]-1-Math.floor(e.pads[1]+e.pads[3])/2],_=!1,w=e.group,v=t[1].dims,S=v[0]/w,$=v[1],E=[{type:12,data:i},{type:12,data:u},{type:12,data:p},{type:12,data:h},{type:12,data:m},{type:6,data:d},{type:12,data:S},{type:12,data:$},...ye(t[0].dims,t[1].dims)];n&&(E.push(...ye(t[2].dims)),l.push("rank")),E.push(...ye(a));let T=s[1]===1&&s[2]===1,A=P=>{let B=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:u.length},{name:"filter_dims",type:"u32",length:p.length},{name:"dilations",type:"u32",length:p.length},{name:"effective_filter_dims",type:"u32",length:m.length},{name:"pads",type:"i32",length:d.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],D=mt(t[0].dataType);return`${Ed(P,t,a,n,T,_,D,B,o)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${e.cacheKey};`,inputDependencies:l},getRunData:()=>({dispatchGroup:{x:s[0],y:s[1],z:s[2]},outputs:[{dims:r?r(a):a,dataType:t[0].dataType}],programUniforms:E}),getShaderSource:A}}}),Cd,Td,Ad,Gs,Zf,Id,Md,Od,zd,Jf,My=J(()=>{Ay(),Iy(),dn(),Ia(),Cd=(t,e,r,n,a,i)=>(t-1)*e+r+(n-1)*a+1-i,Td=(t,e,r,n,a)=>{let i=Math.floor(t/2);e==="SAME_UPPER"?(r[n]=i,r[a]=t-i):e==="SAME_LOWER"&&(r[n]=t-i,r[a]=i)},Ad=(t,e,r,n,a,i,s,o,l,u)=>{let p=t.length-2,h=u.length===0;if(l.length===0)for(let _=0;_{let r=t.kernelShape.slice();if(t.kernelShape.length===0||t.kernelShape.reduce((h,m)=>h*m,1)===0){r.length=0;for(let h=2;hh+m,0)===0){let h=e[0].dims.length-2;l=new Array(h).fill(1)}let u=t.strides.slice();if(u.reduce((h,m)=>h+m,0)===0){let h=e[0].dims.length-2;u=new Array(h).fill(1)}Ad(o,r,l,t.autoPad,t.group,a,u,n,s,i);let p=Object.assign({},t);return Object.assign(p,{kernelShape:r,pads:a,outputPadding:s,outputShape:i,dilations:l,strides:u}),p},Zf=t=>{let e=rl(t),r=t.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof t.autoPad>"u"?0:t.autoPad],a=t.dilations,i=t.group,s=t.kernelShape,o=t.pads,l=t.strides,u=t.wIsConst(),p=t.outputPadding,h=t.outputShape;return{autoPad:n,format:r,dilations:a,group:i,kernelShape:s,outputPadding:p,outputShape:h,pads:o,strides:l,wIsConst:u,...e,cacheKey:`${t.format};${e.activation};`}},Id=(t,e)=>{if(!t||t.length!==2&&t.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(t[0].dims.length!==4&&t[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(t[0].dims.length!==t[1].dims.length)throw new Error("filter does not have same dimension as input");let r=t[0].dims[e.format==="NHWC"?t[0].dims.length-1:1],n=t[1].dims[0];if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let a=t[1].dims[1]*e.group;if(t.length===3&&(t[2].dims.length!==1||t[2].dims[0]!==a))throw new Error("invalid bias");let i=t[0].dims.length-2;if(e.dilations.reduce((s,o)=>s+o,0)>0&&e.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(e.strides.reduce((s,o)=>s+o,0)>0&&e.strides.length!==i)throw new Error(`strides should be ${i}D`);if(e.pads.reduce((s,o)=>s+o,0)>0&&e.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(e.outputPadding.length!==i&&e.outputPadding.length!==0)throw new Error(`output_padding should be ${i}D`);if(e.kernelShape.reduce((s,o)=>s+o,0)>0&&e.kernelShape.length!==0&&e.kernelShape.length!==t[1].dims.length-2)throw new Error("invalid kernel shape");if(e.outputShape.length!==0&&e.outputShape.length!==t[0].dims.length-2)throw new Error("invalid output shape")},Md=[2,3,1,0],Od=(t,e,r)=>{let n=Gs(r,e),a=r.format==="NHWC",i=n.outputShape,s=i[a?3:1],o=e[0].dims[a?3:1];if(n.group!==1||s===1&&o===1){t.compute(Mo(e,n));return}let l=i[a?1:2],u=i[a?2:3],p=e[1].dims[2],h=e[1].dims[3],m=a?l*u:s,d=a?s:l*u,_=p*h*o,w=!0,v=t.kernelCustomData.wT??t.compute(Er(e[1],Md),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=v);let S=[e[0],v],$=e.length===3;$&&(!a&&e[2].dims.length===1?S.push(e[2].reshape([e[2].dims[0],1,1])):S.push(e[2])),t.compute(Qf(S,n,i,m,d,_,$,w),{inputs:S})},zd=(t,e)=>{let r=e.format==="NHWC",n=[t.inputs[0].reshape(r?[t.inputs[0].dims[0],1,t.inputs[0].dims[1],t.inputs[0].dims[2]]:[t.inputs[0].dims[0],t.inputs[0].dims[1],1,t.inputs[0].dims[2]]),t.inputs[1].reshape([t.inputs[1].dims[0],t.inputs[1].dims[1],1,t.inputs[1].dims[2]])];t.inputs.length===3&&n.push(t.inputs[2]);let a=e.kernelShape;(a.length===0||a[0]===0)&&(a=[t.inputs[1].dims[2]]);let i=e.dilations;(i.length===0||i[0]===0)&&(i=[1]);let s=e.strides;(s.length===0||s[0]===0)&&(s=[1]);let o=e.pads;o.length===0&&(o=[0,0]),o=[0,o[0],0,o[1]],s=[1].concat(s),i=[1].concat(i),a=[1].concat(a);let l=Gs({...e,pads:o,strides:s,dilations:i,kernelShape:a},n);t.compute(Mo(n,l,u=>r?[u[0],u[2],u[3]]:[u[0],u[1],u[3]]))},Jf=(t,e)=>{Id(t.inputs,e),t.inputs[0].dims.length===3?zd(t,e):Od(t,t.inputs,e)}}),Pd,em,tm,Oy=J(()=>{$e(),Me(),ct(),Te(),Pd=(t,e,r,n)=>{let a=K.size(e),i=e.length,s=X("input",t,i),o=ge("output",t,i),l=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),u=K.normalizeAxis(l,i),p=h=>{let m=` i32(${s.indicesGet("inputIndices","uniforms.axis")}) `,d=Se("uniforms.input_shape","uniforms.axis",i),_=n.reverse?m+(n.exclusive?" + 1":""):"0",w=n.reverse?d:m+(n.exclusive?"":" + 1");return` + ${h.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(s,o)} + ${h.mainStart()} + ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var inputIndices = ${o.offsetToIndices("global_idx")}; + var sum = ${o.type.value}(0); + let first : i32 = ${_}; + let last : i32 = ${w}; + for (var i : i32 = first; i < last; i++) { + ${s.indicesSet("inputIndices","uniforms.axis","u32(i)")}; + sum = sum + ${s.getByIndices("inputIndices")}; + } + ${o.setByOffset("global_idx","sum")}; + }`};return{name:"CumSum",shaderCache:{hint:n.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:e,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},{type:12,data:u},...ye(e,e)]}),getShaderSource:p}},em=(t,e)=>{let r=t.inputs[0].dims,n=t.inputs[0].dataType,a=t.inputs[1];t.compute(Pd(n,r,a,e),{inputs:[0]})},tm=t=>{let e=t.exclusive===1,r=t.reverse===1;return Ke({exclusive:e,reverse:r})}}),Rd,Bd,Dd,rm,nm,zy=J(()=>{$e(),Me(),ct(),Te(),Rd=t=>{if(!t||t.length!==1)throw new Error("DepthToSpace requires 1 input.");if(t[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},Bd=(t,e,r,n)=>{let a=[];a.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} { + var a: ${r.type.indices};`);for(let i=0;i{let r,n,a,i,s,o,l=e.format==="NHWC",u=e.blocksize,p=e.mode==="DCR";l?([r,n,a,i]=t.dims,s=p?[r,n,a,u,u,i/u**2]:[r,n,a,i/u**2,u,u],o=p?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,n,a,i]=[t.dims[0],t.dims[2],t.dims[3],t.dims[1]],s=p?[r,u,u,i/u**2,n,a]:[r,i/u**2,u,u,n,a],o=p?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let h=t.reshape(s),m=h.dims.length,d=t.dataType,_=X("a",d,m),w=ge("output",d,m),v=S=>` + ${S.registerUniform("output_size","u32").declareVariables(_,w)} + + ${Bd(o,m,_,w)} + + ${S.mainStart()} + ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${w.offsetToIndices("global_idx")}; + let aIndices = perm(indices); + + ${w.setByOffset("global_idx",_.getByIndices("aIndices"))} + }`;return{name:"DepthToSpace",shaderCache:{hint:`${t.dims};${e.blocksize};${e.mode}`,inputDependencies:["rank"]},getRunData:S=>{let $=l?[r,n*u,a*u,i/u**2]:[r,i/u**2,n*u,a*u],E=K.size($),T=h.dims,A=K.sortBasedOnPerm(T,o);return{outputs:[{dims:$,dataType:S[0].dataType}],dispatchGroup:{x:Math.ceil(E/64)},programUniforms:[{type:12,data:E},...ye(T,A)]}},getShaderSource:v}},rm=(t,e)=>{Rd(t.inputs),t.compute(Dd(t.inputs[0],e))},nm=t=>Ke({blocksize:t.blocksize,mode:t.mode,format:t.format})}),Si,sa,Hs,Nd,Fd,Ld,Ud,js,Wd,am,im,Py=J(()=>{$e(),Me(),ct(),Te(),Si="[a-zA-Z]|\\.\\.\\.",sa="("+Si+")+",Hs="^"+sa+"$",Nd="("+sa+",)*"+sa,Fd="^"+Nd+"$",Ld=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,e){let r=this.symbolToIndices.get(t);r===void 0?r=[e]:r.push(e),this.symbolToIndices.set(t,r)}},Ud=class{constructor(t,e){this.equation=e,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,n]=e.includes("->")?e.split("->",2):[e,""];if(!r.match(RegExp(Fd)))throw new Error("Invalid LHS term");if(r.split(",").forEach((a,i)=>{let s=t[i].dims.slice();if(!a.match(RegExp(Hs)))throw new Error("Invalid LHS term");let o=this.processTerm(a,!0,s,i);this.lhs.push(o)}),n==="")n+=[...this.symbolToInfo.entries()].filter(([a,i])=>i.count===1||a==="...").map(([a])=>a).join("");else if(!n.match(RegExp(sa)))throw new Error("Invalid RHS");n.match(RegExp(Si,"g"))?.forEach(a=>{if(a==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let i=this.symbolToInfo.get(a);if(i===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(i.dimValue)}}),this.rhs=this.processTerm(n,!1,this.outputDims)}addSymbol(t,e,r){let n=this.symbolToInfo.get(t);if(n!==void 0){if(n.dimValue!==e&&n.count!==1)throw new Error("Dimension mismatch");n.count++,n.inputIndices.push(r)}else n={count:1,dimValue:e,inputIndices:[r]};this.symbolToInfo.set(t,n)}processTerm(t,e,r,n=-1){let a=r.length,i=!1,s=[],o=0;if(!t.match(RegExp(Hs))&&!e&&t!=="")throw new Error("Invalid LHS term");let l=t.match(RegExp(Si,"g")),u=new Ld(n);return l?.forEach((p,h)=>{if(p==="..."){if(i)throw new Error("Only one ellipsis is allowed per input term");i=!0;let m=a-l.length+1;if(m<0)throw new Error("Ellipsis out of bounds");if(s=r.slice(o,o+m),this.hasEllipsis){if(this.ellipsisDims.length!==s.length||this.ellipsisDims.toString()!==s.toString())throw new Error("Ellipsis dimensions mismatch")}else if(e)this.hasEllipsis=!0,this.ellipsisDims=s;else throw new Error("Ellipsis must be specified in the LHS");for(let d=0;dt+"_max",Wd=(t,e,r,n)=>{let a=t.map(u=>u.length).map((u,p)=>X(`input${p}`,e,u)),i=K.size(n),s=ge("output",e,n.length),o=[...r.symbolToInfo.keys()].filter(u=>!r.rhs.symbolToIndices.has(u)),l=u=>{let p=[],h="var prod = 1.0;",m="var sum = 0.0;",d="sum += prod;",_=[],w=[],v=[],S=[],$=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((T,A)=>{if(r.rhs.symbolToIndices.has(A)){let P=r.rhs.symbolToIndices.get(A)?.[0];P!==void 0&&r.lhs.forEach((B,D)=>{if(T.inputIndices.includes(D)){let q=B.symbolToIndices.get(A);if(q===void 0)throw new Error("Invalid symbol error");q.forEach(H=>{p.push(`${a[D].indicesSet(`input${D}Indices`,H,s.indicesGet("outputIndices",P))}`)})}})}else r.lhs.forEach((P,B)=>{if(T.inputIndices.includes(B)){let D=P.symbolToIndices.get(A);if(D===void 0)throw new Error("Invalid symbol error");D.forEach(q=>{_.push(`${a[B].indicesSet(`input${B}Indices`,q,`${A}`)}`)}),S.push(`prod *= ${a[B].getByIndices(`input${B}Indices`)};`)}}),w.push(`for(var ${A}: u32 = 0; ${A} < uniforms.${js(A)}; ${A}++) {`),v.push("}")});let E=$?[...p,`let sum = ${a.map((T,A)=>T.getByIndices(`input${A}Indices`)).join(" * ")};`]:[...p,m,...w,..._,h,...S,d,...v];return` + ${u.registerUniforms(o.map(T=>({name:`${js(T)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...a,s)} + + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var outputIndices = ${s.offsetToIndices("global_idx")}; + ${a.map((T,A)=>`var input${A}Indices: ${a[A].type.indices};`).join(` +`)} + ${E.join(` +`)}; + ${s.setByOffset("global_idx","sum")}; + }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:t.map(()=>"rank")},getRunData:()=>{let u=o.filter(h=>r.symbolToInfo.has(h)).map(h=>({type:12,data:r.symbolToInfo.get(h)?.dimValue||0}));u.push({type:12,data:i});let p=t.map((h,m)=>[...ye(h)]).reduce((h,m)=>h.concat(m),u);return p.push(...ye(n)),{outputs:[{dims:n,dataType:e}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:p}},getShaderSource:l}},am=(t,e)=>{let r=new Ud(t.inputs,e.equation),n=r.outputDims,a=t.inputs.map((i,s)=>i.dims);t.compute(Wd(a,t.inputs[0].dataType,r,n))},im=t=>{let e=t.equation.replace(/\s+/g,"");return Ke({equation:e})}}),Vd,qs,Gd,Hd,sm,Ry=J(()=>{$e(),Me(),Te(),Vd=t=>{if(!t||t.length!==2)throw new Error("Expand requires 2 input.");let e=t[0].dims,r=Array.from(t[1].getBigInt64Array(),Number),n=r.length{let r=t.length-e.length,n=[];for(let a=0;at.length>e.length?qs(t,e):qs(e,t),Hd=t=>{let e=t[0].dims,r=Array.from(t[1].getBigInt64Array(),Number),n=Gd(e,r),a=t[0].dataType,i=a===9?4:1,s=Math.ceil(K.size(n)/i),o=u=>{let p=X("input",a,e.length,i),h=ge("output",a,n.length,i),m;if(a===9){let d=(_,w,v="")=>` + let outputIndices${w} = ${h.offsetToIndices(`outputOffset + ${w}u`)}; + let offset${w} = ${p.broadcastedIndicesToOffset(`outputIndices${w}`,h)}; + let index${w} = offset${w} / 4u; + let component${w} = offset${w} % 4u; + ${_}[${w}] = ${v}(${p.getByOffset(`index${w}`)}[component${w}]); + `;m=` + let outputOffset = global_idx * ${i}; + var data = vec4(0); + ${d("data",0,"u32")} + ${d("data",1,"u32")} + ${d("data",2,"u32")} + ${d("data",3,"u32")} + ${h.setByOffset("global_idx","data")} + }`}else m=` + let outputIndices = ${h.offsetToIndices("global_idx")}; + let inputOffset = ${p.broadcastedIndicesToOffset("outputIndices",h)}; + ${h.setByOffset("global_idx",p.getByOffset("inputOffset"))} + }`;return` + ${u.registerUniform("vec_size","u32").declareVariables(p,h)} + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${m}`},l=[{type:12,data:s},...ye(e,n)];return{name:"Expand",shaderCache:{hint:`${n.length}`,inputDependencies:["rank"]},getShaderSource:o,getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:l})}},sm=t=>{Vd(t.inputs),t.compute(Hd(t.inputs),{inputs:[0]})}}),jd,om,By=J(()=>{$e(),Me(),Te(),tl(),jd=t=>{let e=t[0].dataType,r=K.size(t[0].dims),n=K.size(t[1].dims),a=n%4===0,i=s=>{let o=X("x",e,[1],4),l=X("bias",e,[1],4),u=ge("y",e,[1],4),p=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],h=d=>` + let bias${d}_offset: u32 = (global_idx * 4 + ${d}) % uniforms.bias_size; + let bias${d} = ${l.getByOffset(`bias${d}_offset / 4`)}[bias${d}_offset % 4];`,m=a?` + let bias = ${l.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${h(0)}${h(1)}${h(2)}${h(3)} + let bias = ${o.type.value}(bias0, bias1, bias2, bias3);`;return`${s.registerUniforms(p).declareVariables(o,l,u)} + + ${ko(Pt(e))} + + ${s.mainStart(Pn)} + ${s.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")} + + let x = ${o.getByOffset("global_idx")}; + ${m} + let x_in = x + bias; + ${u.setByOffset("global_idx",Eo("x_in"))} + }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${a}`,inputDependencies:["type","type"]},getShaderSource:i,getRunData:s=>({outputs:[{dims:s[0].dims,dataType:s[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:n}],dispatchGroup:{x:Math.ceil(r/Pn/4)}})}},om=t=>{t.inputs.length<2||K.size(t.inputs[1].dims)===0?Mf(t):t.compute(jd(t.inputs))}}),qd,Kd,lm,um,Dy=J(()=>{$e(),Me(),ct(),Te(),qd=t=>{if(!t||t.length!==2)throw new Error("Gather requires 2 inputs.")},Kd=(t,e)=>{let r=t[0].dims,n=t[1].dims,a=r.length,i=K.normalizeAxis(e.axis,a),s=r.slice(0);s.splice(i,1,...n);let o=r[i],l=t[0].dataType===9?4:1,u=Math.ceil(K.size(s)/l),p=[{type:12,data:u},{type:6,data:o},{type:12,data:i},...ye(t[0].dims,t[1].dims,s)],h=m=>{let d=X("data",t[0].dataType,t[0].dims.length,l),_=X("inputIndices",t[1].dataType,t[1].dims.length),w=ge("output",t[0].dataType,s.length,l),v=$=>{let E=n.length,T=`var indicesIndices${$} = ${_.type.indices}(0);`;for(let A=0;A1?`indicesIndices${$}[${A}]`:`indicesIndices${$}`} = ${s.length>1?`outputIndices${$}[uniforms.axis + ${A}]`:`outputIndices${$}`};`;T+=` + var idx${$} = ${_.getByIndices(`indicesIndices${$}`)}; + if (idx${$} < 0) { + idx${$} = idx${$} + uniforms.axisDimLimit; + } + var dataIndices${$} : ${d.type.indices}; + `;for(let A=0,P=0;A1?`dataIndices${$}[${A}]`:`dataIndices${$}`} = u32(idx${$});`,P+=E):(T+=`${a>1?`dataIndices${$}[${A}]`:`dataIndices${$}`} = ${s.length>1?`outputIndices${$}[${P}]`:`outputIndices${$}`};`,P++);return T},S;if(t[0].dataType===9){let $=(E,T,A="")=>` + let outputIndices${T} = ${w.offsetToIndices(`outputOffset + ${T}u`)}; + ${v(T)}; + let offset${T} = ${d.indicesToOffset(`dataIndices${T}`)}; + let index${T} = offset${T} / 4u; + let component${T} = offset${T} % 4u; + ${E}[${T}] = ${A}(${d.getByOffset(`index${T}`)}[component${T}]); + `;S=` + let outputOffset = global_idx * ${l}; + var value = vec4(0); + ${$("value",0,"u32")} + ${$("value",1,"u32")} + ${$("value",2,"u32")} + ${$("value",3,"u32")} + ${w.setByOffset("global_idx","value")} + `}else S=` + let outputIndices = ${w.offsetToIndices("global_idx")}; + ${v("")}; + let value = ${d.getByIndices("dataIndices")}; + ${w.setByOffset("global_idx","value")}; + `;return` + ${m.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(d,_,w)} + ${m.mainStart()} + ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + ${S} + }`};return{name:"Gather",shaderCache:{hint:e.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:s,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:p}),getShaderSource:h}},lm=t=>Ke({axis:t.axis}),um=(t,e)=>{let r=t.inputs;qd(r),t.compute(Kd(t.inputs,e))}}),Yd,Xd,dm,cm,Ny=J(()=>{$e(),Me(),ct(),Te(),Yd=t=>{if(!t||t.length!==2)throw new Error("GatherElements requires 2 inputs.");if(t[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(t[0].dims.length!==t[1].dims.length)throw new Error(`GatherElements requires that the data input and + indices input tensors be of same rank.`)},Xd=(t,e)=>{let r=t[0].dims,n=t[0].dataType,a=r.length,i=t[1].dims,s=t[1].dataType,o=K.normalizeAxis(e.axis,a),l=r[o],u=i.slice(0),p=K.size(u),h=X("input",n,a),m=X("indicesInput",s,i.length),d=ge("output",n,u.length),_=[{type:12,data:p},{type:6,data:l},{type:12,data:o}];return _.push(...ye(r,i,u)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:u,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:_}),getShaderSource:w=>` + ${w.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(h,m,d)} + ${w.mainStart()} + ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + + let outputIndices = ${d.offsetToIndices("global_idx")}; + + var idx = ${m.getByOffset("global_idx")}; + if (idx < 0) { + idx = idx + uniforms.axisDimLimit; + } + var inputIndices = ${h.type.indices}(outputIndices); + ${h.indicesSet("inputIndices","uniforms.axis","u32(idx)")}; + let value = ${h.getByIndices("inputIndices")}; + + ${d.setByOffset("global_idx","value")}; + }`}},dm=t=>Ke({axis:t.axis}),cm=(t,e)=>{let r=t.inputs;Yd(r),t.compute(Xd(t.inputs,e))}}),Qd,Zd,pm,hm,Fy=J(()=>{$e(),Me(),Te(),Qd=t=>{if(!t)throw new Error("Input is missing");if(t.length<2||t.length>3)throw new Error("Invaid input number.");if(t.length===3&&t[2].dims.length>2)throw new Error("Invalid input shape of C");if(t[0].dataType!==t[1].dataType||t.length===3&&t[0].dataType!==t[2].dataType)throw new Error("Input types are mismatched")},Zd=(t,e)=>{let r=t[0].dims.slice(),n=t[1].dims.slice(),[a,i,s]=bh.getShapeOfGemmResult(r,e.transA,n,e.transB,t.length===3?t[2].dims:void 0),o=[a,i];if(!o)throw new Error("Can't use gemm on the given tensors");let l=K.size(o),u=[{type:12,data:l},{type:12,data:a},{type:12,data:i},{type:12,data:s},{type:1,data:e.alpha},{type:1,data:e.beta}],p=["type","type"];t.length===3&&(u.push(...ye(t[2].dims)),p.push("rank")),u.push(...ye(o));let h=m=>{let d="";e.transA&&e.transB?d="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":e.transA&&!e.transB?d="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!e.transA&&e.transB?d="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!e.transA&&!e.transB&&(d="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let _=e.alpha===1?"":"value *= uniforms.alpha;",w=X("a",t[0].dataType,t[0].dims),v=X("b",t[1].dataType,t[1].dims),S=w.type.value,$=null,E=[w,v];t.length===3&&($=X("c",t[2].dataType,t[2].dims.length),E.push($));let T=ge("output",t[0].dataType,o.length);E.push(T);let A=[{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` + ${m.registerUniforms(A).declareVariables(...E)} + + ${m.mainStart()} + ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let m = global_idx / uniforms.N; + let n = global_idx % uniforms.N; + + var value = ${S}(0); + for (var k: u32 = 0u; k < uniforms.K; k++) { + ${d} + } + + ${_} + ${$!=null?`let cOffset = ${$.broadcastedIndicesToOffset("vec2(m, n)",T)}; value += ${S}(uniforms.beta) * ${$.getByOffset("cOffset")};`:""} + output[global_idx] = value; + }`};return{name:"Gemm",shaderCache:{hint:`${e.cacheKey}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u}),getShaderSource:h}},pm=t=>{let e=t.transA,r=t.transB,n=t.alpha,a=t.beta;return{transA:e,transB:r,alpha:n,beta:a,cacheKey:`${t.transA};${t.transB};${t.alpha===1}`}},hm=(t,e)=>{Qd(t.inputs),t.compute(Zd(t.inputs,e))}}),Jd,ec,tc,fm,Ly=J(()=>{$e(),Me(),Te(),Jd=(t,e)=>{let r=t[0].dims,n=r,a=2,i=K.sizeToDimension(r,a),s=K.sizeFromDimension(r,a),o=ot(s),l=s/o,u=[r[0],r[1],l],p=["rank","type","type"],h=[{type:12,data:s},{type:12,data:l}];h.push(...ye(u,u));let m=d=>{let _=X("x",t[0].dataType,u.length,o),w=X("scale",t[1].dataType,t[1].dims),v=X("bias",t[2].dataType,t[2].dims),S=ge("output",t[0].dataType,u.length,o),$=[_,w,v,S],E=_.type.value,T=o===1?"f32":`vec${o}`,A=64,P=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return` + var meanShared : f32; + var squaredNormShared : f32; + var workgroupShared : array<${T}, ${A}>; + const workgroupSize = ${A}u; + ${d.registerUniforms(P).declareVariables(...$)} + ${d.mainStart(A)} + let norm = global_idx / workgroupSize; + let batch = norm / uniforms.x_shape[1]; + let channel = norm % uniforms.x_shape[1]; + let localIndex = local_id.x; + + // initialize workgroup memory + var initial = ${T}(0); + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + initial = initial + ${T}(${_.get("batch","channel","h")}); + } + workgroupShared[localIndex] = initial; + workgroupBarrier(); + + // Calculate the mean of current channel data. + for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) { + if (localIndex < currSize) { + workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize]; + } + workgroupBarrier(); + } + if (localIndex == 0) { + meanShared = ${Dr("workgroupShared[0]",o)} / f32(uniforms.normSize); + } + workgroupBarrier(); + + // reinitialize workgroup memory. + initial = ${T}(0); + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + let deviation = ${T}(${_.get("batch","channel","h")}) - ${T}(meanShared); + initial = initial + deviation * deviation; + } + workgroupShared[localIndex] = initial; + workgroupBarrier(); + + // Calculate the sum of square of deviation of current channel data. + for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) { + if (localIndex < currSize) { + workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize]; + } + workgroupBarrier(); + } + if (localIndex == 0) { + squaredNormShared = ${Dr("workgroupShared[0]",o)}; + } + workgroupBarrier(); + + let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${e.epsilon})); + let channelScale = invStdDev * f32(${w.getByOffset("channel")}); + let channelShift = f32(${v.getByOffset("channel")}) - meanShared * channelScale; + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + let value = ${_.get("batch","channel","h")} * ${E}(${T}(channelScale)) + ${E}(${T}(channelShift)); + ${S.set("batch","channel","h","value")}; + } + }`};return{name:"InstanceNormalization",shaderCache:{hint:`${e.epsilon};${o}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:i},programUniforms:h}),getShaderSource:m}},ec=(t,e,r,n,a,i,s,o)=>{let l=ot(s),u=64,p=l===1?"vec2f":`mat2x${l}f`,h=l===1?"f32":`vec${l}f`,m=(P,B)=>`${p}(${P}, ${B})`,d=a*s/l,_=Math.ceil(i/u),w=["type"],v=[{type:12,data:_},{type:12,data:i},{type:12,data:Math.floor(s/l)},{type:12,data:Math.floor(i*s/l)}],S=P=>{let B=X("input",e.dataType,e.dims,l);return` + ${P.declareVariables(B)} + @group(0) @binding(1) var output : array<${p}>; + struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32}; + @group(0) @binding(2) var uniforms: Uniforms; + + ${P.mainStart(u)} + let currentImageNumber = global_idx / ${u} / uniforms.C; + let currentChannelNumber = (global_idx / ${u}) % uniforms.C; + let wgOffset = local_id.x * uniforms.wg_size; + if (wgOffset >= uniforms.H) { + return; + } + let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H); + + let offset = currentImageNumber * uniforms.image_size + currentChannelNumber; + var sum = ${Sr("f32",l)}; + var squaredSum = ${Sr("f32",l)}; + for (var i: u32 = wgOffset; i < wgMax; i++) { + let value = ${h}(input[offset + i * uniforms.C]); + sum += value; + squaredSum += value * value; + } + output[global_idx] = ${m("sum","squaredSum")}; + }`},$=t.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${l}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:[a,s,u,2],dataType:1}],dispatchGroup:{x:a*s/l},programUniforms:v}),getShaderSource:S},{inputs:[e],outputs:[-1]})[0],E=[{type:12,data:d},{type:12,data:i},{type:12,data:Math.floor(s/l)},{type:12,data:Math.floor(u*s/l)}],T=["type","type","type"],A=P=>{let B=X("scale",r.dataType,r.dims,l),D=X("bias",n.dataType,n.dims,l);return` + @group(0) @binding(0) var input : array<${p}>; + @group(0) @binding(1) var scale : array<${B.type.storage}>; + @group(0) @binding(2) var bias : array<${D.type.storage}>; + @group(0) @binding(3) var output : array<${p}>; + struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32}; + @group(0) @binding(4) var uniforms: Uniforms; + + ${P.mainStart()} + ${P.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.units_of_work")} + let currentImageNumber = global_idx / uniforms.C; + let currentChannelNumber = global_idx % uniforms.C; + + let offset = currentImageNumber * uniforms.image_size; + var sum = ${Sr("f32",l)}; + var squaredSum = ${Sr("f32",l)}; + for (var i: u32 = 0; i < min(${u}, uniforms.H); i++) { + let value = input[offset + i + currentChannelNumber * ${u}]; + sum += value[0]; + squaredSum += value[1]; + } + sum = sum / f32(uniforms.H); + squaredSum = squaredSum / f32(uniforms.H); + let invStdDev = inverseSqrt(squaredSum - sum * sum + f32(${o})); + let channelScale = invStdDev * ${h}(scale[currentChannelNumber]); + let channelShift = ${h}(bias[currentChannelNumber]) - sum * channelScale; + + output[global_idx] = ${m("channelScale","channelShift")}; + }`};return t.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${o}`,inputDependencies:T},getRunData:()=>({outputs:[{dims:[a,s,2],dataType:1}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:E}),getShaderSource:A},{inputs:[$,r,n],outputs:[-1]})[0]},tc=(t,e,r)=>{let n=e[0].dims,a=n,i=n[0],s=n[n.length-1],o=K.sizeFromDimension(n,1)/s,l=ot(s),u=K.size(a)/l,p=[{type:12,data:o},{type:12,data:Math.floor(s/l)}],h=["type","type"],m=ec(t,e[0],e[1],e[2],i,o,s,r.epsilon),d=_=>{let w=mt(e[0].dataType),v=l===1?"vec2f":`mat2x${l}f`,S=l===1?w:`vec${l}<${w}>`,$=X("input",e[0].dataType,e[0].dims,l),E=ge("output",e[0].dataType,a,l);return` + @group(0) @binding(0) var input : array<${$.type.storage}>; + @group(0) @binding(1) var scaleInput : array<${v}>; + @group(0) @binding(2) var output : array<${E.type.storage}>; + struct Uniforms {H: u32, C : u32}; + @group(0) @binding(3) var uniforms: Uniforms; + + ${_.mainStart()} + let currentImageNumber = global_idx / (uniforms.C * uniforms.H); + let currentChannelNumber = global_idx % uniforms.C; + + let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber; + let scale = scaleInput[scaleOffset]; + output[global_idx] = fma(input[global_idx], ${S}(scale[0]), ${S}(scale[1])); + }`};t.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${l}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:p}),getShaderSource:d},{inputs:[e[0],m]})},fm=(t,e)=>{e.format==="NHWC"?tc(t,t.inputs,e):t.compute(Jd(t.inputs,e))}}),rc,nc,mm,Uy=J(()=>{$e(),Me(),Te(),rc=t=>{if(!t||t.length<2)throw new Error("layerNorm requires at least 2 inputs.")},nc=(t,e,r)=>{let n=e.simplified,a=t[0].dims,i=t[1],s=!n&&t[2],o=a,l=K.normalizeAxis(e.axis,a.length),u=K.sizeToDimension(a,l),p=K.sizeFromDimension(a,l),h=K.size(i.dims),m=s?K.size(s.dims):0;if(h!==p||s&&m!==p)throw new Error(`Size of X.shape()[axis:] == ${p}. + Size of scale and bias (if provided) must match this. + Got scale size of ${h} and bias size of ${m}`);let d=[];for(let A=0;A1,$=r>2,E=A=>{let P=mt(t[0].dataType),B=[X("x",t[0].dataType,t[0].dims,_),X("scale",i.dataType,i.dims,_)];s&&B.push(X("bias",s.dataType,s.dims,_)),B.push(ge("output",t[0].dataType,o,_)),S&&B.push(ge("mean_data_output",1,d)),$&&B.push(ge("inv_std_output",1,d));let D=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return` + ${A.registerUniforms(D).declareVariables(...B)} + ${A.mainStart()} + ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")} + let offset = global_idx * uniforms.norm_size_vectorized; + var mean_vector = ${Sr("f32",_)}; + var mean_square_vector = ${Sr("f32",_)}; + + for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) { + let value = ${Mn(P,_,"x[h + offset]")}; + mean_vector += value; + mean_square_vector += value * value; + } + let mean = ${Dr("mean_vector",_)} / uniforms.norm_size; + let inv_std_dev = inverseSqrt(${Dr("mean_square_vector",_)} / uniforms.norm_size ${n?"":"- mean * mean"} + uniforms.epsilon); + + for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) { + let f32input = ${Mn(P,_,"x[j + offset]")}; + let f32scale = ${Mn(P,_,"scale[j]")}; + output[j + offset] = ${B[0].type.value}((f32input ${n?"":"- mean"}) * inv_std_dev * f32scale + ${s?`+ ${Mn(P,_,"bias[j]")}`:""} + ); + } + + ${S?"mean_data_output[global_idx] = mean":""}; + ${$?"inv_std_output[global_idx] = inv_std_dev":""}; + }`},T=[{dims:o,dataType:t[0].dataType}];return S&&T.push({dims:d,dataType:1}),$&&T.push({dims:d,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${_};${r};${n}`,inputDependencies:w},getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(u/64)},programUniforms:v}),getShaderSource:E}},mm=(t,e)=>{rc(t.inputs),t.compute(nc(t.inputs,e,t.outputCount))}}),ac,ic,gm,_m,Wy=J(()=>{$e(),Me(),ct(),Te(),ac=(t,e)=>{if(t.length<3||t.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=t[0],n=r.dims.length;if(r.dims[n-1]!==e.k)throw new Error("The last dim of input shape does not match the k value");let a=Math.floor((e.k+e.blockSize-1)/e.blockSize),i=e.blockSize/8*e.bits,s=t[1];if(!K.areEqual(s.dims,[e.n,a,i]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let o=t[2].dims;if(K.size(o)!==e.n*a)throw new Error("scales input size error.");if(t.length===4){let l=t[3].dims,u=e.bits>4?e.n*a:e.n*Math.floor((a+1)/2);if(K.size(l)!==u)throw new Error("zeroPoints input size error.")}},ic=(t,e,r,n)=>{let a=t[0].dims,i=a.length,s=Math.floor((e.k+e.blockSize-1)/e.blockSize),o=a[i-2],l=e.k,u=e.n,p=a.slice(0,i-2),h=K.size(p),m=e.blockSize/8*e.bits/4,d=t[0].dataType,_=ot(o),w=ot(e.k),v=ot(m),S=$a(d),$=o*s*S,E=Math.floor(n/$),T=s<=r[0]&&E>0,A=!T||E>=4?ot(u):E>=2&&ot(u)>=2?2:1,P=p.concat([o,u]),B=K.size(P)/A/_,D=T?[]:[{type:12,data:B},{type:12,data:e.blockSize}],q=[h,o,l/w],H=K.convertShape(t[1].dims).slice();H.splice(-1,1,m/v),D.push(...ye(q)),D.push(...ye(H)),D.push(...ye(t[2].dims)),t.length===4&&D.push(...ye(K.convertShape(t[3].dims)));let ie=[h,o,u/A];D.push(...ye(ie));let te=de=>{let se=q.length,M=X("a",t[0].dataType,se,w),R=X("b",12,H.length,v),G=X("scales",t[2].dataType,t[2].dims.length),re=[M,R,G],ee=t.length===4?X("zero_points",12,t[3].dims.length):void 0;ee&&re.push(ee);let ne=ie.length,W=ge("output",t[0].dataType,ne,A),ae=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],fe=mt(t[0].dataType),Ce=(()=>{switch(w){case 1:return`array<${fe}, 8>`;case 2:return`mat4x2<${fe}>`;case 4:return`mat2x4<${fe}>`;default:throw new Error(`${w}-component is not supported.`)}})(),Be=` + for (var word: u32 = 0; word < ${m}; word += ${v}) { + ${R.indicesSet("b_indices","2","word")}; + let b_data = ${R.getByIndices("b_indices")}; + for (var i: u32 = 0; i < ${v}; i++) { + let b_value: u32 = ${v===1?"b_data":"b_data[word + i]"}; + let b_mask: u32 = 0x0F0F0F0Fu; + let b_value_lower: vec4 = unpack4xU8(b_value & b_mask); + let b_value_upper: vec4 = unpack4xU8((b_value >> 4) & b_mask); + let b_quantized_values = ${Ce}(${Array.from({length:4},(Qe,We)=>`${fe}(b_value_lower[${We}]), ${fe}(b_value_upper[${We}])`).join(", ")}); + let b_dequantized_values = ${w===1?`${Ce}(${Array.from({length:8},(Qe,We)=>`(b_quantized_values[${We}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${Ce}(${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?o:_}u; m++) { + ${M.indicesSet("a_indices",se-2,T?"m":`row * ${_} + m`)}; + ${M.indicesSet("a_indices",se-1,"word_offset")}; + var input_offset = ${M.indicesToOffset("a_indices")}; + var a_data: ${Ce}; + for (var j: u32 = 0; j < ${8/w}; j++) { + a_data[j] = ${M.getByOffset("input_offset")}; + input_offset++; + } + ${T?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${A>1?"[c]":""} += ${Array.from({length:8/w},(Qe,We)=>`${w===1?`a_data[${We}] * b_dequantized_values[${We}]`:`dot(a_data[${We}], b_dequantized_values[${We}])`}`).join(" + ")}; + } + word_offset += ${8/w}; + } + }`,Ye=ee?` + zero_point_offset += 4; + if (zero_point_offset == 32) { + zero_point_offset = 0; + zero_point_index++; + zero_point_word = ${ee.getByOffset("zero_point_index")}; + }`:"";return T?` + var workgroup_shared: array<${W.type.value}, ${o*s}>; + ${de.declareVariables(...re,W)} + ${de.mainStart([s,1,1])} + var a_indices: ${M.type.indices}; + var block = local_id.x; + var col = workgroup_id.y; + var batch = workgroup_id.z; + ${M.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; + ${ee?` + var zero_point_bytes_per_col: u32 = (${s} + 1) / 2; + var zero_point_byte_count: u32 = col_times_components_plus_c * zero_point_bytes_per_col + (block >> 0x1u); + var zero_point_word_index: u32 = zero_point_byte_count >> 0x2u; + var zero_point_byte_offset: u32 = zero_point_byte_count & 0x3u; + 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 = ${ee.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;`:""} + var b_indices: ${R.type.indices}; + ${R.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 * ${s} + block; + let scale = ${G.getByOffset("scales_index")}; + // The default zero point is 8 for unsigned 4-bit quantization. + let zero_point = ${fe}(${ee?"(zero_point_word) & 0xFu":8}); + ${R.indicesSet("b_indices","1","block")}; + var word_offset: u32 = block * ${e.blockSize/w}; + var workgroup_shared_offset: u32 = block * ${o}; + ${Be} + } + workgroupBarrier(); + if (local_id.x == 0u) { + var output_indices: ${W.type.indices}; + ${W.indicesSet("output_indices","0","batch")}; + ${W.indicesSet("output_indices",ne-1,"col")}; + ${W.indicesSet("output_indices",ne-2,"0")}; + var output_offset = ${W.indicesToOffset("output_indices")}; + for (var m: u32 = 0u; m < ${o}u; m++) { + var output_value: ${W.type.value} = ${W.type.value}(0); + var workgroup_shared_offset: u32 = m; + for (var b: u32 = 0u; b < ${s}u; b++) { + output_value += workgroup_shared[workgroup_shared_offset]; + workgroup_shared_offset += ${o}; + } + ${W.setByOffset("output_offset","output_value")}; + output_offset += ${u/A}; + } + } + }`:` + ${de.registerUniforms(ae).declareVariables(...re,W)} + ${de.mainStart()} + ${de.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + var output_values: array<${W.type.value}, ${_}>; + var output_indices = ${W.offsetToIndices("global_idx")}; + var col = ${W.indicesGet("output_indices",ne-1)}; + var row = ${W.indicesGet("output_indices",ne-2)}; + var a_indices: ${M.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 + ${ee?` + var zero_point_abs_offset = col * ${A} * ((${s} + 1) / 2); + var zero_point_index: u32 = zero_point_abs_offset / 4; + var zero_point_word: u32 = ${ee.getByOffset("zero_point_index")}; + var zero_point_offset: u32 = (zero_point_abs_offset % 4) * 8;`:""} + var scale_index = col * ${s*A}; + var b_indices: ${R.type.indices}; + for (var c: u32 = 0; c < ${A}; c++) { + ${R.indicesSet("b_indices","0",`col * ${A} + c`)}; + var block_offset: u32 = 0; + for (var block: u32 = 0; block < ${s}; block++) { + // The scale and zero points are computed per block. + let scale = ${G.getByOffset("scale_index")}; + // The default zero point is 8 for unsigned 4-bit quantization. + let zero_point = ${fe}(${ee?"extractBits(zero_point_word, zero_point_offset, 4)":8}); + ${R.indicesSet("b_indices","1","block")}; + var word_offset: u32 = block_offset; + ${Be} + scale_index++; + ${Ye} + block_offset += uniforms.block_size / ${w}; + } + // Drop the trailing 4 bits if the zero_poit_offset is not a byte boundary to align with the next byte. + ${ee?`if (zero_point_offset % 8 > 0) { + ${Ye} + }`:""} + } + for (var k: u32 = 0u; k < ${_}u; k++) { + ${W.indicesSet("output_indices",ne-2,`${_} * row + k`)}; + ${W.setByIndices("output_indices","output_values[k]")} + } + }`};return{name:T?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${e.cacheKey};${o};${d};${t.length}`,inputDependencies:Array(t.length).fill("rank")},getRunData:()=>({outputs:[{dims:P,dataType:d}],name:T?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:T?{x:1,y:Math.ceil(u/A),z:h}:{x:Math.ceil(B/64)},programUniforms:D}),getShaderSource:te}},gm=(t,e)=>{ac(t.inputs,e);let r=t.getMaxComputeWorkgroupSizes(),n=t.getMaxComputeWorkgroupStoragesize();t.compute(ic(t.inputs,e,r,n))},_m=t=>Ke(t)}),St,sc,ym,Ks,oc,ki,wm,Vy=J(()=>{$e(),Me(),ct(),Xo(),Xh(),Te(),Ia(),St=(t,e)=>t.length>e&&t[e].dims.length>0&&K.size(t[e].dims)>0?t[e]:void 0,sc=(t,e)=>{let r=t[0],n=St(t,1),a=St(t,2),i=St(t,3),s=St(t,4),o=St(t,5),l=St(t,6),u=St(t,7);if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let p=!1,h=r.dims[0],m=r.dims[1],d=r.dims.length===3?p?r.dims[2]/3:r.dims[2]:e.numHeads*r.dims[4],_=m,w=0,v=0,S=Math.floor(d/e.numHeads);if(l&&u){if(l.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(l.dims[0]!==h||l.dims[1]!==e.numHeads||l.dims[3]!==S)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(u.dims[0]!==h||u.dims[1]!==e.numHeads||u.dims[3]!==S)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(l.dims[2]!==u.dims[2])throw new Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(u.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');w=l.dims[2],v=l.dims[2]}else if(l||u)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,_=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==e.numHeads||n.dims[3]!==2||n.dims[4]!==S)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(a)throw new Error('Expect "value" be none when "key" has packed kv format.');$=5,_=n.dims[1]}else{if(n.dims[1]!==e.numHeads||n.dims[3]!==S)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,_=n.dims[2]}}else{if(r.dims.length!==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]!==e.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');$=3}if(i){if(i.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(a&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let E=0;if(s){E=8;let D=s.dims;throw D.length===1?D[0]===h?E=1:D[0]===3*h+2&&(E=3):D.length===2&&D[0]===h&&D[1]===_&&(E=5),E===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)'):new Error("Mask not supported")}let T=!1,A=d;if(a){if(a.dims.length!==3&&a.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==a.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(a.dims.length===3){if(_!==a.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');A=a.dims[2]}else{if(_!==a.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');A=a.dims[1]*a.dims[3],T=!0}}let P=w+_,B=!1;if(s)throw new Error("Key padding mask is not supported");if(o){if(o.dims.length!==4)throw new Error('Input "relative_position_bias" is expected to have 4 dimensions');if(o.dims[0]!==h&&o.dims[0]!==1||o.dims[1]!==e.numHeads||o.dims[2]!==m||o.dims[3]!==P)throw new Error('Input "relative_position_bias" shape (batch_size, 1, sequence_length, kv_sequence_length)')}return{batchSize:h,sequenceLength:m,pastSequenceLength:w,kvSequenceLength:_,totalSequenceLength:P,maxSequenceLength:v,inputHiddenSize:0,hiddenSize:d,vHiddenSize:A,headSize:S,vHeadSize:Math.floor(A/e.numHeads),numHeads:e.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:e.maskFilterValue,maskType:E,scale:e.scale,broadcastResPosBias:B,passPastInKv:T,qkvFormat:$}},ym=t=>Ke({...t}),Ks=Ke({perm:[0,2,1,3]}),oc=(t,e,r,n,a,i,s)=>{let o=[n,a,i],l=K.size(o),u=[{type:12,data:l},{type:12,data:s},{type:12,data:i}],p=h=>{let m=ge("qkv_with_bias",e.dataType,o),d=X("qkv",e.dataType,o),_=X("bias",r.dataType,o),w=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return` + ${h.registerUniforms(w).declareVariables(d,_,m)} + ${h.mainStart()} + ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset; + + qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx]; + }`};return t.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:o,dataType:e.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u}),getShaderSource:p},{inputs:[e,r],outputs:[-1]})[0]},ki=(t,e,r,n,a,i,s,o)=>{let l=i;if(s){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=oc(t,i,s,e,n,r*a,o),l=l.reshape([e,n,r,a]),t.compute(Er(l,Ks.perm),{inputs:[l],outputs:[-1]})[0]}else return i.dims.length===3&&(l=i.reshape([e,n,r,a])),t.compute(Er(l,Ks.perm),{inputs:[l],outputs:[-1]})[0]},wm=(t,e)=>{let r=sc(t.inputs,e),n=t.inputs[0],a=St(t.inputs,1),i=St(t.inputs,2),s=St(t.inputs,3),o=St(t.inputs,4),l=St(t.inputs,5),u=St(t.inputs,6),p=St(t.inputs,7);if(n.dims.length===5)throw new Error("Packed QKV is not implemented");if(a?.dims.length===5)throw new Error("Packed KV is not implemented");let h=a&&i&&a.dims.length===4&&i.dims.length===4,m=ki(t,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,n,s,0);if(h)return Ki(t,m,a,i,o,void 0,u,p,l,r,e);if(!a||!i)throw new Error("key and value must be provided");let d=ki(t,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,a,s,r.hiddenSize),_=ki(t,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,s,2*r.hiddenSize);Ki(t,m,d,_,o,void 0,u,p,l,r,e)}}),lc,uc,dc,cc,pc,hc,fc,mc,bm,Gy=J(()=>{$e(),Me(),Te(),lc=t=>{if(!t||t.length<1)throw new Error("Too few inputs");if(t[0].dataType!==1&&t[0].dataType!==10)throw new Error("Input type must be float or float16.");if(t.length>=2){let e=t[0].dims.length*2===t[1].dims[0];if(t.length===4&&(e=t[3].dims[0]*2===t[1].dims[0]),!e)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},uc=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=` + k = i32(${t.indicesGet("indices",a)}) - ${Se("uniforms.pads",a,r)}; + if (k < 0) { + break; + } + if (k >= i32(${Se("uniforms.x_shape",a,e)})) { + break; + } + offset += k * i32(${Se("uniforms.x_strides",a,e)}); + `;return` + value = ${t.type.value}(uniforms.constant_value); + for (var i = 0; i < 1; i++) { + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + } + `},dc=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=` + k = i32(${t.indicesGet("indices",a)}) - ${Se("uniforms.pads",a,r)}; + if (k < 0) { + k = -k; + } + { + let _2n_1 = 2 * (i32(${Se("uniforms.x_shape",a,e)}) - 1); + k = k % _2n_1; + if(k >= i32(${Se("uniforms.x_shape",a,e)})) { + k = _2n_1 - k; + } + } + offset += k * i32(${Se("uniforms.x_strides",a,e)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},cc=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=` + k = i32(${t.indicesGet("indices",a)}) - ${Se("uniforms.pads",a,r)}; + if (k < 0) { + k = 0; + } + if (k >= i32(${Se("uniforms.x_shape",a,e)})) { + k = i32(${Se("uniforms.x_shape",a,e)}) - 1; + } + offset += k * i32(${Se("uniforms.x_strides",a,e)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},pc=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=` + k = i32(${t.indicesGet("indices",a)}) - ${Se("uniforms.pads",a,r)}; + if (k < 0) { + k += i32(${Se("uniforms.x_shape",a,e)}]); + } + if (k >= i32(${Se("uniforms.x_shape",a,e)})) { + k -= i32(${Se("uniforms.x_shape",a,e)}); + } + offset += k * i32(${Se("uniforms.x_strides",a,e)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},hc=(t,e,r)=>{switch(r.mode){case 0:return uc(t,e,r.pads.length);case 1:return dc(t,e,r.pads.length);case 2:return cc(t,e,r.pads.length);case 3:return pc(t,e,r.pads.length);default:throw new Error("Invalid mode")}},fc=(t,e)=>{let r=K.padShape(t[0].dims.slice(),e.pads),n=t[0].dims,a=K.size(r),i=[{type:12,data:a},{type:6,data:e.pads}];e.mode===0&&i.push({type:t[0].dataType,data:e.value}),i.push(...ye(t[0].dims,r));let s=["rank"],o=l=>{let u=ge("output",t[0].dataType,r.length),p=X("x",t[0].dataType,n.length),h=p.type.value,m=hc(u,n.length,e),d=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:e.pads.length}];return e.mode===0&&d.push({name:"constant_value",type:h}),` + ${l.registerUniforms(d).declareVariables(p,u)} + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${u.offsetToIndices("global_idx")}; + + var value = ${h}(0); + ${m} + output[global_idx] = value; + }`};return{name:"Pad",shaderCache:{hint:`${e.mode}`,inputDependencies:s},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(K.size(r)/64)},programUniforms:i}),getShaderSource:o}},mc=(t,e)=>{if(t.length>1){let r=t[1].getBigInt64Array(),n=t.length>=3&&t[2].data?t[2].getFloat32Array()[0]:0,a=t[0].dims.length,i=new Int32Array(2*a).fill(0);if(t.length>=4){let o=t[3].getBigInt64Array();for(let l=0;li[Number(l)]=Number(o));let s=[];return i.forEach(o=>s.push(o)),{mode:e.mode,value:n,pads:s}}else return e},bm=(t,e)=>{lc(t.inputs);let r=mc(t.inputs,e);t.compute(fc(t.inputs,r),{inputs:[0]})}}),oa,Ys,Xs,Qs,Zs,gc,_c,Js,eo,vm,$m,to,xm,Sm,ro,km,Em,Cm,Tm,Hy=J(()=>{dr(),$e(),Me(),Te(),oa=t=>{if(Ue.webgpu.validateInputContent&&(!t||t.length!==1))throw new Error("Pool ops requires 1 input.")},Ys=(t,e,r)=>{let n=e.format==="NHWC",a=t.dims.slice();n&&a.splice(1,0,a.pop());let i=Object.hasOwnProperty.call(e,"dilations"),s=e.kernelShape.slice(),o=e.strides.slice(),l=i?e.dilations.slice():[],u=e.pads.slice();Hi.adjustPoolAttributes(r,a,s,o,l,u);let p=Hi.computePoolOutputShape(r,a,o,l,s,u,e.autoPad),h=Object.assign({},e);i?Object.assign(h,{kernelShape:s,strides:o,pads:u,dilations:l,cacheKey:e.cacheKey}):Object.assign(h,{kernelShape:s,strides:o,pads:u,cacheKey:e.cacheKey});let m=p.slice();return m.push(m.splice(1,1)[0]),[h,n?m:p]},Xs=(t,e)=>{let r=e.format==="NHWC",n=K.size(t),a=K.size(e.kernelShape),i=[{type:12,data:n},{type:12,data:a}],s=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(e.kernelShape.length<=2){let o=e.kernelShape[e.kernelShape.length-1],l=e.strides[e.strides.length-1],u=e.pads[e.pads.length/2-1],p=e.pads[e.pads.length-1],h=!!(u+p);i.push({type:12,data:o},{type:12,data:l},{type:12,data:u},{type:12,data:p}),s.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let m=!1;if(e.kernelShape.length===2){let d=e.kernelShape[e.kernelShape.length-2],_=e.strides[e.strides.length-2],w=e.pads[e.pads.length/2-2],v=e.pads[e.pads.length-2];m=!!(w+v),i.push({type:12,data:d},{type:12,data:_},{type:12,data:w},{type:12,data:v}),s.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[i,s,!0,h,m]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let o=K.computeStrides(e.kernelShape);i.push({type:12,data:o},{type:12,data:e.pads},{type:12,data:e.strides}),s.push({name:"kernelStrides",type:"u32",length:o.length},{name:"pads",type:"u32",length:e.pads.length},{name:"strides",type:"u32",length:e.strides.length});let l=e.pads.reduce((u,p)=>u+p);return[i,s,!!l,!1,!1]}},Qs=(t,e,r,n,a,i,s,o,l,u,p,h)=>{let m=a.format==="NHWC",d=e.type.value,_=ge("output",e.type.tensor,n);if(a.kernelShape.length<=2){let w="",v="",S="",$=r-(m?2:1);if(p?w=` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i; + if (xIndices[${$}] < 0 || xIndices[${$}] + >= uniforms.x_shape[${$}]) { + pad++; + continue; + } + let x_val = x[${e.indicesToOffset("xIndices")}]; + ${i} + }`:w=` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i; + let x_val = x[${e.indicesToOffset("xIndices")}]; + ${i} + }`,a.kernelShape.length===2){let E=r-(m?3:2);h?v=` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${E}] = indices[${E}] * uniforms.sh - uniforms.phStart + j; + if (xIndices[${E}] < 0 || xIndices[${E}] >= uniforms.x_shape[${E}]) { + pad += i32(uniforms.kw); + continue; + } + `:v=` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${E}] = indices[${E}] * uniforms.sh - uniforms.phStart + j; + `,S=` + } + `}return` + ${t.registerUniforms(l).declareVariables(e,_)} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + + let indices = ${_.offsetToIndices("global_idx")}; + var xIndices = ${_.offsetToIndices("global_idx")}; + + var value = ${d}(${o}); + var pad = 0; + ${v} + ${w} + ${S} + ${s} + + output[global_idx] = value; + }`}else{if(m)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let w=a.kernelShape.length,v=a.pads.length,S="";return u?S=` + if (xIndices[j] >= uniforms.x_shape[j]) { + pad++; + isPad = true; + break; + } + } + if (!isPad) { + let x_val = x[${e.indicesToOffset("xIndices")}]; + ${i} + }`:S=` + } + let x_val = x[${e.indicesToOffset("xIndices")}]; + ${i} + `,` + ${t.registerUniforms(l).declareVariables(e,_)} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + let indices = ${_.offsetToIndices("global_idx")}; + var xIndices = ${_.offsetToIndices("global_idx")}; + + var offsets: array; + + var value = ${d}(${o}); + var pad = 0; + var isPad = false; + + for (var i: u32 = 0u; i < uniforms.kernelSize; i++) { + var offset = i; + for (var j = 0u; j < ${w-1}u; j++) { + offsets[j] = offset / ${Se("uniforms.kernelStrides","j",w)}; + offset -= offsets[j] * ${Se("uniforms.kernelStrides","j",w)}; + } + offsets[${w-1}] = offset; + + isPad = false; + for (var j = ${r-w}u; j < ${r}u; j++) { + xIndices[j] = indices[j] * ${Se("uniforms.strides",`j - ${r-w}u`,w)} + + offsets[j - ${r-w}u] - ${Se("uniforms.pads","j - 2u",v)}; + ${S} + } + ${s} + + output[global_idx] = value; + }`}},Zs=t=>`${t.format};${t.ceilMode};${t.autoPad};${t.kernelShape.length}`,gc=t=>`${Zs(t)};${t.countIncludePad}`,_c=t=>`${Zs(t)};${t.storageOrder};${t.dilations}`,Js=t=>({format:t.format,autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][t.auto_pad],ceilMode:t.ceil_mode,kernelShape:t.kernel_shape,strides:t.strides,pads:t.pads}),eo=(t,e,r,n)=>{let[a,i]=Ys(e,n,r),s=X("x",e.dataType,e.dims.length),o=s.type.value,l="value += x_val;",u="";a.countIncludePad?u+=`value /= ${o}(uniforms.kernelSize);`:u+=`value /= ${o}(i32(uniforms.kernelSize) - pad);`;let[p,h,m,d,_]=Xs(i,a);p.push(...ye(e.dims,i));let w=["rank"];return{name:t,shaderCache:{hint:`${n.cacheKey};${m};${d};${_}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(K.size(i)/64)},programUniforms:p}),getShaderSource:v=>Qs(v,s,e.dims.length,i.length,a,l,u,0,h,m,d,_)}},vm=t=>{let e=t.count_include_pad!==0,r=Js(t);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let n={countIncludePad:e,...r,cacheKey:""};return{...n,cacheKey:gc(n)}},$m=(t,e)=>{oa(t.inputs),t.compute(eo("AveragePool",t.inputs[0],!1,e))},to={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},xm=t=>{let e=t.format;return{format:e,...to,cacheKey:e}},Sm=(t,e)=>{oa(t.inputs),t.compute(eo("GlobalAveragePool",t.inputs[0],!0,e))},ro=(t,e,r,n)=>{let[a,i]=Ys(e,n,r),s=` + value = max(x_val, value); + `,o="",l=X("x",e.dataType,e.dims.length),u=["rank"],[p,h,m,d,_]=Xs(i,a);return p.push(...ye(e.dims,i)),{name:t,shaderCache:{hint:`${n.cacheKey};${m};${d};${_}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(K.size(i)/64)},programUniforms:p}),getShaderSource:w=>Qs(w,l,e.dims.length,i.length,a,s,o,e.dataType===10?-65504:-1e5,h,m,d,_)}},km=(t,e)=>{oa(t.inputs),t.compute(ro("MaxPool",t.inputs[0],!1,e))},Em=t=>{let e=t.storage_order,r=t.dilations,n=Js(t);if(e!==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 a={storageOrder:e,dilations:r,...n,cacheKey:""};return{...a,cacheKey:_c(a)}},Cm=t=>{let e=t.format;return{format:e,...to,cacheKey:e}},Tm=(t,e)=>{oa(t.inputs),t.compute(ro("GlobalMaxPool",t.inputs[0],!0,e))}}),yc,wc,Am,jy=J(()=>{dr(),$e(),Te(),yc=(t,e,r)=>{let n=t===e,a=te&&r>0;if(n||a||i)throw new Error("Range these inputs' contents are invalid.")},wc=(t,e,r,n)=>{let a=Math.abs(Math.ceil((e-t)/r)),i=[a],s=a,o=[{type:12,data:s},{type:n,data:t},{type:n,data:r},...ye(i)],l=u=>{let p=ge("output",n,i.length),h=p.type.value,m=[{name:"outputSize",type:"u32"},{name:"start",type:h},{name:"delta",type:h}];return` + ${u.registerUniforms(m).declareVariables(p)} + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + output[global_idx] = uniforms.start + ${h}(global_idx) * uniforms.delta; + }`};return{name:"Range",shaderCache:{hint:`${n}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:i,dataType:n}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:o})}},Am=t=>{let e=0,r=0,n=0;t.inputs[0].dataType===6?(e=t.inputs[0].getInt32Array()[0],r=t.inputs[1].getInt32Array()[0],n=t.inputs[2].getInt32Array()[0]):t.inputs[0].dataType===1&&(e=t.inputs[0].getFloat32Array()[0],r=t.inputs[1].getFloat32Array()[0],n=t.inputs[2].getFloat32Array()[0]),Ue.webgpu.validateInputContent&&yc(e,r,n),t.compute(wc(e,r,n,t.inputs[0].dataType),{inputs:[]})}}),bc,vc,$c,xc,Sc,kc,Ec,Cc,Tc,Ac,Ic,no,Mc,Oc,zc,Pc,Rc,Im,Mm,qy=J(()=>{$e(),Me(),ct(),Te(),bc=(t,e)=>{if(t.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),t.length>0){if(e.mode==="linear"){if(!(t.length===2||t.length===3||t.length===4&&t[0]===1&&t[1]===1||t.length===4&&t[0]===1&&t[3]===1||t.length===5&&t[0]===1&&t[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(e.mode==="cubic"&&!(t.length===2||t.length===4&&t[0]===1&&t[1]===1||t.length===4&&t[0]===1&&t[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},vc=(t,e,r)=>{e.every(a=>a>=0&&a{throw new Error("Resize requires axes input values to be positive and less than rank")}));let n=new Array(r).fill(1);return e.forEach((a,i)=>n[a]=t[i]),n},$c=(t,e,r,n,a,i)=>{let[s,o,l]=r>10?[1,2,3]:[-1,t.length>1?1:-1,-1],u=t[0].dims.length;if(s>0&&t.length>s&&t[s].dims.length>0)t[s].getFloat32Array().forEach(p=>i.push(p));else if(e.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(o>0&&t.length>o&&t[o].dims.length>0){if(t[o].getFloat32Array().forEach(p=>n.push(p)),n.length!==0&&n.length!==u&&r>=18&&n.length!==e.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");bc(n,e),e.axes.length>0&&vc(n,e.axes,u).forEach((p,h)=>n[h]=p)}if(l>0&&t.length>l&&(t[l].getBigInt64Array().forEach(p=>a.push(Number(p))),a.length!==u||r>=18&&a.length===e.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(e.axes.length>0){if(n.length!==e.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(a.length!==e.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 a<"u"&&n.length>0&&a.length>u)throw new Error("Resize requires only of scales or sizes to be specified")},xc=(t,e)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32, + lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${e} { `+(()=>{switch(t){case"asymmetric":return`return ${e}(xResized) / ${e}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) { + return (${e}(xResized) + 0.5) / ${e}(xScale) - 0.5; + } else { + return 0.0; + }`;case"tf_half_pixel_for_nn":return`return (${e}(xResized) + 0.5) / ${e}(xScale);`;case"align_corners":return`if (lengthResized == 1) { + return 0.0; + } else { + // The whole part and the fractional part are calculated separately due to inaccuracy of floating + // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an + // offset-by-one error later in floor(). + let whole = ${e}(xResized * (lengthOriginal - 1) / (lengthResized - 1)); + let fract = + ${e}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${e}(lengthResized - 1); + return whole + fract; + }`;case"tf_crop_and_resize":return`if (lengthResized > 1) { + return ${e}(roiStart) * ${e}(lengthOriginal - 1) + + (${e}(xResized) * ${e}(roiEnd - roiStart) * ${e}(lengthOriginal - 1)) / + ${e}(lengthResized - 1); + } else { + return 0.5 * ${e}(roiStart + roiEnd) * ${e}(lengthOriginal - 1); + }`;case"half_pixel_symmetric":return`const outputWidth = ${e}xScale * ${e}(lengthResized); + const adjustment = ${e}(lengthResized) / outputWidth; + const center = ${e}(lengthOriginal) / 2; + const offset = center * (1 - adjustment); + return offset + ((${e}(xResized) + 0.5) / ${e}(xScale)) - 0.5;`;case"half_pixel":return`return ((${e}(xResized) + 0.5) / ${e}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${t} is not supported`)}})()+"}",Sc=(t,e,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(t){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(e<11)return"if (isDownSample) { return ceil(xOriginal); } else { return xOriginal; }";throw new Error(`Nearest mode ${t} is not supported`)}})()+"}",kc=(t,e,r)=>{let n=new Array(r).fill(0).concat(new Array(r).fill(1)),a=t.length===0?n:t.slice();return e.length>0?(e.forEach((i,s)=>{n[i]=a[s],n[s+r]=a[e.length+s]}),n):a},Ec=(t,e,r,n)=>{let a=[];if(r.length>0)if(n.length>0){if(t.forEach(i=>a.push(i)),Math.max(...n)>t.length)throw new Error("axes is out of bound");n.forEach((i,s)=>a[i]=r[s])}else r.forEach(i=>a.push(i));else{if(e.length===0)throw new Error("Resize requires either scales or sizes.");a=t.map((i,s)=>Math.round(i*e[s]))}return a},Cc=(t,e,r)=>{let n=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(i=>e[i]),Number.MAX_VALUE):Math.min(...e,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(i=>e[i]),Number.MIN_VALUE):Math.max(...e,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();e.fill(1,0,e.length);let a=t.slice();return r.axes.length>0?(r.axes.forEach(i=>e[i]=n),r.axes.forEach(i=>a[i]=Math.round(t[i]*e[i]))):(e.fill(n,0,e.length),a.forEach((i,s)=>a[s]=Math.round(i*e[s]))),a},Tc=(t,e,r,n,a)=>` + fn calculateOriginalIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> array<${t.type.value}, ${r.length}> { + var original_indices: array<${t.type.value}, ${r.length}>; + for (var i:u32 = 0; i < ${r.length}; i++) { + var output_index = ${t.indicesGet("output_indices","i")}; + var scale = ${Se("uniforms.scales","i",n)}; + var roi_low = ${Se("uniforms.roi","i",a)}; + var roi_hi = ${Se("uniforms.roi",`i + ${e.length}`,a)}; + if (scale == 1.0) { + original_indices[i] = ${t.type.value}(output_index); + } else { + var input_shape_i = ${Se("uniforms.input_shape","i",e.length)}; + var output_shape_i = ${Se("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; + }`,Ac=(t,e,r,n,a,i,s)=>` + fn calculateInputIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> ${t.type.indices} { + var input_indices: ${t.type.indices}; + for (var i:u32 = 0; i < ${n.length}; i++) { + var output_index = ${e.indicesGet("output_indices","i")}; + var input_index: u32; + var scale = ${Se("uniforms.scales","i",a)}; + if (scale == 1.0) { + input_index = output_index; + } else { + var roi_low = ${Se("uniforms.roi","i",i)}; + var roi_hi = ${Se("uniforms.roi",`i + ${r.length}`,i)}; + var input_shape_i = ${Se("uniforms.input_shape","i",r.length)}; + var output_shape_i = ${Se("uniforms.output_shape","i",n.length)}; + var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + if (!${s} || (original_idx >= 0 && original_idx < ${e.type.value}(input_shape_i))) { + if (original_idx < 0) { + input_index = 0; + } else if (original_idx > ${e.type.value}(input_shape_i - 1)) { + input_index = input_shape_i - 1; + } else { + input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1)); + } + } else { + input_index = u32(original_idx); + } + } + ${t.indicesSet("input_indices","i"," input_index")} + } + return input_indices; + }`,Ic=(t,e)=>` + fn checkInputIndices(input_indices: ${t.type.indices}) -> bool { + for (var i:u32 = 0; i < ${e.length}; i++) { + var input_index = ${t.indicesGet("input_indices","i")}; + if (input_index < 0 || input_index >= ${Se("uniforms.input_shape","i",e.length)}) { + return false; + } + } + return true; + }`,no=(t,e,r,n)=>t.rank>n?` + ${t.indicesSet("input_indices",e,"channel")}; + ${t.indicesSet("input_indices",r,"batch")}; +`:"",Mc=(t,e,r,n,a)=>{let[i,s,o,l]=r.length===2?[-1,0,1,-1]:[0,2,3,1],u=t.type.value;return` + fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${u} { + var input_indices: ${t.type.indices}; + ${t.indicesSet("input_indices",s,`max(0, min(row, ${r[s]} - 1))`)}; + ${t.indicesSet("input_indices",o,`max(0, min(col, ${r[o]} - 1))`)}; + ${no(t,l,i,2)} + return ${t.getByIndices("input_indices")}; + } + + fn bilinearInterpolation(output_indices: ${e.type.indices}) -> ${u} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var row:${u} = originalIndices[${s}]; + var col:${u} = originalIndices[${o}]; + ${n?`if (row < 0 || row > (${r[s]} - 1) || col < 0 || col > (${r[o]} - 1)) { + return ${a}; + }`:""}; + row = max(0, min(row, ${r[s]} - 1)); + col = max(0, min(col, ${r[o]} - 1)); + var row1: u32 = u32(row); + var col1: u32 = u32(col); + var row2: u32 = u32(row + 1); + var col2: u32 = u32(col + 1); + var channel: u32 = ${r.length>2?`u32(originalIndices[${l}])`:"0"}; + var batch: u32 = ${r.length>2?`u32(originalIndices[${i}])`:"0"}; + var x11: ${u} = getInputValue(batch, channel, row1, col1); + var x12: ${u} = getInputValue(batch, channel, row1, col2); + var x21: ${u} = getInputValue(batch, channel, row2, col1); + var x22: ${u} = getInputValue(batch, channel, row2, col2); + var dx1: ${u} = abs(row - ${u}(row1)); + var dx2: ${u} = abs(${u}(row2) - row); + var dy1: ${u} = abs(col - ${u}(col1)); + var dy2: ${u} = abs(${u}(col2) - col); + if (row1 == row2) { + dx1 = 0.5; + dx2 = 0.5; + } + if (col1 == col2) { + dy1 = 0.5; + dy2 = 0.5; + } + return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1); + }`},Oc=(t,e,r,n,a,i,s,o,l,u)=>{let p=r.length===2,[h,m]=p?[0,1]:[2,3],d=t.type.value,_=w=>{let v=w===h?"row":"col";return` + fn ${v}CubicInterpolation(input_indices: ${t.type.indices}, output_indices: ${e.type.indices}) -> ${d} { + var output_index = ${e.indicesGet("output_indices",w)}; + var originalIdx: ${d} = getOriginalCoordinateFromResizedCoordinate(output_index, ${a[w]}, + ${n[w]}, ${r[w]}, ${i[w]}, ${i[w]} + ${r.length}); + var fractOriginalIdx: ${d} = originalIdx - floor(originalIdx); + var coefs = getCubicInterpolationCoefs(fractOriginalIdx); + + if (${o} && (originalIdx < 0 || originalIdx > (${r[w]} - 1))) { + return ${l}; + } + var data: array<${d}, 4> = array<${d}, 4>(0.0, 0.0, 0.0, 0.0); + for (var i: i32 = -1; i < 3; i++) { + var ${v}: ${d} = originalIdx + ${d}(i); + if (${v} < 0 || ${v} >= ${r[w]}) { + ${u?`coefs[i + 1] = 0.0; + continue;`:o?`return ${l};`:`${v} = max(0, min(${v}, ${r[w]} - 1));`}; + } + var input_indices_copy: ${t.type.indices} = input_indices; + ${t.indicesSet("input_indices_copy",w,`u32(${v})`)}; + data[i + 1] = ${w===h?t.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"}; + } + return cubicInterpolation1D(data, coefs); + }`};return` + ${_(h)}; + ${_(m)}; + fn getCubicInterpolationCoefs(s: ${d}) -> array<${d}, 4> { + var absS = abs(s); + var coeffs: array<${d}, 4> = array<${d}, 4>(0.0, 0.0, 0.0, 0.0); + var oneMinusAbsS: ${d} = 1.0 - absS; + var twoMinusAbsS: ${d} = 2.0 - absS; + var onePlusAbsS: ${d} = 1.0 + absS; + coeffs[0] = ((${s} * onePlusAbsS - 5 * ${s}) * onePlusAbsS + 8 * ${s}) * onePlusAbsS - 4 * ${s}; + coeffs[1] = ((${s} + 2) * absS - (${s} + 3)) * absS * absS + 1; + coeffs[2] = ((${s} + 2) * oneMinusAbsS - (${s} + 3)) * oneMinusAbsS * oneMinusAbsS + 1; + coeffs[3] = ((${s} * twoMinusAbsS - 5 * ${s}) * twoMinusAbsS + 8 * ${s}) * twoMinusAbsS - 4 * ${s}; + return coeffs; + } + + fn cubicInterpolation1D(x: array<${d}, 4>, coefs: array<${d}, 4>) -> ${d} { + var coefsSum: ${d} = coefs[0] + coefs[1] + coefs[2] + coefs[3]; + return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum; + } + + fn bicubicInterpolation(output_indices: ${e.type.indices}) -> ${d} { + var input_indices: ${t.type.indices} = output_indices; + return colCubicInterpolation(input_indices, output_indices); + } + `},zc=(t,e,r,n,a)=>{let[i,s,o,l,u]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],p=t.type.value;return` + fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${p} { + var input_indices: ${t.type.indices}; + ${t.indicesSet("input_indices",s,`max(0, min(depth, ${r[s]} - 1))`)}; + ${t.indicesSet("input_indices",o,`max(0, min(height, ${r[o]} - 1))`)}; + ${t.indicesSet("input_indices",l,`max(0, min(width, ${r[l]} - 1))`)}; + ${no(t,u,i,3)} + return ${t.getByIndices("input_indices")}; + } + + fn trilinearInterpolation(output_indices: ${e.type.indices}) -> ${p} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var depth:${p} = originalIndices[${s}]; + var height:${p} = originalIndices[${o}]; + var width:${p} = originalIndices[${l}]; + ${n?`if (depth < 0 || depth > (${r[s]} - 1) || height < 0 || height > (${r[o]} - 1) || width < 0 || (width > ${r[l]} - 1)) { + return ${a}; + }`:""}; + + depth = max(0, min(depth, ${r[s]} - 1)); + height = max(0, min(height, ${r[o]} - 1)); + width = max(0, min(width, ${r[l]} - 1)); + var depth1: u32 = u32(depth); + var height1: u32 = u32(height); + var width1: u32 = u32(width); + var depth2: u32 = u32(depth + 1); + var height2: u32 = u32(height + 1); + var width2: u32 = u32(width + 1); + var channel: u32 = ${r.length>3?`u32(originalIndices[${u}])`:"0"}; + var batch: u32 = ${r.length>3?`u32(originalIndices[${i}])`:"0"}; + + var x111: ${p} = getInputValue(batch, channel, depth1, height1, width1); + var x112: ${p} = getInputValue(batch, channel, depth1, height1, width2); + var x121: ${p} = getInputValue(batch, channel, depth1, height2, width1); + var x122: ${p} = getInputValue(batch, channel, depth1, height2, width2); + var x211: ${p} = getInputValue(batch, channel, depth2, height1, width1); + var x212: ${p} = getInputValue(batch, channel, depth2, height1, width2); + var x221: ${p} = getInputValue(batch, channel, depth2, height2, width1); + var x222: ${p} = getInputValue(batch, channel, depth2, height2, width2); + var dx1: ${p} = abs(depth - ${p}(depth1)); + var dx2: ${p} = abs(${p}(depth2) - depth); + var dy1: ${p} = abs(height - ${p}(height1)); + var dy2: ${p} = abs(${p}(height2) - height); + var dz1: ${p} = abs(width - ${p}(width1)); + var dz2: ${p} = abs(${p}(width2) - width); + if (depth1 == depth2) { + dx1 = 0.5; + dx2 = 0.5; + } + if (height1 == height2) { + dy1 = 0.5; + dy2 = 0.5; + } + if (width1 == width2) { + dz1 = 0.5; + dz2 = 0.5; + } + 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); + }`},Pc=(t,e,r,n,a,i)=>{let s=t.dims,o=kc(i,e.axes,s.length),l=Ec(s,n,a,e.axes),u=n.slice();n.length===0&&(u=s.map(($,E)=>$===0?1:l[E]/$),e.keepAspectRatioPolicy!=="stretch"&&(l=Cc(s,u,e)));let p=ge("output",t.dataType,l.length),h=X("input",t.dataType,s.length),m=K.size(l),d=s.length===l.length&&s.every(($,E)=>$===l[E]),_=e.coordinateTransformMode==="tf_crop_and_resize",w=e.extrapolationValue,v=h.type.value,S=$=>` + ${d?"":` + ${xc(e.coordinateTransformMode,v)}; + ${(()=>{switch(e.mode){case"nearest":return` + ${Ic(h,s)}; + ${Sc(e.nearestMode,r,v)}; + ${Ac(h,p,s,l,u.length,o.length,_)}; + `;case"linear":return` + ${Tc(p,s,l,u.length,o.length)}; + ${(()=>{if(s.length===2||s.length===4)return`${Mc(h,p,s,_,w)}`;if(s.length===3||s.length===5)return`${zc(h,p,s,_,w)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()}; + `;case"cubic":return` + ${(()=>{if(s.length===2||s.length===4)return`${Oc(h,p,s,l,u,o,e.cubicCoeffA,_,e.extrapolationValue,e.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",u.length).registerUniform("roi","f32",o.length).declareVariables(h,p)} + ${$.mainStart()} + ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + ${d?"output[global_idx] = input[global_idx];":` + let output_indices = ${p.offsetToIndices("global_idx")}; + var input_indices: ${h.type.indices}; + ${(()=>{switch(e.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices); + if (checkInputIndices(input_indices)) { + output[global_idx] = ${h.getByIndices("input_indices")}; + } else { + output[global_idx] = ${e.extrapolationValue}; + }`;case"linear":return`output[global_idx] = ${s.length===2||s.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${e.mode}`)}})()}; +`} + }`;return{name:"Resize",shaderCache:{hint:`${e.cacheKey}|${r}|${u.length>0?u:""}|${a.length>0?a:""}|${o.length>0?o:""}|${d}|${s}`,inputDependencies:["rank"]},getShaderSource:S,getRunData:()=>({outputs:[{dims:l,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},{type:1,data:u},{type:1,data:o},...ye(s,l)]})}},Rc=t=>{let e=t.customDataBuffer;return new Uint32Array(e,e.byteOffset,1)[0]},Im=(t,e)=>{let r=[],n=[],a=[],i=Rc(t);if(e.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");$c(t.inputs,e,i,r,n,a),t.compute(Pc(t.inputs[0],e,i,r,n,a),{inputs:[0]})},Mm=t=>{let e=t.antialias,r=t.axes,n=t.coordinateTransformMode,a=t.cubicCoeffA,i=t.excludeOutside!==0,s=t.extrapolationValue,o=t.keepAspectRatioPolicy,l=t.mode,u=t.nearestMode===""?"simple":t.nearestMode;return Ke({antialias:e,axes:r,coordinateTransformMode:n,cubicCoeffA:a,excludeOutside:i,extrapolationValue:s,keepAspectRatioPolicy:o,mode:l,nearestMode:u})}}),Bc,Dc,Om,Ky=J(()=>{$e(),Me(),ct(),Te(),Bc=(t,e)=>{let[r,n,a,i]=t,{numHeads:s,rotaryEmbeddingDim:o}=e;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(a.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${a.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(a.dims,i.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(o>0&&s===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let l=r.dims[0],u=r.dims[r.dims.length-2],p=a.dims[0],h=K.sizeFromDimension(r.dims,1)/u,m=o===0?a.dims[1]*2:h/s;if(o>m)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(n.dims.length===2){if(l!==n.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${n.dims[0]}`);if(u!==n.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${n.dims[1]}`)}if(m/2!==a.dims[1]&&o/2!==a.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${a.dims[1]}`);if(u>p)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Dc=(t,e)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:a,scale:i}=e,s=t[0].dims[0],o=K.sizeFromDimension(t[0].dims,1),l=t[0].dims[t[0].dims.length-2],u=o/l,p=t[2].dims[1],h=a===0?p*2:u/n,m=new Array(s,l,u/h,h-p),d=K.computeStrides(m),_=[{type:1,data:i},{type:12,data:m},{type:12,data:d},...t[0].dims.length===3?new Array({type:12,data:[o,u,h,1]}):[],...t[0].dims.length===4?new Array({type:12,data:[o,h,l*h,1]}):[],...ye(t[0].dims,t[1].dims,t[2].dims,t[3].dims,t[0].dims)],w=v=>{let S=X("input",t[0].dataType,t[0].dims.length),$=X("position_ids",t[1].dataType,t[1].dims.length),E=X("cos_cache",t[2].dataType,t[2].dims.length),T=X("sin_cache",t[3].dataType,t[3].dims.length),A=ge("output",t[0].dataType,t[0].dims.length);return v.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:m.length},{name:"global_strides",type:"u32",length:d.length},{name:"input_output_strides",type:"u32",length:d.length}]),` + ${v.declareVariables(S,$,E,T,A)} + + ${v.mainStart(Pn)} + let half_rotary_emb_dim = uniforms.${E.name}_shape[1]; + let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape; + let size = uniforms.global_shape[0] * uniforms.global_strides[0]; + ${v.guardAgainstOutOfBoundsWorkgroupSizes("size")} + + if (bsnh[3] < half_rotary_emb_dim) { + let position_ids_idx = + ${$.broadcastedIndicesToOffset("bsnh.xy",ge("",$.type.tensor,2))}; + let position_id = + u32(${$.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}); + let j = i + select(half_rotary_emb_dim, 1, ${r}); + let re = ${S.getByOffset("i")} * ${E.get("position_id","bsnh[3]")} - + ${S.getByOffset("j")} * ${T.get("position_id","bsnh[3]")}; + ${A.setByOffset("i","re")} + let im = ${S.getByOffset("i")} * ${T.get("position_id","bsnh[3]")} + + ${S.getByOffset("j")} * ${E.get("position_id","bsnh[3]")}; + ${A.setByOffset("j","im")} + } else { + let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim; + ${A.setByOffset("k",S.getByOffset("k"))} + } + }`};return{name:"RotaryEmbedding",shaderCache:{hint:Ke({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:w,getRunData:()=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(K.size(m)/Pn)},programUniforms:_})}},Om=(t,e)=>{Bc(t.inputs,e),t.compute(Dc(t.inputs,e))}}),Nc,Fc,zm,Yy=J(()=>{$e(),Me(),Te(),Nc=t=>{if(!t||t.length<3)throw new Error("layerNorm requires at least 3 inputs.");let e=t[0],r=t[1],n=t[2];if(e.dataType!==r.dataType||e.dataType!==n.dataType)throw new Error("All inputs must have the same data type");if(e.dims.length!==3&&e.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 a=e.dims[e.dims.length-1],i=e.dims[e.dims.length-2];if(r.dims[r.dims.length-1]!==a)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]!==a)throw new Error("Gamma must have the same hidden size as input");if(t.length>3){let s=t[3];if(s.dims.length!==1)throw new Error("Beta must be 1D");if(s.dims[s.dims.length-1]!==a)throw new Error("Beta must have the same hidden size as input")}if(t.length>4){let s=t[4];if(s.dims.length!==1)throw new Error("Bias must be 1D");if(s.dims[s.dims.length-1]!==a)throw new Error("Bias must have the same hidden size as input")}},Fc=(t,e,r,n)=>{let a=e.simplified,i=t[0].dims,s=K.size(i),o=i,l=s,u=i.slice(-1)[0],p=n?i.slice(0,-1).concat(1):[],h=!a&&t.length>3,m=t.length>4,d=n&&r>1,_=n&&r>2,w=r>3,v=ot(u),S=[{type:12,data:l},{type:12,data:v},{type:12,data:u},{type:1,data:e.epsilon}],$=T=>{let A=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],P=[X("x",t[0].dataType,t[0].dims,v),X("skip",t[1].dataType,t[1].dims,v),X("gamma",t[2].dataType,t[2].dims,v)];h&&P.push(X("beta",t[3].dataType,t[3].dims,v)),m&&P.push(X("bias",t[4].dataType,t[4].dims,v)),P.push(ge("output",t[0].dataType,o,v)),d&&P.push(ge("mean_output",1,p)),_&&P.push(ge("inv_std_output",1,p)),w&&P.push(ge("input_skip_bias_sum",t[0].dataType,o,v));let B=mt(t[0].dataType);return` + + ${T.registerUniforms(A).declareVariables(...P)} + + ${T.mainStart()} + ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size / uniforms.hidden_size")} + let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components; + let offset = global_idx * hidden_size_vectorized; + var sum = ${Sr("f32",v)}; + var squareSum = ${Sr("f32",v)}; + for (var i: u32 = 0; i < hidden_size_vectorized; i++) { + let skip_value = skip[offset + i]; + let bias_value = ${m?"bias[i]":B+"(0.0)"}; + let input_value = x[offset + i]; + let value = input_value + skip_value + bias_value; + ${w?"input_skip_bias_sum[offset + i] = value;":""} + output[offset + i] = value; + let f32_value = ${Mn(B,v,"value")}; + sum += f32_value; + squareSum += f32_value * f32_value; + } + let mean = ${Dr("sum",v)} / f32(uniforms.hidden_size); + let inv_std_dev = inverseSqrt(${Dr("squareSum",v)} / f32(uniforms.hidden_size) ${a?"":"- mean * mean"} + uniforms.epsilon); + ${d?"mean_output[global_idx] = mean;":""} + ${_?"inv_std_output[global_idx] = inv_std_dev;":""} + for (var i: u32 = 0; i < hidden_size_vectorized; i++) { + output[offset + i] = (output[offset + i] ${a?"":`- ${B}(mean)`}) * ${B}(inv_std_dev) * gamma[i] ${h?"+ beta[i]":""}; + } + }`},E=[{dims:o,dataType:t[0].dataType}];return r>1&&E.push({dims:p,dataType:1}),r>2&&E.push({dims:p,dataType:1}),r>3&&E.push({dims:i,dataType:t[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${v};${d};${_};${w}`,inputDependencies:t.map((T,A)=>"type")},getShaderSource:$,getRunData:()=>({outputs:E,dispatchGroup:{x:Math.ceil(l/u/64)},programUniforms:S})}},zm=(t,e)=>{Nc(t.inputs);let r=[0];t.outputCount>1&&r.push(-3),t.outputCount>2&&r.push(-3),t.outputCount>3&&r.push(3),t.compute(Fc(t.inputs,e,t.outputCount,!1),{outputs:r})}}),Lc,la,Uc,ao,Wc,Vc,Pm,Rm,Xy=J(()=>{$e(),Me(),ct(),Te(),Lc=(t,e)=>{if(!t||t.length<1)throw new Error("too few inputs");if(e.axes.length!==0){if(e.axes.length!==e.starts.length||e.axes.length!==e.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(e.starts.length!==e.ends.length)throw new Error("starts and ends must have the same length");t.slice(1).forEach((r,n)=>{if(t[n+1].dataType!==6&&t[n+1].dataType!==7)throw new Error(`Input ${n} must be an array of int32 or int64`)})},la=(t,e)=>{let r=[];if(t.length>e)if(t[e].dataType===7)t[e].getBigInt64Array().forEach(n=>r.push(Number(n)));else if(t[e].dataType===6)t[e].getInt32Array().forEach(n=>r.push(Number(n)));else throw new Error(`Input ${e} must be an array of int32 or int64`);return r},Uc=(t,e)=>{if(t.length>1){let r=la(t,1),n=la(t,2),a=la(t,3);return a.length===0&&(a=[...Array(t[0].dims.length).keys()]),Ke({starts:r,ends:n,axes:a})}else return e},ao=(t,e,r,n,a)=>{let i=t;return t<0&&(i+=r[n[e]]),a[e]<0?Math.max(0,Math.min(i,r[n[e]]-1)):Math.max(0,Math.min(i,r[n[e]]))},Wc=(t,e,r)=>`fn calculateInputIndices(output_indices: ${e.type.indices}) -> ${t.type.indices} { + var input_indices: ${t.type.indices}; + var carry = 0u; + for (var i = ${r.length}; i >= 0; i--) { + let input_shape_i = ${Se("uniforms.input_shape","i",r.length)}; + let steps_i = ${Se("uniforms.steps","i",r.length)}; + let signs_i = ${Se("uniforms.signs","i",r.length)}; + let starts_i = ${Se("uniforms.starts","i",r.length)}; + var output_index = ${e.indicesGet("output_indices","i")}; + var input_index = output_index * steps_i + starts_i + carry; + carry = input_index / input_shape_i; + input_index = input_index % input_shape_i; + if (signs_i < 0) { + input_index = input_shape_i - input_index - 1u + starts_i; + } + ${t.indicesSet("input_indices","i","input_index")}; + } + return input_indices; + }`,Vc=(t,e)=>{let r=t[0].dims,n=K.size(r),a=e.axes.length>0?K.normalizeAxes(e.axes,r.length):[...Array(r.length).keys()],i=la(t,4);i.forEach(S=>S!==0||(()=>{throw new Error("step cannot be 0")})),i.length===0&&(i=Array(a.length).fill(1));let s=e.starts.map((S,$)=>ao(S,$,r,a,i)),o=e.ends.map((S,$)=>ao(S,$,r,a,i));if(a.length!==s.length||a.length!==o.length)throw new Error("start, ends and axes should have the same number of elements");if(a.length!==r.length)for(let S=0;SMath.sign(S));i.forEach((S,$,E)=>{if(S<0){let T=(o[$]-s[$])/S,A=s[$],P=A+T*i[$];s[$]=P,o[$]=A,E[$]=-S}});let u=r.slice(0);a.forEach((S,$)=>{u[S]=Math.ceil((o[S]-s[S])/i[S])});let p={dims:u,dataType:t[0].dataType},h=ge("output",t[0].dataType,u.length),m=X("input",t[0].dataType,t[0].dims.length),d=K.size(u),_=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:s.length},{name:"signs",type:"i32",length:l.length},{name:"steps",type:"u32",length:i.length}],w=[{type:12,data:d},{type:12,data:s},{type:6,data:l},{type:12,data:i},...ye(t[0].dims,u)],v=S=>` + ${S.registerUniforms(_).declareVariables(m,h)} + ${Wc(m,h,r)} + ${S.mainStart()} + ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + let output_indices = ${h.offsetToIndices("global_idx")}; + let input_indices = calculateInputIndices(output_indices); + ${h.setByOffset("global_idx",m.getByIndices("input_indices"))} + }`;return{name:"Slice",shaderCache:{hint:`${l.length}_${s.length}_${i.length}`,inputDependencies:["rank"]},getShaderSource:v,getRunData:()=>({outputs:[p],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:w})}},Pm=(t,e)=>{Lc(t.inputs,e);let r=Uc(t.inputs,e);t.compute(Vc(t.inputs,r),{inputs:[0]})},Rm=t=>{let e=t.starts,r=t.ends,n=t.axes;return Ke({starts:e,ends:r,axes:n})}}),Gc,Hc,Bm,Dm,Qy=J(()=>{$e(),Me(),ct(),Te(),Gc=t=>{if(!t||t.length!==1)throw new Error("Softmax op requires 1 input.")},Hc=(t,e)=>{let r=t.dims,n=K.size(r),a=64,i=e.axis;if(i<0&&(i=r.length+i),iS===4?`max(max(${v}.x, ${v}.y), max(${v}.z, ${v}.w))`:S===2?`max(${v}.x, ${v}.y)`:S===3?`max(max(${v}.x, ${v}.y), ${v}.z)`:v,h=X("x",t.dataType,t.dims,l),m=ge("result",t.dataType,t.dims,l),d=h.type.value,_=mt(t.dataType)==="f32"?`var threadMax = ${d}(-3.402823e+38f);`:`var threadMax = ${d}(-65504.0h);`,w=v=>` + var rowMaxShared : ${d}; + var rowSumShared : ${d}; + var threadShared : array<${d}, ${a}>; + + fn getValue(row: i32, col: i32, row_stride: i32) -> ${d} { + let index = row * row_stride + col; + return x[index]; + } + + fn setValue(row: i32, col: i32, row_stride: i32, value: ${d}) { + let index = row * row_stride + col; + result[index] = value; + } + ${v.registerUniform("packedCols","i32").declareVariables(h,m)} + ${v.mainStart()} + let gindex = i32(global_idx); + let lindex = i32(local_idx); + const wg = ${a}; + let row = gindex / wg; + let cols = uniforms.packedCols; + let row_stride : i32 = uniforms.packedCols; + + // find the rows max + ${_} + for (var col = lindex; col < cols; col += wg) { + let value = getValue(row, col, row_stride); + threadMax = max(threadMax, value); + } + if (lindex < cols) { + threadShared[lindex] = threadMax; + } + workgroupBarrier(); + + var reduceSize = min(cols, wg); + for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) { + reduceSize = currSize + (reduceSize & 1); + if (lindex < currSize) { + threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]); + } + workgroupBarrier(); + } + if (lindex == 0) { + rowMaxShared = ${d}(${p("threadShared[0]",l)}); + } + workgroupBarrier(); + + // find the rows sum + var threadSum = ${d}(0.0); + for (var col = lindex; col < cols; col += wg) { + let subExp = exp(getValue(row, col, row_stride) - rowMaxShared); + threadSum += subExp; + } + threadShared[lindex] = threadSum; + workgroupBarrier(); + + for (var currSize = wg >> 1; currSize > 0; currSize = currSize >> 1) { + if (lindex < currSize) { + threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize]; + } + workgroupBarrier(); + } + if (lindex == 0) { + rowSumShared = ${d}(${Dr("threadShared[0]",l)}); + } + workgroupBarrier(); + + // calculate final value for each element in the row + for (var col = lindex; col < cols; col += wg) { + let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared; + setValue(row, col, row_stride, value); + } + }`;return{name:"Softmax",shaderCache:{hint:`${l}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:t.dataType}],dispatchGroup:{x:o},programUniforms:[{type:6,data:u}]}),getShaderSource:w}},Bm=(t,e)=>{Gc(t.inputs),t.compute(Hc(t.inputs[0],e))},Dm=t=>Ke({axis:t.axis})}),jc,qc,Kc,Yc,Xc,Nm,Fm,Zy=J(()=>{$e(),Me(),ct(),Te(),jc=t=>{if(!t||t.length<1)throw new Error("too few inputs")},qc=(t,e)=>{let r=[],n=e.numOutputs;return t[1].dims[0]>0&&(t[1].getBigInt64Array().forEach(a=>r.push(Number(a))),n=r.length),Ke({numOutputs:n,axis:e.axis,splitSizes:r})},Kc=t=>` +fn calculateOutputIndex(index: u32) -> u32 { + for (var i: u32 = 0u; i < ${t}u; i += 1u ) { + if (index < ${Se("uniforms.size_in_split_axis","i",t)}) { + return i; + } + } + return ${t}u; +}`,Yc=t=>{let e=t.length,r=[];for(let n=0;n{let r=t[0].dims,n=K.size(r),a=t[0].dataType,i=K.normalizeAxis(e.axis,r.length),s=new Array(e.numOutputs),o=X("input",a,r.length),l=new Array(e.numOutputs),u=[],p=[],h=0,m=[{type:12,data:n}];for(let _=0;_` + ${_.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",l.length).declareVariables(o,...s)} + ${Kc(l.length)} + ${Yc(s)} + + ${_.mainStart()} + ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")} + + var indices = ${o.offsetToIndices("global_idx")}; + var index = ${o.indicesGet("indices",i)}; + let output_number = calculateOutputIndex(index); + if (output_number != 0) { + index -= ${Se("uniforms.size_in_split_axis","output_number - 1u",l.length)}; + ${o.indicesSet("indices",i,"index")}; + } + writeBufferData(output_number, indices, global_idx); + }`;return{name:"Split",shaderCache:{hint:e.cacheKey,inputDependencies:["rank"]},getShaderSource:d,getRunData:()=>({outputs:u,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:m})}},Nm=(t,e)=>{jc(t.inputs);let r=t.inputs.length===1?e:qc(t.inputs,e);t.compute(Xc(t.inputs,r),{inputs:[0]})},Fm=t=>{let e=t.axis,r=t.splitSizes,n=t.numOutputs<0?r.length:t.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return Ke({axis:e,numOutputs:n,splitSizes:r})}}),io,Qc,Zc,Jc,Lm,Jy=J(()=>{$e(),Me(),Te(),io=t=>Array.from(t.getBigInt64Array(),Number),Qc=t=>{if(!t||t.length!==2)throw new Error("Tile requires 2 inputs.");if(t[0].dataType!==1&&t[0].dataType!==6&&t[0].dataType!==12)throw new Error("Tile only support float, int32, and uint32 data types");if(t[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(t[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(io(t[1]).length!==t[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},Zc=(t,e)=>{let r=[];for(let n=0;n{let e=t[0].dims,r=io(t[1]),n=Zc(e,r),a=K.size(n),i=t[0].dataType,s=X("input",i,e.length),o=ge("output",i,n.length),l=u=>` + const inputShape = ${s.indices(...e)}; + ${u.registerUniform("output_size","u32").declareVariables(s,o)} + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let output_indices = ${o.offsetToIndices("global_idx")}; + var input_indices: ${s.type.indices}; + for (var i = 0; i < ${e.length}; i++) { + let input_dim_i = ${s.indicesGet("uniforms.input_shape","i")}; + let input_dim_value = ${o.indicesGet("output_indices","i")} % input_dim_i; + + ${s.indicesSet("input_indices","i","input_dim_value")} + } + ${o.setByOffset("global_idx",s.getByIndices("input_indices"))} + }`;return{name:"Tile",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},...ye(t[0].dims,n)]}),getShaderSource:l}},Lm=t=>{Qc(t.inputs),t.compute(Jc(t.inputs),{inputs:[0]})}}),ep,tp,Um,ew=J(()=>{$e(),Me(),Te(),ep=(t,e,r,n,a)=>{let i=ge("output_data",a,r.length,4),s=X("a_data",e[1].dataType,e[1].dims.length,4),o=X("b_data",e[2].dataType,e[2].dims.length,4),l=X("c_data",e[0].dataType,e[0].dims.length,4),u,p=(h,m,d)=>`select(${m}, ${h}, ${d})`;if(!n)u=i.setByOffset("global_idx",p(s.getByOffset("global_idx"),o.getByOffset("global_idx"),l.getByOffset("global_idx")));else{let h=(m,d,_="")=>{let w=`a_data[index_a${d}][component_a${d}]`,v=`b_data[index_b${d}][component_b${d}]`,S=`bool(c_data[index_c${d}] & (0xffu << (component_c${d} * 8)))`;return` + let output_indices${d} = ${i.offsetToIndices(`global_idx * 4u + ${d}u`)}; + let offset_a${d} = ${s.broadcastedIndicesToOffset(`output_indices${d}`,i)}; + let offset_b${d} = ${o.broadcastedIndicesToOffset(`output_indices${d}`,i)}; + let offset_c${d} = ${l.broadcastedIndicesToOffset(`output_indices${d}`,i)}; + let index_a${d} = offset_a${d} / 4u; + let index_b${d} = offset_b${d} / 4u; + let index_c${d} = offset_c${d} / 4u; + let component_a${d} = offset_a${d} % 4u; + let component_b${d} = offset_b${d} % 4u; + let component_c${d} = offset_c${d} % 4u; + ${m}[${d}] = ${_}(${p(w,v,S)}); + `};a===9?u=` + var data = vec4(0); + ${h("data",0,"u32")} + ${h("data",1,"u32")} + ${h("data",2,"u32")} + ${h("data",3,"u32")} + output_data[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:u=` + ${h("output_data[global_idx]",0)} + ${h("output_data[global_idx]",1)} + ${h("output_data[global_idx]",2)} + ${h("output_data[global_idx]",3)} + `}return` + ${t.registerUniform("vec_size","u32").declareVariables(l,s,o,i)} + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${u} + }`},tp=t=>{let e=t[1].dims,r=t[2].dims,n=t[0].dims,a=t[1].dataType,i=!(K.areEqual(e,r)&&K.areEqual(r,n)),s=e,o=K.size(e);if(i){let u=zn.calcShape(zn.calcShape(e,r,!1),n,!1);if(!u)throw new Error("Can't perform where op on the given tensors");s=u,o=K.size(s)}let l=Math.ceil(o/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:u=>ep(u,t,s,i,a),getRunData:()=>({outputs:[{dims:s,dataType:a}],dispatchGroup:{x:Math.ceil(o/64/4)},programUniforms:[{type:12,data:l},...ye(n,e,r,s)]})}},Um=t=>{t.compute(tp(t.inputs))}}),Wm,tw=J(()=>{$y(),Xh(),xy(),Sy(),ky(),Ey(),Kh(),Xf(),My(),Oy(),zy(),Py(),Ry(),By(),Dy(),Ny(),Fy(),Ly(),Uy(),Yf(),Wy(),Vy(),Gy(),Hy(),jy(),el(),qy(),Ky(),Yy(),Xy(),Qy(),Zy(),Jy(),Ia(),tl(),ew(),Wm=new Map([["Abs",[Jh]],["Acos",[ef]],["Acosh",[tf]],["Add",[Rf]],["ArgMax",[Hh,So]],["ArgMin",[Gh,So]],["Asin",[rf]],["Asinh",[nf]],["Atan",[af]],["Atanh",[sf]],["Attention",[Yh]],["AveragePool",[$m,vm]],["BatchNormalization",[Qh]],["BiasAdd",[Zh]],["BiasSplitGelu",[Pf]],["Cast",[lf,of]],["Ceil",[df]],["Clip",[uf]],["Concat",[jh,qh]],["Conv",[Io,Ao]],["ConvTranspose",[Jf,Zf]],["Cos",[cf]],["Cosh",[pf]],["CumSum",[em,tm]],["DepthToSpace",[rm,nm]],["Div",[Bf]],["Einsum",[am,im]],["Elu",[hf,zi]],["Equal",[Df]],["Erf",[ff]],["Exp",[mf]],["Expand",[sm]],["FastGelu",[om]],["Floor",[gf]],["FusedConv",[Io,Ao]],["Gather",[um,lm]],["GatherElements",[cm,dm]],["Gelu",[_f]],["Gemm",[hm,pm]],["GlobalAveragePool",[Sm,xm]],["GlobalMaxPool",[Tm,Cm]],["Greater",[Uf]],["GreaterOrEqual",[Vf]],["HardSigmoid",[kf,Sf]],["InstanceNormalization",[fm]],["LayerNormalization",[mm]],["LeakyRelu",[yf,zi]],["Less",[Wf]],["LessOrEqual",[Gf]],["Log",[zf]],["MatMul",[Kf]],["MatMulNBits",[gm,_m]],["MaxPool",[km,Em]],["Mul",[Nf]],["MultiHeadAttention",[wm,ym]],["Neg",[bf]],["Not",[wf]],["Pad",[bm]],["Pow",[Ff]],["Range",[Am]],["Reciprocal",[vf]],["ReduceMin",[Fh]],["ReduceMean",[Ph]],["ReduceMax",[Nh]],["ReduceSum",[Uh]],["ReduceProd",[Lh]],["ReduceL1",[Rh]],["ReduceL2",[Bh]],["ReduceLogSum",[Vh]],["ReduceLogSumExp",[Dh]],["ReduceSumSquare",[Wh]],["Relu",[$f]],["Resize",[Im,Mm]],["RotaryEmbedding",[Om]],["Sigmoid",[xf]],["Sin",[Ef]],["Sinh",[Cf]],["Slice",[Pm,Rm]],["SkipLayerNormalization",[zm]],["Split",[Nm,Fm]],["Sqrt",[Tf]],["Softmax",[Bm,Dm]],["Sub",[Lf]],["Tan",[Af]],["Tanh",[If]],["ThresholdedRelu",[Of,zi]],["Tile",[Lm]],["Transpose",[$h,xh]],["Where",[Um]]])}),Vm,rw=J(()=>{dr(),un(),Te(),Vm=class{constructor(t){this.backend=t,this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,e){this.repo.set(t,e)}run(t,e,r,n,a){ur(t.programInfo.name);let i=this.backend.device,s=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let o=[];for(let u of e)o.push({binding:o.length,resource:{buffer:u.buffer}});for(let u of r)o.push({binding:o.length,resource:{buffer:u.buffer}});a&&o.push({binding:o.length,resource:a});let l=i.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:o,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let u={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:l,dispatchGroup:n};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(u)}s.setPipeline(t.computePipeline),s.setBindGroup(0,l),s.dispatchWorkgroups(...n),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),jt(t.programInfo.name)}dispose(){}build(t,e){ur(t.name);let r=this.backend.device,n=[];r.features.has("shader-f16")&&n.push("enable f16;");let a=vh(e,this.backend.device.limits),i=t.getShaderSource(a),s=`${n.join(` +`)} +${a.additionalImplementations} +${i}`,o=r.createShaderModule({code:s,label:t.name});at("verbose",()=>`[WebGPU] ${t.name} shader code: ${s}`);let l=r.createComputePipeline({compute:{module:o,entryPoint:"main"},layout:"auto",label:t.name});return jt(t.name),{programInfo:t,computePipeline:l,uniformVariablesInfo:a.variablesInfo}}normalizeDispatchGroupSize(t){let e=typeof t=="number"?t:t.x,r=typeof t=="number"?1:t.y||1,n=typeof t=="number"?1:t.z||1,a=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(e<=a&&r<=a&&n<=a)return[e,r,n];let i=e*r*n,s=Math.ceil(Math.sqrt(i));if(s>a){if(s=Math.ceil(Math.cbrt(i)),s>a)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[s,s,s]}else return[s,s,1]}}}),rp,np,ap,Gm,nw=J(()=>{dr(),$e(),un(),wy(),by(),tw(),rw(),rp=(t,e)=>{if(e.length!==t.length)throw new Error(`inputDependencies length ${e.length} is not equal to inputTensors length ${t.length}.`);let r=[];for(let n=0;n{let n=t.name;return t.shaderCache?.hint&&(n+="["+t.shaderCache.hint+"]"),n+=":"+r+`:${rp(e,t.shaderCache?.inputDependencies??new Array(e.length).fill("dims"))}`,n},ap=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},Gm=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,e){this.env=t;let r=[],n={requiredLimits:{maxComputeWorkgroupStorageSize:e.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:e.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:e.limits.maxStorageBufferBindingSize,maxBufferSize:e.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:e.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:e.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:e.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:e.limits.maxComputeWorkgroupSizeZ},requiredFeatures:r};e.features.has("chromium-experimental-timestamp-query-inside-passes")?r.push("chromium-experimental-timestamp-query-inside-passes"):e.features.has("timestamp-query")&&r.push("timestamp-query"),e.features.has("shader-f16")&&r.push("shader-f16"),this.device=await e.requestDevice(n),this.adapterInfo=new ap(await e.requestAdapterInfo()),this.gpuDataManager=wh(this),this.programManager=new Vm(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,_h(t.logLevel,!!t.debug),this.device.onuncapturederror=a=>{a.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${a.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value:e,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(),e={};this.queryType==="at-passes"&&(e.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=t.beginComputePass(e)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;ur(),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 e=new BigUint64Array(t.getMappedRange()),r=this.pendingQueries.get(t);for(let n=0;n"u"&&(this.queryTimeBase=m);let _=Number(m-this.queryTimeBase),w=Number(d-this.queryTimeBase);if(!Number.isSafeInteger(_)||!Number.isSafeInteger(w))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:p.map(v=>({dims:v.dims,dataType:Jr(v.dataType)})),outputsMetadata:h.map(v=>({dims:v.dims,dataType:Jr(v.dataType)})),kernelId:i,kernelType:o,kernelName:l,programName:u,startTime:_,endTime:w});else{let v="";p.forEach(($,E)=>{v+=`input[${E}]: [${$.dims}] | ${Jr($.dataType)}, `});let S="";h.forEach(($,E)=>{S+=`output[${E}]: [${$.dims}] | ${Jr($.dataType)}, `}),console.log(`[profiling] kernel "${i}|${o}|${l}|${u}" ${v}${S}execution time: ${w-_} ns`)}va("GPU",`${u}::${m}::${d}`)}t.unmap(),this.pendingQueries.delete(t)}),jt()}run(t,e,r,n,a,i){ur(t.name);let s=[];for(let $=0;$E):r;if(p.length!==o.length)throw new Error(`Output size ${p.length} must be equal to ${o.length}.`);let h=[],m=[];for(let $=0;$=i)throw new Error(`Invalid output index: ${p[$]}`);if(p[$]===-3)continue;let E=p[$]===-1,T=p[$]===-2,A=E||T?a(o[$].dataType,o[$].dims):n(p[$],o[$].dataType,o[$].dims);if(h.push(A),A.data===0)continue;let P=this.gpuDataManager.get(A.data);if(!P)throw new Error(`no GPU data for output: ${A.data}`);if(E&&this.temporaryData.push(P),T){let B=this.kernelPersistentData.get(this.currentKernelId);B||(B=[],this.kernelPersistentData.set(this.currentKernelId,B)),B.push(P)}m.push(P)}if(s.length!==e.length||m.length!==h.length){if(m.length===0)return jt(t.name),h;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let d;if(u){let $=0,E=[];u.forEach(B=>{let D=typeof B.data=="number"?[B.data]:B.data;if(D.length===0)return;let q=B.type===10?2:4,H,ie;B.type===10?(ie=D.length>4?16:D.length>2?8:D.length*q,H=D.length>4?16:q*D.length):(ie=D.length<=2?D.length*q:16,H=16),$=Math.ceil($/ie)*ie,E.push($);let te=B.type===10?8:4;$+=D.length>4?Math.ceil(D.length/te)*H:D.length*q});let T=16;$=Math.ceil($/T)*T;let A=new ArrayBuffer($);u.forEach((B,D)=>{let q=E[D],H=typeof B.data=="number"?[B.data]:B.data;if(B.type===6)new Int32Array(A,q,H.length).set(H);else if(B.type===12)new Uint32Array(A,q,H.length).set(H);else if(B.type===10)new Uint16Array(A,q,H.length).set(H);else if(B.type===1)new Float32Array(A,q,H.length).set(H);else throw new Error(`Unsupported uniform type: ${Jr(B.type)}`)});let P=this.gpuDataManager.create($,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(P.buffer,0,A,0,$),this.gpuDataManager.release(P.id),d={offset:0,size:$,buffer:P.buffer}}let _=this.programManager.normalizeDispatchGroupSize(l),w=_[1]===1&&_[2]===1,v=np(t,e,w),S=this.programManager.getArtifact(v);if(S||(S=this.programManager.build(t,_),this.programManager.setArtifact(v,S),at("info",()=>`[artifact] key: ${v}, programName: ${t.name}`)),u&&S.uniformVariablesInfo){if(u.length!==S.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${S.uniformVariablesInfo.length}, got ${u.length} in program "${S.programInfo.name}".`);for(let $=0;$`[ProgramManager] run "${t.name}" (key=${v}) with ${_[0]}x${_[1]}x${_[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let $={kernelId:this.currentKernelId,programName:S.programInfo.name,inputTensorViews:e,outputTensorViews:h};this.pendingKernels.push($),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push($)}return this.programManager.run(S,s,m,_,d),jt(t.name),h}upload(t,e){this.gpuDataManager.upload(t,e)}memcpy(t,e){this.gpuDataManager.memcpy(t,e)}async download(t,e){await this.gpuDataManager.download(t,e)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,e,r,n){let a=Wm.get(t);if(!a)throw new Error(`kernel not implemented: ${t}`);let i={kernelType:t,kernelName:n,kernelEntry:a[0],attributes:[a[1],r]};this.kernels.set(e,i)}releaseKernel(t){let e=this.kernelPersistentData.get(t);if(e){for(let r of e)this.gpuDataManager.release(r.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,e,r){let n=this.kernels.get(t);if(!n)throw new Error(`kernel not created: ${t}`);let a=n.kernelType,i=n.kernelName,s=n.kernelEntry,o=n.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${a}] ${i}" is not allowed to be called recursively`);this.currentKernelId=t,o[0]&&(o[1]=o[0](o[1]),o[0]=void 0),at("info",()=>`[WebGPU] Start to run kernel "[${a}] ${i}"...`);let l=this.env.debug;this.temporaryData=[];try{return l&&this.device.pushErrorScope("validation"),s(e,o[1]),0}catch(u){return r.push(Promise.resolve(`[WebGPU] Kernel "[${a}] ${i}" failed. ${u}`)),1}finally{l&&r.push(this.device.popErrorScope().then(u=>u?`GPU validation error for kernel "[${a}] ${i}": ${u.message}`:null));for(let u of this.temporaryData)this.gpuDataManager.release(u.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,e,r,n){let a=this.sessionExternalDataMapping.get(t);a||(a=new Map,this.sessionExternalDataMapping.set(t,a));let i=a.get(e),s=this.gpuDataManager.registerExternalBuffer(r,n,i?.[1]);return a.set(e,[s,r]),s}unregisterBuffers(t){let e=this.sessionExternalDataMapping.get(t);e&&(e.forEach(r=>this.gpuDataManager.unregisterExternalBuffer(r[1])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let e=this.gpuDataManager.get(t);if(!e)throw new Error(`no GPU data for buffer: ${t}`);return e.buffer}createDownloader(t,e,r){return async()=>{let n=await $o(this,t,e);return yh(n.buffer,r)}}writeTimestamp(t){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,t)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||(typeof this.env.trace>"u"?this.env.wasm.trace:this.env.trace))&&(this.device.features.has("chromium-experimental-timestamp-query-inside-passes")?this.queryType="inside-passes":this.device.features.has("timestamp-query")&&(this.queryType="at-passes"),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){at("info","captureBegin"),this.capturedCommandList.get(this.currentSessionId)||this.capturedCommandList.set(this.currentSessionId,[]),this.capturedPendingKernels.get(this.currentSessionId)||this.capturedPendingKernels.set(this.currentSessionId,[]),this.flush(),this.sessionStatus="capturing"}captureEnd(){at("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){at("info","replay"),this.sessionStatus="replaying";let t=this.capturedCommandList.get(this.currentSessionId),e=this.capturedPendingKernels.get(this.currentSessionId),r=t.length;this.pendingKernels=[];for(let n=0;n=this.maxDispatchNumber||this.queryType==="at-passes")&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}onReleaseSession(t){this.unregisterBuffers(t),this.capturedCommandList.has(t)&&this.capturedCommandList.delete(t),this.capturedPendingKernels.has(t)&&this.capturedPendingKernels.delete(t),this.gpuDataManager.onReleaseSession(t)}onRunStart(t){this.currentSessionId=t,this.setQueryType()}}}),Hm={};Dn(Hm,{init:()=>jm});var Ei,ip,jm,aw=J(()=>{$e(),nw(),un(),Me(),Ei=class qm{constructor(e,r,n,a){this.module=e,this.dataType=r,this.data=n,this.dims=a}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let e=K.size(this.dims);return e===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,e)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let e=K.size(this.dims);return e===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,e)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let e=K.size(this.dims);return e===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,e)}reshape(e){if(K.size(e)!==K.size(this.dims))throw new Error("Invalid new shape");return new qm(this.module,this.dataType,this.data,e)}},ip=class{constructor(t,e,r){this.module=t,this.backend=e,this.customDataOffset=0,this.customDataSize=0,this.adapterInfo=e.adapterInfo;let n=t.HEAPU32,a=r>>>2;this.opKernelContext=n[a++];let i=n[a++];this.outputCount=n[a++],this.customDataOffset=n[a++],this.customDataSize=n[a++];let s=[];for(let o=0;otypeof s=="number"?this.inputs[s]:s)??this.inputs,n=e?.outputs??[],a=(s,o,l)=>new Ei(this.module,o,this.output(s,l),l),i=(s,o)=>{let l=$a(s);if(!l)throw new Error(`Unsupported data type: ${s}`);let u=l*K.size(o),p=u>0?this.backend.gpuDataManager.create(u).id:0;return new Ei(this.module,s,p,o)};return this.backend.run(t,r,n,a,i,this.outputCount)}output(t,e){let r=this.module.stackSave();try{let n=this.module.stackAlloc((1+e.length)*4),a=n>>2;this.module.HEAPU32[a++]=e.length;for(let i=0;i{let a=e.jsepInit;if(!a)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(t==="webgpu"){let i=new Gm;await i.initialize(r,n),a("webgpu",[i,s=>i.alloc(s),s=>i.free(s),(s,o,l,u=!1)=>{if(u)at("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${s}, dst=${o}, size=${l}`),i.memcpy(s,o);else{at("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${s}, gpuDataId=${o}, size=${l}`);let p=e.HEAPU8.subarray(s>>>0,(s>>>0)+l);i.upload(o,p)}},async(s,o,l)=>{at("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${s}, dataOffset=${o}, size=${l}`),await i.download(s,()=>e.HEAPU8.subarray(o>>>0,(o>>>0)+l))},(s,o,l)=>i.createKernel(s,o,l,e.UTF8ToString(e._JsepGetNodeName(o))),s=>i.releaseKernel(s),(s,o,l,u)=>{at("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${l}, kernel=${s}, contextDataOffset=${o}`);let p=new ip(e,i,o);return i.computeKernel(s,p,u)},()=>i.captureBegin(),()=>i.captureEnd(),()=>i.replay()])}else a("webnn")}}),sp,Km,Ym,zr,op,Oo,Xm,Qm,so,Zm,Jm,eg,iw=J(()=>{_y(),yy(),$e(),Aa(),qo(),gh(),sp=(t,e)=>{st()._OrtInit(t,e)!==0&&qe("Can't initialize onnxruntime.")},Km=async t=>{sp(t.wasm.numThreads,Vi(t.logLevel))},Ym=async(t,e)=>{{let r=(aw(),nn(Hm)).init;if(e==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let n=t.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 a=t.webgpu.powerPreference;if(a!==void 0&&a!=="low-power"&&a!=="high-performance")throw new Error(`Invalid powerPreference setting: "${a}"`);let i=t.webgpu.forceFallbackAdapter;if(i!==void 0&&typeof i!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${i}"`);if(n=await navigator.gpu.requestAdapter({powerPreference:a,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.')}if(!t.wasm.simd)throw new Error("Not supported for WebGPU=ON and SIMD=OFF. Please set `env.wasm.simd` to true when using `webgpu` EP");await r("webgpu",st(),t,n)}if(e==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",st(),t)}}},zr=new Map,op=t=>{let e=st(),r=e.stackSave();try{let n=e.stackAlloc(8);return e._OrtGetInputOutputCount(t,n,n+4)!==0&&qe("Can't get session input/output count."),[e.HEAP32[n/4],e.HEAP32[n/4+1]]}finally{e.stackRestore(r)}},Oo=t=>{let e=st(),r=e._malloc(t.byteLength);if(r===0)throw new Error(`Can't create a session. failed to allocate a buffer of size ${t.byteLength}.`);return e.HEAPU8.set(t,r),[r,t.byteLength]},Xm=async(t,e)=>{let r,n,a=st();Array.isArray(t)?[r,n]=t:t.buffer===a.HEAPU8.buffer?[r,n]=[t.byteOffset,t.byteLength]:[r,n]=Oo(t);let i=0,s=0,o=0,l=[],u=[],p=[];try{if([s,l]=mh(e),e?.externalData&&a.mountExternalData){let $=[];for(let E of e.externalData){let T=typeof E=="string"?E:E.path;$.push(Gi(typeof E=="string"?E:E.data).then(A=>{a.mountExternalData(T,A)}))}await Promise.all($)}i=await a._OrtCreateSession(r,n,s),i===0&&qe("Can't create a session.");let[h,m]=op(i),d=!!e?.enableGraphCapture,_=[],w=[],v=[];for(let $=0;$$==="gpu-buffer")&&(o=a._OrtCreateBinding(i),o===0&&qe("Can't create IO binding."),S={handle:o,outputPreferredLocations:v,outputPreferredLocationsEncoded:v.map($=>vo($))}),zr.set(i,[i,u,p,S,d,!1]),[i,_,w]}catch(h){throw u.forEach(m=>a._OrtFree(m)),p.forEach(m=>a._OrtFree(m)),o!==0&&a._OrtReleaseBinding(o),i!==0&&a._OrtReleaseSession(i),h}finally{a._free(r),s!==0&&a._OrtReleaseSessionOptions(s),l.forEach(h=>a._free(h)),a.unmountExternalData?.()}},Qm=t=>{let e=st(),r=zr.get(t);if(!r)throw new Error(`cannot release session. invalid session id: ${t}`);let[n,a,i,s,o]=r;s&&(o&&e._OrtClearBoundOutputs(s.handle),e._OrtReleaseBinding(s.handle)),e.jsepOnReleaseSession?.(t),a.forEach(l=>e._OrtFree(l)),i.forEach(l=>e._OrtFree(l)),e._OrtReleaseSession(n),zr.delete(t)},so=(t,e,r,n,a,i=!1)=>{if(!t){e.push(0);return}let s=st(),o=t[0],l=t[1],u=t[3],p,h;if(o==="string"&&u==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(i&&u!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${a} when enableGraphCapture is true.`);if(u==="gpu-buffer"){let _=t[2].gpuBuffer,w=$a(bo(o));h=l.reduce((S,$)=>S*$,1)*w;let v=s.jsepRegisterBuffer;if(!v)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');p=v(n,a,_,h)}else{let _=t[2];if(Array.isArray(_)){h=4*_.length,p=s._malloc(h),r.push(p);let w=p/4;for(let v=0;v<_.length;v++){if(typeof _[v]!="string")throw new TypeError(`tensor data at index ${v} is not a string`);s.HEAPU32[w++]=nt(_[v],r)}}else h=_.byteLength,p=s._malloc(h),r.push(p),s.HEAPU8.set(new Uint8Array(_.buffer,_.byteOffset,h),p)}let m=s.stackSave(),d=s.stackAlloc(4*l.length);try{let _=d/4;l.forEach(v=>s.HEAP32[_++]=v);let w=s._OrtCreateTensor(bo(o),p,h,d,l.length,vo(u));w===0&&qe(`Can't create tensor for input/output. session=${n}, index=${a}.`),e.push(w)}finally{s.stackRestore(m)}},Zm=async(t,e,r,n,a,i)=>{let s=st(),o=zr.get(t);if(!o)throw new Error(`cannot run inference. invalid session id: ${t}`);let l=o[0],u=o[1],p=o[2],h=o[3],m=o[4],d=o[5],_=e.length,w=n.length,v=0,S=[],$=[],E=[],T=[],A=s.stackSave(),P=s.stackAlloc(_*4),B=s.stackAlloc(_*4),D=s.stackAlloc(w*4),q=s.stackAlloc(w*4);try{[v,S]=fh(i);for(let R=0;R<_;R++)so(r[R],$,T,t,e[R],m);for(let R=0;RVe*Ne,1);W=Jr(Ce);let pt=h?.outputPreferredLocations[n[R]];if(W==="string"){if(pt==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Ve=[],Ne=ae/4;for(let gt=0;gt0){let Ve=s.jsepGetBuffer;if(!Ve)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let Ne=Ve(ae),gt=$a(Ce);if(gt===void 0||!Yo(W))throw new Error(`Unsupported data type: ${W}`);ne=!0,M.push([W,Qe,{gpuBuffer:Ne,download:s.jsepCreateDownloader(Ne,We*gt,W),dispose:()=>{s._OrtReleaseTensor(G)}},"gpu-buffer"])}else{let Ve=Ko(W),Ne=new Ve(We);new Uint8Array(Ne.buffer,Ne.byteOffset,Ne.byteLength).set(s.HEAPU8.subarray(ae,ae+Ne.byteLength)),M.push([W,Qe,Ne,"cpu"])}}finally{s.stackRestore(re),W==="string"&&ae&&s._free(ae),ne||s._OrtReleaseTensor(G)}}return h&&!m&&(s._OrtClearBoundOutputs(h.handle),zr.set(t,[l,u,p,h,m,!1])),M}finally{s.stackRestore(A),$.forEach(H=>s._OrtReleaseTensor(H)),E.forEach(H=>s._OrtReleaseTensor(H)),T.forEach(H=>s._free(H)),v!==0&&s._OrtReleaseRunOptions(v),S.forEach(H=>s._free(H))}},Jm=t=>{let e=st(),r=zr.get(t);if(!r)throw new Error("invalid session id");let n=r[0],a=e._OrtEndProfiling(n);a===0&&qe("Can't get an profile file name."),e._OrtFree(a)},eg=t=>{let e=[];for(let r of t){let n=r[2];!Array.isArray(n)&&"buffer"in n&&e.push(n.buffer)}return e}}),sw=Bn((t,e)=>{e.exports='/*!\n * ONNX Runtime Web v1.18.0\n * Copyright (c) Microsoft Corporation. All rights reserved.\n * Licensed under the MIT License.\n */\n"use strict";(()=>{var ao=Object.defineProperty;var nl=Object.getOwnPropertyDescriptor;var ol=Object.getOwnPropertyNames;var il=Object.prototype.hasOwnProperty;var Y=(e,t)=>()=>(e&&(t=e(e=0)),t);var Wr=(e,t)=>()=>(t||e((t={exports:{}}).exports,t),t.exports),gn=(e,t)=>{for(var r in t)ao(e,r,{get:t[r],enumerable:!0})},al=(e,t,r,o)=>{if(t&&typeof t=="object"||typeof t=="function")for(let i of ol(t))!il.call(e,i)&&i!==r&&ao(e,i,{get:()=>t[i],enumerable:!(o=nl(t,i))||o.enumerable});return e};var wr=e=>al(ao({},"__esModule",{value:!0}),e);var so={};gn(so,{createReadStream:()=>Ai,readFile:()=>sl,readFileSync:()=>ul});var sl,ul,Ai,uo=Y(()=>{sl=void 0,ul=void 0,Ai=void 0});var lo={};gn(lo,{join:()=>dl});var dl,co=Y(()=>{dl=void 0});var Ei=Wr((Ti,po)=>{"use strict";var Ii=(()=>{var e=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(e||=__filename),function(t={}){var r=t,o,i,u=new Promise((s,m)=>{o=s,i=m});r.mountExternalData=(s,m)=>{(r.eb||(r.eb=new Map)).set(s,m)},r.unmountExternalData=()=>{delete r.eb};let a=()=>{let s=(g,$,T)=>(...B)=>{let H=dt,q=$?.();B=g(...B);let te=$?.();return q!==te&&(g=te,T(q),$=T=null),dt!=H?tn():B},m=g=>async(...$)=>{try{if(r.cb)throw Error("Session already started");let T=r.cb={xb:$[0],errors:[]},B=await g(...$);if(r.cb!==T)throw Error("Session mismatch");r.kb?.flush();let H=T.errors;if(0te),0r._OrtCreateSession,g=>r._OrtCreateSession=g),r._OrtRun=m(s(r._OrtRun,()=>r._OrtRun,g=>r._OrtRun=g)),r._OrtRunWithBinding=m(s(r._OrtRunWithBinding,()=>r._OrtRunWithBinding,g=>r._OrtRunWithBinding=g)),r._OrtBindInput=s(r._OrtBindInput,()=>r._OrtBindInput,g=>r._OrtBindInput=g),a=void 0};r.jsepInit=(s,m)=>{if(a?.(),s==="webgpu"){[r.kb,r.pb,r.tb,r.lb,r.sb,r.Ra,r.ub,r.wb,r.qb,r.rb,r.vb]=m;let g=r.kb;r.jsepRegisterBuffer=($,T,B,H)=>g.registerBuffer($,T,B,H),r.jsepGetBuffer=$=>g.getBuffer($),r.jsepCreateDownloader=($,T,B)=>g.createDownloader($,T,B),r.jsepOnReleaseSession=$=>{g.onReleaseSession($)},r.jsepOnRunStart=$=>g.onRunStart($)}};var c=Object.assign({},r),p="./this.program",h=(s,m)=>{throw m},d=typeof window=="object",y=typeof importScripts=="function",w=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",_="",v,S,A;if(w){var I=(uo(),wr(so)),x=(co(),wr(lo));_=y?x.dirname(_)+"/":__dirname+"/",v=(s,m)=>(s=Ke(s)?new URL(s):x.normalize(s),I.readFileSync(s,m?void 0:"utf8")),A=s=>(s=v(s,!0),s.buffer||(s=new Uint8Array(s)),s),S=(s,m,g,$=!0)=>{s=Ke(s)?new URL(s):x.normalize(s),I.readFile(s,$?void 0:"utf8",(T,B)=>{T?g(T):m($?B.buffer:B)})},!r.thisProgram&&1{throw process.exitCode=s,m}}else(d||y)&&(y?_=self.location.href:typeof document<"u"&&document.currentScript&&(_=document.currentScript.src),e&&(_=e),_.startsWith("blob:")?_="":_=_.substr(0,_.replace(/[?#].*/,"").lastIndexOf("/")+1),v=s=>{var m=new XMLHttpRequest;return m.open("GET",s,!1),m.send(null),m.responseText},y&&(A=s=>{var m=new XMLHttpRequest;return m.open("GET",s,!1),m.responseType="arraybuffer",m.send(null),new Uint8Array(m.response)}),S=(s,m,g)=>{var $=new XMLHttpRequest;$.open("GET",s,!0),$.responseType="arraybuffer",$.onload=()=>{$.status==200||$.status==0&&$.response?m($.response):g()},$.onerror=g,$.send(null)});var E=console.log.bind(console),P=console.error.bind(console);Object.assign(r,c),c=null;var O,R=!1,L,N,K,Q,he,W,se,Ce,We,ee,ae;function Ae(){var s=O.buffer;r.HEAP8=N=new Int8Array(s),r.HEAP16=Q=new Int16Array(s),r.HEAPU8=K=new Uint8Array(s),r.HEAPU16=he=new Uint16Array(s),r.HEAP32=W=new Int32Array(s),r.HEAPU32=se=new Uint32Array(s),r.HEAPF32=Ce=new Float32Array(s),r.HEAPF64=ae=new Float64Array(s),r.HEAP64=We=new BigInt64Array(s),r.HEAPU64=ee=new BigUint64Array(s)}var me=[],ie=[],ue=[],le=0,qe=null,G=null;function ne(s){throw s="Aborted("+s+")",P(s),R=!0,L=1,s=new WebAssembly.RuntimeError(s+". Build with -sASSERTIONS for more info."),i(s),s}var xe=s=>s.startsWith("data:application/octet-stream;base64,"),Ke=s=>s.startsWith("file://"),Be;if(Be="ort-wasm-simd.wasm",!xe(Be)){var Ge=Be;Be=r.locateFile?r.locateFile(Ge,_):_+Ge}function Ut(s){if(A)return A(s);throw"both async and sync fetching of the wasm failed"}function Ne(s){if(d||y){if(typeof fetch=="function"&&!Ke(s))return fetch(s,{credentials:"same-origin"}).then(m=>{if(!m.ok)throw`failed to load wasm binary file at \'${s}\'`;return m.arrayBuffer()}).catch(()=>Ut(s));if(S)return new Promise((m,g)=>{S(s,$=>m(new Uint8Array($)),g)})}return Promise.resolve().then(()=>Ut(s))}function Ye(s,m,g){return Ne(s).then($=>WebAssembly.instantiate($,m)).then(g,$=>{P(`failed to asynchronously prepare wasm: ${$}`),ne($)})}function mt(s,m){var g=Be;return typeof WebAssembly.instantiateStreaming!="function"||xe(g)||Ke(g)||w||typeof fetch!="function"?Ye(g,s,m):fetch(g,{credentials:"same-origin"}).then($=>WebAssembly.instantiateStreaming($,s).then(m,function(T){return P(`wasm streaming compile failed: ${T}`),P("falling back to ArrayBuffer instantiation"),Ye(g,s,m)}))}var Rt={824920:(s,m,g,$)=>{if(typeof r>"u"||!r.eb)return 1;if(s=je(s>>>0),s.startsWith("./")&&(s=s.substring(2)),s=r.eb.get(s),!s)return 2;if(m>>>=0,g>>>=0,m+g>s.byteLength)return 3;try{return K.set(s.subarray(m,m+g),$>>>0>>>0),0}catch{return 4}},825421:()=>{r.qb()},825452:()=>{r.rb()},825481:()=>{r.vb()},825506:s=>r.pb(s),825539:s=>r.tb(s),825571:(s,m,g)=>{r.lb(s,m,g,!0)},825610:(s,m,g)=>{r.lb(s,m,g)},825643:s=>{r.Ra("Abs",s,void 0)},825694:s=>{r.Ra("Neg",s,void 0)},825745:s=>{r.Ra("Floor",s,void 0)},825798:s=>{r.Ra("Ceil",s,void 0)},825850:s=>{r.Ra("Reciprocal",s,void 0)},825908:s=>{r.Ra("Sqrt",s,void 0)},825960:s=>{r.Ra("Exp",s,void 0)},826011:s=>{r.Ra("Erf",s,void 0)},826062:s=>{r.Ra("Sigmoid",s,void 0)},826117:(s,m,g)=>{r.Ra("HardSigmoid",s,{alpha:m,beta:g})},826196:s=>{r.Ra("Log",s,void 0)},826247:s=>{r.Ra("Sin",s,void 0)},826298:s=>{r.Ra("Cos",s,void 0)},826349:s=>{r.Ra("Tan",s,void 0)},826400:s=>{r.Ra("Asin",s,void 0)},826452:s=>{r.Ra("Acos",s,void 0)},826504:s=>{r.Ra("Atan",s,void 0)},826556:s=>{r.Ra("Sinh",s,void 0)},826608:s=>{r.Ra("Cosh",s,void 0)},826660:s=>{r.Ra("Asinh",s,void 0)},826713:s=>{r.Ra("Acosh",s,void 0)},826766:s=>{r.Ra("Atanh",s,void 0)},826819:s=>{r.Ra("Tanh",s,void 0)},826871:s=>{r.Ra("Not",s,void 0)},826922:(s,m,g)=>{r.Ra("Clip",s,{min:m,max:g})},826991:s=>{r.Ra("Clip",s,void 0)},827043:(s,m)=>{r.Ra("Elu",s,{alpha:m})},827101:s=>{r.Ra("Relu",s,void 0)},827153:(s,m)=>{r.Ra("LeakyRelu",s,{alpha:m})},827217:(s,m)=>{r.Ra("ThresholdedRelu",s,{alpha:m})},827287:(s,m)=>{r.Ra("Cast",s,{to:m})},827345:s=>{r.Ra("Add",s,void 0)},827396:s=>{r.Ra("Sub",s,void 0)},827447:s=>{r.Ra("Mul",s,void 0)},827498:s=>{r.Ra("Div",s,void 0)},827549:s=>{r.Ra("Pow",s,void 0)},827600:s=>{r.Ra("Equal",s,void 0)},827653:s=>{r.Ra("Greater",s,void 0)},827708:s=>{r.Ra("GreaterOrEqual",s,void 0)},827770:s=>{r.Ra("Less",s,void 0)},827822:s=>{r.Ra("LessOrEqual",s,void 0)},827881:(s,m,g,$,T)=>{r.Ra("ReduceMean",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828040:(s,m,g,$,T)=>{r.Ra("ReduceMax",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828198:(s,m,g,$,T)=>{r.Ra("ReduceMin",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828356:(s,m,g,$,T)=>{r.Ra("ReduceProd",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828515:(s,m,g,$,T)=>{r.Ra("ReduceSum",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828673:(s,m,g,$,T)=>{r.Ra("ReduceL1",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828830:(s,m,g,$,T)=>{r.Ra("ReduceL2",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828987:(s,m,g,$,T)=>{r.Ra("ReduceLogSum",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829148:(s,m,g,$,T)=>{r.Ra("ReduceSumSquare",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829312:(s,m,g,$,T)=>{r.Ra("ReduceLogSumExp",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829476:s=>{r.Ra("Where",s,void 0)},829529:(s,m,g)=>{r.Ra("Transpose",s,{perm:m?Array.from(W.subarray(m>>>0,g>>>0)):[]})},829637:(s,m,g,$)=>{r.Ra("DepthToSpace",s,{blocksize:m,mode:je(g),format:$?"NHWC":"NCHW"})},829770:(s,m,g,$)=>{r.Ra("DepthToSpace",s,{blocksize:m,mode:je(g),format:$?"NHWC":"NCHW"})},829903:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be)=>{r.Ra("ConvTranspose",s,{format:te?"NHWC":"NCHW",autoPad:m,dilations:[g],group:$,kernelShape:[T],pads:[B,H],strides:[q],wIsConst:()=>!!N[X>>>0],outputPadding:de?Array.from(W.subarray(de>>>0,Ee>>>0)):[],outputShape:Oe?Array.from(W.subarray(Oe>>>0,D>>>0)):[],activation:je(be)})},830304:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D)=>{r.Ra("ConvTranspose",s,{format:q?"NHWC":"NCHW",autoPad:m,dilations:Array.from(W.subarray(g>>>0,(g>>>0)+2>>>0)),group:$,kernelShape:Array.from(W.subarray(T>>>0,(T>>>0)+2>>>0)),pads:Array.from(W.subarray(B>>>0,(B>>>0)+4>>>0)),strides:Array.from(W.subarray(H>>>0,(H>>>0)+2>>>0)),wIsConst:()=>!!N[te>>>0],outputPadding:X?Array.from(W.subarray(X>>>0,de>>>0)):[],outputShape:Ee?Array.from(W.subarray(Ee>>>0,Oe>>>0)):[],activation:je(D)})},830869:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be)=>{r.Ra("ConvTranspose",s,{format:te?"NHWC":"NCHW",autoPad:m,dilations:[g],group:$,kernelShape:[T],pads:[B,H],strides:[q],wIsConst:()=>!!N[X>>>0],outputPadding:de?Array.from(W.subarray(de>>>0,Ee>>>0)):[],outputShape:Oe?Array.from(W.subarray(Oe>>>0,D>>>0)):[],activation:je(be)})},831270:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D)=>{r.Ra("ConvTranspose",s,{format:q?"NHWC":"NCHW",autoPad:m,dilations:Array.from(W.subarray(g>>>0,(g>>>0)+2>>>0)),group:$,kernelShape:Array.from(W.subarray(T>>>0,(T>>>0)+2>>>0)),pads:Array.from(W.subarray(B>>>0,(B>>>0)+4>>>0)),strides:Array.from(W.subarray(H>>>0,(H>>>0)+2>>>0)),wIsConst:()=>!!N[te>>>0],outputPadding:X?Array.from(W.subarray(X>>>0,de>>>0)):[],outputShape:Ee?Array.from(W.subarray(Ee>>>0,Oe>>>0)):[],activation:je(D)})},831835:(s,m)=>{r.Ra("GlobalAveragePool",s,{format:m?"NHWC":"NCHW"})},831926:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("AveragePool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832210:(s,m)=>{r.Ra("GlobalAveragePool",s,{format:m?"NHWC":"NCHW"})},832301:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("AveragePool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832585:(s,m)=>{r.Ra("GlobalMaxPool",s,{format:m?"NHWC":"NCHW"})},832672:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("MaxPool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832952:(s,m)=>{r.Ra("GlobalMaxPool",s,{format:m?"NHWC":"NCHW"})},833039:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("MaxPool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},833319:(s,m,g,$,T)=>{r.Ra("Gemm",s,{alpha:m,beta:g,transA:$,transB:T})},833423:s=>{r.Ra("MatMul",s,void 0)},833477:(s,m,g,$)=>{r.Ra("ArgMax",s,{keepDims:!!m,selectLastIndex:!!g,axis:$})},833585:(s,m,g,$)=>{r.Ra("ArgMin",s,{keepDims:!!m,selectLastIndex:!!g,axis:$})},833693:(s,m)=>{r.Ra("Softmax",s,{axis:m})},833756:(s,m)=>{r.Ra("Concat",s,{axis:m})},833816:(s,m,g,$,T)=>{r.Ra("Split",s,{axis:m,numOutputs:g,splitSizes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},833956:s=>{r.Ra("Expand",s,void 0)},834010:(s,m)=>{r.Ra("Gather",s,{axis:Number(m)})},834081:(s,m)=>{r.Ra("GatherElements",s,{axis:Number(m)})},834160:(s,m,g,$,T,B,H,q,te,X,de)=>{r.Ra("Resize",s,{antialias:m,axes:g?Array.from(W.subarray(g>>>0,$>>>0)):[],coordinateTransformMode:je(T),cubicCoeffA:B,excludeOutside:H,extrapolationValue:q,keepAspectRatioPolicy:je(te),mode:je(X),nearestMode:je(de)})},834506:(s,m,g,$,T,B,H)=>{r.Ra("Slice",s,{starts:m?Array.from(W.subarray(m>>>0,g>>>0)):[],ends:$?Array.from(W.subarray($>>>0,T>>>0)):[],axes:B?Array.from(W.subarray(B>>>0,H>>>0)):[]})},834722:s=>{r.Ra("Tile",s,void 0)},834774:(s,m,g,$)=>{r.Ra("LayerNormalization",s,{axis:m,epsilon:g,simplified:!!$})},834885:(s,m,g)=>{r.Ra("InstanceNormalization",s,{epsilon:m,format:g?"NHWC":"NCHW"})},834999:(s,m,g)=>{r.Ra("InstanceNormalization",s,{epsilon:m,format:g?"NHWC":"NCHW"})},835113:s=>{r.Ra("Range",s,void 0)},835166:(s,m)=>{r.Ra("Einsum",s,{equation:je(m)})},835247:(s,m,g,$,T)=>{r.Ra("Pad",s,{mode:m,value:g,pads:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},835374:(s,m,g,$,T,B)=>{r.Ra("BatchNormalization",s,{epsilon:m,momentum:g,spatial:!!T,trainingMode:!!$,format:B?"NHWC":"NCHW"})},835543:(s,m,g,$,T,B)=>{r.Ra("BatchNormalization",s,{epsilon:m,momentum:g,spatial:!!T,trainingMode:!!$,format:B?"NHWC":"NCHW"})},835712:(s,m,g)=>{r.Ra("CumSum",s,{exclusive:Number(m),reverse:Number(g)})},835809:(s,m,g,$,T,B,H,q,te)=>{r.Ra("Attention",s,{numHeads:m,isUnidirectional:g,maskFilterValue:$,scale:T,doRotary:B,qkvHiddenSizes:H?Array.from(W.subarray(Number(q)>>>0,Number(q)+H>>>0)):[],pastPresentShareBuffer:!!te})},836081:s=>{r.Ra("BiasAdd",s,void 0)},836136:s=>{r.Ra("BiasSplitGelu",s,void 0)},836197:s=>{r.Ra("FastGelu",s,void 0)},836253:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe)=>{r.Ra("Conv",s,{format:te?"NHWC":"NCHW",auto_pad:m,dilations:[g],group:$,kernel_shape:[T],pads:B?Array.from(W.subarray(B>>>0,H>>>0)):[],strides:[q],w_is_const:()=>!!N[X>>>0],activation:je(de),activation_params:Ee?Array.from(Ce.subarray(Ee>>>0,Oe>>>0)):[]})},836623:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("Conv",s,{format:Ee?"NHWC":"NCHW",auto_pad:m,dilations:[g,$],group:T,kernel_shape:[B,H],pads:q?Array.from(W.subarray(q>>>0,te>>>0)):[],strides:[X,de],w_is_const:()=>!!N[Oe>>>0],activation:je(D),activation_params:be?Array.from(Ce.subarray(be>>>0,Ie>>>0)):[]})},837014:s=>{r.Ra("Gelu",s,void 0)},837066:(s,m,g,$,T,B)=>{r.Ra("MatMulNBits",s,{k:m,n:g,accuracyLevel:$,bits:T,blockSize:B})},837193:(s,m,g,$,T,B)=>{r.Ra("MultiHeadAttention",s,{numHeads:m,isUnidirectional:g,maskFilterValue:$,scale:T,doRotary:B})},837352:(s,m,g,$,T)=>{r.Ra("RotaryEmbedding",s,{interleaved:!!m,numHeads:g,rotaryEmbeddingDim:$,scale:T})},837491:(s,m,g)=>{r.Ra("SkipLayerNormalization",s,{epsilon:m,simplified:!!g})},837593:(s,m,g)=>{r.Ra("SkipLayerNormalization",s,{epsilon:m,simplified:!!g})},837695:(s,m,g,$)=>{r.Ra("LayerNormalization",s,{axis:m,epsilon:g,simplified:!!$})},837806:s=>{r.ub(s)},837840:(s,m)=>r.wb(s,m,r.cb.xb,r.cb.errors)};function qt(s){this.name="ExitStatus",this.message=`Program terminated with exit(${s})`,this.status=s}class Vt{constructor(m){this.hb=m-24}}var sr=0,jt=0,xr=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Kt=(s,m,g)=>{m>>>=0;var $=m+g;for(g=m;s[g]&&!(g>=$);)++g;if(16T?$+=String.fromCharCode(T):(T-=65536,$+=String.fromCharCode(55296|T>>10,56320|T&1023))}}else $+=String.fromCharCode(T)}return $},je=(s,m)=>(s>>>=0)?Kt(K,s,m):"",Wt=s=>{for(var m=0,g=0;g=$?m++:2047>=$?m+=2:55296<=$&&57343>=$?(m+=4,++g):m+=3}return m},at=(s,m,g,$)=>{if(g>>>=0,!(0<$))return 0;var T=g;$=g+$-1;for(var B=0;B=H){var q=s.charCodeAt(++B);H=65536+((H&1023)<<10)|q&1023}if(127>=H){if(g>=$)break;m[g++>>>0]=H}else{if(2047>=H){if(g+1>=$)break;m[g++>>>0]=192|H>>6}else{if(65535>=H){if(g+2>=$)break;m[g++>>>0]=224|H>>12}else{if(g+3>=$)break;m[g++>>>0]=240|H>>18,m[g++>>>0]=128|H>>12&63}m[g++>>>0]=128|H>>6&63}m[g++>>>0]=128|H&63}}return m[g>>>0]=0,g-T},Cr,ft=s=>{for(var m="";K[s>>>0];)m+=Cr[K[s++>>>0]];return m},ur={},Ar={},Ir={},st;function Tr(s,m,g={}){var $=m.name;if(!s)throw new st(`type "${$}" must have a positive integer typeid pointer`);if(Ar.hasOwnProperty(s)){if(g.nb)return;throw new st(`Cannot register type \'${$}\' twice`)}Ar[s]=m,delete Ir[s],ur.hasOwnProperty(s)&&(m=ur[s],delete ur[s],m.forEach(T=>T()))}function ht(s,m,g={}){if(!("argPackAdvance"in m))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Tr(s,m,g)}var bt=(s,m,g)=>{switch(m){case 1:return g?$=>N[$>>>0]:$=>K[$>>>0];case 2:return g?$=>Q[$>>>1>>>0]:$=>he[$>>>1>>>0];case 4:return g?$=>W[$>>>2>>>0]:$=>se[$>>>2>>>0];case 8:return g?$=>We[$>>>3]:$=>ee[$>>>3];default:throw new TypeError(`invalid integer width (${m}): ${s}`)}},Er=[],Te=[];function dr(s){s>>>=0,9{if(!s)throw new st("Cannot use deleted val. handle = "+s);return Te[s]},ut=s=>{switch(s){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let m=Er.pop()||Te.length;return Te[m]=s,Te[m+1]=1,m}};function lr(s){return this.fromWireType(se[s>>>2>>>0])}var Fn={name:"emscripten::val",fromWireType:s=>{var m=Xe(s);return dr(s),m},toWireType:(s,m)=>ut(m),argPackAdvance:8,readValueFromPointer:lr,bb:null},ge=(s,m)=>{switch(m){case 4:return function(g){return this.fromWireType(Ce[g>>>2>>>0])};case 8:return function(g){return this.fromWireType(ae[g>>>3>>>0])};default:throw new TypeError(`invalid float width (${m}): ${s}`)}},Yt=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,jr=(s,m)=>{for(var g=s>>1,$=g+m/2;!(g>=$)&&he[g>>>0];)++g;if(g<<=1,32>>0,g>>>0));for(g="",$=0;!($>=m/2);++$){var T=Q[s+2*$>>>1>>>0];if(T==0)break;g+=String.fromCharCode(T)}return g},qn=(s,m,g)=>{if(g??=2147483647,2>g)return 0;g-=2;var $=m;g=g<2*s.length?g/2:s.length;for(var T=0;T>>1>>>0]=s.charCodeAt(T),m+=2;return Q[m>>>1>>>0]=0,m-$},Kr=s=>2*s.length,jn=(s,m)=>{for(var g=0,$="";!(g>=m/4);){var T=W[s+4*g>>>2>>>0];if(T==0)break;++g,65536<=T?(T-=65536,$+=String.fromCharCode(55296|T>>10,56320|T&1023)):$+=String.fromCharCode(T)}return $},Yr=(s,m,g)=>{if(m>>>=0,g??=2147483647,4>g)return 0;var $=m;g=$+g-4;for(var T=0;T=B){var H=s.charCodeAt(++T);B=65536+((B&1023)<<10)|H&1023}if(W[m>>>2>>>0]=B,m+=4,m+4>g)break}return W[m>>>2>>>0]=0,m-$},Zr=s=>{for(var m=0,g=0;g=$&&++g,m+=4}return m},cr=(s,m)=>{var g=Ar[s];if(g===void 0)throw s=Mr(s),g=ft(s),rt(s),new st(`${m} has unknown type ${g}`);return g},Pr=(s,m,g)=>{var $=[];return s=s.toWireType($,g),$.length&&(se[m>>>2>>>0]=ut($)),s},He=s=>{try{s()}catch(m){ne(m)}},Xr=s=>{if(!R)try{s();try{L=L=s=L,r.onExit?.(s),R=!0,h(s,new qt(s))}catch(m){m instanceof qt||m=="unwind"||h(1,m)}}catch(m){m instanceof qt||m=="unwind"||h(1,m)}};function Qr(){var s=pe,m={};for(let[g,$]of Object.entries(s))m[g]=typeof $=="function"?(...T)=>{Zt.push(g);try{return $(...T)}finally{R||(Zt.pop(),dt&>===1&&Zt.length===0&&(gt=0,He(dn),typeof Fibers<"u"&&Fibers.Db()))}}:$;return m}var gt=0,dt=null,Bt=0,Zt=[],kr={},Or={},Jr=0,pr=null,en=[];function tn(){return new Promise((s,m)=>{pr={resolve:s,reject:m}})}function rn(){var s=Nt(65548),m=s+12;se[s>>>2>>>0]=m,se[s+4>>>2>>>0]=m+65536,m=Zt[0];var g=kr[m];return g===void 0&&(g=Jr++,kr[m]=g,Or[g]=m),W[s+8>>>2>>>0]=g,s}function nn(s){if(!R){if(gt===0){var m=!1,g=!1;s(($=0)=>{if(!R&&(Bt=$,m=!0,g)){gt=2,He(()=>vt(dt)),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.resume(),$=!1;try{var T=(0,pe[Or[W[dt+8>>>2>>>0]]])()}catch(q){T=q,$=!0}var B=!1;if(!dt){var H=pr;H&&(pr=null,($?H.reject:H.resolve)(T),B=!0)}if($&&!B)throw T}}),g=!0,m||(gt=1,dt=rn(),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.pause(),He(()=>Jt(dt)))}else gt===2?(gt=0,He(ct),rt(dt),dt=null,en.forEach(Xr)):ne(`invalid state: ${gt}`);return Bt}}function Rr(s){return nn(m=>{s().then(m)})}var Xt=[],on={},nt=s=>{var m=on[s];return m===void 0?ft(s):m},mr=()=>typeof globalThis=="object"?globalThis:Function("return this")(),Br=s=>{var m=Xt.length;return Xt.push(s),m},Kn=(s,m)=>{for(var g=Array(s),$=0;$>>2>>>0],"parameter "+$);return g},wt=(s,m)=>Object.defineProperty(m,"name",{value:s});function Yn(s){var m=Function;if(!(m instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof m} which is not a function`);var g=wt(m.name||"unknownFunctionName",function(){});return g.prototype=m.prototype,g=new g,s=m.apply(g,s),s instanceof Object?s:g}var Je=s=>s%4===0&&(s%100!==0||s%400===0),Dr=[0,31,60,91,121,152,182,213,244,274,305,335],fr=[0,31,59,90,120,151,181,212,243,273,304,334],lt=[],hr=(s,m)=>{lt.length=0;for(var g;g=K[s++>>>0];){var $=g!=105;$&=g!=112,m+=$&&m%8?4:0,lt.push(g==112?se[m>>>2>>>0]:g==106?We[m>>>3]:g==105?W[m>>>2>>>0]:ae[m>>>3>>>0]),m+=$?8:4}return lt},Qe={},ot=()=>{if(!Qt){var s={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",_:p||"./this.program"},m;for(m in Qe)Qe[m]===void 0?delete s[m]:s[m]=Qe[m];var g=[];for(m in s)g.push(`${m}=${s[m]}`);Qt=g}return Qt},Qt,Zn=[null,[],[]],an=[31,29,31,30,31,30,31,31,30,31,30,31],zr=[31,28,31,30,31,30,31,31,30,31,30,31];function Xn(s){var m=Array(Wt(s)+1);return at(s,m,0,m.length),m}function sn(s,m,g,$){function T(D,be,Ie){for(D=typeof D=="number"?D.toString():D||"";D.lengthMt?-1:0zt-D.getDate())be-=zt-D.getDate()+1,D.setDate(1),11>Ie?D.setMonth(Ie+1):(D.setMonth(0),D.setFullYear(D.getFullYear()+1));else{D.setDate(D.getDate()+be);break}}return Ie=new Date(D.getFullYear()+1,0,4),be=q(new Date(D.getFullYear(),0,4)),Ie=q(Ie),0>=H(be,D)?0>=H(Ie,D)?D.getFullYear()+1:D.getFullYear():D.getFullYear()-1}s>>>=0,m>>>=0,g>>>=0,$>>>=0;var X=se[$+40>>>2>>>0];$={Ab:W[$>>>2>>>0],zb:W[$+4>>>2>>>0],fb:W[$+8>>>2>>>0],jb:W[$+12>>>2>>>0],gb:W[$+16>>>2>>>0],ab:W[$+20>>>2>>>0],Va:W[$+24>>>2>>>0],$a:W[$+28>>>2>>>0],Cb:W[$+32>>>2>>>0],yb:W[$+36>>>2>>>0],Bb:X?je(X):""},g=je(g),X={"%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"};for(var de in X)g=g.replace(new RegExp(de,"g"),X[de]);var Ee="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Oe="January February March April May June July August September October November December".split(" ");X={"%a":D=>Ee[D.Va].substring(0,3),"%A":D=>Ee[D.Va],"%b":D=>Oe[D.gb].substring(0,3),"%B":D=>Oe[D.gb],"%C":D=>B((D.ab+1900)/100|0,2),"%d":D=>B(D.jb,2),"%e":D=>T(D.jb,2," "),"%g":D=>te(D).toString().substring(2),"%G":te,"%H":D=>B(D.fb,2),"%I":D=>(D=D.fb,D==0?D=12:12{for(var be=0,Ie=0;Ie<=D.gb-1;be+=(Je(D.ab+1900)?an:zr)[Ie++]);return B(D.jb+be,3)},"%m":D=>B(D.gb+1,2),"%M":D=>B(D.zb,2),"%n":()=>`\n`,"%p":D=>0<=D.fb&&12>D.fb?"AM":"PM","%S":D=>B(D.Ab,2),"%t":()=>" ","%u":D=>D.Va||7,"%U":D=>B(Math.floor((D.$a+7-D.Va)/7),2),"%V":D=>{var be=Math.floor((D.$a+7-(D.Va+6)%7)/7);if(2>=(D.Va+371-D.$a-2)%7&&be++,be)be==53&&(Ie=(D.Va+371-D.$a)%7,Ie==4||Ie==3&&Je(D.ab)||(be=1));else{be=52;var Ie=(D.Va+7-D.$a-1)%7;(Ie==4||Ie==5&&Je(D.ab%400-1))&&be++}return B(be,2)},"%w":D=>D.Va,"%W":D=>B(Math.floor((D.$a+7-(D.Va+6)%7)/7),2),"%y":D=>(D.ab+1900).toString().substring(2),"%Y":D=>D.ab+1900,"%z":D=>{D=D.yb;var be=0<=D;return D=Math.abs(D)/60,(be?"+":"-")+("0000"+(D/60*100+D%60)).slice(-4)},"%Z":D=>D.Bb,"%%":()=>"%"},g=g.replace(/%%/g,"\\0\\0");for(de in X)g.includes(de)&&(g=g.replace(new RegExp(de,"g"),X[de]($)));return g=g.replace(/\\0\\0/g,"%"),de=Xn(g),de.length>m?0:(N.set(de,s>>>0),de.length-1)}for(var un=Array(256),gr=0;256>gr;++gr)un[gr]=String.fromCharCode(gr);Cr=un,st=r.BindingError=class extends Error{constructor(s){super(s),this.name="BindingError"}},r.InternalError=class extends Error{constructor(s){super(s),this.name="InternalError"}},Te.push(0,1,void 0,1,null,1,!0,1,!1,1),r.count_emval_handles=()=>Te.length/2-5-Er.length;var Qn={ia:function(s,m,g){return Rr(async()=>{await r.sb(s,m,g)})},a:function(s,m,g){s>>>=0;var $=new Vt(s);throw se[$.hb+16>>>2>>>0]=0,se[$.hb+4>>>2>>>0]=m>>>0,se[$.hb+8>>>2>>>0]=g>>>0,sr=s,jt++,sr},y:function(){return 0},ea:function(){},R:function(){},T:function(){},ga:function(){return 0},ca:function(){},Z:function(){},ba:function(){},G:function(){},S:function(){},P:function(){},da:function(){},Q:function(){},C:function(s,m,g){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:$=>$,toWireType:function($,T){if(typeof T!="bigint"&&typeof T!="number")throw T===null?T="null":($=typeof T,T=$==="object"||$==="array"||$==="function"?T.toString():""+T),new TypeError(`Cannot convert "${T}" to ${this.name}`);return typeof T=="number"&&(T=BigInt(T)),T},argPackAdvance:8,readValueFromPointer:bt(m,g>>>0,m.indexOf("u")==-1),bb:null})},K:function(s,m,g,$){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:function(T){return!!T},toWireType:function(T,B){return B?g:$},argPackAdvance:8,readValueFromPointer:function(T){return this.fromWireType(K[T>>>0])},bb:null})},J:function(s){return ht(s>>>0,Fn)},B:function(s,m,g){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:$=>$,toWireType:($,T)=>T,argPackAdvance:8,readValueFromPointer:ge(m,g>>>0),bb:null})},s:function(s,m,g,$,T){if(s>>>=0,g>>>=0,m=ft(m>>>0),T===-1&&(T=4294967295),T=q=>q,$===0){var B=32-8*g;T=q=>q<>>B}var H=m.includes("unsigned")?function(q,te){return te>>>0}:function(q,te){return te};ht(s,{name:m,fromWireType:T,toWireType:H,argPackAdvance:8,readValueFromPointer:bt(m,g,$!==0),bb:null})},o:function(s,m,g){function $(B){return new T(N.buffer,se[B+4>>>2>>>0],se[B>>>2>>>0])}var T=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][m];g=ft(g>>>0),ht(s>>>0,{name:g,fromWireType:$,argPackAdvance:8,readValueFromPointer:$},{nb:!0})},D:function(s,m){m=ft(m>>>0);var g=m==="std::string";ht(s>>>0,{name:m,fromWireType:function($){var T=se[$>>>2>>>0],B=$+4;if(g)for(var H=B,q=0;q<=T;++q){var te=B+q;if(q==T||K[te>>>0]==0){if(H=je(H,te-H),X===void 0)var X=H;else X+=String.fromCharCode(0),X+=H;H=te+1}}else{for(X=Array(T),q=0;q>>0]);X=X.join("")}return rt($),X},toWireType:function($,T){T instanceof ArrayBuffer&&(T=new Uint8Array(T));var B=typeof T=="string";if(!(B||T instanceof Uint8Array||T instanceof Uint8ClampedArray||T instanceof Int8Array))throw new st("Cannot pass non-string to std::string");var H=g&&B?Wt(T):T.length,q=Nt(4+H+1),te=q+4;if(se[q>>>2>>>0]=H,g&&B)at(T,K,te,H+1);else if(B)for(B=0;B>>0]=X}else for(B=0;B>>0]=T[B];return $!==null&&$.push(rt,q),q},argPackAdvance:8,readValueFromPointer:lr,bb($){rt($)}})},x:function(s,m,g){if(m>>>=0,g>>>=0,g=ft(g),m===2)var $=jr,T=qn,B=Kr,H=q=>he[q>>>1>>>0];else m===4&&($=jn,T=Yr,B=Zr,H=q=>se[q>>>2>>>0]);ht(s>>>0,{name:g,fromWireType:q=>{for(var te=se[q>>>2>>>0],X,de=q+4,Ee=0;Ee<=te;++Ee){var Oe=q+4+Ee*m;(Ee==te||H(Oe)==0)&&(de=$(de,Oe-de),X===void 0?X=de:(X+=String.fromCharCode(0),X+=de),de=Oe+m)}return rt(q),X},toWireType:(q,te)=>{if(typeof te!="string")throw new st(`Cannot pass non-string to C++ string type ${g}`);var X=B(te),de=Nt(4+X+m);return se[de>>>2>>>0]=X/m,T(te,de+4,X+m),q!==null&&q.push(rt,de),de},argPackAdvance:8,readValueFromPointer:lr,bb(q){rt(q)}})},L:function(s,m){m=ft(m>>>0),ht(s>>>0,{ob:!0,name:m,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ha:()=>1,u:function(s,m,g){return m>>>=0,g>>>=0,s=Xe(s>>>0),m=cr(m,"emval::as"),Pr(m,g,s)},w:function(s){return s>>>=0,Rr(()=>(s=Xe(s),s.then(ut)))},n:function(s,m,g,$){return g>>>=0,$>>>=0,s=Xt[s>>>0],m=Xe(m>>>0),s(null,m,g,$)},j:function(s,m,g,$,T){return g>>>=0,$>>>=0,T>>>=0,s=Xt[s>>>0],m=Xe(m>>>0),g=nt(g),s(m,m[g],$,T)},b:dr,A:function(s,m){return m>>>=0,s=Xe(s>>>0),m=Xe(m),s==m},m:function(s){return s>>>=0,s===0?ut(mr()):(s=nt(s),ut(mr()[s]))},i:function(s,m,g){m=Kn(s,m>>>0);var $=m.shift();s--;var T=`return function (obj, func, destructorsRef, args) {\n`,B=0,H=[];g===0&&H.push("obj");for(var q=["retType"],te=[$],X=0;Xde.name).join(", ")}) => ${$.name}>`,Br(wt(g,s))},r:function(s,m){return m>>>=0,s=Xe(s>>>0),m=Xe(m),ut(s[m])},e:function(s){s>>>=0,9>>0);for(var m=Array(s.length),g=0;g>>0))},k:function(){return ut({})},h:function(s){s>>>=0;for(var m=Xe(s);m.length;){var g=m.pop();m.pop()(g)}dr(s)},g:function(s,m,g){m>>>=0,g>>>=0,s=Xe(s>>>0),m=Xe(m),g=Xe(g),s[m]=g},c:function(s,m){return m>>>=0,s=cr(s>>>0,"_emval_take_value"),s=s.readValueFromPointer(m),ut(s)},W:function(s,m){s=-9007199254740992>s||9007199254740992>>=0,s=new Date(1e3*s),W[m>>>2>>>0]=s.getUTCSeconds(),W[m+4>>>2>>>0]=s.getUTCMinutes(),W[m+8>>>2>>>0]=s.getUTCHours(),W[m+12>>>2>>>0]=s.getUTCDate(),W[m+16>>>2>>>0]=s.getUTCMonth(),W[m+20>>>2>>>0]=s.getUTCFullYear()-1900,W[m+24>>>2>>>0]=s.getUTCDay(),W[m+28>>>2>>>0]=(s.getTime()-Date.UTC(s.getUTCFullYear(),0,1,0,0,0,0))/864e5|0},X:function(s,m){s=-9007199254740992>s||9007199254740992>>=0,s=new Date(1e3*s),W[m>>>2>>>0]=s.getSeconds(),W[m+4>>>2>>>0]=s.getMinutes(),W[m+8>>>2>>>0]=s.getHours(),W[m+12>>>2>>>0]=s.getDate(),W[m+16>>>2>>>0]=s.getMonth(),W[m+20>>>2>>>0]=s.getFullYear()-1900,W[m+24>>>2>>>0]=s.getDay(),W[m+28>>>2>>>0]=(Je(s.getFullYear())?Dr:fr)[s.getMonth()]+s.getDate()-1|0,W[m+36>>>2>>>0]=-(60*s.getTimezoneOffset());var g=new Date(s.getFullYear(),6,1).getTimezoneOffset(),$=new Date(s.getFullYear(),0,1).getTimezoneOffset();W[m+32>>>2>>>0]=(g!=$&&s.getTimezoneOffset()==Math.min($,g))|0},Y:function(s){s>>>=0;var m=new Date(W[s+20>>>2>>>0]+1900,W[s+16>>>2>>>0],W[s+12>>>2>>>0],W[s+8>>>2>>>0],W[s+4>>>2>>>0],W[s>>>2>>>0],0),g=W[s+32>>>2>>>0],$=m.getTimezoneOffset(),T=new Date(m.getFullYear(),6,1).getTimezoneOffset(),B=new Date(m.getFullYear(),0,1).getTimezoneOffset(),H=Math.min(B,T);return 0>g?W[s+32>>>2>>>0]=+(T!=B&&H==$):0>>2>>>0]=m.getDay(),W[s+28>>>2>>>0]=(Je(m.getFullYear())?Dr:fr)[m.getMonth()]+m.getDate()-1|0,W[s>>>2>>>0]=m.getSeconds(),W[s+4>>>2>>>0]=m.getMinutes(),W[s+8>>>2>>>0]=m.getHours(),W[s+12>>>2>>>0]=m.getDate(),W[s+16>>>2>>>0]=m.getMonth(),W[s+20>>>2>>>0]=m.getYear(),s=m.getTime(),BigInt(isNaN(s)?-1:s/1e3)},U:function(){return-52},V:function(){},N:function(s,m,g,$){g>>>=0,$>>>=0;var T=new Date().getFullYear(),B=new Date(T,0,1),H=new Date(T,6,1);T=B.getTimezoneOffset();var q=H.getTimezoneOffset();se[s>>>0>>>2>>>0]=60*Math.max(T,q),W[m>>>0>>>2>>>0]=+(T!=q),s=te=>te.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],B=s(B),H=s(H),q{ne("")},d:function(s,m,g){return s>>>=0,m=hr(m>>>0,g>>>0),Rt[s](...m)},I:function(s,m,g){return s>>>=0,m=hr(m>>>0,g>>>0),Rt[s](...m)},H:()=>Date.now(),O:function(){return 4294901760},q:()=>performance.now(),M:function(s){s>>>=0;var m=K.length;if(4294901760=g;g*=2){var $=m*(1+.2/g);$=Math.min($,s+100663296);var T=Math;$=Math.max(s,$);e:{T=(T.min.call(T,4294901760,$+(65536-$%65536)%65536)-O.buffer.byteLength+65535)/65536;try{O.grow(T),Ae();var B=1;break e}catch{}B=void 0}if(B)return!0}return!1},$:function(s,m){s>>>=0,m>>>=0;var g=0;return ot().forEach(($,T)=>{var B=m+g;for(T=se[s+4*T>>>2>>>0]=B,B=0;B<$.length;++B)N[T++>>>0]=$.charCodeAt(B);N[T>>>0]=0,g+=$.length+1}),0},aa:function(s,m){s>>>=0,m>>>=0;var g=ot();se[s>>>2>>>0]=g.length;var $=0;return g.forEach(T=>$+=T.length+1),se[m>>>2>>>0]=$,0},z:()=>52,F:function(){return 52},_:function(){return 70},E:function(s,m,g,$){m>>>=0,g>>>=0,$>>>=0;for(var T=0,B=0;B>>2>>>0],q=se[m+4>>>2>>>0];m+=8;for(var te=0;te>>0],de=Zn[s];X===0||X===10?((s===1?E:P)(Kt(de,0)),de.length=0):de.push(X)}T+=q}return se[$>>>2>>>0]=T,0},fa:sn,p:function(s,m,g,$){return sn(s>>>0,m>>>0,g>>>0,$>>>0)}},pe=function(){function s(g){return pe=g.exports,pe=Qr(),pe=ln(),O=pe.ja,Ae(),ie.unshift(pe.ka),le--,le==0&&(qe!==null&&(clearInterval(qe),qe=null),G&&(g=G,G=null,g())),pe}var m={a:Qn};if(le++,r.instantiateWasm)try{return r.instantiateWasm(m,s)}catch(g){P(`Module.instantiateWasm callback failed with error: ${g}`),i(g)}return mt(m,function(g){s(g.instance)}).catch(i),{}}(),Mr=s=>(Mr=pe.la)(s);r._OrtInit=(s,m)=>(r._OrtInit=pe.ma)(s,m),r._OrtGetLastError=(s,m)=>(r._OrtGetLastError=pe.na)(s,m),r._OrtCreateSessionOptions=(s,m,g,$,T,B,H,q,te,X)=>(r._OrtCreateSessionOptions=pe.oa)(s,m,g,$,T,B,H,q,te,X),r._OrtAppendExecutionProvider=(s,m)=>(r._OrtAppendExecutionProvider=pe.pa)(s,m),r._OrtAddFreeDimensionOverride=(s,m,g)=>(r._OrtAddFreeDimensionOverride=pe.qa)(s,m,g),r._OrtAddSessionConfigEntry=(s,m,g)=>(r._OrtAddSessionConfigEntry=pe.ra)(s,m,g),r._OrtReleaseSessionOptions=s=>(r._OrtReleaseSessionOptions=pe.sa)(s),r._OrtCreateSession=(s,m,g)=>(r._OrtCreateSession=pe.ta)(s,m,g),r._OrtReleaseSession=s=>(r._OrtReleaseSession=pe.ua)(s),r._OrtGetInputOutputCount=(s,m,g)=>(r._OrtGetInputOutputCount=pe.va)(s,m,g),r._OrtGetInputName=(s,m)=>(r._OrtGetInputName=pe.wa)(s,m),r._OrtGetOutputName=(s,m)=>(r._OrtGetOutputName=pe.xa)(s,m),r._OrtFree=s=>(r._OrtFree=pe.ya)(s),r._OrtCreateTensor=(s,m,g,$,T,B)=>(r._OrtCreateTensor=pe.za)(s,m,g,$,T,B),r._OrtGetTensorData=(s,m,g,$,T)=>(r._OrtGetTensorData=pe.Aa)(s,m,g,$,T),r._OrtReleaseTensor=s=>(r._OrtReleaseTensor=pe.Ba)(s),r._OrtCreateRunOptions=(s,m,g,$)=>(r._OrtCreateRunOptions=pe.Ca)(s,m,g,$),r._OrtAddRunConfigEntry=(s,m,g)=>(r._OrtAddRunConfigEntry=pe.Da)(s,m,g),r._OrtReleaseRunOptions=s=>(r._OrtReleaseRunOptions=pe.Ea)(s),r._OrtCreateBinding=s=>(r._OrtCreateBinding=pe.Fa)(s),r._OrtBindInput=(s,m,g)=>(r._OrtBindInput=pe.Ga)(s,m,g),r._OrtBindOutput=(s,m,g,$)=>(r._OrtBindOutput=pe.Ha)(s,m,g,$),r._OrtClearBoundOutputs=s=>(r._OrtClearBoundOutputs=pe.Ia)(s),r._OrtReleaseBinding=s=>(r._OrtReleaseBinding=pe.Ja)(s),r._OrtRunWithBinding=(s,m,g,$,T)=>(r._OrtRunWithBinding=pe.Ka)(s,m,g,$,T),r._OrtRun=(s,m,g,$,T,B,H,q)=>(r._OrtRun=pe.La)(s,m,g,$,T,B,H,q),r._OrtEndProfiling=s=>(r._OrtEndProfiling=pe.Ma)(s),r._JsepOutput=(s,m,g)=>(r._JsepOutput=pe.Na)(s,m,g),r._JsepGetNodeName=s=>(r._JsepGetNodeName=pe.Oa)(s);var Nt=r._malloc=s=>(Nt=r._malloc=pe.Pa)(s),rt=r._free=s=>(rt=r._free=pe.Qa)(s),yr=s=>(yr=pe.Sa)(s),br=s=>(br=pe.Ta)(s),Ur=()=>(Ur=pe.Ua)(),Jt=s=>(Jt=pe.Wa)(s),dn=()=>(dn=pe.Xa)(),vt=s=>(vt=pe.Ya)(s),ct=()=>(ct=pe.Za)();r.___start_em_js=837952,r.___stop_em_js=838113;function ln(){var s=pe;s=Object.assign({},s);var m=g=>$=>g($)>>>0;return s.la=m(s.la),s.Pa=m(s.Pa),s.Ta=m(s.Ta),s.Ua=(g=>()=>g()>>>0)(s.Ua),s}r.stackSave=()=>Ur(),r.stackRestore=s=>yr(s),r.stackAlloc=s=>br(s),r.UTF8ToString=je,r.stringToUTF8=(s,m,g)=>at(s,K,m,g),r.lengthBytesUTF8=Wt;var Dt;G=function s(){Dt||Vr(),Dt||(G=s)};function Vr(){if(!(0Ii)});var Pi=Wr(()=>{});var ki=Wr(()=>{});var Oi={};gn(Oi,{cpus:()=>ll});var ll,Ri=Y(()=>{ll=void 0});var zi=Wr((Di,mo)=>{"use strict";var Bi=(()=>{var e=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(e||=__filename),function(t={}){function r(){return ae.buffer!=ue.buffer&&Ne(),ue}function o(){return ae.buffer!=ue.buffer&&Ne(),le}function i(){return ae.buffer!=ue.buffer&&Ne(),qe}function u(){return ae.buffer!=ue.buffer&&Ne(),G}function a(){return ae.buffer!=ue.buffer&&Ne(),ne}function c(){return ae.buffer!=ue.buffer&&Ne(),xe}function p(){return ae.buffer!=ue.buffer&&Ne(),Ke}function h(){return ae.buffer!=ue.buffer&&Ne(),Ut}var d=t,y,w,_=new Promise((n,l)=>{y=n,w=l});d.mountExternalData=(n,l)=>{(d.Db||(d.Db=new Map)).set(n,l)},d.unmountExternalData=()=>{delete d.Db};let v=()=>{let n=(f,b,C)=>(...k)=>{let V=ct,F=b?.();k=f(...k);let re=b?.();return F!==re&&(f=re,C(F),b=C=null),ct!=V?T():k},l=f=>async(...b)=>{try{if(d.Cb)throw Error("Session already started");let C=d.Cb={ec:b[0],errors:[]},k=await f(...b);if(d.Cb!==C)throw Error("Session mismatch");d.Kb?.flush();let V=C.errors;if(0re),0d._OrtCreateSession,f=>d._OrtCreateSession=f),d._OrtRun=l(n(d._OrtRun,()=>d._OrtRun,f=>d._OrtRun=f)),d._OrtRunWithBinding=l(n(d._OrtRunWithBinding,()=>d._OrtRunWithBinding,f=>d._OrtRunWithBinding=f)),d._OrtBindInput=n(d._OrtBindInput,()=>d._OrtBindInput,f=>d._OrtBindInput=f),v=void 0};d.jsepInit=(n,l)=>{if(v?.(),n==="webgpu"){[d.Kb,d.Wb,d.$b,d.Lb,d.Zb,d.ob,d.ac,d.cc,d.Xb,d.Yb,d.bc]=l;let f=d.Kb;d.jsepRegisterBuffer=(b,C,k,V)=>f.registerBuffer(b,C,k,V),d.jsepGetBuffer=b=>f.getBuffer(b),d.jsepCreateDownloader=(b,C,k)=>f.createDownloader(b,C,k),d.jsepOnReleaseSession=b=>{f.onReleaseSession(b)},d.jsepOnRunStart=b=>f.onRunStart(b)}};var S=Object.assign({},d),A="./this.program",I=(n,l)=>{throw l},x=typeof window=="object",E=typeof importScripts=="function",P=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",O=d.ENVIRONMENT_IS_PTHREAD||!1,R="";function L(n){return d.locateFile?d.locateFile(n,R):R+n}var N,K,Q;if(P){var he=(uo(),wr(so)),W=(co(),wr(lo));R=E?W.dirname(R)+"/":__dirname+"/",N=(n,l)=>(n=Wt(n)?new URL(n):W.normalize(n),he.readFileSync(n,l?void 0:"utf8")),Q=n=>(n=N(n,!0),n.buffer||(n=new Uint8Array(n)),n),K=(n,l,f,b=!0)=>{n=Wt(n)?new URL(n):W.normalize(n),he.readFile(n,b?void 0:"utf8",(C,k)=>{C?f(C):l(b?k.buffer:k)})},!d.thisProgram&&1{throw process.exitCode=n,l},global.Worker=Pi().Worker}else(x||E)&&(E?R=self.location.href:typeof document<"u"&&document.currentScript&&(R=document.currentScript.src),typeof e<"u"&&e&&(R=e),R.startsWith("blob:")?R="":R=R.substr(0,R.replace(/[?#].*/,"").lastIndexOf("/")+1),P||(N=n=>{var l=new XMLHttpRequest;return l.open("GET",n,!1),l.send(null),l.responseText},E&&(Q=n=>{var l=new XMLHttpRequest;return l.open("GET",n,!1),l.responseType="arraybuffer",l.send(null),new Uint8Array(l.response)}),K=(n,l,f)=>{var b=new XMLHttpRequest;b.open("GET",n,!0),b.responseType="arraybuffer",b.onload=()=>{b.status==200||b.status==0&&b.response?l(b.response):f()},b.onerror=f,b.send(null)}));P&&typeof performance>"u"&&(global.performance=ki().performance);var se=console.log.bind(console),Ce=console.error.bind(console);P&&(se=(...n)=>he.writeSync(1,n.join(" ")+`\n`),Ce=(...n)=>he.writeSync(2,n.join(" ")+`\n`));var We=se,ee=Ce;Object.assign(d,S),S=null;var ae,Ae,me=!1,ie,ue,le,qe,G,ne,xe,Ke,Be,Ge,Ut;function Ne(){var n=ae.buffer;d.HEAP8=ue=new Int8Array(n),d.HEAP16=qe=new Int16Array(n),d.HEAPU8=le=new Uint8Array(n),d.HEAPU16=G=new Uint16Array(n),d.HEAP32=ne=new Int32Array(n),d.HEAPU32=xe=new Uint32Array(n),d.HEAPF32=Ke=new Float32Array(n),d.HEAPF64=Ut=new Float64Array(n),d.HEAP64=Be=new BigInt64Array(n),d.HEAPU64=Ge=new BigUint64Array(n)}var Ye=16777216;if(O)ae=d.wasmMemory;else if(d.wasmMemory)ae=d.wasmMemory;else if(ae=new WebAssembly.Memory({initial:Ye/65536,maximum:65536,shared:!0}),!(ae.buffer instanceof SharedArrayBuffer))throw ee("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"),P&&ee("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");Ne(),Ye=ae.buffer.byteLength;var mt=[],Rt=[],qt=[],Vt=0,sr=null,jt=null;function xr(){if(Vt--,Vt==0&&(sr!==null&&(clearInterval(sr),sr=null),jt)){var n=jt;jt=null,n()}}function Kt(n){throw n="Aborted("+n+")",ee(n),me=!0,ie=1,n=new WebAssembly.RuntimeError(n+". Build with -sASSERTIONS for more info."),w(n),n}var je=n=>n.startsWith("data:application/octet-stream;base64,"),Wt=n=>n.startsWith("file://"),at;at="ort-wasm-simd-threaded.wasm",je(at)||(at=L(at));function Cr(n){if(Q)return Q(n);throw"both async and sync fetching of the wasm failed"}function ft(n){if(x||E){if(typeof fetch=="function"&&!Wt(n))return fetch(n,{credentials:"same-origin"}).then(l=>{if(!l.ok)throw`failed to load wasm binary file at \'${n}\'`;return l.arrayBuffer()}).catch(()=>Cr(n));if(K)return new Promise((l,f)=>{K(n,b=>l(new Uint8Array(b)),f)})}return Promise.resolve().then(()=>Cr(n))}function ur(n,l,f){return ft(n).then(b=>WebAssembly.instantiate(b,l)).then(f,b=>{ee(`failed to asynchronously prepare wasm: ${b}`),Kt(b)})}function Ar(n,l){var f=at;return typeof WebAssembly.instantiateStreaming!="function"||je(f)||Wt(f)||P||typeof fetch!="function"?ur(f,n,l):fetch(f,{credentials:"same-origin"}).then(b=>WebAssembly.instantiateStreaming(b,n).then(l,function(C){return ee(`wasm streaming compile failed: ${C}`),ee("falling back to ArrayBuffer instantiation"),ur(f,n,l)}))}var Ir={826468:(n,l,f,b)=>{if(typeof d>"u"||!d.Db)return 1;if(n=He(n>>>0),n.startsWith("./")&&(n=n.substring(2)),n=d.Db.get(n),!n)return 2;if(l>>>=0,f>>>=0,b>>>=0,l+f>n.byteLength)return 3;try{return o().set(n.subarray(l,l+f),b>>>0),0}catch{return 4}},826969:()=>{d.Xb()},827e3:()=>{d.Yb()},827029:()=>{d.bc()},827054:n=>d.Wb(n),827087:n=>d.$b(n),827119:(n,l,f)=>{d.Lb(n,l,f,!0)},827158:(n,l,f)=>{d.Lb(n,l,f)},827191:n=>{d.ob("Abs",n,void 0)},827242:n=>{d.ob("Neg",n,void 0)},827293:n=>{d.ob("Floor",n,void 0)},827346:n=>{d.ob("Ceil",n,void 0)},827398:n=>{d.ob("Reciprocal",n,void 0)},827456:n=>{d.ob("Sqrt",n,void 0)},827508:n=>{d.ob("Exp",n,void 0)},827559:n=>{d.ob("Erf",n,void 0)},827610:n=>{d.ob("Sigmoid",n,void 0)},827665:(n,l,f)=>{d.ob("HardSigmoid",n,{alpha:l,beta:f})},827744:n=>{d.ob("Log",n,void 0)},827795:n=>{d.ob("Sin",n,void 0)},827846:n=>{d.ob("Cos",n,void 0)},827897:n=>{d.ob("Tan",n,void 0)},827948:n=>{d.ob("Asin",n,void 0)},828e3:n=>{d.ob("Acos",n,void 0)},828052:n=>{d.ob("Atan",n,void 0)},828104:n=>{d.ob("Sinh",n,void 0)},828156:n=>{d.ob("Cosh",n,void 0)},828208:n=>{d.ob("Asinh",n,void 0)},828261:n=>{d.ob("Acosh",n,void 0)},828314:n=>{d.ob("Atanh",n,void 0)},828367:n=>{d.ob("Tanh",n,void 0)},828419:n=>{d.ob("Not",n,void 0)},828470:(n,l,f)=>{d.ob("Clip",n,{min:l,max:f})},828539:n=>{d.ob("Clip",n,void 0)},828591:(n,l)=>{d.ob("Elu",n,{alpha:l})},828649:n=>{d.ob("Relu",n,void 0)},828701:(n,l)=>{d.ob("LeakyRelu",n,{alpha:l})},828765:(n,l)=>{d.ob("ThresholdedRelu",n,{alpha:l})},828835:(n,l)=>{d.ob("Cast",n,{to:l})},828893:n=>{d.ob("Add",n,void 0)},828944:n=>{d.ob("Sub",n,void 0)},828995:n=>{d.ob("Mul",n,void 0)},829046:n=>{d.ob("Div",n,void 0)},829097:n=>{d.ob("Pow",n,void 0)},829148:n=>{d.ob("Equal",n,void 0)},829201:n=>{d.ob("Greater",n,void 0)},829256:n=>{d.ob("GreaterOrEqual",n,void 0)},829318:n=>{d.ob("Less",n,void 0)},829370:n=>{d.ob("LessOrEqual",n,void 0)},829429:(n,l,f,b,C)=>{d.ob("ReduceMean",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829588:(n,l,f,b,C)=>{d.ob("ReduceMax",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829746:(n,l,f,b,C)=>{d.ob("ReduceMin",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829904:(n,l,f,b,C)=>{d.ob("ReduceProd",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830063:(n,l,f,b,C)=>{d.ob("ReduceSum",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830221:(n,l,f,b,C)=>{d.ob("ReduceL1",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830378:(n,l,f,b,C)=>{d.ob("ReduceL2",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830535:(n,l,f,b,C)=>{d.ob("ReduceLogSum",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830696:(n,l,f,b,C)=>{d.ob("ReduceSumSquare",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830860:(n,l,f,b,C)=>{d.ob("ReduceLogSumExp",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},831024:n=>{d.ob("Where",n,void 0)},831077:(n,l,f)=>{d.ob("Transpose",n,{perm:l?Array.from(a().subarray(l>>>0,f>>>0)):[]})},831185:(n,l,f,b)=>{d.ob("DepthToSpace",n,{blocksize:l,mode:He(f),format:b?"NHWC":"NCHW"})},831318:(n,l,f,b)=>{d.ob("DepthToSpace",n,{blocksize:l,mode:He(f),format:b?"NHWC":"NCHW"})},831451:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we)=>{d.ob("ConvTranspose",n,{format:re?"NHWC":"NCHW",autoPad:l,dilations:[f],group:b,kernelShape:[C],pads:[k,V],strides:[F],wIsConst:()=>!!r()[J>>>0],outputPadding:ce?Array.from(a().subarray(ce>>>0,Re>>>0)):[],outputShape:ze?Array.from(a().subarray(ze>>>0,z>>>0)):[],activation:He(we)})},831852:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z)=>{d.ob("ConvTranspose",n,{format:F?"NHWC":"NCHW",autoPad:l,dilations:Array.from(a().subarray(f>>>0,(f>>>0)+2>>>0)),group:b,kernelShape:Array.from(a().subarray(C>>>0,(C>>>0)+2>>>0)),pads:Array.from(a().subarray(k>>>0,(k>>>0)+4>>>0)),strides:Array.from(a().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!r()[re>>>0],outputPadding:J?Array.from(a().subarray(J>>>0,ce>>>0)):[],outputShape:Re?Array.from(a().subarray(Re>>>0,ze>>>0)):[],activation:He(z)})},832417:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we)=>{d.ob("ConvTranspose",n,{format:re?"NHWC":"NCHW",autoPad:l,dilations:[f],group:b,kernelShape:[C],pads:[k,V],strides:[F],wIsConst:()=>!!r()[J>>>0],outputPadding:ce?Array.from(a().subarray(ce>>>0,Re>>>0)):[],outputShape:ze?Array.from(a().subarray(ze>>>0,z>>>0)):[],activation:He(we)})},832818:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z)=>{d.ob("ConvTranspose",n,{format:F?"NHWC":"NCHW",autoPad:l,dilations:Array.from(a().subarray(f>>>0,(f>>>0)+2>>>0)),group:b,kernelShape:Array.from(a().subarray(C>>>0,(C>>>0)+2>>>0)),pads:Array.from(a().subarray(k>>>0,(k>>>0)+4>>>0)),strides:Array.from(a().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!r()[re>>>0],outputPadding:J?Array.from(a().subarray(J>>>0,ce>>>0)):[],outputShape:Re?Array.from(a().subarray(Re>>>0,ze>>>0)):[],activation:He(z)})},833383:(n,l)=>{d.ob("GlobalAveragePool",n,{format:l?"NHWC":"NCHW"})},833474:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("AveragePool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},833758:(n,l)=>{d.ob("GlobalAveragePool",n,{format:l?"NHWC":"NCHW"})},833849:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("AveragePool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834133:(n,l)=>{d.ob("GlobalMaxPool",n,{format:l?"NHWC":"NCHW"})},834220:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("MaxPool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834500:(n,l)=>{d.ob("GlobalMaxPool",n,{format:l?"NHWC":"NCHW"})},834587:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("MaxPool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834867:(n,l,f,b,C)=>{d.ob("Gemm",n,{alpha:l,beta:f,transA:b,transB:C})},834971:n=>{d.ob("MatMul",n,void 0)},835025:(n,l,f,b)=>{d.ob("ArgMax",n,{keepDims:!!l,selectLastIndex:!!f,axis:b})},835133:(n,l,f,b)=>{d.ob("ArgMin",n,{keepDims:!!l,selectLastIndex:!!f,axis:b})},835241:(n,l)=>{d.ob("Softmax",n,{axis:l})},835304:(n,l)=>{d.ob("Concat",n,{axis:l})},835364:(n,l,f,b,C)=>{d.ob("Split",n,{axis:l,numOutputs:f,splitSizes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},835504:n=>{d.ob("Expand",n,void 0)},835558:(n,l)=>{d.ob("Gather",n,{axis:Number(l)})},835629:(n,l)=>{d.ob("GatherElements",n,{axis:Number(l)})},835708:(n,l,f,b,C,k,V,F,re,J,ce)=>{d.ob("Resize",n,{antialias:l,axes:f?Array.from(a().subarray(f>>>0,b>>>0)):[],coordinateTransformMode:He(C),cubicCoeffA:k,excludeOutside:V,extrapolationValue:F,keepAspectRatioPolicy:He(re),mode:He(J),nearestMode:He(ce)})},836054:(n,l,f,b,C,k,V)=>{d.ob("Slice",n,{starts:l?Array.from(a().subarray(l>>>0,f>>>0)):[],ends:b?Array.from(a().subarray(b>>>0,C>>>0)):[],axes:k?Array.from(a().subarray(k>>>0,V>>>0)):[]})},836270:n=>{d.ob("Tile",n,void 0)},836322:(n,l,f,b)=>{d.ob("LayerNormalization",n,{axis:l,epsilon:f,simplified:!!b})},836433:(n,l,f)=>{d.ob("InstanceNormalization",n,{epsilon:l,format:f?"NHWC":"NCHW"})},836547:(n,l,f)=>{d.ob("InstanceNormalization",n,{epsilon:l,format:f?"NHWC":"NCHW"})},836661:n=>{d.ob("Range",n,void 0)},836714:(n,l)=>{d.ob("Einsum",n,{equation:He(l)})},836795:(n,l,f,b,C)=>{d.ob("Pad",n,{mode:l,value:f,pads:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},836922:(n,l,f,b,C,k)=>{d.ob("BatchNormalization",n,{epsilon:l,momentum:f,spatial:!!C,trainingMode:!!b,format:k?"NHWC":"NCHW"})},837091:(n,l,f,b,C,k)=>{d.ob("BatchNormalization",n,{epsilon:l,momentum:f,spatial:!!C,trainingMode:!!b,format:k?"NHWC":"NCHW"})},837260:(n,l,f)=>{d.ob("CumSum",n,{exclusive:Number(l),reverse:Number(f)})},837357:(n,l,f,b,C,k,V,F,re)=>{d.ob("Attention",n,{numHeads:l,isUnidirectional:f,maskFilterValue:b,scale:C,doRotary:k,qkvHiddenSizes:V?Array.from(a().subarray(Number(F)>>>0,Number(F)+V>>>0)):[],pastPresentShareBuffer:!!re})},837629:n=>{d.ob("BiasAdd",n,void 0)},837684:n=>{d.ob("BiasSplitGelu",n,void 0)},837745:n=>{d.ob("FastGelu",n,void 0)},837801:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze)=>{d.ob("Conv",n,{format:re?"NHWC":"NCHW",auto_pad:l,dilations:[f],group:b,kernel_shape:[C],pads:k?Array.from(a().subarray(k>>>0,V>>>0)):[],strides:[F],w_is_const:()=>!!r()[J>>>0],activation:He(ce),activation_params:Re?Array.from(p().subarray(Re>>>0,ze>>>0)):[]})},838171:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("Conv",n,{format:Re?"NHWC":"NCHW",auto_pad:l,dilations:[f,b],group:C,kernel_shape:[k,V],pads:F?Array.from(a().subarray(F>>>0,re>>>0)):[],strides:[J,ce],w_is_const:()=>!!r()[ze>>>0],activation:He(z),activation_params:we?Array.from(p().subarray(we>>>0,Pe>>>0)):[]})},838562:n=>{d.ob("Gelu",n,void 0)},838614:(n,l,f,b,C,k)=>{d.ob("MatMulNBits",n,{k:l,n:f,accuracyLevel:b,bits:C,blockSize:k})},838741:(n,l,f,b,C,k)=>{d.ob("MultiHeadAttention",n,{numHeads:l,isUnidirectional:f,maskFilterValue:b,scale:C,doRotary:k})},838900:(n,l,f,b,C)=>{d.ob("RotaryEmbedding",n,{interleaved:!!l,numHeads:f,rotaryEmbeddingDim:b,scale:C})},839039:(n,l,f)=>{d.ob("SkipLayerNormalization",n,{epsilon:l,simplified:!!f})},839141:(n,l,f)=>{d.ob("SkipLayerNormalization",n,{epsilon:l,simplified:!!f})},839243:(n,l,f,b)=>{d.ob("LayerNormalization",n,{axis:l,epsilon:f,simplified:!!b})},839354:n=>{d.ac(n)},839388:(n,l)=>d.cc(n,l,d.Cb.ec,d.Cb.errors)};function st(n){this.name="ExitStatus",this.message=`Program terminated with exit(${n})`,this.status=n}var Tr=n=>{n.terminate(),n.onmessage=()=>{}},ht=n=>{ge.xb.length==0&&(lr(),ge.Mb(ge.xb[0]));var l=ge.xb.pop();if(!l)return 6;ge.yb.push(l),ge.ub[n.wb]=l,l.wb=n.wb;var f={cmd:"run",start_routine:n.fc,arg:n.Rb,pthread_ptr:n.wb};return P&&l.unref(),l.postMessage(f,n.mc),0},bt=0,Er=n=>{var l=io();return n=n(),fn(l),n},Te=(n,l,...f)=>Er(()=>{for(var b=2*f.length,C=oo(8*b),k=C>>>3,V=0;V>>0]=F)}return gi(n,0,b,C,l)});function dr(n){if(O)return Te(0,1,n);ie=n,0{if(ie=n,O)throw jr(n),"unwind";dr(n)};function ut(){for(var n=d.numThreads;n--;)lr();mt.unshift(()=>{Vt++,Fn(()=>xr())})}function lr(){var n=L("ort-wasm-simd-threaded.worker.js");n=new Worker(n),ge.xb.push(n)}function Fn(n){O?n():Promise.all(ge.xb.map(ge.Mb)).then(n)}var ge={xb:[],yb:[],Qb:[],ub:{},Gb(){O?(ge.receiveObjectTransfer=ge.dc,ge.threadInitTLS=ge.Pb,ge.setExitStatus=ge.Ob):ut()},Ob:n=>ie=n,pc:["$terminateWorker"],hc:()=>{for(var n of ge.yb)Tr(n);for(n of ge.xb)Tr(n);ge.xb=[],ge.yb=[],ge.ub=[]},Nb:n=>{var l=n.wb;delete ge.ub[l],ge.xb.push(n),ge.yb.splice(ge.yb.indexOf(n),1),n.wb=0,ro(l)},dc(){},Pb(){ge.Qb.forEach(n=>n())},Mb:n=>new Promise(l=>{n.onmessage=k=>{k=k.data;var V=k.cmd;if(k.targetThread&&k.targetThread!=pn()){var F=ge.ub[k.targetThread];F?F.postMessage(k,k.transferList):ee(`Internal error! Worker sent a message "${V}" to target pthread ${k.targetThread}, but that thread no longer exists!`)}else V==="checkMailbox"?rt():V==="spawnThread"?ht(k):V==="cleanupThread"?ge.Nb(ge.ub[k.thread]):V==="killThread"?(k=k.thread,V=ge.ub[k],delete ge.ub[k],Tr(V),ro(k),ge.yb.splice(ge.yb.indexOf(V),1),V.wb=0):V==="cancelThread"?ge.ub[k.thread].postMessage({cmd:"cancel"}):V==="loaded"?(n.loaded=!0,P&&!n.wb&&n.unref(),l(n)):V==="alert"?alert(`Thread ${k.threadId}: ${k.text}`):k.target==="setimmediate"?n.postMessage(k):V==="callHandler"?d[k.handler](...k.args):V&&ee(`worker sent an unknown command ${V}`)},n.onerror=k=>{throw ee(`worker sent an error! ${k.filename}:${k.lineno}: ${k.message}`),k},P&&(n.on("message",k=>n.onmessage({data:k})),n.on("error",k=>n.onerror(k)));var f=[],b=["onExit"],C;for(C of b)d.hasOwnProperty(C)&&f.push(C);n.postMessage({cmd:"load",handlers:f,urlOrBlob:d.mainScriptUrlOrBlob||e,wasmMemory:ae,wasmModule:Ae})})};d.PThread=ge;var Yt=n=>{for(;0{var n=pn(),l=c()[n+52>>>2>>>0];n=c()[n+56>>>2>>>0],bi(l,l-n),fn(l)};function jr(n){if(O)return Te(1,0,n);Xe(n)}d.invokeEntryPoint=(n,l)=>{bt=0,n=wi(n,l),0>>2>>>0]=0,c()[this.Ib+4>>>2>>>0]=l,c()[this.Ib+8>>>2>>>0]=f}}var Kr=0,jn=0;function Yr(n,l,f,b){return O?Te(2,1,n,l,f,b):Zr(n,l,f,b)}function Zr(n,l,f,b){if(n>>>=0,l>>>=0,f>>>=0,b>>>=0,typeof SharedArrayBuffer>"u")return ee("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var C=[];return O&&C.length===0?Yr(n,l,f,b):(n={fc:f,wb:n,Rb:b,mc:C},O?(n.oc="spawnThread",postMessage(n,C),0):ht(n))}var cr=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Pr=(n,l,f)=>{l>>>=0;var b=l+f;for(f=l;n[f]&&!(f>=b);)++f;if(16C?b+=String.fromCharCode(C):(C-=65536,b+=String.fromCharCode(55296|C>>10,56320|C&1023))}}else b+=String.fromCharCode(C)}return b},He=(n,l)=>(n>>>=0)?Pr(o(),n,l):"";function Xr(n,l,f){return O?Te(3,1,n,l,f):0}function Qr(n,l){if(O)return Te(4,1,n,l)}var gt=n=>{for(var l=0,f=0;f=b?l++:2047>=b?l+=2:55296<=b&&57343>=b?(l+=4,++f):l+=3}return l},dt=(n,l,f,b)=>{if(f>>>=0,!(0=V){var F=n.charCodeAt(++k);V=65536+((V&1023)<<10)|F&1023}if(127>=V){if(f>=b)break;l[f++>>>0]=V}else{if(2047>=V){if(f+1>=b)break;l[f++>>>0]=192|V>>6}else{if(65535>=V){if(f+2>=b)break;l[f++>>>0]=224|V>>12}else{if(f+3>=b)break;l[f++>>>0]=240|V>>18,l[f++>>>0]=128|V>>12&63}l[f++>>>0]=128|V>>6&63}l[f++>>>0]=128|V&63}}return l[f>>>0]=0,f-C},Bt=(n,l,f)=>dt(n,o(),l,f);function Zt(n,l){if(O)return Te(5,1,n,l)}function kr(n,l,f){if(O)return Te(6,1,n,l,f)}function Or(n,l,f){return O?Te(7,1,n,l,f):0}function Jr(n,l){if(O)return Te(8,1,n,l)}function pr(n,l,f){if(O)return Te(9,1,n,l,f)}function en(n,l,f,b){if(O)return Te(10,1,n,l,f,b)}function tn(n,l,f,b){if(O)return Te(11,1,n,l,f,b)}function rn(n,l,f,b){if(O)return Te(12,1,n,l,f,b)}function nn(n){if(O)return Te(13,1,n)}function Rr(n,l){if(O)return Te(14,1,n,l)}function Xt(n,l,f){if(O)return Te(15,1,n,l,f)}var on,nt=n=>{for(var l="";o()[n>>>0];)l+=on[o()[n++>>>0]];return l},mr={},Br={},Kn={},wt;function Yn(n,l,f={}){var b=l.name;if(!n)throw new wt(`type "${b}" must have a positive integer typeid pointer`);if(Br.hasOwnProperty(n)){if(f.Ub)return;throw new wt(`Cannot register type \'${b}\' twice`)}Br[n]=l,delete Kn[n],mr.hasOwnProperty(n)&&(l=mr[n],delete mr[n],l.forEach(C=>C()))}function Je(n,l,f={}){if(!("argPackAdvance"in l))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Yn(n,l,f)}var Dr=(n,l,f)=>{switch(l){case 1:return f?b=>r()[b>>>0]:b=>o()[b>>>0];case 2:return f?b=>i()[b>>>1>>>0]:b=>u()[b>>>1>>>0];case 4:return f?b=>a()[b>>>2>>>0]:b=>c()[b>>>2>>>0];case 8:return f?b=>Be[b>>>3]:b=>Ge[b>>>3];default:throw new TypeError(`invalid integer width (${l}): ${n}`)}},fr=[],lt=[];function hr(n){n>>>=0,9{if(!n)throw new wt("Cannot use deleted val. handle = "+n);return lt[n]},ot=n=>{switch(n){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let l=fr.pop()||lt.length;return lt[l]=n,lt[l+1]=1,l}};function Qt(n){return this.fromWireType(c()[n>>>2>>>0])}var Zn={name:"emscripten::val",fromWireType:n=>{var l=Qe(n);return hr(n),l},toWireType:(n,l)=>ot(l),argPackAdvance:8,readValueFromPointer:Qt,Bb:null},an=(n,l)=>{switch(l){case 4:return function(f){return this.fromWireType(p()[f>>>2>>>0])};case 8:return function(f){return this.fromWireType(h()[f>>>3>>>0])};default:throw new TypeError(`invalid float width (${l}): ${n}`)}},zr=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,Xn=(n,l)=>{for(var f=n>>1,b=f+l/2;!(f>=b)&&u()[f>>>0];)++f;if(f<<=1,32=l/2);++b){var C=i()[n+2*b>>>1>>>0];if(C==0)break;f+=String.fromCharCode(C)}return f},sn=(n,l,f)=>{if(f??=2147483647,2>f)return 0;f-=2;var b=l;f=f<2*n.length?f/2:n.length;for(var C=0;C>>1>>>0]=k,l+=2}return i()[l>>>1>>>0]=0,l-b},un=n=>2*n.length,gr=(n,l)=>{for(var f=0,b="";!(f>=l/4);){var C=a()[n+4*f>>>2>>>0];if(C==0)break;++f,65536<=C?(C-=65536,b+=String.fromCharCode(55296|C>>10,56320|C&1023)):b+=String.fromCharCode(C)}return b},Qn=(n,l,f)=>{if(l>>>=0,f??=2147483647,4>f)return 0;var b=l;f=b+f-4;for(var C=0;C=k){var V=n.charCodeAt(++C);k=65536+((k&1023)<<10)|V&1023}if(a()[l>>>2>>>0]=k,l+=4,l+4>f)break}return a()[l>>>2>>>0]=0,l-b},pe=n=>{for(var l=0,f=0;f=b&&++f,l+=4}return l},Mr=n=>{if(!me)try{if(n(),!(0>>=0,typeof Atomics.nc=="function"&&(Atomics.nc(a(),n>>>2,n).value.then(rt),n+=128,Atomics.store(a(),n>>>2,1))}d.__emscripten_thread_mailbox_await=Nt;var rt=()=>{var n=pn();n&&(Nt(n),Mr(yi))};d.checkMailbox=rt;var yr=[],br=(n,l)=>{var f=Br[n];if(f===void 0)throw n=fi(n),f=nt(n),At(n),new wt(`${l} has unknown type ${f}`);return f},Ur=(n,l,f)=>{var b=[];return n=n.toWireType(b,f),b.length&&(c()[l>>>2>>>0]=ot(b)),n},Jt=n=>{try{n()}catch(l){Kt(l)}};function dn(){var n=oe,l={};for(let[f,b]of Object.entries(n))l[f]=typeof b=="function"?(...C)=>{Dt.push(f);try{return b(...C)}finally{me||(Dt.pop(),ct&&vt===1&&Dt.length===0&&(vt=0,bt+=1,Jt($i),typeof Fibers<"u"&&Fibers.rc()))}}:b;return l}var vt=0,ct=null,ln=0,Dt=[],Vr={},s={},m=0,g=null,$=[];function T(){return new Promise((n,l)=>{g={resolve:n,reject:l}})}function B(){var n=mn(65548),l=n+12;c()[n>>>2>>>0]=l,c()[n+4>>>2>>>0]=l+65536,l=Dt[0];var f=Vr[l];return f===void 0&&(f=m++,Vr[l]=f,s[f]=l),l=f,a()[n+8>>>2>>>0]=l,n}function H(){var n=a()[ct+8>>>2>>>0];return n=oe[s[n]],--bt,n()}function q(n){if(!me){if(vt===0){var l=!1,f=!1;n((b=0)=>{if(!me&&(ln=b,l=!0,f)){vt=2,Jt(()=>_i(ct)),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.resume(),b=!1;try{var C=H()}catch(F){C=F,b=!0}var k=!1;if(!ct){var V=g;V&&(g=null,(b?V.reject:V.resolve)(C),k=!0)}if(b&&!k)throw C}}),f=!0,l||(vt=1,ct=B(),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.pause(),Jt(()=>vi(ct)))}else vt===2?(vt=0,Jt(Si),At(ct),ct=null,$.forEach(Mr)):Kt(`invalid state: ${vt}`);return ln}}function te(n){return q(l=>{n().then(l)})}var X=[],de={},Ee=n=>{var l=de[n];return l===void 0?nt(n):l},Oe=()=>typeof globalThis=="object"?globalThis:Function("return this")(),D=n=>{var l=X.length;return X.push(n),l},be=(n,l)=>{for(var f=Array(n),b=0;b>>2>>>0],"parameter "+b);return f},Ie=(n,l)=>Object.defineProperty(l,"name",{value:n});function zt(n){var l=Function;if(!(l instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof l} which is not a function`);var f=Ie(l.name||"unknownFunctionName",function(){});return f.prototype=l.prototype,f=new f,n=l.apply(f,n),n instanceof Object?n:f}var Mt=n=>n%4===0&&(n%100!==0||n%400===0),Qo=[0,31,60,91,121,152,182,213,244,274,305,335],Jo=[0,31,59,90,120,151,181,212,243,273,304,334];function ei(n,l,f,b,C,k,V){return O?Te(16,1,n,l,f,b,C,k,V):-52}function ti(n,l,f,b,C,k){if(O)return Te(17,1,n,l,f,b,C,k)}var Jn=[],ri=(n,l)=>{Jn.length=0;for(var f;f=o()[n++>>>0];){var b=f!=105;b&=f!=112,l+=b&&l%8?4:0,Jn.push(f==112?c()[l>>>2>>>0]:f==106?Be[l>>>3]:f==105?a()[l>>>2>>>0]:h()[l>>>3>>>0]),l+=b?8:4}return Jn},eo={},ni=()=>{if(!to){var n={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:A||"./this.program"},l;for(l in eo)eo[l]===void 0?delete n[l]:n[l]=eo[l];var f=[];for(l in n)f.push(`${l}=${n[l]}`);to=f}return to},to;function oi(n,l){if(O)return Te(18,1,n,l);n>>>=0,l>>>=0;var f=0;return ni().forEach((b,C)=>{var k=l+f;for(C=c()[n+4*C>>>2>>>0]=k,k=0;k>>0]=b.charCodeAt(k);r()[C>>>0]=0,f+=b.length+1}),0}function ii(n,l){if(O)return Te(19,1,n,l);n>>>=0,l>>>=0;var f=ni();c()[n>>>2>>>0]=f.length;var b=0;return f.forEach(C=>b+=C.length+1),c()[l>>>2>>>0]=b,0}function ai(n){return O?Te(20,1,n):52}function si(n,l,f,b){return O?Te(21,1,n,l,f,b):52}function ui(n,l,f,b){return O?Te(22,1,n,l,f,b):70}var Xd=[null,[],[]];function di(n,l,f,b){if(O)return Te(23,1,n,l,f,b);l>>>=0,f>>>=0,b>>>=0;for(var C=0,k=0;k>>2>>>0],F=c()[l+4>>>2>>>0];l+=8;for(var re=0;re>>0],ce=Xd[n];J===0||J===10?((n===1?We:ee)(Pr(ce,0)),ce.length=0):ce.push(J)}C+=F}return c()[b>>>2>>>0]=C,0}var li=[31,29,31,30,31,30,31,31,30,31,30,31],ci=[31,28,31,30,31,30,31,31,30,31,30,31];function Qd(n){var l=Array(gt(n)+1);return dt(n,l,0,l.length),l}var Jd=(n,l)=>{r().set(n,l>>>0)};function pi(n,l,f,b){function C(z,we,Pe){for(z=typeof z=="number"?z.toString():z||"";z.lengthCi?-1:0er-z.getDate())we-=er-z.getDate()+1,z.setDate(1),11>Pe?z.setMonth(Pe+1):(z.setMonth(0),z.setFullYear(z.getFullYear()+1));else{z.setDate(z.getDate()+we);break}}return Pe=new Date(z.getFullYear()+1,0,4),we=F(new Date(z.getFullYear(),0,4)),Pe=F(Pe),0>=V(we,z)?0>=V(Pe,z)?z.getFullYear()+1:z.getFullYear():z.getFullYear()-1}n>>>=0,l>>>=0,f>>>=0,b>>>=0;var J=c()[b+40>>>2>>>0];b={kc:a()[b>>>2>>>0],jc:a()[b+4>>>2>>>0],Eb:a()[b+8>>>2>>>0],Jb:a()[b+12>>>2>>>0],Fb:a()[b+16>>>2>>>0],Ab:a()[b+20>>>2>>>0],vb:a()[b+24>>>2>>>0],zb:a()[b+28>>>2>>>0],qc:a()[b+32>>>2>>>0],ic:a()[b+36>>>2>>>0],lc:J?He(J):""},f=He(f),J={"%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"};for(var ce in J)f=f.replace(new RegExp(ce,"g"),J[ce]);var Re="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),ze="January February March April May June July August September October November December".split(" ");J={"%a":z=>Re[z.vb].substring(0,3),"%A":z=>Re[z.vb],"%b":z=>ze[z.Fb].substring(0,3),"%B":z=>ze[z.Fb],"%C":z=>k((z.Ab+1900)/100|0,2),"%d":z=>k(z.Jb,2),"%e":z=>C(z.Jb,2," "),"%g":z=>re(z).toString().substring(2),"%G":re,"%H":z=>k(z.Eb,2),"%I":z=>(z=z.Eb,z==0?z=12:12{for(var we=0,Pe=0;Pe<=z.Fb-1;we+=(Mt(z.Ab+1900)?li:ci)[Pe++]);return k(z.Jb+we,3)},"%m":z=>k(z.Fb+1,2),"%M":z=>k(z.jc,2),"%n":()=>`\n`,"%p":z=>0<=z.Eb&&12>z.Eb?"AM":"PM","%S":z=>k(z.kc,2),"%t":()=>" ","%u":z=>z.vb||7,"%U":z=>k(Math.floor((z.zb+7-z.vb)/7),2),"%V":z=>{var we=Math.floor((z.zb+7-(z.vb+6)%7)/7);if(2>=(z.vb+371-z.zb-2)%7&&we++,we)we==53&&(Pe=(z.vb+371-z.zb)%7,Pe==4||Pe==3&&Mt(z.Ab)||(we=1));else{we=52;var Pe=(z.vb+7-z.zb-1)%7;(Pe==4||Pe==5&&Mt(z.Ab%400-1))&&we++}return k(we,2)},"%w":z=>z.vb,"%W":z=>k(Math.floor((z.zb+7-(z.vb+6)%7)/7),2),"%y":z=>(z.Ab+1900).toString().substring(2),"%Y":z=>z.Ab+1900,"%z":z=>{z=z.ic;var we=0<=z;return z=Math.abs(z)/60,(we?"+":"-")+("0000"+(z/60*100+z%60)).slice(-4)},"%Z":z=>z.lc,"%%":()=>"%"},f=f.replace(/%%/g,"\\0\\0");for(ce in J)f.includes(ce)&&(f=f.replace(new RegExp(ce,"g"),J[ce](b)));return f=f.replace(/\\0\\0/g,"%"),ce=Qd(f),ce.length>l?0:(Jd(ce,n),ce.length-1)}ge.Gb();for(var mi=Array(256),cn=0;256>cn;++cn)mi[cn]=String.fromCharCode(cn);on=mi,wt=d.BindingError=class extends Error{constructor(n){super(n),this.name="BindingError"}},d.InternalError=class extends Error{constructor(n){super(n),this.name="InternalError"}},lt.push(0,1,void 0,1,null,1,!0,1,!1,1),d.count_emval_handles=()=>lt.length/2-5-fr.length;var el=[dr,jr,Yr,Xr,Qr,Zt,kr,Or,Jr,pr,en,tn,rn,nn,Rr,Xt,ei,ti,oi,ii,ai,si,ui,di],tl={ua:function(n,l,f){return te(async()=>{await d.Zb(n,l,f)})},b:function(n,l,f){throw n>>>=0,new qn(n).Gb(l>>>0,f>>>0),Kr=n,jn++,Kr},ia:function(n){hi(n>>>0,!E,1,!x,131072,!1),ge.Pb()},G:function(n){n>>>=0,O?postMessage({cmd:"cleanupThread",thread:n}):ge.Nb(ge.ub[n])},_:Zr,A:Xr,pa:Qr,W:Zt,Y:kr,qa:Or,na:Jr,fa:pr,ma:en,K:tn,X:rn,U:nn,oa:Rr,V:Xt,E:function(n,l,f){n>>>=0,f>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:b=>b,toWireType:function(b,C){if(typeof C!="bigint"&&typeof C!="number")throw C===null?C="null":(b=typeof C,C=b==="object"||b==="array"||b==="function"?C.toString():""+C),new TypeError(`Cannot convert "${C}" to ${this.name}`);return typeof C=="number"&&(C=BigInt(C)),C},argPackAdvance:8,readValueFromPointer:Dr(l,f,l.indexOf("u")==-1),Bb:null})},O:function(n,l,f,b){n>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:function(C){return!!C},toWireType:function(C,k){return k?f:b},argPackAdvance:8,readValueFromPointer:function(C){return this.fromWireType(o()[C>>>0])},Bb:null})},N:function(n){return Je(n>>>0,Zn)},D:function(n,l,f){n>>>=0,f>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:b=>b,toWireType:(b,C)=>C,argPackAdvance:8,readValueFromPointer:an(l,f),Bb:null})},t:function(n,l,f,b,C){if(n>>>=0,f>>>=0,l=nt(l>>>0),C===-1&&(C=4294967295),C=F=>F,b===0){var k=32-8*f;C=F=>F<>>k}var V=l.includes("unsigned")?function(F,re){return re>>>0}:function(F,re){return re};Je(n,{name:l,fromWireType:C,toWireType:V,argPackAdvance:8,readValueFromPointer:Dr(l,f,b!==0),Bb:null})},p:function(n,l,f){function b(k){var V=c()[k>>>2>>>0];return k=c()[k+4>>>2>>>0],new C(r().buffer,k,V)}n>>>=0;var C=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][l];f=nt(f>>>0),Je(n,{name:f,fromWireType:b,argPackAdvance:8,readValueFromPointer:b},{Ub:!0})},F:function(n,l){n>>>=0,l=nt(l>>>0);var f=l==="std::string";Je(n,{name:l,fromWireType:function(b){var C=c()[b>>>2>>>0],k=b+4;if(f)for(var V=k,F=0;F<=C;++F){var re=k+F;if(F==C||o()[re>>>0]==0){if(V=He(V,re-V),J===void 0)var J=V;else J+=String.fromCharCode(0),J+=V;V=re+1}}else{for(J=Array(C),F=0;F>>0]);J=J.join("")}return At(b),J},toWireType:function(b,C){C instanceof ArrayBuffer&&(C=new Uint8Array(C));var k=typeof C=="string";if(!(k||C instanceof Uint8Array||C instanceof Uint8ClampedArray||C instanceof Int8Array))throw new wt("Cannot pass non-string to std::string");var V=f&&k?gt(C):C.length,F=mn(4+V+1),re=F+4;if(c()[F>>>2>>>0]=V,f&&k)Bt(C,re,V+1);else if(k)for(k=0;k>>0]=J}else for(k=0;k>>0]=C[k];return b!==null&&b.push(At,F),F},argPackAdvance:8,readValueFromPointer:Qt,Bb(b){At(b)}})},z:function(n,l,f){if(n>>>=0,l>>>=0,f>>>=0,f=nt(f),l===2)var b=Xn,C=sn,k=un,V=F=>u()[F>>>1>>>0];else l===4&&(b=gr,C=Qn,k=pe,V=F=>c()[F>>>2>>>0]);Je(n,{name:f,fromWireType:F=>{for(var re=c()[F>>>2>>>0],J,ce=F+4,Re=0;Re<=re;++Re){var ze=F+4+Re*l;(Re==re||V(ze)==0)&&(ce=b(ce,ze-ce),J===void 0?J=ce:(J+=String.fromCharCode(0),J+=ce),ce=ze+l)}return At(F),J},toWireType:(F,re)=>{if(typeof re!="string")throw new wt(`Cannot pass non-string to C++ string type ${f}`);var J=k(re),ce=mn(4+J+l);return c()[ce>>>2>>>0]=J/l,C(re,ce+4,J+l),F!==null&&F.push(At,ce),ce},argPackAdvance:8,readValueFromPointer:Qt,Bb(F){At(F)}})},P:function(n,l){n>>>=0,l=nt(l>>>0),Je(n,{Vb:!0,name:l,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ta:()=>1,S:function(n,l){n>>>=0,n==l>>>0?setTimeout(rt):O?postMessage({targetThread:n,cmd:"checkMailbox"}):(n=ge.ub[n])&&n.postMessage({cmd:"checkMailbox"})},$:function(n,l,f,b,C){l>>>=0,f>>>=0,b/=2,yr.length=b,C=C>>>0>>>3;for(var k=0;k>>0];return n=l?Ir[l]:el[n],ge.Sb=f,f=n(...yr),ge.Sb=0,f},ha:Nt,sa:function(n){P&&ge.ub[n>>>0].ref()},w:function(n,l,f){return l>>>=0,f>>>=0,n=Qe(n>>>0),l=br(l,"emval::as"),Ur(l,f,n)},y:function(n){return n>>>=0,te(()=>(n=Qe(n),n.then(ot)))},o:function(n,l,f,b){return f>>>=0,b>>>=0,n=X[n>>>0],l=Qe(l>>>0),n(null,l,f,b)},k:function(n,l,f,b,C){return f>>>=0,b>>>=0,C>>>=0,n=X[n>>>0],l=Qe(l>>>0),f=Ee(f),n(l,l[f],b,C)},c:hr,C:function(n,l){return l>>>=0,n=Qe(n>>>0),l=Qe(l),n==l},n:function(n){return n>>>=0,n===0?ot(Oe()):(n=Ee(n),ot(Oe()[n]))},j:function(n,l,f){l=be(n,l>>>0);var b=l.shift();n--;var C=`return function (obj, func, destructorsRef, args) {\n`,k=0,V=[];f===0&&V.push("obj");for(var F=["retType"],re=[b],J=0;Jce.name).join(", ")}) => ${b.name}>`,D(Ie(f,n))},s:function(n,l){return l>>>=0,n=Qe(n>>>0),l=Qe(l),ot(n[l])},d:function(n){n>>>=0,9>>0);for(var l=Array(n.length),f=0;f>>0))},l:function(){return ot({})},i:function(n){n>>>=0;for(var l=Qe(n);l.length;){var f=l.pop();l.pop()(f)}hr(n)},h:function(n,l,f){l>>>=0,f>>>=0,n=Qe(n>>>0),l=Qe(l),f=Qe(f),n[l]=f},e:function(n,l){return l>>>=0,n=br(n>>>0,"_emval_take_value"),n=n.readValueFromPointer(l),ot(n)},ca:function(n,l){n=-9007199254740992>n||9007199254740992>>=0,n=new Date(1e3*n),a()[l>>>2>>>0]=n.getUTCSeconds(),a()[l+4>>>2>>>0]=n.getUTCMinutes(),a()[l+8>>>2>>>0]=n.getUTCHours(),a()[l+12>>>2>>>0]=n.getUTCDate(),a()[l+16>>>2>>>0]=n.getUTCMonth(),a()[l+20>>>2>>>0]=n.getUTCFullYear()-1900,a()[l+24>>>2>>>0]=n.getUTCDay(),n=(n.getTime()-Date.UTC(n.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,a()[l+28>>>2>>>0]=n},da:function(n,l){n=-9007199254740992>n||9007199254740992>>=0,n=new Date(1e3*n),a()[l>>>2>>>0]=n.getSeconds(),a()[l+4>>>2>>>0]=n.getMinutes(),a()[l+8>>>2>>>0]=n.getHours(),a()[l+12>>>2>>>0]=n.getDate(),a()[l+16>>>2>>>0]=n.getMonth(),a()[l+20>>>2>>>0]=n.getFullYear()-1900,a()[l+24>>>2>>>0]=n.getDay();var f=(Mt(n.getFullYear())?Qo:Jo)[n.getMonth()]+n.getDate()-1|0;a()[l+28>>>2>>>0]=f,a()[l+36>>>2>>>0]=-(60*n.getTimezoneOffset()),f=new Date(n.getFullYear(),6,1).getTimezoneOffset();var b=new Date(n.getFullYear(),0,1).getTimezoneOffset();n=(f!=b&&n.getTimezoneOffset()==Math.min(b,f))|0,a()[l+32>>>2>>>0]=n},ea:function(n){n>>>=0;var l=new Date(a()[n+20>>>2>>>0]+1900,a()[n+16>>>2>>>0],a()[n+12>>>2>>>0],a()[n+8>>>2>>>0],a()[n+4>>>2>>>0],a()[n>>>2>>>0],0),f=a()[n+32>>>2>>>0],b=l.getTimezoneOffset(),C=new Date(l.getFullYear(),6,1).getTimezoneOffset(),k=new Date(l.getFullYear(),0,1).getTimezoneOffset(),V=Math.min(k,C);return 0>f?a()[n+32>>>2>>>0]=+(C!=k&&V==b):0>>2>>>0]=l.getDay(),f=(Mt(l.getFullYear())?Qo:Jo)[l.getMonth()]+l.getDate()-1|0,a()[n+28>>>2>>>0]=f,a()[n>>>2>>>0]=l.getSeconds(),a()[n+4>>>2>>>0]=l.getMinutes(),a()[n+8>>>2>>>0]=l.getHours(),a()[n+12>>>2>>>0]=l.getDate(),a()[n+16>>>2>>>0]=l.getMonth(),a()[n+20>>>2>>>0]=l.getYear(),n=l.getTime(),BigInt(isNaN(n)?-1:n/1e3)},aa:ei,ba:ti,R:function(n,l,f,b){n>>>=0,l>>>=0,f>>>=0,b>>>=0;var C=new Date().getFullYear(),k=new Date(C,0,1),V=new Date(C,6,1);C=k.getTimezoneOffset();var F=V.getTimezoneOffset(),re=Math.max(C,F);c()[n>>>2>>>0]=60*re,a()[l>>>2>>>0]=+(C!=F),n=J=>J.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],k=n(k),V=n(V),F{Kt("")},f:function(n,l,f){return n>>>=0,l=ri(l>>>0,f>>>0),Ir[n](...l)},M:function(n,l,f){return n>>>=0,l=ri(l>>>0,f>>>0),Ir[n](...l)},H:()=>{},L:()=>Date.now(),ra:()=>{throw bt+=1,"unwind"},T:function(){return 4294901760},u:()=>performance.timeOrigin+performance.now(),x:()=>P?(Ri(),wr(Oi)).cpus().length:navigator.hardwareConcurrency,Q:function(n){n>>>=0;var l=o().length;if(n<=l||4294901760=f;f*=2){var b=l*(1+.2/f);b=Math.min(b,n+100663296);var C=Math;b=Math.max(n,b);e:{C=(C.min.call(C,4294901760,b+(65536-b%65536)%65536)-ae.buffer.byteLength+65535)/65536;try{ae.grow(C),Ne();var k=1;break e}catch{}k=void 0}if(k)return!0}return!1},ja:oi,la:ii,Z:Xe,B:ai,J:si,ga:ui,I:di,a:ae||d.wasmMemory,ka:pi,r:function(n,l,f,b){return pi(n>>>0,l>>>0,f>>>0,b>>>0)}},oe=function(){function n(f,b){return oe=f.exports,oe=dn(),oe=rl(),ge.Qb.push(oe.cb),Rt.unshift(oe.va),Ae=b,xr(),oe}var l={a:tl};if(Vt++,d.instantiateWasm)try{return d.instantiateWasm(l,n)}catch(f){ee(`Module.instantiateWasm callback failed with error: ${f}`),w(f)}return Ar(l,function(f){n(f.instance,f.module)}).catch(w),{}}(),fi=n=>(fi=oe.wa)(n);d.__embind_initialize_bindings=()=>(d.__embind_initialize_bindings=oe.xa)(),d._OrtInit=(n,l)=>(d._OrtInit=oe.ya)(n,l),d._OrtGetLastError=(n,l)=>(d._OrtGetLastError=oe.za)(n,l),d._OrtCreateSessionOptions=(n,l,f,b,C,k,V,F,re,J)=>(d._OrtCreateSessionOptions=oe.Aa)(n,l,f,b,C,k,V,F,re,J),d._OrtAppendExecutionProvider=(n,l)=>(d._OrtAppendExecutionProvider=oe.Ba)(n,l),d._OrtAddFreeDimensionOverride=(n,l,f)=>(d._OrtAddFreeDimensionOverride=oe.Ca)(n,l,f),d._OrtAddSessionConfigEntry=(n,l,f)=>(d._OrtAddSessionConfigEntry=oe.Da)(n,l,f),d._OrtReleaseSessionOptions=n=>(d._OrtReleaseSessionOptions=oe.Ea)(n),d._OrtCreateSession=(n,l,f)=>(d._OrtCreateSession=oe.Fa)(n,l,f),d._OrtReleaseSession=n=>(d._OrtReleaseSession=oe.Ga)(n),d._OrtGetInputOutputCount=(n,l,f)=>(d._OrtGetInputOutputCount=oe.Ha)(n,l,f),d._OrtGetInputName=(n,l)=>(d._OrtGetInputName=oe.Ia)(n,l),d._OrtGetOutputName=(n,l)=>(d._OrtGetOutputName=oe.Ja)(n,l),d._OrtFree=n=>(d._OrtFree=oe.Ka)(n),d._OrtCreateTensor=(n,l,f,b,C,k)=>(d._OrtCreateTensor=oe.La)(n,l,f,b,C,k),d._OrtGetTensorData=(n,l,f,b,C)=>(d._OrtGetTensorData=oe.Ma)(n,l,f,b,C),d._OrtReleaseTensor=n=>(d._OrtReleaseTensor=oe.Na)(n),d._OrtCreateRunOptions=(n,l,f,b)=>(d._OrtCreateRunOptions=oe.Oa)(n,l,f,b),d._OrtAddRunConfigEntry=(n,l,f)=>(d._OrtAddRunConfigEntry=oe.Pa)(n,l,f),d._OrtReleaseRunOptions=n=>(d._OrtReleaseRunOptions=oe.Qa)(n),d._OrtCreateBinding=n=>(d._OrtCreateBinding=oe.Ra)(n),d._OrtBindInput=(n,l,f)=>(d._OrtBindInput=oe.Sa)(n,l,f),d._OrtBindOutput=(n,l,f,b)=>(d._OrtBindOutput=oe.Ta)(n,l,f,b),d._OrtClearBoundOutputs=n=>(d._OrtClearBoundOutputs=oe.Ua)(n),d._OrtReleaseBinding=n=>(d._OrtReleaseBinding=oe.Va)(n),d._OrtRunWithBinding=(n,l,f,b,C)=>(d._OrtRunWithBinding=oe.Wa)(n,l,f,b,C),d._OrtRun=(n,l,f,b,C,k,V,F)=>(d._OrtRun=oe.Xa)(n,l,f,b,C,k,V,F),d._OrtEndProfiling=n=>(d._OrtEndProfiling=oe.Ya)(n),d._JsepOutput=(n,l,f)=>(d._JsepOutput=oe.Za)(n,l,f),d._JsepGetNodeName=n=>(d._JsepGetNodeName=oe._a)(n);var pn=d._pthread_self=()=>(pn=d._pthread_self=oe.$a)(),mn=d._malloc=n=>(mn=d._malloc=oe.ab)(n),At=d._free=n=>(At=d._free=oe.bb)(n);d.__emscripten_tls_init=()=>(d.__emscripten_tls_init=oe.cb)();var hi=d.__emscripten_thread_init=(n,l,f,b,C,k)=>(hi=d.__emscripten_thread_init=oe.eb)(n,l,f,b,C,k);d.__emscripten_thread_crashed=()=>(d.__emscripten_thread_crashed=oe.fb)();var gi=(n,l,f,b,C)=>(gi=oe.gb)(n,l,f,b,C),ro=n=>(ro=oe.hb)(n),no=d.__emscripten_thread_exit=n=>(no=d.__emscripten_thread_exit=oe.ib)(n),yi=()=>(yi=oe.jb)(),bi=(n,l)=>(bi=oe.kb)(n,l),fn=n=>(fn=oe.lb)(n),oo=n=>(oo=oe.mb)(n),io=()=>(io=oe.nb)(),wi=d.dynCall_ii=(n,l)=>(wi=d.dynCall_ii=oe.pb)(n,l),vi=n=>(vi=oe.qb)(n),$i=()=>($i=oe.rb)(),_i=n=>(_i=oe.sb)(n),Si=()=>(Si=oe.tb)();d.___start_em_js=839500,d.___stop_em_js=839661;function rl(){var n=oe;n=Object.assign({},n);var l=b=>C=>b(C)>>>0,f=b=>()=>b()>>>0;return n.wa=l(n.wa),n.$a=f(n.$a),n.ab=l(n.ab),n.emscripten_main_runtime_thread_id=f(n.emscripten_main_runtime_thread_id),n.mb=l(n.mb),n.nb=f(n.nb),n}d.wasmMemory=ae,d.stackSave=()=>io(),d.stackRestore=n=>fn(n),d.stackAlloc=n=>oo(n),d.keepRuntimeAlive=()=>0Bi)});var Mi=Wr((gm,cl)=>{cl.exports=\'"use strict";var Module={},ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads"),parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",e=>onmessage({data:e}));var fs=require("fs"),vm=require("vm");Object.assign(global,{self:global,require,Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:e=>vm.runInThisContext(fs.readFileSync(e,"utf8"),{filename:e}),postMessage:e=>parentPort.postMessage(e),performance:global.performance||{now:Date.now}})}var initializedJS=!1;function threadPrintErr(...e){var a=e.join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,a+`\\n`);return}console.error(a)}function threadAlert(...e){var a=e.join(" ");postMessage({cmd:"alert",text:a,threadId:Module._pthread_self()})}var err=threadPrintErr;self.alert=threadAlert,Module.instantiateWasm=(e,a)=>{var r=Module.wasmModule;Module.wasmModule=null;var t=new WebAssembly.Instance(r,e);return a(t)},self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let r=[];self.onmessage=t=>r.push(t),self.startWorker=t=>{Module=t,postMessage({cmd:"loaded"});for(let s of r)handleMessage(s);self.onmessage=handleMessage},Module.wasmModule=e.data.wasmModule;for(const t of e.data.handlers)Module[t]=(...s)=>{postMessage({cmd:"callHandler",handler:t,args:s})};if(Module.wasmMemory=e.data.wasmMemory,Module.buffer=Module.wasmMemory.buffer,Module.ENVIRONMENT_IS_PTHREAD=!0,typeof e.data.urlOrBlob=="string")importScripts(e.data.urlOrBlob);else{var a=URL.createObjectURL(e.data.urlOrBlob);importScripts(a),URL.revokeObjectURL(a)}ortWasmThreaded(Module)}else if(e.data.cmd==="run"){Module.__emscripten_thread_init(e.data.pthread_ptr,0,0,1),Module.__emscripten_thread_mailbox_await(e.data.pthread_ptr),Module.establishStackSpace(),Module.PThread.receiveObjectTransfer(e.data),Module.PThread.threadInitTLS(),initializedJS||(initializedJS=!0);try{Module.invokeEntryPoint(e.data.start_routine,e.data.arg)}catch(r){if(r!="unwind")throw r}}else e.data.cmd==="cancel"?Module._pthread_self()&&Module.__emscripten_thread_exit(-1):e.data.target==="setimmediate"||(e.data.cmd==="checkMailbox"?initializedJS&&Module.checkMailbox():e.data.cmd&&(err(`worker.js received unknown command ${e.data.cmd}`),err(e.data)))}catch(r){throw Module.__emscripten_thread_crashed?.(),r}}self.onmessage=handleMessage;\\n\'});var go,Gt,tr,bn,Gr,Hi,yo,ye=Y(()=>{"use strict";go=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}`)}},Gt=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}`)}},tr=e=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][e],bn=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}`)}},Gr=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}`)}},Hi=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool",yo=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 wn=Y(()=>{"use strict"});var Fi=Y(()=>{"use strict";wn()});var qi,ji=Y(()=>{"use strict";qi="1.18.0"});var Ki,pt,bo=Y(()=>{"use strict";ji();Ki="warning",pt={wasm:{},webgl:{},webgpu:{},versions:{common:qi},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}`);Ki=e}},get logLevel(){return Ki}};Object.defineProperty(pt,"logLevel",{enumerable:!0})});var vr,Yi=Y(()=>{"use strict";bo();vr=pt});var Zi=Y(()=>{"use strict"});var Xi=Y(()=>{"use strict";vn()});var Ji=Y(()=>{"use strict"});var ea=Y(()=>{"use strict";vn()});var vn=Y(()=>{"use strict";Zi();Xi();Ji();ea()});var $n=Y(()=>{"use strict";vn()});var wo,ta,rr,Ht,vo=Y(()=>{"use strict";bo();wo=(e,t)=>{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||console.timeStamp(`${e}::ORT::${t}`)},ta=(e,t)=>{let r=new Error().stack?.split(/\\r\\n|\\r|\\n/g)||[],o=!1;for(let i=0;i{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||ta("BEGIN",e)},Ht=e=>{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||ta("END",e)}});var ra=Y(()=>{"use strict";wn();$n();vo()});var na=Y(()=>{"use strict";ra()});var oa=Y(()=>{"use strict"});var ia=Y(()=>{"use strict"});var aa=Y(()=>{"use strict"});var sa=Y(()=>{"use strict"});var ua=Y(()=>{"use strict";wn();$n()});var da=Y(()=>{"use strict";ua()});var $r=Y(()=>{"use strict";Fi();Yi();na();$n();oa();ia();vo();aa();sa();da()});var _l,Sl,la,ca,pa,xl,Ve,Lt=Y(()=>{"use strict";ye();_l=["V","I","W","E","F"],Sl=(e,t)=>{console.log(`[${_l[e]},${new Date().toISOString()}]${t}`)},pa=(e,t)=>{la=e,ca=t},xl=(e,t)=>{let r=Gr(e),o=Gr(la);r>=o&&Sl(r,typeof t=="function"?t():t)},Ve=(...e)=>{ca&&xl(...e)}});var ma,fa=Y(()=>{"use strict";ye();ma=(e,t)=>new(bn(t))(e)});var _n=Y(()=>{"use strict"});var ha,$o,_o,Cl,Al,ga,xo,So,ba,wa=Y(()=>{"use strict";Lt();_n();ha=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]]),$o=[],_o=e=>Math.ceil(e/16)*16,Cl=e=>{for(let t=0;t<$o.length;t++){let r=$o[t];if(e<=r)return r}return Math.ceil(e/16)*16},Al=1,ga=()=>Al++,xo=async(e,t,r,o)=>{let i=_o(r),u=e.device.createBuffer({size:i,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let a=e.getCommandEncoder();e.endComputePass(),a.copyBufferToBuffer(t,0,u,0,i),e.flush(),await u.mapAsync(GPUMapMode.READ);let c=u.getMappedRange();if(o){let p=o();return p.set(new Uint8Array(c,0,r)),p}else return new Uint8Array(c.slice(0,r))}finally{u.destroy()}},So=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 ha)$o.push(r),this.freeBuffers.set(r,[]),this.freeUniformBuffers.set(r,[])}upload(t,r){let o=r.buffer,i=r.byteOffset,u=r.byteLength,a=_o(u),c=this.storageCache.get(t);if(!c)throw new Error("gpu data for uploading does not exist");if(c.originalSize!==u)throw new Error(`inconsistent data size. gpu data size=${c.originalSize}, data size=${u}`);let p=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),h=p.getMappedRange();new Uint8Array(h).set(new Uint8Array(o,i,u)),p.unmap();let d=this.backend.getCommandEncoder();this.backend.endComputePass(),d.copyBufferToBuffer(p,0,c.gpuData.buffer,0,a),Ve("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(p)}memcpy(t,r){let o=this.storageCache.get(t);if(!o)throw new Error("source gpu data for memcpy does not exist");let i=this.storageCache.get(r);if(!i)throw new Error("destination gpu data for memcpy does not exist");if(o.originalSize!==i.originalSize)throw new Error("inconsistent source and destination gpu data size");let u=_o(o.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(o.gpuData.buffer,0,i.gpuData.buffer,0,u)}registerExternalBuffer(t,r,o){let i;if(o){if(i=this.externalBuffers.get(o),i===void 0)throw new Error("previous buffer is not registered");if(t===o)return Ve("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${i}, buffer is the same, skip.`),i;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.\n Please use the previous external buffer!`);this.externalBuffers.delete(o)}else i=ga();return this.storageCache.set(i,{gpuData:{id:i,type:0,buffer:t},originalSize:r}),this.externalBuffers.set(t,i),Ve("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${i}, registered.`),i}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 o=Cl(t),i,u=(r&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(r&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(u||a){let h=(u?this.freeBuffers:this.freeUniformBuffers).get(o);h?h.length>0?i=h.pop():i=this.backend.device.createBuffer({size:o,usage:r}):i=this.backend.device.createBuffer({size:o,usage:r})}else i=this.backend.device.createBuffer({size:o,usage:r});let c={id:ga(),type:0,buffer:i};return this.storageCache.set(c.id,{gpuData:c,originalSize:t}),Ve("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${c.id}`),c}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 o=this.storageCache.get(t);if(!o)throw new Error("data does not exist");await xo(this.backend,o.gpuData.buffer,o.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=ha.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let o=this.freeBuffers.get(t.size)||[];r===void 0||o.length>=r?t.destroy():o.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let o=this.freeUniformBuffers.get(t.size)||[];r===void 0||o.length>=r?t.destroy():o.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(o=>{o.destroy()}),this.capturedPendingBuffers.delete(t))}},ba=(...e)=>new So(...e)});var Co,ve,Ze=Y(()=>{"use strict";Co=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}},ve=e=>new Co(e)});var Ao,It,M,nr,Sn,xn,Cn,Se=Y(()=>{"use strict";Ao=class{static calcMatMulShape(t,r){return t[1]!==r[0]?void 0:[t[0],r[1]]}},It=class{static calcShape(t,r,o=!1){let i=t.length,u=r.length;if(i===0)return r;if(u===0)return t;let a=Math.max(t.length,r.length),c=new Array(a);if(o){if(i<2||u<2)return;let p=Ao.calcMatMulShape([t[i-2],t[i-1]],[r[u-2],r[u-1]]);if(p===void 0)return;[c[a-2],c[a-1]]=p}for(let p=o?3:1;p<=a;p++){let h=i-p<0?1:t[i-p],d=u-p<0?1:r[u-p];if(h!==d&&h>1&&d>1)return;let y=Math.max(h,d);if(h&&d)c[a-p]=Math.max(h,d);else{if(y>1)return;c[a-p]=0}}return c}static isValidBroadcast(t,r){let o=t.length,i=r.length;if(o>i)return!1;for(let u=1;u<=o;u++)if(t[o-u]!==1&&t[o-u]!==r[i-u])return!1;return!0}},M=class e{static size(t){return e.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let o=t.length;if(o===0)return[];let i=new Array(o),u=o-1;for(;u>=0;){if(t[u]%r===0){i[u]=t[u]/r;break}if(r%t[u]!==0)throw new Error("cannot convert shape");i[u]=1,r/=t[u],u--}for(u--;u>=0;u--)i[u]=t[u];return i}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,o){let i=1;for(let u=r;u=0;--i)o[i]=o[i+1]*t[i+1];return o}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(o=>this.normalizeAxis(o,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(o=>t[o]):t.slice().reverse()}static padShape(t,r){let o=t.length;return t.map((i,u)=>i+r[u]+r[u+o])}static areEqual(t,r){return t.length!==r.length?!1:t.every((o,i)=>o===r[i])}},nr=class e{static adjustPoolAttributes(t,r,o,i,u,a){if(!t&&o.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 c=0;c=o.length?o.push(r[c+2]):o[c]=r[c+2];for(let c=0;c=o[c]||a[c+o.length]>=o[c])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,o,i,u,a,c){if(c){if(u.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(i.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let p=0;p{"use strict";ye();Se();or=64,To=(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";case 1:return t>1?`vec${t}`:"f32";case 6:return t>1?`vec${t}`:"i32";case 12:return t>1?`vec${t}`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4"];default:throw new Error(`Unknown data type: ${e}`)}},De=(e,t=1)=>{let r=To(e,t);return typeof r=="string"?r:r[0]},et=(e,t=1)=>{let r=To(e,t);return typeof r=="string"?r:r[1]},Z=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:M.computeStrides(r)})}),t},Me=e=>e%4===0?4:e%2===0?2:1,$t=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,ir=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}(${r})`,_t=(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,fe=(e,t,r,o)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?o==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:o==="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,Eo=(e,t,r,o,i)=>{let u=typeof r=="number",a=u?r:r.length,c=[...new Array(a).keys()],p=a<2?"u32":a<=4?`vec${a}`:`array`,h=To(t,i),d=typeof h=="string"?h:h[1],y=typeof h=="string"?h:h[0],w={indices:p,value:d,storage:y,tensor:t},_=G=>typeof G=="string"?G:`${G}u`,v={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},S=u?"uniforms.":"",A=`${S}${e}_shape`,I=`${S}${e}_strides`,x="";for(let G=0;G ${w.indices} {\n var indices: ${w.indices};\n var current = offset;\n ${x}\n return indices;\n }`,P=G=>(v.offsetToIndices=!0,a<2?G:`o2i_${e}(${G})`),O=[];if(a>=2)for(let G=a-1;G>=0;G--)O.push(`${fe(I,G,a)} * (indices[${G}])`);let R=a<2?"":`\n fn i2o_${e}(indices: ${w.indices}) -> u32 {\n return ${O.join("+")};\n }`,L=G=>(v.indicesToOffset=!0,a<2?G:`i2o_${e}(${G})`),N=(...G)=>a===0?"0u":`${w.indices}(${G.map(_).join(",")})`,K=(G,ne)=>a<2?`${G}`:`${fe(G,ne,a)}`,Q=(G,ne,xe)=>a<2?`${G}=${xe};`:`${fe(G,ne,a)}=${xe};`,he={},W=(G,ne)=>{v.broadcastedIndicesToOffset=!0;let xe=`${ne.name}broadcastedIndicesTo${e}Offset`;if(xe in he)return`${xe}(${G})`;let Ke=[];for(let Be=a-1;Be>=0;Be--){let Ge=ne.indicesGet("outputIndices",Be+ne.rank-a);Ke.push(`${K(I,Be)} * (${Ge} % ${K(A,Be)})`)}return he[xe]=`fn ${xe}(outputIndices: ${ne.type.indices}) -> u32 {\n return ${Ke.length>0?Ke.join("+"):"0u"};\n }`,`${xe}(${G})`},se=(G,ne)=>(()=>{if(w.storage===w.value)return`${e}[${G}]=${ne};`;if(w.storage==="vec2"&&w.value==="i32")return`${e}[${G}]=vec2(u32(${ne}), select(0u, 0xFFFFFFFFu, ${ne} < 0));`;if(w.storage==="vec2"&&w.value==="u32")return`${e}[${G}]=vec2(u32(${ne}), 0u);`;if(w.storage==="u32"&&w.value==="vec4")return`${e}[${G}]=dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(${ne}));`;throw new Error(`not supported combination of storage type ${w.storage} and value type ${w.value} yet`)})(),Ce=G=>(()=>{if(w.storage===w.value)return`${e}[${G}]`;if(w.storage==="vec2"&&w.value==="i32")return`i32(${e}[${G}].x)`;if(w.storage==="vec2"&&w.value==="u32")return`u32(${e}[${G}].x)`;if(w.storage==="u32"&&w.value==="vec4")return`vec4(bool(${e}[${G}] & 0xFFu), bool(${e}[${G}] & 0xFF00u), bool(${e}[${G}] & 0xFF0000u), bool(${e}[${G}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${w.storage} and value type ${w.value} yet`)})(),We=a<2?"":`\n fn get_${e}ByIndices(indices: ${w.indices}) -> ${d} {\n return ${Ce(`i2o_${e}(indices)`)};\n }`,ee=a<2?"":(()=>{let G=c.map(xe=>`d${xe}: u32`).join(", "),ne=c.map(xe=>`d${xe}`).join(", ");return`\n fn get_${e}(${G}) -> ${d} {\n return get_${e}ByIndices(${N(ne)});\n }`})(),ae=(...G)=>{if(G.length!==a)throw new Error(`indices length must be ${a}`);let ne=G.map(_).join(",");return a===0?Ce("0u"):a===1?Ce(ne[0]):(v.get=!0,v.getByIndices=!0,v.indicesToOffset=!0,`get_${e}(${ne})`)},Ae=G=>a<2?Ce(G):(v.getByIndices=!0,v.indicesToOffset=!0,`get_${e}ByIndices(${G})`),me=a<2?"":`\n fn set_${e}ByIndices(indices: ${w.indices}, value: ${d}) {\n ${se(`i2o_${e}(indices)`,"value")}\n }`,ie=a<2?"":(()=>{let G=c.map(xe=>`d${xe}: u32`).join(", "),ne=c.map(xe=>`d${xe}`).join(", ");return`\n fn set_${e}(${G}, value: ${d}) {\n set_${e}ByIndices(${N(ne)}, value);\n }`})();return{impl:()=>{let G=[],ne=!1;return v.offsetToIndices&&(G.push(E),ne=!0),v.indicesToOffset&&(G.push(R),ne=!0),v.broadcastedIndicesToOffset&&(Object.values(he).forEach(xe=>G.push(xe)),ne=!0),v.set&&(G.push(ie),ne=!0),v.setByIndices&&(G.push(me),ne=!0),v.get&&(G.push(ee),ne=!0),v.getByIndices&&(G.push(We),ne=!0),!u&&ne&&G.unshift(`const ${A} = ${w.indices}(${r.join(",")});`,`const ${I} = ${w.indices}(${M.computeStrides(r).join(",")});`),G.join(`\n`)},type:w,offsetToIndices:P,indicesToOffset:L,broadcastedIndicesToOffset:W,indices:N,indicesGet:K,indicesSet:Q,set:(...G)=>{if(G.length!==a+1)throw new Error(`indices length must be ${a}`);let ne=G[a];if(typeof ne!="string")throw new Error("value must be string");let xe=G.slice(0,a).map(_).join(",");return a===0?se("0u",ne):a===1?se(xe[0],ne):(v.set=!0,v.setByIndices=!0,v.indicesToOffset=!0,`set_${e}(${xe}, ${ne})`)},setByOffset:se,setByIndices:(G,ne)=>a<2?se(G,ne):(v.setByIndices=!0,v.indicesToOffset=!0,`set_${e}ByIndices(${G}, ${ne});`),get:ae,getByOffset:Ce,getByIndices:Ae,usage:o,name:e,strides:I,shape:A,rank:a}},U=(e,t,r,o=1)=>Eo(e,t,r,"input",o),j=(e,t,r,o=1)=>Eo(e,t,r,"output",o),An=(e,t,r,o=1)=>Eo(e,t,r,"internal",o),Io=class{constructor(t,r){this.normalizedDispatchGroup=t;this.limits=r;this.internalVariables=[];this.variables=[];this.uniforms=[];this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=or){let r=typeof t=="number"?t:t[0],o=typeof t=="number"?1:t[1],i=typeof t=="number"?1:t[2];if(r>this.limits.maxComputeWorkgroupSizeX||o>this.limits.maxComputeWorkgroupSizeY||i>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${r}, ${o}, ${i}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(r*o*i>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${r}, ${o}, ${i}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let u=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,a=u?`@builtin(global_invocation_id) global_id : vec3,\n @builtin(workgroup_id) workgroup_id : vec3,\n @builtin(local_invocation_id) local_id : vec3`:`@builtin(global_invocation_id) global_id : vec3,\n @builtin(local_invocation_id) local_id : vec3,\n @builtin(local_invocation_index) local_idx : u32,\n @builtin(workgroup_id) workgroup_id : vec3,\n @builtin(num_workgroups) num_workgroups : vec3`,c=u?"let global_idx = global_id.x; let local_idx = local_id.x;":`let global_idx = (workgroup_id.z * num_workgroups[0] * num_workgroups[1] +\n workgroup_id.y * num_workgroups[0] + workgroup_id.x) * ${r*o*i}u + local_idx;`;return`@compute @workgroup_size(${r}, ${o}, ${i})\n fn main(${a}) {\n ${c}\n `}appendVariableUniforms(t){t.rank!==0&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,r){if(t.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let o=t.usage==="input"?"read":"read_write",i=t.type.storage;return`@group(0) @binding(${r}) var ${t.name}: array<${i}>;`}declareVariables(...t){return t.map(r=>this.declareVariable(r,this.variableIndex++)).join(`\n`)}registerInternalVariable(t){if(t.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(r=>this.registerInternalVariable(r)),this}registerUniform(t,r,o=1){return this.uniforms.push({name:t,type:r,length:o}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let t=[];for(let{name:r,type:o,length:i}of this.uniforms)if(i&&i>4)o==="f16"?t.push(`@align(16) ${r}:array, ${Math.ceil(i/8)}>`):t.push(`${r}:array, ${Math.ceil(i/4)}>`);else{let u=i==null||i===1?o:`vec${i}<${o}>`;t.push(`${r}:${u}`)}return`\n struct Uniforms { ${t.join(", ")} };\n @group(0) @binding(${this.variableIndex}) var uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(`\n`)+this.internalVariables.map(t=>t.impl()).join(`\n`)}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])}},va=(e,t)=>new Io(e,t),_r=(e,t)=>{let r=e.length,o=[];for(let i=0;i1&&a===1&&o.unshift(u)}return o}});var Il,$a,Tl,El,yt,_a,Sa,Sr=Y(()=>{"use strict";ye();Se();Ze();_e();Il=e=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.")},$a=(e,t)=>t&&t.length!==e?[...new Array(e).keys()].reverse():t,Tl=(e,t)=>M.sortBasedOnPerm(e,$a(e.length,t)),El=(e,t,r,o)=>{let i=[];i.push(`fn perm(i: ${o.type.indices}) -> ${r.type.indices} {\n var a: ${r.type.indices};`);for(let u=0;u{let r=e.dataType,o=e.dims.length,i=$a(o,t),u=Tl(e.dims,i),a=j("output",r,u.length),c=U("a",r,o),p=h=>`\n ${h.registerUniform("output_size","u32").declareVariables(c,a)}\n\n ${El(i,o,c,a)}\n\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let indices = ${a.offsetToIndices("global_idx")};\n let aIndices = perm(indices);\n\n ${a.setByOffset("global_idx",c.getByIndices("aIndices"))}\n }`;return{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:h=>{let d=M.size(u);return{outputs:[{dims:u,dataType:h[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:[{type:12,data:d},...Z(h[0].dims,u)]}},getShaderSource:p}},_a=(e,t)=>{Il(e.inputs),e.compute(yt(e.inputs[0],t.perm))},Sa=e=>ve({perm:e.perm})});var Pl,kl,Ol,Rl,Bl,Dl,zl,Ml,Ul,Vl,Tt,xa,Ca,Aa,Ia,Ta,Ea,Pa,ka,Oa,Ra,Ba=Y(()=>{"use strict";ye();Se();_e();In();Sr();Pl={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"},kl={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"},Ol={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Rl={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Bl=(e,t)=>{let r=[];for(let o=t-e;o{let r=[],o=e.length;for(let u=0;ue[u]);return[r,i]},zl=(e,t)=>{let r=e.length+t.length,o=[],i=0;for(let u=0;u{for(let r=0;r{let r=[];if(!Ml(e,t)){for(let o=0;or.push(o))}return r},Vl=(e,t,r,o,i,u,a)=>{let c=r[0].dims,p=M.size(u),h=M.size(a),d=U("_A",r[0].dataType,c),y=j("output",i,u),w=32,_=`\n var aBestValues : array;\n `;return{name:e,shaderCache:t,getShaderSource:S=>`\n ${S.registerUniform("reduceSize","u32").declareVariables(d,y)}\n ${_}\n fn DIV_CEIL(a : u32, b : u32) -> u32 {\n return ((a - 1u) / b + 1u);\n }\n ${S.mainStart(w)}\n\n let outputIndex = global_idx / ${w};\n let offset = outputIndex * uniforms.reduceSize;\n\n var bestValue = f32(${Ol[o]});\n let Length = uniforms.reduceSize;\n for (var k = local_idx; k < Length; k = k + ${w}) {\n let candidate = f32(${d.getByOffset("offset + k")});\n bestValue = ${Pl[o]};\n }\n aBestValues[local_idx] = bestValue;\n workgroupBarrier();\n\n var reduceSize = min(Length, ${w}u);\n for (var currentSize = reduceSize / 2u; reduceSize > 1u;\n currentSize = reduceSize / 2u) {\n let interval = DIV_CEIL(reduceSize, 2u);\n if (local_idx < currentSize) {\n let candidate = aBestValues[local_idx + interval];\n bestValue = ${kl[o]};\n aBestValues[local_idx] = bestValue;\n }\n reduceSize = interval;\n workgroupBarrier();\n }\n\n if (local_idx == 0u) {\n ${y.setByOffset("outputIndex",`${o==="mean"?`${y.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${y.type.storage}(${Rl[o]})`}`)};\n }\n }`,getRunData:()=>({outputs:[{dims:u,dataType:i}],dispatchGroup:{x:p},programUniforms:[{type:12,data:h}]})}},Tt=(e,t,r,o)=>{let i=e.inputs.length===1?r:Po(e.inputs,r),u=i.axes;u.length===0&&!i.noopWithEmptyAxes&&(u=e.inputs[0].dims.map((_,v)=>v));let a=M.normalizeAxes(u,e.inputs[0].dims.length),c=a,p=e.inputs[0],h=Ul(c,e.inputs[0].dims.length);h.length>0&&(p=e.compute(yt(e.inputs[0],h),{inputs:[0],outputs:[-1]})[0],c=Bl(c.length,p.dims.length));let[d,y]=Dl(p.dims,c),w=d;i.keepDims&&(w=zl(d,a)),e.compute(Vl(t,{hint:i.cacheKey,inputDependencies:["type"]},[p],o,e.inputs[0].dataType,w,y),{inputs:[p]})},xa=(e,t)=>{Tt(e,"ReduceMeanShared",t,"mean")},Ca=(e,t)=>{Tt(e,"ReduceL1Shared",t,"l1")},Aa=(e,t)=>{Tt(e,"ReduceL2Shared",t,"l2")},Ia=(e,t)=>{Tt(e,"ReduceLogSumExpShared",t,"logSumExp")},Ta=(e,t)=>{Tt(e,"ReduceMaxShared",t,"max")},Ea=(e,t)=>{Tt(e,"ReduceMinShared",t,"min")},Pa=(e,t)=>{Tt(e,"ReduceProdShared",t,"prod")},ka=(e,t)=>{Tt(e,"ReduceSumShared",t,"sum")},Oa=(e,t)=>{Tt(e,"ReduceSumSquareShared",t,"sumSquare")},Ra=(e,t)=>{Tt(e,"ReduceLogSumShared",t,"logSum")}});var Et,Wl,Tn,Po,Pt,Nl,Gl,Hl,Ll,Fl,ql,jl,Kl,Yl,Zl,kt,Da,za,Ma,Ua,Va,Wa,Na,Ga,Ha,La,In=Y(()=>{"use strict";ye();Se();Ze();_e();Ba();Et=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.")},Wl=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Tn=(e,t,r,o,i,u,a=!1,c=!1)=>{let p=[],h=r[0].dims,d=h.length,y=M.normalizeAxes(i,d),w=!c&&y.length===0;h.forEach((A,I)=>{w||y.indexOf(I)>=0?a&&p.push(1):p.push(A)});let _=p.length,v=M.size(p);return{name:e,shaderCache:t,getShaderSource:A=>{let I=[],x=U("_A",r[0].dataType,d),E=j("output",u,_),P=o(x,E,y),O=P[2];for(let R=0,L=0;R=0?(a&&L++,O=`for(var j${R}: u32 = 0; j${R} < ${h[R]}; j${R}++) {\n ${P[2].includes("last_index")?`let last_index = j${R};`:""}\n ${x.indicesSet("input_indices",R,`j${R}`)}\n ${O}\n }`):(I.push(`${x.indicesSet("input_indices",R,E.indicesGet("output_indices",L))};`),L++);return`\n\n ${A.registerUniform("output_size","u32").declareVariables(x,E)}\n\n ${A.mainStart()}\n ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n var input_indices: ${x.type.indices};\n let output_indices = ${E.offsetToIndices("global_idx")};\n\n ${I.join(`\n`)}\n ${P[0]} // init ops for reduce max/min\n ${P[1]}\n ${O}\n ${P[3]}\n ${P.length===4?E.setByOffset("global_idx","value"):P.slice(4).join(`\n`)}\n }`},getRunData:()=>({outputs:[{dims:p,dataType:u}],dispatchGroup:{x:Math.ceil(v/64)},programUniforms:[{type:12,data:v},...Z(h,p)]})}},Po=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(o=>r.push(Number(o))),ve({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},Pt=(e,t,r,o)=>{let i=e.inputs,u=i.length===1?r:Po(i,r);e.compute(Tn(t,{hint:u.cacheKey,inputDependencies:["rank"]},[i[0]],u.noopWithEmptyAxes&&u.axes.length===0?Wl:o,u.axes,i[0].dataType,u.keepDims,u.noopWithEmptyAxes),{inputs:[0]})},Nl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceLogSum",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${o.getByIndices("input_indices")};`,"value = log(value);"])},Gl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceL1",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += abs(${o.getByIndices("input_indices")});`,""])},Hl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceL2",t,(o,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${o.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Ll=(e,t)=>{Et(e.inputs),Pt(e,"ReduceLogSumExp",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += exp(${o.getByIndices("input_indices")});`,"value = log(value);"])},Fl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMax",t,(o,i,u)=>{let a=[];for(let c=0;c=0||u.length===0)&&a.push(o.indicesSet("input_indices",c,0));return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};`,`value = max(value, ${o.getByIndices("input_indices")});`,""]})},ql=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMean",t,(o,i,u)=>{let a=1;for(let c=0;c=0||u.length===0)&&(a*=e.inputs[0].dims[c]);return["var sum = f32(0);","",`sum += f32(${o.getByIndices("input_indices")});`,`let value = ${i.type.value}(sum / ${a});`]})},jl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMin",t,(o,i,u)=>{let a=[];for(let c=0;c=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};`,`value = min(value, ${o.getByIndices("input_indices")});`,""]})},Kl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceProd",t,(o,i)=>[`var value = ${i.type.storage}(1);`,"",`value *= ${o.getByIndices("input_indices")};`,""])},Yl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceSum",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${o.getByIndices("input_indices")};`,""])},Zl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceSumSquare",t,(o,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${o.getByIndices("input_indices")}; value += t * t;`,""])},kt=(e,t,r)=>{if(t.length===0)return r;let o=1,i=1;for(let u=0;u1024},Da=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ql(e,t):xa(e,t)},za=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Gl(e,t):Ca(e,t)},Ma=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Hl(e,t):Aa(e,t)},Ua=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ll(e,t):Ia(e,t)},Va=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Fl(e,t):Ta(e,t)},Wa=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?jl(e,t):Ea(e,t)},Na=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Kl(e,t):Pa(e,t)},Ga=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Yl(e,t):ka(e,t)},Ha=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Zl(e,t):Oa(e,t)},La=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Nl(e,t):Ra(e,t)}});var Fa,qa,ja,ko,Ka=Y(()=>{"use strict";ye();Ze();In();Fa=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.")},qa=(e,t)=>{Fa(e.inputs);let r=(o,i,u)=>{let a=[];for(let c=0;c=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};\nvar best_index : i32 = 0;`,`if (${o.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {\n value = ${o.getByIndices("input_indices")};\n best_index = i32(last_index);\n }`,"",i.setByOffset("global_idx","best_index")]};e.compute(Tn("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ja=(e,t)=>{Fa(e.inputs);let r=(o,i,u)=>{let a=[];for(let c=0;c=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};\nvar best_index : i32 = 0;`,`if (${o.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {\n value = ${o.getByIndices("input_indices")};\n best_index = i32(last_index);\n }`,"",i.setByOffset("global_idx","best_index")]};e.compute(Tn("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ko=e=>ve(e)});var Xl,Ql,Jl,En,Ya,Za,Oo=Y(()=>{"use strict";ye();Se();Ze();_e();Xl=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,o=e[r],i=o.dataType,u=o.dims.length;e.forEach((a,c)=>{if(c!==r){if(a.dataType!==i)throw new Error("input tensors should be one type");if(a.dims.length!==u)throw new Error("input tensors should have the same shape");a.dims.forEach((p,h)=>{if(h!==t&&p!==o.dims[h])throw new Error("non concat dimensions must match")})}})},Ql=(e,t)=>`\n fn calculateInputIndex(index: u32) -> u32 {\n let sizeInConcatAxis = array(${t});\n for (var i: u32 = 0u; i < ${e}; i += 1u ) {\n if (index < sizeInConcatAxis[i]) {\n return i;\n }\n }\n return ${e}u;\n }`,Jl=(e,t)=>{let r=e.length,o=[];for(let i=0;i{let i=M.size(r),u=new Array(e.length),a=new Array(e.length),c=0,p=[],h=[],d=[{type:12,data:i}];for(let S=0;S`uniforms.sizeInConcatAxis${S}`).join(","),v=S=>`\n\n ${(()=>{S.registerUniform("outputSize","u32");for(let A=0;A(${_});\n ${w} -= sizeInConcatAxis[inputIndex - 1u];\n }\n\n ${Jl(a,y)}\n }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:r,dataType:o}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:d}),getShaderSource:v}},Ya=(e,t)=>{let r=e.inputs,o=r[0].dims,i=M.normalizeAxis(t.axis,o.length);Xl(r,i);let u=o.slice();u[i]=r.reduce((c,p)=>c+(p.dims.length>i?p.dims[i]:0),0);let a=r.filter(c=>M.size(c.dims)>0);e.compute(En(a,i,u,r[0].dataType),{inputs:a})},Za=e=>ve({axis:e.axis})});var ec,tc,rc,nc,Pn,oc,Xa,Ro=Y(()=>{"use strict";ye();_n();_e();Oo();ec=(e,t)=>{let r=e[0],o=e[1],i=e[2],u=e[3],a=e[4],c=e[5];if(a&&c)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 p=r.dims[0],h=r.dims[1],d=r.dims[2];if(i.dims.length!==1)throw new Error(\'Input "bias" is expected to have 1 dimensions\');if(o.dims.length!==2)throw new Error(\'Input "weights" is expected to have 2 dimensions\');if(o.dims[0]!==d)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(i.dims[0]!==o.dims[1])throw new Error(\'Input "bias" dimension 0 should have same length as dimension 1 of input "weights"\');let y=i.dims[0]/3,w=y,_=w;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let E of t.qkvHiddenSizes)if(E%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");y=t.qkvHiddenSizes[0],w=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let v=h;if(y!==w)throw new Error("qkv_hidden_sizes first element should be same as the second");if(i.dims[0]!==y+w+_)throw new Error(\'Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes\');let S=0;if(a){if(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]!==p)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]!==w/t.numHeads)throw new Error(\'Input "past" fifth dimension must be k_hidden_size / num_heads\');t.pastPresentShareBuffer||(S=a.dims[3])}let A=v+S,I=-1,x=0;if(u)throw new Error("Mask not supported");if(a)throw new Error("past is not supported");return{batchSize:p,sequenceLength:h,pastSequenceLength:S,kvSequenceLength:v,totalSequenceLength:A,maxSequenceLength:I,inputHiddenSize:d,hiddenSize:y,vHiddenSize:_,headSize:Math.floor(y/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:x,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},tc=(e,t,r,o)=>{let i=Me(o),u=64,a=o/i;a{let _=j("x",t.dataType,t.dims,i),S=[{name:"d_inv",type:et(t.dataType)},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`\n var thread_max: array;\n var thread_sum: array;\n ${w.registerUniforms(S).declareVariables(_)}\n ${w.mainStart([u,1,1])}\n let local_offset = local_idx * uniforms.elements_per_thread;\n let offset = workgroup_id.x * uniforms.d_comp + local_offset;\n\n var thread_max_vector = ${d}(-3.402823e+38f);\n for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n thread_max_vector = max(${d}(x[offset + i]), thread_max_vector);\n }\n thread_max[local_idx] = ${(()=>{switch(i){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${i}`)}})()};\n workgroupBarrier();\n\n var max_value = f32(-3.402823e+38f);\n for (var i = 0u; i < ${u}; i++) {\n max_value = max(thread_max[i], max_value);\n }\n\n var sum_vector = ${d}(0);\n for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n sum_vector += exp(${d}(x[offset + i]) - max_value);\n }\n thread_sum[local_idx] = ${(()=>{switch(i){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${i}`)}})()};\n workgroupBarrier();\n\n var sum: f32 = 0;\n for (var i = 0u; i < ${u}; i++) {\n sum += thread_sum[i];\n }\n\n if (sum == 0) {\n for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n x[offset + i] = ${_.type.value}(uniforms.d_inv);\n }\n } else {\n for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n var f32input = ${d}(x[offset + i]);\n x[offset + i] = ${_.type.value}(exp(f32input - max_value) / sum);\n }\n }\n }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${u};${h};${i}`},getShaderSource:y,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:p})}},rc=(e,t,r,o,i,u,a)=>{let c=a+i.kvSequenceLength,p=[i.batchSize,i.numHeads,i.sequenceLength,c],h=u.scale===0?1/Math.sqrt(i.headSize):u.scale,d=Me(i.headSize),y=i.headSize/d,w=12,_={x:Math.ceil(c/w),y:Math.ceil(i.sequenceLength/w),z:i.batchSize*i.numHeads},v=[{type:12,data:i.sequenceLength},{type:12,data:y},{type:12,data:c},{type:12,data:i.numHeads},{type:1,data:h}],S=o?["type","type","type"]:["type","type"],A=I=>{let x=U("q",t.dataType,t.dims,d),E=U("key",r.dataType,r.dims,d),P=[x,E];o&&P.push(U("relative_position_bias",o.dataType,o.dims));let O=j("output",t.dataType,p),R=et(1,d),L=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"alpha",type:"f32"}];return`\n const TILE_SIZE = ${w}u;\n\n var tileQ: array<${x.type.storage}, ${w*w}>;\n var tileK: array<${x.type.storage}, ${w*w}>;\n ${I.registerUniforms(L).declareVariables(...P,O)}\n ${I.mainStart([w,w,1])}\n // x holds the N and y holds the M\n let headIdx = workgroup_id.z;\n let m = workgroup_id.y * TILE_SIZE;\n let n = workgroup_id.x * TILE_SIZE;\n let qOffset = uniforms.M * uniforms.K * headIdx + m * uniforms.K;\n let kOffset = uniforms.N * uniforms.K * headIdx + n * uniforms.K;\n\n var value = ${R}(0);\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {\n tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];\n }\n if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {\n tileK[TILE_SIZE * local_id.y + local_id.x] = key[kOffset + local_id.y * uniforms.K + w + local_id.x];\n }\n workgroupBarrier();\n\n for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {\n value += ${R}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);\n }\n\n workgroupBarrier();\n }\n\n let headOffset = headIdx * uniforms.M * uniforms.N;\n if (global_id.y < uniforms.M && global_id.x < uniforms.N) {\n let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;\n var sum: f32 = ${(()=>{switch(d){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: ${d}`)}})()};\n output[outputIdx] = ${O.type.value} (sum * uniforms.alpha) + ${o?"relative_position_bias[outputIdx]":"0.0"};\n }\n }`};return{name:"AttentionProbs",shaderCache:{hint:`${d}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:p,dataType:t.dataType,gpuDataType:0}],dispatchGroup:_,programUniforms:v}),getShaderSource:A}},nc=(e,t,r,o,i)=>{let u=i+o.kvSequenceLength,a=[o.batchSize,o.sequenceLength,o.vHiddenSize],c=12,p={x:Math.ceil(o.vHeadSize/c),y:Math.ceil(o.sequenceLength/c),z:o.batchSize*o.numHeads},h=[{type:12,data:o.sequenceLength},{type:12,data:u},{type:12,data:o.vHeadSize},{type:12,data:o.numHeads},{type:12,data:o.vHiddenSize}];return{name:"AttentionScore",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:a,dataType:t.dataType,gpuDataType:0}],dispatchGroup:p,programUniforms:h}),getShaderSource:w=>{let _=U("probs",t.dataType,t.dims),v=U("v",r.dataType,r.dims),S=j("output",t.dataType,a),A=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"}];return`\n const TILE_SIZE = ${c}u;\n var tileQ: array<${_.type.value}, ${c*c}>;\n var tileK: array<${_.type.value}, ${c*c}>;\n ${w.registerUniforms(A).declareVariables(_,v,S)}\n ${w.mainStart([c,c,1])}\n let headIdx = workgroup_id.z;\n let m = global_id.y;\n let n = global_id.x;\n\n let offsetA = headIdx * (uniforms.M * uniforms.K) + m * uniforms.K;\n let offsetB = headIdx * (uniforms.N * uniforms.K) + n;\n\n var value = ${_.type.storage}(0);\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (m < uniforms.M && w + local_id.x < uniforms.K) {\n tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x];\n }\n if (n < uniforms.N && w + local_id.y < uniforms.K) {\n tileK[TILE_SIZE * local_id.y + local_id.x] = v[offsetB + (w + local_id.y) * uniforms.N];\n }\n workgroupBarrier();\n for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {\n value += tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * k + local_id.x];\n }\n workgroupBarrier();\n }\n\n // we need to transpose output from BNSH_v to BSND_v\n let batchIdx = workgroup_id.z / uniforms.num_heads;\n let currentBatchHeadNumber = workgroup_id.z % uniforms.num_heads;\n if (m < uniforms.M && n < uniforms.N) {\n let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size\n + currentBatchHeadNumber * uniforms.N + n;\n output[outputIdx] = value;\n }\n }`}}},Pn=(e,t,r,o,i,u,a,c,p,h,d)=>{let y=e.outputCount>1,w=e.outputCount>2,_=y&&w?h.pastSequenceLength:0,v=_+h.kvSequenceLength,S=[h.batchSize,h.numHeads,v,h.headSize],A=a?[a,r]:[r],I=y?e.compute(En(A,2,S,r.dataType),{inputs:A,outputs:[1]})[0]:r,x=[h.batchSize,h.numHeads,v,h.headSize],E=c?[c,o]:[o],P=w?e.compute(En(E,2,x,o.dataType),{inputs:E,outputs:[2]})[0]:o,O=[t,I];p&&O.push(p);let R=e.compute(rc(e,t,I,p,h,d,_),{inputs:O,outputs:[-1]})[0];e.compute(tc(e,R,h.batchSize*h.numHeads*h.sequenceLength,v),{inputs:[R],outputs:[]});let L=[R,P];e.compute(nc(e,R,P,h,_),{inputs:L,outputs:[0]})},oc=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],o=t.sequenceLength,i=t.inputHiddenSize,u=t.headSize,a=12,c={x:Math.ceil(t.headSize/a),y:Math.ceil(t.sequenceLength/a),z:t.batchSize*t.numHeads},p=[e.inputs[0],e.inputs[1],e.inputs[2]],h=[{type:12,data:o},{type:12,data:i},{type:12,data:u},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],d=y=>{let w=j("output_q",p[0].dataType,r),_=j("output_k",p[0].dataType,r),v=j("output_v",p[0].dataType,r),S=U("input",p[0].dataType,p[0].dims),A=U("weight",p[1].dataType,p[1].dims),I=U("bias",p[2].dataType,p[2].dims),x=S.type.storage,E=[{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`\n const TILE_SIZE = ${a}u;\n var tileInput: array<${x}, ${a*a}>;\n var tileWeightQ: array<${x}, ${a*a}>;\n var tileWeightK: array<${x}, ${a*a}>;\n var tileWeightV: array<${x}, ${a*a}>;\n ${y.registerUniforms(E).declareVariables(S,A,I,w,_,v)}\n ${y.mainStart([a,a,1])}\n let batchIndex = workgroup_id.z / uniforms.num_heads;\n let headNumber = workgroup_id.z % uniforms.num_heads;\n let m = global_id.y;\n let n = global_id.x;\n\n let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K;\n let biasOffsetQ = headNumber * uniforms.head_size;\n let biasOffsetK = uniforms.hidden_size + biasOffsetQ;\n let biasOffsetV = uniforms.hidden_size + biasOffsetK;\n\n var valueQ = ${x}(0);\n var valueK = ${x}(0);\n var valueV = ${x}(0);\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (m < uniforms.M && w + local_id.x < uniforms.K) {\n tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x];\n }\n if (n < uniforms.N && w + local_id.y < uniforms.K) {\n let offset = n + (w + local_id.y) * uniforms.ldb;\n tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset];\n tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset];\n tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset];\n }\n workgroupBarrier();\n for (var k: u32 = 0u; k({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:c,programUniforms:h}),getShaderSource:d},{inputs:p,outputs:[-1,-1,-1]})},Xa=(e,t)=>{let r=ec(e.inputs,t),[o,i,u]=oc(e,r);return Pn(e,o,i,u,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t)}});var ic,ac,sc,Qa,Ja=Y(()=>{"use strict";$r();ye();Se();Ze();_e();ic=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(o,i,u)=>{let a=i.length;if(a!==o.length)throw new Error(`${u}: num dimensions != ${a}`);i.forEach((c,p)=>{if(c!==o[p])throw new Error(`${u}: dim[${p}] do not match`)})};if(e[0].dims.length>1){let o=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,o,"Invalid input scale"),r(e[2].dims,o,"Invalid input B"),r(e[3].dims,o,"Invalid input mean"),r(e[4].dims,o,"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")},ac=(e,t)=>{let{epsilon:r,spatial:o,format:i}=t,u=e[0].dims,a=o?Me(u[u.length-1]):1,c=i==="NHWC"&&u.length>1?a:1,p=M.size(u)/a,h=o,d=h?u.length:u,y=U("x",e[0].dataType,e[0].dims,a),w=U("scale",e[1].dataType,e[1].dims,c),_=U("bias",e[2].dataType,e[2].dims,c),v=U("inputMean",e[3].dataType,e[3].dims,c),S=U("inputVar",e[4].dataType,e[4].dims,c),A=j("y",e[0].dataType,d,a),I=()=>{let E="";if(o)E=`let cOffset = ${u.length===1?"0u":i==="NHWC"?`outputIndices[${u.length-1}] / ${a}`:"outputIndices[1]"};`;else if(i==="NCHW")E=`\n ${A.indicesSet("outputIndices","0","0")}\n let cOffset = ${A.indicesToOffset("outputIndices")};`;else{E=`var cIndices = ${w.type.indices}(0);\n cIndices[0] = outputIndices[${u.length-1}];`;for(let P=1;P`\n const epsilon = ${r};\n ${E.registerUniform("outputSize","u32").declareVariables(y,w,_,v,S,A)}\n ${E.mainStart()}\n ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var outputIndices = ${A.offsetToIndices(`global_idx * ${a}`)};\n ${I()}\n let scale = ${w.getByOffset("cOffset")};\n let bias = ${_.getByOffset("cOffset")};\n let inputMean = ${v.getByOffset("cOffset")};\n let inputVar = ${S.getByOffset("cOffset")};\n let x = ${y.getByOffset("global_idx")};\n let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;\n ${A.setByOffset("global_idx","value")}\n }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${o}_${a}`,inputDependencies:h?["rank","type","type","type","type"]:void 0},getShaderSource:x,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h?[{type:12,data:p},...Z(u)]:[{type:12,data:p}]})}},sc=e=>ve(e),Qa=(e,t)=>{let{inputs:r,outputCount:o}=e,i=sc({...t,outputCount:o});if(vr.webgpu.validateInputContent&&ic(r,i),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(ac(r,i))}});var uc,dc,es,ts=Y(()=>{"use strict";Se();_e();uc=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")},dc=e=>{let t=e[0].dims,r=e[0].dims[2],o=M.size(t)/4,i=e[0].dataType,u=U("input",i,t,4),a=U("bias",i,[r],4),c=U("residual",i,t,4),p=j("output",i,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(o/64)}}),getShaderSource:d=>`\n const channels = ${r}u / 4;\n ${d.declareVariables(u,a,c,p)}\n\n ${d.mainStart()}\n ${d.guardAgainstOutOfBoundsWorkgroupSizes(o)}\n let value = ${u.getByOffset("global_idx")}\n + ${a.getByOffset("global_idx % channels")} + ${c.getByOffset("global_idx")};\n ${p.setByOffset("global_idx","value")}\n }`}},es=e=>{uc(e.inputs),e.compute(dc(e.inputs))}});var lc,ke,rs,ns,os,is,as,ss,us,ds,ls,cc,cs,ps,ms,fs,kn,hs,On,gs,ys,bs,ws,vs,$s,_s,Ss,xs,Cs,As,Is,Ts,Es,Ps,ks,Os,Rs,Bo,Do,Bs,Ds,zs,Rn=Y(()=>{"use strict";ye();Se();Ze();_e();lc=(e,t,r,o,i,u)=>{let a=Math.ceil(t/4),c="";typeof i=="string"?c=`${i}(a)`:c=i("a");let p=U("inputData",r,[a],4),h=j("outputData",o,[a],4);return`\n ${e.registerUniform("vec_size","u32").declareVariables(p,h)}\n\n ${u??""}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n\n let a = ${p.getByOffset("global_idx")};\n ${h.setByOffset("global_idx",c)}\n }`},ke=(e,t,r,o,i,u=e.dataType)=>({name:t,shaderCache:{hint:i,inputDependencies:["type"]},getShaderSource:a=>lc(a,M.size(e.dims),e.dataType,u,r,o),getRunData:a=>({outputs:[{dims:e.dims,dataType:u}],dispatchGroup:{x:Math.ceil(M.size(a[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(M.size(e.dims)/4)}]})}),rs=e=>{e.compute(ke(e.inputs[0],"Abs","abs"))},ns=e=>{e.compute(ke(e.inputs[0],"Acos","acos"))},os=e=>{e.compute(ke(e.inputs[0],"Acosh","acosh"))},is=e=>{e.compute(ke(e.inputs[0],"Asin","asin"))},as=e=>{e.compute(ke(e.inputs[0],"Asinh","asinh"))},ss=e=>{e.compute(ke(e.inputs[0],"Atan","atan"))},us=e=>{e.compute(ke(e.inputs[0],"Atanh","atanh"))},ds=e=>ve(e),ls=(e,t)=>{let r;switch(t.to){case 10:r="vec4";break;case 1:r="vec4";break;case 12:r="vec4";break;case 6:r="vec4";break;case 9:r="vec4";break;default:throw new RangeError(`not supported type (specified in attribute \'to\' from \'Cast\' operator): ${t.to}`)}e.compute(ke(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},cc=e=>{let t=e.length>=2&&e[1].data!==0?e[1].getFloat32Array()[0]:xn,r=e.length>=3&&e[2].data!==0?e[2].getFloat32Array()[0]:Cn;return ve({min:t,max:r})},cs=(e,t)=>{let r=e.inputs.length===1?t:cc(e.inputs),o=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Clip",i=>`clamp(${i}, clip_min_, clip_max_)`,`\n const clip_min_: vec4<${o}> = vec4(${o}(${r.min}));\n const clip_max_: vec4<${o}> = vec4(${o}(${r.max}));\n`,r.cacheKey),{inputs:[0]})},ps=e=>{e.compute(ke(e.inputs[0],"Ceil","ceil"))},ms=e=>{e.compute(ke(e.inputs[0],"Cos","cos"))},fs=e=>{e.compute(ke(e.inputs[0],"Cosh","cosh"))},kn=e=>ve(e),hs=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Elu",o=>`elu_vf32(${o})`,`\n const elu_alpha_ = ${r}(${t.alpha});\n\n fn elu_f32(a: ${r}) -> ${r} {\n return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);\n }\n\n fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {\n return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));\n }`,t.cacheKey))},On=(e="f32")=>`\nconst r0: ${e} = 0.3275911;\nconst r1: ${e} = 0.254829592;\nconst r2: ${e} = -0.284496736;\nconst r3: ${e} = 1.421413741;\nconst r4: ${e} = -1.453152027;\nconst r5: ${e} = 1.061405429;\n\nfn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {\n let absv = abs(v);\n let x = 1.0 / (1.0 + r0 * absv);\n return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));\n}`,gs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,On(t)))},ys=e=>{e.compute(ke(e.inputs[0],"Exp","exp"))},bs=e=>{e.compute(ke(e.inputs[0],"Floor","floor"))},ws=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,On(t)))},vs=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"LeakyRelu",o=>`select(leaky_relu_alpha_ * ${o}, ${o}, ${o} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},$s=e=>{e.compute(ke(e.inputs[0],"Not",t=>`!${t}`))},_s=e=>{e.compute(ke(e.inputs[0],"Neg",t=>`-${t}`))},Ss=e=>{e.compute(ke(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},xs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},Cs=e=>{e.compute(ke(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},As=e=>ve(e),Is=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"HardSigmoid",o=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${o} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},Ts=e=>{e.compute(ke(e.inputs[0],"Sin","sin"))},Es=e=>{e.compute(ke(e.inputs[0],"Sinh","sinh"))},Ps=e=>{e.compute(ke(e.inputs[0],"Sqrt","sqrt"))},ks=e=>{e.compute(ke(e.inputs[0],"Tan","tan"))},Os=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Rs=e=>{e.compute(ke(e.inputs[0],"Tanh",Os))},Bo=(e="f32")=>`\nconst fast_gelu_a: ${e} = 0.5;\nconst fast_gelu_b: ${e} = 0.7978845608028654;\nconst fast_gelu_c: ${e} = 0.035677408136300125;\n\nfn tanh_v(v: vec4<${e}>) -> vec4<${e}> {\n return ${Os("v")};\n}\n`,Do=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,Bs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"FastGelu",Do,Bo(t),void 0,e.inputs[0].dataType))},Ds=(e,t)=>{let r=et(e.inputs[0].dataType);return e.compute(ke(e.inputs[0],"ThresholdedRelu",o=>`select(vec4<${r}>(0.0), ${o}, ${o} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},zs=e=>{e.compute(ke(e.inputs[0],"Log","log"))}});var pc,mc,Us,Vs=Y(()=>{"use strict";Se();_e();Rn();pc=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")},mc=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=U("input",e[0].dataType,e[0].dims,4),o=U("bias",e[0].dataType,[e[0].dims[2]],4),i=j("output",e[0].dataType,t,4),u=M.size(t)/4,a=De(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)}}),getShaderSource:p=>`\n const M_SQRT2 = sqrt(2.0);\n const halfChannels = ${e[0].dims[2]/4/2}u;\n\n ${p.declareVariables(r,o,i)}\n\n ${On(a)}\n\n ${p.mainStart()}\n ${p.guardAgainstOutOfBoundsWorkgroupSizes(u)}\n let biasIdx = global_idx % halfChannels;\n let batchIndex = global_idx / halfChannels;\n let inputOffset = biasIdx + batchIndex * halfChannels * 2;\n let valueLeft = input[inputOffset] + bias[biasIdx];\n let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];\n let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);\n\n ${i.setByOffset("global_idx","valueLeft * geluRight")}\n }`}},Us=e=>{pc(e.inputs),e.compute(mc(e.inputs))}});var fc,hc,Ot,Ws,Ns,Gs,Hs,Ls,Fs,qs,js,Ks,Ys,Zs=Y(()=>{"use strict";ye();Se();_e();fc=(e,t,r,o,i,u,a,c,p,h,d,y)=>{let w,_;typeof c=="string"?w=_=(x,E)=>`${c}((${x}),(${E}))`:typeof c=="function"?w=_=c:(w=c.scalar,_=c.vector);let v=j("outputData",d,o.length,4),S=U("aData",p,t.length,4),A=U("bData",h,r.length,4),I;if(i)if(u){let x=M.size(t)===1,E=M.size(r)===1,P=t.length>0&&t[t.length-1]%4===0,O=r.length>0&&r[r.length-1]%4===0;x||E?I=v.setByOffset("global_idx",_(x?`${S.type.value}(${S.getByOffset("0")}.x)`:S.getByOffset("global_idx"),E?`${A.type.value}(${A.getByOffset("0")}.x)`:A.getByOffset("global_idx"))):I=`\n let outputIndices = ${v.offsetToIndices("global_idx * 4u")};\n let offsetA = ${S.broadcastedIndicesToOffset("outputIndices",v)};\n let offsetB = ${A.broadcastedIndicesToOffset("outputIndices",v)};\n ${v.setByOffset("global_idx",_(a||P?S.getByOffset("offsetA / 4u"):`${S.type.value}(${S.getByOffset("offsetA / 4u")}[offsetA % 4u])`,a||O?A.getByOffset("offsetB / 4u"):`${A.type.value}(${A.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}\n `}else I=v.setByOffset("global_idx",_(S.getByOffset("global_idx"),A.getByOffset("global_idx")));else{if(!u)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let x=(E,P,O="")=>{let R=`aData[indexA${P}][componentA${P}]`,L=`bData[indexB${P}][componentB${P}]`;return`\n let outputIndices${P} = ${v.offsetToIndices(`global_idx * 4u + ${P}u`)};\n let offsetA${P} = ${S.broadcastedIndicesToOffset(`outputIndices${P}`,v)};\n let offsetB${P} = ${A.broadcastedIndicesToOffset(`outputIndices${P}`,v)};\n let indexA${P} = offsetA${P} / 4u;\n let indexB${P} = offsetB${P} / 4u;\n let componentA${P} = offsetA${P} % 4u;\n let componentB${P} = offsetB${P} % 4u;\n ${E}[${P}] = ${O}(${w(R,L)});\n `};d===9?I=`\n var data = vec4(0);\n ${x("data",0,"u32")}\n ${x("data",1,"u32")}\n ${x("data",2,"u32")}\n ${x("data",3,"u32")}\n outputData[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:I=`\n ${x("outputData[global_idx]",0)}\n ${x("outputData[global_idx]",1)}\n ${x("outputData[global_idx]",2)}\n ${x("outputData[global_idx]",3)}\n `}return`\n ${e.registerUniform("vec_size","u32").declareVariables(S,A,v)}\n\n ${y??""}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${I}\n }`},hc=(e,t,r,o,i,u,a=r.dataType)=>{let c=!M.areEqual(r.dims,o.dims),p=r.dims,h=M.size(r.dims),d=!1,y=!1,w=[c];if(c){let _=It.calcShape(r.dims,o.dims,!1);if(!_)throw new Error("Can\'t perform binary op on the given tensors");p=_,h=M.size(p);let v=M.size(r.dims)===1,S=M.size(o.dims)===1,A=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,I=o.dims.length>0&&o.dims[o.dims.length-1]%4===0;w.push(v),w.push(S),w.push(A),w.push(I);let x=1;for(let E=1;E_.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:_=>fc(_,r.dims,o.dims,p,d,c,y,i,r.dataType,o.dataType,a,u),getRunData:()=>({outputs:[{dims:p,dataType:a}],dispatchGroup:{x:Math.ceil(h/64/4)},programUniforms:[{type:12,data:Math.ceil(M.size(p)/4)},...Z(r.dims,o.dims,p)]})}},Ot=(e,t,r,o,i,u)=>{e.compute(hc(t,i??"",e.inputs[0],e.inputs[1],r,o,u))},Ws=e=>{Ot(e,"Add",(t,r)=>`${t}+${r}`)},Ns=e=>{Ot(e,"Div",(t,r)=>`${t}/${r}`)},Gs=e=>{Ot(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4(${t}==${r})`},void 0,void 0,9)},Hs=e=>{Ot(e,"Mul",(t,r)=>`${t}*${r}`)},Ls=e=>{let t=U("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;Ot(e,"Pow",{scalar:(o,i)=>`pow_custom(${o},${i})`,vector:(o,i)=>`pow_vector_custom(${o},${i})`},`\n fn pow_custom(a : ${t}, b : ${t}) -> ${t} {\n if (b == ${t}(0.0)) {\n return ${t}(1.0);\n } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) {\n return ${t}(pow(f32(a), f32(b))); // NaN\n }\n return select(sign(a), ${t}(1.0), round(f32(abs(b) % ${t}(2.0))) != 1.0) * ${t}(${t==="i32"?"round":""}(pow(f32(abs(a)), f32(b))));\n }\n fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> {\n // TODO: implement vectorized pow\n 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));\n }\n `)},Fs=e=>{Ot(e,"Sub",(t,r)=>`${t}-${r}`)},qs=e=>{Ot(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4(${t}>${r})`},void 0,void 0,9)},js=e=>{Ot(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4(${t}<${r})`},void 0,void 0,9)},Ks=e=>{Ot(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4(${t}>=${r})`},void 0,void 0,9)},Ys=e=>{Ot(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4(${t}<=${r})`},void 0,void 0,9)}});var St,xt,Ct,Bn,Ft=Y(()=>{"use strict";ye();Se();St=(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"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},xt=(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})},Ct=(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"})},Bn=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,o]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:o}}else if(t==="Clip"){let[r,o]=e?.activation_params||[xn,Cn];return{activation:t,clipMax:o,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}});var tt,Dn,zn=Y(()=>{"use strict";tt=(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.`)}},Dn=e=>`\n ${e?"value = value + getBiasByOutputCoords(coords);":""}\n `});var Mn,zo=Y(()=>{"use strict";Mn=e=>`\nfn getIndexFromCoords4D(coords : vec4, shape : vec4) -> i32 {\n return dot(coords, vec4(\n shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));\n}\nfn getOutputIndexFromCoords(coords : vec4) -> i32 {\n return dot(coords, vec4(\n i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));\n}\n`});var yc,bc,Hr,Xs,wc,Lr,vc,Un,Fr=Y(()=>{"use strict";ye();Se();_e();Ft();zn();yc=(e,t)=>e?`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n kStart + inputRow,\n globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});\n `:`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n globalRow + innerRow,\n kStart / innerElementSize + inputCol${t?", batchIndices":""});\n `,bc=(e,t)=>e?`\n let ACached0 = mm_Asub[k * innerElementSize][localRow];\n let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];\n let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];\n ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}\n for (var i = 0; i < rowPerThread; i = i + 1) {\n acc[i] = BCached0 * ACached0[i] + acc[i];\n acc[i] = BCached1 * ACached1[i] + acc[i];\n acc[i] = BCached2 * ACached2[i] + acc[i];\n ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}\n }`:`\n for (var i = 0; i < rowPerThread; i = i + 1) {\n let ACached = mm_Asub[tileRow + i][k];\n acc[i] = BCached0 * ACached.x + acc[i];\n acc[i] = BCached1 * ACached.y + acc[i];\n acc[i] = BCached2 * ACached.z + acc[i];\n ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}\n }`,Hr=(e,t,r="f32",o,i=!1,u=32,a=!1,c=32)=>{let p=t[1]*e[1],h=t[0]*e[0],d=i?p:u,y=i?u:p,w=d/t[0],_=u/t[1];if(!((i&&w===4&&e[1]===4||!i&&(w===3||w===4))&&d%t[0]===0&&u%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${i} is true, innerElementSize ${w} and workPerThread[1] ${e[1]} must be 4.\n Otherwise, innerElementSize ${w} must be 3 or 4.\n tileAWidth ${d} must be divisible by workgroupSize[0]${t[0]}. tileInner ${u} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`\nvar mm_Asub: array, ${d/w}>, ${y}>;\nvar mm_Bsub: array, ${h/e[0]}>, ${u}>;\n\nconst rowPerThread = ${e[1]};\nconst colPerThread = ${e[0]};\nconst innerElementSize = ${w};\nconst tileInner = ${u};\n\n@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})\nfn main(@builtin(local_invocation_id) localId : vec3,\n @builtin(global_invocation_id) globalId : vec3,\n @builtin(workgroup_id) workgroupId : vec3) {\n let localRow = i32(localId.y);\n let tileRow = localRow * rowPerThread;\n let tileCol = i32(localId.x);\n\n let globalRow =i32(globalId.y) * rowPerThread;\n let globalCol = i32(globalId.x);\n let batch = ${a?"0":"i32(globalId.z)"};\n ${o?`let batchIndices = ${o.offsetToIndices("u32(batch)")};`:""}\n let globalRowStart = i32(workgroupId.y) * ${p};\n\n let num_tiles = ${a?`${Math.ceil(c/u)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};\n var kStart = ${a?`i32(globalId.z) * ${c}`:"0"};\n\n var acc: array, rowPerThread>;\n\n // Loop over shared dimension.\n let tileRowB = localRow * ${_};\n for (var t = 0; t < num_tiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let inputRow = tileRow + innerRow;\n let inputCol = tileCol;\n ${yc(i,o)}\n }\n\n // Load one tile of B into local memory.\n for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {\n let inputRow = tileRowB + innerRow;\n let inputCol = tileCol;\n mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${o?", batchIndices":""});\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {\n let BCached0 = mm_Bsub[k * innerElementSize][tileCol];\n let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];\n let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];\n ${w===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}\n\n ${bc(i,w)}\n }\n\n workgroupBarrier();\n }\n\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);\n }\n}`},Xs=(e,t)=>e?`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n kStart + inputRow,\n globalRowStart + inputCol${t?", batchIndices":""});\n `:`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n globalRowStart + inputRow,\n kStart + inputCol${t?", batchIndices":""});\n `,wc=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",Lr=(e,t,r="f32",o,i=!1,u=32,a=!1,c=32,p=!1)=>{let h=e[1]*t[1],d=e[0]*t[0],y=i?h:u,w=i?u:h;if(!(w%t[1]===0&&y%t[0]===0&&u%t[1]===0))throw new Error(`tileAHight ${w} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${y} must be divisible by workgroupSize[0]${t[0]}, tileInner ${u} must be divisible by workgroupSize[1]${t[1]}`);let _=w/t[1],v=y/t[0],S=u/t[1],A=p?`\n let localRow = i32(localId.y);\n let localCol = i32(localId.x);\n let globalRowStart = i32(workgroupId.y) * ${h};\n let globalColStart = i32(workgroupId.x) * ${d};\n\n // Loop over shared dimension.\n for (var t = 0; t < num_tiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var inputRow = localRow; inputRow < ${w}; inputRow = inputRow + ${t[1]}) {\n for (var inputCol = localCol; inputCol < ${y}; inputCol = inputCol + ${t[0]}) {\n ${Xs(i,o)}\n }\n }\n // Load one tile of B into local memory.\n for (var inputRow = localRow; inputRow < ${u}; inputRow = inputRow + ${t[1]}) {\n for (var inputCol = localCol; inputCol < ${d}; inputCol = inputCol + ${t[0]}) {\n mm_Bsub[inputRow][inputCol] = mm_readB(batch,\n kStart + inputRow,\n globalColStart + inputCol${o?", batchIndices":""});\n }\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n var BCached : array<${r}, colPerThread>;\n for (var k = 0; k < tileInner; k = k + 1) {\n for (var inner = 0; inner < colPerThread; inner = inner + 1) {\n BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];\n }\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let ACached = ${i?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = acc[innerRow][innerCol] +\n ACached * BCached[innerCol];\n }\n }\n }\n workgroupBarrier();\n }\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let gRow = globalRowStart + localRow + innerRow * ${t[1]};\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n let gCol = globalColStart + localCol + innerCol * ${t[0]};\n mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);\n }\n }\n `:`\nlet tileRow = i32(localId.y) * rowPerThread;\nlet tileCol = i32(localId.x) * colPerThread;\n\nlet globalRow = i32(globalId.y) * rowPerThread;\nlet globalCol = i32(globalId.x) * colPerThread;\nlet globalRowStart = i32(workgroupId.y) * ${h};\n\nlet tileRowA = i32(localId.y) * ${_};\nlet tileColA = i32(localId.x) * ${v};\nlet tileRowB = i32(localId.y) * ${S};\n// Loop over shared dimension.\nfor (var t = 0; t < num_tiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < ${v}; innerCol = innerCol + 1) {\n let inputRow = tileRowA + innerRow;\n let inputCol = tileColA + innerCol;\n ${Xs(i,o)}\n }\n }\n\n // Load one tile of B into local memory.\n for (var innerRow = 0; innerRow < ${S}; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n let inputRow = tileRowB + innerRow;\n let inputCol = tileCol + innerCol;\n mm_Bsub[inputRow][inputCol] = mm_readB(batch,\n kStart + inputRow,\n globalCol + innerCol${o?", batchIndices":""});\n }\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n var BCached : array<${r}, colPerThread>;\n for (var k = 0; k < tileInner; k = k + 1) {\n for (var inner = 0; inner < colPerThread; inner = inner + 1) {\n BCached[inner] = mm_Bsub[k][tileCol + inner];\n }\n\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n ${wc(i)}\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];\n }\n }\n }\n\n workgroupBarrier();\n}\n\nfor (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n mm_write(batch, globalRow + innerRow, globalCol + innerCol,\n acc[innerRow][innerCol]);\n }\n}\n`;return`\n var mm_Asub : array, ${w}>;\n var mm_Bsub : array, ${u}>;\n const rowPerThread = ${e[1]};\n const colPerThread = ${e[0]};\n const tileInner = ${u};\n\n@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})\nfn main(@builtin(local_invocation_id) localId : vec3,\n @builtin(global_invocation_id) globalId : vec3,\n @builtin(workgroup_id) workgroupId : vec3) {\n let batch = ${a?"0":"i32(globalId.z)"};\n ${o?`let batchIndices = ${o.offsetToIndices("u32(batch)")};`:""}\n let num_tiles = ${a?`${Math.ceil(c/u)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};\n var kStart = ${a?`i32(globalId.z) * ${c}`:"0"};\n\n var acc : array, rowPerThread>;\n\n // Without this initialization strange values show up in acc.\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = 0.0;\n }\n }\n ${A}\n }\n`},vc=(e,t,r,o,i,u=!1)=>{let[a,c,p]=i,[h,d,y,w]=o,_=_r(a,p),v=_r(c,p),S=De(o[0].type.tensor),A=()=>{let E=d.rank,P=h.rank,O=`var aIndices: ${d.type.indices};`;for(let R=E-2-1,L=P-1;R>=0;R--,L--)O+=`\naIndices[${R}] = ${P>1?`batchIndices[${L}]`:"batchIndices"};`;return _.forEach(R=>{O+=`\naIndices[${R}] = 0;`}),O+=`\naIndices[${E-2}] = u32(row);\n aIndices[${E-1}] = u32(colIn);`,O},I=()=>{let E=y.rank,P=h.rank,O=`var bIndices: ${y.type.indices};`;for(let R=E-2-1,L=P-1;R>=0;R--,L--)O+=`\nbIndices[${R}] = ${P>1?`batchIndices[${L}]`:"batchIndices"};`;return v.forEach(R=>{O+=`\nbIndices[${R}] = 0;`}),O+=`\nbIndices[${E-2}] = u32(row);\n bIndices[${E-1}] = u32(colIn);`,O};return`\n fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${h.type.indices}) -> ${tt(e,S)} {\n var value = ${tt(e,S)}(0.0);\n let col = colIn * ${e};\n if(row < uniforms.dim_a_outer && col < uniforms.dim_inner)\n {\n ${A()}\n value = ${d.getByIndices("aIndices")};\n }\n return value;\n }\n\n fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${h.type.indices}) -> ${tt(e,S)} {\n var value = ${tt(e,S)}(0.0);\n let col = colIn * ${e};\n if(row < uniforms.dim_inner && col < uniforms.dim_b_outer)\n {\n ${I()}\n value = ${y.getByIndices("bIndices")};\n }\n return value;\n }\n\n fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${tt(e,S)}) {\n let col = colIn * ${e};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {\n var value = valueIn;\n let coords = vec3(batch, row, colIn);\n ${t?`value = value + ${u?"bias[colIn]":`${tt(e,S)}(bias[row])`};`:""}\n ${r}\n ${w.setByIndices("vec3(coords)","value")}\n }\n }\n `},Un=(e,t,r,o,i=!1)=>{let u=e[0].dims,a=e[1].dims,c=u.slice(0,-2),p=a.slice(0,-2),h=o?o.slice(0,-2):r.slice(0,-2),d=M.size(h),y=u[u.length-2],w=u[u.length-1],_=a[a.length-1],v=w%4===0&&_%4===0,S=y<=8?[4,1,1]:[4,4,1],A=[8,8,1],I=[Math.ceil(_/A[0]/S[0]),Math.ceil(y/A[1]/S[1]),Math.ceil(d/A[2]/S[2])],x=v?4:1,E=[...c,y,w/x],P=E.length,O=[...p,w,_/x],R=O.length,L=[d,y,_/x],N=[{type:6,data:y},{type:6,data:_},{type:6,data:w}];xt(t,N),N.push(...Z(h,E,O));let K=["rank","rank"],Q=e.length>2;Q&&(N.push(...Z(e[2].dims)),K.push("rank")),N.push(...Z(L));let he=W=>{let se=h.length,Ce=An("batchDims",e[0].dataType,se,1),We=De(e[0].dataType),ee=U("a",e[0].dataType,P,x),ae=U("b",e[1].dataType,R,x),Ae=j("result",e[0].dataType,L.length,x),me=[ee,ae];if(Q){let G=i?x:1;me.push(U("bias",e[2].dataType,e[2].dims.length,G))}let ie=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];Ct(t,ie);let ue=De(Ae.type.tensor),le=St(t,Ae.type.value,ue),qe=vc(x,Q,le,[Ce,ee,ae,Ae],[c,p,h],i);return`\n ${W.registerUniforms(ie).registerInternalVariables(Ce).declareVariables(...me,Ae)}\n ${qe}\n ${v?Hr(S,A,We,Ce):Lr(S,A,We,Ce)}\n `};return{name:"MatMul",shaderCache:{hint:`${S};${t.activation};${v};${i}`,inputDependencies:K},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:I[0],y:I[1],z:I[2]},programUniforms:N}),getShaderSource:he}}});var $c,Qs,Js=Y(()=>{"use strict";ye();Lt();_e();Ft();zn();zo();Fr();$c=(e,t,r,o,i=!1,u,a=4,c=4,p=4,h="f32")=>{let d=Q=>{switch(Q){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${h}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${Q} is not supported.`)}},y=Q=>{switch(Q){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 ${Q} is not supported.`)}},w=e?`\n let coord = vec4(batch, xRow, xCol, xCh);\n `:`\n let coord = vec4(batch, xCh, xRow, xCol);\n `,_=e?`\n let coords = vec4(\n batch,\n row / outWidth,\n row % outWidth,\n col);\n `:`\n let coords = vec4(\n batch,\n row,\n col / outWidth,\n col % outWidth);\n `,v=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",S=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",A=e?"row":"col",I=e?"col":"row",x=`\n let inChannels = i32(uniforms.w_shape[2]);\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n let outRow = ${A} / outWidth;\n let outCol = ${A} % outWidth;\n\n let WRow = ${I} / (i32(uniforms.w_shape[1]) * inChannels);\n let WCol = ${I} / inChannels % i32(uniforms.w_shape[1]);\n let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0];\n let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1];\n let xCh = ${I} % inChannels;\n var resData = ${tt(a,h)}(0.0);\n // The bounds checking is always needed since we use it to pad zero for\n // the \'same\' padding type.\n if (xRow >= 0 && xRow < ${v} && xCol >= 0 && xCol < ${S}) {\n ${w}\n let xIndex = getIndexFromCoords4D(coord, vec4(uniforms.x_shape));\n ${d(a)}\n }\n return resData;`,E=e?t&&o?`\n let col = colIn * ${a};\n ${x}`:`\n let col = colIn * ${a};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {\n ${x}\n }\n return ${tt(a,h)}(0.0);`:o&&r?`\n let col = colIn * ${a};\n ${x}`:`\n let col = colIn * ${a};\n if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {\n ${x}\n }\n return ${tt(a,h)}(0.0);`,P=`${y(c)}`,O=tt(p,h),R=e?tt(a,h):tt(c,h),L=e?tt(c,h):tt(a,h),N=St(u,O,h);return`\n fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${R} {\n ${e?E:P}\n }\n\n fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${L} {\n ${e?P:E}\n }\n\n fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${O}) {\n let col = colIn * ${p};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer)\n {\n var value = valueIn;\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n ${_}\n ${Dn(i)}\n ${N}\n setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);\n }\n }`},Qs=(e,t,r,o,i,u,a,c)=>{let p=t.format==="NHWC",h=p?e[0].dims[3]:e[0].dims[1],d=r[0],y=p?r[2]:r[3],w=p?r[1]:r[2],_=p?r[3]:r[1],v=p&&(h%4===0||h%3===0)&&_%4===0,S=p?_:y*w,A=p?y*w:_,I=[8,8,1],x=o<=8?[4,1,1]:[4,4,1],E=[Math.ceil(S/I[0]/x[0]),Math.ceil(A/I[1]/x[1]),Math.ceil(d/I[2]/x[2])];Ve("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${E}`);let P=v?p&&h%4!==0?3:4:1,O=I[1]*x[1],R=I[0]*x[0],L=Math.max(I[0]*P,I[1]),N=o%O===0,K=i%R===0,Q=u%L===0,he=v?[P,4,4]:[1,1,1],W=[{type:6,data:o},{type:6,data:i},{type:6,data:u},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];xt(t,W),W.push(...Z(e[0].dims,e[1].dims));let se=["rank","rank"];a&&(W.push(...Z(e[2].dims)),se.push("rank")),W.push(...Z(r));let Ce=We=>{let ee=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"pad",type:"i32",length:2},{name:"stride",type:"i32",length:2},{name:"dilation",type:"i32",length:2}];Ct(t,ee);let ae=v?4:1,Ae=De(e[0].dataType),me=`\n fn setOutputAtIndex(flatIndex : i32, value : ${v?`vec4<${Ae}>`:Ae}) {\n result[flatIndex] = ${v?`vec4<${Ae}>`:Ae}(value);\n }\n fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${v?`vec4<${Ae}>`:Ae}) {\n let flatIndex = getOutputIndexFromCoords(vec4(d0, d1, d2, d3));\n setOutputAtIndex(flatIndex ${v?"/ 4":""}, value);\n }`,ie=U("x",e[0].dataType,e[0].dims.length,P===3?1:P),ue=U("w",e[1].dataType,e[1].dims.length,ae),le=[ie,ue],qe=j("result",e[0].dataType,r.length,ae);if(a){let G=U("bias",e[2].dataType,e[2].dims.length,ae);le.push(G),me+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${v?`vec4<${Ae}>`:Ae} {\n return bias[coords.${p?"w":"y"}${v?"/ 4":""}];\n }`}return`\n ${Mn("uniforms.result_strides")}\n //struct Uniforms { xShape : vec4, wShape : vec4, outShape : vec4,\n // outShapeStrides: vec3, filterDims : vec2, pad : vec2, stride : vec2,\n // dilation : vec2, dimAOuter : i32, dimBOuter : i32, dimInner : i32 };\n ${We.registerUniforms(ee).declareVariables(...le,qe)}\n ${me}\n ${$c(p,N,K,Q,a,t,he[0],he[1],he[2],Ae)}\n ${v?Hr(x,I,Ae,void 0,!p,L):Lr(x,I,Ae,void 0,!p,L,!1,void 0,c)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${P};${v};${N};${K};${Q};${O};${R};${L}`,inputDependencies:se},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:W}),getShaderSource:Ce}}});var Mo,eu,tu=Y(()=>{"use strict";ye();Se();_e();Uo();Ft();Mo=(e,t,r)=>{let o=e.length>2,i=o?"value += b[output_channel];":"",u=e[0].dims,a=e[1].dims,c=a[0]/t.group,p=t.format==="NHWC",h=Vn(u,a,t.dilations,t.pads,t.strides,p),d=M.size(h),y=[{type:12,data:d},{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:c}];xt(t,y),y.push(...Z(u,a));let w=["rank","rank"];o&&(y.push(...Z(e[2].dims)),w.push("rank")),y.push(...Z(h));let _=v=>{let S=j("output",e[0].dataType,h.length),A=De(S.type.tensor),I=St(t,S.type.value,A),x=U("x",e[0].dataType,u.length),E=U("w",e[1].dataType,a.length),P=[x,E];o&&P.push(U("b",e[2].dataType,e[2].dims.length));let O=[{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 Ct(t,O),`\n ${v.registerUniforms(O).declareVariables(...P,S)}\n\n ${v.mainStart()}\n ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let outputIndices = ${S.offsetToIndices("global_idx")};\n let batch: u32 = outputIndices[0];\n let output_channel: u32 = outputIndices[${p?3:1}];\n let xRCCorner: vec2 = vec2(outputIndices[${p?1:2}], outputIndices[${p?2:3}]) * uniforms.strides - uniforms.pads;\n let group_id: u32 = output_channel / uniforms.output_channels_per_group;\n\n var value: ${S.type.value} = ${S.type.value}(0);\n for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {\n let input_channel = group_id * uniforms.w_shape[1] + wInChannel;\n for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) {\n let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];\n\n if (xHeight < 0u || xHeight >= uniforms.x_shape[${p?1:2}]) {\n continue;\n }\n\n for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) {\n let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];\n if (xWidth < 0u || xWidth >= uniforms.x_shape[${p?2:3}]) {\n continue;\n }\n\n let xVal = ${p?x.get("batch","xHeight","xWidth","input_channel"):x.get("batch","input_channel","xHeight","xWidth")};\n let wVal = ${E.get("output_channel","wInChannel","wHeight","wWidth")};\n value += xVal*wVal;\n }\n }\n }\n ${i}\n ${I}\n ${S.setByOffset("global_idx","value")}\n }`};return{name:"GroupedConv",shaderCache:{hint:t.cacheKey,inputDependencies:w},getRunData:()=>({outputs:[{dims:r?r(h):h,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:y}),getShaderSource:_}},eu=(e,t,r)=>{let o=e.length>2,i=Me(r[3]),u=Me(r[2]),a=M.size(r)/i/u,c=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/i],p=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/i],h=[r[0],r[1],r[2],r[3]/i],d=[{type:12,data:a},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];xt(t,d),d.push(...Z(c,p,h));let y=(u-1)*t.strides[1]+p[1],w=_=>{let v=j("output",e[0].dataType,h.length,i),S=De(v.type.tensor),A=St(t,v.type.value,S),I=U("x",e[0].dataType,c.length,i),x=U("w",e[1].dataType,p.length,i),E=[I,x];o&&E.push(U("b",e[2].dataType,e[2].dims,i));let P=o?"value += b[output_channel];":"",O=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Ct(t,O),`\n ${_.registerUniforms(O).declareVariables(...E,v)}\n ${_.mainStart()}\n ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n let width0 = uniforms.output_shape[3];\n let output_channel = global_idx % width0;\n var index1 = global_idx / width0;\n let width1 = uniforms.output_shape[2] / ${u}u;\n let col = (index1 % width1) * ${u}u;\n index1 = index1 / width1;\n let row = index1 % uniforms.output_shape[1];\n let batch = index1 / uniforms.output_shape[1];\n\n let x_corner = vec2(i32(row), i32(col)) * uniforms.strides - uniforms.pads;\n\n var x_vals: array<${I.type.value}, ${y}>;\n var values: array<${v.type.value}, ${u}>;\n let input_channel = output_channel;\n // Use constant instead of uniform can give better performance for w\'s height/width.\n for (var w_height: u32 = 0u; w_height < ${p[0]}; w_height++) {\n let x_height = x_corner.x + i32(w_height);\n if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) {\n for (var i = 0; i < ${y}; i++) {\n let x_width = x_corner.y + i;\n if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) {\n x_vals[i] = ${I.get("batch","u32(x_height)","u32(x_width)","input_channel")};\n } else {\n x_vals[i] = ${I.type.value}(0);\n }\n }\n for (var w_width: u32 = 0u; w_width < ${p[1]}; w_width++) {\n let w_val = ${x.get("w_height","w_width","0","output_channel")};\n for (var i = 0u; i < ${u}u; i++) {\n values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]);\n }\n }\n }\n }\n\n for (var i = 0u; i < ${u}u; i++) {\n var value = values[i];\n ${P}\n ${A}\n ${v.set("batch","row","col + i","output_channel","value")};\n }\n }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${i};${u};${y};${p[0]};${p[1]}`,inputDependencies:o?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d}),getShaderSource:w}}});var Vo,_c,ru,Wo=Y(()=>{"use strict";ye();Se();Fr();_e();Ft();Vo=(e,t,r,o,i=!1)=>{let u=e[0].dims,a=e[1].dims,c=u[u.length-2],p=a[a.length-1],h=u[u.length-1],d=Me(p),y=Me(h),w=Me(c),_=M.size(r)/d/w,v=e.length>2,S=o?o.slice(0,-2):r.slice(0,-2),I=[M.size(S),c,p],x=[{type:12,data:_},{type:12,data:c},{type:12,data:p},{type:12,data:h}];xt(t,x),x.push(...Z(S,u,a)),v&&x.push(...Z(e[2].dims)),x.push(...Z(I));let E=P=>{let O=An("batch_dims",e[0].dataType,S.length),R=U("a",e[0].dataType,u.length,y),L=U("b",e[1].dataType,a.length,d),N=j("output",e[0].dataType,I.length,d),K=De(N.type.tensor),Q=St(t,N.type.value,K),he=[R,L],W="";if(v){let ie=i?d:1;he.push(U("bias",e[2].dataType,e[2].dims.length,ie)),W=`${i?`value += bias[col / ${ie}];`:`value += ${N.type.value}(bias[row + i]);`}`}let se=u.slice(0,-2),Ce=a.slice(0,-2),We=_r(se,S),ee=_r(Ce,S),ae=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Ct(t,ae);let Ae=(ie,ue)=>{let le=ie.rank,qe=ie.name;if(le===2)return`var ${qe}_indices = ${ie.type.indices}(0u, 0u);`;let G=O.rank,ne=`var ${qe}_indices: ${ie.type.indices};`;for(let xe=le-2-1,Ke=G-1;xe>=0;xe--,Ke--)ne+=`\n${qe}_indices[${xe}] = ${G>1?`batch_indices[${Ke}]`:"batch_indices"};`;return ue.forEach(xe=>{ne+=`\n${qe}_indices[${xe}] = 0;`}),ne+=`${qe}_indices[${le-2}] = 0u;\n ${qe}_indices[${le-1}] = 0u;`,ne},me=()=>{let ie=`var a_data: ${R.type.value};`;for(let ue=0;ue;\n for (var k: u32 = 0u; k < uniforms.K; k = k + ${y}) {\n ${me()}\n }\n for (var i = 0u; i < ${w}u; i++) {\n var value = values[i];\n ${W}\n ${Q}\n let cur_indices = ${N.type.indices}(batch, row + i, col);\n let offset = ${N.indicesToOffset("cur_indices")};\n ${N.setByOffset(`offset / ${d}`,"value")};\n }\n }\n `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${d};${y};${w};${i}`,inputDependencies:v?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:x}),getShaderSource:E}},_c=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.")},ru=e=>{_c(e.inputs);let t=It.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],o=e.inputs[0].dims[e.inputs[0].dims.length-1];r<8&&o<8?e.compute(Vo(e.inputs,{activation:""},t)):e.compute(Un(e.inputs,{activation:""},t))}});var Vn,No,Sc,nu,Go,xc,Cc,Ho,Uo=Y(()=>{"use strict";Se();Js();Fr();tu();Ft();Wo();Sr();Vn=(e,t,r,o,i,u)=>{let a=e[0],c=e.slice(u?1:2,u?3:4),p=c.length,h=t[0],y=t.slice(2).map((v,S)=>v+(v-1)*(r[S]-1)),_=c.map((v,S)=>v+o[S]+o[S+p]).map((v,S)=>Math.floor((v-y[S]+i[S])/i[S]));return _.splice(0,0,a),_.splice(u?3:1,0,h),_},No=[2,3,1,0],Sc=(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 conv 1D and 2D");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],o=e[1].dims[1]*t.group;if(r!==o)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 i=e[0].dims.length-2;if(t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},nu=(e,t)=>{let r=e.kernelShape.slice();for(let u=2;u{let t=Bn(e),r=e.format,o=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],i=e.dilations,u=e.group,a=e.kernel_shape,c=e.pads,p=e.strides,h=e.w_is_const();return{autoPad:o,format:r,dilations:i,group:u,kernelShape:a,pads:c,strides:p,wIsConst:h,...t,cacheKey:`${e.format};${t.activation};`}},xc=(e,t,r)=>{let o=nu(r,t),i=r.format==="NHWC";if(r.group!==1){if(!e.adapterInfo.isArchitecture("ampere")&&i&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let L=Vn(t[0].dims,t[1].dims,r.dilations,o.pads,r.strides,i),N=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=N);let K=[t[0],N];t.length===3&&K.push(t[2]),e.compute(eu(K,o,L),{inputs:K})}else e.compute(Mo(t,o));return}let u=t.length===3,a=t[0].dims[i?1:2],c=t[0].dims[i?2:3],p=t[0].dims[i?3:1],h=t[1].dims[2],d=t[1].dims[3],y=Vn(t[0].dims,t[1].dims,r.dilations,o.pads,r.strides,i),w=y[i?1:2],_=y[i?2:3],v=y[i?3:1],S=i&&h===a&&d===c&&r.pads[0]===0&&r.pads[1]===0;if(S||h===1&&d===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 R=y[0],L,N,K,Q=[];if(i){let se=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=se),S){let Ce=a*c*p;L=t[0].reshape([1,R,Ce]),N=se.reshape([1,Ce,v]),K=[1,R,v]}else L=t[0].reshape([R,a*c,p]),N=se.reshape([1,p,v]),K=[R,w*_,v];Q.push(L),Q.push(N)}else L=t[0].reshape([R,p,a*c]),N=t[1].reshape([1,v,p]),K=[R,v,w*_],Q.push(N),Q.push(L);u&&Q.push(t[2]);let he=K[2],W=Q[0].dims[Q[0].dims.length-1];he<8&&W<8?e.compute(Vo(Q,o,y,K,i),{inputs:Q}):e.compute(Un(Q,o,y,K,i),{inputs:Q});return}let A=!0,I=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=I);let x=[t[0],I];u&&x.push(t[2]);let E=i?w*_:v,P=i?v:w*_,O=h*d*p;e.compute(Qs(x,o,y,E,P,O,u,A),{inputs:x})},Cc=(e,t)=>{let r=t.format==="NHWC",o=[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&&o.push(e.inputs[2]);let i=[0,t.pads[0],0,t.pads[1]],u=[1].concat(t.strides),a=[1].concat(t.dilations),c=[1].concat(t.kernelShape),p=nu({...t,pads:i,strides:u,dilations:a,kernelShape:c},o);e.compute(Mo(o,p,h=>r?[h[0],h[2],h[3]]:[]))},Ho=(e,t)=>{Sc(e.inputs,t),e.inputs[0].dims.length===3?Cc(e,t):xc(e,e.inputs,t)}});var Ac,ou,iu=Y(()=>{"use strict";ye();Lt();_e();Ft();zn();zo();Fr();Ac=(e,t=!1,r,o,i=4)=>{let u=I=>{switch(I){case 1:return"return w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];";case 4:return`\n let coord1 = vec4(coordX, coordY, col + 1, rowInner);\n let coord2 = vec4(coordX, coordY, col + 2, rowInner);\n let coord3 = vec4(coordX, coordY, col + 3, rowInner);\n let v0 = w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];\n let v1 = w[getIndexFromCoords4D(coord1, vec4(uniforms.w_shape))];\n let v2 = w[getIndexFromCoords4D(coord2, vec4(uniforms.w_shape))];\n let v3 = w[getIndexFromCoords4D(coord3, vec4(uniforms.w_shape))];\n return ${o}(v0, v1, v2, v3);\n `;default:throw new Error(`innerElementSize ${I} is not supported.`)}},a=e?`\n let coord = vec4(batch, iXR, iXC, xCh);\n `:`\n let coord = vec4(batch, xCh, iXR, iXC);\n `,c=e?`\n let coords = vec4(\n batch,\n row / outWidth,\n row % outWidth,\n col);\n `:`\n let coords = vec4(\n batch,\n row,\n col / outWidth,\n col % outWidth);\n `,p=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",h=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",d=e?"row":"col",y=e?"col":"row",w=`\n let inChannels = ${e?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"};\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n let outRow = ${d} / outWidth;\n let outCol = ${d} % outWidth;\n\n let WRow = ${y} / (uniforms.filter_dims[1] * inChannels);\n let WCol = ${y} / inChannels % uniforms.filter_dims[1];\n let xR = f32(outRow - uniforms.pads[0] + uniforms.dilations[0] * WRow) / f32(uniforms.strides[0]);\n let xC = f32(outCol - uniforms.pads[1] + uniforms.dilations[1] * WCol) / f32(uniforms.strides[1]);\n if (xR < 0.0 || xR >= f32(${p}) || fract(xR) > 0.0) {\n return ${o}(0.0);\n }\n if (xC < 0.0 || xC >= f32(${h}) || fract(xC) > 0.0) {\n return ${o}(0.0);\n }\n let iXR = i32(xR);\n let iXC = i32(xC);\n let xCh = ${y} % inChannels;\n ${a}\n return x[getIndexFromCoords4D(coord, vec4(uniforms.x_shape))/${i}];`,_=e?`\n let col = colIn * ${i};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {\n ${w}\n }\n return ${o}(0.0);`:`\n let col = colIn * ${i};\n if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {\n ${w}\n }\n return ${o}(0.0);`,v=`\n let col = colIn * ${i};\n let inChannels = ${e?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"};\n let coordX = uniforms.filter_dims[0] - 1 - row / (uniforms.filter_dims[1] * inChannels);\n let coordY = uniforms.filter_dims[1] - 1 - (row / inChannels) % uniforms.filter_dims[1];\n if (${e?"row < uniforms.dim_inner && col < uniforms.dim_b_outer":"row < uniforms.dim_inner && col < uniforms.dim_a_outer"} && coordX >= 0 && coordY >= 0) {\n let rowInner = row % inChannels;\n let coord = vec4(coordX, coordY, col, rowInner);\n ${u(i)}\n }\n return ${o}(0.0);\n `,S=St(r,o);return`\n fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${o} {\n ${e?_:v}\n }\n\n fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${o} {\n ${e?v:_}\n }\n\n fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${o}) {\n let col = colIn * ${i};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {\n var value = valueInput;\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n ${c}\n ${Dn(t)}\n ${S}\n result[getIndexFromCoords4D(coords, vec4(uniforms.result_shape))/${i}] = value;\n }\n }`},ou=(e,t,r,o,i,u,a,c)=>{let p=t.format==="NHWC",h=p?e[0].dims[3]:e[0].dims[1],d=r[0],y=p?r[2]:r[3],w=p?r[1]:r[2],_=p?r[3]:r[1],v=p&&h%4===0&&h%3&&_%4===0,S=p?_:y*w,A=p?y*w:_,I=[8,8,1],x=o<=8?[4,1,1]:[4,4,1],E=[Math.ceil(S/I[0]/x[0]),Math.ceil(A/I[1]/x[1]),Math.ceil(d/I[2]/x[2])];Ve("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${E}`);let P=v?4:1,O=Math.max(I[0]*P,I[1]),R=v?4:1,L=[t.kernelShape[p?1:2],t.kernelShape[p?2:3]],N=[L[0]+(t.dilations[0]<=1?0:(L[0]-1)*(t.dilations[0]-1)),L[1]+(t.dilations[1]<=1?0:(L[1]-1)*(t.dilations[1]-1))],K=[N[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),N[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],Q=[{type:6,data:o},{type:6,data:i},{type:6,data:u},{type:6,data:t.strides},{type:6,data:t.dilations},{type:6,data:L},{type:6,data:K}];xt(t,Q),Q.push(...Z(e[0].dims,e[1].dims));let he=["rank","rank"];a&&(Q.push(...Z(e[2].dims)),he.push("rank")),Q.push(...Z(r));let W=se=>{let Ce=U("x",e[0].dataType,e[0].dims.length,R),We=U("w",e[1].dataType,e[1].dims.length,1),ee=j("result",e[0].dataType,r.length,R),ae=[Ce,We],Ae="";if(a){let ue=U("bias",e[2].dataType,e[2].dims.length,R);ae.push(ue),Ae+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${ue.type.value} {\n return bias[coords.${p?"w":"y"}${v?"/ 4":""}];\n }`}let me=[{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:L.length},{name:"pads",type:"i32",length:K.length}];Ct(t,me);let ie=De(e[0].dataType,1);if(ie!=="f16"&&ie!=="f32")throw new Error(`elemType ${ie} is not supported.`);return`\n ${Mn("uniforms.result_strides")}\n ${se.registerUniforms(me).declareVariables(...ae,ee)};\n ${Ae}\n ${Ac(p,a,t,Ce.type.value,P)}\n ${v?Hr(x,I,ie,void 0,!p,O):Lr(x,I,ie,void 0,!p,O,!1,void 0,c)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${t.cacheKey};${x};${I};${v}`,inputDependencies:he},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:Q}),getShaderSource:W}}});var Ic,Lo,au=Y(()=>{"use strict";ye();Lt();Se();_e();Ic=(e,t,r,o,i,u=!1,a,c,p=!1)=>{let h=p?1:2,d=p?2:3,y=p?3:1,w=u?2:1,_=`\n fn setOutputAtIndex(flatIndex : u32, value : ${u?`vec4<${a}>`:a}) {\n result[flatIndex] = ${u?`vec4<${a}>`:a}(value);\n }`;o&&(_+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${u?`vec4<${a}>`:a} {\n return bias[coords.${p?"w":"y"}${u?"/ 4":""}];\n }`);let v=u?4:1,S=U("W",t[1].dataType,t[1].dims.length,v),A=U("Dy",t[0].dataType,t[0].dims.length,v),I=[A,S];o&&I.push(U("bias",t[2].dataType,[r[y]].length,v));let x=j("result",t[0].dataType,r.length,v),E=`{\n let batch: u32 = ${i?"global_id.z":"workgroup_id.z"} / uniforms.result_shape[1];\n let r = ${i?"global_id.z":"workgroup_id.z"} % uniforms.result_shape[1];\n let c = ${i?"global_id.y":"workgroup_id.y"} * ${w};\n let d1: u32 = ${i?"global_id.x":"workgroup_id.x"} * 4;\n\n let dyCorner = vec2(i32(r), i32(c)) - vec2(uniforms.pads);\n\n // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).\n // ? = to be determined. : = across all values in that axis.\n var dotProd: array, ${w}>;\n for (var i = 0; i < ${w}; i++) {\n dotProd[i] = vec4<${a}>(0.0);\n }\n for (var wR: u32 = 0; wR < uniforms.filter_dims[0]; wR = wR + 1) {\n var dyR = (${a}(dyCorner.x) + ${a}(wR)) / ${a}(uniforms.strides.x);\n let wRPerm = uniforms.filter_dims[0] - 1 - wR;\n if (dyR < 0.0 || dyR >= ${a}(uniforms.Dy_shape[1]) ||\n fract(dyR) > 0.0 || wRPerm < 0) {\n continue;\n }\n let idyR: u32 = u32(dyR);\n\n for (var wC: u32 = 0; wC < uniforms.filter_dims[1]; wC = wC + 1) {\n let dyC = (${a}(dyCorner.y) + ${a}(wC)) / ${a}(uniforms.strides.y);\n let dyC2 = (${a}(dyCorner.y) + 1.0 + ${a}(wC)) / ${a}(uniforms.strides.y);\n let wCPerm = uniforms.filter_dims[1] - 1 - wC;\n if (wCPerm < 0) {\n continue;\n }\n var bDyCVal = true;\n var bDyCVal2 = true;\n if (dyC < 0.0 || dyC >= ${a}(uniforms.Dy_shape[2]) ||\n fract(dyC) > 0.0) {\n bDyCVal = false;\n }\n if (dyC2 < 0.0 || dyC2 >= ${a}(uniforms.Dy_shape[2]) ||\n fract(dyC2) > 0.0) {\n bDyCVal2 = false;\n }\n\n let idyC: u32 = u32(dyC);\n let idyC2: u32 = u32(dyC2);\n if (bDyCVal && bDyCVal2) {\n let d2Length = uniforms.Dy_shape[3];\n for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${A.get("batch","idyR","idyC","d2")};\n let tmpval = vec4<${a}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[0] = dotProd[0] + tmpval;\n\n xValue = ${A.get("batch","idyR","idyC2","d2")};\n\n dotProd[1] = dotProd[1] + vec4<${a}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n }\n } else if (bDyCVal) {\n let d2Length = uniforms.Dy_shape[${y}];\n for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${A.get("batch","idyR","idyC","d2")};\n let tmpval = vec4<${a}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[0] = dotProd[0] + tmpval;\n }\n } else if (bDyCVal2) {\n let d2Length = uniforms.Dy_shape[3];\n for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${A.get("batch","idyR","idyC2","d2")};\n let tmpval = vec4<${a}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[1] = dotProd[1] + tmpval;\n }\n }\n }\n }\n\n for (var i: u32 = 0; i < ${w}; i = i + 1) {\n let value = dotProd[i] + ${o?"bias[c+i]":`vec4<${a}>(0.0)`};\n ${x.set("batch","r","c + i","d1","value")};\n }\n }`,P=`\n let outputIndices = ${x.offsetToIndices("global_idx")};\n let batch = ${x.indicesGet("outputIndices",0)};\n let d1 = ${x.indicesGet("outputIndices",y)};\n let r = ${x.indicesGet("outputIndices",h)};\n let c = ${x.indicesGet("outputIndices",d)};\n let dyCorner = vec2(i32(r), i32(c)) - uniforms.pads;\n let dyRCorner = dyCorner.x;\n let dyCCorner = dyCorner.y;\n let groupId = d1 / uniforms.output_channels_per_group;\n let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;\n // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).\n // ? = to be determined. : = across all values in that axis.\n var dotProd = ${a}(0.0);\n for (var wR: u32 = 0; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {\n if (wR % uniforms.dilations.x != 0) {\n continue;\n }\n let dyR = (${a}(dyRCorner) + ${a}(wR)) / ${a}(uniforms.strides[0]);\n let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;\n if (dyR < 0.0 || dyR >= ${a}(uniforms.Dy_shape[${h}]) || fract(dyR) > 0.0 ||\n wRPerm < 0) {\n continue;\n }\n let idyR: u32 = u32(dyR);\n\n for (var wC: u32 = 0; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {\n if (wC % uniforms.dilations.y != 0) {\n continue;\n }\n let dyC = (${a}(dyCCorner) + ${a}(wC)) / ${a}(uniforms.strides.y);\n let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;\n if (dyC < 0.0 || dyC >= ${a}(uniforms.Dy_shape[${d}]) ||\n fract(dyC) > 0.0 || wCPerm < 0) {\n continue;\n }\n let idyC: u32 = u32(dyC);\n var inputChannel = groupId * uniforms.input_channels_per_group;\n for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + 1) {\n let xValue = ${p?A.get("batch","idyR","idyC","inputChannel"):A.get("batch","inputChannel","idyR","idyC")};\n let wValue = ${S.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")};\n dotProd = dotProd + xValue * wValue;\n inputChannel = inputChannel + 1;\n }\n }\n }\n let value = dotProd + ${o?"bias[d1]":`${a}(0.0)`};\n ${x.setByOffset("global_idx","value")};\n `;return`\n ${e.registerUniforms(c).declareVariables(...I,x)}\n ${_}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};\n ${u?E:P}}`},Lo=(e,t,r)=>{let o=e.length>2,i=t.outputShape,u=M.size(i),a=[Math.ceil(u/64),1,1];Ve("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${a}`);let c=t.format==="NHWC",p=["rank","rank"],h=[t.strides[0],t.strides[1]],d=[t.kernelShape[c?1:2],t.kernelShape[c?2:3]],y=[t.dilations[0],t.dilations[1]],w=[d[0]+(t.dilations[0]<=1?0:(t.kernelShape[c?1:2]-1)*(t.dilations[0]-1)),d[1]+(t.dilations[1]<=1?0:(t.kernelShape[c?2:3]-1)*(t.dilations[1]-1))],_=[w[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),w[1]-1-Math.floor(t.pads[1]+t.pads[3])/2],v=!1,S=t.group,A=e[1].dims,I=A[0]/S,x=A[1],E=[{type:12,data:u},{type:12,data:h},{type:12,data:d},{type:12,data:y},{type:12,data:w},{type:6,data:_},{type:12,data:I},{type:12,data:x},...Z(e[0].dims,e[1].dims)];o&&(E.push(...Z(e[2].dims)),p.push("rank")),E.push(...Z(i));let P=a[1]===1&&a[2]===1,O=R=>{let L=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:h.length},{name:"filter_dims",type:"u32",length:d.length},{name:"dilations",type:"u32",length:d.length},{name:"effective_filter_dims",type:"u32",length:w.length},{name:"pads",type:"i32",length:_.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],N=De(e[0].dataType);return`${Ic(R,e,i,o,P,v,N,L,c)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};`,inputDependencies:p},getRunData:()=>({dispatchGroup:{x:a[0],y:a[1],z:a[2]},outputs:[{dims:r?r(i):i,dataType:e[0].dataType}],programUniforms:E}),getShaderSource:O}}});var Tc,Ec,Pc,su,uu,kc,Oc,Rc,Bc,du,lu=Y(()=>{"use strict";iu();au();Ft();Sr();Tc=(e,t,r,o,i,u)=>(e-1)*t+r+(o-1)*i+1-u,Ec=(e,t,r,o,i)=>{let u=Math.floor(e/2);t==="SAME_UPPER"?(r[o]=u,r[i]=e-u):t==="SAME_LOWER"&&(r[o]=e-u,r[i]=u)},Pc=(e,t,r,o,i,u,a,c,p,h)=>{let d=e.length-2,y=h.length===0;if(p.length===0)for(let v=0;v{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((y,w)=>y*w,1)===0){r.length=0;for(let y=2;yy+w,0)===0){let y=t[0].dims.length-2;p=new Array(y).fill(1)}let h=e.strides.slice();if(h.reduce((y,w)=>y+w,0)===0){let y=t[0].dims.length-2;h=new Array(y).fill(1)}Pc(c,r,p,e.autoPad,e.group,i,h,o,a,u);let d=Object.assign({},e);return Object.assign(d,{kernelShape:r,pads:i,outputPadding:a,outputShape:u,dilations:p,strides:h}),d},uu=e=>{let t=Bn(e),r=e.format,o=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],i=e.dilations,u=e.group,a=e.kernelShape,c=e.pads,p=e.strides,h=e.wIsConst(),d=e.outputPadding,y=e.outputShape;return{autoPad:o,format:r,dilations:i,group:u,kernelShape:a,outputPadding:d,outputShape:y,pads:c,strides:p,wIsConst:h,...t,cacheKey:`${e.format};${t.activation};`}},kc=(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],o=e[1].dims[0];if(r!==o)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let i=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==i))throw new Error("invalid bias");let u=e[0].dims.length-2;if(t.dilations.reduce((d,y)=>d+y,0)>0&&t.dilations.length!==u)throw new Error(`dilations should be ${u}D`);if(t.strides.reduce((d,y)=>d+y,0)>0&&t.strides.length!==u)throw new Error(`strides should be ${u}D`);if(t.pads.reduce((d,y)=>d+y,0)>0&&t.pads.length!==u*2)throw new Error(`pads should be ${u*2}D`);if(t.outputPadding.length!==u&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${u}D`);if(t.kernelShape.reduce((d,y)=>d+y,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")},Oc=[2,3,1,0],Rc=(e,t,r)=>{let o=su(r,t),i=r.format==="NHWC",u=o.outputShape,a=u[i?3:1],c=t[0].dims[i?3:1];if(o.group!==1||a===1&&c===1){e.compute(Lo(t,o));return}let p=u[i?1:2],h=u[i?2:3],d=t[1].dims[2],y=t[1].dims[3],w=i?p*h:a,_=i?a:p*h,v=d*y*c,S=!0,A=e.kernelCustomData.wT??e.compute(yt(t[1],Oc),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=A);let I=[t[0],A],x=t.length===3;x&&(!i&&t[2].dims.length===1?I.push(t[2].reshape([t[2].dims[0],1,1])):I.push(t[2])),e.compute(ou(I,o,u,w,_,v,x,S),{inputs:I})},Bc=(e,t)=>{let r=t.format==="NHWC",o=[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&&o.push(e.inputs[2]);let i=t.kernelShape;(i.length===0||i[0]===0)&&(i=[e.inputs[1].dims[2]]);let u=t.dilations;(u.length===0||u[0]===0)&&(u=[1]);let a=t.strides;(a.length===0||a[0]===0)&&(a=[1]);let c=t.pads;c.length===0&&(c=[0,0]),c=[0,c[0],0,c[1]],a=[1].concat(a),u=[1].concat(u),i=[1].concat(i);let p=su({...t,pads:c,strides:a,dilations:u,kernelShape:i},o);e.compute(Lo(o,p,h=>r?[h[0],h[2],h[3]]:[h[0],h[1],h[3]]))},du=(e,t)=>{kc(e.inputs,t),e.inputs[0].dims.length===3?Bc(e,t):Rc(e,e.inputs,t)}});var Dc,cu,pu,mu=Y(()=>{"use strict";ye();Se();Ze();_e();Dc=(e,t,r,o)=>{let i=M.size(t),u=t.length,a=U("input",e,u),c=j("output",e,u),p=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),h=M.normalizeAxis(p,u),d=y=>{let w=` i32(${a.indicesGet("inputIndices","uniforms.axis")}) `,_=fe("uniforms.input_shape","uniforms.axis",u),v=o.reverse?w+(o.exclusive?" + 1":""):"0",S=o.reverse?_:w+(o.exclusive?"":" + 1");return`\n ${y.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(a,c)}\n ${y.mainStart()}\n ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var inputIndices = ${c.offsetToIndices("global_idx")};\n var sum = ${c.type.value}(0);\n let first : i32 = ${v};\n let last : i32 = ${S};\n for (var i : i32 = first; i < last; i++) {\n ${a.indicesSet("inputIndices","uniforms.axis","u32(i)")};\n sum = sum + ${a.getByIndices("inputIndices")};\n }\n ${c.setByOffset("global_idx","sum")};\n }`};return{name:"CumSum",shaderCache:{hint:o.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},{type:12,data:h},...Z(t,t)]}),getShaderSource:d}},cu=(e,t)=>{let r=e.inputs[0].dims,o=e.inputs[0].dataType,i=e.inputs[1];e.compute(Dc(o,r,i,t),{inputs:[0]})},pu=e=>{let t=e.exclusive===1,r=e.reverse===1;return ve({exclusive:t,reverse:r})}});var zc,Mc,Uc,fu,hu,gu=Y(()=>{"use strict";ye();Se();Ze();_e();zc=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.")},Mc=(e,t,r,o)=>{let i=[];i.push(`fn perm(i: ${o.type.indices}) -> ${r.type.indices} {\n var a: ${r.type.indices};`);for(let u=0;u{let r,o,i,u,a,c,p=t.format==="NHWC",h=t.blocksize,d=t.mode==="DCR";p?([r,o,i,u]=e.dims,a=d?[r,o,i,h,h,u/h**2]:[r,o,i,u/h**2,h,h],c=d?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,o,i,u]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],a=d?[r,h,h,u/h**2,o,i]:[r,u/h**2,h,h,o,i],c=d?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let y=e.reshape(a),w=y.dims.length,_=e.dataType,v=U("a",_,w),S=j("output",_,w),A=I=>`\n ${I.registerUniform("output_size","u32").declareVariables(v,S)}\n\n ${Mc(c,w,v,S)}\n\n ${I.mainStart()}\n ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let indices = ${S.offsetToIndices("global_idx")};\n let aIndices = perm(indices);\n\n ${S.setByOffset("global_idx",v.getByIndices("aIndices"))}\n }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:I=>{let x=p?[r,o*h,i*h,u/h**2]:[r,u/h**2,o*h,i*h],E=M.size(x),P=y.dims,O=M.sortBasedOnPerm(P,c);return{outputs:[{dims:x,dataType:I[0].dataType}],dispatchGroup:{x:Math.ceil(E/64)},programUniforms:[{type:12,data:E},...Z(P,O)]}},getShaderSource:A}},fu=(e,t)=>{zc(e.inputs),e.compute(Uc(e.inputs[0],t))},hu=e=>ve({blocksize:e.blocksize,mode:e.mode,format:e.format})});var Fo,Wn,yu,Vc,Wc,qo,jo,bu,Nc,wu,vu,$u=Y(()=>{"use strict";ye();Se();Ze();_e();Fo="[a-zA-Z]|\\\\.\\\\.\\\\.",Wn="("+Fo+")+",yu="^"+Wn+"$",Vc="("+Wn+",)*"+Wn,Wc="^"+Vc+"$",qo=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,r){let o=this.symbolToIndices.get(t);o===void 0?o=[r]:o.push(r),this.symbolToIndices.set(t,o)}},jo=class{constructor(t,r){this.equation=r;this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[o,i]=r.includes("->")?r.split("->",2):[r,""];if(!o.match(RegExp(Wc)))throw new Error("Invalid LHS term");if(o.split(",").forEach((c,p)=>{let h=t[p].dims.slice();if(!c.match(RegExp(yu)))throw new Error("Invalid LHS term");let d=this.processTerm(c,!0,h,p);this.lhs.push(d)}),i==="")i+=[...this.symbolToInfo.entries()].filter(([c,p])=>p.count===1||c==="...").map(([c])=>c).join("");else if(!i.match(RegExp(Wn)))throw new Error("Invalid RHS");i.match(RegExp(Fo,"g"))?.forEach(c=>{if(c==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let p=this.symbolToInfo.get(c);if(p===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(p.dimValue)}}),this.rhs=this.processTerm(i,!1,this.outputDims)}addSymbol(t,r,o){let i=this.symbolToInfo.get(t);if(i!==void 0){if(i.dimValue!==r&&i.count!==1)throw new Error("Dimension mismatch");i.count++,i.inputIndices.push(o)}else i={count:1,dimValue:r,inputIndices:[o]};this.symbolToInfo.set(t,i)}processTerm(t,r,o,i=-1){let u=o.length,a=!1,c=[],p=0;if(!t.match(RegExp(yu))&&!r&&t!=="")throw new Error("Invalid LHS term");let h=t.match(RegExp(Fo,"g")),d=new qo(i);return h?.forEach((y,w)=>{if(y==="..."){if(a)throw new Error("Only one ellipsis is allowed per input term");a=!0;let _=u-h.length+1;if(_<0)throw new Error("Ellipsis out of bounds");if(c=o.slice(p,p+_),this.hasEllipsis){if(this.ellipsisDims.length!==c.length||this.ellipsisDims.toString()!==c.toString())throw new Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=c;else throw new Error("Ellipsis must be specified in the LHS");for(let v=0;ve+"_max",Nc=(e,t,r,o)=>{let u=e.map(d=>d.length).map((d,y)=>U(`input${y}`,t,d)),a=M.size(o),c=j("output",t,o.length),p=[...r.symbolToInfo.keys()].filter(d=>!r.rhs.symbolToIndices.has(d)),h=d=>{let y=[],w="var prod = 1.0;",_="var sum = 0.0;",v="sum += prod;",S=[],A=[],I=[],x=[],E=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((O,R)=>{if(r.rhs.symbolToIndices.has(R)){let L=r.rhs.symbolToIndices.get(R)?.[0];L!==void 0&&r.lhs.forEach((N,K)=>{if(O.inputIndices.includes(K)){let Q=N.symbolToIndices.get(R);if(Q===void 0)throw new Error("Invalid symbol error");Q.forEach(he=>{y.push(`${u[K].indicesSet(`input${K}Indices`,he,c.indicesGet("outputIndices",L))}`)})}})}else r.lhs.forEach((L,N)=>{if(O.inputIndices.includes(N)){let K=L.symbolToIndices.get(R);if(K===void 0)throw new Error("Invalid symbol error");K.forEach(Q=>{S.push(`${u[N].indicesSet(`input${N}Indices`,Q,`${R}`)}`)}),x.push(`prod *= ${u[N].getByIndices(`input${N}Indices`)};`)}}),A.push(`for(var ${R}: u32 = 0; ${R} < uniforms.${bu(R)}; ${R}++) {`),I.push("}")});let P=E?[...y,`let sum = ${u.map((O,R)=>O.getByIndices(`input${R}Indices`)).join(" * ")};`]:[...y,_,...A,...S,w,...x,v,...I];return`\n ${d.registerUniforms(p.map(O=>({name:`${bu(O)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...u,c)}\n\n ${d.mainStart()}\n ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var outputIndices = ${c.offsetToIndices("global_idx")};\n ${u.map((O,R)=>`var input${R}Indices: ${u[R].type.indices};`).join(`\n`)}\n ${P.join(`\n`)};\n ${c.setByOffset("global_idx","sum")};\n }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let d=p.filter(w=>r.symbolToInfo.has(w)).map(w=>({type:12,data:r.symbolToInfo.get(w)?.dimValue||0}));d.push({type:12,data:a});let y=e.map((w,_)=>[...Z(w)]).reduce((w,_)=>w.concat(_),d);return y.push(...Z(o)),{outputs:[{dims:o,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:y}},getShaderSource:h}},wu=(e,t)=>{let r=new jo(e.inputs,t.equation),o=r.outputDims,i=e.inputs.map((u,a)=>u.dims);e.compute(Nc(i,e.inputs[0].dataType,r,o))},vu=e=>{let t=e.equation.replace(/\\s+/g,"");return ve({equation:t})}});var Gc,_u,Hc,Lc,Su,xu=Y(()=>{"use strict";ye();Se();_e();Gc=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),o=r.length{let r=e.length-t.length,o=[];for(let i=0;ie.length>t.length?_u(e,t):_u(t,e),Lc=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),o=Hc(t,r),i=e[0].dataType,u=i===9?4:1,a=Math.ceil(M.size(o)/u),c=h=>{let d=U("input",i,t.length,u),y=j("output",i,o.length,u),w;if(i===9){let _=(v,S,A="")=>`\n let outputIndices${S} = ${y.offsetToIndices(`outputOffset + ${S}u`)};\n let offset${S} = ${d.broadcastedIndicesToOffset(`outputIndices${S}`,y)};\n let index${S} = offset${S} / 4u;\n let component${S} = offset${S} % 4u;\n ${v}[${S}] = ${A}(${d.getByOffset(`index${S}`)}[component${S}]);\n `;w=`\n let outputOffset = global_idx * ${u};\n var data = vec4(0);\n ${_("data",0,"u32")}\n ${_("data",1,"u32")}\n ${_("data",2,"u32")}\n ${_("data",3,"u32")}\n ${y.setByOffset("global_idx","data")}\n }`}else w=`\n let outputIndices = ${y.offsetToIndices("global_idx")};\n let inputOffset = ${d.broadcastedIndicesToOffset("outputIndices",y)};\n ${y.setByOffset("global_idx",d.getByOffset("inputOffset"))}\n }`;return`\n ${h.registerUniform("vec_size","u32").declareVariables(d,y)}\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${w}`},p=[{type:12,data:a},...Z(t,o)];return{name:"Expand",shaderCache:{hint:`${o.length}`,inputDependencies:["rank"]},getShaderSource:c,getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p})}},Su=e=>{Gc(e.inputs),e.compute(Lc(e.inputs),{inputs:[0]})}});var Fc,Cu,Au=Y(()=>{"use strict";ye();Se();_e();Rn();Fc=e=>{let t=e[0].dataType,r=M.size(e[0].dims),o=M.size(e[1].dims),i=o%4===0,u=a=>{let c=U("x",t,[1],4),p=U("bias",t,[1],4),h=j("y",t,[1],4),d=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],y=_=>`\n let bias${_}_offset: u32 = (global_idx * 4 + ${_}) % uniforms.bias_size;\n let bias${_} = ${p.getByOffset(`bias${_}_offset / 4`)}[bias${_}_offset % 4];`,w=i?`\n let bias = ${p.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${y(0)}${y(1)}${y(2)}${y(3)}\n let bias = ${c.type.value}(bias0, bias1, bias2, bias3);`;return`${a.registerUniforms(d).declareVariables(c,p,h)}\n\n ${Bo(et(t))}\n\n ${a.mainStart(or)}\n ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}\n\n let x = ${c.getByOffset("global_idx")};\n ${w}\n let x_in = x + bias;\n ${h.setByOffset("global_idx",Do("x_in"))}\n }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${i}`,inputDependencies:["type","type"]},getShaderSource:u,getRunData:a=>({outputs:[{dims:a[0].dims,dataType:a[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:o}],dispatchGroup:{x:Math.ceil(r/or/4)}})}},Cu=e=>{e.inputs.length<2||M.size(e.inputs[1].dims)===0?Bs(e):e.compute(Fc(e.inputs))}});var qc,jc,Iu,Tu,Eu=Y(()=>{"use strict";ye();Se();Ze();_e();qc=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},jc=(e,t)=>{let r=e[0].dims,o=e[1].dims,i=r.length,u=M.normalizeAxis(t.axis,i),a=r.slice(0);a.splice(u,1,...o);let c=r[u],p=e[0].dataType===9?4:1,h=Math.ceil(M.size(a)/p),d=[{type:12,data:h},{type:6,data:c},{type:12,data:u},...Z(e[0].dims,e[1].dims,a)],y=w=>{let _=U("data",e[0].dataType,e[0].dims.length,p),v=U("inputIndices",e[1].dataType,e[1].dims.length),S=j("output",e[0].dataType,a.length,p),A=x=>{let E=o.length,P=`var indicesIndices${x} = ${v.type.indices}(0);`;for(let O=0;O1?`indicesIndices${x}[${O}]`:`indicesIndices${x}`} = ${a.length>1?`outputIndices${x}[uniforms.axis + ${O}]`:`outputIndices${x}`};`;P+=`\n var idx${x} = ${v.getByIndices(`indicesIndices${x}`)};\n if (idx${x} < 0) {\n idx${x} = idx${x} + uniforms.axisDimLimit;\n }\n var dataIndices${x} : ${_.type.indices};\n `;for(let O=0,R=0;O1?`dataIndices${x}[${O}]`:`dataIndices${x}`} = u32(idx${x});`,R+=E):(P+=`${i>1?`dataIndices${x}[${O}]`:`dataIndices${x}`} = ${a.length>1?`outputIndices${x}[${R}]`:`outputIndices${x}`};`,R++);return P},I;if(e[0].dataType===9){let x=(E,P,O="")=>`\n let outputIndices${P} = ${S.offsetToIndices(`outputOffset + ${P}u`)};\n ${A(P)};\n let offset${P} = ${_.indicesToOffset(`dataIndices${P}`)};\n let index${P} = offset${P} / 4u;\n let component${P} = offset${P} % 4u;\n ${E}[${P}] = ${O}(${_.getByOffset(`index${P}`)}[component${P}]);\n `;I=`\n let outputOffset = global_idx * ${p};\n var value = vec4(0);\n ${x("value",0,"u32")}\n ${x("value",1,"u32")}\n ${x("value",2,"u32")}\n ${x("value",3,"u32")}\n ${S.setByOffset("global_idx","value")}\n `}else I=`\n let outputIndices = ${S.offsetToIndices("global_idx")};\n ${A("")};\n let value = ${_.getByIndices("dataIndices")};\n ${S.setByOffset("global_idx","value")};\n `;return`\n ${w.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(_,v,S)}\n ${w.mainStart()}\n ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n ${I}\n }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:d}),getShaderSource:y}},Iu=e=>ve({axis:e.axis}),Tu=(e,t)=>{let r=e.inputs;qc(r),e.compute(jc(e.inputs,t))}});var Kc,Yc,Pu,ku,Ou=Y(()=>{"use strict";ye();Se();Ze();_e();Kc=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\n indices input tensors be of same rank.`)},Yc=(e,t)=>{let r=e[0].dims,o=e[0].dataType,i=r.length,u=e[1].dims,a=e[1].dataType,c=M.normalizeAxis(t.axis,i),p=r[c],h=u.slice(0),d=M.size(h),y=U("input",o,i),w=U("indicesInput",a,u.length),_=j("output",o,h.length),v=[{type:12,data:d},{type:6,data:p},{type:12,data:c}];return v.push(...Z(r,u,h)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:h,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:v}),getShaderSource:I=>`\n ${I.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(y,w,_)}\n ${I.mainStart()}\n ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n let outputIndices = ${_.offsetToIndices("global_idx")};\n\n var idx = ${w.getByOffset("global_idx")};\n if (idx < 0) {\n idx = idx + uniforms.axisDimLimit;\n }\n var inputIndices = ${y.type.indices}(outputIndices);\n ${y.indicesSet("inputIndices","uniforms.axis","u32(idx)")};\n let value = ${y.getByIndices("inputIndices")};\n\n ${_.setByOffset("global_idx","value")};\n }`}},Pu=e=>ve({axis:e.axis}),ku=(e,t)=>{let r=e.inputs;Kc(r),e.compute(Yc(e.inputs,t))}});var Zc,Xc,Ru,Bu,Du=Y(()=>{"use strict";ye();Se();_e();Zc=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")},Xc=(e,t)=>{let r=e[0].dims.slice(),o=e[1].dims.slice(),[i,u,a]=Sn.getShapeOfGemmResult(r,t.transA,o,t.transB,e.length===3?e[2].dims:void 0),c=[i,u];if(!c)throw new Error("Can\'t use gemm on the given tensors");let p=M.size(c),h=[{type:12,data:p},{type:12,data:i},{type:12,data:u},{type:12,data:a},{type:1,data:t.alpha},{type:1,data:t.beta}],d=["type","type"];e.length===3&&(h.push(...Z(e[2].dims)),d.push("rank")),h.push(...Z(c));let y=w=>{let _="";t.transA&&t.transB?_="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?_="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?_="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(_="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let v=t.alpha===1?"":"value *= uniforms.alpha;",S=U("a",e[0].dataType,e[0].dims),A=U("b",e[1].dataType,e[1].dims),I=S.type.value,x=null,E=[S,A];e.length===3&&(x=U("c",e[2].dataType,e[2].dims.length),E.push(x));let P=j("output",e[0].dataType,c.length);E.push(P);let O=[{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`\n ${w.registerUniforms(O).declareVariables(...E)}\n\n ${w.mainStart()}\n ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let m = global_idx / uniforms.N;\n let n = global_idx % uniforms.N;\n\n var value = ${I}(0);\n for (var k: u32 = 0u; k < uniforms.K; k++) {\n ${_}\n }\n\n ${v}\n ${(()=>x!=null?`let cOffset = ${x.broadcastedIndicesToOffset("vec2(m, n)",P)}; value += ${I}(uniforms.beta) * ${x.getByOffset("cOffset")};`:"")()}\n output[global_idx] = value;\n }`};return{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:c,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:y}},Ru=e=>{let t=e.transA,r=e.transB,o=e.alpha,i=e.beta;return{transA:t,transB:r,alpha:o,beta:i,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},Bu=(e,t)=>{Zc(e.inputs),e.compute(Xc(e.inputs,t))}});var Qc,Jc,ep,zu,Mu=Y(()=>{"use strict";ye();Se();_e();Qc=(e,t)=>{let r=e[0].dims,o=r,i=2,u=M.sizeToDimension(r,i),a=M.sizeFromDimension(r,i),c=Me(a),p=a/c,h=[r[0],r[1],p],d=["rank","type","type"],y=[{type:12,data:a},{type:12,data:p}];y.push(...Z(h,h));let w=_=>{let v=U("x",e[0].dataType,h.length,c),S=U("scale",e[1].dataType,e[1].dims),A=U("bias",e[2].dataType,e[2].dims),I=j("output",e[0].dataType,h.length,c),x=[v,S,A,I],E=v.type.value,P=c===1?"f32":`vec${c}`,O=64,R=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return`\n var meanShared : f32;\n var squaredNormShared : f32;\n var workgroupShared : array<${P}, ${O}>;\n const workgroupSize = ${O}u;\n ${_.registerUniforms(R).declareVariables(...x)}\n ${_.mainStart(O)}\n let norm = global_idx / workgroupSize;\n let batch = norm / uniforms.x_shape[1];\n let channel = norm % uniforms.x_shape[1];\n let localIndex = local_id.x;\n\n // initialize workgroup memory\n var initial = ${P}(0);\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n initial = initial + ${P}(${v.get("batch","channel","h")});\n }\n workgroupShared[localIndex] = initial;\n workgroupBarrier();\n\n // Calculate the mean of current channel data.\n for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) {\n if (localIndex < currSize) {\n workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];\n }\n workgroupBarrier();\n }\n if (localIndex == 0) {\n meanShared = ${_t("workgroupShared[0]",c)} / f32(uniforms.normSize);\n }\n workgroupBarrier();\n\n // reinitialize workgroup memory.\n initial = ${P}(0);\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n let deviation = ${P}(${v.get("batch","channel","h")}) - ${P}(meanShared);\n initial = initial + deviation * deviation;\n }\n workgroupShared[localIndex] = initial;\n workgroupBarrier();\n\n // Calculate the sum of square of deviation of current channel data.\n for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) {\n if (localIndex < currSize) {\n workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];\n }\n workgroupBarrier();\n }\n if (localIndex == 0) {\n squaredNormShared = ${_t("workgroupShared[0]",c)};\n }\n workgroupBarrier();\n\n let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${t.epsilon}));\n let channelScale = invStdDev * f32(${S.getByOffset("channel")});\n let channelShift = f32(${A.getByOffset("channel")}) - meanShared * channelScale;\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n let value = ${v.get("batch","channel","h")} * ${E}(${P}(channelScale)) + ${E}(${P}(channelShift));\n ${I.set("batch","channel","h","value")};\n }\n }`};return{name:"InstanceNormalization",shaderCache:{hint:`${t.epsilon};${c}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:u},programUniforms:y}),getShaderSource:w}},Jc=(e,t,r,o,i,u,a,c)=>{let p=Me(a),h=64,d=p===1?"vec2f":`mat2x${p}f`,y=p===1?"f32":`vec${p}f`,w=(R,L)=>`${d}(${R}, ${L})`,_=i*a/p,v=Math.ceil(u/h),S=["type"],A=[{type:12,data:v},{type:12,data:u},{type:12,data:Math.floor(a/p)},{type:12,data:Math.floor(u*a/p)}],I=R=>{let L=U("input",t.dataType,t.dims,p);return`\n ${R.declareVariables(L)}\n @group(0) @binding(1) var output : array<${d}>;\n struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32};\n @group(0) @binding(2) var uniforms: Uniforms;\n\n ${R.mainStart(h)}\n let currentImageNumber = global_idx / ${h} / uniforms.C;\n let currentChannelNumber = (global_idx / ${h}) % uniforms.C;\n let wgOffset = local_id.x * uniforms.wg_size;\n if (wgOffset >= uniforms.H) {\n return;\n }\n let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H);\n\n let offset = currentImageNumber * uniforms.image_size + currentChannelNumber;\n var sum = ${$t("f32",p)};\n var squaredSum = ${$t("f32",p)};\n for (var i: u32 = wgOffset; i < wgMax; i++) {\n let value = ${y}(input[offset + i * uniforms.C]);\n sum += value;\n squaredSum += value * value;\n }\n output[global_idx] = ${w("sum","squaredSum")};\n }`},x=e.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${p}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:[i,a,h,2],dataType:1}],dispatchGroup:{x:i*a/p},programUniforms:A}),getShaderSource:I},{inputs:[t],outputs:[-1]})[0],E=[{type:12,data:_},{type:12,data:u},{type:12,data:Math.floor(a/p)},{type:12,data:Math.floor(h*a/p)}],P=["type","type","type"],O=R=>{let L=U("scale",r.dataType,r.dims,p),N=U("bias",o.dataType,o.dims,p);return`\n @group(0) @binding(0) var input : array<${d}>;\n @group(0) @binding(1) var scale : array<${L.type.storage}>;\n @group(0) @binding(2) var bias : array<${N.type.storage}>;\n @group(0) @binding(3) var output : array<${d}>;\n struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32};\n @group(0) @binding(4) var uniforms: Uniforms;\n\n ${R.mainStart()}\n ${R.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.units_of_work")}\n let currentImageNumber = global_idx / uniforms.C;\n let currentChannelNumber = global_idx % uniforms.C;\n\n let offset = currentImageNumber * uniforms.image_size;\n var sum = ${$t("f32",p)};\n var squaredSum = ${$t("f32",p)};\n for (var i: u32 = 0; i < min(${h}, uniforms.H); i++) {\n let value = input[offset + i + currentChannelNumber * ${h}];\n sum += value[0];\n squaredSum += value[1];\n }\n sum = sum / f32(uniforms.H);\n squaredSum = squaredSum / f32(uniforms.H);\n let invStdDev = inverseSqrt(squaredSum - sum * sum + f32(${c}));\n let channelScale = invStdDev * ${y}(scale[currentChannelNumber]);\n let channelShift = ${y}(bias[currentChannelNumber]) - sum * channelScale;\n\n output[global_idx] = ${w("channelScale","channelShift")};\n }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${p};${c}`,inputDependencies:P},getRunData:()=>({outputs:[{dims:[i,a,2],dataType:1}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:E}),getShaderSource:O},{inputs:[x,r,o],outputs:[-1]})[0]},ep=(e,t,r)=>{let o=t[0].dims,i=o,u=o[0],a=o[o.length-1],c=M.sizeFromDimension(o,1)/a,p=Me(a),h=M.size(i)/p,d=[{type:12,data:c},{type:12,data:Math.floor(a/p)}],y=["type","type"],w=Jc(e,t[0],t[1],t[2],u,c,a,r.epsilon),_=v=>{let S=De(t[0].dataType),A=p===1?"vec2f":`mat2x${p}f`,I=p===1?S:`vec${p}<${S}>`,x=U("input",t[0].dataType,t[0].dims,p),E=j("output",t[0].dataType,i,p);return`\n @group(0) @binding(0) var input : array<${x.type.storage}>;\n @group(0) @binding(1) var scaleInput : array<${A}>;\n @group(0) @binding(2) var output : array<${E.type.storage}>;\n struct Uniforms {H: u32, C : u32};\n @group(0) @binding(3) var uniforms: Uniforms;\n\n ${v.mainStart()}\n let currentImageNumber = global_idx / (uniforms.C * uniforms.H);\n let currentChannelNumber = global_idx % uniforms.C;\n\n let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber;\n let scale = scaleInput[scaleOffset];\n output[global_idx] = fma(input[global_idx], ${I}(scale[0]), ${I}(scale[1]));\n }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${p}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:d}),getShaderSource:_},{inputs:[t[0],w]})},zu=(e,t)=>{t.format==="NHWC"?ep(e,e.inputs,t):e.compute(Qc(e.inputs,t))}});var tp,rp,Uu,Vu=Y(()=>{"use strict";ye();Se();_e();tp=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},rp=(e,t,r)=>{let o=t.simplified,i=e[0].dims,u=e[1],a=!o&&e[2],c=i,p=M.normalizeAxis(t.axis,i.length),h=M.sizeToDimension(i,p),d=M.sizeFromDimension(i,p),y=M.size(u.dims),w=a?M.size(a.dims):0;if(y!==d||a&&w!==d)throw new Error(`Size of X.shape()[axis:] == ${d}.\n Size of scale and bias (if provided) must match this.\n Got scale size of ${y} and bias size of ${w}`);let _=[];for(let O=0;O1,x=r>2,E=O=>{let R=De(e[0].dataType),L=[U("x",e[0].dataType,e[0].dims,v),U("scale",u.dataType,u.dims,v)];a&&L.push(U("bias",a.dataType,a.dims,v)),L.push(j("output",e[0].dataType,c,v)),I&&L.push(j("mean_data_output",1,_)),x&&L.push(j("inv_std_output",1,_));let N=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`\n ${O.registerUniforms(N).declareVariables(...L)}\n ${O.mainStart()}\n ${O.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}\n let offset = global_idx * uniforms.norm_size_vectorized;\n var mean_vector = ${$t("f32",v)};\n var mean_square_vector = ${$t("f32",v)};\n\n for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {\n let value = ${ir(R,v,"x[h + offset]")};\n mean_vector += value;\n mean_square_vector += value * value;\n }\n let mean = ${_t("mean_vector",v)} / uniforms.norm_size;\n let inv_std_dev = inverseSqrt(${_t("mean_square_vector",v)} / uniforms.norm_size ${o?"":"- mean * mean"} + uniforms.epsilon);\n\n for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {\n let f32input = ${ir(R,v,"x[j + offset]")};\n let f32scale = ${ir(R,v,"scale[j]")};\n output[j + offset] = ${L[0].type.value}((f32input ${o?"":"- mean"}) * inv_std_dev * f32scale\n ${a?`+ ${ir(R,v,"bias[j]")}`:""}\n );\n }\n\n ${I?"mean_data_output[global_idx] = mean":""};\n ${x?"inv_std_output[global_idx] = inv_std_dev":""};\n }`},P=[{dims:c,dataType:e[0].dataType}];return I&&P.push({dims:_,dataType:1}),x&&P.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${v};${r};${o}`,inputDependencies:S},getRunData:()=>({outputs:P,dispatchGroup:{x:Math.ceil(h/64)},programUniforms:A}),getShaderSource:E}},Uu=(e,t)=>{tp(e.inputs),e.compute(rp(e.inputs,t,e.outputCount))}});var np,op,Wu,Nu,Gu=Y(()=>{"use strict";ye();Se();Ze();_e();np=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],o=r.dims.length;if(r.dims[o-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let i=Math.floor((t.k+t.blockSize-1)/t.blockSize),u=t.blockSize/8*t.bits,a=e[1];if(!M.areEqual(a.dims,[t.n,i,u]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let p=e[2].dims;if(M.size(p)!==t.n*i)throw new Error("scales input size error.");if(e.length===4){let d=e[3].dims,y=t.bits>4?t.n*i:t.n*Math.floor((i+1)/2);if(M.size(d)!==y)throw new Error("zeroPoints input size error.")}},op=(e,t,r,o)=>{let i=e[0].dims,u=i.length,a=Math.floor((t.k+t.blockSize-1)/t.blockSize),c=i[u-2],p=t.k,h=t.n,d=i.slice(0,u-2),y=M.size(d),_=t.blockSize/8*t.bits/4,v=e[0].dataType,S=Me(c),A=Me(t.k),I=Me(_),x=tr(v),E=c*a*x,P=Math.floor(o/E),O=a<=r[0]&&P>0,R=!O||P>=4?Me(h):P>=2&&Me(h)>=2?2:1,L=d.concat([c,h]),N=M.size(L)/R/S,K=O?[]:[{type:12,data:N},{type:12,data:t.blockSize}],Q=[y,c,p/A],he=M.convertShape(e[1].dims).slice();he.splice(-1,1,_/I),K.push(...Z(Q)),K.push(...Z(he)),K.push(...Z(e[2].dims)),e.length===4&&K.push(...Z(M.convertShape(e[3].dims)));let W=[y,c,h/R];K.push(...Z(W));let se=Ce=>{let We=Q.length,ee=U("a",e[0].dataType,We,A),ae=U("b",12,he.length,I),Ae=U("scales",e[2].dataType,e[2].dims.length),me=[ee,ae,Ae],ie=e.length===4?U("zero_points",12,e[3].dims.length):void 0;ie&&me.push(ie);let ue=W.length,le=j("output",e[0].dataType,ue,R),qe=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],G=De(e[0].dataType),ne=(()=>{switch(A){case 1:return`array<${G}, 8>`;case 2:return`mat4x2<${G}>`;case 4:return`mat2x4<${G}>`;default:throw new Error(`${A}-component is not supported.`)}})(),xe=`\n for (var word: u32 = 0; word < ${_}; word += ${I}) {\n ${ae.indicesSet("b_indices","2","word")};\n let b_data = ${ae.getByIndices("b_indices")};\n for (var i: u32 = 0; i < ${I}; i++) {\n let b_value: u32 = ${I===1?"b_data":"b_data[word + i]"};\n let b_mask: u32 = 0x0F0F0F0Fu;\n let b_value_lower: vec4 = unpack4xU8(b_value & b_mask);\n let b_value_upper: vec4 = unpack4xU8((b_value >> 4) & b_mask);\n let b_quantized_values = ${ne}(${Array.from({length:4},(Be,Ge)=>`${G}(b_value_lower[${Ge}]), ${G}(b_value_upper[${Ge}])`).join(", ")});\n let b_dequantized_values = ${(()=>A===1?`${ne}(${Array.from({length:8},(Be,Ge)=>`(b_quantized_values[${Ge}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${ne}(${Array(8).fill("zero_point").join(",")})) * scale;`)()};\n // Number of B elements per 32-bit word is 32/bits = 32/4 = 8\n for (var m: u32 = 0; m < ${O?c:S}u; m++) {\n ${ee.indicesSet("a_indices",We-2,O?"m":`row * ${S} + m`)};\n ${ee.indicesSet("a_indices",We-1,"word_offset")};\n var input_offset = ${ee.indicesToOffset("a_indices")};\n var a_data: ${ne};\n for (var j: u32 = 0; j < ${8/A}; j++) {\n a_data[j] = ${ee.getByOffset("input_offset")};\n input_offset++;\n }\n ${O?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${R>1?"[c]":""} += ${Array.from({length:8/A},(Be,Ge)=>`${A===1?`a_data[${Ge}] * b_dequantized_values[${Ge}]`:`dot(a_data[${Ge}], b_dequantized_values[${Ge}])`}`).join(" + ")};\n }\n word_offset += ${8/A};\n }\n }`,Ke=ie?`\n zero_point_offset += 4;\n if (zero_point_offset == 32) {\n zero_point_offset = 0;\n zero_point_index++;\n zero_point_word = ${ie.getByOffset("zero_point_index")};\n }`:"";return O?`\n var workgroup_shared: array<${le.type.value}, ${c*a}>;\n ${Ce.declareVariables(...me,le)}\n ${Ce.mainStart([a,1,1])}\n var a_indices: ${ee.type.indices};\n var block = local_id.x;\n var col = workgroup_id.y;\n var batch = workgroup_id.z;\n ${ee.indicesSet("a_indices","0","batch")};\n // Two zero points are packed into one byte when uniforms.bits is 4.\n for (var c: u32 = 0; c < ${R}; c++) {\n let col_times_components_plus_c = col * ${R} + c;\n ${ie?`\n var zero_point_bytes_per_col: u32 = (${a} + 1) / 2;\n var zero_point_byte_count: u32 = col_times_components_plus_c * zero_point_bytes_per_col + (block >> 0x1u);\n var zero_point_word_index: u32 = zero_point_byte_count >> 0x2u;\n var zero_point_byte_offset: u32 = zero_point_byte_count & 0x3u;\n var zero_point_nibble_offset: u32 = block & 0x1u;\n var zero_point_bits_offset: u32 = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);\n var zero_point_word: u32 = ${ie.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;`:""}\n var b_indices: ${ae.type.indices};\n ${ae.indicesSet("b_indices","0","col_times_components_plus_c")};\n // The scale and zero points are computed per block.\n var scales_index = col_times_components_plus_c * ${a} + block;\n let scale = ${Ae.getByOffset("scales_index")};\n // The default zero point is 8 for unsigned 4-bit quantization.\n let zero_point = ${G}(${ie?"(zero_point_word) & 0xFu":8});\n ${ae.indicesSet("b_indices","1","block")};\n var word_offset: u32 = block * ${t.blockSize/A};\n var workgroup_shared_offset: u32 = block * ${c};\n ${xe}\n }\n workgroupBarrier();\n if (local_id.x == 0u) {\n var output_indices: ${le.type.indices};\n ${le.indicesSet("output_indices","0","batch")};\n ${le.indicesSet("output_indices",ue-1,"col")};\n ${le.indicesSet("output_indices",ue-2,"0")};\n var output_offset = ${le.indicesToOffset("output_indices")};\n for (var m: u32 = 0u; m < ${c}u; m++) {\n var output_value: ${le.type.value} = ${le.type.value}(0);\n var workgroup_shared_offset: u32 = m;\n for (var b: u32 = 0u; b < ${a}u; b++) {\n output_value += workgroup_shared[workgroup_shared_offset];\n workgroup_shared_offset += ${c};\n }\n ${le.setByOffset("output_offset","output_value")};\n output_offset += ${h/R};\n }\n }\n }`:`\n ${Ce.registerUniforms(qe).declareVariables(...me,le)}\n ${Ce.mainStart()}\n ${Ce.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n var output_values: array<${le.type.value}, ${S}>;\n var output_indices = ${le.offsetToIndices("global_idx")};\n var col = ${le.indicesGet("output_indices",ue-1)};\n var row = ${le.indicesGet("output_indices",ue-2)};\n var a_indices: ${ee.type.indices} = output_indices;\n // Two zero points are packed into one byte because uniforms.bits <= 4.\n // zero_point_offset is either 0 or 4. It is bit offset within one byte.\n // TODO support zero_point_offset for bits > 4\n ${ie?`\n var zero_point_abs_offset = col * ${R} * ((${a} + 1) / 2);\n var zero_point_index: u32 = zero_point_abs_offset / 4;\n var zero_point_word: u32 = ${ie.getByOffset("zero_point_index")};\n var zero_point_offset: u32 = (zero_point_abs_offset % 4) * 8;`:""}\n var scale_index = col * ${a*R};\n var b_indices: ${ae.type.indices};\n for (var c: u32 = 0; c < ${R}; c++) {\n ${ae.indicesSet("b_indices","0",`col * ${R} + c`)};\n var block_offset: u32 = 0;\n for (var block: u32 = 0; block < ${a}; block++) {\n // The scale and zero points are computed per block.\n let scale = ${Ae.getByOffset("scale_index")};\n // The default zero point is 8 for unsigned 4-bit quantization.\n let zero_point = ${G}(${ie?"extractBits(zero_point_word, zero_point_offset, 4)":8});\n ${ae.indicesSet("b_indices","1","block")};\n var word_offset: u32 = block_offset;\n ${xe}\n scale_index++;\n ${Ke}\n block_offset += uniforms.block_size / ${A};\n }\n // Drop the trailing 4 bits if the zero_poit_offset is not a byte boundary to align with the next byte.\n ${ie?`if (zero_point_offset % 8 > 0) {\n ${Ke}\n }`:""}\n }\n for (var k: u32 = 0u; k < ${S}u; k++) {\n ${le.indicesSet("output_indices",ue-2,`${S} * row + k`)};\n ${le.setByIndices("output_indices","output_values[k]")}\n }\n }`};return{name:O?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${t.cacheKey};${c};${v};${e.length}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:L,dataType:v}],name:O?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:O?{x:1,y:Math.ceil(h/R),z:y}:{x:Math.ceil(N/64)},programUniforms:K}),getShaderSource:se}},Wu=(e,t)=>{np(e.inputs,t);let r=e.getMaxComputeWorkgroupSizes(),o=e.getMaxComputeWorkgroupStoragesize();e.compute(op(e.inputs,t,r,o))},Nu=e=>ve(e)});var it,ip,Lu,Hu,ap,Ko,Fu,qu=Y(()=>{"use strict";ye();Se();Ze();_n();Ro();_e();Sr();it=(e,t)=>e.length>t&&e[t].dims.length>0&&M.size(e[t].dims)>0?e[t]:void 0,ip=(e,t)=>{let r=e[0],o=it(e,1),i=it(e,2),u=it(e,3),a=it(e,4),c=it(e,5),p=it(e,6),h=it(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 d=!1,y=r.dims[0],w=r.dims[1],_=r.dims.length===3?d?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],v=w,S=0,A=0,I=Math.floor(_/t.numHeads);if(p&&h){if(p.dims.length!==4)throw new Error(\'Input "past_key" is expected to have 4 dimensions\');if(p.dims[0]!==y||p.dims[1]!==t.numHeads||p.dims[3]!==I)throw new Error(\'Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)\');if(h.dims[0]!==y||h.dims[1]!==t.numHeads||h.dims[3]!==I)throw new Error(\'Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)\');if(p.dims[2]!==h.dims[2])throw new Error(\'Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)\');if(h.dims.length!==4)throw new Error(\'Input "past_value" is expected to have 4 dimensions\');S=p.dims[2],A=p.dims[2]}else if(p||h)throw new Error(\'Input "past_key" and "past_value" shall be both present or both absent\');let x;if(o){if(r.dims.length!==3)throw new Error(\'Input "query" is expected to have 3 dimensions when key is given\');if(o.dims.length<3||o.dims.length>5)throw new Error(\'Input "key" is expected to have 3, 4, or 5 dimensions\');if(r.dims[0]!==o.dims[0])throw new Error(\'Input "query" and "key" shall have same dim 0 (batch size)\');if(o.dims.length===3){if(o.dims[2]!==r.dims[2])throw new Error(\'Input "query" and "key" shall have same dim 2 (hidden_size)\');x=2,v=o.dims[1]}else if(o.dims.length===5){if(o.dims[2]!==t.numHeads||o.dims[3]!==2||o.dims[4]!==I)throw new Error(\'Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv\');if(i)throw new Error(\'Expect "value" be none when "key" has packed kv format.\');x=5,v=o.dims[1]}else{if(o.dims[1]!==t.numHeads||o.dims[3]!==I)throw new Error(\'Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key\');x=0,v=o.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\');x=3}if(u){if(u.dims.length!==1)throw new Error(\'Input "bias" is expected to have 1 dimension\');if(i&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let E=0;if(a){E=8;let N=a.dims;throw N.length===1?N[0]===y?E=1:N[0]===3*y+2&&(E=3):N.length===2&&N[0]===y&&N[1]===v&&(E=5),E===8?new Error(\'Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)\'):new Error("Mask not supported")}let P=!1,O=_;if(i){if(i.dims.length!==3&&i.dims.length!==4)throw new Error(\'Input "value" is expected to have 3 or 4 dimensions\');if(r.dims[0]!==i.dims[0])throw new Error(\'Input "query" and "value" shall have same dim 0 (batch_size)\');if(i.dims.length===3){if(v!==i.dims[1])throw new Error(\'Input "key" and "value" shall have the same dim 1 (kv_sequence_length)\');O=i.dims[2]}else{if(v!==i.dims[2])throw new Error(\'Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)\');O=i.dims[1]*i.dims[3],P=!0}}let R=S+v,L=!1;if(a)throw new Error("Key padding mask is not supported");if(c){if(c.dims.length!==4)throw new Error(\'Input "relative_position_bias" is expected to have 4 dimensions\');if(c.dims[0]!==y&&c.dims[0]!==1||c.dims[1]!==t.numHeads||c.dims[2]!==w||c.dims[3]!==R)throw new Error(\'Input "relative_position_bias" shape (batch_size, 1, sequence_length, kv_sequence_length)\')}return{batchSize:y,sequenceLength:w,pastSequenceLength:S,kvSequenceLength:v,totalSequenceLength:R,maxSequenceLength:A,inputHiddenSize:0,hiddenSize:_,vHiddenSize:O,headSize:I,vHeadSize:Math.floor(O/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:E,scale:t.scale,broadcastResPosBias:L,passPastInKv:P,qkvFormat:x}},Lu=e=>ve({...e}),Hu=ve({perm:[0,2,1,3]}),ap=(e,t,r,o,i,u,a)=>{let c=[o,i,u],p=M.size(c),h=[{type:12,data:p},{type:12,data:a},{type:12,data:u}],d=y=>{let w=j("qkv_with_bias",t.dataType,c),_=U("qkv",t.dataType,c),v=U("bias",r.dataType,c),S=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`\n ${y.registerUniforms(S).declareVariables(_,v,w)}\n ${y.mainStart()}\n ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;\n\n qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];\n }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:c,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:d},{inputs:[t,r],outputs:[-1]})[0]},Ko=(e,t,r,o,i,u,a,c)=>{let p=u;if(a){if(o===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return p=ap(e,u,a,t,o,r*i,c),p=p.reshape([t,o,r,i]),e.compute(yt(p,Hu.perm),{inputs:[p],outputs:[-1]})[0]}else return u.dims.length===3&&(p=u.reshape([t,o,r,i])),e.compute(yt(p,Hu.perm),{inputs:[p],outputs:[-1]})[0]},Fu=(e,t)=>{let r=ip(e.inputs,t),o=e.inputs[0],i=it(e.inputs,1),u=it(e.inputs,2),a=it(e.inputs,3),c=it(e.inputs,4),p=it(e.inputs,5),h=it(e.inputs,6),d=it(e.inputs,7);if(o.dims.length===5)throw new Error("Packed QKV is not implemented");if(i?.dims.length===5)throw new Error("Packed KV is not implemented");let y=i&&u&&i.dims.length===4&&u.dims.length===4,w=Ko(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,o,a,0);if(y)return Pn(e,w,i,u,c,void 0,h,d,p,r,t);if(!i||!u)throw new Error("key and value must be provided");let _=Ko(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,i,a,r.hiddenSize),v=Ko(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,u,a,2*r.hiddenSize);Pn(e,w,_,v,c,void 0,h,d,p,r,t)}});var sp,up,dp,lp,cp,pp,mp,fp,ju,Ku=Y(()=>{"use strict";ye();Se();_e();sp=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].")}},up=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n if (k < 0) {\n break;\n }\n if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n break;\n }\n offset += k * i32(${fe("uniforms.x_strides",i,t)});\n `;return`\n value = ${e.type.value}(uniforms.constant_value);\n for (var i = 0; i < 1; i++) {\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n }\n `},dp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n if (k < 0) {\n k = -k;\n }\n {\n let _2n_1 = 2 * (i32(${fe("uniforms.x_shape",i,t)}) - 1);\n k = k % _2n_1;\n if(k >= i32(${fe("uniforms.x_shape",i,t)})) {\n k = _2n_1 - k;\n }\n }\n offset += k * i32(${fe("uniforms.x_strides",i,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},lp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n if (k < 0) {\n k = 0;\n }\n if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n k = i32(${fe("uniforms.x_shape",i,t)}) - 1;\n }\n offset += k * i32(${fe("uniforms.x_strides",i,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},cp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n if (k < 0) {\n k += i32(${fe("uniforms.x_shape",i,t)}]);\n }\n if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n k -= i32(${fe("uniforms.x_shape",i,t)});\n }\n offset += k * i32(${fe("uniforms.x_strides",i,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},pp=(e,t,r)=>{switch(r.mode){case 0:return up(e,t,r.pads.length);case 1:return dp(e,t,r.pads.length);case 2:return lp(e,t,r.pads.length);case 3:return cp(e,t,r.pads.length);default:throw new Error("Invalid mode")}},mp=(e,t)=>{let r=M.padShape(e[0].dims.slice(),t.pads),o=e[0].dims,i=M.size(r),u=[{type:12,data:i},{type:6,data:t.pads}];t.mode===0&&u.push({type:e[0].dataType,data:t.value}),u.push(...Z(e[0].dims,r));let a=["rank"],c=p=>{let h=j("output",e[0].dataType,r.length),d=U("x",e[0].dataType,o.length),y=d.type.value,w=pp(h,o.length,t),_=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&_.push({name:"constant_value",type:y}),`\n ${p.registerUniforms(_).declareVariables(d,h)}\n ${p.mainStart()}\n ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let indices = ${h.offsetToIndices("global_idx")};\n\n var value = ${y}(0);\n ${w}\n output[global_idx] = value;\n }`};return{name:"Pad",shaderCache:{hint:`${t.mode}`,inputDependencies:a},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(M.size(r)/64)},programUniforms:u}),getShaderSource:c}},fp=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),o=e.length>=3&&e[2].data?e[2].getFloat32Array()[0]:0,i=e[0].dims.length,u=new Int32Array(2*i).fill(0);if(e.length>=4){let c=e[3].getBigInt64Array();for(let p=0;pu[Number(p)]=Number(c));let a=[];return u.forEach(c=>a.push(c)),{mode:t.mode,value:o,pads:a}}else return t},ju=(e,t)=>{sp(e.inputs);let r=fp(e.inputs,t);e.compute(mp(e.inputs,r),{inputs:[0]})}});var Nn,Yu,Zu,Xu,Qu,hp,gp,Ju,ed,td,rd,nd,od,id,ad,sd,ud,dd,ld,cd=Y(()=>{"use strict";$r();ye();Se();_e();Nn=e=>{if(vr.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Yu=(e,t,r)=>{let o=t.format==="NHWC",i=e.dims.slice();o&&i.splice(1,0,i.pop());let u=Object.hasOwnProperty.call(t,"dilations"),a=t.kernelShape.slice(),c=t.strides.slice(),p=u?t.dilations.slice():[],h=t.pads.slice();nr.adjustPoolAttributes(r,i,a,c,p,h);let d=nr.computePoolOutputShape(r,i,c,p,a,h,t.autoPad),y=Object.assign({},t);u?Object.assign(y,{kernelShape:a,strides:c,pads:h,dilations:p,cacheKey:t.cacheKey}):Object.assign(y,{kernelShape:a,strides:c,pads:h,cacheKey:t.cacheKey});let w=d.slice();return w.push(w.splice(1,1)[0]),[y,o?w:d]},Zu=(e,t)=>{let r=t.format==="NHWC",o=M.size(e),i=M.size(t.kernelShape),u=[{type:12,data:o},{type:12,data:i}],a=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let c=t.kernelShape[t.kernelShape.length-1],p=t.strides[t.strides.length-1],h=t.pads[t.pads.length/2-1],d=t.pads[t.pads.length-1],y=!!(h+d);u.push({type:12,data:c},{type:12,data:p},{type:12,data:h},{type:12,data:d}),a.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let w=!1;if(t.kernelShape.length===2){let _=t.kernelShape[t.kernelShape.length-2],v=t.strides[t.strides.length-2],S=t.pads[t.pads.length/2-2],A=t.pads[t.pads.length-2];w=!!(S+A),u.push({type:12,data:_},{type:12,data:v},{type:12,data:S},{type:12,data:A}),a.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[u,a,!0,y,w]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let c=M.computeStrides(t.kernelShape);u.push({type:12,data:c},{type:12,data:t.pads},{type:12,data:t.strides}),a.push({name:"kernelStrides",type:"u32",length:c.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let p=t.pads.reduce((h,d)=>h+d);return[u,a,!!p,!1,!1]}},Xu=(e,t,r,o,i,u,a,c,p,h,d,y)=>{let w=i.format==="NHWC",_=t.type.value,v=j("output",t.type.tensor,o);if(i.kernelShape.length<=2){let S="",A="",I="",x=r-(w?2:1);if(d?S=`\n for (var i: u32 = 0u; i < uniforms.kw; i++) {\n xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i;\n if (xIndices[${x}] < 0 || xIndices[${x}]\n >= uniforms.x_shape[${x}]) {\n pad++;\n continue;\n }\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${u}\n }`:S=`\n for (var i: u32 = 0u; i < uniforms.kw; i++) {\n xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i;\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${u}\n }`,i.kernelShape.length===2){let P=r-(w?3:2);y?A=`\n for (var j: u32 = 0u; j < uniforms.kh; j++) {\n xIndices[${P}] = indices[${P}] * uniforms.sh - uniforms.phStart + j;\n if (xIndices[${P}] < 0 || xIndices[${P}] >= uniforms.x_shape[${P}]) {\n pad += i32(uniforms.kw);\n continue;\n }\n `:A=`\n for (var j: u32 = 0u; j < uniforms.kh; j++) {\n xIndices[${P}] = indices[${P}] * uniforms.sh - uniforms.phStart + j;\n `,I=`\n }\n `}return`\n ${e.registerUniforms(p).declareVariables(t,v)}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n let indices = ${v.offsetToIndices("global_idx")};\n var xIndices = ${v.offsetToIndices("global_idx")};\n\n var value = ${_}(${c});\n var pad = 0;\n ${A}\n ${S}\n ${I}\n ${a}\n\n output[global_idx] = value;\n }`}else{if(w)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let S=i.kernelShape.length,A=i.pads.length,I="";return h?I=`\n if (xIndices[j] >= uniforms.x_shape[j]) {\n pad++;\n isPad = true;\n break;\n }\n }\n if (!isPad) {\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${u}\n }`:I=`\n }\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${u}\n `,`\n ${e.registerUniforms(p).declareVariables(t,v)}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n let indices = ${v.offsetToIndices("global_idx")};\n var xIndices = ${v.offsetToIndices("global_idx")};\n\n var offsets: array;\n\n var value = ${_}(${c});\n var pad = 0;\n var isPad = false;\n\n for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {\n var offset = i;\n for (var j = 0u; j < ${S-1}u; j++) {\n offsets[j] = offset / ${fe("uniforms.kernelStrides","j",S)};\n offset -= offsets[j] * ${fe("uniforms.kernelStrides","j",S)};\n }\n offsets[${S-1}] = offset;\n\n isPad = false;\n for (var j = ${r-S}u; j < ${r}u; j++) {\n xIndices[j] = indices[j] * ${fe("uniforms.strides",`j - ${r-S}u`,S)}\n + offsets[j - ${r-S}u] - ${fe("uniforms.pads","j - 2u",A)};\n ${I}\n }\n ${a}\n\n output[global_idx] = value;\n }`}},Qu=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,hp=e=>`${Qu(e)};${e.countIncludePad}`,gp=e=>`${Qu(e)};${e.storageOrder};${e.dilations}`,Ju=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}),ed=(e,t,r,o)=>{let[i,u]=Yu(t,o,r),a=U("x",t.dataType,t.dims.length),c=a.type.value,p="value += x_val;",h="";i.countIncludePad?h+=`value /= ${c}(uniforms.kernelSize);`:h+=`value /= ${c}(i32(uniforms.kernelSize) - pad);`;let[d,y,w,_,v]=Zu(u,i);d.push(...Z(t.dims,u));let S=["rank"];return{name:e,shaderCache:{hint:`${o.cacheKey};${w};${_};${v}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(M.size(u)/64)},programUniforms:d}),getShaderSource:A=>Xu(A,a,t.dims.length,u.length,i,p,h,0,y,w,_,v)}},td=e=>{let t=e.count_include_pad!==0,r=Ju(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let o={countIncludePad:t,...r,cacheKey:""};return{...o,cacheKey:hp(o)}},rd=(e,t)=>{Nn(e.inputs),e.compute(ed("AveragePool",e.inputs[0],!1,t))},nd={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},od=e=>{let t=e.format;return{format:t,...nd,cacheKey:t}},id=(e,t)=>{Nn(e.inputs),e.compute(ed("GlobalAveragePool",e.inputs[0],!0,t))},ad=(e,t,r,o)=>{let[i,u]=Yu(t,o,r),a=`\n value = max(x_val, value);\n `,c="",p=U("x",t.dataType,t.dims.length),h=["rank"],[d,y,w,_,v]=Zu(u,i);return d.push(...Z(t.dims,u)),{name:e,shaderCache:{hint:`${o.cacheKey};${w};${_};${v}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(M.size(u)/64)},programUniforms:d}),getShaderSource:S=>Xu(S,p,t.dims.length,u.length,i,a,c,t.dataType===10?-65504:-1e5,y,w,_,v)}},sd=(e,t)=>{Nn(e.inputs),e.compute(ad("MaxPool",e.inputs[0],!1,t))},ud=e=>{let t=e.storage_order,r=e.dilations,o=Ju(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(o.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let i={storageOrder:t,dilations:r,...o,cacheKey:""};return{...i,cacheKey:gp(i)}},dd=e=>{let t=e.format;return{format:t,...nd,cacheKey:t}},ld=(e,t)=>{Nn(e.inputs),e.compute(ad("GlobalMaxPool",e.inputs[0],!0,t))}});var bp,wp,pd,md=Y(()=>{"use strict";$r();ye();_e();bp=(e,t,r)=>{let o=e===t,i=et&&r>0;if(o||i||u)throw new Error("Range these inputs\' contents are invalid.")},wp=(e,t,r,o)=>{let i=Math.abs(Math.ceil((t-e)/r)),u=[i],a=i,c=[{type:12,data:a},{type:o,data:e},{type:o,data:r},...Z(u)],p=h=>{let d=j("output",o,u.length),y=d.type.value,w=[{name:"outputSize",type:"u32"},{name:"start",type:y},{name:"delta",type:y}];return`\n ${h.registerUniforms(w).declareVariables(d)}\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n output[global_idx] = uniforms.start + ${y}(global_idx) * uniforms.delta;\n }`};return{name:"Range",shaderCache:{hint:`${o}`},getShaderSource:p,getRunData:()=>({outputs:[{dims:u,dataType:o}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:c})}},pd=e=>{let t=0,r=0,o=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],o=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],o=e.inputs[2].getFloat32Array()[0]),vr.webgpu.validateInputContent&&bp(t,r,o),e.compute(wp(t,r,o,e.inputs[0].dataType),{inputs:[]})}});var vp,$p,_p,Sp,xp,Cp,Ap,Ip,Tp,Ep,Pp,fd,kp,Op,Rp,Bp,Dp,hd,gd,yd=Y(()=>{"use strict";ye();Se();Ze();_e();vp=(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\n 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")}},$p=(e,t,r)=>{t.every(i=>i>=0&&i{throw new Error("Resize requires axes input values to be positive and less than rank")}));let o=new Array(r).fill(1);return t.forEach((i,u)=>o[i]=e[u]),o},_p=(e,t,r,o,i,u)=>{let[a,c,p]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],h=e[0].dims.length;if(a>0&&e.length>a&&e[a].dims.length>0)e[a].getFloat32Array().forEach(d=>u.push(d));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(c>0&&e.length>c&&e[c].dims.length>0){if(e[c].getFloat32Array().forEach(d=>o.push(d)),o.length!==0&&o.length!==h&&r>=18&&o.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");vp(o,t),t.axes.length>0&&$p(o,t.axes,h).forEach((d,y)=>o[y]=d)}if(p>0&&e.length>p&&(e[p].getBigInt64Array().forEach(d=>i.push(Number(d))),i.length!==h||r>=18&&i.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(o.length!==t.axes.length)throw new Error(\'Resize requires "scales" input size to be of axes rank when axes attributes is specified\');if(i.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 o<"u"&&typeof i<"u"&&o.length>0&&i.length>h)throw new Error("Resize requires only of scales or sizes to be specified")},Sp=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,\n 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) {\n return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;\n } else {\n return 0.0;\n }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {\n return 0.0;\n } else {\n // The whole part and the fractional part are calculated separately due to inaccuracy of floating\n // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an\n // offset-by-one error later in floor().\n let whole = ${t}(xResized * (lengthOriginal - 1) / (lengthResized - 1));\n let fract =\n ${t}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${t}(lengthResized - 1);\n return whole + fract;\n }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {\n return ${t}(roiStart) * ${t}(lengthOriginal - 1) +\n (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /\n ${t}(lengthResized - 1);\n } else {\n return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);\n }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);\n const adjustment = ${t}(lengthResized) / outputWidth;\n const center = ${t}(lengthOriginal) / 2;\n const offset = center * (1 - adjustment);\n 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`)}})()+"}",xp=(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`)}})()+"}",Cp=(e,t,r)=>{let o=new Array(r).fill(0).concat(new Array(r).fill(1)),i=e.length===0?o:e.slice();return t.length>0?(t.forEach((u,a)=>{o[u]=i[a],o[a+r]=i[t.length+a]}),o):i},Ap=(e,t,r,o)=>{let i=[];if(r.length>0)if(o.length>0){if(e.forEach(u=>i.push(u)),Math.max(...o)>e.length)throw new Error("axes is out of bound");o.forEach((u,a)=>i[u]=r[a])}else r.forEach(u=>i.push(u));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");i=e.map((u,a)=>Math.round(u*t[a]))}return i},Ip=(e,t,r)=>{let o=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(u=>t[u]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(u=>t[u]),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 i=e.slice();return r.axes.length>0?(r.axes.forEach(u=>t[u]=o),r.axes.forEach(u=>i[u]=Math.round(e[u]*t[u]))):(t.fill(o,0,t.length),i.forEach((u,a)=>i[a]=Math.round(u*t[a]))),i},Tp=(e,t,r,o,i)=>`\n fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {\n var original_indices: array<${e.type.value}, ${r.length}>;\n for (var i:u32 = 0; i < ${r.length}; i++) {\n var output_index = ${e.indicesGet("output_indices","i")};\n var scale = ${fe("uniforms.scales","i",o)};\n var roi_low = ${fe("uniforms.roi","i",i)};\n var roi_hi = ${fe("uniforms.roi",`i + ${t.length}`,i)};\n if (scale == 1.0) {\n original_indices[i] = ${e.type.value}(output_index);\n } else {\n var input_shape_i = ${fe("uniforms.input_shape","i",t.length)};\n var output_shape_i = ${fe("uniforms.output_shape","i",r.length)};\n original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,\n input_shape_i, roi_low, roi_hi);\n }\n }\n return original_indices;\n }`,Ep=(e,t,r,o,i,u,a)=>`\n fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {\n var input_indices: ${e.type.indices};\n for (var i:u32 = 0; i < ${o.length}; i++) {\n var output_index = ${t.indicesGet("output_indices","i")};\n var input_index: u32;\n var scale = ${fe("uniforms.scales","i",i)};\n if (scale == 1.0) {\n input_index = output_index;\n } else {\n var roi_low = ${fe("uniforms.roi","i",u)};\n var roi_hi = ${fe("uniforms.roi",`i + ${r.length}`,u)};\n var input_shape_i = ${fe("uniforms.input_shape","i",r.length)};\n var output_shape_i = ${fe("uniforms.output_shape","i",o.length)};\n var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,\n input_shape_i, roi_low, roi_hi);\n if (!${a} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {\n if (original_idx < 0) {\n input_index = 0;\n } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {\n input_index = input_shape_i - 1;\n } else {\n input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));\n }\n } else {\n input_index = u32(original_idx);\n }\n }\n ${e.indicesSet("input_indices","i"," input_index")}\n }\n return input_indices;\n }`,Pp=(e,t)=>`\n fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {\n for (var i:u32 = 0; i < ${t.length}; i++) {\n var input_index = ${e.indicesGet("input_indices","i")};\n if (input_index < 0 || input_index >= ${fe("uniforms.input_shape","i",t.length)}) {\n return false;\n }\n }\n return true;\n }`,fd=(e,t,r,o)=>e.rank>o?`\n ${e.indicesSet("input_indices",t,"channel")};\n ${e.indicesSet("input_indices",r,"batch")};\n`:"",kp=(e,t,r,o,i)=>{let[a,c,p,h]=r.length===2?[-1,0,1,-1]:[0,2,3,1],d=e.type.value;return`\n fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${d} {\n var input_indices: ${e.type.indices};\n ${e.indicesSet("input_indices",c,`max(0, min(row, ${r[c]} - 1))`)};\n ${e.indicesSet("input_indices",p,`max(0, min(col, ${r[p]} - 1))`)};\n ${fd(e,h,a,2)}\n return ${e.getByIndices("input_indices")};\n }\n\n fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${d} {\n var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);\n var row:${d} = originalIndices[${c}];\n var col:${d} = originalIndices[${p}];\n ${o?`if (row < 0 || row > (${r[c]} - 1) || col < 0 || col > (${r[p]} - 1)) {\n return ${i};\n }`:""};\n row = max(0, min(row, ${r[c]} - 1));\n col = max(0, min(col, ${r[p]} - 1));\n var row1: u32 = u32(row);\n var col1: u32 = u32(col);\n var row2: u32 = u32(row + 1);\n var col2: u32 = u32(col + 1);\n var channel: u32 = ${r.length>2?`u32(originalIndices[${h}])`:"0"};\n var batch: u32 = ${r.length>2?`u32(originalIndices[${a}])`:"0"};\n var x11: ${d} = getInputValue(batch, channel, row1, col1);\n var x12: ${d} = getInputValue(batch, channel, row1, col2);\n var x21: ${d} = getInputValue(batch, channel, row2, col1);\n var x22: ${d} = getInputValue(batch, channel, row2, col2);\n var dx1: ${d} = abs(row - ${d}(row1));\n var dx2: ${d} = abs(${d}(row2) - row);\n var dy1: ${d} = abs(col - ${d}(col1));\n var dy2: ${d} = abs(${d}(col2) - col);\n if (row1 == row2) {\n dx1 = 0.5;\n dx2 = 0.5;\n }\n if (col1 == col2) {\n dy1 = 0.5;\n dy2 = 0.5;\n }\n return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);\n }`},Op=(e,t,r,o,i,u,a,c,p,h)=>{let d=r.length===2,y=!0,[w,_]=d?[0,1]:y?[2,3]:[1,2],v=e.type.value,S=A=>{let I=A===w?"row":"col";return`\n fn ${I}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${v} {\n var output_index = ${t.indicesGet("output_indices",A)};\n var originalIdx: ${v} = getOriginalCoordinateFromResizedCoordinate(output_index, ${i[A]},\n ${o[A]}, ${r[A]}, ${u[A]}, ${u[A]} + ${r.length});\n var fractOriginalIdx: ${v} = originalIdx - floor(originalIdx);\n var coefs = getCubicInterpolationCoefs(fractOriginalIdx);\n\n if (${c} && (originalIdx < 0 || originalIdx > (${r[A]} - 1))) {\n return ${p};\n }\n var data: array<${v}, 4> = array<${v}, 4>(0.0, 0.0, 0.0, 0.0);\n for (var i: i32 = -1; i < 3; i++) {\n var ${I}: ${v} = originalIdx + ${v}(i);\n if (${I} < 0 || ${I} >= ${r[A]}) {\n ${(()=>h?`coefs[i + 1] = 0.0;\n continue;`:c?`return ${p};`:`${I} = max(0, min(${I}, ${r[A]} - 1));`)()};\n }\n var input_indices_copy: ${e.type.indices} = input_indices;\n ${e.indicesSet("input_indices_copy",A,`u32(${I})`)};\n data[i + 1] = ${A===w?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};\n }\n return cubicInterpolation1D(data, coefs);\n }`};return`\n ${S(w)};\n ${S(_)};\n fn getCubicInterpolationCoefs(s: ${v}) -> array<${v}, 4> {\n var absS = abs(s);\n var coeffs: array<${v}, 4> = array<${v}, 4>(0.0, 0.0, 0.0, 0.0);\n var oneMinusAbsS: ${v} = 1.0 - absS;\n var twoMinusAbsS: ${v} = 2.0 - absS;\n var onePlusAbsS: ${v} = 1.0 + absS;\n coeffs[0] = ((${a} * onePlusAbsS - 5 * ${a}) * onePlusAbsS + 8 * ${a}) * onePlusAbsS - 4 * ${a};\n coeffs[1] = ((${a} + 2) * absS - (${a} + 3)) * absS * absS + 1;\n coeffs[2] = ((${a} + 2) * oneMinusAbsS - (${a} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;\n coeffs[3] = ((${a} * twoMinusAbsS - 5 * ${a}) * twoMinusAbsS + 8 * ${a}) * twoMinusAbsS - 4 * ${a};\n return coeffs;\n }\n\n fn cubicInterpolation1D(x: array<${v}, 4>, coefs: array<${v}, 4>) -> ${v} {\n var coefsSum: ${v} = coefs[0] + coefs[1] + coefs[2] + coefs[3];\n return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;\n }\n\n fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${v} {\n var input_indices: ${e.type.indices} = output_indices;\n return colCubicInterpolation(input_indices, output_indices);\n }\n `},Rp=(e,t,r,o,i)=>{let[a,c,p,h,d]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],y=e.type.value;return`\n fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${y} {\n var input_indices: ${e.type.indices};\n ${e.indicesSet("input_indices",c,`max(0, min(depth, ${r[c]} - 1))`)};\n ${e.indicesSet("input_indices",p,`max(0, min(height, ${r[p]} - 1))`)};\n ${e.indicesSet("input_indices",h,`max(0, min(width, ${r[h]} - 1))`)};\n ${fd(e,d,a,3)}\n return ${e.getByIndices("input_indices")};\n }\n\n fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${y} {\n var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);\n var depth:${y} = originalIndices[${c}];\n var height:${y} = originalIndices[${p}];\n var width:${y} = originalIndices[${h}];\n ${o?`if (depth < 0 || depth > (${r[c]} - 1) || height < 0 || height > (${r[p]} - 1) || width < 0 || (width > ${r[h]} - 1)) {\n return ${i};\n }`:""};\n\n depth = max(0, min(depth, ${r[c]} - 1));\n height = max(0, min(height, ${r[p]} - 1));\n width = max(0, min(width, ${r[h]} - 1));\n var depth1: u32 = u32(depth);\n var height1: u32 = u32(height);\n var width1: u32 = u32(width);\n var depth2: u32 = u32(depth + 1);\n var height2: u32 = u32(height + 1);\n var width2: u32 = u32(width + 1);\n var channel: u32 = ${r.length>3?`u32(originalIndices[${d}])`:"0"};\n var batch: u32 = ${r.length>3?`u32(originalIndices[${a}])`:"0"};\n\n var x111: ${y} = getInputValue(batch, channel, depth1, height1, width1);\n var x112: ${y} = getInputValue(batch, channel, depth1, height1, width2);\n var x121: ${y} = getInputValue(batch, channel, depth1, height2, width1);\n var x122: ${y} = getInputValue(batch, channel, depth1, height2, width2);\n var x211: ${y} = getInputValue(batch, channel, depth2, height1, width1);\n var x212: ${y} = getInputValue(batch, channel, depth2, height1, width2);\n var x221: ${y} = getInputValue(batch, channel, depth2, height2, width1);\n var x222: ${y} = getInputValue(batch, channel, depth2, height2, width2);\n var dx1: ${y} = abs(depth - ${y}(depth1));\n var dx2: ${y} = abs(${y}(depth2) - depth);\n var dy1: ${y} = abs(height - ${y}(height1));\n var dy2: ${y} = abs(${y}(height2) - height);\n var dz1: ${y} = abs(width - ${y}(width1));\n var dz2: ${y} = abs(${y}(width2) - width);\n if (depth1 == depth2) {\n dx1 = 0.5;\n dx2 = 0.5;\n }\n if (height1 == height2) {\n dy1 = 0.5;\n dy2 = 0.5;\n }\n if (width1 == width2) {\n dz1 = 0.5;\n dz2 = 0.5;\n }\n return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +\n x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);\n }`},Bp=(e,t,r,o,i,u)=>{let a=e.dims,c=Cp(u,t.axes,a.length),p=Ap(a,o,i,t.axes),h=o.slice();o.length===0&&(h=a.map((x,E)=>x===0?1:p[E]/x),t.keepAspectRatioPolicy!=="stretch"&&(p=Ip(a,h,t)));let d=j("output",e.dataType,p.length),y=U("input",e.dataType,a.length),w=M.size(p),_=a.length===p.length&&a.every((x,E)=>x===p[E]),v=t.coordinateTransformMode==="tf_crop_and_resize",S=t.extrapolationValue,A=y.type.value,I=x=>`\n ${_?"":`\n ${Sp(t.coordinateTransformMode,A)};\n ${(()=>{switch(t.mode){case"nearest":return`\n ${Pp(y,a)};\n ${xp(t.nearestMode,r,A)};\n ${Ep(y,d,a,p,h.length,c.length,v)};\n `;case"linear":return`\n ${Tp(d,a,p,h.length,c.length)};\n ${(()=>{if(a.length===2||a.length===4)return`${kp(y,d,a,v,S)}`;if(a.length===3||a.length===5)return`${Rp(y,d,a,v,S)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};\n `;case"cubic":return`\n ${(()=>{if(a.length===2||a.length===4)return`${Op(y,d,a,p,h,c,t.cubicCoeffA,v,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};\n `;default:throw Error("Invalid resize mode")}})()};\n `}\n ${x.registerUniform("output_size","u32").registerUniform("scales","f32",h.length).registerUniform("roi","f32",c.length).declareVariables(y,d)}\n ${x.mainStart()}\n ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n ${_?"output[global_idx] = input[global_idx];":`\n let output_indices = ${d.offsetToIndices("global_idx")};\n var input_indices: ${y.type.indices};\n ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);\n if (checkInputIndices(input_indices)) {\n output[global_idx] = ${y.getByIndices("input_indices")};\n } else {\n output[global_idx] = ${t.extrapolationValue};\n }`;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}`)}})()};\n`}\n }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${h.length>0?h:""}|${i.length>0?i:""}|${c.length>0?c:""}|${_}|${a}`,inputDependencies:["rank"]},getShaderSource:I,getRunData:()=>({outputs:[{dims:p,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:[{type:12,data:w},{type:1,data:h},{type:1,data:c},...Z(a,p)]})}},Dp=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},hd=(e,t)=>{let r=[],o=[],i=[],u=Dp(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");_p(e.inputs,t,u,r,o,i),e.compute(Bp(e.inputs[0],t,u,r,o,i),{inputs:[0]})},gd=e=>{let t=e.antialias,r=e.axes,o=e.coordinateTransformMode,i=e.cubicCoeffA,u=e.excludeOutside!==0,a=e.extrapolationValue,c=e.keepAspectRatioPolicy,p=e.mode,h=e.nearestMode===""?"simple":e.nearestMode;return ve({antialias:t,axes:r,coordinateTransformMode:o,cubicCoeffA:i,excludeOutside:u,extrapolationValue:a,keepAspectRatioPolicy:c,mode:p,nearestMode:h})}});var zp,Mp,bd,wd=Y(()=>{"use strict";ye();Se();Ze();_e();zp=(e,t)=>{let[r,o,i,u]=e,{numHeads:a,rotaryEmbeddingDim:c}=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(!M.areEqual(o.dims,[])&&!M.areEqual(o.dims,[1])&&o.dims.length!==2)throw new Error(`Input \'position_ids\' is expected to have 0, 1, or 2 dimensions, got ${o.dims.length}`);if(i.dims.length!==2)throw new Error(`Input \'cos_cache\' is expected to have 2 dimensions, got ${i.dims.length}`);if(u.dims.length!==2)throw new Error(`Input \'sin_cache\' is expected to have 2 dimensions, got ${u.dims.length}`);if(!M.areEqual(i.dims,u.dims))throw new Error("Inputs \'cos_cache\' and \'sin_cache\' are expected to have the same shape");if(c>0&&a===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let p=r.dims[0],h=r.dims[r.dims.length-2],d=i.dims[0],y=M.sizeFromDimension(r.dims,1)/h,w=c===0?i.dims[1]*2:y/a;if(c>w)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(o.dims.length===2){if(p!==o.dims[0])throw new Error(`Input \'position_ids\' dimension 0 should be of size batch_size, got ${o.dims[0]}`);if(h!==o.dims[1])throw new Error(`Input \'position_ids\' dimension 1 should be of size sequence_length, got ${o.dims[1]}`)}if(w/2!==i.dims[1]&&c/2!==i.dims[1])throw new Error(`Input \'cos_cache\' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${i.dims[1]}`);if(h>d)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Mp=(e,t)=>{let{interleaved:r,numHeads:o,rotaryEmbeddingDim:i,scale:u}=t,a=e[0].dims[0],c=M.sizeFromDimension(e[0].dims,1),p=e[0].dims[e[0].dims.length-2],h=c/p,d=e[2].dims[1],y=i===0?d*2:h/o,w=new Array(a,p,h/y,y-d),_=M.computeStrides(w),v=[{type:1,data:u},{type:12,data:w},{type:12,data:_},...e[0].dims.length===3?new Array({type:12,data:[c,h,y,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[c,y,p*y,1]}):[],...Z(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],S=A=>{let I=U("input",e[0].dataType,e[0].dims.length),x=U("position_ids",e[1].dataType,e[1].dims.length),E=U("cos_cache",e[2].dataType,e[2].dims.length),P=U("sin_cache",e[3].dataType,e[3].dims.length),O=j("output",e[0].dataType,e[0].dims.length);return A.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:w.length},{name:"global_strides",type:"u32",length:_.length},{name:"input_output_strides",type:"u32",length:_.length}]),`\n ${A.declareVariables(I,x,E,P,O)}\n\n ${A.mainStart(or)}\n let half_rotary_emb_dim = uniforms.${E.name}_shape[1];\n let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;\n let size = uniforms.global_shape[0] * uniforms.global_strides[0];\n ${A.guardAgainstOutOfBoundsWorkgroupSizes("size")}\n\n if (bsnh[3] < half_rotary_emb_dim) {\n let position_ids_idx =\n ${x.broadcastedIndicesToOffset("bsnh.xy",j("",x.type.tensor,2))};\n let position_id =\n u32(${x.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);\n let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${r});\n let j = i + select(half_rotary_emb_dim, 1, ${r});\n let re = ${I.getByOffset("i")} * ${E.get("position_id","bsnh[3]")} -\n ${I.getByOffset("j")} * ${P.get("position_id","bsnh[3]")};\n ${O.setByOffset("i","re")}\n let im = ${I.getByOffset("i")} * ${P.get("position_id","bsnh[3]")} +\n ${I.getByOffset("j")} * ${E.get("position_id","bsnh[3]")};\n ${O.setByOffset("j","im")}\n } else {\n let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;\n ${O.setByOffset("k",I.getByOffset("k"))}\n }\n }`};return{name:"RotaryEmbedding",shaderCache:{hint:ve({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:S,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(M.size(w)/or)},programUniforms:v})}},bd=(e,t)=>{zp(e.inputs,t),e.compute(Mp(e.inputs,t))}});var Up,Vp,vd,$d=Y(()=>{"use strict";ye();Se();_e();Up=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],o=e[2];if(t.dataType!==r.dataType||t.dataType!==o.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 i=t.dims[t.dims.length-1],u=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==i)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==u)throw new Error("Skip must have the same sequence length as input");if(o.dims.length!==1)throw new Error("Gamma must be 1D");if(o.dims[o.dims.length-1]!==i)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]!==i)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]!==i)throw new Error("Bias must have the same hidden size as input")}},Vp=(e,t,r,o)=>{let i=t.simplified,u=e[0].dims,a=M.size(u),c=u,p=a,h=u.slice(-1)[0],d=o?u.slice(0,-1).concat(1):[],y=!i&&e.length>3,w=e.length>4,_=o&&r>1,v=o&&r>2,S=r>3,A=Me(h),I=[{type:12,data:p},{type:12,data:A},{type:12,data:h},{type:1,data:t.epsilon}],x=P=>{let O=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],R=[U("x",e[0].dataType,e[0].dims,A),U("skip",e[1].dataType,e[1].dims,A),U("gamma",e[2].dataType,e[2].dims,A)];y&&R.push(U("beta",e[3].dataType,e[3].dims,A)),w&&R.push(U("bias",e[4].dataType,e[4].dims,A)),R.push(j("output",e[0].dataType,c,A)),_&&R.push(j("mean_output",1,d)),v&&R.push(j("inv_std_output",1,d)),S&&R.push(j("input_skip_bias_sum",e[0].dataType,c,A));let L=De(e[0].dataType);return`\n\n ${P.registerUniforms(O).declareVariables(...R)}\n\n ${P.mainStart()}\n ${P.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size / uniforms.hidden_size")}\n let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components;\n let offset = global_idx * hidden_size_vectorized;\n var sum = ${$t("f32",A)};\n var squareSum = ${$t("f32",A)};\n for (var i: u32 = 0; i < hidden_size_vectorized; i++) {\n let skip_value = skip[offset + i];\n let bias_value = ${w?"bias[i]":L+"(0.0)"};\n let input_value = x[offset + i];\n let value = input_value + skip_value + bias_value;\n ${S?"input_skip_bias_sum[offset + i] = value;":""}\n output[offset + i] = value;\n let f32_value = ${ir(L,A,"value")};\n sum += f32_value;\n squareSum += f32_value * f32_value;\n }\n let mean = ${_t("sum",A)} / f32(uniforms.hidden_size);\n let inv_std_dev = inverseSqrt(${_t("squareSum",A)} / f32(uniforms.hidden_size) ${i?"":"- mean * mean"} + uniforms.epsilon);\n ${_?"mean_output[global_idx] = mean;":""}\n ${v?"inv_std_output[global_idx] = inv_std_dev;":""}\n for (var i: u32 = 0; i < hidden_size_vectorized; i++) {\n output[offset + i] = (output[offset + i] ${i?"":`- ${L}(mean)`}) * ${L}(inv_std_dev) * gamma[i] ${y?"+ beta[i]":""};\n }\n }`},E=[{dims:c,dataType:e[0].dataType}];return r>1&&E.push({dims:d,dataType:1}),r>2&&E.push({dims:d,dataType:1}),r>3&&E.push({dims:u,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${A};${_};${v};${S}`,inputDependencies:e.map((P,O)=>"type")},getShaderSource:x,getRunData:()=>({outputs:E,dispatchGroup:{x:Math.ceil(p/h/64)},programUniforms:I})}},vd=(e,t)=>{Up(e.inputs);let o=[0];e.outputCount>1&&o.push(-3),e.outputCount>2&&o.push(-3),e.outputCount>3&&o.push(3),e.compute(Vp(e.inputs,t,e.outputCount,!1),{outputs:o})}});var Wp,Gn,Np,_d,Gp,Hp,Sd,xd,Cd=Y(()=>{"use strict";ye();Se();Ze();_e();Wp=(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,o)=>{if(e[o+1].dataType!==6&&e[o+1].dataType!==7)throw new Error(`Input ${o} must be an array of int32 or int64`)})},Gn=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(o=>r.push(Number(o)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(o=>r.push(Number(o)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Np=(e,t)=>{if(e.length>1){let r=Gn(e,1),o=Gn(e,2),i=Gn(e,3);return i.length===0&&(i=[...Array(e[0].dims.length).keys()]),ve({starts:r,ends:o,axes:i})}else return t},_d=(e,t,r,o,i)=>{let u=e;return e<0&&(u+=r[o[t]]),i[t]<0?Math.max(0,Math.min(u,r[o[t]]-1)):Math.max(0,Math.min(u,r[o[t]]))},Gp=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {\n var input_indices: ${e.type.indices};\n var carry = 0u;\n for (var i = ${r.length}; i >= 0; i--) {\n let input_shape_i = ${fe("uniforms.input_shape","i",r.length)};\n let steps_i = ${fe("uniforms.steps","i",r.length)};\n let signs_i = ${fe("uniforms.signs","i",r.length)};\n let starts_i = ${fe("uniforms.starts","i",r.length)};\n var output_index = ${t.indicesGet("output_indices","i")};\n var input_index = output_index * steps_i + starts_i + carry;\n carry = input_index / input_shape_i;\n input_index = input_index % input_shape_i;\n if (signs_i < 0) {\n input_index = input_shape_i - input_index - 1u + starts_i;\n }\n ${e.indicesSet("input_indices","i","input_index")};\n }\n return input_indices;\n }`,Hp=(e,t)=>{let r=e[0].dims,o=M.size(r),i=t.axes.length>0?M.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],u=Gn(e,4);u.forEach(I=>I!==0||(()=>{throw new Error("step cannot be 0")})),u.length===0&&(u=Array(i.length).fill(1));let a=t.starts.map((I,x)=>_d(I,x,r,i,u)),c=t.ends.map((I,x)=>_d(I,x,r,i,u));if(i.length!==a.length||i.length!==c.length)throw new Error("start, ends and axes should have the same number of elements");if(i.length!==r.length)for(let I=0;IMath.sign(I));u.forEach((I,x,E)=>{if(I<0){let P=(c[x]-a[x])/I,O=a[x],R=O+P*u[x];a[x]=R,c[x]=O,E[x]=-I}});let h=r.slice(0);i.forEach((I,x)=>{h[I]=Math.ceil((c[I]-a[I])/u[I])});let d={dims:h,dataType:e[0].dataType},y=j("output",e[0].dataType,h.length),w=U("input",e[0].dataType,e[0].dims.length),_=M.size(h),v=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:a.length},{name:"signs",type:"i32",length:p.length},{name:"steps",type:"u32",length:u.length}],S=[{type:12,data:_},{type:12,data:a},{type:6,data:p},{type:12,data:u},...Z(e[0].dims,h)],A=I=>`\n ${I.registerUniforms(v).declareVariables(w,y)}\n ${Gp(w,y,r)}\n ${I.mainStart()}\n ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n let output_indices = ${y.offsetToIndices("global_idx")};\n let input_indices = calculateInputIndices(output_indices);\n ${y.setByOffset("global_idx",w.getByIndices("input_indices"))}\n }`;return{name:"Slice",shaderCache:{hint:`${p.length}_${a.length}_${u.length}`,inputDependencies:["rank"]},getShaderSource:A,getRunData:()=>({outputs:[d],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:S})}},Sd=(e,t)=>{Wp(e.inputs,t);let r=Np(e.inputs,t);e.compute(Hp(e.inputs,r),{inputs:[0]})},xd=e=>{let t=e.starts,r=e.ends,o=e.axes;return ve({starts:t,ends:r,axes:o})}});var Lp,Fp,Ad,Id,Td=Y(()=>{"use strict";ye();Se();Ze();_e();Lp=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},Fp=(e,t)=>{let r=e.dims,o=M.size(r),i=64,u=t.axis;if(u<0&&(u=r.length+u),uI===4?`max(max(${A}.x, ${A}.y), max(${A}.z, ${A}.w))`:I===2?`max(${A}.x, ${A}.y)`:I===3?`max(max(${A}.x, ${A}.y), ${A}.z)`:A,y=U("x",e.dataType,e.dims,p),w=j("result",e.dataType,e.dims,p),_=y.type.value,v=De(e.dataType)==="f32"?`var threadMax = ${_}(-3.402823e+38f);`:`var threadMax = ${_}(-65504.0h);`,S=A=>`\n var rowMaxShared : ${_};\n var rowSumShared : ${_};\n var threadShared : array<${_}, ${i}>;\n\n fn getValue(row: i32, col: i32, row_stride: i32) -> ${_} {\n let index = row * row_stride + col;\n return x[index];\n }\n\n fn setValue(row: i32, col: i32, row_stride: i32, value: ${_}) {\n let index = row * row_stride + col;\n result[index] = value;\n }\n ${A.registerUniform("packedCols","i32").declareVariables(y,w)}\n ${A.mainStart()}\n let gindex = i32(global_idx);\n let lindex = i32(local_idx);\n const wg = ${i};\n let row = gindex / wg;\n let cols = uniforms.packedCols;\n let row_stride : i32 = uniforms.packedCols;\n\n // find the rows max\n ${v}\n for (var col = lindex; col < cols; col += wg) {\n let value = getValue(row, col, row_stride);\n threadMax = max(threadMax, value);\n }\n if (lindex < cols) {\n threadShared[lindex] = threadMax;\n }\n workgroupBarrier();\n\n var reduceSize = min(cols, wg);\n for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) {\n reduceSize = currSize + (reduceSize & 1);\n if (lindex < currSize) {\n threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]);\n }\n workgroupBarrier();\n }\n if (lindex == 0) {\n rowMaxShared = ${_}(${d("threadShared[0]",p)});\n }\n workgroupBarrier();\n\n // find the rows sum\n var threadSum = ${_}(0.0);\n for (var col = lindex; col < cols; col += wg) {\n let subExp = exp(getValue(row, col, row_stride) - rowMaxShared);\n threadSum += subExp;\n }\n threadShared[lindex] = threadSum;\n workgroupBarrier();\n\n for (var currSize = wg >> 1; currSize > 0; currSize = currSize >> 1) {\n if (lindex < currSize) {\n threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize];\n }\n workgroupBarrier();\n }\n if (lindex == 0) {\n rowSumShared = ${_}(${_t("threadShared[0]",p)});\n }\n workgroupBarrier();\n\n // calculate final value for each element in the row\n for (var col = lindex; col < cols; col += wg) {\n let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;\n setValue(row, col, row_stride, value);\n }\n }`;return{name:"Softmax",shaderCache:{hint:`${p}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.dataType}],dispatchGroup:{x:c},programUniforms:[{type:6,data:h}]}),getShaderSource:S}},Ad=(e,t)=>{Lp(e.inputs),e.compute(Fp(e.inputs[0],t))},Id=e=>ve({axis:e.axis})});var qp,jp,Kp,Yp,Zp,Ed,Pd,kd=Y(()=>{"use strict";ye();Se();Ze();_e();qp=e=>{if(!e||e.length<1)throw new Error("too few inputs")},jp=(e,t)=>{let r=[],o=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(i=>r.push(Number(i))),o=r.length),ve({numOutputs:o,axis:t.axis,splitSizes:r})},Kp=e=>`\nfn calculateOutputIndex(index: u32) -> u32 {\n for (var i: u32 = 0u; i < ${e}u; i += 1u ) {\n if (index < ${fe("uniforms.size_in_split_axis","i",e)}) {\n return i;\n }\n }\n return ${e}u;\n}`,Yp=e=>{let t=e.length,r=[];for(let o=0;o{let r=e[0].dims,o=M.size(r),i=e[0].dataType,u=M.normalizeAxis(t.axis,r.length),a=new Array(t.numOutputs),c=U("input",i,r.length),p=new Array(t.numOutputs),h=[],d=[],y=0,w=[{type:12,data:o}];for(let v=0;v`\n ${v.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",p.length).declareVariables(c,...a)}\n ${Kp(p.length)}\n ${Yp(a)}\n\n ${v.mainStart()}\n ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}\n\n var indices = ${c.offsetToIndices("global_idx")};\n var index = ${c.indicesGet("indices",u)};\n let output_number = calculateOutputIndex(index);\n if (output_number != 0) {\n index -= ${fe("uniforms.size_in_split_axis","output_number - 1u",p.length)};\n ${c.indicesSet("indices",u,"index")};\n }\n writeBufferData(output_number, indices, global_idx);\n }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:h,dispatchGroup:{x:Math.ceil(o/64)},programUniforms:w})}},Ed=(e,t)=>{qp(e.inputs);let r=e.inputs.length===1?t:jp(e.inputs,t);e.compute(Zp(e.inputs,r),{inputs:[0]})},Pd=e=>{let t=e.axis,r=e.splitSizes,o=e.numOutputs<0?r.length:e.numOutputs;if(o!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return ve({axis:t,numOutputs:o,splitSizes:r})}});var Od,Xp,Qp,Jp,Rd,Bd=Y(()=>{"use strict";ye();Se();_e();Od=e=>Array.from(e.getBigInt64Array(),Number),Xp=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, 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(Od(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")},Qp=(e,t)=>{let r=[];for(let o=0;o{let t=e[0].dims,r=Od(e[1]),o=Qp(t,r),i=M.size(o),u=e[0].dataType,a=U("input",u,t.length),c=j("output",u,o.length),p=h=>`\n const inputShape = ${a.indices(...t)};\n ${h.registerUniform("output_size","u32").declareVariables(a,c)}\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n let output_indices = ${c.offsetToIndices("global_idx")};\n var input_indices: ${a.type.indices};\n for (var i = 0; i < ${t.length}; i++) {\n let input_dim_i = ${a.indicesGet("uniforms.input_shape","i")};\n let input_dim_value = ${c.indicesGet("output_indices","i")} % input_dim_i;\n\n ${a.indicesSet("input_indices","i","input_dim_value")}\n }\n ${c.setByOffset("global_idx",a.getByIndices("input_indices"))}\n }`;return{name:"Tile",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},...Z(e[0].dims,o)]}),getShaderSource:p}},Rd=e=>{Xp(e.inputs),e.compute(Jp(e.inputs),{inputs:[0]})}});var em,tm,Dd,zd=Y(()=>{"use strict";ye();Se();_e();em=(e,t,r,o,i)=>{let u=j("output_data",i,r.length,4),a=U("a_data",t[1].dataType,t[1].dims.length,4),c=U("b_data",t[2].dataType,t[2].dims.length,4),p=U("c_data",t[0].dataType,t[0].dims.length,4),h,d=(y,w,_)=>`select(${w}, ${y}, ${_})`;if(!o)h=u.setByOffset("global_idx",d(a.getByOffset("global_idx"),c.getByOffset("global_idx"),p.getByOffset("global_idx")));else{let y=(w,_,v="")=>{let S=`a_data[index_a${_}][component_a${_}]`,A=`b_data[index_b${_}][component_b${_}]`,I=`bool(c_data[index_c${_}] & (0xffu << (component_c${_} * 8)))`;return`\n let output_indices${_} = ${u.offsetToIndices(`global_idx * 4u + ${_}u`)};\n let offset_a${_} = ${a.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n let offset_b${_} = ${c.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n let offset_c${_} = ${p.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n let index_a${_} = offset_a${_} / 4u;\n let index_b${_} = offset_b${_} / 4u;\n let index_c${_} = offset_c${_} / 4u;\n let component_a${_} = offset_a${_} % 4u;\n let component_b${_} = offset_b${_} % 4u;\n let component_c${_} = offset_c${_} % 4u;\n ${w}[${_}] = ${v}(${d(S,A,I)});\n `};i===9?h=`\n var data = vec4(0);\n ${y("data",0,"u32")}\n ${y("data",1,"u32")}\n ${y("data",2,"u32")}\n ${y("data",3,"u32")}\n output_data[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:h=`\n ${y("output_data[global_idx]",0)}\n ${y("output_data[global_idx]",1)}\n ${y("output_data[global_idx]",2)}\n ${y("output_data[global_idx]",3)}\n `}return`\n ${e.registerUniform("vec_size","u32").declareVariables(p,a,c,u)}\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${h}\n }`},tm=e=>{let t=e[1].dims,r=e[2].dims,o=e[0].dims,i=e[1].dataType,u=!(M.areEqual(t,r)&&M.areEqual(r,o)),a=t,c=M.size(t);if(u){let h=It.calcShape(It.calcShape(t,r,!1),o,!1);if(!h)throw new Error("Can\'t perform where op on the given tensors");a=h,c=M.size(a)}let p=Math.ceil(c/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:h=>em(h,e,a,u,i),getRunData:()=>({outputs:[{dims:a,dataType:i}],dispatchGroup:{x:Math.ceil(c/64/4)},programUniforms:[{type:12,data:p},...Z(o,t,r,a)]})}},Dd=e=>{e.compute(tm(e.inputs))}});var Md,Ud=Y(()=>{"use strict";Ka();Ro();Ja();ts();Vs();Zs();Oo();Uo();lu();mu();gu();$u();xu();Au();Eu();Ou();Du();Mu();Vu();Wo();Gu();qu();Ku();cd();md();In();yd();wd();$d();Cd();Td();kd();Bd();Sr();Rn();zd();Md=new Map([["Abs",[rs]],["Acos",[ns]],["Acosh",[os]],["Add",[Ws]],["ArgMax",[ja,ko]],["ArgMin",[qa,ko]],["Asin",[is]],["Asinh",[as]],["Atan",[ss]],["Atanh",[us]],["Attention",[Xa]],["AveragePool",[rd,td]],["BatchNormalization",[Qa]],["BiasAdd",[es]],["BiasSplitGelu",[Us]],["Cast",[ls,ds]],["Ceil",[ps]],["Clip",[cs]],["Concat",[Ya,Za]],["Conv",[Ho,Go]],["ConvTranspose",[du,uu]],["Cos",[ms]],["Cosh",[fs]],["CumSum",[cu,pu]],["DepthToSpace",[fu,hu]],["Div",[Ns]],["Einsum",[wu,vu]],["Elu",[hs,kn]],["Equal",[Gs]],["Erf",[gs]],["Exp",[ys]],["Expand",[Su]],["FastGelu",[Cu]],["Floor",[bs]],["FusedConv",[Ho,Go]],["Gather",[Tu,Iu]],["GatherElements",[ku,Pu]],["Gelu",[ws]],["Gemm",[Bu,Ru]],["GlobalAveragePool",[id,od]],["GlobalMaxPool",[ld,dd]],["Greater",[qs]],["GreaterOrEqual",[Ks]],["HardSigmoid",[Is,As]],["InstanceNormalization",[zu]],["LayerNormalization",[Uu]],["LeakyRelu",[vs,kn]],["Less",[js]],["LessOrEqual",[Ys]],["Log",[zs]],["MatMul",[ru]],["MatMulNBits",[Wu,Nu]],["MaxPool",[sd,ud]],["Mul",[Hs]],["MultiHeadAttention",[Fu,Lu]],["Neg",[_s]],["Not",[$s]],["Pad",[ju]],["Pow",[Ls]],["Range",[pd]],["Reciprocal",[Ss]],["ReduceMin",[Wa]],["ReduceMean",[Da]],["ReduceMax",[Va]],["ReduceSum",[Ga]],["ReduceProd",[Na]],["ReduceL1",[za]],["ReduceL2",[Ma]],["ReduceLogSum",[La]],["ReduceLogSumExp",[Ua]],["ReduceSumSquare",[Ha]],["Relu",[xs]],["Resize",[hd,gd]],["RotaryEmbedding",[bd]],["Sigmoid",[Cs]],["Sin",[Ts]],["Sinh",[Es]],["Slice",[Sd,xd]],["SkipLayerNormalization",[vd]],["Split",[Ed,Pd]],["Sqrt",[Ps]],["Softmax",[Ad,Id]],["Sub",[Fs]],["Tan",[ks]],["Tanh",[Rs]],["ThresholdedRelu",[Ds,kn]],["Tile",[Rd]],["Transpose",[_a,Sa]],["Where",[Dd]]])});var Hn,Vd=Y(()=>{"use strict";$r();Lt();_e();Hn=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,o,i,u){rr(t.programInfo.name);let a=this.backend.device,c=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let p=[];for(let d of r)p.push({binding:p.length,resource:{buffer:d.buffer}});for(let d of o)p.push({binding:p.length,resource:{buffer:d.buffer}});u&&p.push({binding:p.length,resource:u});let h=a.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:p,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let d={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:h,dispatchGroup:i};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(d)}c.setPipeline(t.computePipeline),c.setBindGroup(0,h),c.dispatchWorkgroups(...i),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(),Ht(t.programInfo.name)}dispose(){}build(t,r){rr(t.name);let o=this.backend.device,i=[];o.features.has("shader-f16")&&i.push("enable f16;");let u=va(r,this.backend.device.limits),a=t.getShaderSource(u),c=`${i.join(`\n`)}\n${u.additionalImplementations}\n${a}`,p=o.createShaderModule({code:c,label:t.name});Ve("verbose",()=>`[WebGPU] ${t.name} shader code: ${c}`);let h=o.createComputePipeline({compute:{module:p,entryPoint:"main"},layout:"auto",label:t.name});return Ht(t.name),{programInfo:t,computePipeline:h,uniformVariablesInfo:u.variablesInfo}}normalizeDispatchGroupSize(t){let r=typeof t=="number"?t:t.x,o=typeof t=="number"?1:t.y||1,i=typeof t=="number"?1:t.z||1,u=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(r<=u&&o<=u&&i<=u)return[r,o,i];let a=r*o*i,c=Math.ceil(Math.sqrt(a));if(c>u){if(c=Math.ceil(Math.cbrt(a)),c>u)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[c,c,c]}else return[c,c,1]}}});var rm,nm,Yo,Ln,Wd=Y(()=>{"use strict";$r();ye();Lt();fa();wa();Ud();Vd();rm=(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 o=0;o{let o=e.name;return e.shaderCache?.hint&&(o+="["+e.shaderCache.hint+"]"),o+=":"+r+`:${rm(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,o},Yo=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},Ln=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 o=[],i={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:o};r.features.has("chromium-experimental-timestamp-query-inside-passes")?o.push("chromium-experimental-timestamp-query-inside-passes"):r.features.has("timestamp-query")&&o.push("timestamp-query"),r.features.has("shader-f16")&&o.push("shader-f16"),this.device=await r.requestDevice(i),this.adapterInfo=new Yo(await r.requestAdapterInfo()),this.gpuDataManager=ba(this),this.programManager=new Hn(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,pa(t.logLevel,!!t.debug),this.device.onuncapturederror=u=>{u.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${u.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;rr(),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()),o=this.pendingQueries.get(t);for(let i=0;i"u"&&(this.queryTimeBase=_);let S=Number(_-this.queryTimeBase),A=Number(v-this.queryTimeBase);if(!Number.isSafeInteger(S)||!Number.isSafeInteger(A))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:y.map(I=>({dims:I.dims,dataType:Gt(I.dataType)})),outputsMetadata:w.map(I=>({dims:I.dims,dataType:Gt(I.dataType)})),kernelId:a,kernelType:p,kernelName:h,programName:d,startTime:S,endTime:A});else{let I="";y.forEach((E,P)=>{I+=`input[${P}]: [${E.dims}] | ${Gt(E.dataType)}, `});let x="";w.forEach((E,P)=>{x+=`output[${P}]: [${E.dims}] | ${Gt(E.dataType)}, `}),console.log(`[profiling] kernel "${a}|${p}|${h}|${d}" ${I}${x}execution time: ${A-S} ns`)}wo("GPU",`${d}::${_}::${v}`)}t.unmap(),this.pendingQueries.delete(t)}),Ht()}run(t,r,o,i,u,a){rr(t.name);let c=[];for(let E=0;EP):o;if(y.length!==p.length)throw new Error(`Output size ${y.length} must be equal to ${p.length}.`);let w=[],_=[];for(let E=0;E=a)throw new Error(`Invalid output index: ${y[E]}`);if(y[E]===-3)continue;let P=y[E]===-1,O=y[E]===-2,R=P||O?u(p[E].dataType,p[E].dims):i(y[E],p[E].dataType,p[E].dims);if(w.push(R),R.data===0)continue;let L=this.gpuDataManager.get(R.data);if(!L)throw new Error(`no GPU data for output: ${R.data}`);if(P&&this.temporaryData.push(L),O){let N=this.kernelPersistentData.get(this.currentKernelId);N||(N=[],this.kernelPersistentData.set(this.currentKernelId,N)),N.push(L)}_.push(L)}if(c.length!==r.length||_.length!==w.length){if(_.length===0)return Ht(t.name),w;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let v;if(d){let E=0,P=[];d.forEach(N=>{let K=typeof N.data=="number"?[N.data]:N.data;if(K.length===0)return;let Q=N.type===10?2:4,he,W;N.type===10?(W=K.length>4?16:K.length>2?8:K.length*Q,he=K.length>4?16:Q*K.length):(W=K.length<=2?K.length*Q:16,he=16),E=Math.ceil(E/W)*W,P.push(E);let se=N.type===10?8:4;E+=K.length>4?Math.ceil(K.length/se)*he:K.length*Q});let O=16;E=Math.ceil(E/O)*O;let R=new ArrayBuffer(E);d.forEach((N,K)=>{let Q=P[K],he=typeof N.data=="number"?[N.data]:N.data;if(N.type===6)new Int32Array(R,Q,he.length).set(he);else if(N.type===12)new Uint32Array(R,Q,he.length).set(he);else if(N.type===10)new Uint16Array(R,Q,he.length).set(he);else if(N.type===1)new Float32Array(R,Q,he.length).set(he);else throw new Error(`Unsupported uniform type: ${Gt(N.type)}`)});let L=this.gpuDataManager.create(E,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(L.buffer,0,R,0,E),this.gpuDataManager.release(L.id),v={offset:0,size:E,buffer:L.buffer}}let S=this.programManager.normalizeDispatchGroupSize(h),A=S[1]===1&&S[2]===1,I=nm(t,r,A),x=this.programManager.getArtifact(I);if(x||(x=this.programManager.build(t,S),this.programManager.setArtifact(I,x),Ve("info",()=>`[artifact] key: ${I}, programName: ${t.name}`)),d&&x.uniformVariablesInfo){if(d.length!==x.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${x.uniformVariablesInfo.length}, got ${d.length} in program "${x.programInfo.name}".`);for(let E=0;E`[ProgramManager] run "${t.name}" (key=${I}) with ${S[0]}x${S[1]}x${S[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let E={kernelId:this.currentKernelId,programName:x.programInfo.name,inputTensorViews:r,outputTensorViews:w};this.pendingKernels.push(E),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(E)}return this.programManager.run(x,c,_,S,v),Ht(t.name),w}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,o,i){let u=Md.get(t);if(!u)throw new Error(`kernel not implemented: ${t}`);let a={kernelType:t,kernelName:i,kernelEntry:u[0],attributes:[u[1],o]};this.kernels.set(r,a)}releaseKernel(t){let r=this.kernelPersistentData.get(t);if(r){for(let o of r)this.gpuDataManager.release(o.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,r,o){let i=this.kernels.get(t);if(!i)throw new Error(`kernel not created: ${t}`);let u=i.kernelType,a=i.kernelName,c=i.kernelEntry,p=i.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${u}] ${a}" is not allowed to be called recursively`);this.currentKernelId=t,p[0]&&(p[1]=p[0](p[1]),p[0]=void 0),Ve("info",()=>`[WebGPU] Start to run kernel "[${u}] ${a}"...`);let h=this.env.debug;this.temporaryData=[];try{return h&&this.device.pushErrorScope("validation"),c(r,p[1]),0}catch(d){return o.push(Promise.resolve(`[WebGPU] Kernel "[${u}] ${a}" failed. ${d}`)),1}finally{h&&o.push(this.device.popErrorScope().then(d=>d?`GPU validation error for kernel "[${u}] ${a}": ${d.message}`:null));for(let d of this.temporaryData)this.gpuDataManager.release(d.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,r,o,i){let u=this.sessionExternalDataMapping.get(t);u||(u=new Map,this.sessionExternalDataMapping.set(t,u));let a=u.get(r),c=this.gpuDataManager.registerExternalBuffer(o,i,a?.[1]);return u.set(r,[c,o]),c}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(o=>this.gpuDataManager.unregisterExternalBuffer(o[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,o){return async()=>{let i=await xo(this,t,r);return ma(i.buffer,o)}}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),o=t.length;this.pendingKernels=[];for(let i=0;i=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 Nd={};gn(Nd,{init:()=>om});var qr,Zo,om,Gd=Y(()=>{"use strict";ye();Wd();Lt();Se();qr=class e{constructor(t,r,o,i){this.module=t;this.dataType=r;this.data=o;this.dims=i}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=M.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=M.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=M.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(M.size(t)!==M.size(this.dims))throw new Error("Invalid new shape");return new e(this.module,this.dataType,this.data,t)}},Zo=class{constructor(t,r,o){this.module=t;this.backend=r;this.customDataOffset=0;this.customDataSize=0;this.adapterInfo=r.adapterInfo;let i=t.HEAPU32,u=o>>>2;this.opKernelContext=i[u++];let a=i[u++];this.outputCount=i[u++],this.customDataOffset=i[u++],this.customDataSize=i[u++];let c=[];for(let p=0;ptypeof c=="number"?this.inputs[c]:c)??this.inputs,i=r?.outputs??[],u=(c,p,h)=>new qr(this.module,p,this.output(c,h),h),a=(c,p)=>{let h=tr(c);if(!h)throw new Error(`Unsupported data type: ${c}`);let d=h*M.size(p),y=d>0?this.backend.gpuDataManager.create(d).id:0;return new qr(this.module,c,y,p)};return this.backend.run(t,o,i,u,a,this.outputCount)}output(t,r){let o=this.module.stackSave();try{let i=this.module.stackAlloc((1+r.length)*4),u=i>>2;this.module.HEAPU32[u++]=r.length;for(let a=0;a{let i=t.jsepInit;if(!i)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let u=new Ln;await u.initialize(r,o),i("webgpu",[u,a=>u.alloc(a),a=>u.free(a),(a,c,p,h=!1)=>{if(h)Ve("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${a}, dst=${c}, size=${p}`),u.memcpy(a,c);else{Ve("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${a}, gpuDataId=${c}, size=${p}`);let d=t.HEAPU8.subarray(a>>>0,(a>>>0)+p);u.upload(c,d)}},async(a,c,p)=>{Ve("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${a}, dataOffset=${c}, size=${p}`),await u.download(a,()=>t.HEAPU8.subarray(c>>>0,(c>>>0)+p))},(a,c,p)=>u.createKernel(a,c,p,t.UTF8ToString(t._JsepGetNodeName(c))),a=>u.releaseKernel(a),(a,c,p,h)=>{Ve("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${p}, kernel=${a}, contextDataOffset=${c}`);let d=new Zo(t,u,c);return u.computeKernel(a,d,h)},()=>u.captureBegin(),()=>u.captureEnd(),()=>u.replay()])}else i("webnn")}});var Vi;Vi=Ei();var pl=zi(),fo,ho=!1,yn=!1,Ui=!1,ml=e=>{if(e===1)return!1;if(typeof SharedArrayBuffer>"u")return typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+e+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),!1;typeof process<"u"&&process.versions&&process.versions.node&&console.warn("env.wasm.numThreads is set to "+e+", however, currently onnxruntime-web does not support multi-threads in Node.js. Please consider using onnxruntime-node for performance critical scenarios.");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}},fl=()=>{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}},hl=(e,t)=>e?t?"ort-wasm-simd-threaded.wasm":"ort-wasm-simd.wasm":t?"ort-wasm-threaded.wasm":"ort-wasm.wasm",Wi=async e=>{if(ho)return Promise.resolve();if(yn)throw new Error("multiple calls to \'initializeWebAssembly()\' detected.");if(Ui)throw new Error("previous call to \'initializeWebAssembly()\' failed.");yn=!0;let t=e.initTimeout,r=e.numThreads,o=e.simd,i=ml(r),u=o&&fl(),a=e.wasmPaths,c=typeof a=="string"?a:void 0,p=hl(u,i),h=typeof a=="object"?a[p]:void 0,d=!1,y=[];if(t>0&&y.push(new Promise(w=>{setTimeout(()=>{d=!0,w()},t)})),y.push(new Promise((w,_)=>{let v=i?pl:Vi,S={locateFile:(A,I)=>{if(i&&A.endsWith(".worker.js")&&typeof Blob<"u")return URL.createObjectURL(new Blob([Mi()],{type:"text/javascript"}));if(A.endsWith(".wasm")){if(h)return h;let x=c??I;return p==="ort-wasm-simd.wasm"?x+"ort-wasm-simd.jsep.wasm":p==="ort-wasm-simd-threaded.wasm"?x+"ort-wasm-simd-threaded.jsep.wasm":x+p}return I+A}};if(i)if(S.numThreads=r,typeof Blob>"u")S.mainScriptUrlOrBlob=(void 0)(__dirname,"ort-wasm-threaded.js");else{let A=`var ortWasmThreaded=${v.toString()};`;S.mainScriptUrlOrBlob=new Blob([A],{type:"text/javascript"})}v(S).then(A=>{yn=!1,ho=!0,fo=A,w()},A=>{yn=!1,Ui=!0,_(A)})})),await Promise.race(y),d)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Le=()=>{if(ho&&fo)return fo;throw new Error("WebAssembly is not initialized yet.")};var Fe=(e,t)=>{let r=Le(),o=r.lengthBytesUTF8(e)+1,i=r._malloc(o);return r.stringToUTF8(e,i,o),t.push(i),i},Nr=(e,t,r,o)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([i,u])=>{let a=t?t+i:i;if(typeof u=="object")Nr(u,a+".",r,o);else if(typeof u=="string"||typeof u=="number")o(a,u.toString());else if(typeof u=="boolean")o(a,u?"1":"0");else throw new Error(`Can\'t handle extra config type: ${typeof u}`)})},Ue=e=>{let t=Le(),r=t.stackSave();try{let o=t.stackAlloc(8);t._OrtGetLastError(o,o+4);let i=t.HEAP32[o/4],u=t.HEAPU32[o/4+1],a=u?t.UTF8ToString(u):"";throw new Error(`${e} ERROR_CODE: ${i}, ERROR_MESSAGE: ${a}`)}finally{t.stackRestore(r)}};var Ni=e=>{let t=Le(),r=0,o=[],i=e||{};try{if(e?.logSeverityLevel===void 0)i.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)i.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&&(i.terminate=!1);let u=0;return e?.tag!==void 0&&(u=Fe(e.tag,o)),r=t._OrtCreateRunOptions(i.logSeverityLevel,i.logVerbosityLevel,!!i.terminate,u),r===0&&Ue("Can\'t create run options."),e?.extra!==void 0&&Nr(e.extra,"",new WeakSet,(a,c)=>{let p=Fe(a,o),h=Fe(c,o);t._OrtAddRunConfigEntry(r,p,h)!==0&&Ue(`Can\'t set a run config entry: ${a} - ${c}.`)}),[r,o]}catch(u){throw r!==0&&t._OrtReleaseRunOptions(r),o.forEach(a=>t._free(a)),u}};var gl=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}`)}},yl=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},bl=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)},wl=(e,t,r)=>{for(let o of t){let i=typeof o=="string"?o:o.name;switch(i){case"webnn":if(i="WEBNN",typeof o!="string"){let a=o;if(a?.deviceType){let c=Fe("deviceType",r),p=Fe(a.deviceType,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'deviceType\' - ${a.deviceType}.`)}if(a?.numThreads){let c=a.numThreads;(typeof c!="number"||!Number.isInteger(c)||c<0)&&(c=0);let p=Fe("numThreads",r),h=Fe(c.toString(),r);Le()._OrtAddSessionConfigEntry(e,p,h)!==0&&Ue(`Can\'t set a session config entry: \'numThreads\' - ${a.numThreads}.`)}if(a?.powerPreference){let c=Fe("powerPreference",r),p=Fe(a.powerPreference,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'powerPreference\' - ${a.powerPreference}.`)}}break;case"webgpu":if(i="JS",typeof o!="string"){let a=o;if(a?.preferredLayout){if(a.preferredLayout!=="NCHW"&&a.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either \'NCHW\' or \'NHWC\': ${a.preferredLayout}`);let c=Fe("preferredLayout",r),p=Fe(a.preferredLayout,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'preferredLayout\' - ${a.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${i}`)}let u=Fe(i,r);Le()._OrtAppendExecutionProvider(e,u)!==0&&Ue(`Can\'t append execution provider: ${i}.`)}},Gi=e=>{let t=Le(),r=0,o=[],i=e||{};bl(i);try{let u=gl(i.graphOptimizationLevel??"all"),a=yl(i.executionMode??"sequential"),c=typeof i.logId=="string"?Fe(i.logId,o):0,p=i.logSeverityLevel??2;if(!Number.isInteger(p)||p<0||p>4)throw new Error(`log serverity level is not valid: ${p}`);let h=i.logVerbosityLevel??0;if(!Number.isInteger(h)||h<0||h>4)throw new Error(`log verbosity level is not valid: ${h}`);let d=typeof i.optimizedModelFilePath=="string"?Fe(i.optimizedModelFilePath,o):0;if(r=t._OrtCreateSessionOptions(u,!!i.enableCpuMemArena,!!i.enableMemPattern,a,!!i.enableProfiling,0,c,p,h,d),r===0&&Ue("Can\'t create session options."),i.executionProviders&&wl(r,i.executionProviders,o),i.enableGraphCapture!==void 0){if(typeof i.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${i.enableGraphCapture}`);let y=Fe("enableGraphCapture",o),w=Fe(i.enableGraphCapture.toString(),o);t._OrtAddSessionConfigEntry(r,y,w)!==0&&Ue(`Can\'t set a session config entry: \'enableGraphCapture\' - ${i.enableGraphCapture}.`)}if(i.freeDimensionOverrides)for(let[y,w]of Object.entries(i.freeDimensionOverrides)){if(typeof y!="string")throw new Error(`free dimension override name must be a string: ${y}`);if(typeof w!="number"||!Number.isInteger(w)||w<0)throw new Error(`free dimension override value must be a non-negative integer: ${w}`);let _=Fe(y,o);t._OrtAddFreeDimensionOverride(r,_,w)!==0&&Ue(`Can\'t set a free dimension override: ${y} - ${w}.`)}return i.extra!==void 0&&Nr(i.extra,"",new WeakSet,(y,w)=>{let _=Fe(y,o),v=Fe(w,o);t._OrtAddSessionConfigEntry(r,_,v)!==0&&Ue(`Can\'t set a session config entry: ${y} - ${w}.`)}),[r,o]}catch(u){throw r!==0&&t._OrtReleaseSessionOptions(r),o.forEach(a=>t._free(a)),u}};ye();var Li=async e=>{if(typeof e=="string")if(typeof process<"u"&&process.versions&&process.versions.node)try{return new Uint8Array(await(void 0)(e))}catch(t){if(t.code==="ERR_FS_FILE_TOO_LARGE"){let r=(void 0)(e),o=[];for await(let i of r)o.push(i);return new Uint8Array(Buffer.concat(o))}throw t}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"),o=r?parseInt(r,10):0;if(o<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let i=t.body.getReader(),u;try{u=new ArrayBuffer(o)}catch(c){if(c instanceof RangeError){let p=Math.ceil(o/65536);u=new WebAssembly.Memory({initial:p,maximum:p}).buffer}else throw c}let a=0;for(;;){let{done:c,value:p}=await i.read();if(c)break;let h=p.byteLength;new Uint8Array(u,a,h).set(p),a+=h}return new Uint8Array(u,0,o)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)};var im=(e,t)=>{Le()._OrtInit(e,t)!==0&&Ue("Can\'t initialize onnxruntime.")},Ld=async e=>{im(e.wasm.numThreads,Gr(e.logLevel))},Fd=async(e,t)=>{{let r=(Gd(),wr(Nd)).init;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let o=e.webgpu.adapter;if(o){if(typeof o.limits!="object"||typeof o.features!="object"||typeof o.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let i=e.webgpu.powerPreference;if(i!==void 0&&i!=="low-power"&&i!=="high-performance")throw new Error(`Invalid powerPreference setting: "${i}"`);let u=e.webgpu.forceFallbackAdapter;if(u!==void 0&&typeof u!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${u}"`);if(o=await navigator.gpu.requestAdapter({powerPreference:i,forceFallbackAdapter:u}),!o)throw new Error(\'Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.\')}if(!e.wasm.simd)throw new Error("Not supported for WebGPU=ON and SIMD=OFF. Please set `env.wasm.simd` to true when using `webgpu` EP");await r("webgpu",Le(),e,o)}if(t==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",Le(),e)}}},ar=new Map,am=e=>{let t=Le(),r=t.stackSave();try{let o=t.stackAlloc(8);return t._OrtGetInputOutputCount(e,o,o+4)!==0&&Ue("Can\'t get session input/output count."),[t.HEAP32[o/4],t.HEAP32[o/4+1]]}finally{t.stackRestore(r)}},Xo=e=>{let t=Le(),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]},qd=async(e,t)=>{let r,o,i=Le();Array.isArray(e)?[r,o]=e:e.buffer===i.HEAPU8.buffer?[r,o]=[e.byteOffset,e.byteLength]:[r,o]=Xo(e);let u=0,a=0,c=0,p=[],h=[],d=[];try{if([a,p]=Gi(t),t?.externalData&&i.mountExternalData){let x=[];for(let E of t.externalData){let P=typeof E=="string"?E:E.path;x.push(Li(typeof E=="string"?E:E.data).then(O=>{i.mountExternalData(P,O)}))}await Promise.all(x)}u=await i._OrtCreateSession(r,o,a),u===0&&Ue("Can\'t create a session.");let[y,w]=am(u),_=!!t?.enableGraphCapture,v=[],S=[],A=[];for(let x=0;xx==="gpu-buffer")&&(c=i._OrtCreateBinding(u),c===0&&Ue("Can\'t create IO binding."),I={handle:c,outputPreferredLocations:A,outputPreferredLocationsEncoded:A.map(x=>yo(x))}),ar.set(u,[u,h,d,I,_,!1]),[u,v,S]}catch(y){throw h.forEach(w=>i._OrtFree(w)),d.forEach(w=>i._OrtFree(w)),c!==0&&i._OrtReleaseBinding(c),u!==0&&i._OrtReleaseSession(u),y}finally{i._free(r),a!==0&&i._OrtReleaseSessionOptions(a),p.forEach(y=>i._free(y)),i.unmountExternalData?.()}},jd=e=>{let t=Le(),r=ar.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[o,i,u,a,c]=r;a&&(c&&t._OrtClearBoundOutputs(a.handle),t._OrtReleaseBinding(a.handle)),t.jsepOnReleaseSession?.(e),i.forEach(p=>t._OrtFree(p)),u.forEach(p=>t._OrtFree(p)),t._OrtReleaseSession(o),ar.delete(e)},Hd=(e,t,r,o,i,u=!1)=>{if(!e){t.push(0);return}let a=Le(),c=e[0],p=e[1],h=e[3],d,y;if(c==="string"&&h==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(u&&h!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${i} when enableGraphCapture is true.`);if(h==="gpu-buffer"){let v=e[2].gpuBuffer,S=tr(go(c));y=p.reduce((I,x)=>I*x,1)*S;let A=a.jsepRegisterBuffer;if(!A)throw new Error(\'Tensor location "gpu-buffer" is not supported without using WebGPU.\');d=A(o,i,v,y)}else{let v=e[2];if(Array.isArray(v)){y=4*v.length,d=a._malloc(y),r.push(d);let S=d/4;for(let A=0;Aa.HEAP32[v++]=A);let S=a._OrtCreateTensor(go(c),d,y,_,p.length,yo(h));S===0&&Ue(`Can\'t create tensor for input/output. session=${o}, index=${i}.`),t.push(S)}finally{a.stackRestore(w)}},Kd=async(e,t,r,o,i,u)=>{let a=Le(),c=ar.get(e);if(!c)throw new Error(`cannot run inference. invalid session id: ${e}`);let p=c[0],h=c[1],d=c[2],y=c[3],w=c[4],_=c[5],v=t.length,S=o.length,A=0,I=[],x=[],E=[],P=[],O=a.stackSave(),R=a.stackAlloc(v*4),L=a.stackAlloc(v*4),N=a.stackAlloc(S*4),K=a.stackAlloc(S*4);try{[A,I]=Ni(u);for(let ee=0;eeNe*Ye,1);ue=Gt(ne);let Ut=y?.outputPreferredLocations[o[ee]];if(ue==="string"){if(Ut==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Ne=[],Ye=le/4;for(let mt=0;mt0){let Ne=a.jsepGetBuffer;if(!Ne)throw new Error(\'preferredLocation "gpu-buffer" is not supported without using WebGPU.\');let Ye=Ne(le),mt=tr(ne);if(mt===void 0||!Hi(ue))throw new Error(`Unsupported data type: ${ue}`);ie=!0,We.push([ue,Be,{gpuBuffer:Ye,download:a.jsepCreateDownloader(Ye,Ge*mt,ue),dispose:()=>{a._OrtReleaseTensor(ae)}},"gpu-buffer"])}else{let Ne=bn(ue),Ye=new Ne(Ge);new Uint8Array(Ye.buffer,Ye.byteOffset,Ye.byteLength).set(a.HEAPU8.subarray(le,le+Ye.byteLength)),We.push([ue,Be,Ye,"cpu"])}}finally{a.stackRestore(Ae),ue==="string"&&le&&a._free(le),ie||a._OrtReleaseTensor(ae)}}return y&&!w&&(a._OrtClearBoundOutputs(y.handle),ar.set(e,[p,h,d,y,w,!1])),We}finally{a.stackRestore(O),x.forEach(Q=>a._OrtReleaseTensor(Q)),E.forEach(Q=>a._OrtReleaseTensor(Q)),P.forEach(Q=>a._free(Q)),A!==0&&a._OrtReleaseRunOptions(A),I.forEach(Q=>a._free(Q))}},Yd=e=>{let t=Le(),r=ar.get(e);if(!r)throw new Error("invalid session id");let o=r[0],i=t._OrtEndProfiling(o);i===0&&Ue("Can\'t get an profile file name."),t._OrtFree(i)},Zd=e=>{let t=[];for(let r of e){let o=r[2];!Array.isArray(o)&&"buffer"in o&&t.push(o.buffer)}return t};self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":Wi(r.wasm).then(()=>{Ld(r).then(()=>{postMessage({type:t})},o=>{postMessage({type:t,err:o})})},o=>{postMessage({type:t,err:o})});break;case"init-ep":{let{epName:o,env:i}=r;Fd(i,o).then(()=>{postMessage({type:t})},u=>{postMessage({type:t,err:u})});break}case"copy-from":{let{buffer:o}=r,i=Xo(o);postMessage({type:t,out:i});break}case"create":{let{model:o,options:i}=r;qd(o,i).then(u=>{postMessage({type:t,out:u})},u=>{postMessage({type:t,err:u})});break}case"release":jd(r),postMessage({type:t});break;case"run":{let{sessionId:o,inputIndices:i,inputs:u,outputIndices:a,options:c}=r;Kd(o,i,u,a,new Array(a.length).fill(null),c).then(p=>{p.some(h=>h[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:p},Zd([...u,...p]))},p=>{postMessage({type:t,err:p})});break}case"end-profiling":Yd(r),postMessage({type:t});break;default:}}catch(o){postMessage({type:t,err:o})}};})();\n/**\n * @license\n * Copyright 2021 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n/**\n * @license\n * Copyright 2020 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n/**\n * @license\n * Copyright 2019 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n'}),Pr,Vt,En,ua,da,oo,Ci,jr,qr,lp,ca,tg,rg,ng,ag,ig,sg,og,lg=J(()=>{dr(),iw(),Aa(),Pr=()=>!!Ue.wasm.proxy&&typeof document<"u",En=!1,ua=!1,da=!1,Ci=new Map,jr=(t,e)=>{let r=Ci.get(t);r?r.push(e):Ci.set(t,[e])},qr=()=>{if(En||!ua||da||!Vt)throw new Error("worker not ready")},lp=t=>{switch(t.data.type){case"init-wasm":En=!1,t.data.err?(da=!0,oo[1](t.data.err)):(ua=!0,oo[0]());break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let e=Ci.get(t.data.type);t.data.err?e.shift()[1](t.data.err):e.shift()[0](t.data.out);break}}},ca=typeof document<"u"?document?.currentScript?.src:void 0,tg=async()=>{if(!ua){if(En)throw new Error("multiple calls to 'initWasm()' detected.");if(da)throw new Error("previous call to 'initWasm()' failed.");if(En=!0,Pr())return Ue.wasm.wasmPaths===void 0&&ca&&ca.indexOf("blob:")!==0&&(Ue.wasm.wasmPaths=ca.substr(0,+ca.lastIndexOf("/")+1)),new Promise((t,e)=>{Vt?.terminate();let r=URL.createObjectURL(new Blob([sw()],{type:"text/javascript"}));Vt=new Worker(r,{name:"ort-wasm-proxy-worker"}),Vt.onerror=a=>e(a),Vt.onmessage=lp,URL.revokeObjectURL(r),oo=[t,e];let n={type:"init-wasm",in:Ue};Vt.postMessage(n)});try{await hh(Ue.wasm),await Km(Ue),ua=!0}catch(t){throw da=!0,t}finally{En=!1}}},rg=async t=>{if(Pr())return qr(),new Promise((e,r)=>{jr("init-ep",[e,r]);let n={type:"init-ep",in:{epName:t,env:Ue}};Vt.postMessage(n)});await Ym(Ue,t)},ng=async t=>Pr()?(qr(),new Promise((e,r)=>{jr("copy-from",[e,r]);let n={type:"copy-from",in:{buffer:t}};Vt.postMessage(n,[t.buffer])})):Oo(t),ag=async(t,e)=>{if(Pr()){if(e?.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return qr(),new Promise((r,n)=>{jr("create",[r,n]);let a={type:"create",in:{model:t,options:{...e}}},i=[];t instanceof Uint8Array&&i.push(t.buffer),Vt.postMessage(a,i)})}else return Xm(t,e)},ig=async t=>{if(Pr())return qr(),new Promise((e,r)=>{jr("release",[e,r]);let n={type:"release",in:t};Vt.postMessage(n)});Qm(t)},sg=async(t,e,r,n,a,i)=>{if(Pr()){if(r.some(s=>s[3]!=="cpu"))throw new Error("input tensor on GPU is not supported for proxy.");if(a.some(s=>s))throw new Error("pre-allocated output tensor is not supported for proxy.");return qr(),new Promise((s,o)=>{jr("run",[s,o]);let l=r,u={type:"run",in:{sessionId:t,inputIndices:e,inputs:l,outputIndices:n,options:i}};Vt.postMessage(u,eg(l))})}else return Zm(t,e,r,n,a,i)},og=async t=>{if(Pr())return qr(),new Promise((e,r)=>{jr("end-profiling",[e,r]);let n={type:"end-profiling",in:t};Vt.postMessage(n)});Jm(t)}}),lo,up,ug,ow=J(()=>{dr(),lg(),$e(),gh(),lo=(t,e)=>{switch(t.location){case"cpu":return[t.type,t.dims,t.data,"cpu"];case"gpu-buffer":return[t.type,t.dims,{gpuBuffer:t.gpuBuffer},"gpu-buffer"];default:throw new Error(`invalid data location: ${t.location} for ${e()}`)}},up=t=>{switch(t[3]){case"cpu":return new Et(t[0],t[2],t[1]);case"gpu-buffer":{let e=t[0];if(!Yo(e))throw new Error(`not supported data type: ${e} for deserializing GPU tensor`);let{gpuBuffer:r,download:n,dispose:a}=t[2];return Et.fromGpuBuffer(r,{dataType:e,dims:t[1],download:n,dispose:a})}default:throw new Error(`invalid data location: ${t[3]}`)}},ug=class{async fetchModelAndCopyToWasmMemory(t){return ng(await Gi(t))}async loadModel(t,e){ur();let r;typeof t=="string"?typeof process<"u"&&process.versions&&process.versions.node?r=await Gi(t):r=await this.fetchModelAndCopyToWasmMemory(t):r=t,[this.sessionId,this.inputNames,this.outputNames]=await ag(r,e),jt()}async dispose(){return ig(this.sessionId)}async run(t,e,r){ur();let n=[],a=[];Object.entries(t).forEach(h=>{let m=h[0],d=h[1],_=this.inputNames.indexOf(m);if(_===-1)throw new Error(`invalid input '${m}'`);n.push(d),a.push(_)});let i=[],s=[];Object.entries(e).forEach(h=>{let m=h[0],d=h[1],_=this.outputNames.indexOf(m);if(_===-1)throw new Error(`invalid output '${m}'`);i.push(d),s.push(_)});let o=n.map((h,m)=>lo(h,()=>`input "${this.inputNames[a[m]]}"`)),l=i.map((h,m)=>h?lo(h,()=>`output "${this.outputNames[s[m]]}"`):null),u=await sg(this.sessionId,a,o,s,l,r),p={};for(let h=0;h{dr(),lg(),ow(),dp=()=>{if((typeof Ue.wasm.initTimeout!="number"||Ue.wasm.initTimeout<0)&&(Ue.wasm.initTimeout=0),typeof Ue.wasm.simd!="boolean"&&(Ue.wasm.simd=!0),typeof Ue.wasm.proxy!="boolean"&&(Ue.wasm.proxy=!1),typeof Ue.wasm.trace!="boolean"&&(Ue.wasm.trace=!1),typeof Ue.wasm.numThreads!="number"||!Number.isInteger(Ue.wasm.numThreads)||Ue.wasm.numThreads<=0){(typeof self<"u"&&!self.crossOriginIsolated||typeof process<"u"&&process.versions&&process.versions.node)&&(Ue.wasm.numThreads=1);let t=typeof navigator>"u"?(void 0)().length:navigator.hardwareConcurrency;Ue.wasm.numThreads=Math.min(4,Math.ceil((t||1)/2))}},dg=class{async init(t){dp(),await tg(),await rg(t)}async createInferenceSessionHandler(t,e){let r=new ug;return await r.loadModel(t,e),Promise.resolve(r)}}}),cg={};Dn(cg,{wasmBackend:()=>pg});var pg,uw=J(()=>{lw(),pg=new dg});dr();dr();dr();var dw="1.18.0",cw=ah;{let t=(uw(),nn(cg)).wasmBackend;tn("webgpu",t,5),tn("webnn",t,5),tn("cpu",t,10),tn("wasm",t,10)}Object.defineProperty(Ue.versions,"web",{value:dw,enumerable:!0});/** +* @license +* Copyright 2021 Google LLC. All Rights Reserved. +* Licensed under the Apache License, Version 2.0 (the "License"); +* you may not use this file except in compliance with the License. +* You may obtain a copy of the License at +* +* http://www.apache.org/licenses/LICENSE-2.0 +* +* Unless required by applicable law or agreed to in writing, software +* distributed under the License is distributed on an "AS IS" BASIS, +* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +* See the License for the specific language governing permissions and +* limitations under the License. +* ============================================================================= +*//** + * @license + * Copyright 2020 Google LLC. All Rights Reserved. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * ============================================================================= + *//** + * @license + * Copyright 2019 Google LLC. All Rights Reserved. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * ============================================================================= + */const pw=Object.freeze(Object.defineProperty({__proto__:null,get InferenceSession(){return Vo},get TRACE(){return va},get TRACE_FUNC_BEGIN(){return ur},get TRACE_FUNC_END(){return jt},get Tensor(){return Et},get TrainingSession(){return Go},default:cw,get env(){return Ue},get registerBackend(){return tn}},Symbol.toStringTag,{value:"Module"})),hw=(t,e)=>{const r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=t.dims[3],r.height=t.dims[2];const n=r.getContext("2d");if(n!=null){let a,i;e?.tensorLayout!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],i=t.dims[3]):(a=t.dims[3],i=t.dims[2]);const s=e?.format!==void 0?e.format:"RGB",o=e?.norm;let l,u;o===void 0||o.mean===void 0?l=[255,255,255,255]:typeof o.mean=="number"?l=[o.mean,o.mean,o.mean,o.mean]:(l=[o.mean[0],o.mean[1],o.mean[2],0],o.mean[3]!==void 0&&(l[3]=o.mean[3])),o===void 0||o.bias===void 0?u=[0,0,0,0]:typeof o.bias=="number"?u=[o.bias,o.bias,o.bias,o.bias]:(u=[o.bias[0],o.bias[1],o.bias[2],0],o.bias[3]!==void 0&&(u[3]=o.bias[3]));const p=i*a;let h=0,m=p,d=p*2,_=-1;s==="RGBA"?(h=0,m=p,d=p*2,_=p*3):s==="RGB"?(h=0,m=p,d=p*2):s==="RBG"&&(h=0,d=p,m=p*2);for(let w=0;w{const r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d");let n;if(r!=null){let a,i,s;e?.tensorLayout!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],i=t.dims[1],s=t.dims[3]):(a=t.dims[3],i=t.dims[2],s=t.dims[1]);const o=e!==void 0&&e.format!==void 0?e.format:"RGB",l=e?.norm;let u,p;l===void 0||l.mean===void 0?u=[255,255,255,255]:typeof l.mean=="number"?u=[l.mean,l.mean,l.mean,l.mean]:(u=[l.mean[0],l.mean[1],l.mean[2],255],l.mean[3]!==void 0&&(u[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]));const h=i*a;if(e!==void 0&&(e.format!==void 0&&s===4&&e.format!=="RGBA"||s===3&&e.format!=="RGB"&&e.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");const m=4;let d=0,_=1,w=2,v=3,S=0,$=h,E=h*2,T=-1;o==="RGBA"?(S=0,$=h,E=h*2,T=h*3):o==="RGB"?(S=0,$=h,E=h*2):o==="RBG"&&(S=0,E=h,$=h*2),n=r.createImageData(a,i);for(let A=0;A{if(t===void 0)throw new Error("Image buffer must be defined");if(e.height===void 0||e.width===void 0)throw new Error("Image height and width must be defined");if(e.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");const{height:r,width:n}=e,a=e.norm??{mean:255,bias:0};let i,s;typeof a.mean=="number"?i=[a.mean,a.mean,a.mean,a.mean]:i=[a.mean[0],a.mean[1],a.mean[2],a.mean[3]??255],typeof a.bias=="number"?s=[a.bias,a.bias,a.bias,a.bias]:s=[a.bias[0],a.bias[1],a.bias[2],a.bias[3]??0];const o=e.format!==void 0?e.format:"RGBA",l=e.tensorFormat!==void 0&&e.tensorFormat!==void 0?e.tensorFormat:"RGB",u=r*n,p=l==="RGBA"?new Float32Array(u*4):new Float32Array(u*3);let h=4,m=0,d=1,_=2,w=3,v=0,S=u,$=u*2,E=-1;o==="RGB"&&(h=3,m=0,d=1,_=2,w=-1),l==="RGBA"?E=u*3:l==="RBG"?(v=0,$=u,S=u*2):l==="BGR"&&($=0,S=u,v=u*2);for(let A=0;A{const r=typeof HTMLImageElement<"u"&&t instanceof HTMLImageElement,n=typeof ImageData<"u"&&t instanceof ImageData,a=typeof ImageBitmap<"u"&&t instanceof ImageBitmap,i=typeof t=="string";let s,o=e??{};const l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},u=p=>p instanceof HTMLCanvasElement||p instanceof OffscreenCanvas?p.getContext("2d"):null;if(r){const p=l();p.width=t.width,p.height=t.height;const h=u(p);if(h!=null){let m=t.height,d=t.width;if(e!==void 0&&e.resizedHeight!==void 0&&e.resizedWidth!==void 0&&(m=e.resizedHeight,d=e.resizedWidth),e!==void 0){if(o=e,e.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=m,o.width=d}else o.tensorFormat="RGBA",o.height=m,o.width=d;h.drawImage(t,0,0),s=h.getImageData(0,0,d,m).data}else throw new Error("Can not access image data")}else if(n){let p,h;if(e!==void 0&&e.resizedWidth!==void 0&&e.resizedHeight!==void 0?(p=e.resizedHeight,h=e.resizedWidth):(p=t.height,h=t.width),e!==void 0&&(o=e),o.format="RGBA",o.height=p,o.width=h,e!==void 0){const m=l();m.width=h,m.height=p;const d=u(m);if(d!=null)d.putImageData(t,0,0),s=d.getImageData(0,0,h,p).data;else throw new Error("Can not access image data")}else s=t.data}else if(a){if(e===void 0)throw new Error("Please provide image config with format for Imagebitmap");const p=l();p.width=t.width,p.height=t.height;const h=u(p);if(h!=null){const m=t.height,d=t.width;return h.drawImage(t,0,0,d,m),s=h.getImageData(0,0,d,m).data,o.height=m,o.width=d,uo(s,o)}else throw new Error("Can not access image data")}else{if(i)return new Promise((p,h)=>{const m=l(),d=u(m);if(!t||!d)return h();const _=new Image;_.crossOrigin="Anonymous",_.src=t,_.onload=()=>{m.width=_.width,m.height=_.height,d.drawImage(_,0,0,m.width,m.height);const w=d.getImageData(0,0,m.width,m.height);o.height=m.height,o.width=m.width,p(uo(w.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return uo(s,o);throw new Error("Input data provided is not supported - aborted tensor creation")},gw=(t,e)=>{const{width:r,height:n,download:a,dispose:i}=e,s=[1,n,r,4];return new mr({location:"texture",type:"float32",texture:t,dims:s,download:a,dispose:i})},_w=(t,e)=>{const{dataType:r,dims:n,download:a,dispose:i}=e;return new mr({location:"gpu-buffer",type:r??"float32",gpuBuffer:t,dims:n,download:a,dispose:i})},yw=(t,e,r)=>new mr({location:"cpu-pinned",type:t,data:e,dims:r??[e.length]}),On=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),Bi=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]);let cp=!1;const ww=()=>{if(!cp){cp=!0;const t=typeof BigInt64Array<"u"&&BigInt64Array.from,e=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;t&&(On.set("int64",BigInt64Array),Bi.set(BigInt64Array,"int64")),e&&(On.set("uint64",BigUint64Array),Bi.set(BigUint64Array,"uint64")),r?(On.set("float16",Float16Array),Bi.set(Float16Array,"float16")):On.set("float16",Uint16Array)}},bw=t=>{let e=1;for(let r=0;r{switch(t.location){case"cpu":return new mr(t.type,t.data,e);case"cpu-pinned":return new mr({location:"cpu-pinned",data:t.data,type:t.type,dims:e});case"texture":return new mr({location:"texture",texture:t.texture,type:t.type,dims:e});case"gpu-buffer":return new mr({location:"gpu-buffer",gpuBuffer:t.gpuBuffer,type:t.type,dims:e});default:throw new Error(`tensorReshape: tensor location ${t.location} is not supported`)}};let mr=class{constructor(e,r,n){ww();let a,i;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,a=e.type,i=e.dims,e.location){case"cpu-pinned":{const o=On.get(a);if(!o)throw new TypeError(`unsupported type "${a}" to create tensor from pinned buffer`);if(!(e.data instanceof o))throw new TypeError(`buffer should be of type ${o.name}`);this.cpuData=e.data;break}case"texture":{if(a!=="float32")throw new TypeError(`unsupported type "${a}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(a!=="float32"&&a!=="float16"&&a!=="int32"&&a!=="int64"&&a!=="uint32"&&a!=="uint8"&&a!=="bool")throw new TypeError(`unsupported type "${a}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let o,l;if(typeof e=="string")if(a=e,l=n,e==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");o=r}else{const u=On.get(e);if(u===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(r)){if(e==="float16"&&u===Uint16Array)throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");e==="uint64"||e==="int64"?o=u.from(r,BigInt):o=u.from(r)}else if(r instanceof u)o=r;else throw new TypeError(`A ${a} tensor's data must be type of ${u}`)}else if(l=r,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");const u=typeof e[0];if(u==="string")a="string",o=e;else if(u==="boolean")a="bool",o=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${u}.`)}else{const u=Bi.get(e.constructor);if(u===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);a=u,o=e}if(l===void 0)l=[o.length];else if(!Array.isArray(l))throw new TypeError("A tensor's dims must be a number array");i=l,this.cpuData=o,this.dataLocation="cpu"}const s=bw(i);if(this.cpuData&&s!==this.cpuData.length)throw new Error(`Tensor's size(${s}) does not match data length(${this.cpuData.length}).`);this.type=a,this.dims=i,this.size=s}static async fromImage(e,r){return mw(e,r)}static fromTexture(e,r){return gw(e,r)}static fromGpuBuffer(e,r){return _w(e,r)}static fromPinnedBuffer(e,r,n){return yw(e,r,n)}toDataURL(e){return hw(this,e)}toImageData(e){return fw(this,e)}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(e){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;const r=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=r,e&&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(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return vw(this,e)}};const $w=mr,_a=[];let zo,Sa;an.IS_NODE_ENV?(Sa=Je??k0,_a.push("cpu"),zo=["cpu"]):(Sa=pw,an.IS_WEBGPU_AVAILABLE&&_a.push("webgpu"),_a.push("wasm"),zo=["wasm"]);const xw=Sa.InferenceSession;function Sw(t){let e=zo;if(t){if(!_a.includes(t))throw new Error(`Unsupported device: "${t}". Should be one of: ${_a.join(", ")}.`);e=[t]}return e}async function hg(t,e){return await xw.create(t,e)}function fg(t){return t instanceof Sa.Tensor}const en=Sa?.env;en?.wasm&&(en.wasm.wasmPaths="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.18.0/dist/",en.wasm.proxy=!an.IS_WEBWORKER_ENV,(typeof crossOriginIsolated>"u"||!crossOriginIsolated)&&(en.wasm.numThreads=1),typeof navigator<"u"&&/iP(hone|od|ad).+16_4.+AppleWebKit/.test(navigator.userAgent)&&(en.wasm.simd=!1));function kw(){return en?.wasm?.proxy}zt.backends.onnx=en;const Cn=async(t,e,r)=>{const n=await hg(new Uint8Array(t),e);return async a=>{const i=Object.fromEntries(Object.entries(a).map(([o,l])=>[o,l.ort_tensor])),s=await n.run(i);return Array.isArray(r)?r.map(o=>new ue(s[o])):new ue(s[r])}};class Qi{static session_options={};static get bilinear_interpolate_4d(){return this._bilinear_interpolate_4d||(this._bilinear_interpolate_4d=Cn([8,9,18,0,58,128,1,10,40,10,1,120,10,0,10,0,10,1,115,18,1,121,34,6,82,101,115,105,122,101,42,17,10,4,109,111,100,101,34,6,108,105,110,101,97,114,160,1,3,18,1,114,90,31,10,1,120,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,90,15,10,1,115,18,10,10,8,8,7,18,4,10,2,8,4,98,31,10,1,121,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,66,2,16,20],this.session_options,"y")),this._bilinear_interpolate_4d}static get bicubic_interpolate_4d(){return this._bicubic_interpolate_4d||(this._bicubic_interpolate_4d=Cn([8,9,18,0,58,127,10,39,10,1,120,10,0,10,0,10,1,115,18,1,121,34,6,82,101,115,105,122,101,42,16,10,4,109,111,100,101,34,5,99,117,98,105,99,160,1,3,18,1,114,90,31,10,1,120,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,90,15,10,1,115,18,10,10,8,8,7,18,4,10,2,8,4,98,31,10,1,121,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,66,2,16,20],this.session_options,"y")),this._bicubic_interpolate_4d}static get matmul(){return this._matmul||(this._matmul=Cn([8,9,18,0,58,55,10,17,10,1,97,10,1,98,18,1,99,34,6,77,97,116,77,117,108,18,1,114,90,9,10,1,97,18,4,10,2,8,1,90,9,10,1,98,18,4,10,2,8,1,98,9,10,1,99,18,4,10,2,8,1,66,2,16,20],this.session_options,"c")),this._matmul}static get stft(){return this._stft||(this._stft=Cn([8,7,18,0,58,148,1,10,38,10,1,115,10,1,106,10,1,119,10,1,108,18,1,111,34,4,83,84,70,84,42,15,10,8,111,110,101,115,105,100,101,100,24,1,160,1,2,18,1,115,90,26,10,1,115,18,21,10,19,8,1,18,15,10,3,18,1,98,10,3,18,1,115,10,3,18,1,99,90,11,10,1,106,18,6,10,4,8,7,18,0,90,16,10,1,119,18,11,10,9,8,1,18,5,10,3,18,1,119,90,11,10,1,108,18,6,10,4,8,7,18,0,98,31,10,1,111,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,102,10,3,18,1,100,10,3,18,1,99,66,2,16,17],this.session_options,"o")),this._stft}static get rfft(){return this._rfft||(this._rfft=Cn([8,9,18,0,58,97,10,33,10,1,120,10,0,10,1,97,18,1,121,34,3,68,70,84,42,15,10,8,111,110,101,115,105,100,101,100,24,1,160,1,2,18,1,100,90,21,10,1,120,18,16,10,14,8,1,18,10,10,3,18,1,115,10,3,18,1,99,90,11,10,1,97,18,6,10,4,8,7,18,0,98,21,10,1,121,18,16,10,14,8,1,18,10,10,3,18,1,115,10,3,18,1,99,66,2,16,20],this.session_options,"y")),this._rfft}static get top_k(){return this._top_k||(this._top_k=Cn([8,10,18,0,58,73,10,18,10,1,120,10,1,107,18,1,118,18,1,105,34,4,84,111,112,75,18,1,116,90,9,10,1,120,18,4,10,2,8,1,90,15,10,1,107,18,10,10,8,8,7,18,4,10,2,8,1,98,9,10,1,118,18,4,10,2,8,1,98,9,10,1,105,18,4,10,2,8,7,66,2,16,21],this.session_options,["v","i"])),this._top_k}}const pp=Object.freeze({float32:Float32Array,float16:Uint16Array,float64:Float64Array,string:Array,int8:Int8Array,uint8:Uint8Array,int16:Int16Array,uint16:Uint16Array,int32:Int32Array,uint32:Uint32Array,int64:BigInt64Array,uint64:BigUint64Array,bool:Uint8Array});class ue{get dims(){return this.ort_tensor.dims}set dims(e){this.ort_tensor.dims=e}get type(){return this.ort_tensor.type}get data(){return this.ort_tensor.data}get size(){return this.ort_tensor.size}get location(){return this.ort_tensor.location}ort_tensor;constructor(...e){return fg(e[0])?this.ort_tensor=e[0]:this.ort_tensor=new $w(e[0],e[1],e[2]),new Proxy(this,{get:(r,n)=>{if(typeof n=="string"){let a=Number(n);if(Number.isInteger(a))return r._getitem(a)}return r[n]},set:(r,n,a)=>r[n]=a})}dispose(){this.ort_tensor.dispose()}*[Symbol.iterator](){const[e,...r]=this.dims;if(r.length>0){const n=r.reduce((a,i)=>a*i);for(let a=0;a0){const a=n.reduce((i,s)=>i*s);return this._subarray(e,a,n)}else return new ue(this.type,[this.data[e]],n)}indexOf(e){const r=this.data;for(let n=0;nm)throw new Error(`Invalid slice: ${p}`);let d=[Math.max(h,0),Math.min(m,this.dims[u])];n.push(d),r.push(d[1]-d[0])}else throw new Error(`Invalid slice: ${p}`)}let a=n.map(([u,p])=>p-u),i=a.reduce((u,p)=>u*p);const s=this.data;let o=new s.constructor(i);const l=this.stride();for(let u=0;u=0;--h){const d=a[h];p+=(m%d+n[h][0])*l[h],m=Math.floor(m/d)}o[u]=s[p]}return new ue(this.type,o,r)}permute(...e){return Cw(this,e)}transpose(...e){return this.permute(...e)}sum(e=null,r=!1){return this.norm(1,e,r)}norm(e="fro",r=null,n=!1){if(e==="fro")e=2;else if(typeof e=="string")throw Error(`Unsupported norm: ${e}`);const a=this.data;if(r===null){let o=a.reduce((l,u)=>l+u**e,0)**(1/e);return new ue(this.type,[o],[])}r=fr(r,this.dims.length);const i=this.dims.slice();i[r]=1;const s=new a.constructor(a.length/this.dims[r]);for(let o=0;o=0;--u){const m=this.dims[u];if(u!==r){const d=p%m;l+=d*h,h*=i[u]}p=Math.floor(p/m)}s[l]+=a[o]**e}if(e!==1)for(let o=0;o=0;--o){const p=this.dims[o];if(o!==r){const h=l%p;s+=h*u,u*=this.dims[o]}l=Math.floor(l/p)}a[i]/=n.data[s]}return this}normalize(e=2,r=1){return this.clone().normalize_(e,r)}stride(){return Ow(this.dims)}squeeze(e=null){return new ue(this.type,this.data,fp(this.dims,e))}squeeze_(e=null){return this.dims=fp(this.dims,e),this}unsqueeze(e=null){return new ue(this.type,this.data,mp(this.dims,e))}unsqueeze_(e=null){return this.dims=mp(this.dims,e),this}flatten_(e=0,r=-1){r=(r+this.dims.length)%this.dims.length;let n=this.dims.slice(0,e),a=this.dims.slice(e,r+1),i=this.dims.slice(r+1);return this.dims=[...n,a.reduce((s,o)=>s*o,1),...i],this}flatten(e=0,r=-1){return this.clone().flatten_(e,r)}view(...e){let r=-1;for(let n=0;ns!==r?a*i:a,1);e[r]=this.data.length/n}return new ue(this.type,this.data,e)}neg_(){const e=this.data;for(let r=0;ri*s);if(r!==n)throw Error(`cannot reshape array of size ${r} into shape (${e})`);let a=t;for(let i=e.length-1;i>=0;i--)a=a.reduce((s,o)=>{let l=s[s.length-1];return l.lengthr!==1):typeof e=="number"?t[e]===1&&t.splice(e,1):Array.isArray(e)&&(t=t.filter((r,n)=>r!==1||!e.includes(n))),t}function mp(t,e){return e=fr(e,t.length+1),t=t.slice(),t.splice(e,0,1),t}function fr(t,e,r=null,n=!0){if(n&&(t<-e||t>=e))throw new Error(`IndexError: index ${t} is out of bounds for dimension${r===null?"":" "+r} with size ${e}`);return t<0&&(t=(t%e+e)%e),t}function gr(t,e=0){e=fr(e,t[0].dims.length);const r=t[0].dims.slice();r[e]=t.reduce((s,o)=>s+o.dims[e],0);const n=r.reduce((s,o)=>s*o,1),a=new t[0].data.constructor(n),i=t[0].type;if(e===0){let s=0;for(let o of t)a.set(o.data,s),s+=o.data.length}else{let s=0;for(let o=0;o=0;--h){const _=l.dims[h];let w=m%_;h===e&&(w+=s),p+=w*d,d*=r[h],m=Math.floor(m/_)}a[p]=l.data[u]}s+=l.dims[e]}}return new ue(i,a,r)}function ka(t,e=0){return gr(t.map(r=>r.unsqueeze(e)),e)}function Iw(t,e=null,r=1,n=!1){if(e===null){const u=t.data.reduce((d,_)=>d+_,0)/t.data.length,p=Math.sqrt(t.data.reduce((d,_)=>d+(_-u)**2,0)/(t.data.length-r)),h=new ue(t.type,[u],[]);return[new ue(t.type,[p],[]),h]}e=fr(e,t.dims.length);const a=ll(t,e,n),i=t.dims.slice();i[e]=1;const s=new t.data.constructor(t.data.length/t.dims[e]);for(let l=0;l=0;--p){const d=t.dims[p];if(p!==e){const _=h%d;u+=_*m,m*=i[p]}h=Math.floor(h/d)}s[u]+=(t.data[l]-a.data[u])**2}for(let l=0;ls+o,0);return new ue(t.type,[i/t.data.length],[])}e=fr(e,t.dims.length);const n=t.dims.slice();n[e]=1;const a=new t.data.constructor(t.data.length/t.dims[e]);for(let i=0;i=0;--o){const p=t.dims[o];if(o!==e){const h=l%p;s+=h*u,u*=n[o]}l=Math.floor(l/p)}a[s]+=t.data[i]}if(t.dims[e]!==1)for(let i=0;i0||o>0;)switch(l.push(s-1),u.push(o-1),i[s][o].item()){case 0:--s,--o;break;case 1:--s;break;case 2:--o;break;default:throw new Error(`Internal error in dynamic time warping. Unexpected trace[${s}, ${o}]. Please file a bug report.`)}return l.reverse(),u.reverse(),[l,u]}function Ow(t){const e=new Array(t.length);for(let r=t.length-1,n=1;r>=0;--r)e[r]=n,n*=t[r];return e}function ul(t,e,r,n){const a=t.reduce((i,s)=>i*s,1);return new ue(r,new n(a).fill(e),t)}function zw(t,e){let r,n;return r="float32",n=Float32Array,ul(t,e,r,n)}function Pw(t,e){return zw(t.dims,e)}function Ma(t){return ul(t,1n,"int64",BigInt64Array)}function Rw(t){return Ma(t.dims)}function Bw(t){return ul(t,0n,"int64",BigInt64Array)}function Dw(t){return Bw(t.dims)}function Nw(t,e){if(t.dims.length!==2)throw new Error("The tensor must have 2 dimensions");if(t.dims.at(-1)%8!==0)throw new Error("The last dimension of the tensor must be a multiple of 8");if(!["binary","ubinary"].includes(e))throw new Error("The precision must be either 'binary' or 'ubinary'");const r=e==="binary",n=r?"int8":"uint8",a=r?Int8Array:Uint8Array,i=t.data,s=new a(i.length/8);for(let o=0;o0?1:0,u=Math.floor(o/8),p=o%8;s[u]|=l<<7-p,r&&p===0&&(s[u]-=128)}return new ue(n,s,[t.dims[0],t.dims[1]/8])}class Fw{constructor(e=(n,a)=>n>a,r=1/0){this._heap=[],this._comparator=e,this._maxSize=r}get size(){return this._heap.length}isEmpty(){return this.size===0}peek(){return this._heap[0]}push(...e){return this.extend(e)}extend(e){for(const r of e)if(this.size0&&this._swap(0,r),this._heap.pop(),this._siftDown(),e}replace(e){const r=this.peek();return this._heap[0]=e,this._siftDown(),r}_parent(e){return(e+1>>>1)-1}_left(e){return(e<<1)+1}_right(e){return e+1<<1}_greater(e,r){return this._comparator(this._heap[e],this._heap[r])}_swap(e,r){const n=this._heap[e];this._heap[e]=this._heap[r],this._heap[r]=n}_siftUp(){this._siftUpFrom(this.size-1)}_siftUpFrom(e){for(;e>0&&this._greater(e,this._parent(e));)this._swap(e,this._parent(e)),e=this._parent(e)}_siftDown(){let e=0;for(;this._left(e)[]),this.endNodes=Array.from({length:this.len+1},()=>[]);const a=new ya(this.bosTokenId,0,0,0,0),i=new ya(this.eosTokenId,1,this.len,0,0);this.nodes.push(a.clone()),this.nodes.push(i.clone()),this.beginNodes[this.len].push(i),this.endNodes[0].push(a)}insert(e,r,n,a){const i=this.nodes.length,s=new ya(a,i,e,r,n);this.beginNodes[e].push(s),this.endNodes[e+r].push(s),this.nodes.push(s)}viterbi(){const e=this.len;let r=0;for(;r<=e;){if(this.beginNodes[r].length==0)return[];for(let o of this.beginNodes[r]){o.prev=null;let l=0,u=null;for(let p of this.endNodes[r]){const h=p.backtraceScore+o.score;(u===null||h>l)&&(u=p.clone(),l=h)}if(u!==null)o.prev=u,o.backtraceScore=l;else return[]}++r}const n=[],i=this.beginNodes[e][0].prev;if(i===null)return[];let s=i.clone();for(;s.prev!==null;)n.push(s.clone()),s=s.clone().prev.clone();return n.reverse(),n}piece(e){return this.sentence.slice(e.pos,e.pos+e.length)}tokens(){return this.viterbi().map(r=>this.piece(r))}tokenIds(){return this.viterbi().map(r=>r.tokenId)}}class ya{constructor(e,r,n,a,i){this.tokenId=e,this.nodeId=r,this.pos=n,this.length=a,this.score=i,this.prev=null,this.backtraceScore=0}clone(){const e=new ya(this.tokenId,this.nodeId,this.pos,this.length,this.score);return e.prev=this.prev,e.backtraceScore=this.backtraceScore,e}}var U=Object.freeze({Text:"Text",NumericLiteral:"NumericLiteral",BooleanLiteral:"BooleanLiteral",StringLiteral:"StringLiteral",Identifier:"Identifier",Equals:"Equals",OpenParen:"OpenParen",CloseParen:"CloseParen",OpenStatement:"OpenStatement",CloseStatement:"CloseStatement",OpenExpression:"OpenExpression",CloseExpression:"CloseExpression",OpenSquareBracket:"OpenSquareBracket",CloseSquareBracket:"CloseSquareBracket",OpenCurlyBracket:"OpenCurlyBracket",CloseCurlyBracket:"CloseCurlyBracket",Comma:"Comma",Dot:"Dot",Colon:"Colon",Pipe:"Pipe",CallOperator:"CallOperator",AdditiveBinaryOperator:"AdditiveBinaryOperator",MultiplicativeBinaryOperator:"MultiplicativeBinaryOperator",ComparisonBinaryOperator:"ComparisonBinaryOperator",UnaryOperator:"UnaryOperator",Set:"Set",If:"If",For:"For",In:"In",Is:"Is",NotIn:"NotIn",Else:"Else",EndIf:"EndIf",ElseIf:"ElseIf",EndFor:"EndFor",And:"And",Or:"Or",Not:"UnaryOperator"}),gp=Object.freeze({set:U.Set,for:U.For,in:U.In,is:U.Is,if:U.If,else:U.Else,endif:U.EndIf,elif:U.ElseIf,endfor:U.EndFor,and:U.And,or:U.Or,not:U.Not,"not in":U.NotIn,true:U.BooleanLiteral,false:U.BooleanLiteral}),Kr=class{constructor(t,e){this.value=t,this.type=e}};function _p(t){return/\w/.test(t)}function co(t){return/[0-9]/.test(t)}var Ww=[["{%",U.OpenStatement],["%}",U.CloseStatement],["{{",U.OpenExpression],["}}",U.CloseExpression],["(",U.OpenParen],[")",U.CloseParen],["{",U.OpenCurlyBracket],["}",U.CloseCurlyBracket],["[",U.OpenSquareBracket],["]",U.CloseSquareBracket],[",",U.Comma],[".",U.Dot],[":",U.Colon],["|",U.Pipe],["<=",U.ComparisonBinaryOperator],[">=",U.ComparisonBinaryOperator],["==",U.ComparisonBinaryOperator],["!=",U.ComparisonBinaryOperator],["<",U.ComparisonBinaryOperator],[">",U.ComparisonBinaryOperator],["+",U.AdditiveBinaryOperator],["-",U.AdditiveBinaryOperator],["*",U.MultiplicativeBinaryOperator],["/",U.MultiplicativeBinaryOperator],["%",U.MultiplicativeBinaryOperator],["=",U.Equals]],Vw=new Map([["n",` +`],["t"," "],["r","\r"],["b","\b"],["f","\f"],["v","\v"],["'","'"],['"','"'],["\\","\\"]]);function Gw(t,e={}){return t.endsWith(` +`)&&(t=t.slice(0,-1)),t=t.replace(/{#.*?#}/gs,"{##}"),e.lstrip_blocks&&(t=t.replace(/^[ \t]*({[#%])/gm,"$1")),e.trim_blocks&&(t=t.replace(/([#%]})\n/g,"$1")),t.replace(/{##}/g,"").replace(/-%}\s*/g,"%}").replace(/\s*{%-/g,"{%").replace(/-}}\s*/g,"}}").replace(/\s*{{-/g,"{{")}function Hw(t,e={}){const r=[],n=Gw(t,e);let a=0;const i=s=>{let o="";for(;s(n[a]);){if(n[a]==="\\"){if(++a,a>=n.length)throw new SyntaxError("Unexpected end of input");const l=n[a++],u=Vw.get(l);if(u===void 0)throw new SyntaxError(`Unexpected escaped character: ${l}`);o+=u;continue}if(o+=n[a++],a>=n.length)throw new SyntaxError("Unexpected end of input")}return o};e:for(;a0){r.push(new Kr(l,U.Text));continue}}i(l=>/\s/.test(l));const o=n[a];if(o==="-"||o==="+"){const l=r.at(-1)?.type;if(l===U.Text||l===void 0)throw new SyntaxError(`Unexpected character: ${o}`);switch(l){case U.Identifier:case U.NumericLiteral:case U.BooleanLiteral:case U.StringLiteral:case U.CloseParen:case U.CloseSquareBracket:break;default:{++a;const u=i(co);r.push(new Kr(`${o}${u}`,u.length>0?U.NumericLiteral:U.UnaryOperator));continue}}}for(const[l,u]of Ww)if(n.slice(a,a+l.length)===l){r.push(new Kr(l,u)),a+=l.length;continue e}if(o==="'"||o==='"'){++a;const l=i(u=>u!==o);r.push(new Kr(l,U.StringLiteral)),++a;continue}if(co(o)){const l=i(co);r.push(new Kr(l,U.NumericLiteral));continue}if(_p(o)){const l=i(_p),u=Object.hasOwn(gp,l)?gp[l]:U.Identifier;u===U.In&&r.at(-1)?.type===U.Not?(r.pop(),r.push(new Kr("not in",U.NotIn))):r.push(new Kr(l,u));continue}throw new SyntaxError(`Unexpected character: ${o}`)}return r}var Oa=class{type="Statement"},jw=class extends Oa{constructor(t){super(),this.body=t}type="Program"},yp=class extends Oa{constructor(t,e,r){super(),this.test=t,this.body=e,this.alternate=r}type="If"},qw=class extends Oa{constructor(t,e,r){super(),this.loopvar=t,this.iterable=e,this.body=r}type="For"},Kw=class extends Oa{constructor(t,e){super(),this.assignee=t,this.value=e}type="Set"},_r=class extends Oa{type="Expression"},Yw=class extends _r{constructor(t,e,r){super(),this.object=t,this.property=e,this.computed=r}type="MemberExpression"},Xw=class extends _r{constructor(t,e){super(),this.callee=t,this.args=e}type="CallExpression"},Tn=class extends _r{constructor(t){super(),this.value=t}type="Identifier"},Fn=class extends _r{constructor(t){super(),this.value=t}type="Literal"},Qw=class extends Fn{type="NumericLiteral"},wp=class extends Fn{type="StringLiteral"},bp=class extends Fn{type="BooleanLiteral"},Zw=class extends Fn{type="ArrayLiteral"},vp=class extends Fn{type="TupleLiteral"},Jw=class extends Fn{type="ObjectLiteral"},pa=class extends _r{constructor(t,e,r){super(),this.operator=t,this.left=e,this.right=r}type="BinaryExpression"},eb=class extends _r{constructor(t,e){super(),this.operand=t,this.filter=e}type="FilterExpression"},tb=class extends _r{constructor(t,e,r){super(),this.operand=t,this.negate=e,this.test=r}type="TestExpression"},rb=class extends _r{constructor(t,e){super(),this.operator=t,this.argument=e}type="UnaryExpression"},nb=class extends _r{constructor(t=void 0,e=void 0,r=void 0){super(),this.start=t,this.stop=e,this.step=r}type="SliceExpression"},ab=class extends _r{constructor(t,e){super(),this.key=t,this.value=e}type="KeywordArgumentExpression"};function ib(t){const e=new jw([]);let r=0;function n(M,R){const G=t[r++];if(!G||G.type!==M)throw new Error(`Parser Error: ${R}. ${G.type} !== ${M}.`);return G}function a(){switch(t[r].type){case U.Text:return o();case U.OpenStatement:return l();case U.OpenExpression:return u();default:throw new SyntaxError(`Unexpected token type: ${t[r].type}`)}}function i(...M){return r+M.length<=t.length&&M.some((R,G)=>R!==t[r+G].type)}function s(...M){return r+M.length<=t.length&&M.every((R,G)=>R===t[r+G].type)}function o(){return new wp(n(U.Text,"Expected text token").value)}function l(){n(U.OpenStatement,"Expected opening statement token");let M;switch(t[r].type){case U.Set:++r,M=p(),n(U.CloseStatement,"Expected closing statement token");break;case U.If:++r,M=h(),n(U.OpenStatement,"Expected {% token"),n(U.EndIf,"Expected endif token"),n(U.CloseStatement,"Expected %} token");break;case U.For:++r,M=d(),n(U.OpenStatement,"Expected {% token"),n(U.EndFor,"Expected endfor token"),n(U.CloseStatement,"Expected %} token");break;default:throw new SyntaxError(`Unknown statement type: ${t[r].type}`)}return M}function u(){n(U.OpenExpression,"Expected opening expression token");const M=_();return n(U.CloseExpression,"Expected closing expression token"),M}function p(){const M=_();if(s(U.Equals)){++r;const R=p();return new Kw(M,R)}return M}function h(){const M=_();n(U.CloseStatement,"Expected closing statement token");const R=[],G=[];for(;!(t[r]?.type===U.OpenStatement&&(t[r+1]?.type===U.ElseIf||t[r+1]?.type===U.Else||t[r+1]?.type===U.EndIf));)R.push(a());if(t[r]?.type===U.OpenStatement&&t[r+1]?.type!==U.EndIf)if(++r,s(U.ElseIf))n(U.ElseIf,"Expected elseif token"),G.push(h());else for(n(U.Else,"Expected else token"),n(U.CloseStatement,"Expected closing statement token");!(t[r]?.type===U.OpenStatement&&t[r+1]?.type===U.EndIf);)G.push(a());return new yp(M,R,G)}function m(M=!1){const R=M?se:_,G=[R()],re=s(U.Comma);for(;re&&(++r,G.push(R()),!!s(U.Comma)););return re?new vp(G):G[0]}function d(){const M=m(!0);if(!(M instanceof Tn||M instanceof vp))throw new SyntaxError(`Expected identifier/tuple for the loop variable, got ${M.type} instead`);n(U.In,"Expected `in` keyword following loop variable");const R=_();n(U.CloseStatement,"Expected closing statement token");const G=[];for(;i(U.OpenStatement,U.EndFor);)G.push(a());return new qw(M,R,G)}function _(){return w()}function w(){const M=v();if(s(U.If)){++r;const R=v();n(U.Else,"Expected else token");const G=v();return new yp(R,[M],[G])}return M}function v(){let M=S();for(;s(U.Or);){const R=t[r];++r;const G=S();M=new pa(R,M,G)}return M}function S(){let M=$();for(;s(U.And);){const R=t[r];++r;const G=$();M=new pa(R,M,G)}return M}function $(){let M;for(;s(U.Not);){const R=t[r];++r;const G=$();M=new rb(R,G)}return M??E()}function E(){let M=T();for(;s(U.ComparisonBinaryOperator)||s(U.In)||s(U.NotIn);){const R=t[r];++r;const G=T();M=new pa(R,M,G)}return M}function T(){let M=ie();for(;s(U.AdditiveBinaryOperator);){const R=t[r];++r;const G=ie();M=new pa(R,M,G)}return M}function A(){const M=H();return s(U.OpenParen)?P(M):M}function P(M){let R=new Xw(M,B());return s(U.OpenParen)&&(R=P(R)),R}function B(){n(U.OpenParen,"Expected opening parenthesis for arguments list");const M=D();return n(U.CloseParen,"Expected closing parenthesis for arguments list"),M}function D(){const M=[];for(;!s(U.CloseParen);){let R=_();if(s(U.Equals)){if(++r,!(R instanceof Tn))throw new SyntaxError("Expected identifier for keyword argument");const G=_();R=new ab(R,G)}M.push(R),s(U.Comma)&&++r}return M}function q(){const M=[];let R=!1;for(;!s(U.CloseSquareBracket);)s(U.Colon)?(M.push(void 0),++r,R=!0):(M.push(_()),s(U.Colon)&&(++r,R=!0));if(M.length===0)throw new SyntaxError("Expected at least one argument for member/slice expression");if(R){if(M.length>3)throw new SyntaxError("Expected 0-3 arguments for slice expression");return new nb(...M)}return M[0]}function H(){let M=se();for(;s(U.Dot)||s(U.OpenSquareBracket);){const R=t[r];++r;let G;const re=R.type!==U.Dot;if(re)G=q(),n(U.CloseSquareBracket,"Expected closing square bracket");else if(G=se(),G.type!=="Identifier")throw new SyntaxError("Expected identifier following dot operator");M=new Yw(M,G,re)}return M}function ie(){let M=te();for(;s(U.MultiplicativeBinaryOperator);){const R=t[r];++r;const G=te();M=new pa(R,M,G)}return M}function te(){let M=de();for(;s(U.Is);){++r;const R=s(U.Not);R&&++r;let G=se();if(G instanceof bp&&(G=new Tn(G.value.toString())),!(G instanceof Tn))throw new SyntaxError("Expected identifier for the test");M=new tb(M,R,G)}return M}function de(){let M=A();for(;s(U.Pipe);){++r;let R=se();if(!(R instanceof Tn))throw new SyntaxError("Expected identifier for the filter");s(U.OpenParen)&&(R=P(R)),M=new eb(M,R)}return M}function se(){const M=t[r];switch(M.type){case U.NumericLiteral:return++r,new Qw(Number(M.value));case U.StringLiteral:return++r,new wp(M.value);case U.BooleanLiteral:return++r,new bp(M.value==="true");case U.Identifier:return++r,new Tn(M.value);case U.OpenParen:{++r;const R=m();if(t[r].type!==U.CloseParen)throw new SyntaxError(`Expected closing parenthesis, got ${t[r].type} instead`);return++r,R}case U.OpenSquareBracket:{++r;const R=[];for(;!s(U.CloseSquareBracket);)R.push(_()),s(U.Comma)&&++r;return++r,new Zw(R)}case U.OpenCurlyBracket:{++r;const R=new Map;for(;!s(U.CloseCurlyBracket);){const G=_();n(U.Colon,"Expected colon between key and value in object literal");const re=_();R.set(G,re),s(U.Comma)&&++r}return++r,new Jw(R)}default:throw new SyntaxError(`Unexpected token: ${M.type}`)}}for(;r=0?(e=(e??=0)<0?Math.max(t.length+e,0):Math.min(e,t.length),r=(r??=t.length)<0?Math.max(t.length+r,0):Math.min(r,t.length)):(e=(e??=t.length-1)<0?Math.max(t.length+e,-1):Math.min(e,t.length-1),r=(r??=-1)<-1?Math.max(t.length+r,-1):Math.min(r,t.length-1));const i=[];for(let s=e;a*se.toUpperCase())}var Cr=class{type="RuntimeValue";value;builtins=new Map;constructor(t=void 0){this.value=t}__bool__(){return new rt(!!this.value)}},je=class extends Cr{type="NumericValue"},Pe=class extends Cr{type="StringValue";builtins=new Map([["upper",new kr(()=>new Pe(this.value.toUpperCase()))],["lower",new kr(()=>new Pe(this.value.toLowerCase()))],["strip",new kr(()=>new Pe(this.value.trim()))],["title",new kr(()=>new Pe(mg(this.value)))],["length",new je(this.value.length)]])},rt=class extends Cr{type="BooleanValue"},or=class extends Cr{type="ObjectValue";__bool__(){return new rt(this.value.size>0)}builtins=new Map([["get",new kr(([t,e])=>{if(!(t instanceof Pe))throw new Error(`Object key must be a string: got ${t.type}`);return this.value.get(t.value)??e??new wa})],["items",new kr(()=>new Ze(Array.from(this.value.entries()).map(([t,e])=>new Ze([new Pe(t),e]))))]])},Ze=class extends Cr{type="ArrayValue";builtins=new Map([["length",new je(this.value.length)]]);__bool__(){return new rt(this.value.length>0)}},ob=class extends Ze{type="TupleValue"},kr=class extends Cr{type="FunctionValue"},wa=class extends Cr{type="NullValue"},sr=class extends Cr{type="UndefinedValue"},Po=class{constructor(t){this.parent=t}variables=new Map([["namespace",new kr(t=>{if(t.length===0)return new or(new Map);if(t.length!==1||!(t[0]instanceof or))throw new Error("`namespace` expects either zero arguments or a single object argument");return t[0]})]]);tests=new Map([["boolean",t=>t.type==="BooleanValue"],["callable",t=>t instanceof kr],["odd",t=>{if(t.type!=="NumericValue")throw new Error(`Cannot apply test "odd" to type: ${t.type}`);return t.value%2!==0}],["even",t=>{if(t.type!=="NumericValue")throw new Error(`Cannot apply test "even" to type: ${t.type}`);return t.value%2===0}],["false",t=>t.type==="BooleanValue"&&!t.value],["true",t=>t.type==="BooleanValue"&&t.value],["number",t=>t.type==="NumericValue"],["integer",t=>t.type==="NumericValue"&&Number.isInteger(t.value)],["iterable",t=>t instanceof Ze||t instanceof Pe],["lower",t=>{const e=t.value;return t.type==="StringValue"&&e===e.toLowerCase()}],["upper",t=>{const e=t.value;return t.type==="StringValue"&&e===e.toUpperCase()}],["none",t=>t.type==="NullValue"],["defined",t=>t.type!=="UndefinedValue"],["undefined",t=>t.type==="UndefinedValue"],["equalto",(t,e)=>t.value===e.value]]);set(t,e){return this.declareVariable(t,Di(e))}declareVariable(t,e){if(this.variables.has(t))throw new SyntaxError(`Variable already declared: ${t}`);return this.variables.set(t,e),e}setVariable(t,e){return this.variables.set(t,e),e}resolve(t){if(this.variables.has(t))return this;if(this.parent)return this.parent.resolve(t);throw new Error(`Unknown variable: ${t}`)}lookupVariable(t){try{return this.resolve(t).variables.get(t)??new sr}catch{return new sr}}},lb=class{global;constructor(t){this.global=t??new Po}run(t){return this.evaluate(t,this.global)}evaluateBinaryExpression(t,e){const r=this.evaluate(t.left,e);switch(t.operator.value){case"and":return r.__bool__().value?this.evaluate(t.right,e):r;case"or":return r.__bool__().value?r:this.evaluate(t.right,e)}const n=this.evaluate(t.right,e);switch(t.operator.value){case"==":return new rt(r.value==n.value);case"!=":return new rt(r.value!=n.value)}if(r instanceof sr||n instanceof sr)throw new Error("Cannot perform operation on undefined values");if(r instanceof wa||n instanceof wa)throw new Error("Cannot perform operation on null values");if(r instanceof je&&n instanceof je)switch(t.operator.value){case"+":return new je(r.value+n.value);case"-":return new je(r.value-n.value);case"*":return new je(r.value*n.value);case"/":return new je(r.value/n.value);case"%":return new je(r.value%n.value);case"<":return new rt(r.value":return new rt(r.value>n.value);case">=":return new rt(r.value>=n.value);case"<=":return new rt(r.value<=n.value)}else if(r instanceof Ze&&n instanceof Ze)switch(t.operator.value){case"+":return new Ze(r.value.concat(n.value))}else if(n instanceof Ze){const a=n.value.find(i=>i.value===r.value)!==void 0;switch(t.operator.value){case"in":return new rt(a);case"not in":return new rt(!a)}}if(r instanceof Pe||n instanceof Pe)switch(t.operator.value){case"+":return new Pe(r.value.toString()+n.value.toString())}if(r instanceof Pe&&n instanceof Pe)switch(t.operator.value){case"in":return new rt(n.value.includes(r.value));case"not in":return new rt(!n.value.includes(r.value))}if(r instanceof Pe&&n instanceof or)switch(t.operator.value){case"in":return new rt(n.value.has(r.value));case"not in":return new rt(!n.value.has(r.value))}throw new SyntaxError(`Unknown operator "${t.operator.value}" between ${r.type} and ${n.type}`)}evaluateFilterExpression(t,e){const r=this.evaluate(t.operand,e);if(t.filter.type==="Identifier"){const n=t.filter;if(r instanceof Ze)switch(n.value){case"list":return r;case"first":return r.value[0];case"last":return r.value[r.value.length-1];case"length":return new je(r.value.length);case"reverse":return new Ze(r.value.reverse());case"sort":return new Ze(r.value.sort((a,i)=>{if(a.type!==i.type)throw new Error(`Cannot compare different types: ${a.type} and ${i.type}`);switch(a.type){case"NumericValue":return a.value-i.value;case"StringValue":return a.value.localeCompare(i.value);default:throw new Error(`Cannot compare type: ${a.type}`)}}));default:throw new Error(`Unknown ArrayValue filter: ${n.value}`)}else if(r instanceof Pe)switch(n.value){case"length":return new je(r.value.length);case"upper":return new Pe(r.value.toUpperCase());case"lower":return new Pe(r.value.toLowerCase());case"title":return new Pe(mg(r.value));case"capitalize":return new Pe(r.value.charAt(0).toUpperCase()+r.value.slice(1));case"trim":return new Pe(r.value.trim());default:throw new Error(`Unknown StringValue filter: ${n.value}`)}else if(r instanceof je)switch(n.value){case"abs":return new je(Math.abs(r.value));default:throw new Error(`Unknown NumericValue filter: ${n.value}`)}else if(r instanceof or)switch(n.value){case"items":return new Ze(Array.from(r.value.entries()).map(([a,i])=>new Ze([new Pe(a),i])));case"length":return new je(r.value.size);default:throw new Error(`Unknown ObjectValue filter: ${n.value}`)}throw new Error(`Cannot apply filter "${n.value}" to type: ${r.type}`)}else if(t.filter.type==="CallExpression"){const n=t.filter;if(n.callee.type!=="Identifier")throw new Error(`Unknown filter: ${n.callee.type}`);const a=n.callee.value;if(r instanceof Ze){switch(a){case"selectattr":{if(r.value.some(p=>!(p instanceof or)))throw new Error("`selectattr` can only be applied to array of objects");if(n.args.some(p=>p.type!=="StringLiteral"))throw new Error("arguments of `selectattr` must be strings");const[i,s,o]=n.args.map(p=>this.evaluate(p,e));let l;if(s){const p=e.tests.get(s.value);if(!p)throw new Error(`Unknown test: ${s.value}`);l=p}else l=(...p)=>p[0].__bool__().value;const u=r.value.filter(p=>{const h=p.value.get(i.value);return h?l(h,o):!1});return new Ze(u)}}throw new Error(`Unknown ArrayValue filter: ${a}`)}else throw new Error(`Cannot apply filter "${a}" to type: ${r.type}`)}throw new Error(`Unknown filter: ${t.filter.type}`)}evaluateTestExpression(t,e){const r=this.evaluate(t.operand,e),n=e.tests.get(t.test.value);if(!n)throw new Error(`Unknown test: ${t.test.value}`);const a=n(r);return new rt(t.negate?!a:a)}evaluateUnaryExpression(t,e){const r=this.evaluate(t.argument,e);switch(t.operator.value){case"not":return new rt(!r.value);default:throw new SyntaxError(`Unknown operator: ${t.operator.value}`)}}evalProgram(t,e){return this.evaluateBlock(t.body,e)}evaluateBlock(t,e){let r="";for(const n of t){const a=this.evaluate(n,e);a.type!=="NullValue"&&a.type!=="UndefinedValue"&&(r+=a.value)}return new Pe(r)}evaluateIdentifier(t,e){return e.lookupVariable(t.value)}evaluateCallExpression(t,e){const r=[],n=new Map;for(const i of t.args)if(i.type==="KeywordArgumentExpression"){const s=i;n.set(s.key.value,this.evaluate(s.value,e))}else r.push(this.evaluate(i,e));n.size>0&&r.push(new or(n));const a=this.evaluate(t.callee,e);if(a.type!=="FunctionValue")throw new Error(`Cannot call something that is not a function: got ${a.type}`);return a.value(r,e)}evaluateSliceExpression(t,e,r){if(!(t instanceof Ze||t instanceof Pe))throw new Error("Slice object must be an array or string");const n=this.evaluate(e.start,r),a=this.evaluate(e.stop,r),i=this.evaluate(e.step,r);if(!(n instanceof je||n instanceof sr))throw new Error("Slice start must be numeric or undefined");if(!(a instanceof je||a instanceof sr))throw new Error("Slice stop must be numeric or undefined");if(!(i instanceof je||i instanceof sr))throw new Error("Slice step must be numeric or undefined");return t instanceof Ze?new Ze($p(t.value,n.value,a.value,i.value)):new Pe($p(Array.from(t.value),n.value,a.value,i.value).join(""))}evaluateMemberExpression(t,e){const r=this.evaluate(t.object,e);let n;if(t.computed){if(t.property.type==="SliceExpression")return this.evaluateSliceExpression(r,t.property,e);n=this.evaluate(t.property,e)}else n=new Pe(t.property.value);let a;if(r instanceof or){if(!(n instanceof Pe))throw new Error(`Cannot access property with non-string: got ${n.type}`);a=r.value.get(n.value)??r.builtins.get(n.value)}else if(r instanceof Ze||r instanceof Pe)if(n instanceof je)a=r.value.at(n.value),r instanceof Pe&&(a=new Pe(r.value.at(n.value)));else if(n instanceof Pe)a=r.builtins.get(n.value);else throw new Error(`Cannot access property with non-string/non-number: got ${n.type}`);else{if(!(n instanceof Pe))throw new Error(`Cannot access property with non-string: got ${n.type}`);a=r.builtins.get(n.value)}return a instanceof Cr?a:new sr}evaluateSet(t,e){const r=this.evaluate(t.value,e);if(t.assignee.type==="Identifier"){const n=t.assignee.value;e.setVariable(n,r)}else if(t.assignee.type==="MemberExpression"){const n=t.assignee,a=this.evaluate(n.object,e);if(!(a instanceof or))throw new Error("Cannot assign to member of non-object");if(n.property.type!=="Identifier")throw new Error("Cannot assign to member with non-identifier property");a.value.set(n.property.value,r)}else throw new Error(`Invalid LHS inside assignment expression: ${JSON.stringify(t.assignee)}`);return new wa}evaluateIf(t,e){const r=this.evaluate(t.test,e);return this.evaluateBlock(r.__bool__().value?t.body:t.alternate,e)}evaluateFor(t,e){const r=new Po(e),n=this.evaluate(t.iterable,r);if(!(n instanceof Ze))throw new Error(`Expected iterable type in for loop: got ${n.type}`);let a="";for(let i=0;i0?n.value[i-1]:new sr],["nextitem",ip.value.length?"few":"many"} items to unpack`);for(let h=0;hthis.evaluate(r,e)));case"TupleLiteral":return new ob(t.value.map(r=>this.evaluate(r,e)));case"ObjectLiteral":{const r=new Map;for(const[n,a]of t.value){const i=this.evaluate(n,e);if(!(i instanceof Pe))throw new Error(`Object keys must be strings: got ${i.type}`);r.set(i.value,this.evaluate(a,e))}return new or(r)}case"Identifier":return this.evaluateIdentifier(t,e);case"CallExpression":return this.evaluateCallExpression(t,e);case"MemberExpression":return this.evaluateMemberExpression(t,e);case"UnaryExpression":return this.evaluateUnaryExpression(t,e);case"BinaryExpression":return this.evaluateBinaryExpression(t,e);case"FilterExpression":return this.evaluateFilterExpression(t,e);case"TestExpression":return this.evaluateTestExpression(t,e);default:throw new SyntaxError(`Unknown node type: ${t.type}`)}}};function Di(t){switch(typeof t){case"number":return new je(t);case"string":return new Pe(t);case"boolean":return new rt(t);case"object":return t===null?new wa:Array.isArray(t)?new Ze(t.map(Di)):new or(new Map(Object.entries(t).map(([e,r])=>[e,Di(r)])));case"function":return new kr((e,r)=>{const n=t(...e.map(a=>a.value))??null;return Di(n)});default:throw new Error(`Cannot convert to runtime value: ${t}`)}}var ub=class{parsed;constructor(t){const e=Hw(t,{lstrip_blocks:!0,trim_blocks:!0});this.parsed=ib(e)}render(t){const e=new Po;e.set("false",!1),e.set("true",!0),e.set("raise_exception",a=>{throw new Error(a)}),e.set("range",sb);for(const[a,i]of Object.entries(t))e.set(a,i);return new lb(e).run(this.parsed).value}};const gg=[["en","english"],["zh","chinese"],["de","german"],["es","spanish"],["ru","russian"],["ko","korean"],["fr","french"],["ja","japanese"],["pt","portuguese"],["tr","turkish"],["pl","polish"],["ca","catalan"],["nl","dutch"],["ar","arabic"],["sv","swedish"],["it","italian"],["id","indonesian"],["hi","hindi"],["fi","finnish"],["vi","vietnamese"],["he","hebrew"],["uk","ukrainian"],["el","greek"],["ms","malay"],["cs","czech"],["ro","romanian"],["da","danish"],["hu","hungarian"],["ta","tamil"],["no","norwegian"],["th","thai"],["ur","urdu"],["hr","croatian"],["bg","bulgarian"],["lt","lithuanian"],["la","latin"],["mi","maori"],["ml","malayalam"],["cy","welsh"],["sk","slovak"],["te","telugu"],["fa","persian"],["lv","latvian"],["bn","bengali"],["sr","serbian"],["az","azerbaijani"],["sl","slovenian"],["kn","kannada"],["et","estonian"],["mk","macedonian"],["br","breton"],["eu","basque"],["is","icelandic"],["hy","armenian"],["ne","nepali"],["mn","mongolian"],["bs","bosnian"],["kk","kazakh"],["sq","albanian"],["sw","swahili"],["gl","galician"],["mr","marathi"],["pa","punjabi"],["si","sinhala"],["km","khmer"],["sn","shona"],["yo","yoruba"],["so","somali"],["af","afrikaans"],["oc","occitan"],["ka","georgian"],["be","belarusian"],["tg","tajik"],["sd","sindhi"],["gu","gujarati"],["am","amharic"],["yi","yiddish"],["lo","lao"],["uz","uzbek"],["fo","faroese"],["ht","haitian creole"],["ps","pashto"],["tk","turkmen"],["nn","nynorsk"],["mt","maltese"],["sa","sanskrit"],["lb","luxembourgish"],["my","myanmar"],["bo","tibetan"],["tl","tagalog"],["mg","malagasy"],["as","assamese"],["tt","tatar"],["haw","hawaiian"],["ln","lingala"],["ha","hausa"],["ba","bashkir"],["jw","javanese"],["su","sundanese"]],Ni=new Map(gg),db=new Map([...gg.map(([t,e])=>[e,t]),["burmese","my"],["valencian","ca"],["flemish","nl"],["haitian","ht"],["letzeburgesch","lb"],["pushto","ps"],["panjabi","pa"],["moldavian","ro"],["moldovan","ro"],["sinhalese","si"],["castilian","es"]]);function _g(t){t=t.toLowerCase();let e=db.get(t);if(e===void 0)if(Ni.has(t))e=t;else{const n=t.length===2?Ni.keys():Ni.values();throw new Error(`Language "${t}" is not supported. Must be one of: ${JSON.stringify(n)}`)}return e}const po="https://github.com/xenova/transformers.js/issues/new/choose";async function yg(t,e){const r=await Promise.all([Br(t,"tokenizer.json",!0,e),Br(t,"tokenizer_config.json",!0,e)]);return e.legacy!==null&&(r[1].legacy=e.legacy),r}function cb(t,e){const r=[];let n=0;for(const a of t.matchAll(e)){const i=a[0];n0&&r.push(i),n=a.index+i.length}return n=19968&&t<=40959||t>=13312&&t<=19903||t>=131072&&t<=173791||t>=173824&&t<=177983||t>=177984&&t<=178207||t>=178208&&t<=183983||t>=63744&&t<=64255||t>=194560&&t<=195103}function fb(t,e,r){const n=[];let a=0;for(;athis.tokens_to_ids.get(r)??this.unk_token_id)}convert_ids_to_tokens(e){return e.map(r=>this.vocab[r]??this.unk_token)}}class yb extends za{constructor(e){super(e),this.tokens_to_ids=dl(e.vocab),this.unk_token_id=this.tokens_to_ids.get(e.unk_token),this.unk_token=e.unk_token,this.max_input_chars_per_word=e.max_input_chars_per_word??100,this.vocab=new Array(this.tokens_to_ids.size);for(const[r,n]of this.tokens_to_ids)this.vocab[n]=r}encode(e){const r=[];for(const n of e){const a=[...n];if(a.length>this.max_input_chars_per_word){r.push(this.unk_token);continue}let i=!1,s=0;const o=[];for(;s0&&(p=this.config.continuing_subword_prefix+p),this.tokens_to_ids.has(p)){u=p;break}--l}if(u===null){i=!0;break}o.push(u),s=l}i?r.push(this.unk_token):r.push(...o)}return r}}class wb extends za{constructor(e,r){super(e);const n=e.vocab.length;this.vocab=new Array(n),this.scores=new Array(n);for(let a=0;a[a,i])),this.bosToken=" ",this.bosTokenId=this.tokens_to_ids.get(this.bosToken),this.eosToken=r.eos_token,this.eosTokenId=this.tokens_to_ids.get(this.eosToken),this.unkToken=this.vocab[this.unk_token_id],this.minScore=Fp(this.scores)[0],this.unkScore=this.minScore-10,this.scores[this.unk_token_id]=this.unkScore,this.trie=new Lw,this.trie.extend(this.vocab),this.fuse_unk=!0}populateNodes(e){const r=e.sentence,n=r.length;let a=0;for(;a{const t=[...Array.from({length:94},(a,i)=>i+33),...Array.from({length:12},(a,i)=>i+161),...Array.from({length:82},(a,i)=>i+174)],e=t.slice();let r=0;for(let a=0;a<256;++a)t.includes(a)||(t.push(a),e.push(256+r),r+=1);const n=e.map(a=>String.fromCharCode(a));return Object.fromEntries(t.map((a,i)=>[a,n[i]]))})(),bb=O0(vg);class vb extends za{constructor(e){super(e),this.BPE_SPLIT_TOKEN=" ",this.tokens_to_ids=dl(e.vocab),this.unk_token_id=this.tokens_to_ids.get(e.unk_token),this.unk_token=e.unk_token,this.vocab=new Array(this.tokens_to_ids.size);for(const[r,n]of this.tokens_to_ids)this.vocab[n]=r;this.bpe_ranks=new Map(e.merges.map((r,n)=>[r,n])),this.merges=e.merges.map(r=>r.split(this.BPE_SPLIT_TOKEN)),this.end_of_word_suffix=e.end_of_word_suffix,this.continuing_subword_suffix=e.continuing_subword_suffix??null,this.byte_fallback=this.config.byte_fallback??!1,this.byte_fallback&&(this.text_encoder=new TextEncoder),this.ignore_merges=this.config.ignore_merges??!1,this.cache=new Map}bpe(e){if(e.length===0)return[];const r=this.cache.get(e);if(r!==void 0)return r;const n=Array.from(e);this.end_of_word_suffix&&(n[n.length-1]+=this.end_of_word_suffix);let a=[];if(n.length>1){const i=new Fw((l,u)=>l.score`<0x${s.toString(16).toUpperCase().padStart(2,"0")}>`)):r.push(this.unk_token)}return r}}class $b extends za{constructor(e,r){super(e),this.tokens_to_ids=dl(r.target_lang?e.vocab[r.target_lang]:e.vocab),this.bos_token=r.bos_token,this.bos_token_id=this.tokens_to_ids.get(this.bos_token),this.eos_token=r.eos_token,this.eos_token_id=this.tokens_to_ids.get(this.eos_token),this.pad_token=r.pad_token,this.pad_token_id=this.tokens_to_ids.get(this.pad_token),this.unk_token=r.unk_token,this.unk_token_id=this.tokens_to_ids.get(this.unk_token),this.vocab=new Array(this.tokens_to_ids.size);for(const[n,a]of this.tokens_to_ids)this.vocab[a]=n}encode(e){return e}}class Lt extends bt{constructor(e){super(),this.config=e}static fromConfig(e){if(e===null)return null;switch(e.type){case"BertNormalizer":return new Ob(e);case"Precompiled":return new Qb(e);case"Sequence":return new Mb(e);case"Replace":return new xb(e);case"NFC":return new Sb(e);case"NFKC":return new kb(e);case"NFKD":return new Eb(e);case"Strip":return new Cb(e);case"StripAccents":return new Tb(e);case"Lowercase":return new Ab(e);case"Prepend":return new Ib(e);default:throw new Error(`Unknown Normalizer type: ${e.type}`)}}normalize(e){throw Error("normalize should be implemented in subclass.")}_call(e){return this.normalize(e)}}class xb extends Lt{normalize(e){const r=is(this.config.pattern);return r===null?e:e.replaceAll(r,this.config.content)}}class Sb extends Lt{normalize(e){return e=e.normalize("NFC"),e}}class kb extends Lt{normalize(e){return e=e.normalize("NFKC"),e}}class Eb extends Lt{normalize(e){return e=e.normalize("NFKD"),e}}class Cb extends Lt{normalize(e){return this.config.strip_left&&this.config.strip_right?e=e.trim():(this.config.strip_left&&(e=e.trimStart()),this.config.strip_right&&(e=e.trimEnd())),e}}class Tb extends Lt{normalize(e){return e=bg(e),e}}class Ab extends Lt{normalize(e){return e=e.toLowerCase(),e}}class Ib extends Lt{normalize(e){return e=this.config.prepend+e,e}}class Mb extends Lt{constructor(e){super(e),this.normalizers=e.normalizers.map(r=>Lt.fromConfig(r))}normalize(e){return this.normalizers.reduce((r,n)=>n.normalize(r),e)}}class Ob extends Lt{_tokenize_chinese_chars(e){const r=[];for(let n=0;nthis.pre_tokenize_text(n,r)):this.pre_tokenize_text(e,r)).flat()}_call(e,r){return this.pre_tokenize(e,r)}}class zb extends qt{constructor(e){super(),this.pattern=new RegExp(`[^\\s${Ea}]+|[${Ea}]`,"gu")}pre_tokenize_text(e,r){return e.trim().match(this.pattern)||[]}}class Pb extends qt{constructor(e){super(),this.config=e,this.add_prefix_space=this.config.add_prefix_space,this.trim_offsets=this.config.trim_offsets,this.use_regex=this.config.use_regex??!0,this.pattern=/'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+/gu,this.byte_encoder=vg,this.text_encoder=new TextEncoder}pre_tokenize_text(e,r){return this.add_prefix_space&&!e.startsWith(" ")&&(e=" "+e),(this.use_regex?e.match(this.pattern)||[]:[e]).map(a=>Array.from(this.text_encoder.encode(a),i=>this.byte_encoder[i]).join(""))}}class Rb extends qt{constructor(e){super(),this.config=e,this.pattern=is(this.config.pattern,this.config.invert)}pre_tokenize_text(e,r){return this.pattern===null?[]:this.config.invert?e.match(this.pattern)||[]:cb(e,this.pattern)}}class Bb extends qt{constructor(e){super(),this.config=e,this.pattern=new RegExp(`[^${Ea}]+|[${Ea}]+`,"gu")}pre_tokenize_text(e,r){return e.match(this.pattern)||[]}}class Db extends qt{constructor(e){super(),this.config=e;const r=`[^\\d]+|\\d${this.config.individual_digits?"":"+"}`;this.pattern=new RegExp(r,"gu")}pre_tokenize_text(e,r){return e.match(this.pattern)||[]}}class Rn extends bt{constructor(e){super(),this.config=e}static fromConfig(e){if(e===null)return null;switch(e.type){case"TemplateProcessing":return new Nb(e);case"ByteLevel":return new Sg(e);case"RobertaProcessing":return new xg(e);case"BertProcessing":return new $g(e);case"Sequence":return new Fb(e);default:throw new Error(`Unknown PostProcessor type: ${e.type}`)}}post_process(e,...r){throw Error("post_process should be implemented in subclass.")}_call(e,...r){return this.post_process(e,...r)}}class $g extends Rn{constructor(e){super(e),this.cls=e.cls[0],this.sep=e.sep[0]}post_process(e,r=null,{add_special_tokens:n=!0}={}){n&&(e=ft([this.cls],e,[this.sep]));let a=new Array(e.length).fill(0);if(r!==null){const i=n&&this instanceof xg?[this.sep]:[],s=n?[this.sep]:[];e=ft(e,i,r,s),a=ft(a,new Array(r.length+i.length+s.length).fill(1))}return{tokens:e,token_type_ids:a}}}class xg extends $g{}class Nb extends Rn{constructor(e){super(e),this.single=e.single,this.pair=e.pair}post_process(e,r=null,{add_special_tokens:n=!0}={}){const a=r===null?this.single:this.pair;let i=[],s=[];for(const o of a)"SpecialToken"in o?n&&(i.push(o.SpecialToken.id),s.push(o.SpecialToken.type_id)):"Sequence"in o&&(o.Sequence.id==="A"?(i=ft(i,e),s=ft(s,new Array(e.length).fill(o.Sequence.type_id))):o.Sequence.id==="B"&&(i=ft(i,r),s=ft(s,new Array(r.length).fill(o.Sequence.type_id))));return{tokens:i,token_type_ids:s}}}class Sg extends Rn{post_process(e,r=null){return r&&(e=ft(e,r)),{tokens:e}}}class Fb extends Rn{constructor(e){super(e),this.processors=e.processors.map(r=>Rn.fromConfig(r))}post_process(e,r=null,n={}){let a;for(const i of this.processors)if(i instanceof Sg)e=i.post_process(e).tokens,r&&(r=i.post_process(r).tokens);else{const s=i.post_process(e,r,n);e=s.tokens,a=s.token_type_ids}return{tokens:e,token_type_ids:a}}}class Ut extends bt{constructor(e){super(),this.config=e,this.added_tokens=[],this.end_of_word_suffix=null,this.trim_offsets=e.trim_offsets}static fromConfig(e){if(e===null)return null;switch(e.type){case"WordPiece":return new Gb(e);case"Metaspace":return new Xb(e);case"ByteLevel":return new Hb(e);case"Replace":return new Lb(e);case"ByteFallback":return new Ub(e);case"Fuse":return new Wb(e);case"Strip":return new Vb(e);case"Sequence":return new qb(e);case"CTC":return new jb(e);case"BPEDecoder":return new Kb(e);default:throw new Error(`Unknown Decoder type: ${e.type}`)}}_call(e){return this.decode(e)}decode(e){return this.decode_chain(e).join("")}decode_chain(e){throw Error("`decode_chain` should be implemented in subclass.")}}class Lb extends Ut{decode_chain(e){const r=is(this.config.pattern);return r===null?e:e.map(n=>n.replaceAll(r,this.config.content))}}class Ub extends Ut{constructor(e){super(e),this.text_decoder=new TextDecoder}decode_chain(e){const r=[];let n=[];for(const a of e){let i=null;if(a.length===6&&a.startsWith("<0x")&&a.endsWith(">")){const s=parseInt(a.slice(3,5),16);isNaN(s)||(i=s)}if(i!==null)n.push(i);else{if(n.length>0){const s=this.text_decoder.decode(Uint8Array.from(n));r.push(s),n=[]}r.push(a)}}if(n.length>0){const a=this.text_decoder.decode(Uint8Array.from(n));r.push(a),n=[]}return r}}class Wb extends Ut{decode_chain(e){return[e.join("")]}}class Vb extends Ut{constructor(e){super(e),this.content=this.config.content,this.start=this.config.start,this.stop=this.config.stop}decode_chain(e){return e.map(r=>{let n=0;for(let i=0;i(n!==0&&(r.startsWith(this.config.prefix)?r=r.replace(this.config.prefix,""):r=" "+r),this.cleanup&&(r=cl(r)),r))}}class Hb extends Ut{constructor(e){super(e),this.byte_decoder=bb,this.text_decoder=new TextDecoder("utf-8",{fatal:!1,ignoreBOM:!0}),this.end_of_word_suffix=null}convert_tokens_to_string(e){const r=e.join(""),n=new Uint8Array([...r].map(i=>this.byte_decoder[i]));return this.text_decoder.decode(n)}decode_chain(e){const r=[];let n=[];for(const a of e)this.added_tokens.find(i=>i.content===a)!==void 0?(n.length>0&&(r.push(this.convert_tokens_to_string(n)),n=[]),r.push(a)):n.push(a);return n.length>0&&r.push(this.convert_tokens_to_string(n)),r}}class jb extends Ut{constructor(e){super(e),this.pad_token=this.config.pad_token,this.word_delimiter_token=this.config.word_delimiter_token,this.cleanup=this.config.cleanup}convert_tokens_to_string(e){if(e.length===0)return"";const r=[e[0]];for(let i=1;ii!==this.pad_token).join("");return this.cleanup&&(a=cl(a).replaceAll(this.word_delimiter_token," ").trim()),a}decode_chain(e){return[this.convert_tokens_to_string(e)]}}class qb extends Ut{constructor(e){super(e),this.decoders=e.decoders.map(r=>Ut.fromConfig(r))}decode_chain(e){return this.decoders.reduce((r,n)=>n.decode_chain(r),e)}}class Kb extends Ut{constructor(e){super(e),this.suffix=this.config.suffix}decode_chain(e){return e.map((r,n)=>r.replaceAll(this.suffix,n===e.length-1?"":" "))}}class Yb extends Ut{decode_chain(e){let r="";for(let n=1;nn.normalize("NFKC")).join("～"):e=e.normalize("NFKC"),e}}class Zb extends qt{constructor(e){super(),this.tokenizers=e.pretokenizers.map(r=>qt.fromConfig(r))}pre_tokenize_text(e,r){return this.tokenizers.reduce((n,a)=>a.pre_tokenize(n,r),[e])}}class Jb extends qt{constructor(e){super()}pre_tokenize_text(e,r){return e.match(/\w+|[^\w\s]+/g)||[]}}class ev extends qt{constructor(e){super()}pre_tokenize_text(e,r){return mb(e)}}class tv extends qt{constructor(e){super(),this.config=e,this.pattern=is(this.config.pattern),this.content=this.config.content}pre_tokenize_text(e,r){return this.pattern===null?[e]:[e.replaceAll(this.pattern,this.config.content)]}}const rv=["bos_token","eos_token","unk_token","sep_token","pad_token","cls_token","mask_token"];function nv(t,e,r,n){for(const a of Object.keys(t)){const i=e-t[a].length,s=r(a),o=new Array(i).fill(s);t[a]=n==="right"?ft(t[a],o):ft(o,t[a])}}function av(t,e){for(const r of Object.keys(t))t[r].length=e}class Ee extends bt{return_token_type_ids=!1;_default_chat_template=`{% for message in messages %}{{'<|im_start|>' + message['role'] + ' +' + message['content'] + '<|im_end|>' + ' +'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant +' }}{% endif %}`;padding_side="right";constructor(e,r){super(),this._tokenizer_config=r,this.normalizer=Lt.fromConfig(e.normalizer),this.pre_tokenizer=qt.fromConfig(e.pre_tokenizer),this.model=za.fromConfig(e.model,r),this.post_processor=Rn.fromConfig(e.post_processor),this.decoder=Ut.fromConfig(e.decoder),this.special_tokens=[],this.all_special_ids=[],this.added_tokens=[];for(const n of e.added_tokens){const a=new _b(n);this.added_tokens.push(a),this.model.tokens_to_ids.set(a.content,a.id),this.model.vocab[a.id]=a.content,a.special&&(this.special_tokens.push(a.content),this.all_special_ids.push(a.id))}if(this.additional_special_tokens=r.additional_special_tokens??[],this.special_tokens.push(...this.additional_special_tokens),this.special_tokens=[...new Set(this.special_tokens)],this.decoder&&(this.decoder.added_tokens=this.added_tokens,this.decoder.end_of_word_suffix=this.model.end_of_word_suffix),this.added_tokens_regex=this.added_tokens.length>0?new RegExp(this.added_tokens.map(n=>`${n.lstrip?"\\s*":""}(${Np(n.content)})${n.rstrip?"\\s*":""}`).join("|")):null,this.mask_token=this.getToken("mask_token"),this.mask_token_id=this.model.tokens_to_ids.get(this.mask_token),this.pad_token=this.getToken("pad_token","eos_token"),this.pad_token_id=this.model.tokens_to_ids.get(this.pad_token),this.sep_token=this.getToken("sep_token"),this.sep_token_id=this.model.tokens_to_ids.get(this.sep_token),this.unk_token=this.getToken("unk_token"),this.unk_token_id=this.model.tokens_to_ids.get(this.unk_token),this.model_max_length=r.model_max_length,this.remove_space=r.remove_space,this.clean_up_tokenization_spaces=r.clean_up_tokenization_spaces??!0,this.do_lowercase_and_remove_accent=r.do_lowercase_and_remove_accent??!1,r.padding_side&&(this.padding_side=r.padding_side),this.legacy=!1,this.chat_template=r.chat_template??null,Array.isArray(this.chat_template)){const n=Object.create(null);for(const{name:a,template:i}of this.chat_template){if(typeof a!="string"||typeof i!="string")throw new Error('Chat template must be a list of objects with "name" and "template" properties');n[a]=i}this.chat_template=n}this._compiled_template_cache=new Map}getToken(...e){for(const r of e){const n=this._tokenizer_config[r];if(n)if(typeof n=="object"){if(n.__type==="AddedToken")return n.content;throw Error(`Unknown token: ${n}`)}else return n}return null}static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",legacy:o=null}={}){const l=await yg(e,{progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,legacy:o});return new this(...l)}_call(e,{text_pair:r=null,add_special_tokens:n=!0,padding:a=!1,truncation:i=null,max_length:s=null,return_tensor:o=!0,return_token_type_ids:l=null}={}){const u=Array.isArray(e);let p;if(u){if(e.length===0)throw Error("text array must be non-empty");if(r!==null){if(Array.isArray(r)){if(e.length!==r.length)throw Error("text and text_pair must have the same length")}else throw Error("text_pair must also be an array");p=e.map((m,d)=>this._encode_plus(m,{text_pair:r[d],add_special_tokens:n,return_token_type_ids:l}))}else p=e.map(m=>this._encode_plus(m,{add_special_tokens:n,return_token_type_ids:l}))}else{if(e==null)throw Error("text may not be null or undefined");if(Array.isArray(r))throw Error("When specifying `text_pair`, since `text` is a string, `text_pair` must also be a string (i.e., not an array).");p=[this._encode_plus(e,{text_pair:r,add_special_tokens:n,return_token_type_ids:l})]}if(s===null?a==="max_length"?s=this.model_max_length:s=Kt(p.map(m=>m.input_ids.length))[0]:i||console.warn("Truncation was not explicitly activated but `max_length` is provided a specific value, please use `truncation=true` to explicitly truncate examples to max length."),s=Math.min(s,this.model_max_length??1/0),a||i)for(let m=0;ms?i&&av(p[m],s):a&&nv(p[m],s,d=>d==="input_ids"?this.pad_token_id:0,this.padding_side));const h={};if(o){if(!(a&&i)&&p.some(d=>{for(const _ of Object.keys(d))if(d[_].length!==p[0][_]?.length)return!0;return!1}))throw Error("Unable to create tensor, you should probably activate truncation and/or padding with 'padding=true' and 'truncation=true' to have batched tensors with the same length.");const m=[p.length,p[0].input_ids.length];for(const d of Object.keys(p[0]))h[d]=new ue("int64",BigInt64Array.from(p.flatMap(_=>_[d]).map(BigInt)),m)}else{for(const m of Object.keys(p[0]))h[m]=p.map(d=>d[m]);if(!u)for(const m of Object.keys(h))h[m]=h[m][0]}return h}_encode_text(e){return e===null?null:(this.added_tokens_regex?e.split(this.added_tokens_regex).filter(a=>a):[e]).map((a,i)=>{if(this.added_tokens.find(o=>o.content===a)!==void 0)return a;{if(this.remove_space===!0&&(a=a.trim().split(/\s+/).join(" ")),this.do_lowercase_and_remove_accent&&(a=pb(a)),this.normalizer!==null&&(a=this.normalizer(a)),a.length===0)return[];const o=this.pre_tokenizer!==null?this.pre_tokenizer(a,{section_index:i}):[a];return this.model(o)}}).flat()}_encode_plus(e,{text_pair:r=null,add_special_tokens:n=!0,return_token_type_ids:a=null}={}){const{tokens:i,token_type_ids:s}=this._tokenize_helper(e,{pair:r,add_special_tokens:n}),o=this.model.convert_tokens_to_ids(i),l={input_ids:o,attention_mask:new Array(o.length).fill(1)};return(a??this.return_token_type_ids)&&s&&(l.token_type_ids=s),l}_tokenize_helper(e,{pair:r=null,add_special_tokens:n=!1}={}){const a=this._encode_text(e),i=this._encode_text(r);return this.post_processor?this.post_processor(a,i,{add_special_tokens:n}):{tokens:ft(a??[],i??[])}}tokenize(e,{pair:r=null,add_special_tokens:n=!1}={}){return this._tokenize_helper(e,{pair:r,add_special_tokens:n}).tokens}encode(e,{text_pair:r=null,add_special_tokens:n=!0,return_token_type_ids:a=null}={}){return this._encode_plus(e,{text_pair:r,add_special_tokens:n,return_token_type_ids:a}).input_ids}batch_decode(e,r={}){return e instanceof ue&&(e=e.tolist()),e.map(n=>this.decode(n,r))}decode(e,r={}){if(e instanceof ue&&(e=wg(e)),!Array.isArray(e)||e.length===0||!z0(e[0]))throw Error("token_ids must be a non-empty array of integers.");return this.decode_single(e,r)}decode_single(e,{skip_special_tokens:r=!1,clean_up_tokenization_spaces:n=null}){let a=this.model.convert_ids_to_tokens(e);r&&(a=a.filter(s=>!this.special_tokens.includes(s)));let i=this.decoder?this.decoder(a):a.join(" ");return this.decoder&&this.decoder.end_of_word_suffix&&(i=i.replaceAll(this.decoder.end_of_word_suffix," "),r&&(i=i.trim())),(n??this.clean_up_tokenization_spaces)&&(i=cl(i)),i}get default_chat_template(){return this._warned_about_chat_template||(console.warn("No chat template is defined for this tokenizer - using a default chat template that implements the ChatML format. If the default is not appropriate for your model, please set `tokenizer.chat_template` to an appropriate template. See https://huggingface.co/docs/transformers/main/chat_templating for more information."),this._warned_about_chat_template=!0),this._default_chat_template}apply_chat_template(e,{chat_template:r=null,add_generation_prompt:n=!1,tokenize:a=!0,padding:i=!1,truncation:s=!1,max_length:o=null,return_tensor:l=!0,return_dict:u=!1,tokenizer_kwargs:p={},...h}={}){if(this.chat_template&&typeof this.chat_template=="object"||this.chat_template===null&&this.default_chat_template&&typeof this.default_chat_template=="object"){const w=this.chat_template??this.default_chat_template;if(r!==null&&Object.hasOwn(w,r))r=w[r];else if(r===null&&"default"in w)r=w.default;else if(r===null)throw Error(`This model has multiple chat templates with no default specified! Please either pass a chat template or the name of the template you wish to use to the 'chat_template' argument. Available template names are ${Object.keys(w).sort()}.`)}else r??=this.chat_template??this.default_chat_template;if(typeof r!="string")throw Error(`chat_template must be a string, but got ${typeof r}`);let m=this._compiled_template_cache.get(r);m===void 0&&(m=new ub(r),this._compiled_template_cache.set(r,m));const d=Object.create(null);for(const w of rv){const v=this.getToken(w);v&&(d[w]=v)}const _=m.render({messages:e,add_generation_prompt:n,...d,...h});if(a){const w=this._call(_,{add_special_tokens:!1,padding:i,truncation:s,max_length:o,return_tensor:l,...p});return u?w:w.input_ids}return _}}class iv extends Ee{return_token_type_ids=!0}class sv extends Ee{return_token_type_ids=!0}class ov extends Ee{return_token_type_ids=!0}class lv extends Ee{return_token_type_ids=!0}class uv extends Ee{return_token_type_ids=!0}class dv extends Ee{return_token_type_ids=!0}class cv extends Ee{return_token_type_ids=!0}class pv extends Ee{return_token_type_ids=!0}class hv extends Ee{return_token_type_ids=!0}class fv extends Ee{}class mv extends Ee{}class gv extends Ee{return_token_type_ids=!0;constructor(e,r){super(e,r),console.warn('WARNING: `XLMTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.')}}class _v extends Ee{return_token_type_ids=!0}class yv extends Ee{}class Eg extends Ee{_default_chat_template='{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}'}class wv extends Ee{}class Cg extends Ee{constructor(e,r){super(e,r),this.languageRegex=/^[a-z]{2}_[A-Z]{2}$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)),this.lang_to_token=n=>n}_build_translation_inputs(e,r,n){return pl(this,e,r,n)}}class bv extends Cg{}class vv extends Ee{}class $v extends Eg{constructor(e,r){const n=".,!?…。，、।۔،",a=e.pre_tokenizer?.pretokenizers[0]?.pattern;a&&a.Regex===` ?[^(\\s|[${n}])]+`&&(a.Regex=` ?[^\\s${n}]+`),super(e,r)}}const Ti="▁";class Tg extends Ee{_default_chat_template=`{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif USE_DEFAULT_PROMPT == true and not '<>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'DEFAULT_SYSTEM_MESSAGE' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<> +' + system_message + ' +<> + +' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ bos_token + '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<> +' + content.strip() + ' +<> + +' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}`;DEFAULT_SYSTEM_PROMPT=`You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature. + +If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.`;padding_side="left";constructor(e,r){super(e,r),this.use_default_system_prompt=r.use_default_system_prompt??!1,this.legacy=r.legacy??!0,this.legacy||(this.normalizer=null,this.pre_tokenizer=new kg({replacement:Ti,add_prefix_space:!0,prepend_scheme:"first"}))}_encode_text(e){if(e===null)return null;if(this.legacy||e.length===0)return super._encode_text(e);let r=super._encode_text(Ti+e.replaceAll(Ti," "));return r.length>1&&r[0]===Ti&&this.special_tokens.includes(r[1])&&(r=r.slice(1)),r}get default_chat_template(){return super.default_chat_template.replaceAll("USE_DEFAULT_PROMPT",this.use_default_system_prompt?"true":"false").replaceAll("DEFAULT_SYSTEM_MESSAGE",this.DEFAULT_SYSTEM_PROMPT.replaceAll(` +`,"\\n").replaceAll("'","\\'"))}}class xv extends Tg{}class Sv extends Ee{}class kv extends Ee{}class Ev extends Ee{}class Cv extends Ee{}class Tv extends Ee{}class Av extends Ee{}class Iv extends Ee{_default_chat_template=`{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '' + role + ' +' + message['content'] | trim + ' +' }}{% endfor %}{% if add_generation_prompt %}{{'model +'}}{% endif %}`}class Mv extends Ee{}function pl(t,e,r,n){if(!("language_codes"in t)||!Array.isArray(t.language_codes))throw new Error("Tokenizer must have `language_codes` attribute set and it should be an array of language ids.");if(!("languageRegex"in t)||!(t.languageRegex instanceof RegExp))throw new Error("Tokenizer must have `languageRegex` attribute set and it should be a regular expression.");if(!("lang_to_token"in t)||typeof t.lang_to_token!="function")throw new Error("Tokenizer must have `lang_to_token` attribute set and it should be a function.");const a=n.src_lang,i=n.tgt_lang;if(!t.language_codes.includes(i))throw new Error(`Target language code "${i}" is not valid. Must be one of: {${t.language_codes.join(", ")}}`);if(a!==void 0){if(!t.language_codes.includes(a))throw new Error(`Source language code "${a}" is not valid. Must be one of: {${t.language_codes.join(", ")}}`);for(const s of t.post_processor.config.single)if("SpecialToken"in s&&t.languageRegex.test(s.SpecialToken.id)){s.SpecialToken.id=t.lang_to_token(a);break}}return n.forced_bos_token_id=t.model.convert_tokens_to_ids([t.lang_to_token(i)])[0],t._call(e,r)}class Ov extends Ee{constructor(e,r){super(e,r),this.languageRegex=/^[a-z]{3}_[A-Z][a-z]{3}$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)),this.lang_to_token=n=>n}_build_translation_inputs(e,r,n){return pl(this,e,r,n)}}class zv extends Ee{constructor(e,r){super(e,r),this.languageRegex=/^__[a-z]{2,3}__$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)).map(n=>n.slice(2,-2)),this.lang_to_token=n=>`__${n}__`}_build_translation_inputs(e,r,n){return pl(this,e,r,n)}}class Pv extends Ee{_default_chat_template='{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}';get timestamp_begin(){return this.model.convert_tokens_to_ids(["<|notimestamps|>"])[0]+1}_decode_asr(e,{return_timestamps:r=!1,return_language:n=!1,time_precision:a=null,force_full_sequences:i=!0}={}){if(a===null)throw Error("Must specify time_precision");let s=null;const o=r==="word";function l(){return{language:s,timestamp:[null,null],text:""}}const u=[];let p=l(),h=0;const m=this.timestamp_begin;let d=[],_=[],w=!1,v=null;const S=new Set(this.all_special_ids);for(const T of e){const A=T.tokens,P=o?T.token_timestamps:null;let B=null,D=m;if("stride"in T){const[ie,te,de]=T.stride;if(h-=te,v=ie-de,te&&(D=te/a+m),de)for(let se=A.length-1;se>=0;--se){const M=Number(A[se]);if(M>=m){if(B!==null&&(M-m)*a=m){const de=(te-m)*a+h,se=Mi(de,2);if(B!==null&&te>=B)w=!0;else if(w||d.length>0&&te0?(d.push(q),o&&_.push(H)):d.every(ie=>ie.length===0)&&(p=l(),d=[],q=[],_=[],H=[])}if(d.length>0){if(i&&r)throw new Error("Whisper did not predict an ending timestamp, which can happen if audio is cut off in the middle of a word. Also make sure WhisperTimeStampLogitsProcessor was used during generation.");const[T,A]=this.findLongestCommonSequence(d,_),P=this.decode(T);p.text=P,o&&(p.words=this.collateWordTimestamps(T,A,s)),u.push(p)}let $=Object.create(null);const E=u.map(T=>T.text).join("");if(r||n){for(let T=0;T0;let o=s?[]:null,l=s?r[0]:null;for(let u=1;use===ie[M]).length,de=te/T+A;te>1&&de>h&&(h=de,m=[P,B,q,H])}const[_,w,v,S]=m,$=Math.floor((w+_)/2),E=Math.floor((S+v)/2);i.push(...n.slice(0,$)),n=p.slice(E),a=n.length,s&&(o.push(...l.slice(0,$)),l=r[u].slice(E))}return i.push(...n),s?(o.push(...l),[i,o]):[i,[]]}collateWordTimestamps(e,r,n){const[a,i,s]=this.combineTokensIntoWords(e,n),o=[];for(let l=0;l=a){const o=((s-a)*n).toFixed(2);i.push(`<|${o}|>`),i.push([])}else i[i.length-1].push(s);return i=i.map(s=>typeof s=="string"?s:super.decode(s,r)),i.join("")}splitTokensOnUnicode(e){const r=this.decode(e,{decode_with_timestamps:!0}),n="�",a=[],i=[],s=[];let o=[],l=[],u=0;for(let p=0;p=this.model.tokens_to_ids.get("<|endoftext|>"),_=p.startsWith(" "),w=p.trim(),v=l.test(w);if(d||_||v||i.length===0)i.push(p),s.push(h),o.push(m);else{const S=i.length-1;i[S]+=p,s[S].push(...h),o[S].push(...m)}}return[i,s,o]}mergePunctuations(e,r,n,a,i){const s=structuredClone(e),o=structuredClone(r),l=structuredClone(n);let u=s.length-2,p=s.length-1;for(;u>=0;)s[u].startsWith(" ")&&a.includes(s[u].trim())?(s[p]=s[u]+s[p],o[p]=ft(o[u],o[p]),l[p]=ft(l[u],l[p]),s[u]="",o[u]=[],l[u]=[]):p=u,--u;for(u=0,p=1;ph),o.filter(h=>h.length>0),l.filter(h=>h.length>0)]}get_decoder_prompt_ids({language:e=null,task:r=null,no_timestamps:n=!0}={}){const a=[];if(e){const i=_g(e),s=this.model.tokens_to_ids.get(`<|${i}|>`);if(s===void 0)throw new Error(`Unable to find language "${i}" in model vocabulary. Please report this issue at ${po}.`);a.push(s)}else a.push(null);if(r){if(r=r.toLowerCase(),r!=="transcribe"&&r!=="translate")throw new Error(`Task "${r}" is not supported. Must be one of: ["transcribe", "translate"]`);const i=this.model.tokens_to_ids.get(`<|${r}|>`);if(i===void 0)throw new Error(`Unable to find task "${r}" in model vocabulary. Please report this issue at ${po}.`);a.push(i)}else a.push(null);if(n){const i=this.model.tokens_to_ids.get("<|notimestamps|>");if(i===void 0)throw new Error(`Unable to find "<|notimestamps|>" in model vocabulary. Please report this issue at ${po}.`);a.push(i)}return a.map((i,s)=>[s+1,i]).filter(i=>i[1]!==null)}}class Rv extends Ee{}class Bv extends Ee{}class Dv extends Ee{}class Nv extends Ee{constructor(e,r){super(e,r),this.languageRegex=/^(>>\w+<<)\s*/g,this.supported_language_codes=this.model.vocab.filter(n=>this.languageRegex.test(n)),console.warn('WARNING: `MarianTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.')}_encode_text(e){if(e===null)return null;const[r,...n]=e.trim().split(this.languageRegex);if(n.length===0)return super._encode_text(r);if(n.length===2){const[a,i]=n;return this.supported_language_codes.includes(a)||console.warn(`Unsupported language code "${a}" detected, which may lead to unexpected behavior. Should be one of: ${JSON.stringify(this.supported_language_codes)}`),ft([a],super._encode_text(i))}}}class Fv extends Ee{}class Ag extends Ee{_default_chat_template="{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ ' ' }}{% endif %}{% endfor %}{{ eos_token }}"}class Lv extends Ag{}class Uv extends Ee{}class Wv extends Ee{}class Vv extends Ee{constructor(e,r){super(e,r),this.decoder=new Yb({})}}class Gv extends Ee{}class ht{static TOKENIZER_CLASS_MAPPING={T5Tokenizer:yv,DistilBertTokenizer:fv,CamembertTokenizer:mv,DebertaTokenizer:uv,DebertaV2Tokenizer:dv,BertTokenizer:iv,HerbertTokenizer:cv,ConvBertTokenizer:pv,RoFormerTokenizer:hv,XLMTokenizer:gv,ElectraTokenizer:_v,MobileBertTokenizer:ov,SqueezeBertTokenizer:lv,AlbertTokenizer:sv,GPT2Tokenizer:Eg,BartTokenizer:wv,MBartTokenizer:Cg,MBart50Tokenizer:bv,RobertaTokenizer:vv,WhisperTokenizer:Pv,CodeGenTokenizer:Rv,CLIPTokenizer:Bv,SiglipTokenizer:Dv,MarianTokenizer:Nv,BloomTokenizer:$v,NllbTokenizer:Ov,M2M100Tokenizer:zv,LlamaTokenizer:Tg,CodeLlamaTokenizer:xv,XLMRobertaTokenizer:Sv,MPNetTokenizer:kv,FalconTokenizer:Ev,GPTNeoXTokenizer:Cv,EsmTokenizer:Tv,Wav2Vec2CTCTokenizer:Fv,BlenderbotTokenizer:Ag,BlenderbotSmallTokenizer:Lv,SpeechT5Tokenizer:Uv,NougatTokenizer:Wv,VitsTokenizer:Vv,Qwen2Tokenizer:Av,GemmaTokenizer:Iv,Grok1Tokenizer:Mv,CohereTokenizer:Gv,PreTrainedTokenizer:Ee};static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",legacy:o=null}={}){const[l,u]=await yg(e,{progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,legacy:o}),p=u.tokenizer_class?.replace(/Fast$/,"")??"PreTrainedTokenizer";let h=this.TOKENIZER_CLASS_MAPPING[p];return h||(console.warn(`Unknown tokenizer class "${p}", attempting to construct from base class.`),h=Ee),new h(l,u)}}async function Hv(t,e){return await Br(t,"config.json",!0,e)}function In(t){const e={};let r={};switch(t.model_type){case"llava":case"paligemma":r=In(t.text_config);break;case"moondream1":r=In(t.phi_config);break;case"musicgen":r=In(t.decoder);break;case"gpt2":case"gptj":case"codegen":case"gpt_bigcode":e.num_heads="n_head",e.num_layers="n_layer",e.hidden_size="n_embd";break;case"gpt_neox":case"stablelm":case"opt":case"phi":case"phi3":case"falcon":e.num_heads="num_attention_heads",e.num_layers="num_hidden_layers",e.hidden_size="hidden_size";break;case"llama":case"cohere":case"mistral":case"starcoder2":case"qwen2":e.num_heads="num_key_value_heads",e.num_layers="num_hidden_layers",e.hidden_size="hidden_size",e.num_attention_heads="num_attention_heads";break;case"gemma":e.num_heads="num_key_value_heads",e.num_layers="num_hidden_layers",e.dim_kv="head_dim";break;case"openelm":e.num_heads="num_kv_heads",e.num_layers="num_transformer_layers",e.dim_kv="head_dim";break;case"gpt_neo":e.num_heads="num_heads",e.num_layers="num_layers",e.hidden_size="hidden_size";break;case"bloom":e.num_heads="n_head",e.num_layers="n_layer",e.hidden_size="hidden_size";break;case"mpt":e.num_heads="n_heads",e.num_layers="n_layers",e.hidden_size="d_model";break;case"t5":case"mt5":case"longt5":e.num_decoder_layers="num_decoder_layers",e.num_decoder_heads="num_heads",e.decoder_dim_kv="d_kv",e.num_encoder_layers="num_layers",e.num_encoder_heads="num_heads",e.encoder_dim_kv="d_kv";break;case"bart":case"mbart":case"marian":case"whisper":case"m2m_100":case"blenderbot":case"blenderbot-small":e.num_decoder_layers="decoder_layers",e.num_decoder_heads="decoder_attention_heads",e.decoder_hidden_size="d_model",e.num_encoder_layers="encoder_layers",e.num_encoder_heads="encoder_attention_heads",e.encoder_hidden_size="d_model";break;case"speecht5":e.num_decoder_layers="decoder_layers",e.num_decoder_heads="decoder_attention_heads",e.decoder_hidden_size="hidden_size",e.num_encoder_layers="encoder_layers",e.num_encoder_heads="encoder_attention_heads",e.encoder_hidden_size="hidden_size";break;case"trocr":e.num_encoder_layers=e.num_decoder_layers="decoder_layers",e.num_encoder_heads=e.num_decoder_heads="decoder_attention_heads",e.encoder_hidden_size=e.decoder_hidden_size="d_model";break;case"musicgen_decoder":e.num_encoder_layers=e.num_decoder_layers="num_hidden_layers",e.num_encoder_heads=e.num_decoder_heads="num_attention_heads",e.encoder_hidden_size=e.decoder_hidden_size="hidden_size";break;case"vision-encoder-decoder":const a=In(t.encoder),i=In(t.decoder),s="num_decoder_layers"in i,o={};return s?(o.num_decoder_layers=i.num_layers,o.num_decoder_heads=i.num_heads,o.decoder_hidden_size=i.hidden_size,o.num_encoder_layers=a.num_layers,o.num_encoder_heads=a.num_heads,o.encoder_hidden_size=a.hidden_size):(o.num_layers=i.num_layers,o.num_heads=i.num_heads,o.hidden_size=i.hidden_size),o}const n={...r,...rn(t,["model_type","multi_query","is_encoder_decoder"])};for(const a in e)n[a]=t[e[a]];return n}function Ig(t,{prefix:e="past_key_values",encoder_add_pkv:r=!0}={}){const n={},a=t.normalized_config,i=1;if(a.is_encoder_decoder&&r){const s=a.encoder_dim_kv??a.encoder_hidden_size/a.num_encoder_heads,o=a.decoder_dim_kv??a.decoder_hidden_size/a.num_decoder_heads,l=[i,a.num_encoder_heads,0,s],u=[i,a.num_decoder_heads,0,o];for(let p=0;p=1&&s[s.length-1]>=this.timestamp_begin,l=s.length<2||s[s.length-2]>=this.timestamp_begin;if(o&&(l?i.subarray(this.timestamp_begin).fill(-1/0):i.subarray(0,this.eos_token_id).fill(-1/0)),e[n].length===this.begin_index&&this.max_initial_timestamp_index!==null){const m=this.timestamp_begin+this.max_initial_timestamp_index;i.subarray(m+1).fill(-1/0)}const u=U0(i),p=Math.log(u.subarray(this.timestamp_begin).map(Math.exp).reduce((m,d)=>m+d)),h=Kt(u.subarray(0,this.timestamp_begin))[0];p>h&&i.subarray(0,this.timestamp_begin).fill(-1/0)}return r}}class Zv extends yr{constructor(e){super(),this.no_repeat_ngram_size=e}getNgrams(e){const r=e.length,n=[];for(let i=0;i1 to use the classifier free guidance processor, got guidance scale ${e}.`);this.guidance_scale=e}_call(e,r){if(r.dims[0]!==2*e.length)throw new Error(`Logits should have twice the batch size of the input ids, the first half of batches corresponding to the conditional inputs, and the second half of batches corresponding to the unconditional inputs. Got batch size ${r.dims[0]} for the logits and ${e.length} for the input ids.`);const n=e.length,a=r.slice([0,n],null),i=r.slice([n,r.dims[0]],null);for(let s=0;s1)throw new Error(`\`top_p\` must be a float > 0 and < 1, but is ${e}`);if(!Number.isInteger(n)||n<1)throw new Error(`\`min_tokens_to_keep\` must be a positive integer, but is ${n}`);this.top_p=e,this.filter_value=r,this.min_tokens_to_keep=n}}class s1 extends hl{constructor(e,{filter_value:r=-1/0,min_tokens_to_keep:n=1}={}){if(super(),!Number.isInteger(e)||e<0)throw new Error(`\`top_k\` must be a positive integer, but is ${e}`);this.top_k=Math.max(e,n),this.filter_value=r}}class Og{max_length=20;max_new_tokens=null;min_length=0;min_new_tokens=null;early_stopping=!1;max_time=null;do_sample=!1;num_beams=1;num_beam_groups=1;penalty_alpha=null;use_cache=!0;temperature=1;top_k=50;top_p=1;typical_p=1;epsilon_cutoff=0;eta_cutoff=0;diversity_penalty=0;repetition_penalty=1;encoder_repetition_penalty=1;length_penalty=1;no_repeat_ngram_size=0;bad_words_ids=null;force_words_ids=null;renormalize_logits=!1;constraints=null;forced_bos_token_id=null;forced_eos_token_id=null;remove_invalid_values=!1;exponential_decay_length_penalty=null;suppress_tokens=null;begin_suppress_tokens=null;forced_decoder_ids=null;guidance_scale=null;num_return_sequences=1;output_attentions=!1;output_hidden_states=!1;output_scores=!1;return_dict_in_generate=!1;pad_token_id=null;bos_token_id=null;eos_token_id=null;encoder_no_repeat_ngram_size=0;decoder_start_token_id=null;generation_kwargs={};constructor(e){Object.assign(this,rn(e,Object.getOwnPropertyNames(this)))}}class fl extends bt{_call(e,r){throw Error("StoppingCriteria needs to be subclassed")}}class ml extends bt{constructor(){super(),this.criteria=[]}push(e){this.criteria.push(e)}extend(e){e instanceof ml?e=e.criteria:e instanceof fl&&(e=[e]),this.criteria.push(...e)}_call(e,r){const n=new Array(e.length).fill(!1);for(const a of this.criteria){const i=a(e,r);for(let s=0;sr.length>=this.max_length)}}class l1 extends fl{constructor(e){super(),Array.isArray(e)||(e=[e]),this.eos_token_id=e}_call(e,r){return e.map(n=>{const a=n.at(-1);return this.eos_token_id.some(i=>a==i)})}}class ss extends bt{constructor(e){super(),this.generation_config=e}async _call(e){return this.sample(e)}async sample(e){throw Error("sample should be implemented in subclasses.")}getLogits(e,r){let n=e.dims.at(-1),a=e.data;if(r===-1)a=a.slice(-n);else{let i=r*n;a=a.slice(i,i+n)}return a}randomSelect(e){let r=0;for(let a=0;a1)return new c1(e);if(e.num_return_sequences>1)throw Error(`num_return_sequences has to be 1 when doing greedy search, but is ${e.num_return_sequences}.`);return new u1(e)}}class u1 extends ss{async sample(e){const r=Kt(e.data)[1];return[[BigInt(r),0]]}}class d1 extends ss{async sample(e){let r=e.dims.at(-1);this.generation_config.top_k>0&&(r=Math.min(this.generation_config.top_k,r));const[n,a]=await Nn(e,r),i=wt(n.data);return Array.from({length:this.generation_config.num_beams},()=>{const s=this.randomSelect(i);return[a.data[s],Math.log(i[s])]})}}class c1 extends ss{async sample(e){let r=e.dims.at(-1);this.generation_config.top_k>0&&(r=Math.min(this.generation_config.top_k,r));const[n,a]=await Nn(e,r),i=wt(n.data);return Array.from({length:this.generation_config.num_beams},(s,o)=>[a.data[o],Math.log(i[o])])}}class p1 extends Og{return_timestamps=null;return_token_timestamps=null;num_frames=null;alignment_heads=null;task=null;language=null;no_timestamps_token_id=null;prompt_ids=null;is_multilingual=null;lang_to_id=null;task_to_id=null;max_initial_timestamp_index=1}const ve={EncoderOnly:0,EncoderDecoder:1,Seq2Seq:2,Vision2Seq:3,DecoderOnly:4,MaskGeneration:5,ImageTextToText:6,Musicgen:7},es=new Map,zg=new Map,ba=new Map;async function h1(t,e,r){let n=r.device;n&&typeof n!="string"&&(n.hasOwnProperty(e)?n=n[e]:(console.warn(`Device not specified for ${e}. Using the default device.`),n=null));const a=Sw(n);let i=r.dtype;if(typeof i!="string"&&(i&&i.hasOwnProperty(e)?i=i[e]:(i=qv[a[0]],console.warn(`Dtype not specified for ${e}. Using the default dtype: ${i}.`))),xp.hasOwnProperty(i)){if(i===Rt.fp16&&!await jv())throw new Error("The device does not support fp16.")}else throw new Error(`Invalid dtype: ${i}. Should be one of: ${Object.keys(Rt).join(", ")}`);const s=xp[i],o=`${r.subfolder??""}/${e}${s}.onnx`,l={...r.session_options};l.executionProviders??=a;const u=Ii(t,o,!0,r);let p=[];if(r.use_external_data_format){if(an.IS_NODE_ENV)throw new Error("External data format is not yet supported in Node.js");const m=`${e}${s}.onnx_data`,d=`${r.subfolder??""}/${m}`;p.push(new Promise(async(_,w)=>{const v=await Ii(t,d,!0,r);_({path:m,data:v})}))}else l.externalData!==void 0&&(p=l.externalData.map(async m=>{if(typeof m.data=="string"){const d=await Ii(t,m.data,!0,r);return{...m,data:d}}return m}));if(p.length>0&&(l.externalData=await Promise.all(p)),n==="webgpu"){const m=Ig(r.config,{prefix:"present"});if(Object.keys(m).length>0){const d={};for(const _ in m)d[_]="gpu-buffer";l.preferredOutputLocation=d}}return{buffer:await u,session_options:l}}async function Yr(t,e,r){const n=Object.keys(e),a=await Promise.all(n.map(async s=>h1(t,e[s],r))),i={};for(let s=0;s0)throw new Error(`An error occurred during model execution: "Missing the following inputs: ${n.join(", ")}.`);const a=Object.keys(e).length,i=t.inputNames.length;if(a>i){let s=Object.keys(e).filter(o=>!t.inputNames.includes(o));console.warn(`WARNING: Too many inputs were provided (${a} > ${i}). The following inputs will be ignored: "${s.join(", ")}".`)}return r}async function Nr(t,e){const r=f1(t,e);try{const n=Object.fromEntries(Object.entries(r).map(([i,s])=>[i,s.ort_tensor]));let a=await t.run(n);return a=Pg(a),a}catch(n){throw console.error(`An error occurred during model execution: "${n}".`),console.error("Inputs given to model:",r),n}}function Pg(t){for(let e in t)fg(t[e])?t[e]=new ue(t[e]):typeof t[e]=="object"&&Pg(t[e]);return t}function Rg(t){if(t instanceof ue)return t;if(t.length===0)throw Error("items must be non-empty");if(Array.isArray(t[0])){if(t.some(e=>e.length!==t[0].length))throw Error("Unable to create tensor, you should probably activate truncation and/or padding with 'padding=True' and/or 'truncation=True' to have batched tensors with the same length.");return new ue("int64",BigInt64Array.from(t.flat().map(e=>BigInt(e))),[t.length,t[0].length])}else return new ue("int64",BigInt64Array.from(t.map(e=>BigInt(e))),[1,t.length])}function Bg(t){return new ue("bool",[t],[1])}async function Sp(t,e){let{encoder_outputs:r,past_key_values:n}=e;if(!r){const l=rn(e,t.sessions.model.inputNames);r=(await Ca(t,l)).last_hidden_state}const{input_ids:a,decoder_input_ids:i,...s}=e;return s.input_ids=i,s.encoder_hidden_states=r,t.sessions.decoder_model_merged.inputNames.includes("encoder_attention_mask")&&(s.encoder_attention_mask=e.attention_mask),await gl(t,s,!0)}async function Ca(t,e){const r=t.sessions.model,n=Object.create(null);for(const a of r.inputNames)n[a]=e[a];return r.inputNames.includes("token_type_ids")&&!n.token_type_ids&&(n.token_type_ids=new ue("int64",new BigInt64Array(n.input_ids.data.length),n.input_ids.dims)),await Nr(r,n)}async function gl(t,e,r=!1){const n=t.sessions[r?"decoder_model_merged":"model"],{past_key_values:a,...i}=e;n.inputNames.includes("use_cache_branch")&&(i.use_cache_branch=Bg(!!a)),n.inputNames.includes("position_ids")&&i.attention_mask&&!i.position_ids&&(i.position_ids=g1(i,a)),t.addPastKeyValues(i,a);const s=rn(i,n.inputNames);return await Nr(n,s)}async function m1(t,{input_ids:e=null,attention_mask:r=null,pixel_values:n=null,position_ids:a=null,inputs_embeds:i=null,past_key_values:s=null,generation_config:o=null,logits_processor:l=null,...u}){if(!i){if(i=await t.encode_text({input_ids:e}),n&&e.dims[1]!==1){const h=await t.encode_image({pixel_values:n});({inputs_embeds:i,attention_mask:r}=t._merge_input_ids_with_image_features({image_features:h,inputs_embeds:i,input_ids:e,attention_mask:r}))}else if(s&&n&&e.dims[1]===1){const h=e.dims[1],m=Object.values(s)[0].dims.at(-2);r=gr([Ma([e.dims[0],m]),r.slice(null,[r.dims[1]-h,r.dims[1]])],1)}}return await gl(t,{inputs_embeds:i,past_key_values:s,attention_mask:r,position_ids:a,generation_config:o,logits_processor:l},!0)}function g1(t,e=null){const{input_ids:r,inputs_embeds:n,attention_mask:a}=t,[i,s]=a.dims,o=new BigInt64Array(a.data.length);for(let u=0;ui.dims[1])){if(ao==t.config.image_token_index)){const o=t.config.num_image_tokens;if(!o)throw new Error("`num_image_tokens` is missing in the model configuration.");const l=i.dims[1]-(a-o);r.input_ids=i.slice(null,[-l,null]),r.attention_mask=Ma([1,a+l])}}}return r}function _1(t,e,r,n){const{...a}=r;return r.past_key_values&&(e=e.map(s=>[s.at(-1)])),a.decoder_input_ids=Rg(e),a}class Q extends bt{main_input_name="input_ids";forward_params=["input_ids","attention_mask"];constructor(e,r){super(),this.config=e,this.sessions=r;const n=ba.get(this.constructor),a=es.get(n);this.can_generate=!1,this._forward=null,this._prepare_inputs_for_generation=null,a===ve.DecoderOnly?(this.can_generate=!0,this._forward=gl,this._prepare_inputs_for_generation=kp):a===ve.Seq2Seq||a===ve.Vision2Seq||a===ve.Musicgen?(this.can_generate=!0,this._forward=Sp,this._prepare_inputs_for_generation=_1):a===ve.EncoderDecoder?this._forward=Sp:a===ve.ImageTextToText?(this.can_generate=!0,this._forward=m1,this._prepare_inputs_for_generation=kp):this._forward=Ca,this.can_generate&&this.forward_params.push("past_key_values"),this.custom_config=this.config["transformers.js_config"]??{}}async dispose(){const e=[];for(const r of Object.values(this.sessions))r?.handler?.dispose&&e.push(r.handler.dispose());return await Promise.all(e)}static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",model_file_name:o=null,subfolder:l="onnx",device:u=null,dtype:p=null,use_external_data_format:h=null,session_options:m={}}={}){let d={progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,model_file_name:o,subfolder:l,device:u,dtype:p,use_external_data_format:h,session_options:m};const _=ba.get(this),w=es.get(_);d.config=await Mg.from_pretrained(e,d);let v;return w===ve.DecoderOnly?v=await Promise.all([Yr(e,{model:d.model_file_name??"model"},d),Br(e,"generation_config.json",!1,d)]):w===ve.Seq2Seq||w===ve.Vision2Seq?v=await Promise.all([Yr(e,{model:"encoder_model",decoder_model_merged:"decoder_model_merged"},d),Br(e,"generation_config.json",!1,d)]):w===ve.MaskGeneration?v=await Promise.all([Yr(e,{model:"vision_encoder",prompt_encoder_mask_decoder:"prompt_encoder_mask_decoder"},d)]):w===ve.EncoderDecoder?v=await Promise.all([Yr(e,{model:"encoder_model",decoder_model_merged:"decoder_model_merged"},d)]):w===ve.ImageTextToText?v=await Promise.all([Yr(e,{embed_tokens:"embed_tokens",vision_encoder:"vision_encoder",decoder_model_merged:"decoder_model_merged"},d),Br(e,"generation_config.json",!1,d)]):w===ve.Musicgen?v=await Promise.all([Yr(e,{model:"text_encoder",decoder_model_merged:"decoder_model_merged",encodec_decode:"encodec_decode"},d),Br(e,"generation_config.json",!1,d)]):(w!==ve.EncoderOnly&&console.warn(`Model type for '${_??n?.model_type}' not found, assuming encoder-only architecture. Please report this at https://github.com/xenova/transformers.js/issues/new/choose.`),v=await Promise.all([Yr(e,{model:d.model_file_name??"model"},d)])),new this(d.config,...v)}async _call(e){return await this.forward(e)}async forward(e){return await this._forward(this,e)}_get_logits_warper(e){const r=new Ro;return e.temperature!==null&&e.temperature!==1&&r.push(new a1(e.temperature)),e.top_k!==null&&e.top_k!==0&&r.push(new s1(e.top_k)),e.top_p!==null&&e.top_p<1&&r.push(new i1(e.top_p)),r}_get_logits_processor(e,r,n=null){const a=new Ro;if(e.repetition_penalty!==null&&e.repetition_penalty!==1&&a.push(new Jv(e.repetition_penalty)),e.no_repeat_ngram_size!==null&&e.no_repeat_ngram_size>0&&a.push(new Zv(e.no_repeat_ngram_size)),e.bad_words_ids!==null&&a.push(new r1(e.bad_words_ids,e.eos_token_id)),e.min_length!==null&&e.eos_token_id!==null&&e.min_length>0&&a.push(new e1(e.min_length,e.eos_token_id)),e.min_new_tokens!==null&&e.eos_token_id!==null&&e.min_new_tokens>0&&a.push(new t1(r,e.min_new_tokens,e.eos_token_id)),e.forced_bos_token_id!==null&&a.push(new Kv(e.forced_bos_token_id)),e.forced_eos_token_id!==null&&a.push(new Yv(e.max_length,e.forced_eos_token_id)),e.begin_suppress_tokens!==null){const i=r>1||e.forced_bos_token_id===null?r:r+1;a.push(new Xv(e.begin_suppress_tokens,i))}return e.guidance_scale!==null&&e.guidance_scale>1&&a.push(new n1(e.guidance_scale)),n!==null&&a.extend(n),a}_prepare_generation_config(e,r,n=Og){const a={...this.config};for(const s of["decoder","generator","text_config"])s in a&&Object.assign(a,a[s]);const i=new n(a);return"generation_config"in this&&Object.assign(i,this.generation_config),e&&Object.assign(i,e),r&&Object.assign(i,rn(r,Object.getOwnPropertyNames(i))),i}_get_stopping_criteria(e,r=null){const n=new ml;return e.max_length!==null&&n.push(new o1(e.max_length,this.config.max_position_embeddings??null)),e.eos_token_id!==null&&n.push(new l1(e.eos_token_id)),r&&n.extend(r),n}_validate_model_class(){if(!this.can_generate){const e=[Sl,kl,xl,$l],r=ba.get(this.constructor),n=new Set,a=this.config.model_type;for(const s of e){const o=s.get(a);o&&n.add(o[0])}let i=`The current model class (${r}) is not compatible with \`.generate()\`, as it doesn't have a language model head.`;throw n.size>0&&(i+=` Please use the following class instead: ${[...n].join(", ")}`),Error(i)}}prepare_inputs_for_generation(...e){return this._prepare_inputs_for_generation(this,...e)}_update_model_kwargs_for_generation({generated_input_ids:e,outputs:r,model_inputs:n,is_encoder_decoder:a}){return n.past_key_values=this.getPastKeyValues(r,n.past_key_values),n.input_ids=new ue("int64",e.flat(),[e.length,1]),a||(n.attention_mask=gr([n.attention_mask,Ma([n.attention_mask.dims[0],1])],1)),n.position_ids=null,n}_prepare_model_inputs({inputs:e,bos_token_id:r,model_kwargs:n}){const a=rn(n,this.forward_params),i=this.main_input_name;if(i in a){if(e)throw new Error("`inputs`: {inputs}` were passed alongside {input_name} which is not allowed. Make sure to either pass {inputs} or {input_name}=...")}else a[i]=e;return{inputs_tensor:a[i],model_inputs:a,model_input_name:i}}async _prepare_encoder_decoder_kwargs_for_generation({inputs_tensor:e,model_inputs:r,model_input_name:n,generation_config:a}){const i=rn(r,this.sessions.model.inputNames);let{last_hidden_state:s}=await Ca(this,i);return a.guidance_scale!==null&&a.guidance_scale>1&&(s=gr([s,Pw(s,0)],0),"attention_mask"in r&&(r.attention_mask=gr([r.attention_mask,Dw(r.attention_mask)],0))),r.encoder_outputs=s,r}_prepare_decoder_input_ids_for_generation({batch_size:e,model_input_name:r,model_kwargs:n,decoder_start_token_id:a,bos_token_id:i,generation_config:s}){let{decoder_input_ids:o,...l}=n;if(!o)if(a??=i,this.config.model_type==="musicgen")o=Array.from({length:e*this.config.decoder.num_codebooks},()=>[a]);else if(Array.isArray(a)){if(a.length!==e)throw new Error(`\`decoder_start_token_id\` expcted to have length ${e} but got ${a.length}`);o=a}else o=Array.from({length:e},()=>[a]);return o=Rg(o),n.decoder_attention_mask=Rw(o),{input_ids:o,model_inputs:l}}async generate({inputs:e=null,generation_config:r=null,logits_processor:n=null,stopping_criteria:a=null,streamer:i=null,...s}){this._validate_model_class(),r=this._prepare_generation_config(r,s);let{inputs_tensor:o,model_inputs:l,model_input_name:u}=this._prepare_model_inputs({inputs:e,model_kwargs:s});const p=this.config.is_encoder_decoder;p&&("encoder_outputs"in l||(l=await this._prepare_encoder_decoder_kwargs_for_generation({inputs_tensor:o,model_inputs:l,model_input_name:u,generation_config:r})));let h;p?{input_ids:h,model_inputs:l}=this._prepare_decoder_input_ids_for_generation({batch_size:l[u].dims.at(0),model_input_name:u,model_kwargs:l,decoder_start_token_id:r.decoder_start_token_id,bos_token_id:r.bos_token_id,generation_config:r}):h=l[u];let m=h.dims.at(-1);r.max_new_tokens!==null&&(r.max_length=m+r.max_new_tokens);const d=this._get_logits_processor(r,m,n),_=this._get_stopping_criteria(r,a),w=l[u].dims.at(0),v=ss.getSampler(r),S=new Array(w).fill(0),$=h.tolist();i&&i.put($);let E=null;for(;;){l=this.prepare_inputs_for_generation($,l,r);const A=await this.forward(l),P=A.logits.slice(null,-1,null),B=d($,P),D=[];for(let H=0;HH)){r.return_dict_in_generate&&(E=this.getPastKeyValues(A,l.past_key_values,!1));break}l=this._update_model_kwargs_for_generation({generated_input_ids:D,outputs:A,model_inputs:l,is_encoder_decoder:p})}i&&i.end();const T=new ue("int64",$.flat(),[$.length,$[0].length]);return r.return_dict_in_generate?{sequences:T,past_key_values:E}:T}addAttentionsToBeam(e,r){if(this.config.is_encoder_decoder){if(!r.cross_attentions||r.cross_attentions.length===0)throw Error("`output_attentions` is true, but the model did not produce cross-attentions. This is most likely because the model was not exported with `output_attentions=True`.");e.cross_attentions||(e.cross_attentions=[]),e.cross_attentions.push(r.cross_attentions)}if(!r.decoder_attentions||r.decoder_attentions.length===0)throw Error("`output_attentions` is true, but the model did not produce decoder-attentions. This is most likely because the model was not exported with `output_attentions=True`.");e.decoder_attentions||(e.decoder_attentions=[]),e.decoder_attentions.push(r.decoder_attentions)}groupBeams(e){const r=Object.create(null);for(const n of e)r[n.id]===void 0?r[n.id]=[n]:r[n.id].push(n);return Object.values(r)}getPastKeyValues(e,r,n=!0){const a=Object.create(null);for(const i in e)if(i.startsWith("present")){let s=i.replace("present","past_key_values");if(r&&i.includes("encoder"))a[s]=r[s];else{if(n&&r){const o=r[s];o.location==="gpu-buffer"&&o.dispose()}a[s]=e[i]}}return a}getAttentions(e){const r=Object.create(null);for(const n of["cross_attentions","decoder_attentions"]){const a=[];for(const i in e)if(i.startsWith(n)){const s=i.split(".").pop();a[s]=e[i]}r[n]=a}return r}addPastKeyValues(e,r){if(r)Object.assign(e,r);else{const n=this.custom_config.kv_cache_dtype??"float32",a=n==="float16"?new Uint16Array:[],i=Ig(this.config);for(const s in i)e[s]=new ue(n,a,i[s])}}}class Yt{}class Pa extends Q{}class y1 extends Pa{}class w1 extends Pa{async _call(e){return new $t(await super._call(e))}}class b1 extends Pa{async _call(e){return new Ae(await super._call(e))}}class v1 extends Pa{async _call(e){return new vt(await super._call(e))}}class $1 extends Pa{async _call(e){return new Ct(await super._call(e))}}class x1 extends Q{}class S1 extends x1{}class Ra extends Q{}class k1 extends Ra{}class E1 extends Ra{async _call(e){return new $t(await super._call(e))}}class C1 extends Ra{async _call(e){return new Ae(await super._call(e))}}class T1 extends Ra{async _call(e){return new vt(await super._call(e))}}class A1 extends Ra{async _call(e){return new Ct(await super._call(e))}}class Ba extends Q{}class I1 extends Ba{}class M1 extends Ba{async _call(e){return new $t(await super._call(e))}}class O1 extends Ba{async _call(e){return new Ae(await super._call(e))}}class z1 extends Ba{async _call(e){return new vt(await super._call(e))}}class P1 extends Ba{async _call(e){return new Ct(await super._call(e))}}class Da extends Q{}class R1 extends Da{}class B1 extends Da{async _call(e){return new $t(await super._call(e))}}class D1 extends Da{async _call(e){return new Ae(await super._call(e))}}class N1 extends Da{async _call(e){return new vt(await super._call(e))}}class F1 extends Da{async _call(e){return new Ct(await super._call(e))}}class Na extends Q{}class L1 extends Na{}class U1 extends Na{async _call(e){return new $t(await super._call(e))}}class W1 extends Na{async _call(e){return new Ae(await super._call(e))}}class V1 extends Na{async _call(e){return new vt(await super._call(e))}}class G1 extends Na{async _call(e){return new Ct(await super._call(e))}}class Fa extends Q{}class H1 extends Fa{}class j1 extends Fa{async _call(e){return new $t(await super._call(e))}}class q1 extends Fa{async _call(e){return new Ae(await super._call(e))}}class K1 extends Fa{async _call(e){return new vt(await super._call(e))}}class Y1 extends Fa{async _call(e){return new Ct(await super._call(e))}}class La extends Q{}class X1 extends La{}class Q1 extends La{async _call(e){return new $t(await super._call(e))}}class Z1 extends La{async _call(e){return new Ae(await super._call(e))}}class J1 extends La{async _call(e){return new vt(await super._call(e))}}class e2 extends La{async _call(e){return new Ct(await super._call(e))}}class Ua extends Q{}class t2 extends Ua{}class r2 extends Ua{async _call(e){return new Ae(await super._call(e))}}class n2 extends Ua{async _call(e){return new vt(await super._call(e))}}class a2 extends Ua{async _call(e){return new Ct(await super._call(e))}}class i2 extends Ua{async _call(e){return new $t(await super._call(e))}}class os extends Q{}class s2 extends os{}class o2 extends os{async _call(e){return new $t(await super._call(e))}}class l2 extends os{async _call(e){return new Ae(await super._call(e))}}class u2 extends os{async _call(e){return new vt(await super._call(e))}}class ls extends Q{}class d2 extends ls{}class c2 extends ls{async _call(e){return new $t(await super._call(e))}}class p2 extends ls{async _call(e){return new Ae(await super._call(e))}}class h2 extends ls{async _call(e){return new Ct(await super._call(e))}}class Wa extends Q{}class f2 extends Wa{}class m2 extends Wa{async _call(e){return new $t(await super._call(e))}}class g2 extends Wa{async _call(e){return new Ae(await super._call(e))}}class _2 extends Wa{async _call(e){return new vt(await super._call(e))}}class y2 extends Wa{async _call(e){return new Ct(await super._call(e))}}class us extends Q{}class w2 extends us{}class b2 extends us{async _call(e){return new $t(await super._call(e))}}class v2 extends us{async _call(e){return new Ae(await super._call(e))}}class $2 extends us{async _call(e){return new Ct(await super._call(e))}}class ds extends Q{}class x2 extends ds{}class S2 extends ds{async _call(e){return new Ae(await super._call(e))}}class k2 extends ds{async _call(e){return new Ct(await super._call(e))}}class E2 extends ds{async _call(e){return new $t(await super._call(e))}}class Dg extends Q{forward_params=["input_ids","attention_mask","encoder_outputs","decoder_input_ids","decoder_attention_mask","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}}class C2 extends Dg{}class T2 extends Dg{}class Ng extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class A2 extends Ng{}class I2 extends Ng{}class Fg extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class M2 extends Fg{}class O2 extends Fg{}class _l extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class z2 extends _l{}class P2 extends _l{}class R2 extends _l{async _call(e){return new Ae(await super._call(e))}}class cs extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class B2 extends cs{}class D2 extends cs{}class N2 extends cs{async _call(e){return new Ae(await super._call(e))}}class F2 extends cs{}class Lg extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class L2 extends Lg{}class U2 extends Lg{}class Ug extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class W2 extends Ug{}class V2 extends Ug{}class Va extends Q{}class G2 extends Va{}class H2 extends Va{async _call(e){return new $t(await super._call(e))}}class j2 extends Va{async _call(e){return new Ae(await super._call(e))}}class q2 extends Va{async _call(e){return new vt(await super._call(e))}}class K2 extends Va{async _call(e){return new Ct(await super._call(e))}}class Ga extends Q{}class Y2 extends Ga{}class X2 extends Ga{async _call(e){return new $t(await super._call(e))}}class Q2 extends Ga{async _call(e){return new Ae(await super._call(e))}}class Z2 extends Ga{async _call(e){return new vt(await super._call(e))}}class J2 extends Ga{async _call(e){return new Ct(await super._call(e))}}class Ha extends Q{}class e$ extends Ha{}class t$ extends Ha{async _call(e){return new $t(await super._call(e))}}class r$ extends Ha{async _call(e){return new Ae(await super._call(e))}}class n$ extends Ha{async _call(e){return new vt(await super._call(e))}}class a$ extends Ha{async _call(e){return new Ct(await super._call(e))}}class Wg extends Q{}class i$ extends Wg{}class s$ extends Wg{}class Vg extends Q{requires_attention_mask=!1;main_input_name="input_features";forward_params=["input_features","attention_mask","decoder_input_ids","decoder_attention_mask","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}}class o$ extends Vg{}class l$ extends Vg{_prepare_generation_config(e,r){return super._prepare_generation_config(e,r,p1)}_retrieve_init_tokens(e){const r=[e.decoder_start_token_id];let n=e.language;const a=e.task;if(e.is_multilingual){n||(console.warn("No language specified - defaulting to English (en)."),n="en");const s=`<|${_g(n)}|>`;r.push(e.lang_to_id[s]),r.push(e.task_to_id[a??"transcribe"])}else if(n||a)throw new Error("Cannot specify `task` or `language` for an English-only model. If the model is intended to be multilingual, pass `is_multilingual=true` to generate, or update the generation config.");return!e.return_timestamps&&e.no_timestamps_token_id&&r.at(-1)!==e.no_timestamps_token_id?r.push(e.no_timestamps_token_id):e.return_timestamps&&r.at(-1)===e.no_timestamps_token_id&&(console.warn("<|notimestamps|> prompt token is removed from generation_config since `return_timestamps` is set to `true`."),r.pop()),r.filter(i=>i!=null)}async generate({inputs:e=null,generation_config:r=null,logits_processor:n=null,stopping_criteria:a=null,...i}){r=this._prepare_generation_config(r,i);const s=this._retrieve_init_tokens(r);return r.return_timestamps&&(n??=new Ro,n.push(new Qv(r,s))),await super.generate({inputs:e,generation_config:r,logits_processor:n,decoder_input_ids:s,...i})}_extract_token_timestamps(e,r,n=null,a=.02){if(!e.cross_attentions)throw new Error("Model outputs must contain cross attentions to extract timestamps. This is most likely because the model was not exported with `output_attentions=True`.");let i=this.config.median_filter_width;i===void 0&&(console.warn("Model config has no `median_filter_width`, using default value of 7."),i=7);const s=e.cross_attentions.map(u=>{let p=Array.from({length:this.config.decoder_layers},(v,S)=>gr(u.map($=>$[S]),2)),h=ka(r.map(([v,S])=>n?p[v].slice(null,S,null,[0,n]):p[v].slice(null,S)));h=h.transpose(1,0,2,3);let[m,d]=Iw(h,-2,0,!0),_=h.clone();for(let v=0;v<_.dims[0];++v){let S=_[v];for(let $=0;$h[S+1]-h[S]),_=ft([1],d).map(v=>!!v),w=[];for(let v=0;v<_.length;++v)_[v]&&w.push(m[v]*a);l[u].data.set(w,1)}return l}}class Gg extends Q{main_input_name="pixel_values";forward_params=["pixel_values","input_ids","encoder_hidden_states","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}}class u$ extends Q{forward_params=["input_ids","pixel_values","attention_mask","position_ids","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}}class Hg extends u${async encode_image({pixel_values:e}){const r=(await Nr(this.sessions.vision_encoder,{pixel_values:e})).image_features;return this.config.num_image_tokens||(console.warn(`The number of image tokens was not set in the model configuration. Setting it to the number of features detected by the vision encoder (${r.dims[1]}).`),this.config.num_image_tokens=r.dims[1]),r}async encode_text({input_ids:e}){return(await Nr(this.sessions.embed_tokens,{input_ids:e})).inputs_embeds}_merge_input_ids_with_image_features({inputs_embeds:e,image_features:r,input_ids:n,attention_mask:a}){const i=this.config.image_token_index,o=n.tolist().map(m=>m.findIndex(d=>d==i)),l=o.every(m=>m===-1),u=o.every(m=>m!==-1);if(!l&&!u)throw new Error("Every input should contain either 0 or 1 image token.");if(l)return{inputs_embeds:e,attention_mask:a};const p=[],h=[];for(let m=0;mi*s,1);e.input_labels=new ue("int64",new BigInt64Array(a).fill(1n),n)}const r={image_embeddings:e.image_embeddings,image_positional_embeddings:e.image_positional_embeddings};return e.input_points&&(r.input_points=e.input_points),e.input_labels&&(r.input_labels=e.input_labels),e.input_boxes&&(r.input_boxes=e.input_boxes),await Nr(this.sessions.prompt_encoder_mask_decoder,r)}async _call(e){return new Yx(await super._call(e))}}class Yx extends Yt{constructor({iou_scores:e,pred_masks:r}){super(),this.iou_scores=e,this.pred_masks=r}}class A_ extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class Xx extends A_{}class Qx extends A_{}class I_ extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class Zx extends I_{}class Jx extends I_{}class cn extends Q{}class eS extends cn{}class tS extends cn{async _call(e){return new Ln(await super._call(e))}}class rS extends cn{async _call(e){return new Ae(await super._call(e))}}class nS extends cn{async _call(e){return new vt(await super._call(e))}}class wl extends Q{}class aS extends wl{}class iS extends wl{async _call(e){return new Ln(await super._call(e))}}class sS extends wl{async _call(e){return new Ae(await super._call(e))}}class hs extends Q{}class oS extends hs{}class lS extends hs{async _call(e){return new Ln(await super._call(e))}}class uS extends hs{async _call(e){return new Ae(await super._call(e))}}class dS extends hs{async _call(e){return new vt(await super._call(e))}}class bl extends Q{}class cS extends bl{}class pS extends bl{async _call(e){return new Ln(await super._call(e))}}class hS extends bl{async _call(e){return new Ae(await super._call(e))}}class fS extends cn{}class mS extends cn{async _call(e){return new Ln(await super._call(e))}}class gS extends cn{async _call(e){return new Ae(await super._call(e))}}class ja extends Q{}class _S extends ja{}class yS extends ja{async _call(e){return new Ln(await super._call(e))}}class wS extends ja{async _call(e){return new Ae(await super._call(e))}}class bS extends ja{async _call(e){return new E3(await super._call(e))}}class vS extends ja{async _call(e){return new vt(await super._call(e))}}class M_ extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class $S extends M_{}class xS extends M_{async generate_speech(e,r,{threshold:n=.5,minlenratio:a=0,maxlenratio:i=20,vocoder:s=null}={}){const o={input_ids:e},{encoder_outputs:l,encoder_attention_mask:u}=await Ca(this,o),p=l.dims[1]/this.config.reduction_factor,h=Math.floor(p*i),m=Math.floor(p*a),d=this.config.num_mel_bins;let _=[],w=null,v=null,S=0;for(;;){++S;const T=Bg(!!v);let A;v?A=v.output_sequence_out:A=new ue("float32",new Float32Array(d),[1,1,d]);let P={use_cache_branch:T,output_sequence:A,encoder_attention_mask:u,speaker_embeddings:r,encoder_hidden_states:l};this.addPastKeyValues(P,w),v=await Nr(this.sessions.decoder_model_merged,P),w=this.getPastKeyValues(v,w);const{prob:B,spectrum:D}=v;if(_.push(D),S>=m&&(Array.from(B.data).filter(q=>q>=n).length>0||S>=h))break}const $=gr(_),{waveform:E}=await Nr(s.sessions.model,{spectrogram:$});return{spectrogram:$,waveform:E}}}class SS extends Q{main_input_name="spectrogram"}class kS extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class ES extends kS{}class O_ extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class CS extends O_{}class TS extends O_{}class z_ extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class AS extends z_{}class IS extends z_{}class P_ extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class MS extends P_{}class OS extends P_{}class vl extends Q{}class zS extends vl{}class PS extends vl{static async from_pretrained(e,r={}){return r.model_file_name??="text_model",super.from_pretrained(e,r)}}class RS extends vl{static async from_pretrained(e,r={}){return r.model_file_name??="audio_model",super.from_pretrained(e,r)}}class BS extends Q{}class R_ extends BS{async _call(e){return new T3(await super._call(e))}}class B_ extends Q{}class DS extends B_{}class NS extends B_{}class D_ extends Q{constructor(e,r,n){super(e,r),this.generation_config=n}}class FS extends D_{}class LS extends D_{}class N_ extends Q{}class US extends N_{}class WS extends N_{async _call(e){return new Ae(await super._call(e))}}class F_ extends Q{forward_params=["input_ids","attention_mask","encoder_outputs","decoder_input_ids","decoder_attention_mask","past_key_values"];constructor(e,r,n){super(e,r),this.generation_config=n}_apply_and_filter_by_delay_pattern_mask(e){const[r,n]=e.dims,a=this.config.decoder.num_codebooks,i=n-a;let s=0;for(let u=0;u0&&m<=i&&(e.data[s++]=e.data[u])}const o=Math.floor(r/a),l=s/(o*a);return new ue(e.type,e.data.slice(0,s),[o,a,l])}prepare_inputs_for_generation(e,r,n){let a=structuredClone(e);for(let s=0;s=o&&(a[s][o]=BigInt(this.config.decoder.pad_token_id));return n.guidance_scale!==null&&n.guidance_scale>1&&(a=a.concat(a)),super.prepare_inputs_for_generation(a,r,n)}async generate(e){const r=await super.generate(e),n=this._apply_and_filter_by_delay_pattern_mask(r).unsqueeze_(0),{audio_values:a}=await Nr(this.sessions.encodec_decode,{audio_codes:n});return a}}class L_ extends Q{}class VS extends L_{}class GS extends L_{async _call(e){return new Ae(await super._call(e))}}class U_ extends Q{}class HS extends U_{}class jS extends U_{async _call(e){return new Ae(await super._call(e))}}class W_ extends Q{}class qS extends W_{}class KS extends W_{async _call(e){return new Ae(await super._call(e))}}class V_ extends Q{}class YS extends V_{}class XS extends V_{async _call(e){return new Ae(await super._call(e))}}class et{static MODEL_CLASS_MAPPINGS=null;static BASE_IF_FAIL=!1;static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",model_file_name:o=null,subfolder:l="onnx",device:u=null,dtype:p=null,use_external_data_format:h=null,session_options:m={}}={}){let d={progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,model_file_name:o,subfolder:l,device:u,dtype:p,use_external_data_format:h,session_options:m};if(d.config=await Mg.from_pretrained(e,d),!this.MODEL_CLASS_MAPPINGS)throw new Error("`MODEL_CLASS_MAPPINGS` not implemented for this type of `AutoClass`: "+this.name);for(let _ of this.MODEL_CLASS_MAPPINGS){const w=_.get(d.config.model_type);if(w)return await w[1].from_pretrained(e,d)}if(this.BASE_IF_FAIL)return console.warn(`Unknown model class "${d.config.model_type}", attempting to construct from base class.`),await Q.from_pretrained(e,d);throw Error(`Unsupported model type: ${d.config.model_type}`)}}const QS=new Map([["bert",["BertModel",y1]],["nomic_bert",["NomicBertModel",S1]],["roformer",["RoFormerModel",k1]],["electra",["ElectraModel",R1]],["esm",["EsmModel",s2]],["convbert",["ConvBertModel",I1]],["camembert",["CamembertModel",L1]],["deberta",["DebertaModel",H1]],["deberta-v2",["DebertaV2Model",X1]],["mpnet",["MPNetModel",f2]],["albert",["AlbertModel",x2]],["distilbert",["DistilBertModel",t2]],["roberta",["RobertaModel",G2]],["xlm",["XLMModel",Y2]],["xlm-roberta",["XLMRobertaModel",e$]],["clap",["ClapModel",zS]],["clip",["CLIPModel",c$]],["clipseg",["CLIPSegModel",w$]],["chinese_clip",["ChineseCLIPModel",y$]],["siglip",["SiglipModel",f$]],["mobilebert",["MobileBertModel",d2]],["squeezebert",["SqueezeBertModel",w2]],["wav2vec2",["Wav2Vec2Model",eS]],["wav2vec2-bert",["Wav2Vec2BertModel",cS]],["unispeech",["UniSpeechModel",aS]],["unispeech-sat",["UniSpeechSatModel",oS]],["hubert",["HubertModel",fS]],["wavlm",["WavLMModel",_S]],["audio-spectrogram-transformer",["ASTModel",i$]],["vits",["VitsModel",R_]],["detr",["DetrModel",mx]],["table-transformer",["TableTransformerModel",wx]],["vit",["ViTModel",J$]],["fastvit",["FastViTModel",tx]],["mobilevit",["MobileViTModel",ix]],["mobilevitv2",["MobileViTV2Model",ox]],["owlvit",["OwlViTModel",ux]],["owlv2",["Owlv2Model",cx]],["beit",["BeitModel",hx]],["deit",["DeiTModel",$x]],["convnext",["ConvNextModel",Nx]],["convnextv2",["ConvNextV2Model",Lx]],["dinov2",["Dinov2Model",Wx]],["resnet",["ResNetModel",Sx]],["swin",["SwinModel",Ex]],["swin2sr",["Swin2SRModel",Tx]],["donut-swin",["DonutSwinModel",Dx]],["yolos",["YolosModel",Gx]],["dpt",["DPTModel",Ix]],["glpn",["GLPNModel",Px]],["hifigan",["SpeechT5HifiGan",SS]],["efficientnet",["EfficientNetModel",US]],["mobilenet_v1",["MobileNetV1Model",VS]],["mobilenet_v2",["MobileNetV2Model",HS]],["mobilenet_v3",["MobileNetV3Model",qS]],["mobilenet_v4",["MobileNetV4Model",YS]]]),ZS=new Map([["t5",["T5Model",C2]],["longt5",["LongT5Model",A2]],["mt5",["MT5Model",M2]],["bart",["BartModel",z2]],["mbart",["MBartModel",B2]],["marian",["MarianModel",Xx]],["whisper",["WhisperModel",o$]],["m2m_100",["M2M100Model",Zx]],["blenderbot",["BlenderbotModel",L2]],["blenderbot-small",["BlenderbotSmallModel",W2]]]),JS=new Map([["bloom",["BloomModel",q$]],["gpt2",["GPT2Model",v$]],["gptj",["GPTJModel",C$]],["gpt_bigcode",["GPTBigCodeModel",A$]],["gpt_neo",["GPTNeoModel",x$]],["gpt_neox",["GPTNeoXModel",k$]],["codegen",["CodeGenModel",M$]],["llama",["LlamaModel",z$]],["cohere",["CohereModel",R$]],["gemma",["GemmaModel",D$]],["openelm",["OpenELMModel",F$]],["qwen2",["Qwen2Model",U$]],["phi",["PhiModel",V$]],["phi3",["Phi3Model",H$]],["mpt",["MptModel",Y$]],["opt",["OPTModel",Q$]],["mistral",["MistralModel",CS]],["starcoder2",["Starcoder2Model",AS]],["falcon",["FalconModel",MS]],["stablelm",["StableLmModel",FS]]]),$l=new Map([["speecht5",["SpeechT5ForSpeechToText",$S]],["whisper",["WhisperForConditionalGeneration",l$]]]),G_=new Map([["speecht5",["SpeechT5ForTextToSpeech",xS]]]),H_=new Map([["vits",["VitsModel",R_]],["musicgen",["MusicgenForConditionalGeneration",F_]]]),j_=new Map([["bert",["BertForSequenceClassification",b1]],["roformer",["RoFormerForSequenceClassification",C1]],["electra",["ElectraForSequenceClassification",D1]],["esm",["EsmForSequenceClassification",l2]],["convbert",["ConvBertForSequenceClassification",O1]],["camembert",["CamembertForSequenceClassification",W1]],["deberta",["DebertaForSequenceClassification",q1]],["deberta-v2",["DebertaV2ForSequenceClassification",Z1]],["mpnet",["MPNetForSequenceClassification",g2]],["albert",["AlbertForSequenceClassification",S2]],["distilbert",["DistilBertForSequenceClassification",r2]],["roberta",["RobertaForSequenceClassification",j2]],["xlm",["XLMForSequenceClassification",Q2]],["xlm-roberta",["XLMRobertaForSequenceClassification",r$]],["bart",["BartForSequenceClassification",R2]],["mbart",["MBartForSequenceClassification",N2]],["mobilebert",["MobileBertForSequenceClassification",p2]],["squeezebert",["SqueezeBertForSequenceClassification",v2]]]),q_=new Map([["bert",["BertForTokenClassification",v1]],["roformer",["RoFormerForTokenClassification",T1]],["electra",["ElectraForTokenClassification",N1]],["esm",["EsmForTokenClassification",u2]],["convbert",["ConvBertForTokenClassification",z1]],["camembert",["CamembertForTokenClassification",V1]],["deberta",["DebertaForTokenClassification",K1]],["deberta-v2",["DebertaV2ForTokenClassification",J1]],["mpnet",["MPNetForTokenClassification",_2]],["distilbert",["DistilBertForTokenClassification",n2]],["roberta",["RobertaForTokenClassification",q2]],["xlm",["XLMForTokenClassification",Z2]],["xlm-roberta",["XLMRobertaForTokenClassification",n$]]]),xl=new Map([["t5",["T5ForConditionalGeneration",T2]],["longt5",["LongT5ForConditionalGeneration",I2]],["mt5",["MT5ForConditionalGeneration",O2]],["bart",["BartForConditionalGeneration",P2]],["mbart",["MBartForConditionalGeneration",D2]],["marian",["MarianMTModel",Qx]],["m2m_100",["M2M100ForConditionalGeneration",Jx]],["blenderbot",["BlenderbotForConditionalGeneration",U2]],["blenderbot-small",["BlenderbotSmallForConditionalGeneration",V2]]]),Sl=new Map([["bloom",["BloomForCausalLM",K$]],["gpt2",["GPT2LMHeadModel",$$]],["gptj",["GPTJForCausalLM",T$]],["gpt_bigcode",["GPTBigCodeForCausalLM",I$]],["gpt_neo",["GPTNeoForCausalLM",S$]],["gpt_neox",["GPTNeoXForCausalLM",E$]],["codegen",["CodeGenForCausalLM",O$]],["llama",["LlamaForCausalLM",P$]],["cohere",["CohereForCausalLM",B$]],["gemma",["GemmaForCausalLM",N$]],["openelm",["OpenELMForCausalLM",L$]],["qwen2",["Qwen2ForCausalLM",W$]],["phi",["PhiForCausalLM",G$]],["phi3",["Phi3ForCausalLM",j$]],["mpt",["MptForCausalLM",X$]],["opt",["OPTForCausalLM",Z$]],["mbart",["MBartForCausalLM",F2]],["mistral",["MistralForCausalLM",TS]],["starcoder2",["Starcoder2ForCausalLM",IS]],["falcon",["FalconForCausalLM",OS]],["trocr",["TrOCRForCausalLM",ES]],["stablelm",["StableLmForCausalLM",LS]]]),K_=new Map([["bert",["BertForMaskedLM",w1]],["roformer",["RoFormerForMaskedLM",E1]],["electra",["ElectraForMaskedLM",B1]],["esm",["EsmForMaskedLM",o2]],["convbert",["ConvBertForMaskedLM",M1]],["camembert",["CamembertForMaskedLM",U1]],["deberta",["DebertaForMaskedLM",j1]],["deberta-v2",["DebertaV2ForMaskedLM",Q1]],["mpnet",["MPNetForMaskedLM",m2]],["albert",["AlbertForMaskedLM",E2]],["distilbert",["DistilBertForMaskedLM",i2]],["roberta",["RobertaForMaskedLM",H2]],["xlm",["XLMWithLMHeadModel",X2]],["xlm-roberta",["XLMRobertaForMaskedLM",t$]],["mobilebert",["MobileBertForMaskedLM",c2]],["squeezebert",["SqueezeBertForMaskedLM",b2]]]),Y_=new Map([["bert",["BertForQuestionAnswering",$1]],["roformer",["RoFormerForQuestionAnswering",A1]],["electra",["ElectraForQuestionAnswering",F1]],["convbert",["ConvBertForQuestionAnswering",P1]],["camembert",["CamembertForQuestionAnswering",G1]],["deberta",["DebertaForQuestionAnswering",Y1]],["deberta-v2",["DebertaV2ForQuestionAnswering",e2]],["mpnet",["MPNetForQuestionAnswering",y2]],["albert",["AlbertForQuestionAnswering",k2]],["distilbert",["DistilBertForQuestionAnswering",a2]],["roberta",["RobertaForQuestionAnswering",K2]],["xlm",["XLMForQuestionAnswering",J2]],["xlm-roberta",["XLMRobertaForQuestionAnswering",a$]],["mobilebert",["MobileBertForQuestionAnswering",h2]],["squeezebert",["SqueezeBertForQuestionAnswering",$2]]]),kl=new Map([["vision-encoder-decoder",["VisionEncoderDecoderModel",Gg]]]),e3=new Map([["llava",["LlavaForConditionalGeneration",Hg]],["moondream1",["Moondream1ForConditionalGeneration",d$]]]),t3=new Map([["vision-encoder-decoder",["VisionEncoderDecoderModel",Gg]]]),X_=new Map([["vit",["ViTForImageClassification",ex]],["fastvit",["FastViTForImageClassification",rx]],["mobilevit",["MobileViTForImageClassification",sx]],["mobilevitv2",["MobileViTV2ForImageClassification",lx]],["beit",["BeitForImageClassification",fx]],["deit",["DeiTForImageClassification",xx]],["convnext",["ConvNextForImageClassification",Fx]],["convnextv2",["ConvNextV2ForImageClassification",Ux]],["dinov2",["Dinov2ForImageClassification",Vx]],["resnet",["ResNetForImageClassification",kx]],["swin",["SwinForImageClassification",Cx]],["segformer",["SegformerForImageClassification",DS]],["efficientnet",["EfficientNetForImageClassification",WS]],["mobilenet_v1",["MobileNetV1ForImageClassification",GS]],["mobilenet_v2",["MobileNetV2ForImageClassification",jS]],["mobilenet_v3",["MobileNetV3ForImageClassification",KS]],["mobilenet_v4",["MobileNetV4ForImageClassification",XS]]]),Q_=new Map([["detr",["DetrForObjectDetection",gx]],["table-transformer",["TableTransformerForObjectDetection",bx]],["yolos",["YolosForObjectDetection",Hx]]]),Z_=new Map([["owlvit",["OwlViTForObjectDetection",dx]],["owlv2",["Owlv2ForObjectDetection",px]]]),J_=new Map([["detr",["DetrForSegmentation",_x]],["clipseg",["CLIPSegForImageSegmentation",b$]]]),e0=new Map([["segformer",["SegformerForSemanticSegmentation",NS]]]),r3=new Map([["sam",["SamModel",Kx]]]),t0=new Map([["wav2vec2",["Wav2Vec2ForCTC",tS]],["wav2vec2-bert",["Wav2Vec2BertForCTC",pS]],["unispeech",["UniSpeechForCTC",iS]],["unispeech-sat",["UniSpeechSatForCTC",lS]],["wavlm",["WavLMForCTC",yS]],["hubert",["HubertForCTC",mS]]]),r0=new Map([["wav2vec2",["Wav2Vec2ForSequenceClassification",rS]],["wav2vec2-bert",["Wav2Vec2BertForSequenceClassification",hS]],["unispeech",["UniSpeechForSequenceClassification",sS]],["unispeech-sat",["UniSpeechSatForSequenceClassification",uS]],["wavlm",["WavLMForSequenceClassification",wS]],["hubert",["HubertForSequenceClassification",gS]],["audio-spectrogram-transformer",["ASTForAudioClassification",s$]]]),n3=new Map([["wavlm",["WavLMForXVector",bS]]]),a3=new Map([["unispeech-sat",["UniSpeechSatForAudioFrameClassification",dS]],["wavlm",["WavLMForAudioFrameClassification",vS]],["wav2vec2",["Wav2Vec2ForAudioFrameClassification",nS]]]),i3=new Map([["vitmatte",["VitMatteForImageMatting",ax]]]),n0=new Map([["swin2sr",["Swin2SRForImageSuperResolution",Ax]]]),a0=new Map([["dpt",["DPTForDepthEstimation",Mx]],["depth_anything",["DepthAnythingForDepthEstimation",zx]],["glpn",["GLPNForDepthEstimation",Rx]]]),i0=new Map([["clip",["CLIPVisionModelWithProjection",h$]],["siglip",["SiglipVisionModel",g$]]]),s0=[[QS,ve.EncoderOnly],[ZS,ve.EncoderDecoder],[JS,ve.DecoderOnly],[j_,ve.EncoderOnly],[q_,ve.EncoderOnly],[xl,ve.Seq2Seq],[$l,ve.Seq2Seq],[Sl,ve.DecoderOnly],[K_,ve.EncoderOnly],[Y_,ve.EncoderOnly],[kl,ve.Vision2Seq],[e3,ve.ImageTextToText],[X_,ve.EncoderOnly],[J_,ve.EncoderOnly],[e0,ve.EncoderOnly],[i3,ve.EncoderOnly],[n0,ve.EncoderOnly],[a0,ve.EncoderOnly],[Q_,ve.EncoderOnly],[Z_,ve.EncoderOnly],[r3,ve.MaskGeneration],[t0,ve.EncoderOnly],[r0,ve.EncoderOnly],[G_,ve.Seq2Seq],[H_,ve.EncoderOnly],[n3,ve.EncoderOnly],[a3,ve.EncoderOnly],[i0,ve.EncoderOnly]];for(const[t,e]of s0)for(const[r,n]of t.values())es.set(r,e),ba.set(n,r),zg.set(r,n);const s3=[["MusicgenForConditionalGeneration",F_,ve.Musicgen],["CLIPTextModelWithProjection",p$,ve.EncoderOnly],["SiglipTextModel",m$,ve.EncoderOnly],["ClapTextModelWithProjection",PS,ve.EncoderOnly],["ClapAudioModelWithProjection",RS,ve.EncoderOnly]];for(const[t,e,r]of s3)es.set(t,r),ba.set(e,t),zg.set(t,e);class ma extends et{static MODEL_CLASS_MAPPINGS=s0.map(e=>e[0]);static BASE_IF_FAIL=!0}class Ep extends et{static MODEL_CLASS_MAPPINGS=[j_]}class o3 extends et{static MODEL_CLASS_MAPPINGS=[q_]}class ho extends et{static MODEL_CLASS_MAPPINGS=[xl]}class l3 extends et{static MODEL_CLASS_MAPPINGS=[$l]}class u3 extends et{static MODEL_CLASS_MAPPINGS=[G_]}class d3 extends et{static MODEL_CLASS_MAPPINGS=[H_]}class c3 extends et{static MODEL_CLASS_MAPPINGS=[Sl]}class p3 extends et{static MODEL_CLASS_MAPPINGS=[K_]}class h3 extends et{static MODEL_CLASS_MAPPINGS=[Y_]}class f3 extends et{static MODEL_CLASS_MAPPINGS=[kl]}class m3 extends et{static MODEL_CLASS_MAPPINGS=[X_]}class g3 extends et{static MODEL_CLASS_MAPPINGS=[J_]}class _3 extends et{static MODEL_CLASS_MAPPINGS=[e0]}class y3 extends et{static MODEL_CLASS_MAPPINGS=[Q_]}class w3 extends et{static MODEL_CLASS_MAPPINGS=[Z_]}class b3 extends et{static MODEL_CLASS_MAPPINGS=[t0]}class v3 extends et{static MODEL_CLASS_MAPPINGS=[r0]}class $3 extends et{static MODEL_CLASS_MAPPINGS=[t3]}class x3 extends et{static MODEL_CLASS_MAPPINGS=[n0]}class S3 extends et{static MODEL_CLASS_MAPPINGS=[a0]}class k3 extends et{static MODEL_CLASS_MAPPINGS=[i0]}class Ae extends Yt{constructor({logits:e}){super(),this.logits=e}}class E3 extends Yt{constructor({logits:e,embeddings:r}){super(),this.logits=e,this.embeddings=r}}class vt extends Yt{constructor({logits:e}){super(),this.logits=e}}class $t extends Yt{constructor({logits:e}){super(),this.logits=e}}class Ct extends Yt{constructor({start_logits:e,end_logits:r}){super(),this.start_logits=e,this.end_logits=r}}class Ln extends Yt{constructor({logits:e}){super(),this.logits=e}}class C3 extends Yt{constructor({alphas:e}){super(),this.alphas=e}}class T3 extends Yt{constructor({waveform:e,spectrogram:r}){super(),this.waveform=e,this.spectrogram=r}}const Gt=typeof self<"u",A3=Gt&&self.constructor.name==="DedicatedWorkerGlobalScope";let Xr,o0,Rr;if(Gt)Xr=(t,e)=>{if(!self.OffscreenCanvas)throw new Error("OffscreenCanvas not supported by this browser.");return new self.OffscreenCanvas(t,e)},Rr=self.createImageBitmap,o0=self.ImageData;else if(Je)Rr=async t=>{const r=(await t.metadata()).channels,{data:n,info:a}=await t.rotate().raw().toBuffer({resolveWithObject:!0}),i=new yt(new Uint8ClampedArray(n),a.width,a.height,a.channels);return r!==void 0&&r!==a.channels&&i.convert(r),i};else throw new Error("Unable to load image processing library.");const I3={0:"nearest",1:"lanczos",2:"bilinear",3:"bicubic",4:"box",5:"hamming"},M3=new Map([["png","image/png"],["jpg","image/jpeg"],["jpeg","image/jpeg"],["gif","image/gif"]]);class yt{constructor(e,r,n,a){this.data=e,this.width=r,this.height=n,this.channels=a}get size(){return[this.width,this.height]}static async read(e){if(e instanceof yt)return e;if(typeof e=="string"||e instanceof URL)return await this.fromURL(e);throw new Error(`Unsupported input type: ${typeof e}`)}static fromCanvas(e){if(!Gt)throw new Error("fromCanvas() is only supported in browser environments.");const n=e.getContext("2d").getImageData(0,0,e.width,e.height).data;return new yt(n,e.width,e.height,4)}static async fromURL(e){const r=await Ui(e);if(r.status!==200)throw new Error(`Unable to read image from "${e}" (${r.status} ${r.statusText})`);const n=await r.blob();return this.fromBlob(n)}static async fromBlob(e){if(Gt){const r=await Rr(e),n=Xr(r.width,r.height).getContext("2d");return n.drawImage(r,0,0),new this(n.getImageData(0,0,r.width,r.height).data,r.width,r.height,4)}else{const r=Je(await e.arrayBuffer());return await Rr(r)}}static fromTensor(e,r="CHW"){if(e.dims.length!==3)throw new Error(`Tensor should have 3 dimensions, but has ${e.dims.length} dimensions.`);if(r==="CHW")e=e.transpose(1,2,0);else if(r!=="HWC")throw new Error(`Unsupported channel format: ${r}`);if(!(e.data instanceof Uint8ClampedArray||e.data instanceof Uint8Array))throw new Error(`Unsupported tensor type: ${e.type}`);switch(e.dims[2]){case 1:case 2:case 3:case 4:return new yt(e.data,e.dims[1],e.dims[0],e.dims[2]);default:throw new Error(`Unsupported number of channels: ${e.dims[2]}`)}}grayscale(){if(this.channels===1)return this;const e=new Uint8ClampedArray(this.width*this.height*1);switch(this.channels){case 3:case 4:for(let r=0,n=0;r=0?l=n:p=-n,a>=0?u=a:h=-a,o.drawImage(s,l,u,e,r,p,h,e,r),new yt(o.getImageData(0,0,e,r).data,e,r,4).convert(i)}else{let i=this.toSharp();if(n>=0&&a>=0)i=i.extract({left:Math.floor(n),top:Math.floor(a),width:e,height:r});else if(n<=0&&a<=0){const s=Math.floor(-a),o=Math.floor(-n);i=i.extend({top:s,left:o,right:e-this.width-o,bottom:r-this.height-s})}else{let s=[0,0],o=0;a<0?(s[0]=Math.floor(-a),s[1]=r-this.height-s[0]):o=Math.floor(a);let l=[0,0],u=0;n<0?(l[0]=Math.floor(-n),l[1]=e-this.width-l[0]):u=Math.floor(n),i=i.extend({top:s[0],bottom:s[1],left:l[0],right:l[1]}).extract({left:u,top:o,width:e,height:r})}return await Rr(i)}}async toBlob(e="image/png",r=1){if(!Gt)throw new Error("toBlob() is only supported in browser environments.");return await this.toCanvas().convertToBlob({type:e,quality:r})}toTensor(e="CHW"){let r=new ue("uint8",new Uint8Array(this.data),[this.height,this.width,this.channels]);if(e!=="HWC")if(e==="CHW")r=r.permute(2,0,1);else throw new Error(`Unsupported channel format: ${e}`);return r}toCanvas(){if(!Gt)throw new Error("toCanvas() is only supported in browser environments.");const e=this.clone().rgba(),r=Xr(e.width,e.height),n=new o0(e.data,e.width,e.height);return r.getContext("2d").putImageData(n,0,0),r}_update(e,r,n,a=null){return this.data=e,this.width=r,this.height=n,a!==null&&(this.channels=a),this}clone(){return new yt(this.data.slice(),this.width,this.height,this.channels)}convert(e){if(this.channels===e)return this;switch(e){case 1:this.grayscale();break;case 3:this.rgb();break;case 4:this.rgba();break;default:throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`)}return this}async save(e){if(Gt){if(A3)throw new Error("Unable to save an image from a Web Worker.");const r=e.split(".").pop().toLowerCase(),n=M3.get(r)??"image/png",a=await this.toBlob(n),i=URL.createObjectURL(a),s=document.createElement("a");s.href=i,s.download=e,s.click(),s.remove()}else{if(zt.useFS)return await this.toSharp().toFile(e);throw new Error("Unable to save the image because filesystem is disabled in this environment.")}}toSharp(){if(Gt)throw new Error("toSharp() is only supported in server-side environments.");return Je(this.data,{raw:{width:this.width,height:this.height,channels:this.channels}})}}async function O3(t,e){if(typeof AudioContext>"u")throw Error("Unable to load audio from path/URL since `AudioContext` is not available in your environment. Instead, audio data should be passed directly to the pipeline/processor. For more information and some example code, see https://huggingface.co/docs/transformers.js/guides/node-audio-processing.");const r=await(await Ui(t)).arrayBuffer(),n=new AudioContext({sampleRate:e});typeof e>"u"&&console.warn(`No sampling rate provided, using default of ${n.sampleRate}Hz.`);const a=await n.decodeAudioData(r);let i;if(a.numberOfChannels===2){const s=Math.sqrt(2),o=a.getChannelData(0),l=a.getChannelData(1);i=new Float32Array(o.length);for(let u=0;u2595*Math.log10(1+t/700),kaldi:t=>1127*Math.log(1+t/700),slaney:(t,e=1e3,r=15,n=27/Math.log(6.4))=>t>=e?r+Math.log(t/e)*n:3*t/200};function fo(t,e="htk"){const r=z3[e];if(!r)throw new Error('mel_scale should be one of "htk", "slaney" or "kaldi".');return typeof t=="number"?r(t):t.map(n=>r(n))}const P3={htk:t=>700*(10**(t/2595)-1),kaldi:t=>700*(Math.exp(t/1127)-1),slaney:(t,e=1e3,r=15,n=Math.log(6.4)/27)=>t>=r?e*Math.exp(n*(t-r)):200*t/3};function R3(t,e="htk"){const r=P3[e];if(!r)throw new Error('mel_scale should be one of "htk", "slaney" or "kaldi".');return typeof t=="number"?r(t):t.map(n=>r(n))}function B3(t,e){const r=Float64Array.from({length:e.length-1},(s,o)=>e[o+1]-e[o]),n=Array.from({length:t.length},()=>new Array(e.length));for(let s=0;snew Array(t.length));for(let s=0;st+n*i)}function Ta(t,e,r,n,a,i=null,s="htk",o=!1){if(i!==null&&i!=="slaney")throw new Error('norm must be one of null or "slaney"');const l=fo(r,s),u=fo(n,s),p=Tp(l,u,e+2);let h=R3(p,s),m;if(o){const _=a/(t*2);m=fo(Float64Array.from({length:t},(w,v)=>v*_),s),h=p}else m=Tp(0,Math.floor(a/2),t);const d=B3(m,h);if(i!==null&&i==="slaney")for(let _=0;_a)throw Error(`frame_length (${r}) may not be larger than fft_length (${a})`);if(T!==r)throw new Error(`Length of the window (${T}) must equal frame_length (${r})`);if(n<=0)throw new Error("hop_length must be greater than zero");if(i===null&&p!==null)throw new Error("You have provided `mel_filters` but `power` is `None`. Mel spectrogram computation is not yet supported for complex-valued spectrogram. Specify `power` to fix this issue.");if(s){if(o!=="reflect")throw new Error(`pad_mode="${o}" not implemented yet.`);const R=Math.floor((a-1)/2)+1;t=D3(t,R,R)}const A=Math.floor(1+Math.floor((t.length-r)/n)),P=l?Math.floor(a/2)+1:a;let B=A,D=A;S!==null&&(S>A?$&&(D=S):D=B=S);const q=new V0(a),H=new Float64Array(a),ie=new Float64Array(q.outputBufferSize),te=new Float32Array(P*D);for(let R=0;R=1;--re)H[re]-=u*H[re-1];H[0]*=1-u}for(let re=0;reMath.pow(o,.85));break;default:throw new Error(`Unknown window type ${e}.`)}if(r&&(s=s.subarray(0,t)),n===null)return s;if(t>n)throw new Error(`Length of the window (${t}) may not be larger than frame_length (${n})`);return s}function L3([t,e,r,n]){return[t-r/2,e-n/2,t+r/2,e+n/2]}function El(t,e=.5,r=null,n=!1){const a=t.logits,i=t.pred_boxes,[s,o,l]=a.dims;if(r!==null&&r.length!==s)throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");let u=[];for(let p=0;pe&&S.push(E)}else{let E=Kt(v.data)[1];if(E===l-1||($=wt(v.data),$[E]A*h[(P+1)%2])),m.boxes.push(T),m.classes.push(E),m.scores.push($[E])}}u.push(m)}return u}function qa(t,e){if(!(t instanceof Float32Array||t instanceof Float64Array))throw new Error(`${e} expects input to be a Float32Array or a Float64Array, but got ${t?.constructor?.name??typeof t} instead. If using the feature extractor directly, remember to use \`read_audio(url, sampling_rate)\` to obtain the raw audio data of the file/url.`)}function Ap(t,e,r=0,n=null){const a=t/e;let i=H0(a)*e;return n!==null&&i>n&&(i=Math.floor(a)*e),ii?u=Math.floor(i*l/a):i>a&&(l=Math.floor(a*u/i)),await e.resize(u,l,{resample:n}))}async crop_margin(e,r=200){const n=e.clone().grayscale(),a=Fp(n.data)[0],s=Kt(n.data)[0]-a;if(s===0)return e;const o=r/255;let l=n.width,u=n.height,p=0,h=0;const m=n.data;for(let d=0;dthis.preprocess(i)));return{pixel_values:ka(n.map(i=>i.pixel_values),0),original_sizes:n.map(i=>i.original_size),reshaped_input_sizes:n.map(i=>i.reshaped_input_size)}}}class U3 extends He{post_process_semantic_segmentation(e,r=null){const n=e.logits,a=n.dims[0];if(r!==null&&r.length!==a)throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");const i=[];for(let s=0;sm[E]&&(m[E]=$[E],d[E]=S)}const _=new Array(l.dims[0]),w=h.data;for(let S=0;SS!==void 0);i.push({segmentation:h,labels:v})}return i}}class u0 extends He{}class W3 extends u0{}class V3 extends He{}class G3 extends He{}class d0 extends He{}class H3 extends d0{}class j3 extends He{}class q3 extends He{}class c0 extends He{constructor(e){super(e),this.crop_pct=this.config.crop_pct??224/256}async resize(e){const r=this.size?.shortest_edge;if(r===void 0)throw new Error("Size dictionary must contain 'shortest_edge' key.");if(r<384){const n=Math.floor(r/this.crop_pct),[a,i]=this.get_resize_output_image_size(e,{shortest_edge:n});e=await e.resize(a,i,{resample:this.resample}),e=await e.center_crop(r,r)}else e=await e.resize(r,r,{resample:this.resample});return e}}class K3 extends c0{}class Y3 extends He{}class X3 extends He{}class Q3 extends He{constructor(e){super(e),this.include_top=this.config.include_top??!0,this.include_top&&(this.image_std=this.image_std.map(r=>r*r))}}class Z3 extends He{}class J3 extends He{}class ek extends He{}class tk extends He{}class p0 extends He{}class rk extends p0{}class h0 extends He{post_process_object_detection(...e){return El(...e)}}class nk extends h0{}class ak extends He{}class ik extends He{}class f0 extends He{pad_image(e,r,n,a={}){const[i,s,o]=r;let l=this.image_mean;Array.isArray(this.image_mean)||(l=new Array(o).fill(l));let u=this.image_std;Array.isArray(u)||(u=new Array(o).fill(l));const p=l.map((h,m)=>-h/u[m]);return super.pad_image(e,r,n,{center:!0,constant_values:p,...a})}}class sk extends f0{}class ok extends He{async _call(e){const r=await super._call(e),n=[r.pixel_values.dims[0],64,64],a=new ue("int64",new BigInt64Array(n.reduce((i,s)=>i*s)).fill(1n),n);return{...r,pixel_mask:a}}post_process_object_detection(...e){return El(...e)}remove_low_and_no_objects(e,r,n,a){let i=[],s=[],o=[];for(let l=0;ln&&(i.push(p),s.push(d),o.push(h))}return[i,s,o]}check_segment_validity(e,r,n,a=.5,i=.8){let s=[],o=0,l=0;const u=r[n].data;for(let h=0;h=a&&++l;let p=o>0&&l>0;return p&&(p=o/l>i),[p,s]}compute_segments(e,r,n,a,i,s=null,o=null){let[l,u]=o??e[0].dims,p=new ue("int32",new Int32Array(l*u),[l,u]),h=[];if(o!==null)for(let v=0;vd[E]&&(m[E]=v,d[E]=$[E])}let _=0;const w=p.data;for(let v=0;va!==r.dims[i]))throw Error(`The first ${n.length} dimensions of 'input_points' and 'input_labels' must be the same.`);return new ue("int64",e.flat(1/0).map(BigInt),n)}async _call(e,{input_points:r=null,input_labels:n=null,input_boxes:a=null}={}){const i=await super._call(e);if(r&&(i.input_points=this.reshape_input_points(r,i.original_sizes,i.reshaped_input_sizes)),n){if(!i.input_points)throw Error("`input_points` must be provided if `input_labels` are provided.");i.input_labels=this.add_input_labels(n,i.input_points)}return a&&(i.input_boxes=this.reshape_input_points(a,i.original_sizes,i.reshaped_input_sizes,!0)),i}async post_process_masks(e,r,n,{mask_threshold:a=0,binarize:i=!0,pad_size:s=null}={}){const o=[];s=s??this.pad_size;const l=[s.height,s.width];for(let u=0;ua&&(_[w]=1);m=new ue("bool",_,m.dims)}o.push(m)}return o}generate_crop_boxes(e,r,{crop_n_layers:n=0,overlap_ratio:a=512/1500,points_per_crop:i=32,crop_n_points_downscale_factor:s=1}={}){}}class dk extends He{pad_image(e,r,n,a={}){const[i,s,o]=r;return super.pad_image(e,r,{width:s+(n-s%n)%n,height:i+(n-i%n)%n},{mode:"symmetric",center:!1,constant_values:-1,...a})}}class ck extends He{async _call(e,r){Array.isArray(e)||(e=[e]),Array.isArray(r)||(r=[r]);const n=await Promise.all(e.map(s=>this.preprocess(s))),a=await Promise.all(r.map(s=>this.preprocess(s,{do_normalize:!1,do_convert_rgb:!1,do_convert_grayscale:!0})));return{pixel_values:ka(n.map((s,o)=>gr([s.pixel_values,a[o].pixel_values],0)),0),original_sizes:n.map(s=>s.original_size),reshaped_input_sizes:n.map(s=>s.reshaped_input_size)}}}class pk extends pn{constructor(e){super(e),this.config.mel_filters??=Ta(Math.floor(1+this.config.n_fft/2),this.config.feature_size,0,8e3,this.config.sampling_rate,"slaney","slaney"),this.window=ms(this.config.n_fft,"hann")}async _extract_fbank_features(e){const r=await fs(e,this.window,this.config.n_fft,this.config.hop_length,{power:2,mel_filters:this.config.mel_filters,log_mel:"log10",max_num_frames:this.config.nb_max_frames}),n=r.data,a=Kt(n)[0];for(let i=0;ithis.config.n_samples?(console.warn("Attempting to extract features for audio longer than 30 seconds. If using a pipeline to extract transcript from a long audio clip, remember to specify `chunk_length_s` and/or `stride_length_s`."),r=e.slice(0,this.config.n_samples)):(r=new Float32Array(this.config.n_samples),r.set(e)),{input_features:(await this._extract_fbank_features(r)).unsqueeze_(0)}}}class hk extends pn{_zero_mean_unit_var_norm(e){const n=e.reduce((i,s)=>i+s,0)/e.length,a=e.reduce((i,s)=>i+(s-n)**2,0)/e.length;return e.map(i=>(i-n)/Math.sqrt(a+1e-7))}async _call(e){qa(e,"Wav2Vec2FeatureExtractor"),e instanceof Float64Array&&(e=new Float32Array(e));let r=e;this.config.do_normalize&&(r=this._zero_mean_unit_var_norm(r));const n=[1,r.length];return{input_values:new ue("float32",r,n),attention_mask:new ue("int64",new BigInt64Array(r.length).fill(1n),n)}}}class fk extends pn{constructor(e){super(e);const r=this.config.sampling_rate,n=Ta(256,this.config.num_mel_bins,20,Math.floor(r/2),r,null,"kaldi",!0);for(let a=0;an*32768),fs(e,this.window,400,160,{fft_length:512,power:2,center:!1,preemphasis:.97,mel_filters:this.mel_filters,log_mel:"log",mel_floor:1192092955078125e-22,remove_dc_offset:!0,max_num_frames:r,transpose:!0})}async _call(e,{padding:r=!0,pad_to_multiple_of:n=2,do_normalize_per_mel_bins:a=!0,return_attention_mask:i=!0}={}){qa(e,"SeamlessM4TFeatureExtractor");let s=await this._extract_fbank_features(e,this.config.max_length);if(a){const[_,w]=s.dims,v=s.data;for(let S=0;S0){const $=new Float32Array(w*(_+S));$.set(v),$.fill(this.config.padding_value,v.length);const E=_+S;s=new ue(s.type,$,[E,w]),i&&(o=new ue("int64",new BigInt64Array(E),[1,E]),o.data.fill(1n,0,_))}}const[l,u]=s.dims,p=this.config.stride;if(l%p!==0)throw new Error(`The number of frames (${l}) must be a multiple of the stride (${p}).`);const m=s.view(1,Math.floor(l/p),u*p),d={input_features:m};if(i){const _=m.dims[1],w=new BigInt64Array(_);if(o){const v=o.data;for(let S=1,$=0;S0)if(n==="rand_trunc"){const o=Math.floor(Math.random()*(s+1));e=e.subarray(o,o+r),i=await this._extract_fbank_features(e,this.mel_filters_slaney,this.config.nb_max_samples)}else throw new Error(`Truncation strategy "${n}" not implemented`);else{if(s<0){let o=new Float64Array(r);if(o.set(e),a==="repeat")for(let l=e.length;lyt.read(e)))}async function ts(t,e){return Array.isArray(t)||(t=[t]),await Promise.all(t.map(r=>typeof r=="string"||r instanceof URL?O3(r,e):r instanceof Float64Array?new Float32Array(r):r))}function m0(t,e){e&&(t=t.map(s=>s|0));const[r,n,a,i]=t;return{xmin:r,ymin:n,xmax:a,ymax:i}}class tt extends bt{constructor({task:e,model:r,tokenizer:n=null,processor:a=null}){super(),this.task=e,this.model=r,this.tokenizer=n,this.processor=a}async dispose(){await this.model.dispose()}}class xk extends tt{constructor(e){super(e)}async _call(e,{top_k:r=1}={}){const n=this.tokenizer(e,{padding:!0,truncation:!0}),a=await this.model(n),i=this.model.config.problem_type==="multi_label_classification"?l=>l.sigmoid():l=>new ue("float32",wt(l.data),l.dims),s=this.model.config.id2label,o=[];for(const l of a.logits){const u=i(l),p=await Nn(u,r),h=p[0].tolist(),d=p[1].tolist().map((_,w)=>({label:s?s[_]:`LABEL_${_}`,score:h[w]}));r===1?o.push(...d):o.push(d)}return Array.isArray(e)||r===1?o:o[0]}}class Sk extends tt{constructor(e){super(e)}async _call(e,{ignore_labels:r=["O"]}={}){const n=Array.isArray(e),a=this.tokenizer(n?e:[e],{padding:!0,truncation:!0}),s=(await this.model(a)).logits,o=this.model.config.id2label,l=[];for(let u=0;uE==this.tokenizer.sep_token_id);l[h].map((E,T)=>E==1&&(T===0||T>d&&u.findIndex(A=>A==m[T])===-1));const _=i[h].tolist(),w=s[h].tolist();for(let E=1;E<_.length;++E)(l[h]==0||E<=d||u.findIndex(T=>T==m[E])!==-1)&&(_[E]=-1/0,w[E]=-1/0);const v=wt(_).map((E,T)=>[E,T]),S=wt(w).map((E,T)=>[E,T]);v[0][0]=0,S[0][0]=0;const $=P0(v,S).filter(E=>E[0][1]<=E[1][1]).map(E=>[E[0][1],E[1][1],E[0][0]*E[1][0]]).sort((E,T)=>T[2]-E[2]);for(let E=0;E_==this.tokenizer.mask_token_id);if(u===-1)throw Error(`Mask token (${this.tokenizer.mask_token}) not found in text.`);const p=a[o][u],h=await Nn(new ue("float32",wt(p.data),p.dims),r),m=h[0].tolist(),d=h[1].tolist();i.push(d.map((_,w)=>{const v=l.slice();return v[u]=_,{score:m[w],token:Number(_),token_str:this.tokenizer.model.vocab[_],sequence:this.tokenizer.decode(v,{skip_special_tokens:!0})}}))}return Array.isArray(e)?i:i[0]}}class Cl extends tt{_key="generated_text";constructor(e){super(e)}async _call(e,r={}){Array.isArray(e)||(e=[e]),this.model.config.prefix&&(e=e.map(l=>this.model.config.prefix+l));const n=this.model.config.task_specific_params;n&&n[this.task]&&n[this.task].prefix&&(e=e.map(l=>n[this.task].prefix+l));const a=this.tokenizer,i={padding:!0,truncation:!0};let s;this instanceof g0&&"_build_translation_inputs"in a?s=a._build_translation_inputs(e,i,r):s=a(e,i);const o=await this.model.generate({...s,...r});return a.batch_decode(o,{skip_special_tokens:!0}).map(l=>({[this._key]:l}))}}class Ck extends Cl{_key="summary_text";constructor(e){super(e)}}class g0 extends Cl{_key="translation_text";constructor(e){super(e)}}class Tk extends tt{constructor(e){super(e)}async _call(e,r={}){throw new Error("This pipeline is not yet supported in Transformers.js v3.")}}class Ak extends tt{constructor(e){super(e),this.label2id=Object.fromEntries(Object.entries(this.model.config.label2id).map(([r,n])=>[r.toLowerCase(),n])),this.entailment_id=this.label2id.entailment,this.entailment_id===void 0&&(console.warn("Could not find 'entailment' in label2id mapping. Using 2 as entailment_id."),this.entailment_id=2),this.contradiction_id=this.label2id.contradiction??this.label2id.not_entailment,this.contradiction_id===void 0&&(console.warn("Could not find 'contradiction' in label2id mapping. Using 0 as contradiction_id."),this.contradiction_id=0)}async _call(e,r,{hypothesis_template:n="This example is {}.",multi_label:a=!1}={}){const i=Array.isArray(e);i||(e=[e]),Array.isArray(r)||(r=[r]);const s=r.map(u=>n.replace("{}",u)),o=a||r.length===1,l=[];for(const u of e){const p=[];for(const d of s){const _=this.tokenizer(u,{text_pair:d,padding:!0,truncation:!0}),w=await this.model(_);o?p.push([w.logits.data[this.contradiction_id],w.logits.data[this.entailment_id]]):p.push(w.logits.data[this.entailment_id])}const m=(o?p.map(d=>wt(d)[1]):wt(p)).map((d,_)=>[d,_]).sort((d,_)=>_[0]-d[0]);l.push({sequence:u,labels:m.map(d=>r[d[1]]),scores:m.map(d=>d[0])})}return i?l:l[0]}}class Ik extends tt{constructor(e){super(e)}async _call(e,{pooling:r="none",normalize:n=!1,quantize:a=!1,precision:i="binary"}={}){const s=this.tokenizer(e,{padding:!0,truncation:!0}),o=await this.model(s);let l=o.last_hidden_state??o.logits??o.token_embeddings;if(r!=="none")if(r==="mean")l=Aw(l,s.attention_mask);else if(r==="cls")l=l.slice(null,0);else throw Error(`Pooling method '${r}' not supported.`);return n&&(l=l.normalize(2,-1)),a&&(l=Nw(l,i)),l}}class Mk extends tt{constructor(e){super(e)}async _call(e,{pool:r=null}={}){const n=await Tr(e),{pixel_values:a}=await this.processor(n),i=await this.model({pixel_values:a});let s;if(r){if(!("pooler_output"in i))throw Error("No pooled output was returned. Make sure the model has a 'pooler' layer when using the 'pool' option.");s=i.pooler_output}else s=i.last_hidden_state??i.logits??i.image_embeds;return s}}class Ok extends tt{constructor(e){super(e)}async _call(e,{top_k:r=5}={}){const n=this.processor.feature_extractor.config.sampling_rate,a=await ts(e,n),i=this.model.config.id2label,s=[];for(const o of a){const l=await this.processor(o),p=(await this.model(l)).logits[0],h=await Nn(new ue("float32",wt(p.data),p.dims),r),m=h[0].tolist(),_=h[1].tolist().map((w,v)=>({label:i?i[w]:`LABEL_${w}`,score:m[v]}));s.push(_)}return Array.isArray(e)?s:s[0]}}class zk extends tt{constructor(e){super(e)}async _call(e,r,{hypothesis_template:n="This is a sound of {}."}={}){const a=!Array.isArray(e);a&&(e=[e]);const i=r.map(p=>n.replace("{}",p)),s=this.tokenizer(i,{padding:!0,truncation:!0}),o=this.processor.feature_extractor.config.sampling_rate,l=await ts(e,o),u=[];for(const p of l){const h=await this.processor(p),m=await this.model({...s,...h}),d=wt(m.logits_per_audio.data);u.push([...d].map((_,w)=>({score:_,label:r[w]})))}return a?u[0]:u}}class Pk extends tt{constructor(e){super(e)}async _call(e,r={}){switch(this.model.config.model_type){case"whisper":return this._call_whisper(e,r);case"wav2vec2":case"wav2vec2-bert":case"unispeech":case"unispeech-sat":case"hubert":return this._call_wav2vec2(e,r);default:throw new Error(`AutomaticSpeechRecognitionPipeline does not support model type '${this.model.config.model_type}'.`)}}async _call_wav2vec2(e,r){r.language&&console.warn('`language` parameter is not yet supported for `wav2vec2` models, defaulting to "English".'),r.task&&console.warn('`task` parameter is not yet supported for `wav2vec2` models, defaulting to "transcribe".');const n=!Array.isArray(e);n&&(e=[e]);const a=this.processor.feature_extractor.config.sampling_rate,i=await ts(e,a),s=[];for(const o of i){const l=await this.processor(o),p=(await this.model(l)).logits[0],h=[];for(const d of p)h.push(Kt(d.data)[1]);const m=this.tokenizer.decode(h);s.push({text:m})}return n?s[0]:s}async _call_whisper(e,r){const n=r.return_timestamps??!1,a=r.chunk_length_s??0,i=r.force_full_sequences??!1;let s=r.stride_length_s??null;n==="word"&&(r.return_token_timestamps=!0);const o=!Array.isArray(e);o&&(e=[e]);const l=this.processor.feature_extractor.config.chunk_length/this.model.config.max_source_positions,u=this.processor.feature_extractor.config.hop_length,p=this.processor.feature_extractor.config.sampling_rate,h=await ts(e,p),m=[];for(const d of h){let _=[];if(a>0){if(s===null)s=a/6;else if(a<=s)throw Error("`chunk_length_s` must be larger than `stride_length_s`.");const S=p*a,$=p*s,E=S-2*$;let T=0;for(;T=d.length;_.push({stride:[A.length,B?0:$,D?0:$],input_features:P.input_features,is_last:D}),T+=E}}else _=[{stride:[d.length,0,0],input_features:(await this.processor(d)).input_features,is_last:!0}];for(const S of _){r.num_frames=Math.floor(S.stride[0]/u);const $=await this.model.generate({inputs:S.input_features,...r});n==="word"?(S.tokens=$.sequences[0].tolist(),S.token_timestamps=$.token_timestamps.tolist()[0].map(E=>Mi(E,2))):S.tokens=$[0].tolist(),S.stride=S.stride.map(E=>E/p)}const[w,v]=this.tokenizer._decode_asr(_,{time_precision:l,return_timestamps:n,force_full_sequences:i});m.push({text:w,...v})}return o?m[0]:m}}class Rk extends tt{constructor(e){super(e)}async _call(e,r={}){const n=Array.isArray(e),a=await Tr(e),{pixel_values:i}=await this.processor(a),s=[];for(const o of i){o.dims=[1,...o.dims];const l=await this.model.generate({inputs:o,...r}),u=this.tokenizer.batch_decode(l,{skip_special_tokens:!0}).map(p=>({generated_text:p.trim()}));s.push(u)}return n?s:s[0]}}class Bk extends tt{constructor(e){super(e)}async _call(e,{top_k:r=5}={}){const n=await Tr(e),{pixel_values:a}=await this.processor(n),i=await this.model({pixel_values:a}),s=this.model.config.id2label,o=[];for(const l of i.logits){const u=await Nn(new ue("float32",wt(l.data),l.dims),r),p=u[0].tolist(),m=u[1].tolist().map((d,_)=>({label:s?s[d]:`LABEL_${d}`,score:p[_]}));o.push(m)}return Array.isArray(e)?o:o[0]}}class Dk extends tt{constructor(e){super(e),this.subtasks_mapping={panoptic:"post_process_panoptic_segmentation",instance:"post_process_instance_segmentation",semantic:"post_process_semantic_segmentation"}}async _call(e,{threshold:r=.5,mask_threshold:n=.5,overlap_mask_area_threshold:a=.8,label_ids_to_fuse:i=null,target_sizes:s=null,subtask:o=null}={}){if(Array.isArray(e)&&e.length!==1)throw Error("Image segmentation pipeline currently only supports a batch size of 1.");const u=await Tr(e),p=u.map(S=>[S.height,S.width]),{pixel_values:h,pixel_mask:m}=await this.processor(u),d=await this.model({pixel_values:h,pixel_mask:m});let _=null;if(o!==null)_=this.subtasks_mapping[o];else for(let[S,$]of Object.entries(this.subtasks_mapping))if($ in this.processor.feature_extractor){_=this.processor.feature_extractor[$].bind(this.processor.feature_extractor),o=S;break}const w=this.model.config.id2label,v=[];if(o==="panoptic"||o==="instance"){const S=_(d,r,n,a,i,s??p)[0],$=S.segmentation;for(const E of S.segments_info){const T=new Uint8ClampedArray($.data.length);for(let P=0;P<$.data.length;++P)$.data[P]===E.id&&(T[P]=255);const A=new yt(T,$.dims[1],$.dims[0],1);v.push({score:E.score,label:w[E.label_id],mask:A})}}else if(o==="semantic"){const{segmentation:S,labels:$}=_(d,s??p)[0];for(const E of $){const T=new Uint8ClampedArray(S.data.length);for(let P=0;Pn.replace("{}",m)),o=this.tokenizer(s,{padding:this.model.config.model_type==="siglip"?"max_length":!0,truncation:!0}),{pixel_values:l}=await this.processor(i),u=await this.model({...o,pixel_values:l}),p=this.model.config.model_type==="siglip"?m=>m.sigmoid().data:m=>wt(m.data),h=[];for(const m of u.logits_per_image){const _=[...p(m)].map((w,v)=>({score:w,label:r[v]}));_.sort((w,v)=>v.score-w.score),h.push(_)}return a?h:h[0]}}class Fk extends tt{constructor(e){super(e)}async _call(e,{threshold:r=.9,percentage:n=!1}={}){const a=Array.isArray(e);if(a&&e.length!==1)throw Error("Object detection pipeline currently only supports a batch size of 1.");const i=await Tr(e),s=n?null:i.map(d=>[d.height,d.width]),{pixel_values:o,pixel_mask:l}=await this.processor(i),u=await this.model({pixel_values:o,pixel_mask:l}),p=this.processor.feature_extractor.post_process_object_detection(u,r,s),h=this.model.config.id2label,m=p.map(d=>d.boxes.map((_,w)=>({score:d.scores[w],label:h[d.classes[w]],box:m0(_,!n)})));return a?m:m[0]}}class Lk extends tt{constructor(e){super(e)}async _call(e,r,{threshold:n=.1,top_k:a=null,percentage:i=!1}={}){const s=Array.isArray(e),o=await Tr(e),l=this.tokenizer(r,{padding:!0,truncation:!0}),u=await this.processor(o),p=[];for(let h=0;h({score:v.scores[E],label:r[v.classes[E]],box:m0($,!i)})).sort(($,E)=>E.score-$.score);a!==null&&(S=S.slice(0,a)),p.push(S)}return s?p:p[0]}}class Uk extends tt{constructor(e){super(e)}async _call(e,r,n={}){throw new Error("This pipeline is not yet supported in Transformers.js v3.")}}class Wk extends tt{DEFAULT_VOCODER_ID="Xenova/speecht5_hifigan";constructor(e){super(e),this.vocoder=e.vocoder??null}async _call(e,{speaker_embeddings:r=null}={}){throw new Error("This pipeline is not yet supported in Transformers.js v3.")}async _call_text_to_waveform(e){const r=this.tokenizer(e,{padding:!0,truncation:!0}),{waveform:n}=await this.model(r),a=this.model.config.sampling_rate;return{audio:n.data,sampling_rate:a}}async _call_text_to_spectrogram(e,{speaker_embeddings:r}){if(this.vocoder||(console.log("No vocoder specified, using default HifiGan vocoder."),this.vocoder=await ma.from_pretrained(this.DEFAULT_VOCODER_ID,{dtype:"fp32"})),(typeof r=="string"||r instanceof URL)&&(r=new Float32Array(await(await fetch(r)).arrayBuffer())),r instanceof Float32Array)r=new ue("float32",r,[1,r.length]);else if(!(r instanceof ue))throw new Error("Speaker embeddings must be a `Tensor`, `Float32Array`, `string`, or `URL`.");const{input_ids:n}=this.tokenizer(e,{padding:!0,truncation:!0}),{waveform:a}=await this.model.generate_speech(n,r,{vocoder:this.vocoder}),i=this.processor.feature_extractor.config.sampling_rate;return{audio:a.data,sampling_rate:i}}}class Vk extends tt{constructor(e){super(e)}async _call(e){const r=await Tr(e),n=await this.processor(r),a=await this.model(n),i=[];for(const s of a.reconstruction){const o=s.squeeze().clamp_(0,1).mul_(255).round_().to("uint8");i.push(yt.fromTensor(o))}return i.length>1?i:i[0]}}class Gk extends tt{constructor(e){super(e)}async _call(e){const r=await Tr(e),n=await this.processor(r),{predicted_depth:a}=await this.model(n),i=[];for(let s=0;s1?i:i[0]}}const Ip=Object.freeze({"text-classification":{tokenizer:ht,pipeline:xk,model:Ep,default:{model:"Xenova/distilbert-base-uncased-finetuned-sst-2-english"},type:"text"},"token-classification":{tokenizer:ht,pipeline:Sk,model:o3,default:{model:"Xenova/bert-base-multilingual-cased-ner-hrl"},type:"text"},"question-answering":{tokenizer:ht,pipeline:kk,model:h3,default:{model:"Xenova/distilbert-base-cased-distilled-squad"},type:"text"},"fill-mask":{tokenizer:ht,pipeline:Ek,model:p3,default:{model:"Xenova/bert-base-uncased"},type:"text"},summarization:{tokenizer:ht,pipeline:Ck,model:ho,default:{model:"Xenova/distilbart-cnn-6-6"},type:"text"},translation:{tokenizer:ht,pipeline:g0,model:ho,default:{model:"Xenova/t5-small"},type:"text"},"text2text-generation":{tokenizer:ht,pipeline:Cl,model:ho,default:{model:"Xenova/flan-t5-small"},type:"text"},"text-generation":{tokenizer:ht,pipeline:Tk,model:c3,default:{model:"Xenova/gpt2"},type:"text"},"zero-shot-classification":{tokenizer:ht,pipeline:Ak,model:Ep,default:{model:"Xenova/distilbert-base-uncased-mnli"},type:"text"},"audio-classification":{pipeline:Ok,model:v3,processor:Ot,default:{model:"Xenova/wav2vec2-base-superb-ks"},type:"audio"},"zero-shot-audio-classification":{tokenizer:ht,pipeline:zk,model:ma,processor:Ot,default:{model:"Xenova/clap-htsat-unfused"},type:"multimodal"},"automatic-speech-recognition":{tokenizer:ht,pipeline:Pk,model:[l3,b3],processor:Ot,default:{model:"Xenova/whisper-tiny.en"},type:"multimodal"},"text-to-audio":{tokenizer:ht,pipeline:Wk,model:[d3,u3],processor:[Ot,null],default:{model:"Xenova/speecht5_tts"},type:"text"},"image-to-text":{tokenizer:ht,pipeline:Rk,model:f3,processor:Ot,default:{model:"Xenova/vit-gpt2-image-captioning"},type:"multimodal"},"image-classification":{pipeline:Bk,model:m3,processor:Ot,default:{model:"Xenova/vit-base-patch16-224"},type:"multimodal"},"image-segmentation":{pipeline:Dk,model:[g3,_3],processor:Ot,default:{model:"Xenova/detr-resnet-50-panoptic"},type:"multimodal"},"zero-shot-image-classification":{tokenizer:ht,pipeline:Nk,model:ma,processor:Ot,default:{model:"Xenova/clip-vit-base-patch32"},type:"multimodal"},"object-detection":{pipeline:Fk,model:y3,processor:Ot,default:{model:"Xenova/detr-resnet-50"},type:"multimodal"},"zero-shot-object-detection":{tokenizer:ht,pipeline:Lk,model:w3,processor:Ot,default:{model:"Xenova/owlvit-base-patch32"},type:"multimodal"},"document-question-answering":{tokenizer:ht,pipeline:Uk,model:$3,processor:Ot,default:{model:"Xenova/donut-base-finetuned-docvqa"},type:"multimodal"},"image-to-image":{pipeline:Vk,model:x3,processor:Ot,default:{model:"Xenova/swin2SR-classical-sr-x2-64"},type:"image"},"depth-estimation":{pipeline:Gk,model:S3,processor:Ot,default:{model:"Xenova/dpt-large"},type:"image"},"feature-extraction":{tokenizer:ht,pipeline:Ik,model:ma,default:{model:"Xenova/all-MiniLM-L6-v2"},type:"text"},"image-feature-extraction":{processor:Ot,pipeline:Mk,model:[k3,ma],default:{model:"Xenova/vit-base-patch16-224-in21k"},type:"image"}}),Hk=Object.freeze({"sentiment-analysis":"text-classification",ner:"token-classification",asr:"automatic-speech-recognition","text-to-speech":"text-to-audio",embeddings:"feature-extraction"});async function jk(t,e=null,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:i=!1,revision:s="main",device:o=null,dtype:l=null,model_file_name:u=null,session_options:p={}}={}){t=Hk[t]??t;const h=Ip[t.split("_",1)[0]];if(!h)throw Error(`Unsupported pipeline: ${t}. Must be one of [${Object.keys(Ip)}]`);e||(e=h.default.model,console.log(`No model specified. Using default model: "${e}".`));const m={progress_callback:r,config:n,cache_dir:a,local_files_only:i,revision:s,device:o,dtype:l,model_file_name:u,session_options:p},d=new Map([["tokenizer",h.tokenizer],["model",h.model],["processor",h.processor]]),_=await qk(d,e,m);_.task=t,An(r,{status:"ready",task:t,model:e});const w=h.pipeline;return new w(_)}async function qk(t,e,r){const n=Object.create(null),a=[];for(let[i,s]of t.entries()){if(!s)continue;let o;Array.isArray(s)?o=new Promise(async(l,u)=>{let p;for(let h of s){if(h===null){l(null);return}try{l(await h.from_pretrained(e,r));return}catch(m){if(m.message?.includes("Unsupported model type"))p=m;else{u(m);return}}}u(p)}):o=s.from_pretrained(e,r),n[i]=o,a.push(o)}await Promise.all(a);for(let[i,s]of Object.entries(n))n[i]=await s;return n}an.IS_PROCESS_AVAILABLE;const gs=document.getElementById("status"),Mp=document.getElementById("container"),ga=document.getElementById("video"),go=document.getElementById("overlay");gs.textContent="Loading model (8 MB)...";async function Kk(){try{return(await navigator.gpu.requestAdapter()).features.has("shader-f16")}catch{return!1}}const Yk="onnx-community/mobilenetv4_conv_small.e1200_r224_in1k";let _0;try{_0=await jk("image-classification",Yk,{dtype:await Kk()?"fp16":"fp32",device:"webgpu"})}catch(t){throw gs.textContent=t.message,alert(t.message),t}gs.textContent="Ready";const Xk=.1,Qr=256,Bo=document.createElement("canvas");Bo.width=Bo.height=Qr;const Op=Bo.getContext("2d",{willReadFrequently:!0});let _o=!1,yo;function y0(){_o||(_o=!0,async function(){Op.drawImage(ga,0,0,Qr,Qr);const t=Op.getImageData(0,0,Qr,Qr).data,e=new yt(t,Qr,Qr,4),r=await _0(e,{top_k:null});go.innerHTML="";for(const{label:n,score:a}of r){if(a{ga.srcObject=t,ga.play();const e=t.getVideoTracks()[0],{width:r,height:n}=e.getSettings();ga.width=r,ga.height=n;const a=r/n,[i,s]=a>720/405?[720,720/a]:[405*a,405];Mp.style.width=`${i}px`,Mp.style.height=`${s}px`,window.requestAnimationFrame(y0)}).catch(t=>{alert(t)});