diff --git "a/worker-CnHDnJ52.js" "b/worker-CnHDnJ52.js" new file mode 100644--- /dev/null +++ "b/worker-CnHDnJ52.js" @@ -0,0 +1,2431 @@ +var R$=Object.defineProperty;var D$=(oe,nt,ft)=>nt in oe?R$(oe,nt,{enumerable:!0,configurable:!0,writable:!0,value:ft}):oe[nt]=ft;var T=(oe,nt,ft)=>(D$(oe,typeof nt!="symbol"?nt+"":nt,ft),ft);(function(){"use strict";var Zh;var oe={},nt=Object.freeze({__proto__:null,default:oe});const ft="3.0.0-alpha.0",zr=typeof self<"u",sf=zr&&self.constructor.name==="DedicatedWorkerGlobalScope",zi=zr&&"caches"in self,af=typeof navigator<"u"&&"gpu"in navigator,Ai=typeof process<"u",of=Ai&&((Zh=process==null?void 0:process.release)==null?void 0:Zh.name)==="node",Ar=!Ri(oe),Pi=!Ri(oe),mt=Object.freeze({IS_BROWSER_ENV:zr,IS_WEBWORKER_ENV:sf,IS_WEB_CACHE_AVAILABLE:zi,IS_WEBGPU_AVAILABLE:af,IS_PROCESS_AVAILABLE:Ai,IS_NODE_ENV:of,IS_FS_AVAILABLE:Ar,IS_PATH_AVAILABLE:Pi}),An=Ar&&Pi,Oi=An?oe.dirname(oe.dirname(oe.fileURLToPath(self.location.href))):"./",lf=An?oe.join(Oi,"/.cache/"):null,Bi="/models/",uf=An?oe.join(Oi,Bi):Bi,Ee={version:ft,backends:{onnx:{},tfjs:{}},allowRemoteModels:!0,remoteHost:"https://huggingface.co/",remotePathTemplate:"{model}/resolve/{revision}/",allowLocalModels:!zr,localModelPath:uf,useFS:Ar,useBrowserCache:zi,useFSCache:Ar,cacheDir:lf,useCustomCache:!1,customCache:null};function Ri(e){return Object.keys(e).length===0}const Ce=class{constructor(){let e=function(...t){return e._call(...t)};return Object.setPrototypeOf(e,new.target.prototype)}_call(...e){throw Error("Must implement _call method in subclass")}};function Vt(e,t){e&&e(t)}function df(e){return Object.fromEntries(Object.entries(e).map(([t,r])=>[r,t]))}function Di(e){return e.replace(/[.*+?^${}()|[\]\\]/g,"\\$&")}function cf(e){return Number.isInteger(e)||typeof e=="bigint"}function we(...e){return Array.prototype.concat.apply([],e)}function gt(e,t){return Object.assign({},...t.map(r=>{if(e[r]!==void 0)return{[r]:e[r]}}))}var st={};class Pr{constructor(t){T(this,"_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"});if(this.filePath=t,this.headers=new Headers,this.exists=oe.existsSync(t),this.exists){this.status=200,this.statusText="OK";let r=oe.statSync(t);this.headers.set("content-length",r.size.toString()),this.updateContentType();let n=this;this.body=new ReadableStream({start(s){n.arrayBuffer().then(i=>{s.enqueue(new Uint8Array(i)),s.close()})}})}else this.status=404,this.statusText="Not Found",this.body=null}updateContentType(){const t=this.filePath.toString().split(".").pop().toLowerCase();this.headers.set("content-type",this._CONTENT_TYPE_MAP[t]??"application/octet-stream")}clone(){let t=new Pr(this.filePath);return t.exists=this.exists,t.status=this.status,t.statusText=this.statusText,t.headers=new Headers(this.headers),t}async arrayBuffer(){return(await oe.promises.readFile(this.filePath)).buffer}async blob(){const t=await oe.promises.readFile(this.filePath);return new Blob([t],{type:this.headers.get("content-type")})}async text(){return await oe.promises.readFile(this.filePath,"utf8")}async json(){return JSON.parse(await this.text())}}function Pn(e,t=null){let r;try{r=new URL(e)}catch{return!1}return t&&!t.includes(r.hostname)?!1:r.protocol==="http:"||r.protocol==="https:"}async function Li(e){var t;if(Ee.useFS&&!Pn(e))return new Pr(e);if(typeof process<"u"&&((t=process==null?void 0:process.release)==null?void 0:t.name)==="node"){const r=!!(st!=null&&st.TESTING_REMOTELY),n=Ee.version,s=new Headers;if(s.set("User-Agent",`transformers.js/${n}; is_ci/${r};`),Pn(e,["huggingface.co","hf.co"])){const a=(st==null?void 0:st.HF_TOKEN)??(st==null?void 0:st.HF_ACCESS_TOKEN);a&&s.set("Authorization",`Bearer ${a}`)}return fetch(e,{headers:s})}else return fetch(e)}const pf={400:"Bad request error occurred while trying to load file",401:"Unauthorized access to file",403:"Forbidden access to file",404:"Could not locate file",408:"Request timeout error occurred while trying to load file",500:"Internal server error error occurred while trying to load file",502:"Bad gateway error occurred while trying to load file",503:"Service unavailable error occurred while trying to load file",504:"Gateway timeout error occurred while trying to load file"};function hf(e,t,r){if(!r)return null;const n=pf[e]??`Error (${e}) occurred while trying to load file`;throw Error(`${n}: "${t}".`)}class Fi{constructor(t){this.path=t}async match(t){let r=oe.join(this.path,t),n=new Pr(r);if(n.exists)return n}async put(t,r){const n=Buffer.from(await r.arrayBuffer());let s=oe.join(this.path,t);try{await oe.promises.mkdir(oe.dirname(s),{recursive:!0}),await oe.promises.writeFile(s,n)}catch(i){console.warn("An error occurred while writing the file to cache:",i)}}}async function ff(e,...t){for(let r of t)try{let n=await e.match(r);if(n)return n}catch{continue}}async function Or(e,t,r=!0,n={}){if(!Ee.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`).");Vt(n.progress_callback,{status:"initiate",name:e,file:t});let s;if(!s&&Ee.useBrowserCache){if(typeof caches>"u")throw Error("Browser cache is not available in this environment.");try{s=await caches.open("transformers-cache")}catch(w){console.warn("An error occurred while opening the browser cache:",w)}}if(!s&&Ee.useFSCache&&(s=new Fi(n.cache_dir??Ee.cacheDir)),!s&&Ee.useCustomCache)throw Error("`env.useCustomCache=true`, but `env.customCache` is not defined.");const i=n.revision??"main";let a=Br(e,t),o=Br(Ee.localModelPath,a),l=Br(Ee.remoteHost,Ee.remotePathTemplate.replaceAll("{model}",e).replaceAll("{revision}",encodeURIComponent(i)),t),u=i==="main"?a:Br(e,i,t),d,c=s instanceof Fi?u:l,p=!1,h;s&&(h=await ff(s,o,c));const f=h!==void 0;if(h===void 0){if(Ee.allowLocalModels)if(Pn(a)){if(n.local_files_only)throw new Error(`\`local_files_only=true\`, but attempted to load a remote file from: ${a}.`)}else try{h=await Li(o),d=o}catch(_){console.warn(`Unable to load from local path "${o}": "${_}"`)}if(h===void 0||h.status===404){if(n.local_files_only||!Ee.allowRemoteModels){if(r)throw Error(`\`local_files_only=true\` or \`env.allowRemoteModels=false\` and file was not found locally at "${o}".`);return null}if(h=await Li(l),h.status!==200)return hf(h.status,l,r);d=c}p=s&&typeof Response<"u"&&h instanceof Response&&h.status===200}Vt(n.progress_callback,{status:"download",name:e,file:t});const m={status:"progress",name:e,file:t};let g;return n.progress_callback?f&&typeof navigator<"u"&&/firefox/i.test(navigator.userAgent)?(g=new Uint8Array(await h.arrayBuffer()),Vt(n.progress_callback,{...m,progress:100,loaded:g.length,total:g.length})):g=await mf(h,w=>{Vt(n.progress_callback,{...m,...w})}):g=new Uint8Array(await h.arrayBuffer()),p&&d&&await s.match(d)===void 0&&await s.put(d,new Response(g,{headers:h.headers})).catch(w=>{console.warn(`Unable to add response to browser cache: ${w}.`)}),Vt(n.progress_callback,{status:"done",name:e,file:t}),g}async function _t(e,t,r=!0,n={}){let s=await Or(e,t,r,n);if(s===null)return{};let a=new TextDecoder("utf-8").decode(s);return JSON.parse(a)}async function mf(e,t){const r=e.headers.get("Content-Length");r===null&&console.warn("Unable to determine content-length from response headers. Will expand buffer when needed.");let n=parseInt(r??"0"),s=new Uint8Array(n),i=0;const a=e.body.getReader();async function o(){const{done:l,value:u}=await a.read();if(l)return;let d=i+u.length;if(d>n){n=d;let p=new Uint8Array(n);p.set(s),s=p}s.set(u,i),i=d;const c=i/n*100;return t({progress:c,loaded:i,total:n}),o()}return await o(),s}function Br(...e){return e=e.map((t,r)=>(r&&(t=t.replace(new RegExp("^/"),"")),r!==e.length-1&&(t=t.replace(new RegExp("/$"),"")),t)),e.join("/")}function gf(e,t,r){const n=new Array(r.length),s=new Array(r.length);for(let o=r.length-1,l=1;o>=0;--o)s[o]=l,n[o]=t[r[o]],l*=n[o];const i=r.map((o,l)=>s[r.indexOf(l)]),a=new e.constructor(e.length);for(let o=0;o=0;--u)l+=d%t[u]*i[u],d=Math.floor(d/t[u]);a[l]=e[o]}return[a,n]}function Ni(e){const t=On(e)[0],r=e.map(i=>Math.exp(i-t)),n=r.reduce((i,a)=>i+a,0);return r.map(i=>i/n)}function Ui(e,t=0){return e=Array.from(e).map((r,n)=>[n,r]).sort((r,n)=>n[1]-r[1]),t!==null&&t>0&&(e=e.slice(0,t)),e}function _f(e){if(e.length===0)throw Error("Array must not be empty");let t=e[0],r=0;for(let n=1;nt&&(t=e[n],r=n);return[Number(t),r]}function wf(e,t){if(t%2===0||t<=0)throw new Error("Window size must be a positive odd number");const r=new e.constructor(e.length),n=new e.constructor(t),s=Math.floor(t/2);for(let i=0;i=e.length&&(l=2*(e.length-1)-l),n[a++]=e[l]}n.sort(),r[i]=n[s]}return r}function Rr(e,t){const r=Math.pow(10,t);return Math.round(e*r)/r}/*! + * ONNX Runtime Web v1.19.0-esmtest.20240513-a16cd2bd21 + * Copyright (c) Microsoft Corporation. All rights reserved. + * Licensed under the MIT License. + */var Bn=Object.defineProperty,yf=Object.getOwnPropertyDescriptor,bf=Object.getOwnPropertyNames,vf=Object.prototype.hasOwnProperty,$f=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,r)=>(typeof require<"u"?require:t)[r]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')}),A=(e,t)=>()=>(e&&(t=e(e=0)),t),Dr=(e,t)=>{for(var r in t)Bn(e,r,{get:t[r],enumerable:!0})},xf=(e,t,r,n)=>{if(t&&typeof t=="object"||typeof t=="function")for(let s of bf(t))!vf.call(e,s)&&s!==r&&Bn(e,s,{get:()=>t[s],enumerable:!(n=yf(t,s))||n.enumerable});return e},Rn=e=>xf(Bn({},"__esModule",{value:!0}),e),jt,ut,wt,qi,Dn,Ln=A(()=>{jt=new Map,ut=[],wt=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let n=jt.get(e);if(n===void 0)jt.set(e,{backend:t,priority:r});else{if(n.priority>r)return;if(n.priority===r&&n.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${r}`)}if(r>=0){let s=ut.indexOf(e);s!==-1&&ut.splice(s,1);for(let i=0;i{let t=jt.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(n){return r||(t.error=`${n}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},Dn=async e=>{let t=e.executionProviders||[],r=t.map(l=>typeof l=="string"?l:l.name),n=r.length===0?ut:r,s,i=[],a=new Set;for(let l of n){let u=await qi(l);typeof u=="string"?i.push({name:l,err:u}):(s||(s=u),s===u&&a.add(l))}if(!s)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=t.filter(l=>a.has(typeof l=="string"?l:l.name));return[s,new Proxy(e,{get:(l,u)=>u==="executionProviders"?o:Reflect.get(l,u)})]}}),kf=A(()=>{Ln()}),Gi,Sf=A(()=>{Gi="1.19.0-esmtest.20240513-a16cd2bd21"}),Fn,Be,Vi=A(()=>{Sf(),Fn="warning",Be={wasm:{},webgl:{},webgpu:{},versions:{common:Gi},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}`);Fn=e}},get logLevel(){return Fn}},Object.defineProperty(Be,"logLevel",{enumerable:!0})}),se,Ef=A(()=>{Vi(),se=Be}),ji,Wi,Tf=A(()=>{ji=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let n=r.getContext("2d");if(n!=null){let s,i;(t==null?void 0:t.tensorLayout)!==void 0&&t.tensorLayout==="NHWC"?(s=e.dims[2],i=e.dims[3]):(s=e.dims[3],i=e.dims[2]);let a=(t==null?void 0:t.format)!==void 0?t.format:"RGB",o=t==null?void 0:t.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 d=i*s,c=0,p=d,h=d*2,f=-1;a==="RGBA"?(c=0,p=d,h=d*2,f=d*3):a==="RGB"?(c=0,p=d,h=d*2):a==="RBG"&&(c=0,h=d,p=d*2);for(let m=0;m{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let s,i,a;(t==null?void 0:t.tensorLayout)!==void 0&&t.tensorLayout==="NHWC"?(s=e.dims[2],i=e.dims[1],a=e.dims[3]):(s=e.dims[3],i=e.dims[2],a=e.dims[1]);let o=t!==void 0&&t.format!==void 0?t.format:"RGB",l=t==null?void 0:t.norm,u,d;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?d=[0,0,0,0]:typeof l.bias=="number"?d=[l.bias,l.bias,l.bias,l.bias]:(d=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(d[3]=l.bias[3]));let c=i*s;if(t!==void 0&&(t.format!==void 0&&a===4&&t.format!=="RGBA"||a===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let p=4,h=0,f=1,m=2,g=3,w=0,_=c,v=c*2,b=-1;o==="RGBA"?(w=0,_=c,v=c*2,b=c*3):o==="RGB"?(w=0,_=c,v=c*2):o==="RBG"&&(w=0,v=c,_=c*2),n=r.createImageData(s,i);for(let y=0;y{Un(),Lr=(e,t)=>{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:n}=t,s=t.norm??{mean:255,bias:0},i,a;typeof s.mean=="number"?i=[s.mean,s.mean,s.mean,s.mean]:i=[s.mean[0],s.mean[1],s.mean[2],s.mean[3]??255],typeof s.bias=="number"?a=[s.bias,s.bias,s.bias,s.bias]:a=[s.bias[0],s.bias[1],s.bias[2],s.bias[3]??0];let o=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",u=r*n,d=l==="RGBA"?new Float32Array(u*4):new Float32Array(u*3),c=4,p=0,h=1,f=2,m=3,g=0,w=u,_=u*2,v=-1;o==="RGB"&&(c=3,p=0,h=1,f=2,m=-1),l==="RGBA"?v=u*3:l==="RBG"?(g=0,_=u,w=u*2):l==="BGR"&&(_=0,w=u,g=u*2);for(let b=0;b{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,n=typeof ImageData<"u"&&e instanceof ImageData,s=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,i=typeof e=="string",a,o=t??{},l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},u=d=>d instanceof HTMLCanvasElement||d instanceof OffscreenCanvas?d.getContext("2d"):null;if(r){let d=l();d.width=e.width,d.height=e.height;let c=u(d);if(c!=null){let p=e.height,h=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(p=t.resizedHeight,h=t.resizedWidth),t!==void 0){if(o=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=p,o.width=h}else o.tensorFormat="RGBA",o.height=p,o.width=h;c.drawImage(e,0,0),a=c.getImageData(0,0,h,p).data}else throw new Error("Can not access image data")}else if(n){let d,c;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(d=t.resizedHeight,c=t.resizedWidth):(d=e.height,c=e.width),t!==void 0&&(o=t),o.format="RGBA",o.height=d,o.width=c,t!==void 0){let p=l();p.width=c,p.height=d;let h=u(p);if(h!=null)h.putImageData(e,0,0),a=h.getImageData(0,0,c,d).data;else throw new Error("Can not access image data")}else a=e.data}else if(s){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let d=l();d.width=e.width,d.height=e.height;let c=u(d);if(c!=null){let p=e.height,h=e.width;return c.drawImage(e,0,0,h,p),a=c.getImageData(0,0,h,p).data,o.height=p,o.width=h,Lr(a,o)}else throw new Error("Can not access image data")}else{if(i)return new Promise((d,c)=>{let p=l(),h=u(p);if(!e||!h)return c();let f=new Image;f.crossOrigin="Anonymous",f.src=e,f.onload=()=>{p.width=f.width,p.height=f.height,h.drawImage(f,0,0,p.width,p.height);let m=h.getImageData(0,0,p.width,p.height);o.height=p.height,o.width=p.width,d(Lr(m.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(a!==void 0)return Lr(a,o);throw new Error("Input data provided is not supported - aborted tensor creation")},Ki=(e,t)=>{let{width:r,height:n,download:s,dispose:i}=t,a=[1,n,r,4];return new Fe({location:"texture",type:"float32",texture:e,dims:a,download:s,dispose:i})},Qi=(e,t)=>{let{dataType:r,dims:n,download:s,dispose:i}=t;return new Fe({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:n,download:s,dispose:i})},Xi=(e,t,r)=>new Fe({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})}),yt,Wt,Nn,Yi,Mf=A(()=>{yt=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),Wt=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),Nn=!1,Yi=()=>{if(!Nn){Nn=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;e&&(yt.set("int64",BigInt64Array),Wt.set(BigInt64Array,"int64")),t&&(yt.set("uint64",BigUint64Array),Wt.set(BigUint64Array,"uint64")),r?(yt.set("float16",Float16Array),Wt.set(Float16Array,"float16")):yt.set("float16",Uint16Array)}}}),Zi,Ji,Cf=A(()=>{Un(),Zi=e=>{let t=1;for(let r=0;r{switch(e.location){case"cpu":return new Fe(e.type,e.data,t);case"cpu-pinned":return new Fe({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new Fe({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new Fe({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),Fe,Un=A(()=>{Tf(),If(),Mf(),Cf(),Fe=class{constructor(e,t,r){Yi();let n,s;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,n=e.type,s=e.dims,e.location){case"cpu-pinned":{let a=yt.get(n);if(!a)throw new TypeError(`unsupported type "${n}" to create tensor from pinned buffer`);if(!(e.data instanceof a))throw new TypeError(`buffer should be of type ${a.name}`);this.cpuData=e.data;break}case"texture":{if(n!=="float32")throw new TypeError(`unsupported type "${n}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.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 "${n}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let a,o;if(typeof e=="string")if(n=e,o=r,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");a=t}else{let l=yt.get(e);if(l===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16"&&l===Uint16Array)throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");e==="uint64"||e==="int64"?a=l.from(t,BigInt):a=l.from(t)}else if(t instanceof l)a=t;else throw new TypeError(`A ${n} tensor's data must be type of ${l}`)}else if(o=t,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let l=typeof e[0];if(l==="string")n="string",a=e;else if(l==="boolean")n="bool",a=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${l}.`)}else{let l=Wt.get(e.constructor);if(l===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);n=l,a=e}if(o===void 0)o=[a.length];else if(!Array.isArray(o))throw new TypeError("A tensor's dims must be a number array");s=o,this.cpuData=a,this.dataLocation="cpu"}let i=Zi(s);if(this.cpuData&&i!==this.cpuData.length)throw new Error(`Tensor's size(${i}) does not match data length(${this.cpuData.length}).`);this.type=n,this.dims=s,this.size=i}static async fromImage(e,t){return Hi(e,t)}static fromTexture(e,t){return Ki(e,t)}static fromGpuBuffer(e,t){return Qi(e,t)}static fromPinnedBuffer(e,t,r){return Xi(e,t,r)}toDataURL(e){return ji(this,e)}toImageData(e){return Wi(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;let t=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=t,e&&this.disposer&&(this.disposer(),this.disposer=void 0),t}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Ji(this,e)}}}),$e,qn=A(()=>{Un(),$e=Fe}),Ht,Gn,Ne,Re,ea=A(()=>{Vi(),Ht=(e,t)=>{(typeof Be.trace>"u"?!Be.wasm.trace:!Be.trace)||console.timeStamp(`${e}::ORT::${t}`)},Gn=(e,t)=>{var s;let r=((s=new Error().stack)==null?void 0:s.split(/\r\n|\r|\n/g))||[],n=!1;for(let i=0;i{(typeof Be.trace>"u"?!Be.wasm.trace:!Be.trace)||Gn("BEGIN",e)},Re=e=>{(typeof Be.trace>"u"?!Be.wasm.trace:!Be.trace)||Gn("END",e)}}),ta,zf=A(()=>{Ln(),qn(),ea(),ta=class tf{constructor(t){this.handler=t}async run(t,r,n){Ne();let s={},i={};if(typeof t!="object"||t===null||t instanceof $e||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let a=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof $e)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");a=!1;for(let 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}.`);s[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,d=Object.getOwnPropertyNames(r);for(let c of this.outputNames)if(d.indexOf(c)!==-1){let p=r[c];(p===null||p instanceof $e)&&(u=!0,a=!1,s[c]=p)}if(u){if(typeof n=="object"&&n!==null)i=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else i=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let u of this.inputNames)if(typeof t[u]>"u")throw new Error(`input '${u}' is missing in 'feeds'.`);if(a)for(let u of this.outputNames)s[u]=null;let o=await this.handler.run(t,s,i),l={};for(let u in o)if(Object.hasOwnProperty.call(o,u)){let d=o[u];d instanceof $e?l[u]=d:l[u]=new $e(d.type,d.data,d.dims)}return Re(),l}async release(){return this.handler.dispose()}static async create(t,r,n,s){Ne();let i,a={};if(typeof t=="string"){if(i=t,typeof r=="object"&&r!==null)a=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof Uint8Array){if(i=t,typeof r=="object"&&r!==null)a=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof ArrayBuffer||typeof SharedArrayBuffer<"u"&&t instanceof SharedArrayBuffer){let d=t,c=0,p=t.byteLength;if(typeof r=="object"&&r!==null)a=r;else if(typeof r=="number"){if(c=r,!Number.isSafeInteger(c))throw new RangeError("'byteOffset' must be an integer.");if(c<0||c>=d.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${d.byteLength}).`);if(p=t.byteLength-c,typeof n=="number"){if(p=n,!Number.isSafeInteger(p))throw new RangeError("'byteLength' must be an integer.");if(p<=0||c+p>d.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${d.byteLength-c}].`);if(typeof s=="object"&&s!==null)a=s;else if(typeof s<"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(d,c,p)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[o,l]=await Dn(a),u=await o.createInferenceSessionHandler(i,l);return Re(),new tf(u)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}}),Vn,Af=A(()=>{zf(),Vn=ta}),Pf=A(()=>{}),Of=A(()=>{}),Bf=A(()=>{}),Rf=A(()=>{}),ra,na,Df=A(()=>{Ln(),qn(),ra="Training backend could not be resolved. Make sure you're using the correct configuration & WebAssembly files.",na=class rf{constructor(t,r,n){this.handler=t,this.hasOptimizerModel=r,this.hasEvalModel=n}get trainingInputNames(){return this.handler.inputNames}get trainingOutputNames(){return this.handler.outputNames}get evalInputNames(){if(this.hasEvalModel)return this.handler.evalInputNames;throw new Error("This training session has no evalModel loaded.")}get evalOutputNames(){if(this.hasEvalModel)return this.handler.evalOutputNames;throw new Error("This training session has no evalModel loaded.")}static async create(t,r){let n=t.evalModel||"",s=t.optimizerModel||"",i=r||{},[a,o]=await Dn(i);if(a.createTrainingSessionHandler){let l=await a.createTrainingSessionHandler(t.checkpointState,t.trainModel,n,s,o);return new rf(l,!!t.optimizerModel,!!t.evalModel)}else throw new Error(ra)}typeNarrowingForRunStep(t,r,n,s,i){let a={},o={};if(typeof n!="object"||n===null||n instanceof $e||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 s=="object"){if(s===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(s instanceof $e)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(s)){if(s.length===0)throw new TypeError("'fetches' cannot be an empty array.");l=!1;for(let u of s){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}.`);a[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,d=Object.getOwnPropertyNames(s);for(let c of r)if(d.indexOf(c)!==-1){let p=s[c];(p===null||p instanceof $e)&&(u=!0,l=!1,a[c]=p)}if(u){if(typeof i=="object"&&i!==null)o=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else o=s}}else if(typeof s<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let u of t)if(typeof n[u]>"u")throw new Error(`input '${u}' is missing in 'feeds'.`);if(l)for(let u of r)a[u]=null;return[a,o]}convertHandlerReturnTypeToMapOfTensors(t){let r={};for(let n in t)if(Object.hasOwnProperty.call(t,n)){let s=t[n];s instanceof $e?r[n]=s:r[n]=new $e(s.type,s.data,s.dims)}return r}async lazyResetGrad(){await this.handler.lazyResetGrad()}async runTrainStep(t,r,n){let[s,i]=this.typeNarrowingForRunStep(this.trainingInputNames,this.trainingOutputNames,t,r,n),a=await this.handler.runTrainStep(t,s,i);return this.convertHandlerReturnTypeToMapOfTensors(a)}async runOptimizerStep(t){if(this.hasOptimizerModel)await this.handler.runOptimizerStep(t||{});else throw new Error("This TrainingSession has no OptimizerModel loaded.")}async runEvalStep(t,r,n){if(this.hasEvalModel){let[s,i]=this.typeNarrowingForRunStep(this.evalInputNames,this.evalOutputNames,t,r,n),a=await this.handler.runEvalStep(t,s,i);return this.convertHandlerReturnTypeToMapOfTensors(a)}else throw new Error("This TrainingSession has no EvalModel loaded.")}async getParametersSize(t=!0){return this.handler.getParametersSize(t)}async loadParametersBuffer(t,r=!0){let n=await this.getParametersSize(r);if(t.length!==4*n)throw new Error("Size of the buffer passed into loadParametersBuffer must match the number of parameters in the model. Please use getParametersSize method to check.");return this.handler.loadParametersBuffer(t,r)}async getContiguousParameters(t=!0){return this.handler.getContiguousParameters(t)}async release(){return this.handler.dispose()}}}),jn,Lf=A(()=>{Df(),jn=na}),sa={};Dr(sa,{InferenceSession:()=>Vn,TRACE:()=>Ht,TRACE_FUNC_BEGIN:()=>Ne,TRACE_FUNC_END:()=>Re,Tensor:()=>$e,TrainingSession:()=>jn,env:()=>se,registerBackend:()=>wt});var Ue=A(()=>{kf(),Ef(),Af(),qn(),Pf(),Of(),ea(),Bf(),Rf(),Lf()}),Wn=A(()=>{}),ia={};Dr(ia,{default:()=>aa});var Hn,Kn,aa,Ff=A(()=>{var e;Kc(),Pt(),Fr(),Hn="ort-wasm-proxy-worker",Kn=((e=globalThis.self)==null?void 0:e.name)===Hn,Kn&&(self.onmessage=t=>{let{type:r,in:n}=t.data;try{switch(r){case"init-wasm":es(n.wasm).then(()=>{ti(n).then(()=>{postMessage({type:r})},s=>{postMessage({type:r,err:s})})},s=>{postMessage({type:r,err:s})});break;case"init-ep":{let{epName:s,env:i}=n;ri(i,s).then(()=>{postMessage({type:r})},a=>{postMessage({type:r,err:a})});break}case"copy-from":{let{buffer:s}=n,i=an(s);postMessage({type:r,out:i});break}case"create":{let{model:s,options:i}=n;ni(s,i).then(a=>{postMessage({type:r,out:a})},a=>{postMessage({type:r,err:a})});break}case"release":si(n),postMessage({type:r});break;case"run":{let{sessionId:s,inputIndices:i,inputs:a,outputIndices:o,options:l}=n;ai(s,i,a,o,new Array(o.length).fill(null),l).then(u=>{u.some(d=>d[3]!=="cpu")?postMessage({type:r,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:r,out:u},li([...a,...u]))},u=>{postMessage({type:r,err:u})});break}case"end-profiling":oi(n),postMessage({type:r});break;default:}}catch(s){postMessage({type:r,err:s})}}),aa=Kn?null:t=>new Worker(t??qe,{type:"module",name:Hn})}),qe,oa,Qn,la,ua,Xn,da,Yn,ca,pa,Fr=A(()=>{var e,t;Wn(),qe=self.location.href??(typeof document<"u"?(e=document.currentScript)==null?void 0:e.src:typeof self<"u"?(t=self.location)==null?void 0:t.href:void 0),oa=typeof location>"u"?void 0:location.origin,Qn=(r,n)=>{try{let s=n??qe;return(s?new URL(r,s):new URL(r)).origin===oa}catch{return!1}},la=(r,n)=>{let s=n??qe;try{return(s?new URL(r,s):new URL(r)).href}catch{return}},ua=(r,n)=>`${n??"./"}${r}`,Xn=async r=>{let n=await(await fetch(r,{credentials:"same-origin"})).blob();return URL.createObjectURL(n)},da=async r=>(await import(r)).default,Yn=(Ff(),Rn(ia)).default,ca=async()=>{if(!qe)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(Qn(qe))return[void 0,Yn()];let r=await Xn(qe);return[r,Yn(r)]},pa=async(r,n,s)=>{let i="ort-wasm-simd-threaded.jsep.mjs",a=r??la(i,n),o=s&&a&&!Qn(a,n),l=o?await Xn(a):a??ua(i,n);return[o?l:void 0,await da(l)]}}),Zn,Nr,Kt,Jn,ha,fa,es,me,Pt=A(()=>{Fr(),Nr=!1,Kt=!1,Jn=!1,ha=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},fa=()=>{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}},es=async e=>{if(Nr)return Promise.resolve();if(Kt)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Jn)throw new Error("previous call to 'initializeWebAssembly()' failed.");Kt=!0;let t=e.initTimeout,r=e.numThreads;if(!fa())throw new Error("WebAssembly SIMD is not supported in the current environment.");let n=ha();r>1&&!n&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+r+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=r=1);let s=e.wasmPaths,i=typeof s=="string"?s:void 0,a=s==null?void 0:s.mjs,o=(a==null?void 0:a.href)??a,l=s==null?void 0:s.wasm,u=(l==null?void 0:l.href)??l,[d,c]=await pa(o,i,r>1),p=!1,h=[];if(t>0&&h.push(new Promise(f=>{setTimeout(()=>{p=!0,f()},t)})),h.push(new Promise((f,m)=>{c({numThreads:r,locateFile:(g,w)=>u??(i??w)+g}).then(g=>{Kt=!1,Nr=!0,Zn=g,f(),d&&URL.revokeObjectURL(d)},g=>{Kt=!1,Jn=!0,m(g)})})),await Promise.race(h),p)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},me=()=>{if(Nr&&Zn)return Zn;throw new Error("WebAssembly is not initialized yet.")}}),pe,Ur,le,ts=A(()=>{Pt(),pe=(e,t)=>{let r=me(),n=r.lengthBytesUTF8(e)+1,s=r._malloc(n);return r.stringToUTF8(e,s,n),t.push(s),s},Ur=(e,t,r,n)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([s,i])=>{let a=t?t+s:s;if(typeof i=="object")Ur(i,a+".",r,n);else if(typeof i=="string"||typeof i=="number")n(a,i.toString());else if(typeof i=="boolean")n(a,i?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof i}`)})},le=e=>{let t=me(),r=t.stackSave();try{let n=t.stackAlloc(8);t._OrtGetLastError(n,n+4);let s=t.HEAP32[n/4],i=t.HEAPU32[n/4+1],a=i?t.UTF8ToString(i):"";throw new Error(`${e} ERROR_CODE: ${s}, ERROR_MESSAGE: ${a}`)}finally{t.stackRestore(r)}}}),ma,Nf=A(()=>{Pt(),ts(),ma=e=>{let t=me(),r=0,n=[],s=e||{};try{if((e==null?void 0:e.logSeverityLevel)===void 0)s.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==null?void 0:e.logVerbosityLevel)===void 0)s.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);(e==null?void 0:e.terminate)===void 0&&(s.terminate=!1);let i=0;return(e==null?void 0:e.tag)!==void 0&&(i=pe(e.tag,n)),r=t._OrtCreateRunOptions(s.logSeverityLevel,s.logVerbosityLevel,!!s.terminate,i),r===0&&le("Can't create run options."),(e==null?void 0:e.extra)!==void 0&&Ur(e.extra,"",new WeakSet,(a,o)=>{let l=pe(a,n),u=pe(o,n);t._OrtAddRunConfigEntry(r,l,u)!==0&&le(`Can't set a run config entry: ${a} - ${o}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseRunOptions(r),n.forEach(a=>t._free(a)),i}}}),ga,_a,wa,ya,ba,Uf=A(()=>{Pt(),ts(),ga=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}`)}},_a=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},wa=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)},ya=(e,t,r)=>{for(let n of t){let s=typeof n=="string"?n:n.name;switch(s){case"webnn":if(s="WEBNN",typeof n!="string"){let a=n;if(a!=null&&a.deviceType){let o=pe("deviceType",r),l=pe(a.deviceType,r);me()._OrtAddSessionConfigEntry(e,o,l)!==0&&le(`Can't set a session config entry: 'deviceType' - ${a.deviceType}.`)}if(a!=null&&a.numThreads){let o=a.numThreads;(typeof o!="number"||!Number.isInteger(o)||o<0)&&(o=0);let l=pe("numThreads",r),u=pe(o.toString(),r);me()._OrtAddSessionConfigEntry(e,l,u)!==0&&le(`Can't set a session config entry: 'numThreads' - ${a.numThreads}.`)}if(a!=null&&a.powerPreference){let o=pe("powerPreference",r),l=pe(a.powerPreference,r);me()._OrtAddSessionConfigEntry(e,o,l)!==0&&le(`Can't set a session config entry: 'powerPreference' - ${a.powerPreference}.`)}}break;case"webgpu":if(s="JS",typeof n!="string"){let a=n;if(a!=null&&a.preferredLayout){if(a.preferredLayout!=="NCHW"&&a.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${a.preferredLayout}`);let o=pe("preferredLayout",r),l=pe(a.preferredLayout,r);me()._OrtAddSessionConfigEntry(e,o,l)!==0&&le(`Can't set a session config entry: 'preferredLayout' - ${a.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${s}`)}let i=pe(s,r);me()._OrtAppendExecutionProvider(e,i)!==0&&le(`Can't append execution provider: ${s}.`)}},ba=e=>{let t=me(),r=0,n=[],s=e||{};wa(s);try{let i=ga(s.graphOptimizationLevel??"all"),a=_a(s.executionMode??"sequential"),o=typeof s.logId=="string"?pe(s.logId,n):0,l=s.logSeverityLevel??2;if(!Number.isInteger(l)||l<0||l>4)throw new Error(`log serverity level is not valid: ${l}`);let u=s.logVerbosityLevel??0;if(!Number.isInteger(u)||u<0||u>4)throw new Error(`log verbosity level is not valid: ${u}`);let d=typeof s.optimizedModelFilePath=="string"?pe(s.optimizedModelFilePath,n):0;if(r=t._OrtCreateSessionOptions(i,!!s.enableCpuMemArena,!!s.enableMemPattern,a,!!s.enableProfiling,0,o,l,u,d),r===0&&le("Can't create session options."),s.executionProviders&&ya(r,s.executionProviders,n),s.enableGraphCapture!==void 0){if(typeof s.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${s.enableGraphCapture}`);let c=pe("enableGraphCapture",n),p=pe(s.enableGraphCapture.toString(),n);t._OrtAddSessionConfigEntry(r,c,p)!==0&&le(`Can't set a session config entry: 'enableGraphCapture' - ${s.enableGraphCapture}.`)}if(s.freeDimensionOverrides)for(let[c,p]of Object.entries(s.freeDimensionOverrides)){if(typeof c!="string")throw new Error(`free dimension override name must be a string: ${c}`);if(typeof p!="number"||!Number.isInteger(p)||p<0)throw new Error(`free dimension override value must be a non-negative integer: ${p}`);let h=pe(c,n);t._OrtAddFreeDimensionOverride(r,h,p)!==0&&le(`Can't set a free dimension override: ${c} - ${p}.`)}return s.extra!==void 0&&Ur(s.extra,"",new WeakSet,(c,p)=>{let h=pe(c,n),f=pe(p,n);t._OrtAddSessionConfigEntry(r,h,f)!==0&&le(`Can't set a session config entry: ${c} - ${p}.`)}),[r,n]}catch(i){throw r!==0&&t._OrtReleaseSessionOptions(r),n.forEach(a=>t._free(a)),i}}}),rs,bt,Qt,ns,qr,ss,is,j=A(()=>{rs=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}`)}},bt=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";default:throw new Error(`unsupported data type: ${e}`)}},Qt=e=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][e],ns=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}`)}},qr=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}`)}},ss=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool",is=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}`)}}}),as,va=A(()=>{Wn(),as=async e=>{if(typeof e=="string"){let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),n=r?parseInt(r,10):0;if(n<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let s=t.body.getReader(),i;try{i=new ArrayBuffer(n)}catch(o){if(o instanceof RangeError){let l=Math.ceil(n/65536);i=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw o}let a=0;for(;;){let{done:o,value:l}=await s.read();if(o)break;let u=l.byteLength;new Uint8Array(i,a,u).set(l),a+=u}return new Uint8Array(i,0,n)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),$a,xa,ka,Sa,Ea,Ta,he,vt=A(()=>{j(),$a=["V","I","W","E","F"],xa=(e,t)=>{console.log(`[${$a[e]},${new Date().toISOString()}]${t}`)},Ea=(e,t)=>{ka=e,Sa=t},Ta=(e,t)=>{let r=qr(e),n=qr(ka);r>=n&&xa(r,typeof t=="function"?t():t)},he=(...e)=>{Sa&&Ta(...e)}}),Ia,qf=A(()=>{j(),Ia=(e,t)=>new(ns(t))(e)}),os=A(()=>{}),ls,Gr,Vr,Ma,Ca,us,ds,za,Aa,Gf=A(()=>{vt(),os(),ls=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),Gr=[],Vr=e=>Math.ceil(e/16)*16,Ma=e=>{for(let t=0;tCa++,ds=async(e,t,r,n)=>{let s=Vr(r),i=e.device.createBuffer({size:s,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let a=e.getCommandEncoder();e.endComputePass(),a.copyBufferToBuffer(t,0,i,0,s),e.flush(),await i.mapAsync(GPUMapMode.READ);let o=i.getMappedRange();if(n){let l=n();return l.set(new Uint8Array(o,0,r)),l}else return new Uint8Array(o.slice(0,r))}finally{i.destroy()}},za=class{constructor(e){this.backend=e,this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map,this.capturedPendingBuffers=new Map;for(let[t]of ls)Gr.push(t),this.freeBuffers.set(t,[]),this.freeUniformBuffers.set(t,[])}upload(e,t){let r=t.buffer,n=t.byteOffset,s=t.byteLength,i=Vr(s),a=this.storageCache.get(e);if(!a)throw new Error("gpu data for uploading does not exist");if(a.originalSize!==s)throw new Error(`inconsistent data size. gpu data size=${a.originalSize}, data size=${s}`);let o=this.backend.device.createBuffer({mappedAtCreation:!0,size:i,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),l=o.getMappedRange();new Uint8Array(l).set(new Uint8Array(r,n,s)),o.unmap();let u=this.backend.getCommandEncoder();this.backend.endComputePass(),u.copyBufferToBuffer(o,0,a.gpuData.buffer,0,i),he("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${e})`),this.buffersForUploadingPending.push(o)}memcpy(e,t){let r=this.storageCache.get(e);if(!r)throw new Error("source gpu data for memcpy does not exist");let n=this.storageCache.get(t);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 s=Vr(r.originalSize),i=this.backend.getCommandEncoder();this.backend.endComputePass(),i.copyBufferToBuffer(r.gpuData.buffer,0,n.gpuData.buffer,0,s)}registerExternalBuffer(e,t,r){let n;if(r){if(n=this.externalBuffers.get(r),n===void 0)throw new Error("previous buffer is not registered");if(e===r)return he("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${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=us();return this.storageCache.set(n,{gpuData:{id:n,type:0,buffer:e},originalSize:t}),this.externalBuffers.set(e,n),he("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${n}, registered.`),n}unregisterExternalBuffer(e){let t=this.externalBuffers.get(e);t!==void 0&&(this.storageCache.delete(t),this.externalBuffers.delete(e),he("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${t}`))}create(e,t=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=Ma(e),n,s=(t&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,i=(t&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(s||i){let o=(s?this.freeBuffers:this.freeUniformBuffers).get(r);o?o.length>0?n=o.pop():n=this.backend.device.createBuffer({size:r,usage:t}):n=this.backend.device.createBuffer({size:r,usage:t})}else n=this.backend.device.createBuffer({size:r,usage:t});let a={id:us(),type:0,buffer:n};return this.storageCache.set(a.id,{gpuData:a,originalSize:e}),he("verbose",()=>`[WebGPU] GpuDataManager.create(size=${e}) => id=${a.id}`),a}get(e){var t;return(t=this.storageCache.get(e))==null?void 0:t.gpuData}release(e){let t=this.storageCache.get(e);if(!t)throw new Error("releasing data does not exist");return he("verbose",()=>`[WebGPU] GpuDataManager.release(id=${e}), gpuDataId=${t.gpuData.id}`),this.storageCache.delete(e),this.buffersPending.push(t.gpuData.buffer),t.originalSize}async download(e,t){let r=this.storageCache.get(e);if(!r)throw new Error("data does not exist");await ds(this.backend,r.gpuData.buffer,r.originalSize,t)}refreshPendingBuffers(){for(let e of this.buffersForUploadingPending)e.destroy();if(this.buffersForUploadingPending=[],this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let e of this.buffersPending){let t=ls.get(e.size);if((e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else if((e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else e.destroy()}this.buffersPending=[]}else{let e=this.capturedPendingBuffers.get(this.backend.currentSessionId);e||(e=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,e));for(let t of this.buffersPending)e.push(t);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.freeUniformBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache.forEach(e=>{e.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onReleaseSession(e){let t=this.capturedPendingBuffers.get(e);t&&(t.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(e))}},Aa=(...e)=>new za(...e)}),Pa,ae,_e=A(()=>{Pa=class{constructor(e){Object.assign(this,e)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(e=>`${this[e]}`).join(";")),this.key}},ae=e=>new Pa(e)}),Oa,Ot,C,jr,Ba,cs,ps,J=A(()=>{Oa=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},Ot=class{static calcShape(e,t,r=!1){let n=e.length,s=t.length;if(n===0)return t;if(s===0)return e;let i=Math.max(e.length,t.length),a=new Array(i);if(r){if(n<2||s<2)return;let o=Oa.calcMatMulShape([e[n-2],e[n-1]],[t[s-2],t[s-1]]);if(o===void 0)return;[a[i-2],a[i-1]]=o}for(let o=r?3:1;o<=i;o++){let l=n-o<0?1:e[n-o],u=s-o<0?1:t[s-o];if(l!==u&&l>1&&u>1)return;let d=Math.max(l,u);if(l&&u)a[i-o]=Math.max(l,u);else{if(d>1)return;a[i-o]=0}}return a}static isValidBroadcast(e,t){let r=e.length,n=t.length;if(r>n)return!1;for(let s=1;s<=r;s++)if(e[r-s]!==1&&e[r-s]!==t[n-s])return!1;return!0}},C=class zn{static size(t){return zn.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let n=t.length;if(n===0)return[];let s=new Array(n),i=n-1;for(;i>=0;){if(t[i]%r===0){s[i]=t[i]/r;break}if(r%t[i]!==0)throw new Error("cannot convert shape");s[i]=1,r/=t[i],i--}for(i--;i>=0;i--)s[i]=t[i];return s}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 zn.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 ${t.length} dimensions.`);return zn.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,n){let s=1;for(let i=r;i=0;--s)n[s]=n[s+1]*t[s+1];return n}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(n=>this.normalizeAxis(n,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(n=>t[n]):t.slice().reverse()}static padShape(t,r){let n=t.length;return t.map((s,i)=>s+r[i]+r[i+n])}static areEqual(t,r){return t.length!==r.length?!1:t.every((n,s)=>n===r[s])}},jr=class Cr{static adjustPoolAttributes(t,r,n,s,i,a){if(!t&&n.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let o=0;o=n.length?n.push(r[o+2]):n[o]=r[o+2];for(let o=0;o=n[o]||a[o+n.length]>=n[o])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,n,s,i,a,o){if(o){if(i.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(s.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let l=0;l{j(),J(),Bt=64,Wr=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(e){case 10:return t>1?`vec${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: ${e}`)}},ye=(e,t=1)=>{let r=Wr(e,t);return typeof r=="string"?r:r[0]},Te=(e,t=1)=>{let r=Wr(e,t);return typeof r=="string"?r:r[1]},F=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:C.computeStrides(r)})}),t},ge=e=>e%4===0?4:e%2===0?2:1,$t=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,Rt=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}(${r})`,dt=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,Q=(e,t,r,n)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?n==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:n==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Hr=(e,t,r,n,s)=>{let i=typeof r=="number",a=i?r:r.length,o=[...new Array(a).keys()],l=a<2?"u32":a<=4?`vec${a}`:`array`,u=Wr(t,s),d=typeof u=="string"?u:u[1],c=typeof u=="string"?u:u[0],p={indices:l,value:d,storage:c,tensor:t},h=k=>typeof k=="string"?k:`${k}u`,f={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},m=i?"uniforms.":"",g=`${m}${e}_shape`,w=`${m}${e}_strides`,_="";for(let k=0;k ${p.indices} { + var indices: ${p.indices}; + var current = offset; + ${_} + return indices; + }`,b=k=>(f.offsetToIndices=!0,a<2?k:`o2i_${e}(${k})`),y=[];if(a>=2)for(let k=a-1;k>=0;k--)y.push(`${Q(w,k,a)} * (indices[${k}])`);let S=a<2?"":` + fn i2o_${e}(indices: ${p.indices}) -> u32 { + return ${y.join("+")}; + }`,x=k=>(f.indicesToOffset=!0,a<2?k:`i2o_${e}(${k})`),I=(...k)=>a===0?"0u":`${p.indices}(${k.map(h).join(",")})`,D=(k,L)=>a<2?`${k}`:`${Q(k,L,a)}`,R=(k,L,G)=>a<2?`${k}=${G};`:`${Q(k,L,a)}=${G};`,U={},q=(k,L)=>{f.broadcastedIndicesToOffset=!0;let G=`${L.name}broadcastedIndicesTo${e}Offset`;if(G in U)return`${G}(${k})`;let ie=[];for(let ce=a-1;ce>=0;ce--){let rt=L.indicesGet("outputIndices",ce+L.rank-a);ie.push(`${D(w,ce)} * (${rt} % ${D(g,ce)})`)}return U[G]=`fn ${G}(outputIndices: ${L.type.indices}) -> u32 { + return ${ie.length>0?ie.join("+"):"0u"}; + }`,`${G}(${k})`},H=(k,L)=>(()=>{if(p.storage===p.value)return`${e}[${k}]=${L};`;if(p.storage==="vec2"&&p.value==="i32")return`${e}[${k}]=vec2(u32(${L}), select(0u, 0xFFFFFFFFu, ${L} < 0));`;if(p.storage==="vec2"&&p.value==="u32")return`${e}[${k}]=vec2(u32(${L}), 0u);`;if(p.storage==="u32"&&p.value==="vec4")return`${e}[${k}]=dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(${L}));`;throw new Error(`not supported combination of storage type ${p.storage} and value type ${p.value} yet`)})(),V=k=>(()=>{if(p.storage===p.value)return`${e}[${k}]`;if(p.storage==="vec2"&&p.value==="i32")return`i32(${e}[${k}].x)`;if(p.storage==="vec2"&&p.value==="u32")return`u32(${e}[${k}].x)`;if(p.storage==="u32"&&p.value==="vec4")return`vec4(bool(${e}[${k}] & 0xFFu), bool(${e}[${k}] & 0xFF00u), bool(${e}[${k}] & 0xFF0000u), bool(${e}[${k}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${p.storage} and value type ${p.value} yet`)})(),E=a<2?"":` + fn get_${e}ByIndices(indices: ${p.indices}) -> ${d} { + return ${V(`i2o_${e}(indices)`)}; + }`,M=a<2?"":(()=>{let k=o.map(G=>`d${G}: u32`).join(", "),L=o.map(G=>`d${G}`).join(", ");return` + fn get_${e}(${k}) -> ${d} { + return get_${e}ByIndices(${I(L)}); + }`})(),O=(...k)=>{if(k.length!==a)throw new Error(`indices length must be ${a}`);let L=k.map(h).join(",");return a===0?V("0u"):a===1?V(L[0]):(f.get=!0,f.getByIndices=!0,f.indicesToOffset=!0,`get_${e}(${L})`)},B=k=>a<2?V(k):(f.getByIndices=!0,f.indicesToOffset=!0,`get_${e}ByIndices(${k})`),K=a<2?"":` + fn set_${e}ByIndices(indices: ${p.indices}, value: ${d}) { + ${H(`i2o_${e}(indices)`,"value")} + }`,ee=a<2?"":(()=>{let k=o.map(G=>`d${G}: u32`).join(", "),L=o.map(G=>`d${G}`).join(", ");return` + fn set_${e}(${k}, value: ${d}) { + set_${e}ByIndices(${I(L)}, value); + }`})();return{impl:()=>{let k=[],L=!1;return f.offsetToIndices&&(k.push(v),L=!0),f.indicesToOffset&&(k.push(S),L=!0),f.broadcastedIndicesToOffset&&(Object.values(U).forEach(G=>k.push(G)),L=!0),f.set&&(k.push(ee),L=!0),f.setByIndices&&(k.push(K),L=!0),f.get&&(k.push(M),L=!0),f.getByIndices&&(k.push(E),L=!0),!i&&L&&k.unshift(`const ${g} = ${p.indices}(${r.join(",")});`,`const ${w} = ${p.indices}(${C.computeStrides(r).join(",")});`),k.join(` +`)},type:p,offsetToIndices:b,indicesToOffset:x,broadcastedIndicesToOffset:q,indices:I,indicesGet:D,indicesSet:R,set:(...k)=>{if(k.length!==a+1)throw new Error(`indices length must be ${a}`);let L=k[a];if(typeof L!="string")throw new Error("value must be string");let G=k.slice(0,a).map(h).join(",");return a===0?H("0u",L):a===1?H(G[0],L):(f.set=!0,f.setByIndices=!0,f.indicesToOffset=!0,`set_${e}(${G}, ${L})`)},setByOffset:H,setByIndices:(k,L)=>a<2?H(k,L):(f.setByIndices=!0,f.indicesToOffset=!0,`set_${e}ByIndices(${k}, ${L});`),get:O,getByOffset:V,getByIndices:B,usage:n,name:e,strides:w,shape:g,rank:a}},z=(e,t,r,n=1)=>Hr(e,t,r,"input",n),N=(e,t,r,n=1)=>Hr(e,t,r,"output",n),hs=(e,t,r,n=1)=>Hr(e,t,r,"internal",n),Ra=class{constructor(e,t){this.normalizedDispatchGroup=e,this.limits=t,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(e){return`if (global_idx >= ${typeof e=="number"?`${e}u`:e}) { return; }`}mainStart(e=Bt){let t=typeof e=="number"?e:e[0],r=typeof e=="number"?1:e[1],n=typeof e=="number"?1:e[2];if(t>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||n>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${t}, ${r}, ${n}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(t*r*n>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${t}, ${r}, ${n}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let s=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,i=s?`@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`,a=s?"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) * ${t*r*n}u + local_idx;`;return`@compute @workgroup_size(${t}, ${r}, ${n}) + fn main(${i}) { + ${a} + `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith("uniforms.")&&this.uniforms.push({name:e.shape.replace("uniforms.",""),type:"u32",length:e.rank}),e.strides.startsWith("uniforms.")&&this.uniforms.push({name:e.strides.replace("uniforms.",""),type:"u32",length:e.rank}))}declareVariable(e,t){if(e.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(e),this.appendVariableUniforms(e);let r=e.usage==="input"?"read":"read_write",n=e.type.storage;return`@group(0) @binding(${t}) var ${e.name}: array<${n}>;`}declareVariables(...e){return e.map(t=>this.declareVariable(t,this.variableIndex++)).join(` +`)}registerInternalVariable(e){if(e.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(e),this.appendVariableUniforms(e)}registerInternalVariables(...e){return e.forEach(t=>this.registerInternalVariable(t)),this}registerUniform(e,t,r=1){return this.uniforms.push({name:e,type:t,length:r}),this}registerUniforms(e){return this.uniforms=this.uniforms.concat(e),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let e=[];for(let{name:t,type:r,length:n}of this.uniforms)if(n&&n>4)r==="f16"?e.push(`@align(16) ${t}:array, ${Math.ceil(n/8)}>`):e.push(`${t}:array, ${Math.ceil(n/4)}>`);else{let s=n==null||n===1?r:`vec${n}<${r}>`;e.push(`${t}:${s}`)}return` + struct Uniforms { ${e.join(", ")} }; + @group(0) @binding(${this.variableIndex}) var uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(e=>e.impl()).join(` +`)+this.internalVariables.map(e=>e.impl()).join(` +`)}get variablesInfo(){if(this.uniforms.length===0)return;let e=t=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(t)];return this.uniforms.map(t=>[e(t.type),t.length??1])}},Da=(e,t)=>new Ra(e,t),Xt=(e,t)=>{let r=e.length,n=[];for(let s=0;s1&&a===1&&n.unshift(i)}return n}}),La,fs,Fa,Na,Ze,Ua,qa,Dt=A(()=>{j(),J(),_e(),Y(),La=e=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.")},fs=(e,t)=>t&&t.length!==e?[...new Array(e).keys()].reverse():t,Fa=(e,t)=>C.sortBasedOnPerm(e,fs(e.length,t)),Na=(e,t,r,n)=>{let s=[];s.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} { + var a: ${r.type.indices};`);for(let i=0;i{let r=e.dataType,n=e.dims.length,s=fs(n,t),i=Fa(e.dims,s),a=N("output",r,i.length),o=z("a",r,n),l=u=>` + ${u.registerUniform("output_size","u32").declareVariables(o,a)} + + ${Na(s,n,o,a)} + + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${a.offsetToIndices("global_idx")}; + let aIndices = perm(indices); + + ${a.setByOffset("global_idx",o.getByIndices("aIndices"))} + }`;return{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:u=>{let d=C.size(i);return{outputs:[{dims:i,dataType:u[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:[{type:12,data:d},...F(u[0].dims,i)]}},getShaderSource:l}},Ua=(e,t)=>{La(e.inputs),e.compute(Ze(e.inputs[0],t.perm))},qa=e=>ae({perm:e.perm})}),Ga,Va,ja,Wa,Ha,Ka,Qa,Xa,Ya,Za,Ge,Ja,eo,to,ro,no,so,io,ao,oo,lo,Vf=A(()=>{j(),J(),Y(),gs(),Dt(),Ga={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"},Va={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"},ja={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Wa={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Ha=(e,t)=>{let r=[];for(let n=t-e;n{let r=[],n=e.length;for(let i=0;ie[i]);return[r,s]},Qa=(e,t)=>{let r=e.length+t.length,n=[],s=0;for(let i=0;i{for(let r=0;r{let r=[];if(!Xa(e,t)){for(let n=0;nr.push(n))}return r},Za=(e,t,r,n,s,i,a)=>{let o=r[0].dims,l=C.size(i),u=C.size(a),d=z("_A",r[0].dataType,o),c=N("output",s,i),p=32,h=` + var aBestValues : array; + `;return{name:e,shaderCache:t,getShaderSource:f=>` + ${f.registerUniform("reduceSize","u32").declareVariables(d,c)} + ${h} + fn DIV_CEIL(a : u32, b : u32) -> u32 { + return ((a - 1u) / b + 1u); + } + ${f.mainStart(p)} + + let outputIndex = global_idx / ${p}; + let offset = outputIndex * uniforms.reduceSize; + + var bestValue = f32(${ja[n]}); + let Length = uniforms.reduceSize; + for (var k = local_idx; k < Length; k = k + ${p}) { + let candidate = f32(${d.getByOffset("offset + k")}); + bestValue = ${Ga[n]}; + } + aBestValues[local_idx] = bestValue; + workgroupBarrier(); + + var reduceSize = min(Length, ${p}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 = ${Va[n]}; + aBestValues[local_idx] = bestValue; + } + reduceSize = interval; + workgroupBarrier(); + } + + if (local_idx == 0u) { + ${c.setByOffset("outputIndex",`${n==="mean"?`${c.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${c.type.storage}(${Wa[n]})`}`)}; + } + }`,getRunData:()=>({outputs:[{dims:i,dataType:s}],dispatchGroup:{x:l},programUniforms:[{type:12,data:u}]})}},Ge=(e,t,r,n)=>{let s=e.inputs.length===1?r:ms(e.inputs,r),i=s.axes;i.length===0&&!s.noopWithEmptyAxes&&(i=e.inputs[0].dims.map((h,f)=>f));let a=C.normalizeAxes(i,e.inputs[0].dims.length),o=a,l=e.inputs[0],u=Ya(o,e.inputs[0].dims.length);u.length>0&&(l=e.compute(Ze(e.inputs[0],u),{inputs:[0],outputs:[-1]})[0],o=Ha(o.length,l.dims.length));let[d,c]=Ka(l.dims,o),p=d;s.keepDims&&(p=Qa(d,a)),e.compute(Za(t,{hint:s.cacheKey,inputDependencies:["type"]},[l],n,e.inputs[0].dataType,p,c),{inputs:[l]})},Ja=(e,t)=>{Ge(e,"ReduceMeanShared",t,"mean")},eo=(e,t)=>{Ge(e,"ReduceL1Shared",t,"l1")},to=(e,t)=>{Ge(e,"ReduceL2Shared",t,"l2")},ro=(e,t)=>{Ge(e,"ReduceLogSumExpShared",t,"logSumExp")},no=(e,t)=>{Ge(e,"ReduceMaxShared",t,"max")},so=(e,t)=>{Ge(e,"ReduceMinShared",t,"min")},io=(e,t)=>{Ge(e,"ReduceProdShared",t,"prod")},ao=(e,t)=>{Ge(e,"ReduceSumShared",t,"sum")},oo=(e,t)=>{Ge(e,"ReduceSumSquareShared",t,"sumSquare")},lo=(e,t)=>{Ge(e,"ReduceLogSumShared",t,"logSum")}}),Ve,uo,Kr,ms,je,co,po,ho,fo,mo,go,_o,wo,yo,bo,We,vo,$o,xo,ko,So,Eo,To,Io,Mo,Co,gs=A(()=>{j(),J(),_e(),Y(),Vf(),Ve=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.")},uo=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Kr=(e,t,r,n,s,i,a=!1,o=!1)=>{let l=[],u=r[0].dims,d=u.length,c=C.normalizeAxes(s,d),p=!o&&c.length===0;u.forEach((m,g)=>{p||c.indexOf(g)>=0?a&&l.push(1):l.push(m)});let h=l.length,f=C.size(l);return{name:e,shaderCache:t,getShaderSource:m=>{let g=[],w=z("_A",r[0].dataType,d),_=N("output",i,h),v=n(w,_,c),b=v[2];for(let y=0,S=0;y=0?(a&&S++,b=`for(var j${y}: u32 = 0; j${y} < ${u[y]}; j${y}++) { + ${v[2].includes("last_index")?`let last_index = j${y};`:""} + ${w.indicesSet("input_indices",y,`j${y}`)} + ${b} + }`):(g.push(`${w.indicesSet("input_indices",y,_.indicesGet("output_indices",S))};`),S++);return` + + ${m.registerUniform("output_size","u32").declareVariables(w,_)} + + ${m.mainStart()} + ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + var input_indices: ${w.type.indices}; + let output_indices = ${_.offsetToIndices("global_idx")}; + + ${g.join(` +`)} + ${v[0]} // init ops for reduce max/min + ${v[1]} + ${b} + ${v[3]} + ${v.length===4?_.setByOffset("global_idx","value"):v.slice(4).join(` +`)} + }`},getRunData:()=>({outputs:[{dims:l,dataType:i}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:[{type:12,data:f},...F(u,l)]})}},ms=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),ae({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},je=(e,t,r,n)=>{let s=e.inputs,i=s.length===1?r:ms(s,r);e.compute(Kr(t,{hint:i.cacheKey,inputDependencies:["rank"]},[s[0]],i.noopWithEmptyAxes&&i.axes.length===0?uo:n,i.axes,s[0].dataType,i.keepDims,i.noopWithEmptyAxes),{inputs:[0]})},co=(e,t)=>{Ve(e.inputs),je(e,"ReduceLogSum",t,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},po=(e,t)=>{Ve(e.inputs),je(e,"ReduceL1",t,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},ho=(e,t)=>{Ve(e.inputs),je(e,"ReduceL2",t,(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);"])},fo=(e,t)=>{Ve(e.inputs),je(e,"ReduceLogSumExp",t,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},mo=(e,t)=>{Ve(e.inputs),je(e,"ReduceMax",t,(r,n,s)=>{let i=[];for(let a=0;a=0||s.length===0)&&i.push(r.indicesSet("input_indices",a,0));return[`${i.join(` +`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},go=(e,t)=>{Ve(e.inputs),je(e,"ReduceMean",t,(r,n,s)=>{let i=1;for(let a=0;a=0||s.length===0)&&(i*=e.inputs[0].dims[a]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${n.type.value}(sum / ${i});`]})},_o=(e,t)=>{Ve(e.inputs),je(e,"ReduceMin",t,(r,n,s)=>{let i=[];for(let a=0;a=0||s.length===0)&&i.push(`input_indices[${a}] = 0;`);return[`${i.join(` +`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},wo=(e,t)=>{Ve(e.inputs),je(e,"ReduceProd",t,(r,n)=>[`var value = ${n.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},yo=(e,t)=>{Ve(e.inputs),je(e,"ReduceSum",t,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},bo=(e,t)=>{Ve(e.inputs),je(e,"ReduceSumSquare",t,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},We=(e,t,r)=>{if(t.length===0)return r;let n=1,s=1;for(let i=0;i1024},vo=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?go(e,t):Ja(e,t)},$o=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?po(e,t):eo(e,t)},xo=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ho(e,t):to(e,t)},ko=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?fo(e,t):ro(e,t)},So=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?mo(e,t):no(e,t)},Eo=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?_o(e,t):so(e,t)},To=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?wo(e,t):io(e,t)},Io=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?yo(e,t):ao(e,t)},Mo=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?bo(e,t):oo(e,t)},Co=(e,t)=>{We(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?co(e,t):lo(e,t)}}),_s,zo,Ao,ws,jf=A(()=>{j(),_e(),gs(),_s=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.")},zo=(e,t)=>{_s(e.inputs);let r=(n,s,i)=>{let a=[];for(let o=0;o=0||i.length===0)&&a.push(`input_indices[${o}] = 0;`);return[`${a.join(` +`)}`,`var value = ${n.getByIndices("input_indices")}; +var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) { + value = ${n.getByIndices("input_indices")}; + best_index = i32(last_index); + }`,"",s.setByOffset("global_idx","best_index")]};e.compute(Kr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},Ao=(e,t)=>{_s(e.inputs);let r=(n,s,i)=>{let a=[];for(let o=0;o=0||i.length===0)&&a.push(`input_indices[${o}] = 0;`);return[`${a.join(` +`)}`,`var value = ${n.getByIndices("input_indices")}; +var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) { + value = ${n.getByIndices("input_indices")}; + best_index = i32(last_index); + }`,"",s.setByOffset("global_idx","best_index")]};e.compute(Kr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ws=e=>ae(e)}),Po,Oo,Bo,Qr,Ro,Do,Lo=A(()=>{j(),J(),_e(),Y(),Po=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,n=e[r],s=n.dataType,i=n.dims.length;e.forEach((a,o)=>{if(o!==r){if(a.dataType!==s)throw new Error("input tensors should be one type");if(a.dims.length!==i)throw new Error("input tensors should have the same shape");a.dims.forEach((l,u)=>{if(u!==t&&l!==n.dims[u])throw new Error("non concat dimensions must match")})}})},Oo=(e,t)=>` + fn calculateInputIndex(index: u32) -> u32 { + let sizeInConcatAxis = array(${t}); + for (var i: u32 = 0u; i < ${e}; i += 1u ) { + if (index < sizeInConcatAxis[i]) { + return i; + } + } + return ${e}u; + }`,Bo=(e,t)=>{let r=e.length,n=[];for(let s=0;s{let s=C.size(r),i=new Array(e.length),a=new Array(e.length),o=0,l=[],u=[],d=[{type:12,data:s}];for(let m=0;m`uniforms.sizeInConcatAxis${m}`).join(","),f=m=>` + + ${(()=>{m.registerUniform("outputSize","u32");for(let g=0;g(${h}); + ${p} -= sizeInConcatAxis[inputIndex - 1u]; + } + + ${Bo(a,c)} + }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:d}),getShaderSource:f}},Ro=(e,t)=>{let r=e.inputs,n=r[0].dims,s=C.normalizeAxis(t.axis,n.length);Po(r,s);let i=n.slice();i[s]=r.reduce((o,l)=>o+(l.dims.length>s?l.dims[s]:0),0);let a=r.filter(o=>C.size(o.dims)>0);e.compute(Qr(a,s,i,r[0].dataType),{inputs:a})},Do=e=>ae({axis:e.axis})}),Fo,No,Uo,qo,Yt,Go,Vo,ys=A(()=>{j(),os(),Y(),Lo(),Fo=(e,t)=>{let r=e[0],n=e[1],s=e[2],i=e[3],a=e[4],o=e[5];if(a&&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],d=r.dims[2];if(s.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]!==d)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(s.dims[0]!==n.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let c=s.dims[0]/3,p=c,h=p;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let v of t.qkvHiddenSizes)if(v%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");c=t.qkvHiddenSizes[0],p=t.qkvHiddenSizes[1],h=t.qkvHiddenSizes[2]}let f=u;if(c!==p)throw new Error("qkv_hidden_sizes first element should be same as the second");if(s.dims[0]!==c+p+h)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let m=0;if(a){if(p!==h)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(a.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(a.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(a.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(a.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(a.dims[4]!==p/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(m=a.dims[3])}let g=f+m,w=-1,_=0;if(i)throw new Error("Mask not supported");if(a)throw new Error("past is not supported");return{batchSize:l,sequenceLength:u,pastSequenceLength:m,kvSequenceLength:f,totalSequenceLength:g,maxSequenceLength:w,inputHiddenSize:d,hiddenSize:c,vHiddenSize:h,headSize:Math.floor(c/t.numHeads),vHeadSize:Math.floor(h/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:_,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},No=(e,t,r,n)=>{let s=ge(n),i=64,a=n/s;a{let h=N("x",t.dataType,t.dims,s),f=[{name:"d_inv",type:Te(t.dataType)},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return` + var thread_max: array; + var thread_sum: array; + ${p.registerUniforms(f).declareVariables(h)} + ${p.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 = ${d}(-3.402823e+38f); + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + thread_max_vector = max(${d}(x[offset + i]), thread_max_vector); + } + thread_max[local_idx] = ${(()=>{switch(s){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: ${s}`)}})()}; + 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 = ${d}(0); + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + sum_vector += exp(${d}(x[offset + i]) - max_value); + } + thread_sum[local_idx] = ${(()=>{switch(s){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: ${s}`)}})()}; + 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] = ${h.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 = ${d}(x[offset + i]); + x[offset + i] = ${h.type.value}(exp(f32input - max_value) / sum); + } + } + }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${i};${u};${s}`},getShaderSource:c,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:l})}},Uo=(e,t,r,n,s,i,a)=>{let o=a+s.kvSequenceLength,l=[s.batchSize,s.numHeads,s.sequenceLength,o],u=i.scale===0?1/Math.sqrt(s.headSize):i.scale,d=ge(s.headSize),c=s.headSize/d,p=12,h={x:Math.ceil(o/p),y:Math.ceil(s.sequenceLength/p),z:s.batchSize*s.numHeads},f=[{type:12,data:s.sequenceLength},{type:12,data:c},{type:12,data:o},{type:12,data:s.numHeads},{type:1,data:u}],m=n?["type","type","type"]:["type","type"],g=w=>{let _=z("q",t.dataType,t.dims,d),v=z("key",r.dataType,r.dims,d),b=[_,v];n&&b.push(z("relative_position_bias",n.dataType,n.dims));let y=N("output",t.dataType,l),S=Te(1,d),x=[{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 = ${p}u; + + var tileQ: array<${_.type.storage}, ${p*p}>; + var tileK: array<${_.type.storage}, ${p*p}>; + ${w.registerUniforms(x).declareVariables(...b,y)} + ${w.mainStart([p,p,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 = ${S}(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 += ${S}(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(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}`)}})()}; + output[outputIdx] = ${y.type.value} (sum * uniforms.alpha) + ${n?"relative_position_bias[outputIdx]":"0.0"}; + } + }`};return{name:"AttentionProbs",shaderCache:{hint:`${d}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:l,dataType:t.dataType,gpuDataType:0}],dispatchGroup:h,programUniforms:f}),getShaderSource:g}},qo=(e,t,r,n,s)=>{let i=s+n.kvSequenceLength,a=n.nReps?n.nReps:1,o=n.vHiddenSize*a,l=[n.batchSize,n.sequenceLength,o],u=12,d={x:Math.ceil(n.vHeadSize/u),y:Math.ceil(n.sequenceLength/u),z:n.batchSize*n.numHeads},c=[{type:12,data:n.sequenceLength},{type:12,data:i},{type:12,data:n.vHeadSize},{type:12,data:n.numHeads},{type:12,data:o}];return{name:"AttentionScore",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:l,dataType:t.dataType,gpuDataType:0}],dispatchGroup:d,programUniforms:c}),getShaderSource:p=>{let h=z("probs",t.dataType,t.dims),f=z("v",r.dataType,r.dims),m=N("output",t.dataType,l),g=[{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 = ${u}u; + var tileQ: array<${h.type.value}, ${u*u}>; + var tileK: array<${h.type.value}, ${u*u}>; + ${p.registerUniforms(g).declareVariables(h,f,m)} + ${p.mainStart([u,u,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; + } + }`}}},Yt=(e,t,r,n,s,i,a,o,l,u,d)=>{let c=e.outputCount>1,p=e.outputCount>2,h=u.kvNumHeads!=null||c&&p?u.pastSequenceLength:0,f=h+u.kvSequenceLength,m=[u.batchSize,u.numHeads,f,u.headSize],g=a?[a,r]:[r],w=u.kvNumHeads==null&&c?e.compute(Qr(g,2,m,r.dataType),{inputs:g,outputs:[1]})[0]:r,_=[u.batchSize,u.numHeads,f,u.headSize],v=o?[o,n]:[n],b=u.kvNumHeads==null&&p?e.compute(Qr(v,2,_,n.dataType),{inputs:v,outputs:[2]})[0]:n,y=[t,w];l&&y.push(l);let S=e.compute(Uo(e,t,w,l,u,d,h),{inputs:y,outputs:[-1]})[0];e.compute(No(e,S,u.batchSize*u.numHeads*u.sequenceLength,f),{inputs:[S],outputs:[]});let x=[S,b];e.compute(qo(e,S,b,u,h),{inputs:x,outputs:[0]})},Go=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],n=t.sequenceLength,s=t.inputHiddenSize,i=t.headSize,a=12,o={x:Math.ceil(t.headSize/a),y:Math.ceil(t.sequenceLength/a),z:t.batchSize*t.numHeads},l=[e.inputs[0],e.inputs[1],e.inputs[2]],u=[{type:12,data:n},{type:12,data:s},{type:12,data:i},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],d=c=>{let p=N("output_q",l[0].dataType,r),h=N("output_k",l[0].dataType,r),f=N("output_v",l[0].dataType,r),m=z("input",l[0].dataType,l[0].dims),g=z("weight",l[1].dataType,l[1].dims),w=z("bias",l[2].dataType,l[2].dims),_=m.type.storage,v=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return` + const TILE_SIZE = ${a}u; + var tileInput: array<${_}, ${a*a}>; + var tileWeightQ: array<${_}, ${a*a}>; + var tileWeightK: array<${_}, ${a*a}>; + var tileWeightV: array<${_}, ${a*a}>; + ${c.registerUniforms(v).declareVariables(m,g,w,p,h,f)} + ${c.mainStart([a,a,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 = ${_}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + let offset = n + (w + local_id.y) * uniforms.ldb; + tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset]; + tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset]; + tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset]; + } + workgroupBarrier(); + for (var k: u32 = 0u; k({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:o,programUniforms:u}),getShaderSource:d},{inputs:l,outputs:[-1,-1,-1]})},Vo=(e,t)=>{let r=Fo(e.inputs,t),[n,s,i]=Go(e,r);return Yt(e,n,s,i,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t)}}),jo,Wo,Ho,Ko,Wf=A(()=>{Ue(),j(),J(),_e(),Y(),jo=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(n,s,i)=>{let a=s.length;if(a!==n.length)throw new Error(`${i}: num dimensions != ${a}`);s.forEach((o,l)=>{if(o!==n[l])throw new Error(`${i}: dim[${l}] do not match`)})};if(e[0].dims.length>1){let n=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,n,"Invalid input scale"),r(e[2].dims,n,"Invalid input B"),r(e[3].dims,n,"Invalid input mean"),r(e[4].dims,n,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},Wo=(e,t)=>{let{epsilon:r,spatial:n,format:s}=t,i=e[0].dims,a=n?ge(i[i.length-1]):1,o=s==="NHWC"&&i.length>1?a:1,l=C.size(i)/a,u=n,d=u?i.length:i,c=z("x",e[0].dataType,e[0].dims,a),p=z("scale",e[1].dataType,e[1].dims,o),h=z("bias",e[2].dataType,e[2].dims,o),f=z("inputMean",e[3].dataType,e[3].dims,o),m=z("inputVar",e[4].dataType,e[4].dims,o),g=N("y",e[0].dataType,d,a),w=()=>{let v="";if(n)v=`let cOffset = ${i.length===1?"0u":s==="NHWC"?`outputIndices[${i.length-1}] / ${a}`:"outputIndices[1]"};`;else if(s==="NCHW")v=` + ${g.indicesSet("outputIndices","0","0")} + let cOffset = ${g.indicesToOffset("outputIndices")};`;else{v=`var cIndices = ${p.type.indices}(0); + cIndices[0] = outputIndices[${i.length-1}];`;for(let b=1;b` + const epsilon = ${r}; + ${v.registerUniform("outputSize","u32").declareVariables(c,p,h,f,m,g)} + ${v.mainStart()} + ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var outputIndices = ${g.offsetToIndices(`global_idx * ${a}`)}; + ${w()} + let scale = ${p.getByOffset("cOffset")}; + let bias = ${h.getByOffset("cOffset")}; + let inputMean = ${f.getByOffset("cOffset")}; + let inputVar = ${m.getByOffset("cOffset")}; + let x = ${c.getByOffset("global_idx")}; + let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias; + ${g.setByOffset("global_idx","value")} + }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${n}_${a}`,inputDependencies:u?["rank","type","type","type","type"]:void 0},getShaderSource:_,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u?[{type:12,data:l},...F(i)]:[{type:12,data:l}]})}},Ho=e=>ae(e),Ko=(e,t)=>{let{inputs:r,outputCount:n}=e,s=Ho({...t,outputCount:n});if(se.webgpu.validateInputContent&&jo(r,s),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(Wo(r,s))}}),Qo,Xo,Yo,Hf=A(()=>{J(),Y(),Qo=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")},Xo=e=>{let t=e[0].dims,r=e[0].dims[2],n=C.size(t)/4,s=e[0].dataType,i=z("input",s,t,4),a=z("bias",s,[r],4),o=z("residual",s,t,4),l=N("output",s,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)}}),getShaderSource:u=>` + const channels = ${r}u / 4; + ${u.declareVariables(i,a,o,l)} + + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes(n)} + let value = ${i.getByOffset("global_idx")} + + ${a.getByOffset("global_idx % channels")} + ${o.getByOffset("global_idx")}; + ${l.setByOffset("global_idx","value")} + }`}},Yo=e=>{Qo(e.inputs),e.compute(Xo(e.inputs))}}),Zo,ne,Jo,el,tl,rl,nl,sl,il,al,ol,ll,ul,dl,cl,pl,Xr,hl,Yr,fl,ml,gl,_l,wl,yl,bl,vl,$l,xl,kl,Sl,El,Tl,Il,Ml,bs,Cl,vs,$s,zl,Al,Pl,xs=A(()=>{j(),J(),_e(),Y(),Zo=(e,t,r,n,s,i)=>{let a=Math.ceil(t/4),o="";typeof s=="string"?o=`${s}(a)`:o=s("a");let l=z("inputData",r,[a],4),u=N("outputData",n,[a],4);return` + ${e.registerUniform("vec_size","u32").declareVariables(l,u)} + + ${i??""} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + + let a = ${l.getByOffset("global_idx")}; + ${u.setByOffset("global_idx",o)} + }`},ne=(e,t,r,n,s,i=e.dataType)=>({name:t,shaderCache:{hint:s,inputDependencies:["type"]},getShaderSource:a=>Zo(a,C.size(e.dims),e.dataType,i,r,n),getRunData:a=>({outputs:[{dims:e.dims,dataType:i}],dispatchGroup:{x:Math.ceil(C.size(a[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(C.size(e.dims)/4)}]})}),Jo=e=>{e.compute(ne(e.inputs[0],"Abs","abs"))},el=e=>{e.compute(ne(e.inputs[0],"Acos","acos"))},tl=e=>{e.compute(ne(e.inputs[0],"Acosh","acosh"))},rl=e=>{e.compute(ne(e.inputs[0],"Asin","asin"))},nl=e=>{e.compute(ne(e.inputs[0],"Asinh","asinh"))},sl=e=>{e.compute(ne(e.inputs[0],"Atan","atan"))},il=e=>{e.compute(ne(e.inputs[0],"Atanh","atanh"))},al=e=>ae(e),ol=(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): ${t.to}`)}e.compute(ne(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},ll=e=>{let t=e.length>=2&&e[1].data!==0?e[1].getFloat32Array()[0]:cs,r=e.length>=3&&e[2].data!==0?e[2].getFloat32Array()[0]:ps;return ae({min:t,max:r})},ul=(e,t)=>{let r=e.inputs.length===1?t:ll(e.inputs),n=Te(e.inputs[0].dataType);e.compute(ne(e.inputs[0],"Clip",s=>`clamp(${s}, 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]})},dl=e=>{e.compute(ne(e.inputs[0],"Ceil","ceil"))},cl=e=>{e.compute(ne(e.inputs[0],"Cos","cos"))},pl=e=>{e.compute(ne(e.inputs[0],"Cosh","cosh"))},Xr=e=>ae(e),hl=(e,t)=>{let r=Te(e.inputs[0].dataType);e.compute(ne(e.inputs[0],"Elu",n=>`elu_vf32(${n})`,` + const elu_alpha_ = ${r}(${t.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)); + }`,t.cacheKey))},Yr=(e="f32")=>` +const r0: ${e} = 0.3275911; +const r1: ${e} = 0.254829592; +const r2: ${e} = -0.284496736; +const r3: ${e} = 1.421413741; +const r4: ${e} = -1.453152027; +const r5: ${e} = 1.061405429; + +fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> { + let absv = abs(v); + let x = 1.0 / (1.0 + r0 * absv); + return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv)); +}`,fl=e=>{let t=Te(e.inputs[0].dataType);e.compute(ne(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,Yr(t)))},ml=e=>{e.compute(ne(e.inputs[0],"Exp","exp"))},gl=e=>{e.compute(ne(e.inputs[0],"Floor","floor"))},_l=e=>{let t=Te(e.inputs[0].dataType);e.compute(ne(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Yr(t)))},wl=(e,t)=>{let r=Te(e.inputs[0].dataType);e.compute(ne(e.inputs[0],"LeakyRelu",n=>`select(leaky_relu_alpha_ * ${n}, ${n}, ${n} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},yl=e=>{e.compute(ne(e.inputs[0],"Not",t=>`!${t}`))},bl=e=>{e.compute(ne(e.inputs[0],"Neg",t=>`-${t}`))},vl=e=>{e.compute(ne(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},$l=e=>{let t=Te(e.inputs[0].dataType);e.compute(ne(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},xl=e=>{e.compute(ne(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},kl=e=>ae(e),Sl=(e,t)=>{let r=Te(e.inputs[0].dataType);e.compute(ne(e.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${n} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},El=e=>{e.compute(ne(e.inputs[0],"Sin","sin"))},Tl=e=>{e.compute(ne(e.inputs[0],"Sinh","sinh"))},Il=e=>{e.compute(ne(e.inputs[0],"Sqrt","sqrt"))},Ml=e=>{e.compute(ne(e.inputs[0],"Tan","tan"))},bs=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Cl=e=>{e.compute(ne(e.inputs[0],"Tanh",bs))},vs=(e="f32")=>` +const fast_gelu_a: ${e} = 0.5; +const fast_gelu_b: ${e} = 0.7978845608028654; +const fast_gelu_c: ${e} = 0.035677408136300125; + +fn tanh_v(v: vec4<${e}>) -> vec4<${e}> { + return ${bs("v")}; +} +`,$s=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,zl=e=>{let t=Te(e.inputs[0].dataType);e.compute(ne(e.inputs[0],"FastGelu",$s,vs(t),void 0,e.inputs[0].dataType))},Al=(e,t)=>{let r=Te(e.inputs[0].dataType);return e.compute(ne(e.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},Pl=e=>{e.compute(ne(e.inputs[0],"Log","log"))}}),Ol,Bl,Rl,Kf=A(()=>{J(),Y(),xs(),Ol=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")},Bl=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=z("input",e[0].dataType,e[0].dims,4),n=z("bias",e[0].dataType,[e[0].dims[2]],4),s=N("output",e[0].dataType,t,4),i=C.size(t)/4,a=ye(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:o=>` + const M_SQRT2 = sqrt(2.0); + const halfChannels = ${e[0].dims[2]/4/2}u; + + ${o.declareVariables(r,n,s)} + + ${Yr(a)} + + ${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); + + ${s.setByOffset("global_idx","valueLeft * geluRight")} + }`}},Rl=e=>{Ol(e.inputs),e.compute(Bl(e.inputs))}}),Dl,Ll,He,Fl,Nl,Ul,ql,Gl,Vl,jl,Wl,Hl,Kl,Qf=A(()=>{j(),J(),Y(),Dl=(e,t,r,n,s,i,a,o,l,u,d,c)=>{let p,h;typeof o=="string"?p=h=(_,v)=>`${o}((${_}),(${v}))`:typeof o=="function"?p=h=o:(p=o.scalar,h=o.vector);let f=N("outputData",d,n.length,4),m=z("aData",l,t.length,4),g=z("bData",u,r.length,4),w;if(s)if(i){let _=C.size(t)===1,v=C.size(r)===1,b=t.length>0&&t[t.length-1]%4===0,y=r.length>0&&r[r.length-1]%4===0;_||v?w=f.setByOffset("global_idx",h(_?`${m.type.value}(${m.getByOffset("0")}.x)`:m.getByOffset("global_idx"),v?`${g.type.value}(${g.getByOffset("0")}.x)`:g.getByOffset("global_idx"))):w=` + let outputIndices = ${f.offsetToIndices("global_idx * 4u")}; + let offsetA = ${m.broadcastedIndicesToOffset("outputIndices",f)}; + let offsetB = ${g.broadcastedIndicesToOffset("outputIndices",f)}; + ${f.setByOffset("global_idx",h(a||b?m.getByOffset("offsetA / 4u"):`${m.type.value}(${m.getByOffset("offsetA / 4u")}[offsetA % 4u])`,a||y?g.getByOffset("offsetB / 4u"):`${g.type.value}(${g.getByOffset("offsetB / 4u")}[offsetB % 4u])`))} + `}else w=f.setByOffset("global_idx",h(m.getByOffset("global_idx"),g.getByOffset("global_idx")));else{if(!i)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let _=(v,b,y="")=>{let S=`aData[indexA${b}][componentA${b}]`,x=`bData[indexB${b}][componentB${b}]`;return` + let outputIndices${b} = ${f.offsetToIndices(`global_idx * 4u + ${b}u`)}; + let offsetA${b} = ${m.broadcastedIndicesToOffset(`outputIndices${b}`,f)}; + let offsetB${b} = ${g.broadcastedIndicesToOffset(`outputIndices${b}`,f)}; + let indexA${b} = offsetA${b} / 4u; + let indexB${b} = offsetB${b} / 4u; + let componentA${b} = offsetA${b} % 4u; + let componentB${b} = offsetB${b} % 4u; + ${v}[${b}] = ${y}(${p(S,x)}); + `};d===9?w=` + 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));`:w=` + ${_("outputData[global_idx]",0)} + ${_("outputData[global_idx]",1)} + ${_("outputData[global_idx]",2)} + ${_("outputData[global_idx]",3)} + `}return` + ${e.registerUniform("vec_size","u32").declareVariables(m,g,f)} + + ${c??""} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${w} + }`},Ll=(e,t,r,n,s,i,a=r.dataType)=>{let o=!C.areEqual(r.dims,n.dims),l=r.dims,u=C.size(r.dims),d=!1,c=!1,p=[o];if(o){let h=Ot.calcShape(r.dims,n.dims,!1);if(!h)throw new Error("Can't perform binary op on the given tensors");l=h,u=C.size(l);let f=C.size(r.dims)===1,m=C.size(n.dims)===1,g=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,w=n.dims.length>0&&n.dims[n.dims.length-1]%4===0;p.push(f),p.push(m),p.push(g),p.push(w);let _=1;for(let v=1;vh.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:h=>Dl(h,r.dims,n.dims,l,d,o,c,s,r.dataType,n.dataType,a,i),getRunData:()=>({outputs:[{dims:l,dataType:a}],dispatchGroup:{x:Math.ceil(u/64/4)},programUniforms:[{type:12,data:Math.ceil(C.size(l)/4)},...F(r.dims,n.dims,l)]})}},He=(e,t,r,n,s,i)=>{e.compute(Ll(t,s??"",e.inputs[0],e.inputs[1],r,n,i))},Fl=e=>{He(e,"Add",(t,r)=>`${t}+${r}`)},Nl=e=>{He(e,"Div",(t,r)=>`${t}/${r}`)},Ul=e=>{He(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4(${t}==${r})`},void 0,void 0,9)},ql=e=>{He(e,"Mul",(t,r)=>`${t}*${r}`)},Gl=e=>{let t=z("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;He(e,"Pow",{scalar:(r,n)=>`pow_custom(${r},${n})`,vector:(r,n)=>`pow_vector_custom(${r},${n})`},` + fn pow_custom(a : ${t}, b : ${t}) -> ${t} { + if (b == ${t}(0.0)) { + return ${t}(1.0); + } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) { + return ${t}(pow(f32(a), f32(b))); // NaN + } + 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)))); + } + fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> { + // TODO: implement vectorized pow + return vec4<${t}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w)); + } + `)},Vl=e=>{He(e,"Sub",(t,r)=>`${t}-${r}`)},jl=e=>{He(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4(${t}>${r})`},void 0,void 0,9)},Wl=e=>{He(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4(${t}<${r})`},void 0,void 0,9)},Hl=e=>{He(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4(${t}>=${r})`},void 0,void 0,9)},Kl=e=>{He(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4(${t}<=${r})`},void 0,void 0,9)}}),xt,kt,St,ks,Et=A(()=>{j(),J(),xt=(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 ${e.activation}`)}},kt=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},St=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},ks=e=>{let t=(e==null?void 0:e.activation)||"";if(t==="HardSigmoid"){let[r,n]=(e==null?void 0:e.activation_params)||[.2,.5];return{activation:t,alpha:r,beta:n}}else if(t==="Clip"){let[r,n]=(e==null?void 0:e.activation_params)||[cs,ps];return{activation:t,clipMax:n,clipMin:r}}else if(t==="LeakyRelu"){let[r]=(e==null?void 0:e.activation_params)||[.01];return{activation:t,alpha:r}}return{activation:t}}}),xe,Ss,Es=A(()=>{xe=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Ss=e=>` + ${e?"value = value + getBiasByOutputCoords(coords);":""} + `}),Ts,Ql=A(()=>{Ts=e=>` +fn getIndexFromCoords4D(coords : vec4, shape : vec4) -> i32 { + return dot(coords, vec4( + shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1)); +} +fn getOutputIndexFromCoords(coords : vec4) -> i32 { + return dot(coords, vec4( + i32(${e}.x), i32(${e}.y), i32(${e}.z), 1)); +} +`}),Xl,Yl,Zr,Is,Zl,Jr,Jl,Ms,en=A(()=>{j(),J(),Y(),Et(),Es(),Xl=(e,t)=>e?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart / innerElementSize + inputCol${t?", batchIndices":""}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRow + innerRow, + kStart / innerElementSize + inputCol${t?", batchIndices":""}); + `,Yl=(e,t)=>e?` + let ACached0 = mm_Asub[k * innerElementSize][localRow]; + let ACached1 = mm_Asub[k * innerElementSize + 1][localRow]; + let ACached2 = mm_Asub[k * innerElementSize + 2][localRow]; + ${t===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]; + ${t===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]; + ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"} + }`,Zr=(e,t,r="f32",n,s=!1,i=32,a=!1,o=32)=>{let l=t[1]*e[1],u=t[0]*e[0],d=s?l:i,c=s?i:l,p=d/t[0],h=i/t[1];if(!((s&&p===4&&e[1]===4||!s&&(p===3||p===4))&&d%t[0]===0&&i%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${s} is true, innerElementSize ${p} and workPerThread[1] ${e[1]} must be 4. + Otherwise, innerElementSize ${p} must be 3 or 4. + tileAWidth ${d} must be divisible by workgroupSize[0]${t[0]}. tileInner ${i} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return` +var mm_Asub: array, ${d/p}>, ${c}>; +var mm_Bsub: array, ${u/e[0]}>, ${i}>; + +const rowPerThread = ${e[1]}; +const colPerThread = ${e[0]}; +const innerElementSize = ${p}; +const tileInner = ${i}; + +@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[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 = ${a?"0":"i32(globalId.z)"}; + ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""} + let globalRowStart = i32(workgroupId.y) * ${l}; + + let num_tiles = ${a?`${Math.ceil(o/i)}`:"(uniforms.dim_inner - 1) / tileInner + 1"}; + var kStart = ${a?`i32(globalId.z) * ${o}`:"0"}; + + var acc: array, rowPerThread>; + + // Loop over shared dimension. + let tileRowB = localRow * ${h}; + 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; + ${Xl(s,n)} + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${h}; innerRow = innerRow + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${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]; + ${p===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"} + + ${Yl(s,p)} + } + + workgroupBarrier(); + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]); + } +}`},Is=(e,t)=>e?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart + inputCol${t?", batchIndices":""}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRowStart + inputRow, + kStart + inputCol${t?", batchIndices":""}); + `,Zl=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",Jr=(e,t,r="f32",n,s=!1,i=32,a=!1,o=32,l=!1)=>{let u=e[1]*t[1],d=e[0]*t[0],c=s?u:i,p=s?i:u;if(!(p%t[1]===0&&c%t[0]===0&&i%t[1]===0))throw new Error(`tileAHight ${p} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${c} must be divisible by workgroupSize[0]${t[0]}, tileInner ${i} must be divisible by workgroupSize[1]${t[1]}`);let h=p/t[1],f=c/t[0],m=i/t[1],g=l?` + let localRow = i32(localId.y); + let localCol = i32(localId.x); + let globalRowStart = i32(workgroupId.y) * ${u}; + let globalColStart = i32(workgroupId.x) * ${d}; + + // 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 < ${p}; inputRow = inputRow + ${t[1]}) { + for (var inputCol = localCol; inputCol < ${c}; inputCol = inputCol + ${t[0]}) { + ${Is(s,n)} + } + } + // Load one tile of B into local memory. + for (var inputRow = localRow; inputRow < ${i}; inputRow = inputRow + ${t[1]}) { + for (var inputCol = localCol; inputCol < ${d}; inputCol = inputCol + ${t[0]}) { + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalColStart + inputCol${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 * ${t[0]}]; + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let ACached = ${s?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[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 * ${t[1]}; + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let gCol = globalColStart + localCol + innerCol * ${t[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) * ${h}; +let tileColA = i32(localId.x) * ${f}; +let tileRowB = i32(localId.y) * ${m}; +// 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 < ${h}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < ${f}; innerCol = innerCol + 1) { + let inputRow = tileRowA + innerRow; + let inputCol = tileColA + innerCol; + ${Is(s,n)} + } + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${m}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol + innerCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalCol + innerCol${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) { + ${Zl(s)} + 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, ${p}>; + var mm_Bsub : array, ${i}>; + const rowPerThread = ${e[1]}; + const colPerThread = ${e[0]}; + const tileInner = ${i}; + +@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let batch = ${a?"0":"i32(globalId.z)"}; + ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""} + let num_tiles = ${a?`${Math.ceil(o/i)}`:"(uniforms.dim_inner - 1) / tileInner + 1"}; + var kStart = ${a?`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; + } + } + ${g} + } +`},Jl=(e,t,r,n,s,i=!1)=>{let[a,o,l]=s,[u,d,c,p]=n,h=Xt(a,l),f=Xt(o,l),m=ye(n[0].type.tensor),g=()=>{let _=d.rank,v=u.rank,b=`var aIndices: ${d.type.indices};`;for(let y=_-2-1,S=v-1;y>=0;y--,S--)b+=` +aIndices[${y}] = ${v>1?`batchIndices[${S}]`:"batchIndices"};`;return h.forEach(y=>{b+=` +aIndices[${y}] = 0;`}),b+=` +aIndices[${_-2}] = u32(row); + aIndices[${_-1}] = u32(colIn);`,b},w=()=>{let _=c.rank,v=u.rank,b=`var bIndices: ${c.type.indices};`;for(let y=_-2-1,S=v-1;y>=0;y--,S--)b+=` +bIndices[${y}] = ${v>1?`batchIndices[${S}]`:"batchIndices"};`;return f.forEach(y=>{b+=` +bIndices[${y}] = 0;`}),b+=` +bIndices[${_-2}] = u32(row); + bIndices[${_-1}] = u32(colIn);`,b};return` + fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${u.type.indices}) -> ${xe(e,m)} { + var value = ${xe(e,m)}(0.0); + let col = colIn * ${e}; + if(row < uniforms.dim_a_outer && col < uniforms.dim_inner) + { + ${g()} + value = ${d.getByIndices("aIndices")}; + } + return value; + } + + fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${u.type.indices}) -> ${xe(e,m)} { + var value = ${xe(e,m)}(0.0); + let col = colIn * ${e}; + if(row < uniforms.dim_inner && col < uniforms.dim_b_outer) + { + ${w()} + value = ${c.getByIndices("bIndices")}; + } + return value; + } + + fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${xe(e,m)}) { + let col = colIn * ${e}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) { + var value = valueIn; + let coords = vec3(batch, row, colIn); + ${t?`value = value + ${i?"bias[colIn]":`${xe(e,m)}(bias[row])`};`:""} + ${r} + ${p.setByIndices("vec3(coords)","value")} + } + } + `},Ms=(e,t,r,n,s=!1)=>{let i=e[0].dims,a=e[1].dims,o=i.slice(0,-2),l=a.slice(0,-2),u=n?n.slice(0,-2):r.slice(0,-2),d=C.size(u),c=i[i.length-2],p=i[i.length-1],h=a[a.length-1],f=p%4===0&&h%4===0,m=c<=8?[4,1,1]:[4,4,1],g=[8,8,1],w=[Math.ceil(h/g[0]/m[0]),Math.ceil(c/g[1]/m[1]),Math.ceil(d/g[2]/m[2])],_=f?4:1,v=[...o,c,p/_],b=v.length,y=[...l,p,h/_],S=y.length,x=[d,c,h/_],I=[{type:6,data:c},{type:6,data:h},{type:6,data:p}];kt(t,I),I.push(...F(u,v,y));let D=["rank","rank"],R=e.length>2;R&&(I.push(...F(e[2].dims)),D.push("rank")),I.push(...F(x));let U=q=>{let H=u.length,V=hs("batchDims",e[0].dataType,H,1),E=ye(e[0].dataType),M=z("a",e[0].dataType,b,_),O=z("b",e[1].dataType,S,_),B=N("result",e[0].dataType,x.length,_),K=[M,O];if(R){let ie=s?_:1;K.push(z("bias",e[2].dataType,e[2].dims.length,ie))}let ee=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];St(t,ee);let k=ye(B.type.tensor),L=xt(t,B.type.value,k),G=Jl(_,R,L,[V,M,O,B],[o,l,u],s);return` + ${q.registerUniforms(ee).registerInternalVariables(V).declareVariables(...K,B)} + ${G} + ${f?Zr(m,g,E,V):Jr(m,g,E,V)} + `};return{name:"MatMul",shaderCache:{hint:`${m};${t.activation};${f};${s}`,inputDependencies:D},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:w[0],y:w[1],z:w[2]},programUniforms:I}),getShaderSource:U}}}),eu,tu,Xf=A(()=>{j(),vt(),Y(),Et(),Es(),Ql(),en(),eu=(e,t,r,n,s=!1,i,a=4,o=4,l=4,u="f32")=>{let d=D=>{switch(D){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${u}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${D} is not supported.`)}},c=D=>{switch(D){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 ${D} is not supported.`)}},p=e?` + let coord = vec4(batch, xRow, xCol, xCh); + `:` + let coord = vec4(batch, xCh, xRow, xCol); + `,h=e?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,f=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",m=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",g=e?"row":"col",w=e?"col":"row",_=` + let inChannels = i32(uniforms.w_shape[2]); + let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + let outRow = ${g} / outWidth; + let outCol = ${g} % outWidth; + + let WRow = ${w} / (i32(uniforms.w_shape[1]) * inChannels); + let WCol = ${w} / 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 = ${w} % inChannels; + var resData = ${xe(a,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 < ${f} && xCol >= 0 && xCol < ${m}) { + ${p} + let xIndex = getIndexFromCoords4D(coord, vec4(uniforms.x_shape)); + ${d(a)} + } + return resData;`,v=e?t&&n?` + let col = colIn * ${a}; + ${_}`:` + let col = colIn * ${a}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) { + ${_} + } + return ${xe(a,u)}(0.0);`:n&&r?` + let col = colIn * ${a}; + ${_}`:` + let col = colIn * ${a}; + if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { + ${_} + } + return ${xe(a,u)}(0.0);`,b=`${c(o)}`,y=xe(l,u),S=xe(e?a:o,u),x=xe(e?o:a,u),I=xt(i,y,u);return` + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${S} { + ${e?v:b} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${x} { + ${e?b:v} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${y}) { + let col = colIn * ${l}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) + { + var value = valueIn; + let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + ${h} + ${Ss(s)} + ${I} + setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value); + } + }`},tu=(e,t,r,n,s,i,a,o)=>{let l=t.format==="NHWC",u=l?e[0].dims[3]:e[0].dims[1],d=r[0],c=l?r[2]:r[3],p=l?r[1]:r[2],h=l?r[3]:r[1],f=l&&(u%4===0||u%3===0)&&h%4===0,m=l?h:c*p,g=l?c*p:h,w=[8,8,1],_=n<=8?[4,1,1]:[4,4,1],v=[Math.ceil(m/w[0]/_[0]),Math.ceil(g/w[1]/_[1]),Math.ceil(d/w[2]/_[2])];he("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${v}`);let b=f?l&&u%4!==0?3:4:1,y=w[1]*_[1],S=w[0]*_[0],x=Math.max(w[0]*b,w[1]),I=n%y===0,D=s%S===0,R=i%x===0,U=f?[b,4,4]:[1,1,1],q=[{type:6,data:n},{type:6,data:s},{type:6,data:i},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];kt(t,q),q.push(...F(e[0].dims,e[1].dims));let H=["rank","rank"];a&&(q.push(...F(e[2].dims)),H.push("rank")),q.push(...F(r));let V=E=>{let M=[{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}];St(t,M);let O=f?4:1,B=ye(e[0].dataType),K=` + fn setOutputAtIndex(flatIndex : i32, value : ${f?`vec4<${B}>`:B}) { + result[flatIndex] = ${f?`vec4<${B}>`:B}(value); + } + fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${f?`vec4<${B}>`:B}) { + let flatIndex = getOutputIndexFromCoords(vec4(d0, d1, d2, d3)); + setOutputAtIndex(flatIndex ${f?"/ 4":""}, value); + }`,ee=z("x",e[0].dataType,e[0].dims.length,b===3?1:b),k=z("w",e[1].dataType,e[1].dims.length,O),L=[ee,k],G=N("result",e[0].dataType,r.length,O);if(a){let ie=z("bias",e[2].dataType,e[2].dims.length,O);L.push(ie),K+=` + fn getBiasByOutputCoords(coords : vec4) -> ${f?`vec4<${B}>`:B} { + return bias[coords.${l?"w":"y"}${f?"/ 4":""}]; + }`}return` + ${Ts("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 }; + ${E.registerUniforms(M).declareVariables(...L,G)} + ${K} + ${eu(l,I,D,R,a,t,U[0],U[1],U[2],B)} + ${f?Zr(_,w,B,void 0,!l,x):Jr(_,w,B,void 0,!l,x,!1,void 0,o)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${b};${f};${I};${D};${R};${y};${S};${x}`,inputDependencies:H},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:v[0],y:v[1],z:v[2]},programUniforms:q}),getShaderSource:V}}}),Cs,ru,Yf=A(()=>{j(),J(),Y(),uu(),Et(),Cs=(e,t,r)=>{let n=e.length>2,s=n?"value += b[output_channel];":"",i=e[0].dims,a=e[1].dims,o=a[0]/t.group,l=t.format==="NHWC",u=tn(i,a,t.dilations,t.pads,t.strides,l),d=C.size(u),c=[{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:o}];kt(t,c),c.push(...F(i,a));let p=["rank","rank"];n&&(c.push(...F(e[2].dims)),p.push("rank")),c.push(...F(u));let h=f=>{let m=N("output",e[0].dataType,u.length),g=ye(m.type.tensor),w=xt(t,m.type.value,g),_=z("x",e[0].dataType,i.length),v=z("w",e[1].dataType,a.length),b=[_,v];n&&b.push(z("b",e[2].dataType,e[2].dims.length));let y=[{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 St(t,y),` + ${f.registerUniforms(y).declareVariables(...b,m)} + + ${f.mainStart()} + ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let outputIndices = ${m.offsetToIndices("global_idx")}; + let batch: u32 = outputIndices[0]; + let output_channel: u32 = outputIndices[${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: ${m.type.value} = ${m.type.value}(0); + for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) { + let input_channel = group_id * uniforms.w_shape[1] + wInChannel; + for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) { + let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0]; + + if (xHeight < 0u || xHeight >= uniforms.x_shape[${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 = ${v.get("output_channel","wInChannel","wHeight","wWidth")}; + value += xVal*wVal; + } + } + } + ${s} + ${w} + ${m.setByOffset("global_idx","value")} + }`};return{name:"GroupedConv",shaderCache:{hint:t.cacheKey,inputDependencies:p},getRunData:()=>({outputs:[{dims:r?r(u):u,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:c}),getShaderSource:h}},ru=(e,t,r)=>{let n=e.length>2,s=ge(r[3]),i=ge(r[2]),a=C.size(r)/s/i,o=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/s],l=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/s],u=[r[0],r[1],r[2],r[3]/s],d=[{type:12,data:a},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];kt(t,d),d.push(...F(o,l,u));let c=(i-1)*t.strides[1]+l[1],p=h=>{let f=N("output",e[0].dataType,u.length,s),m=ye(f.type.tensor),g=xt(t,f.type.value,m),w=z("x",e[0].dataType,o.length,s),_=z("w",e[1].dataType,l.length,s),v=[w,_];n&&v.push(z("b",e[2].dataType,e[2].dims,s));let b=n?"value += b[output_channel];":"",y=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return St(t,y),` + ${h.registerUniforms(y).declareVariables(...v,f)} + ${h.mainStart()} + ${h.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) * ${i}u; + index1 = index1 / width1; + let row = index1 % uniforms.output_shape[1]; + let batch = index1 / uniforms.output_shape[1]; + + let x_corner = vec2(i32(row), i32(col)) * uniforms.strides - uniforms.pads; + + var x_vals: array<${w.type.value}, ${c}>; + var values: array<${f.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 < ${c}; i++) { + let x_width = x_corner.y + i; + if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) { + x_vals[i] = ${w.get("batch","u32(x_height)","u32(x_width)","input_channel")}; + } else { + x_vals[i] = ${w.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]; + ${b} + ${g} + ${f.set("batch","row","col + i","output_channel","value")}; + } + }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${s};${i};${c};${l[0]};${l[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d}),getShaderSource:p}}}),zs,nu,su,iu=A(()=>{j(),J(),en(),Y(),Et(),zs=(e,t,r,n,s=!1)=>{let i=e[0].dims,a=e[1].dims,o=i[i.length-2],l=a[a.length-1],u=i[i.length-1],d=ge(l),c=ge(u),p=ge(o),h=C.size(r)/d/p,f=e.length>2,m=n?n.slice(0,-2):r.slice(0,-2),g=[C.size(m),o,l],w=[{type:12,data:h},{type:12,data:o},{type:12,data:l},{type:12,data:u}];kt(t,w),w.push(...F(m,i,a)),f&&w.push(...F(e[2].dims)),w.push(...F(g));let _=v=>{let b=hs("batch_dims",e[0].dataType,m.length),y=z("a",e[0].dataType,i.length,c),S=z("b",e[1].dataType,a.length,d),x=N("output",e[0].dataType,g.length,d),I=ye(x.type.tensor),D=xt(t,x.type.value,I),R=[y,S],U="";if(f){let K=s?d:1;R.push(z("bias",e[2].dataType,e[2].dims.length,K)),U=`${s?`value += bias[col / ${K}];`:`value += ${x.type.value}(bias[row + i]);`}`}let q=i.slice(0,-2),H=a.slice(0,-2),V=Xt(q,m),E=Xt(H,m),M=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];St(t,M);let O=(K,ee)=>{let k=K.rank,L=K.name;if(k===2)return`var ${L}_indices = ${K.type.indices}(0u, 0u);`;let G=b.rank,ie=`var ${L}_indices: ${K.type.indices};`;for(let ce=k-2-1,rt=G-1;ce>=0;ce--,rt--)ie+=` +${L}_indices[${ce}] = ${G>1?`batch_indices[${rt}]`:"batch_indices"};`;return ee.forEach(ce=>{ie+=` +${L}_indices[${ce}] = 0;`}),ie+=`${L}_indices[${k-2}] = 0u; + ${L}_indices[${k-1}] = 0u;`,ie},B=()=>{let K=`var a_data: ${y.type.value};`;for(let ee=0;ee; + for (var k: u32 = 0u; k < uniforms.K; k = k + ${c}) { + ${B()} + } + for (var i = 0u; i < ${p}u; i++) { + var value = values[i]; + ${U} + ${D} + let cur_indices = ${x.type.indices}(batch, row + i, col); + let offset = ${x.indicesToOffset("cur_indices")}; + ${x.setByOffset(`offset / ${d}`,"value")}; + } + } + `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${d};${c};${p};${s}`,inputDependencies:f?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:w}),getShaderSource:_}},nu=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.")},su=e=>{nu(e.inputs);let t=Ot.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can't use matmul on the given tensors");let r=t[t.length-1],n=e.inputs[0].dims[e.inputs[0].dims.length-1];r<8&&n<8?e.compute(zs(e.inputs,{activation:""},t)):e.compute(Ms(e.inputs,{activation:""},t))}}),tn,rn,au,As,Ps,ou,lu,Os,uu=A(()=>{J(),Xf(),en(),Yf(),Et(),iu(),Dt(),tn=(e,t,r,n,s,i)=>{let a=e[0],o=e.slice(i?1:2,i?3:4),l=o.length,u=t[0],d=t.slice(2).map((p,h)=>p+(p-1)*(r[h]-1)),c=o.map((p,h)=>p+n[h]+n[h+l]).map((p,h)=>Math.floor((p-d[h]+s[h])/s[h]));return c.splice(0,0,a),c.splice(i?3:1,0,u),c},rn=[2,3,1,0],au=(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],n=e[1].dims[1]*t.group;if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let s=e[0].dims.length-2;if(t.dilations.length!==s)throw new Error(`dilations should be ${s}D`);if(t.strides.length!==s)throw new Error(`strides should be ${s}D`);if(t.pads.length!==s*2)throw new Error(`pads should be ${s*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},As=(e,t)=>{let r=e.kernelShape.slice();for(let i=2;i{let t=ks(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],s=e.dilations,i=e.group,a=e.kernel_shape,o=e.pads,l=e.strides,u=e.w_is_const();return{autoPad:n,format:r,dilations:s,group:i,kernelShape:a,pads:o,strides:l,wIsConst:u,...t,cacheKey:`${e.format};${t.activation};`}},ou=(e,t,r)=>{let n=As(r,t),s=r.format==="NHWC";if(r.group!==1){if(!e.adapterInfo.isArchitecture("ampere")&&s&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let S=tn(t[0].dims,t[1].dims,r.dilations,n.pads,r.strides,s),x=e.kernelCustomData.wT??e.compute(Ze(t[1],rn),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=x);let I=[t[0],x];t.length===3&&I.push(t[2]),e.compute(ru(I,n,S),{inputs:I})}else e.compute(Cs(t,n));return}let i=t.length===3,a=t[0].dims[s?1:2],o=t[0].dims[s?2:3],l=t[0].dims[s?3:1],u=t[1].dims[2],d=t[1].dims[3],c=tn(t[0].dims,t[1].dims,r.dilations,n.pads,r.strides,s),p=c[s?1:2],h=c[s?2:3],f=c[s?3:1],m=s&&u===a&&d===o&&r.pads[0]===0&&r.pads[1]===0;if(m||u===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 S=c[0],x,I,D,R=[];if(s){let H=e.kernelCustomData.wT??e.compute(Ze(t[1],rn),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=H),m){let V=a*o*l;x=t[0].reshape([1,S,V]),I=H.reshape([1,V,f]),D=[1,S,f]}else x=t[0].reshape([S,a*o,l]),I=H.reshape([1,l,f]),D=[S,p*h,f];R.push(x),R.push(I)}else x=t[0].reshape([S,l,a*o]),I=t[1].reshape([1,f,l]),D=[S,f,p*h],R.push(I),R.push(x);i&&R.push(t[2]);let U=D[2],q=R[0].dims[R[0].dims.length-1];U<8&&q<8?e.compute(zs(R,n,c,D,s),{inputs:R}):e.compute(Ms(R,n,c,D,s),{inputs:R});return}let g=!0,w=e.kernelCustomData.wT??e.compute(Ze(t[1],rn),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=w);let _=[t[0],w];i&&_.push(t[2]);let v=s?p*h:f,b=s?f:p*h,y=u*d*l;e.compute(tu(_,n,c,v,b,y,i,g),{inputs:_})},lu=(e,t)=>{let r=t.format==="NHWC",n=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&n.push(e.inputs[2]);let s=[0,t.pads[0],0,t.pads[1]],i=[1].concat(t.strides),a=[1].concat(t.dilations),o=[1].concat(t.kernelShape),l=As({...t,pads:s,strides:i,dilations:a,kernelShape:o},n);e.compute(Cs(n,l,u=>r?[u[0],u[2],u[3]]:[]))},Os=(e,t)=>{au(e.inputs,t),e.inputs[0].dims.length===3?lu(e,t):ou(e,e.inputs,t)}}),du,cu,Zf=A(()=>{j(),vt(),Y(),Et(),Es(),Ql(),en(),du=(e,t=!1,r,n,s=4)=>{let i=g=>{switch(g){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 ${g} is not supported.`)}},a=e?` + let coord = vec4(batch, iXR, iXC, xCh); + `:` + let coord = vec4(batch, xCh, iXR, iXC); + `,o=e?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,l=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",u=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",d=e?"row":"col",c=e?"col":"row",p=` + let inChannels = ${e?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"}; + let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + let outRow = ${d} / outWidth; + let outCol = ${d} % outWidth; + + let WRow = ${c} / (uniforms.filter_dims[1] * inChannels); + let WCol = ${c} / inChannels % uniforms.filter_dims[1]; + let xR = f32(outRow - uniforms.pads[0] + uniforms.dilations[0] * WRow) / f32(uniforms.strides[0]); + let xC = f32(outCol - uniforms.pads[1] + uniforms.dilations[1] * WCol) / f32(uniforms.strides[1]); + if (xR < 0.0 || xR >= f32(${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 = ${c} % inChannels; + ${a} + return x[getIndexFromCoords4D(coord, vec4(uniforms.x_shape))/${s}];`,h=e?` + let col = colIn * ${s}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) { + ${p} + } + return ${n}(0.0);`:` + let col = colIn * ${s}; + if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { + ${p} + } + return ${n}(0.0);`,f=` + let col = colIn * ${s}; + let inChannels = ${e?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"}; + let coordX = uniforms.filter_dims[0] - 1 - row / (uniforms.filter_dims[1] * inChannels); + let coordY = uniforms.filter_dims[1] - 1 - (row / inChannels) % uniforms.filter_dims[1]; + if (${e?"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(s)} + } + return ${n}(0.0); + `,m=xt(r,n);return` + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${n} { + ${e?h:f} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${n} { + ${e?f:h} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${n}) { + let col = colIn * ${s}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) { + var value = valueInput; + let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + ${o} + ${Ss(t)} + ${m} + result[getIndexFromCoords4D(coords, vec4(uniforms.result_shape))/${s}] = value; + } + }`},cu=(e,t,r,n,s,i,a,o)=>{let l=t.format==="NHWC",u=l?e[0].dims[3]:e[0].dims[1],d=r[0],c=l?r[2]:r[3],p=l?r[1]:r[2],h=l?r[3]:r[1],f=l&&u%4===0&&u%3&&h%4===0,m=l?h:c*p,g=l?c*p:h,w=[8,8,1],_=n<=8?[4,1,1]:[4,4,1],v=[Math.ceil(m/w[0]/_[0]),Math.ceil(g/w[1]/_[1]),Math.ceil(d/w[2]/_[2])];he("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${v}`);let b=f?4:1,y=Math.max(w[0]*b,w[1]),S=f?4:1,x=[t.kernelShape[l?1:2],t.kernelShape[l?2:3]],I=[x[0]+(t.dilations[0]<=1?0:(x[0]-1)*(t.dilations[0]-1)),x[1]+(t.dilations[1]<=1?0:(x[1]-1)*(t.dilations[1]-1))],D=[I[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),I[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],R=[{type:6,data:n},{type:6,data:s},{type:6,data:i},{type:6,data:t.strides},{type:6,data:t.dilations},{type:6,data:x},{type:6,data:D}];kt(t,R),R.push(...F(e[0].dims,e[1].dims));let U=["rank","rank"];a&&(R.push(...F(e[2].dims)),U.push("rank")),R.push(...F(r));let q=H=>{let V=z("x",e[0].dataType,e[0].dims.length,S),E=z("w",e[1].dataType,e[1].dims.length,1),M=N("result",e[0].dataType,r.length,S),O=[V,E],B="";if(a){let k=z("bias",e[2].dataType,e[2].dims.length,S);O.push(k),B+=` + fn getBiasByOutputCoords(coords : vec4) -> ${k.type.value} { + return bias[coords.${l?"w":"y"}${f?"/ 4":""}]; + }`}let K=[{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:x.length},{name:"pads",type:"i32",length:D.length}];St(t,K);let ee=ye(e[0].dataType,1);if(ee!=="f16"&&ee!=="f32")throw new Error(`elemType ${ee} is not supported.`);return` + ${Ts("uniforms.result_strides")} + ${H.registerUniforms(K).declareVariables(...O,M)}; + ${B} + ${du(l,a,t,V.type.value,b)} + ${f?Zr(_,w,ee,void 0,!l,y):Jr(_,w,ee,void 0,!l,y,!1,void 0,o)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${t.cacheKey};${_};${w};${f}`,inputDependencies:U},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:v[0],y:v[1],z:v[2]},programUniforms:R}),getShaderSource:q}}}),pu,Bs,Jf=A(()=>{j(),vt(),J(),Y(),pu=(e,t,r,n,s,i=!1,a,o,l=!1)=>{let u=l?1:2,d=l?2:3,c=l?3:1,p=i?2:1,h=` + fn setOutputAtIndex(flatIndex : u32, value : ${i?`vec4<${a}>`:a}) { + result[flatIndex] = ${i?`vec4<${a}>`:a}(value); + }`;n&&(h+=` + fn getBiasByOutputCoords(coords : vec4) -> ${i?`vec4<${a}>`:a} { + return bias[coords.${l?"w":"y"}${i?"/ 4":""}]; + }`);let f=i?4:1,m=z("W",t[1].dataType,t[1].dims.length,f),g=z("Dy",t[0].dataType,t[0].dims.length,f),w=[g,m];n&&w.push(z("bias",t[2].dataType,[r[c]].length,f));let _=N("result",t[0].dataType,r.length,f),v=`{ + let batch: u32 = ${s?"global_id.z":"workgroup_id.z"} / uniforms.result_shape[1]; + let r = ${s?"global_id.z":"workgroup_id.z"} % uniforms.result_shape[1]; + let c = ${s?"global_id.y":"workgroup_id.y"} * ${p}; + let d1: u32 = ${s?"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, ${p}>; + for (var i = 0; i < ${p}; i++) { + dotProd[i] = vec4<${a}>(0.0); + } + for (var wR: u32 = 0; wR < uniforms.filter_dims[0]; wR = wR + 1) { + var dyR = (${a}(dyCorner.x) + ${a}(wR)) / ${a}(uniforms.strides.x); + let wRPerm = uniforms.filter_dims[0] - 1 - wR; + if (dyR < 0.0 || dyR >= ${a}(uniforms.Dy_shape[1]) || + fract(dyR) > 0.0 || wRPerm < 0) { + continue; + } + let idyR: u32 = u32(dyR); + + for (var wC: u32 = 0; wC < uniforms.filter_dims[1]; wC = wC + 1) { + let dyC = (${a}(dyCorner.y) + ${a}(wC)) / ${a}(uniforms.strides.y); + let dyC2 = (${a}(dyCorner.y) + 1.0 + ${a}(wC)) / ${a}(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 >= ${a}(uniforms.Dy_shape[2]) || + fract(dyC) > 0.0) { + bDyCVal = false; + } + if (dyC2 < 0.0 || dyC2 >= ${a}(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 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${g.get("batch","idyR","idyC","d2")}; + let tmpval = vec4<${a}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[0] = dotProd[0] + tmpval; + + xValue = ${g.get("batch","idyR","idyC2","d2")}; + + dotProd[1] = dotProd[1] + vec4<${a}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + } + } else if (bDyCVal) { + let d2Length = uniforms.Dy_shape[${c}]; + for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${g.get("batch","idyR","idyC","d2")}; + let tmpval = vec4<${a}>(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 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${m.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${g.get("batch","idyR","idyC2","d2")}; + let tmpval = vec4<${a}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[1] = dotProd[1] + tmpval; + } + } + } + } + + for (var i: u32 = 0; i < ${p}; i = i + 1) { + let value = dotProd[i] + ${n?"bias[c+i]":`vec4<${a}>(0.0)`}; + ${_.set("batch","r","c + i","d1","value")}; + } + }`,b=` + let outputIndices = ${_.offsetToIndices("global_idx")}; + let batch = ${_.indicesGet("outputIndices",0)}; + let d1 = ${_.indicesGet("outputIndices",c)}; + let r = ${_.indicesGet("outputIndices",u)}; + let c = ${_.indicesGet("outputIndices",d)}; + 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 = ${a}(0.0); + for (var wR: u32 = 0; wR < uniforms.effective_filter_dims.x; wR = wR + 1) { + if (wR % uniforms.dilations.x != 0) { + continue; + } + let dyR = (${a}(dyRCorner) + ${a}(wR)) / ${a}(uniforms.strides[0]); + let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x; + if (dyR < 0.0 || dyR >= ${a}(uniforms.Dy_shape[${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 = (${a}(dyCCorner) + ${a}(wC)) / ${a}(uniforms.strides.y); + let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y; + if (dyC < 0.0 || dyC >= ${a}(uniforms.Dy_shape[${d}]) || + 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?g.get("batch","idyR","idyC","inputChannel"):g.get("batch","inputChannel","idyR","idyC")}; + let wValue = ${m.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")}; + dotProd = dotProd + xValue * wValue; + inputChannel = inputChannel + 1; + } + } + } + let value = dotProd + ${n?"bias[d1]":`${a}(0.0)`}; + ${_.setByOffset("global_idx","value")}; + `;return` + ${e.registerUniforms(o).declareVariables(...w,_)} + ${h} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}; + ${i?v:b}}`},Bs=(e,t,r)=>{let n=e.length>2,s=t.outputShape,i=C.size(s),a=[Math.ceil(i/64),1,1];he("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${a}`);let o=t.format==="NHWC",l=["rank","rank"],u=[t.strides[0],t.strides[1]],d=[t.kernelShape[o?1:2],t.kernelShape[o?2:3]],c=[t.dilations[0],t.dilations[1]],p=[d[0]+(t.dilations[0]<=1?0:(t.kernelShape[o?1:2]-1)*(t.dilations[0]-1)),d[1]+(t.dilations[1]<=1?0:(t.kernelShape[o?2:3]-1)*(t.dilations[1]-1))],h=[p[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),p[1]-1-Math.floor(t.pads[1]+t.pads[3])/2],f=!1,m=t.group,g=e[1].dims,w=g[0]/m,_=g[1],v=[{type:12,data:i},{type:12,data:u},{type:12,data:d},{type:12,data:c},{type:12,data:p},{type:6,data:h},{type:12,data:w},{type:12,data:_},...F(e[0].dims,e[1].dims)];n&&(v.push(...F(e[2].dims)),l.push("rank")),v.push(...F(s));let b=a[1]===1&&a[2]===1,y=S=>{let x=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:u.length},{name:"filter_dims",type:"u32",length:d.length},{name:"dilations",type:"u32",length:d.length},{name:"effective_filter_dims",type:"u32",length:p.length},{name:"pads",type:"i32",length:h.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],I=ye(e[0].dataType);return`${pu(S,e,s,n,b,f,I,x,o)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};`,inputDependencies:l},getRunData:()=>({dispatchGroup:{x:a[0],y:a[1],z:a[2]},outputs:[{dims:r?r(s):s,dataType:e[0].dataType}],programUniforms:v}),getShaderSource:y}}}),hu,fu,mu,Rs,gu,_u,wu,yu,bu,vu,em=A(()=>{Zf(),Jf(),Et(),Dt(),hu=(e,t,r,n,s,i)=>(e-1)*t+r+(n-1)*s+1-i,fu=(e,t,r,n,s)=>{let i=Math.floor(e/2);t==="SAME_UPPER"?(r[n]=i,r[s]=e-i):t==="SAME_LOWER"&&(r[n]=e-i,r[s]=i)},mu=(e,t,r,n,s,i,a,o,l,u)=>{let d=e.length-2,c=u.length===0;if(l.length===0)for(let f=0;f{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((c,p)=>c*p,1)===0){r.length=0;for(let c=2;cc+p,0)===0){let c=t[0].dims.length-2;l=new Array(c).fill(1)}let u=e.strides.slice();if(u.reduce((c,p)=>c+p,0)===0){let c=t[0].dims.length-2;u=new Array(c).fill(1)}mu(o,r,l,e.autoPad,e.group,s,u,n,a,i);let d=Object.assign({},e);return Object.assign(d,{kernelShape:r,pads:s,outputPadding:a,outputShape:i,dilations:l,strides:u}),d},gu=e=>{let t=ks(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],s=e.dilations,i=e.group,a=e.kernelShape,o=e.pads,l=e.strides,u=e.wIsConst(),d=e.outputPadding,c=e.outputShape;return{autoPad:n,format:r,dilations:s,group:i,kernelShape:a,outputPadding:d,outputShape:c,pads:o,strides:l,wIsConst:u,...t,cacheKey:`${e.format};${t.activation};`}},_u=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],n=e[1].dims[0];if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let s=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==s))throw new Error("invalid bias");let i=e[0].dims.length-2;if(t.dilations.reduce((a,o)=>a+o,0)>0&&t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.reduce((a,o)=>a+o,0)>0&&t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.reduce((a,o)=>a+o,0)>0&&t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.outputPadding.length!==i&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${i}D`);if(t.kernelShape.reduce((a,o)=>a+o,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},wu=[2,3,1,0],yu=(e,t,r)=>{let n=Rs(r,t),s=r.format==="NHWC",i=n.outputShape,a=i[s?3:1],o=t[0].dims[s?3:1];if(n.group!==1||a===1&&o===1){e.compute(Bs(t,n));return}let l=i[s?1:2],u=i[s?2:3],d=t[1].dims[2],c=t[1].dims[3],p=s?l*u:a,h=s?a:l*u,f=d*c*o,m=!0,g=e.kernelCustomData.wT??e.compute(Ze(t[1],wu),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=g);let w=[t[0],g],_=t.length===3;_&&(!s&&t[2].dims.length===1?w.push(t[2].reshape([t[2].dims[0],1,1])):w.push(t[2])),e.compute(cu(w,n,i,p,h,f,_,m),{inputs:w})},bu=(e,t)=>{let r=t.format==="NHWC",n=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&n.push(e.inputs[2]);let s=t.kernelShape;(s.length===0||s[0]===0)&&(s=[e.inputs[1].dims[2]]);let i=t.dilations;(i.length===0||i[0]===0)&&(i=[1]);let a=t.strides;(a.length===0||a[0]===0)&&(a=[1]);let o=t.pads;o.length===0&&(o=[0,0]),o=[0,o[0],0,o[1]],a=[1].concat(a),i=[1].concat(i),s=[1].concat(s);let l=Rs({...t,pads:o,strides:a,dilations:i,kernelShape:s},n);e.compute(Bs(n,l,u=>r?[u[0],u[2],u[3]]:[u[0],u[1],u[3]]))},vu=(e,t)=>{_u(e.inputs,t),e.inputs[0].dims.length===3?bu(e,t):yu(e,e.inputs,t)}}),$u,xu,ku,tm=A(()=>{j(),J(),_e(),Y(),$u=(e,t,r,n)=>{let s=C.size(t),i=t.length,a=z("input",e,i),o=N("output",e,i),l=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),u=C.normalizeAxis(l,i),d=c=>{let p=` i32(${a.indicesGet("inputIndices","uniforms.axis")}) `,h=Q("uniforms.input_shape","uniforms.axis",i),f=n.reverse?p+(n.exclusive?" + 1":""):"0",m=n.reverse?h:p+(n.exclusive?"":" + 1");return` + ${c.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(a,o)} + ${c.mainStart()} + ${c.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var inputIndices = ${o.offsetToIndices("global_idx")}; + var sum = ${o.type.value}(0); + let first : i32 = ${f}; + let last : i32 = ${m}; + for (var i : i32 = first; i < last; i++) { + ${a.indicesSet("inputIndices","uniforms.axis","u32(i)")}; + sum = sum + ${a.getByIndices("inputIndices")}; + } + ${o.setByOffset("global_idx","sum")}; + }`};return{name:"CumSum",shaderCache:{hint:n.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:[{type:12,data:s},{type:12,data:u},...F(t,t)]}),getShaderSource:d}},xu=(e,t)=>{let r=e.inputs[0].dims,n=e.inputs[0].dataType,s=e.inputs[1];e.compute($u(n,r,s,t),{inputs:[0]})},ku=e=>{let t=e.exclusive===1,r=e.reverse===1;return ae({exclusive:t,reverse:r})}}),Su,Eu,Tu,Iu,Mu,rm=A(()=>{j(),J(),_e(),Y(),Su=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.")},Eu=(e,t,r,n)=>{let s=[];s.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} { + var a: ${r.type.indices};`);for(let i=0;i{let r,n,s,i,a,o,l=t.format==="NHWC",u=t.blocksize,d=t.mode==="DCR";l?([r,n,s,i]=e.dims,a=d?[r,n,s,u,u,i/u**2]:[r,n,s,i/u**2,u,u],o=d?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,n,s,i]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],a=d?[r,u,u,i/u**2,n,s]:[r,i/u**2,u,u,n,s],o=d?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let c=e.reshape(a),p=c.dims.length,h=e.dataType,f=z("a",h,p),m=N("output",h,p),g=w=>` + ${w.registerUniform("output_size","u32").declareVariables(f,m)} + + ${Eu(o,p,f,m)} + + ${w.mainStart()} + ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${m.offsetToIndices("global_idx")}; + let aIndices = perm(indices); + + ${m.setByOffset("global_idx",f.getByIndices("aIndices"))} + }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:w=>{let _=l?[r,n*u,s*u,i/u**2]:[r,i/u**2,n*u,s*u],v=C.size(_),b=c.dims,y=C.sortBasedOnPerm(b,o);return{outputs:[{dims:_,dataType:w[0].dataType}],dispatchGroup:{x:Math.ceil(v/64)},programUniforms:[{type:12,data:v},...F(b,y)]}},getShaderSource:g}},Iu=(e,t)=>{Su(e.inputs),e.compute(Tu(e.inputs[0],t))},Mu=e=>ae({blocksize:e.blocksize,mode:e.mode,format:e.format})}),nn,Zt,Ds,Cu,zu,Au,Pu,Ls,Ou,Bu,Ru,nm=A(()=>{j(),J(),_e(),Y(),nn="[a-zA-Z]|\\.\\.\\.",Zt="("+nn+")+",Ds="^"+Zt+"$",Cu="("+Zt+",)*"+Zt,zu="^"+Cu+"$",Au=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let r=this.symbolToIndices.get(e);r===void 0?r=[t]:r.push(t),this.symbolToIndices.set(e,r)}},Pu=class{constructor(e,t){var s;this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,n]=t.includes("->")?t.split("->",2):[t,""];if(!r.match(RegExp(zu)))throw new Error("Invalid LHS term");if(r.split(",").forEach((i,a)=>{let o=e[a].dims.slice();if(!i.match(RegExp(Ds)))throw new Error("Invalid LHS term");let l=this.processTerm(i,!0,o,a);this.lhs.push(l)}),n==="")n+=[...this.symbolToInfo.entries()].filter(([i,a])=>a.count===1||i==="...").map(([i])=>i).join("");else if(!n.match(RegExp(Zt)))throw new Error("Invalid RHS");(s=n.match(RegExp(nn,"g")))==null||s.forEach(i=>{if(i==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let a=this.symbolToInfo.get(i);if(a===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(a.dimValue)}}),this.rhs=this.processTerm(n,!1,this.outputDims)}addSymbol(e,t,r){let n=this.symbolToInfo.get(e);if(n!==void 0){if(n.dimValue!==t&&n.count!==1)throw new Error("Dimension mismatch");n.count++,n.inputIndices.push(r)}else n={count:1,dimValue:t,inputIndices:[r]};this.symbolToInfo.set(e,n)}processTerm(e,t,r,n=-1){let s=r.length,i=!1,a=[],o=0;if(!e.match(RegExp(Ds))&&!t&&e!=="")throw new Error("Invalid LHS term");let l=e.match(RegExp(nn,"g")),u=new Au(n);return l==null||l.forEach((d,c)=>{if(d==="..."){if(i)throw new Error("Only one ellipsis is allowed per input term");i=!0;let p=s-l.length+1;if(p<0)throw new Error("Ellipsis out of bounds");if(a=r.slice(o,o+p),this.hasEllipsis){if(this.ellipsisDims.length!==a.length||this.ellipsisDims.toString()!==a.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=a;else throw new Error("Ellipsis must be specified in the LHS");for(let h=0;he+"_max",Ou=(e,t,r,n)=>{let s=e.map(u=>u.length).map((u,d)=>z(`input${d}`,t,u)),i=C.size(n),a=N("output",t,n.length),o=[...r.symbolToInfo.keys()].filter(u=>!r.rhs.symbolToIndices.has(u)),l=u=>{let d=[],c="var prod = 1.0;",p="var sum = 0.0;",h="sum += prod;",f=[],m=[],g=[],w=[],_=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((b,y)=>{var S;if(r.rhs.symbolToIndices.has(y)){let x=(S=r.rhs.symbolToIndices.get(y))==null?void 0:S[0];x!==void 0&&r.lhs.forEach((I,D)=>{if(b.inputIndices.includes(D)){let R=I.symbolToIndices.get(y);if(R===void 0)throw new Error("Invalid symbol error");R.forEach(U=>{d.push(`${s[D].indicesSet(`input${D}Indices`,U,a.indicesGet("outputIndices",x))}`)})}})}else r.lhs.forEach((x,I)=>{if(b.inputIndices.includes(I)){let D=x.symbolToIndices.get(y);if(D===void 0)throw new Error("Invalid symbol error");D.forEach(R=>{f.push(`${s[I].indicesSet(`input${I}Indices`,R,`${y}`)}`)}),w.push(`prod *= ${s[I].getByIndices(`input${I}Indices`)};`)}}),m.push(`for(var ${y}: u32 = 0; ${y} < uniforms.${Ls(y)}; ${y}++) {`),g.push("}")});let v=_?[...d,`let sum = ${s.map((b,y)=>b.getByIndices(`input${y}Indices`)).join(" * ")};`]:[...d,p,...m,...f,c,...w,h,...g];return` + ${u.registerUniforms(o.map(b=>({name:`${Ls(b)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...s,a)} + + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var outputIndices = ${a.offsetToIndices("global_idx")}; + ${s.map((b,y)=>`var input${y}Indices: ${s[y].type.indices};`).join(` +`)} + ${v.join(` +`)}; + ${a.setByOffset("global_idx","sum")}; + }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let u=o.filter(c=>r.symbolToInfo.has(c)).map(c=>{var p;return{type:12,data:((p=r.symbolToInfo.get(c))==null?void 0:p.dimValue)||0}});u.push({type:12,data:i});let d=e.map((c,p)=>[...F(c)]).reduce((c,p)=>c.concat(p),u);return d.push(...F(n)),{outputs:[{dims:n,dataType:t}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:d}},getShaderSource:l}},Bu=(e,t)=>{let r=new Pu(e.inputs,t.equation),n=r.outputDims,s=e.inputs.map((i,a)=>i.dims);e.compute(Ou(s,e.inputs[0].dataType,r,n))},Ru=e=>{let t=e.equation.replace(/\s+/g,"");return ae({equation:t})}}),Du,Fs,Lu,Fu,Nu,sm=A(()=>{j(),J(),Y(),Du=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=r.length{let r=e.length-t.length,n=[];for(let s=0;se.length>t.length?Fs(e,t):Fs(t,e),Fu=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),n=Lu(t,r),s=e[0].dataType,i=s===9?4:1,a=Math.ceil(C.size(n)/i),o=u=>{let d=z("input",s,t.length,i),c=N("output",s,n.length,i),p;if(s===9){let h=(f,m,g="")=>` + let outputIndices${m} = ${c.offsetToIndices(`outputOffset + ${m}u`)}; + let offset${m} = ${d.broadcastedIndicesToOffset(`outputIndices${m}`,c)}; + let index${m} = offset${m} / 4u; + let component${m} = offset${m} % 4u; + ${f}[${m}] = ${g}(${d.getByOffset(`index${m}`)}[component${m}]); + `;p=` + let outputOffset = global_idx * ${i}; + var data = vec4(0); + ${h("data",0,"u32")} + ${h("data",1,"u32")} + ${h("data",2,"u32")} + ${h("data",3,"u32")} + ${c.setByOffset("global_idx","data")} + }`}else p=` + let outputIndices = ${c.offsetToIndices("global_idx")}; + let inputOffset = ${d.broadcastedIndicesToOffset("outputIndices",c)}; + ${c.setByOffset("global_idx",d.getByOffset("inputOffset"))} + }`;return` + ${u.registerUniform("vec_size","u32").declareVariables(d,c)} + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${p}`},l=[{type:12,data:a},...F(t,n)];return{name:"Expand",shaderCache:{hint:`${n.length}`,inputDependencies:["rank"]},getShaderSource:o,getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:l})}},Nu=e=>{Du(e.inputs),e.compute(Fu(e.inputs),{inputs:[0]})}}),Uu,qu,im=A(()=>{j(),J(),Y(),xs(),Uu=e=>{let t=e[0].dataType,r=C.size(e[0].dims),n=C.size(e[1].dims),s=n%4===0,i=a=>{let o=z("x",t,[1],4),l=z("bias",t,[1],4),u=N("y",t,[1],4),d=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],c=h=>` + let bias${h}_offset: u32 = (global_idx * 4 + ${h}) % uniforms.bias_size; + let bias${h} = ${l.getByOffset(`bias${h}_offset / 4`)}[bias${h}_offset % 4];`,p=s?` + let bias = ${l.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${c(0)}${c(1)}${c(2)}${c(3)} + let bias = ${o.type.value}(bias0, bias1, bias2, bias3);`;return`${a.registerUniforms(d).declareVariables(o,l,u)} + + ${vs(Te(t))} + + ${a.mainStart(Bt)} + ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")} + + let x = ${o.getByOffset("global_idx")}; + ${p} + let x_in = x + bias; + ${u.setByOffset("global_idx",$s("x_in"))} + }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${s}`,inputDependencies:["type","type"]},getShaderSource:i,getRunData:a=>({outputs:[{dims:a[0].dims,dataType:a[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:n}],dispatchGroup:{x:Math.ceil(r/Bt/4)}})}},qu=e=>{e.inputs.length<2||C.size(e.inputs[1].dims)===0?zl(e):e.compute(Uu(e.inputs))}}),Gu,Vu,ju,Wu,am=A(()=>{j(),J(),_e(),Y(),Gu=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},Vu=(e,t)=>{let r=e[0].dims,n=e[1].dims,s=r.length,i=C.normalizeAxis(t.axis,s),a=r.slice(0);a.splice(i,1,...n);let o=r[i],l=e[0].dataType===9?4:1,u=Math.ceil(C.size(a)/l),d=[{type:12,data:u},{type:6,data:o},{type:12,data:i},...F(e[0].dims,e[1].dims,a)],c=p=>{let h=z("data",e[0].dataType,e[0].dims.length,l),f=z("inputIndices",e[1].dataType,e[1].dims.length),m=N("output",e[0].dataType,a.length,l),g=_=>{let v=n.length,b=`var indicesIndices${_} = ${f.type.indices}(0);`;for(let y=0;y1?`indicesIndices${_}[${y}]`:`indicesIndices${_}`} = ${a.length>1?`outputIndices${_}[uniforms.axis + ${y}]`:`outputIndices${_}`};`;b+=` + var idx${_} = ${f.getByIndices(`indicesIndices${_}`)}; + if (idx${_} < 0) { + idx${_} = idx${_} + uniforms.axisDimLimit; + } + var dataIndices${_} : ${h.type.indices}; + `;for(let y=0,S=0;y1?`dataIndices${_}[${y}]`:`dataIndices${_}`} = u32(idx${_});`,S+=v):(b+=`${s>1?`dataIndices${_}[${y}]`:`dataIndices${_}`} = ${a.length>1?`outputIndices${_}[${S}]`:`outputIndices${_}`};`,S++);return b},w;if(e[0].dataType===9){let _=(v,b,y="")=>` + let outputIndices${b} = ${m.offsetToIndices(`outputOffset + ${b}u`)}; + ${g(b)}; + let offset${b} = ${h.indicesToOffset(`dataIndices${b}`)}; + let index${b} = offset${b} / 4u; + let component${b} = offset${b} % 4u; + ${v}[${b}] = ${y}(${h.getByOffset(`index${b}`)}[component${b}]); + `;w=` + let outputOffset = global_idx * ${l}; + var value = vec4(0); + ${_("value",0,"u32")} + ${_("value",1,"u32")} + ${_("value",2,"u32")} + ${_("value",3,"u32")} + ${m.setByOffset("global_idx","value")} + `}else w=` + let outputIndices = ${m.offsetToIndices("global_idx")}; + ${g("")}; + let value = ${h.getByIndices("dataIndices")}; + ${m.setByOffset("global_idx","value")}; + `;return` + ${p.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(h,f,m)} + ${p.mainStart()} + ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + ${w} + }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:d}),getShaderSource:c}},ju=e=>ae({axis:e.axis}),Wu=(e,t)=>{let r=e.inputs;Gu(r),e.compute(Vu(e.inputs,t))}}),Hu,Ku,Qu,Xu,om=A(()=>{j(),J(),_e(),Y(),Hu=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and + indices input tensors be of same rank.`)},Ku=(e,t)=>{let r=e[0].dims,n=e[0].dataType,s=r.length,i=e[1].dims,a=e[1].dataType,o=C.normalizeAxis(t.axis,s),l=r[o],u=i.slice(0),d=C.size(u),c=z("input",n,s),p=z("indicesInput",a,i.length),h=N("output",n,u.length),f=[{type:12,data:d},{type:6,data:l},{type:12,data:o}];return f.push(...F(r,i,u)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:u,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:f}),getShaderSource:m=>` + ${m.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(c,p,h)} + ${m.mainStart()} + ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + + let outputIndices = ${h.offsetToIndices("global_idx")}; + + var idx = ${p.getByOffset("global_idx")}; + if (idx < 0) { + idx = idx + uniforms.axisDimLimit; + } + var inputIndices = ${c.type.indices}(outputIndices); + ${c.indicesSet("inputIndices","uniforms.axis","u32(idx)")}; + let value = ${c.getByIndices("inputIndices")}; + + ${h.setByOffset("global_idx","value")}; + }`}},Qu=e=>ae({axis:e.axis}),Xu=(e,t)=>{let r=e.inputs;Hu(r),e.compute(Ku(e.inputs,t))}}),Yu,Zu,Ju,ed,lm=A(()=>{j(),J(),Y(),Yu=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")},Zu=(e,t)=>{let r=e[0].dims.slice(),n=e[1].dims.slice(),[s,i,a]=Ba.getShapeOfGemmResult(r,t.transA,n,t.transB,e.length===3?e[2].dims:void 0),o=[s,i];if(!o)throw new Error("Can't use gemm on the given tensors");let l=C.size(o),u=[{type:12,data:l},{type:12,data:s},{type:12,data:i},{type:12,data:a},{type:1,data:t.alpha},{type:1,data:t.beta}],d=["type","type"];e.length===3&&(u.push(...F(e[2].dims)),d.push("rank")),u.push(...F(o));let c=p=>{let h="";t.transA&&t.transB?h="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?h="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?h="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(h="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let f=t.alpha===1?"":"value *= uniforms.alpha;",m=z("a",e[0].dataType,e[0].dims),g=z("b",e[1].dataType,e[1].dims),w=m.type.value,_=null,v=[m,g];e.length===3&&(_=z("c",e[2].dataType,e[2].dims.length),v.push(_));let b=N("output",e[0].dataType,o.length);v.push(b);let y=[{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` + ${p.registerUniforms(y).declareVariables(...v)} + + ${p.mainStart()} + ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let m = global_idx / uniforms.N; + let n = global_idx % uniforms.N; + + var value = ${w}(0); + for (var k: u32 = 0u; k < uniforms.K; k++) { + ${h} + } + + ${f} + ${_!=null?`let cOffset = ${_.broadcastedIndicesToOffset("vec2(m, n)",b)}; value += ${w}(uniforms.beta) * ${_.getByOffset("cOffset")};`:""} + output[global_idx] = value; + }`};return{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u}),getShaderSource:c}},Ju=e=>{let t=e.transA,r=e.transB,n=e.alpha,s=e.beta;return{transA:t,transB:r,alpha:n,beta:s,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},ed=(e,t)=>{Yu(e.inputs),e.compute(Zu(e.inputs,t))}}),ke,td,rd,Ns,nd,Jt,sd,id=A(()=>{j(),J(),_e(),os(),ys(),Y(),Dt(),ke=(e,t)=>e.length>t&&e[t].dims.length>0&&C.size(e[t].dims)>0?e[t]:void 0,td=(e,t)=>{let r=e[0],n=ke(e,1),s=ke(e,2),i=ke(e,3),a=ke(e,4),o=ke(e,5),l=ke(e,6),u=ke(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,c=r.dims[0],p=r.dims[1],h=r.dims.length===3?d?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],f=p,m=0,g=0,w=Math.floor(h/t.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]!==c||l.dims[1]!==t.numHeads||l.dims[3]!==w)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(u.dims[0]!==c||u.dims[1]!==t.numHeads||u.dims[3]!==w)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');m=l.dims[2],g=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,f=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==w)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(s)throw new Error('Expect "value" be none when "key" has packed kv format.');_=5,f=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==w)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');_=0,f=n.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');_=3}if(i){if(i.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(s&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let v=0;if(a){v=8;let I=a.dims;throw I.length===1?I[0]===c?v=1:I[0]===3*c+2&&(v=3):I.length===2&&I[0]===c&&I[1]===f&&(v=5),v===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)'):new Error("Mask not supported")}let b=!1,y=h;if(s){if(s.dims.length!==3&&s.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==s.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(s.dims.length===3){if(f!==s.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');y=s.dims[2]}else{if(f!==s.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');y=s.dims[1]*s.dims[3],b=!0}}let S=m+f,x=!1;if(a)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]!==c&&o.dims[0]!==1||o.dims[1]!==t.numHeads||o.dims[2]!==p||o.dims[3]!==S)throw new Error('Input "relative_position_bias" shape (batch_size, 1, sequence_length, kv_sequence_length)')}return{batchSize:c,sequenceLength:p,pastSequenceLength:m,kvSequenceLength:f,totalSequenceLength:S,maxSequenceLength:g,inputHiddenSize:0,hiddenSize:h,vHiddenSize:y,headSize:w,vHeadSize:Math.floor(y/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:v,scale:t.scale,broadcastResPosBias:x,passPastInKv:b,qkvFormat:_}},rd=e=>ae({...e}),Ns=ae({perm:[0,2,1,3]}),nd=(e,t,r,n,s,i,a)=>{let o=[n,s,i],l=C.size(o),u=[{type:12,data:l},{type:12,data:a},{type:12,data:i}],d=c=>{let p=N("qkv_with_bias",t.dataType,o),h=z("qkv",t.dataType,o),f=z("bias",r.dataType,o),m=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return` + ${c.registerUniforms(m).declareVariables(h,f,p)} + ${c.mainStart()} + ${c.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset; + + qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx]; + }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:o,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u}),getShaderSource:d},{inputs:[t,r],outputs:[-1]})[0]},Jt=(e,t,r,n,s,i,a,o)=>{let l=i;if(a){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=nd(e,i,a,t,n,r*s,o),l=l.reshape([t,n,r,s]),e.compute(Ze(l,Ns.perm),{inputs:[l],outputs:[-1]})[0]}else return i.dims.length===3&&(l=i.reshape([t,n,r,s])),e.compute(Ze(l,Ns.perm),{inputs:[l],outputs:[-1]})[0]},sd=(e,t)=>{let r=td(e.inputs,t),n=e.inputs[0],s=ke(e.inputs,1),i=ke(e.inputs,2),a=ke(e.inputs,3),o=ke(e.inputs,4),l=ke(e.inputs,5),u=ke(e.inputs,6),d=ke(e.inputs,7);if(n.dims.length===5)throw new Error("Packed QKV is not implemented");if((s==null?void 0:s.dims.length)===5)throw new Error("Packed KV is not implemented");let c=s&&i&&s.dims.length===4&&i.dims.length===4,p=Jt(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,n,a,0);if(c)return Yt(e,p,s,i,o,void 0,u,d,l,r,t);if(!s||!i)throw new Error("key and value must be provided");let h=Jt(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,s,a,r.hiddenSize),f=Jt(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,a,2*r.hiddenSize);Yt(e,p,h,f,o,void 0,u,d,l,r,t)}}),Us,ad,od,qs,ld,ud=A(()=>{j(),J(),Y(),Us=e=>Array.from(e.getBigInt64Array(),Number),ad=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(Us(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")},od=(e,t)=>{let r=[];for(let n=0;n{let r=e[0].dims,n=t??Us(e[1]),s=od(r,n),i=C.size(s),a=e[0].dataType,o=z("input",a,r.length),l=N("output",a,s.length),u=d=>` + const inputShape = ${o.indices(...r)}; + ${d.registerUniform("output_size","u32").declareVariables(o,l)} + ${d.mainStart()} + ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let output_indices = ${l.offsetToIndices("global_idx")}; + var input_indices: ${o.type.indices}; + for (var i = 0; i < ${r.length}; i++) { + let input_dim_i = ${o.indicesGet("uniforms.input_shape","i")}; + let input_dim_value = ${l.indicesGet("output_indices","i")} % input_dim_i; + + ${o.indicesSet("input_indices","i","input_dim_value")} + } + ${l.setByOffset("global_idx",o.getByIndices("input_indices"))} + }`;return{name:"Tile",shaderCache:{hint:`${n}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:s,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},...F(e[0].dims,s)]}),getShaderSource:u}},ld=e=>{ad(e.inputs),e.compute(qs(e.inputs),{inputs:[0]})}}),dd,Gs,cd,pd,Vs,hd,um=A(()=>{j(),J(),_e(),ys(),Y(),id(),ud(),Dt(),dd=(e,t)=>{let r=e[0],n=e[1],s=e[2],i=e[3],a=e[4];if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let o=!1,l=r.dims[0],u=r.dims[1],d=r.dims.length===3?o?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],c=u,p=0,h=0,f=Math.floor(d/t.numHeads),m=i&&i.dims.length!==0,g=a&&a.dims.length!==0,w=!0;if(m&&g){if(i.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(a.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');p=i.dims[1],h=i.dims[1]}else if(m||g)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let _;if(n){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(r.dims[2]%n.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');_=2,c=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==t.numHeads||n.dims[3]!==2||n.dims[4]!==f)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(s)throw new Error('Expect "value" be none when "key" has packed kv format.');_=5,c=n.dims[1]}else{if(n.dims[1]!==t.numHeads||n.dims[3]!==f)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');_=0,c=n.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');_=3}let v=0,b=!1,y=d;if(s){if(s.dims.length!==3&&s.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==s.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(s.dims.length===3){if(c!==s.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');y=s.dims[2]}else{if(c!==s.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');y=s.dims[1]*s.dims[3],b=!0}}let S=p+c;return{batchSize:l,sequenceLength:u,pastSequenceLength:p,kvSequenceLength:c,totalSequenceLength:S,maxSequenceLength:h,inputHiddenSize:0,hiddenSize:d,vHiddenSize:y,headSize:f,vHeadSize:Math.floor(y/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:v,scale:t.scale,broadcastResPosBias:!1,passPastInKv:b,qkvFormat:_,isPastkvBSNH:w}},Gs=(e,t,r,n)=>{let s=[n.batchSize,n.totalSequenceLength,n.kvNumHeads,n.headSize],i=4,a=C.size(s)/i,o=n.totalSequenceLength,l=N("present_kv",r,s.length,i),u=z("new_kv",e.dataType,e.dims.length,i),d=t?z("past_kv",t.dataType,t.dims.length,i):void 0,c=Math.ceil(n.headSize/i),p={x:o,y:e.dims[0],z:1},h=t?["rank","rank"]:["rank"],f=[{type:12,data:a},{type:12,data:n.pastSequenceLength},{type:12,data:n.kvSequenceLength},{type:12,data:n.totalSequenceLength}],m=[u];d?(f.push(...F(e.dims),...F(t.dims),...F(s)),m.push(d)):f.push(...F(e.dims),...F(s));let g=[{name:"output_size",type:"u32"},{name:"past_seqlen",type:"u32"},{name:"new_seqlen",type:"u32"},{name:"present_seqlen",type:"u32"}],w=` let past_batch_stride = uniforms.past_seqlen * num_heads * H; + var past_head_stride = uniforms.past_seqlen * H; + if (is_bsnh) { + past_head_stride = H; + } + let in_offset = b * past_batch_stride + s * row_stride + n * past_head_stride + h; + present_kv[out_offset] = past_kv[in_offset];`,_=` let new_batch_stride = uniforms.new_seqlen * num_heads * H; + let new_row_stride = num_heads * H; + let new_head_stride = H; + let in_offset = b * new_batch_stride + (s - past_seqlen) * new_row_stride + n * new_head_stride + h; + present_kv[out_offset] = new_kv[in_offset];`,v=t?`if (s < past_seqlen) { + ${w} + } else if (s < past_seqlen + uniforms.new_seqlen) { + ${_} + }`:`if (s < past_seqlen + uniforms.new_seqlen) { + ${_} + }`,b=y=>` + + ${y.registerUniforms(g).declareVariables(...m,l)} + ${y.mainStart([c,n.kvNumHeads,1])} + ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + var indices = ${l.offsetToIndices("global_idx")}; + let h = local_id.x; + let n = local_id.y; + let s = workgroup_id.x; + let b = workgroup_id.y; + let num_heads = ${n.kvNumHeads}u; + let H = ${c}u; + + let present_seqlen = uniforms.present_seqlen; + let present_batch_stride = present_seqlen * num_heads * H; + var row_stride = H; + let is_bsnh = ${n.isPastkvBSNH}; + + if (is_bsnh) { + row_stride = num_heads * H; + } + var present_head_stride = present_seqlen * H; + if (is_bsnh) { + present_head_stride = H; + } + + let past_seqlen = uniforms.past_seqlen; + + let out_offset = b * present_batch_stride + s * row_stride + n * present_head_stride + h; + ${v} + }`;return{name:"ConcatPastNew",shaderCache:{hint:`${n.kvNumHeads}${c}${!!t}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:s,dataType:r}],dispatchGroup:p,programUniforms:f}),getShaderSource:b}},cd=e=>ae({...e}),pd=ae({perm:[0,2,1,3]}),Vs=(e,t,r,n,s)=>{let i=t,a=n.kvNumHeads,o=n.nReps;return t.dims.length===3&&n.kvSequenceLength!==0&&(i=t.reshape([n.batchSize,n.kvSequenceLength,a,n.headSize])),r?i=e.compute(Gs(i,r,i.dataType,n),{inputs:[i,r],outputs:[n.isPastkvBSNH?s:-1]})[0]:i=e.compute(Gs(i,void 0,i.dataType,n),{inputs:[i],outputs:[n.isPastkvBSNH?s:-1]})[0],o!==1&&(i=e.compute(qs([i],[1,1,1,o]),{inputs:[i],outputs:[-1]})[0],i=i.reshape([n.batchSize,n.totalSequenceLength,a*o,n.headSize])),e.compute(Ze(i,pd.perm),{inputs:[i],outputs:[-1]})[0]},hd=(e,t)=>{var l;let r=dd(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(((l=e.inputs[1])==null?void 0:l.dims.length)===5)throw new Error("Packed KV is not implemented");let n=Jt(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,e.inputs[0],void 0,0),s=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,i=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,a=Vs(e,e.inputs[1],s,r,1),o=Vs(e,e.inputs[2],i,r,2);Yt(e,n,a,o,void 0,void 0,void 0,void 0,void 0,r,t)}}),fd,md,gd,_d,dm=A(()=>{j(),J(),Y(),fd=(e,t)=>{let r=e[0].dims,n=r,s=2,i=C.sizeToDimension(r,s),a=C.sizeFromDimension(r,s),o=ge(a),l=a/o,u=[r[0],r[1],l],d=["rank","type","type"],c=[{type:12,data:a},{type:12,data:l}];c.push(...F(u,u));let p=h=>{let f=z("x",e[0].dataType,u.length,o),m=z("scale",e[1].dataType,e[1].dims),g=z("bias",e[2].dataType,e[2].dims),w=N("output",e[0].dataType,u.length,o),_=[f,m,g,w],v=f.type.value,b=o===1?"f32":`vec${o}`,y=64,S=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return` + var meanShared : f32; + var squaredNormShared : f32; + var workgroupShared : array<${b}, ${y}>; + const workgroupSize = ${y}u; + ${h.registerUniforms(S).declareVariables(..._)} + ${h.mainStart(y)} + 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 = ${b}(0); + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + initial = initial + ${b}(${f.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 = ${dt("workgroupShared[0]",o)} / f32(uniforms.normSize); + } + workgroupBarrier(); + + // reinitialize workgroup memory. + initial = ${b}(0); + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + let deviation = ${b}(${f.get("batch","channel","h")}) - ${b}(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 = ${dt("workgroupShared[0]",o)}; + } + workgroupBarrier(); + + let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${t.epsilon})); + let channelScale = invStdDev * f32(${m.getByOffset("channel")}); + let channelShift = f32(${g.getByOffset("channel")}) - meanShared * channelScale; + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + let value = ${f.get("batch","channel","h")} * ${v}(${b}(channelScale)) + ${v}(${b}(channelShift)); + ${w.set("batch","channel","h","value")}; + } + }`};return{name:"InstanceNormalization",shaderCache:{hint:`${t.epsilon};${o}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:i},programUniforms:c}),getShaderSource:p}},md=(e,t,r,n,s,i,a,o)=>{let l=ge(a),u=64,d=l===1?"vec2f":`mat2x${l}f`,c=l===1?"f32":`vec${l}f`,p=(S,x)=>`${d}(${S}, ${x})`,h=s*a/l,f=Math.ceil(i/u),m=["type"],g=[{type:12,data:f},{type:12,data:i},{type:12,data:Math.floor(a/l)},{type:12,data:Math.floor(i*a/l)}],w=S=>{let x=z("input",t.dataType,t.dims,l);return` + ${S.declareVariables(x)} + @group(0) @binding(1) var output : array<${d}>; + struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32}; + @group(0) @binding(2) var uniforms: Uniforms; + + ${S.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 = ${$t("f32",l)}; + var squaredSum = ${$t("f32",l)}; + for (var i: u32 = wgOffset; i < wgMax; i++) { + let value = ${c}(input[offset + i * uniforms.C]); + sum += value; + squaredSum += value * value; + } + output[global_idx] = ${p("sum","squaredSum")}; + }`},_=e.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${l}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:[s,a,u,2],dataType:1}],dispatchGroup:{x:s*a/l},programUniforms:g}),getShaderSource:w},{inputs:[t],outputs:[-1]})[0],v=[{type:12,data:h},{type:12,data:i},{type:12,data:Math.floor(a/l)},{type:12,data:Math.floor(u*a/l)}],b=["type","type","type"],y=S=>{let x=z("scale",r.dataType,r.dims,l),I=z("bias",n.dataType,n.dims,l);return` + @group(0) @binding(0) var input : array<${d}>; + @group(0) @binding(1) var scale : array<${x.type.storage}>; + @group(0) @binding(2) var bias : array<${I.type.storage}>; + @group(0) @binding(3) var output : array<${d}>; + struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32}; + @group(0) @binding(4) var uniforms: Uniforms; + + ${S.mainStart()} + ${S.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 = ${$t("f32",l)}; + var squaredSum = ${$t("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 * ${c}(scale[currentChannelNumber]); + let channelShift = ${c}(bias[currentChannelNumber]) - sum * channelScale; + + output[global_idx] = ${p("channelScale","channelShift")}; + }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${o}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:[s,a,2],dataType:1}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:v}),getShaderSource:y},{inputs:[_,r,n],outputs:[-1]})[0]},gd=(e,t,r)=>{let n=t[0].dims,s=n,i=n[0],a=n[n.length-1],o=C.sizeFromDimension(n,1)/a,l=ge(a),u=C.size(s)/l,d=[{type:12,data:o},{type:12,data:Math.floor(a/l)}],c=["type","type"],p=md(e,t[0],t[1],t[2],i,o,a,r.epsilon),h=f=>{let m=ye(t[0].dataType),g=l===1?"vec2f":`mat2x${l}f`,w=l===1?m:`vec${l}<${m}>`,_=z("input",t[0].dataType,t[0].dims,l),v=N("output",t[0].dataType,s,l);return` + @group(0) @binding(0) var input : array<${_.type.storage}>; + @group(0) @binding(1) var scaleInput : array<${g}>; + @group(0) @binding(2) var output : array<${v.type.storage}>; + struct Uniforms {H: u32, C : u32}; + @group(0) @binding(3) var uniforms: Uniforms; + + ${f.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], ${w}(scale[0]), ${w}(scale[1])); + }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${l}`,inputDependencies:c},getRunData:()=>({outputs:[{dims:s,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:d}),getShaderSource:h},{inputs:[t[0],p]})},_d=(e,t)=>{t.format==="NHWC"?gd(e,e.inputs,t):e.compute(fd(e.inputs,t))}}),wd,yd,bd,cm=A(()=>{j(),J(),Y(),wd=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},yd=(e,t,r)=>{let n=t.simplified,s=e[0].dims,i=e[1],a=!n&&e[2],o=s,l=C.normalizeAxis(t.axis,s.length),u=C.sizeToDimension(s,l),d=C.sizeFromDimension(s,l),c=C.size(i.dims),p=a?C.size(a.dims):0;if(c!==d||a&&p!==d)throw new Error(`Size of X.shape()[axis:] == ${d}. + Size of scale and bias (if provided) must match this. + Got scale size of ${c} and bias size of ${p}`);let h=[];for(let y=0;y1,_=r>2,v=y=>{let S=ye(e[0].dataType),x=[z("x",e[0].dataType,e[0].dims,f),z("scale",i.dataType,i.dims,f)];a&&x.push(z("bias",a.dataType,a.dims,f)),x.push(N("output",e[0].dataType,o,f)),w&&x.push(N("mean_data_output",1,h)),_&&x.push(N("inv_std_output",1,h));let I=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return` + ${y.registerUniforms(I).declareVariables(...x)} + ${y.mainStart()} + ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")} + let offset = global_idx * uniforms.norm_size_vectorized; + var mean_vector = ${$t("f32",f)}; + var mean_square_vector = ${$t("f32",f)}; + + for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) { + let value = ${Rt(S,f,"x[h + offset]")}; + mean_vector += value; + mean_square_vector += value * value; + } + let mean = ${dt("mean_vector",f)} / uniforms.norm_size; + let inv_std_dev = inverseSqrt(${dt("mean_square_vector",f)} / uniforms.norm_size ${n?"":"- mean * mean"} + uniforms.epsilon); + + for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) { + let f32input = ${Rt(S,f,"x[j + offset]")}; + let f32scale = ${Rt(S,f,"scale[j]")}; + output[j + offset] = ${x[0].type.value}((f32input ${n?"":"- mean"}) * inv_std_dev * f32scale + ${a?`+ ${Rt(S,f,"bias[j]")}`:""} + ); + } + + ${w?"mean_data_output[global_idx] = mean":""}; + ${_?"inv_std_output[global_idx] = inv_std_dev":""}; + }`},b=[{dims:o,dataType:e[0].dataType}];return w&&b.push({dims:h,dataType:1}),_&&b.push({dims:h,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${f};${r};${n}`,inputDependencies:m},getRunData:()=>({outputs:b,dispatchGroup:{x:Math.ceil(u/64)},programUniforms:g}),getShaderSource:v}},bd=(e,t)=>{wd(e.inputs),e.compute(yd(e.inputs,t,e.outputCount))}}),vd,$d,xd,kd,pm=A(()=>{j(),J(),_e(),Y(),vd=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],n=r.dims.length;if(r.dims[n-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let s=Math.floor((t.k+t.blockSize-1)/t.blockSize),i=t.blockSize/8*t.bits,a=e[1];if(!C.areEqual(a.dims,[t.n,s,i]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let o=e[2].dims;if(C.size(o)!==t.n*s)throw new Error("scales input size error.");if(e.length===4){let l=e[3].dims,u=t.bits>4?t.n*s:t.n*Math.floor((s+1)/2);if(C.size(l)!==u)throw new Error("zeroPoints input size error.")}},$d=(e,t,r,n)=>{let s=e[0].dims,i=s.length,a=Math.floor((t.k+t.blockSize-1)/t.blockSize),o=s[i-2],l=t.k,u=t.n,d=s.slice(0,i-2),c=C.size(d),p=t.blockSize/8*t.bits/4,h=e[0].dataType,f=ge(o),m=ge(t.k),g=ge(p),w=Qt(h),_=o*a*w,v=Math.floor(n/_),b=a<=r[0]&&v>0,y=!b||v>=4?ge(u):v>=2&&ge(u)>=2?2:1,S=d.concat([o,u]),x=C.size(S)/y/f,I=b?[]:[{type:12,data:x},{type:12,data:t.blockSize}],D=[c,o,l/m],R=C.convertShape(e[1].dims).slice();R.splice(-1,1,p/g),I.push(...F(D)),I.push(...F(R)),I.push(...F(e[2].dims)),e.length===4&&I.push(...F(C.convertShape(e[3].dims)));let U=[c,o,u/y];I.push(...F(U));let q=H=>{let V=D.length,E=z("a",e[0].dataType,V,m),M=z("b",12,R.length,g),O=z("scales",e[2].dataType,e[2].dims.length),B=[E,M,O],K=e.length===4?z("zero_points",12,e[3].dims.length):void 0;K&&B.push(K);let ee=U.length,k=N("output",e[0].dataType,ee,y),L=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],G=ye(e[0].dataType),ie=(()=>{switch(m){case 1:return`array<${G}, 8>`;case 2:return`mat4x2<${G}>`;case 4:return`mat2x4<${G}>`;default:throw new Error(`${m}-component is not supported.`)}})(),ce=` + for (var word: u32 = 0; word < ${p}; word += ${g}) { + ${M.indicesSet("b_indices","2","word")}; + let b_data = ${M.getByIndices("b_indices")}; + for (var i: u32 = 0; i < ${g}; i++) { + let b_value: u32 = ${g===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 = ${ie}(${Array.from({length:4},(Tr,Ye)=>`${G}(b_value_lower[${Ye}]), ${G}(b_value_upper[${Ye}])`).join(", ")}); + let b_dequantized_values = ${m===1?`${ie}(${Array.from({length:8},(Tr,Ye)=>`(b_quantized_values[${Ye}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${ie}(${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 < ${b?o:f}u; m++) { + ${E.indicesSet("a_indices",V-2,b?"m":`row * ${f} + m`)}; + ${E.indicesSet("a_indices",V-1,"word_offset")}; + var input_offset = ${E.indicesToOffset("a_indices")}; + var a_data: ${ie}; + for (var j: u32 = 0; j < ${8/m}; j++) { + a_data[j] = ${E.getByOffset("input_offset")}; + input_offset++; + } + ${b?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${y>1?"[c]":""} += ${Array.from({length:8/m},(Tr,Ye)=>`${m===1?`a_data[${Ye}] * b_dequantized_values[${Ye}]`:`dot(a_data[${Ye}], b_dequantized_values[${Ye}])`}`).join(" + ")}; + } + word_offset += ${8/m}; + } + }`,rt=K?` + zero_point_offset += 4; + if (zero_point_offset == 32) { + zero_point_offset = 0; + zero_point_index++; + zero_point_word = ${K.getByOffset("zero_point_index")}; + }`:"";return b?` + var workgroup_shared: array<${k.type.value}, ${o*a}>; + ${H.declareVariables(...B,k)} + ${H.mainStart([a,1,1])} + var a_indices: ${E.type.indices}; + var block = local_id.x; + var col = workgroup_id.y; + var batch = workgroup_id.z; + ${E.indicesSet("a_indices","0","batch")}; + // Two zero points are packed into one byte when uniforms.bits is 4. + for (var c: u32 = 0; c < ${y}; c++) { + let col_times_components_plus_c = col * ${y} + c; + ${K?` + var zero_point_bytes_per_col: u32 = (${a} + 1) / 2; + var zero_point_byte_count: u32 = col_times_components_plus_c * zero_point_bytes_per_col + (block >> 0x1u); + 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 = ${K.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;`:""} + var b_indices: ${M.type.indices}; + ${M.indicesSet("b_indices","0","col_times_components_plus_c")}; + // The scale and zero points are computed per block. + var scales_index = col_times_components_plus_c * ${a} + block; + let scale = ${O.getByOffset("scales_index")}; + // The default zero point is 8 for unsigned 4-bit quantization. + let zero_point = ${G}(${K?"(zero_point_word) & 0xFu":8}); + ${M.indicesSet("b_indices","1","block")}; + var word_offset: u32 = block * ${t.blockSize/m}; + var workgroup_shared_offset: u32 = block * ${o}; + ${ce} + } + workgroupBarrier(); + var output_indices: ${k.type.indices}; + var elements_per_thread: u32 = ${Math.ceil(o/a)}; + ${k.indicesSet("output_indices","0","batch")}; + ${k.indicesSet("output_indices",ee-1,"col")}; + ${k.indicesSet("output_indices",ee-2,"local_id.x * elements_per_thread")}; + var output_offset = ${k.indicesToOffset("output_indices")}; + for (var m: u32 = 0u; m < elements_per_thread; m++) { + var row = m + local_id.x * elements_per_thread; + if (row < ${o}) { + var output_value: ${k.type.value} = ${k.type.value}(0); + var workgroup_shared_offset: u32 = row; + for (var b: u32 = 0u; b < ${a}u; b++) { + output_value += workgroup_shared[workgroup_shared_offset]; + workgroup_shared_offset += ${o}; + } + ${k.setByOffset("output_offset","output_value")}; + output_offset += ${u/y}; + } + } + }`:` + ${H.registerUniforms(L).declareVariables(...B,k)} + ${H.mainStart()} + ${H.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + var output_values: array<${k.type.value}, ${f}>; + var output_indices = ${k.offsetToIndices("global_idx")}; + var col = ${k.indicesGet("output_indices",ee-1)}; + var row = ${k.indicesGet("output_indices",ee-2)}; + var a_indices: ${E.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 + ${K?` + var zero_point_abs_offset = col * ${y} * ((${a} + 1) / 2); + var zero_point_index: u32 = zero_point_abs_offset / 4; + var zero_point_word: u32 = ${K.getByOffset("zero_point_index")}; + var zero_point_offset: u32 = (zero_point_abs_offset % 4) * 8;`:""} + var scale_index = col * ${a*y}; + var b_indices: ${M.type.indices}; + for (var c: u32 = 0; c < ${y}; c++) { + ${M.indicesSet("b_indices","0",`col * ${y} + c`)}; + var block_offset: u32 = 0; + for (var block: u32 = 0; block < ${a}; block++) { + // The scale and zero points are computed per block. + let scale = ${O.getByOffset("scale_index")}; + // The default zero point is 8 for unsigned 4-bit quantization. + let zero_point = ${G}(${K?"extractBits(zero_point_word, zero_point_offset, 4)":8}); + ${M.indicesSet("b_indices","1","block")}; + var word_offset: u32 = block_offset; + ${ce} + scale_index++; + ${rt} + block_offset += uniforms.block_size / ${m}; + } + // Drop the trailing 4 bits if the zero_poit_offset is not a byte boundary to align with the next byte. + ${K?`if (zero_point_offset % 8 > 0) { + ${rt} + }`:""} + } + for (var k: u32 = 0u; k < ${f}u; k++) { + ${k.indicesSet("output_indices",ee-2,`${f} * row + k`)}; + ${k.setByIndices("output_indices","output_values[k]")} + } + }`};return{name:b?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${t.cacheKey};${o};${h};${e.length}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:S,dataType:h}],name:b?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:b?{x:1,y:Math.ceil(u/y),z:c}:{x:Math.ceil(x/64)},programUniforms:I}),getShaderSource:q}},xd=(e,t)=>{vd(e.inputs,t);let r=e.getMaxComputeWorkgroupSizes(),n=e.getMaxComputeWorkgroupStoragesize();e.compute($d(e.inputs,t,r,n))},kd=e=>ae(e)}),Sd,Ed,Td,Id,Md,Cd,zd,Ad,Pd,hm=A(()=>{j(),J(),Y(),Sd=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].")}},Ed=(e,t,r)=>{let n="";for(let s=t-1;s>=0;--s)n+=` + k = i32(${e.indicesGet("indices",s)}) - ${Q("uniforms.pads",s,r)}; + if (k < 0) { + break; + } + if (k >= i32(${Q("uniforms.x_shape",s,t)})) { + break; + } + offset += k * i32(${Q("uniforms.x_strides",s,t)}); + `;return` + value = ${e.type.value}(uniforms.constant_value); + for (var i = 0; i < 1; i++) { + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + } + `},Td=(e,t,r)=>{let n="";for(let s=t-1;s>=0;--s)n+=` + k = i32(${e.indicesGet("indices",s)}) - ${Q("uniforms.pads",s,r)}; + if (k < 0) { + k = -k; + } + { + let _2n_1 = 2 * (i32(${Q("uniforms.x_shape",s,t)}) - 1); + k = k % _2n_1; + if(k >= i32(${Q("uniforms.x_shape",s,t)})) { + k = _2n_1 - k; + } + } + offset += k * i32(${Q("uniforms.x_strides",s,t)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},Id=(e,t,r)=>{let n="";for(let s=t-1;s>=0;--s)n+=` + k = i32(${e.indicesGet("indices",s)}) - ${Q("uniforms.pads",s,r)}; + if (k < 0) { + k = 0; + } + if (k >= i32(${Q("uniforms.x_shape",s,t)})) { + k = i32(${Q("uniforms.x_shape",s,t)}) - 1; + } + offset += k * i32(${Q("uniforms.x_strides",s,t)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},Md=(e,t,r)=>{let n="";for(let s=t-1;s>=0;--s)n+=` + k = i32(${e.indicesGet("indices",s)}) - ${Q("uniforms.pads",s,r)}; + if (k < 0) { + k += i32(${Q("uniforms.x_shape",s,t)}]); + } + if (k >= i32(${Q("uniforms.x_shape",s,t)})) { + k -= i32(${Q("uniforms.x_shape",s,t)}); + } + offset += k * i32(${Q("uniforms.x_strides",s,t)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},Cd=(e,t,r)=>{switch(r.mode){case 0:return Ed(e,t,r.pads.length);case 1:return Td(e,t,r.pads.length);case 2:return Id(e,t,r.pads.length);case 3:return Md(e,t,r.pads.length);default:throw new Error("Invalid mode")}},zd=(e,t)=>{let r=C.padShape(e[0].dims.slice(),t.pads),n=e[0].dims,s=C.size(r),i=[{type:12,data:s},{type:6,data:t.pads}];t.mode===0&&i.push({type:e[0].dataType,data:t.value}),i.push(...F(e[0].dims,r));let a=["rank"],o=l=>{let u=N("output",e[0].dataType,r.length),d=z("x",e[0].dataType,n.length),c=d.type.value,p=Cd(u,n.length,t),h=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&h.push({name:"constant_value",type:c}),` + ${l.registerUniforms(h).declareVariables(d,u)} + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${u.offsetToIndices("global_idx")}; + + var value = ${c}(0); + ${p} + output[global_idx] = value; + }`};return{name:"Pad",shaderCache:{hint:`${t.mode}`,inputDependencies:a},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(C.size(r)/64)},programUniforms:i}),getShaderSource:o}},Ad=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),n=e.length>=3&&e[2].data?e[2].getFloat32Array()[0]:0,s=e[0].dims.length,i=new Int32Array(2*s).fill(0);if(e.length>=4){let o=e[3].getBigInt64Array();for(let l=0;li[Number(l)]=Number(o));let a=[];return i.forEach(o=>a.push(o)),{mode:t.mode,value:n,pads:a}}else return t},Pd=(e,t)=>{Sd(e.inputs);let r=Ad(e.inputs,t);e.compute(zd(e.inputs,r),{inputs:[0]})}}),er,js,Ws,Hs,Ks,Od,Bd,Qs,Xs,Rd,Dd,Ys,Ld,Fd,Zs,Nd,Ud,qd,Gd,fm=A(()=>{Ue(),j(),J(),Y(),er=e=>{if(se.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},js=(e,t,r)=>{let n=t.format==="NHWC",s=e.dims.slice();n&&s.splice(1,0,s.pop());let i=Object.hasOwnProperty.call(t,"dilations"),a=t.kernelShape.slice(),o=t.strides.slice(),l=i?t.dilations.slice():[],u=t.pads.slice();jr.adjustPoolAttributes(r,s,a,o,l,u);let d=jr.computePoolOutputShape(r,s,o,l,a,u,t.autoPad),c=Object.assign({},t);i?Object.assign(c,{kernelShape:a,strides:o,pads:u,dilations:l,cacheKey:t.cacheKey}):Object.assign(c,{kernelShape:a,strides:o,pads:u,cacheKey:t.cacheKey});let p=d.slice();return p.push(p.splice(1,1)[0]),[c,n?p:d]},Ws=(e,t)=>{let r=t.format==="NHWC",n=C.size(e),s=C.size(t.kernelShape),i=[{type:12,data:n},{type:12,data:s}],a=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let o=t.kernelShape[t.kernelShape.length-1],l=t.strides[t.strides.length-1],u=t.pads[t.pads.length/2-1],d=t.pads[t.pads.length-1],c=!!(u+d);i.push({type:12,data:o},{type:12,data:l},{type:12,data:u},{type:12,data:d}),a.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let p=!1;if(t.kernelShape.length===2){let h=t.kernelShape[t.kernelShape.length-2],f=t.strides[t.strides.length-2],m=t.pads[t.pads.length/2-2],g=t.pads[t.pads.length-2];p=!!(m+g),i.push({type:12,data:h},{type:12,data:f},{type:12,data:m},{type:12,data:g}),a.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[i,a,!0,c,p]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let o=C.computeStrides(t.kernelShape);i.push({type:12,data:o},{type:12,data:t.pads},{type:12,data:t.strides}),a.push({name:"kernelStrides",type:"u32",length:o.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let l=t.pads.reduce((u,d)=>u+d);return[i,a,!!l,!1,!1]}},Hs=(e,t,r,n,s,i,a,o,l,u,d,c)=>{let p=s.format==="NHWC",h=t.type.value,f=N("output",t.type.tensor,n);if(s.kernelShape.length<=2){let m="",g="",w="",_=r-(p?2:1);if(d?m=` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${_}] = indices[${_}] * uniforms.sw - uniforms.pwStart + i; + if (xIndices[${_}] < 0 || xIndices[${_}] + >= uniforms.x_shape[${_}]) { + pad++; + continue; + } + let x_val = x[${t.indicesToOffset("xIndices")}]; + ${i} + }`:m=` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${_}] = indices[${_}] * uniforms.sw - uniforms.pwStart + i; + let x_val = x[${t.indicesToOffset("xIndices")}]; + ${i} + }`,s.kernelShape.length===2){let v=r-(p?3:2);c?g=` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${v}] = indices[${v}] * uniforms.sh - uniforms.phStart + j; + if (xIndices[${v}] < 0 || xIndices[${v}] >= uniforms.x_shape[${v}]) { + pad += i32(uniforms.kw); + continue; + } + `:g=` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${v}] = indices[${v}] * uniforms.sh - uniforms.phStart + j; + `,w=` + } + `}return` + ${e.registerUniforms(l).declareVariables(t,f)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + + let indices = ${f.offsetToIndices("global_idx")}; + var xIndices = ${f.offsetToIndices("global_idx")}; + + var value = ${h}(${o}); + var pad = 0; + ${g} + ${m} + ${w} + ${a} + + output[global_idx] = value; + }`}else{if(p)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let m=s.kernelShape.length,g=s.pads.length,w="";return u?w=` + if (xIndices[j] >= uniforms.x_shape[j]) { + pad++; + isPad = true; + break; + } + } + if (!isPad) { + let x_val = x[${t.indicesToOffset("xIndices")}]; + ${i} + }`:w=` + } + let x_val = x[${t.indicesToOffset("xIndices")}]; + ${i} + `,` + ${e.registerUniforms(l).declareVariables(t,f)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + let indices = ${f.offsetToIndices("global_idx")}; + var xIndices = ${f.offsetToIndices("global_idx")}; + + var offsets: array; + + var value = ${h}(${o}); + var pad = 0; + var isPad = false; + + for (var i: u32 = 0u; i < uniforms.kernelSize; i++) { + var offset = i; + for (var j = 0u; j < ${m-1}u; j++) { + offsets[j] = offset / ${Q("uniforms.kernelStrides","j",m)}; + offset -= offsets[j] * ${Q("uniforms.kernelStrides","j",m)}; + } + offsets[${m-1}] = offset; + + isPad = false; + for (var j = ${r-m}u; j < ${r}u; j++) { + xIndices[j] = indices[j] * ${Q("uniforms.strides",`j - ${r-m}u`,m)} + + offsets[j - ${r-m}u] - ${Q("uniforms.pads","j - 2u",g)}; + ${w} + } + ${a} + + output[global_idx] = value; + }`}},Ks=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,Od=e=>`${Ks(e)};${e.countIncludePad}`,Bd=e=>`${Ks(e)};${e.storageOrder};${e.dilations}`,Qs=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}),Xs=(e,t,r,n)=>{let[s,i]=js(t,n,r),a=z("x",t.dataType,t.dims.length),o=a.type.value,l="value += x_val;",u="";s.countIncludePad?u+=`value /= ${o}(uniforms.kernelSize);`:u+=`value /= ${o}(i32(uniforms.kernelSize) - pad);`;let[d,c,p,h,f]=Ws(i,s);d.push(...F(t.dims,i));let m=["rank"];return{name:e,shaderCache:{hint:`${n.cacheKey};${p};${h};${f}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(C.size(i)/64)},programUniforms:d}),getShaderSource:g=>Hs(g,a,t.dims.length,i.length,s,l,u,0,c,p,h,f)}},Rd=e=>{let t=e.count_include_pad!==0,r=Qs(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let n={countIncludePad:t,...r,cacheKey:""};return{...n,cacheKey:Od(n)}},Dd=(e,t)=>{er(e.inputs),e.compute(Xs("AveragePool",e.inputs[0],!1,t))},Ys={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},Ld=e=>{let t=e.format;return{format:t,...Ys,cacheKey:t}},Fd=(e,t)=>{er(e.inputs),e.compute(Xs("GlobalAveragePool",e.inputs[0],!0,t))},Zs=(e,t,r,n)=>{let[s,i]=js(t,n,r),a=` + value = max(x_val, value); + `,o="",l=z("x",t.dataType,t.dims.length),u=["rank"],[d,c,p,h,f]=Ws(i,s);return d.push(...F(t.dims,i)),{name:e,shaderCache:{hint:`${n.cacheKey};${p};${h};${f}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(C.size(i)/64)},programUniforms:d}),getShaderSource:m=>Hs(m,l,t.dims.length,i.length,s,a,o,t.dataType===10?-65504:-1e5,c,p,h,f)}},Nd=(e,t)=>{er(e.inputs),e.compute(Zs("MaxPool",e.inputs[0],!1,t))},Ud=e=>{let t=e.storage_order,r=e.dilations,n=Qs(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(n.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let s={storageOrder:t,dilations:r,...n,cacheKey:""};return{...s,cacheKey:Bd(s)}},qd=e=>{let t=e.format;return{format:t,...Ys,cacheKey:t}},Gd=(e,t)=>{er(e.inputs),e.compute(Zs("GlobalMaxPool",e.inputs[0],!0,t))}}),Vd,jd,Wd,mm=A(()=>{Ue(),j(),Y(),Vd=(e,t,r)=>{let n=e===t,s=et&&r>0;if(n||s||i)throw new Error("Range these inputs' contents are invalid.")},jd=(e,t,r,n)=>{let s=Math.abs(Math.ceil((t-e)/r)),i=[s],a=s,o=[{type:12,data:a},{type:n,data:e},{type:n,data:r},...F(i)],l=u=>{let d=N("output",n,i.length),c=d.type.value,p=[{name:"outputSize",type:"u32"},{name:"start",type:c},{name:"delta",type:c}];return` + ${u.registerUniforms(p).declareVariables(d)} + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + output[global_idx] = uniforms.start + ${c}(global_idx) * uniforms.delta; + }`};return{name:"Range",shaderCache:{hint:`${n}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:i,dataType:n}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:o})}},Wd=e=>{let t=0,r=0,n=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],n=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],n=e.inputs[2].getFloat32Array()[0]),se.webgpu.validateInputContent&&Vd(t,r,n),e.compute(jd(t,r,n,e.inputs[0].dataType),{inputs:[]})}}),Hd,Kd,Qd,Xd,Yd,Zd,Jd,ec,tc,rc,nc,Js,sc,ic,ac,oc,lc,uc,dc,gm=A(()=>{j(),J(),_e(),Y(),Hd=(e,t)=>{if(e.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),e.length>0){if(t.mode==="linear"){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and + one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(t.mode==="cubic"&&!(e.length===2||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},Kd=(e,t,r)=>{t.every(s=>s>=0&&s{throw new Error("Resize requires axes input values to be positive and less than rank")}));let n=new Array(r).fill(1);return t.forEach((s,i)=>n[s]=e[i]),n},Qd=(e,t,r,n,s,i)=>{let[a,o,l]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],u=e[0].dims.length;if(a>0&&e.length>a&&e[a].dims.length>0)e[a].getFloat32Array().forEach(d=>i.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(o>0&&e.length>o&&e[o].dims.length>0){if(e[o].getFloat32Array().forEach(d=>n.push(d)),n.length!==0&&n.length!==u&&r>=18&&n.length!==t.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");Hd(n,t),t.axes.length>0&&Kd(n,t.axes,u).forEach((d,c)=>n[c]=d)}if(l>0&&e.length>l&&(e[l].getBigInt64Array().forEach(d=>s.push(Number(d))),s.length!==u||r>=18&&s.length===t.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(t.axes.length>0){if(n.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(s.length!==t.axes.length)throw new Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if(typeof n<"u"&&typeof s<"u"&&n.length>0&&s.length>u)throw new Error("Resize requires only of scales or sizes to be specified")},Xd=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32, + lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`return ${t}(xResized) / ${t}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) { + return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5; + } else { + return 0.0; + }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) { + return 0.0; + } else { + // The whole part and the fractional part are calculated separately due to inaccuracy of floating + // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an + // offset-by-one error later in floor(). + let whole = ${t}(xResized * (lengthOriginal - 1) / (lengthResized - 1)); + let fract = + ${t}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${t}(lengthResized - 1); + return whole + fract; + }`;case"tf_crop_and_resize":return`if (lengthResized > 1) { + return ${t}(roiStart) * ${t}(lengthOriginal - 1) + + (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) / + ${t}(lengthResized - 1); + } else { + return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1); + }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized); + const adjustment = ${t}(lengthResized) / outputWidth; + const center = ${t}(lengthOriginal) / 2; + const offset = center * (1 - adjustment); + return offset + ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;case"half_pixel":return`return ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${e} is not supported`)}})()+"}",Yd=(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`)}})()+"}",Zd=(e,t,r)=>{let n=new Array(r).fill(0).concat(new Array(r).fill(1)),s=e.length===0?n:e.slice();return t.length>0?(t.forEach((i,a)=>{n[i]=s[a],n[a+r]=s[t.length+a]}),n):s},Jd=(e,t,r,n)=>{let s=[];if(r.length>0)if(n.length>0){if(e.forEach(i=>s.push(i)),Math.max(...n)>e.length)throw new Error("axes is out of bound");n.forEach((i,a)=>s[i]=r[a])}else r.forEach(i=>s.push(i));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");s=e.map((i,a)=>Math.round(i*t[a]))}return s},ec=(e,t,r)=>{let n=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(i=>t[i]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(i=>t[i]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let s=e.slice();return r.axes.length>0?(r.axes.forEach(i=>t[i]=n),r.axes.forEach(i=>s[i]=Math.round(e[i]*t[i]))):(t.fill(n,0,t.length),s.forEach((i,a)=>s[a]=Math.round(i*t[a]))),s},tc=(e,t,r,n,s)=>` + fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> { + var original_indices: array<${e.type.value}, ${r.length}>; + for (var i:u32 = 0; i < ${r.length}; i++) { + var output_index = ${e.indicesGet("output_indices","i")}; + var scale = ${Q("uniforms.scales","i",n)}; + var roi_low = ${Q("uniforms.roi","i",s)}; + var roi_hi = ${Q("uniforms.roi",`i + ${t.length}`,s)}; + if (scale == 1.0) { + original_indices[i] = ${e.type.value}(output_index); + } else { + var input_shape_i = ${Q("uniforms.input_shape","i",t.length)}; + var output_shape_i = ${Q("uniforms.output_shape","i",r.length)}; + original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + } + } + return original_indices; + }`,rc=(e,t,r,n,s,i,a)=>` + fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} { + var input_indices: ${e.type.indices}; + for (var i:u32 = 0; i < ${n.length}; i++) { + var output_index = ${t.indicesGet("output_indices","i")}; + var input_index: u32; + var scale = ${Q("uniforms.scales","i",s)}; + if (scale == 1.0) { + input_index = output_index; + } else { + var roi_low = ${Q("uniforms.roi","i",i)}; + var roi_hi = ${Q("uniforms.roi",`i + ${r.length}`,i)}; + var input_shape_i = ${Q("uniforms.input_shape","i",r.length)}; + var output_shape_i = ${Q("uniforms.output_shape","i",n.length)}; + var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + if (!${a} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) { + if (original_idx < 0) { + input_index = 0; + } else if (original_idx > ${t.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); + } + } + ${e.indicesSet("input_indices","i"," input_index")} + } + return input_indices; + }`,nc=(e,t)=>` + fn checkInputIndices(input_indices: ${e.type.indices}) -> bool { + for (var i:u32 = 0; i < ${t.length}; i++) { + var input_index = ${e.indicesGet("input_indices","i")}; + if (input_index < 0 || input_index >= ${Q("uniforms.input_shape","i",t.length)}) { + return false; + } + } + return true; + }`,Js=(e,t,r,n)=>e.rank>n?` + ${e.indicesSet("input_indices",t,"channel")}; + ${e.indicesSet("input_indices",r,"batch")}; +`:"",sc=(e,t,r,n,s)=>{let[i,a,o,l]=r.length===2?[-1,0,1,-1]:[0,2,3,1],u=e.type.value;return` + fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${u} { + var input_indices: ${e.type.indices}; + ${e.indicesSet("input_indices",a,`max(0, min(row, ${r[a]} - 1))`)}; + ${e.indicesSet("input_indices",o,`max(0, min(col, ${r[o]} - 1))`)}; + ${Js(e,l,i,2)} + return ${e.getByIndices("input_indices")}; + } + + fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${u} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var row:${u} = originalIndices[${a}]; + var col:${u} = originalIndices[${o}]; + ${n?`if (row < 0 || row > (${r[a]} - 1) || col < 0 || col > (${r[o]} - 1)) { + return ${s}; + }`:""}; + row = max(0, min(row, ${r[a]} - 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); + }`},ic=(e,t,r,n,s,i,a,o,l,u)=>{let d=r.length===2,[c,p]=d?[0,1]:[2,3],h=e.type.value,f=m=>{let g=m===c?"row":"col";return` + fn ${g}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${h} { + var output_index = ${t.indicesGet("output_indices",m)}; + var originalIdx: ${h} = getOriginalCoordinateFromResizedCoordinate(output_index, ${s[m]}, + ${n[m]}, ${r[m]}, ${i[m]}, ${i[m]} + ${r.length}); + var fractOriginalIdx: ${h} = originalIdx - floor(originalIdx); + var coefs = getCubicInterpolationCoefs(fractOriginalIdx); + + if (${o} && (originalIdx < 0 || originalIdx > (${r[m]} - 1))) { + return ${l}; + } + var data: array<${h}, 4> = array<${h}, 4>(0.0, 0.0, 0.0, 0.0); + for (var i: i32 = -1; i < 3; i++) { + var ${g}: ${h} = originalIdx + ${h}(i); + if (${g} < 0 || ${g} >= ${r[m]}) { + ${u?`coefs[i + 1] = 0.0; + continue;`:o?`return ${l};`:`${g} = max(0, min(${g}, ${r[m]} - 1));`}; + } + var input_indices_copy: ${e.type.indices} = input_indices; + ${e.indicesSet("input_indices_copy",m,`u32(${g})`)}; + data[i + 1] = ${m===c?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"}; + } + return cubicInterpolation1D(data, coefs); + }`};return` + ${f(c)}; + ${f(p)}; + fn getCubicInterpolationCoefs(s: ${h}) -> array<${h}, 4> { + var absS = abs(s); + var coeffs: array<${h}, 4> = array<${h}, 4>(0.0, 0.0, 0.0, 0.0); + var oneMinusAbsS: ${h} = 1.0 - absS; + var twoMinusAbsS: ${h} = 2.0 - absS; + var onePlusAbsS: ${h} = 1.0 + absS; + coeffs[0] = ((${a} * onePlusAbsS - 5 * ${a}) * onePlusAbsS + 8 * ${a}) * onePlusAbsS - 4 * ${a}; + coeffs[1] = ((${a} + 2) * absS - (${a} + 3)) * absS * absS + 1; + coeffs[2] = ((${a} + 2) * oneMinusAbsS - (${a} + 3)) * oneMinusAbsS * oneMinusAbsS + 1; + coeffs[3] = ((${a} * twoMinusAbsS - 5 * ${a}) * twoMinusAbsS + 8 * ${a}) * twoMinusAbsS - 4 * ${a}; + return coeffs; + } + + fn cubicInterpolation1D(x: array<${h}, 4>, coefs: array<${h}, 4>) -> ${h} { + var coefsSum: ${h} = 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: ${t.type.indices}) -> ${h} { + var input_indices: ${e.type.indices} = output_indices; + return colCubicInterpolation(input_indices, output_indices); + } + `},ac=(e,t,r,n,s)=>{let[i,a,o,l,u]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],d=e.type.value;return` + fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${d} { + var input_indices: ${e.type.indices}; + ${e.indicesSet("input_indices",a,`max(0, min(depth, ${r[a]} - 1))`)}; + ${e.indicesSet("input_indices",o,`max(0, min(height, ${r[o]} - 1))`)}; + ${e.indicesSet("input_indices",l,`max(0, min(width, ${r[l]} - 1))`)}; + ${Js(e,u,i,3)} + return ${e.getByIndices("input_indices")}; + } + + fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${d} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var depth:${d} = originalIndices[${a}]; + var height:${d} = originalIndices[${o}]; + var width:${d} = originalIndices[${l}]; + ${n?`if (depth < 0 || depth > (${r[a]} - 1) || height < 0 || height > (${r[o]} - 1) || width < 0 || (width > ${r[l]} - 1)) { + return ${s}; + }`:""}; + + depth = max(0, min(depth, ${r[a]} - 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: ${d} = getInputValue(batch, channel, depth1, height1, width1); + var x112: ${d} = getInputValue(batch, channel, depth1, height1, width2); + var x121: ${d} = getInputValue(batch, channel, depth1, height2, width1); + var x122: ${d} = getInputValue(batch, channel, depth1, height2, width2); + var x211: ${d} = getInputValue(batch, channel, depth2, height1, width1); + var x212: ${d} = getInputValue(batch, channel, depth2, height1, width2); + var x221: ${d} = getInputValue(batch, channel, depth2, height2, width1); + var x222: ${d} = getInputValue(batch, channel, depth2, height2, width2); + var dx1: ${d} = abs(depth - ${d}(depth1)); + var dx2: ${d} = abs(${d}(depth2) - depth); + var dy1: ${d} = abs(height - ${d}(height1)); + var dy2: ${d} = abs(${d}(height2) - height); + var dz1: ${d} = abs(width - ${d}(width1)); + var dz2: ${d} = abs(${d}(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); + }`},oc=(e,t,r,n,s,i)=>{let a=e.dims,o=Zd(i,t.axes,a.length),l=Jd(a,n,s,t.axes),u=n.slice();n.length===0&&(u=a.map((_,v)=>_===0?1:l[v]/_),t.keepAspectRatioPolicy!=="stretch"&&(l=ec(a,u,t)));let d=N("output",e.dataType,l.length),c=z("input",e.dataType,a.length),p=C.size(l),h=a.length===l.length&&a.every((_,v)=>_===l[v]),f=t.coordinateTransformMode==="tf_crop_and_resize",m=t.extrapolationValue,g=c.type.value,w=_=>` + ${h?"":` + ${Xd(t.coordinateTransformMode,g)}; + ${(()=>{switch(t.mode){case"nearest":return` + ${nc(c,a)}; + ${Yd(t.nearestMode,r,g)}; + ${rc(c,d,a,l,u.length,o.length,f)}; + `;case"linear":return` + ${tc(d,a,l,u.length,o.length)}; + ${(()=>{if(a.length===2||a.length===4)return`${sc(c,d,a,f,m)}`;if(a.length===3||a.length===5)return`${ac(c,d,a,f,m)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()}; + `;case"cubic":return` + ${(()=>{if(a.length===2||a.length===4)return`${ic(c,d,a,l,u,o,t.cubicCoeffA,f,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()}; + `;default:throw Error("Invalid resize mode")}})()}; + `} + ${_.registerUniform("output_size","u32").registerUniform("scales","f32",u.length).registerUniform("roi","f32",o.length).declareVariables(c,d)} + ${_.mainStart()} + ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + ${h?"output[global_idx] = input[global_idx];":` + let output_indices = ${d.offsetToIndices("global_idx")}; + var input_indices: ${c.type.indices}; + ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices); + if (checkInputIndices(input_indices)) { + output[global_idx] = ${c.getByIndices("input_indices")}; + } else { + output[global_idx] = ${t.extrapolationValue}; + }`;case"linear":return`output[global_idx] = ${a.length===2||a.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()}; +`} + }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${u.length>0?u:""}|${s.length>0?s:""}|${o.length>0?o:""}|${h}|${a}`,inputDependencies:["rank"]},getShaderSource:w,getRunData:()=>({outputs:[{dims:l,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:[{type:12,data:p},{type:1,data:u},{type:1,data:o},...F(a,l)]})}},lc=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},uc=(e,t)=>{let r=[],n=[],s=[],i=lc(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");Qd(e.inputs,t,i,r,n,s),e.compute(oc(e.inputs[0],t,i,r,n,s),{inputs:[0]})},dc=e=>{let t=e.antialias,r=e.axes,n=e.coordinateTransformMode,s=e.cubicCoeffA,i=e.excludeOutside!==0,a=e.extrapolationValue,o=e.keepAspectRatioPolicy,l=e.mode,u=e.nearestMode===""?"simple":e.nearestMode;return ae({antialias:t,axes:r,coordinateTransformMode:n,cubicCoeffA:s,excludeOutside:i,extrapolationValue:a,keepAspectRatioPolicy:o,mode:l,nearestMode:u})}}),cc,pc,hc,_m=A(()=>{j(),J(),_e(),Y(),cc=(e,t)=>{let[r,n,s,i]=e,{numHeads:a,rotaryEmbeddingDim:o}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!C.areEqual(n.dims,[])&&!C.areEqual(n.dims,[1])&&n.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${n.dims.length}`);if(s.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${s.dims.length}`);if(i.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${i.dims.length}`);if(!C.areEqual(s.dims,i.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(o>0&&a===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],d=s.dims[0],c=C.sizeFromDimension(r.dims,1)/u,p=o===0?s.dims[1]*2:c/a;if(o>p)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(p/2!==s.dims[1]&&o/2!==s.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${s.dims[1]}`);if(u>d)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},pc=(e,t)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:s,scale:i}=t,a=e[0].dims[0],o=C.sizeFromDimension(e[0].dims,1),l=e[0].dims[e[0].dims.length-2],u=o/l,d=e[2].dims[1],c=s===0?d*2:u/n,p=new Array(a,l,u/c,c-d),h=C.computeStrides(p),f=[{type:1,data:i},{type:12,data:p},{type:12,data:h},...e[0].dims.length===3?new Array({type:12,data:[o,u,c,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[o,c,l*c,1]}):[],...F(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],m=g=>{let w=z("input",e[0].dataType,e[0].dims.length),_=z("position_ids",e[1].dataType,e[1].dims.length),v=z("cos_cache",e[2].dataType,e[2].dims.length),b=z("sin_cache",e[3].dataType,e[3].dims.length),y=N("output",e[0].dataType,e[0].dims.length);return g.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:p.length},{name:"global_strides",type:"u32",length:h.length},{name:"input_output_strides",type:"u32",length:h.length}]),` + ${g.declareVariables(w,_,v,b,y)} + + ${g.mainStart(Bt)} + let half_rotary_emb_dim = uniforms.${v.name}_shape[1]; + let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape; + let size = uniforms.global_shape[0] * uniforms.global_strides[0]; + ${g.guardAgainstOutOfBoundsWorkgroupSizes("size")} + + if (bsnh[3] < half_rotary_emb_dim) { + let position_ids_idx = + ${_.broadcastedIndicesToOffset("bsnh.xy",N("",_.type.tensor,2))}; + let position_id = + u32(${_.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 = ${w.getByOffset("i")} * ${v.get("position_id","bsnh[3]")} - + ${w.getByOffset("j")} * ${b.get("position_id","bsnh[3]")}; + ${y.setByOffset("i","re")} + let im = ${w.getByOffset("i")} * ${b.get("position_id","bsnh[3]")} + + ${w.getByOffset("j")} * ${v.get("position_id","bsnh[3]")}; + ${y.setByOffset("j","im")} + } else { + let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim; + ${y.setByOffset("k",w.getByOffset("k"))} + } + }`};return{name:"RotaryEmbedding",shaderCache:{hint:ae({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:m,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(C.size(p)/Bt)},programUniforms:f})}},hc=(e,t)=>{cc(e.inputs,t),e.compute(pc(e.inputs,t))}}),fc,mc,gc,wm=A(()=>{j(),J(),Y(),fc=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],n=e[2];if(t.dataType!==r.dataType||t.dataType!==n.dataType)throw new Error("All inputs must have the same data type");if(t.dims.length!==3&&t.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let s=t.dims[t.dims.length-1],i=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==s)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]!==s)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]!==s)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]!==s)throw new Error("Bias must have the same hidden size as input")}},mc=(e,t,r,n)=>{let s=t.simplified,i=e[0].dims,a=C.size(i),o=i,l=a,u=i.slice(-1)[0],d=n?i.slice(0,-1).concat(1):[],c=!s&&e.length>3,p=e.length>4,h=n&&r>1,f=n&&r>2,m=r>3,g=64,w=ge(u),_=[{type:12,data:l},{type:12,data:w},{type:12,data:u},{type:1,data:t.epsilon}],v=y=>{let S=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],x=[z("x",e[0].dataType,e[0].dims,w),z("skip",e[1].dataType,e[1].dims,w),z("gamma",e[2].dataType,e[2].dims,w)];c&&x.push(z("beta",e[3].dataType,e[3].dims,w)),p&&x.push(z("bias",e[4].dataType,e[4].dims,w)),x.push(N("output",e[0].dataType,o,w)),h&&x.push(N("mean_output",1,d)),f&&x.push(N("inv_std_output",1,d)),m&&x.push(N("input_skip_bias_sum",e[0].dataType,o,w));let I=ye(e[0].dataType),D=ye(1,w);return` + + ${y.registerUniforms(S).declareVariables(...x)} + var sum_shared : array<${D}, ${g}>; + var sum_squared_shared : array<${D}, ${g}>; + + ${y.mainStart([g,1,1])} + let ix = local_id.x; + let iy = global_id.x / ${g}; + + let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components; + var stride = hidden_size_vectorized / ${g}; + let offset = ix * stride + iy * hidden_size_vectorized; + let offset1d = stride * ix; + if (ix == ${g-1}) { + stride = hidden_size_vectorized - stride * ix; + } + for (var i: u32 = 0; i < stride; i++) { + let skip_value = skip[offset + i]; + let bias_value = ${p?"bias[offset1d + i]":I+"(0.0)"}; + let input_value = x[offset + i]; + let value = input_value + skip_value + bias_value; + ${m?"input_skip_bias_sum[offset + i] = value;":""} + output[offset + i] = value; + let f32_value = ${Rt(I,w,"value")}; + sum_shared[ix] += f32_value; + sum_squared_shared[ix] += f32_value * f32_value; + } + workgroupBarrier(); + + var reduce_size : u32 = ${g}; + for (var curr_size = reduce_size >> 1; curr_size > 0; curr_size = reduce_size >> 1) { + reduce_size = curr_size + (reduce_size & 1); + if (ix < curr_size) { + sum_shared[ix] += sum_shared[ix + reduce_size]; + sum_squared_shared[ix] += sum_squared_shared[ix + reduce_size]; + } + workgroupBarrier(); + } + + let sum = sum_shared[0]; + let square_sum = sum_squared_shared[0]; + let mean = ${dt("sum",w)} / f32(uniforms.hidden_size); + let inv_std_dev = inverseSqrt(${dt("square_sum",w)} / f32(uniforms.hidden_size) ${s?"":"- mean * mean"} + uniforms.epsilon); + ${h?"mean_output[global_idx] = mean;":""} + ${f?"inv_std_output[global_idx] = inv_std_dev;":""} + + for (var i: u32 = 0; i < stride; i++) { + output[offset + i] = (output[offset + i] ${s?"":`- ${I}(mean)`}) * + ${I}(inv_std_dev) * gamma[offset1d + i] + ${c?"+ beta[offset1d + i]":""}; + } + }`},b=[{dims:o,dataType:e[0].dataType}];return r>1&&b.push({dims:d,dataType:1}),r>2&&b.push({dims:d,dataType:1}),r>3&&b.push({dims:i,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${w};${h};${f};${m}`,inputDependencies:e.map((y,S)=>"type")},getShaderSource:v,getRunData:()=>({outputs:b,dispatchGroup:{x:Math.ceil(l/u)},programUniforms:_})}},gc=(e,t)=>{fc(e.inputs);let r=[0];e.outputCount>1&&r.push(-3),e.outputCount>2&&r.push(-3),e.outputCount>3&&r.push(3),e.compute(mc(e.inputs,t,e.outputCount,!1),{outputs:r})}}),_c,tr,wc,ei,yc,bc,vc,$c,ym=A(()=>{j(),J(),_e(),Y(),_c=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(t.starts.length!==t.ends.length)throw new Error("starts and ends must have the same length");e.slice(1).forEach((r,n)=>{if(e[n+1].dataType!==6&&e[n+1].dataType!==7)throw new Error(`Input ${n} must be an array of int32 or int64`)})},tr=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(n=>r.push(Number(n)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(n=>r.push(Number(n)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},wc=(e,t)=>{if(e.length>1){let r=tr(e,1),n=tr(e,2),s=tr(e,3);return s.length===0&&(s=[...Array(e[0].dims.length).keys()]),ae({starts:r,ends:n,axes:s})}else return t},ei=(e,t,r,n,s)=>{let i=e;return e<0&&(i+=r[n[t]]),s[t]<0?Math.max(0,Math.min(i,r[n[t]]-1)):Math.max(0,Math.min(i,r[n[t]]))},yc=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} { + var input_indices: ${e.type.indices}; + var carry = 0u; + for (var i = ${r.length}; i >= 0; i--) { + let input_shape_i = ${Q("uniforms.input_shape","i",r.length)}; + let steps_i = ${Q("uniforms.steps","i",r.length)}; + let signs_i = ${Q("uniforms.signs","i",r.length)}; + let starts_i = ${Q("uniforms.starts","i",r.length)}; + var output_index = ${t.indicesGet("output_indices","i")}; + var input_index = output_index * steps_i + starts_i + carry; + carry = input_index / input_shape_i; + input_index = input_index % input_shape_i; + if (signs_i < 0) { + input_index = input_shape_i - input_index - 1u + starts_i; + } + ${e.indicesSet("input_indices","i","input_index")}; + } + return input_indices; + }`,bc=(e,t)=>{let r=e[0].dims,n=C.size(r),s=t.axes.length>0?C.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],i=tr(e,4);i.forEach(w=>w!==0||(()=>{throw new Error("step cannot be 0")})),i.length===0&&(i=Array(s.length).fill(1));let a=t.starts.map((w,_)=>ei(w,_,r,s,i)),o=t.ends.map((w,_)=>ei(w,_,r,s,i));if(s.length!==a.length||s.length!==o.length)throw new Error("start, ends and axes should have the same number of elements");if(s.length!==r.length)for(let w=0;wMath.sign(w));i.forEach((w,_,v)=>{if(w<0){let b=(o[_]-a[_])/w,y=a[_],S=y+b*i[_];a[_]=S,o[_]=y,v[_]=-w}});let u=r.slice(0);s.forEach((w,_)=>{u[w]=Math.ceil((o[w]-a[w])/i[w])});let d={dims:u,dataType:e[0].dataType},c=N("output",e[0].dataType,u.length),p=z("input",e[0].dataType,e[0].dims.length),h=C.size(u),f=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:a.length},{name:"signs",type:"i32",length:l.length},{name:"steps",type:"u32",length:i.length}],m=[{type:12,data:h},{type:12,data:a},{type:6,data:l},{type:12,data:i},...F(e[0].dims,u)],g=w=>` + ${w.registerUniforms(f).declareVariables(p,c)} + ${yc(p,c,r)} + ${w.mainStart()} + ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + let output_indices = ${c.offsetToIndices("global_idx")}; + let input_indices = calculateInputIndices(output_indices); + ${c.setByOffset("global_idx",p.getByIndices("input_indices"))} + }`;return{name:"Slice",shaderCache:{hint:`${l.length}_${a.length}_${i.length}`,inputDependencies:["rank"]},getShaderSource:g,getRunData:()=>({outputs:[d],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:m})}},vc=(e,t)=>{_c(e.inputs,t);let r=wc(e.inputs,t);e.compute(bc(e.inputs,r),{inputs:[0]})},$c=e=>{let t=e.starts,r=e.ends,n=e.axes;return ae({starts:t,ends:r,axes:n})}}),xc,kc,Sc,Ec,bm=A(()=>{j(),J(),_e(),Y(),xc=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},kc=(e,t)=>{let r=e.dims,n=C.size(r),s=64,i=t.axis;if(i<0&&(i=r.length+i),iw===4?`max(max(${g}.x, ${g}.y), max(${g}.z, ${g}.w))`:w===2?`max(${g}.x, ${g}.y)`:w===3?`max(max(${g}.x, ${g}.y), ${g}.z)`:g,c=z("x",e.dataType,e.dims,l),p=N("result",e.dataType,e.dims,l),h=c.type.value,f=ye(e.dataType)==="f32"?`var threadMax = ${h}(-3.402823e+38f);`:`var threadMax = ${h}(-65504.0h);`,m=g=>` + var rowMaxShared : ${h}; + var rowSumShared : ${h}; + var threadShared : array<${h}, ${s}>; + + fn getValue(row: i32, col: i32, row_stride: i32) -> ${h} { + let index = row * row_stride + col; + return x[index]; + } + + fn setValue(row: i32, col: i32, row_stride: i32, value: ${h}) { + let index = row * row_stride + col; + result[index] = value; + } + ${g.registerUniform("packedCols","i32").declareVariables(c,p)} + ${g.mainStart()} + let gindex = i32(global_idx); + let lindex = i32(local_idx); + const wg = ${s}; + let row = gindex / wg; + let cols = uniforms.packedCols; + let row_stride : i32 = uniforms.packedCols; + + // find the rows max + ${f} + 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 = ${h}(${d("threadShared[0]",l)}); + } + workgroupBarrier(); + + // find the rows sum + var threadSum = ${h}(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 = ${h}(${dt("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:e.dataType}],dispatchGroup:{x:o},programUniforms:[{type:6,data:u}]}),getShaderSource:m}},Sc=(e,t)=>{xc(e.inputs),e.compute(kc(e.inputs[0],t))},Ec=e=>ae({axis:e.axis})}),Tc,Ic,Mc,Cc,zc,Ac,Pc,vm=A(()=>{j(),J(),_e(),Y(),Tc=e=>{if(!e||e.length<1)throw new Error("too few inputs")},Ic=(e,t)=>{let r=[],n=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(s=>r.push(Number(s))),n=r.length),ae({numOutputs:n,axis:t.axis,splitSizes:r})},Mc=e=>` +fn calculateOutputIndex(index: u32) -> u32 { + for (var i: u32 = 0u; i < ${e}u; i += 1u ) { + if (index < ${Q("uniforms.size_in_split_axis","i",e)}) { + return i; + } + } + return ${e}u; +}`,Cc=e=>{let t=e.length,r=[];for(let n=0;n{let r=e[0].dims,n=C.size(r),s=e[0].dataType,i=C.normalizeAxis(t.axis,r.length),a=new Array(t.numOutputs),o=z("input",s,r.length),l=new Array(t.numOutputs),u=[],d=[],c=0,p=[{type:12,data:n}];for(let f=0;f` + ${f.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",l.length).declareVariables(o,...a)} + ${Mc(l.length)} + ${Cc(a)} + + ${f.mainStart()} + ${f.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 -= ${Q("uniforms.size_in_split_axis","output_number - 1u",l.length)}; + ${o.indicesSet("indices",i,"index")}; + } + writeBufferData(output_number, indices, global_idx); + }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:h,getRunData:()=>({outputs:u,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:p})}},Ac=(e,t)=>{Tc(e.inputs);let r=e.inputs.length===1?t:Ic(e.inputs,t);e.compute(zc(e.inputs,r),{inputs:[0]})},Pc=e=>{let t=e.axis,r=e.splitSizes,n=e.numOutputs<0?r.length:e.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return ae({axis:t,numOutputs:n,splitSizes:r})}}),Oc,Bc,Rc,$m=A(()=>{j(),J(),Y(),Oc=(e,t,r,n,s)=>{let i=N("output_data",s,r.length,4),a=z("a_data",t[1].dataType,t[1].dims.length,4),o=z("b_data",t[2].dataType,t[2].dims.length,4),l=z("c_data",t[0].dataType,t[0].dims.length,4),u,d=(c,p,h)=>`select(${p}, ${c}, ${h})`;if(!n)u=i.setByOffset("global_idx",d(a.getByOffset("global_idx"),o.getByOffset("global_idx"),l.getByOffset("global_idx")));else{let c=(p,h,f="")=>{let m=`a_data[index_a${h}][component_a${h}]`,g=`b_data[index_b${h}][component_b${h}]`,w=`bool(c_data[index_c${h}] & (0xffu << (component_c${h} * 8)))`;return` + let output_indices${h} = ${i.offsetToIndices(`global_idx * 4u + ${h}u`)}; + let offset_a${h} = ${a.broadcastedIndicesToOffset(`output_indices${h}`,i)}; + let offset_b${h} = ${o.broadcastedIndicesToOffset(`output_indices${h}`,i)}; + let offset_c${h} = ${l.broadcastedIndicesToOffset(`output_indices${h}`,i)}; + let index_a${h} = offset_a${h} / 4u; + let index_b${h} = offset_b${h} / 4u; + let index_c${h} = offset_c${h} / 4u; + let component_a${h} = offset_a${h} % 4u; + let component_b${h} = offset_b${h} % 4u; + let component_c${h} = offset_c${h} % 4u; + ${p}[${h}] = ${f}(${d(m,g,w)}); + `};s===9?u=` + var data = vec4(0); + ${c("data",0,"u32")} + ${c("data",1,"u32")} + ${c("data",2,"u32")} + ${c("data",3,"u32")} + output_data[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:u=` + ${c("output_data[global_idx]",0)} + ${c("output_data[global_idx]",1)} + ${c("output_data[global_idx]",2)} + ${c("output_data[global_idx]",3)} + `}return` + ${e.registerUniform("vec_size","u32").declareVariables(l,a,o,i)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${u} + }`},Bc=e=>{let t=e[1].dims,r=e[2].dims,n=e[0].dims,s=e[1].dataType,i=!(C.areEqual(t,r)&&C.areEqual(r,n)),a=t,o=C.size(t);if(i){let u=Ot.calcShape(Ot.calcShape(t,r,!1),n,!1);if(!u)throw new Error("Can't perform where op on the given tensors");a=u,o=C.size(a)}let l=Math.ceil(o/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:u=>Oc(u,e,a,i,s),getRunData:()=>({outputs:[{dims:a,dataType:s}],dispatchGroup:{x:Math.ceil(o/64/4)},programUniforms:[{type:12,data:l},...F(n,t,r,a)]})}},Rc=e=>{e.compute(Bc(e.inputs))}}),Dc,xm=A(()=>{jf(),ys(),Wf(),Hf(),Kf(),Qf(),Lo(),uu(),em(),tm(),rm(),nm(),sm(),im(),am(),om(),lm(),um(),dm(),cm(),iu(),pm(),id(),hm(),fm(),mm(),gs(),gm(),_m(),wm(),ym(),bm(),vm(),ud(),Dt(),xs(),$m(),Dc=new Map([["Abs",[Jo]],["Acos",[el]],["Acosh",[tl]],["Add",[Fl]],["ArgMax",[Ao,ws]],["ArgMin",[zo,ws]],["Asin",[rl]],["Asinh",[nl]],["Atan",[sl]],["Atanh",[il]],["Attention",[Vo]],["AveragePool",[Dd,Rd]],["BatchNormalization",[Ko]],["BiasAdd",[Yo]],["BiasSplitGelu",[Rl]],["Cast",[ol,al]],["Ceil",[dl]],["Clip",[ul]],["Concat",[Ro,Do]],["Conv",[Os,Ps]],["ConvTranspose",[vu,gu]],["Cos",[cl]],["Cosh",[pl]],["CumSum",[xu,ku]],["DepthToSpace",[Iu,Mu]],["Div",[Nl]],["Einsum",[Bu,Ru]],["Elu",[hl,Xr]],["Equal",[Ul]],["Erf",[fl]],["Exp",[ml]],["Expand",[Nu]],["FastGelu",[qu]],["Floor",[gl]],["FusedConv",[Os,Ps]],["Gather",[Wu,ju]],["GatherElements",[Xu,Qu]],["Gelu",[_l]],["Gemm",[ed,Ju]],["GlobalAveragePool",[Fd,Ld]],["GlobalMaxPool",[Gd,qd]],["Greater",[jl]],["GreaterOrEqual",[Hl]],["GroupQueryAttention",[hd,cd]],["HardSigmoid",[Sl,kl]],["InstanceNormalization",[_d]],["LayerNormalization",[bd]],["LeakyRelu",[wl,Xr]],["Less",[Wl]],["LessOrEqual",[Kl]],["Log",[Pl]],["MatMul",[su]],["MatMulNBits",[xd,kd]],["MaxPool",[Nd,Ud]],["Mul",[ql]],["MultiHeadAttention",[sd,rd]],["Neg",[bl]],["Not",[yl]],["Pad",[Pd]],["Pow",[Gl]],["Range",[Wd]],["Reciprocal",[vl]],["ReduceMin",[Eo]],["ReduceMean",[vo]],["ReduceMax",[So]],["ReduceSum",[Io]],["ReduceProd",[To]],["ReduceL1",[$o]],["ReduceL2",[xo]],["ReduceLogSum",[Co]],["ReduceLogSumExp",[ko]],["ReduceSumSquare",[Mo]],["Relu",[$l]],["Resize",[uc,dc]],["RotaryEmbedding",[hc]],["Sigmoid",[xl]],["Sin",[El]],["Sinh",[Tl]],["Slice",[vc,$c]],["SkipLayerNormalization",[gc]],["Split",[Ac,Pc]],["Sqrt",[Il]],["Softmax",[Sc,Ec]],["Sub",[Vl]],["Tan",[Ml]],["Tanh",[Cl]],["ThresholdedRelu",[Al,Xr]],["Tile",[ld]],["Transpose",[Ua,qa]],["Where",[Rc]]])}),Lc,km=A(()=>{Ue(),vt(),Y(),Lc=class{constructor(e){this.backend=e,this.repo=new Map,this.attributesBound=!1}getArtifact(e){return this.repo.get(e)}setArtifact(e,t){this.repo.set(e,t)}run(e,t,r,n,s){Ne(e.programInfo.name);let i=this.backend.device,a=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let o=[];for(let u of t)o.push({binding:o.length,resource:{buffer:u.buffer}});for(let u of r)o.push({binding:o.length,resource:{buffer:u.buffer}});s&&o.push({binding:o.length,resource:s});let l=i.createBindGroup({layout:e.computePipeline.getBindGroupLayout(0),entries:o,label:e.programInfo.name});if(this.backend.sessionStatus==="capturing"){let u={kernelId:this.backend.currentKernelId,computePipeline:e.computePipeline,bindGroup:l,dispatchGroup:n};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(u)}a.setPipeline(e.computePipeline),a.setBindGroup(0,l),a.dispatchWorkgroups(...n),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),Re(e.programInfo.name)}dispose(){}build(e,t){Ne(e.name);let r=this.backend.device,n=[];r.features.has("shader-f16")&&n.push("enable f16;");let s=Da(t,this.backend.device.limits),i=e.getShaderSource(s),a=`${n.join(` +`)} +${s.additionalImplementations} +${i}`,o=r.createShaderModule({code:a,label:e.name});he("verbose",()=>`[WebGPU] ${e.name} shader code: ${a}`);let l=r.createComputePipeline({compute:{module:o,entryPoint:"main"},layout:"auto",label:e.name});return Re(e.name),{programInfo:e,computePipeline:l,uniformVariablesInfo:s.variablesInfo}}normalizeDispatchGroupSize(e){let t=typeof e=="number"?e:e.x,r=typeof e=="number"?1:e.y||1,n=typeof e=="number"?1:e.z||1,s=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(t<=s&&r<=s&&n<=s)return[t,r,n];let i=t*r*n,a=Math.ceil(Math.sqrt(i));if(a>s){if(a=Math.ceil(Math.cbrt(i)),a>s)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[a,a,a]}else return[a,a,1]}}}),Fc,Nc,Uc,qc,Sm=A(()=>{Ue(),j(),vt(),qf(),Gf(),xm(),km(),Fc=(e,t)=>{if(t.length!==e.length)throw new Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let r=[];for(let n=0;n{var s,i;let n=e.name;return(s=e.shaderCache)!=null&&s.hint&&(n+="["+e.shaderCache.hint+"]"),n+=":"+r+`:${Fc(t,((i=e.shaderCache)==null?void 0:i.inputDependencies)??new Array(t.length).fill("dims"))}`,n},Uc=class{constructor(e){e&&(this.architecture=e.architecture,this.vendor=e.vendor)}isArchitecture(e){return this.architecture===e}isVendor(e){return this.vendor===e}},qc=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 e=this.kernelCustomData.get(this.currentKernelId);return e||(e={},this.kernelCustomData.set(this.currentKernelId,e)),e}async initialize(e,t){this.env=e;let r=[],n={requiredLimits:{maxComputeWorkgroupStorageSize:t.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:t.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:t.limits.maxStorageBufferBindingSize,maxBufferSize:t.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:t.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:t.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:t.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:t.limits.maxComputeWorkgroupSizeZ},requiredFeatures:r};t.features.has("chromium-experimental-timestamp-query-inside-passes")?r.push("chromium-experimental-timestamp-query-inside-passes"):t.features.has("timestamp-query")&&r.push("timestamp-query"),t.features.has("shader-f16")&&r.push("shader-f16"),this.device=await t.requestDevice(n),this.adapterInfo=new Uc(await t.requestAdapterInfo()),this.gpuDataManager=Aa(this),this.programManager=new Lc(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,Ea(e.logLevel,!!e.debug),this.device.onuncapturederror=s=>{s.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${s.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value:t,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 e=this.getCommandEncoder(),t={};this.queryType==="at-passes"&&(t.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=e.beginComputePass(t)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;Ne(),this.endComputePass();let e;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),e=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(e,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,e,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&e.mapAsync(GPUMapMode.READ).then(()=>{var n;let t=new BigUint64Array(e.getMappedRange()),r=this.pendingQueries.get(e);for(let s=0;s"u"&&(this.queryTimeBase=h);let m=Number(h-this.queryTimeBase),g=Number(f-this.queryTimeBase);if(!Number.isSafeInteger(m)||!Number.isSafeInteger(g))throw new RangeError("incorrect timestamp range");if((n=this.env.webgpu.profiling)!=null&&n.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:c.map(w=>({dims:w.dims,dataType:bt(w.dataType)})),outputsMetadata:p.map(w=>({dims:w.dims,dataType:bt(w.dataType)})),kernelId:a,kernelType:l,kernelName:u,programName:d,startTime:m,endTime:g});else{let w="";c.forEach((v,b)=>{w+=`input[${b}]: [${v.dims}] | ${bt(v.dataType)}, `});let _="";p.forEach((v,b)=>{_+=`output[${b}]: [${v.dims}] | ${bt(v.dataType)}, `}),console.log(`[profiling] kernel "${a}|${l}|${u}|${d}" ${w}${_}execution time: ${g-m} ns`)}Ht("GPU",`${d}::${h}::${f}`)}e.unmap(),this.pendingQueries.delete(e)}),Re()}run(e,t,r,n,s,i){Ne(e.name);let a=[];for(let _=0;_v):r;if(d.length!==o.length)throw new Error(`Output size ${d.length} must be equal to ${o.length}.`);let c=[],p=[];for(let _=0;_=i)throw new Error(`Invalid output index: ${d[_]}`);if(d[_]===-3)continue;let v=d[_]===-1,b=d[_]===-2,y=v||b?s(o[_].dataType,o[_].dims):n(d[_],o[_].dataType,o[_].dims);if(c.push(y),y.data===0)continue;let S=this.gpuDataManager.get(y.data);if(!S)throw new Error(`no GPU data for output: ${y.data}`);if(v&&this.temporaryData.push(S),b){let x=this.kernelPersistentData.get(this.currentKernelId);x||(x=[],this.kernelPersistentData.set(this.currentKernelId,x)),x.push(S)}p.push(S)}if(a.length!==t.length||p.length!==c.length){if(p.length===0)return Re(e.name),c;throw new Error(`Program ${e.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let h;if(u){let _=0,v=[];u.forEach(x=>{let I=typeof x.data=="number"?[x.data]:x.data;if(I.length===0)return;let D=x.type===10?2:4,R,U;x.type===10?(U=I.length>4?16:I.length>2?8:I.length*D,R=I.length>4?16:D*I.length):(U=I.length<=2?I.length*D:16,R=16),_=Math.ceil(_/U)*U,v.push(_);let q=x.type===10?8:4;_+=I.length>4?Math.ceil(I.length/q)*R:I.length*D});let b=16;_=Math.ceil(_/b)*b;let y=new ArrayBuffer(_);u.forEach((x,I)=>{let D=v[I],R=typeof x.data=="number"?[x.data]:x.data;if(x.type===6)new Int32Array(y,D,R.length).set(R);else if(x.type===12)new Uint32Array(y,D,R.length).set(R);else if(x.type===10)new Uint16Array(y,D,R.length).set(R);else if(x.type===1)new Float32Array(y,D,R.length).set(R);else throw new Error(`Unsupported uniform type: ${bt(x.type)}`)});let S=this.gpuDataManager.create(_,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(S.buffer,0,y,0,_),this.gpuDataManager.release(S.id),h={offset:0,size:_,buffer:S.buffer}}let f=this.programManager.normalizeDispatchGroupSize(l),m=f[1]===1&&f[2]===1,g=Nc(e,t,m),w=this.programManager.getArtifact(g);if(w||(w=this.programManager.build(e,f),this.programManager.setArtifact(g,w),he("info",()=>`[artifact] key: ${g}, programName: ${e.name}`)),u&&w.uniformVariablesInfo){if(u.length!==w.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${w.uniformVariablesInfo.length}, got ${u.length} in program "${w.programInfo.name}".`);for(let _=0;_`[ProgramManager] run "${e.name}" (key=${g}) with ${f[0]}x${f[1]}x${f[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let _={kernelId:this.currentKernelId,programName:w.programInfo.name,inputTensorViews:t,outputTensorViews:c};this.pendingKernels.push(_),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(_)}return this.programManager.run(w,a,p,f,h),Re(e.name),c}upload(e,t){this.gpuDataManager.upload(e,t)}memcpy(e,t){this.gpuDataManager.memcpy(e,t)}async download(e,t){await this.gpuDataManager.download(e,t)}alloc(e){return this.gpuDataManager.create(e).id}free(e){return this.gpuDataManager.release(e)}createKernel(e,t,r,n){let s=Dc.get(e);if(!s)throw new Error(`kernel not implemented: ${e}`);let i={kernelType:e,kernelName:n,kernelEntry:s[0],attributes:[s[1],r]};this.kernels.set(t,i)}releaseKernel(e){let t=this.kernelPersistentData.get(e);if(t){for(let r of t)this.gpuDataManager.release(r.id);this.kernelPersistentData.delete(e)}this.kernelCustomData.delete(e),this.kernels.delete(e)}computeKernel(e,t,r){let n=this.kernels.get(e);if(!n)throw new Error(`kernel not created: ${e}`);let s=n.kernelType,i=n.kernelName,a=n.kernelEntry,o=n.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${s}] ${i}" is not allowed to be called recursively`);this.currentKernelId=e,o[0]&&(o[1]=o[0](o[1]),o[0]=void 0),he("info",()=>`[WebGPU] Start to run kernel "[${s}] ${i}"...`);let l=this.env.debug;this.temporaryData=[];try{return l&&this.device.pushErrorScope("validation"),a(t,o[1]),0}catch(u){return r.push(Promise.resolve(`[WebGPU] Kernel "[${s}] ${i}" failed. ${u}`)),1}finally{l&&r.push(this.device.popErrorScope().then(u=>u?`GPU validation error for kernel "[${s}] ${i}": ${u.message}`:null));for(let u of this.temporaryData)this.gpuDataManager.release(u.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(e,t,r,n){let s=this.sessionExternalDataMapping.get(e);s||(s=new Map,this.sessionExternalDataMapping.set(e,s));let i=s.get(t),a=this.gpuDataManager.registerExternalBuffer(r,n,i==null?void 0:i[1]);return s.set(t,[a,r]),a}unregisterBuffers(e){let t=this.sessionExternalDataMapping.get(e);t&&(t.forEach(r=>this.gpuDataManager.unregisterExternalBuffer(r[1])),this.sessionExternalDataMapping.delete(e))}getBuffer(e){let t=this.gpuDataManager.get(e);if(!t)throw new Error(`no GPU data for buffer: ${e}`);return t.buffer}createDownloader(e,t,r){return async()=>{let n=await ds(this,e,t);return Ia(n.buffer,r)}}writeTimestamp(e){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,e)}setQueryType(){var e;this.queryType="none",(((e=this.env.webgpu.profiling)==null?void 0:e.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(){he("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(){he("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){he("info","replay"),this.sessionStatus="replaying";let e=this.capturedCommandList.get(this.currentSessionId),t=this.capturedPendingKernels.get(this.currentSessionId),r=e.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(e){this.unregisterBuffers(e),this.capturedCommandList.has(e)&&this.capturedCommandList.delete(e),this.capturedPendingKernels.has(e)&&this.capturedPendingKernels.delete(e),this.gpuDataManager.onReleaseSession(e)}onRunStart(e){this.currentSessionId=e,this.setQueryType()}}}),Gc={};Dr(Gc,{init:()=>jc});var sn,Vc,jc,Em=A(()=>{j(),Sm(),vt(),J(),sn=class nf{constructor(t,r,n,s){this.module=t,this.dataType=r,this.data=n,this.dims=s}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=C.size(this.dims);return t===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,t)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let t=C.size(this.dims);return t===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,t)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let t=C.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(C.size(t)!==C.size(this.dims))throw new Error("Invalid new shape");return new nf(this.module,this.dataType,this.data,t)}},Vc=class{constructor(e,t,r){this.module=e,this.backend=t,this.customDataOffset=0,this.customDataSize=0,this.adapterInfo=t.adapterInfo;let n=e.HEAPU32,s=r>>>2;this.opKernelContext=n[s++];let i=n[s++];this.outputCount=n[s++],this.customDataOffset=n[s++],this.customDataSize=n[s++];let a=[];for(let o=0;otypeof o=="number"?this.inputs[o]:o))??this.inputs,n=(t==null?void 0:t.outputs)??[],s=(o,l,u)=>new sn(this.module,l,this.output(o,u),u),i=(o,l)=>{let u=Qt(o);if(!u)throw new Error(`Unsupported data type: ${o}`);let d=u*C.size(l),c=d>0?this.backend.gpuDataManager.create(d).id:0;return new sn(this.module,o,c,l)};return this.backend.run(e,r,n,s,i,this.outputCount)}output(e,t){let r=this.module.stackSave();try{let n=this.module.stackAlloc((1+t.length)*4),s=n>>2;this.module.HEAPU32[s++]=t.length;for(let i=0;i{let s=t.jsepInit;if(!s)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let i=new qc;await i.initialize(r,n),s("webgpu",[i,a=>i.alloc(a),a=>i.free(a),(a,o,l,u=!1)=>{if(u)he("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${a}, dst=${o}, size=${l}`),i.memcpy(a,o);else{he("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${a}, gpuDataId=${o}, size=${l}`);let d=t.HEAPU8.subarray(a>>>0,(a>>>0)+l);i.upload(o,d)}},async(a,o,l)=>{he("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${a}, dataOffset=${o}, size=${l}`),await i.download(a,()=>t.HEAPU8.subarray(o>>>0,(o>>>0)+l))},(a,o,l)=>i.createKernel(a,o,l,t.UTF8ToString(t._JsepGetNodeName(o))),a=>i.releaseKernel(a),(a,o,l,u)=>{he("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${l}, kernel=${a}, contextDataOffset=${o}`);let d=new Vc(t,i,o);return i.computeKernel(a,d,u)},()=>i.captureBegin(),()=>i.captureEnd(),()=>i.replay()])}else s("webnn")}}),Wc,ti,ri,ct,Hc,an,ni,si,ii,ai,oi,li,Kc=A(()=>{Nf(),Uf(),j(),Pt(),ts(),va(),Wc=(e,t)=>{me()._OrtInit(e,t)!==0&&le("Can't initialize onnxruntime.")},ti=async e=>{Wc(e.wasm.numThreads,qr(e.logLevel))},ri=async(e,t)=>{{let r=(Em(),Rn(Gc)).init;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let n=e.webgpu.adapter;if(n){if(typeof n.limits!="object"||typeof n.features!="object"||typeof n.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let s=e.webgpu.powerPreference;if(s!==void 0&&s!=="low-power"&&s!=="high-performance")throw new Error(`Invalid powerPreference setting: "${s}"`);let i=e.webgpu.forceFallbackAdapter;if(i!==void 0&&typeof i!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${i}"`);if(n=await navigator.gpu.requestAdapter({powerPreference:s,forceFallbackAdapter:i}),!n)throw new Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.')}await r("webgpu",me(),e,n)}if(t==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",me(),e)}}},ct=new Map,Hc=e=>{let t=me(),r=t.stackSave();try{let n=t.stackAlloc(8);return t._OrtGetInputOutputCount(e,n,n+4)!==0&&le("Can't get session input/output count."),[t.HEAP32[n/4],t.HEAP32[n/4+1]]}finally{t.stackRestore(r)}},an=e=>{let t=me(),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]},ni=async(e,t)=>{var c,p;let r,n,s=me();Array.isArray(e)?[r,n]=e:e.buffer===s.HEAPU8.buffer?[r,n]=[e.byteOffset,e.byteLength]:[r,n]=an(e);let i=0,a=0,o=0,l=[],u=[],d=[];try{if([a,l]=ba(t),(t==null?void 0:t.externalData)&&s.mountExternalData){let b=[];for(let y of t.externalData){let S=typeof y=="string"?y:y.path;b.push(as(typeof y=="string"?y:y.data).then(x=>{s.mountExternalData(S,x)}))}await Promise.all(b)}i=await s._OrtCreateSession(r,n,a),i===0&&le("Can't create a session.");let[h,f]=Hc(i),m=!!(t!=null&&t.enableGraphCapture),g=[],w=[],_=[];for(let b=0;bb==="gpu-buffer")&&(o=s._OrtCreateBinding(i),o===0&&le("Can't create IO binding."),v={handle:o,outputPreferredLocations:_,outputPreferredLocationsEncoded:_.map(b=>is(b))}),ct.set(i,[i,u,d,v,m,!1]),[i,g,w]}catch(h){throw u.forEach(f=>s._OrtFree(f)),d.forEach(f=>s._OrtFree(f)),o!==0&&s._OrtReleaseBinding(o),i!==0&&s._OrtReleaseSession(i),h}finally{s._free(r),a!==0&&s._OrtReleaseSessionOptions(a),l.forEach(h=>s._free(h)),(p=s.unmountExternalData)==null||p.call(s)}},si=e=>{var l;let t=me(),r=ct.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[n,s,i,a,o]=r;a&&(o&&t._OrtClearBoundOutputs(a.handle),t._OrtReleaseBinding(a.handle)),(l=t.jsepOnReleaseSession)==null||l.call(t,e),s.forEach(u=>t._OrtFree(u)),i.forEach(u=>t._OrtFree(u)),t._OrtReleaseSession(n),ct.delete(e)},ii=(e,t,r,n,s,i=!1)=>{if(!e){t.push(0);return}let a=me(),o=e[0],l=e[1],u=e[3],d,c;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 ${s} when enableGraphCapture is true.`);if(u==="gpu-buffer"){let f=e[2].gpuBuffer,m=Qt(rs(o));c=l.reduce((w,_)=>w*_,1)*m;let g=a.jsepRegisterBuffer;if(!g)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');d=g(n,s,f,c)}else{let f=e[2];if(Array.isArray(f)){c=4*f.length,d=a._malloc(c),r.push(d);let m=d/4;for(let g=0;ga.HEAP32[f++]=g);let m=a._OrtCreateTensor(rs(o),d,c,h,l.length,is(u));m===0&&le(`Can't create tensor for input/output. session=${n}, index=${s}.`),t.push(m)}finally{a.stackRestore(p)}},ai=async(e,t,r,n,s,i)=>{var R,U;let a=me(),o=ct.get(e);if(!o)throw new Error(`cannot run inference. invalid session id: ${e}`);let l=o[0],u=o[1],d=o[2],c=o[3],p=o[4],h=o[5],f=t.length,m=n.length,g=0,w=[],_=[],v=[],b=[],y=a.stackSave(),S=a.stackAlloc(f*4),x=a.stackAlloc(f*4),I=a.stackAlloc(m*4),D=a.stackAlloc(m*4);try{[g,w]=ma(i);for(let B=0;BPe*Oe,1);G=bt(rt);let Jh=c==null?void 0:c.outputPreferredLocations[n[B]];if(G==="string"){if(Jh==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Pe=[],Oe=ie/4;for(let Gt=0;Gt0){let Pe=a.jsepGetBuffer;if(!Pe)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let Oe=Pe(ie),Gt=Qt(rt);if(Gt===void 0||!ss(G))throw new Error(`Unsupported data type: ${G}`);L=!0,O.push([G,Ir,{gpuBuffer:Oe,download:a.jsepCreateDownloader(Oe,Mr*Gt,G),dispose:()=>{a._OrtReleaseTensor(K)}},"gpu-buffer"])}else{let Pe=ns(G),Oe=new Pe(Mr);new Uint8Array(Oe.buffer,Oe.byteOffset,Oe.byteLength).set(a.HEAPU8.subarray(ie,ie+Oe.byteLength)),O.push([G,Ir,Oe,"cpu"])}}finally{a.stackRestore(ee),G==="string"&&ie&&a._free(ie),L||a._OrtReleaseTensor(K)}}return c&&!p&&(a._OrtClearBoundOutputs(c.handle),ct.set(e,[l,u,d,c,p,!1])),O}finally{a.stackRestore(y),_.forEach(q=>a._OrtReleaseTensor(q)),v.forEach(q=>a._OrtReleaseTensor(q)),b.forEach(q=>a._free(q)),g!==0&&a._OrtReleaseRunOptions(g),w.forEach(q=>a._free(q))}},oi=e=>{let t=me(),r=ct.get(e);if(!r)throw new Error("invalid session id");let n=r[0],s=t._OrtEndProfiling(n);s===0&&le("Can't get an profile file name."),t._OrtFree(s)},li=e=>{let t=[];for(let r of e){let n=r[2];!Array.isArray(n)&&"buffer"in n&&t.push(n.buffer)}return t}}),pt,Ie,Lt,rr,nr,on,ui,ln,Tt,It,Qc,Xc,Yc,Zc,Jc,ep,tp,rp,np=A(()=>{Ue(),Kc(),Pt(),Fr(),pt=()=>!!se.wasm.proxy&&typeof document<"u",Lt=!1,rr=!1,nr=!1,ln=new Map,Tt=(e,t)=>{let r=ln.get(e);r?r.push(t):ln.set(e,[t])},It=()=>{if(Lt||!rr||nr||!Ie)throw new Error("worker not ready")},Qc=e=>{switch(e.data.type){case"init-wasm":Lt=!1,e.data.err?(nr=!0,ui[1](e.data.err)):(rr=!0,ui[0]()),on&&(URL.revokeObjectURL(on),on=void 0);break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let t=ln.get(e.data.type);e.data.err?t.shift()[1](e.data.err):t.shift()[0](e.data.out);break}}},Xc=async()=>{if(!rr){if(Lt)throw new Error("multiple calls to 'initWasm()' detected.");if(nr)throw new Error("previous call to 'initWasm()' failed.");if(Lt=!0,pt())return new Promise((e,t)=>{Ie==null||Ie.terminate(),ca().then(([r,n])=>{try{Ie=n,Ie.onerror=i=>t(i),Ie.onmessage=Qc,ui=[e,t];let s={type:"init-wasm",in:se};Ie.postMessage(s),on=r}catch(s){t(s)}},t)});try{await es(se.wasm),await ti(se),rr=!0}catch(e){throw nr=!0,e}finally{Lt=!1}}},Yc=async e=>{if(pt())return It(),new Promise((t,r)=>{Tt("init-ep",[t,r]);let n={type:"init-ep",in:{epName:e,env:se}};Ie.postMessage(n)});await ri(se,e)},Zc=async e=>pt()?(It(),new Promise((t,r)=>{Tt("copy-from",[t,r]);let n={type:"copy-from",in:{buffer:e}};Ie.postMessage(n,[e.buffer])})):an(e),Jc=async(e,t)=>{if(pt()){if(t!=null&&t.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return It(),new Promise((r,n)=>{Tt("create",[r,n]);let s={type:"create",in:{model:e,options:{...t}}},i=[];e instanceof Uint8Array&&i.push(e.buffer),Ie.postMessage(s,i)})}else return ni(e,t)},ep=async e=>{if(pt())return It(),new Promise((t,r)=>{Tt("release",[t,r]);let n={type:"release",in:e};Ie.postMessage(n)});si(e)},tp=async(e,t,r,n,s,i)=>{if(pt()){if(r.some(a=>a[3]!=="cpu"))throw new Error("input tensor on GPU is not supported for proxy.");if(s.some(a=>a))throw new Error("pre-allocated output tensor is not supported for proxy.");return It(),new Promise((a,o)=>{Tt("run",[a,o]);let l=r,u={type:"run",in:{sessionId:e,inputIndices:t,inputs:l,outputIndices:n,options:i}};Ie.postMessage(u,li(l))})}else return ai(e,t,r,n,s,i)},rp=async e=>{if(pt())return It(),new Promise((t,r)=>{Tt("end-profiling",[t,r]);let n={type:"end-profiling",in:e};Ie.postMessage(n)});oi(e)}}),di,sp,ip,Tm=A(()=>{Ue(),np(),j(),Wn(),va(),di=(e,t)=>{switch(e.location){case"cpu":return[e.type,e.dims,e.data,"cpu"];case"gpu-buffer":return[e.type,e.dims,{gpuBuffer:e.gpuBuffer},"gpu-buffer"];default:throw new Error(`invalid data location: ${e.location} for ${t()}`)}},sp=e=>{switch(e[3]){case"cpu":return new $e(e[0],e[2],e[1]);case"gpu-buffer":{let t=e[0];if(!ss(t))throw new Error(`not supported data type: ${t} for deserializing GPU tensor`);let{gpuBuffer:r,download:n,dispose:s}=e[2];return $e.fromGpuBuffer(r,{dataType:t,dims:e[1],download:n,dispose:s})}default:throw new Error(`invalid data location: ${e[3]}`)}},ip=class{async fetchModelAndCopyToWasmMemory(e){return Zc(await as(e))}async loadModel(e,t){Ne();let r;typeof e=="string"?r=await this.fetchModelAndCopyToWasmMemory(e):r=e,[this.sessionId,this.inputNames,this.outputNames]=await Jc(r,t),Re()}async dispose(){return ep(this.sessionId)}async run(e,t,r){Ne();let n=[],s=[];Object.entries(e).forEach(c=>{let p=c[0],h=c[1],f=this.inputNames.indexOf(p);if(f===-1)throw new Error(`invalid input '${p}'`);n.push(h),s.push(f)});let i=[],a=[];Object.entries(t).forEach(c=>{let p=c[0],h=c[1],f=this.outputNames.indexOf(p);if(f===-1)throw new Error(`invalid output '${p}'`);i.push(h),a.push(f)});let o=n.map((c,p)=>di(c,()=>`input "${this.inputNames[s[p]]}"`)),l=i.map((c,p)=>c?di(c,()=>`output "${this.outputNames[a[p]]}"`):null),u=await tp(this.sessionId,s,o,a,l,r),d={};for(let c=0;c{Ue(),np(),Tm(),Fr(),ap=()=>{if((typeof se.wasm.initTimeout!="number"||se.wasm.initTimeout<0)&&(se.wasm.initTimeout=0),se.wasm.simd===!1&&console.warn('Deprecated property "env.wasm.simd" is set to false. non-SIMD build is no longer provided, and this setting will be ignored.'),typeof se.wasm.proxy!="boolean"&&(se.wasm.proxy=!1),typeof se.wasm.trace!="boolean"&&(se.wasm.trace=!1),typeof se.wasm.numThreads!="number"||!Number.isInteger(se.wasm.numThreads)||se.wasm.numThreads<=0)if(typeof self<"u"&&!self.crossOriginIsolated)se.wasm.numThreads=1;else{let e=typeof navigator>"u"?$f("node:os").cpus().length:navigator.hardwareConcurrency;se.wasm.numThreads=Math.min(4,Math.ceil((e||1)/2))}se.wasm.wasmPaths===void 0&&qe&&qe.indexOf("blob:")!==0&&(se.wasm.wasmPaths=qe.substring(0,qe.lastIndexOf("/")+1))},op=class{async init(e){ap(),await Xc(),await Yc(e)}async createInferenceSessionHandler(e,t){let r=new ip;return await r.loadModel(e,t),Promise.resolve(r)}}}),lp={};Dr(lp,{wasmBackend:()=>up});var up,Mm=A(()=>{Im(),up=new op});Ue(),Ue(),Ue();var Cm="1.19.0-esmtest.20240513-a16cd2bd21",zm=sa;{let e=(Mm(),Rn(lp)).wasmBackend;wt("webgpu",e,5),wt("webnn",e,5),wt("cpu",e,10),wt("wasm",e,10)}Object.defineProperty(se.versions,"web",{value:Cm,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. + * ============================================================================= + */var Am=Object.freeze({__proto__:null,get InferenceSession(){return Vn},get TRACE(){return Ht},get TRACE_FUNC_BEGIN(){return Ne},get TRACE_FUNC_END(){return Re},get Tensor(){return $e},get TrainingSession(){return jn},default:zm,get env(){return se},get registerBackend(){return wt}});const Pm=(e,t)=>{const r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];const n=r.getContext("2d");if(n!=null){let s,i;(t==null?void 0:t.tensorLayout)!==void 0&&t.tensorLayout==="NHWC"?(s=e.dims[2],i=e.dims[3]):(s=e.dims[3],i=e.dims[2]);const a=(t==null?void 0:t.format)!==void 0?t.format:"RGB",o=t==null?void 0:t.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 d=i*s;let c=0,p=d,h=d*2,f=-1;a==="RGBA"?(c=0,p=d,h=d*2,f=d*3):a==="RGB"?(c=0,p=d,h=d*2):a==="RBG"&&(c=0,h=d,p=d*2);for(let m=0;m{const r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d");let n;if(r!=null){let s,i,a;(t==null?void 0:t.tensorLayout)!==void 0&&t.tensorLayout==="NHWC"?(s=e.dims[2],i=e.dims[1],a=e.dims[3]):(s=e.dims[3],i=e.dims[2],a=e.dims[1]);const o=t!==void 0&&t.format!==void 0?t.format:"RGB",l=t==null?void 0:t.norm;let u,d;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?d=[0,0,0,0]:typeof l.bias=="number"?d=[l.bias,l.bias,l.bias,l.bias]:(d=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(d[3]=l.bias[3]));const c=i*s;if(t!==void 0&&(t.format!==void 0&&a===4&&t.format!=="RGBA"||a===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");const p=4;let h=0,f=1,m=2,g=3,w=0,_=c,v=c*2,b=-1;o==="RGBA"?(w=0,_=c,v=c*2,b=c*3):o==="RGB"?(w=0,_=c,v=c*2):o==="RBG"&&(w=0,v=c,_=c*2),n=r.createImageData(s,i);for(let y=0;y{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");const{height:r,width:n}=t,s=t.norm??{mean:255,bias:0};let i,a;typeof s.mean=="number"?i=[s.mean,s.mean,s.mean,s.mean]:i=[s.mean[0],s.mean[1],s.mean[2],s.mean[3]??255],typeof s.bias=="number"?a=[s.bias,s.bias,s.bias,s.bias]:a=[s.bias[0],s.bias[1],s.bias[2],s.bias[3]??0];const o=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",u=r*n,d=l==="RGBA"?new Float32Array(u*4):new Float32Array(u*3);let c=4,p=0,h=1,f=2,m=3,g=0,w=u,_=u*2,v=-1;o==="RGB"&&(c=3,p=0,h=1,f=2,m=-1),l==="RGBA"?v=u*3:l==="RBG"?(g=0,_=u,w=u*2):l==="BGR"&&(_=0,w=u,g=u*2);for(let y=0;y{const r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,n=typeof ImageData<"u"&&e instanceof ImageData,s=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,i=typeof e=="string";let a,o=t??{};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=d=>d instanceof HTMLCanvasElement||d instanceof OffscreenCanvas?d.getContext("2d"):null;if(r){const d=l();d.width=e.width,d.height=e.height;const c=u(d);if(c!=null){let p=e.height,h=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(p=t.resizedHeight,h=t.resizedWidth),t!==void 0){if(o=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=p,o.width=h}else o.tensorFormat="RGBA",o.height=p,o.width=h;c.drawImage(e,0,0),a=c.getImageData(0,0,h,p).data}else throw new Error("Can not access image data")}else if(n){let d,c;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(d=t.resizedHeight,c=t.resizedWidth):(d=e.height,c=e.width),t!==void 0&&(o=t),o.format="RGBA",o.height=d,o.width=c,t!==void 0){const p=l();p.width=c,p.height=d;const h=u(p);if(h!=null)h.putImageData(e,0,0),a=h.getImageData(0,0,c,d).data;else throw new Error("Can not access image data")}else a=e.data}else if(s){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");const d=l();d.width=e.width,d.height=e.height;const c=u(d);if(c!=null){const p=e.height,h=e.width;return c.drawImage(e,0,0,h,p),a=c.getImageData(0,0,h,p).data,o.height=p,o.width=h,ci(a,o)}else throw new Error("Can not access image data")}else{if(i)return new Promise((d,c)=>{const p=l(),h=u(p);if(!e||!h)return c();const f=new Image;f.crossOrigin="Anonymous",f.src=e,f.onload=()=>{p.width=f.width,p.height=f.height,h.drawImage(f,0,0,p.width,p.height);const m=h.getImageData(0,0,p.width,p.height);o.height=p.height,o.width=p.width,d(ci(m.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(a!==void 0)return ci(a,o);throw new Error("Input data provided is not supported - aborted tensor creation")},Rm=(e,t)=>{const{width:r,height:n,download:s,dispose:i}=t,a=[1,n,r,4];return new Je({location:"texture",type:"float32",texture:e,dims:a,download:s,dispose:i})},Dm=(e,t)=>{const{dataType:r,dims:n,download:s,dispose:i}=t;return new Je({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:n,download:s,dispose:i})},Lm=(e,t,r)=>new Je({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]}),un=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["float16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),pi=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]);let dp=!1;const Fm=()=>{if(!dp){dp=!0;const e=typeof BigInt64Array<"u"&&typeof BigInt64Array.from=="function",t=typeof BigUint64Array<"u"&&typeof BigUint64Array.from=="function";e&&(un.set("int64",BigInt64Array),pi.set(BigInt64Array,"int64")),t&&(un.set("uint64",BigUint64Array),pi.set(BigUint64Array,"uint64"))}},Nm=e=>{let t=1;for(let r=0;r{switch(e.location){case"cpu":return new Je(e.type,e.data,t);case"cpu-pinned":return new Je({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new Je({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new Je({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}};let Je=class{constructor(t,r,n){Fm();let s,i;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,s=t.type,i=t.dims,t.location){case"cpu-pinned":{const o=un.get(s);if(!o)throw new TypeError(`unsupported type "${s}" to create tensor from pinned buffer`);if(!(t.data instanceof o))throw new TypeError(`buffer should be of type ${o.name}`);this.cpuData=t.data;break}case"texture":{if(s!=="float32")throw new TypeError(`unsupported type "${s}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break}case"gpu-buffer":{if(s!=="float32"&&s!=="float16"&&s!=="int32"&&s!=="int64"&&s!=="uint32"&&s!=="bool")throw new TypeError(`unsupported type "${s}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let o,l;if(typeof t=="string")if(s=t,l=n,t==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");o=r}else{const u=un.get(t);if(u===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(r)){if(t==="float16")throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");t==="uint64"||t==="int64"?o=u.from(r,BigInt):o=u.from(r)}else if(r instanceof u)o=r;else throw new TypeError(`A ${s} tensor's data must be type of ${u}`)}else if(l=r,Array.isArray(t)){if(t.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");const u=typeof t[0];if(u==="string")s="string",o=t;else if(u==="boolean")s="bool",o=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${u}.`)}else{const u=pi.get(t.constructor);if(u===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);s=u,o=t}if(l===void 0)l=[o.length];else if(!Array.isArray(l))throw new TypeError("A tensor's dims must be a number array");i=l,this.cpuData=o,this.dataLocation="cpu"}const a=Nm(i);if(this.cpuData&&a!==this.cpuData.length)throw new Error(`Tensor's size(${a}) does not match data length(${this.cpuData.length}).`);this.type=s,this.dims=i,this.size=a}static async fromImage(t,r){return Bm(t,r)}static fromTexture(t,r){return Rm(t,r)}static fromGpuBuffer(t,r){return Dm(t,r)}static fromPinnedBuffer(t,r,n){return Lm(t,r,n)}toDataURL(t){return Pm(this,t)}toImageData(t){return Om(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;const r=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=r,t&&this.disposer&&(this.disposer(),this.disposer=void 0),r}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Um(this,t)}};const qm=Je,sr=[];let hi,Mt;mt.IS_NODE_ENV?(Mt=oe??nt,sr.push("cpu"),hi=["cpu"]):(Mt=Am,mt.IS_WEBGPU_AVAILABLE&&sr.push("webgpu"),sr.push("wasm"),hi=["wasm"]);const Gm=Mt.InferenceSession;function Vm(e){let t=hi;if(e){if(!sr.includes(e))throw new Error(`Unsupported device: "${e}". Should be one of: ${sr.join(", ")}.`);t=[e]}return t}async function jm(e,t){return await Gm.create(e,t)}function cp(e){return e instanceof Mt.Tensor}const Ke=Mt==null?void 0:Mt.env;Ke!=null&&Ke.wasm&&(Ke.wasm.wasmPaths="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.19.0-dev.20240509-69cfcba38a/dist/",Ke.wasm.proxy=!mt.IS_WEBWORKER_ENV,(typeof crossOriginIsolated>"u"||!crossOriginIsolated)&&(Ke.wasm.numThreads=1),typeof navigator<"u"&&/iP(hone|od|ad).+16_4.+AppleWebKit/.test(navigator.userAgent)&&(Ke.wasm.simd=!1));function Wm(){var e;return(e=Ke==null?void 0:Ke.wasm)==null?void 0:e.proxy}Ee.backends.onnx=Ke;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 Z{constructor(...t){T(this,"ort_tensor");return cp(t[0])?this.ort_tensor=t[0]:this.ort_tensor=new qm(t[0],t[1],t[2]),new Proxy(this,{get:(r,n)=>{if(typeof n=="string"){let s=Number(n);if(Number.isInteger(s))return r._getitem(s)}return r[n]},set:(r,n,s)=>r[n]=s})}get dims(){return this.ort_tensor.dims}set dims(t){this.ort_tensor.dims=t}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}dispose(){this.ort_tensor.dispose()}*[Symbol.iterator](){const[t,...r]=this.dims;if(r.length>0){const n=r.reduce((s,i)=>s*i);for(let s=0;s0){const s=n.reduce((i,a)=>i*a);return this._subarray(t,s,n)}else return new Z(this.type,[this.data[t]],n)}indexOf(t){const r=this.data;for(let n=0;np)throw new Error(`Invalid slice: ${d}`);let h=[Math.max(c,0),Math.min(p,this.dims[u])];n.push(h),r.push(h[1]-h[0])}else throw new Error(`Invalid slice: ${d}`)}let s=n.map(([u,d])=>d-u),i=s.reduce((u,d)=>u*d);const a=this.data;let o=new a.constructor(i);const l=this.stride();for(let u=0;u=0;--c){const h=s[c];d+=(p%h+n[c][0])*l[c],p=Math.floor(p/h)}o[u]=a[d]}return new Z(this.type,o,r)}permute(...t){return Km(this,t)}transpose(...t){return this.permute(...t)}sum(t=null,r=!1){return this.norm(1,t,r)}norm(t="fro",r=null,n=!1){if(t==="fro")t=2;else if(typeof t=="string")throw Error(`Unsupported norm: ${t}`);const s=this.data;if(r===null){let o=s.reduce((l,u)=>l+u**t,0)**(1/t);return new Z(this.type,[o],[])}r=et(r,this.dims.length);const i=this.dims.slice();i[r]=1;const a=new s.constructor(s.length/this.dims[r]);for(let o=0;o=0;--u){const p=this.dims[u];if(u!==r){const h=d%p;l+=h*c,c*=i[u]}d=Math.floor(d/p)}a[l]+=s[o]**t}if(t!==1)for(let o=0;o=0;--o){const d=this.dims[o];if(o!==r){const c=l%d;a+=c*u,u*=this.dims[o]}l=Math.floor(l/d)}s[i]/=n.data[a]}return this}normalize(t=2,r=1){return this.clone().normalize_(t,r)}stride(){return Ym(this.dims)}squeeze(t=null){return new Z(this.type,this.data,hp(this.dims,t))}squeeze_(t=null){return this.dims=hp(this.dims,t),this}unsqueeze(t=null){return new Z(this.type,this.data,fp(this.dims,t))}unsqueeze_(t=null){return this.dims=fp(this.dims,t),this}flatten_(t=0,r=-1){r=(r+this.dims.length)%this.dims.length;let n=this.dims.slice(0,t),s=this.dims.slice(t,r+1),i=this.dims.slice(r+1);return this.dims=[...n,s.reduce((a,o)=>a*o,1),...i],this}flatten(t=0,r=-1){return this.clone().flatten_(t,r)}view(...t){let r=-1;for(let n=0;na!==r?s*i:s,1);t[r]=this.data.length/n}return new Z(this.type,this.data,t)}neg_(){const t=this.data;for(let r=0;ri*a);if(r!==n)throw Error(`cannot reshape array of size ${r} into shape (${t})`);let s=e;for(let i=t.length-1;i>=0;i--)s=s.reduce((a,o)=>{let l=a[a.length-1];return l.lengthr!==1):typeof t=="number"?e[t]===1&&e.splice(t,1):Array.isArray(t)&&(e=e.filter((r,n)=>r!==1||!t.includes(n))),e}function fp(e,t){return t=et(t,e.length+1),e=e.slice(),e.splice(t,0,1),e}function et(e,t,r=null,n=!0){if(n&&(e<-t||e>=t))throw new Error(`IndexError: index ${e} is out of bounds for dimension${r===null?"":" "+r} with size ${t}`);return e<0&&(e=(e%t+t)%t),e}function it(e,t=0){t=et(t,e[0].dims.length);const r=e[0].dims.slice();r[t]=e.reduce((a,o)=>a+o.dims[t],0);const n=r.reduce((a,o)=>a*o,1),s=new e[0].data.constructor(n),i=e[0].type;if(t===0){let a=0;for(let o of e)s.set(o.data,a),a+=o.data.length}else{let a=0;for(let o=0;o=0;--c){const f=l.dims[c];let m=p%f;c===t&&(m+=a),d+=m*h,h*=r[c],p=Math.floor(p/f)}s[d]=l.data[u]}a+=l.dims[t]}}return new Z(i,s,r)}function fi(e,t=0){return it(e.map(r=>r.unsqueeze(t)),t)}function Qm(e,t=null,r=1,n=!1){if(t===null){const u=e.data.reduce((h,f)=>h+f,0)/e.data.length,d=Math.sqrt(e.data.reduce((h,f)=>h+(f-u)**2,0)/(e.data.length-r)),c=new Z(e.type,[u],[]);return[new Z(e.type,[d],[]),c]}t=et(t,e.dims.length);const s=mi(e,t,n),i=e.dims.slice();i[t]=1;const a=new e.data.constructor(e.data.length/e.dims[t]);for(let l=0;l=0;--d){const h=e.dims[d];if(d!==t){const f=c%h;u+=f*p,p*=i[d]}c=Math.floor(c/h)}a[u]+=(e.data[l]-s.data[u])**2}for(let l=0;la+o,0);return new Z(e.type,[i/e.data.length],[])}t=et(t,e.dims.length);const n=e.dims.slice();n[t]=1;const s=new e.data.constructor(e.data.length/e.dims[t]);for(let i=0;i=0;--o){const d=e.dims[o];if(o!==t){const c=l%d;a+=c*u,u*=n[o]}l=Math.floor(l/d)}s[a]+=e.data[i]}if(e.dims[t]!==1)for(let i=0;i0||o>0;)switch(l.push(a-1),u.push(o-1),i[a][o].item()){case 0:--a,--o;break;case 1:--a;break;case 2:--o;break;default:throw new Error(`Internal error in dynamic time warping. Unexpected trace[${a}, ${o}]. Please file a bug report.`)}return l.reverse(),u.reverse(),[l,u]}function Ym(e){const t=new Array(e.length);for(let r=e.length-1,n=1;r>=0;--r)t[r]=n,n*=e[r];return t}function gi(e,t,r,n){const s=e.reduce((i,a)=>i*a,1);return new Z(r,new n(s).fill(t),e)}function Zm(e,t){let r,n;return r="float32",n=Float32Array,gi(e,t,r,n)}function Jm(e,t){return Zm(e.dims,t)}function ir(e){return gi(e,1n,"int64",BigInt64Array)}function eg(e){return ir(e.dims)}function tg(e){return gi(e,0n,"int64",BigInt64Array)}function rg(e){return tg(e.dims)}class ng{constructor(t=(r,n)=>r>n){this._heap=[],this._comparator=t}get size(){return this._heap.length}isEmpty(){return this.size===0}peek(){return this._heap[0]}push(...t){return this.extend(t)}extend(t){for(const r of t)this._heap.push(r),this._siftUp();return this.size}pop(){const t=this.peek(),r=this.size-1;return r>0&&this._swap(0,r),this._heap.pop(),this._siftDown(),t}replace(t){const r=this.peek();return this._heap[0]=t,this._siftDown(),r}_parent(t){return(t+1>>>1)-1}_left(t){return(t<<1)+1}_right(t){return t+1<<1}_greater(t,r){return this._comparator(this._heap[t],this._heap[r])}_swap(t,r){const n=this._heap[t];this._heap[t]=this._heap[r],this._heap[r]=n}_siftUp(){let t=this.size-1;for(;t>0&&this._greater(t,this._parent(t));)this._swap(t,this._parent(t)),t=this._parent(t)}_siftDown(){let t=0;for(;this._left(t)[]),this.endNodes=Array.from({length:this.len+1},()=>[]);const s=new ar(this.bosTokenId,0,0,0,0),i=new ar(this.eosTokenId,1,this.len,0,0);this.nodes.push(s.clone()),this.nodes.push(i.clone()),this.beginNodes[this.len].push(i),this.endNodes[0].push(s)}insert(t,r,n,s){const i=this.nodes.length,a=new ar(s,i,t,r,n);this.beginNodes[t].push(a),this.endNodes[t+r].push(a),this.nodes.push(a)}viterbi(){const t=this.len;let r=0;for(;r<=t;){if(this.beginNodes[r].length==0)return[];for(let o of this.beginNodes[r]){o.prev=null;let l=0,u=null;for(let d of this.endNodes[r]){const c=d.backtraceScore+o.score;(u===null||c>l)&&(u=d.clone(),l=c)}if(u!==null)o.prev=u,o.backtraceScore=l;else return[]}++r}const n=[],i=this.beginNodes[t][0].prev;if(i===null)return[];let a=i.clone();for(;a.prev!==null;)n.push(a.clone()),a=a.clone().prev.clone();return n.reverse(),n}piece(t){return this.sentence.slice(t.pos,t.pos+t.length)}tokens(){return this.viterbi().map(r=>this.piece(r))}tokenIds(){return this.viterbi().map(r=>r.tokenId)}}class ar{constructor(t,r,n,s,i){this.tokenId=t,this.nodeId=r,this.pos=n,this.length=s,this.score=i,this.prev=null,this.backtraceScore=0}clone(){const t=new ar(this.tokenId,this.nodeId,this.pos,this.length,this.score);return t.prev=this.prev,t.backtraceScore=this.backtraceScore,t}}var $=Object.freeze({Text:"Text",NumericLiteral:"NumericLiteral",BooleanLiteral:"BooleanLiteral",StringLiteral:"StringLiteral",Identifier:"Identifier",Equals:"Equals",OpenParen:"OpenParen",CloseParen:"CloseParen",OpenStatement:"OpenStatement",CloseStatement:"CloseStatement",OpenExpression:"OpenExpression",CloseExpression:"CloseExpression",OpenSquareBracket:"OpenSquareBracket",CloseSquareBracket:"CloseSquareBracket",OpenCurlyBracket:"OpenCurlyBracket",CloseCurlyBracket:"CloseCurlyBracket",Comma:"Comma",Dot:"Dot",Colon:"Colon",Pipe:"Pipe",CallOperator:"CallOperator",AdditiveBinaryOperator:"AdditiveBinaryOperator",MultiplicativeBinaryOperator:"MultiplicativeBinaryOperator",ComparisonBinaryOperator:"ComparisonBinaryOperator",UnaryOperator:"UnaryOperator",Set:"Set",If:"If",For:"For",In:"In",Is:"Is",NotIn:"NotIn",Else:"Else",EndIf:"EndIf",ElseIf:"ElseIf",EndFor:"EndFor",And:"And",Or:"Or",Not:"UnaryOperator"}),mp=Object.freeze({set:$.Set,for:$.For,in:$.In,is:$.Is,if:$.If,else:$.Else,endif:$.EndIf,elif:$.ElseIf,endfor:$.EndFor,and:$.And,or:$.Or,not:$.Not,"not in":$.NotIn,true:$.BooleanLiteral,false:$.BooleanLiteral}),Ct=class{constructor(e,t){this.value=e,this.type=t}};function gp(e){return/\w/.test(e)}function _i(e){return/[0-9]/.test(e)}var ag=[["{%",$.OpenStatement],["%}",$.CloseStatement],["{{",$.OpenExpression],["}}",$.CloseExpression],["(",$.OpenParen],[")",$.CloseParen],["{",$.OpenCurlyBracket],["}",$.CloseCurlyBracket],["[",$.OpenSquareBracket],["]",$.CloseSquareBracket],[",",$.Comma],[".",$.Dot],[":",$.Colon],["|",$.Pipe],["<=",$.ComparisonBinaryOperator],[">=",$.ComparisonBinaryOperator],["==",$.ComparisonBinaryOperator],["!=",$.ComparisonBinaryOperator],["<",$.ComparisonBinaryOperator],[">",$.ComparisonBinaryOperator],["+",$.AdditiveBinaryOperator],["-",$.AdditiveBinaryOperator],["*",$.MultiplicativeBinaryOperator],["/",$.MultiplicativeBinaryOperator],["%",$.MultiplicativeBinaryOperator],["=",$.Equals]],og=new Map([["n",` +`],["t"," "],["r","\r"],["b","\b"],["f","\f"],["v","\v"],["'","'"],['"','"'],["\\","\\"]]);function lg(e,t={}){return e.endsWith(` +`)&&(e=e.slice(0,-1)),e=e.replace(/{#.*?#}/gs,"{##}"),t.lstrip_blocks&&(e=e.replace(/^[ \t]*({[#%])/gm,"$1")),t.trim_blocks&&(e=e.replace(/([#%]})\n/g,"$1")),e.replace(/{##}/g,"").replace(/-%}\s*/g,"%}").replace(/\s*{%-/g,"{%").replace(/-}}\s*/g,"}}").replace(/\s*{{-/g,"{{")}function ug(e,t={}){var a,o,l;const r=[],n=lg(e,t);let s=0;const i=u=>{let d="";for(;u(n[s]);){if(n[s]==="\\"){if(++s,s>=n.length)throw new SyntaxError("Unexpected end of input");const c=n[s++],p=og.get(c);if(p===void 0)throw new SyntaxError(`Unexpected escaped character: ${c}`);d+=p;continue}if(d+=n[s++],s>=n.length)throw new SyntaxError("Unexpected end of input")}return d};e:for(;s0){r.push(new Ct(c,$.Text));continue}}i(c=>/\s/.test(c));const d=n[s];if(d==="-"||d==="+"){const c=(o=r.at(-1))==null?void 0:o.type;if(c===$.Text||c===void 0)throw new SyntaxError(`Unexpected character: ${d}`);switch(c){case $.Identifier:case $.NumericLiteral:case $.BooleanLiteral:case $.StringLiteral:case $.CloseParen:case $.CloseSquareBracket:break;default:{++s;const p=i(_i);r.push(new Ct(`${d}${p}`,p.length>0?$.NumericLiteral:$.UnaryOperator));continue}}}for(const[c,p]of ag)if(n.slice(s,s+c.length)===c){r.push(new Ct(c,p)),s+=c.length;continue e}if(d==="'"||d==='"'){++s;const c=i(p=>p!==d);r.push(new Ct(c,$.StringLiteral)),++s;continue}if(_i(d)){const c=i(_i);r.push(new Ct(c,$.NumericLiteral));continue}if(gp(d)){const c=i(gp),p=Object.hasOwn(mp,c)?mp[c]:$.Identifier;p===$.In&&((l=r.at(-1))==null?void 0:l.type)===$.Not?(r.pop(),r.push(new Ct("not in",$.NotIn))):r.push(new Ct(c,p));continue}throw new SyntaxError(`Unexpected character: ${d}`)}return r}var or=class{constructor(){T(this,"type","Statement")}},dg=class extends or{constructor(t){super();T(this,"type","Program");this.body=t}},_p=class extends or{constructor(t,r,n){super();T(this,"type","If");this.test=t,this.body=r,this.alternate=n}},cg=class extends or{constructor(t,r,n){super();T(this,"type","For");this.loopvar=t,this.iterable=r,this.body=n}},pg=class extends or{constructor(t,r){super();T(this,"type","Set");this.assignee=t,this.value=r}},tt=class extends or{constructor(){super(...arguments);T(this,"type","Expression")}},hg=class extends tt{constructor(t,r,n){super();T(this,"type","MemberExpression");this.object=t,this.property=r,this.computed=n}},fg=class extends tt{constructor(t,r){super();T(this,"type","CallExpression");this.callee=t,this.args=r}},Ft=class extends tt{constructor(t){super();T(this,"type","Identifier");this.value=t}},Nt=class extends tt{constructor(t){super();T(this,"type","Literal");this.value=t}},mg=class extends Nt{constructor(){super(...arguments);T(this,"type","NumericLiteral")}},wp=class extends Nt{constructor(){super(...arguments);T(this,"type","StringLiteral")}},yp=class extends Nt{constructor(){super(...arguments);T(this,"type","BooleanLiteral")}},gg=class extends Nt{constructor(){super(...arguments);T(this,"type","ArrayLiteral")}},bp=class extends Nt{constructor(){super(...arguments);T(this,"type","TupleLiteral")}},_g=class extends Nt{constructor(){super(...arguments);T(this,"type","ObjectLiteral")}},lr=class extends tt{constructor(t,r,n){super();T(this,"type","BinaryExpression");this.operator=t,this.left=r,this.right=n}},wg=class extends tt{constructor(t,r){super();T(this,"type","FilterExpression");this.operand=t,this.filter=r}},yg=class extends tt{constructor(t,r,n){super();T(this,"type","TestExpression");this.operand=t,this.negate=r,this.test=n}},bg=class extends tt{constructor(t,r){super();T(this,"type","UnaryExpression");this.operator=t,this.argument=r}},vg=class extends tt{constructor(t=void 0,r=void 0,n=void 0){super();T(this,"type","SliceExpression");this.start=t,this.stop=r,this.step=n}},$g=class extends tt{constructor(t,r){super();T(this,"type","KeywordArgumentExpression");this.key=t,this.value=r}};function xg(e){const t=new dg([]);let r=0;function n(E,M){const O=e[r++];if(!O||O.type!==E)throw new Error(`Parser Error: ${M}. ${O.type} !== ${E}.`);return O}function s(){switch(e[r].type){case $.Text:return o();case $.OpenStatement:return l();case $.OpenExpression:return u();default:throw new SyntaxError(`Unexpected token type: ${e[r].type}`)}}function i(...E){return r+E.length<=e.length&&E.some((M,O)=>M!==e[r+O].type)}function a(...E){return r+E.length<=e.length&&E.every((M,O)=>M===e[r+O].type)}function o(){return new wp(n($.Text,"Expected text token").value)}function l(){n($.OpenStatement,"Expected opening statement token");let E;switch(e[r].type){case $.Set:++r,E=d(),n($.CloseStatement,"Expected closing statement token");break;case $.If:++r,E=c(),n($.OpenStatement,"Expected {% token"),n($.EndIf,"Expected endif token"),n($.CloseStatement,"Expected %} token");break;case $.For:++r,E=h(),n($.OpenStatement,"Expected {% token"),n($.EndFor,"Expected endfor token"),n($.CloseStatement,"Expected %} token");break;default:throw new SyntaxError(`Unknown statement type: ${e[r].type}`)}return E}function u(){n($.OpenExpression,"Expected opening expression token");const E=f();return n($.CloseExpression,"Expected closing expression token"),E}function d(){const E=f();if(a($.Equals)){++r;const M=d();return new pg(E,M)}return E}function c(){var B,K,ee,k,L,G,ie,ce;const E=f();n($.CloseStatement,"Expected closing statement token");const M=[],O=[];for(;!(((B=e[r])==null?void 0:B.type)===$.OpenStatement&&(((K=e[r+1])==null?void 0:K.type)===$.ElseIf||((ee=e[r+1])==null?void 0:ee.type)===$.Else||((k=e[r+1])==null?void 0:k.type)===$.EndIf));)M.push(s());if(((L=e[r])==null?void 0:L.type)===$.OpenStatement&&((G=e[r+1])==null?void 0:G.type)!==$.EndIf)if(++r,a($.ElseIf))n($.ElseIf,"Expected elseif token"),O.push(c());else for(n($.Else,"Expected else token"),n($.CloseStatement,"Expected closing statement token");!(((ie=e[r])==null?void 0:ie.type)===$.OpenStatement&&((ce=e[r+1])==null?void 0:ce.type)===$.EndIf);)O.push(s());return new _p(E,M,O)}function p(E=!1){const M=E?V:f,O=[M()],B=a($.Comma);for(;B&&(++r,O.push(M()),!!a($.Comma)););return B?new bp(O):O[0]}function h(){const E=p(!0);if(!(E instanceof Ft||E instanceof bp))throw new SyntaxError(`Expected identifier/tuple for the loop variable, got ${E.type} instead`);n($.In,"Expected `in` keyword following loop variable");const M=f();n($.CloseStatement,"Expected closing statement token");const O=[];for(;i($.OpenStatement,$.EndFor);)O.push(s());return new cg(E,M,O)}function f(){return m()}function m(){const E=g();if(a($.If)){++r;const M=g();n($.Else,"Expected else token");const O=g();return new _p(M,[E],[O])}return E}function g(){let E=w();for(;a($.Or);){const M=e[r];++r;const O=w();E=new lr(M,E,O)}return E}function w(){let E=_();for(;a($.And);){const M=e[r];++r;const O=_();E=new lr(M,E,O)}return E}function _(){let E;for(;a($.Not);){const M=e[r];++r;const O=_();E=new bg(M,O)}return E??v()}function v(){let E=b();for(;a($.ComparisonBinaryOperator)||a($.In)||a($.NotIn);){const M=e[r];++r;const O=b();E=new lr(M,E,O)}return E}function b(){let E=U();for(;a($.AdditiveBinaryOperator);){const M=e[r];++r;const O=U();E=new lr(M,E,O)}return E}function y(){const E=R();return a($.OpenParen)?S(E):E}function S(E){let M=new fg(E,x());return a($.OpenParen)&&(M=S(M)),M}function x(){n($.OpenParen,"Expected opening parenthesis for arguments list");const E=I();return n($.CloseParen,"Expected closing parenthesis for arguments list"),E}function I(){const E=[];for(;!a($.CloseParen);){let M=f();if(a($.Equals)){if(++r,!(M instanceof Ft))throw new SyntaxError("Expected identifier for keyword argument");const O=f();M=new $g(M,O)}E.push(M),a($.Comma)&&++r}return E}function D(){const E=[];let M=!1;for(;!a($.CloseSquareBracket);)a($.Colon)?(E.push(void 0),++r,M=!0):(E.push(f()),a($.Colon)&&(++r,M=!0));if(E.length===0)throw new SyntaxError("Expected at least one argument for member/slice expression");if(M){if(E.length>3)throw new SyntaxError("Expected 0-3 arguments for slice expression");return new vg(...E)}return E[0]}function R(){let E=V();for(;a($.Dot)||a($.OpenSquareBracket);){const M=e[r];++r;let O;const B=M.type!==$.Dot;if(B)O=D(),n($.CloseSquareBracket,"Expected closing square bracket");else if(O=V(),O.type!=="Identifier")throw new SyntaxError("Expected identifier following dot operator");E=new hg(E,O,B)}return E}function U(){let E=q();for(;a($.MultiplicativeBinaryOperator);){const M=e[r];++r;const O=q();E=new lr(M,E,O)}return E}function q(){let E=H();for(;a($.Is);){++r;const M=a($.Not);M&&++r;let O=V();if(O instanceof yp&&(O=new Ft(O.value.toString())),!(O instanceof Ft))throw new SyntaxError("Expected identifier for the test");E=new yg(E,M,O)}return E}function H(){let E=y();for(;a($.Pipe);){++r;let M=V();if(!(M instanceof Ft))throw new SyntaxError("Expected identifier for the filter");a($.OpenParen)&&(M=S(M)),E=new wg(E,M)}return E}function V(){const E=e[r];switch(E.type){case $.NumericLiteral:return++r,new mg(Number(E.value));case $.StringLiteral:return++r,new wp(E.value);case $.BooleanLiteral:return++r,new yp(E.value==="true");case $.Identifier:return++r,new Ft(E.value);case $.OpenParen:{++r;const M=p();if(e[r].type!==$.CloseParen)throw new SyntaxError(`Expected closing parenthesis, got ${e[r].type} instead`);return++r,M}case $.OpenSquareBracket:{++r;const M=[];for(;!a($.CloseSquareBracket);)M.push(f()),a($.Comma)&&++r;return++r,new gg(M)}case $.OpenCurlyBracket:{++r;const M=new Map;for(;!a($.CloseCurlyBracket);){const O=f();n($.Colon,"Expected colon between key and value in object literal");const B=f();M.set(O,B),a($.Comma)&&++r}return++r,new _g(M)}default:throw new SyntaxError(`Unexpected token: ${E.type}`)}}for(;r=0?(t=(t??(t=0))<0?Math.max(e.length+t,0):Math.min(t,e.length),r=(r??(r=e.length))<0?Math.max(e.length+r,0):Math.min(r,e.length)):(t=(t??(t=e.length-1))<0?Math.max(e.length+t,-1):Math.min(t,e.length-1),r=(r??(r=-1))<-1?Math.max(e.length+r,-1):Math.min(r,e.length-1));const i=[];for(let a=t;s*at.toUpperCase())}var at=class{constructor(e=void 0){T(this,"type","RuntimeValue");T(this,"value");T(this,"builtins",new Map);this.value=e}__bool__(){return new fe(!!this.value)}},ue=class extends at{constructor(){super(...arguments);T(this,"type","NumericValue")}},te=class extends at{constructor(){super(...arguments);T(this,"type","StringValue");T(this,"builtins",new Map([["upper",new ot(()=>new te(this.value.toUpperCase()))],["lower",new ot(()=>new te(this.value.toLowerCase()))],["strip",new ot(()=>new te(this.value.trim()))],["title",new ot(()=>new te($p(this.value)))],["length",new ue(this.value.length)]]))}},fe=class extends at{constructor(){super(...arguments);T(this,"type","BooleanValue")}},Qe=class extends at{constructor(){super(...arguments);T(this,"type","ObjectValue");T(this,"builtins",new Map([["get",new ot(([t,r])=>{if(!(t instanceof te))throw new Error(`Object key must be a string: got ${t.type}`);return this.value.get(t.value)??r??new ur})],["items",new ot(()=>new de(Array.from(this.value.entries()).map(([t,r])=>new de([new te(t),r]))))]]))}__bool__(){return new fe(this.value.size>0)}},de=class extends at{constructor(){super(...arguments);T(this,"type","ArrayValue");T(this,"builtins",new Map([["length",new ue(this.value.length)]]))}__bool__(){return new fe(this.value.length>0)}},Sg=class extends de{constructor(){super(...arguments);T(this,"type","TupleValue")}},ot=class extends at{constructor(){super(...arguments);T(this,"type","FunctionValue")}},ur=class extends at{constructor(){super(...arguments);T(this,"type","NullValue")}},Xe=class extends at{constructor(){super(...arguments);T(this,"type","UndefinedValue")}},wi=class{constructor(e){T(this,"variables",new Map([["namespace",new ot(e=>{if(e.length===0)return new Qe(new Map);if(e.length!==1||!(e[0]instanceof Qe))throw new Error("`namespace` expects either zero arguments or a single object argument");return e[0]})]]));T(this,"tests",new Map([["boolean",e=>e.type==="BooleanValue"],["callable",e=>e instanceof ot],["odd",e=>{if(e.type!=="NumericValue")throw new Error(`Cannot apply test "odd" to type: ${e.type}`);return e.value%2!==0}],["even",e=>{if(e.type!=="NumericValue")throw new Error(`Cannot apply test "even" to type: ${e.type}`);return e.value%2===0}],["false",e=>e.type==="BooleanValue"&&!e.value],["true",e=>e.type==="BooleanValue"&&e.value],["number",e=>e.type==="NumericValue"],["integer",e=>e.type==="NumericValue"&&Number.isInteger(e.value)],["iterable",e=>e instanceof de||e instanceof te],["lower",e=>{const t=e.value;return e.type==="StringValue"&&t===t.toLowerCase()}],["upper",e=>{const t=e.value;return e.type==="StringValue"&&t===t.toUpperCase()}],["none",e=>e.type==="NullValue"],["defined",e=>e.type!=="UndefinedValue"],["undefined",e=>e.type==="UndefinedValue"],["equalto",(e,t)=>e.value===t.value]]));this.parent=e}set(e,t){return this.declareVariable(e,cn(t))}declareVariable(e,t){if(this.variables.has(e))throw new SyntaxError(`Variable already declared: ${e}`);return this.variables.set(e,t),t}setVariable(e,t){return this.variables.set(e,t),t}resolve(e){if(this.variables.has(e))return this;if(this.parent)return this.parent.resolve(e);throw new Error(`Unknown variable: ${e}`)}lookupVariable(e){try{return this.resolve(e).variables.get(e)??new Xe}catch{return new Xe}}},Eg=class{constructor(e){T(this,"global");this.global=e??new wi}run(e){return this.evaluate(e,this.global)}evaluateBinaryExpression(e,t){const r=this.evaluate(e.left,t);switch(e.operator.value){case"and":return r.__bool__().value?this.evaluate(e.right,t):r;case"or":return r.__bool__().value?r:this.evaluate(e.right,t)}const n=this.evaluate(e.right,t);switch(e.operator.value){case"==":return new fe(r.value==n.value);case"!=":return new fe(r.value!=n.value)}if(r instanceof Xe||n instanceof Xe)throw new Error("Cannot perform operation on undefined values");if(r instanceof ur||n instanceof ur)throw new Error("Cannot perform operation on null values");if(r instanceof ue&&n instanceof ue)switch(e.operator.value){case"+":return new ue(r.value+n.value);case"-":return new ue(r.value-n.value);case"*":return new ue(r.value*n.value);case"/":return new ue(r.value/n.value);case"%":return new ue(r.value%n.value);case"<":return new fe(r.value":return new fe(r.value>n.value);case">=":return new fe(r.value>=n.value);case"<=":return new fe(r.value<=n.value)}else if(r instanceof de&&n instanceof de)switch(e.operator.value){case"+":return new de(r.value.concat(n.value))}else if(n instanceof de){const s=n.value.find(i=>i.value===r.value)!==void 0;switch(e.operator.value){case"in":return new fe(s);case"not in":return new fe(!s)}}if(r instanceof te||n instanceof te)switch(e.operator.value){case"+":return new te(r.value.toString()+n.value.toString())}if(r instanceof te&&n instanceof te)switch(e.operator.value){case"in":return new fe(n.value.includes(r.value));case"not in":return new fe(!n.value.includes(r.value))}if(r instanceof te&&n instanceof Qe)switch(e.operator.value){case"in":return new fe(n.value.has(r.value));case"not in":return new fe(!n.value.has(r.value))}throw new SyntaxError(`Unknown operator "${e.operator.value}" between ${r.type} and ${n.type}`)}evaluateFilterExpression(e,t){const r=this.evaluate(e.operand,t);if(e.filter.type==="Identifier"){const n=e.filter;if(r instanceof de)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 ue(r.value.length);case"reverse":return new de(r.value.reverse());case"sort":return new de(r.value.sort((s,i)=>{if(s.type!==i.type)throw new Error(`Cannot compare different types: ${s.type} and ${i.type}`);switch(s.type){case"NumericValue":return s.value-i.value;case"StringValue":return s.value.localeCompare(i.value);default:throw new Error(`Cannot compare type: ${s.type}`)}}));default:throw new Error(`Unknown ArrayValue filter: ${n.value}`)}else if(r instanceof te)switch(n.value){case"length":return new ue(r.value.length);case"upper":return new te(r.value.toUpperCase());case"lower":return new te(r.value.toLowerCase());case"title":return new te($p(r.value));case"capitalize":return new te(r.value.charAt(0).toUpperCase()+r.value.slice(1));case"trim":return new te(r.value.trim());default:throw new Error(`Unknown StringValue filter: ${n.value}`)}else if(r instanceof ue)switch(n.value){case"abs":return new ue(Math.abs(r.value));default:throw new Error(`Unknown NumericValue filter: ${n.value}`)}else if(r instanceof Qe)switch(n.value){case"items":return new de(Array.from(r.value.entries()).map(([s,i])=>new de([new te(s),i])));case"length":return new ue(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(e.filter.type==="CallExpression"){const n=e.filter;if(n.callee.type!=="Identifier")throw new Error(`Unknown filter: ${n.callee.type}`);const s=n.callee.value;if(r instanceof de){switch(s){case"selectattr":{if(r.value.some(d=>!(d instanceof Qe)))throw new Error("`selectattr` can only be applied to array of objects");if(n.args.some(d=>d.type!=="StringLiteral"))throw new Error("arguments of `selectattr` must be strings");const[i,a,o]=n.args.map(d=>this.evaluate(d,t));let l;if(a){const d=t.tests.get(a.value);if(!d)throw new Error(`Unknown test: ${a.value}`);l=d}else l=(...d)=>d[0].__bool__().value;const u=r.value.filter(d=>{const c=d.value.get(i.value);return c?l(c,o):!1});return new de(u)}}throw new Error(`Unknown ArrayValue filter: ${s}`)}else throw new Error(`Cannot apply filter "${s}" to type: ${r.type}`)}throw new Error(`Unknown filter: ${e.filter.type}`)}evaluateTestExpression(e,t){const r=this.evaluate(e.operand,t),n=t.tests.get(e.test.value);if(!n)throw new Error(`Unknown test: ${e.test.value}`);const s=n(r);return new fe(e.negate?!s:s)}evaluateUnaryExpression(e,t){const r=this.evaluate(e.argument,t);switch(e.operator.value){case"not":return new fe(!r.value);default:throw new SyntaxError(`Unknown operator: ${e.operator.value}`)}}evalProgram(e,t){return this.evaluateBlock(e.body,t)}evaluateBlock(e,t){let r="";for(const n of e){const s=this.evaluate(n,t);s.type!=="NullValue"&&s.type!=="UndefinedValue"&&(r+=s.value)}return new te(r)}evaluateIdentifier(e,t){return t.lookupVariable(e.value)}evaluateCallExpression(e,t){const r=[],n=new Map;for(const i of e.args)if(i.type==="KeywordArgumentExpression"){const a=i;n.set(a.key.value,this.evaluate(a.value,t))}else r.push(this.evaluate(i,t));n.size>0&&r.push(new Qe(n));const s=this.evaluate(e.callee,t);if(s.type!=="FunctionValue")throw new Error(`Cannot call something that is not a function: got ${s.type}`);return s.value(r,t)}evaluateSliceExpression(e,t,r){if(!(e instanceof de||e instanceof te))throw new Error("Slice object must be an array or string");const n=this.evaluate(t.start,r),s=this.evaluate(t.stop,r),i=this.evaluate(t.step,r);if(!(n instanceof ue||n instanceof Xe))throw new Error("Slice start must be numeric or undefined");if(!(s instanceof ue||s instanceof Xe))throw new Error("Slice stop must be numeric or undefined");if(!(i instanceof ue||i instanceof Xe))throw new Error("Slice step must be numeric or undefined");return e instanceof de?new de(vp(e.value,n.value,s.value,i.value)):new te(vp(Array.from(e.value),n.value,s.value,i.value).join(""))}evaluateMemberExpression(e,t){const r=this.evaluate(e.object,t);let n;if(e.computed){if(e.property.type==="SliceExpression")return this.evaluateSliceExpression(r,e.property,t);n=this.evaluate(e.property,t)}else n=new te(e.property.value);let s;if(r instanceof Qe){if(!(n instanceof te))throw new Error(`Cannot access property with non-string: got ${n.type}`);s=r.value.get(n.value)??r.builtins.get(n.value)}else if(r instanceof de||r instanceof te)if(n instanceof ue)s=r.value.at(n.value),r instanceof te&&(s=new te(r.value.at(n.value)));else if(n instanceof te)s=r.builtins.get(n.value);else throw new Error(`Cannot access property with non-string/non-number: got ${n.type}`);else{if(!(n instanceof te))throw new Error(`Cannot access property with non-string: got ${n.type}`);s=r.builtins.get(n.value)}return s instanceof at?s:new Xe}evaluateSet(e,t){const r=this.evaluate(e.value,t);if(e.assignee.type==="Identifier"){const n=e.assignee.value;t.setVariable(n,r)}else if(e.assignee.type==="MemberExpression"){const n=e.assignee,s=this.evaluate(n.object,t);if(!(s instanceof Qe))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");s.value.set(n.property.value,r)}else throw new Error(`Invalid LHS inside assignment expression: ${JSON.stringify(e.assignee)}`);return new ur}evaluateIf(e,t){const r=this.evaluate(e.test,t);return this.evaluateBlock(r.__bool__().value?e.body:e.alternate,t)}evaluateFor(e,t){const r=new wi(t),n=this.evaluate(e.iterable,r);if(!(n instanceof de))throw new Error(`Expected iterable type in for loop: got ${n.type}`);let s="";for(let i=0;i0?n.value[i-1]:new Xe],["nextitem",id.value.length?"few":"many"} items to unpack`);for(let c=0;cthis.evaluate(r,t)));case"TupleLiteral":return new Sg(e.value.map(r=>this.evaluate(r,t)));case"ObjectLiteral":{const r=new Map;for(const[n,s]of e.value){const i=this.evaluate(n,t);if(!(i instanceof te))throw new Error(`Object keys must be strings: got ${i.type}`);r.set(i.value,this.evaluate(s,t))}return new Qe(r)}case"Identifier":return this.evaluateIdentifier(e,t);case"CallExpression":return this.evaluateCallExpression(e,t);case"MemberExpression":return this.evaluateMemberExpression(e,t);case"UnaryExpression":return this.evaluateUnaryExpression(e,t);case"BinaryExpression":return this.evaluateBinaryExpression(e,t);case"FilterExpression":return this.evaluateFilterExpression(e,t);case"TestExpression":return this.evaluateTestExpression(e,t);default:throw new SyntaxError(`Unknown node type: ${e.type}`)}}};function cn(e){switch(typeof e){case"number":return new ue(e);case"string":return new te(e);case"boolean":return new fe(e);case"object":return e===null?new ur:Array.isArray(e)?new de(e.map(cn)):new Qe(new Map(Object.entries(e).map(([t,r])=>[t,cn(r)])));case"function":return new ot((t,r)=>{const n=e(...t.map(s=>s.value))??null;return cn(n)});default:throw new Error(`Cannot convert to runtime value: ${e}`)}}var Tg=class{constructor(e){T(this,"parsed");const t=ug(e,{lstrip_blocks:!0,trim_blocks:!0});this.parsed=xg(t)}render(e){const t=new wi;t.set("false",!1),t.set("true",!0),t.set("raise_exception",s=>{throw new Error(s)}),t.set("range",kg);for(const[s,i]of Object.entries(e))t.set(s,i);return new Eg(t).run(this.parsed).value}};async function xp(e,t){const r=await Promise.all([_t(e,"tokenizer.json",!0,t),_t(e,"tokenizer_config.json",!0,t)]);return t.legacy!==null&&(r[1].legacy=t.legacy),r}function Ig(e,t){const r=[];let n=0;for(const s of e.matchAll(t)){const i=s[0];n0&&r.push(i),n=s.index+i.length}return n=19968&&e<=40959||e>=13312&&e<=19903||e>=131072&&e<=173791||e>=173824&&e<=177983||e>=177984&&e<=178207||e>=178208&&e<=183983||e>=63744&&e<=64255||e>=194560&&e<=195103}function Cg(e,t,r){const n=[];let s=0;for(;sthis.tokens_to_ids.get(r)??this.unk_token_id)}convert_ids_to_tokens(t){return t.map(r=>this.vocab[r]??this.unk_token)}}class Og extends cr{constructor(t){super(t),this.tokens_to_ids=yi(t.vocab),this.unk_token_id=this.tokens_to_ids.get(t.unk_token),this.unk_token=t.unk_token,this.max_input_chars_per_word=t.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(t){const r=[];for(const n of t){const s=[...n];if(s.length>this.max_input_chars_per_word){r.push(this.unk_token);continue}let i=!1,a=0;const o=[];for(;a0&&(d=this.config.continuing_subword_prefix+d),this.tokens_to_ids.has(d)){u=d;break}--l}if(u===null){i=!0;break}o.push(u),a=l}i?r.push(this.unk_token):r.push(...o)}return r}}class Bg extends cr{constructor(t,r){super(t);const n=t.vocab.length;this.vocab=new Array(n),this.scores=new Array(n);for(let s=0;s[s,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=_f(this.scores)[0],this.unkScore=this.minScore-10,this.scores[this.unk_token_id]=this.unkScore,this.trie=new sg,this.trie.extend(this.vocab),this.fuse_unk=!0}populateNodes(t){const r=t.sentence,n=r.length;let s=0;for(;s{const e=[...Array.from({length:94},(s,i)=>i+33),...Array.from({length:12},(s,i)=>i+161),...Array.from({length:82},(s,i)=>i+174)],t=e.slice();let r=0;for(let s=0;s<256;++s)e.includes(s)||(e.push(s),t.push(256+r),r+=1);const n=t.map(s=>String.fromCharCode(s));return Object.fromEntries(e.map((s,i)=>[s,n[i]]))})(),Rg=df(Tp);class Dg extends cr{constructor(t){super(t),this.BPE_SPLIT_TOKEN=" ",this.tokens_to_ids=yi(t.vocab),this.unk_token_id=this.tokens_to_ids.get(t.unk_token),this.unk_token=t.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(t.merges.map((r,n)=>[r,n])),this.merges=t.merges.map(r=>r.split(this.BPE_SPLIT_TOKEN)),this.end_of_word_suffix=t.end_of_word_suffix,this.continuing_subword_suffix=t.continuing_subword_suffix??null,this.byte_fallback=this.config.byte_fallback??!1,this.byte_fallback&&(this.text_encoder=new TextEncoder),this.cache=new Map}bpe(t){if(t.length===0)return[];const r=this.cache.get(t);if(r!==void 0)return r;const n=Array.from(t);this.end_of_word_suffix&&(n[n.length-1]+=this.end_of_word_suffix);let s=[];if(n.length>1){const i=new ng((l,u)=>l.score`<0x${a.toString(16).toUpperCase().padStart(2,"0")}>`)):r.push(this.unk_token)}return r}}class Lg extends cr{constructor(t,r){super(t),this.tokens_to_ids=yi(r.target_lang?t.vocab[r.target_lang]:t.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,s]of this.tokens_to_ids)this.vocab[s]=n}encode(t){return t}}class ze extends Ce{constructor(t){super(),this.config=t}static fromConfig(t){if(t===null)return null;switch(t.type){case"BertNormalizer":return new Kg(t);case"Precompiled":return new h_(t);case"Sequence":return new Hg(t);case"Replace":return new Fg(t);case"NFC":return new Ng(t);case"NFKC":return new Ug(t);case"NFKD":return new qg(t);case"Strip":return new Gg(t);case"StripAccents":return new Vg(t);case"Lowercase":return new jg(t);case"Prepend":return new Wg(t);default:throw new Error(`Unknown Normalizer type: ${t.type}`)}}normalize(t){throw Error("normalize should be implemented in subclass.")}_call(t){return this.normalize(t)}}class Fg extends ze{normalize(t){const r=pn(this.config.pattern);return r===null?t:t.replaceAll(r,this.config.content)}}class Ng extends ze{normalize(t){return t=t.normalize("NFC"),t}}class Ug extends ze{normalize(t){return t=t.normalize("NFKC"),t}}class qg extends ze{normalize(t){return t=t.normalize("NFKD"),t}}class Gg extends ze{normalize(t){return this.config.strip_left&&this.config.strip_right?t=t.trim():(this.config.strip_left&&(t=t.trimStart()),this.config.strip_right&&(t=t.trimEnd())),t}}class Vg extends ze{normalize(t){return t=Sp(t),t}}class jg extends ze{normalize(t){return t=t.toLowerCase(),t}}class Wg extends ze{normalize(t){return t=this.config.prepend+t,t}}class Hg extends ze{constructor(t){super(t),this.normalizers=t.normalizers.map(r=>ze.fromConfig(r))}normalize(t){return this.normalizers.reduce((r,n)=>n.normalize(r),t)}}class Kg extends ze{_tokenize_chinese_chars(t){const r=[];for(let n=0;nthis.pre_tokenize_text(n,r)):this.pre_tokenize_text(t,r)).flat()}_call(t,r){return this.pre_tokenize(t,r)}}class Qg extends De{constructor(t){super(),this.pattern=new RegExp(`[^\\s${dr}]+|[${dr}]`,"gu")}pre_tokenize_text(t,r){return t.trim().match(this.pattern)||[]}}class Xg extends De{constructor(t){super(),this.config=t,this.add_prefix_space=this.config.add_prefix_space,this.trim_offsets=this.config.trim_offsets,this.use_regex=this.config.use_regex??!0,this.pattern=new RegExp("'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)|\\s+","gu"),this.byte_encoder=Tp,this.text_encoder=new TextEncoder}pre_tokenize_text(t,r){return this.add_prefix_space&&!t.startsWith(" ")&&(t=" "+t),(this.use_regex?t.match(this.pattern)||[]:[t]).map(s=>Array.from(this.text_encoder.encode(s),i=>this.byte_encoder[i]).join(""))}}class Yg extends De{constructor(t){super(),this.config=t,this.pattern=pn(this.config.pattern,this.config.invert)}pre_tokenize_text(t,r){return this.pattern===null?[]:this.config.invert?t.match(this.pattern)||[]:Ig(t,this.pattern)}}class Zg extends De{constructor(t){super(),this.config=t,this.pattern=new RegExp(`[^${dr}]+|[${dr}]+`,"gu")}pre_tokenize_text(t,r){return t.match(this.pattern)||[]}}class Jg extends De{constructor(t){super(),this.config=t;const r=`[^\\d]+|\\d${this.config.individual_digits?"":"+"}`;this.pattern=new RegExp(r,"gu")}pre_tokenize_text(t,r){return t.match(this.pattern)||[]}}class hn extends Ce{constructor(t){super(),this.config=t}static fromConfig(t){if(t===null)return null;switch(t.type){case"TemplateProcessing":return new e_(t);case"ByteLevel":return new t_(t);case"RobertaProcessing":return new Mp(t);case"BertProcessing":return new Ip(t);default:throw new Error(`Unknown PostProcessor type: ${t.type}`)}}post_process(t,...r){throw Error("post_process should be implemented in subclass.")}_call(t,...r){return this.post_process(t,...r)}}class Ip extends hn{constructor(t){super(t),this.cls=t.cls[0],this.sep=t.sep[0]}post_process(t,r=null,{add_special_tokens:n=!0}={}){n&&(t=we([this.cls],t,[this.sep]));let s=new Array(t.length).fill(0);if(r!==null){const i=n&&this instanceof Mp?[this.sep]:[],a=n?[this.sep]:[];t=we(t,i,r,a),s=we(s,new Array(r.length+i.length+a.length).fill(1))}return{tokens:t,token_type_ids:s}}}class Mp extends Ip{}class e_ extends hn{constructor(t){super(t),this.single=t.single,this.pair=t.pair}post_process(t,r=null,{add_special_tokens:n=!0}={}){const s=r===null?this.single:this.pair;let i=[],a=[];for(const o of s)"SpecialToken"in o?n&&(i.push(o.SpecialToken.id),a.push(o.SpecialToken.type_id)):"Sequence"in o&&(o.Sequence.id==="A"?(i=we(i,t),a=we(a,new Array(t.length).fill(o.Sequence.type_id))):o.Sequence.id==="B"&&(i=we(i,r),a=we(a,new Array(r.length).fill(o.Sequence.type_id))));return{tokens:i,token_type_ids:a}}}class t_ extends hn{post_process(t,r=null){return r&&(t=we(t,r)),{tokens:t}}}class Ae extends Ce{constructor(t){super(),this.config=t,this.added_tokens=[],this.end_of_word_suffix=null,this.trim_offsets=t.trim_offsets}static fromConfig(t){if(t===null)return null;switch(t.type){case"WordPiece":return new a_(t);case"Metaspace":return new p_(t);case"ByteLevel":return new o_(t);case"Replace":return new r_(t);case"ByteFallback":return new n_(t);case"Fuse":return new s_(t);case"Strip":return new i_(t);case"Sequence":return new u_(t);case"CTC":return new l_(t);case"BPEDecoder":return new d_(t);default:throw new Error(`Unknown Decoder type: ${t.type}`)}}_call(t){return this.decode(t)}decode(t){return this.decode_chain(t).join("")}decode_chain(t){throw Error("`decode_chain` should be implemented in subclass.")}}class r_ extends Ae{decode_chain(t){const r=pn(this.config.pattern);return r===null?t:t.map(n=>n.replaceAll(r,this.config.content))}}class n_ extends Ae{constructor(t){super(t),this.text_decoder=new TextDecoder}decode_chain(t){const r=[];let n=[];for(const s of t){let i=null;if(s.length===6&&s.startsWith("<0x")&&s.endsWith(">")){const a=parseInt(s.slice(3,5),16);isNaN(a)||(i=a)}if(i!==null)n.push(i);else{if(n.length>0){const a=this.text_decoder.decode(Uint8Array.from(n));r.push(a),n=[]}r.push(s)}}if(n.length>0){const s=this.text_decoder.decode(Uint8Array.from(n));r.push(s),n=[]}return r}}class s_ extends Ae{decode_chain(t){return[t.join("")]}}class i_ extends Ae{constructor(t){super(t),this.content=this.config.content,this.start=this.config.start,this.stop=this.config.stop}decode_chain(t){return t.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=bi(r)),r))}}class o_ extends Ae{constructor(t){super(t),this.byte_decoder=Rg,this.text_decoder=new TextDecoder("utf-8",{fatal:!1,ignoreBOM:!0}),this.end_of_word_suffix=null}convert_tokens_to_string(t){const r=t.join(""),n=new Uint8Array([...r].map(i=>this.byte_decoder[i]));return this.text_decoder.decode(n)}decode_chain(t){const r=[];let n=[];for(const s of t)this.added_tokens.find(i=>i.content===s)!==void 0?(n.length>0&&(r.push(this.convert_tokens_to_string(n)),n=[]),r.push(s)):n.push(s);return n.length>0&&r.push(this.convert_tokens_to_string(n)),r}}class l_ extends Ae{constructor(t){super(t),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(t){if(t.length===0)return"";const r=[t[0]];for(let i=1;ii!==this.pad_token).join("");return this.cleanup&&(s=bi(s).replaceAll(this.word_delimiter_token," ").trim()),s}decode_chain(t){return[this.convert_tokens_to_string(t)]}}class u_ extends Ae{constructor(t){super(t),this.decoders=t.decoders.map(r=>Ae.fromConfig(r))}decode_chain(t){return this.decoders.reduce((r,n)=>n.decode_chain(r),t)}}class d_ extends Ae{constructor(t){super(t),this.suffix=this.config.suffix}decode_chain(t){return t.map((r,n)=>r.replaceAll(this.suffix,n===t.length-1?"":" "))}}class c_ extends Ae{decode_chain(t){let r="";for(let n=1;nn.normalize("NFKC")).join("～"):t=t.normalize("NFKC"),t}}class f_ extends De{constructor(t){super(),this.tokenizers=t.pretokenizers.map(r=>De.fromConfig(r))}pre_tokenize_text(t,r){return this.tokenizers.reduce((n,s)=>s.pre_tokenize(n,r),[t])}}class m_ extends De{constructor(t){super()}pre_tokenize_text(t,r){return t.match(/\w+|[^\w\s]+/g)||[]}}class g_ extends De{constructor(t){super()}pre_tokenize_text(t,r){return zg(t)}}class __ extends De{constructor(t){super(),this.config=t,this.pattern=pn(this.config.pattern),this.content=this.config.content}pre_tokenize_text(t,r){return this.pattern===null?[t]:[t.replaceAll(this.pattern,this.config.content)]}}const w_=["bos_token","eos_token","unk_token","sep_token","pad_token","cls_token","mask_token"];function y_(e,t,r,n){for(const s of Object.keys(e)){const i=t-e[s].length,a=r(s),o=new Array(i).fill(a);e[s]=n==="right"?we(e[s],o):we(o,e[s])}}function b_(e,t){for(const r of Object.keys(e))e[r].length=t}class X extends Ce{constructor(r,n){super();T(this,"return_token_type_ids",!1);T(this,"_default_chat_template",`{% for message in messages %}{{'<|im_start|>' + message['role'] + ' +' + message['content'] + '<|im_end|>' + ' +'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant +' }}{% endif %}`);T(this,"padding_side","right");this._tokenizer_config=n,this.normalizer=ze.fromConfig(r.normalizer),this.pre_tokenizer=De.fromConfig(r.pre_tokenizer),this.model=cr.fromConfig(r.model,n),this.post_processor=hn.fromConfig(r.post_processor),this.decoder=Ae.fromConfig(r.decoder),this.special_tokens=[],this.all_special_ids=[],this.added_tokens=[];for(const s of r.added_tokens){const i=new Pg(s);this.added_tokens.push(i),this.model.tokens_to_ids.set(i.content,i.id),this.model.vocab[i.id]=i.content,i.special&&(this.special_tokens.push(i.content),this.all_special_ids.push(i.id))}if(this.additional_special_tokens=n.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(s=>`${s.lstrip?"\\s*":""}(${Di(s.content)})${s.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=n.model_max_length,this.remove_space=n.remove_space,this.clean_up_tokenization_spaces=n.clean_up_tokenization_spaces??!0,this.do_lowercase_and_remove_accent=n.do_lowercase_and_remove_accent??!1,n.padding_side&&(this.padding_side=n.padding_side),this.legacy=!1,this.chat_template=n.chat_template??null,Array.isArray(this.chat_template)){const s=Object.create(null);for(const{name:i,template:a}of this.chat_template){if(typeof i!="string"||typeof a!="string")throw new Error('Chat template must be a list of objects with "name" and "template" properties');s[i]=a}this.chat_template=s}this._compiled_template_cache=new Map}getToken(...r){for(const n of r){const s=this._tokenizer_config[n];if(s)if(typeof s=="object"){if(s.__type==="AddedToken")return s.content;throw Error(`Unknown token: ${s}`)}else return s}return null}static async from_pretrained(r,{progress_callback:n=null,config:s=null,cache_dir:i=null,local_files_only:a=!1,revision:o="main",legacy:l=null}={}){const u=await xp(r,{progress_callback:n,config:s,cache_dir:i,local_files_only:a,revision:o,legacy:l});return new this(...u)}_call(r,{text_pair:n=null,add_special_tokens:s=!0,padding:i=!1,truncation:a=null,max_length:o=null,return_tensor:l=!0}={}){const u=Array.isArray(r);let d;if(u){if(r.length===0)throw Error("text array must be non-empty");if(n!==null){if(Array.isArray(n)){if(r.length!==n.length)throw Error("text and text_pair must have the same length")}else throw Error("text_pair must also be an array");d=r.map((p,h)=>this._encode_plus(p,{text_pair:n[h],add_special_tokens:s}))}else d=r.map(p=>this._encode_plus(p,{add_special_tokens:s}))}else{if(r==null)throw Error("text may not be null or undefined");if(Array.isArray(n))throw Error("When specifying `text_pair`, since `text` is a string, `text_pair` must also be a string (i.e., not an array).");d=[this._encode_plus(r,{text_pair:n,add_special_tokens:s})]}if(o===null?i==="max_length"?o=this.model_max_length:o=On(d.map(p=>p.input_ids.length))[0]:a||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."),o=Math.min(o,this.model_max_length),i||a)for(let p=0;po?a&&b_(d[p],o):i&&y_(d[p],o,h=>h==="input_ids"?this.pad_token_id:0,this.padding_side));const c={};if(l){if(!(i&&a)&&d.some(h=>{var f;for(const m of Object.keys(h))if(h[m].length!==((f=d[0][m])==null?void 0:f.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 p=[d.length,d[0].input_ids.length];for(const h of Object.keys(d[0]))c[h]=new Z("int64",BigInt64Array.from(d.flatMap(f=>f[h]).map(BigInt)),p)}else{for(const p of Object.keys(d[0]))c[p]=d.map(h=>h[p]);if(!u)for(const p of Object.keys(c))c[p]=c[p][0]}return c}_encode_text(r){return r===null?null:(this.added_tokens_regex?r.split(this.added_tokens_regex).filter(i=>i):[r]).map((i,a)=>{if(this.added_tokens.find(l=>l.content===i)!==void 0)return i;{if(this.remove_space===!0&&(i=i.trim().split(/\s+/).join(" ")),this.do_lowercase_and_remove_accent&&(i=Mg(i)),this.normalizer!==null&&(i=this.normalizer(i)),i.length===0)return[];const l=this.pre_tokenizer!==null?this.pre_tokenizer(i,{section_index:a}):[i];return this.model(l)}}).flat()}_encode_plus(r,{text_pair:n=null,add_special_tokens:s=!0}={}){const{tokens:i,token_type_ids:a}=this._tokenize_helper(r,{pair:n,add_special_tokens:s}),o=this.model.convert_tokens_to_ids(i),l={input_ids:o,attention_mask:new Array(o.length).fill(1)};return this.return_token_type_ids&&a&&(l.token_type_ids=a),l}_tokenize_helper(r,{pair:n=null,add_special_tokens:s=!1}={}){const i=this._encode_text(r),a=this._encode_text(n);return this.post_processor?this.post_processor(i,a,{add_special_tokens:s}):{tokens:we(i??[],a??[])}}tokenize(r,{pair:n=null,add_special_tokens:s=!1}={}){return this._tokenize_helper(r,{pair:n,add_special_tokens:s}).tokens}encode(r,{text_pair:n=null,add_special_tokens:s=!0}={}){return this._encode_plus(r,{text_pair:n,add_special_tokens:s}).input_ids}batch_decode(r,n={}){return r instanceof Z&&(r=r.tolist()),r.map(s=>this.decode(s,n))}decode(r,n={}){if(r instanceof Z&&(r=kp(r)),!Array.isArray(r)||r.length===0||!cf(r[0]))throw Error("token_ids must be a non-empty array of integers.");return this.decode_single(r,n)}decode_single(r,{skip_special_tokens:n=!1,clean_up_tokenization_spaces:s=null}){let i=this.model.convert_ids_to_tokens(r);n&&(i=i.filter(o=>!this.special_tokens.includes(o)));let a=this.decoder?this.decoder(i):i.join(" ");return this.decoder&&this.decoder.end_of_word_suffix&&(a=a.replaceAll(this.decoder.end_of_word_suffix," "),n&&(a=a.trim())),(s??this.clean_up_tokenization_spaces)&&(a=bi(a)),a}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(r,{chat_template:n=null,add_generation_prompt:s=!1,tokenize:i=!0,padding:a=!1,truncation:o=!1,max_length:l=null,return_tensor:u=!0,return_dict:d=!1,tokenizer_kwargs:c={},...p}={}){if(this.chat_template&&typeof this.chat_template=="object"||this.chat_template===null&&this.default_chat_template&&typeof this.default_chat_template=="object"){const g=this.chat_template??this.default_chat_template;if(n!==null&&Object.hasOwn(g,n))n=g[n];else if(n===null&&"default"in g)n=g.default;else if(n===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(g).sort()}.`)}else n??(n=this.chat_template??this.default_chat_template);if(typeof n!="string")throw Error(`chat_template must be a string, but got ${typeof n}`);let h=this._compiled_template_cache.get(n);h===void 0&&(h=new Tg(n),this._compiled_template_cache.set(n,h));const f=Object.create(null);for(const g of w_){const w=this.getToken(g);w&&(f[g]=w)}const m=h.render({messages:r,add_generation_prompt:s,...f,...p});if(i){const g=this._call(m,{add_special_tokens:!1,padding:a,truncation:o,max_length:l,return_tensor:u,...c});return d?g:g.input_ids}return m}}class v_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class $_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class x_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class k_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class S_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class E_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class T_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class I_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class M_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class C_ extends X{}class z_ extends X{}class A_ extends X{constructor(r,n){super(r,n);T(this,"return_token_type_ids",!0);console.warn('WARNING: `XLMTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.')}}class P_ extends X{constructor(){super(...arguments);T(this,"return_token_type_ids",!0)}}class O_ extends X{}class zp extends X{constructor(){super(...arguments);T(this,"_default_chat_template",'{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}')}}class B_ extends X{}class Ap extends X{constructor(t,r){super(t,r),this.languageRegex=/^[a-z]{2}_[A-Z]{2}$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)),this.lang_to_token=n=>n}_build_translation_inputs(t,r,n){return vi(this,t,r,n)}}class R_ extends Ap{}class D_ extends X{}class L_ extends zp{constructor(t,r){var i,a;const n=".,!?…。，、।۔،",s=(a=(i=t.pre_tokenizer)==null?void 0:i.pretokenizers[0])==null?void 0:a.pattern;s&&s.Regex===` ?[^(\\s|[${n}])]+`&&(s.Regex=` ?[^\\s${n}]+`),super(t,r)}}const fn="▁";class Pp extends X{constructor(r,n){super(r,n);T(this,"_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 %}`);T(this,"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.`);T(this,"padding_side","left");this.use_default_system_prompt=n.use_default_system_prompt??!1,this.legacy=n.legacy??!0,this.legacy||(this.normalizer=null,this.pre_tokenizer=new Cp({replacement:fn,add_prefix_space:!0,prepend_scheme:"first"}))}_encode_text(r){if(r===null)return null;if(this.legacy||r.length===0)return super._encode_text(r);let n=super._encode_text(fn+r.replaceAll(fn," "));return n.length>1&&n[0]===fn&&this.special_tokens.includes(n[1])&&(n=n.slice(1)),n}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 F_ extends Pp{}class N_ extends X{}class U_ extends X{}class q_ extends X{}class G_ extends X{}class V_ extends X{}class j_ extends X{}class W_ extends X{constructor(){super(...arguments);T(this,"_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 H_ extends X{}function vi(e,t,r,n){if(!("language_codes"in e)||!Array.isArray(e.language_codes))throw new Error("Tokenizer must have `language_codes` attribute set and it should be an array of language ids.");if(!("languageRegex"in e)||!(e.languageRegex instanceof RegExp))throw new Error("Tokenizer must have `languageRegex` attribute set and it should be a regular expression.");if(!("lang_to_token"in e)||typeof e.lang_to_token!="function")throw new Error("Tokenizer must have `lang_to_token` attribute set and it should be a function.");const s=n.src_lang,i=n.tgt_lang;if(!e.language_codes.includes(i))throw new Error(`Target language code "${i}" is not valid. Must be one of: {${e.language_codes.join(", ")}}`);if(s!==void 0){if(!e.language_codes.includes(s))throw new Error(`Source language code "${s}" is not valid. Must be one of: {${e.language_codes.join(", ")}}`);for(const a of e.post_processor.config.single)if("SpecialToken"in a&&e.languageRegex.test(a.SpecialToken.id)){a.SpecialToken.id=e.lang_to_token(s);break}}return n.forced_bos_token_id=e.model.convert_tokens_to_ids([e.lang_to_token(i)])[0],e._call(t,r)}class K_ extends X{constructor(t,r){super(t,r),this.languageRegex=/^[a-z]{3}_[A-Z][a-z]{3}$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)),this.lang_to_token=n=>n}_build_translation_inputs(t,r,n){return vi(this,t,r,n)}}class Q_ extends X{constructor(t,r){super(t,r),this.languageRegex=/^__[a-z]{2,3}__$/,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(t,r,n){return vi(this,t,r,n)}}const Op=[["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"]],mn=new Map(Op),X_=new Map([...Op.map(([e,t])=>[t,e]),["burmese","my"],["valencian","ca"],["flemish","nl"],["haitian","ht"],["letzeburgesch","lb"],["pushto","ps"],["panjabi","pa"],["moldavian","ro"],["moldovan","ro"],["sinhalese","si"],["castilian","es"]]);class Y_ extends X{constructor(){super(...arguments);T(this,"_default_chat_template",'{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}')}_decode_asr(r,{return_timestamps:n=!1,return_language:s=!1,time_precision:i=null,force_full_sequences:a=!0}={}){if(i===null)throw Error("Must specify time_precision");let o=null;const l=n==="word";function u(){return{language:o,timestamp:[null,null],text:""}}const d=[];let c=u(),p=0;const h=this.model.convert_tokens_to_ids(["<|notimestamps|>"])[0]+1;let f=[],m=[],g=!1,w=null;const _=new Set(this.all_special_ids);for(const y of r){const S=y.tokens,x=l?y.token_timestamps:null;let I=null,D=h;if("stride"in y){const[q,H,V]=y.stride;if(p-=H,w=q-V,H&&(D=H/i+h),V)for(let E=S.length-1;E>=0;--E){const M=S[E];if(M>=h){if(I!==null&&(M-h)*i=h){const V=(H-h)*i+p,E=Rr(V,2);if(I!==null&&H>=I)g=!0;else if(g||f.length>0&&H0?(f.push(R),l&&m.push(U)):f.every(q=>q.length===0)&&(c=u(),f=[],R=[],m=[],U=[])}if(f.length>0){if(a&&n)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[y,S]=this.findLongestCommonSequence(f,m),x=this.decode(y);c.text=x,l&&(c.words=this.collateWordTimestamps(y,S,o)),d.push(c)}let v=Object.create(null);const b=d.map(y=>y.text).join("");if(n||s){for(let y=0;y0;let l=o?[]:null,u=o?n[0]:null;for(let d=1;dE===q[M]).length,V=H/y+S;H>1&&V>p&&(p=V,h=[x,I,R,U])}const[m,g,w,_]=h,v=Math.floor((g+m)/2),b=Math.floor((_+w)/2);a.push(...s.slice(0,v)),s=c.slice(b),i=s.length,o&&(l.push(...u.slice(0,v)),u=n[d].slice(b))}return a.push(...s),o?(l.push(...u),[a,l]):[a,[]]}collateWordTimestamps(r,n,s){const[i,a,o]=this.combineTokensIntoWords(r,s),l=[];for(let u=0;u=i){const l=Rr((Number(o)-i)*s,2);a.push(`<|${l}|>`),a.push([])}else a[a.length-1].push(o);return a=a.map(o=>typeof o=="string"?o:super.decode(o,n)),a.join("")}splitTokensOnUnicode(r){const n=this.decode(r,{decode_with_timestamps:!0}),s="�",i=[],a=[],o=[];let l=[],u=[],d=0;for(let c=0;c=this.model.tokens_to_ids.get("<|endoftext|>"),m=c.startsWith(" "),g=c.trim(),w=u.test(g);if(f||m||w||a.length===0)a.push(c),o.push(p),l.push(h);else{const _=a.length-1;a[_]+=c,o[_].push(...p),l[_].push(...h)}}return[a,o,l]}mergePunctuations(r,n,s,i,a){const o=structuredClone(r),l=structuredClone(n),u=structuredClone(s);let d=o.length-2,c=o.length-1;for(;d>=0;)o[d].startsWith(" ")&&i.includes(o[d].trim())?(o[c]=o[d]+o[c],l[c]=we(l[d],l[c]),u[c]=we(u[d],u[c]),o[d]="",l[d]=[],u[d]=[]):c=d,--d;for(d=0,c=1;cp),l.filter(p=>p.length>0),u.filter(p=>p.length>0)]}get_decoder_prompt_ids({language:r=null,task:n=null,no_timestamps:s=!0}={}){const i=[];if(r){r=r.toLowerCase();let a=X_.get(r);if(a===void 0)if(mn.has(r))a=r;else{const u=r.length===2?mn.keys():mn.values();throw new Error(`Language "${r}" is not supported. Must be one of: ${JSON.stringify(u)}`)}const o=this.model.tokens_to_ids.get(`<|${a}|>`);if(o===void 0)throw new Error(`Unable to find language "${a}" in model vocabulary. Please report this issue at https://github.com/xenova/transformers.js/issues/new/choose.`);i.push(o)}else i.push(null);if(n){if(n=n.toLowerCase(),n!=="transcribe"&&n!=="translate")throw new Error(`Task "${n}" is not supported. Must be one of: ["transcribe", "translate"]`);const a=this.model.tokens_to_ids.get(`<|${n}|>`);if(a===void 0)throw new Error(`Unable to find task "${n}" in model vocabulary. Please report this issue at https://github.com/xenova/transformers.js/issues/new/choose.`);i.push(a)}else i.push(null);if(s){const a=this.model.tokens_to_ids.get("<|notimestamps|>");if(a===void 0)throw new Error('Unable to find "<|notimestamps|>" in model vocabulary. Please report this issue at https://github.com/xenova/transformers.js/issues/new/choose.');i.push(a)}return i.map((a,o)=>[o+1,a]).filter(a=>a[1]!==null)}}class Z_ extends X{}class J_ extends X{}class ew extends X{}class tw extends X{constructor(t,r){super(t,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(t){if(t===null)return null;const[r,...n]=t.trim().split(this.languageRegex);if(n.length===0)return super._encode_text(r);if(n.length===2){const[s,i]=n;return this.supported_language_codes.includes(s)||console.warn(`Unsupported language code "${s}" detected, which may lead to unexpected behavior. Should be one of: ${JSON.stringify(this.supported_language_codes)}`),we([s],super._encode_text(i))}}}class rw extends X{}class Bp extends X{constructor(){super(...arguments);T(this,"_default_chat_template","{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ ' ' }}{% endif %}{% endfor %}{{ eos_token }}")}}class nw extends Bp{}class sw extends X{}class iw extends X{}class aw extends X{constructor(t,r){super(t,r),this.decoder=new c_({})}}class ow extends X{}class Rp{static async from_pretrained(t,{progress_callback:r=null,config:n=null,cache_dir:s=null,local_files_only:i=!1,revision:a="main",legacy:o=null}={}){var p;const[l,u]=await xp(t,{progress_callback:r,config:n,cache_dir:s,local_files_only:i,revision:a,legacy:o}),d=((p=u.tokenizer_class)==null?void 0:p.replace(/Fast$/,""))??"PreTrainedTokenizer";let c=this.TOKENIZER_CLASS_MAPPING[d];return c||(console.warn(`Unknown tokenizer class "${d}", attempting to construct from base class.`),c=X),new c(l,u)}}T(Rp,"TOKENIZER_CLASS_MAPPING",{T5Tokenizer:O_,DistilBertTokenizer:C_,CamembertTokenizer:z_,DebertaTokenizer:S_,DebertaV2Tokenizer:E_,BertTokenizer:v_,HerbertTokenizer:T_,ConvBertTokenizer:I_,RoFormerTokenizer:M_,XLMTokenizer:A_,ElectraTokenizer:P_,MobileBertTokenizer:x_,SqueezeBertTokenizer:k_,AlbertTokenizer:$_,GPT2Tokenizer:zp,BartTokenizer:B_,MBartTokenizer:Ap,MBart50Tokenizer:R_,RobertaTokenizer:D_,WhisperTokenizer:Y_,CodeGenTokenizer:Z_,CLIPTokenizer:J_,SiglipTokenizer:ew,MarianTokenizer:tw,BloomTokenizer:L_,NllbTokenizer:K_,M2M100Tokenizer:Q_,LlamaTokenizer:Pp,CodeLlamaTokenizer:F_,XLMRobertaTokenizer:N_,MPNetTokenizer:U_,FalconTokenizer:q_,GPTNeoXTokenizer:G_,EsmTokenizer:V_,Wav2Vec2CTCTokenizer:rw,BlenderbotTokenizer:Bp,BlenderbotSmallTokenizer:nw,SpeechT5Tokenizer:sw,NougatTokenizer:iw,VitsTokenizer:aw,Qwen2Tokenizer:j_,GemmaTokenizer:W_,Grok1Tokenizer:H_,CohereTokenizer:ow,PreTrainedTokenizer:X});async function lw(e,t){return await _t(e,"config.json",!0,t)}function gn(e){const t={};let r={};switch(e.model_type){case"llava":case"paligemma":r=gn(e.text_config);break;case"moondream1":r=gn(e.phi_config);break;case"musicgen":r=gn(e.decoder);break;case"gpt2":case"gptj":case"codegen":case"gpt_bigcode":t.num_heads="n_head",t.num_layers="n_layer",t.hidden_size="n_embd";break;case"gpt_neox":case"stablelm":case"opt":case"phi":case"phi3":case"falcon":t.num_heads="num_attention_heads",t.num_layers="num_hidden_layers",t.hidden_size="hidden_size";break;case"llama":case"mistral":case"starcoder2":case"qwen2":t.num_heads="num_key_value_heads",t.num_layers="num_hidden_layers",t.hidden_size="hidden_size",t.num_attention_heads="num_attention_heads";break;case"gemma":t.num_heads="num_key_value_heads",t.num_layers="num_hidden_layers",t.dim_kv="head_dim";break;case"openelm":t.num_heads="num_kv_heads",t.num_layers="num_transformer_layers",t.dim_kv="head_dim";break;case"gpt_neo":t.num_heads="num_heads",t.num_layers="num_layers",t.hidden_size="hidden_size";break;case"bloom":t.num_heads="n_head",t.num_layers="n_layer",t.hidden_size="hidden_size";break;case"mpt":t.num_heads="n_heads",t.num_layers="n_layers",t.hidden_size="d_model";break;case"t5":case"mt5":case"longt5":t.num_decoder_layers="num_decoder_layers",t.num_decoder_heads="num_heads",t.decoder_dim_kv="d_kv",t.num_encoder_layers="num_layers",t.num_encoder_heads="num_heads",t.encoder_dim_kv="d_kv";break;case"bart":case"mbart":case"marian":case"whisper":case"m2m_100":case"blenderbot":case"blenderbot-small":t.num_decoder_layers="decoder_layers",t.num_decoder_heads="decoder_attention_heads",t.decoder_hidden_size="d_model",t.num_encoder_layers="encoder_layers",t.num_encoder_heads="encoder_attention_heads",t.encoder_hidden_size="d_model";break;case"speecht5":t.num_decoder_layers="decoder_layers",t.num_decoder_heads="decoder_attention_heads",t.decoder_hidden_size="hidden_size",t.num_encoder_layers="encoder_layers",t.num_encoder_heads="encoder_attention_heads",t.encoder_hidden_size="hidden_size";break;case"trocr":t.num_encoder_layers=t.num_decoder_layers="decoder_layers",t.num_encoder_heads=t.num_decoder_heads="decoder_attention_heads",t.encoder_hidden_size=t.decoder_hidden_size="d_model";break;case"musicgen_decoder":t.num_encoder_layers=t.num_decoder_layers="num_hidden_layers",t.num_encoder_heads=t.num_decoder_heads="num_attention_heads",t.encoder_hidden_size=t.decoder_hidden_size="hidden_size";break}const n={...r,...gt(e,["model_type","multi_query","is_encoder_decoder"])};for(const s in t)n[s]=e[t[s]];return n}function Dp(e,{prefix:t="past_key_values",encoder_add_pkv:r=!0}={}){const n={},s=e.normalized_config,i=1;if(s.is_encoder_decoder&&r){const a=s.encoder_dim_kv??s.encoder_hidden_size/s.num_encoder_heads,o=s.decoder_dim_kv??s.decoder_hidden_size/s.num_decoder_heads,l=[i,s.num_encoder_heads,0,a],u=[i,s.num_decoder_heads,0,o];for(let d=0;d1 to use the classifier free guidance processor, got guidance scale ${t}.`);this.guidance_scale=t}_call(t,r){if(r.dims[0]!==2*t.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 ${t.length} for the input ids.`);const n=t.length,s=r.slice([0,n],null),i=r.slice([n,r.dims[0]],null);for(let a=0;a1)throw new Error(`\`top_p\` must be a float > 0 and < 1, but is ${t}`);if(!Number.isInteger(n)||n<1)throw new Error(`\`min_tokens_to_keep\` must be a positive integer, but is ${n}`);this.top_p=t,this.filter_value=r,this.min_tokens_to_keep=n}}class $w extends $i{constructor(t,{filter_value:r=-1/0,min_tokens_to_keep:n=1}={}){if(super(),!Number.isInteger(t)||t<0)throw new Error(`\`top_k\` must be a positive integer, but is ${t}`);this.top_k=Math.max(t,n),this.filter_value=r}}class xw{constructor(t){T(this,"max_length",20);T(this,"max_new_tokens",null);T(this,"min_length",0);T(this,"min_new_tokens",null);T(this,"early_stopping",!1);T(this,"max_time",null);T(this,"do_sample",!1);T(this,"num_beams",1);T(this,"num_beam_groups",1);T(this,"penalty_alpha",null);T(this,"use_cache",!0);T(this,"temperature",1);T(this,"top_k",50);T(this,"top_p",1);T(this,"typical_p",1);T(this,"epsilon_cutoff",0);T(this,"eta_cutoff",0);T(this,"diversity_penalty",0);T(this,"repetition_penalty",1);T(this,"encoder_repetition_penalty",1);T(this,"length_penalty",1);T(this,"no_repeat_ngram_size",0);T(this,"bad_words_ids",null);T(this,"force_words_ids",null);T(this,"renormalize_logits",!1);T(this,"constraints",null);T(this,"forced_bos_token_id",null);T(this,"forced_eos_token_id",null);T(this,"remove_invalid_values",!1);T(this,"exponential_decay_length_penalty",null);T(this,"suppress_tokens",null);T(this,"begin_suppress_tokens",null);T(this,"forced_decoder_ids",null);T(this,"guidance_scale",null);T(this,"num_return_sequences",1);T(this,"output_attentions",!1);T(this,"output_hidden_states",!1);T(this,"output_scores",!1);T(this,"return_dict_in_generate",!1);T(this,"pad_token_id",null);T(this,"bos_token_id",null);T(this,"eos_token_id",null);T(this,"encoder_no_repeat_ngram_size",0);T(this,"decoder_start_token_id",null);T(this,"generation_kwargs",{});Object.assign(this,gt(t,Object.getOwnPropertyNames(this)))}}class yn extends Ce{_call(t,r){throw Error("StoppingCriteria needs to be subclassed")}}class xi extends Ce{constructor(){super(),this.criteria=[]}push(t){this.criteria.push(t)}extend(t){t instanceof xi?t=t.criteria:t instanceof yn&&(t=[t]),this.criteria.push(...t)}_call(t,r){const n=new Array(t.length).fill(!1);for(const s of this.criteria){const i=s(t,r);for(let a=0;ar.length>=this.max_length)}}class Sw extends yn{constructor(t){super(),Array.isArray(t)||(t=[t]),this.eos_token_id=t}_call(t,r){return t.map(n=>{const s=n.at(-1);return this.eos_token_id.some(i=>s==i)})}}class bn extends Ce{constructor(t){super(),this.generation_config=t}_call(t,r=-1){return this.sample(t,r)}sample(t,r){throw Error("sample should be implemented in subclasses.")}getLogits(t,r){let n=t.dims.at(-1),s=t.data;if(r===-1)s=s.slice(-n);else{let i=r*n;s=s.slice(i,i+n)}return s}randomSelect(t){let r=t.reduce((s,i)=>s+i,0),n=Math.random()*r;for(let s=0;s1)return new Iw(t);if(t.num_return_sequences>1)throw Error(`num_return_sequences has to be 1 when doing greedy search, but is ${t.num_return_sequences}.`);return new Ew(t)}}class Ew extends bn{sample(t,r=-1){let n=this.getLogits(t,r);return[[On(n)[1],0]]}}class Tw extends bn{sample(t,r=-1){let n=t.dims.at(-1);this.generation_config.top_k>0&&(n=Math.min(this.generation_config.top_k,n));const s=this.getLogits(t,r),i=Ui(s,n),a=Ni(i.map(o=>o[1]));return Array.from({length:this.generation_config.num_beams},()=>{const o=this.randomSelect(a);return[i[o][0],Math.log(a[o])]})}}class Iw extends bn{sample(t,r=-1){let n=t.dims.at(-1);this.generation_config.top_k>0&&(n=Math.min(this.generation_config.top_k,n));const s=this.getLogits(t,r),i=Ui(s,n),a=Ni(i.map(o=>o[1]));return Array.from({length:this.generation_config.num_beams},(o,l)=>[i[l][0],Math.log(a[l])])}}const W={EncoderOnly:0,EncoderDecoder:1,Seq2Seq:2,Vision2Seq:3,DecoderOnly:4,MaskGeneration:5,ImageTextToText:6,Musicgen:7},vn=new Map,Up=new Map,pr=new Map;async function Mw(e,t,r){let n=r.device;n&&typeof n!="string"&&(n.hasOwnProperty(t)?n=n[t]:(console.warn(`Device not specified for ${t}. Using the default device.`),n=null));const s=Vm(n);let i=r.dtype;if(typeof i!="string"&&(i&&i.hasOwnProperty(t)?i=i[t]:(i=dw[s[0]],console.warn(`Dtype not specified for ${t}. Using the default dtype: ${i}.`))),Fp.hasOwnProperty(i)){if(i===Me.fp16&&!await uw())throw new Error("The device does not support fp16.")}else throw new Error(`Invalid dtype: ${i}. Should be one of: ${Object.keys(Me).join(", ")}`);const a=Fp[i],o=`${r.subfolder??""}/${t}${a}.onnx`,l={...r.session_options};l.executionProviders??(l.executionProviders=s);const u=Or(e,o,!0,r);let d=[];if(r.use_external_data_format){if(mt.IS_NODE_ENV)throw new Error("External data format is not yet supported in Node.js");const p=`${t}${a}.onnx_data`,h=`${r.subfolder??""}/${p}`;d.push(new Promise(async(f,m)=>{const g=await Or(e,h,!0,r);f({path:p,data:g})}))}else l.externalData!==void 0&&(d=l.externalData.map(async p=>{if(typeof p.data=="string"){const h=await Or(e,p.data,!0,r);return{...p,data:h}}return p}));if(d.length>0&&(l.externalData=await Promise.all(d)),n==="webgpu"){const p=Dp(r.config,{prefix:"present"}),h={};for(const f in p)h[f]="gpu-buffer";l.preferredOutputLocation=h}return{buffer:await u,session_options:l}}async function zt(e,t,r){const n=Object.keys(t),s=await Promise.all(n.map(async a=>Mw(e,t[a],r))),i={};for(let a=0;a0)throw new Error(`An error occurred during model execution: "Missing the following inputs: ${n.join(", ")}.`);const s=Object.keys(t).length,i=e.inputNames.length;if(s>i){let a=Object.keys(t).filter(o=>!e.inputNames.includes(o));console.warn(`WARNING: Too many inputs were provided (${s} > ${i}). The following inputs will be ignored: "${a.join(", ")}".`)}return r}async function ht(e,t){const r=Cw(e,t);try{const n=Object.fromEntries(Object.entries(r).map(([i,a])=>[i,a.ort_tensor]));let s=await e.run(n);return s=qp(s),s}catch(n){throw console.error(`An error occurred during model execution: "${n}".`),console.error("Inputs given to model:",r),n}}function qp(e){for(let t in e)cp(e[t])?e[t]=new Z(e[t]):typeof e[t]=="object"&&qp(e[t]);return e}function zw(e){if(e instanceof Z)return e;if(e.length===0)throw Error("items must be non-empty");if(Array.isArray(e[0])){if(e.some(t=>t.length!==e[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 Z("int64",BigInt64Array.from(e.flat().map(t=>BigInt(t))),[e.length,e[0].length])}else return new Z("int64",BigInt64Array.from(e.map(t=>BigInt(t))),[1,e.length])}function Gp(e){return new Z("bool",[e],[1])}async function Vp(e,t){let{encoder_outputs:r,past_key_values:n}=t;if(!r){const l=gt(t,e.sessions.model.inputNames);r=(await hr(e,l)).last_hidden_state}const{input_ids:s,decoder_input_ids:i,...a}=t;return a.input_ids=i,a.encoder_hidden_states=r,e.sessions.decoder_model_merged.inputNames.includes("encoder_attention_mask")&&(a.encoder_attention_mask=t.attention_mask),await ki(e,a,!0)}async function hr(e,t){const r=e.sessions.model,n=Object.create(null);for(const s of r.inputNames)n[s]=t[s];return r.inputNames.includes("token_type_ids")&&!n.token_type_ids&&(n.token_type_ids=new Z("int64",new BigInt64Array(n.input_ids.data.length),n.input_ids.dims)),await ht(r,n)}async function ki(e,t,r=!1){const n=e.sessions[r?"decoder_model_merged":"model"],{past_key_values:s,...i}=t;n.inputNames.includes("use_cache_branch")&&(i.use_cache_branch=Gp(!!s)),n.inputNames.includes("position_ids")&&i.attention_mask&&!i.position_ids&&(i.position_ids=Pw(i,s)),e.addPastKeyValues(i,s);const a=gt(i,n.inputNames);return await ht(n,a)}async function Aw(e,{input_ids:t=null,attention_mask:r=null,pixel_values:n=null,position_ids:s=null,inputs_embeds:i=null,past_key_values:a=null,generation_config:o=null,logits_processor:l=null,...u}){if(!i){if(i=await e.encode_text({input_ids:t}),n&&t.dims[1]!==1){const c=await e.encode_image({pixel_values:n});({inputs_embeds:i,attention_mask:r}=e._merge_input_ids_with_image_features({image_features:c,inputs_embeds:i,input_ids:t,attention_mask:r}))}else if(a&&n&&t.dims[1]===1){const c=t.dims[1],p=Object.values(a)[0].dims.at(-2);r=it([ir([t.dims[0],p]),r.slice(null,[r.dims[1]-c,r.dims[1]])],1)}}return await ki(e,{inputs_embeds:i,past_key_values:a,attention_mask:r,position_ids:s,generation_config:o,logits_processor:l},!0)}function Pw(e,t=null){const{input_ids:r,inputs_embeds:n,attention_mask:s}=e,[i,a]=s.dims,o=new BigInt64Array(s.data.length);for(let u=0;ui.dims[1])){if(so==e.config.image_token_index)){const o=e.config.num_image_tokens;if(!o)throw new Error("`num_image_tokens` is missing in the model configuration.");const l=i.dims[1]-(s-o);r.input_ids=i.slice(null,[-l,null]),r.attention_mask=ir([1,s+l])}}}return r}function Ow(e,t,r,n){const{...s}=r;return r.past_key_values&&(t=t.map(a=>[a.at(-1)])),s.decoder_input_ids=zw(t),s}class P extends Ce{constructor(r,n){super();T(this,"main_input_name","input_ids");T(this,"forward_params",["input_ids","attention_mask"]);this.config=r,this.sessions=n;const s=pr.get(this.constructor),i=vn.get(s);this.can_generate=!1,this._forward=null,this._prepare_inputs_for_generation=null,i===W.DecoderOnly?(this.can_generate=!0,this._forward=ki,this._prepare_inputs_for_generation=jp):i===W.Seq2Seq||i===W.Vision2Seq||i===W.Musicgen?(this.can_generate=!0,this._forward=Vp,this._prepare_inputs_for_generation=Ow):i===W.EncoderDecoder?this._forward=Vp:i===W.ImageTextToText?(this.can_generate=!0,this._forward=Aw,this._prepare_inputs_for_generation=jp):this._forward=hr,this.can_generate&&this.forward_params.push("past_key_values"),this.custom_config=this.config["transformers.js_config"]??{}}async dispose(){var n;const r=[];for(const s of Object.values(this.sessions))(n=s==null?void 0:s.handler)!=null&&n.dispose&&r.push(s.handler.dispose());return await Promise.all(r)}static async from_pretrained(r,{progress_callback:n=null,config:s=null,cache_dir:i=null,local_files_only:a=!1,revision:o="main",model_file_name:l=null,subfolder:u="onnx",device:d=null,dtype:c=null,use_external_data_format:p=null,session_options:h={}}={}){let f={progress_callback:n,config:s,cache_dir:i,local_files_only:a,revision:o,model_file_name:l,subfolder:u,device:d,dtype:c,use_external_data_format:p,session_options:h};const m=pr.get(this),g=vn.get(m);f.config=await Lp.from_pretrained(r,f);let w;return g===W.DecoderOnly?w=await Promise.all([zt(r,{model:f.model_file_name??"model"},f),_t(r,"generation_config.json",!1,f)]):g===W.Seq2Seq||g===W.Vision2Seq?w=await Promise.all([zt(r,{model:"encoder_model",decoder_model_merged:"decoder_model_merged"},f),_t(r,"generation_config.json",!1,f)]):g===W.MaskGeneration?w=await Promise.all([zt(r,{model:"vision_encoder",prompt_encoder_mask_decoder:"prompt_encoder_mask_decoder"},f)]):g===W.EncoderDecoder?w=await Promise.all([zt(r,{model:"encoder_model",decoder_model_merged:"decoder_model_merged"},f)]):g===W.ImageTextToText?w=await Promise.all([zt(r,{embed_tokens:"embed_tokens",vision_encoder:"vision_encoder",decoder_model_merged:"decoder_model_merged"},f),_t(r,"generation_config.json",!1,f)]):g===W.Musicgen?w=await Promise.all([zt(r,{model:"text_encoder",decoder_model_merged:"decoder_model_merged",encodec_decode:"encodec_decode"},f),_t(r,"generation_config.json",!1,f)]):(g!==W.EncoderOnly&&console.warn(`Model type for '${m??(s==null?void 0:s.model_type)}' not found, assuming encoder-only architecture. Please report this at https://github.com/xenova/transformers.js/issues/new/choose.`),w=await Promise.all([zt(r,{model:f.model_file_name??"model"},f)])),new this(f.config,...w)}async _call(r){return await this.forward(r)}async forward(r){return await this._forward(this,r)}_get_logits_warper(r){const n=new Np;return r.temperature!==null&&r.temperature!==1&&n.push(new bw(r.temperature)),r.top_k!==null&&r.top_k!==0&&n.push(new $w(r.top_k)),r.top_p!==null&&r.top_p<1&&n.push(new vw(r.top_p)),n}_get_logits_processor(r,n,s=null){const i=new Np;if(r.repetition_penalty!==null&&r.repetition_penalty!==1&&i.push(new mw(r.repetition_penalty)),r.no_repeat_ngram_size!==null&&r.no_repeat_ngram_size>0&&i.push(new fw(r.no_repeat_ngram_size)),r.bad_words_ids!==null&&i.push(new ww(r.bad_words_ids,r.eos_token_id)),r.min_length!==null&&r.eos_token_id!==null&&r.min_length>0&&i.push(new gw(r.min_length,r.eos_token_id)),r.min_new_tokens!==null&&r.eos_token_id!==null&&r.min_new_tokens>0&&i.push(new _w(n,r.min_new_tokens,r.eos_token_id)),r.forced_bos_token_id!==null&&i.push(new cw(r.forced_bos_token_id)),r.forced_eos_token_id!==null&&i.push(new pw(r.max_length,r.forced_eos_token_id)),r.begin_suppress_tokens!==null){let a=n>1||r.forced_bos_token_id===null?n:n+1;r.forced_decoder_ids!==null&&(a+=r.forced_decoder_ids[r.forced_decoder_ids.length-1][0]),i.push(new hw(r.begin_suppress_tokens,a))}return r.guidance_scale!==null&&r.guidance_scale>1&&i.push(new yw(r.guidance_scale)),s!==null&&i.extend(s),i}_prepare_generation_config(r,n){const s=new xw(this.config);return"generation_config"in this&&Object.assign(s,this.generation_config),r&&Object.assign(s,r),n&&Object.assign(s,gt(n,Object.getOwnPropertyNames(s))),s}_get_stopping_criteria(r,n=null){const s=new xi;return r.max_length!==null&&s.push(new kw(r.max_length,this.config.max_position_embeddings??null)),r.eos_token_id!==null&&s.push(new Sw(r.eos_token_id)),n&&s.extend(n),s}_validate_model_class(){if(!this.can_generate){const r=[Ci,Hh,Wh,jh],n=pr.get(this.constructor),s=new Set,i=this.config.model_type;for(const o of r){const l=o.get(i);l&&s.add(l[0])}let a=`The current model class (${n}) is not compatible with \`.generate()\`, as it doesn't have a language model head.`;throw s.size>0&&(a+=` Please use the following class instead: ${[...s].join(", ")}`),Error(a)}}prepare_inputs_for_generation(...r){return this._prepare_inputs_for_generation(this,...r)}_update_model_kwargs_for_generation({generated_input_ids:r,outputs:n,model_inputs:s,is_encoder_decoder:i}){return s.past_key_values=this.getPastKeyValues(n,s.past_key_values),s.input_ids=new Z("int64",r.flat(),[r.length,1]),i||(s.attention_mask=it([s.attention_mask,ir([s.attention_mask.dims[0],1])],1)),s.position_ids=null,s}_prepare_model_inputs({inputs:r,bos_token_id:n,model_kwargs:s}){const i=gt(s,this.forward_params),a=this.main_input_name;if(a in i){if(r)throw new Error("`inputs`: {inputs}` were passed alongside {input_name} which is not allowed. Make sure to either pass {inputs} or {input_name}=...")}else i[a]=r;return{inputs_tensor:i[a],model_inputs:i,model_input_name:a}}async _prepare_encoder_decoder_kwargs_for_generation({inputs_tensor:r,model_inputs:n,model_input_name:s,generation_config:i}){const a=gt(n,this.sessions.model.inputNames);let{last_hidden_state:o}=await hr(this,a);return i.guidance_scale!==null&&i.guidance_scale>1&&(o=it([o,Jm(o,0)],0),"attention_mask"in n&&(n.attention_mask=it([n.attention_mask,rg(n.attention_mask)],0))),n.encoder_outputs=o,n}_prepare_decoder_input_ids_for_generation({batch_size:r,model_input_name:n,model_kwargs:s,decoder_start_token_id:i,bos_token_id:a,generation_config:o}){i=i??a;let l;if(this.config.model_type==="musicgen")l=new Array(r*this.config.decoder.num_codebooks).fill(i);else if(Array.isArray(i)){if(i.length!==r)throw new Error(`\`decoder_start_token_id\` expcted to have length ${r} but got ${i.length}`);l=i}else l=new Array(r).fill(i);const d=new Z("int64",l,[l.length,1]);return s.decoder_attention_mask=eg(d),{input_ids:d,model_inputs:s}}async generate({inputs:r=null,generation_config:n=null,logits_processor:s=null,stopping_criteria:i=null,streamer:a=null,...o}){this._validate_model_class(),n=this._prepare_generation_config(n,o);let{inputs_tensor:l,model_inputs:u,model_input_name:d}=this._prepare_model_inputs({inputs:r,model_kwargs:o});const c=this.config.is_encoder_decoder;c&&("encoder_outputs"in u||(u=await this._prepare_encoder_decoder_kwargs_for_generation({inputs_tensor:l,model_inputs:u,model_input_name:d,generation_config:n})));let p;c?{input_ids:p,model_inputs:u}=this._prepare_decoder_input_ids_for_generation({batch_size:u[d].dims.at(0),model_input_name:d,model_kwargs:u,decoder_start_token_id:n.decoder_start_token_id,bos_token_id:n.bos_token_id,generation_config:n}):p=u[d];let h=p.dims.at(-1);n.max_new_tokens!==null&&(n.max_length=h+n.max_new_tokens);const f=this._get_logits_processor(n,h,s),m=this._get_stopping_criteria(n,i),g=u[d].dims.at(0),w=bn.getSampler(n),_=new Array(g).fill(0),v=p.tolist();a&&a.put(v);let b=null;for(;;){u=this.prepare_inputs_for_generation(v,u,n);const S=await this.forward(u),x=S.logits.slice(null,-1,null),I=f(v,x),D=[];for(let U=0;UU)){n.return_dict_in_generate&&(b=this.getPastKeyValues(S,u.past_key_values,!1));break}u=this._update_model_kwargs_for_generation({generated_input_ids:D,outputs:S,model_inputs:u,is_encoder_decoder:c})}a&&a.end();const y=new Z("int64",v.flat(),[v.length,v[0].length]);return n.return_dict_in_generate?{sequences:y,past_key_values:b}:y}addAttentionsToBeam(r,n){if(this.config.is_encoder_decoder){if(!n.cross_attentions||n.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`.");r.cross_attentions||(r.cross_attentions=[]),r.cross_attentions.push(n.cross_attentions)}if(!n.decoder_attentions||n.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`.");r.decoder_attentions||(r.decoder_attentions=[]),r.decoder_attentions.push(n.decoder_attentions)}groupBeams(r){const n=Object.create(null);for(const s of r)n[s.id]===void 0?n[s.id]=[s]:n[s.id].push(s);return Object.values(n)}getPastKeyValues(r,n,s=!0){const i=Object.create(null);for(const a in r)if(a.startsWith("present")){let o=a.replace("present","past_key_values");if(n&&a.includes("encoder"))i[o]=n[o];else{if(s&&n){const l=n[o];l.location==="gpu-buffer"&&l.dispose()}i[o]=r[a]}}return i}getAttentions(r){const n=Object.create(null);for(const s of["cross_attentions","decoder_attentions"]){const i=[];for(const a in r)if(a.startsWith(s)){const o=a.split(".").pop();i[o]=r[a]}n[s]=i}return n}addPastKeyValues(r,n){if(n)Object.assign(r,n);else{const s=this.custom_config.kv_cache_dtype??"float32",i=s==="float16"?new Uint16Array:[],a=Dp(this.config,{encoder_add_pkv:this.add_encoder_pkv??!0});for(const o in a)r[o]=new Z(s,i,a[o])}}}class Le{}class fr extends P{}class Bw extends fr{}class Rw extends fr{async _call(t){return new ve(await super._call(t))}}class Dw extends fr{async _call(t){return new re(await super._call(t))}}class Lw extends fr{async _call(t){return new be(await super._call(t))}}class Fw extends fr{async _call(t){return new Se(await super._call(t))}}class Nw extends P{}class Uw extends Nw{}class mr extends P{}class qw extends mr{}class Gw extends mr{async _call(t){return new ve(await super._call(t))}}class Vw extends mr{async _call(t){return new re(await super._call(t))}}class jw extends mr{async _call(t){return new be(await super._call(t))}}class Ww extends mr{async _call(t){return new Se(await super._call(t))}}class gr extends P{}class Hw extends gr{}class Kw extends gr{async _call(t){return new ve(await super._call(t))}}class Qw extends gr{async _call(t){return new re(await super._call(t))}}class Xw extends gr{async _call(t){return new be(await super._call(t))}}class Yw extends gr{async _call(t){return new Se(await super._call(t))}}class _r extends P{}class Zw extends _r{}class Jw extends _r{async _call(t){return new ve(await super._call(t))}}class ey extends _r{async _call(t){return new re(await super._call(t))}}class ty extends _r{async _call(t){return new be(await super._call(t))}}class ry extends _r{async _call(t){return new Se(await super._call(t))}}class wr extends P{}class ny extends wr{}class sy extends wr{async _call(t){return new ve(await super._call(t))}}class iy extends wr{async _call(t){return new re(await super._call(t))}}class ay extends wr{async _call(t){return new be(await super._call(t))}}class oy extends wr{async _call(t){return new Se(await super._call(t))}}class yr extends P{}class ly extends yr{}class uy extends yr{async _call(t){return new ve(await super._call(t))}}class dy extends yr{async _call(t){return new re(await super._call(t))}}class cy extends yr{async _call(t){return new be(await super._call(t))}}class py extends yr{async _call(t){return new Se(await super._call(t))}}class br extends P{}class hy extends br{}class fy extends br{async _call(t){return new ve(await super._call(t))}}class my extends br{async _call(t){return new re(await super._call(t))}}class gy extends br{async _call(t){return new be(await super._call(t))}}class _y extends br{async _call(t){return new Se(await super._call(t))}}class vr extends P{}class wy extends vr{}class yy extends vr{async _call(t){return new re(await super._call(t))}}class by extends vr{async _call(t){return new be(await super._call(t))}}class vy extends vr{async _call(t){return new Se(await super._call(t))}}class $y extends vr{async _call(t){return new ve(await super._call(t))}}class $n extends P{}class xy extends $n{}class ky extends $n{async _call(t){return new ve(await super._call(t))}}class Sy extends $n{async _call(t){return new re(await super._call(t))}}class Ey extends $n{async _call(t){return new be(await super._call(t))}}class xn extends P{}class Ty extends xn{}class Iy extends xn{async _call(t){return new ve(await super._call(t))}}class My extends xn{async _call(t){return new re(await super._call(t))}}class Cy extends xn{async _call(t){return new Se(await super._call(t))}}class $r extends P{}class zy extends $r{}class Ay extends $r{async _call(t){return new ve(await super._call(t))}}class Py extends $r{async _call(t){return new re(await super._call(t))}}class Oy extends $r{async _call(t){return new be(await super._call(t))}}class By extends $r{async _call(t){return new Se(await super._call(t))}}class kn extends P{}class Ry extends kn{}class Dy extends kn{async _call(t){return new ve(await super._call(t))}}class Ly extends kn{async _call(t){return new re(await super._call(t))}}class Fy extends kn{async _call(t){return new Se(await super._call(t))}}class Sn extends P{}class Ny extends Sn{}class Uy extends Sn{async _call(t){return new re(await super._call(t))}}class qy extends Sn{async _call(t){return new Se(await super._call(t))}}class Gy extends Sn{async _call(t){return new ve(await super._call(t))}}class Wp extends P{constructor(r,n,s){super(r,n);T(this,"forward_params",["input_ids","attention_mask","encoder_outputs","decoder_input_ids","decoder_attention_mask","past_key_values"]);this.generation_config=s}}class Vy extends Wp{}class jy extends Wp{}class Hp extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Wy extends Hp{}class Hy extends Hp{}class Kp extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Ky extends Kp{}class Qy extends Kp{}class Si extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Xy extends Si{}class Yy extends Si{}class Zy extends Si{async _call(t){return new re(await super._call(t))}}class En extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Jy extends En{}class eb extends En{}class tb extends En{async _call(t){return new re(await super._call(t))}}class rb extends En{}class Qp extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class nb extends Qp{}class sb extends Qp{}class Xp extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class ib extends Xp{}class ab extends Xp{}class xr extends P{}class ob extends xr{}class lb extends xr{async _call(t){return new ve(await super._call(t))}}class ub extends xr{async _call(t){return new re(await super._call(t))}}class db extends xr{async _call(t){return new be(await super._call(t))}}class cb extends xr{async _call(t){return new Se(await super._call(t))}}class kr extends P{}class pb extends kr{}class hb extends kr{async _call(t){return new ve(await super._call(t))}}class fb extends kr{async _call(t){return new re(await super._call(t))}}class mb extends kr{async _call(t){return new be(await super._call(t))}}class gb extends kr{async _call(t){return new Se(await super._call(t))}}class Sr extends P{}class _b extends Sr{}class wb extends Sr{async _call(t){return new ve(await super._call(t))}}class yb extends Sr{async _call(t){return new re(await super._call(t))}}class bb extends Sr{async _call(t){return new be(await super._call(t))}}class vb extends Sr{async _call(t){return new Se(await super._call(t))}}class Yp extends P{}class $b extends Yp{}class xb extends Yp{}class Zp extends P{constructor(r,n,s){super(r,n);T(this,"requires_attention_mask",!1);T(this,"main_input_name","input_features");T(this,"forward_params",["input_features","attention_mask","decoder_input_ids","decoder_attention_mask","past_key_values"]);this.generation_config=s}}class kb extends Zp{}class Sb extends Zp{_retrieve_init_tokens(t){throw t.decoder_start_token_id,new Error("Not implemented yet")}async generate({inputs:t=null,generation_config:r=null,logits_processor:n=null,stopping_criteria:s=null,language:i=null,task:a=null,...o}){throw new Error("WhisperForConditionalGeneration.generate is not yet in Transformers.js v3.")}_extract_token_timestamps(t,r,n=null,s=.02){if(!t.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 a=t.cross_attentions.map(u=>{let d=Array.from({length:this.config.decoder_layers},(g,w)=>it(u.map(_=>_[w]),2)),c=fi(r.map(([g,w])=>n?d[g].slice(null,w,null,[0,n]):d[g].slice(null,w)));c=c.transpose(1,0,2,3);let[p,h]=Qm(c,-2,0,!0),f=c.clone();for(let g=0;gc[w+1]-c[w]),f=we([1],h).map(g=>!!g),m=[];for(let g=0;gp.findIndex(h=>h==i)),l=o.every(p=>p===-1),u=o.every(p=>p!==-1);if(!l&&!u)throw new Error("Every input should contain either 0 or 1 image token.");if(l)return{inputs_embeds:t,attention_mask:s};const d=[],c=[];for(let p=0;pi*a,1);t.input_labels=new Z("int64",new BigInt64Array(s).fill(1n),n)}const r={image_embeddings:t.image_embeddings,image_positional_embeddings:t.image_positional_embeddings};return t.input_points&&(r.input_points=t.input_points),t.input_labels&&(r.input_labels=t.input_labels),t.input_boxes&&(r.input_boxes=t.input_boxes),await ht(this.sessions.prompt_encoder_mask_decoder,r)}async _call(t){return new ov(await super._call(t))}}class ov extends Le{constructor({iou_scores:t,pred_masks:r}){super(),this.iou_scores=t,this.pred_masks=r}}class Oh extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class lv extends Oh{}class uv extends Oh{}class Bh extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class dv extends Bh{}class cv extends Bh{}class At extends P{}class pv extends At{}class hv extends At{async _call(t){return new Ut(await super._call(t))}}class fv extends At{async _call(t){return new re(await super._call(t))}}class mv extends At{async _call(t){return new be(await super._call(t))}}class Ti extends P{}class gv extends Ti{}class _v extends Ti{async _call(t){return new Ut(await super._call(t))}}class wv extends Ti{async _call(t){return new re(await super._call(t))}}class In extends P{}class yv extends In{}class bv extends In{async _call(t){return new Ut(await super._call(t))}}class vv extends In{async _call(t){return new re(await super._call(t))}}class $v extends In{async _call(t){return new be(await super._call(t))}}class Ii extends P{}class xv extends Ii{}class kv extends Ii{async _call(t){return new Ut(await super._call(t))}}class Sv extends Ii{async _call(t){return new re(await super._call(t))}}class Ev extends At{}class Tv extends At{async _call(t){return new Ut(await super._call(t))}}class Iv extends At{async _call(t){return new re(await super._call(t))}}class Er extends P{}class Mv extends Er{}class Cv extends Er{async _call(t){return new Ut(await super._call(t))}}class zv extends Er{async _call(t){return new re(await super._call(t))}}class Av extends Er{async _call(t){return new E$(await super._call(t))}}class Pv extends Er{async _call(t){return new be(await super._call(t))}}class Rh extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Ov extends Rh{}class Bv extends Rh{async generate_speech(t,r,{threshold:n=.5,minlenratio:s=0,maxlenratio:i=20,vocoder:a=null}={}){const o={input_ids:t},{encoder_outputs:l,encoder_attention_mask:u}=await hr(this,o),d=l.dims[1]/this.config.reduction_factor,c=Math.floor(d*i),p=Math.floor(d*s),h=this.config.num_mel_bins;let f=[],m=null,g=null,w=0;for(;;){++w;const b=Gp(!!g);let y;g?y=g.output_sequence_out:y=new Z("float32",new Float32Array(h),[1,1,h]);let S={use_cache_branch:b,output_sequence:y,encoder_attention_mask:u,speaker_embeddings:r,encoder_hidden_states:l};this.addPastKeyValues(S,m),g=await ht(this.sessions.decoder_model_merged,S),m=this.getPastKeyValues(g,m);const{prob:x,spectrum:I}=g;if(f.push(I),w>=p&&(Array.from(x.data).filter(D=>D>=n).length>0||w>=c))break}const _=it(f),{waveform:v}=await ht(a.sessions.model,{spectrogram:_});return{spectrogram:_,waveform:v}}}class Rv extends P{constructor(){super(...arguments);T(this,"main_input_name","spectrogram")}}class Dv extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Lv extends Dv{}class Dh extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Fv extends Dh{}class Nv extends Dh{}class Lh extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Uv extends Lh{}class qv extends Lh{}class Fh extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Gv extends Fh{}class Vv extends Fh{}class Mi extends P{}class jv extends Mi{}class Wv extends Mi{static async from_pretrained(t,r={}){return r.model_file_name??(r.model_file_name="text_model"),super.from_pretrained(t,r)}}class Hv extends Mi{static async from_pretrained(t,r={}){return r.model_file_name??(r.model_file_name="audio_model"),super.from_pretrained(t,r)}}class Kv extends P{}class Nh extends Kv{async _call(t){return new I$(await super._call(t))}}class Uh extends P{}class Qv extends Uh{}class Xv extends Uh{}class qh extends P{constructor(t,r,n){super(t,r),this.generation_config=n}}class Yv extends qh{}class Zv extends qh{}class Gh extends P{}class Jv extends Gh{}class e$ extends Gh{async _call(t){return new re(await super._call(t))}}class Vh extends P{constructor(r,n,s){super(r,n);T(this,"forward_params",["input_ids","attention_mask","encoder_outputs","decoder_input_ids","decoder_attention_mask","past_key_values"]);this.generation_config=s}_apply_and_filter_by_delay_pattern_mask(r){const[n,s]=r.dims,i=this.config.decoder.num_codebooks,a=s-i;let o=0;for(let d=0;d0&&h<=a&&(r.data[o++]=r.data[d])}const l=Math.floor(n/i),u=o/(l*i);return new Z(r.type,r.data.slice(0,o),[l,i,u])}prepare_inputs_for_generation(r,n,s){let i=structuredClone(r);for(let o=0;o=l&&(i[o][l]=BigInt(this.config.decoder.pad_token_id));return s.guidance_scale!==null&&s.guidance_scale>1&&(i=i.concat(i)),super.prepare_inputs_for_generation(i,n,s)}async generate(r){const n=await super.generate(r),s=this._apply_and_filter_by_delay_pattern_mask(n).unsqueeze_(0),{audio_values:i}=await ht(this.sessions.encodec_decode,{audio_codes:s});return i}}class Mn{static async from_pretrained(t,{progress_callback:r=null,config:n=null,cache_dir:s=null,local_files_only:i=!1,revision:a="main",model_file_name:o=null,subfolder:l="onnx",device:u=null,dtype:d=null,use_external_data_format:c=null,session_options:p={}}={}){let h={progress_callback:r,config:n,cache_dir:s,local_files_only:i,revision:a,model_file_name:o,subfolder:l,device:u,dtype:d,use_external_data_format:c,session_options:p};if(h.config=await Lp.from_pretrained(t,h),!this.MODEL_CLASS_MAPPINGS)throw new Error("`MODEL_CLASS_MAPPINGS` not implemented for this type of `AutoClass`: "+this.name);for(let f of this.MODEL_CLASS_MAPPINGS){const m=f.get(h.config.model_type);if(m)return await m[1].from_pretrained(t,h)}if(this.BASE_IF_FAIL)return console.warn(`Unknown model class "${h.config.model_type}", attempting to construct from base class.`),await P.from_pretrained(t,h);throw Error(`Unsupported model type: ${h.config.model_type}`)}}T(Mn,"MODEL_CLASS_MAPPINGS",null),T(Mn,"BASE_IF_FAIL",!1);const t$=new Map([["bert",["BertModel",Bw]],["nomic_bert",["NomicBertModel",Uw]],["roformer",["RoFormerModel",qw]],["electra",["ElectraModel",Zw]],["esm",["EsmModel",xy]],["convbert",["ConvBertModel",Hw]],["camembert",["CamembertModel",ny]],["deberta",["DebertaModel",ly]],["deberta-v2",["DebertaV2Model",hy]],["mpnet",["MPNetModel",zy]],["albert",["AlbertModel",Ny]],["distilbert",["DistilBertModel",wy]],["roberta",["RobertaModel",ob]],["xlm",["XLMModel",pb]],["xlm-roberta",["XLMRobertaModel",_b]],["clap",["ClapModel",jv]],["clip",["CLIPModel",Mb]],["clipseg",["CLIPSegModel",Db]],["chinese_clip",["ChineseCLIPModel",Rb]],["siglip",["SiglipModel",Ab]],["mobilebert",["MobileBertModel",Ty]],["squeezebert",["SqueezeBertModel",Ry]],["wav2vec2",["Wav2Vec2Model",pv]],["wav2vec2-bert",["Wav2Vec2BertModel",xv]],["unispeech",["UniSpeechModel",gv]],["unispeech-sat",["UniSpeechSatModel",yv]],["hubert",["HubertModel",Ev]],["wavlm",["WavLMModel",Mv]],["audio-spectrogram-transformer",["ASTModel",$b]],["vits",["VitsModel",Nh]],["detr",["DetrModel",T0]],["table-transformer",["TableTransformerModel",z0]],["vit",["ViTModel",m0]],["mobilevit",["MobileViTModel",y0]],["owlvit",["OwlViTModel",v0]],["owlv2",["Owlv2Model",x0]],["beit",["BeitModel",S0]],["deit",["DeiTModel",O0]],["convnext",["ConvNextModel",X0]],["convnextv2",["ConvNextV2Model",Z0]],["dinov2",["Dinov2Model",ev]],["resnet",["ResNetModel",R0]],["swin",["SwinModel",L0]],["swin2sr",["Swin2SRModel",N0]],["donut-swin",["DonutSwinModel",Q0]],["yolos",["YolosModel",rv]],["dpt",["DPTModel",q0]],["glpn",["GLPNModel",W0]],["hifigan",["SpeechT5HifiGan",Rv]],["efficientnet",["EfficientNetModel",Jv]]]),r$=new Map([["t5",["T5Model",Vy]],["longt5",["LongT5Model",Wy]],["mt5",["MT5Model",Ky]],["bart",["BartModel",Xy]],["mbart",["MBartModel",Jy]],["marian",["MarianModel",lv]],["whisper",["WhisperModel",kb]],["m2m_100",["M2M100Model",dv]],["blenderbot",["BlenderbotModel",nb]],["blenderbot-small",["BlenderbotSmallModel",ib]]]),n$=new Map([["bloom",["BloomModel",u0]],["gpt2",["GPT2Model",Fb]],["gptj",["GPTJModel",jb]],["gpt_bigcode",["GPTBigCodeModel",Hb]],["gpt_neo",["GPTNeoModel",Ub]],["gpt_neox",["GPTNeoXModel",Gb]],["codegen",["CodeGenModel",Qb]],["llama",["LlamaModel",Yb]],["gemma",["GemmaModel",Jb]],["openelm",["OpenELMModel",t0]],["qwen2",["Qwen2Model",n0]],["phi",["PhiModel",i0]],["phi3",["Phi3Model",o0]],["mpt",["MptModel",c0]],["opt",["OPTModel",h0]],["mistral",["MistralModel",Fv]],["starcoder2",["Starcoder2Model",Uv]],["falcon",["FalconModel",Gv]],["stablelm",["StableLmModel",Yv]]]),jh=new Map([["speecht5",["SpeechT5ForSpeechToText",Ov]],["whisper",["WhisperForConditionalGeneration",Sb]]]),s$=new Map([["speecht5",["SpeechT5ForTextToSpeech",Bv]]]),i$=new Map([["vits",["VitsModel",Nh]],["musicgen",["MusicgenForConditionalGeneration",Vh]]]),a$=new Map([["bert",["BertForSequenceClassification",Dw]],["roformer",["RoFormerForSequenceClassification",Vw]],["electra",["ElectraForSequenceClassification",ey]],["esm",["EsmForSequenceClassification",Sy]],["convbert",["ConvBertForSequenceClassification",Qw]],["camembert",["CamembertForSequenceClassification",iy]],["deberta",["DebertaForSequenceClassification",dy]],["deberta-v2",["DebertaV2ForSequenceClassification",my]],["mpnet",["MPNetForSequenceClassification",Py]],["albert",["AlbertForSequenceClassification",Uy]],["distilbert",["DistilBertForSequenceClassification",yy]],["roberta",["RobertaForSequenceClassification",ub]],["xlm",["XLMForSequenceClassification",fb]],["xlm-roberta",["XLMRobertaForSequenceClassification",yb]],["bart",["BartForSequenceClassification",Zy]],["mbart",["MBartForSequenceClassification",tb]],["mobilebert",["MobileBertForSequenceClassification",My]],["squeezebert",["SqueezeBertForSequenceClassification",Ly]]]),o$=new Map([["bert",["BertForTokenClassification",Lw]],["roformer",["RoFormerForTokenClassification",jw]],["electra",["ElectraForTokenClassification",ty]],["esm",["EsmForTokenClassification",Ey]],["convbert",["ConvBertForTokenClassification",Xw]],["camembert",["CamembertForTokenClassification",ay]],["deberta",["DebertaForTokenClassification",cy]],["deberta-v2",["DebertaV2ForTokenClassification",gy]],["mpnet",["MPNetForTokenClassification",Oy]],["distilbert",["DistilBertForTokenClassification",by]],["roberta",["RobertaForTokenClassification",db]],["xlm",["XLMForTokenClassification",mb]],["xlm-roberta",["XLMRobertaForTokenClassification",bb]]]),Wh=new Map([["t5",["T5ForConditionalGeneration",jy]],["longt5",["LongT5ForConditionalGeneration",Hy]],["mt5",["MT5ForConditionalGeneration",Qy]],["bart",["BartForConditionalGeneration",Yy]],["mbart",["MBartForConditionalGeneration",eb]],["marian",["MarianMTModel",uv]],["m2m_100",["M2M100ForConditionalGeneration",cv]],["blenderbot",["BlenderbotForConditionalGeneration",sb]],["blenderbot-small",["BlenderbotSmallForConditionalGeneration",ab]]]),Ci=new Map([["bloom",["BloomForCausalLM",d0]],["gpt2",["GPT2LMHeadModel",Nb]],["gptj",["GPTJForCausalLM",Wb]],["gpt_bigcode",["GPTBigCodeForCausalLM",Kb]],["gpt_neo",["GPTNeoForCausalLM",qb]],["gpt_neox",["GPTNeoXForCausalLM",Vb]],["codegen",["CodeGenForCausalLM",Xb]],["llama",["LlamaForCausalLM",Zb]],["gemma",["GemmaForCausalLM",e0]],["openelm",["OpenELMForCausalLM",r0]],["qwen2",["Qwen2ForCausalLM",s0]],["phi",["PhiForCausalLM",a0]],["phi3",["Phi3ForCausalLM",l0]],["mpt",["MptForCausalLM",p0]],["opt",["OPTForCausalLM",f0]],["mbart",["MBartForCausalLM",rb]],["mistral",["MistralForCausalLM",Nv]],["starcoder2",["Starcoder2ForCausalLM",qv]],["falcon",["FalconForCausalLM",Vv]],["trocr",["TrOCRForCausalLM",Lv]],["stablelm",["StableLmForCausalLM",Zv]]]),l$=new Map([["bert",["BertForMaskedLM",Rw]],["roformer",["RoFormerForMaskedLM",Gw]],["electra",["ElectraForMaskedLM",Jw]],["esm",["EsmForMaskedLM",ky]],["convbert",["ConvBertForMaskedLM",Kw]],["camembert",["CamembertForMaskedLM",sy]],["deberta",["DebertaForMaskedLM",uy]],["deberta-v2",["DebertaV2ForMaskedLM",fy]],["mpnet",["MPNetForMaskedLM",Ay]],["albert",["AlbertForMaskedLM",Gy]],["distilbert",["DistilBertForMaskedLM",$y]],["roberta",["RobertaForMaskedLM",lb]],["xlm",["XLMWithLMHeadModel",hb]],["xlm-roberta",["XLMRobertaForMaskedLM",wb]],["mobilebert",["MobileBertForMaskedLM",Iy]],["squeezebert",["SqueezeBertForMaskedLM",Dy]]]),u$=new Map([["bert",["BertForQuestionAnswering",Fw]],["roformer",["RoFormerForQuestionAnswering",Ww]],["electra",["ElectraForQuestionAnswering",ry]],["convbert",["ConvBertForQuestionAnswering",Yw]],["camembert",["CamembertForQuestionAnswering",oy]],["deberta",["DebertaForQuestionAnswering",py]],["deberta-v2",["DebertaV2ForQuestionAnswering",_y]],["mpnet",["MPNetForQuestionAnswering",By]],["albert",["AlbertForQuestionAnswering",qy]],["distilbert",["DistilBertForQuestionAnswering",vy]],["roberta",["RobertaForQuestionAnswering",cb]],["xlm",["XLMForQuestionAnswering",gb]],["xlm-roberta",["XLMRobertaForQuestionAnswering",vb]],["mobilebert",["MobileBertForQuestionAnswering",Cy]],["squeezebert",["SqueezeBertForQuestionAnswering",Fy]]]),Hh=new Map([["vision-encoder-decoder",["VisionEncoderDecoderModel",Eb]]]),d$=new Map([["llava",["LlavaForConditionalGeneration",Jp]],["moondream1",["Moondream1ForConditionalGeneration",Ib]]]),c$=new Map([["vit",["ViTForImageClassification",g0]],["mobilevit",["MobileViTForImageClassification",b0]],["beit",["BeitForImageClassification",E0]],["deit",["DeiTForImageClassification",B0]],["convnext",["ConvNextForImageClassification",Y0]],["convnextv2",["ConvNextV2ForImageClassification",J0]],["dinov2",["Dinov2ForImageClassification",tv]],["resnet",["ResNetForImageClassification",D0]],["swin",["SwinForImageClassification",F0]],["segformer",["SegformerForImageClassification",Qv]],["efficientnet",["EfficientNetForImageClassification",e$]]]),p$=new Map([["detr",["DetrForObjectDetection",I0]],["table-transformer",["TableTransformerForObjectDetection",A0]],["yolos",["YolosForObjectDetection",nv]]]),h$=new Map([["owlvit",["OwlViTForObjectDetection",$0]],["owlv2",["Owlv2ForObjectDetection",k0]]]),f$=new Map([["detr",["DetrForSegmentation",M0]],["clipseg",["CLIPSegForImageSegmentation",Lb]]]),m$=new Map([["segformer",["SegformerForSemanticSegmentation",Xv]]]),g$=new Map([["sam",["SamModel",av]]]),_$=new Map([["wav2vec2",["Wav2Vec2ForCTC",hv]],["wav2vec2-bert",["Wav2Vec2BertForCTC",kv]],["unispeech",["UniSpeechForCTC",_v]],["unispeech-sat",["UniSpeechSatForCTC",bv]],["wavlm",["WavLMForCTC",Cv]],["hubert",["HubertForCTC",Tv]]]),w$=new Map([["wav2vec2",["Wav2Vec2ForSequenceClassification",fv]],["wav2vec2-bert",["Wav2Vec2BertForSequenceClassification",Sv]],["unispeech",["UniSpeechForSequenceClassification",wv]],["unispeech-sat",["UniSpeechSatForSequenceClassification",vv]],["wavlm",["WavLMForSequenceClassification",zv]],["hubert",["HubertForSequenceClassification",Iv]],["audio-spectrogram-transformer",["ASTForAudioClassification",xb]]]),y$=new Map([["wavlm",["WavLMForXVector",Av]]]),b$=new Map([["unispeech-sat",["UniSpeechSatForAudioFrameClassification",$v]],["wavlm",["WavLMForAudioFrameClassification",Pv]],["wav2vec2",["Wav2Vec2ForAudioFrameClassification",mv]]]),v$=new Map([["vitmatte",["VitMatteForImageMatting",w0]]]),$$=new Map([["swin2sr",["Swin2SRForImageSuperResolution",U0]]]),x$=new Map([["dpt",["DPTForDepthEstimation",G0]],["depth_anything",["DepthAnythingForDepthEstimation",j0]],["glpn",["GLPNForDepthEstimation",H0]]]),k$=new Map([["clip",["CLIPVisionModelWithProjection",zb]],["siglip",["SiglipVisionModel",Ob]]]),Kh=[[t$,W.EncoderOnly],[r$,W.EncoderDecoder],[n$,W.DecoderOnly],[a$,W.EncoderOnly],[o$,W.EncoderOnly],[Wh,W.Seq2Seq],[jh,W.Seq2Seq],[Ci,W.DecoderOnly],[l$,W.EncoderOnly],[u$,W.EncoderOnly],[Hh,W.Vision2Seq],[d$,W.ImageTextToText],[c$,W.EncoderOnly],[f$,W.EncoderOnly],[m$,W.EncoderOnly],[v$,W.EncoderOnly],[$$,W.EncoderOnly],[x$,W.EncoderOnly],[p$,W.EncoderOnly],[h$,W.EncoderOnly],[g$,W.MaskGeneration],[_$,W.EncoderOnly],[w$,W.EncoderOnly],[s$,W.Seq2Seq],[i$,W.EncoderOnly],[y$,W.EncoderOnly],[b$,W.EncoderOnly],[k$,W.EncoderOnly]];for(const[e,t]of Kh)for(const[r,n]of e.values())vn.set(r,t),pr.set(n,r),Up.set(r,n);const S$=[["MusicgenForConditionalGeneration",Vh,W.Musicgen],["CLIPTextModelWithProjection",Cb,W.EncoderOnly],["SiglipTextModel",Pb,W.EncoderOnly],["ClapTextModelWithProjection",Wv,W.EncoderOnly],["ClapAudioModelWithProjection",Hv,W.EncoderOnly]];for(const[e,t,r]of S$)vn.set(e,r),pr.set(t,e),Up.set(e,t);class Qh extends Mn{}T(Qh,"MODEL_CLASS_MAPPINGS",Kh.map(t=>t[0])),T(Qh,"BASE_IF_FAIL",!0);class Xh extends Mn{}T(Xh,"MODEL_CLASS_MAPPINGS",[Ci]);class re extends Le{constructor({logits:t}){super(),this.logits=t}}class E$ extends Le{constructor({logits:t,embeddings:r}){super(),this.logits=t,this.embeddings=r}}class be extends Le{constructor({logits:t}){super(),this.logits=t}}class ve extends Le{constructor({logits:t}){super(),this.logits=t}}class Se extends Le{constructor({start_logits:t,end_logits:r}){super(),this.start_logits=t,this.end_logits=r}}class Ut extends Le{constructor({logits:t}){super(),this.logits=t}}class T$ extends Le{constructor({alphas:t}){super(),this.alphas=t}}class I$ extends Le{constructor({waveform:t,spectrogram:r}){super(),this.waveform=t,this.spectrogram=r}}if(!(typeof self<"u")){if(!oe)throw new Error("Unable to load image processing library.")}class M${put(t){throw Error("Not implemented")}end(){throw Error("Not implemented")}}const Yh=mt.IS_PROCESS_AVAILABLE?e=>process.stdout.write(e):e=>console.log(e);class C$ extends M${constructor(t,{skip_prompt:r=!1,...n}={}){super(),this.tokenizer=t,this.skip_prompt=r,this.decode_kwargs=n,this.token_cache=[],this.print_len=0,this.next_tokens_are_prompt=!0}put(t){if(t.length>1)throw Error("TextStreamer only supports batch size of 1");const r=t[0];if(this.skip_prompt&&this.next_tokens_are_prompt){this.next_tokens_are_prompt=!1;return}this.token_cache=we(this.token_cache,r);const n=this.tokenizer.decode(this.token_cache,this.decode_kwargs);let s;n.endsWith(` +`)?(s=n.slice(this.print_len),this.token_cache=[],this.print_len=0):n.length>0&&Ep(n.charCodeAt(n.length-1))?(s=n.slice(this.print_len),this.print_len+=s.length):(s=n.slice(this.print_len,n.lastIndexOf(" ")+1),this.print_len+=s.length),this.on_finalized_text(s,!1)}end(){let t;this.token_cache.length>0?(t=this.tokenizer.decode(this.token_cache,this.decode_kwargs).slice(this.print_len),this.token_cache=[],this.print_len=0):t="",this.next_tokens_are_prompt=!0,this.on_finalized_text(t,!0)}on_finalized_text(t,r){Yh(t),r&&Yh(` +`)}}Ee.backends.onnx.wasm.wasmPaths="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.19.0-esmtest.20240513-a16cd2bd21/dist/";class z$ extends C${constructor(t,r){super(t,{skip_prompt:!0,skip_special_tokens:!0}),this.cb=r}on_finalized_text(t){this.cb(t)}}class A$ extends yn{constructor(){super(),this.interrupted=!1}interrupt(){this.interrupted=!0}reset(){this.interrupted=!1}_call(t,r){return new Array(t.length).fill(this.interrupted)}}const Cn=new A$;class qt{static async getInstance(t=null){return this.model_id??(this.model_id="ucalyptus/prem-1B-chat-onnx-q4"),this.tokenizer??(this.tokenizer=Rp.from_pretrained(this.model_id,{legacy:!0,progress_callback:t})),this.model??(this.model=Xh.from_pretrained(this.model_id,{dtype:"q4",device:"webgpu",use_external_data_format:!0,progress_callback:t})),Promise.all([this.tokenizer,this.model])}}T(qt,"model_id",null),T(qt,"model",null),T(qt,"tokenizer",null),T(qt,"streamer",null);async function P$(e){const[t,r]=await qt.getInstance(),n=t.apply_chat_template(e,{add_generation_prompt:!0,return_dict:!0});let s,i=0;const a=d=>{s??(s=performance.now());let c;i++>0&&(c=i/(performance.now()-s)*1e3),self.postMessage({status:"update",output:d,tps:c,numTokens:i})},o=new z$(t,a);self.postMessage({status:"start"});const l=await r.generate({...n,max_new_tokens:512,streamer:o,stopping_criteria:Cn}),u=t.batch_decode(l,{skip_special_tokens:!1});self.postMessage({status:"complete",output:u})}async function O$(){self.postMessage({status:"loading",data:"Loading model..."});const[e,t]=await qt.getInstance(n=>{self.postMessage(n)});self.postMessage({status:"loading",data:"Compiling shaders and warming up model..."});const r=e("a");await t.generate({...r,max_new_tokens:1}),self.postMessage({status:"ready"})}self.addEventListener("message",async e=>{const{type:t,data:r}=e.data;switch(t){case"load":O$();break;case"generate":Cn.reset(),P$(r);break;case"interrupt":Cn.interrupt();break;case"reset":Cn.reset();break}})})();