@octomil/browser 1.0.0

package/dist/octomil.min.js

@@ -0,0 +1,2849 @@
+"use strict";var __Octomil_ns=(()=>{var Ur=Object.defineProperty;var yy=Object.getOwnPropertyDescriptor;var wy=Object.getOwnPropertyNames;var _y=Object.prototype.hasOwnProperty;var by=(e,t,r)=>t in e?Ur(e,t,{enumerable:!0,configurable:!0,writable:!0,value:r}):e[t]=r;var It=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,r)=>(typeof require<"u"?require:t)[r]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')});var vy=(e,t)=>()=>(e&&(t=e(e=0)),t);var zo=(e,t)=>{for(var r in t)Ur(e,r,{get:t[r],enumerable:!0})},$y=(e,t,r,i)=>{if(t&&typeof t=="object"||typeof t=="function")for(let n of wy(t))!_y.call(e,n)&&n!==r&&Ur(e,n,{get:()=>t[n],enumerable:!(i=yy(t,n))||i.enumerable});return e};var xy=e=>$y(Ur({},"__esModule",{value:!0}),e);var Q=(e,t,r)=>by(e,typeof t!="symbol"?t+"":t,r);var jf={};zo(jf,{InferenceSession:()=>na,TRACE:()=>xr,TRACE_EVENT_BEGIN:()=>_t,TRACE_EVENT_END:()=>bt,TRACE_FUNC_BEGIN:()=>Je,TRACE_FUNC_END:()=>We,Tensor:()=>Xe,default:()=>H0,env:()=>be,registerBackend:()=>Bt});async function Oo(e={}){var t=e,r=!!globalThis.window,i=!!globalThis.WorkerGlobalScope,n=i&&self.name?.startsWith("em-pthread");t.mountExternalData=(o,p)=>{o.startsWith("./")&&(o=o.substring(2)),(t.Zc||(t.Zc=new Map)).set(o,p)},t.unmountExternalData=()=>{delete t.Zc},globalThis.SharedArrayBuffer??new WebAssembly.Memory({initial:0,maximum:0,ae:!0}).buffer.constructor;let a=o=>async(...p)=>{try{if(t.$c)throw Error("Session already started");let m=t.$c={Nd:p[0],errors:[]},c=await o(...p);if(t.$c!==m)throw Error("Session mismatch");t.gd?.flush();let b=m.errors;if(0<b.length){let S=await Promise.all(b);if(S=S.filter(z=>z),0<S.length)throw Error(S.join(`
+`))}return c}finally{t.$c=null}};t.jsepInit=(o,p)=>{if(o==="webgpu"){[t.gd,t.Dd,t.Hd,t.jd,t.Gd,t.ac,t.Id,t.Kd,t.Ed,t.Fd,t.Jd]=p;let m=t.gd;t.jsepRegisterBuffer=(c,b,S,z)=>m.registerBuffer(c,b,S,z),t.jsepGetBuffer=c=>m.getBuffer(c),t.jsepCreateDownloader=(c,b,S)=>m.createDownloader(c,b,S),t.jsepOnCreateSession=c=>{m.onCreateSession(c)},t.jsepOnReleaseSession=c=>{m.onReleaseSession(c)},t.jsepOnRunStart=c=>m.onRunStart(c),t.Ld=(c,b)=>{m.upload(c,b)}}else if(o==="webnn"){let m=p[0];[t.Zd,t.vd,t.webnnEnsureTensor,t.xd,t.webnnDownloadTensor,t.Yd,t.webnnEnableTraceEvent]=p.slice(1),t.webnnReleaseTensorId=t.vd,t.webnnUploadTensor=t.xd,t.webnnRegisterMLContext=t.Yd,t.webnnOnRunStart=c=>m.onRunStart(c),t.webnnOnRunEnd=m.onRunEnd.bind(m),t.webnnOnReleaseSession=c=>{m.onReleaseSession(c)},t.webnnCreateMLTensorDownloader=(c,b)=>m.createMLTensorDownloader(c,b),t.webnnRegisterMLTensor=(c,b,S,z)=>m.registerMLTensor(c,b,S,z),t.webnnCreateMLContext=c=>m.createMLContext(c),t.webnnRegisterMLConstant=(c,b,S,z,M,W)=>m.registerMLConstant(c,b,S,z,M,t.Zc,W),t.webnnRegisterGraphInput=m.registerGraphInput.bind(m),t.webnnIsGraphInput=m.isGraphInput.bind(m),t.webnnRegisterGraphOutput=m.registerGraphOutput.bind(m),t.webnnIsGraphOutput=m.isGraphOutput.bind(m),t.webnnCreateTemporaryTensor=m.createTemporaryTensor.bind(m),t.webnnIsGraphInputOutputTypeSupported=m.isGraphInputOutputTypeSupported.bind(m)}};let s=()=>{let o=p=>(...m)=>{let c=tt;return m=p(...m),tt!=c?new Promise((b,S)=>{Ci={resolve:b,reject:S}}):m};(()=>{for(let p of["_OrtAppendExecutionProvider","_OrtCreateSession","_OrtRun","_OrtRunWithBinding","_OrtBindInput"])t[p]=o(t[p])})(),a!==void 0&&(t._OrtRun=a(t._OrtRun),t._OrtRunWithBinding=a(t._OrtRunWithBinding)),s=void 0};t.asyncInit=()=>{s?.()};var u,l,d=(o,p)=>{throw p},f=Ye.url,h="";if(r||i){try{h=new URL(".",f).href}catch{}i&&(l=o=>{var p=new XMLHttpRequest;return p.open("GET",o,!1),p.responseType="arraybuffer",p.send(null),new Uint8Array(p.response)}),u=async o=>{if(O(o))return new Promise((m,c)=>{var b=new XMLHttpRequest;b.open("GET",o,!0),b.responseType="arraybuffer",b.onload=()=>{b.status==200||b.status==0&&b.response?m(b.response):c(b.status)},b.onerror=c,b.send(null)});var p=await fetch(o,{credentials:"same-origin"});if(p.ok)return p.arrayBuffer();throw Error(p.status+" : "+p.url)}}var g,w,y,v,T,$,_=console.log.bind(console),I=console.error.bind(console),k=_,E=I,C=!1,O=o=>o.startsWith("file://");function x(){ct.buffer!=Y.buffer&&te()}if(n){let o=function(p){try{var m=p.data,c=m.Uc;if(c==="load"){let b=[];self.onmessage=S=>b.push(S),$=()=>{postMessage({Uc:"loaded"});for(let S of b)o(S);self.onmessage=o};for(let S of m.Ad)t[S]&&!t[S].proxy||(t[S]=(...z)=>{postMessage({Uc:"callHandler",zd:S,args:z})},S=="print"&&(k=t[S]),S=="printErr"&&(E=t[S]));ct=m.Vd,te(),w=m.Wd,Ve(),Pr()}else if(c==="run"){(function(b){var S=(x(),U)[b+52>>>2>>>0];b=(x(),U)[b+56>>>2>>>0],Ps(S,S-b),se(S)})(m.Tc),Mi(m.Tc,0,0,1,0,0),Ua(),ki(m.Tc),L||(Os(),L=!0);try{sm(m.Pd,m.dd)}catch(b){if(b!="unwind")throw b}}else m.target!=="setimmediate"&&(c==="checkMailbox"?L&&Ar():c&&(E(`worker: received unknown command ${c}`),E(m)))}catch(b){throw Rs(),b}};var D=o,L=!1;self.onunhandledrejection=p=>{throw p.reason||p},self.onmessage=o}var Y,q,j,ue,A,U,re,J,H,pe,N,V=!1;function te(){var o=ct.buffer;t.HEAP8=Y=new Int8Array(o),j=new Int16Array(o),t.HEAPU8=q=new Uint8Array(o),ue=new Uint16Array(o),t.HEAP32=A=new Int32Array(o),t.HEAPU32=U=new Uint32Array(o),re=new Float32Array(o),J=new Float64Array(o),H=new BigInt64Array(o),pe=new BigUint64Array(o)}function ce(){V=!0,n?$():st.tb()}function Ie(o){throw E(o="Aborted("+o+")"),C=!0,o=new WebAssembly.RuntimeError(o+". Build with -sASSERTIONS for more info."),T?.(o),o}function Me(){return{a:{ma:Ag,hb:zg,g:om,J:um,f:lm,o:dm,h:pm,ha:cm,b:hm,T:fm,Ia:Fa,n:mm,_:Qa,Ya:Za,Ea:Ya,Ga:Xa,Za:Ja,Wa:es,Pa:ts,Va:rs,ka:is,Fa:ns,Ca:as,Xa:ss,Da:os,cb:gm,ea:wm,xa:_m,va:vm,da:xm,O:Tm,H:Sm,wa:km,Z:Rm,ya:Mm,Sa:Bm,Aa:Dm,Ja:Pm,ta:Um,fa:Wm,Ra:ki,$a:Lm,R:Fm,s:Zm,c:Ti,ib:Ym,y:Xm,M:Jm,D:eg,m:tg,t:ms,jb:rg,I:ig,S:ng,j:ag,v:sg,r:og,l:ug,Ma:lg,Na:dg,Oa:pg,Ka:_s,La:bs,ua:vs,eb:hg,bb:gg,u:yg,aa:wg,ga:_g,ab:fg,V:bg,_a:vg,Ba:$g,F:cg,U:xg,la:Nr,za:Sg,gb:Tg,fb:kg,Ta:Ss,Ua:ks,Ha:_i,$:Is,ja:Es,Qa:Cs,ia:zs,lb:fy,na:oy,mb:hy,oa:sy,G:Yg,d:Bg,q:Rg,w:Og,B:Fg,pb:iy,K:jg,x:Dg,pa:ny,X:uy,ba:ry,nb:cy,ob:py,ra:Xg,qa:ty,qb:Jg,N:Qg,Y:ay,e:Ng,A:Pg,k:Mg,kb:my,p:Wg,z:Lg,C:Ug,E:qg,L:Hg,rb:Zg,Q:ly,ca:Kg,W:dy,sb:Gg,sa:Vg,P:ey,i:Eg,a:ct,db:wi}}}async function Ve(){function o(c,b){var S=st=c.exports;c={};for(let[z,M]of Object.entries(S))typeof M=="function"?(S=qm(M),c[z]=S):c[z]=M;return st=c,st=function(){var z=st,M=G=>ae=>G(ae)>>>0,W=G=>()=>G()>>>0;return(z=Object.assign({},z)).ub=M(z.ub),z.Yb=W(z.Yb),z._b=M(z._b),z.mc=M(z.mc),z.nc=W(z.nc),z.rc=M(z.rc),z}(),Da.push(st.$b),As=(c=st).ub,Os=c.vb,t._OrtInit=c.wb,t._OrtGetLastError=c.xb,t._OrtCreateSessionOptions=c.yb,t._OrtAppendExecutionProvider=c.zb,t._OrtAddFreeDimensionOverride=c.Ab,t._OrtAddSessionConfigEntry=c.Bb,t._OrtReleaseSessionOptions=c.Cb,t._OrtCreateSession=c.Db,t._OrtReleaseSession=c.Eb,t._OrtGetInputOutputCount=c.Fb,t._OrtGetInputOutputMetadata=c.Gb,t._OrtFree=c.Hb,t._OrtCreateTensor=c.Ib,t._OrtGetTensorData=c.Jb,t._OrtReleaseTensor=c.Kb,t._OrtCreateRunOptions=c.Lb,t._OrtAddRunConfigEntry=c.Mb,t._OrtReleaseRunOptions=c.Nb,t._OrtCreateBinding=c.Ob,t._OrtBindInput=c.Pb,t._OrtBindOutput=c.Qb,t._OrtClearBoundOutputs=c.Rb,t._OrtReleaseBinding=c.Sb,t._OrtRunWithBinding=c.Tb,t._OrtRun=c.Ub,t._OrtEndProfiling=c.Vb,t._JsepOutput=c.Wb,t._JsepGetNodeName=c.Xb,Dr=c.Yb,rt=t._free=c.Zb,ir=t._malloc=c._b,Mi=c.bc,Rs=c.cc,Ms=c.dc,Bs=c.ec,Bi=c.fc,Ns=c.gc,Ds=c.hc,le=c.ic,nr=c.jc,Ps=c.kc,se=c.lc,Ni=c.mc,oe=c.nc,Us=c.oc,Di=c.pc,Ws=c.qc,Ls=c.rc,qs=c.sc,Pi=c.tc,Vs=c.uc,Gs=c.vc,Fs=c.wc,Hs=c.xc,Ks=c.yc,js=c.zc,Qs=c.Ac,Zs=c.Bc,Ys=c.Cc,Xs=c.Dc,Js=c.Ec,eo=c.Fc,to=c.Gc,ro=c.Hc,io=c.Ic,no=c.Jc,ao=c.Kc,so=c.Lc,oo=c.Mc,uo=c.Nc,lo=c.Oc,po=c.Pc,co=c.Rc,ho=c.Sc,fo=c.bd,mo=c.cd,go=c.hd,yo=c.kd,wo=c.ld,_o=c.md,bo=c.nd,vo=c.od,$o=c.pd,xo=c.qd,To=c.rd,So=c.wd,ko=c.Rd,Io=c.Sd,Eo=c.Td,Co=c.Ud,w=b,st}var p,m=Me();return t.instantiateWasm?new Promise(c=>{t.instantiateWasm(m,(b,S)=>{c(o(b,S))})}):n?o(new WebAssembly.Instance(w,Me()),w):(N??(N=t.locateFile?t.locateFile?t.locateFile("ort-wasm-simd-threaded.jsep.wasm",h):h+"ort-wasm-simd-threaded.jsep.wasm":new URL("ort-wasm-simd-threaded.jsep.wasm",Ye.url).href),p=await async function(c){var b=N;if(!g&&!O(b))try{var S=fetch(b,{credentials:"same-origin"});return await WebAssembly.instantiateStreaming(S,c)}catch(z){E(`wasm streaming compile failed: ${z}`),E("falling back to ArrayBuffer instantiation")}return async function(z,M){try{var W=await async function(G){if(!g)try{var ae=await u(G);return new Uint8Array(ae)}catch{}if(G==N&&g)G=new Uint8Array(g);else{if(!l)throw"both async and sync fetching of the wasm failed";G=l(G)}return G}(z);return await WebAssembly.instantiate(W,M)}catch(G){E(`failed to asynchronously prepare wasm: ${G}`),Ie(G)}}(b,c)}(m),o(p.instance,p.module))}class nt{constructor(p){Q(this,"name","ExitStatus");this.message=`Program terminated with exit(${p})`,this.status=p}}var Se=o=>{o.terminate(),o.onmessage=()=>{}},$e=[],Re=0,xt=null,kr=o=>{pt.length==0&&(La(),Wa(pt[0]));var p=pt.pop();if(!p)return 6;tr.push(p),Tt[o.Tc]=p,p.Tc=o.Tc;var m={Uc:"run",Pd:o.Od,dd:o.dd,Tc:o.Tc};return p.postMessage(m,o.ud),0},dt=0,ve=(o,p,...m)=>{var c,b=16*m.length,S=oe(),z=Ni(b),M=z>>>3;for(c of m)typeof c=="bigint"?((x(),H)[M++>>>0]=1n,(x(),H)[M++>>>0]=c):((x(),H)[M++>>>0]=0n,(x(),J)[M++>>>0]=c);return o=Ms(o,0,b,z,p),se(S),o};function wi(o){if(n)return ve(0,1,o);if(y=o,!(0<dt)){for(var p of tr)Se(p);for(p of pt)Se(p);pt=[],tr=[],Tt={},C=!0}d(0,new nt(o))}function Na(o){if(n)return ve(1,0,o);_i(o)}var _i=o=>{if(y=o,n)throw Na(o),"unwind";wi(o)},pt=[],tr=[],Da=[],Tt={},Pa=o=>{var p=o.Tc;delete Tt[p],pt.push(o),tr.splice(tr.indexOf(o),1),o.Tc=0,Bs(p)};function Ua(){Da.forEach(o=>o())}var Wa=o=>new Promise(p=>{o.onmessage=b=>{var S=b.data;if(b=S.Uc,S.ad&&S.ad!=Dr()){var z=Tt[S.ad];z?z.postMessage(S,S.ud):E(`Internal error! Worker sent a message "${b}" to target pthread ${S.ad}, but that thread no longer exists!`)}else b==="checkMailbox"?Ar():b==="spawnThread"?kr(S):b==="cleanupThread"?zr(()=>{Pa(Tt[S.Qd])}):b==="loaded"?(o.loaded=!0,p(o)):S.target==="setimmediate"?o.postMessage(S):b==="uncaughtException"?o.onerror(S.error):b==="callHandler"?t[S.zd](...S.args):b&&E(`worker sent an unknown command ${b}`)},o.onerror=b=>{throw E(`worker sent an error! ${b.filename}:${b.lineno}: ${b.message}`),b};var m,c=[];for(m of[])t.propertyIsEnumerable(m)&&c.push(m);o.postMessage({Uc:"load",Ad:c,Vd:ct,Wd:w})});function La(){var o=new Worker((()=>{let p=URL;return Ye.url>"file:"&&Ye.url<"file;"?new p("ort.bundle.min.mjs",Ye.url):new URL(Ye.url)})(),{type:"module",workerData:"em-pthread",name:"em-pthread"});pt.push(o)}var ct,sm=(o,p)=>{dt=0,o=Pi(o,p),0<dt?y=o:Bi(o)},Ir=[],Er=0;function om(o){var p=new bi(o>>>=0);return(x(),Y)[p.Vc+12>>>0]==0&&(qa(p,!0),Er--),Va(p,!1),Ir.push(p),Ls(o)}var Ft=0,um=()=>{le(0,0);var o=Ir.pop();Us(o.ed),Ft=0};function qa(o,p){p=p?1:0,(x(),Y)[o.Vc+12>>>0]=p}function Va(o,p){p=p?1:0,(x(),Y)[o.Vc+13>>>0]=p}class bi{constructor(p){this.ed=p,this.Vc=p-24}}var vi=o=>{var p=Ft;if(!p)return nr(0),0;var m=new bi(p);(x(),U)[m.Vc+16>>>2>>>0]=p;var c=(x(),U)[m.Vc+4>>>2>>>0];if(!c)return nr(0),p;for(var b of o){if(b===0||b===c)break;if(Ws(b,c,m.Vc+16))return nr(b),p}return nr(c),p};function lm(){return vi([])}function dm(o){return vi([o>>>0])}function pm(o,p,m,c){return vi([o>>>0,p>>>0,m>>>0,c>>>0])}var cm=()=>{var o=Ir.pop();o||Ie("no exception to throw");var p=o.ed;throw(x(),Y)[o.Vc+13>>>0]==0&&(Ir.push(o),Va(o,!0),qa(o,!1),Er++),Di(p),Ft=p};function hm(o,p,m){var c=new bi(o>>>=0);throw p>>>=0,m>>>=0,(x(),U)[c.Vc+16>>>2>>>0]=0,(x(),U)[c.Vc+4>>>2>>>0]=p,(x(),U)[c.Vc+8>>>2>>>0]=m,Di(o),Er++,Ft=o}var fm=()=>Er;function Ga(o,p,m,c){return n?ve(2,1,o,p,m,c):Fa(o,p,m,c)}function Fa(o,p,m,c){if(o>>>=0,p>>>=0,m>>>=0,c>>>=0,!globalThis.SharedArrayBuffer)return 6;var b=[];return n&&b.length===0?Ga(o,p,m,c):(o={Od:m,Tc:o,dd:c,ud:b},n?(o.Uc="spawnThread",postMessage(o,b),0):kr(o))}function mm(o){throw Ft||(Ft=o>>>0),Ft}var Ha=globalThis.TextDecoder&&new TextDecoder,Ka=(o,p,m,c)=>{if(m=p+m,c)return m;for(;o[p]&&!(p>=m);)++p;return p},ja=(o,p=0,m,c)=>{if(16<(m=Ka(o,p>>>=0,m,c))-p&&o.buffer&&Ha)return Ha.decode(o.buffer instanceof ArrayBuffer?o.subarray(p,m):o.slice(p,m));for(c="";p<m;){var b=o[p++];if(128&b){var S=63&o[p++];if((224&b)==192)c+=String.fromCharCode((31&b)<<6|S);else{var z=63&o[p++];65536>(b=(240&b)==224?(15&b)<<12|S<<6|z:(7&b)<<18|S<<12|z<<6|63&o[p++])?c+=String.fromCharCode(b):(b-=65536,c+=String.fromCharCode(55296|b>>10,56320|1023&b))}}else c+=String.fromCharCode(b)}return c},ke=(o,p,m)=>(o>>>=0)?ja((x(),q),o,p,m):"";function Qa(o,p,m){return n?ve(3,1,o,p,m):0}function Za(o,p){if(n)return ve(4,1,o,p)}function Ya(o,p){if(n)return ve(5,1,o,p)}function Xa(o,p,m){if(n)return ve(6,1,o,p,m)}function Ja(o,p,m){return n?ve(7,1,o,p,m):0}function es(o,p){if(n)return ve(8,1,o,p)}function ts(o,p,m){if(n)return ve(9,1,o,p,m)}function rs(o,p,m,c){if(n)return ve(10,1,o,p,m,c)}function is(o,p,m,c){if(n)return ve(11,1,o,p,m,c)}function ns(o,p,m,c){if(n)return ve(12,1,o,p,m,c)}function as(o){if(n)return ve(13,1,o)}function ss(o,p){if(n)return ve(14,1,o,p)}function os(o,p,m){if(n)return ve(15,1,o,p,m)}var gm=()=>Ie(""),et=o=>{o>>>=0;for(var p="";;){var m=(x(),q)[o++>>>0];if(!m)return p;p+=String.fromCharCode(m)}},$i={},xi={},ym={},Ht=class extends Error{constructor(o){super(o),this.name="BindingError"}};function at(o,p,m={}){return function(c,b,S={}){var z=b.name;if(!c)throw new Ht(`type "${z}" must have a positive integer typeid pointer`);if(xi.hasOwnProperty(c)){if(S.Bd)return;throw new Ht(`Cannot register type '${z}' twice`)}xi[c]=b,delete ym[c],$i.hasOwnProperty(c)&&(b=$i[c],delete $i[c],b.forEach(M=>M()))}(o,p,m)}var us=(o,p,m)=>{switch(p){case 1:return m?c=>(x(),Y)[c>>>0]:c=>(x(),q)[c>>>0];case 2:return m?c=>(x(),j)[c>>>1>>>0]:c=>(x(),ue)[c>>>1>>>0];case 4:return m?c=>(x(),A)[c>>>2>>>0]:c=>(x(),U)[c>>>2>>>0];case 8:return m?c=>(x(),H)[c>>>3>>>0]:c=>(x(),pe)[c>>>3>>>0];default:throw new TypeError(`invalid integer width (${p}): ${o}`)}};function wm(o,p,m,c,b){o>>>=0,m>>>=0,p=et(p>>>0);let S=z=>z;if(c=c===0n){let z=8*m;S=M=>BigInt.asUintN(z,M),b=S(b)}at(o,{name:p,Qc:S,Xc:(z,M)=>(typeof M=="number"&&(M=BigInt(M)),M),Wc:us(p,m,!c),Yc:null})}function _m(o,p,m,c){at(o>>>=0,{name:p=et(p>>>0),Qc:function(b){return!!b},Xc:function(b,S){return S?m:c},Wc:function(b){return this.Qc((x(),q)[b>>>0])},Yc:null})}var ls=[],St=[0,1,,1,null,1,!0,1,!1,1];function Ti(o){9<(o>>>=0)&&--St[o+1]==0&&(St[o]=void 0,ls.push(o))}var Pe=o=>{if(!o)throw new Ht(`Cannot use deleted val. handle = ${o}`);return St[o]},Ge=o=>{switch(o){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let p=ls.pop()||St.length;return St[p]=o,St[p+1]=1,p}};function Si(o){return this.Qc((x(),U)[o>>>2>>>0])}var bm={name:"emscripten::val",Qc:o=>{var p=Pe(o);return Ti(o),p},Xc:(o,p)=>Ge(p),Wc:Si,Yc:null};function vm(o){return at(o>>>0,bm)}var $m=(o,p)=>{switch(p){case 4:return function(m){return this.Qc((x(),re)[m>>>2>>>0])};case 8:return function(m){return this.Qc((x(),J)[m>>>3>>>0])};default:throw new TypeError(`invalid float width (${p}): ${o}`)}};function xm(o,p,m){m>>>=0,at(o>>>=0,{name:p=et(p>>>0),Qc:c=>c,Xc:(c,b)=>b,Wc:$m(p,m),Yc:null})}function Tm(o,p,m,c,b){o>>>=0,m>>>=0,p=et(p>>>0);let S=M=>M;if(c===0){var z=32-8*m;S=M=>M<<z>>>z,b=S(b)}at(o,{name:p,Qc:S,Xc:(M,W)=>W,Wc:us(p,m,c!==0),Yc:null})}function Sm(o,p,m){function c(S){var z=(x(),U)[S>>>2>>>0];return S=(x(),U)[S+4>>>2>>>0],new b((x(),Y).buffer,S,z)}var b=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][p];at(o>>>=0,{name:m=et(m>>>0),Qc:c,Wc:c},{Bd:!0})}var ht=(o,p,m)=>{var c=(x(),q);if(p>>>=0,0<m){var b=p;m=p+m-1;for(var S=0;S<o.length;++S){var z=o.codePointAt(S);if(127>=z){if(p>=m)break;c[p++>>>0]=z}else if(2047>=z){if(p+1>=m)break;c[p++>>>0]=192|z>>6,c[p++>>>0]=128|63&z}else if(65535>=z){if(p+2>=m)break;c[p++>>>0]=224|z>>12,c[p++>>>0]=128|z>>6&63,c[p++>>>0]=128|63&z}else{if(p+3>=m)break;c[p++>>>0]=240|z>>18,c[p++>>>0]=128|z>>12&63,c[p++>>>0]=128|z>>6&63,c[p++>>>0]=128|63&z,S++}}c[p>>>0]=0,o=p-b}else o=0;return o},Cr=o=>{for(var p=0,m=0;m<o.length;++m){var c=o.charCodeAt(m);127>=c?p++:2047>=c?p+=2:55296<=c&&57343>=c?(p+=4,++m):p+=3}return p};function km(o,p){at(o>>>=0,{name:p=et(p>>>0),Qc(m){var c=(x(),U)[m>>>2>>>0];return c=ke(m+4,c,!0),rt(m),c},Xc(m,c){c instanceof ArrayBuffer&&(c=new Uint8Array(c));var b=typeof c=="string";if(!(b||ArrayBuffer.isView(c)&&c.BYTES_PER_ELEMENT==1))throw new Ht("Cannot pass non-string to std::string");var S=b?Cr(c):c.length,z=ir(4+S+1),M=z+4;return(x(),U)[z>>>2>>>0]=S,b?ht(c,M,S+1):(x(),q).set(c,M>>>0),m!==null&&m.push(rt,z),z},Wc:Si,Yc(m){rt(m)}})}var ds=globalThis.TextDecoder?new TextDecoder("utf-16le"):void 0,Im=(o,p,m)=>{if(o>>>=1,16<(p=Ka((x(),ue),o,p/2,m))-o&&ds)return ds.decode((x(),ue).slice(o,p));for(m="";o<p;++o){var c=(x(),ue)[o>>>0];m+=String.fromCharCode(c)}return m},Em=(o,p,m)=>{if(m??(m=2147483647),2>m)return 0;var c=p;m=(m-=2)<2*o.length?m/2:o.length;for(var b=0;b<m;++b){var S=o.charCodeAt(b);(x(),j)[p>>>1>>>0]=S,p+=2}return(x(),j)[p>>>1>>>0]=0,p-c},Cm=o=>2*o.length,zm=(o,p,m)=>{var c="";o>>>=2;for(var b=0;!(b>=p/4);b++){var S=(x(),U)[o+b>>>0];if(!S&&!m)break;c+=String.fromCodePoint(S)}return c},Am=(o,p,m)=>{if(p>>>=0,m??(m=2147483647),4>m)return 0;var c=p;m=c+m-4;for(var b=0;b<o.length;++b){var S=o.codePointAt(b);if(65535<S&&b++,(x(),A)[p>>>2>>>0]=S,(p+=4)+4>m)break}return(x(),A)[p>>>2>>>0]=0,p-c},Om=o=>{for(var p=0,m=0;m<o.length;++m)65535<o.codePointAt(m)&&m++,p+=4;return p};function Rm(o,p,m){if(o>>>=0,p>>>=0,m=et(m>>>=0),p===2)var c=Im,b=Em,S=Cm;else c=zm,b=Am,S=Om;at(o,{name:m,Qc:z=>{var M=(x(),U)[z>>>2>>>0];return M=c(z+4,M*p,!0),rt(z),M},Xc:(z,M)=>{if(typeof M!="string")throw new Ht(`Cannot pass non-string to C++ string type ${m}`);var W=S(M),G=ir(4+W+p);return(x(),U)[G>>>2>>>0]=W/p,b(M,G+4,W+p),z!==null&&z.push(rt,G),G},Wc:Si,Yc(z){rt(z)}})}function Mm(o,p){at(o>>>=0,{Cd:!0,name:p=et(p>>>0),Qc:()=>{},Xc:()=>{}})}function Bm(o){Mi(o>>>0,!i,1,!r,131072,!1),Ua()}var zr=o=>{if(!C)try{if(o(),!(0<dt))try{n?Dr()&&Bi(y):_i(y)}catch(p){p instanceof nt||p=="unwind"||d(0,p)}}catch(p){p instanceof nt||p=="unwind"||d(0,p)}},Nm=!Atomics.waitAsync||globalThis.navigator?.userAgent&&91>Number((navigator.userAgent.match(/Chrom(e|ium)\/([0-9]+)\./)||[])[2]);function ki(o){o>>>=0,Nm||(Atomics.waitAsync((x(),A),o>>>2,o).value.then(Ar),o+=128,Atomics.store((x(),A),o>>>2,1))}var Ar=()=>zr(()=>{var o=Dr();o&&(ki(o),Ds())});function Dm(o,p){(o>>>=0)==p>>>0?setTimeout(Ar):n?postMessage({ad:o,Uc:"checkMailbox"}):(o=Tt[o])&&o.postMessage({Uc:"checkMailbox"})}var Ii=[];function Pm(o,p,m,c,b){for(p>>>=0,b>>>=0,Ii.length=0,m=b>>>3,c=b+c>>>3;m<c;){var S;S=(x(),H)[m++>>>0]?(x(),H)[m++>>>0]:(x(),J)[m++>>>0],Ii.push(S)}return(p?Ui[p]:Cg[o])(...Ii)}var Um=()=>{dt=0};function Wm(o){o>>>=0,n?postMessage({Uc:"cleanupThread",Qd:o}):Pa(Tt[o])}function Lm(o){}var Or=o=>{try{o()}catch(p){Ie(p)}};function qm(o){var p=(...m)=>{Rr.push(o);try{return o(...m)}finally{C||(Rr.pop(),tt&&ft===1&&Rr.length===0&&(ft=0,dt+=1,Or(Io),typeof Fibers<"u"&&Fibers.ce()))}};return hs.set(o,p),p}var ft=0,tt=null,ps=0,Rr=[],Ei=new Map,cs=new Map,hs=new Map,Vm=0,Ci=null,Gm=[],fs=o=>function(p){if(!C){if(ft===0){var m=!1,c=!1;p((b=0)=>{if(!C&&(ps=b,m=!0,c)){ft=2,Or(()=>Eo(tt)),typeof MainLoop<"u"&&MainLoop.yd&&MainLoop.resume(),b=!1;try{var S=function(){var W=(x(),A)[tt+8>>>2>>>0];return W=cs.get(W),W=hs.get(W),--dt,W()}()}catch(W){S=W,b=!0}var z=!1;if(!tt){var M=Ci;M&&(Ci=null,(b?M.reject:M.resolve)(S),z=!0)}if(b&&!z)throw S}}),c=!0,m||(ft=1,tt=function(){var b=ir(65548),S=b+12;if((x(),U)[b>>>2>>>0]=S,(x(),U)[b+4>>>2>>>0]=S+65536,S=Rr[0],!Ei.has(S)){var z=Vm++;Ei.set(S,z),cs.set(z,S)}return S=Ei.get(S),(x(),A)[b+8>>>2>>>0]=S,b}(),typeof MainLoop<"u"&&MainLoop.yd&&MainLoop.pause(),Or(()=>ko(tt)))}else ft===2?(ft=0,Or(Co),rt(tt),tt=null,Gm.forEach(zr)):Ie(`invalid state: ${ft}`);return ps}}(p=>{o().then(p)});function Fm(o){return o>>>=0,fs(async()=>{var p=await Pe(o);return Ge(p)})}var zi=[],Hm=o=>{var p=zi.length;return zi.push(o),p},Km=(o,p)=>{for(var m=Array(o),c=0;c<o;++c){var b=c,S=(x(),U)[p+4*c>>>2>>>0],z=xi[S];if(z===void 0)throw o=`parameter ${c}`,S=As(S),p=et(S),rt(S),new Ht(`${o} has unknown type ${p}`);m[b]=z}return m},jm=(o,p,m)=>{var c=[];return o=o(c,m),c.length&&((x(),U)[p>>>2>>>0]=Ge(c)),o},Qm={},Mr=o=>{var p=Qm[o];return p===void 0?et(o):p};function Zm(o,p,m){var[c,...b]=Km(o,p>>>0);p=c.Xc.bind(c);var S=b.map(W=>W.Wc.bind(W));o--;var z={toValue:Pe};switch(o=S.map((W,G)=>{var ae=`argFromPtr${G}`;return z[ae]=W,`${ae}(args${G?"+"+8*G:""})`}),m){case 0:var M="toValue(handle)";break;case 2:M="new (toValue(handle))";break;case 3:M="";break;case 1:z.getStringOrSymbol=Mr,M="toValue(handle)[getStringOrSymbol(methodName)]"}return M+=`(${o})`,c.Cd||(z.toReturnWire=p,z.emval_returnValue=jm,M=`return emval_returnValue(toReturnWire, destructorsRef, ${M})`),M=`return function (handle, methodName, destructorsRef, args) {
+  ${M}
+  }`,m=new Function(Object.keys(z),M)(...Object.values(z)),M=`methodCaller<(${b.map(W=>W.name)}) => ${c.name}>`,Hm(Object.defineProperty(m,"name",{value:M}))}function Ym(o,p){return p>>>=0,(o=Pe(o>>>0))==Pe(p)}function Xm(o){return(o>>>=0)?(o=Mr(o),Ge(globalThis[o])):Ge(globalThis)}function Jm(o){return o=Mr(o>>>0),Ge(t[o])}function eg(o,p){return p>>>=0,o=Pe(o>>>0),p=Pe(p),Ge(o[p])}function tg(o){9<(o>>>=0)&&(St[o+1]+=1)}function ms(o,p,m,c,b){return zi[o>>>0](p>>>0,m>>>0,c>>>0,b>>>0)}function rg(o,p,m,c,b){return ms(o>>>0,p>>>0,m>>>0,c>>>0,b>>>0)}function ig(){return Ge([])}function ng(o){o=Pe(o>>>0);for(var p=Array(o.length),m=0;m<o.length;m++)p[m]=o[m];return Ge(p)}function ag(o){return Ge(Mr(o>>>0))}function sg(){return Ge({})}function og(o){for(var p=Pe(o>>>=0);p.length;){var m=p.pop();p.pop()(m)}Ti(o)}function ug(o,p,m){p>>>=0,m>>>=0,o=Pe(o>>>0),p=Pe(p),m=Pe(m),o[p]=m}function lg(o,p){o=-9007199254740992>o||9007199254740992<o?NaN:Number(o),p>>>=0,o=new Date(1e3*o),(x(),A)[p>>>2>>>0]=o.getUTCSeconds(),(x(),A)[p+4>>>2>>>0]=o.getUTCMinutes(),(x(),A)[p+8>>>2>>>0]=o.getUTCHours(),(x(),A)[p+12>>>2>>>0]=o.getUTCDate(),(x(),A)[p+16>>>2>>>0]=o.getUTCMonth(),(x(),A)[p+20>>>2>>>0]=o.getUTCFullYear()-1900,(x(),A)[p+24>>>2>>>0]=o.getUTCDay(),o=(o.getTime()-Date.UTC(o.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,(x(),A)[p+28>>>2>>>0]=o}var gs=o=>o%4==0&&(o%100!=0||o%400==0),ys=[0,31,60,91,121,152,182,213,244,274,305,335],ws=[0,31,59,90,120,151,181,212,243,273,304,334];function dg(o,p){o=-9007199254740992>o||9007199254740992<o?NaN:Number(o),p>>>=0,o=new Date(1e3*o),(x(),A)[p>>>2>>>0]=o.getSeconds(),(x(),A)[p+4>>>2>>>0]=o.getMinutes(),(x(),A)[p+8>>>2>>>0]=o.getHours(),(x(),A)[p+12>>>2>>>0]=o.getDate(),(x(),A)[p+16>>>2>>>0]=o.getMonth(),(x(),A)[p+20>>>2>>>0]=o.getFullYear()-1900,(x(),A)[p+24>>>2>>>0]=o.getDay();var m=(gs(o.getFullYear())?ys:ws)[o.getMonth()]+o.getDate()-1|0;(x(),A)[p+28>>>2>>>0]=m,(x(),A)[p+36>>>2>>>0]=-60*o.getTimezoneOffset(),m=new Date(o.getFullYear(),6,1).getTimezoneOffset();var c=new Date(o.getFullYear(),0,1).getTimezoneOffset();o=0|(m!=c&&o.getTimezoneOffset()==Math.min(c,m)),(x(),A)[p+32>>>2>>>0]=o}function pg(o){o>>>=0;var p=new Date((x(),A)[o+20>>>2>>>0]+1900,(x(),A)[o+16>>>2>>>0],(x(),A)[o+12>>>2>>>0],(x(),A)[o+8>>>2>>>0],(x(),A)[o+4>>>2>>>0],(x(),A)[o>>>2>>>0],0),m=(x(),A)[o+32>>>2>>>0],c=p.getTimezoneOffset(),b=new Date(p.getFullYear(),6,1).getTimezoneOffset(),S=new Date(p.getFullYear(),0,1).getTimezoneOffset(),z=Math.min(S,b);return 0>m?(x(),A)[o+32>>>2>>>0]=+(b!=S&&z==c):0<m!=(z==c)&&(b=Math.max(S,b),p.setTime(p.getTime()+6e4*((0<m?z:b)-c))),(x(),A)[o+24>>>2>>>0]=p.getDay(),m=(gs(p.getFullYear())?ys:ws)[p.getMonth()]+p.getDate()-1|0,(x(),A)[o+28>>>2>>>0]=m,(x(),A)[o>>>2>>>0]=p.getSeconds(),(x(),A)[o+4>>>2>>>0]=p.getMinutes(),(x(),A)[o+8>>>2>>>0]=p.getHours(),(x(),A)[o+12>>>2>>>0]=p.getDate(),(x(),A)[o+16>>>2>>>0]=p.getMonth(),(x(),A)[o+20>>>2>>>0]=p.getYear(),o=p.getTime(),BigInt(isNaN(o)?-1:o/1e3)}function _s(o,p,m,c,b,S,z){return n?ve(16,1,o,p,m,c,b,S,z):-52}function bs(o,p,m,c,b,S){if(n)return ve(17,1,o,p,m,c,b,S)}var rr={},cg=()=>performance.timeOrigin+performance.now();function vs(o,p){if(n)return ve(18,1,o,p);if(rr[o]&&(clearTimeout(rr[o].id),delete rr[o]),!p)return 0;var m=setTimeout(()=>{delete rr[o],zr(()=>Ns(o,performance.timeOrigin+performance.now()))},p);return rr[o]={id:m,be:p},0}function hg(o,p,m,c){o>>>=0,p>>>=0,m>>>=0,c>>>=0;var b=new Date().getFullYear(),S=new Date(b,0,1).getTimezoneOffset();b=new Date(b,6,1).getTimezoneOffset();var z=Math.max(S,b);(x(),U)[o>>>2>>>0]=60*z,(x(),A)[p>>>2>>>0]=+(S!=b),o=(p=M=>{var W=Math.abs(M);return`UTC${0<=M?"-":"+"}${String(Math.floor(W/60)).padStart(2,"0")}${String(W%60).padStart(2,"0")}`})(S),p=p(b),b<S?(ht(o,m,17),ht(p,c,17)):(ht(o,c,17),ht(p,m,17))}var fg=()=>Date.now(),mg=1;function gg(o,p,m){if(m>>>=0,!(0<=o&&3>=o))return 28;if(o===0)o=Date.now();else{if(!mg)return 52;o=performance.timeOrigin+performance.now()}return o=Math.round(1e6*o),(x(),H)[m>>>3>>>0]=BigInt(o),0}var Ai=[],$s=(o,p)=>{Ai.length=0;for(var m;m=(x(),q)[o++>>>0];){var c=m!=105;p+=(c&=m!=112)&&p%8?4:0,Ai.push(m==112?(x(),U)[p>>>2>>>0]:m==106?(x(),H)[p>>>3>>>0]:m==105?(x(),A)[p>>>2>>>0]:(x(),J)[p>>>3>>>0]),p+=c?8:4}return Ai};function yg(o,p,m){return o>>>=0,p=$s(p>>>0,m>>>0),Ui[o](...p)}function wg(o,p,m){return o>>>=0,p=$s(p>>>0,m>>>0),Ui[o](...p)}var _g=()=>{};function bg(o,p){return E(ke(o>>>0,p>>>0))}var vg=()=>{throw dt+=1,"unwind"};function $g(){return 4294901760}var xg=()=>navigator.hardwareConcurrency,kt={},Br=o=>{var p;return(p=/\bwasm-function\[\d+\]:(0x[0-9a-f]+)/.exec(o))?+p[1]:(p=/:(\d+):\d+(?:\)|$)/.exec(o))?2147483648|+p[1]:0},xs=o=>{for(var p of o)(o=Br(p))&&(kt[o]=p)};function Tg(){var o=Error().stack.toString().split(`
+`);return o[0]=="Error"&&o.shift(),xs(o),kt.sd=Br(o[3]),kt.Md=o,kt.sd}function Nr(o){if(!(o=kt[o>>>0]))return 0;var p;if(p=/^\s+at .*\.wasm\.(.*) \(.*\)$/.exec(o))o=p[1];else if(p=/^\s+at (.*) \(.*\)$/.exec(o))o=p[1];else{if(!(p=/^(.+?)@/.exec(o)))return 0;o=p[1]}rt(Nr.td??0),p=Cr(o)+1;var m=ir(p);return m&&ht(o,m,p),Nr.td=m,Nr.td}function Sg(o){o>>>=0;var p=(x(),q).length;if(o<=p||4294901760<o)return!1;for(var m=1;4>=m;m*=2){var c=p*(1+.2/m);c=Math.min(c,o+100663296);e:{c=(Math.min(4294901760,65536*Math.ceil(Math.max(o,c)/65536))-ct.buffer.byteLength+65535)/65536|0;try{ct.grow(c),te();var b=1;break e}catch{}b=void 0}if(b)return!0}return!1}function kg(o,p,m){if(o>>>=0,p>>>=0,kt.sd==o)var c=kt.Md;else(c=Error().stack.toString().split(`
+`))[0]=="Error"&&c.shift(),xs(c);for(var b=3;c[b]&&Br(c[b])!=o;)++b;for(o=0;o<m&&c[o+b];++o)(x(),A)[p+4*o>>>2>>>0]=Br(c[o+b]);return o}var Oi,Ri={},Ts=()=>{if(!Oi){var o,p={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(globalThis.navigator?.language??"C").replace("-","_")+".UTF-8",_:"./this.program"};for(o in Ri)Ri[o]===void 0?delete p[o]:p[o]=Ri[o];var m=[];for(o in p)m.push(`${o}=${p[o]}`);Oi=m}return Oi};function Ss(o,p){if(n)return ve(19,1,o,p);o>>>=0,p>>>=0;var m,c=0,b=0;for(m of Ts()){var S=p+c;(x(),U)[o+b>>>2>>>0]=S,c+=ht(m,S,1/0)+1,b+=4}return 0}function ks(o,p){if(n)return ve(20,1,o,p);o>>>=0,p>>>=0;var m=Ts();for(var c of((x(),U)[o>>>2>>>0]=m.length,o=0,m))o+=Cr(c)+1;return(x(),U)[p>>>2>>>0]=o,0}function Is(o){return n?ve(21,1,o):52}function Es(o,p,m,c){return n?ve(22,1,o,p,m,c):52}function Cs(o,p,m,c){return n?ve(23,1,o,p,m,c):70}var Ig=[null,[],[]];function zs(o,p,m,c){if(n)return ve(24,1,o,p,m,c);p>>>=0,m>>>=0,c>>>=0;for(var b=0,S=0;S<m;S++){var z=(x(),U)[p>>>2>>>0],M=(x(),U)[p+4>>>2>>>0];p+=8;for(var W=0;W<M;W++){var G=o,ae=(x(),q)[z+W>>>0],he=Ig[G];ae===0||ae===10?((G===1?k:E)(ja(he)),he.length=0):he.push(ae)}b+=M}return(x(),U)[c>>>2>>>0]=b,0}function Eg(o){return o>>>0}n||function(){for(var o=t.numThreads-1;o--;)La();$e.push(async()=>{var p=async function(){if(!n)return Promise.all(pt.map(Wa))}();Re++,await p,--Re==0&&xt&&(p=xt,xt=null,p())})}(),n||(ct=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0}),te()),t.wasmBinary&&(g=t.wasmBinary),t.stackSave=()=>oe(),t.stackRestore=o=>se(o),t.stackAlloc=o=>Ni(o),t.setValue=function(o,p,m="i8"){switch(m.endsWith("*")&&(m="*"),m){case"i1":case"i8":(x(),Y)[o>>>0]=p;break;case"i16":(x(),j)[o>>>1>>>0]=p;break;case"i32":(x(),A)[o>>>2>>>0]=p;break;case"i64":(x(),H)[o>>>3>>>0]=BigInt(p);break;case"float":(x(),re)[o>>>2>>>0]=p;break;case"double":(x(),J)[o>>>3>>>0]=p;break;case"*":(x(),U)[o>>>2>>>0]=p;break;default:Ie(`invalid type for setValue: ${m}`)}},t.getValue=function(o,p="i8"){switch(p.endsWith("*")&&(p="*"),p){case"i1":case"i8":return(x(),Y)[o>>>0];case"i16":return(x(),j)[o>>>1>>>0];case"i32":return(x(),A)[o>>>2>>>0];case"i64":return(x(),H)[o>>>3>>>0];case"float":return(x(),re)[o>>>2>>>0];case"double":return(x(),J)[o>>>3>>>0];case"*":return(x(),U)[o>>>2>>>0];default:Ie(`invalid type for getValue: ${p}`)}},t.UTF8ToString=ke,t.stringToUTF8=ht,t.lengthBytesUTF8=Cr;var As,Os,Dr,rt,ir,Mi,Rs,Ms,Bs,Bi,Ns,Ds,le,nr,Ps,se,Ni,oe,Us,Di,Ws,Ls,qs,Pi,Vs,Gs,Fs,Hs,Ks,js,Qs,Zs,Ys,Xs,Js,eo,to,ro,io,no,ao,so,oo,uo,lo,po,co,ho,fo,mo,go,yo,wo,_o,bo,vo,$o,xo,To,So,ko,Io,Eo,Co,st,Cg=[wi,Na,Ga,Qa,Za,Ya,Xa,Ja,es,ts,rs,is,ns,as,ss,os,_s,bs,vs,Ss,ks,Is,Es,Cs,zs],Ui={927820:(o,p,m,c,b)=>{if(t===void 0||!t.Zc)return 1;if((o=ke(Number(o>>>0))).startsWith("./")&&(o=o.substring(2)),!(o=t.Zc.get(o)))return 2;if(p=Number(p>>>0),m=Number(m>>>0),c=Number(c>>>0),p+m>o.byteLength)return 3;try{let S=o.subarray(p,p+m);switch(b){case 0:(x(),q).set(S,c>>>0);break;case 1:t.Xd?t.Xd(c,S):t.Ld(c,S);break;default:return 4}return 0}catch{return 4}},928644:(o,p,m)=>{t.xd(o,(x(),q).subarray(p>>>0,p+m>>>0))},928708:()=>t.Zd(),928750:o=>{t.vd(o)},928787:()=>{t.Ed()},928818:()=>{t.Fd()},928847:()=>{t.Jd()},928872:o=>t.Dd(o),928905:o=>t.Hd(o),928937:(o,p,m)=>{t.jd(Number(o),Number(p),Number(m),!0)},929e3:(o,p,m)=>{t.jd(Number(o),Number(p),Number(m))},929057:()=>typeof wasmOffsetConverter<"u",929114:o=>{t.ac("Abs",o,void 0)},929165:o=>{t.ac("Neg",o,void 0)},929216:o=>{t.ac("Floor",o,void 0)},929269:o=>{t.ac("Ceil",o,void 0)},929321:o=>{t.ac("Reciprocal",o,void 0)},929379:o=>{t.ac("Sqrt",o,void 0)},929431:o=>{t.ac("Exp",o,void 0)},929482:o=>{t.ac("Erf",o,void 0)},929533:o=>{t.ac("Sigmoid",o,void 0)},929588:(o,p,m)=>{t.ac("HardSigmoid",o,{alpha:p,beta:m})},929667:o=>{t.ac("Log",o,void 0)},929718:o=>{t.ac("Sin",o,void 0)},929769:o=>{t.ac("Cos",o,void 0)},929820:o=>{t.ac("Tan",o,void 0)},929871:o=>{t.ac("Asin",o,void 0)},929923:o=>{t.ac("Acos",o,void 0)},929975:o=>{t.ac("Atan",o,void 0)},930027:o=>{t.ac("Sinh",o,void 0)},930079:o=>{t.ac("Cosh",o,void 0)},930131:o=>{t.ac("Asinh",o,void 0)},930184:o=>{t.ac("Acosh",o,void 0)},930237:o=>{t.ac("Atanh",o,void 0)},930290:o=>{t.ac("Tanh",o,void 0)},930342:o=>{t.ac("Not",o,void 0)},930393:(o,p,m)=>{t.ac("Clip",o,{min:p,max:m})},930462:o=>{t.ac("Clip",o,void 0)},930514:(o,p)=>{t.ac("Elu",o,{alpha:p})},930572:o=>{t.ac("Gelu",o,void 0)},930624:o=>{t.ac("Relu",o,void 0)},930676:(o,p)=>{t.ac("LeakyRelu",o,{alpha:p})},930740:(o,p)=>{t.ac("ThresholdedRelu",o,{alpha:p})},930810:(o,p)=>{t.ac("Cast",o,{to:p})},930868:o=>{t.ac("Add",o,void 0)},930919:o=>{t.ac("Sub",o,void 0)},930970:o=>{t.ac("Mul",o,void 0)},931021:o=>{t.ac("Div",o,void 0)},931072:o=>{t.ac("Pow",o,void 0)},931123:o=>{t.ac("Equal",o,void 0)},931176:o=>{t.ac("Greater",o,void 0)},931231:o=>{t.ac("GreaterOrEqual",o,void 0)},931293:o=>{t.ac("Less",o,void 0)},931345:o=>{t.ac("LessOrEqual",o,void 0)},931404:(o,p,m,c,b)=>{t.ac("ReduceMean",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},931579:(o,p,m,c,b)=>{t.ac("ReduceMax",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},931753:(o,p,m,c,b)=>{t.ac("ReduceMin",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},931927:(o,p,m,c,b)=>{t.ac("ReduceProd",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},932102:(o,p,m,c,b)=>{t.ac("ReduceSum",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},932276:(o,p,m,c,b)=>{t.ac("ReduceL1",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},932449:(o,p,m,c,b)=>{t.ac("ReduceL2",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},932622:(o,p,m,c,b)=>{t.ac("ReduceLogSum",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},932799:(o,p,m,c,b)=>{t.ac("ReduceSumSquare",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},932979:(o,p,m,c,b)=>{t.ac("ReduceLogSumExp",o,{keepDims:!!p,noopWithEmptyAxes:!!m,axes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},933159:o=>{t.ac("Where",o,void 0)},933212:(o,p,m)=>{t.ac("Transpose",o,{perm:p?Array.from((x(),A).subarray(Number(p)>>>0,Number(m)>>>0)):[]})},933336:(o,p,m,c)=>{t.ac("DepthToSpace",o,{blocksize:p,mode:ke(m),format:c?"NHWC":"NCHW"})},933469:(o,p,m,c)=>{t.ac("DepthToSpace",o,{blocksize:p,mode:ke(m),format:c?"NHWC":"NCHW"})},933602:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e,mt)=>{t.ac("ConvTranspose",o,{format:W?"NHWC":"NCHW",autoPad:p,dilations:[m],group:c,kernelShape:[b],pads:[S,z],strides:[M],wIsConst:()=>!!(x(),Y)[G>>>0],outputPadding:ae?Array.from((x(),A).subarray(Number(ae)>>>0,Number(he)>>>0)):[],outputShape:ye?Array.from((x(),A).subarray(Number(ye)>>>0,Number(_e)>>>0)):[],activation:ke(mt)})},934035:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e)=>{t.ac("ConvTranspose",o,{format:M?"NHWC":"NCHW",autoPad:p,dilations:Array.from((x(),A).subarray(Number(m)>>>0,2+(Number(m)>>>0)>>>0)),group:c,kernelShape:Array.from((x(),A).subarray(Number(b)>>>0,2+(Number(b)>>>0)>>>0)),pads:Array.from((x(),A).subarray(Number(S)>>>0,4+(Number(S)>>>0)>>>0)),strides:Array.from((x(),A).subarray(Number(z)>>>0,2+(Number(z)>>>0)>>>0)),wIsConst:()=>!!(x(),Y)[W>>>0],outputPadding:G?Array.from((x(),A).subarray(Number(G)>>>0,Number(ae)>>>0)):[],outputShape:he?Array.from((x(),A).subarray(Number(he)>>>0,Number(ye)>>>0)):[],activation:ke(_e)})},934696:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e,mt)=>{t.ac("ConvTranspose",o,{format:W?"NHWC":"NCHW",autoPad:p,dilations:[m],group:c,kernelShape:[b],pads:[S,z],strides:[M],wIsConst:()=>!!(x(),Y)[G>>>0],outputPadding:ae?Array.from((x(),A).subarray(Number(ae)>>>0,Number(he)>>>0)):[],outputShape:ye?Array.from((x(),A).subarray(Number(ye)>>>0,Number(_e)>>>0)):[],activation:ke(mt)})},935129:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e)=>{t.ac("ConvTranspose",o,{format:M?"NHWC":"NCHW",autoPad:p,dilations:Array.from((x(),A).subarray(Number(m)>>>0,2+(Number(m)>>>0)>>>0)),group:c,kernelShape:Array.from((x(),A).subarray(Number(b)>>>0,2+(Number(b)>>>0)>>>0)),pads:Array.from((x(),A).subarray(Number(S)>>>0,4+(Number(S)>>>0)>>>0)),strides:Array.from((x(),A).subarray(Number(z)>>>0,2+(Number(z)>>>0)>>>0)),wIsConst:()=>!!(x(),Y)[W>>>0],outputPadding:G?Array.from((x(),A).subarray(Number(G)>>>0,Number(ae)>>>0)):[],outputShape:he?Array.from((x(),A).subarray(Number(he)>>>0,Number(ye)>>>0)):[],activation:ke(_e)})},935790:(o,p)=>{t.ac("GlobalAveragePool",o,{format:p?"NHWC":"NCHW"})},935881:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e)=>{t.ac("AveragePool",o,{format:_e?"NHWC":"NCHW",auto_pad:p,ceil_mode:m,count_include_pad:c,storage_order:b,dilations:S?Array.from((x(),A).subarray(Number(S)>>>0,Number(z)>>>0)):[],kernel_shape:M?Array.from((x(),A).subarray(Number(M)>>>0,Number(W)>>>0)):[],pads:G?Array.from((x(),A).subarray(Number(G)>>>0,Number(ae)>>>0)):[],strides:he?Array.from((x(),A).subarray(Number(he)>>>0,Number(ye)>>>0)):[]})},936360:(o,p)=>{t.ac("GlobalAveragePool",o,{format:p?"NHWC":"NCHW"})},936451:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e)=>{t.ac("AveragePool",o,{format:_e?"NHWC":"NCHW",auto_pad:p,ceil_mode:m,count_include_pad:c,storage_order:b,dilations:S?Array.from((x(),A).subarray(Number(S)>>>0,Number(z)>>>0)):[],kernel_shape:M?Array.from((x(),A).subarray(Number(M)>>>0,Number(W)>>>0)):[],pads:G?Array.from((x(),A).subarray(Number(G)>>>0,Number(ae)>>>0)):[],strides:he?Array.from((x(),A).subarray(Number(he)>>>0,Number(ye)>>>0)):[]})},936930:(o,p)=>{t.ac("GlobalMaxPool",o,{format:p?"NHWC":"NCHW"})},937017:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e)=>{t.ac("MaxPool",o,{format:_e?"NHWC":"NCHW",auto_pad:p,ceil_mode:m,count_include_pad:c,storage_order:b,dilations:S?Array.from((x(),A).subarray(Number(S)>>>0,Number(z)>>>0)):[],kernel_shape:M?Array.from((x(),A).subarray(Number(M)>>>0,Number(W)>>>0)):[],pads:G?Array.from((x(),A).subarray(Number(G)>>>0,Number(ae)>>>0)):[],strides:he?Array.from((x(),A).subarray(Number(he)>>>0,Number(ye)>>>0)):[]})},937492:(o,p)=>{t.ac("GlobalMaxPool",o,{format:p?"NHWC":"NCHW"})},937579:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e)=>{t.ac("MaxPool",o,{format:_e?"NHWC":"NCHW",auto_pad:p,ceil_mode:m,count_include_pad:c,storage_order:b,dilations:S?Array.from((x(),A).subarray(Number(S)>>>0,Number(z)>>>0)):[],kernel_shape:M?Array.from((x(),A).subarray(Number(M)>>>0,Number(W)>>>0)):[],pads:G?Array.from((x(),A).subarray(Number(G)>>>0,Number(ae)>>>0)):[],strides:he?Array.from((x(),A).subarray(Number(he)>>>0,Number(ye)>>>0)):[]})},938054:(o,p,m,c,b)=>{t.ac("Gemm",o,{alpha:p,beta:m,transA:c,transB:b})},938158:o=>{t.ac("MatMul",o,void 0)},938212:(o,p,m,c)=>{t.ac("ArgMax",o,{keepDims:!!p,selectLastIndex:!!m,axis:c})},938320:(o,p,m,c)=>{t.ac("ArgMin",o,{keepDims:!!p,selectLastIndex:!!m,axis:c})},938428:(o,p)=>{t.ac("Softmax",o,{axis:p})},938491:(o,p)=>{t.ac("Concat",o,{axis:p})},938551:(o,p,m,c,b)=>{t.ac("Split",o,{axis:p,numOutputs:m,splitSizes:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},938707:o=>{t.ac("Expand",o,void 0)},938761:(o,p)=>{t.ac("Gather",o,{axis:Number(p)})},938832:(o,p)=>{t.ac("GatherElements",o,{axis:Number(p)})},938911:(o,p)=>{t.ac("GatherND",o,{batch_dims:Number(p)})},938990:(o,p,m,c,b,S,z,M,W,G,ae)=>{t.ac("Resize",o,{antialias:p,axes:m?Array.from((x(),A).subarray(Number(m)>>>0,Number(c)>>>0)):[],coordinateTransformMode:ke(b),cubicCoeffA:S,excludeOutside:z,extrapolationValue:M,keepAspectRatioPolicy:ke(W),mode:ke(G),nearestMode:ke(ae)})},939352:(o,p,m,c,b,S,z)=>{t.ac("Slice",o,{starts:p?Array.from((x(),A).subarray(Number(p)>>>0,Number(m)>>>0)):[],ends:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[],axes:S?Array.from((x(),A).subarray(Number(S)>>>0,Number(z)>>>0)):[]})},939616:o=>{t.ac("Tile",o,void 0)},939668:(o,p,m)=>{t.ac("InstanceNormalization",o,{epsilon:p,format:m?"NHWC":"NCHW"})},939782:(o,p,m)=>{t.ac("InstanceNormalization",o,{epsilon:p,format:m?"NHWC":"NCHW"})},939896:o=>{t.ac("Range",o,void 0)},939949:(o,p)=>{t.ac("Einsum",o,{equation:ke(p)})},940030:(o,p,m,c,b)=>{t.ac("Pad",o,{mode:p,value:m,pads:c?Array.from((x(),A).subarray(Number(c)>>>0,Number(b)>>>0)):[]})},940173:(o,p,m,c,b,S)=>{t.ac("BatchNormalization",o,{epsilon:p,momentum:m,spatial:!!b,trainingMode:!!c,format:S?"NHWC":"NCHW"})},940342:(o,p,m,c,b,S)=>{t.ac("BatchNormalization",o,{epsilon:p,momentum:m,spatial:!!b,trainingMode:!!c,format:S?"NHWC":"NCHW"})},940511:(o,p,m)=>{t.ac("CumSum",o,{exclusive:Number(p),reverse:Number(m)})},940608:(o,p,m)=>{t.ac("DequantizeLinear",o,{axis:p,blockSize:m})},940698:(o,p,m,c,b)=>{t.ac("GridSample",o,{align_corners:p,mode:ke(m),padding_mode:ke(c),format:b?"NHWC":"NCHW"})},940868:(o,p,m,c,b)=>{t.ac("GridSample",o,{align_corners:p,mode:ke(m),padding_mode:ke(c),format:b?"NHWC":"NCHW"})},941038:(o,p)=>{t.ac("ScatterND",o,{reduction:ke(p)})},941123:(o,p,m,c,b,S,z,M,W)=>{t.ac("Attention",o,{numHeads:p,isUnidirectional:m,maskFilterValue:c,scale:b,doRotary:S,qkvHiddenSizes:z?Array.from((x(),A).subarray(Number(M)>>>0,Number(M)+z>>>0)):[],pastPresentShareBuffer:!!W})},941395:o=>{t.ac("BiasAdd",o,void 0)},941450:o=>{t.ac("BiasSplitGelu",o,void 0)},941511:o=>{t.ac("FastGelu",o,void 0)},941567:(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e,mt,Wi)=>{t.ac("Conv",o,{format:he?"NHWC":"NCHW",auto_pad:p,dilations:m?Array.from((x(),A).subarray(Number(m)>>>0,Number(c)>>>0)):[],group:b,kernel_shape:S?Array.from((x(),A).subarray(Number(S)>>>0,Number(z)>>>0)):[],pads:M?Array.from((x(),A).subarray(Number(M)>>>0,Number(W)>>>0)):[],strides:G?Array.from((x(),A).subarray(Number(G)>>>0,Number(ae)>>>0)):[],w_is_const:()=>!!(x(),Y)[Number(ye)>>>0],activation:ke(_e),activation_params:mt?Array.from((x(),re).subarray(Number(mt)>>>0,Number(Wi)>>>0)):[]})},942151:o=>{t.ac("Gelu",o,void 0)},942203:(o,p,m,c,b,S,z,M,W)=>{t.ac("GroupQueryAttention",o,{numHeads:p,kvNumHeads:m,scale:c,softcap:b,doRotary:S,rotaryInterleaved:z,smoothSoftmax:M,localWindowSize:W})},942420:(o,p,m,c)=>{t.ac("LayerNormalization",o,{axis:p,epsilon:m,simplified:!!c})},942531:(o,p,m,c)=>{t.ac("LayerNormalization",o,{axis:p,epsilon:m,simplified:!!c})},942642:(o,p,m,c,b,S)=>{t.ac("MatMulNBits",o,{k:p,n:m,accuracyLevel:c,bits:b,blockSize:S})},942769:(o,p,m,c,b,S)=>{t.ac("MultiHeadAttention",o,{numHeads:p,isUnidirectional:m,maskFilterValue:c,scale:b,doRotary:S})},942928:(o,p)=>{t.ac("QuickGelu",o,{alpha:p})},942992:(o,p,m,c,b)=>{t.ac("RotaryEmbedding",o,{interleaved:!!p,numHeads:m,rotaryEmbeddingDim:c,scale:b})},943131:(o,p,m)=>{t.ac("SkipLayerNormalization",o,{epsilon:p,simplified:!!m})},943233:(o,p,m)=>{t.ac("SkipLayerNormalization",o,{epsilon:p,simplified:!!m})},943335:(o,p,m,c)=>{t.ac("GatherBlockQuantized",o,{gatherAxis:p,quantizeAxis:m,blockSize:c})},943456:o=>{t.Id(o)},943490:(o,p)=>t.Kd(Number(o),Number(p),t.$c.Nd,t.$c.errors)};function zg(o,p,m){return fs(async()=>{await t.Gd(Number(o),Number(p),Number(m))})}function Ag(){return typeof wasmOffsetConverter<"u"}function Og(o,p,m,c){var b=oe();try{return Zs(o,p,m,c)}catch(S){if(se(b),S!==S+0)throw S;le(1,0)}}function Rg(o,p,m){var c=oe();try{return Hs(o,p,m)}catch(b){if(se(c),b!==b+0)throw b;le(1,0)}}function Mg(o,p,m){var c=oe();try{qs(o,p,m)}catch(b){if(se(c),b!==b+0)throw b;le(1,0)}}function Bg(o,p){var m=oe();try{return Pi(o,p)}catch(c){if(se(m),c!==c+0)throw c;le(1,0)}}function Ng(o){var p=oe();try{Vs(o)}catch(m){if(se(p),m!==m+0)throw m;le(1,0)}}function Dg(o,p,m,c,b,S,z){var M=oe();try{return js(o,p,m,c,b,S,z)}catch(W){if(se(M),W!==W+0)throw W;le(1,0)}}function Pg(o,p){var m=oe();try{Ys(o,p)}catch(c){if(se(m),c!==c+0)throw c;le(1,0)}}function Ug(o,p,m,c,b,S){var z=oe();try{Gs(o,p,m,c,b,S)}catch(M){if(se(z),M!==M+0)throw M;le(1,0)}}function Wg(o,p,m,c){var b=oe();try{Qs(o,p,m,c)}catch(S){if(se(b),S!==S+0)throw S;le(1,0)}}function Lg(o,p,m,c,b){var S=oe();try{Fs(o,p,m,c,b)}catch(z){if(se(S),z!==z+0)throw z;le(1,0)}}function qg(o,p,m,c,b,S,z){var M=oe();try{Js(o,p,m,c,b,S,z)}catch(W){if(se(M),W!==W+0)throw W;le(1,0)}}function Vg(o,p,m,c,b,S,z){var M=oe();try{eo(o,p,m,c,b,S,z)}catch(W){if(se(M),W!==W+0)throw W;le(1,0)}}function Gg(o,p,m,c,b,S,z,M){var W=oe();try{no(o,p,m,c,b,S,z,M)}catch(G){if(se(W),G!==G+0)throw G;le(1,0)}}function Fg(o,p,m,c,b){var S=oe();try{return Xs(o,p,m,c,b)}catch(z){if(se(S),z!==z+0)throw z;le(1,0)}}function Hg(o,p,m,c,b,S,z,M){var W=oe();try{ao(o,p,m,c,b,S,z,M)}catch(G){if(se(W),G!==G+0)throw G;le(1,0)}}function Kg(o,p,m,c,b,S,z,M,W,G,ae,he){var ye=oe();try{to(o,p,m,c,b,S,z,M,W,G,ae,he)}catch(_e){if(se(ye),_e!==_e+0)throw _e;le(1,0)}}function jg(o,p,m,c,b,S){var z=oe();try{return ro(o,p,m,c,b,S)}catch(M){if(se(z),M!==M+0)throw M;le(1,0)}}function Qg(o,p,m){var c=oe();try{return so(o,p,m)}catch(b){if(se(c),b!==b+0)throw b;return le(1,0),0n}}function Zg(o,p,m,c,b,S,z,M,W){var G=oe();try{Ks(o,p,m,c,b,S,z,M,W)}catch(ae){if(se(G),ae!==ae+0)throw ae;le(1,0)}}function Yg(o){var p=oe();try{return oo(o)}catch(m){if(se(p),m!==m+0)throw m;le(1,0)}}function Xg(o,p,m){var c=oe();try{return uo(o,p,m)}catch(b){if(se(c),b!==b+0)throw b;le(1,0)}}function Jg(o,p){var m=oe();try{return So(o,p)}catch(c){if(se(m),c!==c+0)throw c;return le(1,0),0n}}function ey(o,p,m,c,b){var S=oe();try{lo(o,p,m,c,b)}catch(z){if(se(S),z!==z+0)throw z;le(1,0)}}function ty(o){var p=oe();try{return po(o)}catch(m){if(se(p),m!==m+0)throw m;return le(1,0),0n}}function ry(o,p,m,c,b,S){var z=oe();try{return yo(o,p,m,c,b,S)}catch(M){if(se(z),M!==M+0)throw M;le(1,0)}}function iy(o,p,m,c,b,S){var z=oe();try{return wo(o,p,m,c,b,S)}catch(M){if(se(z),M!==M+0)throw M;le(1,0)}}function ny(o,p,m,c,b,S,z,M){var W=oe();try{return io(o,p,m,c,b,S,z,M)}catch(G){if(se(W),G!==G+0)throw G;le(1,0)}}function ay(o,p,m,c,b){var S=oe();try{return _o(o,p,m,c,b)}catch(z){if(se(S),z!==z+0)throw z;return le(1,0),0n}}function sy(o,p,m,c){var b=oe();try{return bo(o,p,m,c)}catch(S){if(se(b),S!==S+0)throw S;le(1,0)}}function oy(o,p,m,c){var b=oe();try{return vo(o,p,m,c)}catch(S){if(se(b),S!==S+0)throw S;le(1,0)}}function uy(o,p,m,c,b,S,z,M,W,G,ae,he){var ye=oe();try{return $o(o,p,m,c,b,S,z,M,W,G,ae,he)}catch(_e){if(se(ye),_e!==_e+0)throw _e;le(1,0)}}function ly(o,p,m,c,b,S,z,M,W,G,ae){var he=oe();try{mo(o,p,m,c,b,S,z,M,W,G,ae)}catch(ye){if(se(he),ye!==ye+0)throw ye;le(1,0)}}function dy(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e,mt,Wi){var gy=oe();try{go(o,p,m,c,b,S,z,M,W,G,ae,he,ye,_e,mt,Wi)}catch(Li){if(se(gy),Li!==Li+0)throw Li;le(1,0)}}function py(o,p,m,c){var b=oe();try{return xo(o,p,m,c)}catch(S){if(se(b),S!==S+0)throw S;le(1,0)}}function cy(o,p,m,c,b){var S=oe();try{return To(o,p,m,c,b)}catch(z){if(se(S),z!==z+0)throw z;le(1,0)}}function hy(o,p,m){var c=oe();try{return co(o,p,m)}catch(b){if(se(c),b!==b+0)throw b;le(1,0)}}function fy(o,p,m){var c=oe();try{return ho(o,p,m)}catch(b){if(se(c),b!==b+0)throw b;le(1,0)}}function my(o,p,m,c){var b=oe();try{fo(o,p,m,c)}catch(S){if(se(b),S!==S+0)throw S;le(1,0)}}function Pr(){if(0<Re)xt=Pr;else if(n)v?.(t),ce();else{for(var o=$e;0<o.length;)o.shift()(t);0<Re?xt=Pr:(t.calledRun=!0,C||(ce(),v?.(t)))}}return n||(st=await Ve(),Pr()),t.PTR_SIZE=4,V?t:new Promise((o,p)=>{v=o,T=p})}var Ye,ra,Iy,Ey,Cy,zy,P,Yt,Ay,$r,sr,gt,Bt,Ao,dp,pp,Oy,cp,Ry,Vi,Ce,hp,be,My,fp,mp,By,Vr,gp,yp,wp,_p,bp,Ny,Ot,yr,Gi,vp,Dy,$p,xp,Py,Ne,ia,Xe,Tp,xr,Fi,Je,We,_t,bt,Sp,kp,Uy,na,Wy,Ly,qy,Vy,Gy,Ep,Le,aa,Cp,Hi,Ki,zp,Fy,Ap,Op,Ro,Hy,ji,Wn,Mo,Be,Rp,Gr,Bo,No,Qi,Do,Zi,Mp,Yi,Bp,sa,Xi,Fr,or,Ji,Po,Uo,Wo,oa,we,Ut,Ze,ni,ge,ua,Np,Ky,Lo,qo,Vo,ur,Go,Dp,jy,Rt,ut,Mt,pi,ai,la,da,Ln,ee,pa,Pp,Fo,Ho,Ko,jo,ca,Qo,de,lt,Zo,Qt,R,si,Up,Wp,Lp,ie,ha,qp,en,qn,tn,Yo,rn,Xo,nn,an,sn,Jo,Vp,Qy,lr,eu,Gp,Zy,fa,on,Hr,Kr,tu,ru,un,Vn,iu,Fp,Yy,nu,me,Te,Zt,jr,Ee,Ae,Z,xe,Gn,jt,vt,K,dr,B,F,Hp,ma,au,Kp,ne,su,ln,ou,uu,lu,du,De,jp,Qp,$t,pu,cu,hu,fu,mu,gu,yu,wu,_u,bu,Fe,Zp,Yp,Xp,Jp,ec,tc,rc,ic,nc,ac,Xy,He,vu,oi,Fn,Ke,$u,xu,Tu,Su,ku,Iu,Eu,Cu,zu,Au,je,sc,oc,uc,lc,dc,pc,cc,hc,fc,mc,ga,dn,gc,yc,Hn,Jy,Ou,Qr,Ru,Mu,Bu,Tr,Nu,wc,ya,Du,Pu,Uu,_c,e0,Wu,Lu,bc,t0,qu,fe,vc,$c,xc,Tc,Sc,kc,Ic,Ec,Cc,Vu,zc,Ac,Oc,Rc,_r,Mc,ii,Bc,Nc,Dc,Pc,Uc,Wc,Lc,qc,Vc,Gc,Fc,Hc,Kc,jc,Qc,Zc,pn,Yc,Kn,jn,Xc,Jc,eh,Gu,Fu,th,wa,Hu,Ku,rh,r0,ju,Qu,Qe,ih,nh,ah,sh,oh,uh,lh,dh,ph,ch,i0,Zu,Yu,Xu,Ju,hh,fh,n0,Nt,Dt,Pt,_a,Wt,ze,mh,ba,gh,a0,vr,va,$a,el,tl,Qn,cn,rl,Zn,il,ui,xa,nl,yh,s0,al,hn,pr,sl,fn,ol,wh,_h,o0,bh,vh,u0,ul,Zr,ll,Yr,Yn,mn,dl,pl,Xn,l0,$h,d0,cl,hl,fl,gn,xh,ml,yn,gl,Th,p0,yl,Sh,kh,c0,wl,_l,bl,Ih,Eh,h0,Xr,cr,wn,vl,$l,xl,Tl,_n,Sl,Ch,zh,f0,kl,bn,Il,El,Ah,m0,Cl,Oh,g0,zl,Al,Rh,Mh,y0,Ol,Bh,Nh,w0,Rl,Ml,Dh,Ph,_0,Bl,Nl,Uh,Wh,b0,Dl,Pl,Lh,qh,v0,it,ot,Et,Ct,Ul,Wl,Ll,ql,Vl,Gl,Fl,Hl,Vh,Gh,$0,Oe,Kl,Fh,vn,jl,br,Hh,Kh,Ql,Zl,Yl,Xl,Jn,jh,Qh,Zh,Jl,li,Yh,Xh,ed,td,$n,rd,Jh,x0,xn,id,nd,ef,T0,ad,sd,tf,S0,od,rf,k0,ud,ld,dd,nf,af,I0,pd,cd,hd,fd,md,gd,yd,wd,sf,E0,hr,Tn,Sn,kn,In,_d,bd,En,Cn,of,uf,zn,lf,df,An,pf,cf,hf,ff,C0,vd,$d,mf,gf,z0,xd,Td,yf,A0,Sd,kd,wf,_f,O0,Id,Ed,Cd,On,zd,Ad,Od,Rd,Md,Bd,Nd,Dd,Rn,Pd,Ud,Wd,Ld,qd,bf,vf,R0,Vd,Gd,$f,M0,Fd,fr,Hd,Mn,Kd,jd,xf,Tf,B0,Qd,Zd,Sf,kf,N0,Bn,Yd,Xd,Jd,If,D0,ep,tp,Ef,P0,Cf,U0,zf,W0,Af,rp,ip,np,Of,L0,Rf,Jr,ap,Mf,q0,sp,Ta,Sa,yt,op,Nn,di,ka,Ia,Dn,Ea,Ca,za,Nf,wt,Ue,Kt,mr,gr,ei,Pn,ti,zt,At,up,Df,Pf,Uf,Wf,Lf,qf,Vf,Gf,Un,lp,Ff,V0,Hf,ea,ta,Kf,G0,F0,H0,Qf=vy(()=>{Ye={};ra=Object.defineProperty,Iy=Object.getOwnPropertyDescriptor,Ey=Object.getOwnPropertyNames,Cy=Object.prototype.hasOwnProperty,zy=(e=>typeof It<"u"?It:typeof Proxy<"u"?new Proxy(e,{get:(t,r)=>(typeof It<"u"?It:t)[r]}):e)(function(e){if(typeof It<"u")return It.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')}),P=(e,t)=>()=>(e&&(t=e(e=0)),t),Yt=(e,t)=>{for(var r in t)ra(e,r,{get:t[r],enumerable:!0})},Ay=(e,t,r,i)=>{if(t&&typeof t=="object"||typeof t=="function")for(let n of Ey(t))!Cy.call(e,n)&&n!==r&&ra(e,n,{get:()=>t[n],enumerable:!(i=Iy(t,n))||i.enumerable});return e},$r=e=>Ay(ra({},"__esModule",{value:!0}),e),pp=P(()=>{"use strict";sr=new Map,gt=[],Bt=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let i=sr.get(e);if(i===void 0)sr.set(e,{backend:t,priority:r});else{if(i.priority>r)return;if(i.priority===r&&i.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${r}`)}if(r>=0){let n=gt.indexOf(e);n!==-1&&gt.splice(n,1);for(let a=0;a<gt.length;a++)if(sr.get(gt[a]).priority<=r){gt.splice(a,0,e);return}gt.push(e)}return}throw new TypeError("not a valid backend")},Ao=async e=>{let t=sr.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(i){return r||(t.error=`${i}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},dp=async e=>{let t=e.executionProviders||[],r=t.map(l=>typeof l=="string"?l:l.name),i=r.length===0?gt:r,n,a=[],s=new Set;for(let l of i){let d=await Ao(l);typeof d=="string"?a.push({name:l,err:d}):(n||(n=d),n===d&&s.add(l))}if(!n)throw new Error(`no available backend found. ERR: ${a.map(l=>`[${l.name}] ${l.err}`).join(", ")}`);for(let{name:l,err:d}of a)r.includes(l)&&console.warn(`removing requested execution provider "${l}" from session options because it is not available: ${d}`);let u=t.filter(l=>s.has(typeof l=="string"?l:l.name));return[n,new Proxy(e,{get:(l,d)=>d==="executionProviders"?u:Reflect.get(l,d)})]}}),Oy=P(()=>{"use strict";pp()}),Ry=P(()=>{"use strict";cp="1.24.2"}),hp=P(()=>{"use strict";Ry(),Vi="warning",Ce={wasm:{},webgl:{},webgpu:{},versions:{common:cp},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}`);Vi=e}},get logLevel(){return Vi}},Object.defineProperty(Ce,"logLevel",{enumerable:!0})}),My=P(()=>{"use strict";hp(),be=Ce}),By=P(()=>{"use strict";fp=(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 i=r.getContext("2d");if(i!=null){let n,a;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(n=e.dims[2],a=e.dims[3]):(n=e.dims[3],a=e.dims[2]);let s=t?.format!==void 0?t.format:"RGB",u=t?.norm,l,d;u===void 0||u.mean===void 0?l=[255,255,255,255]:typeof u.mean=="number"?l=[u.mean,u.mean,u.mean,u.mean]:(l=[u.mean[0],u.mean[1],u.mean[2],0],u.mean[3]!==void 0&&(l[3]=u.mean[3])),u===void 0||u.bias===void 0?d=[0,0,0,0]:typeof u.bias=="number"?d=[u.bias,u.bias,u.bias,u.bias]:(d=[u.bias[0],u.bias[1],u.bias[2],0],u.bias[3]!==void 0&&(d[3]=u.bias[3]));let f=a*n,h=0,g=f,w=f*2,y=-1;s==="RGBA"?(h=0,g=f,w=f*2,y=f*3):s==="RGB"?(h=0,g=f,w=f*2):s==="RBG"&&(h=0,w=f,g=f*2);for(let v=0;v<a;v++)for(let T=0;T<n;T++){let $=(e.data[h++]-d[0])*l[0],_=(e.data[g++]-d[1])*l[1],I=(e.data[w++]-d[2])*l[2],k=y===-1?255:(e.data[y++]-d[3])*l[3];i.fillStyle="rgba("+$+","+_+","+I+","+k+")",i.fillRect(T,v,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},mp=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),i;if(r!=null){let n,a,s;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(n=e.dims[2],a=e.dims[1],s=e.dims[3]):(n=e.dims[3],a=e.dims[2],s=e.dims[1]);let u=t!==void 0&&t.format!==void 0?t.format:"RGB",l=t?.norm,d,f;l===void 0||l.mean===void 0?d=[255,255,255,255]:typeof l.mean=="number"?d=[l.mean,l.mean,l.mean,l.mean]:(d=[l.mean[0],l.mean[1],l.mean[2],255],l.mean[3]!==void 0&&(d[3]=l.mean[3])),l===void 0||l.bias===void 0?f=[0,0,0,0]:typeof l.bias=="number"?f=[l.bias,l.bias,l.bias,l.bias]:(f=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(f[3]=l.bias[3]));let h=a*n;if(t!==void 0&&(t.format!==void 0&&s===4&&t.format!=="RGBA"||s===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let g=4,w=0,y=1,v=2,T=3,$=0,_=h,I=h*2,k=-1;u==="RGBA"?($=0,_=h,I=h*2,k=h*3):u==="RGB"?($=0,_=h,I=h*2):u==="RBG"&&($=0,I=h,_=h*2),i=r.createImageData(n,a);for(let E=0;E<a*n;w+=g,y+=g,v+=g,T+=g,E++)i.data[w]=(e.data[$++]-f[0])*d[0],i.data[y]=(e.data[_++]-f[1])*d[1],i.data[v]=(e.data[I++]-f[2])*d[2],i.data[T]=k===-1?255:(e.data[k++]-f[3])*d[3]}else throw new Error("Can not access image data");return i}}),Ny=P(()=>{"use strict";ia(),Vr=(e,t)=>{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:i}=t,n=t.norm??{mean:255,bias:0},a,s;typeof n.mean=="number"?a=[n.mean,n.mean,n.mean,n.mean]:a=[n.mean[0],n.mean[1],n.mean[2],n.mean[3]??255],typeof n.bias=="number"?s=[n.bias,n.bias,n.bias,n.bias]:s=[n.bias[0],n.bias[1],n.bias[2],n.bias[3]??0];let u=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",d=r*i,f=l==="RGBA"?new Float32Array(d*4):new Float32Array(d*3),h=4,g=0,w=1,y=2,v=3,T=0,$=d,_=d*2,I=-1;u==="RGB"&&(h=3,g=0,w=1,y=2,v=-1),l==="RGBA"?I=d*3:l==="RBG"?(T=0,_=d,$=d*2):l==="BGR"&&(_=0,$=d,T=d*2);for(let k=0;k<d;k++,g+=h,y+=h,w+=h,v+=h)f[T++]=(e[g]+s[0])/a[0],f[$++]=(e[w]+s[1])/a[1],f[_++]=(e[y]+s[2])/a[2],I!==-1&&v!==-1&&(f[I++]=(e[v]+s[3])/a[3]);return l==="RGBA"?new Ne("float32",f,[1,4,r,i]):new Ne("float32",f,[1,3,r,i])},gp=async(e,t)=>{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,i=typeof ImageData<"u"&&e instanceof ImageData,n=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,a=typeof e=="string",s,u=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")},d=f=>typeof HTMLCanvasElement<"u"&&f instanceof HTMLCanvasElement||f instanceof OffscreenCanvas?f.getContext("2d"):null;if(r){let f=l();f.width=e.width,f.height=e.height;let h=d(f);if(h!=null){let g=e.height,w=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(g=t.resizedHeight,w=t.resizedWidth),t!==void 0){if(u=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");u.tensorFormat="RGBA",u.height=g,u.width=w}else u.tensorFormat="RGBA",u.height=g,u.width=w;h.drawImage(e,0,0),s=h.getImageData(0,0,w,g).data}else throw new Error("Can not access image data")}else if(i){let f,h;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(f=t.resizedHeight,h=t.resizedWidth):(f=e.height,h=e.width),t!==void 0&&(u=t),u.format="RGBA",u.height=f,u.width=h,t!==void 0){let g=l();g.width=h,g.height=f;let w=d(g);if(w!=null)w.putImageData(e,0,0),s=w.getImageData(0,0,h,f).data;else throw new Error("Can not access image data")}else s=e.data}else if(n){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let f=l();f.width=e.width,f.height=e.height;let h=d(f);if(h!=null){let g=e.height,w=e.width;return h.drawImage(e,0,0,w,g),s=h.getImageData(0,0,w,g).data,u.height=g,u.width=w,Vr(s,u)}else throw new Error("Can not access image data")}else{if(a)return new Promise((f,h)=>{let g=l(),w=d(g);if(!e||!w)return h();let y=new Image;y.crossOrigin="Anonymous",y.src=e,y.onload=()=>{g.width=y.width,g.height=y.height,w.drawImage(y,0,0,g.width,g.height);let v=w.getImageData(0,0,g.width,g.height);u.height=g.height,u.width=g.width,f(Vr(v.data,u))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return Vr(s,u);throw new Error("Input data provided is not supported - aborted tensor creation")},yp=(e,t)=>{let{width:r,height:i,download:n,dispose:a}=t,s=[1,i,r,4];return new Ne({location:"texture",type:"float32",texture:e,dims:s,download:n,dispose:a})},wp=(e,t)=>{let{dataType:r,dims:i,download:n,dispose:a}=t;return new Ne({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:i,download:n,dispose:a})},_p=(e,t)=>{let{dataType:r,dims:i,download:n,dispose:a}=t;return new Ne({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:i,download:n,dispose:a})},bp=(e,t,r)=>new Ne({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})}),Dy=P(()=>{"use strict";Ot=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),yr=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),Gi=!1,vp=()=>{if(!Gi){Gi=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=globalThis.Float16Array,i=typeof r<"u"&&r.from;e&&(Ot.set("int64",BigInt64Array),yr.set(BigInt64Array,"int64")),t&&(Ot.set("uint64",BigUint64Array),yr.set(BigUint64Array,"uint64")),i?(Ot.set("float16",r),yr.set(r,"float16")):Ot.set("float16",Uint16Array)}}}),Py=P(()=>{"use strict";ia(),$p=e=>{let t=1;for(let r=0;r<e.length;r++){let i=e[r];if(typeof i!="number"||!Number.isSafeInteger(i))throw new TypeError(`dims[${r}] must be an integer, got: ${i}`);if(i<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${i}`);t*=i}return t},xp=(e,t)=>{switch(e.location){case"cpu":return new Ne(e.type,e.data,t);case"cpu-pinned":return new Ne({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new Ne({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new Ne({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new Ne({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),ia=P(()=>{"use strict";By(),Ny(),Dy(),Py(),Ne=class{constructor(e,t,r){vp();let i,n;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,i=e.type,n=e.dims,e.location){case"cpu-pinned":{let s=Ot.get(i);if(!s)throw new TypeError(`unsupported type "${i}" to create tensor from pinned buffer`);if(!(e.data instanceof s))throw new TypeError(`buffer should be of type ${s.name}`);this.cpuData=e.data;break}case"texture":{if(i!=="float32")throw new TypeError(`unsupported type "${i}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(i!=="float32"&&i!=="float16"&&i!=="int32"&&i!=="int64"&&i!=="uint32"&&i!=="uint8"&&i!=="bool"&&i!=="uint4"&&i!=="int4")throw new TypeError(`unsupported type "${i}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}case"ml-tensor":{if(i!=="float32"&&i!=="float16"&&i!=="int32"&&i!=="int64"&&i!=="uint32"&&i!=="uint64"&&i!=="int8"&&i!=="uint8"&&i!=="bool"&&i!=="uint4"&&i!=="int4")throw new TypeError(`unsupported type "${i}" to create tensor from MLTensor`);this.mlTensorData=e.mlTensor,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let s,u;if(typeof e=="string")if(i=e,u=r,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");s=t}else{let l=Ot.get(e);if(l===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16"&&l===Uint16Array||e==="uint4"||e==="int4")throw new TypeError(`Creating a ${e} tensor from number array is not supported. Please use ${l.name} as data.`);e==="uint64"||e==="int64"?s=l.from(t,BigInt):s=l.from(t)}else if(t instanceof l)s=t;else if(t instanceof Uint8ClampedArray)if(e==="uint8")s=Uint8Array.from(t);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(e==="float16"&&t instanceof Uint16Array&&l!==Uint16Array)s=new globalThis.Float16Array(t.buffer,t.byteOffset,t.length);else throw new TypeError(`A ${i} tensor's data must be type of ${l}`)}else if(u=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")i="string",s=e;else if(l==="boolean")i="bool",s=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${l}.`)}else if(e instanceof Uint8ClampedArray)i="uint8",s=Uint8Array.from(e);else{let l=yr.get(e.constructor);if(l===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);i=l,s=e}if(u===void 0)u=[s.length];else if(!Array.isArray(u))throw new TypeError("A tensor's dims must be a number array");n=u,this.cpuData=s,this.dataLocation="cpu"}let a=$p(n);if(this.cpuData&&a!==this.cpuData.length&&!((i==="uint4"||i==="int4")&&Math.ceil(a/2)===this.cpuData.length))throw new Error(`Tensor's size(${a}) does not match data length(${this.cpuData.length}).`);this.type=i,this.dims=n,this.size=a}static async fromImage(e,t){return gp(e,t)}static fromTexture(e,t){return yp(e,t)}static fromGpuBuffer(e,t){return wp(e,t)}static fromMLTensor(e,t){return _p(e,t)}static fromPinnedBuffer(e,t,r){return bp(e,t,r)}toDataURL(e){return fp(this,e)}toImageData(e){return mp(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}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw new Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(e){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let 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: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return xp(this,e)}}}),Tp=P(()=>{"use strict";ia(),Xe=Ne}),Sp=P(()=>{"use strict";hp(),xr=(e,t)=>{(typeof Ce.trace>"u"?!Ce.wasm.trace:!Ce.trace)||console.timeStamp(`${e}::ORT::${t}`)},Fi=(e,t)=>{let r=new Error().stack?.split(/\r\n|\r|\n/g)||[],i=!1;for(let n=0;n<r.length;n++){if(i&&!r[n].includes("TRACE_FUNC")){let a=`FUNC_${e}::${r[n].trim().split(" ")[1]}`;t&&(a+=`::${t}`),xr("CPU",a);return}r[n].includes("TRACE_FUNC")&&(i=!0)}},Je=e=>{(typeof Ce.trace>"u"?!Ce.wasm.trace:!Ce.trace)||Fi("BEGIN",e)},We=e=>{(typeof Ce.trace>"u"?!Ce.wasm.trace:!Ce.trace)||Fi("END",e)},_t=e=>{(typeof Ce.trace>"u"?!Ce.wasm.trace:!Ce.trace)||console.time(`ORT::${e}`)},bt=e=>{(typeof Ce.trace>"u"?!Ce.wasm.trace:!Ce.trace)||console.timeEnd(`ORT::${e}`)}}),Uy=P(()=>{"use strict";pp(),Tp(),Sp(),kp=class Ip{constructor(t){this.handler=t}async run(t,r,i){Je(),_t("InferenceSession.run");let n={},a={};if(typeof t!="object"||t===null||t instanceof Xe||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let s=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof Xe)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");s=!1;for(let d of r){if(typeof d!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(d)===-1)throw new RangeError(`'fetches' contains invalid output name: ${d}.`);n[d]=null}if(typeof i=="object"&&i!==null)a=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else{let d=!1,f=Object.getOwnPropertyNames(r);for(let h of this.outputNames)if(f.indexOf(h)!==-1){let g=r[h];(g===null||g instanceof Xe)&&(d=!0,s=!1,n[h]=g)}if(d){if(typeof i=="object"&&i!==null)a=i;else if(typeof i<"u")throw new TypeError("'options' must be an object.")}else a=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let d of this.inputNames)if(typeof t[d]>"u")throw new Error(`input '${d}' is missing in 'feeds'.`);if(s)for(let d of this.outputNames)n[d]=null;let u=await this.handler.run(t,n,a),l={};for(let d in u)if(Object.hasOwnProperty.call(u,d)){let f=u[d];f instanceof Xe?l[d]=f:l[d]=new Xe(f.type,f.data,f.dims)}return bt("InferenceSession.run"),We(),l}async release(){return this.handler.dispose()}static async create(t,r,i,n){Je(),_t("InferenceSession.create");let a,s={};if(typeof t=="string"){if(a=t,typeof r=="object"&&r!==null)s=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof Uint8Array){if(a=t,typeof r=="object"&&r!==null)s=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 f=t,h=0,g=t.byteLength;if(typeof r=="object"&&r!==null)s=r;else if(typeof r=="number"){if(h=r,!Number.isSafeInteger(h))throw new RangeError("'byteOffset' must be an integer.");if(h<0||h>=f.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${f.byteLength}).`);if(g=t.byteLength-h,typeof i=="number"){if(g=i,!Number.isSafeInteger(g))throw new RangeError("'byteLength' must be an integer.");if(g<=0||h+g>f.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${f.byteLength-h}].`);if(typeof n=="object"&&n!==null)s=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else if(typeof i<"u")throw new TypeError("'byteLength' must be a number.")}else if(typeof r<"u")throw new TypeError("'options' must be an object.");a=new Uint8Array(f,h,g)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[u,l]=await dp(s),d=await u.createInferenceSessionHandler(a,l);return bt("InferenceSession.create"),We(),new Ip(d)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}get inputMetadata(){return this.handler.inputMetadata}get outputMetadata(){return this.handler.outputMetadata}}}),Wy=P(()=>{"use strict";Uy(),na=kp}),Ly=P(()=>{"use strict"}),qy=P(()=>{"use strict"}),Vy=P(()=>{"use strict"}),Gy=P(()=>{"use strict"}),Ep={};Yt(Ep,{InferenceSession:()=>na,TRACE:()=>xr,TRACE_EVENT_BEGIN:()=>_t,TRACE_EVENT_END:()=>bt,TRACE_FUNC_BEGIN:()=>Je,TRACE_FUNC_END:()=>We,Tensor:()=>Xe,env:()=>be,registerBackend:()=>Bt});Le=P(()=>{"use strict";Oy(),My(),Wy(),Tp(),Ly(),qy(),Sp(),Vy(),Gy()}),aa=P(()=>{"use strict"}),Cp={};Yt(Cp,{default:()=>zp});Fy=P(()=>{"use strict";Nf(),Ut(),sa(),Hi="ort-wasm-proxy-worker",Ki=globalThis.self?.name===Hi,Ki&&(self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":oa(r.wasm).then(()=>{Ta(r).then(()=>{postMessage({type:t})},i=>{postMessage({type:t,err:i})})},i=>{postMessage({type:t,err:i})});break;case"init-ep":{let{epName:i,env:n}=r;Sa(n,i).then(()=>{postMessage({type:t})},a=>{postMessage({type:t,err:a})});break}case"copy-from":{let{buffer:i}=r,n=di(i);postMessage({type:t,out:n});break}case"create":{let{model:i,options:n}=r;ka(i,n).then(a=>{postMessage({type:t,out:a})},a=>{postMessage({type:t,err:a})});break}case"release":Ia(r),postMessage({type:t});break;case"run":{let{sessionId:i,inputIndices:n,inputs:a,outputIndices:s,options:u}=r;Ea(i,n,a,s,new Array(s.length).fill(null),u).then(l=>{l.some(d=>d[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:l},za([...a,...l]))},l=>{postMessage({type:t,err:l})});break}case"end-profiling":Ca(r),postMessage({type:t});break;default:}}catch(i){postMessage({type:t,err:i})}}),zp=Ki?null:e=>new Worker(e??Be,{type:"module",name:Hi})}),Ap={};Yt(Ap,{default:()=>Op});Hy=P(()=>{"use strict";Op=Oo,Ro=globalThis.self?.name?.startsWith("em-pthread"),Ro&&Oo()}),sa=P(()=>{"use strict";aa(),ji=typeof location>"u"?void 0:location.origin,Wn=Ye.url>"file:"&&Ye.url<"file;",Mo=()=>{if(Wn){let e=URL;return new URL(new e("ort.bundle.min.mjs",Ye.url).href,ji).href}return Ye.url},Be=Mo(),Rp=()=>{if(Be&&!Be.startsWith("blob:"))return Be.substring(0,Be.lastIndexOf("/")+1)},Gr=(e,t)=>{try{let r=t??Be;return(r?new URL(e,r):new URL(e)).origin===ji}catch{return!1}},Bo=(e,t)=>{let r=t??Be;try{return(r?new URL(e,r):new URL(e)).href}catch{return}},No=(e,t)=>`${t??"./"}${e}`,Qi=async e=>{let t=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(t)},Do=async e=>(await import(e)).default,Zi=(Fy(),$r(Cp)).default,Mp=async()=>{if(!Be)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(Gr(Be))return[void 0,Zi()];let e=await Qi(Be);return[e,Zi(e)]},Yi=(Hy(),$r(Ap)).default,Bp=async(e,t,r,i)=>{let n=Yi&&!(e||t);if(n)if(Be)n=Gr(Be);else if(i&&!r)n=!0;else throw new Error("cannot determine the script source URL.");if(n)return[void 0,Yi];{let a="ort-wasm-simd-threaded.jsep.mjs",s=e??Bo(a,t),u=r&&s&&!Gr(s,t),l=u?await Qi(s):s??No(a,t);return[u?l:void 0,await Do(l)]}}}),Ut=P(()=>{"use strict";sa(),Fr=!1,or=!1,Ji=!1,Po=()=>{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}},Uo=()=>{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}},Wo=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,19,1,17,0,65,1,253,15,65,2,253,15,65,3,253,15,253,147,2,11]))}catch{return!1}},oa=async e=>{if(Fr)return Promise.resolve();if(or)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Ji)throw new Error("previous call to 'initializeWebAssembly()' failed.");or=!0;let t=e.initTimeout,r=e.numThreads;if(e.simd!==!1){if(e.simd==="relaxed"){if(!Wo())throw new Error("Relaxed WebAssembly SIMD is not supported in the current environment.")}else if(!Uo())throw new Error("WebAssembly SIMD is not supported in the current environment.")}let i=Po();r>1&&!i&&(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 n=e.wasmPaths,a=typeof n=="string"?n:void 0,s=n?.mjs,u=s?.href??s,l=n?.wasm,d=l?.href??l,f=e.wasmBinary,[h,g]=await Bp(u,a,r>1,!!f||!!d),w=!1,y=[];if(t>0&&y.push(new Promise(v=>{setTimeout(()=>{w=!0,v()},t)})),y.push(new Promise((v,T)=>{let $={numThreads:r};if(f)$.wasmBinary=f;else if(d||a)$.locateFile=_=>d??a+_;else if(u&&u.indexOf("blob:")!==0)$.locateFile=_=>new URL(_,u).href;else if(h){let _=Rp();_&&($.locateFile=I=>_+I)}g($).then(_=>{or=!1,Fr=!0,Xi=_,v(),h&&URL.revokeObjectURL(h)},_=>{or=!1,Ji=!0,T(_)})})),await Promise.race(y),w)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},we=()=>{if(Fr&&Xi)return Xi;throw new Error("WebAssembly is not initialized yet.")}}),ua=P(()=>{"use strict";Ut(),Ze=(e,t)=>{let r=we(),i=r.lengthBytesUTF8(e)+1,n=r._malloc(i);return r.stringToUTF8(e,n,i),t.push(n),n},ni=(e,t,r,i)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([n,a])=>{let s=t?t+n:n;if(typeof a=="object")ni(a,s+".",r,i);else if(typeof a=="string"||typeof a=="number")i(s,a.toString());else if(typeof a=="boolean")i(s,a?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof a}`)})},ge=e=>{let t=we(),r=t.stackSave();try{let i=t.PTR_SIZE,n=t.stackAlloc(2*i);t._OrtGetLastError(n,n+i);let a=Number(t.getValue(n,i===4?"i32":"i64")),s=t.getValue(n+i,"*"),u=s?t.UTF8ToString(s):"";throw new Error(`${e} ERROR_CODE: ${a}, ERROR_MESSAGE: ${u}`)}finally{t.stackRestore(r)}}}),Ky=P(()=>{"use strict";Ut(),ua(),Np=e=>{let t=we(),r=0,i=[],n=e||{};try{if(e?.logSeverityLevel===void 0)n.logSeverityLevel=2;else if(typeof e.logSeverityLevel!="number"||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw new Error(`log severity level is not valid: ${e.logSeverityLevel}`);if(e?.logVerbosityLevel===void 0)n.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);e?.terminate===void 0&&(n.terminate=!1);let a=0;return e?.tag!==void 0&&(a=Ze(e.tag,i)),r=t._OrtCreateRunOptions(n.logSeverityLevel,n.logVerbosityLevel,!!n.terminate,a),r===0&&ge("Can't create run options."),e?.extra!==void 0&&ni(e.extra,"",new WeakSet,(s,u)=>{let l=Ze(s,i),d=Ze(u,i);t._OrtAddRunConfigEntry(r,l,d)!==0&&ge(`Can't set a run config entry: ${s} - ${u}.`)}),[r,i]}catch(a){throw r!==0&&t._OrtReleaseRunOptions(r),i.forEach(s=>t._free(s)),a}}}),jy=P(()=>{"use strict";Ut(),ua(),Lo=e=>{switch(e){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"layout":return 3;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${e}`)}},qo=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},Vo=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)},ur=(e,t,r,i)=>{let n=Ze(t,i),a=Ze(r,i);we()._OrtAddSessionConfigEntry(e,n,a)!==0&&ge(`Can't set a session config entry: ${t} - ${r}.`)},Go=async(e,t,r)=>{let i=t.executionProviders;for(let n of i){let a=typeof n=="string"?n:n.name,s=[];switch(a){case"webnn":if(a="WEBNN",typeof n!="string"){let h=n?.deviceType;h&&ur(e,"deviceType",h,r)}break;case"webgpu":if(a="JS",typeof n!="string"){let h=n;if(h?.preferredLayout){if(h.preferredLayout!=="NCHW"&&h.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${h.preferredLayout}`);ur(e,"preferredLayout",h.preferredLayout,r)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${a}`)}let u=Ze(a,r),l=s.length,d=0,f=0;if(l>0){d=we()._malloc(l*we().PTR_SIZE),r.push(d),f=we()._malloc(l*we().PTR_SIZE),r.push(f);for(let h=0;h<l;h++)we().setValue(d+h*we().PTR_SIZE,s[h][0],"*"),we().setValue(f+h*we().PTR_SIZE,s[h][1],"*")}await we()._OrtAppendExecutionProvider(e,u,d,f,l)!==0&&ge(`Can't append execution provider: ${a}.`)}},Dp=async e=>{let t=we(),r=0,i=[],n=e||{};Vo(n);try{let a=Lo(n.graphOptimizationLevel??"all"),s=qo(n.executionMode??"sequential"),u=typeof n.logId=="string"?Ze(n.logId,i):0,l=n.logSeverityLevel??2;if(!Number.isInteger(l)||l<0||l>4)throw new Error(`log severity level is not valid: ${l}`);let d=n.logVerbosityLevel??0;if(!Number.isInteger(d)||d<0||d>4)throw new Error(`log verbosity level is not valid: ${d}`);let f=typeof n.optimizedModelFilePath=="string"?Ze(n.optimizedModelFilePath,i):0;if(r=t._OrtCreateSessionOptions(a,!!n.enableCpuMemArena,!!n.enableMemPattern,s,!!n.enableProfiling,0,u,l,d,f),r===0&&ge("Can't create session options."),n.executionProviders&&await Go(r,n,i),n.enableGraphCapture!==void 0){if(typeof n.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${n.enableGraphCapture}`);ur(r,"enableGraphCapture",n.enableGraphCapture.toString(),i)}if(n.freeDimensionOverrides)for(let[h,g]of Object.entries(n.freeDimensionOverrides)){if(typeof h!="string")throw new Error(`free dimension override name must be a string: ${h}`);if(typeof g!="number"||!Number.isInteger(g)||g<0)throw new Error(`free dimension override value must be a non-negative integer: ${g}`);let w=Ze(h,i);t._OrtAddFreeDimensionOverride(r,w,g)!==0&&ge(`Can't set a free dimension override: ${h} - ${g}.`)}return n.extra!==void 0&&ni(n.extra,"",new WeakSet,(h,g)=>{ur(r,h,g,i)}),[r,i]}catch(a){throw r!==0&&t._OrtReleaseSessionOptions(r)!==0&&ge("Can't release session options."),i.forEach(s=>t._free(s)),a}}}),ee=P(()=>{"use strict";Rt=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw new Error(`unsupported data type: ${e}`)}},ut=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";case 22:return"int4";case 21:return"uint4";default:throw new Error(`unsupported data type: ${e}`)}},Mt=(e,t)=>{let r=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],i=typeof t=="number"?t:t.reduce((n,a)=>n*a,1);return r>0?Math.ceil(i*r):void 0},pi=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}`)}},ai=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}`)}},la=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",da=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Ln=e=>{switch(e){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;case"ml-tensor":return 5;default:throw new Error(`unsupported data location: ${e}`)}}}),Pp=P(()=>{"use strict";aa(),pa=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"),i=r?parseInt(r,10):0;if(i<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let n=t.body.getReader(),a;try{a=new ArrayBuffer(i)}catch(u){if(u instanceof RangeError){let l=Math.ceil(i/65536);a=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw u}let s=0;for(;;){let{done:u,value:l}=await n.read();if(u)break;let d=l.byteLength;new Uint8Array(a,s,d).set(l),s+=d}return new Uint8Array(a,0,i)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),lt=P(()=>{"use strict";ee(),Fo=["V","I","W","E","F"],Ho=(e,t)=>{console.log(`[${Fo[e]},${new Date().toISOString()}]${t}`)},ca=(e,t)=>{Ko=e,jo=t},Qo=(e,t)=>{let r=ai(e),i=ai(Ko);r>=i&&Ho(r,typeof t=="function"?t():t)},de=(...e)=>{jo&&Qo(...e)}}),ie=P(()=>{"use strict";Zo=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},Qt=class{static calcShape(e,t,r=!1){let i=e.length,n=t.length;if(i===0)return t;if(n===0)return e;let a=Math.max(e.length,t.length),s=new Array(a);if(r){if(i<2||n<2)return;let u=Zo.calcMatMulShape([e[i-2],e[i-1]],[t[n-2],t[n-1]]);if(u===void 0)return;[s[a-2],s[a-1]]=u}for(let u=r?3:1;u<=a;u++){let l=i-u<0?1:e[i-u],d=n-u<0?1:t[n-u];if(l!==d&&l>1&&d>1)return;let f=Math.max(l,d);if(l&&d)s[a-u]=Math.max(l,d);else{if(f>1)return;s[a-u]=0}}return s}static isValidBroadcast(e,t){let r=e.length,i=t.length;if(r>i)return!1;for(let n=1;n<=r;n++)if(e[r-n]!==1&&e[r-n]!==t[i-n])return!1;return!0}},R=class ri{static size(t){return ri.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let i=t.length;if(i===0)return[];let n=new Array(i),a=i-1;for(;a>=0;){if(t[a]%r===0){n[a]=t[a]/r;break}if(r%t[a]!==0)throw new Error("cannot convert shape");n[a]=1,r/=t[a],a--}for(a--;a>=0;a--)n[a]=t[a];return n}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 ri.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 ri.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,i){let n=1;for(let a=r;a<i;a++){if(t[a]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");n*=Number(t[a])}return n}static computeStrides(t){let r=t.length;if(r===0)return[];if(r===1)return[1];let i=new Array(r);i[r-1]=1,i[r-2]=t[r-1];for(let n=r-3;n>=0;--n)i[n]=i[n+1]*t[n+1];return i}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(i=>this.normalizeAxis(i,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(i=>t[i]):t.slice().reverse()}static padShape(t,r){let i=t.length;return t.map((n,a)=>n+r[a]+r[a+i])}static areEqual(t,r){return t.length!==r.length?!1:t.every((i,n)=>i===r[n])}},si=class wr{static adjustPoolAttributes(t,r,i,n,a,s){if(!t&&i.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 u=0;u<r.length-2;u++)u>=i.length?i.push(r[u+2]):i[u]=r[u+2];for(let u=0;u<i.length;u++)if(u<n.length){if(n[u]<0)throw new Error("strides should be greater than or equal to 1")}else n.push(1);for(let u=0;u<i.length;u++)if(u<a.length){if(a[u]<0)throw new Error("dilations should be greater than or equal to 1")}else a.push(1);for(let u=0;u<i.length*2;u++)if(u<s.length){if(s[u]<0)throw new Error("pad should be greater than or equal to 1")}else s.push(0);for(let u=0;u<i.length;u++){if(i[u]<=0)throw new Error("kernel shapes need to be greater than 0");if(s[u]>=i[u]||s[u+i.length]>=i[u])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,i,n,a,s,u){if(u){if(a.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(n.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let l=0;l<t.length-2;l++)wr.adjustPadAndReturnShape(t[l+(s?1:2)],r[l],i[l],n[l],a,l,l+t.length-2,u)}}static computePoolOutputShape(t,r,i,n,a,s,u){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let l=[r[0],r[1]];return wr.computeShapeHelper(t,r,l,i,n,a,s,u),l}static computeConvOutputShape(t,r,i,n,a,s,u){if(t.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let l=[t[0],r[0]];return wr.computeShapeHelper(!1,t,l,i,n,a,s,u),l}static computeShapeHelper(t,r,i,n,a,s,u,l){if(t)for(let d=0;d<r.length-2;d++)i.push(1);else for(let d=0;d<r.length-2;d++)i.push(wr.adjustPadAndReturnShape(r[d+2],n[d],a[d],s[d],u,d,d+r.length-2,l))}static adjustPadAndReturnShape(t,r,i,n,a,s,u,l){let d=i*(n-1)+1;if(l&&l!=="NOTSET")switch(l){case"VALID":return a[s]=0,a[u]=0,Math.floor((t-d)/r+1);case"SAME_LOWER":case"SAME_UPPER":if(i!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let f=((t+r-1)/r-1)*r+n-t;return a[s]=Math.floor(l==="SAME_LOWER"?(f+1)/2:f/2),a[u]=f-a[s],Math.floor((t+f-n)/r+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((t+a[s]+a[u]-d)/r+1)}},Up=class{static getShapeOfGemmResult(e,t,r,i,n){if(e.length!==2||r.length!==2)throw new Error("shape need to be of size 2");let a,s,u;t?(a=e[1],s=e[0]):(a=e[0],s=e[1]);let l=-1;if(i?(u=r[0],l=1):(u=r[1],l=0),r[l]!==s)throw new Error("dimension mismatch");if(a<=0||u<=0||s<=0)throw new Error("invalid shape specified");if(n&&!Qt.isValidBroadcast(n,[a,u]))throw new Error("gemm: invalid bias shape for broadcast");return[a,u,s]}},Wp=-34028234663852886e22,Lp=34028234663852886e22}),qp=P(()=>{"use strict";ee(),ha=(e,t)=>new(pi(t))(e)}),Qy=P(()=>{"use strict";ee(),lt(),en=new Map([["float32",32],["float16",16],["int32",32],["uint32",32],["int64",64],["uint64",64],["int8",8],["uint8",8],["int4",4],["uint4",4]]),qn=(e,t)=>{if(t==="int32")return e;let r=en.get(t);if(!r)throw new Error(`WebNN backend does not support data type: ${t}`);let i=r/8;if(e.byteLength%i!==0)throw new Error(`Invalid Uint8Array length - must be a multiple of ${i}.`);let n=e.byteLength/i,a=new(pi(t))(e.buffer,e.byteOffset,n);switch(t){case"int64":case"uint64":{let s=new Int32Array(n);for(let u=0;u<n;u++){let l=a[u];if(l>2147483647n||l<-2147483648n)throw new Error("Can not convert int64 data to int32 - value out of range.");s[u]=Number(l)}return new Uint8Array(s.buffer)}case"int8":case"uint8":case"uint32":{if(t==="uint32"&&a.some(u=>u>2147483647))throw new Error("Can not convert uint32 data to int32 - value out of range.");let s=Int32Array.from(a,Number);return new Uint8Array(s.buffer)}default:throw new Error(`Unsupported data conversion from ${t} to 'int32'`)}},tn=(e,t)=>{if(t==="int32")return e;if(e.byteLength%4!==0)throw new Error("Invalid Uint8Array length - must be a multiple of 4 (int32).");let r=e.byteLength/4,i=new Int32Array(e.buffer,e.byteOffset,r);switch(t){case"int64":{let n=BigInt64Array.from(i,BigInt);return new Uint8Array(n.buffer)}case"uint64":{if(i.some(a=>a<0))throw new Error("Can not convert int32 data to uin64 - negative value found.");let n=BigUint64Array.from(i,BigInt);return new Uint8Array(n.buffer)}case"int8":{if(i.some(a=>a<-128||a>127))throw new Error("Can not convert int32 data to int8 - value out of range.");let n=Int8Array.from(i,Number);return new Uint8Array(n.buffer)}case"uint8":{if(i.some(n=>n<0||n>255))throw new Error("Can not convert int32 data to uint8 - value out of range.");return Uint8Array.from(i,Number)}case"uint32":{if(i.some(a=>a<0))throw new Error("Can not convert int32 data to uint32 - negative value found.");let n=Uint32Array.from(i,Number);return new Uint8Array(n.buffer)}default:throw new Error(`Unsupported data conversion from 'int32' to ${t}`)}},Yo=1,rn=()=>Yo++,Xo=new Map([["int8","int32"],["uint8","int32"],["uint32","int32"],["int64","int32"]]),nn=(e,t)=>{let r=en.get(e);if(!r)throw new Error(`WebNN backend does not support data type: ${e}`);return t.length>0?Math.ceil(t.reduce((i,n)=>i*n)*r/8):0},an=class{constructor(e){this.isDataConverted=!1;let{sessionId:t,context:r,tensor:i,dataType:n,shape:a,fallbackDataType:s}=e;this.sessionId=t,this.mlContext=r,this.mlTensor=i,this.dataType=n,this.tensorShape=a,this.fallbackDataType=s}get tensor(){return this.mlTensor}get type(){return this.dataType}get fallbackType(){return this.fallbackDataType}get shape(){return this.tensorShape}get byteLength(){return nn(this.dataType,this.tensorShape)}destroy(){de("verbose",()=>"[WebNN] TensorWrapper.destroy"),this.mlTensor.destroy()}write(e){this.mlContext.writeTensor(this.mlTensor,e)}async read(e){if(this.fallbackDataType){let t=await this.mlContext.readTensor(this.mlTensor),r=tn(new Uint8Array(t),this.dataType);if(e){(e instanceof ArrayBuffer?new Uint8Array(e):new Uint8Array(e.buffer,e.byteOffset,e.byteLength)).set(r);return}else return r.buffer}else return e?this.mlContext.readTensor(this.mlTensor,e):this.mlContext.readTensor(this.mlTensor)}canReuseTensor(e,t,r){return this.mlContext===e&&this.dataType===t&&this.tensorShape.length===r.length&&this.tensorShape.every((i,n)=>i===r[n])}setIsDataConverted(e){this.isDataConverted=e}},sn=class{constructor(e,t){this.tensorManager=e,this.wrapper=t}get tensorWrapper(){return this.wrapper}releaseTensor(){this.tensorWrapper&&(this.tensorManager.releaseTensor(this.tensorWrapper),this.wrapper=void 0)}async ensureTensor(e,t,r,i){let n=this.tensorManager.getMLContext(e),a=this.tensorManager.getMLOpSupportLimits(e),s;if(!a?.input.dataTypes.includes(t)){if(s=Xo.get(t),!s||a?.input.dataTypes.includes(s))throw new Error(`WebNN backend does not support data type: ${t}`);de("verbose",()=>`[WebNN] TensorIdTracker.ensureTensor: fallback dataType from ${t} to ${s}`)}if(this.wrapper){if(this.wrapper.canReuseTensor(n,t,r))return this.wrapper.tensor;if(i){if(this.wrapper.byteLength!==nn(t,r))throw new Error("Unable to copy data to tensor with different size.");this.activeUpload=new Uint8Array(await this.wrapper.read())}this.tensorManager.releaseTensor(this.wrapper)}let u=typeof MLTensorUsage>"u"?void 0:MLTensorUsage.READ|MLTensorUsage.WRITE;return this.wrapper=await this.tensorManager.getCachedTensor(e,t,r,u,!0,!0,s),i&&this.activeUpload&&(this.wrapper.write(this.activeUpload),this.activeUpload=void 0),this.wrapper.tensor}upload(e){let t=e;if(this.wrapper){if(this.wrapper.fallbackType)if(this.wrapper.fallbackType==="int32")t=qn(e,this.wrapper.type),this.wrapper.setIsDataConverted(!0);else throw new Error(`Unsupported fallback data type: ${this.wrapper.fallbackType}`);if(e.byteLength===this.wrapper.byteLength){this.wrapper.write(t);return}else de("verbose",()=>"Data size does not match tensor size. Releasing tensor."),this.releaseTensor()}this.activeUpload?this.activeUpload.set(t):this.activeUpload=new Uint8Array(t)}async download(e){if(this.activeUpload){let t=this.wrapper?.isDataConverted?tn(this.activeUpload,this.wrapper?.type):this.activeUpload;if(e){e instanceof ArrayBuffer?new Uint8Array(e).set(t):new Uint8Array(e.buffer,e.byteOffset,e.byteLength).set(t);return}else return t.buffer}if(!this.wrapper)throw new Error("Tensor has not been created.");return e?this.wrapper.read(e):this.wrapper.read()}},Jo=class{constructor(e){this.backend=e,this.tensorTrackersById=new Map,this.freeTensors=[],this.externalTensors=new Set}getMLContext(e){let t=this.backend.getMLContext(e);if(!t)throw new Error("MLContext not found for session.");return t}getMLOpSupportLimits(e){return this.backend.getMLOpSupportLimits(e)}reserveTensorId(){let e=rn();return this.tensorTrackersById.set(e,new sn(this)),e}releaseTensorId(e){let t=this.tensorTrackersById.get(e);t&&(this.tensorTrackersById.delete(e),t.tensorWrapper&&this.releaseTensor(t.tensorWrapper))}async ensureTensor(e,t,r,i,n){de("verbose",()=>`[WebNN] TensorManager.ensureTensor {tensorId: ${t}, dataType: ${r}, shape: ${i}, copyOld: ${n}}`);let a=this.tensorTrackersById.get(t);if(!a)throw new Error("Tensor not found.");return a.ensureTensor(e,r,i,n)}upload(e,t){let r=this.tensorTrackersById.get(e);if(!r)throw new Error("Tensor not found.");r.upload(t)}async download(e,t){de("verbose",()=>`[WebNN] TensorManager.download {tensorId: ${e}, dstBuffer: ${t?.byteLength}}`);let r=this.tensorTrackersById.get(e);if(!r)throw new Error("Tensor not found.");return r.download(t)}releaseTensorsForSession(e){for(let t of this.freeTensors)t.sessionId===e&&t.destroy();this.freeTensors=this.freeTensors.filter(t=>t.sessionId!==e)}registerTensor(e,t,r,i){let n=this.getMLContext(e),a=rn(),s=new an({sessionId:e,context:n,tensor:t,dataType:r,shape:i});return this.tensorTrackersById.set(a,new sn(this,s)),this.externalTensors.add(s),a}async getCachedTensor(e,t,r,i,n,a,s){let u=this.getMLContext(e);for(let[d,f]of this.freeTensors.entries())if(f.canReuseTensor(u,t,r)){de("verbose",()=>`[WebNN] Reusing tensor {dataType: ${t}, ${s?`fallbackDataType: ${s},`:""} shape: ${r}`);let h=this.freeTensors.splice(d,1)[0];return h.sessionId=e,h}de("verbose",()=>`[WebNN] MLContext.createTensor {dataType: ${t}, ${s?`fallbackDataType: ${s},`:""} shape: ${r}}`);let l=await u.createTensor({dataType:s??t,shape:r,dimensions:r,usage:i,writable:n,readable:a});return new an({sessionId:e,context:u,tensor:l,dataType:t,shape:r,fallbackDataType:s})}releaseTensor(e){this.externalTensors.has(e)&&this.externalTensors.delete(e),this.freeTensors.push(e)}},Vp=(...e)=>new Jo(...e)}),Zy=P(()=>{"use strict";ee(),Ut(),qp(),Qy(),lt(),lr=new Map([[1,"float32"],[10,"float16"],[6,"int32"],[12,"uint32"],[7,"int64"],[13,"uint64"],[22,"int4"],[21,"uint4"],[3,"int8"],[2,"uint8"],[9,"uint8"]]),eu=(e,t)=>{if(e===t)return!0;if(e===void 0||t===void 0)return!1;let r=Object.keys(e).sort(),i=Object.keys(t).sort();return r.length===i.length&&r.every((n,a)=>n===i[a]&&e[n]===t[n])},Gp=class{constructor(e){this.tensorManager=Vp(this),this.mlContextBySessionId=new Map,this.sessionIdsByMLContext=new Map,this.mlContextCache=[],this.sessionGraphInputs=new Map,this.sessionGraphOutputs=new Map,this.temporaryGraphInputs=[],this.temporaryGraphOutputs=[],this.temporarySessionTensorIds=new Map,this.mlOpSupportLimitsBySessionId=new Map,ca(e.logLevel,!!e.debug)}get currentSessionId(){if(this.activeSessionId===void 0)throw new Error("No active session");return this.activeSessionId}onRunStart(e){de("verbose",()=>`[WebNN] onRunStart {sessionId: ${e}}`),this.activeSessionId=e}onRunEnd(e){de("verbose",()=>`[WebNN] onRunEnd {sessionId: ${e}}`);let t=this.temporarySessionTensorIds.get(e);if(t){for(let r of t)de("verbose",()=>`[WebNN] releasing temporary tensor {tensorId: ${r}}`),this.tensorManager.releaseTensorId(r);this.temporarySessionTensorIds.delete(e),this.activeSessionId=void 0}}async createMLContext(e){if(e instanceof GPUDevice){let r=this.mlContextCache.findIndex(i=>i.gpuDevice===e);if(r!==-1)return this.mlContextCache[r].mlContext;{let i=await navigator.ml.createContext(e);return this.mlContextCache.push({gpuDevice:e,mlContext:i}),i}}else if(e===void 0){let r=this.mlContextCache.findIndex(i=>i.options===void 0&&i.gpuDevice===void 0);if(r!==-1)return this.mlContextCache[r].mlContext;{let i=await navigator.ml.createContext();return this.mlContextCache.push({mlContext:i}),i}}let t=this.mlContextCache.findIndex(r=>eu(r.options,e));if(t!==-1)return this.mlContextCache[t].mlContext;{let r=await navigator.ml.createContext(e);return this.mlContextCache.push({options:e,mlContext:r}),r}}registerMLContext(e,t){this.mlContextBySessionId.set(e,t);let r=this.sessionIdsByMLContext.get(t);r||(r=new Set,this.sessionIdsByMLContext.set(t,r)),r.add(e),this.mlOpSupportLimitsBySessionId.has(e)||this.mlOpSupportLimitsBySessionId.set(e,t.opSupportLimits()),this.temporaryGraphInputs.length>0&&(this.sessionGraphInputs.set(e,this.temporaryGraphInputs),this.temporaryGraphInputs=[]),this.temporaryGraphOutputs.length>0&&(this.sessionGraphOutputs.set(e,this.temporaryGraphOutputs),this.temporaryGraphOutputs=[])}onReleaseSession(e){this.sessionGraphInputs.delete(e),this.sessionGraphOutputs.delete(e);let t=this.mlContextBySessionId.get(e);if(!t)return;this.tensorManager.releaseTensorsForSession(e),this.mlContextBySessionId.delete(e),this.mlOpSupportLimitsBySessionId.delete(e);let r=this.sessionIdsByMLContext.get(t);if(r.delete(e),r.size===0){this.sessionIdsByMLContext.delete(t);let i=this.mlContextCache.findIndex(n=>n.mlContext===t);i!==-1&&this.mlContextCache.splice(i,1)}}getMLContext(e){return this.mlContextBySessionId.get(e)}getMLOpSupportLimits(e){return this.mlOpSupportLimitsBySessionId.get(e)}reserveTensorId(){return this.tensorManager.reserveTensorId()}releaseTensorId(e){de("verbose",()=>`[WebNN] releaseTensorId {tensorId: ${e}}`),this.tensorManager.releaseTensorId(e)}async ensureTensor(e,t,r,i,n){let a=lr.get(r);if(!a)throw new Error(`Unsupported ONNX data type: ${r}`);return this.tensorManager.ensureTensor(e??this.currentSessionId,t,a,i,n)}async createTemporaryTensor(e,t,r){de("verbose",()=>`[WebNN] createTemporaryTensor {onnxDataType: ${t}, shape: ${r}}`);let i=lr.get(t);if(!i)throw new Error(`Unsupported ONNX data type: ${t}`);let n=this.tensorManager.reserveTensorId();await this.tensorManager.ensureTensor(e,n,i,r,!1);let a=this.temporarySessionTensorIds.get(e);return a?a.push(n):this.temporarySessionTensorIds.set(e,[n]),n}uploadTensor(e,t){if(!we().shouldTransferToMLTensor)throw new Error("Trying to upload to a MLTensor while shouldTransferToMLTensor is false");de("verbose",()=>`[WebNN] uploadTensor {tensorId: ${e}, data: ${t.byteLength}}`),this.tensorManager.upload(e,t)}async downloadTensor(e,t){return this.tensorManager.download(e,t)}createMLTensorDownloader(e,t){return async()=>{let r=await this.tensorManager.download(e);return ha(r,t)}}registerMLTensor(e,t,r,i){let n=lr.get(r);if(!n)throw new Error(`Unsupported ONNX data type: ${r}`);let a=this.tensorManager.registerTensor(e,t,n,i);return de("verbose",()=>`[WebNN] registerMLTensor {tensor: ${t}, dataType: ${n}, dimensions: ${i}} -> {tensorId: ${a}}`),a}registerMLConstant(e,t,r,i,n,a,s=!1){if(!a)throw new Error("External mounted files are not available.");let u=e;e.startsWith("./")&&(u=e.substring(2));let l=a.get(u);if(!l)throw new Error(`File with name ${u} not found in preloaded files.`);if(t+r>l.byteLength)throw new Error("Out of bounds: data offset and length exceed the external file data size.");let d=l.slice(t,t+r).buffer,f;switch(n.dataType){case"float32":f=new Float32Array(d);break;case"float16":f=typeof Float16Array<"u"&&Float16Array.from?new Float16Array(d):new Uint16Array(d);break;case"int32":f=new Int32Array(d);break;case"uint32":f=new Uint32Array(d);break;case"int64":if(s){let h=qn(new Uint8Array(d),"int64");f=new Int32Array(h.buffer),n.dataType="int32"}else f=new BigInt64Array(d);break;case"uint64":f=new BigUint64Array(d);break;case"int8":f=new Int8Array(d);break;case"int4":case"uint4":case"uint8":f=new Uint8Array(d);break;default:throw new Error(`Unsupported data type: ${n.dataType} in creating WebNN Constant from external data.`)}return de("verbose",()=>`[WebNN] registerMLConstant {dataType: ${n.dataType}, shape: ${n.shape}}} ${s?"(Note: it was int64 data type and registered to int32 as workaround)":""}`),i.constant(n,f)}registerGraphInput(e){this.temporaryGraphInputs.push(e)}registerGraphOutput(e){this.temporaryGraphOutputs.push(e)}isGraphInput(e,t){let r=this.sessionGraphInputs.get(e);return r?r.includes(t):!1}isGraphOutput(e,t){let r=this.sessionGraphOutputs.get(e);return r?r.includes(t):!1}isGraphInputOutputTypeSupported(e,t,r=!0){let i=lr.get(Rt(t)),n=this.mlOpSupportLimitsBySessionId.get(e);return typeof i>"u"?!1:r?!!n?.input.dataTypes.includes(i):!!n?.output.dataTypes.includes(i)}flush(){}}}),fa=P(()=>{"use strict"}),Yy=P(()=>{"use strict";lt(),fa(),on=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]]),Hr=[],Kr=e=>Math.ceil(Number(e)/16)*16,tu=e=>{for(let t=0;t<Hr.length;t++){let r=Hr[t];if(e<=r)return r}return Math.ceil(e/16)*16},ru=1,un=()=>ru++,Vn=async(e,t,r,i)=>{let n=Kr(r),a=e.device.createBuffer({size:n,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let s=e.getCommandEncoder();e.endComputePass(),s.copyBufferToBuffer(t,0,a,0,n),e.flush(),await a.mapAsync(GPUMapMode.READ);let u=a.getMappedRange();if(i){let l=i();return l.set(new Uint8Array(u,0,r)),l}else return new Uint8Array(u.slice(0,r))}finally{a.destroy()}},iu=class{constructor(e){this.backend=e,this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersPending=[],this.capturedPendingBuffers=new Map;for(let[t]of on)Hr.push(t),this.freeBuffers.set(t,[]),this.freeUniformBuffers.set(t,[]);this.sessionCount=0}upload(e,t){let r=t.buffer,i=t.byteOffset,n=t.byteLength,a=Kr(n),s=this.storageCache.get(e);if(!s)throw new Error("gpu data for uploading does not exist");if(Number(s.originalSize)!==n)throw new Error(`inconsistent data size. gpu data size=${s.originalSize}, data size=${n}`);let u=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),l=u.getMappedRange();new Uint8Array(l).set(new Uint8Array(r,i,n)),u.unmap();let d=this.backend.device.createCommandEncoder();d.copyBufferToBuffer(u,0,s.gpuData.buffer,0,a),this.backend.device.queue.submit([d.finish()]),u.destroy(),de("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${e})`)}memcpy(e,t){let r=this.storageCache.get(e);if(!r)throw new Error("source gpu data for memcpy does not exist");let i=this.storageCache.get(t);if(!i)throw new Error("destination gpu data for memcpy does not exist");if(r.originalSize!==i.originalSize)throw new Error("inconsistent source and destination gpu data size");let n=Kr(r.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(r.gpuData.buffer,0,i.gpuData.buffer,0,n)}registerExternalBuffer(e,t,r){let i;if(r){if(i=r[0],e===r[1])return de("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${i}, buffer is the same, skip.`),i;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
+             Please use the previous external buffer!`)}else i=un();return this.storageCache.set(i,{gpuData:{id:i,type:0,buffer:e},originalSize:t}),de("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${i}, registered.`),i}unregisterExternalBuffer(e){e!==void 0&&(this.storageCache.delete(e),de("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${e}`))}create(e,t=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=tu(e),i,n=(t&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(t&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(n||a){let u=(n?this.freeBuffers:this.freeUniformBuffers).get(r);u?u.length>0?i=u.pop():i=this.backend.device.createBuffer({size:r,usage:t}):i=this.backend.device.createBuffer({size:r,usage:t})}else i=this.backend.device.createBuffer({size:r,usage:t});let s={id:un(),type:0,buffer:i};return this.storageCache.set(s.id,{gpuData:s,originalSize:Number(e)}),de("verbose",()=>`[WebGPU] GpuDataManager.create(size=${e}) => id=${s.id}`),s}get(e){return this.storageCache.get(e)?.gpuData}release(e){let t=typeof e=="bigint"?Number(e):e,r=this.storageCache.get(t);if(!r){if(this.storageCache.size===0)return 0;throw new Error("releasing data does not exist")}return de("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(e,t){let r=this.storageCache.get(Number(e));if(!r)throw new Error("data does not exist");await Vn(this.backend,r.gpuData.buffer,r.originalSize,t)}refreshPendingBuffers(){if(this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let e of this.buffersPending){let t=on.get(e.size);if((e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else if((e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else e.destroy()}this.buffersPending=[]}else{let e=this.capturedPendingBuffers.get(this.backend.currentSessionId);e||(e=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,e));for(let t of this.buffersPending)e.push(t);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.freeUniformBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache.forEach(e=>{e.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(e){let t=this.capturedPendingBuffers.get(e);t&&(t.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(e)),this.sessionCount-=1,this.sessionCount===0&&(de("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(r=>{r.gpuData.buffer.destroy()}),this.storageCache=new Map)}},Fp=(...e)=>new iu(...e)}),Te=P(()=>{"use strict";nu=class{constructor(e){Object.assign(this,e)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(e=>`${this[e]}`).join(";")),this.key}},me=e=>new nu(e)}),ne=P(()=>{"use strict";ee(),ie(),Zt=64,jr=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${t}<f16>`:"f16";case 1:return t>1?`vec${t}<f32>`:"f32";case 6:return t>1?`vec${t}<i32>`:"i32";case 12:return t>1?`vec${t}<u32>`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];case 22:return"i32";case 21:return"u32";default:throw new Error(`Unknown data type: ${e}`)}},Ee=(e,t=1)=>{let r=jr(e,t);return typeof r=="string"?r:r[0]},Ae=(e,t=1)=>{let r=jr(e,t);return typeof r=="string"?r:r[1]},Z=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:R.computeStrides(r)})}),t},xe=e=>e%4===0?4:e%2===0?2:1,Gn=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,jt=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,vt=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,K=(e,t,r,i)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?i==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:i==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,dr=(e,t,r,i,n)=>{let a=typeof r=="number",s=a?r:r.length,u=[...new Array(s).keys()],l=s<2?"u32":s<=4?`vec${s}<u32>`:`array<u32, ${s}>`,d=jr(t,n),f=typeof d=="string"?d:d[1],h=typeof d=="string"?d:d[0],g={indices:l,value:f,storage:h,tensor:t},w=N=>typeof N=="string"?N:`${N}u`,y={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},v=a?"uniforms.":"",T=`${v}${e}_shape`,$=`${v}${e}_strides`,_="";for(let N=0;N<s-1;N++)_+=`
+    let dim${N} = current / ${K($,N,s)};
+    let rest${N} = current % ${K($,N,s)};
+    indices[${N}] = dim${N};
+    current = rest${N};
+    `;_+=`indices[${s-1}] = current;`;let I=s<2?"":`
+  fn o2i_${e}(offset: u32) -> ${g.indices} {
+    var indices: ${g.indices};
+    var current = offset;
+    ${_}
+    return indices;
+  }`,k=N=>(y.offsetToIndices=!0,s<2?N:`o2i_${e}(${N})`),E=[];if(s>=2)for(let N=s-1;N>=0;N--)E.push(`${K($,N,s)} * (indices[${N}])`);let C=s<2?"":`
+  fn i2o_${e}(indices: ${g.indices}) -> u32 {
+    return ${E.join("+")};
+  }`,O=N=>(y.indicesToOffset=!0,s<2?N:`i2o_${e}(${N})`),x=(...N)=>s===0?"0u":`${g.indices}(${N.map(w).join(",")})`,D=(N,V)=>s<2?`${N}`:`${K(N,V,s)}`,L=(N,V,te)=>s<2?`${N}=${te};`:`${K(N,V,s)}=${te};`,Y={},q=(N,V)=>{y.broadcastedIndicesToOffset=!0;let te=`${V.name}broadcastedIndicesTo${e}Offset`;if(te in Y)return`${te}(${N})`;let ce=[];for(let Ie=s-1;Ie>=0;Ie--){let Me=V.indicesGet("outputIndices",Ie+V.rank-s);ce.push(`${D($,Ie)} * (${Me} % ${D(T,Ie)})`)}return Y[te]=`fn ${te}(outputIndices: ${V.type.indices}) -> u32 {
+             return ${ce.length>0?ce.join("+"):"0u"};
+           }`,`${te}(${N})`},j=(N,V)=>(()=>{if(g.storage===g.value)return`${e}[${N}]=${V};`;if(g.storage==="vec2<u32>"&&g.value==="i32")return`${e}[${N}]=vec2<u32>(u32(${V}), select(0u, 0xFFFFFFFFu, ${V} < 0));`;if(g.storage==="vec2<u32>"&&g.value==="u32")return`${e}[${N}]=vec2<u32>(u32(${V}), 0u);`;if(g.storage==="u32"&&g.value==="vec4<bool>")return`${e}[${N}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${V}));`;throw new Error(`not supported combination of storage type ${g.storage} and value type ${g.value} yet`)})(),ue=N=>(()=>{if(g.storage===g.value)return`${e}[${N}]`;if(g.storage==="vec2<u32>"&&g.value==="i32")return`i32(${e}[${N}].x)`;if(g.storage==="vec2<u32>"&&g.value==="u32")return`u32(${e}[${N}].x)`;if(g.storage==="u32"&&g.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${N}] & 0xFFu), bool(${e}[${N}] & 0xFF00u), bool(${e}[${N}] & 0xFF0000u), bool(${e}[${N}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${g.storage} and value type ${g.value} yet`)})(),A=s<2?"":`
+  fn get_${e}ByIndices(indices: ${g.indices}) -> ${f} {
+    return ${ue(`i2o_${e}(indices)`)};
+  }`,U=s<2?"":(()=>{let N=u.map(te=>`d${te}: u32`).join(", "),V=u.map(te=>`d${te}`).join(", ");return`
+  fn get_${e}(${N}) -> ${f} {
+    return get_${e}ByIndices(${x(V)});
+  }`})(),re=(...N)=>{if(N.length!==s)throw new Error(`indices length must be ${s}`);let V=N.map(w).join(",");return s===0?ue("0u"):s===1?ue(V[0]):(y.get=!0,y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}(${V})`)},J=N=>s<2?ue(N):(y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}ByIndices(${N})`),H=s<2?"":`
+  fn set_${e}ByIndices(indices: ${g.indices}, value: ${f}) {
+    ${j(`i2o_${e}(indices)`,"value")}
+  }`,pe=s<2?"":(()=>{let N=u.map(te=>`d${te}: u32`).join(", "),V=u.map(te=>`d${te}`).join(", ");return`
+  fn set_${e}(${N}, value: ${f}) {
+    set_${e}ByIndices(${x(V)}, value);
+  }`})();return{impl:()=>{let N=[],V=!1;return y.offsetToIndices&&(N.push(I),V=!0),y.indicesToOffset&&(N.push(C),V=!0),y.broadcastedIndicesToOffset&&(Object.values(Y).forEach(te=>N.push(te)),V=!0),y.set&&(N.push(pe),V=!0),y.setByIndices&&(N.push(H),V=!0),y.get&&(N.push(U),V=!0),y.getByIndices&&(N.push(A),V=!0),!a&&V&&N.unshift(`const ${T} = ${g.indices}(${r.join(",")});`,`const ${$} = ${g.indices}(${R.computeStrides(r).join(",")});`),N.join(`
+`)},type:g,offsetToIndices:k,indicesToOffset:O,broadcastedIndicesToOffset:q,indices:x,indicesGet:D,indicesSet:L,set:(...N)=>{if(N.length!==s+1)throw new Error(`indices length must be ${s}`);let V=N[s];if(typeof V!="string")throw new Error("value must be string");let te=N.slice(0,s).map(w).join(",");return s===0?j("0u",V):s===1?j(te[0],V):(y.set=!0,y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}(${te}, ${V})`)},setByOffset:j,setByIndices:(N,V)=>s<2?j(N,V):(y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}ByIndices(${N}, ${V});`),get:re,getByOffset:ue,getByIndices:J,usage:i,name:e,strides:$,shape:T,rank:s}},B=(e,t,r,i=1)=>dr(e,t,r,"input",i),F=(e,t,r,i=1)=>dr(e,t,r,"output",i),Hp=(e,t,r)=>dr(e,t,r,"atomicOutput",1),ma=(e,t,r,i=1)=>dr(e,t,r,"internal",i),au=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=Zt){let t=typeof e=="number"?e:e[0],r=typeof e=="number"?1:e[1],i=typeof e=="number"?1:e[2];if(t>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||i>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${t}, ${r}, ${i}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(t*r*i>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${t}, ${r}, ${i}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let n=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,a=n?`@builtin(global_invocation_id) global_id : vec3<u32>,
+    @builtin(workgroup_id) workgroup_id : vec3<u32>,
+    @builtin(local_invocation_index) local_idx : u32,
+    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
+                                             @builtin(local_invocation_id) local_id : vec3<u32>,
+    @builtin(local_invocation_index) local_idx : u32,
+    @builtin(workgroup_id) workgroup_id : vec3<u32>,
+    @builtin(num_workgroups) num_workgroups : vec3<u32>`,s=n?`let global_idx = global_id.x;
+         let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] +
+             workgroup_id.y * num_workgroups[0] + workgroup_id.x;
+         let global_idx = workgroup_index * ${t*r*i}u + local_idx;`;return`@compute @workgroup_size(${t}, ${r}, ${i})
+  fn main(${a}) {
+    ${s}
+  `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith("uniforms.")&&this.uniforms.push({name:e.shape.replace("uniforms.",""),type:"u32",length:e.rank}),e.strides.startsWith("uniforms.")&&this.uniforms.push({name:e.strides.replace("uniforms.",""),type:"u32",length:e.rank}))}declareVariable(e,t){if(e.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(e),this.appendVariableUniforms(e);let r=e.usage==="input"?"read":"read_write",i=e.usage==="atomicOutput"?"atomic<i32>":e.type.storage;return`@group(0) @binding(${t}) var<storage, ${r}> ${e.name}: array<${i}>;`}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:i}of this.uniforms)if(i&&i>4)r==="f16"?e.push(`@align(16) ${t}:array<mat2x4<${r}>, ${Math.ceil(i/8)}>`):e.push(`${t}:array<vec4<${r}>, ${Math.ceil(i/4)}>`);else{let n=i==null||i===1?r:`vec${i}<${r}>`;e.push(`${t}:${n}`)}return`
+      struct Uniforms { ${e.join(", ")} };
+      @group(0) @binding(${this.variableIndex}) var<uniform> 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])}},Kp=(e,t)=>new au(e,t)}),$t=P(()=>{"use strict";ee(),ie(),Te(),ne(),su=(e,t)=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.");if(t.length!==0&&t.length!==e[0].dims.length)throw new Error(`perm size ${t.length} does not match input rank ${e[0].dims.length}`)},ln=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),ou=(e,t)=>R.sortBasedOnPerm(e,ln(e.length,t)),uu=(e,t,r,i)=>{let n=`fn perm(i: ${i.type.indices}) -> ${r.type.indices} {
+    var a: ${r.type.indices};`;for(let a=0;a<t;++a)n+=`a[${e[a]}]=i[${a}];`;return n+="return a;}"},lu=(e,t)=>{let r=[],i=[];for(let n=0;n<e.length;++n)e[n]!==1&&r.push(e[n]),e[t[n]]!==1&&i.push(t[n]);return{newShape:r,newPerm:i}},du=(e,t)=>{let r=0;for(let i=0;i<e.length;++i)if(t[e[i]]!==1){if(e[i]<r)return!1;r=e[i]}return!0},De=(e,t)=>{let r=e.dataType,i=e.dims.length,n=ln(i,t),a=ou(e.dims,n),s=e.dims,u=a,l=i<2||du(n,e.dims),d;if(l)return d=y=>{let v=B("input",r,s,4),T=F("output",r,u,4);return`
+  ${y.registerUniform("output_size","u32").declareVariables(v,T)}
+  ${y.mainStart()}
+    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+    output[global_idx] = input[global_idx];
+  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let y=R.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(y/64/4)},programUniforms:[{type:12,data:Math.ceil(y/4)}]}},getShaderSource:d};let{newShape:f,newPerm:h}=lu(e.dims,n),g=R.areEqual(h,[2,3,1]),w=R.areEqual(h,[3,1,2]);if(f.length===2||g||w){s=g?[f[0],f[1]*f[2]]:w?[f[0]*f[1],f[2]]:f,u=[s[1],s[0]];let y=16;return d=v=>{let T=B("a",r,s.length),$=F("output",r,u.length);return`
+  ${v.registerUniform("output_size","u32").declareVariables(T,$)}
+  var<workgroup> tile : array<array<${$.type.value}, ${y+1}>, ${y}>;
+  ${v.mainStart([y,y,1])}
+    let stride = (uniforms.output_shape[1] - 1) / ${y} + 1;
+    let workgroup_id_x = workgroup_index % stride;
+    let workgroup_id_y = workgroup_index / stride;
+    let input_col = workgroup_id_y * ${y}u + local_id.x;
+    let input_row = workgroup_id_x * ${y}u + local_id.y;
+    if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) {
+      tile[local_id.y][local_id.x] = ${T.getByIndices(`${T.type.indices}(input_row, input_col)`)};
+    }
+    workgroupBarrier();
+
+    let output_col = workgroup_id_x * ${y}u + local_id.x;
+    let output_row = workgroup_id_y * ${y}u + local_id.y;
+    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
+      ${$.setByIndices(`${$.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
+    }
+  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let v=R.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(u[1]/y),y:Math.ceil(u[0]/y)},programUniforms:[{type:12,data:v},...Z(s,u)]}},getShaderSource:d}}return d=y=>{let v=B("a",r,s.length),T=F("output",r,u.length);return`
+  ${y.registerUniform("output_size","u32").declareVariables(v,T)}
+
+  ${uu(n,i,v,T)}
+
+  ${y.mainStart()}
+    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+
+    let indices = ${T.offsetToIndices("global_idx")};
+    let aIndices = perm(indices);
+
+    ${T.setByOffset("global_idx",v.getByIndices("aIndices"))}
+  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let y=R.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},...Z(s,u)]}},getShaderSource:d}},jp=(e,t)=>{su(e.inputs,t.perm),e.compute(De(e.inputs[0],t.perm))},Qp=e=>me({perm:e.perm})}),Xy=P(()=>{"use strict";ee(),ie(),ne(),ga(),$t(),pu={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"},cu={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"},hu={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},fu={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},mu=(e,t)=>{let r=[];for(let i=t-e;i<t;++i)r.push(i);return r},gu=(e,t)=>{let r=[],i=e.length;for(let a=0;a<i;a++)t.indexOf(a)===-1&&r.push(e[a]);let n=t.map(a=>e[a]);return[r,n]},yu=(e,t)=>{let r=e.length+t.length,i=[],n=0;for(let a=0;a<r;a++)t.indexOf(a)===-1?i.push(e[n++]):i.push(1);return i},wu=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},_u=(e,t)=>{let r=[];if(!wu(e,t)){for(let i=0;i<t;++i)e.indexOf(i)===-1&&r.push(i);e.forEach(i=>r.push(i))}return r},bu=(e,t,r,i,n,a,s)=>{let u=r[0].dims,l=R.size(a),d=R.size(s),f=B("_A",r[0].dataType,u),h=F("output",n,a),g=64;l===1&&(g=256);let w=`
+          var<workgroup> aBestValues : array<f32, ${g}>;
+       `,y=v=>`
+        ${v.registerUniform("reduceSize","u32").declareVariables(f,h)}
+        ${w}
+        fn DIV_CEIL(a : u32, b : u32) -> u32 {
+          return ((a - 1u) / b + 1u);
+         }
+         ${v.mainStart(g)}
+
+          let outputIndex = global_idx / ${g};
+          let offset = outputIndex * uniforms.reduceSize;
+
+          var bestValue = f32(${hu[i]});
+          let Length = uniforms.reduceSize;
+          for (var k = local_idx; k < Length; k = k + ${g}) {
+           let candidate = f32(${f.getByOffset("offset + k")});
+           bestValue = ${pu[i]};
+          }
+          aBestValues[local_idx] = bestValue;
+          workgroupBarrier();
+
+         var reduceSize = min(Length, ${g}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 = ${cu[i]};
+            aBestValues[local_idx] = bestValue;
+           }
+           reduceSize = interval;
+           workgroupBarrier();
+         }
+
+         if (local_idx == 0u) {
+          ${h.setByOffset("outputIndex",`${i==="mean"?`${h.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${h.type.storage}(${fu[i]})`}`)};
+         }
+        }`;return{name:e,shaderCache:{hint:`${t};${g}`,inputDependencies:["type"]},getShaderSource:y,getRunData:()=>({outputs:[{dims:a,dataType:n}],dispatchGroup:{x:l},programUniforms:[{type:12,data:d}]})}},Fe=(e,t,r,i)=>{let n=e.inputs.length===1?r:Fn(e.inputs,r),a=n.axes;a.length===0&&!n.noopWithEmptyAxes&&(a=e.inputs[0].dims.map((w,y)=>y));let s=R.normalizeAxes(a,e.inputs[0].dims.length),u=s,l=e.inputs[0],d=_u(u,e.inputs[0].dims.length);d.length>0&&(l=e.compute(De(e.inputs[0],d),{inputs:[0],outputs:[-1]})[0],u=mu(u.length,l.dims.length));let[f,h]=gu(l.dims,u),g=f;n.keepDims&&(g=yu(f,s)),e.compute(bu(t,n.cacheKey,[l],i,e.inputs[0].dataType,g,h),{inputs:[l]})},Zp=(e,t)=>{Fe(e,"ReduceMeanShared",t,"mean")},Yp=(e,t)=>{Fe(e,"ReduceL1Shared",t,"l1")},Xp=(e,t)=>{Fe(e,"ReduceL2Shared",t,"l2")},Jp=(e,t)=>{Fe(e,"ReduceLogSumExpShared",t,"logSumExp")},ec=(e,t)=>{Fe(e,"ReduceMaxShared",t,"max")},tc=(e,t)=>{Fe(e,"ReduceMinShared",t,"min")},rc=(e,t)=>{Fe(e,"ReduceProdShared",t,"prod")},ic=(e,t)=>{Fe(e,"ReduceSumShared",t,"sum")},nc=(e,t)=>{Fe(e,"ReduceSumSquareShared",t,"sumSquare")},ac=(e,t)=>{Fe(e,"ReduceLogSumShared",t,"logSum")}}),ga=P(()=>{"use strict";ee(),ie(),Te(),ne(),Xy(),He=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.")},vu=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],oi=(e,t,r,i,n,a,s=!1,u=!1)=>{let l=[],d=r[0].dims,f=d.length,h=R.normalizeAxes(n,f),g=!u&&h.length===0;d.forEach((v,T)=>{g||h.indexOf(T)>=0?s&&l.push(1):l.push(v)});let w=l.length,y=R.size(l);return{name:e,shaderCache:t,getShaderSource:v=>{let T=[],$=B("_A",r[0].dataType,f),_=F("output",a,w),I=i($,_,h),k=I[2];for(let E=0,C=0;E<f;E++)g||h.indexOf(E)>=0?(s&&C++,k=`for(var j${E}: u32 = 0; j${E} < ${d[E]}; j${E}++) {
+                  ${I[2].includes("last_index")?`let last_index = j${E};`:""}
+                  ${$.indicesSet("input_indices",E,`j${E}`)}
+                  ${k}
+                }`):(T.push(`${$.indicesSet("input_indices",E,_.indicesGet("output_indices",C))};`),C++);return`
+
+        ${v.registerUniform("output_size","u32").declareVariables($,_)}
+
+        ${v.mainStart()}
+          ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+          var input_indices: ${$.type.indices};
+          let output_indices = ${_.offsetToIndices("global_idx")};
+
+          ${T.join(`
+`)}
+          ${I[0]}       // init ops for reduce max/min
+          ${I[1]}
+          ${k}
+          ${I[3]}
+          ${I.length===4?_.setByOffset("global_idx","value"):I.slice(4).join(`
+`)}
+        }`},getRunData:()=>({outputs:[{dims:l,dataType:a}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},...Z(d,l)]})}},Fn=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(i=>r.push(Number(i))),me({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},Ke=(e,t,r,i)=>{let n=e.inputs,a=n.length===1?r:Fn(n,r);e.compute(oi(t,{hint:a.cacheKey,inputDependencies:["rank"]},[n[0]],a.noopWithEmptyAxes&&a.axes.length===0?vu:i,a.axes,n[0].dataType,a.keepDims,a.noopWithEmptyAxes),{inputs:[0]})},$u=(e,t)=>{He(e.inputs),Ke(e,"ReduceLogSum",t,(r,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},xu=(e,t)=>{He(e.inputs),Ke(e,"ReduceL1",t,(r,i)=>[`var value = ${i.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},Tu=(e,t)=>{He(e.inputs),Ke(e,"ReduceL2",t,(r,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Su=(e,t)=>{He(e.inputs),Ke(e,"ReduceLogSumExp",t,(r,i)=>[`var value = ${i.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},ku=(e,t)=>{He(e.inputs),Ke(e,"ReduceMax",t,(r,i,n)=>{let a=[];for(let s=0;s<r.rank;s++)(n.indexOf(s)>=0||n.length===0)&&a.push(r.indicesSet("input_indices",s,0));return[`${a.join(`
+`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},Iu=(e,t)=>{He(e.inputs),Ke(e,"ReduceMean",t,(r,i,n)=>{let a=1;for(let s=0;s<r.rank;s++)(n.indexOf(s)>=0||n.length===0)&&(a*=e.inputs[0].dims[s]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${i.type.value}(sum / ${a});`]})},Eu=(e,t)=>{He(e.inputs),Ke(e,"ReduceMin",t,(r,i,n)=>{let a=[];for(let s=0;s<r.rank;s++)(n.indexOf(s)>=0||n.length===0)&&a.push(`input_indices[${s}] = 0;`);return[`${a.join(`
+`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},Cu=(e,t)=>{He(e.inputs),Ke(e,"ReduceProd",t,(r,i)=>[`var value = ${i.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},zu=(e,t)=>{He(e.inputs),Ke(e,"ReduceSum",t,(r,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},Au=(e,t)=>{He(e.inputs),Ke(e,"ReduceSumSquare",t,(r,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},je=(e,t,r)=>{if(t.length===0)return r;let i=1,n=1;for(let a=0;a<t.length;a++)t.indexOf(a)===-1?i*=e[a]:n*=e[a];return n<32&&i>1024},sc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Iu(e,t):Zp(e,t)},oc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?xu(e,t):Yp(e,t)},uc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Tu(e,t):Xp(e,t)},lc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Su(e,t):Jp(e,t)},dc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ku(e,t):ec(e,t)},pc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Eu(e,t):tc(e,t)},cc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Cu(e,t):rc(e,t)},hc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?zu(e,t):ic(e,t)},fc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Au(e,t):nc(e,t)},mc=(e,t)=>{je(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?$u(e,t):ac(e,t)}}),Jy=P(()=>{"use strict";ee(),Te(),ga(),dn=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.")},gc=(e,t)=>{dn(e.inputs);let r=(i,n,a)=>{let s=[];for(let u=0;u<i.rank;u++)(a.indexOf(u)>=0||a.length===0)&&s.push(`input_indices[${u}] = 0;`);return[`${s.join(`
+`)}`,`var value = ${i.getByIndices("input_indices")};
+var best_index : i32 = 0;`,`if (${i.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {
+         value = ${i.getByIndices("input_indices")};
+         best_index = i32(last_index);
+       }`,"",n.setByOffset("global_idx","best_index")]};e.compute(oi("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},yc=(e,t)=>{dn(e.inputs);let r=(i,n,a)=>{let s=[];for(let u=0;u<i.rank;u++)(a.indexOf(u)>=0||a.length===0)&&s.push(`input_indices[${u}] = 0;`);return[`${s.join(`
+`)}`,`var value = ${i.getByIndices("input_indices")};
+var best_index : i32 = 0;`,`if (${i.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {
+         value = ${i.getByIndices("input_indices")};
+         best_index = i32(last_index);
+       }`,"",n.setByOffset("global_idx","best_index")]};e.compute(oi("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},Hn=e=>me(e)}),ya=P(()=>{"use strict";ee(),ie(),fa(),ne(),Ou=(e,t)=>{let r=e[0],i=e[1],n=e[2],a=e[3],s=e[4],u=e[5];if(s&&u)throw new Error("Attention cannot have both past and attention_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let l=r.dims[0],d=r.dims[1],f=r.dims[2];if(n.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(i.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(i.dims[0]!==f)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(n.dims[0]!==i.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let h=n.dims[0]/3,g=h,w=g;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let I of t.qkvHiddenSizes)if(I%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");h=t.qkvHiddenSizes[0],g=t.qkvHiddenSizes[1],w=t.qkvHiddenSizes[2]}let y=d;if(h!==g)throw new Error("qkv_hidden_sizes first element should be same as the second");if(n.dims[0]!==h+g+w)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let v=0;if(s){if(g!==w)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(s.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(s.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(s.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(s.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(s.dims[4]!==g/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(v=s.dims[3])}let T=y+v,$=-1,_=0;if(a)throw new Error("Mask not supported");if(s)throw new Error("past is not supported");if(u){if(u.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(u.dims[0]!==l||u.dims[1]!==t.numHeads||u.dims[2]!==d||u.dims[3]!==T)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:l,sequenceLength:d,pastSequenceLength:v,kvSequenceLength:y,totalSequenceLength:T,maxSequenceLength:$,inputHiddenSize:f,hiddenSize:h,vHiddenSize:w,headSize:Math.floor(h/t.numHeads),vHeadSize:Math.floor(w/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:_,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Qr=(e,t,r)=>t&&e?`
+      let total_sequence_length_input = u32(${t.getByOffset("0")});
+      let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
+      let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
+      let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input;
+      total_sequence_length = u32(${e?.getByOffset("batchIdx")}) + 1;
+      var past_sequence_length: u32 = 0;
+      if (is_first_prompt == false) {
+        past_sequence_length = total_sequence_length - sequence_length;
+      }
+       `:`
+    ${r?"let past_sequence_length = uniforms.past_sequence_length":""};
+    let present_sequence_length = total_sequence_length;
+    `,Ru=(e,t,r,i,n,a,s,u)=>{let l=xe(s?1:a),d=64,f=a/l;f<d&&(d=32);let h=Math.ceil(a/l/d),g=[{type:12,data:t},{type:12,data:r},{type:12,data:i},{type:12,data:n},{type:12,data:f},{type:12,data:h}],w=Ee(e.dataType,l),y=Ae(1,l),v=["type"];s&&v.push("type"),u&&v.push("type");let T=$=>{let _=F("x",e.dataType,e.dims,l),I=[_],k=s?B("seq_lens",s.dataType,s.dims):void 0;k&&I.push(k);let E=u?B("total_sequence_length_input",u.dataType,u.dims):void 0;E&&I.push(E);let C=Ae(e.dataType),O=[{name:"batch_size",type:"u32"},{name:"num_heads",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"sequence_length",type:"u32"},{name:"total_sequence_length",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
+  var<workgroup> thread_max: array<f32, ${d}>;
+  var<workgroup> thread_sum: array<f32, ${d}>;
+  ${$.registerUniforms(O).declareVariables(...I)}
+  ${$.mainStart([d,1,1])}
+    let batchIdx = workgroup_id.z / uniforms.num_heads;
+    let headIdx = workgroup_id.z % uniforms.num_heads;
+    let sequence_length = uniforms.sequence_length;
+    var total_sequence_length = uniforms.total_sequence_length;
+    ${Qr(k,E,!1)}
+    let local_offset = local_idx * uniforms.elements_per_thread;
+    let offset = (global_idx / ${d}) * uniforms.total_sequence_length + local_offset;
+    let seq_causal_length = ${s?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
+    var thread_max_vector = ${y}(-3.4028234663852886e+38f);
+    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
+      thread_max_vector = max(${y}(x[offset + i]), thread_max_vector);
+    }
+    thread_max[local_idx] = ${(()=>{switch(l){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: ${l}`)}})()};
+    workgroupBarrier();
+
+    var max_value =  f32(-3.4028234663852886e+38f);
+    for (var i = 0u; i < ${d}; i++) {
+      max_value = max(thread_max[i], max_value);
+    }
+
+    var sum_vector = ${y}(0);
+    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
+      sum_vector += exp(${y}(x[offset + i]) - max_value);
+    }
+    thread_sum[local_idx] = ${(()=>{switch(l){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: ${l}`)}})()};
+    workgroupBarrier();
+
+    var sum: f32 = 0;
+    for (var i = 0u; i < ${d}; i++) {
+      sum += thread_sum[i];
+    }
+
+    if (sum == 0) {
+      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
+        x[offset + i] = ${_.type.value}(${C}(1.0) / ${C}(seq_causal_length));
+      }
+    } else {
+      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
+        var f32input = ${y}(x[offset + i]);
+        x[offset + i] = ${_.type.value}(exp(f32input - max_value) / sum);
+      }
+    }
+      ${s?`
+        for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) {
+          x[offset + total_seq_id] = ${_.type.value}(${C}(0));
+        }`:""};
+  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${d};${w};${l}`,inputDependencies:v},getShaderSource:T,getRunData:()=>({outputs:[],dispatchGroup:{x:1,y:n,z:t*r},programUniforms:g})}},Mu=(e,t,r,i,n,a,s,u,l)=>{let d=s+a.kvSequenceLength,f=[a.batchSize,a.numHeads,a.sequenceLength,d],h=e>1&&i,g=a.kvNumHeads?a.kvNumHeads:a.numHeads,w=h?[a.batchSize,g,d,a.headSize]:void 0,y=a.nReps?a.nReps:1,v=a.scale===0?1/Math.sqrt(a.headSize):a.scale,T=xe(a.headSize),$=a.headSize/T,_=12,I={x:Math.ceil(d/_),y:Math.ceil(a.sequenceLength/_),z:a.batchSize*a.numHeads},k=[{type:12,data:a.sequenceLength},{type:12,data:$},{type:12,data:d},{type:12,data:a.numHeads},{type:12,data:a.headSize},{type:1,data:v},{type:12,data:s},{type:12,data:a.kvSequenceLength},{type:12,data:y}],E=h&&i&&R.size(i.dims)>0,C=["type","type"];E&&C.push("type"),n&&C.push("type"),u&&C.push("type"),l&&C.push("type");let O=[{dims:f,dataType:t.dataType,gpuDataType:0}];h&&O.push({dims:w,dataType:t.dataType,gpuDataType:0});let x=D=>{let L=B("q",t.dataType,t.dims,T),Y=B("key",r.dataType,r.dims,T),q=[L,Y];if(E){let H=B("past_key",i.dataType,i.dims,T);q.push(H)}n&&q.push(B("attention_bias",n.dataType,n.dims));let j=u?B("seq_lens",u.dataType,u.dims):void 0;j&&q.push(j);let ue=l?B("total_sequence_length_input",l.dataType,l.dims):void 0;ue&&q.push(ue);let A=F("output",t.dataType,f),U=[A];h&&U.push(F("present_key",t.dataType,w,T));let re=Ae(1,T),J=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
+  const TILE_SIZE = ${_}u;
+
+  var<workgroup> tileQ: array<${L.type.storage}, ${_*_}>;
+  var<workgroup> tileK: array<${L.type.storage}, ${_*_}>;
+  ${D.registerUniforms(J).declareVariables(...q,...U)}
+  ${D.mainStart([_,_,1])}
+    // x holds the N and y holds the M
+    let headIdx = workgroup_id.z % uniforms.num_heads;
+    let kvHeadIdx = ${y===1?"headIdx":"headIdx / uniforms.n_reps"};
+    let kv_num_heads = ${y===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
+    let batchIdx = workgroup_id.z / uniforms.num_heads;
+    let m = workgroup_id.y * TILE_SIZE;
+    let n = workgroup_id.x * TILE_SIZE;
+    let sequence_length = uniforms.M;
+    var total_sequence_length = uniforms.N;
+    ${Qr(j,ue,!0)}
+    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
+    let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
+    ${E&&h?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
+    let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
+    ${h?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
+    var value = ${re}(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) {
+        var idx = TILE_SIZE * local_id.y + local_id.x;
+      ${E&&h?`
+              if (n + local_id.y < past_sequence_length) {
+                tileK[idx] = past_key[pastKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
+              } else if (n + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
+                tileK[idx] = key[kOffset + (n + local_id.y - past_sequence_length) * uniforms.K + w + local_id.x];
+              }`:`
+          if (n + local_id.y < uniforms.kv_sequence_length) {
+            tileK[idx] = key[kOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
+          }`}
+      ${h?`if (n + local_id.y < present_sequence_length) {
+        present_key[presentKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x] = tileK[idx];
+      }`:""}
+      }
+      workgroupBarrier();
+
+      for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
+          value += ${re}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
+      }
+
+      workgroupBarrier();
+    }
+
+    if (global_id.y < uniforms.M && global_id.x < total_sequence_length) {
+      let headOffset = workgroup_id.z * uniforms.M * uniforms.N;
+      let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
+      var sum: f32 = ${(()=>{switch(T){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: ${T}`)}})()};
+        output[outputIdx] = ${A.type.value} (sum * uniforms.alpha) + ${n?"attention_bias[outputIdx]":"0.0"};
+    }
+  }`};return{name:"AttentionProbs",shaderCache:{hint:`${T};${n!==void 0};${i!==void 0};${e}`,inputDependencies:C},getRunData:()=>({outputs:O,dispatchGroup:I,programUniforms:k}),getShaderSource:x}},Bu=(e,t,r,i,n,a,s=void 0,u=void 0)=>{let l=a+n.kvSequenceLength,d=n.nReps?n.nReps:1,f=n.vHiddenSize*d,h=e>1&&i,g=n.kvNumHeads?n.kvNumHeads:n.numHeads,w=h?[n.batchSize,g,l,n.headSize]:void 0,y=[n.batchSize,n.sequenceLength,f],v=12,T={x:Math.ceil(n.vHeadSize/v),y:Math.ceil(n.sequenceLength/v),z:n.batchSize*n.numHeads},$=[{type:12,data:n.sequenceLength},{type:12,data:l},{type:12,data:n.vHeadSize},{type:12,data:n.numHeads},{type:12,data:n.headSize},{type:12,data:f},{type:12,data:a},{type:12,data:n.kvSequenceLength},{type:12,data:d}],_=h&&i&&R.size(i.dims)>0,I=["type","type"];_&&I.push("type"),s&&I.push("type"),u&&I.push("type");let k=[{dims:y,dataType:t.dataType,gpuDataType:0}];h&&k.push({dims:w,dataType:t.dataType,gpuDataType:0});let E=C=>{let O=B("probs",t.dataType,t.dims),x=B("v",r.dataType,r.dims),D=[O,x];_&&D.push(B("past_value",i.dataType,i.dims));let L=s?B("seq_lens",s.dataType,s.dims):void 0;s&&D.push(L);let Y=u?B("total_sequence_length_input",u.dataType,u.dims):void 0;u&&D.push(Y);let q=[F("output",t.dataType,y)];h&&q.push(F("present_value",t.dataType,w));let j=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
+  const TILE_SIZE = ${v}u;
+  var<workgroup> tileQ: array<${O.type.value}, ${v*v}>;
+  var<workgroup> tileV: array<${O.type.value}, ${v*v}>;
+  ${C.registerUniforms(j).declareVariables(...D,...q)}
+  ${C.mainStart([v,v,1])}
+   let headIdx = workgroup_id.z % uniforms.num_heads;
+   let batchIdx = workgroup_id.z / uniforms.num_heads;
+   let kvHeadIdx = ${d===1?"headIdx":"headIdx / uniforms.n_reps"};
+   let kv_num_heads = ${d===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
+   let m = global_id.y;
+   let n = global_id.x;
+   let sequence_length = uniforms.M;
+   var total_sequence_length = uniforms.K;
+   ${Qr(L,Y,!0)}
+   let offsetA = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
+   let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; // kvHeadIdx is relative to the batch
+   ${_&&h?"let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;":""};
+   let vOffset = absKvHeadIdx * uniforms.N * uniforms.kv_sequence_length + n;
+   ${h?"let presentValueOffset = absKvHeadIdx * uniforms.N * uniforms.K + n;":""}
+   var value = ${O.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) {
+        var idx = TILE_SIZE * local_id.y + local_id.x;
+        ${_&&h?`
+        if (w + local_id.y < past_sequence_length) {
+          tileV[idx] = past_value[pastValueOffset + (w + local_id.y) * uniforms.N];
+        } else if (w + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
+          tileV[idx] = v[vOffset + (w + local_id.y - past_sequence_length) * uniforms.N];
+        }
+      `:`
+            if (w + local_id.y < uniforms.kv_sequence_length) {
+              tileV[idx] = v[vOffset + (w + local_id.y) * uniforms.N];
+            }`}
+        ${h?`
+            if (w + local_id.y < present_sequence_length) {
+          present_value[presentValueOffset + (w + local_id.y) * uniforms.N] = tileV[idx];
+        }`:""}
+      }
+     workgroupBarrier();
+     for (var k: u32 = 0u; k < TILE_SIZE && w+k < total_sequence_length; k++) {
+       value += tileQ[TILE_SIZE * local_id.y + k] * tileV[TILE_SIZE * k + local_id.x];
+     }
+     workgroupBarrier();
+   }
+
+   // we need to transpose output from BNSH_v to BSND_v
+   if (m < uniforms.M && n < uniforms.N) {
+     let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size
+       + headIdx * uniforms.N + n;
+     output[outputIdx] = value;
+   }
+  }`};return{name:"AttentionScore",shaderCache:{hint:`${i!==void 0};${e}`,inputDependencies:I},getRunData:()=>({outputs:k,dispatchGroup:T,programUniforms:$}),getShaderSource:E}},Tr=(e,t,r,i,n,a,s,u,l,d,f=void 0,h=void 0)=>{let g=Math.min(e.outputCount,1+(s?1:0)+(u?1:0)),w=g>1?d.pastSequenceLength:0,y=w+d.kvSequenceLength,v=l&&R.size(l.dims)>0?l:void 0,T=[t,r];g>1&&s&&R.size(s.dims)>0&&T.push(s),v&&T.push(v),f&&T.push(f),h&&T.push(h);let $=e.compute(Mu(g,t,r,s,v,d,w,f,h),{inputs:T,outputs:g>1?[-1,1]:[-1]})[0];e.compute(Ru($,d.batchSize,d.numHeads,w,d.sequenceLength,y,f,h),{inputs:f&&h?[$,f,h]:[$],outputs:[]});let _=[$,i];g>1&&u&&R.size(u.dims)>0&&_.push(u),f&&_.push(f),h&&_.push(h),e.compute(Bu(g,$,i,u,d,w,f,h),{inputs:_,outputs:g>1?[0,2]:[0]})},Nu=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],i=t.sequenceLength,n=t.inputHiddenSize,a=t.headSize,s=12,u={x:Math.ceil(t.headSize/s),y:Math.ceil(t.sequenceLength/s),z:t.batchSize*t.numHeads},l=[e.inputs[0],e.inputs[1],e.inputs[2]],d=[{type:12,data:i},{type:12,data:n},{type:12,data:a},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],f=h=>{let g=F("output_q",l[0].dataType,r),w=F("output_k",l[0].dataType,r),y=F("output_v",l[0].dataType,r),v=B("input",l[0].dataType,l[0].dims),T=B("weight",l[1].dataType,l[1].dims),$=B("bias",l[2].dataType,l[2].dims),_=v.type.storage,I=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`
+  const TILE_SIZE = ${s}u;
+  var<workgroup> tileInput: array<${_}, ${s*s}>;
+  var<workgroup> tileWeightQ: array<${_}, ${s*s}>;
+  var<workgroup> tileWeightK: array<${_}, ${s*s}>;
+  var<workgroup> tileWeightV: array<${_}, ${s*s}>;
+  ${h.registerUniforms(I).declareVariables(v,T,$,g,w,y)}
+  ${h.mainStart([s,s,1])}
+    let batchIndex = workgroup_id.z / uniforms.num_heads;
+    let headNumber = workgroup_id.z % uniforms.num_heads;
+    let m = global_id.y;
+    let n = global_id.x;
+
+    let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K;
+    let biasOffsetQ = headNumber * uniforms.head_size;
+    let biasOffsetK = uniforms.hidden_size + biasOffsetQ;
+    let biasOffsetV = uniforms.hidden_size + biasOffsetK;
+
+    var valueQ = ${_}(0);
+    var valueK = ${_}(0);
+    var valueV = ${_}(0);
+    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
+      if (m < uniforms.M && w + local_id.x < uniforms.K) {
+        tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x];
+      }
+      if (n < uniforms.N && w + local_id.y < uniforms.K) {
+        let offset = n + (w + local_id.y) * uniforms.ldb;
+        tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset];
+        tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset];
+        tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset];
+      }
+      workgroupBarrier();
+      for (var k: u32 = 0u; k<TILE_SIZE && w+k < uniforms.K; k++) {
+        let inputTileOffset = TILE_SIZE * local_id.y + k;
+        let weightTileOffset = TILE_SIZE * k + local_id.x;
+        valueQ += tileInput[inputTileOffset] * tileWeightQ[weightTileOffset];
+        valueK += tileInput[inputTileOffset] * tileWeightK[weightTileOffset];
+        valueV += tileInput[inputTileOffset] * tileWeightV[weightTileOffset];
+      }
+
+      workgroupBarrier();
+    }
+
+    let headOffset = (m * uniforms.N + n) % uniforms.head_size;
+    valueQ += bias[headOffset + biasOffsetQ];
+    valueK += bias[headOffset + biasOffsetK];
+    valueV += bias[headOffset + biasOffsetV];
+
+    let offset = workgroup_id.z * uniforms.M * uniforms.N;
+    if (m < uniforms.M && n < uniforms.N) {
+      let outputIdx = offset + m * uniforms.N + n;
+      output_q[outputIdx] = valueQ;
+      output_k[outputIdx] = valueK;
+      output_v[outputIdx] = valueV;
+    }
+  }`};return e.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:u,programUniforms:d}),getShaderSource:f},{inputs:l,outputs:[-1,-1,-1]})},wc=(e,t)=>{let r=Ou(e.inputs,t),[i,n,a]=Nu(e,r);return Tr(e,i,n,a,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r)}}),e0=P(()=>{"use strict";Le(),ee(),ie(),Te(),ne(),Du=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(i,n,a)=>{let s=n.length;if(s!==i.length)throw new Error(`${a}: num dimensions != ${s}`);n.forEach((u,l)=>{if(u!==i[l])throw new Error(`${a}: dim[${l}] do not match`)})};if(e[0].dims.length>1){let i=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,i,"Invalid input scale"),r(e[2].dims,i,"Invalid input B"),r(e[3].dims,i,"Invalid input mean"),r(e[4].dims,i,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},Pu=(e,t)=>{let{epsilon:r,spatial:i,format:n}=t,a=e[0].dims,s=i?xe(a[a.length-1]):1,u=n==="NHWC"&&a.length>1?s:1,l=R.size(a)/s,d=i,f=d?a.length:a,h=B("x",e[0].dataType,e[0].dims,s),g=B("scale",e[1].dataType,e[1].dims,u),w=B("bias",e[2].dataType,e[2].dims,u),y=B("inputMean",e[3].dataType,e[3].dims,u),v=B("inputVar",e[4].dataType,e[4].dims,u),T=F("y",e[0].dataType,f,s),$=()=>{let I="";if(i)I=`let cOffset = ${a.length===1?"0u":n==="NHWC"?`outputIndices[${a.length-1}] / ${s}`:"outputIndices[1]"};`;else if(n==="NCHW")I=`
+            ${T.indicesSet("outputIndices","0","0")}
+            let cOffset = ${T.indicesToOffset("outputIndices")};`;else{I=`var cIndices = ${g.type.indices}(0);
+                       cIndices[0] = outputIndices[${a.length-1}];`;for(let k=1;k<g.rank;k++)I+=`cIndices[${k}] = outputIndices[${k}];`;I+=`let cOffset = ${g.indicesToOffset("cIndices")};`}return I},_=I=>`
+  const epsilon = ${r};
+  ${I.registerUniform("outputSize","u32").declareVariables(h,g,w,y,v,T)}
+  ${I.mainStart()}
+  ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+    var outputIndices = ${T.offsetToIndices(`global_idx * ${s}`)};
+    ${$()}
+    let scale = ${g.getByOffset("cOffset")};
+    let bias = ${w.getByOffset("cOffset")};
+    let inputMean = ${y.getByOffset("cOffset")};
+    let inputVar = ${v.getByOffset("cOffset")};
+    let x = ${h.getByOffset("global_idx")};
+    let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;
+    ${T.setByOffset("global_idx","value")}
+  }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${i}_${s}`,inputDependencies:d?["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:d?[{type:12,data:l},...Z(a)]:[{type:12,data:l}]})}},Uu=e=>me(e),_c=(e,t)=>{let{inputs:r,outputCount:i}=e,n=Uu({...t,outputCount:i});if(be.webgpu.validateInputContent&&Du(r,n),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(Pu(r,n))}}),t0=P(()=>{"use strict";ie(),ne(),Wu=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")},Lu=e=>{let t=e[0].dims,r=e[0].dims[2],i=R.size(t)/4,n=e[0].dataType,a=B("input",n,t,4),s=B("bias",n,[r],4),u=B("residual",n,t,4),l=F("output",n,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:d=>`
+  const channels = ${r}u / 4;
+  ${d.declareVariables(a,s,u,l)}
+
+  ${d.mainStart()}
+    ${d.guardAgainstOutOfBoundsWorkgroupSizes(i)}
+    let value = ${a.getByOffset("global_idx")}
+      + ${s.getByOffset("global_idx % channels")} + ${u.getByOffset("global_idx")};
+    ${l.setByOffset("global_idx","value")}
+  }`}},bc=e=>{Wu(e.inputs),e.compute(Lu(e.inputs))}}),wa=P(()=>{"use strict";ee(),ie(),Te(),ne(),qu=(e,t,r,i,n,a,s)=>{let u=Math.ceil(t/4),l="";typeof n=="string"?l=`${n}(a)`:l=n("a");let d=B("inputData",r,[u],4),f=F("outputData",i,[u],4),h=[{name:"vec_size",type:"u32"}];return s&&h.push(...s),`
+      ${e.registerUniforms(h).declareVariables(d,f)}
+
+  ${a??""}
+
+  ${e.mainStart()}
+    ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
+
+    let a = ${d.getByOffset("global_idx")};
+    ${f.setByOffset("global_idx",l)}
+  }`},fe=(e,t,r,i,n,a=e.dataType,s,u)=>{let l=[{type:12,data:Math.ceil(R.size(e.dims)/4)}];return s&&l.push(...s),{name:t,shaderCache:{hint:n,inputDependencies:["type"]},getShaderSource:d=>qu(d,R.size(e.dims),e.dataType,a,r,i,u),getRunData:d=>({outputs:[{dims:e.dims,dataType:a}],dispatchGroup:{x:Math.ceil(R.size(d[0].dims)/64/4)},programUniforms:l})}},vc=e=>{e.compute(fe(e.inputs[0],"Abs","abs"))},$c=e=>{e.compute(fe(e.inputs[0],"Acos","acos"))},xc=e=>{e.compute(fe(e.inputs[0],"Acosh","acosh"))},Tc=e=>{e.compute(fe(e.inputs[0],"Asin","asin"))},Sc=e=>{e.compute(fe(e.inputs[0],"Asinh","asinh"))},kc=e=>{e.compute(fe(e.inputs[0],"Atan","atan"))},Ic=e=>{e.compute(fe(e.inputs[0],"Atanh","atanh"))},Ec=e=>me(e),Cc=(e,t)=>{let r;switch(t.to){case 10:r="vec4<f16>";break;case 1:r="vec4<f32>";break;case 12:r="vec4<u32>";break;case 6:r="vec4<i32>";break;case 9:r="vec4<bool>";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${t.to}`)}e.compute(fe(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},Vu=e=>{let t,r,i=e.length>=2&&e[1].data!==0,n=e.length>=3&&e[2].data!==0;switch(e[0].dataType){case 1:t=i?e[1].getFloat32Array()[0]:-34028234663852886e22,r=n?e[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:t=i?e[1].getUint16Array()[0]:64511,r=n?e[2].getUint16Array()[0]:31743;break;default:throw new Error("Unsupport data type")}return me({min:t,max:r})},zc=(e,t)=>{let r=t||Vu(e.inputs),i=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Clip",n=>`clamp(${n}, vec4<${i}>(uniforms.min), vec4<${i}>(uniforms.max))`,void 0,r.cacheKey,void 0,[{type:e.inputs[0].dataType,data:r.min},{type:e.inputs[0].dataType,data:r.max}],[{name:"min",type:i},{name:"max",type:i}]),{inputs:[0]})},Ac=e=>{e.compute(fe(e.inputs[0],"Ceil","ceil"))},Oc=e=>{e.compute(fe(e.inputs[0],"Cos","cos"))},Rc=e=>{e.compute(fe(e.inputs[0],"Cosh","cosh"))},_r=e=>me(e),Mc=(e,t)=>{let r=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Elu",i=>`elu_vf32(${i})`,`
+  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))},ii=(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<${e}> {
+  let absv = abs(v);
+  let x = 1.0 / (1.0 + r0 * absv);
+  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
+}`,Bc=e=>{let t=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,ii(t)))},Nc=e=>{e.compute(fe(e.inputs[0],"Exp","exp"))},Dc=e=>{e.compute(fe(e.inputs[0],"Floor","floor"))},Pc=e=>{let t=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,ii(t)))},Uc=(e,t)=>{let r=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"LeakyRelu",i=>`select(leaky_relu_alpha_ * ${i}, ${i}, ${i} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},Wc=e=>{e.compute(fe(e.inputs[0],"Not",t=>`!${t}`))},Lc=e=>{e.compute(fe(e.inputs[0],"Neg",t=>`-${t}`))},qc=e=>{e.compute(fe(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},Vc=e=>{let t=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},Gc=e=>{e.compute(fe(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},Fc=e=>me(e),Hc=(e,t)=>{let r=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"HardSigmoid",i=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${i} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},Kc=e=>{e.compute(fe(e.inputs[0],"Sin","sin"))},jc=e=>{e.compute(fe(e.inputs[0],"Sinh","sinh"))},Qc=e=>{e.compute(fe(e.inputs[0],"Sqrt","sqrt"))},Zc=e=>{e.compute(fe(e.inputs[0],"Tan","tan"))},pn=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Yc=e=>{e.compute(fe(e.inputs[0],"Tanh",pn))},Kn=(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 ${pn("v")};
+}
+`,jn=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,Xc=e=>{let t=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"FastGelu",jn,Kn(t),void 0,e.inputs[0].dataType))},Jc=(e,t)=>{let r=Ae(e.inputs[0].dataType);return e.compute(fe(e.inputs[0],"ThresholdedRelu",i=>`select(vec4<${r}>(0.0), ${i}, ${i} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},eh=e=>{e.compute(fe(e.inputs[0],"Log","log"))},Gu=(e,t)=>`
+const alpha = vec4<${e}>(${t});
+const one = ${e}(1.0);
+const zero = ${e}(0.0);
+
+fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
+  let v = x *alpha;
+  var x1 : vec4<${e}>;
+  for (var i = 0; i < 4; i = i + 1) {
+    if (v[i] >= zero) {
+      x1[i] = one / (one + exp(-v[i]));
+    } else {
+      x1[i] = one - one / (one + exp(v[i]));
+    }
+  }
+  return x * x1;
+}
+`,Fu=e=>`quick_gelu_impl(${e})`,th=(e,t)=>{let r=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"QuickGelu",Fu,Gu(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}}),r0=P(()=>{"use strict";ie(),ne(),wa(),Hu=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")},Ku=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=B("input",e[0].dataType,e[0].dims,4),i=B("bias",e[0].dataType,[e[0].dims[2]],4),n=F("output",e[0].dataType,t,4),a=R.size(t)/4,s=Ee(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)}}),getShaderSource:u=>`
+  const M_SQRT2 = sqrt(2.0);
+  const halfChannels = ${e[0].dims[2]/4/2}u;
+
+  ${u.declareVariables(r,i,n)}
+
+  ${ii(s)}
+
+  ${u.mainStart()}
+    ${u.guardAgainstOutOfBoundsWorkgroupSizes(a)}
+    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);
+
+    ${n.setByOffset("global_idx","valueLeft * geluRight")}
+  }`}},rh=e=>{Hu(e.inputs),e.compute(Ku(e.inputs))}}),i0=P(()=>{"use strict";ee(),ie(),ne(),ju=(e,t,r,i,n,a,s,u,l,d,f,h)=>{let g,w;typeof u=="string"?g=w=(_,I)=>`${u}((${_}),(${I}))`:typeof u=="function"?g=w=u:(g=u.scalar,w=u.vector);let y=F("outputData",f,i.length,4),v=B("aData",l,t.length,4),T=B("bData",d,r.length,4),$;if(n)if(a){let _=R.size(t)===1,I=R.size(r)===1,k=t.length>0&&t[t.length-1]%4===0,E=r.length>0&&r[r.length-1]%4===0;_||I?$=y.setByOffset("global_idx",w(_?`${v.type.value}(${v.getByOffset("0")}.x)`:v.getByOffset("global_idx"),I?`${T.type.value}(${T.getByOffset("0")}.x)`:T.getByOffset("global_idx"))):$=`
+            let outputIndices = ${y.offsetToIndices("global_idx * 4u")};
+            let offsetA = ${v.broadcastedIndicesToOffset("outputIndices",y)};
+            let offsetB = ${T.broadcastedIndicesToOffset("outputIndices",y)};
+            ${y.setByOffset("global_idx",w(s||k?v.getByOffset("offsetA / 4u"):`${v.type.value}(${v.getByOffset("offsetA / 4u")}[offsetA % 4u])`,s||E?T.getByOffset("offsetB / 4u"):`${T.type.value}(${T.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
+          `}else $=y.setByOffset("global_idx",w(v.getByOffset("global_idx"),T.getByOffset("global_idx")));else{if(!a)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let _=(I,k,E="")=>{let C=`aData[indexA${k}][componentA${k}]`,O=`bData[indexB${k}][componentB${k}]`;return`
+            let outputIndices${k} = ${y.offsetToIndices(`global_idx * 4u + ${k}u`)};
+            let offsetA${k} = ${v.broadcastedIndicesToOffset(`outputIndices${k}`,y)};
+            let offsetB${k} = ${T.broadcastedIndicesToOffset(`outputIndices${k}`,y)};
+            let indexA${k} = offsetA${k} / 4u;
+            let indexB${k} = offsetB${k} / 4u;
+            let componentA${k} = offsetA${k} % 4u;
+            let componentB${k} = offsetB${k} % 4u;
+            ${I}[${k}] = ${E}(${g(C,O)});
+          `};f===9?$=`
+            var data = vec4<u32>(0);
+            ${_("data",0,"u32")}
+            ${_("data",1,"u32")}
+            ${_("data",2,"u32")}
+            ${_("data",3,"u32")}
+            outputData[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:$=`
+            ${_("outputData[global_idx]",0)}
+            ${_("outputData[global_idx]",1)}
+            ${_("outputData[global_idx]",2)}
+            ${_("outputData[global_idx]",3)}
+          `}return`
+        ${e.registerUniform("vec_size","u32").declareVariables(v,T,y)}
+
+        ${h??""}
+
+        ${e.mainStart()}
+        ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
+        ${$}
+      }`},Qu=(e,t,r,i,n,a,s=r.dataType)=>{let u=r.dims.map(Number),l=i.dims.map(Number),d=!R.areEqual(u,l),f=u,h=R.size(u),g=!1,w=!1,y=[d];if(d){let v=Qt.calcShape(u,l,!1);if(!v)throw new Error("Can't perform binary op on the given tensors");f=v.slice(),h=R.size(f);let T=R.size(u)===1,$=R.size(l)===1,_=u.length>0&&u[u.length-1]%4===0,I=l.length>0&&l[l.length-1]%4===0;y.push(T),y.push($),y.push(_),y.push(I);let k=1;for(let E=1;E<f.length;E++){let C=u[u.length-E],O=l[l.length-E];if(C===O)k*=C;else break}k%4===0?(w=!0,g=!0):(T||$||_||I)&&(g=!0)}else g=!0;return y.push(g),{name:e,shaderCache:{hint:t+y.map(v=>v.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:v=>ju(v,u,l,f,g,d,w,n,r.dataType,i.dataType,s,a),getRunData:()=>({outputs:[{dims:f,dataType:s}],dispatchGroup:{x:Math.ceil(h/64/4)},programUniforms:[{type:12,data:Math.ceil(R.size(f)/4)},...Z(u,l,f)]})}},Qe=(e,t,r,i,n,a)=>{e.compute(Qu(t,n??"",e.inputs[0],e.inputs[1],r,i,a))},ih=e=>{Qe(e,"Add",(t,r)=>`${t}+${r}`)},nh=e=>{Qe(e,"Div",(t,r)=>`${t}/${r}`)},ah=e=>{Qe(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},sh=e=>{Qe(e,"Mul",(t,r)=>`${t}*${r}`)},oh=e=>{let t=B("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;Qe(e,"Pow",{scalar:(r,i)=>`pow_custom(${r},${i})`,vector:(r,i)=>`pow_vector_custom(${r},${i})`},`
+    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));
+    }
+      `)},uh=e=>{Qe(e,"Sub",(t,r)=>`${t}-${r}`)},lh=e=>{Qe(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},dh=e=>{Qe(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},ph=e=>{Qe(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},ch=e=>{Qe(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}}),n0=P(()=>{"use strict";ee(),ie(),Te(),ne(),Zu=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,i=e[r],n=i.dataType,a=i.dims.length;e.forEach((s,u)=>{if(u!==r){if(s.dataType!==n)throw new Error("input tensors should be one type");if(s.dims.length!==a)throw new Error("input tensors should have the same shape");s.dims.forEach((l,d)=>{if(d!==t&&l!==i.dims[d])throw new Error("non concat dimensions must match")})}})},Yu=(e,t)=>`
+  fn calculateInputIndex(index: u32) -> u32 {
+    let sizeInConcatAxis = array<u32, ${e}u>(${t});
+    for (var i: u32 = 0u; i < ${e}; i += 1u ) {
+      if (index < sizeInConcatAxis[i]) {
+        return i;
+      }
+    }
+    return ${e}u;
+  }`,Xu=(e,t)=>{let r=e.length,i=[];for(let n=0;n<r;++n){let a=t.setByOffset("global_idx",e[n].getByIndices("indices"));r===1?i.push(a):n===0?i.push(`if (inputIndex == ${n}u) { ${a} }`):n===r-1?i.push(`else { ${a} }`):i.push(`else if (inputIndex == ${n}) { ${a} }`)}return i.join(`
+`)},Ju=(e,t,r,i)=>{let n=R.size(r),a=new Array(e.length),s=new Array(e.length),u=0,l=[],d=[],f=[{type:12,data:n}];for(let v=0;v<e.length;++v)u+=e[v].dims[t],a[v]=u,d.push(e[v].dims.length),s[v]=B(`input${v}`,i,d[v]),l.push("rank"),f.push({type:12,data:a[v]});for(let v=0;v<e.length;++v)f.push(...Z(e[v].dims));f.push(...Z(r));let h=F("output",i,r.length),g=h.indicesGet("indices",t),w=Array.from(Array(a.length).keys()).map(v=>`uniforms.sizeInConcatAxis${v}`).join(","),y=v=>`
+
+  ${(()=>{v.registerUniform("outputSize","u32");for(let T=0;T<e.length;T++)v.registerUniform(`sizeInConcatAxis${T}`,"u32");return v.declareVariables(...s,h)})()}
+
+  ${Yu(a.length,w)}
+
+  ${v.mainStart()}
+    ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+
+    var indices = ${h.offsetToIndices("global_idx")};
+
+    let inputIndex = calculateInputIndex(${g});
+    if (inputIndex != 0u) {
+      let sizeInConcatAxis = array<u32, ${a.length}u>(${w});
+      ${g} -= sizeInConcatAxis[inputIndex - 1u];
+    }
+
+    ${Xu(s,h)}
+  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:r,dataType:i}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:f}),getShaderSource:y}},hh=(e,t)=>{let r=e.inputs,i=r[0].dims,n=R.normalizeAxis(t.axis,i.length);Zu(r,n);let a=i.slice();a[n]=r.reduce((u,l)=>u+(l.dims.length>n?l.dims[n]:0),0);let s=r.filter(u=>R.size(u.dims)>0);e.compute(Ju(s,n,a,r[0].dataType),{inputs:s})},fh=e=>me({axis:e.axis})}),Wt=P(()=>{"use strict";ee(),ie(),Nt=(e,t,r="f32")=>{switch(e.activation){case"Relu":return`value = max(value, ${t}(0.0));`;case"Sigmoid":return`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${t}(${r}(uniforms.clip_min)), ${t}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${t}(0.0), min(${t}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case"Tanh":return`let e2x = exp(-2.0 * abs(value));
+              value = sign(value) * (1.0 - e2x) / (1.0 + e2x);
+        `;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},Dt=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},Pt=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},_a=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,i]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:i}}else if(t==="Clip"){let[r,i]=e?.activation_params||[Wp,Lp];return{activation:t,clipMax:i,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}}),ba=P(()=>{"use strict";ze=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},mh=e=>`
+      ${e?"value = value + getBiasByOutputCoords(coords);":""}
+      `}),a0=P(()=>{"use strict";gh=e=>`
+fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
+  return dot(coords, vec4<i32>(
+      shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
+}
+fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
+  return dot(coords, vec4<i32>(
+    i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));
+}
+`}),$a=P(()=>{"use strict";ee(),ie(),ne(),Wt(),vr=(e,t,r,i,n)=>{let a=i-r;return`
+      ${Array.from({length:r}).map((s,u)=>`
+      if (${K(t.shape,u,t.rank)} != 1) {
+        ${t.indicesSet(e,u,K(n,u+a,i))}
+      } else {
+        ${t.indicesSet(e,u,0)}
+      }`).join("")}
+`},va=(e,t,r,i,n=!1,a)=>{let s=e[0].dims,u=e[1].dims,l=s[s.length-2],d=u[u.length-1],f=s[s.length-1],h=xe(d),g=xe(f),w=xe(l),y=R.size(r)/h/w,v=e.length>2,T=i?i.slice(0,-2):r.slice(0,-2),$=[R.size(T),l,d],_=[{type:12,data:y},{type:12,data:l},{type:12,data:d},{type:12,data:f}];Dt(t,_),_.push(...Z(T,s,u)),v&&_.push(...Z(e[2].dims)),_.push(...Z($));let I=k=>{let E=ma("batch_dims",e[0].dataType,T.length),C=B("a",e[0].dataType,s.length,g),O=B("b",e[1].dataType,u.length,h),x=F("output",e[0].dataType,$.length,h),D=Ee(x.type.tensor),L=Nt(t,x.type.value,D),Y=[C,O],q="";if(v){let A=n?h:1;Y.push(B("bias",e[2].dataType,e[2].dims.length,A)),q=`${n?`value += bias[col / ${A}];`:`value += ${x.type.value}(bias[row + i]);`}`}let j=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Pt(t,j);let ue=()=>{let A=`var a_data: ${C.type.value};`;for(let U=0;U<g;U++)A+=`
+              let b_data${U} = b[(b_offset + (k + ${U}) * uniforms.N + col) / ${h}];`;for(let U=0;U<w;U++){A+=`a_data = a[(a_offset + (row + ${U}) * uniforms.K + k) / ${g}];`;for(let re=0;re<g;re++)A+=`
+            values[${U}] = fma(${O.type.value}(a_data${g===1?"":`[${re}]`}), b_data${re}, values[${U}]);
+`}return A};return`
+  ${k.registerUniforms(j).registerInternalVariables(E).declareVariables(...Y,x)}
+  ${k.mainStart()}
+    ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+    let col = (global_idx % (uniforms.N / ${h})) * ${h};
+    var index1 = global_idx / (uniforms.N / ${h});
+    let stride1 = uniforms.M / ${w};
+    let row = (index1 % stride1) * ${w};
+    let batch = index1 / stride1;
+
+    ${r.length===2?"":`let batch_indices = ${E.offsetToIndices("batch")};`}
+
+    var a_indices: ${C.type.indices};
+    ${vr("a_indices",C,C.rank-2,E.rank,"batch_indices")}
+    ${C.indicesSet("a_indices",C.rank-2,0)}
+    ${C.indicesSet("a_indices",C.rank-1,0)}
+    let a_offset = ${C.indicesToOffset("a_indices")};
+
+    var b_indices: ${O.type.indices};
+    ${vr("b_indices",O,O.rank-2,E.rank,"batch_indices")}
+    ${O.indicesSet("b_indices",O.rank-2,0)}
+    ${O.indicesSet("b_indices",O.rank-1,0)}
+    let b_offset = ${O.indicesToOffset("b_indices")};
+    var values: array<${x.type.value}, ${w}>;
+    for (var k: u32 = 0u; k < uniforms.K; k = k + ${g}) {
+      ${ue()}
+    }
+    for (var i = 0u; i < ${w}u; i++) {
+      var value = values[i];
+      ${q}
+      ${L}
+      let cur_indices = ${x.type.indices}(batch, row + i, col);
+      let offset = ${x.indicesToOffset("cur_indices")};
+      ${x.setByOffset(`offset / ${h}`,"value")};
+    }
+  }
+  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${h};${g};${w};${n}`,inputDependencies:v?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:a?a(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:_}),getShaderSource:I}}}),xa=P(()=>{"use strict";ee(),ie(),ne(),Wt(),$a(),ba(),el=(e,t)=>e?`
+        mm_Asub[inputRow][inputCol] = mm_readA(batch,
+          kStart + inputRow,
+          globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});
+        `:`
+        mm_Asub[inputRow][inputCol] = mm_readA(batch,
+          globalRow + innerRow,
+          kStart / innerElementSize + inputCol${t?", batchIndices":""});
+        `,tl=(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];"}
+        }`,Qn=(e,t,r="f32",i,n=!1,a=32,s=!1,u=32)=>{let l=t[1]*e[1],d=t[0]*e[0],f=n?l:a,h=n?a:l,g=f/t[0],w=a/t[1];if(!((n&&g===4&&e[1]===4||!n&&(g===3||g===4))&&f%t[0]===0&&a%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${n} is true, innerElementSize ${g} and workPerThread[1] ${e[1]} must be 4.
+      Otherwise, innerElementSize ${g} must be 3 or 4.
+  tileAWidth ${f} must be divisible by workgroupSize[0]${t[0]}. tileInner ${a} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`
+var<workgroup> mm_Asub: array<array<vec${g}<${r}>, ${f/g}>, ${h}>;
+var<workgroup> mm_Bsub: array<array<vec4<${r}>, ${d/e[0]}>, ${a}>;
+
+const rowPerThread = ${e[1]};
+const colPerThread = ${e[0]};
+const innerElementSize = ${g};
+const tileInner = ${a};
+
+@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})
+fn main(@builtin(local_invocation_id) localId : vec3<u32>,
+        @builtin(global_invocation_id) globalId : vec3<u32>,
+        @builtin(workgroup_id) workgroupId : vec3<u32>) {
+  let localRow = i32(localId.y);
+  let tileRow = localRow * rowPerThread;
+  let tileCol = i32(localId.x);
+
+  let globalRow =i32(globalId.y) * rowPerThread;
+  let globalCol = i32(globalId.x);
+  let batch = ${s?"0":"i32(globalId.z)"};
+  ${i?`let batchIndices = ${i.offsetToIndices("u32(batch)")};`:""}
+  let globalRowStart = i32(workgroupId.y) * ${l};
+
+  let num_tiles = ${s?`${Math.ceil(u/a)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
+  var kStart = ${s?`i32(globalId.z) * ${u}`:"0"};
+
+  var acc: array<vec4<${r}>, rowPerThread>;
+
+  // Loop over shared dimension.
+  let tileRowB = localRow * ${w};
+  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;
+          ${el(n,i)}
+      }
+
+      // Load one tile of B into local memory.
+      for (var innerRow = 0; innerRow < ${w}; innerRow = innerRow + 1) {
+          let inputRow = tileRowB + innerRow;
+          let inputCol = tileCol;
+          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${i?", 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];
+          ${g===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}
+
+          ${tl(n,g)}
+      }
+
+      workgroupBarrier();
+  }
+
+  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
+      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
+  }
+}`},cn=(e,t)=>e?`
+            mm_Asub[inputRow][inputCol] = mm_readA(batch,
+              kStart + inputRow,
+              globalRowStart + inputCol${t?", batchIndices":""});
+            `:`
+            mm_Asub[inputRow][inputCol] = mm_readA(batch,
+              globalRowStart + inputRow,
+              kStart + inputCol${t?", batchIndices":""});
+            `,rl=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",Zn=(e,t,r="f32",i,n=!1,a=32,s=!1,u=32,l=!1)=>{let d=e[1]*t[1],f=e[0]*t[0],h=n?d:a,g=n?a:d;if(!(g%t[1]===0&&h%t[0]===0&&a%t[1]===0))throw new Error(`tileAHight ${g} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${h} must be divisible by workgroupSize[0]${t[0]}, tileInner ${a} must be divisible by workgroupSize[1]${t[1]}`);let w=g/t[1],y=h/t[0],v=a/t[1],T=l?`
+    let localRow = i32(localId.y);
+    let localCol = i32(localId.x);
+    let globalRowStart = i32(workgroupId.y) * ${d};
+    let globalColStart = i32(workgroupId.x) * ${f};
+
+    // 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 < ${g}; inputRow = inputRow + ${t[1]}) {
+        for (var inputCol = localCol; inputCol < ${h}; inputCol = inputCol + ${t[0]}) {
+          ${cn(n,i)}
+        }
+      }
+      // Load one tile of B into local memory.
+      for (var inputRow = localRow; inputRow < ${a}; inputRow = inputRow + ${t[1]}) {
+            for (var inputCol = localCol; inputCol < ${f}; inputCol = inputCol + ${t[0]}) {
+          mm_Bsub[inputRow][inputCol] = mm_readB(batch,
+            kStart + inputRow,
+            globalColStart + inputCol${i?", 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 = ${n?`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) * ${d};
+
+let tileRowA = i32(localId.y) * ${w};
+let tileColA = i32(localId.x) * ${y};
+let tileRowB = i32(localId.y) * ${v};
+// 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 < ${w}; innerRow = innerRow + 1) {
+    for (var innerCol = 0; innerCol < ${y}; innerCol = innerCol + 1) {
+      let inputRow = tileRowA + innerRow;
+      let inputCol = tileColA + innerCol;
+      ${cn(n,i)}
+    }
+  }
+
+  // Load one tile of B into local memory.
+  for (var innerRow = 0; innerRow < ${v}; innerRow = innerRow + 1) {
+    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
+      let inputRow = tileRowB + innerRow;
+      let inputCol = tileCol + innerCol;
+      mm_Bsub[inputRow][inputCol] = mm_readB(batch,
+        kStart + inputRow,
+        globalCol + innerCol${i?", 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) {
+      ${rl(n)}
+      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
+        acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
+      }
+    }
+  }
+
+  workgroupBarrier();
+}
+
+for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
+  for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
+    mm_write(batch, globalRow + innerRow, globalCol + innerCol,
+        acc[innerRow][innerCol]);
+  }
+}
+`;return`
+  var<workgroup> mm_Asub : array<array<${r}, ${h}>, ${g}>;
+  var<workgroup> mm_Bsub : array<array<${r}, ${f}>, ${a}>;
+  const rowPerThread = ${e[1]};
+  const colPerThread = ${e[0]};
+  const tileInner = ${a};
+
+@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})
+fn main(@builtin(local_invocation_id) localId : vec3<u32>,
+        @builtin(global_invocation_id) globalId : vec3<u32>,
+        @builtin(workgroup_id) workgroupId : vec3<u32>) {
+    let batch = ${s?"0":"i32(globalId.z)"};
+    ${i?`let batchIndices = ${i.offsetToIndices("u32(batch)")};`:""}
+    let num_tiles = ${s?`${Math.ceil(u/a)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
+    var kStart = ${s?`i32(globalId.z) * ${u}`:"0"};
+
+    var acc : array<array<${r}, colPerThread>, rowPerThread>;
+    ${T}
+  }
+`},il=(e,t,r,i,n=!1)=>{let[a,s,u,l]=i,d=Ee(i[0].type.tensor);return`
+    fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${a.type.indices}) -> ${ze(e,d)} {
+      var value = ${ze(e,d)}(0.0);
+      let col = colIn * ${e};
+      if(row < uniforms.dim_a_outer && col < uniforms.dim_inner)
+      {
+        var aIndices: ${s.type.indices};
+        ${vr("aIndices",s,s.rank-2,a.rank,"batchIndices")}
+        ${s.indicesSet("aIndices",s.rank-2,"u32(row)")}
+        ${s.indicesSet("aIndices",s.rank-1,"u32(colIn)")}
+        value = ${s.getByIndices("aIndices")};
+      }
+      return value;
+    }
+
+    fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${a.type.indices}) -> ${ze(e,d)} {
+      var value = ${ze(e,d)}(0.0);
+      let col = colIn * ${e};
+      if(row < uniforms.dim_inner && col < uniforms.dim_b_outer)
+      {
+        var bIndices: ${u.type.indices};
+        ${vr("bIndices",u,u.rank-2,a.rank,"batchIndices")}
+        ${u.indicesSet("bIndices",u.rank-2,"u32(row)")}
+        ${u.indicesSet("bIndices",u.rank-1,"u32(colIn)")}
+        value = ${u.getByIndices("bIndices")};
+      }
+      return value;
+    }
+
+    fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${ze(e,d)}) {
+      let col = colIn * ${e};
+      if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {
+        var value = valueIn;
+        let coords = vec3<i32>(batch, row, colIn);
+        ${t?`value = value + ${n?"bias[colIn]":`${ze(e,d)}(bias[row])`};`:""}
+        ${r}
+        ${l.setByIndices("vec3<u32>(coords)","value")}
+      }
+    }
+    `},ui=(e,t,r,i,n=!1,a)=>{let s=e[0].dims,u=e[1].dims,l=s.slice(0,-2),d=u.slice(0,-2),f=i?i.slice(0,-2):r.slice(0,-2),h=R.size(f),g=s[s.length-2],w=s[s.length-1],y=u[u.length-1],v=w%4===0&&y%4===0,T=g<=8?[4,1,1]:[4,4,1],$=[8,8,1],_=[Math.ceil(y/$[0]/T[0]),Math.ceil(g/$[1]/T[1]),Math.ceil(h/$[2]/T[2])],I=v?4:1,k=[...l,g,w/I],E=k.length,C=[...d,w,y/I],O=C.length,x=[h,g,y/I],D=[{type:6,data:g},{type:6,data:y},{type:6,data:w}];Dt(t,D),D.push(...Z(f,k,C));let L=["rank","rank"],Y=e.length>2;Y&&(D.push(...Z(e[2].dims)),L.push("rank")),D.push(...Z(x));let q=j=>{let ue=f.length,A=ma("batchDims",e[0].dataType,ue,1),U=Ee(e[0].dataType),re=B("a",e[0].dataType,E,I),J=B("b",e[1].dataType,O,I),H=F("result",e[0].dataType,x.length,I),pe=[re,J];if(Y){let Ie=n?I:1;pe.push(B("bias",e[2].dataType,e[2].dims.length,Ie))}let N=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];Pt(t,N);let V=Ee(H.type.tensor),te=Nt(t,H.type.value,V),ce=il(I,Y,te,[A,re,J,H],n);return`
+  ${j.registerUniforms(N).registerInternalVariables(A).declareVariables(...pe,H)}
+  ${ce}
+  ${v?Qn(T,$,U,A):Zn(T,$,U,A)}
+                   `};return{name:"MatMul",shaderCache:{hint:`${T};${t.activation};${v};${n}`,inputDependencies:L},getRunData:()=>({outputs:[{dims:a?a(r):r,dataType:e[0].dataType}],dispatchGroup:{x:_[0],y:_[1],z:_[2]},programUniforms:D}),getShaderSource:q}}}),s0=P(()=>{"use strict";ee(),lt(),ne(),Wt(),ba(),a0(),xa(),nl=(e,t,r,i,n=!1,a,s=4,u=4,l=4,d="f32")=>{let f=D=>{switch(D){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${d}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${D} is not supported.`)}},h=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 ${D} is not supported.`)}},g=e?`
+    let coord = vec4<i32>(batch, xRow, xCol, xCh);
+    `:`
+    let coord = vec4<i32>(batch, xCh, xRow, xCol);
+    `,w=e?`
+    let coords = vec4<i32>(
+      batch,
+      row / outWidth,
+      row % outWidth,
+      col);
+    `:`
+    let coords = vec4<i32>(
+      batch,
+      row,
+      col / outWidth,
+      col % outWidth);
+    `,y=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",v=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",T=e?"row":"col",$=e?"col":"row",_=`
+    let inChannels = i32(uniforms.w_shape[2]);
+    let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
+    let outRow = ${T} / outWidth;
+    let outCol = ${T} % outWidth;
+
+    let WRow = ${$} / (i32(uniforms.w_shape[1]) * inChannels);
+    let WCol = ${$} / 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 = ${$} % inChannels;
+    var resData = ${ze(s,d)}(0.0);
+    // The bounds checking is always needed since we use it to pad zero for
+    // the 'same' padding type.
+    if (xRow >= 0 && xRow < ${y} && xCol >= 0 && xCol < ${v}) {
+      ${g}
+      let xIndex = getIndexFromCoords4D(coord, vec4<i32>(uniforms.x_shape));
+      ${f(s)}
+    }
+    return resData;`,I=e?t&&i?`
+    let col = colIn * ${s};
+    ${_}`:`
+    let col = colIn * ${s};
+    if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {
+      ${_}
+    }
+    return ${ze(s,d)}(0.0);`:i&&r?`
+    let col = colIn * ${s};
+    ${_}`:`
+    let col = colIn * ${s};
+    if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {
+      ${_}
+    }
+    return ${ze(s,d)}(0.0);`,k=e?i&&r?h(u):`
+    let col = colIn * ${u};
+    if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {
+      ${h(u)}
+    }
+    return ${ze(u,d)}(0.0);`:`
+    let col = colIn * ${u};
+    if (row < uniforms.dim_inner && col < uniforms.dim_a_outer) {
+      ${h(u)}
+    }
+    return ${ze(u,d)}(0.0);`,E=ze(l,d),C=ze(e?s:u,d),O=ze(e?u:s,d),x=Nt(a,E,d);return`
+    fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${C} {
+      ${e?I:k}
+    }
+
+    fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${O} {
+      ${e?k:I}
+    }
+
+    fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${E}) {
+      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])"};
+      ${w}
+      ${mh(n)}
+      ${x}
+      setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
+      }
+    }`},yh=(e,t,r,i,n,a,s,u,l)=>{let d=t.format==="NHWC",f=d?e[0].dims[3]:e[0].dims[1],h=r[0],g=d?r[2]:r[3],w=d?r[1]:r[2],y=d?r[3]:r[1],v=d&&(f%4===0||f%3===0)&&y%4===0,T=d?y:g*w,$=d?g*w:y,_=[8,8,1],I=i<=8?[4,1,1]:[4,4,1],k=[Math.ceil(T/_[0]/I[0]),Math.ceil($/_[1]/I[1]),Math.ceil(h/_[2]/I[2])];de("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${k}`);let E=v?d&&f%4!==0?3:4:1,C=_[1]*I[1],O=_[0]*I[0],x=Math.max(_[0]*E,_[1]),D=i%C===0,L=n%O===0,Y=a%x===0,q=v?[E,4,4]:[1,1,1],j=[{type:6,data:i},{type:6,data:n},{type:6,data:a},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];Dt(t,j),j.push(...Z(e[0].dims,e[1].dims));let ue=["rank","rank"];s&&(j.push(...Z(e[2].dims)),ue.push("rank")),j.push(...Z(r));let A=U=>{let re=[{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}];Pt(t,re);let J=v?4:1,H=Ee(e[0].dataType),pe=`
+      fn setOutputAtIndex(flatIndex : i32, value : ${v?`vec4<${H}>`:H}) {
+        result[flatIndex] = ${v?`vec4<${H}>`:H}(value);
+      }
+      fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${v?`vec4<${H}>`:H}) {
+        let flatIndex = getOutputIndexFromCoords(vec4<i32>(d0, d1, d2, d3));
+        setOutputAtIndex(flatIndex ${v?"/ 4":""}, value);
+      }`,N=B("x",e[0].dataType,e[0].dims.length,E===3?1:E),V=B("w",e[1].dataType,e[1].dims.length,J),te=[N,V],ce=F("result",e[0].dataType,r.length,J);if(s){let Ie=B("bias",e[2].dataType,e[2].dims.length,J);te.push(Ie),pe+=`
+        fn getBiasByOutputCoords(coords : vec4<i32>) -> ${v?`vec4<${H}>`:H} {
+          return bias[coords.${d?"w":"y"}${v?"/ 4":""}];
+        }`}return`
+        ${gh("uniforms.result_strides")}
+        //struct Uniforms { xShape : vec4<i32>, wShape : vec4<i32>, outShape : vec4<i32>,
+        //  outShapeStrides: vec3<i32>, filterDims : vec2<i32>, pad : vec2<i32>, stride : vec2<i32>,
+        //  dilation : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32 };
+        ${U.registerUniforms(re).declareVariables(...te,ce)}
+        ${pe}
+        ${nl(d,D,L,Y,s,t,q[0],q[1],q[2],H)}
+        ${v?Qn(I,_,H,void 0,!d,x):Zn(I,_,H,void 0,!d,x,!1,void 0,u)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${E};${v};${D};${L};${Y};${C};${O};${x}`,inputDependencies:ue},getRunData:()=>({outputs:[{dims:l?l(r):r,dataType:e[0].dataType}],dispatchGroup:{x:k[0],y:k[1],z:k[2]},programUniforms:j}),getShaderSource:A}}}),o0=P(()=>{"use strict";ee(),lt(),ie(),ne(),Wt(),ba(),al=e=>{let t=1;for(let r=0;r<e.length;r++)t*=e[r];return t},hn=e=>typeof e=="number"?[e,e,e]:e,pr=(e,t)=>t<=1?e:e+(e-1)*(t-1),sl=(e,t,r,i=1)=>{let n=pr(t,i);return Math.floor((e[0]*(r-1)-r+n)/2)},fn=(e,t,r,i,n)=>{n==null&&(n=sl(e,t[0],i[0]));let a=[0,0,0,r];for(let s=0;s<3;s++)e[s]+2*n>=t[s]&&(a[s]=Math.trunc((e[s]-t[s]+2*n)/i[s]+1));return a},ol=(e,t,r,i,n,a,s,u,l,d)=>{let f,h,g,w;if(e==="VALID"&&(e=0),typeof e=="number"){f={top:e,bottom:e,left:e,right:e,front:e,back:e};let y=fn([t,r,i,1],[u,l,d],1,[n,a,s],e);h=y[0],g=y[1],w=y[2]}else if(Array.isArray(e)){if(!e.every((v,T,$)=>v===$[0]))throw Error(`Unsupported padding parameter: ${e}`);f={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let y=fn([t,r,i,1],[u,l,d],1,[n,a,s],e[0]);h=y[0],g=y[1],w=y[2]}else if(e==="SAME_UPPER"){h=Math.ceil(t/n),g=Math.ceil(r/a),w=Math.ceil(i/s);let y=(h-1)*n+u-t,v=(g-1)*a+l-r,T=(w-1)*s+d-i,$=Math.floor(y/2),_=y-$,I=Math.floor(v/2),k=v-I,E=Math.floor(T/2),C=T-E;f={top:I,bottom:k,left:E,right:C,front:$,back:_}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:f,outDepth:h,outHeight:g,outWidth:w}},wh=(e,t,r,i,n,a=!1,s="channelsLast")=>{let u,l,d,f,h;if(s==="channelsLast")[u,l,d,f,h]=e;else if(s==="channelsFirst")[u,h,l,d,f]=e;else throw new Error(`Unknown dataFormat ${s}`);let[g,,w,y,v]=t,[T,$,_]=hn(r),[I,k,E]=hn(i),C=pr(w,I),O=pr(y,k),x=pr(v,E),{padInfo:D,outDepth:L,outHeight:Y,outWidth:q}=ol(n,l,d,f,T,$,_,C,O,x),j=a?g*h:g,ue=[0,0,0,0,0];return s==="channelsFirst"?ue=[u,j,L,Y,q]:s==="channelsLast"&&(ue=[u,L,Y,q,j]),{batchSize:u,dataFormat:s,inDepth:l,inHeight:d,inWidth:f,inChannels:h,outDepth:L,outHeight:Y,outWidth:q,outChannels:j,padInfo:D,strideDepth:T,strideHeight:$,strideWidth:_,filterDepth:w,filterHeight:y,filterWidth:v,effectiveFilterDepth:C,effectiveFilterHeight:O,effectiveFilterWidth:x,dilationDepth:I,dilationHeight:k,dilationWidth:E,inShape:e,outShape:ue,filterShape:t}},_h=(e,t,r,i,n,a)=>{let s=a==="channelsLast",u=s?e[0].dims[3]:e[0].dims[1],l=!1,d=[64,1,1],f={x:r.map((_,I)=>I)},h=[Math.ceil(al(f.x.map(_=>r[_]))/d[0]),1,1];de("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${h}`);let g=l?s&&u%4!==0?3:4:1,w=R.size(r),y=[{type:12,data:w},{type:12,data:i},{type:12,data:n},{type:12,data:t.strides},{type:12,data:t.dilations}];Dt(t,y),y.push(...Z(e[0].dims,e[1].dims));let v=["rank","rank"],T=e.length===3;T&&(y.push(...Z(e[2].dims)),v.push("rank")),y.push(...Z(r));let $=_=>{let I=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:i.length},{name:"pads",type:"u32",length:n.length},{name:"strides",type:"u32",length:t.strides.length},{name:"dilations",type:"u32",length:t.dilations.length}];Pt(t,I);let k=l?4:1,E=Ee(e[0].dataType),C=B("x",e[0].dataType,e[0].dims.length,g===3?1:g),O=B("W",e[1].dataType,e[1].dims.length,k),x=[C,O],D=F("result",e[0].dataType,r.length,k),L="";if(T){let j=B("bias",e[2].dataType,e[2].dims.length,k);x.push(j),L+=`
+        fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${l?`vec4<${E}>`:E} {
+          return bias[${s?K("coords",4,5):K("coords",1,5)}${l?"/ 4":""}];
+        }`}let Y=ze(g,E),q=Nt(t,Y,E);return`
+            ${L}
+            fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
+              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
+              return ${C.getByIndices("aIndices")};
+            }
+            fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
+              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
+              return ${O.getByIndices("aIndices")};
+            }
+          ${_.registerUniforms(I).declareVariables(...x,D)}
+          ${_.mainStart()}
+          ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+              let coords = ${D.offsetToIndices("global_idx")};
+              let batch = ${K("coords",0,C.rank)};
+              let d2 = ${s?K("coords",C.rank-1,C.rank):K("coords",1,C.rank)};
+              let xFRCCorner = vec3<u32>(${s?K("coords",1,C.rank):K("coords",2,C.rank)},
+              ${s?K("coords",2,C.rank):K("coords",3,C.rank)},
+              ${s?K("coords",3,C.rank):K("coords",4,C.rank)}) * uniforms.strides - uniforms.pads;
+              let xFCorner = xFRCCorner.x;
+              let xRCorner = xFRCCorner.y;
+              let xCCorner = xFRCCorner.z;
+              let xShapeY = ${s?K("uniforms.x_shape",1,C.rank):K("uniforms.x_shape",2,C.rank)};
+              let xShapeZ = ${s?K("uniforms.x_shape",2,C.rank):K("uniforms.x_shape",3,C.rank)};
+              let xShapeW = ${s?K("uniforms.x_shape",3,C.rank):K("uniforms.x_shape",4,C.rank)};
+              let xShapeU = ${s?K("uniforms.x_shape",4,C.rank):K("uniforms.x_shape",1,C.rank)};
+              let inputDepthNearestVec4 = (xShapeU / 4) * 4;
+              let inputDepthVec4Remainder = xShapeU % 4;
+
+              var value = 0.0;
+              for (var wF = 0u; wF < uniforms.filter_dims[0]; wF++) {
+                let xF = xFCorner + wF * uniforms.dilations[0];
+                if (xF < 0 || xF >= xShapeY) {
+                  continue;
+                }
+
+                for (var wR = 0u; wR < uniforms.filter_dims[1]; wR++) {
+                  let xR = xRCorner + wR * uniforms.dilations[1];
+                  if (xR < 0 || xR >= xShapeZ) {
+                    continue;
+                  }
+
+                  for (var wC = 0u; wC < uniforms.filter_dims[2]; wC++) {
+                    let xC = xCCorner + wC * uniforms.dilations[2];
+                    if (xC < 0 || xC >= xShapeW) {
+                      continue;
+                    }
+
+                    for (var d1 = 0u; d1 < inputDepthNearestVec4; d1 += 4) {
+                      ${s?`let xValues = vec4<f32>(
+                               getX(batch, xF, xR, xC, d1),
+                               getX(batch, xF, xR, xC, d1 + 1),
+                               getX(batch, xF, xR, xC, d1 + 2),
+                               getX(batch, xF, xR, xC, d1 + 3));
+                            `:`let xValues = vec4<f32>(
+                               getX(batch, d1, xF, xR, xC),
+                               getX(batch, d1 + 1, xF, xR, xC),
+                               getX(batch, d1 + 2, xF, xR, xC),
+                               getX(batch, d1 + 3, xF, xR, xC));
+                            `}
+                            let wValues = vec4<f32>(
+                              getW(d2, d1, wF, wR, wC),
+                              getW(d2, d1 + 1, wF, wR, wC),
+                              getW(d2, d1 + 2, wF, wR, wC),
+                              getW(d2, d1 + 3, wF, wR, wC));
+                      value += dot(xValues, wValues);
+                    }
+                    if (inputDepthVec4Remainder == 1) {
+                        ${s?`value += getX(batch, xF, xR, xC, inputDepthNearestVec4)
+                          * getW(d2, inputDepthNearestVec4, wF, wR, wC);`:`value += getX(batch, inputDepthNearestVec4, xF, xR, xC)
+                          * getW(d2, inputDepthNearestVec4, wF, wR, wC);`}
+                    } else if (inputDepthVec4Remainder == 2) {
+                      ${s?`let xValues = vec2<f32>(
+                        getX(batch, xF, xR, xC, inputDepthNearestVec4),
+                        getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1));
+                      `:`let xValues = vec2<f32>(
+                        getX(batch, inputDepthNearestVec4, xF, xR, xC),
+                        getX(batch, inputDepthNearestVec4 + 1, xF, xR, xC));
+                    `}
+                    let wValues = vec2<f32>(
+                      getW(d2, inputDepthNearestVec4, wF, wR, wC),
+                      getW(d2, inputDepthNearestVec4 + 1, wF, wR, wC));
+                      value += dot(xValues, wValues);
+                    } else if (inputDepthVec4Remainder == 3) {
+                      ${s?`let xValues = vec3<f32>(
+                        getX(batch, xF, xR, xC, inputDepthNearestVec4),
+                        getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1),
+                        getX(batch, xF, xR, xC, inputDepthNearestVec4 + 2));
+                      `:`let xValues = vec3<f32>(
+                        getX(batch, inputDepthNearestVec4, xF, xR, xC),
+                        getX(batch, inputDepthNearestVec4 + 1, xF, xR, xC),
+                        getX(batch, inputDepthNearestVec4 + 2, xF, xR, xC));
+                    `}
+                    let wValues = vec3<f32>(
+                      getW(d2, inputDepthNearestVec4, wF, wR, wC),
+                      getW(d2, inputDepthNearestVec4 + 1, wF, wR, wC),
+                      getW(d2, inputDepthNearestVec4 + 2, wF, wR, wC));
+                      value += dot(xValues, wValues);
+                    }
+                  }
+                }
+              }
+              ${T?"value = value + getBiasByOutputCoords(coords)":""};
+              ${q}
+              result[global_idx] = f32(value);
+          }`};return{name:"Conv3DNaive",shaderCache:{hint:`${t.cacheKey};${s};${g};${T}`,inputDependencies:v},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:h[0],y:h[1],z:h[2]},programUniforms:y}),getShaderSource:$}}}),u0=P(()=>{"use strict";ee(),ie(),ne(),Wt(),bh=(e,t,r,i)=>{let n=e.length>2,a=n?"value += b[output_channel];":"",s=e[0].dims,u=e[1].dims,l=t.format==="NHWC",d=l?r[3]:r[1],f=d/t.group,h=l&&f>=4?xe(d):1,g=R.size(r)/h,w=[{type:12,data:g},{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:f}];Dt(t,w),w.push(...Z(s,[u[0],u[1],u[2],u[3]/h]));let y=n?["rank","rank","rank"]:["rank","rank"];w.push(...Z([r[0],r[1],r[2],r[3]/h]));let v=T=>{let $=F("output",e[0].dataType,r.length,h),_=Ee($.type.tensor),I=Nt(t,$.type.value,_),k=B("x",e[0].dataType,s.length),E=B("w",e[1].dataType,u.length,h),C=[k,E];n&&C.push(B("b",e[2].dataType,e[2].dims,h));let O=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:t.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];Pt(t,O);let x=l?`
+      for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[0]; wHeight++) {
+        let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];
+
+        if (xHeight < 0u || xHeight >= uniforms.x_shape[1]) {
+          continue;
+        }
+
+        for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[1]; wWidth++) {
+          let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];
+          if (xWidth < 0u || xWidth >= uniforms.x_shape[2]) {
+            continue;
+          }
+
+          for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[2]; wInChannel++) {
+            let input_channel = in_channel_offset + wInChannel;
+            let xVal = ${k.get("batch","xHeight","xWidth","input_channel")};
+            let wVal = ${E.get("wHeight","wWidth","wInChannel","output_channel")};
+            value += xVal * wVal;
+          }
+        }
+      }
+      `:`
+      for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {
+        let input_channel = in_channel_offset + 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[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[3]) {
+              continue;
+            }
+
+            let xVal = ${k.get("batch","input_channel","xHeight","xWidth")};
+            let wVal = ${E.get("output_channel","wInChannel","wHeight","wWidth")};
+            value += xVal * wVal;
+          }
+        }
+      }
+      `;return`
+  ${T.registerUniforms(O).declareVariables(...C,$)}
+
+  ${T.mainStart()}
+    ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+
+    let outputIndices = ${$.offsetToIndices("global_idx")};
+    let batch: u32 = outputIndices[0];
+    let output_channel: u32 = outputIndices[${l?3:1}];
+    let xRCCorner: vec2<u32> = vec2<u32>(outputIndices[${l?1:2}], outputIndices[${l?2:3}]) * uniforms.strides - uniforms.pads;
+    let group_id: u32 = output_channel * ${h} / uniforms.output_channels_per_group;
+    var in_channel_offset = group_id * uniforms.w_shape[${l?2:1}];
+
+    var value: ${$.type.value} = ${$.type.value}(0);
+    ${x}
+    ${a}
+    ${I}
+    ${$.setByOffset("global_idx","value")}
+  }`};return{name:"GroupedConv",shaderCache:{hint:`${t.cacheKey}_${h}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:w}),getShaderSource:v}},vh=(e,t,r,i)=>{let n=e.length>2,a=xe(r[3]),s=xe(r[2]),u=R.size(r)/a/s,l=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/a],d=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/a],f=[r[0],r[1],r[2],r[3]/a],h=[{type:12,data:u},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];Dt(t,h),h.push(...Z(l,d,f));let g=(s-1)*t.strides[1]+d[1],w=y=>{let v=F("output",e[0].dataType,f.length,a),T=Ee(v.type.tensor),$=Nt(t,v.type.value,T),_=B("x",e[0].dataType,l.length,a),I=B("w",e[1].dataType,d.length,a),k=[_,I];n&&k.push(B("b",e[2].dataType,e[2].dims,a));let E=n?"value += b[output_channel];":"",C=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Pt(t,C),`
+  ${y.registerUniforms(C).declareVariables(...k,v)}
+  ${y.mainStart()}
+    ${y.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] / ${s}u;
+    let col = (index1 % width1) * ${s}u;
+    index1 = index1 / width1;
+    let row = index1 % uniforms.output_shape[1];
+    let batch = index1 / uniforms.output_shape[1];
+
+    let x_corner = vec2<i32>(i32(row), i32(col)) * uniforms.strides - uniforms.pads;
+
+    var x_vals: array<${_.type.value}, ${g}>;
+    var values: array<${v.type.value}, ${s}>;
+    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 < ${d[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 < ${g}; i++) {
+          let x_width = x_corner.y + i;
+          if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) {
+            x_vals[i] = ${_.get("batch","u32(x_height)","u32(x_width)","input_channel")};
+          } else {
+            x_vals[i] = ${_.type.value}(0);
+          }
+        }
+        for (var w_width: u32 = 0u; w_width < ${d[1]}; w_width++) {
+          let w_val = ${I.get("w_height","w_width","0","output_channel")};
+          for (var i = 0u; i < ${s}u; i++) {
+            values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]);
+          }
+        }
+      }
+    }
+
+    for (var i = 0u; i < ${s}u; i++) {
+      var value = values[i];
+      ${E}
+      ${$}
+      ${v.set("batch","row","col + i","output_channel","value")};
+    }
+  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${a};${s};${g};${d[0]};${d[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:h}),getShaderSource:w}}}),l0=P(()=>{"use strict";ie(),s0(),o0(),xa(),u0(),Wt(),$a(),$t(),ul=(e,t,r,i,n,a)=>{let s=e[0],u=e.slice(a?1:2,a?3:4),l=u.length,d=t[0],f=t.slice(2).map((g,w)=>g+(g-1)*(r[w]-1)),h=u.map((g,w)=>g+i[w]+i[w+l]).map((g,w)=>Math.floor((g-f[w]+n[w])/n[w]));return h.splice(0,0,s),h.splice(a?3:1,0,d),h},Zr=[2,3,1,0],ll=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length>5)throw new Error("greater than 5D is not supported");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],i=e[1].dims[1]*t.group;if(r!==i)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 n=e[0].dims.length-2;if(t.dilations.length!==n)throw new Error(`dilations should be ${n}D`);if(t.strides.length!==n)throw new Error(`strides should be ${n}D`);if(t.pads.length!==n*2)throw new Error(`pads should be ${n*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},Yr=(e,t)=>{let r=e.kernelShape.slice();r.length<t[1].dims.length-2&&r.push(...Array(t[1].dims.length-2-r.length).fill(0));for(let a=2;a<t[1].dims.length;++a)r[a-2]===0&&(r[a-2]=t[1].dims[a]);let i=e.pads.slice();si.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,i,e.format==="NHWC",e.autoPad);let n=Object.assign({},e);return Object.assign(n,{kernelShape:r,pads:i}),n},Yn=e=>{let t=_a(e),r=e.format,i=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],n=e.dilations,a=e.group,s=e.kernel_shape,u=e.pads,l=e.strides,d=e.w_is_const();return{autoPad:i,format:r,dilations:n,group:a,kernelShape:s,pads:u,strides:l,wIsConst:d,...t,cacheKey:`${e.format};${t.activation};`}},mn=(e,t,r,i)=>{let n=r.format==="NHWC",a=ul(t[0].dims,t[1].dims,r.dilations,r.pads,r.strides,n);if(r.group!==1){let C=[t[0]];if(n){let O=e.kernelCustomData.wT??e.compute(De(t[1],Zr),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=O),C.push(O)}else C.push(t[1]);t.length===3&&C.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&n&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1?e.compute(vh(C,r,a,i),{inputs:C}):e.compute(bh(C,r,a,i),{inputs:C});return}let s=t.length===3,u=t[0].dims[n?1:2],l=t[0].dims[n?2:3],d=t[0].dims[n?3:1],f=t[1].dims[2],h=t[1].dims[3],g=a[n?1:2],w=a[n?2:3],y=a[n?3:1],v=n&&f===u&&h===l&&r.pads[0]===0&&r.pads[1]===0;if(v||f===1&&h===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let C=a[0],O,x,D,L=[];if(n){let j=e.kernelCustomData.wT??e.compute(De(t[1],Zr),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=j),v){let ue=u*l*d;O=t[0].reshape([1,C,ue]),x=j.reshape([1,ue,y]),D=[1,C,y]}else O=t[0].reshape([C,u*l,d]),x=j.reshape([1,d,y]),D=[C,g*w,y];L.push(O),L.push(x)}else O=t[0].reshape([C,d,u*l]),x=t[1].reshape([1,y,d]),D=[C,y,g*w],L.push(x),L.push(O);s&&L.push(t[2]);let Y=D[2],q=L[0].dims[L[0].dims.length-1];Y<8&&q<8?e.compute(va(L,r,a,D,n,i),{inputs:L}):e.compute(ui(L,r,a,D,n,i),{inputs:L});return}let T=!0,$=e.kernelCustomData.wT??e.compute(De(t[1],Zr),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=$);let _=[t[0],$];s&&_.push(t[2]);let I=n?g*w:y,k=n?y:g*w,E=f*h*d;e.compute(yh(_,r,a,I,k,E,s,T,i),{inputs:_})},dl=(e,t)=>{let r=t.format==="NHWC",i=[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&&i.push(e.inputs[2]);let n=[0,t.pads[0],0,t.pads[1]],a=[1].concat(t.strides),s=[1].concat(t.dilations),u=[1].concat(t.kernelShape),l=Yr({...t,pads:n,strides:a,dilations:s,kernelShape:u},i);mn(e,i,l,d=>r?[d[0],d[2],d[3]]:[d[0],d[1],d[3]])},pl=(e,t,r)=>{let i=r.format==="NHWC"?"channelsLast":"channelsFirst",n=Yr(r,t),a=r.autoPad==="NOTSET"?r.pads:r.autoPad,s=wh(t[0].dims,t[1].dims,r.strides,r.dilations,a,!1,i);e.compute(_h(t,n,s.outShape,[s.filterDepth,s.filterHeight,s.filterWidth],[s.padInfo.front,s.padInfo.top,s.padInfo.left],i))},Xn=(e,t)=>{if(ll(e.inputs,t),e.inputs[0].dims.length===3)dl(e,t);else if(e.inputs[0].dims.length===5)pl(e,e.inputs,t);else{let r=Yr(t,e.inputs);mn(e,e.inputs,r)}}}),d0=P(()=>{"use strict";ee(),lt(),ie(),ne(),$h=(e,t,r)=>{let i=e.length>2,n=t.outputShape,a=t.format==="NHWC",s=t.group,u=e[1].dims,l=u[2]/s,d=u[3],f=a?xe(l):1,h=a&&d===1&&l>=4,g=h?Math.floor(l/4)*4:Math.floor(l/f)*f,w=l-g,y=a?xe(d):1,v=a?d===1?f:y:1,T=R.size(n)/y,$=[Math.ceil(T/64),1,1];de("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${$}`);let _=["rank","rank"],I=[t.strides[0],t.strides[1]],k=[t.kernelShape[a?1:2],t.kernelShape[a?2:3]],E=[t.dilations[0],t.dilations[1]],C=[k[0]+(t.dilations[0]<=1?0:(t.kernelShape[a?1:2]-1)*(t.dilations[0]-1)),k[1]+(t.dilations[1]<=1?0:(t.kernelShape[a?2:3]-1)*(t.dilations[1]-1))],O=[C[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),C[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],x=[{type:12,data:T},{type:12,data:I},{type:12,data:k},{type:12,data:E},{type:12,data:C},{type:6,data:O},{type:12,data:g},{type:12,data:l},{type:12,data:d},...Z(e[0].dims,e[1].dims)];i&&(x.push(...Z(e[2].dims)),_.push("rank")),x.push(...Z(n));let D=L=>{let Y=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:I.length},{name:"filter_dims",type:"u32",length:k.length},{name:"dilations",type:"u32",length:k.length},{name:"effective_filter_dims",type:"u32",length:C.length},{name:"pads",type:"i32",length:O.length},{name:"input_channels_per_group_int",type:"u32"},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],q=Ee(e[0].dataType),j=a?1:2,ue=a?2:3,A=a?3:1,U=B("W",e[1].dataType,e[1].dims.length,v),re=B("Dy",e[0].dataType,e[0].dims.length,f),J=[re,U];i&&J.push(B("bias",e[2].dataType,[n[A]].length,y));let H=F("result",e[0].dataType,n.length,y),pe=()=>{let te="";if(h)f===4?te+=`
+        let xValue = ${re.getByOffset("x_offset")};
+        let wValue = ${U.getByOffset("w_offset")};
+        dotProd = dotProd + dot(xValue, wValue);
+        x_offset += 1u;
+        w_offset += 1u;`:f===2?te+=`
+          dotProd = dotProd + dot(vec4<${q}>(${re.getByOffset("x_offset")}, ${re.getByOffset("x_offset + 1u")}), vec4<${q}>(${U.getByOffset("w_offset")}, ${U.getByOffset("w_offset + 1u")}));
+          x_offset += 2u;
+          w_offset += 2u;`:f===1&&(te+=`
+          dotProd = dotProd + dot(vec4<${q}>(${re.getByOffset("x_offset")}, ${re.getByOffset("x_offset + 1u")}, ${re.getByOffset("x_offset + 2u")}, ${re.getByOffset("x_offset + 3u")}), vec4<${q}>(${U.getByOffset("w_offset")}, ${U.getByOffset("w_offset + 1u")}, ${U.getByOffset("w_offset + 2u")}, ${U.getByOffset("w_offset + 3u")}));
+          x_offset += 4u;
+          w_offset += 4u;`);else if(te+=`
+                  let xValue = ${a?re.getByOffset(`${re.indicesToOffset(`${re.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${f}`):re.get("batch","inputChannel","idyR","idyC")};
+        `,f===1)te+=`
+          let w_offset = ${U.indicesToOffset(`${U.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
+          let wValue = ${U.getByOffset(`w_offset / ${v}`)};
+          dotProd = dotProd + xValue * wValue;`;else for(let ce=0;ce<f;ce++)te+=`
+            let wValue${ce} = ${U.getByOffset(`${U.indicesToOffset(`${U.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${ce}, wOutChannel)`)} / ${v}`)};
+            dotProd = dotProd + xValue[${ce}] * wValue${ce};`;return te},N=()=>{if(w===0)return"";if(!h)throw new Error(`packInputAs4 ${h} is not true.`);let te="";if(f===1){te+="dotProd = dotProd";for(let ce=0;ce<w;ce++)te+=`
+            + ${re.getByOffset(`x_offset + ${ce}`)} * ${U.getByOffset(`w_offset + ${ce}`)}`;te+=";"}else if(f===2){if(w!==2)throw new Error(`Invalid inputChannelsRemainder ${w}.`);te+=`
+          let xValue = ${re.getByOffset("x_offset")};
+          let wValue = ${U.getByOffset("w_offset")};
+          dotProd = dotProd + dot(xValue, wValue);`}return te},V=`
+            let outputIndices = ${H.offsetToIndices(`global_idx * ${y}`)};
+            let batch = ${H.indicesGet("outputIndices",0)};
+            let d1 = ${H.indicesGet("outputIndices",A)};
+            let r = ${H.indicesGet("outputIndices",j)};
+            let c = ${H.indicesGet("outputIndices",ue)};
+            let dyCorner = vec2<i32>(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 = ${H.type.value}(0.0);
+            var wR: u32 = 0;
+            if (uniforms.dilations.x == 1) {
+              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
+              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
+            }
+            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
+              if (wR % uniforms.dilations.x != 0) {
+                continue;
+              }
+              let dyR = (${q}(dyRCorner) + ${q}(wR)) / ${q}(uniforms.strides[0]);
+              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
+              if (dyR < 0.0 || dyR >= ${q}(uniforms.Dy_shape[${j}]) || fract(dyR) > 0.0 ||
+                  wRPerm < 0) {
+                continue;
+              }
+              let idyR: u32 = u32(dyR);
+              var wC: u32 = 0;
+              if (uniforms.dilations.y == 1) {
+                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
+                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
+              }
+              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
+                if (wC % uniforms.dilations.y != 0) {
+                  continue;
+                }
+                let dyC = (${q}(dyCCorner) + ${q}(wC)) / ${q}(uniforms.strides.y);
+                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
+                if (dyC < 0.0 || dyC >= ${q}(uniforms.Dy_shape[${ue}]) ||
+                    fract(dyC) > 0.0 || wCPerm < 0) {
+                  continue;
+                }
+                let idyC: u32 = u32(dyC);
+                var inputChannel = groupId * uniforms.input_channels_per_group;
+                ${h?`
+                var x_offset = ${re.indicesToOffset(`${re.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${f};
+                var w_offset = ${U.indicesToOffset(`${U.type.indices}(wRPerm, wCPerm, inputChannel, wOutChannel)`)} / ${v};
+                  `:""}
+                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group_int; d2 = d2 + ${h?4:f}) {
+                  ${pe()}
+                  inputChannel = inputChannel + ${h?4:f};
+                }
+                ${N()}
+                wC = wC + uniforms.strides.y - 1;
+              }
+              wR = wR + uniforms.strides[0] - 1;
+            }
+            let value = dotProd${i?` + bias[d1 / ${y}]`:""};
+            ${H.setByOffset("global_idx","value")};
+          `;return`
+    ${L.registerUniforms(Y).declareVariables(...J,H)}
+      ${L.mainStart()}
+      ${L.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
+    ${V}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${f}${v}${y}${h}${w}`,inputDependencies:_},getRunData:()=>({dispatchGroup:{x:$[0],y:$[1],z:$[2]},outputs:[{dims:r?r(n):n,dataType:e[0].dataType}],programUniforms:x}),getShaderSource:D}}}),p0=P(()=>{"use strict";d0(),Wt(),$t(),cl=(e,t,r,i,n,a)=>(e-1)*t+r+(i-1)*n+1-a,hl=(e,t,r,i,n)=>{let a=Math.floor(e/2);t==="SAME_UPPER"?(r[i]=a,r[n]=e-a):t==="SAME_LOWER"&&(r[i]=e-a,r[n]=a)},fl=(e,t,r,i,n,a,s,u,l,d)=>{let f=e.length-2,h=d.length===0;l.length<f&&l.push(...Array(f-l.length).fill(0));let g=e[0],w=t[u?3:1]*n;for(let y=0,v=e.length-f-(u?1:0);y<f;++y,++v){let T=e[v],$=h?T*s[y]:d[y],_=cl(T,s[y],a[y],t[v],r[y],$);hl(_,i,a,y,y+f),h&&d.push(s[y]*(T-1)+l[y]+(t[v]-1)*r[y]+1-a[y]-a[y+f])}d.splice(0,0,g),d.splice(u?3:1,0,w)},gn=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((h,g)=>h*g,1)===0){r.length=0;for(let h=2;h<t[1].dims.length;++h)r.push(t[1].dims[h])}let i=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(i?3:1,0,t[1].dims[1]);let n=e.pads.slice(),a=e.outputShape.slice(),s=e.outputPadding.slice(),u=t[0].dims,l=e.dilations.slice();if(l.reduce((h,g)=>h+g,0)===0){let h=t[0].dims.length-2;l=new Array(h).fill(1)}let d=e.strides.slice();if(d.reduce((h,g)=>h+g,0)===0){let h=t[0].dims.length-2;d=new Array(h).fill(1)}fl(u,r,l,e.autoPad,e.group,n,d,i,s,a);let f=Object.assign({},e);return Object.assign(f,{kernelShape:r,pads:n,outputPadding:s,outputShape:a,dilations:l,strides:d}),f},xh=e=>{let t=_a(e),r=e.format,i=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],n=e.dilations,a=e.group??1,s=e.kernelShape,u=e.pads,l=e.strides,d=e.wIsConst(),f=e.outputPadding,h=e.outputShape;return{autoPad:i,format:r,dilations:n,group:a,kernelShape:s,outputPadding:f,outputShape:h,pads:u,strides:l,wIsConst:d,...t,cacheKey:`${e.format};${t.activation};`}},ml=(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],i=e[1].dims[0];if(r!==i)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let n=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==n))throw new Error("invalid bias");let a=e[0].dims.length-2;if(t.dilations.reduce((s,u)=>s+u,0)>0&&t.dilations.length!==a)throw new Error(`dilations should be ${a}D`);if(t.strides.reduce((s,u)=>s+u,0)>0&&t.strides.length!==a)throw new Error(`strides should be ${a}D`);if(t.pads.reduce((s,u)=>s+u,0)>0&&t.pads.length!==a*2)throw new Error(`pads should be ${a*2}D`);if(t.outputPadding.length!==a&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${a}D`);if(t.kernelShape.reduce((s,u)=>s+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},yn=(e,t,r,i)=>{let n=e.kernelCustomData.wT??e.compute(De(t[1],[2,3,0,1]),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=n);let a=[t[0],n];t.length===3&&a.push(t[2]),e.compute($h(a,r,i),{inputs:a})},gl=(e,t)=>{let r=t.format==="NHWC",i=[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&&i.push(e.inputs[2]);let n=t.kernelShape;(n.length===0||n[0]===0)&&(n=[e.inputs[1].dims[2]]);let a=t.dilations;(a.length===0||a[0]===0)&&(a=[1]);let s=t.strides;(s.length===0||s[0]===0)&&(s=[1]);let u=t.pads;u.length===0&&(u=[0,0]),u=[0,u[0],0,u[1]],s=[1].concat(s),a=[1].concat(a),n=[1].concat(n);let l=t.outputPadding;l=[0].concat(l);let d=gn({...t,pads:u,strides:s,dilations:a,kernelShape:n,outputPadding:l},i);yn(e,i,d,f=>r?[f[0],f[2],f[3]]:[f[0],f[1],f[3]])},Th=(e,t)=>{if(ml(e.inputs,t),e.inputs[0].dims.length===3)gl(e,t);else{let r=gn(t,e.inputs);yn(e,e.inputs,r)}}}),c0=P(()=>{"use strict";ee(),ie(),Te(),ne(),yl=(e,t,r,i)=>{let n=R.size(t),a=t.length,s=B("input",e,a),u=F("output",e,a),l=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),d=R.normalizeAxis(l,a),f=h=>{let g=` i32(${s.indicesGet("inputIndices","uniforms.axis")}) `,w=K("uniforms.input_shape","uniforms.axis",a),y=i.reverse?g+(i.exclusive?" + 1":""):"0",v=i.reverse?w:g+(i.exclusive?"":" + 1");return`
+                ${h.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(s,u)}
+                ${h.mainStart()}
+                  ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+                  var inputIndices = ${u.offsetToIndices("global_idx")};
+                  var sum = ${u.type.value}(0);
+                  let first : i32 = ${y};
+                  let last : i32 = ${v};
+                  for (var i : i32 = first; i < last; i++) {
+                    ${s.indicesSet("inputIndices","uniforms.axis","u32(i)")};
+                    sum = sum + ${s.getByIndices("inputIndices")};
+                  }
+                  ${u.setByOffset("global_idx","sum")};
+                }`};return{name:"CumSum",shaderCache:{hint:i.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:[{type:12,data:n},{type:12,data:d},...Z(t,t)]}),getShaderSource:f}},Sh=(e,t)=>{let r=e.inputs[0].dims,i=e.inputs[0].dataType,n=e.inputs[1];e.compute(yl(i,r,n,t),{inputs:[0]})},kh=e=>{let t=e.exclusive===1,r=e.reverse===1;return me({exclusive:t,reverse:r})}}),h0=P(()=>{"use strict";ee(),ie(),Te(),ne(),wl=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.")},_l=(e,t,r,i)=>{let n=[];n.push(`fn perm(i: ${i.type.indices}) -> ${r.type.indices} {
+    var a: ${r.type.indices};`);for(let a=0;a<t;++a)n.push(r.indicesSet("a",e[a],`i[${a}]`));return n.push("return a;}"),n.join(`
+`)},bl=(e,t)=>{let r,i,n,a,s,u,l=t.format==="NHWC",d=t.blocksize,f=t.mode==="DCR";l?([r,i,n,a]=e.dims,s=f?[r,i,n,d,d,a/d**2]:[r,i,n,a/d**2,d,d],u=f?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,i,n,a]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],s=f?[r,d,d,a/d**2,i,n]:[r,a/d**2,d,d,i,n],u=f?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let h=e.reshape(s),g=h.dims.length,w=e.dataType,y=B("a",w,g),v=F("output",w,g),T=$=>`
+  ${$.registerUniform("output_size","u32").declareVariables(y,v)}
+
+  ${_l(u,g,y,v)}
+
+  ${$.mainStart()}
+    ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+
+    let indices = ${v.offsetToIndices("global_idx")};
+    let aIndices = perm(indices);
+
+    ${v.setByOffset("global_idx",y.getByIndices("aIndices"))}
+  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:$=>{let _=l?[r,i*d,n*d,a/d**2]:[r,a/d**2,i*d,n*d],I=R.size(_),k=h.dims,E=R.sortBasedOnPerm(k,u);return{outputs:[{dims:_,dataType:$[0].dataType}],dispatchGroup:{x:Math.ceil(I/64)},programUniforms:[{type:12,data:I},...Z(k,E)]}},getShaderSource:T}},Ih=(e,t)=>{wl(e.inputs),e.compute(bl(e.inputs[0],t))},Eh=e=>me({blocksize:e.blocksize,mode:e.mode,format:e.format})}),f0=P(()=>{"use strict";ee(),ie(),Te(),ne(),Xr="[a-zA-Z]|\\.\\.\\.",cr="("+Xr+")+",wn="^"+cr+"$",vl="("+cr+",)*"+cr,$l="^"+vl+"$",xl=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let r=this.symbolToIndices.get(e);r===void 0?r=[t]:r.push(t),this.symbolToIndices.set(e,r)}},Tl=class{constructor(e,t){this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,i]=t.includes("->")?t.split("->",2):[t,""];if(!r.match(RegExp($l)))throw new Error("Invalid LHS term");if(r.split(",").forEach((n,a)=>{let s=e[a].dims.slice();if(!n.match(RegExp(wn)))throw new Error("Invalid LHS term");let u=this.processTerm(n,!0,s,a);this.lhs.push(u)}),i==="")i+=[...this.symbolToInfo.entries()].filter(([n,a])=>a.count===1||n==="...").map(([n])=>n).join("");else if(!i.match(RegExp(cr)))throw new Error("Invalid RHS");i.match(RegExp(Xr,"g"))?.forEach(n=>{if(n==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let a=this.symbolToInfo.get(n);if(a===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(a.dimValue)}}),this.rhs=this.processTerm(i,!1,this.outputDims)}addSymbol(e,t,r){let i=this.symbolToInfo.get(e);if(i!==void 0){if(i.dimValue!==t&&i.count!==1)throw new Error("Dimension mismatch");i.count++,i.inputIndices.push(r)}else i={count:1,dimValue:t,inputIndices:[r]};this.symbolToInfo.set(e,i)}processTerm(e,t,r,i=-1){let n=r.length,a=!1,s=[],u=0;if(!e.match(RegExp(wn))&&!t&&e!=="")throw new Error("Invalid LHS term");let l=e.match(RegExp(Xr,"g")),d=new xl(i);return l?.forEach((f,h)=>{if(f==="..."){if(a)throw new Error("Only one ellipsis is allowed per input term");a=!0;let g=n-l.length+1;if(g<0)throw new Error("Ellipsis out of bounds");if(s=r.slice(u,u+g),this.hasEllipsis){if(this.ellipsisDims.length!==s.length||this.ellipsisDims.toString()!==s.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=s;else throw new Error("Ellipsis must be specified in the LHS");for(let w=0;w<s.length;w++){let y=String.fromCharCode(48+w);d.addSymbol(y,h+w),this.addSymbol(y,r[u++],i)}}else d.addSymbol(f,h+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(f,r[u++],i)}),d}},_n=e=>e+"_max",Sl=(e,t,r,i)=>{let n=e.map(d=>d.length).map((d,f)=>B(`input${f}`,t,d)),a=R.size(i),s=F("output",t,i.length),u=[...r.symbolToInfo.keys()].filter(d=>!r.rhs.symbolToIndices.has(d)),l=d=>{let f=[],h="var prod = 1.0;",g="var sum = 0.0;",w="sum += prod;",y=[],v=[],T=[],$=[],_=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((k,E)=>{if(r.rhs.symbolToIndices.has(E)){let C=r.rhs.symbolToIndices.get(E)?.[0];C!==void 0&&r.lhs.forEach((O,x)=>{if(k.inputIndices.includes(x)){let D=O.symbolToIndices.get(E);if(D===void 0)throw new Error("Invalid symbol error");D.forEach(L=>{f.push(`${n[x].indicesSet(`input${x}Indices`,L,s.indicesGet("outputIndices",C))}`)})}})}else r.lhs.forEach((C,O)=>{if(k.inputIndices.includes(O)){let x=C.symbolToIndices.get(E);if(x===void 0)throw new Error("Invalid symbol error");x.forEach(D=>{y.push(`${n[O].indicesSet(`input${O}Indices`,D,`${E}`)}`)}),$.push(`prod *= ${n[O].getByIndices(`input${O}Indices`)};`)}}),v.push(`for(var ${E}: u32 = 0; ${E} < uniforms.${_n(E)}; ${E}++) {`),T.push("}")});let I=_?[...f,`let sum = ${n.map((k,E)=>k.getByIndices(`input${E}Indices`)).join(" * ")};`]:[...f,g,...v,...y,h,...$,w,...T];return`
+            ${d.registerUniforms(u.map(k=>({name:`${_n(k)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...n,s)}
+
+            ${d.mainStart()}
+            ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+            var outputIndices = ${s.offsetToIndices("global_idx")};
+            ${n.map((k,E)=>`var input${E}Indices: ${n[E].type.indices};`).join(`
+`)}
+            ${I.join(`
+`)};
+            ${s.setByOffset("global_idx","sum")};
+          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let d=u.filter(h=>r.symbolToInfo.has(h)).map(h=>({type:12,data:r.symbolToInfo.get(h)?.dimValue||0}));d.push({type:12,data:a});let f=e.map((h,g)=>[...Z(h)]).reduce((h,g)=>h.concat(g),d);return f.push(...Z(i)),{outputs:[{dims:i,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:f}},getShaderSource:l}},Ch=(e,t)=>{let r=new Tl(e.inputs,t.equation),i=r.outputDims,n=e.inputs.map((a,s)=>a.dims);e.compute(Sl(n,e.inputs[0].dataType,r,i))},zh=e=>{let t=e.equation.replace(/\s+/g,"");return me({equation:t})}}),m0=P(()=>{"use strict";ee(),ie(),ne(),kl=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),i=r.length<t.length?0:r.length-t.length,n=t.length<r.length?0:t.length-r.length;for(;i<r.length&&n<t.length;++i,++n)if(r[i]!==t[n]&&r[i]!==1&&t[n]!==1)throw new Error("Expand requires shape to be broadcastable to input")},bn=(e,t)=>{let r=e.length-t.length,i=[];for(let n=0;n<r;++n)i.push(e[n]);for(let n=0;n<t.length;++n)i.push(t[n]===1?e[n+r]:t[n]);return i},Il=(e,t)=>e.length>t.length?bn(e,t):bn(t,e),El=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),i=Il(t,r),n=e[0].dataType,a=n===9||R.size(t)===1,s=n===9||t.length>0&&t[t.length-1]%4===0?4:1,u=a||i.length>0&&i[i.length-1]%4===0?4:1,l=Math.ceil(R.size(i)/u),d=h=>{let g=B("input",n,t.length,s),w=F("output",n,i.length,u),y;if(n===9){let v=(T,$,_="")=>`
+          let outputIndices${$} = ${w.offsetToIndices(`outputOffset + ${$}u`)};
+          let offset${$} = ${g.broadcastedIndicesToOffset(`outputIndices${$}`,w)};
+          let index${$} = offset${$} / 4u;
+          let component${$} = offset${$} % 4u;
+          ${T}[${$}] = ${_}(${g.getByOffset(`index${$}`)}[component${$}]);
+        `;y=`
+        let outputOffset = global_idx * ${u};
+        var data = vec4<u32>(0);
+        ${v("data",0,"u32")}
+        ${v("data",1,"u32")}
+        ${v("data",2,"u32")}
+        ${v("data",3,"u32")}
+        ${w.setByOffset("global_idx","data")}
+      }`}else y=`
+        let outputIndices = ${w.offsetToIndices(`global_idx * ${u}`)};
+        let inputOffset = ${g.broadcastedIndicesToOffset("outputIndices",w)};
+        let data = ${w.type.value}(${g.getByOffset(`inputOffset / ${s}`)});
+        ${w.setByOffset("global_idx","data")}
+      }`;return`
+    ${h.registerUniform("vec_size","u32").declareVariables(g,w)}
+    ${h.mainStart()}
+    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
+    ${y}`},f=[{type:12,data:l},...Z(t,i)];return{name:"Expand",shaderCache:{hint:`${i.length};${s}${u}`,inputDependencies:["rank"]},getShaderSource:d,getRunData:()=>({outputs:[{dims:i,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:f})}},Ah=e=>{kl(e.inputs),e.compute(El(e.inputs),{inputs:[0]})}}),g0=P(()=>{"use strict";ee(),ie(),ne(),wa(),Cl=e=>{let t=e[0].dataType,r=R.size(e[0].dims),i=R.size(e[1].dims),n=i%4===0,a=s=>{let u=B("x",t,[1],4),l=B("bias",t,[1],4),d=F("y",t,[1],4),f=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],h=w=>`
+      let bias${w}_offset: u32 = (global_idx * 4 + ${w}) % uniforms.bias_size;
+      let bias${w} = ${l.getByOffset(`bias${w}_offset / 4`)}[bias${w}_offset % 4];`,g=n?`
+      let bias = ${l.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${h(0)}${h(1)}${h(2)}${h(3)}
+      let bias = ${u.type.value}(bias0, bias1, bias2, bias3);`;return`${s.registerUniforms(f).declareVariables(u,l,d)}
+
+    ${Kn(Ae(t))}
+
+    ${s.mainStart(Zt)}
+      ${s.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}
+
+      let x = ${u.getByOffset("global_idx")};
+      ${g}
+      let x_in = x + bias;
+      ${d.setByOffset("global_idx",jn("x_in"))}
+    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${n}`,inputDependencies:["type","type"]},getShaderSource:a,getRunData:s=>({outputs:[{dims:s[0].dims,dataType:s[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:i}],dispatchGroup:{x:Math.ceil(r/Zt/4)}})}},Oh=e=>{e.inputs.length<2||R.size(e.inputs[1].dims)===0?Xc(e):e.compute(Cl(e.inputs))}}),y0=P(()=>{"use strict";ee(),ie(),Te(),ne(),zl=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},Al=(e,t)=>{let r=e[0].dims,i=e[1].dims,n=r.length,a=R.normalizeAxis(t.axis,n),s=r.slice(0);s.splice(a,1,...i);let u=r[a],l=e[0].dataType===9?4:1,d=Math.ceil(R.size(s)/l),f=[{type:12,data:d},{type:6,data:u},{type:12,data:a},...Z(e[0].dims,e[1].dims,s)],h=g=>{let w=B("data",e[0].dataType,e[0].dims.length,l),y=B("inputIndices",e[1].dataType,e[1].dims.length),v=F("output",e[0].dataType,s.length,l),T=_=>{let I=i.length,k=`var indicesIndices${_}  = ${y.type.indices}(0);`;for(let E=0;E<I;E++)k+=`${I>1?`indicesIndices${_}[${E}]`:`indicesIndices${_}`} = ${s.length>1?`outputIndices${_}[uniforms.axis + ${E}]`:`outputIndices${_}`};`;k+=`
+          var idx${_} = ${y.getByIndices(`indicesIndices${_}`)};
+          if (idx${_} < 0) {
+            idx${_} = idx${_} + uniforms.axisDimLimit;
+          }
+          var dataIndices${_} : ${w.type.indices};
+        `;for(let E=0,C=0;E<n;E++)E===a?(k+=`${n>1?`dataIndices${_}[${E}]`:`dataIndices${_}`} = u32(idx${_});`,C+=I):(k+=`${n>1?`dataIndices${_}[${E}]`:`dataIndices${_}`} = ${s.length>1?`outputIndices${_}[${C}]`:`outputIndices${_}`};`,C++);return k},$;if(e[0].dataType===9){let _=(I,k,E="")=>`
+          let outputIndices${k} = ${v.offsetToIndices(`outputOffset + ${k}u`)};
+          ${T(k)};
+          let offset${k} = ${w.indicesToOffset(`dataIndices${k}`)};
+          let index${k} = offset${k} / 4u;
+          let component${k} = offset${k} % 4u;
+          ${I}[${k}] = ${E}(${w.getByOffset(`index${k}`)}[component${k}]);
+        `;$=`
+        let outputOffset = global_idx * ${l};
+        var value = vec4<u32>(0);
+        ${_("value",0,"u32")}
+        ${_("value",1,"u32")}
+        ${_("value",2,"u32")}
+        ${_("value",3,"u32")}
+        ${v.setByOffset("global_idx","value")}
+      `}else $=`
+      let outputIndices = ${v.offsetToIndices("global_idx")};
+      ${T("")};
+      let value = ${w.getByIndices("dataIndices")};
+      ${v.setByOffset("global_idx","value")};
+      `;return`
+      ${g.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(w,y,v)}
+      ${g.mainStart()}
+        ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+        ${$}
+      }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:s,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:f}),getShaderSource:h}},Rh=e=>me({axis:e.axis}),Mh=(e,t)=>{let r=e.inputs;zl(r),e.compute(Al(e.inputs,t))}}),w0=P(()=>{"use strict";ee(),ie(),ne(),Ol=(e,t,r,i,n,a,s,u,l)=>{let d=[{type:12,data:a},{type:12,data:i},{type:12,data:n},{type:12,data:r},{type:12,data:s},{type:12,data:u},{type:12,data:l}],f=[a];d.push(...Z(t.dims,f));let h=g=>{let w=B("indices_data",t.dataType,t.dims.length),y=F("input_slice_offsets_data",12,1,1),v=[w,y],T=[{name:"output_size",type:"u32"},{name:"batch_dims",type:"u32"},{name:"input_dims",type:"u32",length:n.length},{name:"sizes_from_slice_dims_data",type:"u32",length:r.length},{name:"num_slices_per_batch",type:"u32"},{name:"input_batch_stride",type:"u32"},{name:"num_slice_dims",type:"u32"}];return`
+  ${g.registerUniforms(T).declareVariables(...v)}
+  ${g.mainStart()}
+    ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+    let batch_idx = global_idx / uniforms.num_slices_per_batch;
+    let base_offset = batch_idx * uniforms.input_batch_stride;
+
+    let slice_indices_base_offset = global_idx * uniforms.num_slice_dims;
+    var relative_slice_offset = 0;
+    for (var dim_idx = 0u; dim_idx < uniforms.num_slice_dims; dim_idx ++) {
+      var index = i32(indices_data[dim_idx + slice_indices_base_offset].x);
+      let input_dim_idx = uniforms.batch_dims + dim_idx;
+      if (index < 0) {
+        ${n.length===1?"index += i32(uniforms.input_dims);":"index += i32(uniforms.input_dims[input_dim_idx]);"}
+      }
+      ${r.length===1?"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data);":"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data[dim_idx]);"}
+    }
+
+    input_slice_offsets_data[global_idx] =  base_offset + u32(relative_slice_offset);
+  }`};return e.compute({name:"computeSliceOffsets",shaderCache:{hint:`${n.length}_${r.length}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:f,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d}),getShaderSource:h},{inputs:[t],outputs:[-1]})[0]},Bh=(e,t)=>{let r=e.inputs,i=r[0].dims,n=r[0].dataType,a=r[1].dims,s=a[a.length-1],u=R.sizeToDimension(a,a.length-1),l=R.sizeFromDimension(i,t.batchDims+s),d=R.sizeToDimension(i,t.batchDims),f=R.sizeFromDimension(i,t.batchDims),h=u/d,g=new Array(s),w=l;for(let k=0;k<s;++k)g[s-1-k]=w,w*=i[t.batchDims+s-1-k];let y=Ol(e,r[1],g,t.batchDims,i,u,h,f,s),v=t.batchDims+s;if(v>i.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let T=a.slice(0,-1).concat(i.slice(v)),$=R.size(T),_=[{type:12,data:$},{type:12,data:l},...Z(r[0].dims,y.dims,T)],I=k=>{let E=B("data",r[0].dataType,r[0].dims.length),C=B("slice_offsets",12,y.dims.length),O=F("output",r[0].dataType,T.length);return`
+          ${k.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(E,C,O)}
+            ${k.mainStart()}
+            ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+          let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
+          output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
+        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:T,dataType:n}],dispatchGroup:{x:Math.ceil($/64)},programUniforms:_}),getShaderSource:I},{inputs:[r[0],y]})},Nh=e=>({batchDims:e.batch_dims,cacheKey:""})}),_0=P(()=>{"use strict";ee(),ie(),Te(),ne(),Rl=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let r=R.normalizeAxis(t.quantizeAxis,e[0].dims.length),i=t.blockSize,n=e[0],a=e[2],s=e.length===4?e[3]:void 0;if(a.dims.length!==n.dims.length||!n.dims.map((u,l)=>l===r?Math.ceil(u/i)===a.dims[l]:u===a.dims[l]).reduce((u,l)=>u&&l,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(s){if(s.dataType!==n.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(s.dims.length!==a.dims.length||!s.dims.map((u,l)=>u===a.dims[l]).reduce((u,l)=>u&&l,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},Ml=(e,t)=>{let r=e[0].dims,i=e[1].dims,n=r.length,a=R.normalizeAxis(t.gatherAxis,n),s=R.normalizeAxis(t.quantizeAxis,n),u=r.slice(0);u.splice(a,1,...i);let l=R.size(u),d=e[2].dataType,f=e[0].dataType===22,h=[{type:12,data:l},{type:12,data:s},{type:12,data:a},{type:12,data:t.blockSize},...Z(...e.map((w,y)=>w.dims),u)],g=w=>{let y=B("data",e[0].dataType,e[0].dims.length),v=B("inputIndices",e[1].dataType,e[1].dims.length),T=B("scales",e[2].dataType,e[2].dims.length),$=e.length>3?B("zeroPoint",e[3].dataType,e[3].dims.length):void 0,_=F("output",d,u.length),I=[y,v,T];$&&I.push($);let k=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
+        ${w.registerUniforms(k).declareVariables(...I,_)}
+        ${w.mainStart()}
+        let output_indices = ${_.offsetToIndices("global_idx")};
+        var indices_indices = ${v.type.indices}(0);
+        ${i.length>1?`
+          for (var i: u32 = 0; i < ${i.length}; i++) {
+            let index = ${_.indicesGet("output_indices","uniforms.gather_axis + i")};
+            ${v.indicesSet("indices_indices","i","index")};
+          }`:`indices_indices = ${_.indicesGet("output_indices","uniforms.gather_axis")};`};
+        var data_indices = ${y.type.indices}(0);
+        for (var i: u32 = 0; i < uniforms.gather_axis; i++) {
+          let index = ${_.indicesGet("output_indices","i")};
+          ${y.indicesSet("data_indices","i","index")};
+        }
+        var index_from_indices = ${v.getByIndices("indices_indices")};
+        if (index_from_indices < 0) {
+          index_from_indices += ${r[a]};
+        }
+        ${y.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")};
+        for (var i = uniforms.gather_axis + 1; i < ${u.length}; i++) {
+          let index = ${_.indicesGet("output_indices",`i + ${i.length} - 1`)};
+          ${y.indicesSet("data_indices","i","index")};
+        }
+        let data_offset = ${y.indicesToOffset("data_indices")};
+        let data_index = data_offset % 8;
+        // Convert 4-bit packed data to 8-bit packed data.
+        let packed_4bit_quantized_data = ${y.getByOffset("data_offset / 8")};
+        let packed_8bit_quantized_data = (packed_4bit_quantized_data >> (4 * (data_index % 2))) & 0x0f0f0f0f;
+        let quantized_data_vec = ${f?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_quantized_data));
+        let quantized_data = quantized_data_vec[data_index / 2];
+        var scale_indices = data_indices;
+        let quantize_axis_index = ${T.indicesGet("data_indices","uniforms.quantize_axis")} / uniforms.block_size;
+        ${T.indicesSet("scale_indices","uniforms.quantize_axis","quantize_axis_index")};
+        var scale = ${T.getByIndices("scale_indices")};
+        ${$?`
+              let zero_point_indices = scale_indices;
+              let zero_point_offset = ${$.indicesToOffset("zero_point_indices")};
+              let zero_point_index = zero_point_offset % 8;
+              let packed_4bit_zero_points = ${$.getByOffset("zero_point_offset / 8")};
+              let packed_8bit_zero_points = (packed_4bit_zero_points >> (4 * (zero_point_index % 2))) & 0x0f0f0f0f;
+              let zero_point_vec = ${f?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_zero_points));
+              let zero_point = zero_point_vec[zero_point_index / 2];`:"var zero_point = 0"};
+        let dequantized_data = ${Ae(d)}(quantized_data - zero_point) * scale;
+        ${_.setByOffset("global_idx","dequantized_data")};
+    }`};return{name:"GatherBlockQuantized",shaderCache:{hint:`${t.cacheKey};${e.filter((w,y)=>y!==1).map(w=>w.dims.join("_")).join(";")}`,inputDependencies:Array.from({length:e.length},(w,y)=>"rank")},getRunData:()=>({outputs:[{dims:u,dataType:d}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:h}),getShaderSource:g}},Dh=(e,t)=>{let r=e.inputs;Rl(r,t),e.compute(Ml(e.inputs,t))},Ph=e=>me({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})}),b0=P(()=>{"use strict";ee(),ie(),Te(),ne(),Bl=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.`)},Nl=(e,t)=>{let r=e[0].dims,i=e[0].dataType,n=r.length,a=e[1].dims,s=e[1].dataType,u=R.normalizeAxis(t.axis,n),l=r[u],d=a.slice(0),f=R.size(d),h=B("input",i,n),g=B("indicesInput",s,a.length),w=F("output",i,d.length),y=[{type:12,data:f},{type:6,data:l},{type:12,data:u}];return y.push(...Z(r,a,d)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:d,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:y}),getShaderSource:v=>`
+      ${v.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(h,g,w)}
+      ${v.mainStart()}
+      ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+
+      let outputIndices = ${w.offsetToIndices("global_idx")};
+
+      var idx = ${g.getByOffset("global_idx")};
+      if (idx < 0) {
+        idx = idx + uniforms.axisDimLimit;
+      }
+      var inputIndices = ${h.type.indices}(outputIndices);
+      ${h.indicesSet("inputIndices","uniforms.axis","u32(idx)")};
+      let value = ${h.getByIndices("inputIndices")};
+
+      ${w.setByOffset("global_idx","value")};
+  }`}},Uh=e=>me({axis:e.axis}),Wh=(e,t)=>{let r=e.inputs;Bl(r),e.compute(Nl(e.inputs,t))}}),v0=P(()=>{"use strict";ee(),ie(),ne(),Dl=e=>{if(!e)throw new Error("Input is missing");if(e.length<2||e.length>3)throw new Error("Invaid input number.");if(e.length===3&&e[2].dims.length>2)throw new Error("Invalid input shape of C");if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("Input types are mismatched")},Pl=(e,t)=>{let r=e[0].dims.slice(),i=e[1].dims.slice(),[n,a,s]=Up.getShapeOfGemmResult(r,t.transA,i,t.transB,e.length===3?e[2].dims:void 0),u=[n,a];if(!u)throw new Error("Can't use gemm on the given tensors");let l=16,d=Math.ceil(a/l),f=Math.ceil(n/l),h=!0,g=R.size(u),w=[{type:12,data:h?d:g},{type:12,data:n},{type:12,data:a},{type:12,data:s},{type:1,data:t.alpha},{type:1,data:t.beta}],y=["type","type"];e.length===3&&(w.push(...Z(e[2].dims)),y.push("rank")),w.push(...Z(u));let v=$=>{let _="";t.transA&&t.transB?_="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?_="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?_="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(_="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let I=t.alpha===1?"":"value *= uniforms.alpha;",k=B("a",e[0].dataType,e[0].dims),E=B("b",e[1].dataType,e[1].dims),C=k.type.value,O=null,x=[k,E];e.length===3&&(O=B("c",e[2].dataType,e[2].dims.length),x.push(O));let D=F("output",e[0].dataType,u.length);x.push(D);let L=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`
+  ${$.registerUniforms(L).declareVariables(...x)}
+
+  ${$.mainStart()}
+    ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+
+    let m = global_idx / uniforms.N;
+    let n = global_idx % uniforms.N;
+
+    var value = ${C}(0);
+    for (var k: u32 = 0u; k < uniforms.K; k++) {
+      ${_}
+    }
+
+    ${I}
+    ${O!=null?`let cOffset = ${O.broadcastedIndicesToOffset("vec2(m, n)",D)}; value += ${C}(uniforms.beta) * ${O.getByOffset("cOffset")};`:""}
+    output[global_idx] = value;
+  }`},T=$=>{let _=B("a",e[0].dataType,e[0].dims),I=B("b",e[1].dataType,e[1].dims),k=null,E=[_,I];e.length===3&&(k=B("c",e[2].dataType,e[2].dims.length),E.push(k));let C=F("output",e[0].dataType,u.length);E.push(C);let O=[{name:"num_tile_n",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}],x="",D="";t.transA&&t.transB?(D=`
+      var col = tile_row_start + local_id.x;
+      var row = k_start + local_id.y;
+      if (col < uniforms.M && row < uniforms.K) {
+        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
+      } else {
+        tile_a[local_id.y][local_id.x] = ${_.type.value}(0);
+      }
+
+      col = k_start + local_id.x;
+      row = tile_col_start + local_id.y;
+      if (col < uniforms.K && row < uniforms.N) {
+        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
+      } else {
+        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
+      }
+      `,x="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):t.transA&&!t.transB?(D=`
+      var col = tile_row_start + local_id.x;
+      var row = k_start + local_id.y;
+      if (col < uniforms.M && row < uniforms.K) {
+        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
+      } else {
+        tile_a[local_id.y][local_id.x] = ${_.type.value}(0);
+      }
+
+      col = tile_col_start + local_id.x;
+      row = k_start + local_id.y;
+      if (col < uniforms.N && row < uniforms.K) {
+        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
+      } else {
+        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
+      }
+      `,x="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!t.transA&&t.transB?(D=`
+      var col = k_start + local_id.x;
+      var row = tile_row_start + local_id.y;
+      if (col < uniforms.K && row < uniforms.M) {
+        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
+      } else {
+        tile_a[local_id.y][local_id.x] = ${_.type.value}(0);
+      }
+
+      col = k_start + local_id.x;
+      row = tile_col_start + local_id.y;
+      if (col < uniforms.K && row < uniforms.N) {
+        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
+      } else {
+        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
+      }
+      `,x="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):!t.transA&&!t.transB&&(D=`
+      var col = k_start + local_id.x;
+      var row = tile_row_start + local_id.y;
+      if (col < uniforms.K && row < uniforms.M) {
+        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
+      } else {
+        tile_a[local_id.y][local_id.x] = ${_.type.value}(0);
+      }
+
+      col = tile_col_start + local_id.x;
+      row = k_start + local_id.y;
+      if (col < uniforms.N && row < uniforms.K) {
+        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
+      } else {
+        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
+      }
+      `,x="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let L=t.alpha===1?"":"value *= uniforms.alpha;";return`
+  ${$.registerUniforms(O).declareVariables(...E)}
+  var<workgroup> tile_a: array<array<${_.type.storage}, ${l}>, ${l}>;
+  var<workgroup> tile_b: array<array<${I.type.storage}, ${l}>, ${l}>;
+  ${$.mainStart([l,l,1])}
+    let tile_col_start = (workgroup_index % uniforms.num_tile_n) * ${l};
+    let tile_row_start = (workgroup_index / uniforms.num_tile_n) * ${l};
+    let num_tiles = (uniforms.K - 1) / ${l} + 1;
+    var k_start = 0u;
+    var value = ${C.type.value}(0);
+    for (var t: u32 = 0u; t < num_tiles; t++) {
+      ${D}
+      k_start = k_start + ${l};
+      workgroupBarrier();
+
+      for (var k: u32 = 0u; k < ${l}; k++) {
+        ${x}
+      }
+      workgroupBarrier();
+    }
+
+    ${L}
+    let m = tile_row_start + local_id.y;
+    let n = tile_col_start + local_id.x;
+    ${k!=null?`let cOffset = ${k.broadcastedIndicesToOffset("vec2(m, n)",C)}; value += ${C.type.value}(uniforms.beta) * ${k.getByOffset("cOffset")};`:""}
+    if (m < uniforms.M && n < uniforms.N) {
+      output[m * uniforms.N + n] = value;
+    }
+  }`};return h?{name:"GemmShared",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:u,dataType:e[0].dataType}],dispatchGroup:{x:d*f},programUniforms:w}),getShaderSource:T}:{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:u,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:w}),getShaderSource:v}},Lh=e=>{let t=e.transA,r=e.transB,i=e.alpha,n=e.beta;return{transA:t,transB:r,alpha:i,beta:n,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},qh=(e,t)=>{Dl(e.inputs),e.compute(Pl(e.inputs,t))}}),$0=P(()=>{"use strict";ee(),ie(),Te(),ne(),[it,ot,Et,Ct]=[0,1,2,3],Ul=e=>{if(e[0].dims.length!==4)throw new Error("only 4-D tensor is supported.");if(e[0].dims.length!==e[1].dims.length)throw new Error("input dimensions must be equal to grid dimensions");if(e[0].dims.length-2!==e[1].dims[e[1].dims.length-1])throw new Error(`last dimension of grid must be equal to ${e[0].dims.length-2}`);if(e[0].dims[0]!==e[1].dims[0])throw new Error("grid batch size must match input batch size")},Wl=`
+  fn gs_get_cubic_coeffs(x: f32) -> vec4<f32> {
+    let cubic_alpha = -0.75f;
+    let x_abs = abs(x);
+    var coeffs: vec4<f32>;
+    coeffs[0] = (((cubic_alpha * (x_abs + 1) - 5 * cubic_alpha) * (x_abs + 1) + 8 * cubic_alpha) * (x_abs + 1) - 4 * cubic_alpha);
+    coeffs[1] = (((cubic_alpha + 2) * x_abs - (cubic_alpha + 3)) * x_abs * x_abs + 1);
+    coeffs[2] = (((cubic_alpha + 2) * (1 - x_abs) - (cubic_alpha + 3)) * (1 - x_abs) * (1 - x_abs) + 1);
+    coeffs[3] = (((cubic_alpha * (2 - x_abs) - 5 * cubic_alpha) * (2 - x_abs) + 8 * cubic_alpha) * (2 - x_abs) - 4 * cubic_alpha);
+    return coeffs;
+  }
+`,Ll=e=>`
+  fn gs_bicubic_interpolate(p: mat4x4<${e}>, x: f32, y: f32) -> ${e} {
+    var v: vec4<f32>;
+    var coeffs = gs_get_cubic_coeffs(x);
+    for (var i = 0; i < 4; i++) {
+      v[i] = coeffs[0] * p[i][0] + coeffs[1] * p[i][1] + coeffs[2] * p[i][2] + coeffs[3] * p[i][3];
+    }
+    coeffs = gs_get_cubic_coeffs(y);
+    let pixel = ${e}(coeffs[0] * v[0] + coeffs[1] * v[1] + coeffs[2] * v[2] + coeffs[3] * v[3]);
+    return pixel;
+  }
+`,ql=e=>`
+  fn gs_denormalize(n: f32, length: i32) -> f32 {
+    ${e.alignCorners===0?`
+    // alignCorners: false => [-1, 1] to [-0.5, length - 0.5]
+    return ((n + 1.0) * f32(length) - 1.0) / 2.0;
+    `:`
+    // alignCorners: true => [-1, 1] to [0, length - 1]
+    return (n + 1.0) / 2.0 * (f32(length - 1));
+    `}
+  }
+`,Vl=e=>`
+  ${e.paddingMode==="reflection"?`
+      fn gs_reflect(x: i32, x_min: f32, x_max: f32) -> u32 {
+        var dx = 0.0;
+        var fx = f32(x);
+        let range = x_max - x_min;
+        if (fx < x_min) {
+          dx = x_min - fx;
+          let n = u32(dx / range);
+          let r = dx - f32(n) * range;
+          if (n % 2 == 0) {
+            fx = x_min + r;
+          } else {
+            fx = x_max - r;
+          }
+        } else if (fx > x_max) {
+          dx = fx - x_max;
+          let n = u32(dx / range);
+          let r = dx - f32(n) * range;
+          if (n % 2 == 0) {
+            fx = x_max - r;
+          } else {
+            fx = x_min + r;
+          }
+        }
+        return u32(fx);
+      }`:""}
+`,Gl=(e,t,r)=>`
+  fn pixel_at_grid(r: i32, c: i32, H: i32, W: i32, batch: u32, channel: u32, border: vec4<f32>) -> ${t} {
+     var pixel = ${t}(0);
+     var indices = vec4<u32>(0);
+     indices[${it}] = batch;
+     indices[${ot}] = channel;`+(()=>{switch(r.paddingMode){case"zeros":return`
+          if (r >= 0 && r < H && c >=0 && c < W) {
+            indices[${Et}] = u32(r);
+            indices[${Ct}] = u32(c);
+          } else {
+            return ${t}(0);
+          }
+        `;case"border":return`
+          indices[${Et}] = u32(clamp(r, 0, H - 1));
+          indices[${Ct}] = u32(clamp(c, 0, W - 1));
+        `;case"reflection":return`
+          indices[${Et}] = gs_reflect(r, border[1], border[3]);
+          indices[${Ct}] = gs_reflect(c, border[0], border[2]);
+        `;default:throw new Error(`padding mode ${r.paddingMode} is not supported`)}})()+`
+    return ${e.getByIndices("indices")};
+  }
+`,Fl=(e,t,r)=>(()=>{switch(r.mode){case"nearest":return`
+          let result = pixel_at_grid(i32(round(y)), i32(round(x)), H_in, W_in, indices[${it}], indices[${ot}], border);
+        `;case"bilinear":return`
+          let x1 = i32(floor(x));
+          let y1 = i32(floor(y));
+          let x2 = x1 + 1;
+          let y2 = y1 + 1;
+
+          let p11 = pixel_at_grid(y1, x1, H_in, W_in, indices[${it}], indices[${ot}], border);
+          let p12 = pixel_at_grid(y1, x2, H_in, W_in, indices[${it}], indices[${ot}], border);
+          let p21 = pixel_at_grid(y2, x1, H_in, W_in, indices[${it}], indices[${ot}], border);
+          let p22 = pixel_at_grid(y2, x2, H_in, W_in, indices[${it}], indices[${ot}], border);
+
+          let dx2 = ${t}(f32(x2) - x);
+          let dx1 = ${t}(x - f32(x1));
+          let dy2 = ${t}(f32(y2) - y);
+          let dy1 = ${t}(y - f32(y1));
+          let result = dy2 * (dx2 * p11 + dx1 * p12) + dy1 * (dx2 * p21 + dx1 * p22);
+        `;case"bicubic":return`
+          let x0 = i32(floor(x)) - 1;
+          let y0 = i32(floor(y)) - 1;
+          var p: mat4x4<${t}>;
+          for (var h = 0; h < 4; h++) {
+            for (var w = 0; w < 4; w++) {
+              p[h][w] = pixel_at_grid(h + y0, w + x0, H_in, W_in, indices[${it}], indices[${ot}], border);
+            }
+          }
+
+          let dx = x - f32(x0 + 1);
+          let dy = y - f32(y0 + 1);
+          let result = gs_bicubic_interpolate(p, dx, dy);
+        `;default:throw new Error(`mode ${r.mode} is not supported`)}})()+`${e.setByOffset("global_idx","result")}`,Hl=(e,t)=>{let r=B("x",e[0].dataType,e[0].dims.length),i=[e[1].dims[0],e[1].dims[1],e[1].dims[2]],n=B("grid",e[1].dataType,i.length,2),a=[e[0].dims[0],e[0].dims[1],e[1].dims[1],e[1].dims[2]];t.format==="NHWC"&&(a=[e[0].dims[0],e[1].dims[1],e[1].dims[2],e[0].dims[3]],[it,ot,Et,Ct]=[0,3,1,2]);let s=F("output",e[0].dataType,a.length),u=r.type.value,l=R.size(a),d=[{type:12,data:l},...Z(e[0].dims,i,a)],f=h=>`
+  ${h.registerUniform("output_size","u32").declareVariables(r,n,s)}
+  ${Wl}
+  ${Ll(u)}
+  ${ql(t)}
+  ${Vl(t)}
+  ${Gl(r,u,t)}
+
+  ${h.mainStart()}
+    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+      let H_in = i32(uniforms.x_shape[${Et}]);
+      let W_in = i32(uniforms.x_shape[${Ct}]);
+
+      ${t.alignCorners===0?`
+      let x_min = -0.5;
+      let x_max = f32(W_in) - 0.5;
+      let y_min = -0.5;
+      let y_max = f32(H_in) - 0.5;
+      `:`
+      let x_min = 0.0;
+      let x_max = f32(W_in) - 1.0;
+      let y_min = 0.0;
+      let y_max = f32(H_in) - 1.0;
+      `};
+      let border = vec4<f32>(x_min, y_min, x_max, y_max);
+
+      let indices = ${s.offsetToIndices("global_idx")};
+      var grid_indices = vec3<u32>(indices[${it}], indices[${Et}], indices[${Ct}]);
+      let nxy = ${n.getByIndices("grid_indices")};
+      var x = gs_denormalize(f32(nxy[0]), W_in);
+      var y = gs_denormalize(f32(nxy[1]), H_in);
+
+      ${Fl(s,u,t)}
+  }`;return{name:"GridSample",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:["type","type"]},getRunData:h=>{let g=R.size(a);return{outputs:[{dims:a,dataType:h[0].dataType}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:d}},getShaderSource:f}},Vh=(e,t)=>{Ul(e.inputs),e.compute(Hl(e.inputs,t))},Gh=e=>me({alignCorners:e.align_corners,mode:e.mode,paddingMode:e.padding_mode,format:e.format})}),Kh=P(()=>{"use strict";ee(),ie(),Te(),fa(),ya(),ne(),$t(),Oe=(e,t)=>e.length>t&&e[t].dims.length>0?e[t]:void 0,Kl=(e,t)=>{let r=e[0],i=Oe(e,1),n=Oe(e,2),a=Oe(e,3),s=Oe(e,4),u=Oe(e,5),l=Oe(e,6),d=Oe(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 f=r.dims[0],h=r.dims[1],g=r.dims.length===3?r.dims[2]:t.numHeads*r.dims[4],w=h,y=0,v=0,T=Math.floor(g/t.numHeads);if(l&&d&&R.size(l.dims)&&R.size(d.dims)){if(l.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(l.dims[0]!==f||l.dims[1]!==t.numHeads||l.dims[3]!==T)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(d.dims[0]!==f||d.dims[1]!==t.numHeads||d.dims[3]!==T)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(l.dims[2]!==d.dims[2])throw new Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(d.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');y=l.dims[2],v=l.dims[2]}else if(l&&R.size(l.dims)||d&&R.size(d.dims))throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $;if(i&&R.size(i.dims)>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(i.dims.length<3||i.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==i.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(i.dims.length===3){if(i.dims[2]!==r.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');$=2,w=i.dims[1]}else if(i.dims.length===5){if(i.dims[2]!==t.numHeads||i.dims[3]!==2||i.dims[4]!==T)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(n)throw new Error('Expect "value" be none when "key" has packed kv format.');$=5,w=i.dims[1]}else{if(i.dims[1]!==t.numHeads||i.dims[3]!==T)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,w=i.dims[2]}}else{if(r.dims.length!==5)throw new Error('Input "query" is expected to have 5 dimensions when key is empty');if(r.dims[2]!==t.numHeads||r.dims[3]!==3)throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');$=3}if(a&&R.size(a.dims)>0){if(a.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(i&&i.dims.length===5&&i.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let _=y+w,I=0;if(s&&R.size(s.dims)>0){I=8;let O=s.dims;throw O.length===1?O[0]===f?I=1:O[0]===3*f+2&&(I=3):O.length===2&&O[0]===f&&O[1]===_&&(I=5),I===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let k=!1,E=g;if(n&&R.size(n.dims)>0){if(n.dims.length!==3&&n.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(n.dims.length===3){if(w!==n.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');E=n.dims[2]}else{if(w!==n.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');E=n.dims[1]*n.dims[3],k=!0}}let C=!1;if(s&&R.size(s.dims)>0)throw new Error("Key padding mask is not supported");if(u&&R.size(u.dims)>0){if(u.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(u.dims[0]!==f||u.dims[1]!==t.numHeads||u.dims[2]!==h||u.dims[3]!==_)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:f,sequenceLength:h,pastSequenceLength:y,kvSequenceLength:w,totalSequenceLength:_,maxSequenceLength:v,inputHiddenSize:0,hiddenSize:g,vHiddenSize:E,headSize:T,vHeadSize:Math.floor(E/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:I,scale:t.scale,broadcastResPosBias:C,passPastInKv:k,qkvFormat:$}},Fh=e=>me({...e}),vn=me({perm:[0,2,1,3]}),jl=(e,t,r,i,n,a,s)=>{let u=[i,n,a],l=R.size(u),d=[{type:12,data:l},{type:12,data:s},{type:12,data:a}],f=h=>{let g=F("qkv_with_bias",t.dataType,u),w=B("qkv",t.dataType,u),y=B("bias",r.dataType,u),v=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
+  ${h.registerUniforms(v).declareVariables(w,y,g)}
+  ${h.mainStart()}
+    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;
+
+    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
+  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:d}),getShaderSource:f},{inputs:[t,r],outputs:[-1]})[0]},br=(e,t,r,i,n,a,s,u)=>{let l=a;if(s&&R.size(s.dims)>0){if(i===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=jl(e,a,s,t,i,r*n,u),l=l.reshape([t,i,r,n]),r===1||i===1?l:e.compute(De(l,vn.perm),{inputs:[l],outputs:[-1]})[0]}else return a.dims.length===3&&(l=a.reshape([t,i,r,n])),r===1||i===1?l:e.compute(De(l,vn.perm),{inputs:[l],outputs:[-1]})[0]},Hh=(e,t)=>{let r=Kl(e.inputs,t),i=e.inputs[0],n=Oe(e.inputs,1),a=Oe(e.inputs,2),s=Oe(e.inputs,3),u=Oe(e.inputs,4),l=Oe(e.inputs,5),d=Oe(e.inputs,6),f=Oe(e.inputs,7);if(i.dims.length===5)throw new Error("Packed QKV is not implemented");if(n?.dims.length===5)throw new Error("Packed KV is not implemented");let h=n&&a&&n.dims.length===4&&a.dims.length===4,g=br(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,i,s,0);if(h)return Tr(e,g,n,a,u,void 0,d,f,l,r);if(!n||!a)throw new Error("key and value must be provided");let w=br(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,n,s,r.hiddenSize),y=br(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,a,s,2*r.hiddenSize);Tr(e,g,w,y,u,void 0,d,f,l,r)}}),Zh=P(()=>{"use strict";ee(),ie(),Te(),ne(),Ql=e=>{if(!e||e.length<1)throw new Error("too few inputs")},Zl=(e,t)=>{let r=[],i=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),i=r.length),me({numOutputs:i,axis:t.axis,splitSizes:r})},Yl=e=>`
+fn calculateOutputIndex(index: u32) -> u32 {
+    for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
+    if (index < ${K("uniforms.size_in_split_axis","i",e)}) {
+        return i;
+    }
+    }
+    return ${e}u;
+}`,Xl=e=>{let t=e.length,r=[];for(let i=0;i<t;++i){let n=e[i].setByIndices("indices","input[global_idx]");t===1?r.push(n):i===0?r.push(`if (output_number == ${i}u) { ${n} }`):i===t-1?r.push(`else { ${n} }`):r.push(`else if (output_number == ${i}) { ${n} }`)}return`
+      fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
+        ${r.join(`
+`)}
+      }`},Jn=(e,t)=>{let r=e[0].dims,i=R.size(r),n=e[0].dataType,a=R.normalizeAxis(t.axis,r.length),s=new Array(t.numOutputs),u=B("input",n,r.length),l=new Array(t.numOutputs),d=[],f=[],h=0,g=[{type:12,data:i}];for(let y=0;y<t.numOutputs;y++){h+=t.splitSizes[y],l[y]=h;let v=r.slice();v[a]=t.splitSizes[y],f.push(v),s[y]=F(`output${y}`,n,v.length),d.push({dims:f[y],dataType:e[0].dataType})}g.push({type:12,data:l},...Z(r,...f));let w=y=>`
+  ${y.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",l.length).declareVariables(u,...s)}
+  ${Yl(l.length)}
+  ${Xl(s)}
+
+  ${y.mainStart()}
+    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}
+
+    var indices = ${u.offsetToIndices("global_idx")};
+    var index = ${u.indicesGet("indices",a)};
+    let output_number = calculateOutputIndex(index);
+    if (output_number != 0) {
+      index -= ${K("uniforms.size_in_split_axis","output_number - 1u",l.length)};
+      ${u.indicesSet("indices",a,"index")};
+    }
+    writeBufferData(output_number, indices, global_idx);
+  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:w,getRunData:()=>({outputs:d,dispatchGroup:{x:Math.ceil(i/64)},programUniforms:g})}},jh=(e,t)=>{Ql(e.inputs);let r=e.inputs.length===1?t:Zl(e.inputs,t);e.compute(Jn(e.inputs,r),{inputs:[0]})},Qh=e=>{let t=e.axis,r=e.splitSizes,i=e.numOutputs<0?r.length:e.numOutputs;if(i!==r.length)throw new Error("numOutputs and splitSizes length must be equal");return me({axis:t,numOutputs:i,splitSizes:r})}}),Xh=P(()=>{"use strict";ee(),ie(),Te(),ne(),Jl=(e,t)=>{let[r,i,n,a]=e,{numHeads:s,rotaryEmbeddingDim:u}=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(!R.areEqual(i.dims,[])&&!R.areEqual(i.dims,[1])&&i.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${i.dims.length}`);if(n.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${n.dims.length}`);if(a.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${a.dims.length}`);if(!R.areEqual(n.dims,a.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(u>0&&s===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let l=r.dims[0],d=r.dims[r.dims.length-2],f=n.dims[0],h=R.sizeFromDimension(r.dims,1)/d,g=u===0?n.dims[1]*2:h/s;if(u>g)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(i.dims.length===2){if(l!==i.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${i.dims[0]}`);if(d!==i.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${i.dims[1]}`)}if(g/2!==n.dims[1]&&u/2!==n.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${n.dims[1]}`);if(d>f)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},li=(e,t)=>{let{interleaved:r,numHeads:i,rotaryEmbeddingDim:n,scale:a}=t,s=e[0].dims[0],u=R.sizeFromDimension(e[0].dims,1),l=e[0].dims[e[0].dims.length-2],d=u/l,f=e[2].dims[1],h=n===0?f*2:d/i,g=new Array(s,l,d/h,h-f),w=R.computeStrides(g),y=[{type:1,data:a},{type:12,data:g},{type:12,data:w},...e[0].dims.length===3?new Array({type:12,data:[u,d,h,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[u,h,l*h,1]}):[],...Z(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],v=T=>{let $=B("input",e[0].dataType,e[0].dims.length),_=B("position_ids",e[1].dataType,e[1].dims.length),I=B("cos_cache",e[2].dataType,e[2].dims.length),k=B("sin_cache",e[3].dataType,e[3].dims.length),E=F("output",e[0].dataType,e[0].dims.length);return T.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:g.length},{name:"global_strides",type:"u32",length:w.length},{name:"input_output_strides",type:"u32",length:w.length}]),`
+        ${T.declareVariables($,_,I,k,E)}
+
+        ${T.mainStart(Zt)}
+          let half_rotary_emb_dim = uniforms.${I.name}_shape[1];
+          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
+          let size = uniforms.global_shape[0] * uniforms.global_strides[0];
+          ${T.guardAgainstOutOfBoundsWorkgroupSizes("size")}
+
+          if (bsnh[3] < half_rotary_emb_dim) {
+            let position_ids_idx =
+                ${_.broadcastedIndicesToOffset("bsnh.xy",F("",_.type.tensor,2))};
+            let position_id =
+                u32(${_.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 = ${$.getByOffset("i")} * ${I.get("position_id","bsnh[3]")} -
+                ${$.getByOffset("j")} * ${k.get("position_id","bsnh[3]")};
+            ${E.setByOffset("i","re")}
+            let im = ${$.getByOffset("i")} * ${k.get("position_id","bsnh[3]")} +
+                ${$.getByOffset("j")} * ${I.get("position_id","bsnh[3]")};
+            ${E.setByOffset("j","im")}
+          } else {
+            let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;
+            ${E.setByOffset("k",$.getByOffset("k"))}
+          }
+        }`};return{name:"RotaryEmbedding",shaderCache:{hint:me({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:v,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(R.size(g)/Zt)},programUniforms:y})}},Yh=(e,t)=>{Jl(e.inputs,t),e.compute(li(e.inputs,t))}}),x0=P(()=>{"use strict";Te(),ee(),ya(),Kh(),Zh(),$t(),Xh(),ne(),ed=(e,t)=>{if(t.doRotary&&e.length<=7)throw new Error("cos_cache and sin_cache inputs are required if do_rotary is specified");let r=e[0],i=e[1],n=e[2],a=e[3],s=e[4];if(t.doRotary!==0&&e.length<=7)throw new Error("cos_cast and sin_cache are expected if do_rotary attribute is non-zero");if(t.localWindowSize!==-1)throw new Error("Local attention is not supported");if(t.softcap!==0)throw new Error("Softcap is not supported");if(t.rotaryInterleaved!==0)throw new Error("Rotary interleaved is not supported");if(t.smoothSoftmax)throw new Error("Smooth softmax is not supported");if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let u=!1,l=r.dims[0],d=r.dims[1],f=r.dims.length===3?u?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],h=d,g=0,w=!i||i.dims.length===0,y=Math.floor(w?f/(t.numHeads+2*t.kvNumHeads):f/t.numHeads);w&&(f=y*t.numHeads);let v=a&&a.dims.length!==0,T=s&&s.dims.length!==0;if(v&&a.dims.length===4&&a.dims[0]===l&&a.dims[1]!==t.kvNumHeads&&a.dims[2]===t.kvNumHeads&&a.dims[3]===y)throw new Error("BSNH pastKey/pastValue is not supported");if(v&&T){if(a.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(s.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');g=a.dims[2]}else if(v||T)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $=1;if(i&&i.dims.length>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(i.dims.length<3||i.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==i.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(i.dims.length===3){if(r.dims[2]%i.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');h=i.dims[1]}else if(i.dims.length===5){if(i.dims[2]!==t.numHeads||i.dims[3]!==2||i.dims[4]!==y)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(n)throw new Error('Expect "value" be none when "key" has packed kv format.');h=i.dims[1]}else{if(i.dims[1]!==t.numHeads||i.dims[3]!==y)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');h=i.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');$=3}let _=0,I=!1,k=t.kvNumHeads?y*t.kvNumHeads:f;if(n&&n.dims.length>0){if(n.dims.length!==3&&n.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(n.dims.length===3){if(h!==n.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');k=n.dims[2]}else{if(h!==n.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');k=n.dims[1]*n.dims[3],I=!0}}let E=e.length>4?e[5]:void 0;if(E&&E.dims.length!==1&&E.dims[0]!==l)throw new Error('Input "seqlens" is expected to have 1 dimension and the same dim 0 as batch_size');return{batchSize:l,sequenceLength:d,pastSequenceLength:g,kvSequenceLength:h,totalSequenceLength:-1,maxSequenceLength:-1,inputHiddenSize:0,hiddenSize:f,vHiddenSize:k,headSize:y,vHeadSize:Math.floor(k/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:_,scale:t.scale,broadcastResPosBias:!1,passPastInKv:I,qkvFormat:$}},td=me({perm:[0,2,1,3]}),$n=(e,t,r)=>{let i=t,n=r.kvNumHeads;return t.dims.length===3&&r.kvSequenceLength!==0&&(i=t.reshape([r.batchSize,r.kvSequenceLength,n,r.headSize]),i=e.compute(De(i,td.perm),{inputs:[i],outputs:[-1]})[0]),i},rd=(e,t,r,i)=>{let n=7,a=["type","type"],s=[e*t],u=e*t,l=[{type:12,data:u},{type:12,data:t},{type:12,data:e}],d=f=>{let h=B("seq_lens",r.dataType,r.dims),g=B("total_seq_lens",i.dataType,i.dims),w=F("pos_ids",n,s),y=[{name:"output_size",type:"u32"},{name:"sequence_length",type:"u32"},{name:"batch_size",type:"u32"}];return`
+  ${f.registerUniforms(y).declareVariables(h,g,w)}
+  ${f.mainStart()}
+    ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+    let total_sequence_length = u32(${g.getByOffset("0")});
+    let is_subsequent_prompt = uniforms.sequence_length > 1 && uniforms.sequence_length != total_sequence_length;
+    let is_first_prompt = !is_subsequent_prompt && uniforms.sequence_length == total_sequence_length;
+    let batch_idx = global_idx / uniforms.sequence_length;
+    let sequence_idx = i32(global_idx % uniforms.sequence_length);
+    var pos_id: i32 = 0;
+    let seqlen = ${h.getByOffset("batch_idx")};
+    let total_seqlen = seqlen + 1;
+    if (is_first_prompt) {
+      if (sequence_idx < total_seqlen) {
+        pos_id = sequence_idx;
+      } else {
+        pos_id = 1;
+      }
+      ${w.setByOffset("global_idx","pos_id")}
+    } else if (is_subsequent_prompt) {
+      let past_seqlen = total_seqlen - i32(uniforms.sequence_length);
+      if (past_seqlen + sequence_idx < total_seqlen) {
+        pos_id = past_seqlen + sequence_idx;
+      } else {
+        pos_id = 1;
+      }
+      ${w.setByOffset("global_idx","pos_id")}
+    } else if (global_idx < uniforms.batch_size) {
+      ${w.setByOffset("global_idx","seqlen")}
+    };
+  }
+  `};return{name:"GeneratePositionIds",shaderCache:{hint:`${e};${t}`,inputDependencies:a},getRunData:()=>({outputs:[{dims:s,dataType:n}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l}),getShaderSource:d}},Jh=(e,t)=>{let r=ed(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(e.inputs[1]?.dims.length===5)throw new Error("Packed KV is not implemented");let i=e.inputs[0],n=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,a=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,s=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,u=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,l=e.inputs.length>4?e.inputs[5]:void 0,d=e.inputs.length>5?e.inputs[6]:void 0,f=r.kvNumHeads?r.kvNumHeads:r.numHeads,h=me({axis:2,numOutputs:3,splitSizes:[r.numHeads*r.headSize,f*r.headSize,f*r.headSize]}),[g,w,y]=!n&&!a?e.compute(Jn([i],h),{inputs:[i],outputs:[-1,-1,-1]}):[i,n,a],v,T;if(t.doRotary){let k=e.compute(rd(r.batchSize,r.sequenceLength,l,d),{inputs:[l,d],outputs:[-1]})[0],E=e.inputs[7],C=e.inputs[8],O=me({interleaved:t.rotaryInterleaved!==0,numHeads:r.numHeads,rotaryEmbeddingDim:0,scale:t.scale}),x=[g,k,E,C],D=[-1];v=e.compute(li(x,O),{inputs:x,outputs:D})[0],x.splice(0,1,w);let L=me({interleaved:t.rotaryInterleaved!==0,numHeads:r.kvNumHeads,rotaryEmbeddingDim:0,scale:t.scale});T=e.compute(li(x,L),{inputs:x,outputs:D})[0]}let $=br(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,t.doRotary?v:g,void 0,0),_=$n(e,t.doRotary?T:w,r),I=$n(e,y,r);Tr(e,$,_,I,void 0,void 0,s,u,void 0,r,l,d)}}),T0=P(()=>{"use strict";ee(),ie(),$t(),ne(),xn=(e,t,r,i,n,a,s,u)=>{let l=xe(a),d=l===1?"f32":`vec${l}f`,f=l===1?"vec2f":`mat2x${l}f`,h=n*s,g=64;h===1&&(g=256);let w=[n,s,a/l],y=[n,s,2],v=["rank","type","type"],T=[];T.push(...Z(w,y));let $=_=>{let I=B("x",t.dataType,3,l),k=B("scale",r.dataType,r.dims),E=B("bias",i.dataType,i.dims),C=F("output",1,3,2),O=[I,k,E,C];return`
+  var<workgroup> workgroup_shared : array<${f}, ${g}>;
+  const workgroup_size = ${g}u;
+  ${_.declareVariables(...O)}
+  ${_.mainStart(g)}
+    let batch = workgroup_index / uniforms.x_shape[1];
+    let channel = workgroup_index % uniforms.x_shape[1];
+    let hight = uniforms.x_shape[2];
+    // initialize workgroup memory
+    var sum = ${d}(0);
+    var squared_sum = ${d}(0);
+    for (var h = local_idx; h < hight; h += workgroup_size) {
+      let value = ${d}(${I.get("batch","channel","h")});
+      sum += value;
+      squared_sum += value * value;
+    }
+    workgroup_shared[local_idx] = ${f}(sum, squared_sum);
+    workgroupBarrier();
+
+    for (var currSize = workgroup_size >> 1;  currSize > 0; currSize = currSize >> 1) {
+      if (local_idx < currSize) {
+        workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize];
+      }
+      workgroupBarrier();
+    }
+    if (local_idx == 0) {
+      let sum_final = ${vt("workgroup_shared[0][0]",l)} / f32(hight * ${l});
+      let squared_sum_final = ${vt("workgroup_shared[0][1]",l)} / f32(hight * ${l});
+
+      let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${u}));
+      let channel_scale = inv_std_dev * f32(scale[channel]);
+      let channel_shift = f32(bias[channel]) - sum_final * channel_scale;
+      output[workgroup_index] = vec2f(channel_scale, channel_shift);
+    }
+  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${u};${g}`,inputDependencies:v},getRunData:()=>({outputs:[{dims:y,dataType:1}],dispatchGroup:{x:h},programUniforms:T}),getShaderSource:$},{inputs:[t,r,i],outputs:[-1]})[0]},id=(e,t,r)=>{let i=t[0].dims,n=i,a=2,s=i[0],u=i[1],l=R.sizeFromDimension(i,a),d=xe(l),f=R.size(n)/d,h=xn(e,t[0],t[1],t[2],s,l,u,r.epsilon),g=[s,u,l/d],w=[s,u],y=["type","none"],v=T=>{let $=B("x",t[0].dataType,g.length,d),_=B("scale_shift",1,w.length,2),I=F("output",t[0].dataType,g.length,d),k=[$,_,I];return`
+  ${T.registerUniform("output_size","u32").declareVariables(...k)}
+  ${T.mainStart()}
+  ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+      let outputIndices = ${I.offsetToIndices("global_idx")};
+      let batch = outputIndices[0];
+      let channel = outputIndices[1];
+      let scale_shift = ${_.getByIndices("vec2<u32>(batch, channel)")};
+      let value = ${$.getByOffset("global_idx")} * ${I.type.value}(scale_shift.x) + ${I.type.value}(scale_shift.y);
+      ${I.setByOffset("global_idx","value")};
+  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${d}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:[{type:12,data:f},...Z(g,w,g)]}),getShaderSource:v},{inputs:[t[0],h]})},nd=(e,t,r)=>{let i=t[0].dims,n=i,a=i[0],s=i[i.length-1],u=R.sizeFromDimension(i,1)/s,l=xe(s),d=R.size(n)/l,f=[{type:12,data:u},{type:12,data:Math.floor(s/l)}],h=["type","type"],g=!1,w=[0,i.length-1];for(let $=0;$<i.length-2;$++)g=g||i[$+1]!==1,w.push($+1);g=g&&i[i.length-1]!==1;let y=g?e.compute(De(e.inputs[0],w),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:i.length},($,_)=>i[w[_]])),v=xn(e,y,t[1],t[2],a,u,s,r.epsilon),T=$=>{let _=Ee(t[0].dataType),I=l===1?"vec2f":`mat${l}x2f`,k=O=>{let x=O===0?"x":"y",D=l===1?"f32":`vec${l}f`;switch(l){case 1:return`${_}(${D}(scale.${x}))`;case 2:return`vec2<${_}>(${D}(scale[0].${x}, scale[1].${x}))`;case 4:return`vec4<${_}>(${D}(scale[0].${x}, scale[1].${x}, scale[2].${x}, scale[3].${x}))`;default:throw new Error(`Not supported compoents ${l}`)}},E=B("input",t[0].dataType,t[0].dims,l),C=F("output",t[0].dataType,n,l);return`
+  @group(0) @binding(0) var<storage, read> input : array<${E.type.storage}>;
+  @group(0) @binding(1) var<storage, read> scale_input : array<${I}>;
+  @group(0) @binding(2) var<storage, read_write> output : array<${C.type.storage}>;
+  struct Uniforms {H: u32, C : u32};
+  @group(0) @binding(3) var<uniform> uniforms: Uniforms;
+
+  ${$.mainStart()}
+    let current_image_number = global_idx / (uniforms.C * uniforms.H);
+    let current_channel_number = global_idx % uniforms.C;
+
+    let scale_offset = current_image_number * uniforms.C + current_channel_number;
+    let scale = scale_input[scale_offset];
+    output[global_idx] = fma(input[global_idx], ${k(0)}, ${k(1)});
+  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${l}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:f}),getShaderSource:T},{inputs:[t[0],v]})},ef=(e,t)=>{t.format==="NHWC"?nd(e,e.inputs,t):id(e,e.inputs,t)}}),S0=P(()=>{"use strict";ee(),ie(),ne(),ad=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},sd=(e,t,r)=>{let i=t.simplified,n=e[0].dims,a=e[1],s=!i&&e[2],u=n,l=R.normalizeAxis(t.axis,n.length),d=R.sizeToDimension(n,l),f=R.sizeFromDimension(n,l),h=R.size(a.dims),g=s?R.size(s.dims):0;if(h!==f||s&&g!==f)throw new Error(`Size of X.shape()[axis:] == ${f}.
+       Size of scale and bias (if provided) must match this.
+       Got scale size of ${h} and bias size of ${g}`);let w=[];for(let E=0;E<n.length;++E)E<l?w.push(n[E]):w.push(1);let y=xe(f),v=["type","type"],T=[{type:12,data:d},{type:1,data:f},{type:12,data:Math.floor(f/y)},{type:1,data:t.epsilon}];s&&v.push("type");let $=r>1,_=r>2,I=E=>{let C=Ee(e[0].dataType),O=[B("x",e[0].dataType,e[0].dims,y),B("scale",a.dataType,a.dims,y)];s&&O.push(B("bias",s.dataType,s.dims,y)),O.push(F("output",e[0].dataType,u,y)),$&&O.push(F("mean_data_output",1,w)),_&&O.push(F("inv_std_output",1,w));let x=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
+  ${E.registerUniforms(x).declareVariables(...O)}
+  ${E.mainStart()}
+    ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
+    let offset = global_idx * uniforms.norm_size_vectorized;
+    var mean_vector = ${Gn("f32",y)};
+    var mean_square_vector = ${Gn("f32",y)};
+
+    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
+      let value = ${jt(C,y,"x[h + offset]")};
+      mean_vector += value;
+      mean_square_vector += value * value;
+    }
+    let mean = ${vt("mean_vector",y)} / uniforms.norm_size;
+    let inv_std_dev = inverseSqrt(${vt("mean_square_vector",y)} / uniforms.norm_size ${i?"":"- mean * mean"} + uniforms.epsilon);
+
+    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
+      let f32input = ${jt(C,y,"x[j + offset]")};
+      let f32scale = ${jt(C,y,"scale[j]")};
+      output[j + offset] = ${O[0].type.value}((f32input ${i?"":"- mean"}) * inv_std_dev * f32scale
+        ${s?`+ ${jt(C,y,"bias[j]")}`:""}
+      );
+    }
+
+    ${$?"mean_data_output[global_idx] = mean":""};
+    ${_?"inv_std_output[global_idx] = inv_std_dev":""};
+  }`},k=[{dims:u,dataType:e[0].dataType}];return $&&k.push({dims:w,dataType:1}),_&&k.push({dims:w,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${y};${r};${i}`,inputDependencies:v},getRunData:()=>({outputs:k,dispatchGroup:{x:Math.ceil(d/64)},programUniforms:T}),getShaderSource:I}},tf=(e,t)=>{ad(e.inputs),e.compute(sd(e.inputs,t,e.outputCount))}}),k0=P(()=>{"use strict";ie(),$a(),xa(),od=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},rf=e=>{od(e.inputs);let t=Qt.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can't use matmul on the given tensors");let r=t[t.length-1],i=e.inputs[0].dims[e.inputs[0].dims.length-1];if(r<8&&i<8)e.compute(va(e.inputs,{activation:""},t));else{let n=t[t.length-2],a=R.size(e.inputs[0].dims.slice(0,-2)),s=R.size(e.inputs[1].dims.slice(0,-2));if(a!==1&&n===1&&s===1){let u=e.inputs[0].reshape([1,a,i]),l=e.inputs[1].reshape([1,i,r]),d=[1,a,r],f=[u,l];e.compute(ui(f,{activation:""},t,d),{inputs:f})}else e.compute(ui(e.inputs,{activation:""},t))}}}),I0=P(()=>{"use strict";ee(),ie(),Te(),ne(),ud=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],i=r.dims.length;if(r.dims[i-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let n=Math.floor((t.k+t.blockSize-1)/t.blockSize),a=t.blockSize/8*t.bits,s=e[1];if(!R.areEqual(s.dims,[t.n,n,a]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let u=e[2].dims;if(R.size(u)!==t.n*n)throw new Error("scales input size error.");if(e.length===4){let l=e[3].dims,d=t.n*(t.bits===8?n:Math.floor((n*t.bits+7)/8));if(R.size(l)!==d)throw new Error("zeroPoints input size error.")}},ld=(e,t)=>{let r=e[0].dims,i=r.length,n=r[i-2],a=t.k,s=t.n,u=r.slice(0,i-2),l=R.size(u),d=e[1].dims[2]/4,f=e[0].dataType,h=xe(t.k),g=xe(d),w=xe(s),y=u.concat([n,s]),v=n>1&&s/w%2===0?2:1,T=R.size(y)/w/v,$=64,_=[],I=[l,n,a/h],k=R.convertShape(e[1].dims).slice();k.splice(-1,1,d/g),_.push(...Z(I)),_.push(...Z(k)),_.push(...Z(e[2].dims)),e.length===4&&_.push(...Z(R.convertShape(e[3].dims)));let E=[l,n,s/w];_.push(...Z(E));let C=O=>{let x=I.length,D=B("a",e[0].dataType,x,h),L=B("b",12,k.length,g),Y=B("scales",e[2].dataType,e[2].dims.length),q=[D,L,Y],j=e.length===4?B("zero_points",12,e[3].dims.length):void 0;j&&q.push(j);let ue=E.length,A=F("output",e[0].dataType,ue,w),U=Ee(e[0].dataType),re=(()=>{switch(h){case 1:return`array<${U}, 8>`;case 2:return`mat4x2<${U}>`;case 4:return`mat2x4<${U}>`;default:throw new Error(`${h}-component is not supported.`)}})(),J=()=>{let N=`
+          // reuse a data
+            var input_offset = ${D.indicesToOffset(`${D.type.indices}(batch, row, word_offset)`)};
+            var a_data: ${re};
+            for (var j: u32 = 0; j < ${8/h}; j++) {
+              a_data[j] = ${D.getByOffset("input_offset")};
+              input_offset++;
+            }
+          `;for(let V=0;V<w*v;V++)N+=`
+            b_value = ${g===1?`b${V}_data`:`b${V}_data[i]`};
+            b_value_lower = unpack4xU8(b_value & b_mask);
+            b_value_upper = unpack4xU8((b_value >> 4) & b_mask);
+            b_quantized_values = ${re}(${Array.from({length:4},(te,ce)=>`${U}(b_value_lower[${ce}]), ${U}(b_value_upper[${ce}])`).join(", ")});
+            b_dequantized_values = ${h===1?`${re}(${Array.from({length:8},(te,ce)=>`(b_quantized_values[${ce}] - ${j?`zero_point${V}`:"zero_point"}) * scale${V}`).join(", ")});`:`(b_quantized_values - ${re}(${Array(8).fill(`${j?`zero_point${V}`:"zero_point"}`).join(",")})) * scale${V};`};
+            workgroup_shared[local_id.x * ${v} + ${Math.floor(V/w)}]${w>1?`[${V%w}]`:""} += ${Array.from({length:8/h},(te,ce)=>`${h===1?`a_data[${ce}] * b_dequantized_values[${ce}]`:`dot(a_data[${ce}], b_dequantized_values[${ce}])`}`).join(" + ")};
+          `;return N},H=()=>{let N=`
+            var col_index = col * ${w};
+            ${j?`
+            let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2;
+            var zero_point_byte_count: u32;
+            var zero_point_word_index: u32;
+            var zero_point_byte_offset: u32;
+            let zero_point_nibble_offset: u32 = block & 0x1u;
+            var zero_point_bits_offset: u32;
+            var zero_point_word: u32;`:`
+            // The default zero point is 8 for unsigned 4-bit quantization.
+            let zero_point = ${U}(8);`}
+            `;for(let V=0;V<w*v;V++)N+=`
+            let scale${V} = ${Y.getByOffset("col_index * nBlocksPerCol + block")};
+            ${j?`
+            zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u);
+            zero_point_word_index = zero_point_byte_count >> 0x2u;
+            zero_point_byte_offset = zero_point_byte_count & 0x3u;
+            zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
+            zero_point_word = ${j.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
+            let zero_point${V} = ${U}((zero_point_word) & 0xFu);`:""}
+            col_index += 1;`;return N},pe=()=>{let N=`col_index = col * ${w};`;for(let V=0;V<w*v;V++)N+=`
+            let b${V}_data = ${L.getByIndices(`${L.type.indices}(col_index, block, word)`)};
+            col_index += 1;`;return N+=`
+            var b_value: u32;
+            let b_mask: u32 = 0x0F0F0F0Fu;
+            var b_value_lower: vec4<u32>;
+            var b_value_upper: vec4<u32>;
+            var b_quantized_values: ${re};
+            var b_dequantized_values: ${re};`,N};return`
+        var<workgroup> workgroup_shared: array<${A.type.value}, ${v*$}>;
+        ${O.declareVariables(...q,A)}
+        ${O.mainStart([$,1,1])}
+          let output_indices = ${A.offsetToIndices(`(global_idx / ${$}) * ${v}`)};
+          let col = output_indices[2];
+          let row = output_indices[1];
+          let batch = output_indices[0];
+          let nBlocksPerCol = uniforms.b_shape[1];
+
+          for (var block = local_id.x; block < nBlocksPerCol; block += ${$}) {
+            //process one block
+            var word_offset: u32 = block * ${t.blockSize/h};
+            ${H()}
+            for (var word: u32 = 0; word < ${d}; word += ${g}) {
+              ${pe()}
+              for (var i: u32 = 0; i < ${g}; i++) {
+                ${J()}
+                word_offset += ${8/h};
+              }
+            }
+          }
+          workgroupBarrier();
+
+          if (local_id.x < ${v}) {
+            var output_value: ${A.type.value} = ${A.type.value}(0);
+            var workgroup_shared_offset: u32 = local_id.x;
+            for (var b: u32 = 0u; b < ${$}u; b++) {
+              output_value += workgroup_shared[workgroup_shared_offset];
+              workgroup_shared_offset += ${v};
+            }
+            ${A.setByIndices(`${A.type.indices}(batch, row, col + local_id.x)`,"output_value")};
+          }
+        }`};return{name:"MatMulNBits",shaderCache:{hint:`${t.blockSize};${t.bits};${h};${g};${w};${v};${$}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:y,dataType:f}],dispatchGroup:{x:T},programUniforms:_}),getShaderSource:C}},dd=(e,t)=>{let r=e[0].dims,i=r.length,n=r[i-2],a=t.k,s=t.n,u=r.slice(0,i-2),l=R.size(u),d=e[1].dims[2]/4,f=e[0].dataType,h=xe(t.k),g=xe(d),w=u.concat([n,s]),y=128,v=s%8===0?8:s%4===0?4:1,T=y/v,$=T*g*8,_=$/h,I=$/t.blockSize,k=R.size(w)/v,E=[],C=[l,n,a/h],O=R.convertShape(e[1].dims).slice();O.splice(-1,1,d/g),E.push(...Z(C)),E.push(...Z(O)),E.push(...Z(e[2].dims)),e.length===4&&E.push(...Z(R.convertShape(e[3].dims)));let x=[l,n,s];E.push(...Z(x));let D=L=>{let Y=C.length,q=B("a",e[0].dataType,Y,h),j=B("b",12,O.length,g),ue=B("scales",e[2].dataType,e[2].dims.length),A=[q,j,ue],U=e.length===4?B("zero_points",12,e[3].dims.length):void 0;U&&A.push(U);let re=x.length,J=F("output",e[0].dataType,re),H=Ee(e[0].dataType),pe=()=>{switch(h){case 1:return`
+          let a_data0 = vec4<${H}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]);
+          let a_data1 = vec4<${H}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return`
+          let a_data0 = vec4<${H}>(sub_a[word_offset], sub_a[word_offset + 1]);
+          let a_data1 = vec4<${H}>(sub_a[word_offset + 2], sub_a[word_offset + 3]);`;case 4:return`
+          let a_data0 = sub_a[word_offset];
+          let a_data1 = sub_a[word_offset + 1];`;default:throw new Error(`${h}-component is not supported.`)}};return`
+        var<workgroup> sub_a: array<${q.type.value}, ${_}>;
+        var<workgroup> inter_results: array<array<${J.type.value}, ${T}>, ${v}>;
+        ${L.declareVariables(...A,J)}
+        ${L.mainStart([T,v,1])}
+          let output_indices = ${J.offsetToIndices(`workgroup_index * ${v}`)};
+          let col = output_indices[2];
+          let row = output_indices[1];
+          let batch = output_indices[0];
+          let n_blocks_per_col = uniforms.b_shape[1];
+          let num_tiles =  (n_blocks_per_col - 1) / ${I} + 1;
+
+          // Loop over shared dimension.
+          for (var tile: u32 = 0; tile < num_tiles; tile += 1) {
+            let a_col_start = tile * ${_};
+            // load one tile A data into shared memory.
+            for (var a_offset = local_idx; a_offset < ${_}; a_offset += ${y})
+            {
+              let a_col = a_col_start + a_offset;
+              if (a_col < uniforms.a_shape[2])
+              {
+                sub_a[a_offset] = ${q.getByIndices(`${q.type.indices}(batch, row, a_col)`)};
+              } else {
+                sub_a[a_offset] = ${q.type.value}(0);
+              }
+            }
+            workgroupBarrier();
+
+            // each thread process one block
+            let b_row = col + local_id.y;
+            let block = tile * ${I} + local_id.x;
+            ${U?`
+            let zero_point_bytes_per_col = (n_blocks_per_col + 1) / 2;
+            let zero_point_byte_count = b_row * zero_point_bytes_per_col + (block >> 0x1u);
+            let zero_point_word_index = zero_point_byte_count >> 0x2u;
+            let zero_point_byte_offset = zero_point_byte_count & 0x3u;
+            let zero_point_nibble_offset: u32 = block & 0x1u;
+            let zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
+            let zero_point_word = ${U.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
+            let zero_point = ${H}((zero_point_word) & 0xFu);`:`
+            // The default zero point is 8 for unsigned 4-bit quantization.
+            let zero_point = ${H}(8);`}
+            let scale = ${ue.getByOffset("b_row * n_blocks_per_col + block")};
+            let b_data = ${j.getByIndices(`${j.type.indices}(b_row, block, 0)`)};
+            var word_offset = local_id.x * ${t.blockSize/h};
+            for (var i: u32 = 0; i < ${g}; i++) {
+              ${pe()}
+              let b_value = ${g===1?"b_data":"b_data[i]"};
+              let b_value_lower = unpack4xU8(b_value & 0x0F0F0F0Fu);
+              let b_value_upper = unpack4xU8((b_value >> 4) & 0x0F0F0F0Fu);
+              let b_quantized_values = mat2x4<${H}>(${Array.from({length:4},(N,V)=>`${H}(b_value_lower[${V}]), ${H}(b_value_upper[${V}])`).join(", ")});
+              let b_dequantized_values = (b_quantized_values - mat2x4<${H}>(${Array(8).fill("zero_point").join(",")})) * scale;
+              inter_results[local_id.y][local_id.x] += ${Array.from({length:2},(N,V)=>`${`dot(a_data${V}, b_dequantized_values[${V}])`}`).join(" + ")};
+              word_offset += ${8/h};
+            }
+            workgroupBarrier();
+          }
+
+          if (local_idx < ${v}) {
+            var output_value: ${J.type.value} = ${J.type.value}(0);
+            for (var b = 0u; b < ${T}; b++) {
+              output_value += inter_results[local_idx][b];
+            }
+            if (col + local_idx < uniforms.output_shape[2])
+            {
+              ${J.setByIndices(`${J.type.indices}(batch, row, col + local_idx)`,"output_value")}
+            }
+          }
+        }`};return{name:"BlockwiseMatMulNBits32",shaderCache:{hint:`${t.blockSize};${h};${g};${T};${v}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:w,dataType:f}],dispatchGroup:{x:k},programUniforms:E}),getShaderSource:D}},nf=(e,t)=>{ud(e.inputs,t),t.blockSize===32&&e.adapterInfo.isVendor("intel")&&e.adapterInfo.isArchitecture("gen-12lp")?e.compute(dd(e.inputs,t)):e.compute(ld(e.inputs,t))},af=e=>me(e)}),E0=P(()=>{"use strict";ee(),ie(),ne(),pd=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1&&e[0].dataType!==10)throw new Error("Input type must be float or float16.");if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},cd=(e,t,r)=>{let i="";for(let n=t-1;n>=0;--n)i+=`
+            k = i32(${e.indicesGet("indices",n)}) - ${K("uniforms.pads",n,r)};
+            if (k < 0) {
+              break;
+            }
+            if (k >= i32(${K("uniforms.x_shape",n,t)})) {
+              break;
+            }
+            offset += k * i32(${K("uniforms.x_strides",n,t)});
+        `;return`
+          value = ${e.type.value}(uniforms.constant_value);
+          for (var i = 0; i < 1; i++) {
+            var offset = 0;
+            var k = 0;
+            ${i}
+            value = x[offset];
+          }
+      `},hd=(e,t,r)=>{let i="";for(let n=t-1;n>=0;--n)i+=`
+                k = i32(${e.indicesGet("indices",n)}) - ${K("uniforms.pads",n,r)};
+                if (k < 0) {
+                  k = -k;
+                }
+                {
+                  let _2n_1 = 2 * (i32(${K("uniforms.x_shape",n,t)}) - 1);
+                  k = k % _2n_1;
+                  if(k >= i32(${K("uniforms.x_shape",n,t)})) {
+                    k = _2n_1 - k;
+                  }
+                }
+                offset += k * i32(${K("uniforms.x_strides",n,t)});
+            `;return`
+              var offset = 0;
+              var k = 0;
+              ${i}
+              value = x[offset];
+          `},fd=(e,t,r)=>{let i="";for(let n=t-1;n>=0;--n)i+=`
+                k = i32(${e.indicesGet("indices",n)}) - ${K("uniforms.pads",n,r)};
+                if (k < 0) {
+                  k = 0;
+                }
+                if (k >= i32(${K("uniforms.x_shape",n,t)})) {
+                  k = i32(${K("uniforms.x_shape",n,t)}) - 1;
+                }
+                offset += k * i32(${K("uniforms.x_strides",n,t)});
+            `;return`
+              var offset = 0;
+              var k = 0;
+              ${i}
+              value = x[offset];
+          `},md=(e,t,r)=>{let i="";for(let n=t-1;n>=0;--n)i+=`
+                k = i32(${e.indicesGet("indices",n)}) - ${K("uniforms.pads",n,r)};
+                if (k < 0)  {
+                  k += i32(${K("uniforms.x_shape",n,t)}]);
+                }
+                if (k >= i32(${K("uniforms.x_shape",n,t)})) {
+                  k -= i32(${K("uniforms.x_shape",n,t)});
+                }
+                offset += k * i32(${K("uniforms.x_strides",n,t)});
+            `;return`
+              var offset = 0;
+              var k = 0;
+              ${i}
+              value = x[offset];
+          `},gd=(e,t,r)=>{switch(r.mode){case 0:return cd(e,t,r.pads.length);case 1:return hd(e,t,r.pads.length);case 2:return fd(e,t,r.pads.length);case 3:return md(e,t,r.pads.length);default:throw new Error("Invalid mode")}},yd=(e,t)=>{let r=R.padShape(e[0].dims.slice(),t.pads),i=e[0].dims,n=R.size(r),a=[{type:12,data:n},{type:6,data:t.pads}],s=e.length>=3&&e[2].data;t.mode===0&&a.push({type:s?e[2].dataType:1,data:t.value}),a.push(...Z(e[0].dims,r));let u=["rank"],l=d=>{let f=F("output",e[0].dataType,r.length),h=B("x",e[0].dataType,i.length),g=h.type.value,w=gd(f,i.length,t),y=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&y.push({name:"constant_value",type:s?g:"f32"}),`
+            ${d.registerUniforms(y).declareVariables(h,f)}
+            ${d.mainStart()}
+            ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+
+            let indices = ${f.offsetToIndices("global_idx")};
+
+            var value = ${g}(0);
+            ${w}
+            output[global_idx] = value;
+        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}${s}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(R.size(r)/64)},programUniforms:a}),getShaderSource:l}},wd=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),i=e.length>=3&&e[2].data?e[2].dataType===10?e[2].getUint16Array()[0]:e[2].getFloat32Array()[0]:0,n=e[0].dims.length,a=new Int32Array(2*n).fill(0);if(e.length>=4){let u=e[3].getBigInt64Array();for(let l=0;l<u.length;l++)a[Number(u[l])]=Number(r[l]),a[Number(u[l])+n]=Number(r[l+u.length])}else r.forEach((u,l)=>a[Number(l)]=Number(u));let s=[];return a.forEach(u=>s.push(u)),{mode:t.mode,value:i,pads:s}}else return t},sf=(e,t)=>{pd(e.inputs);let r=wd(e.inputs,t);e.compute(yd(e.inputs,r),{inputs:[0]})}}),C0=P(()=>{"use strict";Le(),ee(),ie(),ne(),hr=e=>{if(be.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Tn=(e,t,r)=>{let i=t.format==="NHWC",n=e.dims.slice();i&&n.splice(1,0,n.pop());let a=Object.hasOwnProperty.call(t,"dilations"),s=t.kernelShape.slice(),u=t.strides.slice(),l=a?t.dilations.slice():[],d=t.pads.slice();si.adjustPoolAttributes(r,n,s,u,l,d);let f=si.computePoolOutputShape(r,n,u,l,s,d,t.autoPad),h=Object.assign({},t);a?Object.assign(h,{kernelShape:s,strides:u,pads:d,dilations:l,cacheKey:t.cacheKey}):Object.assign(h,{kernelShape:s,strides:u,pads:d,cacheKey:t.cacheKey});let g=f.slice();return g.push(g.splice(1,1)[0]),[h,i?g:f]},Sn=(e,t)=>{let r=t.format==="NHWC",i=R.size(e),n=R.size(t.kernelShape),a=[{type:12,data:i},{type:12,data:n}],s=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let u=t.kernelShape[t.kernelShape.length-1],l=t.strides[t.strides.length-1],d=t.pads[t.pads.length/2-1],f=t.pads[t.pads.length-1],h=!!(d+f);a.push({type:12,data:u},{type:12,data:l},{type:12,data:d},{type:12,data:f}),s.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let g=!1;if(t.kernelShape.length===2){let w=t.kernelShape[t.kernelShape.length-2],y=t.strides[t.strides.length-2],v=t.pads[t.pads.length/2-2],T=t.pads[t.pads.length-2];g=!!(v+T),a.push({type:12,data:w},{type:12,data:y},{type:12,data:v},{type:12,data:T}),s.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[a,s,!0,h,g]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let u=R.computeStrides(t.kernelShape);a.push({type:12,data:u},{type:12,data:t.pads},{type:12,data:t.strides}),s.push({name:"kernelStrides",type:"u32",length:u.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let l=t.pads.reduce((d,f)=>d+f);return[a,s,!!l,!1,!1]}},kn=(e,t,r,i,n,a,s,u,l,d,f,h)=>{let g=n.format==="NHWC",w=t.type.value,y=F("output",t.type.tensor,i);if(n.kernelShape.length<=2){let v="",T="",$="",_=r-(g?2:1);if(f?v=`
+                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")}];
+                  ${a}
+                }`:v=`
+                for (var i: u32 = 0u; i < uniforms.kw; i++) {
+                  xIndices[${_}] = indices[${_}] * uniforms.sw - uniforms.pwStart + i;
+                  let x_val = x[${t.indicesToOffset("xIndices")}];
+                  ${a}
+                }`,n.kernelShape.length===2){let I=r-(g?3:2);h?T=`
+                for (var j: u32 = 0u; j < uniforms.kh; j++) {
+                  xIndices[${I}] = indices[${I}] * uniforms.sh - uniforms.phStart + j;
+                  if (xIndices[${I}] < 0 || xIndices[${I}] >= uniforms.x_shape[${I}]) {
+                    pad += i32(uniforms.kw);
+                    continue;
+                  }
+              `:T=`
+                for (var j: u32 = 0u; j < uniforms.kh; j++) {
+                  xIndices[${I}] = indices[${I}] * uniforms.sh - uniforms.phStart + j;
+                `,$=`
+              }
+            `}return`
+            ${e.registerUniforms(l).declareVariables(t,y)}
+
+            ${e.mainStart()}
+              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+
+              let indices = ${y.offsetToIndices("global_idx")};
+              var xIndices = ${y.offsetToIndices("global_idx")};
+
+              var value = ${w}(${u});
+              var pad = 0;
+              ${T}
+              ${v}
+              ${$}
+              ${s}
+
+              output[global_idx] = value;
+            }`}else{if(g)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let v=n.kernelShape.length,T=n.pads.length,$="";return d?$=`
+                if (xIndices[j] >= uniforms.x_shape[j]) {
+                  pad++;
+                  isPad = true;
+                  break;
+                }
+              }
+              if (!isPad) {
+                let x_val = x[${t.indicesToOffset("xIndices")}];
+                ${a}
+              }`:$=`
+              }
+              let x_val = x[${t.indicesToOffset("xIndices")}];
+              ${a}
+            `,`
+            ${e.registerUniforms(l).declareVariables(t,y)}
+
+            ${e.mainStart()}
+              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+              let indices = ${y.offsetToIndices("global_idx")};
+              var xIndices = ${y.offsetToIndices("global_idx")};
+
+              var offsets: array<u32, ${v}>;
+
+              var value = ${w}(${u});
+              var pad = 0;
+              var isPad = false;
+
+              for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {
+                var offset = i;
+                for (var j = 0u; j < ${v-1}u; j++) {
+                  offsets[j] = offset / ${K("uniforms.kernelStrides","j",v)};
+                  offset -= offsets[j] * ${K("uniforms.kernelStrides","j",v)};
+                }
+                offsets[${v-1}] = offset;
+
+                isPad = false;
+                for (var j = ${r-v}u; j < ${r}u; j++) {
+                  xIndices[j] = indices[j] * ${K("uniforms.strides",`j - ${r-v}u`,v)}
+                    + offsets[j - ${r-v}u] - ${K("uniforms.pads","j - 2u",T)};
+                  ${$}
+              }
+              ${s}
+
+              output[global_idx] = value;
+            }`}},In=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,_d=e=>`${In(e)};${e.countIncludePad}`,bd=e=>`${In(e)};${e.storageOrder};${e.dilations}`,En=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}),Cn=(e,t,r,i)=>{let[n,a]=Tn(t,i,r),s=B("x",t.dataType,t.dims.length),u=s.type.value,l="value += x_val;",d="";n.countIncludePad?d+=`value /= ${u}(uniforms.kernelSize);`:d+=`value /= ${u}(i32(uniforms.kernelSize) - pad);`;let[f,h,g,w,y]=Sn(a,n);f.push(...Z(t.dims,a));let v=["rank"];return{name:e,shaderCache:{hint:`${i.cacheKey};${g};${w};${y}`,inputDependencies:v},getRunData:()=>({outputs:[{dims:a,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(R.size(a)/64)},programUniforms:f}),getShaderSource:T=>kn(T,s,t.dims.length,a.length,n,l,d,0,h,g,w,y)}},of=e=>{let t=e.count_include_pad!==0,r=En(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let i={countIncludePad:t,...r,cacheKey:""};return{...i,cacheKey:_d(i)}},uf=(e,t)=>{hr(e.inputs),e.compute(Cn("AveragePool",e.inputs[0],!1,t))},zn={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},lf=e=>{let t=e.format;return{format:t,...zn,cacheKey:t}},df=(e,t)=>{hr(e.inputs),e.compute(Cn("GlobalAveragePool",e.inputs[0],!0,t))},An=(e,t,r,i)=>{let[n,a]=Tn(t,i,r),s=`
+      value = max(x_val, value);
+    `,u="",l=B("x",t.dataType,t.dims.length),d=["rank"],[f,h,g,w,y]=Sn(a,n);return f.push(...Z(t.dims,a)),{name:e,shaderCache:{hint:`${i.cacheKey};${g};${w};${y}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:a,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(R.size(a)/64)},programUniforms:f}),getShaderSource:v=>kn(v,l,t.dims.length,a.length,n,s,u,t.dataType===10?-65504:-1e5,h,g,w,y)}},pf=(e,t)=>{hr(e.inputs),e.compute(An("MaxPool",e.inputs[0],!1,t))},cf=e=>{let t=e.storage_order,r=e.dilations,i=En(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(i.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let n={storageOrder:t,dilations:r,...i,cacheKey:""};return{...n,cacheKey:bd(n)}},hf=e=>{let t=e.format;return{format:t,...zn,cacheKey:t}},ff=(e,t)=>{hr(e.inputs),e.compute(An("GlobalMaxPool",e.inputs[0],!0,t))}}),z0=P(()=>{"use strict";ee(),ie(),Te(),ne(),vd=(e,t)=>{if(e.length<2||e.length>3)throw new Error("DequantizeLinear requires 2 or 3 inputs.");if(e.length===3&&e[1].dims===e[2].dims)throw new Error("x-scale and x-zero-point must have the same shape.");if(e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("x and x-zero-point must have the same data type.");if(e[0].dataType===6&&e.length>2)throw new Error("In the case of dequantizing int32 there is no zero point.");if(e[1].dims.length!==0&&e[1].dims.length!==1&&e[1].dims.length!==e[0].dims.length)throw new Error("scale input must be a scalar, a 1D tensor, or have the same rank as the input tensor.");if(e.length>2){if(e[0].dataType!==e[2].dataType)throw new Error("x and x-zero-point must have the same data type.");if(e[1].dims.length!==e[2].dims.length)throw new Error("scale and zero-point inputs must have the same rank.");if(!e[1].dims.map((r,i)=>r===e[2].dims[i]).reduce((r,i)=>r&&i,!0))throw new Error("scale and zero-point inputs must have the same shape.")}if(t.blockSize>0){if(e[1].dims.length===0||e[1].dims.length===1&&e[1].dims[0]===1)throw new Error("blockSize must be set only for block quantization.");if(!e[1].dims.map((n,a)=>a===t.axis||n===e[0].dims[a]).reduce((n,a)=>n&&a,!0))throw new Error("For block qunatization, scale input shape to match the input shape except for the axis");if(e[1].dims.length!==e[0].dims.length)throw new Error("For block qunatization the scale input rank must be the same as the x rank.");let r=e[0].dims[t.axis],i=e[1].dims[t.axis];if(t.blockSize<Math.ceil(r/i)||t.blockSize>Math.ceil(r/(i-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},$d=(e,t)=>{let r=R.normalizeAxis(t.axis,e[0].dims.length),i=e[0].dataType,n=i===3,a=e[0].dims,s=e[1].dataType,u=R.size(a),l=i===3||i===2,d=l?[Math.ceil(R.size(e[0].dims)/4)]:e[0].dims,f=e[1].dims,h=e.length>2?e[2]:void 0,g=h?l?[Math.ceil(R.size(h.dims)/4)]:h.dims:void 0,w=f.length===0||f.length===1&&f[0]===1,y=w===!1&&f.length===1,v=xe(u),T=w&&(!l||v===4),$=T?v:1,_=T&&!l?v:1,I=B("input",l?12:i,d.length,_),k=B("scale",s,f.length),E=h?B("zero_point",l?12:i,g.length):void 0,C=F("output",s,a.length,$),O=[I,k];E&&O.push(E);let x=[d,f];h&&x.push(g);let D=[{type:12,data:u/$},{type:12,data:r},{type:12,data:t.blockSize},...Z(...x,a)],L=Y=>{let q=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
+      ${Y.registerUniforms(q).declareVariables(...O,C)}
+      ${Y.mainStart()}
+          ${Y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+          let output_indices = ${C.offsetToIndices("global_idx")};
+
+          // Set input x
+          ${l?`
+            let input = ${I.getByOffset("global_idx / 4")};
+            let x_vec = ${n?"unpack4xI8(input)":"unpack4xU8(input)"};
+            let x_value = ${$===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${I.getByOffset("global_idx")};`};
+
+          // Set scale input
+          ${w?`let scale_value= ${k.getByOffset("0")}`:y?`
+            let scale_index = ${C.indicesGet("output_indices","uniforms.axis")};
+            let scale_value= ${k.getByOffset("scale_index")};`:`
+            var scale_indices: ${k.type.indices} = output_indices;
+            let index = ${k.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
+            ${k.indicesSet("scale_indices","uniforms.axis","index")};
+            let scale_value= ${k.getByIndices("scale_indices")};`};
+
+          // Set zero-point input
+          ${E?w?l?`
+                let zero_point_input = ${E.getByOffset("0")};
+                let zero_point_vec =  ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
+                let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${E.getByOffset("0")}`:y?l?`
+                let zero_point_index = ${C.indicesGet("output_indices","uniforms.axis")};
+                let zero_point_input = ${E.getByOffset("zero_point_index / 4")};
+                let zero_point_vec =  ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
+                let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
+                let zero_point_index = ${C.indicesGet("output_indices","uniforms.axis")};
+                let zero_point_value = ${E.getByOffset("zero_point_index")};`:l?`
+                let zero_point_offset = ${k.indicesToOffset("scale_indices")};
+                let zero_point_input = ${E.getByOffset("zero_point_offset / 4")};
+                let zero_point_vec = ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
+                let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${E.getByIndices("scale_indices")};`:`let zero_point_value = ${l?n?"i32":"u32":I.type.value}(0);`};
+      // Compute and write output
+      ${C.setByOffset("global_idx",`${C.type.value}(x_value - zero_point_value) * scale_value`)};
+      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:E?["rank","rank","rank"]:["rank","rank"]},getShaderSource:L,getRunData:()=>({outputs:[{dims:a,dataType:s}],dispatchGroup:{x:Math.ceil(u/$/64),y:1,z:1},programUniforms:D})}},mf=(e,t)=>{vd(e.inputs,t),e.compute($d(e.inputs,t))},gf=e=>me({axis:e.axis,blockSize:e.blockSize})}),A0=P(()=>{"use strict";Le(),ee(),ne(),xd=(e,t,r)=>{let i=e===t,n=e<t&&r<0,a=e>t&&r>0;if(i||n||a)throw new Error("Range these inputs' contents are invalid.")},Td=(e,t,r,i)=>{let n=Math.abs(Math.ceil((t-e)/r)),a=[n],s=n,u=[{type:12,data:s},{type:i,data:e},{type:i,data:r},...Z(a)],l=d=>{let f=F("output",i,a.length),h=f.type.value,g=[{name:"outputSize",type:"u32"},{name:"start",type:h},{name:"delta",type:h}];return`
+        ${d.registerUniforms(g).declareVariables(f)}
+        ${d.mainStart()}
+        ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+        output[global_idx] = uniforms.start + ${h}(global_idx) * uniforms.delta;
+      }`};return{name:"Range",shaderCache:{hint:`${i}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:a,dataType:i}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:u})}},yf=e=>{let t=0,r=0,i=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],i=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],i=e.inputs[2].getFloat32Array()[0]),be.webgpu.validateInputContent&&xd(t,r,i),e.compute(Td(t,r,i,e.inputs[0].dataType),{inputs:[]})}}),O0=P(()=>{"use strict";ee(),ie(),Te(),ne(),Sd=(e,t,r,i)=>{if(e!=="none"&&i!=="i32"&&i!=="u32"&&i!=="f32")throw new Error(`Input ${i} is not supported with reduction ${e}.`);let n=`{
+                var oldValue = 0;
+                loop {
+                  let newValueF32 =`,a=`;
+                  let newValue = bitcast<i32>(newValueF32);
+                  let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue);
+                  if res.exchanged {
+                    break;
+                  }
+                  oldValue = res.old_value;
+                }
+              }`;switch(e){case"none":return`${t}=${r};`;case"add":return i==="i32"||i==="u32"?`atomicAdd(&${t}, bitcast<${i}>(${r}));`:`
+              ${n}bitcast<${i}>(oldValue) + (${r})${a}`;case"max":return i==="i32"||i==="u32"?`atomicMax(&${t}, bitcast<${i}>(${r}));`:`
+                ${n}max(bitcast<f32>(oldValue), (${r}))${a}`;case"min":return i==="i32"||i==="u32"?`atomicMin(&${t}, bitcast<${i}>(${r}));`:`${n}min(bitcast<${i}>(oldValue), (${r}))${a}`;case"mul":return`${n}(bitcast<${i}>(oldValue) * (${r}))${a}`;default:throw new Error(`Reduction ${e} is not supported.`)}},kd=(e,t)=>{let r=e[0].dims,i=e[1].dims,n=r,a=1,s=Math.ceil(R.sizeToDimension(i,i.length-1)/a),u=i[i.length-1],l=R.sizeFromDimension(r,u),d=[{type:12,data:s},{type:12,data:u},{type:12,data:l},...Z(e[1].dims,e[2].dims,n)],f=h=>{let g=B("indices",e[1].dataType,e[1].dims.length),w=B("updates",e[2].dataType,e[2].dims.length,a),y=t.reduction!=="none"&&t.reduction!==""?Hp("output",e[0].dataType,n.length):F("output",e[0].dataType,n.length,a);return`
+      ${h.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(g,w,y)}
+      ${h.mainStart()}
+        ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+  var data_offset = 0u;
+  let indices_start = uniforms.last_index_dimension * global_idx;
+  let indices_end = indices_start + uniforms.last_index_dimension;
+  for (var i = indices_start; i < indices_end; i++) {
+    var index = i32(indices[i].x);
+    ${e[0].dims.length===1?`
+    let element_count_dim = uniforms.output_strides;
+    let dim_value = uniforms.output_shape;`:`
+    let element_count_dim = uniforms.output_strides[i - indices_start];
+    let dim_value = uniforms.output_shape[i - indices_start];`}
+    if (index >= 0) {
+      if (index >= i32(dim_value)) {
+        index = i32(dim_value - 1);
+      }
+    } else {
+      if (index < -i32(dim_value)) {
+        index = 0;
+      } else {
+        index += i32(dim_value);
+      }
+    }
+    data_offset += u32((u32(index) * element_count_dim));
+  }
+
+  for (var i = 0u; i < uniforms.num_updates_elements; i++) {
+    let value = updates[uniforms.num_updates_elements * global_idx + i];
+    ${Sd(t.reduction,"output[data_offset + i]","value",y.type.value)}
+  }
+
+      }`};return{name:"ScatterND",shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:d}),getShaderSource:f}},wf=e=>me({reduction:e.reduction}),_f=(e,t)=>{e.compute(kd(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}}),R0=P(()=>{"use strict";ee(),ie(),Te(),ne(),Id=(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")}},Ed=(e,t,r)=>{t.every(n=>n>=0&&n<r||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let i=new Array(r).fill(1);return t.forEach((n,a)=>i[n]=e[a]),i},Cd=(e,t,r,i,n,a)=>{let[s,u,l]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],d=e[0].dims.length;if(s>0&&e.length>s&&e[s].dims.length>0)e[s].getFloat32Array().forEach(f=>a.push(f));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(u>0&&e.length>u&&e[u].dims.length===1&&e[u].dims[0]>0){if(e[u].getFloat32Array().forEach(f=>i.push(f)),i.length!==0&&i.length!==d&&r>=18&&i.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");Id(i,t),t.axes.length>0&&Ed(i,t.axes,d).forEach((f,h)=>i[h]=f)}if(l>0&&e.length>l&&e[l].dims.length===1&&e[l].dims[0]>0&&(e[l].getBigInt64Array().forEach(f=>n.push(Number(f))),n.length!==0&&n.length!==d&&r>=18&&n.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(i.length!==0&&i.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(n.length!==0&&n.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 i<"u"&&typeof n<"u"&&i.length>0&&n.length>d)throw new Error("Resize requires only of scales or sizes to be specified")},On=(e,t,r,i)=>`
+  // The whole part and the fractional part are calculated separately due to inaccuracy of floating
+  // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
+  // offset-by-one error later in floor().
+  let big = (${e}) * (${t});
+  let whole = ${i}(big / (${r}));
+  let fract = ${i}(big % (${r})) / ${i}(${r});
+  return whole + fract;
+`,zd=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
+     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`
+          if (xScale < 1.0 || floor(xScale) != xScale) {
+            return ${t}(xResized) / ${t}(xScale);
+          } else {
+            ${On("xResized","lengthOriginal","lengthResized",t)}
+          }
+        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
+                    return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
+                  } else {
+                    return 0.0;
+                  }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
+                    return 0.0;
+                  } else {
+                    ${On("xResized","lengthOriginal - 1","lengthResized - 1",t)}
+                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
+                    return ${t}(roiStart) * ${t}(lengthOriginal - 1) +
+                        (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /
+                        ${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`)}})()+"}",Ad=(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`)}})()+"}",Od=(e,t,r)=>{let i=new Array(r).fill(0).concat(new Array(r).fill(1)),n=e.length===0?i:e.slice();return t.length>0?(t.forEach((a,s)=>{i[a]=n[s],i[s+r]=n[t.length+s]}),i):n},Rd=(e,t,r,i)=>{let n=[];if(r.length>0)if(i.length>0){if(e.forEach(a=>n.push(a)),Math.max(...i)>e.length)throw new Error("axes is out of bound");i.forEach((a,s)=>n[a]=r[s])}else r.forEach(a=>n.push(a));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");n=e.map((a,s)=>Math.round(a*t[s]))}return n},Md=(e,t,r)=>{let i=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(a=>t[a]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(a=>t[a]),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 n=e.slice();return r.axes.length>0?(r.axes.forEach(a=>t[a]=i),r.axes.forEach(a=>n[a]=Math.round(e[a]*t[a]))):(t.fill(i,0,t.length),n.forEach((a,s)=>n[s]=Math.round(a*t[s]))),n},Bd=(e,t,r,i,n)=>`
+    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 = ${K("uniforms.scales","i",i)};
+        var roi_low = ${K("uniforms.roi","i",n)};
+        var roi_hi = ${K("uniforms.roi",`i + ${t.length}`,n)};
+        if (scale == 1.0) {
+          original_indices[i] = ${e.type.value}(output_index);
+        } else {
+          var input_shape_i = ${K("uniforms.input_shape","i",t.length)};
+          var output_shape_i = ${K("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;
+    }`,Nd=(e,t,r,i,n,a,s)=>`
+    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
+      var input_indices: ${e.type.indices};
+      for (var i:u32 = 0; i < ${i.length}; i++) {
+        var output_index = ${t.indicesGet("output_indices","i")};
+        var input_index: u32;
+        var scale = ${K("uniforms.scales","i",n)};
+        if (scale == 1.0) {
+          input_index = output_index;
+        } else {
+          var roi_low = ${K("uniforms.roi","i",a)};
+          var roi_hi = ${K("uniforms.roi",`i + ${r.length}`,a)};
+          var input_shape_i = ${K("uniforms.input_shape","i",r.length)};
+          var output_shape_i = ${K("uniforms.output_shape","i",i.length)};
+          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
+                                                                        input_shape_i, roi_low, roi_hi);
+          if (!${s} || (original_idx >= 0 && original_idx < ${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;
+    }`,Dd=(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 >= ${K("uniforms.input_shape","i",t.length)}) {
+          return false;
+        }
+      }
+      return true;
+    }`,Rn=(e,t,r,i)=>e.rank>i?`
+    ${e.indicesSet("input_indices",t,"channel")};
+    ${e.indicesSet("input_indices",r,"batch")};
+`:"",Pd=(e,t,r,i,n)=>{let[a,s,u,l]=r.length===2?[-1,0,1,-1]:[0,2,3,1],d=e.type.value;return`
+    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${d} {
+      var input_indices: ${e.type.indices};
+      ${e.indicesSet("input_indices",s,`max(0, min(row, ${r[s]} - 1))`)};
+      ${e.indicesSet("input_indices",u,`max(0, min(col, ${r[u]} - 1))`)};
+      ${Rn(e,l,a,2)}
+      return ${e.getByIndices("input_indices")};
+    }
+
+    fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${d} {
+      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
+      var row:${d} = originalIndices[${s}];
+      var col:${d} = originalIndices[${u}];
+      ${i?`if (row < 0 || row > (${r[s]} - 1) || col < 0 || col > (${r[u]} - 1)) {
+        return ${n};
+      }`:""};
+      row = max(0, min(row, ${r[s]} - 1));
+      col = max(0, min(col, ${r[u]} - 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 =  ${r.length>2?`u32(originalIndices[${a}])`:"0"};
+      var x11: ${d} = getInputValue(batch, channel, row1, col1);
+      var x12: ${d} = getInputValue(batch, channel, row1, col2);
+      var x21: ${d} = getInputValue(batch, channel, row2, col1);
+      var x22: ${d} = getInputValue(batch, channel, row2, col2);
+      var dx1: ${d} = abs(row - ${d}(row1));
+      var dx2: ${d} = abs(${d}(row2) - row);
+      var dy1: ${d} = abs(col - ${d}(col1));
+      var dy2: ${d} = abs(${d}(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);
+    }`},Ud=(e,t,r,i,n,a,s,u,l,d)=>{let f=r.length===2,h=!0,[g,w]=f?[0,1]:h?[2,3]:[1,2],y=e.type.value,v=T=>{let $=T===g?"row":"col";return`
+      fn ${$}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${y} {
+        var output_index = ${t.indicesGet("output_indices",T)};
+        var originalIdx: ${y} = getOriginalCoordinateFromResizedCoordinate(output_index, ${n[T]},
+        ${i[T]}, ${r[T]}, ${a[T]}, ${a[T]} + ${r.length});
+        var fractOriginalIdx: ${y} = originalIdx - floor(originalIdx);
+        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);
+
+        if (${u} && (originalIdx < 0 || originalIdx > (${r[T]} - 1))) {
+          return ${l};
+        }
+        var data: array<${y}, 4> = array<${y}, 4>(0.0, 0.0, 0.0, 0.0);
+        for (var i: i32 = -1; i < 3; i++) {
+          var ${$}: ${y} = originalIdx + ${y}(i);
+          if (${$} < 0 || ${$} >= ${r[T]}) {
+            ${d?`coefs[i + 1] = 0.0;
+                        continue;`:u?`return ${l};`:`${$} = max(0, min(${$}, ${r[T]} - 1));`};
+          }
+        var input_indices_copy: ${e.type.indices} = input_indices;
+          ${e.indicesSet("input_indices_copy",T,`u32(${$})`)};
+          data[i + 1] = ${T===g?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
+        }
+        return cubicInterpolation1D(data, coefs);
+      }`};return`
+    ${v(g)};
+    ${v(w)};
+  fn getCubicInterpolationCoefs(s: ${y}) -> array<${y}, 4> {
+    var absS = abs(s);
+    var coeffs: array<${y}, 4> = array<${y}, 4>(0.0, 0.0, 0.0, 0.0);
+    var oneMinusAbsS: ${y} = 1.0 - absS;
+    var twoMinusAbsS: ${y} = 2.0 - absS;
+    var onePlusAbsS: ${y} = 1.0 + absS;
+    coeffs[0] = ((${s} * onePlusAbsS - 5 * ${s}) * onePlusAbsS + 8 * ${s}) * onePlusAbsS - 4 * ${s};
+    coeffs[1] = ((${s} + 2) * absS - (${s} + 3)) * absS * absS + 1;
+    coeffs[2] = ((${s} + 2) * oneMinusAbsS - (${s} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
+    coeffs[3] = ((${s} * twoMinusAbsS - 5 * ${s}) * twoMinusAbsS + 8 * ${s}) * twoMinusAbsS - 4 * ${s};
+    return coeffs;
+  }
+
+  fn cubicInterpolation1D(x: array<${y}, 4>, coefs: array<${y}, 4>) -> ${y} {
+    var coefsSum: ${y} = 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}) -> ${y} {
+    var input_indices: ${e.type.indices} = output_indices;
+    return colCubicInterpolation(input_indices, output_indices);
+  }
+    `},Wd=(e,t,r,i,n)=>{let[a,s,u,l,d]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],f=e.type.value;return`
+    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${f} {
+      var input_indices: ${e.type.indices};
+      ${e.indicesSet("input_indices",s,`max(0, min(depth, ${r[s]} - 1))`)};
+      ${e.indicesSet("input_indices",u,`max(0, min(height, ${r[u]} - 1))`)};
+      ${e.indicesSet("input_indices",l,`max(0, min(width, ${r[l]} - 1))`)};
+      ${Rn(e,d,a,3)}
+      return ${e.getByIndices("input_indices")};
+    }
+
+    fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${f} {
+      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
+      var depth:${f} = originalIndices[${s}];
+      var height:${f} = originalIndices[${u}];
+      var width:${f} = originalIndices[${l}];
+      ${i?`if (depth < 0 || depth > (${r[s]} - 1) || height < 0 || height > (${r[u]} - 1) || width < 0 || (width > ${r[l]} - 1)) {
+      return ${n};
+        }`:""};
+
+    depth = max(0, min(depth, ${r[s]} - 1));
+      height = max(0, min(height, ${r[u]} - 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 = ${r.length>3?`u32(originalIndices[${d}])`:"0"};
+      var batch: u32 =  ${r.length>3?`u32(originalIndices[${a}])`:"0"};
+
+      var x111: ${f} = getInputValue(batch, channel, depth1, height1, width1);
+      var x112: ${f} = getInputValue(batch, channel, depth1, height1, width2);
+      var x121: ${f} = getInputValue(batch, channel, depth1, height2, width1);
+      var x122: ${f} = getInputValue(batch, channel, depth1, height2, width2);
+      var x211: ${f} = getInputValue(batch, channel, depth2, height1, width1);
+      var x212: ${f} = getInputValue(batch, channel, depth2, height1, width2);
+      var x221: ${f} = getInputValue(batch, channel, depth2, height2, width1);
+      var x222: ${f} = getInputValue(batch, channel, depth2, height2, width2);
+      var dx1: ${f} = abs(depth - ${f}(depth1));
+      var dx2: ${f} = abs(${f}(depth2) - depth);
+      var dy1: ${f} = abs(height - ${f}(height1));
+      var dy2: ${f} = abs(${f}(height2) - height);
+      var dz1: ${f} = abs(width - ${f}(width1));
+      var dz2: ${f} = abs(${f}(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);
+    }`},Ld=(e,t,r,i,n,a)=>{let s=e.dims,u=Od(a,t.axes,s.length),l=Rd(s,i,n,t.axes),d=i.slice();i.length===0&&(d=s.map((_,I)=>_===0?1:l[I]/_),t.keepAspectRatioPolicy!=="stretch"&&(l=Md(s,d,t)));let f=F("output",e.dataType,l.length),h=B("input",e.dataType,s.length),g=R.size(l),w=s.length===l.length&&s.every((_,I)=>_===l[I]),y=t.coordinateTransformMode==="tf_crop_and_resize",v=t.extrapolationValue,T=h.type.value,$=_=>`
+      ${w?"":`
+      ${zd(t.coordinateTransformMode,T)};
+      ${(()=>{switch(t.mode){case"nearest":return`
+              ${Dd(h,s)};
+              ${Ad(t.nearestMode,r,T)};
+              ${Nd(h,f,s,l,d.length,u.length,y)};
+              `;case"linear":return`
+              ${Bd(f,s,l,d.length,u.length)};
+              ${(()=>{if(s.length===2||s.length===4)return`${Pd(h,f,s,y,v)}`;if(s.length===3||s.length===5)return`${Wd(h,f,s,y,v)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
+            `;case"cubic":return`
+            ${(()=>{if(s.length===2||s.length===4)return`${Ud(h,f,s,l,d,u,t.cubicCoeffA,y,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",d.length).registerUniform("roi","f32",u.length).declareVariables(h,f)}
+      ${_.mainStart()}
+        ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+        ${w?"output[global_idx] = input[global_idx];":`
+        let output_indices = ${f.offsetToIndices("global_idx")};
+        var input_indices: ${h.type.indices};
+        ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);
+                if (checkInputIndices(input_indices)) {
+                  output[global_idx] = ${h.getByIndices("input_indices")};
+                } else {
+                  output[global_idx] = ${t.extrapolationValue};
+                }`;case"linear":return`output[global_idx] = ${s.length===2||s.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};
+`}
+      }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${d.length>0?t.mode==="cubic"?d:d.length:""}|${n.length>0?n:""}|${u.length>0?u:""}|${w}|${t.mode==="nearest"?s.length:s}`,inputDependencies:["rank"]},getShaderSource:$,getRunData:()=>({outputs:[{dims:l,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:[{type:12,data:g},{type:1,data:d},{type:1,data:u},...Z(s,l)]})}},qd=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},bf=(e,t)=>{let r=[],i=[],n=[],a=qd(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");Cd(e.inputs,t,a,r,i,n),e.compute(Ld(e.inputs[0],t,a,r,i,n),{inputs:[0]})},vf=e=>{let t=e.antialias,r=e.axes,i=e.coordinateTransformMode,n=e.cubicCoeffA,a=e.excludeOutside!==0,s=e.extrapolationValue,u=e.keepAspectRatioPolicy,l=e.mode,d=e.nearestMode===""?"simple":e.nearestMode;return me({antialias:t,axes:r,coordinateTransformMode:i,cubicCoeffA:n,excludeOutside:a,extrapolationValue:s,keepAspectRatioPolicy:u,mode:l,nearestMode:d})}}),M0=P(()=>{"use strict";ee(),ie(),ne(),Vd=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],i=e[2];if(t.dataType!==r.dataType||t.dataType!==i.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 n=t.dims[t.dims.length-1],a=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==n)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==a)throw new Error("Skip must have the same sequence length as input");if(i.dims.length!==1)throw new Error("Gamma must be 1D");if(i.dims[i.dims.length-1]!==n)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let s=e[3];if(s.dims.length!==1)throw new Error("Beta must be 1D");if(s.dims[s.dims.length-1]!==n)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let s=e[4];if(s.dims.length!==1)throw new Error("Bias must be 1D");if(s.dims[s.dims.length-1]!==n)throw new Error("Bias must have the same hidden size as input")}},Gd=(e,t,r,i)=>{let n=t.simplified,a=e[0].dims,s=R.size(a),u=a,l=s,d=a.slice(-1)[0],f=i?a.slice(0,-1).concat(1):[],h=!n&&e.length>3,g=e.length>4,w=i&&r>1,y=i&&r>2,v=r>3,T=64,$=xe(d),_=[{type:12,data:l},{type:12,data:$},{type:12,data:d},{type:1,data:t.epsilon}],I=E=>{let C=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],O=[B("x",e[0].dataType,e[0].dims,$),B("skip",e[1].dataType,e[1].dims,$),B("gamma",e[2].dataType,e[2].dims,$)];h&&O.push(B("beta",e[3].dataType,e[3].dims,$)),g&&O.push(B("bias",e[4].dataType,e[4].dims,$)),O.push(F("output",e[0].dataType,u,$)),w&&O.push(F("mean_output",1,f)),y&&O.push(F("inv_std_output",1,f)),v&&O.push(F("input_skip_bias_sum",e[0].dataType,u,$));let x=Ee(e[0].dataType),D=Ee(1,$);return`
+
+      ${E.registerUniforms(C).declareVariables(...O)}
+      var<workgroup> sum_shared : array<${D}, ${T}>;
+      var<workgroup> sum_squared_shared : array<${D}, ${T}>;
+
+      ${E.mainStart([T,1,1])}
+        let ix = local_id.x;
+        let iy = global_id.x / ${T};
+
+        let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components;
+        var stride = hidden_size_vectorized / ${T};
+        let offset = ix * stride + iy * hidden_size_vectorized;
+        let offset1d = stride * ix;
+        if (ix == ${T-1}) {
+          stride = hidden_size_vectorized - stride * ix;
+        }
+        for (var i: u32 = 0; i < stride; i++) {
+          let skip_value = skip[offset + i];
+          let bias_value = ${g?"bias[offset1d + i]":x+"(0.0)"};
+          let input_value = x[offset + i];
+          let value = input_value + skip_value + bias_value;
+          ${v?"input_skip_bias_sum[offset + i] = value;":""}
+          output[offset + i] = value;
+          let f32_value = ${jt(x,$,"value")};
+          sum_shared[ix] += f32_value;
+          sum_squared_shared[ix] += f32_value * f32_value;
+        }
+        workgroupBarrier();
+
+        var reduce_size : u32 = ${T};
+        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 = ${vt("sum",$)} / f32(uniforms.hidden_size);
+        let inv_std_dev = inverseSqrt(${vt("square_sum",$)} / f32(uniforms.hidden_size) ${n?"":"- mean * mean"} + uniforms.epsilon);
+        ${w?"mean_output[global_idx] = mean;":""}
+        ${y?"inv_std_output[global_idx] = inv_std_dev;":""}
+
+        for (var i: u32 = 0; i < stride; i++) {
+          output[offset + i] = (output[offset + i] ${n?"":`- ${x}(mean)`}) *
+            ${x}(inv_std_dev) * gamma[offset1d + i]
+            ${h?"+ beta[offset1d + i]":""};
+        }
+      }`},k=[{dims:u,dataType:e[0].dataType}];return r>1&&k.push({dims:f,dataType:1}),r>2&&k.push({dims:f,dataType:1}),r>3&&k.push({dims:a,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${$};${w};${y};${v}`,inputDependencies:e.map((E,C)=>"type")},getShaderSource:I,getRunData:()=>({outputs:k,dispatchGroup:{x:Math.ceil(l/d)},programUniforms:_})}},$f=(e,t)=>{Vd(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(Gd(e.inputs,t,e.outputCount,!1),{outputs:r})}}),B0=P(()=>{"use strict";ee(),ie(),Te(),ne(),Fd=(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,i)=>{if(e[i+1].dataType!==6&&e[i+1].dataType!==7)throw new Error(`Input ${i} must be an array of int32 or int64`)})},fr=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(i=>r.push(Number(i)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(i=>r.push(Number(i)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Hd=(e,t)=>{if(e.length>1){let r=fr(e,1),i=fr(e,2),n=fr(e,3);return n.length===0&&(n=[...Array(e[0].dims.length).keys()]),me({starts:r,ends:i,axes:n})}else return t},Mn=(e,t,r,i,n)=>{let a=e;return e<0&&(a+=r[i[t]]),n[t]<0?Math.max(0,Math.min(a,r[i[t]]-1)):Math.max(0,Math.min(a,r[i[t]]))},Kd=(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-1}; i >= 0; i--) {
+            let input_shape_i = ${K("uniforms.input_shape","i",r.length)};
+            let steps_i = ${K("uniforms.steps","i",r.length)};
+            let signs_i = ${K("uniforms.signs","i",r.length)};
+            let starts_i = ${K("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;
+            }
+            ${e.indicesSet("input_indices","i","input_index")};
+          }
+          return input_indices;
+      }`,jd=(e,t)=>{let r=e[0].dims,i=R.size(r),n=t.axes.length>0?R.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],a=fr(e,4);a.forEach($=>$!==0||(()=>{throw new Error("step cannot be 0")})),a.length===0&&(a=Array(n.length).fill(1));let s=t.starts.map(($,_)=>Mn($,_,r,n,a)),u=t.ends.map(($,_)=>Mn($,_,r,n,a));if(n.length!==s.length||n.length!==u.length)throw new Error("start, ends and axes should have the same number of elements");if(n.length!==r.length)for(let $=0;$<r.length;++$)n.includes($)||(s.splice($,0,0),u.splice($,0,r[$]),a.splice($,0,1));let l=a.map($=>Math.sign($));a.forEach(($,_,I)=>{if($<0){let k=(u[_]-s[_])/$,E=s[_],C=E+k*a[_];s[_]=C,u[_]=E,I[_]=-$}});let d=r.slice(0);n.forEach(($,_)=>{d[$]=Math.ceil((u[$]-s[$])/a[$])});let f={dims:d,dataType:e[0].dataType},h=F("output",e[0].dataType,d.length),g=B("input",e[0].dataType,e[0].dims.length),w=R.size(d),y=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:s.length},{name:"signs",type:"i32",length:l.length},{name:"steps",type:"u32",length:a.length}],v=[{type:12,data:w},{type:12,data:s},{type:6,data:l},{type:12,data:a},...Z(e[0].dims,d)],T=$=>`
+      ${$.registerUniforms(y).declareVariables(g,h)}
+        ${Kd(g,h,r)}
+        ${$.mainStart()}
+          ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
+          let output_indices = ${h.offsetToIndices("global_idx")};
+          let input_indices = calculateInputIndices(output_indices);
+          ${h.setByOffset("global_idx",g.getByIndices("input_indices"))}
+      }`;return{name:"Slice",shaderCache:{hint:`${l.length}_${s.length}_${a.length}`,inputDependencies:["rank"]},getShaderSource:T,getRunData:()=>({outputs:[f],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:v})}},xf=(e,t)=>{Fd(e.inputs,t);let r=Hd(e.inputs,t);e.compute(jd(e.inputs,r),{inputs:[0]})},Tf=e=>{let t=e.starts,r=e.ends,i=e.axes;return me({starts:t,ends:r,axes:i})}}),N0=P(()=>{"use strict";ee(),ie(),Te(),$t(),ne(),Qd=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},Zd=(e,t)=>{let r=e.inputs[0],i=r.dims,n=R.size(i),a=i.length,s=R.normalizeAxis(t.axis,a),u=s<i.length-1,l,d=[];u?(d=Array.from({length:a},(O,x)=>x),d[s]=a-1,d[a-1]=s,l=e.compute(De(r,d),{inputs:[r],outputs:[-1]})[0]):l=r;let f=l.dims,h=f[a-1],g=n/h,w=xe(h),y=h/w,v=64;g===1&&(v=256);let T=(O,x)=>x===4?`max(max(${O}.x, ${O}.y), max(${O}.z, ${O}.w))`:x===2?`max(${O}.x, ${O}.y)`:x===3?`max(max(${O}.x, ${O}.y), ${O}.z)`:O,$=B("x",l.dataType,l.dims,w),_=F("result",l.dataType,l.dims,w),I=$.type.value,k=Ee(l.dataType)==="f32"?`var threadMax = ${I}(-3.4028234663852886e+38f);`:`var threadMax = ${I}(-65504.0h);`,E=O=>`
+      var<workgroup> rowMaxShared : ${I};
+      var<workgroup> rowSumShared : ${I};
+      var<workgroup> threadShared : array<${I}, ${v}>;
+
+      fn getValue(row: i32, col: i32, row_stride: i32) -> ${I} {
+        let index = row * row_stride + col;
+        return x[index];
+      }
+
+      fn setValue(row: i32, col: i32, row_stride: i32, value: ${I}) {
+        let index = row * row_stride + col;
+        result[index] = value;
+      }
+      ${O.registerUniform("packedCols","i32").declareVariables($,_)}
+      ${O.mainStart(v)}
+        let gindex = i32(global_idx);
+        let lindex = i32(local_idx);
+        const wg = ${v};
+        let row = gindex / wg;
+        let cols = uniforms.packedCols;
+        let row_stride : i32 = uniforms.packedCols;
+
+        // find the rows max
+        ${k}
+        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 = ${I}(${T("threadShared[0]",w)});
+        }
+        workgroupBarrier();
+
+        // find the rows sum
+        var threadSum = ${I}(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 = ${I}(${vt("threadShared[0]",w)});
+        }
+        workgroupBarrier();
+
+        // calculate final value for each element in the row
+        for (var col = lindex; col < cols; col += wg) {
+          var value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;
+          // max operation protects against NaN since all values should be >=0
+          value = max(value, ${I}(0.0));
+          setValue(row, col, row_stride, value);
+        }
+      }`,C=e.compute({name:"Softmax",shaderCache:{hint:`${w};${v}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:f,dataType:l.dataType}],dispatchGroup:{x:g},programUniforms:[{type:6,data:y}]}),getShaderSource:E},{inputs:[l],outputs:[u?-1:0]})[0];u&&e.compute(De(C,d),{inputs:[C]})},Sf=(e,t)=>{Qd(e.inputs),Zd(e,t)},kf=e=>me({axis:e.axis})}),D0=P(()=>{"use strict";ee(),ie(),ne(),Bn=e=>Array.from(e.getBigInt64Array(),Number),Yd=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==10&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, float16, int32, and uint32 data types");if(e[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(Bn(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")},Xd=(e,t)=>{let r=[];for(let i=0;i<e.length;++i)r.push(e[i]*t[i]);return r},Jd=(e,t)=>{let r=e[0].dims,i=t??Bn(e[1]),n=Xd(r,i),a=R.size(n),s=e[0].dataType,u=B("input",s,r.length),l=F("output",s,n.length),d=f=>`
+      const inputShape = ${u.indices(...r)};
+      ${f.registerUniform("output_size","u32").declareVariables(u,l)}
+      ${f.mainStart()}
+      ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
+      let output_indices = ${l.offsetToIndices("global_idx")};
+      var input_indices: ${u.type.indices};
+      for (var i = 0; i < ${r.length}; i++) {
+        let input_dim_i = ${u.indicesGet("uniforms.input_shape","i")};
+        let input_dim_value = ${l.indicesGet("output_indices","i")}  % input_dim_i;
+
+        ${u.indicesSet("input_indices","i","input_dim_value")}
+      }
+      ${l.setByOffset("global_idx",u.getByIndices("input_indices"))}
+    }`;return{name:"Tile",shaderCache:{hint:`${i}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},...Z(e[0].dims,n)]}),getShaderSource:d}},If=e=>{Yd(e.inputs),e.compute(Jd(e.inputs),{inputs:[0]})}}),P0=P(()=>{"use strict";ee(),ie(),ne(),ep=(e,t,r,i,n)=>{let a=F("output_data",n,r.length,4),s=B("a_data",t[1].dataType,t[1].dims.length,4),u=B("b_data",t[2].dataType,t[2].dims.length,4),l=B("c_data",t[0].dataType,t[0].dims.length,4),d,f=(h,g,w)=>`select(${g}, ${h}, ${w})`;if(!i)d=a.setByOffset("global_idx",f(s.getByOffset("global_idx"),u.getByOffset("global_idx"),l.getByOffset("global_idx")));else{let h=(g,w,y="")=>{let v=`a_data[index_a${w}][component_a${w}]`,T=`b_data[index_b${w}][component_b${w}]`,$=`bool(c_data[index_c${w}] & (0xffu << (component_c${w} * 8)))`;return`
+            let output_indices${w} = ${a.offsetToIndices(`global_idx * 4u + ${w}u`)};
+            let offset_a${w} = ${s.broadcastedIndicesToOffset(`output_indices${w}`,a)};
+            let offset_b${w} = ${u.broadcastedIndicesToOffset(`output_indices${w}`,a)};
+            let offset_c${w} = ${l.broadcastedIndicesToOffset(`output_indices${w}`,a)};
+            let index_a${w} = offset_a${w} / 4u;
+            let index_b${w} = offset_b${w} / 4u;
+            let index_c${w} = offset_c${w} / 4u;
+            let component_a${w} = offset_a${w} % 4u;
+            let component_b${w} = offset_b${w} % 4u;
+            let component_c${w} = offset_c${w} % 4u;
+            ${g}[${w}] = ${y}(${f(v,T,$)});
+          `};n===9?d=`
+            var data = vec4<u32>(0);
+            ${h("data",0,"u32")}
+            ${h("data",1,"u32")}
+            ${h("data",2,"u32")}
+            ${h("data",3,"u32")}
+            output_data[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:d=`
+            ${h("output_data[global_idx]",0)}
+            ${h("output_data[global_idx]",1)}
+            ${h("output_data[global_idx]",2)}
+            ${h("output_data[global_idx]",3)}
+          `}return`
+        ${e.registerUniform("vec_size","u32").declareVariables(l,s,u,a)}
+        ${e.mainStart()}
+        ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
+        ${d}
+      }`},tp=e=>{let t=e[1].dims,r=e[2].dims,i=e[0].dims,n=e[1].dataType,a=!(R.areEqual(t,r)&&R.areEqual(r,i)),s=t,u=R.size(t);if(a){let d=Qt.calcShape(Qt.calcShape(t,r,!1),i,!1);if(!d)throw new Error("Can't perform where op on the given tensors");s=d,u=R.size(s)}let l=Math.ceil(u/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:d=>ep(d,e,s,a,n),getRunData:()=>({outputs:[{dims:s,dataType:n}],dispatchGroup:{x:Math.ceil(u/64/4)},programUniforms:[{type:12,data:l},...Z(i,t,r,s)]})}},Ef=e=>{e.compute(tp(e.inputs))}}),U0=P(()=>{"use strict";Jy(),ya(),e0(),t0(),r0(),i0(),n0(),l0(),p0(),c0(),h0(),f0(),m0(),g0(),y0(),w0(),_0(),b0(),v0(),$0(),x0(),T0(),S0(),k0(),I0(),Kh(),E0(),C0(),z0(),A0(),O0(),ga(),R0(),Xh(),M0(),B0(),N0(),Zh(),D0(),$t(),wa(),P0(),Cf=new Map([["Abs",[vc]],["Acos",[$c]],["Acosh",[xc]],["Add",[ih]],["ArgMax",[yc,Hn]],["ArgMin",[gc,Hn]],["Asin",[Tc]],["Asinh",[Sc]],["Atan",[kc]],["Atanh",[Ic]],["Attention",[wc]],["AveragePool",[uf,of]],["BatchNormalization",[_c]],["BiasAdd",[bc]],["BiasSplitGelu",[rh]],["Cast",[Cc,Ec]],["Ceil",[Ac]],["Clip",[zc]],["Concat",[hh,fh]],["Conv",[Xn,Yn]],["ConvTranspose",[Th,xh]],["Cos",[Oc]],["Cosh",[Rc]],["CumSum",[Sh,kh]],["DepthToSpace",[Ih,Eh]],["DequantizeLinear",[mf,gf]],["Div",[nh]],["Einsum",[Ch,zh]],["Elu",[Mc,_r]],["Equal",[ah]],["Erf",[Bc]],["Exp",[Nc]],["Expand",[Ah]],["FastGelu",[Oh]],["Floor",[Dc]],["FusedConv",[Xn,Yn]],["Gather",[Mh,Rh]],["GatherElements",[Wh,Uh]],["GatherBlockQuantized",[Dh,Ph]],["GatherND",[Bh,Nh]],["Gelu",[Pc]],["Gemm",[qh,Lh]],["GlobalAveragePool",[df,lf]],["GlobalMaxPool",[ff,hf]],["Greater",[lh]],["GreaterOrEqual",[ph]],["GridSample",[Vh,Gh]],["GroupQueryAttention",[Jh]],["HardSigmoid",[Hc,Fc]],["InstanceNormalization",[ef]],["LayerNormalization",[tf]],["LeakyRelu",[Uc,_r]],["Less",[dh]],["LessOrEqual",[ch]],["Log",[eh]],["MatMul",[rf]],["MatMulNBits",[nf,af]],["MaxPool",[pf,cf]],["Mul",[sh]],["MultiHeadAttention",[Hh,Fh]],["Neg",[Lc]],["Not",[Wc]],["Pad",[sf]],["Pow",[oh]],["QuickGelu",[th,_r]],["Range",[yf]],["Reciprocal",[qc]],["ReduceMin",[pc]],["ReduceMean",[sc]],["ReduceMax",[dc]],["ReduceSum",[hc]],["ReduceProd",[cc]],["ReduceL1",[oc]],["ReduceL2",[uc]],["ReduceLogSum",[mc]],["ReduceLogSumExp",[lc]],["ReduceSumSquare",[fc]],["Relu",[Vc]],["Resize",[bf,vf]],["RotaryEmbedding",[Yh]],["ScatterND",[_f,wf]],["Sigmoid",[Gc]],["Sin",[Kc]],["Sinh",[jc]],["Slice",[xf,Tf]],["SkipLayerNormalization",[$f]],["Split",[jh,Qh]],["Sqrt",[Qc]],["Softmax",[Sf,kf]],["Sub",[uh]],["Tan",[Zc]],["Tanh",[Yc]],["ThresholdedRelu",[Jc,_r]],["Tile",[If]],["Transpose",[jp,Qp]],["Where",[Ef]]])}),W0=P(()=>{"use strict";Le(),lt(),ne(),zf=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,i,n){Je(e.programInfo.name);let a=this.backend.device,s=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let u=[];for(let d of t)u.push({binding:u.length,resource:{buffer:d.buffer}});for(let d of r)u.push({binding:u.length,resource:{buffer:d.buffer}});n&&u.push({binding:u.length,resource:n});let l=a.createBindGroup({layout:e.computePipeline.getBindGroupLayout(0),entries:u,label:e.programInfo.name});if(this.backend.sessionStatus==="capturing"){let d={kernelId:this.backend.currentKernelId,computePipeline:e.computePipeline,bindGroup:l,dispatchGroup:i};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(d)}s.setPipeline(e.computePipeline),s.setBindGroup(0,l),s.dispatchWorkgroups(...i),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),We(e.programInfo.name)}dispose(){}build(e,t){Je(e.name);let r=this.backend.device,i=[];[{feature:"shader-f16",extension:"f16"},{feature:"subgroups",extension:"subgroups"}].forEach(d=>{r.features.has(d.feature)&&i.push(`enable ${d.extension};`)});let n=Kp(t,this.backend.device.limits),a=e.getShaderSource(n),s=`${i.join(`
+`)}
+${n.additionalImplementations}
+${a}`,u=r.createShaderModule({code:s,label:e.name});de("verbose",()=>`[WebGPU] ${e.name} shader code: ${s}`);let l=r.createComputePipeline({compute:{module:u,entryPoint:"main"},layout:"auto",label:e.name});return We(e.name),{programInfo:e,computePipeline:l,uniformVariablesInfo:n.variablesInfo}}normalizeDispatchGroupSize(e){let t=typeof e=="number"?e:e.x,r=typeof e=="number"?1:e.y||1,i=typeof e=="number"?1:e.z||1,n=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(t<=n&&r<=n&&i<=n)return[t,r,i];let a=t*r*i,s=Math.ceil(Math.sqrt(a));if(s>n){if(s=Math.ceil(Math.cbrt(a)),s>n)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[s,s,s]}else return[s,s,1]}}}),Af={};Yt(Af,{WebGpuBackend:()=>Of});L0=P(()=>{"use strict";Le(),ee(),lt(),qp(),Yy(),U0(),W0(),rp=(e,t)=>{if(t.length!==e.length)throw new Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let r=[];for(let i=0;i<e.length;++i){let n=e[i].dataType;switch(t[i]){case"none":{r.push("");break}case"type":{r.push(`${n}`);break}case"rank":{let a=e[i].dims.length;r.push(`${n};${a}`);break}case"dims":{let a=e[i].dims.join(",");r.push(`${n};${a}`);break}default:throw new Error(`unsupported input dependency: ${t[i]}`)}}return r.join("|")},ip=(e,t,r)=>{let i=e.name;return e.shaderCache?.hint&&(i+="["+e.shaderCache.hint+"]"),i+=":"+r+`:${rp(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,i},np=class{constructor(e){e&&(this.architecture=e.architecture,this.vendor=e.vendor)}isArchitecture(e){return this.architecture===e}isVendor(e){return this.vendor===e}},Of=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=[],i={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},n=a=>t.features.has(a)&&r.push(a)&&!0;n("chromium-experimental-timestamp-query-inside-passes")||n("timestamp-query"),n("shader-f16"),n("subgroups"),this.device=await t.requestDevice(i),this.adapterInfo=new np(t.info||await t.requestAdapterInfo()),this.gpuDataManager=Fp(this),this.programManager=new zf(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,ca(e.logLevel,!!e.debug),this.device.onuncapturederror=a=>{a.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${a.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value: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;Je(),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(()=>{let t=new BigUint64Array(e.getMappedRange()),r=this.pendingQueries.get(e);for(let i=0;i<t.length/2;i++){let n=r[i],a=n.kernelId,s=this.kernels.get(a),u=s.kernelType,l=s.kernelName,d=n.programName,f=n.inputTensorViews,h=n.outputTensorViews,g=t[i*2],w=t[i*2+1];typeof this.queryTimeBase>"u"&&(this.queryTimeBase=g);let y=Number(g-this.queryTimeBase),v=Number(w-this.queryTimeBase);if(!Number.isSafeInteger(y)||!Number.isSafeInteger(v))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:f.map(T=>({dims:T.dims,dataType:ut(T.dataType)})),outputsMetadata:h.map(T=>({dims:T.dims,dataType:ut(T.dataType)})),kernelId:a,kernelType:u,kernelName:l,programName:d,startTime:y,endTime:v});else{let T="";f.forEach((_,I)=>{T+=`input[${I}]: [${_.dims}] | ${ut(_.dataType)}, `});let $="";h.forEach((_,I)=>{$+=`output[${I}]: [${_.dims}] | ${ut(_.dataType)}, `}),console.log(`[profiling] kernel "${a}|${u}|${l}|${d}" ${T}${$}start time: ${y} ns, execution time: ${v-y} ns`)}xr("GPU",`${d}::${g}::${w}`)}e.unmap(),this.pendingQueries.delete(e)}),We()}run(e,t,r,i,n,a){Je(e.name);let s=[];for(let _=0;_<t.length;++_){let I=t[_].data;if(I===0)continue;let k=this.gpuDataManager.get(I);if(!k)throw new Error(`no GPU data for input: ${I}`);s.push(k)}let{outputs:u,dispatchGroup:l,programUniforms:d}=e.getRunData(t),f=r.length===0?u.map((_,I)=>I):r;if(f.length!==u.length)throw new Error(`Output size ${f.length} must be equal to ${u.length}.`);let h=[],g=[];for(let _=0;_<u.length;++_){if(!Number.isInteger(f[_])||f[_]<-3||f[_]>=a)throw new Error(`Invalid output index: ${f[_]}`);if(f[_]===-3)continue;let I=f[_]===-1,k=f[_]===-2,E=I||k?n(u[_].dataType,u[_].dims):i(f[_],u[_].dataType,u[_].dims);if(h.push(E),E.data===0)continue;let C=this.gpuDataManager.get(E.data);if(!C)throw new Error(`no GPU data for output: ${E.data}`);if(I&&this.temporaryData.push(C),k){let O=this.kernelPersistentData.get(this.currentKernelId);O||(O=[],this.kernelPersistentData.set(this.currentKernelId,O)),O.push(C)}g.push(C)}if(s.length!==t.length||g.length!==h.length){if(g.length===0)return We(e.name),h;throw new Error(`Program ${e.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let w;if(d){let _=0,I=[];d.forEach(O=>{let x=typeof O.data=="number"?[O.data]:O.data;if(x.length===0)return;let D=O.type===10?2:4,L,Y;O.type===10?(Y=x.length>4?16:x.length>2?8:x.length*D,L=x.length>4?16:D*x.length):(Y=x.length<=2?x.length*D:16,L=16),_=Math.ceil(_/Y)*Y,I.push(_);let q=O.type===10?8:4;_+=x.length>4?Math.ceil(x.length/q)*L:x.length*D});let k=16;_=Math.ceil(_/k)*k;let E=new ArrayBuffer(_);d.forEach((O,x)=>{let D=I[x],L=typeof O.data=="number"?[O.data]:O.data;if(O.type===6)new Int32Array(E,D,L.length).set(L);else if(O.type===12)new Uint32Array(E,D,L.length).set(L);else if(O.type===10)new Uint16Array(E,D,L.length).set(L);else if(O.type===1)new Float32Array(E,D,L.length).set(L);else throw new Error(`Unsupported uniform type: ${ut(O.type)}`)});let C=this.gpuDataManager.create(_,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(C.buffer,0,E,0,_),this.gpuDataManager.release(C.id),w={offset:0,size:_,buffer:C.buffer}}let y=this.programManager.normalizeDispatchGroupSize(l),v=y[1]===1&&y[2]===1,T=ip(e,t,v),$=this.programManager.getArtifact(T);if($||($=this.programManager.build(e,y),this.programManager.setArtifact(T,$),de("info",()=>`[artifact] key: ${T}, programName: ${e.name}`)),d&&$.uniformVariablesInfo){if(d.length!==$.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${$.uniformVariablesInfo.length}, got ${d.length} in program "${$.programInfo.name}".`);for(let _=0;_<d.length;_++){let I=d[_],k=I.type,E=typeof I.data=="number"?1:I.data.length,[C,O]=$.uniformVariablesInfo[_];if(k!==C||E!==O)throw new Error(`Uniform variable ${_} mismatch: expect type ${C} with size ${O}, got type ${k} with size ${E} in program "${$.programInfo.name}".`)}}if(de("info",()=>`[ProgramManager] run "${e.name}" (key=${T}) with ${y[0]}x${y[1]}x${y[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let _={kernelId:this.currentKernelId,programName:$.programInfo.name,inputTensorViews:t,outputTensorViews:h};this.pendingKernels.push(_),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(_)}return this.programManager.run($,s,g,y,w),We(e.name),h}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,i){let n=Cf.get(e);if(!n)throw new Error(`kernel not implemented: ${e}`);let a={kernelType:e,kernelName:i,kernelEntry:n[0],attributes:[n[1],r]};this.kernels.set(t,a)}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 i=this.kernels.get(e);if(!i)throw new Error(`kernel not created: ${e}`);let n=i.kernelType,a=i.kernelName,s=i.kernelEntry,u=i.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${n}] ${a}" is not allowed to be called recursively`);this.currentKernelId=e,u[0]&&(u[1]=u[0](u[1]),u[0]=void 0),de("info",()=>`[WebGPU] Start to run kernel "[${n}] ${a}"...`);let l=this.env.debug;this.temporaryData=[];try{return l&&this.device.pushErrorScope("validation"),s(t,u[1]),0}catch(d){return r.push(Promise.resolve(`[WebGPU] Kernel "[${n}] ${a}" failed. ${d}`)),1}finally{l&&r.push(this.device.popErrorScope().then(d=>d?`GPU validation error for kernel "[${n}] ${a}": ${d.message}`:null));for(let d of this.temporaryData)this.gpuDataManager.release(d.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(e,t,r,i){let n=this.sessionExternalDataMapping.get(e);n||(n=new Map,this.sessionExternalDataMapping.set(e,n));let a=n.get(t),s=this.gpuDataManager.registerExternalBuffer(r,i,a);return n.set(t,[s,r]),s}unregisterBuffers(e){let t=this.sessionExternalDataMapping.get(e);t&&(t.forEach(r=>this.gpuDataManager.unregisterExternalBuffer(r[0])),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 i=await Vn(this,e,t);return ha(i.buffer,r)}}writeTimestamp(e){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,e)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||(typeof this.env.trace>"u"?this.env.wasm.trace:this.env.trace))&&(this.device.features.has("chromium-experimental-timestamp-query-inside-passes")?this.queryType="inside-passes":this.device.features.has("timestamp-query")&&(this.queryType="at-passes"),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){de("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(){de("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){de("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 i=0;i<r;i++){let n=this.getComputePassEncoder(),a=e[i];this.writeTimestamp(this.pendingDispatchNumber*2),n.setPipeline(a.computePipeline),n.setBindGroup(0,a.bindGroup),n.dispatchWorkgroups(...a.dispatchGroup),this.writeTimestamp(this.pendingDispatchNumber*2+1),this.pendingDispatchNumber++,this.queryType!=="none"&&this.pendingKernels.push(t[i]),(this.pendingDispatchNumber>=this.maxDispatchNumber||this.queryType==="at-passes")&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}onCreateSession(){this.gpuDataManager.onCreateSession()}onReleaseSession(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()}}}),Rf={};Yt(Rf,{init:()=>Mf});q0=P(()=>{"use strict";ee(),lt(),ie(),Zy(),Jr=class Bf{constructor(t,r,i,n){this.module=t,this.dataType=r,this.data=i,this.dims=n}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=R.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=R.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=R.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}getUint16Array(){if(this.dataType!==10&&this.dataType!==4)throw new Error("Invalid data type");let t=R.size(this.dims);return t===0?new Uint16Array:new Uint16Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(R.size(t)!==R.size(this.dims))throw new Error("Invalid new shape");return new Bf(this.module,this.dataType,this.data,t)}},ap=class{constructor(e,t,r){this.module=e,this.backend=t,this.customDataOffset=0,this.customDataSize=0,this.adapterInfo=t.adapterInfo;let i=e.PTR_SIZE,n=r/e.PTR_SIZE,a=i===4?"i32":"i64";this.opKernelContext=Number(e.getValue(i*n++,a));let s=Number(e.getValue(i*n++,a));this.outputCount=Number(e.getValue(i*n++,a)),this.customDataOffset=Number(e.getValue(i*n++,"*")),this.customDataSize=Number(e.getValue(i*n++,a));let u=[];for(let l=0;l<s;l++){let d=Number(e.getValue(i*n++,a)),f=Number(e.getValue(i*n++,"*")),h=Number(e.getValue(i*n++,a)),g=[];for(let w=0;w<h;w++)g.push(Number(e.getValue(i*n++,a)));u.push(new Jr(e,d,f,g))}this.inputs=u}get kernelCustomData(){return this.backend.currentKernelCustomData}get customDataBuffer(){return this.module.HEAPU8.subarray(this.customDataOffset,this.customDataOffset+this.customDataSize)}compute(e,t){let r=t?.inputs?.map(s=>typeof s=="number"?this.inputs[s]:s)??this.inputs,i=t?.outputs??[],n=(s,u,l)=>new Jr(this.module,u,this.output(s,l),l),a=(s,u)=>{let l=Mt(s,u);if(!l)throw new Error(`Unsupported data type: ${s}`);let d=l>0?this.backend.gpuDataManager.create(l).id:0;return new Jr(this.module,s,d,u)};return this.backend.run(e,r,i,n,a,this.outputCount)}output(e,t){let r=this.module.stackSave();try{let i=this.module.PTR_SIZE,n=i===4?"i32":"i64",a=this.module.stackAlloc((1+t.length)*i);this.module.setValue(a,t.length,n);for(let s=0;s<t.length;s++)this.module.setValue(a+i*(s+1),t[s],n);return this.module._JsepOutput(this.opKernelContext,e,a)}catch(i){throw new Error(`Failed to generate kernel's output[${e}] with dims [${t}]. If you are running with pre-allocated output, please make sure the output type/dims are correct. Error: ${i}`)}finally{this.module.stackRestore(r)}}},Mf=async(e,t,r,i)=>{let n=t.jsepInit;if(!n)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let a=(L0(),$r(Af)).WebGpuBackend,s=new a;await s.initialize(r,i),n("webgpu",[s,u=>s.alloc(Number(u)),u=>s.free(u),(u,l,d,f=!1)=>{if(f)de("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${Number(u)}, dst=${Number(l)}, size=${Number(d)}`),s.memcpy(Number(u),Number(l));else{de("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${Number(u)}, gpuDataId=${Number(l)}, size=${Number(d)}`);let h=t.HEAPU8.subarray(Number(u>>>0),Number(u>>>0)+Number(d));s.upload(Number(l),h)}},async(u,l,d)=>{de("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${u}, dataOffset=${l}, size=${d}`),await s.download(Number(u),()=>t.HEAPU8.subarray(Number(l)>>>0,Number(l+d)>>>0))},(u,l,d)=>s.createKernel(u,Number(l),d,t.UTF8ToString(t._JsepGetNodeName(Number(l)))),u=>s.releaseKernel(u),(u,l,d,f)=>{de("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${d}, kernel=${u}, contextDataOffset=${l}`);let h=new ap(t,s,Number(l));return s.computeKernel(Number(u),h,f)},()=>s.captureBegin(),()=>s.captureEnd(),()=>s.replay()])}else{let a=new Gp(r);n("webnn",[a,()=>a.reserveTensorId(),s=>a.releaseTensorId(s),async(s,u,l,d,f)=>a.ensureTensor(s,u,l,d,f),(s,u)=>{a.uploadTensor(s,u)},async(s,u)=>a.downloadTensor(s,u),(s,u)=>a.registerMLContext(s,u),!!r.trace])}}}),Nf=P(()=>{"use strict";Le(),Ky(),jy(),ee(),Ut(),ua(),Pp(),sp=(e,t)=>{we()._OrtInit(e,t)!==0&&ge("Can't initialize onnxruntime.")},Ta=async e=>{sp(e.wasm.numThreads,ai(e.logLevel))},Sa=async(e,t)=>{we().asyncInit?.();let r=e.webgpu.adapter;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");if(r){if(typeof r.limits!="object"||typeof r.features!="object"||typeof r.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let i=e.webgpu.powerPreference;if(i!==void 0&&i!=="low-power"&&i!=="high-performance")throw new Error(`Invalid powerPreference setting: "${i}"`);let n=e.webgpu.forceFallbackAdapter;if(n!==void 0&&typeof n!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${n}"`);if(r=await navigator.gpu.requestAdapter({powerPreference:i,forceFallbackAdapter:n}),!r)throw new Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.')}}if(t==="webnn"&&(typeof navigator>"u"||!navigator.ml))throw new Error("WebNN is not supported in current environment");{let i=(q0(),$r(Rf)).init;t==="webgpu"&&await i("webgpu",we(),e,r),t==="webnn"&&await i("webnn",we(),e)}},yt=new Map,op=e=>{let t=we(),r=t.stackSave();try{let i=t.PTR_SIZE,n=t.stackAlloc(2*i);t._OrtGetInputOutputCount(e,n,n+i)!==0&&ge("Can't get session input/output count.");let a=i===4?"i32":"i64";return[Number(t.getValue(n,a)),Number(t.getValue(n+i,a))]}finally{t.stackRestore(r)}},Nn=(e,t)=>{let r=we(),i=r.stackSave(),n=0;try{let a=r.PTR_SIZE,s=r.stackAlloc(2*a);r._OrtGetInputOutputMetadata(e,t,s,s+a)!==0&&ge("Can't get session input/output metadata.");let u=Number(r.getValue(s,"*"));n=Number(r.getValue(s+a,"*"));let l=r.HEAP32[n/4];if(l===0)return[u,0];let d=r.HEAPU32[n/4+1],f=[];for(let h=0;h<d;h++){let g=Number(r.getValue(n+8+h*a,"*"));f.push(g!==0?r.UTF8ToString(g):Number(r.getValue(n+8+(h+d)*a,"*")))}return[u,l,f]}finally{r.stackRestore(i),n!==0&&r._OrtFree(n)}},di=e=>{let t=we(),r=t._malloc(e.byteLength);if(r===0)throw new Error(`Can't create a session. failed to allocate a buffer of size ${e.byteLength}.`);return t.HEAPU8.set(e,r),[r,e.byteLength]},ka=async(e,t)=>{let r,i,n=we();Array.isArray(e)?[r,i]=e:e.buffer===n.HEAPU8.buffer?[r,i]=[e.byteOffset,e.byteLength]:[r,i]=di(e);let a=0,s=0,u=0,l=[],d=[],f=[];try{if([s,l]=await Dp(t),t?.externalData&&n.mountExternalData){let k=[];for(let E of t.externalData){let C=typeof E=="string"?E:E.path;k.push(pa(typeof E=="string"?E:E.data).then(O=>{n.mountExternalData(C,O)}))}await Promise.all(k)}for(let k of t?.executionProviders??[])if((typeof k=="string"?k:k.name)==="webnn"){if(n.shouldTransferToMLTensor=!1,typeof k!="string"){let E=k,C=E?.context,O=E?.gpuDevice,x=E?.deviceType,D=E?.powerPreference;C?n.currentContext=C:O?n.currentContext=await n.webnnCreateMLContext(O):n.currentContext=await n.webnnCreateMLContext({deviceType:x,powerPreference:D})}else n.currentContext=await n.webnnCreateMLContext();break}a=await n._OrtCreateSession(r,i,s),n.webgpuOnCreateSession?.(a),a===0&&ge("Can't create a session."),n.jsepOnCreateSession?.(),n.currentContext&&(n.webnnRegisterMLContext(a,n.currentContext),n.currentContext=void 0,n.shouldTransferToMLTensor=!0);let[h,g]=op(a),w=!!t?.enableGraphCapture,y=[],v=[],T=[],$=[],_=[];for(let k=0;k<h;k++){let[E,C,O]=Nn(a,k);E===0&&ge("Can't get an input name."),d.push(E);let x=n.UTF8ToString(E);y.push(x),T.push(C===0?{name:x,isTensor:!1}:{name:x,isTensor:!0,type:ut(C),shape:O})}for(let k=0;k<g;k++){let[E,C,O]=Nn(a,k+h);E===0&&ge("Can't get an output name."),f.push(E);let x=n.UTF8ToString(E);v.push(x),$.push(C===0?{name:x,isTensor:!1}:{name:x,isTensor:!0,type:ut(C),shape:O});{if(w&&t?.preferredOutputLocation===void 0){_.push("gpu-buffer");continue}let D=typeof t?.preferredOutputLocation=="string"?t.preferredOutputLocation:t?.preferredOutputLocation?.[x]??"cpu",L=n.webnnIsGraphOutput;if(D==="cpu"&&L&&L(a,x)){_.push("ml-tensor-cpu-output");continue}if(D!=="cpu"&&D!=="cpu-pinned"&&D!=="gpu-buffer"&&D!=="ml-tensor")throw new Error(`Not supported preferred output location: ${D}.`);if(w&&D!=="gpu-buffer")throw new Error(`Not supported preferred output location: ${D}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);_.push(D)}}let I=null;return _.some(k=>k==="gpu-buffer"||k==="ml-tensor"||k==="ml-tensor-cpu-output")&&(u=n._OrtCreateBinding(a),u===0&&ge("Can't create IO binding."),I={handle:u,outputPreferredLocations:_,outputPreferredLocationsEncoded:_.map(k=>k==="ml-tensor-cpu-output"?"ml-tensor":k).map(k=>Ln(k))}),yt.set(a,[a,d,f,I,w,!1]),[a,y,v,T,$]}catch(h){throw d.forEach(g=>n._OrtFree(g)),f.forEach(g=>n._OrtFree(g)),u!==0&&n._OrtReleaseBinding(u)!==0&&ge("Can't release IO binding."),a!==0&&n._OrtReleaseSession(a)!==0&&ge("Can't release session."),h}finally{n._free(r),s!==0&&n._OrtReleaseSessionOptions(s)!==0&&ge("Can't release session options."),l.forEach(h=>n._free(h)),n.unmountExternalData?.()}},Ia=e=>{let t=we(),r=yt.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[i,n,a,s,u]=r;s&&(u&&t._OrtClearBoundOutputs(s.handle)!==0&&ge("Can't clear bound outputs."),t._OrtReleaseBinding(s.handle)!==0&&ge("Can't release IO binding.")),t.jsepOnReleaseSession?.(e),t.webnnOnReleaseSession?.(e),t.webgpuOnReleaseSession?.(e),n.forEach(l=>t._OrtFree(l)),a.forEach(l=>t._OrtFree(l)),t._OrtReleaseSession(i)!==0&&ge("Can't release session."),yt.delete(e)},Dn=async(e,t,r,i,n,a,s=!1)=>{if(!e){t.push(0);return}let u=we(),l=u.PTR_SIZE,d=e[0],f=e[1],h=e[3],g=h,w,y;if(d==="string"&&(h==="gpu-buffer"||h==="ml-tensor"))throw new Error("String tensor is not supported on GPU.");if(s&&h!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${a} when enableGraphCapture is true.`);if(h==="gpu-buffer"){let $=e[2].gpuBuffer;y=Mt(Rt(d),f);{let _=u.jsepRegisterBuffer;if(!_)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');w=_(i,a,$,y)}}else if(h==="ml-tensor"){let $=e[2].mlTensor;y=Mt(Rt(d),f);let _=u.webnnRegisterMLTensor;if(!_)throw new Error('Tensor location "ml-tensor" is not supported without using WebNN.');w=_(i,$,Rt(d),f)}else{let $=e[2];if(Array.isArray($)){y=l*$.length,w=u._malloc(y),r.push(w);for(let _=0;_<$.length;_++){if(typeof $[_]!="string")throw new TypeError(`tensor data at index ${_} is not a string`);u.setValue(w+_*l,Ze($[_],r),"*")}}else{let _=u.webnnIsGraphInput,I=u.webnnIsGraphOutput;if(d!=="string"&&_&&I){let k=u.UTF8ToString(n);if(_(i,k)||I(i,k)){let E=Rt(d);y=Mt(E,f),g="ml-tensor";let C=u.webnnCreateTemporaryTensor,O=u.webnnUploadTensor;if(!C||!O)throw new Error('Tensor location "ml-tensor" is not supported without using WebNN.');let x=await C(i,E,f);O(x,new Uint8Array($.buffer,$.byteOffset,$.byteLength)),w=x}else y=$.byteLength,w=u._malloc(y),r.push(w),u.HEAPU8.set(new Uint8Array($.buffer,$.byteOffset,y),w)}else y=$.byteLength,w=u._malloc(y),r.push(w),u.HEAPU8.set(new Uint8Array($.buffer,$.byteOffset,y),w)}}let v=u.stackSave(),T=u.stackAlloc(4*f.length);try{f.forEach((_,I)=>u.setValue(T+I*l,_,l===4?"i32":"i64"));let $=u._OrtCreateTensor(Rt(d),w,y,T,f.length,Ln(g));$===0&&ge(`Can't create tensor for input/output. session=${i}, index=${a}.`),t.push($)}finally{u.stackRestore(v)}},Ea=async(e,t,r,i,n,a)=>{let s=we(),u=s.PTR_SIZE,l=yt.get(e);if(!l)throw new Error(`cannot run inference. invalid session id: ${e}`);let d=l[0],f=l[1],h=l[2],g=l[3],w=l[4],y=l[5],v=t.length,T=i.length,$=0,_=[],I=[],k=[],E=[],C=[],O=s.stackSave(),x=s.stackAlloc(v*u),D=s.stackAlloc(v*u),L=s.stackAlloc(T*u),Y=s.stackAlloc(T*u);try{[$,_]=Np(a),_t("wasm prepareInputOutputTensor");for(let A=0;A<v;A++)await Dn(r[A],I,E,e,f[t[A]],t[A],w);for(let A=0;A<T;A++)await Dn(n[A],k,E,e,h[i[A]],v+i[A],w);bt("wasm prepareInputOutputTensor");for(let A=0;A<v;A++)s.setValue(x+A*u,I[A],"*"),s.setValue(D+A*u,f[t[A]],"*");for(let A=0;A<T;A++)s.setValue(L+A*u,k[A],"*"),s.setValue(Y+A*u,h[i[A]],"*");if(g&&!y){let{handle:A,outputPreferredLocations:U,outputPreferredLocationsEncoded:re}=g;if(f.length!==v)throw new Error(`input count from feeds (${v}) is expected to be always equal to model's input count (${f.length}).`);_t("wasm bindInputsOutputs");for(let J=0;J<v;J++){let H=t[J];await s._OrtBindInput(A,f[H],I[J])!==0&&ge(`Can't bind input[${J}] for session=${e}.`)}for(let J=0;J<T;J++){let H=i[J];n[J]?.[3]?(C.push(k[J]),s._OrtBindOutput(A,h[H],k[J],0)!==0&&ge(`Can't bind pre-allocated output[${J}] for session=${e}.`)):s._OrtBindOutput(A,h[H],0,re[H])!==0&&ge(`Can't bind output[${J}] to ${U[J]} for session=${e}.`)}bt("wasm bindInputsOutputs"),yt.set(e,[d,f,h,g,w,!0])}s.jsepOnRunStart?.(d),s.webnnOnRunStart?.(d);let q;g?q=await s._OrtRunWithBinding(d,g.handle,T,L,$):q=await s._OrtRun(d,D,x,v,Y,T,L,$),q!==0&&ge("failed to call OrtRun().");let j=[],ue=[];_t("wasm ProcessOutputTensor");for(let A=0;A<T;A++){let U=Number(s.getValue(L+A*u,"*"));if(U===k[A]||C.includes(k[A])){j.push(n[A]),U!==k[A]&&s._OrtReleaseTensor(U)!==0&&ge("Can't release tensor.");continue}let re=s.stackSave(),J=s.stackAlloc(4*u),H=!1,pe,N=0;try{s._OrtGetTensorData(U,J,J+u,J+2*u,J+3*u)!==0&&ge(`Can't access output tensor data on index ${A}.`);let V=u===4?"i32":"i64",te=Number(s.getValue(J,V));N=s.getValue(J+u,"*");let ce=s.getValue(J+u*2,"*"),Ie=Number(s.getValue(J+u*3,V)),Me=[];for(let Se=0;Se<Ie;Se++)Me.push(Number(s.getValue(ce+Se*u,V)));s._OrtFree(ce)!==0&&ge("Can't free memory for tensor dims.");let Ve=Me.reduce((Se,$e)=>Se*$e,1);pe=ut(te);let nt=g?.outputPreferredLocations[i[A]];if(pe==="string"){if(nt==="gpu-buffer"||nt==="ml-tensor")throw new Error("String tensor is not supported on GPU.");let Se=[];for(let $e=0;$e<Ve;$e++){let Re=s.getValue(N+$e*u,"*"),xt=s.getValue(N+($e+1)*u,"*"),kr=$e===Ve-1?void 0:xt-Re;Se.push(s.UTF8ToString(Re,kr))}j.push([pe,Me,Se,"cpu"])}else if(nt==="gpu-buffer"&&Ve>0){let Se=s.jsepGetBuffer;if(!Se)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let $e=Se(N),Re=Mt(te,Ve);if(Re===void 0||!la(pe))throw new Error(`Unsupported data type: ${pe}`);H=!0,j.push([pe,Me,{gpuBuffer:$e,download:s.jsepCreateDownloader($e,Re,pe),dispose:()=>{s._OrtReleaseTensor(U)!==0&&ge("Can't release tensor.")}},"gpu-buffer"])}else if(nt==="ml-tensor"&&Ve>0){let Se=s.webnnEnsureTensor,$e=s.webnnIsGraphInputOutputTypeSupported;if(!Se||!$e)throw new Error('preferredLocation "ml-tensor" is not supported without using WebNN.');if(Mt(te,Ve)===void 0||!da(pe))throw new Error(`Unsupported data type: ${pe}`);if(!$e(e,pe,!1))throw new Error(`preferredLocation "ml-tensor" for ${pe} output is not supported by current WebNN Context.`);let Re=await Se(e,N,te,Me,!1);H=!0,j.push([pe,Me,{mlTensor:Re,download:s.webnnCreateMLTensorDownloader(N,pe),dispose:()=>{s.webnnReleaseTensorId(N),s._OrtReleaseTensor(U)}},"ml-tensor"])}else if(nt==="ml-tensor-cpu-output"&&Ve>0){let Se=s.webnnCreateMLTensorDownloader(N,pe)(),$e=j.length;H=!0,ue.push((async()=>{let Re=[$e,await Se];return s.webnnReleaseTensorId(N),s._OrtReleaseTensor(U),Re})()),j.push([pe,Me,[],"cpu"])}else{let Se=pi(pe),$e=new Se(Ve);new Uint8Array($e.buffer,$e.byteOffset,$e.byteLength).set(s.HEAPU8.subarray(N,N+$e.byteLength)),j.push([pe,Me,$e,"cpu"])}}finally{s.stackRestore(re),pe==="string"&&N&&s._free(N),H||s._OrtReleaseTensor(U)}}g&&!w&&(s._OrtClearBoundOutputs(g.handle)!==0&&ge("Can't clear bound outputs."),yt.set(e,[d,f,h,g,w,!1]));for(let[A,U]of await Promise.all(ue))j[A][2]=U;return bt("wasm ProcessOutputTensor"),j}finally{s.webnnOnRunEnd?.(d),s.stackRestore(O),I.forEach(q=>s._OrtReleaseTensor(q)),k.forEach(q=>s._OrtReleaseTensor(q)),E.forEach(q=>s._free(q)),$!==0&&s._OrtReleaseRunOptions($),_.forEach(q=>s._free(q))}},Ca=e=>{let t=we(),r=yt.get(e);if(!r)throw new Error("invalid session id");let i=r[0],n=t._OrtEndProfiling(i);n===0&&ge("Can't get an profile file name."),t._OrtFree(n)},za=e=>{let t=[];for(let r of e){let i=r[2];!Array.isArray(i)&&"buffer"in i&&t.push(i.buffer)}return t}}),Gf=P(()=>{"use strict";Le(),Nf(),Ut(),sa(),wt=()=>!!be.wasm.proxy&&typeof document<"u",Kt=!1,mr=!1,gr=!1,ti=new Map,zt=(e,t)=>{let r=ti.get(e);r?r.push(t):ti.set(e,[t])},At=()=>{if(Kt||!mr||gr||!Ue)throw new Error("worker not ready")},up=e=>{switch(e.data.type){case"init-wasm":Kt=!1,e.data.err?(gr=!0,Pn[1](e.data.err)):(mr=!0,Pn[0]()),ei&&(URL.revokeObjectURL(ei),ei=void 0);break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let t=ti.get(e.data.type);e.data.err?t.shift()[1](e.data.err):t.shift()[0](e.data.out);break}default:}},Df=async()=>{if(!mr){if(Kt)throw new Error("multiple calls to 'initWasm()' detected.");if(gr)throw new Error("previous call to 'initWasm()' failed.");if(Kt=!0,wt())return new Promise((e,t)=>{Ue?.terminate(),Mp().then(([r,i])=>{try{Ue=i,Ue.onerror=a=>t(a),Ue.onmessage=up,Pn=[e,t];let n={type:"init-wasm",in:be};!n.in.wasm.wasmPaths&&(r||Wn)&&(n.in.wasm.wasmPaths={wasm:new URL("ort-wasm-simd-threaded.jsep.wasm",Ye.url).href}),Ue.postMessage(n),ei=r}catch(n){t(n)}},t)});try{await oa(be.wasm),await Ta(be),mr=!0}catch(e){throw gr=!0,e}finally{Kt=!1}}},Pf=async e=>{if(wt())return At(),new Promise((t,r)=>{zt("init-ep",[t,r]);let i={type:"init-ep",in:{epName:e,env:be}};Ue.postMessage(i)});await Sa(be,e)},Uf=async e=>wt()?(At(),new Promise((t,r)=>{zt("copy-from",[t,r]);let i={type:"copy-from",in:{buffer:e}};Ue.postMessage(i,[e.buffer])})):di(e),Wf=async(e,t)=>{if(wt()){if(t?.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return At(),new Promise((r,i)=>{zt("create",[r,i]);let n={type:"create",in:{model:e,options:{...t}}},a=[];e instanceof Uint8Array&&a.push(e.buffer),Ue.postMessage(n,a)})}else return ka(e,t)},Lf=async e=>{if(wt())return At(),new Promise((t,r)=>{zt("release",[t,r]);let i={type:"release",in:e};Ue.postMessage(i)});Ia(e)},qf=async(e,t,r,i,n,a)=>{if(wt()){if(r.some(s=>s[3]!=="cpu"))throw new Error("input tensor on GPU is not supported for proxy.");if(n.some(s=>s))throw new Error("pre-allocated output tensor is not supported for proxy.");return At(),new Promise((s,u)=>{zt("run",[s,u]);let l=r,d={type:"run",in:{sessionId:e,inputIndices:t,inputs:l,outputIndices:i,options:a}};Ue.postMessage(d,za(l))})}else return Ea(e,t,r,i,n,a)},Vf=async e=>{if(wt())return At(),new Promise((t,r)=>{zt("end-profiling",[t,r]);let i={type:"end-profiling",in:e};Ue.postMessage(i)});Ca(e)}}),V0=P(()=>{"use strict";Le(),Gf(),ee(),aa(),Pp(),Un=(e,t)=>{switch(e.location){case"cpu":return[e.type,e.dims,e.data,"cpu"];case"gpu-buffer":return[e.type,e.dims,{gpuBuffer:e.gpuBuffer},"gpu-buffer"];case"ml-tensor":return[e.type,e.dims,{mlTensor:e.mlTensor},"ml-tensor"];default:throw new Error(`invalid data location: ${e.location} for ${t()}`)}},lp=e=>{switch(e[3]){case"cpu":return new Xe(e[0],e[2],e[1]);case"gpu-buffer":{let t=e[0];if(!la(t))throw new Error(`not supported data type: ${t} for deserializing GPU tensor`);let{gpuBuffer:r,download:i,dispose:n}=e[2];return Xe.fromGpuBuffer(r,{dataType:t,dims:e[1],download:i,dispose:n})}case"ml-tensor":{let t=e[0];if(!da(t))throw new Error(`not supported data type: ${t} for deserializing MLTensor tensor`);let{mlTensor:r,download:i,dispose:n}=e[2];return Xe.fromMLTensor(r,{dataType:t,dims:e[1],download:i,dispose:n})}default:throw new Error(`invalid data location: ${e[3]}`)}},Ff=class{async fetchModelAndCopyToWasmMemory(e){return Uf(await pa(e))}async loadModel(e,t){Je();let r;typeof e=="string"?r=await this.fetchModelAndCopyToWasmMemory(e):r=e,[this.sessionId,this.inputNames,this.outputNames,this.inputMetadata,this.outputMetadata]=await Wf(r,t),We()}async dispose(){return Lf(this.sessionId)}async run(e,t,r){Je();let i=[],n=[];Object.entries(e).forEach(h=>{let g=h[0],w=h[1],y=this.inputNames.indexOf(g);if(y===-1)throw new Error(`invalid input '${g}'`);i.push(w),n.push(y)});let a=[],s=[];Object.entries(t).forEach(h=>{let g=h[0],w=h[1],y=this.outputNames.indexOf(g);if(y===-1)throw new Error(`invalid output '${g}'`);a.push(w),s.push(y)});let u=i.map((h,g)=>Un(h,()=>`input "${this.inputNames[n[g]]}"`)),l=a.map((h,g)=>h?Un(h,()=>`output "${this.outputNames[s[g]]}"`):null),d=await qf(this.sessionId,n,u,s,l,r),f={};for(let h=0;h<d.length;h++)f[this.outputNames[s[h]]]=a[h]??lp(d[h]);return We(),f}startProfiling(){}endProfiling(){Vf(this.sessionId)}}}),Hf={};Yt(Hf,{OnnxruntimeWebAssemblyBackend:()=>ta,initializeFlags:()=>ea,wasmBackend:()=>Kf});G0=P(()=>{"use strict";Le(),Gf(),V0(),ea=()=>{(typeof be.wasm.initTimeout!="number"||be.wasm.initTimeout<0)&&(be.wasm.initTimeout=0);let e=be.wasm.simd;if(typeof e!="boolean"&&e!==void 0&&e!=="fixed"&&e!=="relaxed"&&(console.warn(`Property "env.wasm.simd" is set to unknown value "${e}". Reset it to \`false\` and ignore SIMD feature checking.`),be.wasm.simd=!1),typeof be.wasm.proxy!="boolean"&&(be.wasm.proxy=!1),typeof be.wasm.trace!="boolean"&&(be.wasm.trace=!1),typeof be.wasm.numThreads!="number"||!Number.isInteger(be.wasm.numThreads)||be.wasm.numThreads<=0)if(typeof self<"u"&&!self.crossOriginIsolated)be.wasm.numThreads=1;else{let t=typeof navigator>"u"?zy("node:os").cpus().length:navigator.hardwareConcurrency;be.wasm.numThreads=Math.min(4,Math.ceil((t||1)/2))}},ta=class{async init(e){ea(),await Df(),await Pf(e)}async createInferenceSessionHandler(e,t){let r=new Ff;return await r.loadModel(e,t),r}},Kf=new ta});Le();Le();Le();F0="1.24.2",H0=Ep;{let e=(G0(),$r(Hf)).wasmBackend;Bt("webgpu",e,5),Bt("webnn",e,5),Bt("cpu",e,10),Bt("wasm",e,10)}Object.defineProperty(be.versions,"web",{value:F0,enumerable:!0});});var J0={};zo(J0,{DeviceAuthManager:()=>hi,ExperimentsClient:()=>yi,FederatedClient:()=>fi,InferenceEngine:()=>Xt,ModelLoader:()=>Jt,Octomil:()=>ci,OctomilError:()=>X,RolloutsManager:()=>gi,SecAggPlus:()=>mi,SecureAggregation:()=>Sr,StreamingInferenceEngine:()=>Gt,TelemetryReporter:()=>Vt,WeightExtractor:()=>er,addGaussianNoise:()=>im,assertModelIntegrity:()=>em,clipGradients:()=>rm,computeHash:()=>Oa,createModelCache:()=>qr,dequantize:()=>am,disposeTelemetry:()=>Xf,getTelemetry:()=>Yf,initTelemetry:()=>Zf,quantize:()=>nm,shamirReconstruct:()=>Ba,shamirSplit:()=>Ma,verifyModelIntegrity:()=>Ra});var X=class extends Error{constructor(r,i,n){super(i);Q(this,"code");Q(this,"cause");this.name="OctomilError",this.code=r,this.cause=n}};var Ty="octomil-models-v1",Sy="octomil-models",Wr="blobs",ky=1,qi=class{async open(){return caches.open(Ty)}async get(t){let i=await(await this.open()).match(t);return i?i.arrayBuffer():null}async put(t,r){let i=await this.open(),n=new Response(r,{headers:{"x-octomil-cached-at":new Date().toISOString(),"x-octomil-size":String(r.byteLength)}});await i.put(t,n)}async has(t){return await(await this.open()).match(t)!==void 0}async remove(t){await(await this.open()).delete(t)}async info(t){let i=await(await this.open()).match(t);if(!i)return{cached:!1,sizeBytes:0};let n=i.headers.get("x-octomil-size"),a=i.headers.get("x-octomil-cached-at");return{cached:!0,sizeBytes:n?parseInt(n,10):0,cachedAt:a??void 0}}},Lr=class{constructor(){Q(this,"dbPromise",null)}openDB(){return this.dbPromise?this.dbPromise:(this.dbPromise=new Promise((t,r)=>{let i=indexedDB.open(Sy,ky);i.onupgradeneeded=()=>{let n=i.result;n.objectStoreNames.contains(Wr)||n.createObjectStore(Wr,{keyPath:"key"})},i.onsuccess=()=>t(i.result),i.onerror=()=>r(i.error)}),this.dbPromise)}async tx(t){return(await this.openDB()).transaction(Wr,t).objectStore(Wr)}async get(t){let r=await this.tx("readonly");return new Promise((i,n)=>{let a=r.get(t);a.onsuccess=()=>{let s=a.result;i(s?.data??null)},a.onerror=()=>n(a.error)})}async put(t,r){let i=await this.tx("readwrite"),n={key:t,data:r,sizeBytes:r.byteLength,cachedAt:new Date().toISOString()};return new Promise((a,s)=>{let u=i.put(n);u.onsuccess=()=>a(),u.onerror=()=>s(u.error)})}async has(t){let r=await this.tx("readonly");return new Promise((i,n)=>{let a=r.count(t);a.onsuccess=()=>i(a.result>0),a.onerror=()=>n(a.error)})}async remove(t){let r=await this.tx("readwrite");return new Promise((i,n)=>{let a=r.delete(t);a.onsuccess=()=>i(),a.onerror=()=>n(a.error)})}async info(t){let r=await this.tx("readonly");return new Promise((i,n)=>{let a=r.get(t);a.onsuccess=()=>{let s=a.result;i(s?{cached:!0,sizeBytes:s.sizeBytes,cachedAt:s.cachedAt}:{cached:!1,sizeBytes:0})},a.onerror=()=>n(a.error)})}},ar=class{async get(t){return null}async put(t,r){}async has(t){return!1}async remove(t){}async info(t){return{cached:!1,sizeBytes:0}}};function qr(e){if(e==="none")return new ar;if(e==="cache-api")return typeof caches<"u"?new qi:typeof indexedDB<"u"?new Lr:new ar;if(e==="indexeddb")return typeof indexedDB<"u"?new Lr:new ar;throw new X("CACHE_ERROR",`Unknown cache strategy: ${e}`)}var Xt=class{constructor(){Q(this,"session",null);Q(this,"ortModule",null);Q(this,"resolvedBackend",null)}async createSession(t,r){let i=await this.loadOrt(),n=await this.resolveProvider(i,r);this.resolvedBackend=n==="webgpu"?"webgpu":"wasm";let a={executionProviders:[n],graphOptimizationLevel:"all"};try{this.session=await i.InferenceSession.create(t,a)}catch(s){if(n==="webgpu")try{this.session=await i.InferenceSession.create(t,{executionProviders:["wasm"],graphOptimizationLevel:"all"}),this.resolvedBackend="wasm"}catch(u){throw new X("MODEL_LOAD_FAILED","Failed to create ONNX session with both WebGPU and WASM backends.",u)}else throw new X("MODEL_LOAD_FAILED",`Failed to create ONNX session: ${String(s)}`,s)}}async run(t){this.ensureSession();let r=this.ortModule,i=this.session,n={};for(let[f,h]of Object.entries(t))n[f]=new r.Tensor(K0(h.data),h.data,h.dims);let a=performance.now(),s;try{s=await i.run(n)}catch(f){throw new X("INFERENCE_FAILED",`Inference run failed: ${String(f)}`,f)}let u=performance.now()-a,l=this.convertOutputs(s),d=this.extractConvenience(l);return{tensors:l,latencyMs:u,...d}}get inputNames(){return this.ensureSession(),this.session.inputNames}get outputNames(){return this.ensureSession(),this.session.outputNames}get activeBackend(){return this.resolvedBackend}dispose(){this.session&&(this.session.release(),this.session=null)}async loadOrt(){if(this.ortModule)return this.ortModule;try{return this.ortModule=await Promise.resolve().then(()=>(Qf(),jf)),this.ortModule}catch(t){throw new X("BACKEND_UNAVAILABLE","Failed to import onnxruntime-web. Make sure the package is installed: npm i onnxruntime-web",t)}}async resolveProvider(t,r){if(r==="wasm")return"wasm";if(r==="webgpu"||r===void 0){if(await this.detectWebGPU())return"webgpu";if(r==="webgpu")throw new X("BACKEND_UNAVAILABLE","WebGPU was explicitly requested but is not available in this browser.")}return"wasm"}async detectWebGPU(){if(typeof navigator>"u")return!1;try{let t=navigator.gpu;return t?await t.requestAdapter()!==null:!1}catch{return!1}}ensureSession(){if(!this.session)throw new X("SESSION_DISPOSED","No active session. Call load() before running inference.")}convertOutputs(t){let r={};for(let i of Object.keys(t)){let n=t[i];r[i]={data:n.data,dims:Array.from(n.dims)}}return r}extractConvenience(t){let r=Object.keys(t);if(r.length===0)return{};let n=t[r[0]].data;if(!(n instanceof Float32Array))return{};if(n.length===0)return{};let a=Array.from(n),s=0,u=-1/0;for(let l=0;l<a.length;l++)a[l]>u&&(u=a[l],s=l);return{label:String(s),score:u,scores:a}}};function K0(e){return e instanceof Float32Array?"float32":e instanceof Int32Array?"int32":e instanceof BigInt64Array?"int64":e instanceof Uint8Array?"uint8":"float32"}var Lt=3,Aa=1e3,Jt=class{constructor(t,r){Q(this,"modelId");Q(this,"serverUrl");Q(this,"apiKey");Q(this,"onProgress");Q(this,"cache");this.modelId=t.model,this.serverUrl=t.serverUrl,this.apiKey=t.apiKey,this.onProgress=t.onProgress,this.cache=r}async load(){let t=await this.resolveModelUrl(),r=await this.cache.get(t);if(r)return r;let i=await this.download(t);return this.validate(i),await this.cache.put(t,i),i}async isCached(){let t=await this.resolveModelUrl();return this.cache.has(t)}async clearCache(){let t=await this.resolveModelUrl();await this.cache.remove(t)}async getCacheInfo(){let t=await this.resolveModelUrl();return this.cache.info(t)}async resolveModelUrl(){if(this.modelId.startsWith("http://")||this.modelId.startsWith("https://"))return this.modelId;if(!this.serverUrl)throw new X("MODEL_NOT_FOUND",`Cannot resolve model "${this.modelId}": no serverUrl configured.`);return this.fetchRegistryUrl(this.modelId)}async fetchRegistryUrl(t){let r=`${this.serverUrl}/api/v1/models/${encodeURIComponent(t)}/metadata`,i={Accept:"application/json"};this.apiKey&&(i.Authorization=`Bearer ${this.apiKey}`);let n;try{n=await fetch(r,{headers:i})}catch(s){throw new X("NETWORK_ERROR",`Failed to reach model registry at ${r}`,s)}if(!n.ok)throw n.status===404?new X("MODEL_NOT_FOUND",`Model "${t}" not found in registry.`):new X("NETWORK_ERROR",`Registry returned HTTP ${n.status}: ${n.statusText}`);return(await n.json()).url}buildHeaders(t={}){let r={"Accept-Encoding":"gzip, deflate, br",...t};return this.apiKey&&(r.Authorization=`Bearer ${this.apiKey}`),r}async download(t){let r=[],i=0,n=0,a=0;for(;a<=Lt;){let l=this.buildHeaders(i>0?{Range:`bytes=${i}-`}:{}),d;try{d=await fetch(t,{headers:l})}catch(g){if(a++,a>Lt)throw new X("NETWORK_ERROR",`Failed to download model from ${t} after ${Lt} retries`,g);await this.delay(Aa*a);continue}if(d.status===416){if(r.length=0,i=0,a++,a>Lt)throw new X("MODEL_LOAD_FAILED",`Model download failed: range request rejected after ${Lt} retries`);await this.delay(Aa*a);continue}if(!d.ok&&d.status!==206)throw new X("MODEL_LOAD_FAILED",`Model download failed: HTTP ${d.status} ${d.statusText}`);if(n===0){let g=d.headers.get("Content-Range");if(g){let w=g.match(/\/(\d+)$/);w&&(n=parseInt(w[1],10))}else n=parseInt(d.headers.get("Content-Length")??"0",10)}if(!d.body)return await d.arrayBuffer();let f=d.body.getReader(),h=!1;try{for(;;){let{done:g,value:w}=await f.read();if(g)break;r.push(w),i+=w.byteLength,this.onProgress?.({loaded:i,total:n,percent:n>0?i/n*100:NaN})}}catch(g){if(h=!0,a++,a>Lt)throw new X("NETWORK_ERROR",`Model download interrupted after ${Lt} retries`,g);await this.delay(Aa*a)}finally{f.releaseLock()}if(!h)break}let s=new Uint8Array(i),u=0;for(let l of r)s.set(l,u),u+=l.byteLength;return s.buffer}delay(t){return new Promise(r=>setTimeout(r,t))}validate(t){if(t.byteLength===0)throw new X("MODEL_LOAD_FAILED","Downloaded model is empty (0 bytes).");if(new Uint8Array(t,0,Math.min(4,t.byteLength))[0]!==8)throw new X("MODEL_LOAD_FAILED","Downloaded file does not appear to be a valid ONNX model.")}};var j0="https://api.octomil.io/v1/telemetry",Vt=class{constructor(t={}){Q(this,"url");Q(this,"flushIntervalMs");Q(this,"maxBatchSize");Q(this,"apiKey");Q(this,"queue",[]);Q(this,"timerId",null);Q(this,"disposed",!1);this.url=t.url??j0,this.flushIntervalMs=t.flushIntervalMs??3e4,this.maxBatchSize=t.maxBatchSize??50,this.apiKey=t.apiKey,this.startAutoFlush()}track(t){this.disposed||(this.queue.push(t),this.queue.length>=this.maxBatchSize&&this.flush())}async flush(){if(this.queue.length===0)return;let t=this.queue.splice(0,this.maxBatchSize);await this.send(t)}dispose(){this.disposed||(this.disposed=!0,this.timerId!==null&&(clearInterval(this.timerId),this.timerId=null),this.queue.length>0&&this.sendBeacon(this.queue.splice(0)))}startAutoFlush(){typeof setInterval>"u"||(this.timerId=setInterval(()=>{this.flush()},this.flushIntervalMs))}async send(t){let r=JSON.stringify({events:t});try{if(this.sendBeacon(t))return;let i={"Content-Type":"application/json"};this.apiKey&&(i.Authorization=`Bearer ${this.apiKey}`),await fetch(this.url,{method:"POST",headers:i,body:r,keepalive:!0})}catch{}}sendBeacon(t){if(typeof navigator>"u"||!navigator.sendBeacon)return!1;try{let r=new Blob([JSON.stringify({events:t})],{type:"application/json"});return navigator.sendBeacon(this.url,r)}catch{return!1}}},qt=null;function Zf(e={}){return qt&&qt.dispose(),qt=new Vt(e),qt}function Yf(){return qt}function Xf(){qt?.dispose(),qt=null}var Gt=class{constructor(t){Q(this,"serverUrl");Q(this,"apiKey");Q(this,"onTelemetry");this.serverUrl=t.serverUrl,this.apiKey=t.apiKey,this.onTelemetry=t.onTelemetry}async*stream(t,r,i={}){let n=new AbortController,a=i.signal?this.combineSignals(i.signal,n.signal):n.signal,s=`${this.serverUrl}/api/v1/models/${encodeURIComponent(t)}/stream`,u={"Content-Type":"application/json",Accept:"text/event-stream"};this.apiKey&&(u.Authorization=`Bearer ${this.apiKey}`);let l=performance.now(),d=null,f=0,h=0;this.onTelemetry?.({type:"streaming_start",model:t,metadata:{modality:i.modality??"text"},timestamp:Date.now()});let g;try{g=await fetch(s,{method:"POST",headers:u,body:JSON.stringify({input:r,modality:i.modality??"text",...i.params}),signal:a})}catch($){throw this.onTelemetry?.({type:"streaming_error",model:t,metadata:{error:String($)},timestamp:Date.now()}),new X("NETWORK_ERROR",`Streaming request failed: ${String($)}`,$)}if(!g.ok)throw new X("INFERENCE_FAILED",`Streaming inference failed: HTTP ${g.status}`);if(!g.body)throw new X("INFERENCE_FAILED","Server did not return a streaming body.");let w=g.body.getReader(),y=new TextDecoder,v="";try{for(;;){let{done:$,value:_}=await w.read();if($)break;v+=y.decode(_,{stream:!0});let I=v.split(`
+`);v=I.pop()??"";for(let k of I){if(!k.startsWith("data: "))continue;let E=k.slice(6).trim();if(E==="[DONE]")break;let C;try{C=JSON.parse(E)}catch{continue}f++,h+=E.length,d===null&&(d=performance.now()-l,this.onTelemetry?.({type:"streaming_chunk",model:t,durationMs:d,metadata:{chunkIndex:0,ttfc:!0},timestamp:Date.now()})),yield C}}}finally{w.releaseLock()}let T=performance.now()-l;return this.onTelemetry?.({type:"streaming_complete",model:t,durationMs:T,metadata:{chunkCount:f,totalBytes:h,ttfcMs:d},timestamp:Date.now()}),{totalChunks:f,totalBytes:h,durationMs:T,ttfcMs:d??T}}combineSignals(...t){let r=new AbortController;for(let i of t){if(i.aborted)return r.abort(i.reason),r.signal;i.addEventListener("abort",()=>r.abort(i.reason),{once:!0})}return r.signal}};var ci=class{constructor(t){Q(this,"options");Q(this,"cache");Q(this,"loader");Q(this,"engine");Q(this,"telemetry",null);Q(this,"loaded",!1);Q(this,"disposed",!1);this.options={telemetry:!1,cacheStrategy:"cache-api",...t},this.cache=qr(this.options.cacheStrategy),this.loader=new Jt(this.options,this.cache),this.engine=new Xt,this.options.telemetry&&(this.telemetry=new Vt({url:this.options.telemetryUrl,apiKey:this.options.apiKey}))}async load(){this.ensureNotDisposed();let t=performance.now(),r=await this.loader.isCached(),i=await this.loader.load();await this.engine.createSession(i,this.options.backend),this.loaded=!0;let n=performance.now()-t;this.trackEvent({type:"model_load",model:this.options.model,durationMs:n,metadata:{backend:this.engine.activeBackend,cached:r,sizeBytes:i.byteLength},timestamp:Date.now()}),r?this.trackEvent({type:"cache_hit",model:this.options.model,timestamp:Date.now()}):this.trackEvent({type:"cache_miss",model:this.options.model,timestamp:Date.now()})}async predict(t){this.ensureReady();let r=this.prepareTensors(t),i=await this.engine.run(r);return this.trackEvent({type:"inference",model:this.options.model,durationMs:i.latencyMs,metadata:{backend:this.engine.activeBackend},timestamp:Date.now()}),i}async predictBatch(t){this.ensureReady();let r=performance.now(),i=[];for(let a of t){let s=this.prepareTensors(a),u=await this.engine.run(s);i.push(u)}let n=performance.now()-r;return this.trackEvent({type:"inference",model:this.options.model,durationMs:n,metadata:{backend:this.engine.activeBackend,batchSize:t.length},timestamp:Date.now()}),i}async chat(t,r={}){if(this.ensureReady(),!this.options.serverUrl)throw new X("INFERENCE_FAILED","chat() requires serverUrl to be configured.");let i=new Gt({serverUrl:this.options.serverUrl,apiKey:this.options.apiKey,onTelemetry:u=>this.trackEvent(u)}),n=performance.now(),a="",s=i.stream(this.options.model,{messages:t,temperature:r.temperature,max_tokens:r.maxTokens,top_p:r.topP},{modality:"text",signal:r.signal});for await(let u of s)typeof u.data=="string"&&(a+=u.data);return{message:{role:"assistant",content:a},latencyMs:performance.now()-n}}async*chatStream(t,r={}){if(this.ensureReady(),!this.options.serverUrl)throw new X("INFERENCE_FAILED","chatStream() requires serverUrl to be configured.");let n=new Gt({serverUrl:this.options.serverUrl,apiKey:this.options.apiKey,onTelemetry:a=>this.trackEvent(a)}).stream(this.options.model,{messages:t,temperature:r.temperature,max_tokens:r.maxTokens,top_p:r.topP},{modality:"text",signal:r.signal});for await(let a of n)yield{index:a.index,content:typeof a.data=="string"?a.data:JSON.stringify(a.data),done:a.done,role:"assistant"}}async isCached(){return this.ensureNotDisposed(),this.loader.isCached()}async clearCache(){return this.ensureNotDisposed(),this.loader.clearCache()}async cacheInfo(){return this.ensureNotDisposed(),this.loader.getCacheInfo()}get activeBackend(){return this.engine.activeBackend}get inputNames(){return this.ensureReady(),this.engine.inputNames}get outputNames(){return this.ensureReady(),this.engine.outputNames}get isLoaded(){return this.loaded}dispose(){this.disposed||(this.disposed=!0,this.loaded=!1,this.engine.dispose(),this.telemetry?.dispose(),this.telemetry=null)}ensureNotDisposed(){if(this.disposed)throw new X("SESSION_DISPOSED","This Octomil instance has been disposed. Create a new one.")}ensureReady(){if(this.ensureNotDisposed(),!this.loaded)throw new X("NOT_LOADED","Model not loaded. Call load() before predict() or chat().")}prepareTensors(t){if(this.isNamedTensors(t))return t;if("raw"in t&&"dims"in t){let r=this.engine.inputNames[0];if(!r)throw new X("INVALID_INPUT","Model has no input tensors defined.");return{[r]:{data:t.raw,dims:t.dims}}}if("text"in t){let r=this.engine.inputNames[0];if(!r)throw new X("INVALID_INPUT","Model has no input tensors defined.");let i=new Int32Array(Array.from(t.text).map(n=>n.codePointAt(0)??0));return{[r]:{data:i,dims:[1,i.length]}}}if("image"in t)return this.imageToTensors(t.image);throw new X("INVALID_INPUT","Unrecognised PredictInput format. Provide named tensors, { text }, { image }, or { raw, dims }.")}isNamedTensors(t){if("text"in t||"image"in t||"raw"in t)return!1;let r=Object.values(t)[0];return r!==void 0&&typeof r=="object"&&"data"in r&&"dims"in r}imageToTensors(t){let r;if(t instanceof ImageData)r=t;else{let d=t instanceof HTMLCanvasElement?t:(()=>{let h=document.createElement("canvas");return h.width=t.naturalWidth||t.width,h.height=t.naturalHeight||t.height,h.getContext("2d").drawImage(t,0,0),h})(),f=d.getContext("2d");if(!f)throw new X("INVALID_INPUT","Could not get 2D context from canvas.");r=f.getImageData(0,0,d.width,d.height)}let{width:i,height:n,data:a}=r,s=i*n,u=new Float32Array(3*s);for(let d=0;d<s;d++)u[d]=a[d*4]/255,u[s+d]=a[d*4+1]/255,u[2*s+d]=a[d*4+2]/255;let l=this.engine.inputNames[0];if(!l)throw new X("INVALID_INPUT","Model has no input tensors defined.");return{[l]:{data:u,dims:[1,3,n,i]}}}trackEvent(t){this.telemetry?.track(t)}};var Jf=3e4,hi=class{constructor(t){Q(this,"serverUrl");Q(this,"apiKey");Q(this,"token",null);Q(this,"deviceId",null);Q(this,"refreshTimer",null);Q(this,"disposed",!1);this.serverUrl=t.serverUrl,this.apiKey=t.apiKey}async bootstrap(t){this.ensureNotDisposed(),this.deviceId=await this.generateDeviceId();let r=this.collectDeviceInfo(),i=await this.request("/api/v1/devices/register",{method:"POST",body:JSON.stringify({org_id:t,device_id:this.deviceId,platform:"browser",info:r})});if(!i.ok)throw new X("NETWORK_ERROR",`Device registration failed: HTTP ${i.status}`);let n=await i.json();this.token={accessToken:n.token,refreshToken:n.refresh_token,expiresAt:new Date(n.expires_at).getTime()},this.scheduleRefresh()}async getToken(){if(this.ensureNotDisposed(),!this.token)throw new X("NETWORK_ERROR","Not authenticated. Call bootstrap() first.");return this.isTokenExpiringSoon()&&await this.refreshToken(),this.token.accessToken}async refreshToken(){if(this.ensureNotDisposed(),!this.token)throw new X("NETWORK_ERROR","No token to refresh. Call bootstrap() first.");let t=await this.request("/api/v1/auth/refresh",{method:"POST",body:JSON.stringify({refresh_token:this.token.refreshToken,device_id:this.deviceId})});if(!t.ok)throw this.token=null,new X("NETWORK_ERROR",`Token refresh failed: HTTP ${t.status}`);let r=await t.json();this.token={accessToken:r.token,refreshToken:r.refresh_token,expiresAt:new Date(r.expires_at).getTime()},this.scheduleRefresh()}async revokeToken(){if(this.ensureNotDisposed(),!!this.token)try{await this.request("/api/v1/auth/revoke",{method:"POST",body:JSON.stringify({token:this.token.accessToken,device_id:this.deviceId})})}finally{this.clearState()}}get isAuthenticated(){return this.token!==null&&!this.isTokenExpired()}get currentDeviceId(){return this.deviceId}dispose(){this.disposed||(this.disposed=!0,this.clearRefreshTimer())}async request(t,r){let i={"Content-Type":"application/json",Authorization:`Bearer ${this.apiKey}`};return fetch(`${this.serverUrl}${t}`,{...r,headers:{...i,...r.headers}})}isTokenExpired(){return this.token?Date.now()>=this.token.expiresAt:!0}isTokenExpiringSoon(){return this.token?Date.now()>=this.token.expiresAt-Jf:!0}scheduleRefresh(){if(this.clearRefreshTimer(),!this.token)return;let t=Math.max(0,this.token.expiresAt-Date.now()-Jf);this.refreshTimer=setTimeout(()=>{this.refreshToken().catch(()=>{})},t)}clearRefreshTimer(){this.refreshTimer!==null&&(clearTimeout(this.refreshTimer),this.refreshTimer=null)}clearState(){this.token=null,this.clearRefreshTimer()}async generateDeviceId(){let t=[typeof navigator<"u"?navigator.userAgent:"unknown",typeof screen<"u"?`${screen.width}x${screen.height}`:"0x0",typeof Intl<"u"?Intl.DateTimeFormat().resolvedOptions().timeZone:"UTC",typeof navigator<"u"?navigator.language:"en"].join("|"),r=new TextEncoder().encode(t),i=await crypto.subtle.digest("SHA-256",r),n=new Uint8Array(i);return Array.from(n).map(a=>a.toString(16).padStart(2,"0")).join("")}collectDeviceInfo(){return{userAgent:typeof navigator<"u"?navigator.userAgent:"unknown",language:typeof navigator<"u"?navigator.language:"en",screenWidth:typeof screen<"u"?screen.width:0,screenHeight:typeof screen<"u"?screen.height:0,timezone:typeof Intl<"u"?Intl.DateTimeFormat().resolvedOptions().timeZone:"UTC",webgpu:typeof navigator<"u"&&"gpu"in navigator}}ensureNotDisposed(){if(this.disposed)throw new X("SESSION_DISPOSED","DeviceAuthManager has been disposed.")}};async function Oa(e){let t=await crypto.subtle.digest("SHA-256",e),r=new Uint8Array(t);return Array.from(r).map(i=>i.toString(16).padStart(2,"0")).join("")}async function Ra(e,t){return await Oa(e)===t.toLowerCase()}async function em(e,t){if(!await Ra(e,t))throw new X("MODEL_LOAD_FAILED","Model integrity check failed: SHA-256 hash mismatch. The downloaded model may be corrupted or tampered with.")}var er=class{static computeDelta(t,r){let i={};for(let n of Object.keys(t)){let a=t[n],s=r[n];if(!a||!s||a.length!==s.length)throw new X("INVALID_INPUT",`Weight dimension mismatch for "${n}".`);let u=new Float32Array(a.length);for(let l=0;l<a.length;l++)u[l]=s[l]-a[l];i[n]=u}return i}static applyDelta(t,r){let i={};for(let n of Object.keys(t)){let a=t[n],s=r[n];if(!a)continue;if(!s||a.length!==s.length){i[n]=new Float32Array(a);continue}let u=new Float32Array(a.length);for(let l=0;l<a.length;l++)u[l]=a[l]+s[l];i[n]=u}return i}static l2Norm(t){let r=0;for(let i of Object.values(t))if(i)for(let n=0;n<i.length;n++)r+=i[n]*i[n];return Math.sqrt(r)}},fi=class{constructor(t){Q(this,"serverUrl");Q(this,"apiKey");Q(this,"deviceId");Q(this,"onTelemetry");this.serverUrl=t.serverUrl,this.apiKey=t.apiKey,this.deviceId=t.deviceId,this.onTelemetry=t.onTelemetry}async getTrainingRound(t){let r=await this.request(`/api/v1/federations/${encodeURIComponent(t)}/rounds/current`);if(!r.ok)throw new X("NETWORK_ERROR",`Failed to fetch training round: HTTP ${r.status}`);return await r.json()}async joinRound(t,r){let i=await this.request(`/api/v1/federations/${encodeURIComponent(t)}/rounds/${encodeURIComponent(r)}/join`,{method:"POST",body:JSON.stringify({device_id:this.deviceId})});if(!i.ok)throw new X("NETWORK_ERROR",`Failed to join round: HTTP ${i.status}`)}async train(t,r){let i=performance.now(),n=this.cloneWeights(t);for(let u=0;u<r.epochs;u++)n=(await r.onTrainStep(n,{epoch:u,batchSize:r.batchSize,learningRate:r.learningRate})).weights;let a=er.computeDelta(t,n),s=performance.now()-i;return this.onTelemetry?.({type:"training_complete",model:r.modelId??"unknown",durationMs:s,metadata:{epochs:r.epochs,deltaNorm:er.l2Norm(a)},timestamp:Date.now()}),{finalWeights:n,delta:a}}async submitUpdate(t,r,i,n){let a={};for(let[u,l]of Object.entries(i))l&&(a[u]={data:Array.from(l),shape:[l.length]});let s=await this.request(`/api/v1/federations/${encodeURIComponent(t)}/rounds/${encodeURIComponent(r)}/submit`,{method:"POST",body:JSON.stringify({device_id:this.deviceId,delta:a,metrics:n})});if(!s.ok)throw new X("NETWORK_ERROR",`Failed to submit update: HTTP ${s.status}`)}async request(t,r){let i={"Content-Type":"application/json"};return this.apiKey&&(i.Authorization=`Bearer ${this.apiKey}`),fetch(`${this.serverUrl}${t}`,{...r,headers:{...i,...r?.headers}})}cloneWeights(t){let r={};for(let[i,n]of Object.entries(t))n&&(r[i]=new Float32Array(n));return r}};var Sr=class{constructor(){Q(this,"keyPair",null)}async generateKeyPair(){return this.keyPair=await crypto.subtle.generateKey({name:"ECDH",namedCurve:"P-256"},!0,["deriveBits"]),{publicKey:await crypto.subtle.exportKey("jwk",this.keyPair.publicKey)}}async deriveSharedSecret(t){if(!this.keyPair)throw new Error("Call generateKeyPair() first.");let r=await crypto.subtle.importKey("jwk",t,{name:"ECDH",namedCurve:"P-256"},!1,[]);return crypto.subtle.deriveBits({name:"ECDH",public:r},this.keyPair.privateKey,256)}async createMask(t,r){let i=await crypto.subtle.importKey("raw",t,"HKDF",!1,["deriveBits"]),n=r*4*8,a=await crypto.subtle.deriveBits({name:"HKDF",hash:"SHA-256",salt:new Uint8Array(32),info:new TextEncoder().encode("octomil-secagg-mask")},i,Math.min(n,8160)),s=new Float32Array(r),u=new Float32Array(a);for(let l=0;l<r;l++)s[l]=u[l%u.length];return s}maskUpdate(t,r){let i={};for(let[n,a]of Object.entries(t)){if(!a)continue;let s=new Float32Array(a),u=r.get(n);if(u)for(let l=0;l<s.length;l++)s[l]=s[l]+(u[l]??0);i[n]=s}return i}unmask(t,r){let i={};for(let[n,a]of Object.entries(t)){if(!a)continue;let s=new Float32Array(a),u=r.get(n);if(u)for(let l=0;l<s.length;l++)s[l]=s[l]-(u[l]??0);i[n]=s}return i}},qe=2147483647;function tm(e,t,r){let i=1;for(e=e%r;t>0;)t%2===1&&(i=Number(BigInt(i)*BigInt(e)%BigInt(r))),t=Math.floor(t/2),e=Number(BigInt(e)*BigInt(e)%BigInt(r));return i}function Q0(e,t){return tm(e,t-2,t)}function Ma(e,t,r){let i=[e%qe];for(let a=1;a<t;a++){let s=new Uint32Array(1);crypto.getRandomValues(s),i.push(s[0]%qe)}let n=[];for(let a=1;a<=r;a++){let s=0;for(let u=0;u<i.length;u++)s=Number((BigInt(s)+BigInt(i[u])*BigInt(tm(a,u,qe)))%BigInt(qe));n.push({x:a,y:s})}return n}function Ba(e){let t=0,r=e.length;for(let i=0;i<r;i++){let n=1,a=1;for(let u=0;u<r;u++)i!==u&&(n=Number(BigInt(n)*BigInt(qe-e[u].x)%BigInt(qe)),a=Number(BigInt(a)*BigInt((e[i].x-e[u].x+qe)%qe)%BigInt(qe)));let s=Number(BigInt(n)*BigInt(Q0(a,qe))%BigInt(qe));t=Number((BigInt(t)+BigInt(e[i].y)*BigInt(s))%BigInt(qe))}return t}var mi=class extends Sr{constructor(r){super();Q(this,"threshold");this.threshold=r}splitSecret(r,i){return Ma(r,this.threshold,i)}reconstructSecret(r){if(r.length<this.threshold)throw new Error(`Need at least ${this.threshold} shares, got ${r.length}.`);return Ba(r.slice(0,this.threshold))}};function rm(e,t){let r=0;for(let s of Object.values(e))if(s)for(let u=0;u<s.length;u++)r+=s[u]*s[u];let i=Math.sqrt(r);if(i<=t)return e;let n=t/i,a={};for(let[s,u]of Object.entries(e)){if(!u)continue;let l=new Float32Array(u.length);for(let d=0;d<u.length;d++)l[d]=u[d]*n;a[s]=l}return a}function im(e,t,r,i){let n=r*Math.sqrt(2*Math.log(1.25/i))/t,a={};for(let[s,u]of Object.entries(e)){if(!u)continue;let l=new Float32Array(u.length);for(let d=0;d<u.length;d++)l[d]=u[d]+Z0()*n;a[s]=l}return a}function Z0(){let e,t;do e=Math.random(),t=Math.random();while(e===0);return Math.sqrt(-2*Math.log(e))*Math.cos(2*Math.PI*t)}function nm(e,t=8){let r=t===8?127:32767,i={};for(let[n,a]of Object.entries(e)){if(!a)continue;let s=0;for(let d=0;d<a.length;d++){let f=Math.abs(a[d]);f>s&&(s=f)}let u=s>0?s/r:1,l=t===8?new Int8Array(a.length):new Int16Array(a.length);for(let d=0;d<a.length;d++)l[d]=Math.round(a[d]/u);i[n]={data:l,scale:u,zeroPoint:0}}return i}function am(e){let t={};for(let[r,i]of Object.entries(e)){if(!i)continue;let n=new Float32Array(i.data.length);for(let a=0;a<i.data.length;a++)n[a]=(i.data[a]-i.zeroPoint)*i.scale;t[r]=n}return t}var gi=class{constructor(t){Q(this,"serverUrl");Q(this,"apiKey");Q(this,"cacheTtlMs");Q(this,"onTelemetry");Q(this,"configCache",new Map);this.serverUrl=t.serverUrl,this.apiKey=t.apiKey,this.cacheTtlMs=t.cacheTtlMs??5*60*1e3,this.onTelemetry=t.onTelemetry}async resolveVersion(t,r){let i=await this.getRolloutConfig(t),n=i.versions.find(s=>s.status==="canary");if(n&&this.isInCanaryGroup(t,r,n.percentage))return n.version;let a=i.versions.find(s=>s.status==="active");if(a)return a.version;throw new X("MODEL_NOT_FOUND",`No active version found for model "${t}".`)}async getRolloutConfig(t){let r=this.configCache.get(t);if(r&&Date.now()-r.fetchedAt<this.cacheTtlMs)return r.config;let i=`${this.serverUrl}/api/v1/models/${encodeURIComponent(t)}/rollout`,n={Accept:"application/json"};this.apiKey&&(n.Authorization=`Bearer ${this.apiKey}`);let a=await fetch(i,{headers:n});if(!a.ok)throw new X("NETWORK_ERROR",`Failed to fetch rollout config: HTTP ${a.status}`);let s=await a.json();return this.configCache.set(t,{config:s,fetchedAt:Date.now()}),s}isInCanaryGroup(t,r,i){return Y0(r+t)<i}async getAvailableVersions(t){return(await this.getRolloutConfig(t)).versions}async reportRolloutStatus(t,r,i){let n=`${this.serverUrl}/api/v1/models/${encodeURIComponent(t)}/rollout/status`,a={"Content-Type":"application/json"};this.apiKey&&(a.Authorization=`Bearer ${this.apiKey}`),await fetch(n,{method:"POST",headers:a,body:JSON.stringify({version:r,status:i})}),this.onTelemetry?.({type:"rollout_status",model:t,metadata:{version:r,status:i},timestamp:Date.now()})}clearCache(){this.configCache.clear()}};function Y0(e){let t=5381;for(let r=0;r<e.length;r++)t=(t<<5)+t+e.charCodeAt(r)|0;return Math.abs(t)%100}var yi=class{constructor(t){Q(this,"serverUrl");Q(this,"apiKey");Q(this,"cacheTtlMs");Q(this,"onTelemetry");Q(this,"experimentsCache",null);this.serverUrl=t.serverUrl,this.apiKey=t.apiKey,this.cacheTtlMs=t.cacheTtlMs??5*60*1e3,this.onTelemetry=t.onTelemetry}async getActiveExperiments(){if(this.experimentsCache&&Date.now()-this.experimentsCache.fetchedAt<this.cacheTtlMs)return this.experimentsCache.experiments;let t=`${this.serverUrl}/api/v1/experiments?status=active`,r={Accept:"application/json"};this.apiKey&&(r.Authorization=`Bearer ${this.apiKey}`);let i=await fetch(t,{headers:r});if(!i.ok)throw new X("NETWORK_ERROR",`Failed to fetch experiments: HTTP ${i.status}`);let n=await i.json();return this.experimentsCache={experiments:n.experiments,fetchedAt:Date.now()},n.experiments}async getExperimentConfig(t){let r=`${this.serverUrl}/api/v1/experiments/${encodeURIComponent(t)}`,i={Accept:"application/json"};this.apiKey&&(i.Authorization=`Bearer ${this.apiKey}`);let n=await fetch(r,{headers:i});if(!n.ok)throw new X("NETWORK_ERROR",`Failed to fetch experiment: HTTP ${n.status}`);return await n.json()}getVariant(t,r){if(t.variants.length===0)return null;let i=X0(r+t.id),n=0;for(let a of t.variants)if(n+=a.trafficPercentage,i<n)return a;return t.variants[t.variants.length-1]??null}isEnrolled(t,r){return this.getVariant(t,r)!==null}async resolveModelExperiment(t,r){let i=await this.getActiveExperiments();for(let n of i){if(!n.variants.some(u=>u.modelId===t))continue;let s=this.getVariant(n,r);if(s)return{experiment:n,variant:s}}return null}async trackMetric(t,r,i,n){let a=`${this.serverUrl}/api/v1/experiments/${encodeURIComponent(t)}/metrics`,s={"Content-Type":"application/json"};this.apiKey&&(s.Authorization=`Bearer ${this.apiKey}`),await fetch(a,{method:"POST",headers:s,body:JSON.stringify({metric_name:r,value:i,device_id:n,timestamp:Date.now()})}),this.onTelemetry?.({type:"experiment_metric",model:t,metadata:{metricName:r,value:i},timestamp:Date.now()})}clearCache(){this.experimentsCache=null}};function X0(e){let t=5381;for(let r=0;r<e.length;r++)t=(t<<5)+t+e.charCodeAt(r)|0;return Math.abs(t)%100}return xy(J0);})();
+/*! Bundled license information:
+
+onnxruntime-web/dist/ort.bundle.min.mjs:
+  (*!
+   * ONNX Runtime Web v1.24.2
+   * Copyright (c) Microsoft Corporation. All rights reserved.
+   * Licensed under the MIT License.
+   *)
+  (**
+   * @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 Octomil=__Octomil_ns.Octomil;Object.assign(Octomil,__Octomil_ns);
+//# sourceMappingURL=octomil.min.js.map