@slexisvn/mlfw 0.1.2 → 0.1.3

package/dist/index.browser.js

@@ -9,7 +9,7 @@ var SS=Object.defineProperty;var l=(r,e)=>SS(r,"name",{value:e,configurable:!0})
 `,t),t<0&&(t=r.length);continue}if(n==="("&&r[t+1]===";"){t=r.indexOf(";)",t+2),t=t<0?r.length:t+2;continue}if(n==="("||n===")"){e.push(n),t++;continue}if(n==='"'){let o=t+1;for(;o<r.length&&r[o]!=='"';)o++;e.push(r.substring(t,o+1)),t=o+1;continue}let s=t;for(;s<r.length&&r[s]>" "&&r[s]!=="("&&r[s]!==")";)s++;e.push(r.substring(t,s)),t=s}return e}function g2(r){let e=0,t=l(()=>r[e++],"eat"),n=l(()=>r[e],"peek"),s=l(y=>{if(t()!==y)throw new Error("expected "+y+" at "+(e-1))},"expect");s("("),s("module");let o=[],i=1,a=256,u="",p=[],c=[],h=[],f=[],d=-1,m=-1;for(;n()!==")";){if(n()!=="("){e++;continue}e++;let y=t();if(y==="memory"){for(;n()==="(";)g();n()!==")"&&(i=parseInt(t())||1),n()!==")"&&(a=parseInt(t())||256),s(")")}else if(y==="import"){let _=t().replace(/"/g,""),b=t().replace(/"/g,"");s("("),s("func"),n().startsWith("$")&&e++;let w=[],T=[];for(;n()==="(";){e++;let S=t();if(S==="param")for(;n()!==")";){let x=t();x==="f32"?w.push(Rc):x==="f64"?w.push(Bc):x==="i32"?w.push(Oc):x==="v128"&&w.push(Mc)}else if(S==="result")for(;n()!==")";){let x=t();x==="f32"?T.push(Rc):x==="f64"?T.push(Bc):x==="i32"?T.push(Oc):x==="v128"&&T.push(Mc)}s(")")}s(")"),s(")"),o.push({module:_,name:b,params:w,results:T})}else if(y==="func"){for(;n()==="(";){let b=e;e++;let w=t();if(w==="export")u=t().replace(/"/g,""),s(")");else if(w==="param"){let T="";for(;n()!==")";){let S=t();if(S.startsWith("$")){T=S.replace("$","");continue}let x=null;S==="i32"?x=Oc:S==="i64"?x=nw:S==="f32"?x=Rc:S==="f64"?x=Bc:S==="v128"&&(x=Mc),x!==null&&(p.push(x),c.push(T),T="")}s(")")}else if(w==="result"){for(;n()!==")";)t();s(")")}else if(w==="local"){for(;n()!==")";){let T=t();T.startsWith("$")?f.push(T.replace("$","")):T==="i32"?h.push(Oc):T==="i64"?h.push(nw):T==="f32"?h.push(Rc):T==="f64"?h.push(Bc):T==="v128"&&h.push(Mc)}s(")")}else{e=b;break}}d=e;let _=1;for(;_>0;){let b=t();b==="("?_++:b===")"&&_--}m=e-1}else g()}function g(){if(n()!=="("){e++;return}let y=0;do{let _=t();_==="("?y++:_===")"&&y--}while(y>0)}return l(g,"skipSExpr"),{imports:o,memMin:i,memMax:a,funcExportName:u,funcParams:p,funcParamNames:c,funcLocals:h,funcLocalNames:f,bodyTokens:r.slice(d,m)}}function y2(r,e,t){let n=[],s=0,o=l(()=>r[s],"peek"),i=l(()=>r[s++],"eat"),a=[],u=-1;for(let y of e.values())y>u&&(u=y);function p(y){let _=y.replace("$","");if(e.has(_))return e.get(_);let b=parseInt(_,10);if(!isNaN(b))return b>u&&(u=b),b;let w=u+1;return u=w,e.set(_,w),w}l(p,"localIdx");function c(){return o()&&o().startsWith("$")?i().replace("$",""):""}l(c,"eatLabel");function h(y,_){let b=y.indexOf(_);return b>=0?y.substring(b+_.length):""}l(h,"extractName");function f(){for(;s<r.length&&o()!==")";){let y=o();if(y==="("){s++;let _=i();if(_==="i32.const")n.push(65),n.push(...p2(parseInt(i()))),m(")");else if(_==="i64.const")n.push(66),n.push(...f2(i())),m(")");else if(_==="f32.const")n.push(67),n.push(...h2(parseFloat(i()))),m(")");else if(_==="f64.const")n.push(68),n.push(...d2(parseFloat(i()))),m(")");else if(_==="local.get")n.push(32),n.push(...re(p(i()))),m(")");else if(_==="local.set")n.push(33),n.push(...re(p(i()))),m(")");else if(_==="block"){let b=i().replace("$","");a.push(b),n.push(2,ew),f(),n.push(11),a.pop(),m(")")}else if(_==="loop"){let b=i().replace("$","");a.push(b),n.push(3,ew),f(),n.push(11),a.pop(),m(")")}else if(_==="if"){let b=ew;if(o()==="("){let w=s;if(s++,o()==="result"){i();let T=i();b=T==="f32"?Rc:T==="f64"?Bc:T==="v128"?Mc:T==="i64"?nw:Oc,m(")")}else s=w}for(n.push(4,b),a.push("_if");s<r.length&&o()!==")"&&o()==="(";){let w=s;s++;let T=i();if(T==="then")f(),m(")");else if(T==="else")n.push(5),f(),m(")");else{s=w;break}}n.push(11),a.pop(),m(")")}else s-=2,g();continue}if(s++,y==="br_if")n.push(13),n.push(...re(d(c())));else if(y==="br")n.push(12),n.push(...re(d(c())));else if(y.startsWith("call")){n.push(16);let _=h(y,"math_")||(o()&&o().startsWith("$")?h(i(),"math_"):""),b=t.get(_);if(b===void 0)throw new Error(`WAT encoder: call to unresolved import '${_}'`);n.push(...re(b))}else y==="local.get"?(n.push(32),n.push(...re(p(c())))):y==="local.set"?(n.push(33),n.push(...re(p(c())))):y.endsWith(".extract_lane")||y.endsWith(".replace_lane")?(n.push(...tw.get(y)),n.push(parseInt(i(),10))):tw.has(y)&&n.push(...tw.get(y))}}l(f,"emitBlock");function d(y){for(let _=a.length-1;_>=0;_--)if(a[_]===y)return a.length-1-_;return 0}l(d,"resolveBr");function m(y){if(i()!==y)throw new Error("expect "+y)}l(m,"expect");function g(){if(o()!=="("){s++;return}let y=0;do{let _=i();_==="("?y++:_===")"&&y--}while(y>0)}return l(g,"skipSExpr"),f(),n}function Nu(r){let e=m2(r),t=g2(e),n=new Map,s=t.funcParams.length;for(let x=0;x<t.funcParamNames.length;x++)t.funcParamNames[x]&&n.set(t.funcParamNames[x],x);for(let x=0;x<t.funcLocalNames.length;x++)n.set(t.funcLocalNames[x],s+x);let o=new Map;for(let x=0;x<t.imports.length;x++)o.set(t.imports[x].name,x);let i=[],a=new Map;function u(x,k){let A=x.join(",")+">"+k.join(",");if(a.has(A))return a.get(A);let N=i.length;return i.push({params:x,results:k}),a.set(A,N),N}l(u,"getType");for(let x of t.imports)x.typeIdx=u(x.params,x.results);let p=u(t.funcParams,[]),c=vu(s2,Zi(i.map(x=>[c2,...re(x.params.length),...x.params,...re(x.results.length),...x.results]))),h=[];t.imports.length>0&&(h=vu(o2,Zi(t.imports.map(x=>[...Sd(x.module),...Sd(x.name),0,...re(x.typeIdx)]))));let f=vu(i2,Zi([[...re(p)]])),d=vu(a2,Zi([[1,...re(t.memMin),...re(t.memMax)]])),m=t.imports.length,g=vu(u2,Zi([[...Sd("memory"),2,...re(0)],[...Sd(t.funcExportName),0,...re(m)]])),y=[];if(t.funcLocals.length>0){let x=[],k=t.funcLocals[0],A=1;for(let N=1;N<t.funcLocals.length;N++)t.funcLocals[N]===k?A++:(x.push([...re(A),k]),k=t.funcLocals[N],A=1);x.push([...re(A),k]),Ws(y,Zi(x))}else y.push(0);let _=y2(t.bodyTokens,n,o),b=[];Ws(b,y),Ws(b,_),b.push(11);let w=[];Ws(w,re(b.length)),Ws(w,b);let T=vu(l2,Zi([w])),S=[];for(let x of[n2,r2,c,h,f,d,g,T])Ws(S,x);return new Uint8Array(S)}var n2,r2,s2,o2,i2,a2,u2,l2,Oc,nw,Bc,Rc,Mc,c2,ew,tw,kd=M(()=>{"use strict";n2=[0,97,115,109],r2=[1,0,0,0],s2=1,o2=2,i2=3,a2=5,u2=7,l2=10,Oc=127,nw=126,Bc=124,Rc=125,Mc=123,c2=96,ew=64;l(re,"uleb");l(p2,"sleb");l(f2,"slebBig");l(Sd,"encStr");l(h2,"encF32");l(d2,"encF64");l(Ws,"pushAll");l(vu,"section");l(Zi,"vec");tw=new Map([["i32.const",[65]],["f32.const",[67]],["local.get",[32]],["local.set",[33]],["i32.add",[106]],["i32.sub",[107]],["i32.mul",[108]],["i32.div_s",[109]],["i32.rem_s",[111]],["i32.eq",[70]],["i32.ne",[71]],["i32.lt_s",[72]],["i32.gt_s",[74]],["i32.le_s",[76]],["i32.ge_s",[78]],["i32.eqz",[69]],["i32.lt_u",[73]],["i32.gt_u",[75]],["i32.le_u",[77]],["i32.ge_u",[79]],["i32.and",[113]],["i32.or",[114]],["i32.xor",[115]],["i32.shl",[116]],["i32.shr_s",[117]],["i32.shr_u",[118]],["i32.reinterpret_f32",[188]],["f32.reinterpret_i32",[190]],["i32.trunc_f32_s",[168]],["f32.add",[146]],["f32.sub",[147]],["f32.mul",[148]],["f32.div",[149]],["f32.neg",[140]],["f32.abs",[139]],["f32.ceil",[141]],["f32.floor",[142]],["f32.sqrt",[145]],["f32.min",[150]],["f32.max",[151]],["f32.eq",[91]],["f32.ne",[92]],["f32.lt",[93]],["f32.gt",[94]],["f32.le",[95]],["f32.ge",[96]],["f32.convert_i32_s",[178]],["f64.add",[160]],["f64.sub",[161]],["f64.mul",[162]],["f64.div",[163]],["f64.min",[164]],["f64.max",[165]],["f64.neg",[154]],["f64.abs",[153]],["f64.ceil",[155]],["f64.floor",[156]],["f64.sqrt",[159]],["f64.eq",[97]],["f64.ne",[98]],["f64.lt",[99]],["f64.gt",[100]],["f64.le",[101]],["f64.ge",[102]],["f64.convert_i32_s",[183]],["f64.promote_f32",[187]],["f32.demote_f64",[182]],["i32.trunc_f64_s",[170]],["i64.load",[41,3,0]],["i64.store",[55,3,0]],["i64.add",[124]],["i64.sub",[125]],["i64.mul",[126]],["i64.div_s",[127]],["i64.rem_s",[129]],["i64.and",[131]],["i64.or",[132]],["i64.xor",[133]],["i64.shl",[134]],["i64.shr_s",[135]],["i64.shr_u",[136]],["i64.eqz",[80]],["i64.eq",[81]],["i64.ne",[82]],["i64.lt_s",[83]],["i64.gt_s",[85]],["i64.le_s",[87]],["i64.ge_s",[89]],["i64.extend_i32_s",[172]],["i32.wrap_i64",[167]],["f64.convert_i64_s",[185]],["i64.trunc_f64_s",[176]],["f32.convert_i64_s",[180]],["i64.trunc_f32_s",[174]],["select",[27]],["f32.load",[42,2,0]],["f32.store",[56,2,0]],["f64.load",[43,3,0]],["f64.store",[57,3,0]],["i32.load",[40,2,0]],["i32.store",[54,2,0]],["i32.load8_s",[44,0,0]],["i32.load8_u",[45,0,0]],["i32.store8",[58,0,0]],["i32.load16_s",[46,1,0]],["i32.load16_u",[47,1,0]],["i32.store16",[59,1,0]],["v128.load",[253,...re(0),4,0]],["v128.store",[253,...re(11),4,0]],["v128.bitselect",[253,...re(82)]],["v128.and",[253,...re(78)]],["v128.or",[253,...re(80)]],["v128.not",[253,...re(77)]],["f32x4.splat",[253,...re(19)]],["i32x4.splat",[253,...re(17)]],["f32x4.add",[253,...re(228)]],["f32x4.sub",[253,...re(229)]],["f32x4.mul",[253,...re(230)]],["f32x4.div",[253,...re(231)]],["f32x4.neg",[253,...re(225)]],["f32x4.abs",[253,...re(224)]],["f32x4.sqrt",[253,...re(227)]],["f32x4.ceil",[253,...re(103)]],["f32x4.floor",[253,...re(104)]],["f32x4.min",[253,...re(232)]],["f32x4.max",[253,...re(233)]],["f32x4.eq",[253,...re(65)]],["f32x4.ne",[253,...re(66)]],["f32x4.lt",[253,...re(67)]],["f32x4.gt",[253,...re(68)]],["f32x4.le",[253,...re(69)]],["f32x4.ge",[253,...re(70)]],["f32x4.extract_lane",[253,...re(31)]],["f32x4.replace_lane",[253,...re(32)]],["i32x4.add",[253,...re(174)]],["i32x4.sub",[253,...re(177)]],["i32x4.mul",[253,...re(181)]],["i32x4.abs",[253,...re(160)]],["i32x4.min_s",[253,...re(182)]],["i32x4.max_s",[253,...re(184)]],["i32x4.eq",[253,...re(55)]],["i32x4.ne",[253,...re(56)]],["i32x4.lt_s",[253,...re(57)]],["i32x4.gt_s",[253,...re(58)]],["i32x4.le_s",[253,...re(59)]],["i32x4.ge_s",[253,...re(60)]],["i32x4.extract_lane",[253,...re(27)]],["i32x4.replace_lane",[253,...re(28)]],["f64x2.splat",[253,...re(20)]],["f64x2.add",[253,...re(240)]],["f64x2.sub",[253,...re(241)]],["f64x2.mul",[253,...re(242)]],["f64x2.div",[253,...re(243)]],["f64x2.neg",[253,...re(237)]],["f64x2.abs",[253,...re(236)]],["f64x2.sqrt",[253,...re(239)]],["f64x2.min",[253,...re(244)]],["f64x2.max",[253,...re(245)]],["f64x2.eq",[253,...re(71)]],["f64x2.ne",[253,...re(72)]],["f64x2.lt",[253,...re(73)]],["f64x2.gt",[253,...re(74)]],["f64x2.le",[253,...re(75)]],["f64x2.ge",[253,...re(76)]],["f64x2.extract_lane",[253,...re(33)]],["f64x2.replace_lane",[253,...re(34)]]]);l(m2,"tokenize");l(g2,"parseModule");l(y2,"encodeBody");l(Nu,"encodeWat")});function Lv(r){let{source:e,name:t,metadata:n}=r,{bufferOffsets:s,imports:o}=n,i=[...s.entries()],a=Nu(e),u=[];if(u.push("(async () => {"),u.push(""),u.push("const binary = new Uint8Array(["+a.join(",")+"]);"),u.push(""),o&&o.size>0){u.push("const mathImports = {");for(let[c]of o)c==="fmod"?u.push("  fmod: (a, b) => a % b,"):c==="rsqrt"?u.push("  rsqrt: x => 1 / Math.sqrt(x),"):u.push(`  ${c}: Math.${c},`);u.push("};"),u.push("const { instance } = await WebAssembly.instantiate(binary, { math: mathImports });")}else u.push("const { instance } = await WebAssembly.instantiate(binary);");u.push("const memory = instance.exports.memory;"),u.push("");for(let[c,h]of i)u.push(`// ${c}: new Float32Array(memory.buffer, ${h}, N).set(data);`);u.push(""),u.push(`instance.exports.${t}(${i.map(([,c])=>c).join(", ")});`),u.push("");let p=i[i.length-1];return p&&(u.push(`const result = new Float32Array(memory.buffer, ${p[1]}, N);`),u.push("console.log(result);")),u.push("})();"),u.join(`
 `)}var Cv=M(()=>{"use strict";kd();l(Lv,"buildSnippet")});function Dv(r){let{source:e,name:t,metadata:n}=r,{bindings:s,dispatchSize:o}=n,i=[];i.push("(async () => {"),i.push("const adapter = await navigator.gpu.requestAdapter();"),i.push("const device = await adapter.requestDevice();"),i.push(""),i.push("const wgsl = "+JSON.stringify(e)+";"),i.push(""),i.push("const pipeline = device.createComputePipeline({"),i.push('  layout: "auto",'),i.push("  compute: { module: device.createShaderModule({ code: wgsl }), entryPoint: "+JSON.stringify(t)+" }"),i.push("});"),i.push(""),i.push("const entries = [];"),i.push("const gpuBuffers = [];"),i.push("");for(let a of s)a.name==="_shapes"?_2(i,a):a.packed?b2(i,a):a.mode==="read_write"?w2(i,a):T2(i,a),i.push("");return i.push("const bindGroup = device.createBindGroup({ layout: pipeline.getBindGroupLayout(0), entries });"),i.push("const encoder = device.createCommandEncoder();"),i.push("const pass = encoder.beginComputePass();"),i.push("pass.setPipeline(pipeline);"),i.push("pass.setBindGroup(0, bindGroup);"),i.push(`pass.dispatchWorkgroups(${o[0]}, ${o[1]}, ${o[2]});`),i.push("pass.end();"),i.push(""),x2(i,s),i.push(""),i.push("gpuBuffers.forEach(b => b.destroy());"),i.push("device.destroy();"),i.push("})();"),i.join(`
 `)}function _2(r,e){r.push(`// binding ${e.index}: uniform (_shapes)`),r.push("{"),r.push("  const shapeData = new Uint32Array([/* shape values */]);"),r.push("  const size = Math.max(Math.ceil(shapeData.byteLength / 16) * 16, 16);"),r.push("  const buf = device.createBuffer({ size, usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST });"),r.push("  device.queue.writeBuffer(buf, 0, shapeData);"),r.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),r.push("  gpuBuffers.push(buf);"),r.push("}")}function b2(r,e){let t=e.mode==="read_write",n=t?"GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST":"GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST";if(r.push(`// binding ${e.index}: packed ${e.name} (${e.packed.length} buffers, ${e.packedSize} elems)`),r.push("{"),r.push(`  const buf = device.createBuffer({ size: ${e.packedSize*4}, usage: ${n}, mappedAtCreation: true });`),r.push("  const mapped = new Float32Array(buf.getMappedRange());"),t)r.push("  mapped.fill(0);");else for(let s of e.packed)r.push(`  // mapped.set(${s.name}_data, ${s.offset}); // size=${s.size}`);r.push("  buf.unmap();"),r.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),r.push("  gpuBuffers.push(buf);"),r.push("}")}function w2(r,e){r.push(`// binding ${e.index}: output ${e.name}`),r.push("{"),r.push("  const size = N * 4; // N = output element count"),r.push("  const buf = device.createBuffer({ size, usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST, mappedAtCreation: true });"),r.push("  new Float32Array(buf.getMappedRange()).fill(0);"),r.push("  buf.unmap();"),r.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),r.push("  gpuBuffers.push(buf);"),r.push("}")}function T2(r,e){r.push(`// binding ${e.index}: input ${e.name}`),r.push("{"),r.push(`  const data = ${e.name}_data; // Float32Array`),r.push("  const buf = device.createBuffer({ size: Math.max(data.byteLength, 4), usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST, mappedAtCreation: true });"),r.push("  new Float32Array(buf.getMappedRange()).set(data);"),r.push("  buf.unmap();"),r.push("  entries.push({ binding: "+e.index+", resource: { buffer: buf } });"),r.push("  gpuBuffers.push(buf);"),r.push("}")}function x2(r,e){if(e.filter(n=>n.mode==="read_write").length===0){r.push("device.queue.submit([encoder.finish()]);");return}for(let n=0;n<e.length;n++){let s=e[n];if(s.mode==="read_write")if(s.packed)for(let o of s.packed)r.push("{"),r.push(`  const rb = device.createBuffer({ size: ${o.size*4}, usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST });`),r.push(`  encoder.copyBufferToBuffer(gpuBuffers[${n}], ${o.offset*4}, rb, 0, ${o.size*4});`),r.push("  device.queue.submit([encoder.finish()]);"),r.push("  await rb.mapAsync(GPUMapMode.READ);"),r.push(`  console.log("${o.name}:", new Float32Array(rb.getMappedRange()));`),r.push("  rb.unmap(); rb.destroy();"),r.push("}");else r.push("{"),r.push(`  const src = gpuBuffers[${n}];`),r.push("  const rb = device.createBuffer({ size: src.size, usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST });"),r.push("  encoder.copyBufferToBuffer(src, 0, rb, 0, src.size);"),r.push("  device.queue.submit([encoder.finish()]);"),r.push("  await rb.mapAsync(GPUMapMode.READ);"),r.push(`  console.log("${s.name}:", new Float32Array(rb.getMappedRange()));`),r.push("  rb.unmap(); rb.destroy();"),r.push("}")}}var Pv=M(()=>{"use strict";l(Dv,"buildSnippet");l(_2,"emitShapeBinding");l(b2,"emitPackedBinding");l(w2,"emitOutputBinding");l(T2,"emitInputBinding");l(x2,"emitReadback")});function $v(r){let{source:e,name:t,metadata:n}=r,{blockDim:s,gridDim:o,sharedMemBytes:i,params:a}=n,u=[];return u.push(e),u.push(""),u.push("// Launch:"),u.push(`//   dim3 block(${s[0]}, ${s[1]}, ${s[2]});`),u.push(`//   dim3 grid(${o[0]}, ${o[1]}, ${o[2]});`),i>0?u.push(`//   ${t}<<<grid, block, ${i}>>>(${a.join(", ")});`):u.push(`//   ${t}<<<grid, block>>>(${a.join(", ")});`),u.join(`
-`)}var Fv=M(()=>{"use strict";l($v,"buildSnippet")});function vd(r,e){zv.set(r,e),e.runtimeKind&&e.snippet&&Vv.set(e.runtimeKind,e.snippet)}function Gv(r){return zv.get(r)||null}function jv(r){return Vv.get(r)||null}var zv,Vv,qv=M(()=>{"use strict";Ks();Tv();Av();Ov();Mv();Ev();Cv();Pv();Fv();zv=new Map,Vv=new Map;l(vd,"registerCodegen");l(Gv,"getCodegenEntry");l(jv,"getSnippetBuilder");vd(Ot.CPU,{runtimeKind:"js",snippet:Iv,compile(r,e){return{source:new yd(e).generate(r),metadata:{kind:"js",paramCount:r.params.length}}}});vd(Ot.WASM,{runtimeKind:"wasm",snippet:Lv,compile(r,e){let t=new bd(e).generate(r),n={kind:"wasm",memoryPages:t.memoryPages,bufferOffsets:t.bufferOffsets,imports:t.imports,params:t.params,bufferMap:r.bufferMap};return t.parallel&&(n.parallel=t.parallel),{source:t.wat,metadata:n}}});vd(Ot.WEBGPU,{runtimeKind:"webgpu",snippet:Dv,compile(r,e){let t=new xd(e).generate(r);return{source:t.source,metadata:{kind:"webgpu",workgroupSize:t.workgroupSize,dispatchSize:t.dispatchSize,sharedMemBytes:t.sharedMemBytes,params:t.params,bindings:t.bindings}}}});vd(Ot.CUDA,{runtimeKind:"cuda",snippet:$v,compile(r,e,t){if(t&&t.matmulBackend==="cublas"&&r.cublasInfo)return{source:"",metadata:{kind:"cuda",cublas:r.cublasInfo,outputIndices:[r.cublasInfo.cIdx]}};let n=new _d(e).generate(r);return{source:n.source,metadata:{kind:"cuda",blockDim:n.blockDim,gridDim:n.gridDim,sharedMemBytes:n.sharedMemBytes,params:n.params,outputIndices:n.outputIndices,scratch:n.scratch}}}})});function Uv(r){let e=[],t=[r.body];for(;t.length>0;){let m=t.pop();if(m){if(m.type==="BlockNode"){e.push(m),t.push(m.body);continue}if(m.body&&t.push(m.body),m.stmts)for(let g of m.stmts)t.push(g);m.thenBody&&t.push(m.thenBody),m.elseBody&&t.push(m.elseBody)}}let n=null;for(let m of e)if(m.name.includes("matmul"))m.reads.length>=2&&m.writes.length>=1&&(n=m);else return null;if(!n)return null;let s=n.reads[0].buffer,o=n.reads[1].buffer,i=n.writes[0].buffer;if(s.dtype!=="f32"||o.dtype!=="f32"||i.dtype!=="f32"||s.shape.length!==2||o.shape.length!==2||i.shape.length!==2)return null;let a=i.shape[0],u=i.shape[1],p=s.shape[1];if(![a,u,p].every(m=>typeof m=="number"))return null;let c=[];for(let[,m]of r.bufferMap)c.push(m.name);let h=c.indexOf(s.name),f=c.indexOf(o.name),d=c.indexOf(i.name);return h<0||f<0||d<0?null:{M:a,N:u,K:p,aIdx:h,bIdx:f,cIdx:d}}var Ic,Hs,Ec=M(()=>{"use strict";qv();l(Uv,"detectPureMatmul");Ic=class{static{l(this,"CompiledKernel")}constructor(e,t,n,s={}){this.name=e,this.source=t,this.target=n,this.metadata=s}snippet(){let e=jv(this.metadata.kind);if(!e)throw new Error(`No snippet for kind: ${this.metadata.kind}`);return e(this)}},Hs=class{static{l(this,"BackendPipeline")}constructor(e,t={}){this.target=e,this.matmulBackend=t.matmulBackend||"native",this.context=t.context||null}compile(e){let t=this.context&&this.context.getCodegenEntry(this.target.kind)||Gv(this.target.kind);if(!t)throw new Error(`Unsupported target kind: ${this.target.kind}`);let{source:n,metadata:s}=t.compile(e,this.target,this);return new Ic(e.name,n,this.target,s)}compileAll(e){return e.map(t=>this.compile(t))}}});function Wv(r,e){Kv.set(r,e)}function Hv(r){return Kv.get(r)||null}var Kv,rw=M(()=>{"use strict";Kv=new Map;l(Wv,"registerMeasurer");l(Hv,"getMeasurer")});var Yv={};nt(Yv,{runWasmParallel:()=>S2});async function S2(r,e,t,n,s){let{exports:o,memory:i,bufferOffsets:a}=r,u=o[e],p=[...a.values()],c=Math.min(p.length,t.length);for(let f=0;f<c;f++){let d=t[f];ArrayBuffer.isView(d)&&new d.constructor(i.buffer,p[f],d.length).set(d)}let h=p.slice(0,c);if(n)for(let f of n)h.push(f);h.push(0,s.extent),u(...h);for(let f=0;f<c;f++){let d=t[f];ArrayBuffer.isView(d)&&d.set(new d.constructor(i.buffer,p[f],d.length))}}var Jv=M(()=>{"use strict";l(S2,"runWasmParallel")});var Nd={};nt(Nd,{deviceClipGradNorm:()=>A2,measureCudaKernel:()=>v2,runCudaKernel:()=>k2,runCudaPlan:()=>N2});function k2(){throw new Error("CUDA runtime is not available in the browser")}function v2(){throw new Error("CUDA runtime is not available in the browser")}function N2(){throw new Error("CUDA runtime is not available in the browser")}function A2(){throw new Error("CUDA runtime is not available in the browser")}var Ad=M(()=>{"use strict";l(k2,"runCudaKernel");l(v2,"measureCudaKernel");l(N2,"runCudaPlan");l(A2,"deviceClipGradNorm")});function ii(){return Zv}function O2(r){Zv=r}var Qi,ks,Od,sw,Ys,Au,Zv,ar=M(()=>{"use strict";Nn();Qi=class{static{l(this,"Device")}type;index;constructor(e,t=0){this.type=e,this.index=t}dispatchKey(){return ja(this.type)}equals(e){return this.type===e.type&&this.index===e.index}hash(){let e=2166136261;for(let t=0;t<this.type.length;t++)e=(e^this.type.charCodeAt(t))*16777619&2147483647;return e=(e^this.index)*16777619&2147483647,e}toString(){return this.index===0?this.type:`${this.type}:${this.index}`}},ks=new Qi("cpu"),Od=new Qi("gpu"),sw=new Qi("wasm"),Ys=new Qi("webgpu"),Au=new Qi("meta"),Zv=ks;l(ii,"getDefaultDevice");l(O2,"setDefaultDevice")});function Xv(r){let e=Qv.get(r)??null;return e||(e=Nt.findOp(r),e&&Qv.set(r,e)),e||null}var Qv,eN=M(()=>{"use strict";$i();Qv=new Map;l(Xv,"getHandle")});function vs(r,e){return r<0?e+r:r}function xn(r){let e=r.length,t=new Array(e),n=1;for(let s=e-1;s>=0;s--)t[s]=n,n*=r[s];return t}function cn(r){let e=1;for(let t=0;t<r.length;t++)e*=r[t];return e}function ow(r,e){let t=r.length;if(t===0)return!0;let n=1;for(let s=t-1;s>=0;s--){let o=r[s];if(o===0)return!0;if(o!==1&&e[s]!==n)return!1;n*=o}return!0}function Ou(r,e){let t=r.length,n=e.length,s=Math.max(t,n),o=new Array(s);for(let i=0;i<s;i++){let a=i<t?r[t-1-i]:1,u=i<n?e[n-1-i]:1;if(a===u)o[s-1-i]=a;else if(a===1)o[s-1-i]=u;else if(u===1)o[s-1-i]=a;else return null}return o}function Bd(r,e){let t=r.length,n=e.length;if(t===1&&n===1)return[];let s=t===1?[1,r[0]]:[...r],o=n===1?[e[0],1]:[...e],i=s.length,a=o.length,u=Ou(s.slice(0,i-2),o.slice(0,a-2));if(u===null)return null;let p=[...u,s[i-2],o[a-1]],c=new Set;return n===1&&c.add(p.length-1),t===1&&c.add(p.length-2),p.filter((h,f)=>!c.has(f))}function tN(r,e,t){let n=cn(r),s=-1,o=1,i=new Array(t.length);for(let a=0;a<t.length;a++){let u=t[a];if(u===-1){if(s!==-1)return null;s=a,i[a]=-1}else i[a]=u,o*=u}if(s!==-1){if(o===0||(i[s]=n/o|0,i[s]*o!==n))return null}else if(cn(i)!==n)return null;return ow(r,e)?{sizes:i,strides:xn(i),needsCopy:!1}:{sizes:i,strides:xn(i),needsCopy:!0}}var jr=M(()=>{"use strict";l(vs,"normalizeAxis");l(xn,"computeStrides");l(cn,"computeNumel");l(ow,"isContiguous");l(Ou,"broadcastShapes");l(Bd,"matmulOutputShape");l(tN,"inferReshape")});function nN(r,e){if(r.length<4)return!1;let t=r.length,n=1,s=t===4?B2:R2;if(!s)return!1;for(let o=t-1;o>=0;o--){let i=s[o];if(i===void 0||r[i]===0||e[i]!==n)return!1;n*=r[i]}return!0}var B2,R2,iw=M(()=>{"use strict";B2=Object.freeze([0,2,3,1]),R2=Object.freeze([0,2,3,4,1]);l(nN,"isChannelsLast")});var Bt,Oo=M(()=>{"use strict";Nn();jr();iw();iw();Bt=class{static{l(this,"TensorImpl")}_storage;_storageOffset;_sizes;_strides;_dtype;_device;_numel;_keySet;_autogradMeta;_version;_contiguousCache;constructor(e,t,n,s,o,i){this._storage=e,this._storageOffset=t,this._sizes=Object.freeze([...n]),this._strides=Object.freeze(s?[...s]:xn(n)),this._dtype=o,this._device=i,this._numel=cn(n),this._keySet=Pr.fromKey(ja(i.type)),this._autogradMeta=null,this._version=0,this._contiguousCache=null}get storage(){return this._storage}get storageOffset(){return this._storageOffset}size(e){let t=e<0?this._sizes.length+e:e;return this._sizes[t]}stride(e){let t=e<0?this._strides.length+e:e;return this._strides[t]}sizes(){return this._sizes}strides(){return this._strides}dim(){return this._sizes.length}numel(){return this._numel}get dtype(){return this._dtype}get device(){return this._device}isContiguous(e){return!e||e==="contiguous"?(this._contiguousCache===null&&(this._contiguousCache=ow(this._sizes,this._strides)),this._contiguousCache):e==="channels_last"?nN(this._sizes,this._strides):!1}setSizesAndStrides(e,t){this._sizes=Object.freeze([...e]),this._strides=Object.freeze(t?[...t]:xn(e)),this._numel=cn(e),this._contiguousCache=null}setStorageOffset(e){this._storageOffset=e}bumpVersion(){this._version++}get version(){return this._version}get autogradMeta(){return this._autogradMeta}setAutogradMeta(e){this._autogradMeta=e,this._updateKeySet()}keySet(){return this._keySet}addKeyToSet(e){this._keySet=this._keySet.add(e)}removeKeyFromSet(e){this._keySet=this._keySet.remove(e)}_updateKeySet(){let e=Pr.fromKey(ja(this._device.type));if(this._autogradMeta&&this._autogradMeta.requiresGrad){let t=ja(this._device.type);e=e.add(RS(t))}this._keySet=e}get isMeta(){return this._storage&&this._storage.isMeta}shallowCopyFrom(e){this._storage=e._storage,this._storageOffset=e._storageOffset,this._sizes=e._sizes,this._strides=e._strides,this._dtype=e._dtype,this._device=e._device,this._numel=e._numel,this._keySet=e._keySet,this._contiguousCache=null}}});var Bu,aw=M(()=>{"use strict";Bu=class{static{l(this,"AutogradMeta")}_grad;_gradFn;_outputNr;_requiresGrad;_retainGrad;_gradAccumulator;_versionAtCreation;constructor(){this._grad=null,this._gradFn=null,this._outputNr=0,this._requiresGrad=!1,this._retainGrad=!1,this._gradAccumulator=null,this._versionAtCreation=0}get grad(){return this._grad}set grad(e){this._grad=e}get gradFn(){return this._gradFn}setGradFn(e,t){this._gradFn=e,this._outputNr=t??0}get outputNr(){return this._outputNr}get requiresGrad(){return this._requiresGrad}set requiresGrad(e){this._requiresGrad=e}get retainGrad(){return this._retainGrad}set retainGrad(e){this._retainGrad=e}get isLeaf(){return this._gradFn===null}get versionAtCreation(){return this._versionAtCreation}set versionAtCreation(e){this._versionAtCreation=e}getGradAccumulator(){if(this._gradAccumulator){let e=this._gradAccumulator instanceof WeakRef?this._gradAccumulator.deref():this._gradAccumulator;if(e)return e}return null}setGradAccumulator(e){this._gradAccumulator=typeof WeakRef<"u"?new WeakRef(e):e}clearGrad(){this._grad=null}clearGradFn(){this._gradFn=null,this._outputNr=0}}});function sN(r,e,t,n,s,o){let i=e[s];if(s===e.length-1){let u=new Array(i);if(o)for(let p=0;p<i;p++)u[p]=md(o,r[n+p*t[s]]);else for(let p=0;p<i;p++)u[p]=r[n+p*t[s]];return u}let a=new Array(i);for(let u=0;u<i;u++)a[u]=sN(r,e,t,n+u*t[s],s+1,o);return a}function M2(){if(!uw)throw new Error("Autograd engine not initialized. Import autograd/engine.js first.");return uw}function oN(r){uw=r}var rN,We,uw,jn=M(()=>{"use strict";Oo();aw();xu();rN=new Set(["f16","bf16","i64"]),We=class r{static{l(this,"Tensor")}_impl;constructor(e){this._impl=e}get impl(){return this._impl}get shape(){return this._impl.sizes()}get strides(){return this._impl.strides()}get dtype(){return this._impl.dtype}get device(){return this._impl.device}get ndim(){return this._impl.dim()}get rank(){return this._impl.dim()}get numel(){return this._impl.numel()}get length(){return this._impl.numel()}get isContiguous(){return this._impl.isContiguous()}get dispatchKeySet(){return this._impl.keySet()}get storage(){return this._impl.storage}get storageOffset(){return this._impl.storageOffset}get data(){let e=this._impl.storage;if(!e||e.isMeta)return null;let t=e.data,n=this._impl.storageOffset;return t&&n===0&&this._impl.isContiguous()&&t.length===this.numel,t}get requiresGrad(){let e=this._impl.autogradMeta;return e?e.requiresGrad:!1}get gradFn(){let e=this._impl.autogradMeta;return e?e.gradFn:null}get grad(){let e=this._impl.autogradMeta;return e?e.grad:null}set grad(e){this._ensureAutogradMeta().grad=e}get isLeaf(){let e=this._impl.autogradMeta;return e?e.isLeaf:!0}get version(){return this._impl.version}requiresGrad_(e=!0){let t=this._ensureAutogradMeta();return t.requiresGrad=e,this._impl._updateKeySet(),this}retainGrad(){return this._ensureAutogradMeta().retainGrad=!0,this}detach(){let e=new Bt(this._impl.storage,this._impl.storageOffset,this._impl.sizes(),this._impl.strides(),this._impl.dtype,this._impl.device);return new r(e)}item(){if(this.numel!==1)throw new Error(`item() requires tensor with exactly 1 element, got ${this.numel}`);let t=this._impl.storage.data[this._impl.storageOffset];return rN.has(this._impl.dtype)?md(this._impl.dtype,t):t}toArray(){let e=this.shape,t=this.strides,n=this._impl.storage.data,s=this._impl.storageOffset,o=rN.has(this._impl.dtype)?this._impl.dtype:null;return e.length===0?o?md(o,n[s]):n[s]:sN(n,e,t,s,0,o)}toString(){return`Tensor(shape=[${this.shape.join(", ")}], dtype=${this.dtype}, device=${this.device})`}*[Symbol.iterator](){let e=this.shape[0];if(e===void 0)throw new Error("Cannot iterate over a 0-d tensor");for(let t=0;t<e;t++)yield this._select(0,t)}_select(e,t){let n=this._impl.sizes(),s=this._impl.strides(),o=[],i=[];for(let p=0;p<n.length;p++)p!==e&&(o.push(n[p]),i.push(s[p]));let a=this._impl.storageOffset+t*s[e],u=new Bt(this._impl.storage,a,o,i,this._impl.dtype,this._impl.device);return new r(u)}_ensureAutogradMeta(){if(!this._impl.autogradMeta){let e=new Bu;e.versionAtCreation=this._impl.version,this._impl.setAutogradMeta(e)}return this._impl.autogradMeta}backward(e){let{backward:t}=M2();t(this,e)}};l(sN,"_toNestedArray");uw=null;l(M2,"require_autograd_engine");l(oN,"setAutogradEngine")});function Dt(r){return E2[r]||Float32Array}function Rd(r,e){if(r===e)return r;let t=Ru.get(r)??0,n=Ru.get(e)??0;return t>=n?r:e}function Xe(r,e){let t=uN.get(r);t||(t=new Set,uN.set(r,t)),t.add(e)}function Xi(r){return I2(r)}var W,I2,E2,iN,aN,Ru,uN,ts=M(()=>{"use strict";Oe();W=zn,I2=Gs,E2=Object.freeze({[W.F16]:Uint16Array,[W.BF16]:Uint16Array,[W.F32]:Float32Array,[W.F64]:Float64Array,[W.I8]:Int8Array,[W.I16]:Int16Array,[W.I32]:Int32Array,[W.I64]:BigInt64Array,[W.UI8]:Uint8Array,[W.BOOL]:Uint8Array,[W.INDEX]:Int32Array});l(Dt,"typedArrayCtor");iN=[W.F16,W.BF16,W.F32,W.F64],aN=[W.UI8,W.I8,W.I16,W.I32,W.I64],Ru=new Map;for(let r=0;r<iN.length;r++)Ru.set(iN[r],100+r);for(let r=0;r<aN.length;r++)Ru.set(aN[r],r);Ru.set(W.BOOL,-1);Ru.set(W.INDEX,50);l(Rd,"resultDtype");uN=new Map;l(Xe,"_addCastPair");Xe(W.BOOL,W.I8);Xe(W.BOOL,W.I16);Xe(W.BOOL,W.I32);Xe(W.BOOL,W.I64);Xe(W.BOOL,W.F16);Xe(W.BOOL,W.F32);Xe(W.BOOL,W.F64);Xe(W.UI8,W.I16);Xe(W.UI8,W.I32);Xe(W.UI8,W.I64);Xe(W.UI8,W.F16);Xe(W.UI8,W.F32);Xe(W.UI8,W.F64);Xe(W.I8,W.I16);Xe(W.I8,W.I32);Xe(W.I8,W.I64);Xe(W.I8,W.F16);Xe(W.I8,W.F32);Xe(W.I8,W.F64);Xe(W.I16,W.I32);Xe(W.I16,W.I64);Xe(W.I16,W.F32);Xe(W.I16,W.F64);Xe(W.I32,W.I64);Xe(W.I32,W.F64);Xe(W.F16,W.F32);Xe(W.F16,W.F64);Xe(W.BF16,W.F32);Xe(W.BF16,W.F64);Xe(W.BOOL,W.BF16);Xe(W.UI8,W.BF16);Xe(W.I8,W.BF16);Xe(W.F32,W.F64);l(Xi,"dtypeSize")});function lw(r){return lN.get(r)||lN.get("cpu")}var Md,Id,lN,Lc,cN=M(()=>{"use strict";ts();ar();Md=class{static{l(this,"CPUAllocator")}allocate(e,t){let n=Dt(t);return new n(e)}free(e){}},Id=class{static{l(this,"MetaAllocator")}allocate(){return null}free(e){}},lN=new Map([["cpu",new Md],["wasm",new Md],["meta",new Id],["lazy",new Id]]);l(lw,"getAllocator");Lc=class r{static{l(this,"StorageImpl")}static#e=null;_data;_nbytes;_device;_allocator;_refCount;static setHostReadHook(e){r.#e=e}constructor(e,t,n,s){this._data=e,this._nbytes=t,this._device=n,this._allocator=s,this._refCount=1}static allocate(e,t,n){let s=lw(n.type),o=Xi(t),i=Math.max(Math.ceil(e/o),1),a=s.allocate(i,t);return new r(a,e,n,s)}static fromData(e,t){let n=e?e.byteLength:0,s=lw(t.type);return new r(e,n,t,s)}retain(){return this._refCount++,this}release(){this._refCount--,this._refCount===0&&(this._allocator&&this._data&&this._allocator.free(this._data),this._data=null)}get refCount(){return this._refCount}get data(){return r.#e&&this._data&&r.#e(this._data),this._data}get rawData(){return this._data}get nbytes(){return this._nbytes}get device(){return this._device}get isValid(){return this._refCount>0&&this._data!==null}get isMeta(){return this._data===null}resize(e,t){if(e<=this._nbytes&&this._data)return;let n=Xi(t),s=Math.max(Math.ceil(e/n),1),o=this._allocator.allocate(s,t);if(this._data&&o){let i=Math.min(this._data.length,o.length);for(let a=0;a<i;a++)o[a]=this._data[a]}this._allocator&&this._data&&this._allocator.free(this._data),this._data=o,this._nbytes=e}clone(){let e=lw(this._device.type),t=null;if(this._data){r.#e&&r.#e(this._data);let n=this._data.constructor;t=new n(this._data.length),t.set(this._data)}return new r(t,this._nbytes,this._device,e)}}});var En,ea=M(()=>{"use strict";cN();En=class r{static{l(this,"Storage")}_impl;constructor(e){this._impl=e}static allocate(e,t,n){return new r(Lc.allocate(e,t,n))}static fromData(e,t){return new r(Lc.fromData(e,t))}get impl(){return this._impl}get data(){return this._impl.data}get rawData(){return this._impl.rawData}get nbytes(){return this._impl.nbytes}get device(){return this._impl.device}get isValid(){return this._impl.isValid}get isMeta(){return this._impl.isMeta}retain(){return this._impl.retain(),this}release(){this._impl.release()}clone(){return new r(this._impl.clone())}resize(e,t){this._impl.resize(e,t)}}});function Pt(r,e){let t=e?.dtype??W.F32,n=e?.device??ii(),s=e?.requiresGrad??!1;if(ArrayBuffer.isView(r)&&"length"in r)return L2(r,e?.shape,t,n,s);if(Array.isArray(r)){let{flat:o,shape:i}=D2(r),a=e?.shape??i;return C2(o,a,t,n,s)}if(typeof r=="number")return pN(r,t,n,s);throw new Error("Unsupported data type for tensor()")}function ta(r,e,t,n){let s=n?.device??ii(),o=n?.requiresGrad??!1,i=xn(e),a=En.fromData(r,s),u=n?.offset??0,p=new Bt(a,u,e,i,t,s),c=new We(p);return o&&c.requiresGrad_(!0),c}function Dc(r,e){let t=e?.dtype??W.F32,n=e?.device??ii();return pN(r,t,n,e?.requiresGrad??!1)}function pN(r,e,t,n){let s=Dt(e),o=new s(1),i=o;i[0]=cw.has(e)?dd(e,r):r;let a=En.fromData(o,t),u=new Bt(a,0,[],[],e,t),p=new We(u);return n&&p.requiresGrad_(!0),p}function L2(r,e,t,n,s){let o=Dt(t),i=new o(r.length),a=i;if(cw.has(t))for(let d=0;d<r.length;d++)a[d]=dd(t,r[d]);else for(let d=0;d<r.length;d++)a[d]=r[d];let u=e??[r.length],p=xn(u),c=En.fromData(i,n),h=new Bt(c,0,u,p,t,n),f=new We(h);return s&&f.requiresGrad_(!0),f}function C2(r,e,t,n,s){let o=Dt(t),i=new o(r.length),a=i;if(cw.has(t))for(let f=0;f<r.length;f++)a[f]=dd(t,r[f]);else for(let f=0;f<r.length;f++)a[f]=r[f];let u=xn(e),p=En.fromData(i,n),c=new Bt(p,0,e,u,t,n),h=new We(c);return s&&h.requiresGrad_(!0),h}function D2(r){let e=[],t=r;for(;Array.isArray(t);)e.push(t.length),t=t[0];let n=[];return fN(r,n,e,0),{flat:n,shape:e}}function fN(r,e,t,n){if(n===t.length-1){for(let s=0;s<r.length;s++)e.push(r[s]);return}for(let s=0;s<r.length;s++)fN(r[s],e,t,n+1)}var cw,qr=M(()=>{"use strict";jn();Oo();ea();ts();ar();jr();xu();l(Pt,"tensor");l(ta,"fromBuffer");l(Dc,"scalar");cw=new Set(["f16","bf16","i64"]);l(pN,"_fromScalar");l(L2,"_fromTypedArray");l(C2,"_fromFlatArray");l(D2,"_flattenNested");l(fN,"_recurFlatten")});var Ft={};nt(Ft,{_dispatch:()=>ot,abs:()=>Pc,add:()=>ae,argmax:()=>Ud,argmin:()=>kN,argsort:()=>LN,batch_norm:()=>fw,broadcast_in_dim:()=>ON,cat:()=>sa,ceil:()=>wN,chunk:()=>BN,clamp:()=>Vd,clone:()=>AN,contiguous:()=>ci,conv2d:()=>hw,cos:()=>mN,cumsum:()=>IN,div:()=>$t,dot:()=>NN,embedding:()=>z2,eq:()=>$c,erf:()=>gN,erfc:()=>yN,exp:()=>Kr,expand:()=>Lu,fill:()=>F2,flip:()=>MN,floor:()=>bN,gamma:()=>Cd,gather:()=>jd,ge:()=>zd,gelu:()=>Dd,gt:()=>Iu,index_select:()=>li,layer_norm:()=>pw,le:()=>Fd,lgamma:()=>_N,log:()=>Zs,log_softmax:()=>Hd,lt:()=>xN,matmul:()=>lt,max:()=>Qs,maximum:()=>ai,mean:()=>lr,min:()=>SN,minimum:()=>Ld,mul:()=>ee,narrow:()=>Un,ne:()=>TN,neg:()=>Ur,one_hot:()=>Gd,pad:()=>na,permute:()=>Eu,pool2d:()=>Yd,pow:()=>Js,prod:()=>vN,relu:()=>Mu,remainder:()=>$2,repeat:()=>Du,reshape:()=>Cn,roll:()=>RN,rsqrt:()=>hN,scatter:()=>qd,scatter_add:()=>ra,select:()=>Ut,sigmoid:()=>ur,sign:()=>$d,silu:()=>Pd,sin:()=>dN,slice:()=>Kd,softmax:()=>Ro,sort:()=>EN,split:()=>Bo,sqrt:()=>ui,squeeze:()=>Cu,stack:()=>Wr,sub:()=>Pe,sum:()=>ht,tanh:()=>Ns,tile:()=>Wd,topk:()=>CN,transpose:()=>gt,unsqueeze:()=>Sr,where:()=>qn});function P2(r){return typeof r=="object"&&r!==null&&"_impl"in r}function Ln(r,e){return P2(r)?r:Dc(r,{dtype:e.dtype,device:e.device})}function ot(r,...e){let t=Xv(r);if(!t)return Nt.callOp(r,...e);let n=pb(e,t.schema);return Nt.dispatch(t,n,...e)}function se(r,...e){return ot(r,...e)}function Ed(r,...e){return ot(r,...e)}function ae(r,e){return se("add",r,Ln(e,r))}function Pe(r,e){return se("sub",r,Ln(e,r))}function ee(r,e){return se("mul",r,Ln(e,r))}function $t(r,e){return se("div",r,Ln(e,r))}function Ur(r){return se("neg",r)}function Js(r,e){return se("pow",r,Ln(e,r))}function $2(r,e){return se("rem",r,Ln(e,r))}function ai(r,e){return se("maximum",r,Ln(e,r))}function Ld(r,e){return se("minimum",r,Ln(e,r))}function Kr(r){return se("exp",r)}function Zs(r){return se("log",r)}function ui(r){return se("sqrt",r)}function hN(r){return se("rsqrt",r)}function Pc(r){return se("abs",r)}function dN(r){return se("sin",r)}function mN(r){return se("cos",r)}function Ns(r){return se("tanh",r)}function gN(r){return se("erf",r)}function yN(r){return se("erfc",r)}function _N(r){return se("lgamma",r)}function Cd(r){return se("gamma",r)}function ur(r){return se("sigmoid",r)}function Mu(r){return se("relu",r)}function Dd(r){return se("gelu",r)}function Pd(r){return se("silu",r)}function $d(r){return se("sign",r)}function bN(r){return se("floor",r)}function wN(r){return se("ceil",r)}function $c(r,e){return se("eq",r,Ln(e,r))}function TN(r,e){return se("ne",r,Ln(e,r))}function xN(r,e){return se("lt",r,Ln(e,r))}function Fd(r,e){return se("le",r,Ln(e,r))}function Iu(r,e){return se("gt",r,Ln(e,r))}function zd(r,e){return se("ge",r,Ln(e,r))}function qn(r,e,t){return se("where",r,e,t)}function Vd(r,e,t){return se("clamp",r,Ln(e,r),Ln(t,r))}function na(r,e,t,n=0){return se("pad",r,Ln(n,r),e,t)}function Gd(r,e){return se("one_hot",r,e)}function li(r,e,t){return se("index_select",r,t,e)}function jd(r,e,t){return se("gather",r,t,e)}function ra(r,e,t,n){return se("scatter_add",r,t,n,e)}function qd(r,e,t,n){return se("scatter",r,e,t,n)}function ht(r,e,t){return se("sum",r,e,t)}function lr(r,e,t){return se("mean",r,e,t)}function Qs(r,e,t){return se("max",r,e,t)}function SN(r,e,t){return se("min",r,e,t)}function Ud(r,e,t){return se("argmax",r,e,t)}function kN(r,e,t){return se("argmin",r,e,t)}function vN(r,e,t){return se("prod",r,e,t)}function lt(r,e){let t=dw();if(t){let n=t(r,e);if(n!==null)return n}return se("matmul",r,e)}function NN(r,e){return se("dot",r,e)}function sa(r,e){return se("cat",r,e)}function Wr(r,e){return se("stack",r,e)}function AN(r){return se("clone",r)}function F2(r,e){return se("fill",r,e)}function Cn(r,e){return se("reshape",r,e)}function gt(r,e,t){return se("transpose",r,e,t)}function Eu(r,e){return se("permute",r,e)}function ON(r,e,t){return se("broadcast_in_dim",r,e,t)}function Lu(r,e){return se("expand",r,e)}function Kd(r,e,t,n=null,s=1){return se("slice",r,e,t,n,s)}function Sr(r,e){return se("unsqueeze",r,e)}function Cu(r,e=null){return se("squeeze",r,e)}function Un(r,e,t,n){return se("narrow",r,e,t,n)}function Ut(r,e,t){return se("select",r,e,t)}function ci(r){return se("contiguous",r)}function Du(r,e){return se("repeat",r,e)}function Wd(r,e){return se("tile",r,e)}function Bo(r,e,t=0){if(typeof e!="number")return Ed("split",r,e,t);let n=r.shape.length,s=t<0?n+t:t,o=r.shape[s],i=[];for(let a=0;a<o;a+=e)i.push(Math.min(e,o-a));return Ed("split",r,i,t)}function BN(r,e,t=0){return Ed("chunk",r,e,t)}function RN(r,e,t=0){return se("roll",r,e,t)}function MN(r,e){return se("flip",r,Array.isArray(e)?e:[e])}function IN(r,e=0){return se("cumsum",r,e)}function EN(r,e=-1,t=!1){return se("sort",r,e,t)}function LN(r,e=-1,t=!1){return se("argsort",r,e,t)}function CN(r,e,t=-1,n=!0){return Ed("topk",r,e,t,n)}function Ro(r,e){return se("softmax",r,e)}function Hd(r,e){return se("log_softmax",r,e)}function pw(r,e,t,n,s){return se("layer_norm",r,e,t,n,s)}function fw(r,e,t,n,s,o,i){return se("batch_norm",r,e,t,n,s,o,i)}function hw(r,e,t,n,s,o){return se("conv2d",r,e,t,n,s,o)}function Yd(r,e,t,n,s){return se("pool2d",r,e,t,n,s)}function z2(r,e){return se("embedding",r,e)}var be=M(()=>{"use strict";$i();eN();qr();Kn();l(P2,"hasTensorImpl");l(Ln,"_asTensor");l(ot,"_dispatch");l(se,"_dispatchTensor");l(Ed,"_dispatchTensorArray");l(ae,"add");l(Pe,"sub");l(ee,"mul");l($t,"div");l(Ur,"neg");l(Js,"pow");l($2,"remainder");l(ai,"maximum");l(Ld,"minimum");l(Kr,"exp");l(Zs,"log");l(ui,"sqrt");l(hN,"rsqrt");l(Pc,"abs");l(dN,"sin");l(mN,"cos");l(Ns,"tanh");l(gN,"erf");l(yN,"erfc");l(_N,"lgamma");l(Cd,"gamma");l(ur,"sigmoid");l(Mu,"relu");l(Dd,"gelu");l(Pd,"silu");l($d,"sign");l(bN,"floor");l(wN,"ceil");l($c,"eq");l(TN,"ne");l(xN,"lt");l(Fd,"le");l(Iu,"gt");l(zd,"ge");l(qn,"where");l(Vd,"clamp");l(na,"pad");l(Gd,"one_hot");l(li,"index_select");l(jd,"gather");l(ra,"scatter_add");l(qd,"scatter");l(ht,"sum");l(lr,"mean");l(Qs,"max");l(SN,"min");l(Ud,"argmax");l(kN,"argmin");l(vN,"prod");l(lt,"matmul");l(NN,"dot");l(sa,"cat");l(Wr,"stack");l(AN,"clone");l(F2,"fill");l(Cn,"reshape");l(gt,"transpose");l(Eu,"permute");l(ON,"broadcast_in_dim");l(Lu,"expand");l(Kd,"slice");l(Sr,"unsqueeze");l(Cu,"squeeze");l(Un,"narrow");l(Ut,"select");l(ci,"contiguous");l(Du,"repeat");l(Wd,"tile");l(Bo,"split");l(BN,"chunk");l(RN,"roll");l(MN,"flip");l(IN,"cumsum");l(EN,"sort");l(LN,"argsort");l(CN,"topk");l(Ro,"softmax");l(Hd,"log_softmax");l(pw,"layer_norm");l(fw,"batch_norm");l(hw,"conv2d");l(Yd,"pool2d");l(z2,"embedding")});var $N={};nt($N,{create:()=>DN,default:()=>V2,globals:()=>PN});var DN,PN,V2,FN=M(()=>{DN=l(()=>{throw new Error('mlfw: the "webgpu" npm package is not available in the browser; use navigator.gpu instead')},"create"),PN=void 0,V2={create:DN,globals:PN}});var XN={};nt(XN,{ensureWebGPUEager:()=>H2,flushWebGPUEager:()=>ZN,instantiateWebGPU:()=>q2,prewarmPipelines:()=>YN,recordWebGPUEager:()=>Gc,resetDevice:()=>G2,runWebGPUKernel:()=>U2,runWebGPUPlan:()=>K2,webgpuEagerOp:()=>Z2,webgpuEagerReady:()=>Y2,webgpuRNN:()=>tC});function $u(r){switch(Mn(r)){case"i32":return Int32Array;case"u32":return Uint32Array;case"f16":return Uint16Array;default:return Float32Array}}function Qd(r,e,t,n){if(t==="bf16")for(let s=0;s<e.length;s++)r[n+s]=fd(e[s]);else if(t==="i64")for(let s=0;s<e.length;s++)r[n+s]=Number(BigInt.asIntN(32,e[s]));else r.set(e,n)}function bw(r,e,t,n,s){if(t==="bf16")for(let o=0;o<s;o++)r[o]=hd(e[n+o]);else if(t==="i64")for(let o=0;o<s;o++)r[o]=BigInt(e[n+o]);else r.set(e.subarray(n,n+s))}function eo(r){return Math.ceil(r/4)*4}async function ww(){return kt||Fc||(Fc=(async()=>{let r=typeof navigator<"u"&&navigator.gpu;if(!r)try{let i=await Promise.resolve().then(()=>(FN(),$N));qN=i,r=i.create([]),i.globals&&(zc=i.globals.GPUBufferUsage,Jd=i.globals.GPUMapMode,Zd=i.globals.GPUShaderStage)}catch{throw new Error('WebGPU not available: install the "webgpu" npm package or run in a browser with WebGPU support')}!zc&&typeof GPUBufferUsage<"u"&&(zc=GPUBufferUsage,Jd=GPUMapMode,Zd=GPUShaderStage);let e=await r.requestAdapter();if(!e)throw new Error("WebGPU: no adapter found");let t=e.limits||{},n={},s=["maxStorageBuffersPerShaderStage","maxStorageBufferBindingSize","maxBufferSize","maxBindingsPerBindGroup"];for(let i of s)t[i]!==void 0&&(n[i]=t[i]);let o=[];return e.features&&e.features.has("shader-f16")&&o.push("shader-f16"),kt=await e.requestDevice({requiredLimits:n,requiredFeatures:o}),!zN&&typeof process<"u"&&process.on&&(zN=!0,process.on("exit",()=>{kt&&(kt.destroy(),kt=null)})),kt})(),Fc)}function Fu(){return zc}function Tw(){return Jd.READ}function G2(){kt&&(kt.destroy(),kt=null),Fc=null,zc=null,Jd=null,Zd=null,qN=null}function j2(r){return r.name==="_shapes"?"uniform":r.mode==="read_write"?"storage":"read-only-storage"}function UN(r,e){let t=r.createShaderModule({code:e.source}),n=[];for(let i of e.metadata.bindings)n.push({binding:i.index,visibility:Zd.COMPUTE,buffer:{type:j2(i)}});let s=r.createBindGroupLayout({entries:n}),o=r.createPipelineLayout({bindGroupLayouts:[s]});return{shaderModule:t,bindGroupLayout:s,pipelineLayout:o}}function KN(r,e){let{shaderModule:t,bindGroupLayout:n,pipelineLayout:s}=UN(r,e);return{pipeline:r.createComputePipeline({layout:s,compute:{module:t,entryPoint:e.name}}),bindGroupLayout:n}}function WN(r,e){let t=Xd.get(e);return t||(t=KN(r,e),Xd.set(e,t)),t}async function q2(r){let e=await ww(),{pipeline:t,bindGroupLayout:n}=KN(e,r);return{device:e,pipeline:t,bindGroupLayout:n,kernel:r,workgroupSize:r.metadata.workgroupSize,dispatchSize:r.metadata.dispatchSize,bindings:r.metadata.bindings}}function HN(r){let e=new Map,t=0;for(let n of r)if(n.name!=="_shapes")if(n.packed)for(let s of n.packed)e.set(s.name,s.argIndex!==void 0?s.argIndex:t++);else e.set(n.name,n.argIndex!==void 0?n.argIndex:t++);return e}async function U2(r,e,t){let{device:n,pipeline:s,bindGroupLayout:o,bindings:i,dispatchSize:a}=r,u=Fu(),p=[],c=[],h=HN(i);for(let y=0;y<i.length;y++){let _=i[y];if(_.name==="_shapes"){let v=new Uint32Array(t||[]),O=Math.max(Math.ceil(v.byteLength/16)*16,16),R=n.createBuffer({size:O,usage:u.UNIFORM|u.COPY_DST});n.queue.writeBuffer(R,0,v),p.push(R),c.push({binding:_.index,resource:{buffer:R}});continue}let b=bs(_.dtype),w=$u(_.dtype);if(_.packed){let v=eo(_.packedSize*b),R=_.mode==="read_write"?u.STORAGE|u.COPY_SRC|u.COPY_DST:u.STORAGE|u.COPY_DST,B=n.createBuffer({size:Math.max(v,4),usage:R,mappedAtCreation:!0}),E=new w(B.getMappedRange());for(let z of _.packed){let D=h.get(z.name),J=e[D];J&&Qd(E,J,z.dtype,z.offset)}B.unmap(),p.push(B),c.push({binding:_.index,resource:{buffer:B}});continue}let T=h.get(_.name),S=e[T],x=eo(S.length*b),A=_.mode==="read_write"?u.STORAGE|u.COPY_SRC|u.COPY_DST:u.STORAGE|u.COPY_DST,N=n.createBuffer({size:Math.max(x,4),usage:A,mappedAtCreation:!0});Qd(new w(N.getMappedRange()),S,_.dtype,0),N.unmap(),p.push(N),c.push({binding:_.index,resource:{buffer:N}})}let f=n.createBindGroup({layout:o,entries:c}),d=n.createCommandEncoder(),m=d.beginComputePass();m.setPipeline(s),m.setBindGroup(0,f),m.dispatchWorkgroups(a[0],a[1],a[2]),m.end();let g=[];for(let y=0;y<i.length;y++){let _=i[y];if(_.name==="_shapes"||_.mode!=="read_write")continue;let b=p[y],w=bs(_.dtype),T=$u(_.dtype);if(_.packed)for(let S of _.packed){let x=S.offset*w,k=eo(S.size*w),A=n.createBuffer({size:k,usage:u.MAP_READ|u.COPY_DST});d.copyBufferToBuffer(b,x,A,0,k),g.push({readBuf:A,tensorIdx:h.get(S.name),size:S.size,dtype:S.dtype,ViewCtor:T})}else{let S=b.size,x=n.createBuffer({size:S,usage:u.MAP_READ|u.COPY_DST});d.copyBufferToBuffer(b,0,x,0,S),g.push({readBuf:x,tensorIdx:h.get(_.name),size:e[h.get(_.name)].length,dtype:_.dtype,ViewCtor:T})}}n.queue.submit([d.finish()]);for(let y of g){await y.readBuf.mapAsync(Tw());let _=new y.ViewCtor(y.readBuf.getMappedRange());bw(e[y.tensorIdx],_,y.dtype,0,y.size),y.readBuf.unmap(),y.readBuf.destroy()}for(let y of p)y.destroy()}async function YN(r,e){if(typeof r.createComputePipelineAsync!="function")return;let t=[],n=new Set;for(let s of e){if(!s||n.has(s)||Xd.has(s))continue;n.add(s);let{shaderModule:o,bindGroupLayout:i,pipelineLayout:a}=UN(r,s);t.push(r.createComputePipelineAsync({layout:a,compute:{module:o,entryPoint:s.name}}).then(u=>Xd.set(s,{pipeline:u,bindGroupLayout:i}),()=>{}))}t.length>0&&await Promise.all(t)}async function K2(r,e,t){let n=await ww(),s=Fu();await YN(n,t.map(b=>b.kernel));let o=new Set;for(let b of t)for(let w of b.outputSlots)o.add(w);let i=new Array(r.numSlots).fill(null),a=new Array(r.numSlots).fill("f32");for(let b=0;b<r.numSlots;b++){let w=e[b];if(!w)continue;let T=w.dtype||"f32";a[b]=T;let S=Math.max(eo(w.data.length*bs(T)),4),x=!o.has(b),k=n.createBuffer({size:S,usage:s.STORAGE|s.COPY_DST|s.COPY_SRC,mappedAtCreation:x});x&&(Qd(new($u(T))(k.getMappedRange()),w.data,T,0),k.unmap()),i[b]=k}let u=32,p=[],c=[],h={encoder:n.createCommandEncoder(),pending:0},f=l(()=>{++h.pending>=u&&(n.queue.submit([h.encoder.finish()]),h.encoder=n.createCommandEncoder(),h.pending=0)},"maybeFlush"),d=l(b=>{let w=h.encoder,{pipeline:T,bindGroupLayout:S}=WN(n,b.kernel),x=b.inputSlots.concat(b.outputSlots),k=[],A=[],N=0;for(let B of b.kernel.metadata.bindings)if(B.name==="_shapes"){let E=new Uint32Array(b.shapeValues||[]),z=Math.max(Math.ceil(E.byteLength/16)*16,16),D=n.createBuffer({size:z,usage:s.UNIFORM|s.COPY_DST});n.queue.writeBuffer(D,0,E),p.push(D),k.push({binding:B.index,resource:{buffer:D}})}else if(B.packed){let E=bs(B.dtype),z=n.createBuffer({size:Math.max(eo(B.packedSize*E),4),usage:s.STORAGE|s.COPY_DST|s.COPY_SRC});c.push(z);let D=B.mode==="read_write";for(let J of B.packed){let j=x[J.argIndex],U=eo(J.size*E);D?A.push({slot:j,src:z,srcOff:J.offset*E,bytes:U}):w.copyBufferToBuffer(i[j],0,z,J.offset*E,U)}k.push({binding:B.index,resource:{buffer:z}})}else k.push({binding:B.index,resource:{buffer:i[x[N++]]}});let v=n.createBindGroup({layout:S,entries:k}),O=w.beginComputePass();O.setPipeline(T),O.setBindGroup(0,v);let R=b.kernel.metadata.dispatchSize;O.dispatchWorkgroups(R[0],R[1],R[2]),O.end();for(let B of A)w.copyBufferToBuffer(B.src,B.srcOff,i[B.slot],0,B.bytes);f()},"encodeStep"),m=r.scanLoops?[...r.scanLoops].sort((b,w)=>b.loopStart-w.loopStart):r.scanLoop?[r.scanLoop]:null;if(m&&m.length){let b=0;for(let w of m){for(;b<w.loopStart;b++)d(t[b]);for(let T of w.carry)h.encoder.copyBufferToBuffer(i[T.initSlot],0,i[T.a],0,T.bytes);for(let T=0;T<w.T;T++){for(let S of w.xs)h.encoder.copyBufferToBuffer(i[S.xsSlot],T*S.stepBytes,i[S.xtSlot],0,S.stepBytes);for(let S=w.loopStart;S<w.loopEnd;S++)d(t[S]);for(let S of w.ys)h.encoder.copyBufferToBuffer(i[S.ytSlot],0,i[S.ysSlot],T*S.stepBytes,S.stepBytes);for(let S of w.carry){let x=i[S.a];i[S.a]=i[S.b],i[S.b]=x}}for(let T of w.carry)h.encoder.copyBufferToBuffer(i[T.a],0,i[T.finalSlot],0,T.bytes);b=w.loopEnd}for(;b<t.length;b++)d(t[b])}else for(let b of t)d(b);let g=h.encoder,y=new Set(r.argSlots),_=[];for(let b=0;b<r.numSlots;b++){if(!i[b]||!o.has(b)||!y.has(b))continue;let w=e[b],T=Math.max(eo(w.data.length*bs(a[b])),4),S=n.createBuffer({size:T,usage:s.MAP_READ|s.COPY_DST});g.copyBufferToBuffer(i[b],0,S,0,T),_.push({rb:S,dtype:a[b],size:w.data.length,dst:w.data})}n.queue.submit([g.finish()]);for(let b of _)await b.rb.mapAsync(Tw()),bw(b.dst,new($u(b.dtype))(b.rb.getMappedRange()),b.dtype,0,b.size),b.rb.unmap(),b.rb.destroy();for(let b of i)b&&b.destroy();for(let b of p)b.destroy();for(let b of c)b.destroy()}function JN(r){let e=_w.get(r);if(e&&e.length)return e.pop();let t=Fu();return kt.createBuffer({size:r,usage:t.STORAGE|t.COPY_SRC|t.COPY_DST})}function VN(r,e){let t=_w.get(e);t||(t=[],_w.set(e,t)),t.push(r)}async function H2(){return await ww(),As||(As=kt.createCommandEncoder()),kt}function Y2(){return!!kt&&!!As}function GN(r,e,t){let n=Vc.get(r);if(n)return n.buf;let s=Fu(),o=Math.max(eo(r.length*bs(e)),4);if(t){let a=kt.createBuffer({size:o,usage:s.STORAGE|s.COPY_SRC|s.COPY_DST,mappedAtCreation:!0});return Qd(new($u(e))(a.getMappedRange()),r,e,0),a.unmap(),Vc.set(r,{buf:a,bytes:o,dtype:e,persistent:t}),a}let i=JN(o);return Vc.set(r,{buf:i,bytes:o,dtype:e,persistent:!1}),i}function Gc(r,e,t){let{pipeline:n,bindGroupLayout:s}=WN(kt,r),{bindings:o,dispatchSize:i}=r.metadata,a=HN(o),u=Fu(),p=[],c=[];for(let d of o){if(d.name==="_shapes"){let y=new Uint32Array(t||[]),_=Math.max(Math.ceil(y.byteLength/16)*16,16),b=kt.createBuffer({size:_,usage:u.UNIFORM|u.COPY_DST});kt.queue.writeBuffer(b,0,y),gw.push(b),p.push({binding:d.index,resource:{buffer:b}});continue}if(d.packed){let y=bs(d.dtype),_=Math.max(eo(d.packedSize*y),4),b=JN(_);yw.push({buf:b,bytes:_});let w=d.mode==="read_write";for(let T of d.packed){let S=GN(e[a.get(T.name)],T.dtype,!w),x=eo(T.size*y);w?c.push({pbuf:b,off:T.offset*y,sub:S,bytes:x}):As.copyBufferToBuffer(S,0,b,T.offset*y,x)}p.push({binding:d.index,resource:{buffer:b}});continue}let m=e[a.get(d.name)],g=GN(m,d.dtype,d.mode!=="read_write");p.push({binding:d.index,resource:{buffer:g}})}let h=kt.createBindGroup({layout:s,entries:p}),f=As.beginComputePass();f.setPipeline(n),f.setBindGroup(0,h),f.dispatchWorkgroups(i[0],i[1],i[2]),f.end();for(let d of c)As.copyBufferToBuffer(d.pbuf,d.off,d.sub,0,d.bytes);++mw>=W2&&(kt.queue.submit([As.finish()]),As=kt.createCommandEncoder(),mw=0)}async function ZN(){if(!kt)return;let r=Fu(),e=[];for(let[t,n]of Vc){if(n.persistent)continue;let s=kt.createBuffer({size:n.bytes,usage:r.MAP_READ|r.COPY_DST});As.copyBufferToBuffer(n.buf,0,s,0,n.bytes),e.push({rb:s,ha:t,buf:n.buf,bytes:n.bytes,dtype:n.dtype})}kt.queue.submit([As.finish()]),As=kt.createCommandEncoder(),mw=0;for(let t of gw)t.destroy();gw.length=0,await Promise.all(e.map(t=>t.rb.mapAsync(Tw())));for(let t of e)bw(t.ha,new($u(t.dtype))(t.rb.getMappedRange()),t.dtype,0,t.ha.length),t.rb.unmap(),t.rb.destroy();for(let t of yw)VN(t.buf,t.bytes);yw.length=0;for(let t of e)VN(t.buf,t.bytes),Vc.delete(t.ha)}function J2(r,e,t,n,s){let o=`${r.join(",")}|${e.join(",")}|${t}|${n}`,i=jN.get(o);if(i)return i;let a=Mn(n),u="";for(let c=r.length-1;c>=0;c--)u+=`  let idx${c} = rem % ${r[c]}u; rem = rem / ${r[c]}u; src = src + idx${c} * ${e[c]}u;
+`)}var Fv=M(()=>{"use strict";l($v,"buildSnippet")});function vd(r,e){zv.set(r,e),e.runtimeKind&&e.snippet&&Vv.set(e.runtimeKind,e.snippet)}function Gv(r){return zv.get(r)||null}function jv(r){return Vv.get(r)||null}var zv,Vv,qv=M(()=>{"use strict";Ks();Tv();Av();Ov();Mv();Ev();Cv();Pv();Fv();zv=new Map,Vv=new Map;l(vd,"registerCodegen");l(Gv,"getCodegenEntry");l(jv,"getSnippetBuilder");vd(Ot.CPU,{runtimeKind:"js",snippet:Iv,compile(r,e){return{source:new yd(e).generate(r),metadata:{kind:"js",paramCount:r.params.length}}}});vd(Ot.WASM,{runtimeKind:"wasm",snippet:Lv,compile(r,e){let t=new bd(e).generate(r),n={kind:"wasm",memoryPages:t.memoryPages,bufferOffsets:t.bufferOffsets,imports:t.imports,params:t.params,bufferMap:r.bufferMap};return t.parallel&&(n.parallel=t.parallel),{source:t.wat,metadata:n}}});vd(Ot.WEBGPU,{runtimeKind:"webgpu",snippet:Dv,compile(r,e){let t=new xd(e).generate(r);return{source:t.source,metadata:{kind:"webgpu",workgroupSize:t.workgroupSize,dispatchSize:t.dispatchSize,sharedMemBytes:t.sharedMemBytes,params:t.params,bindings:t.bindings}}}});vd(Ot.CUDA,{runtimeKind:"cuda",snippet:$v,compile(r,e,t){if(t&&t.matmulBackend==="cublas"&&r.cublasInfo)return{source:"",metadata:{kind:"cuda",cublas:r.cublasInfo,outputIndices:[r.cublasInfo.cIdx]}};let n=new _d(e).generate(r);return{source:n.source,metadata:{kind:"cuda",blockDim:n.blockDim,gridDim:n.gridDim,sharedMemBytes:n.sharedMemBytes,params:n.params,outputIndices:n.outputIndices,scratch:n.scratch}}}})});function Uv(r){let e=[],t=[r.body];for(;t.length>0;){let m=t.pop();if(m){if(m.type==="BlockNode"){e.push(m),t.push(m.body);continue}if(m.body&&t.push(m.body),m.stmts)for(let g of m.stmts)t.push(g);m.thenBody&&t.push(m.thenBody),m.elseBody&&t.push(m.elseBody)}}let n=null;for(let m of e)if(m.name.includes("matmul"))m.reads.length>=2&&m.writes.length>=1&&(n=m);else return null;if(!n)return null;let s=n.reads[0].buffer,o=n.reads[1].buffer,i=n.writes[0].buffer;if(s.dtype!=="f32"||o.dtype!=="f32"||i.dtype!=="f32"||s.shape.length!==2||o.shape.length!==2||i.shape.length!==2)return null;let a=i.shape[0],u=i.shape[1],p=s.shape[1];if(![a,u,p].every(m=>typeof m=="number"))return null;let c=[];for(let[,m]of r.bufferMap)c.push(m.name);let h=c.indexOf(s.name),f=c.indexOf(o.name),d=c.indexOf(i.name);return h<0||f<0||d<0?null:{M:a,N:u,K:p,aIdx:h,bIdx:f,cIdx:d}}var Ic,Hs,Ec=M(()=>{"use strict";qv();l(Uv,"detectPureMatmul");Ic=class{static{l(this,"CompiledKernel")}constructor(e,t,n,s={}){this.name=e,this.source=t,this.target=n,this.metadata=s}snippet(){let e=jv(this.metadata.kind);if(!e)throw new Error(`No snippet for kind: ${this.metadata.kind}`);return e(this)}},Hs=class{static{l(this,"BackendPipeline")}constructor(e,t={}){this.target=e,this.matmulBackend=t.matmulBackend||"native",this.context=t.context||null}compile(e){let t=this.context&&this.context.getCodegenEntry(this.target.kind)||Gv(this.target.kind);if(!t)throw new Error(`Unsupported target kind: ${this.target.kind}`);let{source:n,metadata:s}=t.compile(e,this.target,this);return new Ic(e.name,n,this.target,s)}compileAll(e){return e.map(t=>this.compile(t))}}});function Wv(r,e){Kv.set(r,e)}function Hv(r){return Kv.get(r)||null}var Kv,rw=M(()=>{"use strict";Kv=new Map;l(Wv,"registerMeasurer");l(Hv,"getMeasurer")});var Yv={};nt(Yv,{runWasmParallel:()=>S2});async function S2(r,e,t,n,s){let{exports:o,memory:i,bufferOffsets:a}=r,u=o[e],p=[...a.values()],c=Math.min(p.length,t.length);for(let f=0;f<c;f++){let d=t[f];ArrayBuffer.isView(d)&&new d.constructor(i.buffer,p[f],d.length).set(d)}let h=p.slice(0,c);if(n)for(let f of n)h.push(f);h.push(0,s.extent),u(...h);for(let f=0;f<c;f++){let d=t[f];ArrayBuffer.isView(d)&&d.set(new d.constructor(i.buffer,p[f],d.length))}}var Jv=M(()=>{"use strict";l(S2,"runWasmParallel")});var Nd={};nt(Nd,{deviceClipGradNorm:()=>A2,measureCudaKernel:()=>v2,runCudaKernel:()=>k2,runCudaPlan:()=>N2});function k2(){throw new Error("CUDA runtime is not available in the browser")}function v2(){throw new Error("CUDA runtime is not available in the browser")}function N2(){throw new Error("CUDA runtime is not available in the browser")}function A2(){throw new Error("CUDA runtime is not available in the browser")}var Ad=M(()=>{"use strict";l(k2,"runCudaKernel");l(v2,"measureCudaKernel");l(N2,"runCudaPlan");l(A2,"deviceClipGradNorm")});function ii(){return Zv}function O2(r){Zv=r}var Qi,ks,Od,sw,Ys,Au,Zv,ar=M(()=>{"use strict";Nn();Qi=class{static{l(this,"Device")}type;index;constructor(e,t=0){this.type=e,this.index=t}dispatchKey(){return ja(this.type)}equals(e){return this.type===e.type&&this.index===e.index}hash(){let e=2166136261;for(let t=0;t<this.type.length;t++)e=(e^this.type.charCodeAt(t))*16777619&2147483647;return e=(e^this.index)*16777619&2147483647,e}toString(){return this.index===0?this.type:`${this.type}:${this.index}`}},ks=new Qi("cpu"),Od=new Qi("gpu"),sw=new Qi("wasm"),Ys=new Qi("webgpu"),Au=new Qi("meta"),Zv=ks;l(ii,"getDefaultDevice");l(O2,"setDefaultDevice")});function Xv(r){let e=Qv.get(r)??null;return e||(e=Nt.findOp(r),e&&Qv.set(r,e)),e||null}var Qv,eN=M(()=>{"use strict";$i();Qv=new Map;l(Xv,"getHandle")});function vs(r,e){return r<0?e+r:r}function xn(r){let e=r.length,t=new Array(e),n=1;for(let s=e-1;s>=0;s--)t[s]=n,n*=r[s];return t}function cn(r){let e=1;for(let t=0;t<r.length;t++)e*=r[t];return e}function ow(r,e){let t=r.length;if(t===0)return!0;let n=1;for(let s=t-1;s>=0;s--){let o=r[s];if(o===0)return!0;if(o!==1&&e[s]!==n)return!1;n*=o}return!0}function Ou(r,e){let t=r.length,n=e.length,s=Math.max(t,n),o=new Array(s);for(let i=0;i<s;i++){let a=i<t?r[t-1-i]:1,u=i<n?e[n-1-i]:1;if(a===u)o[s-1-i]=a;else if(a===1)o[s-1-i]=u;else if(u===1)o[s-1-i]=a;else return null}return o}function Bd(r,e){let t=r.length,n=e.length;if(t===1&&n===1)return[];let s=t===1?[1,r[0]]:[...r],o=n===1?[e[0],1]:[...e],i=s.length,a=o.length,u=Ou(s.slice(0,i-2),o.slice(0,a-2));if(u===null)return null;let p=[...u,s[i-2],o[a-1]],c=new Set;return n===1&&c.add(p.length-1),t===1&&c.add(p.length-2),p.filter((h,f)=>!c.has(f))}function tN(r,e,t){let n=cn(r),s=-1,o=1,i=new Array(t.length);for(let a=0;a<t.length;a++){let u=t[a];if(u===-1){if(s!==-1)return null;s=a,i[a]=-1}else i[a]=u,o*=u}if(s!==-1){if(o===0||(i[s]=n/o|0,i[s]*o!==n))return null}else if(cn(i)!==n)return null;return ow(r,e)?{sizes:i,strides:xn(i),needsCopy:!1}:{sizes:i,strides:xn(i),needsCopy:!0}}var jr=M(()=>{"use strict";l(vs,"normalizeAxis");l(xn,"computeStrides");l(cn,"computeNumel");l(ow,"isContiguous");l(Ou,"broadcastShapes");l(Bd,"matmulOutputShape");l(tN,"inferReshape")});function nN(r,e){if(r.length<4)return!1;let t=r.length,n=1,s=t===4?B2:R2;if(!s)return!1;for(let o=t-1;o>=0;o--){let i=s[o];if(i===void 0||r[i]===0||e[i]!==n)return!1;n*=r[i]}return!0}var B2,R2,iw=M(()=>{"use strict";B2=Object.freeze([0,2,3,1]),R2=Object.freeze([0,2,3,4,1]);l(nN,"isChannelsLast")});var Bt,Oo=M(()=>{"use strict";Nn();jr();iw();iw();Bt=class{static{l(this,"TensorImpl")}_storage;_storageOffset;_sizes;_strides;_dtype;_device;_numel;_keySet;_autogradMeta;_version;_contiguousCache;constructor(e,t,n,s,o,i){this._storage=e,this._storageOffset=t,this._sizes=Object.freeze([...n]),this._strides=Object.freeze(s?[...s]:xn(n)),this._dtype=o,this._device=i,this._numel=cn(n),this._keySet=Pr.fromKey(ja(i.type)),this._autogradMeta=null,this._version=0,this._contiguousCache=null}get storage(){return this._storage}get storageOffset(){return this._storageOffset}size(e){let t=e<0?this._sizes.length+e:e;return this._sizes[t]}stride(e){let t=e<0?this._strides.length+e:e;return this._strides[t]}sizes(){return this._sizes}strides(){return this._strides}dim(){return this._sizes.length}numel(){return this._numel}get dtype(){return this._dtype}get device(){return this._device}isContiguous(e){return!e||e==="contiguous"?(this._contiguousCache===null&&(this._contiguousCache=ow(this._sizes,this._strides)),this._contiguousCache):e==="channels_last"?nN(this._sizes,this._strides):!1}setSizesAndStrides(e,t){this._sizes=Object.freeze([...e]),this._strides=Object.freeze(t?[...t]:xn(e)),this._numel=cn(e),this._contiguousCache=null}setStorageOffset(e){this._storageOffset=e}bumpVersion(){this._version++}get version(){return this._version}get autogradMeta(){return this._autogradMeta}setAutogradMeta(e){this._autogradMeta=e,this._updateKeySet()}keySet(){return this._keySet}addKeyToSet(e){this._keySet=this._keySet.add(e)}removeKeyFromSet(e){this._keySet=this._keySet.remove(e)}_updateKeySet(){let e=Pr.fromKey(ja(this._device.type));if(this._autogradMeta&&this._autogradMeta.requiresGrad){let t=ja(this._device.type);e=e.add(RS(t))}this._keySet=e}get isMeta(){return this._storage&&this._storage.isMeta}shallowCopyFrom(e){this._storage=e._storage,this._storageOffset=e._storageOffset,this._sizes=e._sizes,this._strides=e._strides,this._dtype=e._dtype,this._device=e._device,this._numel=e._numel,this._keySet=e._keySet,this._contiguousCache=null}}});var Bu,aw=M(()=>{"use strict";Bu=class{static{l(this,"AutogradMeta")}_grad;_gradFn;_outputNr;_requiresGrad;_retainGrad;_gradAccumulator;_versionAtCreation;constructor(){this._grad=null,this._gradFn=null,this._outputNr=0,this._requiresGrad=!1,this._retainGrad=!1,this._gradAccumulator=null,this._versionAtCreation=0}get grad(){return this._grad}set grad(e){this._grad=e}get gradFn(){return this._gradFn}setGradFn(e,t){this._gradFn=e,this._outputNr=t??0}get outputNr(){return this._outputNr}get requiresGrad(){return this._requiresGrad}set requiresGrad(e){this._requiresGrad=e}get retainGrad(){return this._retainGrad}set retainGrad(e){this._retainGrad=e}get isLeaf(){return this._gradFn===null}get versionAtCreation(){return this._versionAtCreation}set versionAtCreation(e){this._versionAtCreation=e}getGradAccumulator(){if(this._gradAccumulator){let e=this._gradAccumulator instanceof WeakRef?this._gradAccumulator.deref():this._gradAccumulator;if(e)return e}return null}setGradAccumulator(e){this._gradAccumulator=typeof WeakRef<"u"?new WeakRef(e):e}clearGrad(){this._grad=null}clearGradFn(){this._gradFn=null,this._outputNr=0}}});function sN(r,e,t,n,s,o){let i=e[s];if(s===e.length-1){let u=new Array(i);if(o)for(let p=0;p<i;p++)u[p]=md(o,r[n+p*t[s]]);else for(let p=0;p<i;p++)u[p]=r[n+p*t[s]];return u}let a=new Array(i);for(let u=0;u<i;u++)a[u]=sN(r,e,t,n+u*t[s],s+1,o);return a}function M2(){if(!uw)throw new Error("Autograd engine not initialized. Import autograd/engine.js first.");return uw}function oN(r){uw=r}var rN,We,uw,jn=M(()=>{"use strict";Oo();aw();xu();rN=new Set(["f16","bf16","i64"]),We=class r{static{l(this,"Tensor")}_impl;constructor(e){this._impl=e}get impl(){return this._impl}get shape(){return this._impl.sizes()}get strides(){return this._impl.strides()}get dtype(){return this._impl.dtype}get device(){return this._impl.device}get ndim(){return this._impl.dim()}get rank(){return this._impl.dim()}get numel(){return this._impl.numel()}get length(){return this._impl.numel()}get isContiguous(){return this._impl.isContiguous()}get dispatchKeySet(){return this._impl.keySet()}get storage(){return this._impl.storage}get storageOffset(){return this._impl.storageOffset}get data(){let e=this._impl.storage;if(!e||e.isMeta)return null;let t=e.data,n=this._impl.storageOffset;return t&&n===0&&this._impl.isContiguous()&&t.length===this.numel,t}get requiresGrad(){let e=this._impl.autogradMeta;return e?e.requiresGrad:!1}get gradFn(){let e=this._impl.autogradMeta;return e?e.gradFn:null}get grad(){let e=this._impl.autogradMeta;return e?e.grad:null}set grad(e){this._ensureAutogradMeta().grad=e}get isLeaf(){let e=this._impl.autogradMeta;return e?e.isLeaf:!0}get version(){return this._impl.version}requiresGrad_(e=!0){let t=this._ensureAutogradMeta();return t.requiresGrad=e,this._impl._updateKeySet(),this}retainGrad(){return this._ensureAutogradMeta().retainGrad=!0,this}detach(){let e=new Bt(this._impl.storage,this._impl.storageOffset,this._impl.sizes(),this._impl.strides(),this._impl.dtype,this._impl.device);return new r(e)}item(){if(this.numel!==1)throw new Error(`item() requires tensor with exactly 1 element, got ${this.numel}`);let t=this._impl.storage.data[this._impl.storageOffset];return rN.has(this._impl.dtype)?md(this._impl.dtype,t):t}toArray(){let e=this.shape,t=this.strides,n=this._impl.storage.data,s=this._impl.storageOffset,o=rN.has(this._impl.dtype)?this._impl.dtype:null;return e.length===0?o?md(o,n[s]):n[s]:sN(n,e,t,s,0,o)}toString(){return`Tensor(shape=[${this.shape.join(", ")}], dtype=${this.dtype}, device=${this.device})`}*[Symbol.iterator](){let e=this.shape[0];if(e===void 0)throw new Error("Cannot iterate over a 0-d tensor");for(let t=0;t<e;t++)yield this._select(0,t)}_select(e,t){let n=this._impl.sizes(),s=this._impl.strides(),o=[],i=[];for(let p=0;p<n.length;p++)p!==e&&(o.push(n[p]),i.push(s[p]));let a=this._impl.storageOffset+t*s[e],u=new Bt(this._impl.storage,a,o,i,this._impl.dtype,this._impl.device);return new r(u)}_ensureAutogradMeta(){if(!this._impl.autogradMeta){let e=new Bu;e.versionAtCreation=this._impl.version,this._impl.setAutogradMeta(e)}return this._impl.autogradMeta}backward(e){let{backward:t}=M2();t(this,e)}};l(sN,"_toNestedArray");uw=null;l(M2,"require_autograd_engine");l(oN,"setAutogradEngine")});function Dt(r){return E2[r]||Float32Array}function Rd(r,e){if(r===e)return r;let t=Ru.get(r)??0,n=Ru.get(e)??0;return t>=n?r:e}function Xe(r,e){let t=uN.get(r);t||(t=new Set,uN.set(r,t)),t.add(e)}function Xi(r){return I2(r)}var W,I2,E2,iN,aN,Ru,uN,ts=M(()=>{"use strict";Oe();W=zn,I2=Gs,E2=Object.freeze({[W.F16]:Uint16Array,[W.BF16]:Uint16Array,[W.F32]:Float32Array,[W.F64]:Float64Array,[W.I8]:Int8Array,[W.I16]:Int16Array,[W.I32]:Int32Array,[W.I64]:BigInt64Array,[W.UI8]:Uint8Array,[W.BOOL]:Uint8Array,[W.INDEX]:Int32Array});l(Dt,"typedArrayCtor");iN=[W.F16,W.BF16,W.F32,W.F64],aN=[W.UI8,W.I8,W.I16,W.I32,W.I64],Ru=new Map;for(let r=0;r<iN.length;r++)Ru.set(iN[r],100+r);for(let r=0;r<aN.length;r++)Ru.set(aN[r],r);Ru.set(W.BOOL,-1);Ru.set(W.INDEX,50);l(Rd,"resultDtype");uN=new Map;l(Xe,"_addCastPair");Xe(W.BOOL,W.I8);Xe(W.BOOL,W.I16);Xe(W.BOOL,W.I32);Xe(W.BOOL,W.I64);Xe(W.BOOL,W.F16);Xe(W.BOOL,W.F32);Xe(W.BOOL,W.F64);Xe(W.UI8,W.I16);Xe(W.UI8,W.I32);Xe(W.UI8,W.I64);Xe(W.UI8,W.F16);Xe(W.UI8,W.F32);Xe(W.UI8,W.F64);Xe(W.I8,W.I16);Xe(W.I8,W.I32);Xe(W.I8,W.I64);Xe(W.I8,W.F16);Xe(W.I8,W.F32);Xe(W.I8,W.F64);Xe(W.I16,W.I32);Xe(W.I16,W.I64);Xe(W.I16,W.F32);Xe(W.I16,W.F64);Xe(W.I32,W.I64);Xe(W.I32,W.F64);Xe(W.F16,W.F32);Xe(W.F16,W.F64);Xe(W.BF16,W.F32);Xe(W.BF16,W.F64);Xe(W.BOOL,W.BF16);Xe(W.UI8,W.BF16);Xe(W.I8,W.BF16);Xe(W.F32,W.F64);l(Xi,"dtypeSize")});function lw(r){return lN.get(r)||lN.get("cpu")}var Md,Id,lN,Lc,cN=M(()=>{"use strict";ts();ar();Md=class{static{l(this,"CPUAllocator")}allocate(e,t){let n=Dt(t);return new n(e)}free(e){}},Id=class{static{l(this,"MetaAllocator")}allocate(){return null}free(e){}},lN=new Map([["cpu",new Md],["wasm",new Md],["meta",new Id],["lazy",new Id]]);l(lw,"getAllocator");Lc=class r{static{l(this,"StorageImpl")}static#e=null;_data;_nbytes;_device;_allocator;_refCount;static setHostReadHook(e){r.#e=e}constructor(e,t,n,s){this._data=e,this._nbytes=t,this._device=n,this._allocator=s,this._refCount=1}static allocate(e,t,n){let s=lw(n.type),o=Xi(t),i=Math.max(Math.ceil(e/o),1),a=s.allocate(i,t);return new r(a,e,n,s)}static fromData(e,t){let n=e?e.byteLength:0,s=lw(t.type);return new r(e,n,t,s)}retain(){return this._refCount++,this}release(){this._refCount--,this._refCount===0&&(this._allocator&&this._data&&this._allocator.free(this._data),this._data=null)}get refCount(){return this._refCount}get data(){return r.#e&&this._data&&r.#e(this._data),this._data}get rawData(){return this._data}get nbytes(){return this._nbytes}get device(){return this._device}get isValid(){return this._refCount>0&&this._data!==null}get isMeta(){return this._data===null}resize(e,t){if(e<=this._nbytes&&this._data)return;let n=Xi(t),s=Math.max(Math.ceil(e/n),1),o=this._allocator.allocate(s,t);if(this._data&&o){let i=Math.min(this._data.length,o.length);for(let a=0;a<i;a++)o[a]=this._data[a]}this._allocator&&this._data&&this._allocator.free(this._data),this._data=o,this._nbytes=e}clone(){let e=lw(this._device.type),t=null;if(this._data){r.#e&&r.#e(this._data);let n=this._data.constructor;t=new n(this._data.length),t.set(this._data)}return new r(t,this._nbytes,this._device,e)}}});var En,ea=M(()=>{"use strict";cN();En=class r{static{l(this,"Storage")}_impl;constructor(e){this._impl=e}static allocate(e,t,n){return new r(Lc.allocate(e,t,n))}static fromData(e,t){return new r(Lc.fromData(e,t))}get impl(){return this._impl}get data(){return this._impl.data}get rawData(){return this._impl.rawData}get nbytes(){return this._impl.nbytes}get device(){return this._impl.device}get isValid(){return this._impl.isValid}get isMeta(){return this._impl.isMeta}retain(){return this._impl.retain(),this}release(){this._impl.release()}clone(){return new r(this._impl.clone())}resize(e,t){this._impl.resize(e,t)}}});function Pt(r,e){let t=e?.dtype??W.F32,n=e?.device??ii(),s=e?.requiresGrad??!1;if(ArrayBuffer.isView(r)&&"length"in r)return L2(r,e?.shape,t,n,s);if(Array.isArray(r)){let{flat:o,shape:i}=D2(r),a=e?.shape??i;return C2(o,a,t,n,s)}if(typeof r=="number")return pN(r,t,n,s);throw new Error("Unsupported data type for tensor()")}function ta(r,e,t,n){let s=n?.device??ii(),o=n?.requiresGrad??!1,i=xn(e),a=En.fromData(r,s),u=n?.offset??0,p=new Bt(a,u,e,i,t,s),c=new We(p);return o&&c.requiresGrad_(!0),c}function Dc(r,e){let t=e?.dtype??W.F32,n=e?.device??ii();return pN(r,t,n,e?.requiresGrad??!1)}function pN(r,e,t,n){let s=Dt(e),o=new s(1),i=o;i[0]=cw.has(e)?dd(e,r):r;let a=En.fromData(o,t),u=new Bt(a,0,[],[],e,t),p=new We(u);return n&&p.requiresGrad_(!0),p}function L2(r,e,t,n,s){let o=Dt(t),i=new o(r.length),a=i;if(cw.has(t))for(let d=0;d<r.length;d++)a[d]=dd(t,r[d]);else for(let d=0;d<r.length;d++)a[d]=r[d];let u=e??[r.length],p=xn(u),c=En.fromData(i,n),h=new Bt(c,0,u,p,t,n),f=new We(h);return s&&f.requiresGrad_(!0),f}function C2(r,e,t,n,s){let o=Dt(t),i=new o(r.length),a=i;if(cw.has(t))for(let f=0;f<r.length;f++)a[f]=dd(t,r[f]);else for(let f=0;f<r.length;f++)a[f]=r[f];let u=xn(e),p=En.fromData(i,n),c=new Bt(p,0,e,u,t,n),h=new We(c);return s&&h.requiresGrad_(!0),h}function D2(r){let e=[],t=r;for(;Array.isArray(t);)e.push(t.length),t=t[0];let n=[];return fN(r,n,e,0),{flat:n,shape:e}}function fN(r,e,t,n){if(n===t.length-1){for(let s=0;s<r.length;s++)e.push(r[s]);return}for(let s=0;s<r.length;s++)fN(r[s],e,t,n+1)}var cw,qr=M(()=>{"use strict";jn();Oo();ea();ts();ar();jr();xu();l(Pt,"tensor");l(ta,"fromBuffer");l(Dc,"scalar");cw=new Set(["f16","bf16","i64"]);l(pN,"_fromScalar");l(L2,"_fromTypedArray");l(C2,"_fromFlatArray");l(D2,"_flattenNested");l(fN,"_recurFlatten")});var Ft={};nt(Ft,{_dispatch:()=>ot,abs:()=>Pc,add:()=>ae,argmax:()=>Ud,argmin:()=>kN,argsort:()=>LN,batch_norm:()=>fw,broadcast_in_dim:()=>ON,cat:()=>sa,ceil:()=>wN,chunk:()=>BN,clamp:()=>Vd,clone:()=>AN,contiguous:()=>ci,conv2d:()=>hw,cos:()=>mN,cumsum:()=>IN,div:()=>$t,dot:()=>NN,embedding:()=>z2,eq:()=>$c,erf:()=>gN,erfc:()=>yN,exp:()=>Kr,expand:()=>Lu,fill:()=>F2,flip:()=>MN,floor:()=>bN,gamma:()=>Cd,gather:()=>jd,ge:()=>zd,gelu:()=>Dd,gt:()=>Iu,index_select:()=>li,layer_norm:()=>pw,le:()=>Fd,lgamma:()=>_N,log:()=>Zs,log_softmax:()=>Hd,lt:()=>xN,matmul:()=>lt,max:()=>Qs,maximum:()=>ai,mean:()=>lr,min:()=>SN,minimum:()=>Ld,mul:()=>ee,narrow:()=>Un,ne:()=>TN,neg:()=>Ur,one_hot:()=>Gd,pad:()=>na,permute:()=>Eu,pool2d:()=>Yd,pow:()=>Js,prod:()=>vN,relu:()=>Mu,remainder:()=>$2,repeat:()=>Du,reshape:()=>Cn,roll:()=>RN,rsqrt:()=>hN,scatter:()=>qd,scatter_add:()=>ra,select:()=>Ut,sigmoid:()=>ur,sign:()=>$d,silu:()=>Pd,sin:()=>dN,slice:()=>Kd,softmax:()=>Ro,sort:()=>EN,split:()=>Bo,sqrt:()=>ui,squeeze:()=>Cu,stack:()=>Wr,sub:()=>Pe,sum:()=>ht,tanh:()=>Ns,tile:()=>Wd,topk:()=>CN,transpose:()=>gt,unsqueeze:()=>Sr,where:()=>qn});function P2(r){return typeof r=="object"&&r!==null&&"_impl"in r}function Ln(r,e){return P2(r)?r:Dc(r,{dtype:e.dtype,device:e.device})}function ot(r,...e){let t=Xv(r);if(!t)return Nt.callOp(r,...e);let n=pb(e,t.schema);return Nt.dispatch(t,n,...e)}function se(r,...e){return ot(r,...e)}function Ed(r,...e){return ot(r,...e)}function ae(r,e){return se("add",r,Ln(e,r))}function Pe(r,e){return se("sub",r,Ln(e,r))}function ee(r,e){return se("mul",r,Ln(e,r))}function $t(r,e){return se("div",r,Ln(e,r))}function Ur(r){return se("neg",r)}function Js(r,e){return se("pow",r,Ln(e,r))}function $2(r,e){return se("rem",r,Ln(e,r))}function ai(r,e){return se("maximum",r,Ln(e,r))}function Ld(r,e){return se("minimum",r,Ln(e,r))}function Kr(r){return se("exp",r)}function Zs(r){return se("log",r)}function ui(r){return se("sqrt",r)}function hN(r){return se("rsqrt",r)}function Pc(r){return se("abs",r)}function dN(r){return se("sin",r)}function mN(r){return se("cos",r)}function Ns(r){return se("tanh",r)}function gN(r){return se("erf",r)}function yN(r){return se("erfc",r)}function _N(r){return se("lgamma",r)}function Cd(r){return se("gamma",r)}function ur(r){return se("sigmoid",r)}function Mu(r){return se("relu",r)}function Dd(r){return se("gelu",r)}function Pd(r){return se("silu",r)}function $d(r){return se("sign",r)}function bN(r){return se("floor",r)}function wN(r){return se("ceil",r)}function $c(r,e){return se("eq",r,Ln(e,r))}function TN(r,e){return se("ne",r,Ln(e,r))}function xN(r,e){return se("lt",r,Ln(e,r))}function Fd(r,e){return se("le",r,Ln(e,r))}function Iu(r,e){return se("gt",r,Ln(e,r))}function zd(r,e){return se("ge",r,Ln(e,r))}function qn(r,e,t){return se("where",r,e,t)}function Vd(r,e,t){return se("clamp",r,Ln(e,r),Ln(t,r))}function na(r,e,t,n=0){return se("pad",r,Ln(n,r),e,t)}function Gd(r,e){return se("one_hot",r,e)}function li(r,e,t){return se("index_select",r,t,e)}function jd(r,e,t){return se("gather",r,t,e)}function ra(r,e,t,n){return se("scatter_add",r,t,n,e)}function qd(r,e,t,n){return se("scatter",r,e,t,n)}function ht(r,e,t){return se("sum",r,e,t)}function lr(r,e,t){return se("mean",r,e,t)}function Qs(r,e,t){return se("max",r,e,t)}function SN(r,e,t){return se("min",r,e,t)}function Ud(r,e,t){return se("argmax",r,e,t)}function kN(r,e,t){return se("argmin",r,e,t)}function vN(r,e,t){return se("prod",r,e,t)}function lt(r,e){let t=dw();if(t){let n=t(r,e);if(n!==null)return n}return se("matmul",r,e)}function NN(r,e){return se("dot",r,e)}function sa(r,e){return se("cat",r,e)}function Wr(r,e){return se("stack",r,e)}function AN(r){return se("clone",r)}function F2(r,e){return se("fill",r,e)}function Cn(r,e){return se("reshape",r,e)}function gt(r,e,t){return se("transpose",r,e,t)}function Eu(r,e){return se("permute",r,e)}function ON(r,e,t){return se("broadcast_in_dim",r,e,t)}function Lu(r,e){return se("expand",r,e)}function Kd(r,e,t,n=null,s=1){return se("slice",r,e,t,n,s)}function Sr(r,e){return se("unsqueeze",r,e)}function Cu(r,e=null){return se("squeeze",r,e)}function Un(r,e,t,n){return se("narrow",r,e,t,n)}function Ut(r,e,t){return se("select",r,e,t)}function ci(r){return se("contiguous",r)}function Du(r,e){return se("repeat",r,e)}function Wd(r,e){return se("tile",r,e)}function Bo(r,e,t=0){if(typeof e!="number")return Ed("split",r,e,t);let n=r.shape.length,s=t<0?n+t:t,o=r.shape[s],i=[];for(let a=0;a<o;a+=e)i.push(Math.min(e,o-a));return Ed("split",r,i,t)}function BN(r,e,t=0){return Ed("chunk",r,e,t)}function RN(r,e,t=0){return se("roll",r,e,t)}function MN(r,e){return se("flip",r,Array.isArray(e)?e:[e])}function IN(r,e=0){return se("cumsum",r,e)}function EN(r,e=-1,t=!1){return se("sort",r,e,t)}function LN(r,e=-1,t=!1){return se("argsort",r,e,t)}function CN(r,e,t=-1,n=!0){return Ed("topk",r,e,t,n)}function Ro(r,e){return se("softmax",r,e)}function Hd(r,e){return se("log_softmax",r,e)}function pw(r,e,t,n,s){return se("layer_norm",r,e,t,n,s)}function fw(r,e,t,n,s,o,i){return se("batch_norm",r,e,t,n,s,o,i)}function hw(r,e,t,n,s,o){return se("conv2d",r,e,t,n,s,o)}function Yd(r,e,t,n,s){return se("pool2d",r,e,t,n,s)}function z2(r,e){return se("embedding",r,e)}var be=M(()=>{"use strict";$i();eN();qr();Kn();l(P2,"hasTensorImpl");l(Ln,"_asTensor");l(ot,"_dispatch");l(se,"_dispatchTensor");l(Ed,"_dispatchTensorArray");l(ae,"add");l(Pe,"sub");l(ee,"mul");l($t,"div");l(Ur,"neg");l(Js,"pow");l($2,"remainder");l(ai,"maximum");l(Ld,"minimum");l(Kr,"exp");l(Zs,"log");l(ui,"sqrt");l(hN,"rsqrt");l(Pc,"abs");l(dN,"sin");l(mN,"cos");l(Ns,"tanh");l(gN,"erf");l(yN,"erfc");l(_N,"lgamma");l(Cd,"gamma");l(ur,"sigmoid");l(Mu,"relu");l(Dd,"gelu");l(Pd,"silu");l($d,"sign");l(bN,"floor");l(wN,"ceil");l($c,"eq");l(TN,"ne");l(xN,"lt");l(Fd,"le");l(Iu,"gt");l(zd,"ge");l(qn,"where");l(Vd,"clamp");l(na,"pad");l(Gd,"one_hot");l(li,"index_select");l(jd,"gather");l(ra,"scatter_add");l(qd,"scatter");l(ht,"sum");l(lr,"mean");l(Qs,"max");l(SN,"min");l(Ud,"argmax");l(kN,"argmin");l(vN,"prod");l(lt,"matmul");l(NN,"dot");l(sa,"cat");l(Wr,"stack");l(AN,"clone");l(F2,"fill");l(Cn,"reshape");l(gt,"transpose");l(Eu,"permute");l(ON,"broadcast_in_dim");l(Lu,"expand");l(Kd,"slice");l(Sr,"unsqueeze");l(Cu,"squeeze");l(Un,"narrow");l(Ut,"select");l(ci,"contiguous");l(Du,"repeat");l(Wd,"tile");l(Bo,"split");l(BN,"chunk");l(RN,"roll");l(MN,"flip");l(IN,"cumsum");l(EN,"sort");l(LN,"argsort");l(CN,"topk");l(Ro,"softmax");l(Hd,"log_softmax");l(pw,"layer_norm");l(fw,"batch_norm");l(hw,"conv2d");l(Yd,"pool2d");l(z2,"embedding")});var $N={};nt($N,{create:()=>DN,default:()=>V2,globals:()=>PN});var DN,PN,V2,FN=M(()=>{"use strict";DN=l(()=>{throw new Error('mlfw: the "webgpu" npm package is not available in the browser; use navigator.gpu instead')},"create"),PN=void 0,V2={create:DN,globals:PN}});var XN={};nt(XN,{ensureWebGPUEager:()=>H2,flushWebGPUEager:()=>ZN,instantiateWebGPU:()=>q2,prewarmPipelines:()=>YN,recordWebGPUEager:()=>Gc,resetDevice:()=>G2,runWebGPUKernel:()=>U2,runWebGPUPlan:()=>K2,webgpuEagerOp:()=>Z2,webgpuEagerReady:()=>Y2,webgpuRNN:()=>tC});function $u(r){switch(Mn(r)){case"i32":return Int32Array;case"u32":return Uint32Array;case"f16":return Uint16Array;default:return Float32Array}}function Qd(r,e,t,n){if(t==="bf16")for(let s=0;s<e.length;s++)r[n+s]=fd(e[s]);else if(t==="i64")for(let s=0;s<e.length;s++)r[n+s]=Number(BigInt.asIntN(32,e[s]));else r.set(e,n)}function bw(r,e,t,n,s){if(t==="bf16")for(let o=0;o<s;o++)r[o]=hd(e[n+o]);else if(t==="i64")for(let o=0;o<s;o++)r[o]=BigInt(e[n+o]);else r.set(e.subarray(n,n+s))}function eo(r){return Math.ceil(r/4)*4}async function ww(){return kt||Fc||(Fc=(async()=>{let r=typeof navigator<"u"&&navigator.gpu;if(!r)try{let i=await Promise.resolve().then(()=>(FN(),$N));qN=i,r=i.create([]),i.globals&&(zc=i.globals.GPUBufferUsage,Jd=i.globals.GPUMapMode,Zd=i.globals.GPUShaderStage)}catch{throw new Error('WebGPU not available: install the "webgpu" npm package or run in a browser with WebGPU support')}!zc&&typeof GPUBufferUsage<"u"&&(zc=GPUBufferUsage,Jd=GPUMapMode,Zd=GPUShaderStage);let e=await r.requestAdapter();if(!e)throw new Error("WebGPU: no adapter found");let t=e.limits||{},n={},s=["maxStorageBuffersPerShaderStage","maxStorageBufferBindingSize","maxBufferSize","maxBindingsPerBindGroup"];for(let i of s)t[i]!==void 0&&(n[i]=t[i]);let o=[];return e.features&&e.features.has("shader-f16")&&o.push("shader-f16"),kt=await e.requestDevice({requiredLimits:n,requiredFeatures:o}),!zN&&typeof process<"u"&&process.on&&(zN=!0,process.on("exit",()=>{kt&&(kt.destroy(),kt=null)})),kt})(),Fc)}function Fu(){return zc}function Tw(){return Jd.READ}function G2(){kt&&(kt.destroy(),kt=null),Fc=null,zc=null,Jd=null,Zd=null,qN=null}function j2(r){return r.name==="_shapes"?"uniform":r.mode==="read_write"?"storage":"read-only-storage"}function UN(r,e){let t=r.createShaderModule({code:e.source}),n=[];for(let i of e.metadata.bindings)n.push({binding:i.index,visibility:Zd.COMPUTE,buffer:{type:j2(i)}});let s=r.createBindGroupLayout({entries:n}),o=r.createPipelineLayout({bindGroupLayouts:[s]});return{shaderModule:t,bindGroupLayout:s,pipelineLayout:o}}function KN(r,e){let{shaderModule:t,bindGroupLayout:n,pipelineLayout:s}=UN(r,e);return{pipeline:r.createComputePipeline({layout:s,compute:{module:t,entryPoint:e.name}}),bindGroupLayout:n}}function WN(r,e){let t=Xd.get(e);return t||(t=KN(r,e),Xd.set(e,t)),t}async function q2(r){let e=await ww(),{pipeline:t,bindGroupLayout:n}=KN(e,r);return{device:e,pipeline:t,bindGroupLayout:n,kernel:r,workgroupSize:r.metadata.workgroupSize,dispatchSize:r.metadata.dispatchSize,bindings:r.metadata.bindings}}function HN(r){let e=new Map,t=0;for(let n of r)if(n.name!=="_shapes")if(n.packed)for(let s of n.packed)e.set(s.name,s.argIndex!==void 0?s.argIndex:t++);else e.set(n.name,n.argIndex!==void 0?n.argIndex:t++);return e}async function U2(r,e,t){let{device:n,pipeline:s,bindGroupLayout:o,bindings:i,dispatchSize:a}=r,u=Fu(),p=[],c=[],h=HN(i);for(let y=0;y<i.length;y++){let _=i[y];if(_.name==="_shapes"){let v=new Uint32Array(t||[]),O=Math.max(Math.ceil(v.byteLength/16)*16,16),R=n.createBuffer({size:O,usage:u.UNIFORM|u.COPY_DST});n.queue.writeBuffer(R,0,v),p.push(R),c.push({binding:_.index,resource:{buffer:R}});continue}let b=bs(_.dtype),w=$u(_.dtype);if(_.packed){let v=eo(_.packedSize*b),R=_.mode==="read_write"?u.STORAGE|u.COPY_SRC|u.COPY_DST:u.STORAGE|u.COPY_DST,B=n.createBuffer({size:Math.max(v,4),usage:R,mappedAtCreation:!0}),E=new w(B.getMappedRange());for(let z of _.packed){let D=h.get(z.name),J=e[D];J&&Qd(E,J,z.dtype,z.offset)}B.unmap(),p.push(B),c.push({binding:_.index,resource:{buffer:B}});continue}let T=h.get(_.name),S=e[T],x=eo(S.length*b),A=_.mode==="read_write"?u.STORAGE|u.COPY_SRC|u.COPY_DST:u.STORAGE|u.COPY_DST,N=n.createBuffer({size:Math.max(x,4),usage:A,mappedAtCreation:!0});Qd(new w(N.getMappedRange()),S,_.dtype,0),N.unmap(),p.push(N),c.push({binding:_.index,resource:{buffer:N}})}let f=n.createBindGroup({layout:o,entries:c}),d=n.createCommandEncoder(),m=d.beginComputePass();m.setPipeline(s),m.setBindGroup(0,f),m.dispatchWorkgroups(a[0],a[1],a[2]),m.end();let g=[];for(let y=0;y<i.length;y++){let _=i[y];if(_.name==="_shapes"||_.mode!=="read_write")continue;let b=p[y],w=bs(_.dtype),T=$u(_.dtype);if(_.packed)for(let S of _.packed){let x=S.offset*w,k=eo(S.size*w),A=n.createBuffer({size:k,usage:u.MAP_READ|u.COPY_DST});d.copyBufferToBuffer(b,x,A,0,k),g.push({readBuf:A,tensorIdx:h.get(S.name),size:S.size,dtype:S.dtype,ViewCtor:T})}else{let S=b.size,x=n.createBuffer({size:S,usage:u.MAP_READ|u.COPY_DST});d.copyBufferToBuffer(b,0,x,0,S),g.push({readBuf:x,tensorIdx:h.get(_.name),size:e[h.get(_.name)].length,dtype:_.dtype,ViewCtor:T})}}n.queue.submit([d.finish()]);for(let y of g){await y.readBuf.mapAsync(Tw());let _=new y.ViewCtor(y.readBuf.getMappedRange());bw(e[y.tensorIdx],_,y.dtype,0,y.size),y.readBuf.unmap(),y.readBuf.destroy()}for(let y of p)y.destroy()}async function YN(r,e){if(typeof r.createComputePipelineAsync!="function")return;let t=[],n=new Set;for(let s of e){if(!s||n.has(s)||Xd.has(s))continue;n.add(s);let{shaderModule:o,bindGroupLayout:i,pipelineLayout:a}=UN(r,s);t.push(r.createComputePipelineAsync({layout:a,compute:{module:o,entryPoint:s.name}}).then(u=>Xd.set(s,{pipeline:u,bindGroupLayout:i}),()=>{}))}t.length>0&&await Promise.all(t)}async function K2(r,e,t){let n=await ww(),s=Fu();await YN(n,t.map(b=>b.kernel));let o=new Set;for(let b of t)for(let w of b.outputSlots)o.add(w);let i=new Array(r.numSlots).fill(null),a=new Array(r.numSlots).fill("f32");for(let b=0;b<r.numSlots;b++){let w=e[b];if(!w)continue;let T=w.dtype||"f32";a[b]=T;let S=Math.max(eo(w.data.length*bs(T)),4),x=!o.has(b),k=n.createBuffer({size:S,usage:s.STORAGE|s.COPY_DST|s.COPY_SRC,mappedAtCreation:x});x&&(Qd(new($u(T))(k.getMappedRange()),w.data,T,0),k.unmap()),i[b]=k}let u=32,p=[],c=[],h={encoder:n.createCommandEncoder(),pending:0},f=l(()=>{++h.pending>=u&&(n.queue.submit([h.encoder.finish()]),h.encoder=n.createCommandEncoder(),h.pending=0)},"maybeFlush"),d=l(b=>{let w=h.encoder,{pipeline:T,bindGroupLayout:S}=WN(n,b.kernel),x=b.inputSlots.concat(b.outputSlots),k=[],A=[],N=0;for(let B of b.kernel.metadata.bindings)if(B.name==="_shapes"){let E=new Uint32Array(b.shapeValues||[]),z=Math.max(Math.ceil(E.byteLength/16)*16,16),D=n.createBuffer({size:z,usage:s.UNIFORM|s.COPY_DST});n.queue.writeBuffer(D,0,E),p.push(D),k.push({binding:B.index,resource:{buffer:D}})}else if(B.packed){let E=bs(B.dtype),z=n.createBuffer({size:Math.max(eo(B.packedSize*E),4),usage:s.STORAGE|s.COPY_DST|s.COPY_SRC});c.push(z);let D=B.mode==="read_write";for(let J of B.packed){let j=x[J.argIndex],U=eo(J.size*E);D?A.push({slot:j,src:z,srcOff:J.offset*E,bytes:U}):w.copyBufferToBuffer(i[j],0,z,J.offset*E,U)}k.push({binding:B.index,resource:{buffer:z}})}else k.push({binding:B.index,resource:{buffer:i[x[N++]]}});let v=n.createBindGroup({layout:S,entries:k}),O=w.beginComputePass();O.setPipeline(T),O.setBindGroup(0,v);let R=b.kernel.metadata.dispatchSize;O.dispatchWorkgroups(R[0],R[1],R[2]),O.end();for(let B of A)w.copyBufferToBuffer(B.src,B.srcOff,i[B.slot],0,B.bytes);f()},"encodeStep"),m=r.scanLoops?[...r.scanLoops].sort((b,w)=>b.loopStart-w.loopStart):r.scanLoop?[r.scanLoop]:null;if(m&&m.length){let b=0;for(let w of m){for(;b<w.loopStart;b++)d(t[b]);for(let T of w.carry)h.encoder.copyBufferToBuffer(i[T.initSlot],0,i[T.a],0,T.bytes);for(let T=0;T<w.T;T++){for(let S of w.xs)h.encoder.copyBufferToBuffer(i[S.xsSlot],T*S.stepBytes,i[S.xtSlot],0,S.stepBytes);for(let S=w.loopStart;S<w.loopEnd;S++)d(t[S]);for(let S of w.ys)h.encoder.copyBufferToBuffer(i[S.ytSlot],0,i[S.ysSlot],T*S.stepBytes,S.stepBytes);for(let S of w.carry){let x=i[S.a];i[S.a]=i[S.b],i[S.b]=x}}for(let T of w.carry)h.encoder.copyBufferToBuffer(i[T.a],0,i[T.finalSlot],0,T.bytes);b=w.loopEnd}for(;b<t.length;b++)d(t[b])}else for(let b of t)d(b);let g=h.encoder,y=new Set(r.argSlots),_=[];for(let b=0;b<r.numSlots;b++){if(!i[b]||!o.has(b)||!y.has(b))continue;let w=e[b],T=Math.max(eo(w.data.length*bs(a[b])),4),S=n.createBuffer({size:T,usage:s.MAP_READ|s.COPY_DST});g.copyBufferToBuffer(i[b],0,S,0,T),_.push({rb:S,dtype:a[b],size:w.data.length,dst:w.data})}n.queue.submit([g.finish()]);for(let b of _)await b.rb.mapAsync(Tw()),bw(b.dst,new($u(b.dtype))(b.rb.getMappedRange()),b.dtype,0,b.size),b.rb.unmap(),b.rb.destroy();for(let b of i)b&&b.destroy();for(let b of p)b.destroy();for(let b of c)b.destroy()}function JN(r){let e=_w.get(r);if(e&&e.length)return e.pop();let t=Fu();return kt.createBuffer({size:r,usage:t.STORAGE|t.COPY_SRC|t.COPY_DST})}function VN(r,e){let t=_w.get(e);t||(t=[],_w.set(e,t)),t.push(r)}async function H2(){return await ww(),As||(As=kt.createCommandEncoder()),kt}function Y2(){return!!kt&&!!As}function GN(r,e,t){let n=Vc.get(r);if(n)return n.buf;let s=Fu(),o=Math.max(eo(r.length*bs(e)),4);if(t){let a=kt.createBuffer({size:o,usage:s.STORAGE|s.COPY_SRC|s.COPY_DST,mappedAtCreation:!0});return Qd(new($u(e))(a.getMappedRange()),r,e,0),a.unmap(),Vc.set(r,{buf:a,bytes:o,dtype:e,persistent:t}),a}let i=JN(o);return Vc.set(r,{buf:i,bytes:o,dtype:e,persistent:!1}),i}function Gc(r,e,t){let{pipeline:n,bindGroupLayout:s}=WN(kt,r),{bindings:o,dispatchSize:i}=r.metadata,a=HN(o),u=Fu(),p=[],c=[];for(let d of o){if(d.name==="_shapes"){let y=new Uint32Array(t||[]),_=Math.max(Math.ceil(y.byteLength/16)*16,16),b=kt.createBuffer({size:_,usage:u.UNIFORM|u.COPY_DST});kt.queue.writeBuffer(b,0,y),gw.push(b),p.push({binding:d.index,resource:{buffer:b}});continue}if(d.packed){let y=bs(d.dtype),_=Math.max(eo(d.packedSize*y),4),b=JN(_);yw.push({buf:b,bytes:_});let w=d.mode==="read_write";for(let T of d.packed){let S=GN(e[a.get(T.name)],T.dtype,!w),x=eo(T.size*y);w?c.push({pbuf:b,off:T.offset*y,sub:S,bytes:x}):As.copyBufferToBuffer(S,0,b,T.offset*y,x)}p.push({binding:d.index,resource:{buffer:b}});continue}let m=e[a.get(d.name)],g=GN(m,d.dtype,d.mode!=="read_write");p.push({binding:d.index,resource:{buffer:g}})}let h=kt.createBindGroup({layout:s,entries:p}),f=As.beginComputePass();f.setPipeline(n),f.setBindGroup(0,h),f.dispatchWorkgroups(i[0],i[1],i[2]),f.end();for(let d of c)As.copyBufferToBuffer(d.pbuf,d.off,d.sub,0,d.bytes);++mw>=W2&&(kt.queue.submit([As.finish()]),As=kt.createCommandEncoder(),mw=0)}async function ZN(){if(!kt)return;let r=Fu(),e=[];for(let[t,n]of Vc){if(n.persistent)continue;let s=kt.createBuffer({size:n.bytes,usage:r.MAP_READ|r.COPY_DST});As.copyBufferToBuffer(n.buf,0,s,0,n.bytes),e.push({rb:s,ha:t,buf:n.buf,bytes:n.bytes,dtype:n.dtype})}kt.queue.submit([As.finish()]),As=kt.createCommandEncoder(),mw=0;for(let t of gw)t.destroy();gw.length=0,await Promise.all(e.map(t=>t.rb.mapAsync(Tw())));for(let t of e)bw(t.ha,new($u(t.dtype))(t.rb.getMappedRange()),t.dtype,0,t.ha.length),t.rb.unmap(),t.rb.destroy();for(let t of yw)VN(t.buf,t.bytes);yw.length=0;for(let t of e)VN(t.buf,t.bytes),Vc.delete(t.ha)}function J2(r,e,t,n,s){let o=`${r.join(",")}|${e.join(",")}|${t}|${n}`,i=jN.get(o);if(i)return i;let a=Mn(n),u="";for(let c=r.length-1;c>=0;c--)u+=`  let idx${c} = rem % ${r[c]}u; rem = rem / ${r[c]}u; src = src + idx${c} * ${e[c]}u;
 `;return i={name:"contig",source:`@group(0) @binding(0) var<storage, read> inp : array<${a}>;
 @group(0) @binding(1) var<storage, read_write> outp : array<${a}>;
 @compute @workgroup_size(64)