@genai-fi/nanogpt 0.7.1 → 0.7.2

package/dist/ops/webgpu/attentionMask.js

@@ -1,6 +1,6 @@
-import { f } from "../../index-UdZhlibC.js";
-import { g as m } from "../../webgpu_program-WaoMq-WD.js";
-import { f as k, c as l } from "../../webgpu_util-DhSeP4b6.js";
+import { f, a2 as m } from "../../index-BoWRt-10.js";
+import { g as k } from "../../webgpu_program-PFzf1hAQ.js";
+import { f as l, c as v } from "../../webgpu_util-D____QpY.js";
 class g {
   variableNames = ["q", "k"];
   outputShape;
@@ -14,13 +14,13 @@ class g {
   nh;
   T1;
   T2;
-  constructor(t, e, i, r, s) {
-    if (this.shaderKey = `AttentionMask_${s}`, this.outputShape = [t, e, i, r], this.hs = s, this.nh = e, this.T1 = i, this.T2 = r, this.dispatchLayout = k(this.outputShape), this.dispatch = l(this.dispatchLayout, this.outputShape, this.workgroupSize), s % 4 !== 0)
+  constructor(t, e, s, o, i) {
+    if (this.shaderKey = `AttentionMask_${i}`, this.outputShape = [t, e, s, o], this.hs = i, this.nh = e, this.T1 = s, this.T2 = o, this.dispatchLayout = l(this.outputShape), this.dispatch = v(this.dispatchLayout, this.outputShape, this.workgroupSize), i % 4 !== 0)
       throw new Error("Head size must be a multiple of 4 for AttentionMaskProgram");
   }
   getUserCode() {
     return `
-            ${m("index")} {
+            ${k("index")} {
                 
                 let coords = getCoordsFromIndex(index);
                 let b = coords[0];
@@ -33,13 +33,14 @@ class g {
                         setOutputAtIndex(index, uniforms.inf);
                         return;
                     }
+
+                    let q0 = getIndexFromCoords4D(vec4<i32>(b, h, t1, 0), uniforms.qShape);
+                    let k0 = getIndexFromCoords4D(vec4<i32>(b, h, t2, 0), uniforms.kShape);
                     
                     var sum: f32 = 0.0;
                     for (var i: i32 = 0; i < ${this.hs}; i = i + 4) {
-                        let q0 = getIndexFromCoords4D(vec4<i32>(b, h, t1, i), uniforms.qShape);
-                        let qv = vec4<f32>(q[q0], q[q0 + 1], q[q0 + 2], q[q0 + 3]);
-                        let k0 = getIndexFromCoords4D(vec4<i32>(b, h, t2, i), uniforms.kShape);
-                        let kv = vec4<f32>(k[k0], k[k0 + 1], k[k0 + 2], k[k0 + 3]);
+                        let qv = vec4<f32>(q[q0 + i], q[q0 + i + 1], q[q0 + i + 2], q[q0 + i + 3]);
+                        let kv = vec4<f32>(k[k0 + i], k[k0 + i + 1], k[k0 + i + 2], k[k0 + i + 3]);
                         sum = sum + dot(qv, kv);
                     }
                     let scaled = sum * uniforms.divisor;
@@ -49,17 +50,22 @@ class g {
         `;
   }
 }
-function q(o) {
-  const { q: t, k: e } = o.inputs, { divisor: i, pastLen: r } = o.attrs, s = o.backend, a = t.shape[0], n = t.shape[2], u = e.shape[2], d = t.shape[1], h = t.shape[3], c = new g(a, d, n, u, h), p = [
-    { type: "float32", data: [i] },
-    { type: "int32", data: [r] },
+function b(n) {
+  const { q: t, k: e } = n.inputs, { divisor: s, pastLen: o } = n.attrs, i = n.backend, r = t.shape[0], p = t.shape[2], a = e.shape[2], u = t.shape[1], h = t.shape[3];
+  if (m(e.shape, [r, u, a, h], "Error in AttentionMask: "), s === 0)
+    throw new Error("Divisor must be non-zero in AttentionMask");
+  if (o < 0)
+    throw new Error("pastLen must be non-negative in AttentionMask");
+  const c = new g(r, u, p, a, h), d = [
+    { type: "float32", data: [s] },
+    { type: "int32", data: [o] },
     { type: "float32", data: [Number.NEGATIVE_INFINITY] }
   ];
-  return s.runWebGPUProgram(c, [t, e], "float32", p);
+  return i.runWebGPUProgram(c, [t, e], "float32", d);
 }
-const v = {
+const q = {
   kernelName: "AttentionMask",
   backendName: "webgpu",
-  kernelFunc: q
+  kernelFunc: b
 };
-f(v);
+f(q);

package/dist/ops/webgpu/gatherSub.js

@@ -1,6 +1,6 @@
-import { g as r } from "../../webgpu_program-WaoMq-WD.js";
-import { f as u, c as d } from "../../webgpu_util-DhSeP4b6.js";
-import { f as p } from "../../index-UdZhlibC.js";
+import { g as u } from "../../webgpu_program-PFzf1hAQ.js";
+import { f as h, c as p } from "../../webgpu_util-D____QpY.js";
+import { f as c, a2 as r } from "../../index-BoWRt-10.js";
 class l {
   variableNames = ["labels", "logits", "values"];
   outputShape;
@@ -10,43 +10,30 @@ class l {
   workgroupSize = [64, 1, 1];
   size = !0;
   constructor(e) {
-    this.outputShape = [e], this.dispatchLayout = u(this.outputShape), this.dispatch = d(this.dispatchLayout, this.outputShape, this.workgroupSize);
+    this.outputShape = [e], this.dispatchLayout = h(this.outputShape), this.dispatch = p(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     return `
-        ${r("index")} {
+        ${u("index")} {
             if (index < uniforms.size) {
-                let coords = getCoordsFromIndex(index);
-                let idx = i32(getLabelsByOutputIndex(index));
-                let val = getValuesByOutputIndex(index);
-
-                if (isnan(val)) {
-                    setOutputAtIndex(index, 0.0);
-                    return;
-                }
-
-                if (idx < uniforms.logitsShape[1] && idx >= 0) {
-                    let logit = getLogits(coords, idx);
-                    if (isnan(logit)) {
-                        setOutputAtIndex(index, 0.0);
-                        return;
-                    }
-                    setOutputAtIndex(index, val - logit);
-                } else {
-                    setOutputAtIndex(index, 0.0);
-                }
+                let idx = i32(labels[index]);
+                let val = values[index];
+                let logit = logits[index * uniforms.logitsShape[1] + idx];
+                setOutputAtIndex(index, val - logit);
             }
         }
     `;
   }
 }
-function c(t) {
-  const { logits: e, labels: i, values: s } = t.inputs, a = t.backend, n = i.shape[0], o = new l(n);
-  return a.runWebGPUProgram(o, [i, e, s], "float32");
+function d(t) {
+  const { logits: e, labels: a, values: s } = t.inputs, o = t.backend, i = a.shape[0];
+  r(s.shape, [i], "Error in EfficientGatherSub: "), r(a.shape, [i], "Error in EfficientGatherSub: ");
+  const n = new l(i);
+  return o.runWebGPUProgram(n, [a, e, s], "float32");
 }
-const h = {
+const f = {
   kernelName: "EfficientGatherSub",
   backendName: "webgpu",
-  kernelFunc: c
+  kernelFunc: d
 };
-p(h);
+c(f);

package/dist/ops/webgpu/gelu.js

@@ -1,6 +1,6 @@
-import { f as i } from "../../index-UdZhlibC.js";
-import { g as o } from "../../webgpu_program-WaoMq-WD.js";
-import { f as s, c as p } from "../../webgpu_util-DhSeP4b6.js";
+import { f as i } from "../../index-BoWRt-10.js";
+import { g as o } from "../../webgpu_program-PFzf1hAQ.js";
+import { f as s, c as p } from "../../webgpu_util-D____QpY.js";
 const u = 0.7978845608028654, a = 0.044715;
 class c {
   outputShape;

package/dist/ops/webgpu/normRMS.js

@@ -1,5 +1,5 @@
-import { f as p } from "../../webgpu_util-DhSeP4b6.js";
-import { f as s } from "../../index-UdZhlibC.js";
+import { f as n } from "../../webgpu_util-D____QpY.js";
+import { f as p, a2 as h } from "../../index-BoWRt-10.js";
 import { createReduceInfo as u, reduce as c, createReductionShader as m } from "./utils/reductions.js";
 class d {
   outputShape;
@@ -12,7 +12,7 @@ class d {
   inputShape;
   size = !0;
   constructor(e) {
-    this.inputShape = [e.batchSize, e.inSize], this.outputShape = this.inputShape, this.dispatchLayout = p(this.outputShape), this.dispatch = [e.batchSize, 1, 1];
+    this.inputShape = [e.batchSize, e.inSize], this.outputShape = this.inputShape, this.dispatchLayout = n(this.outputShape), this.dispatch = [e.batchSize, 1, 1];
   }
   getUserCode() {
     const e = this.workgroupSize[0];
@@ -29,13 +29,21 @@ class d {
         `);
   }
 }
-function S(t) {
-  const { x: e, gamma: i } = t.inputs, n = t.backend, a = [e, i], o = u(a, -1), r = new d(o);
-  return c(r, a, n);
+function S(a) {
+  const { x: e, gamma: r } = a.inputs, o = a.backend, s = [e, r], t = u(s, -1), i = new d(t);
+  if (h(r.shape, [e.shape[e.shape.length - 1]], "Error in RMSNorm: "), e.shape.length !== 3)
+    throw new Error(`rmsNormGPU: input rank ${e.shape.length} not supported, only rank 3 is supported`);
+  if (t.inSize !== e.shape[2])
+    throw new Error(`rmsNormGPU: reduction size ${t.inSize} does not match expected size ${e.shape[2]}`);
+  if (t.batchSize !== e.shape[0] * e.shape[1])
+    throw new Error(
+      `rmsNormGPU: batch size ${t.batchSize} does not match expected size ${e.shape[0] * e.shape[1]}`
+    );
+  return c(i, s, o);
 }
-const h = {
+const l = {
   kernelName: "RMSNorm",
   backendName: "webgpu",
   kernelFunc: S
 };
-s(h);
+p(l);

package/dist/ops/webgpu/normRMSGrad.js

@@ -1,10 +1,10 @@
-import { f as h, e as S } from "../../index-UdZhlibC.js";
-import { createReduceInfo as g } from "./utils/reductions.js";
-import { f as k } from "../../webgpu_util-DhSeP4b6.js";
-import { g as x } from "../../webgpu_program-WaoMq-WD.js";
-import { s as c } from "../../slice-BHbDHObE.js";
-import { s as z } from "../../sum-NLYbiDag.js";
-class w {
+import { f, a2 as c, e as g } from "../../index-BoWRt-10.js";
+import { createReduceInfo as k } from "./utils/reductions.js";
+import { f as x } from "../../webgpu_util-D____QpY.js";
+import { g as z } from "../../webgpu_program-PFzf1hAQ.js";
+import { s as d } from "../../slice-BiOsknYS.js";
+import { s as w } from "../../sum-66ew2byf.js";
+class y {
   outputShape;
   shaderKey = "RMSNormGrad";
   dispatchLayout;
@@ -16,7 +16,7 @@ class w {
   size = !1;
   rowsPerWorkgroup;
   constructor(r, e = 4) {
-    if (this.shaderKey = `RMSNormGrad_${e}`, this.rowsPerWorkgroup = e, this.inputShape = [r.batchSize, r.inSize], this.outputShape = [r.batchSize + r.batchSize / this.rowsPerWorkgroup, r.inSize], this.dispatchLayout = k(this.outputShape), this.dispatch = [r.batchSize / this.rowsPerWorkgroup, 1, 1], r.batchSize % this.rowsPerWorkgroup !== 0)
+    if (this.shaderKey = `RMSNormGrad_${e}`, this.rowsPerWorkgroup = e, this.inputShape = [r.batchSize, r.inSize], this.outputShape = [r.batchSize + r.batchSize / this.rowsPerWorkgroup, r.inSize], this.dispatchLayout = x(this.outputShape), this.dispatch = [r.batchSize / this.rowsPerWorkgroup, 1, 1], r.batchSize % this.rowsPerWorkgroup !== 0)
       throw new Error(
         `RMSNormGradProgram: batch size ${r.batchSize} must be divisible by rowsPerWorkgroup ${this.rowsPerWorkgroup}`
       );
@@ -35,7 +35,7 @@ class w {
           var<workgroup> accumulation: array<f32, 1024>;
         `}
 
-          ${x("index")} {
+          ${z("index")} {
             // One workgroup per row (batch).
             let Length = uniforms.reduceSize;
             let BatchSize = uniforms.batchSize;
@@ -105,24 +105,29 @@ class w {
         `;
   }
 }
-function v(t) {
-  const { dy: r, x: e, gamma: o } = t.inputs, i = 4, l = t.backend, a = g([e, o, r], -1), s = new w(a, i), d = [
+function v(o) {
+  const { dy: r, x: e, gamma: t } = o.inputs, i = 4;
+  c(e.shape, r.shape, "Error in RMSNormGrad dy: "), c(t.shape, [e.shape[e.shape.length - 1]], "Error in RMSNormGrad gamma: ");
+  const l = o.backend, a = k([e, t, r], -1), s = new y(a, i), p = [
     { type: "int32", data: [s.inputShape[1]] },
     // Reduce size
     { type: "int32", data: [s.inputShape[0]] }
     // Batch size
-  ], n = S().makeTensorFromTensorInfo(
-    l.runWebGPUProgram(s, [e, o, r], "float32", d)
-  ), u = c(n, [0, 0], [a.batchSize, a.inSize]), m = c(n, [a.batchSize, 0], [a.batchSize / i, a.inSize]);
+  ];
+  if (a.inSize > 1024)
+    throw new Error(`rmsNormGradGPU: inSize ${a.inSize} exceeds max of 1024`);
+  const n = g().makeTensorFromTensorInfo(
+    l.runWebGPUProgram(s, [e, t, r], "float32", p)
+  ), m = d(n, [0, 0], [a.batchSize, a.inSize]), u = d(n, [a.batchSize, 0], [a.batchSize / i, a.inSize]);
   n.dispose();
-  const p = u.reshape(e.shape);
-  u.dispose();
-  const f = z(m, [0]);
-  return m.dispose(), [p, f];
+  const h = m.reshape(e.shape);
+  m.dispose();
+  const S = w(u, [0]);
+  return u.dispose(), [h, S];
 }
-const y = {
+const b = {
   kernelName: "RMSNormGrad",
   backendName: "webgpu",
   kernelFunc: v
 };
-h(y);
+f(b);

package/dist/ops/webgpu/qkv.js

@@ -1,7 +1,7 @@
-import { g as c } from "../../webgpu_program-WaoMq-WD.js";
-import { f as d, c as h } from "../../webgpu_util-DhSeP4b6.js";
-import { f as p } from "../../index-UdZhlibC.js";
-class m {
+import { g as h } from "../../webgpu_program-PFzf1hAQ.js";
+import { f as c, c as d } from "../../webgpu_util-D____QpY.js";
+import { f as p, a2 as m } from "../../index-BoWRt-10.js";
+class l {
   variableNames = ["x", "kernel"];
   outputShape;
   shaderKey = "QKV";
@@ -11,13 +11,13 @@ class m {
   workgroupSize = [64, 1, 1];
   size = !0;
   constructor(t, e, o, s) {
-    const r = s / e;
-    this.shaderKey = `QKV_${e}_${r}`, this.outputShape = [t, e, o, r], this.dispatchLayout = d(this.outputShape), this.dispatch = h(this.dispatchLayout, this.outputShape, this.workgroupSize);
+    const n = s / e;
+    this.shaderKey = `QKV_${e}_${n}`, this.outputShape = [t, e, o, n], this.dispatchLayout = c(this.outputShape), this.dispatch = d(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     const t = this.outputShape[1], e = this.outputShape[3], o = t * e;
     return `
-        ${c("index")} {
+        ${h("index")} {
             if (index < uniforms.size) {
                 let coords = getCoordsFromIndex(index); // [b, h, t, d]
                 let b = coords[0];
@@ -26,13 +26,14 @@ class m {
                 let d = coords[3];
 
                 // Compute output channel index in fused kernel
-                let out_offset = uniforms.mode * ${t} * ${e} + h * ${e} + d;
+                let out_offset = uniforms.mode * ${o} + h * ${e} + d;
 
                 var sum = 0.0;
+                let baseX = b * uniforms.xShape[1] * uniforms.xShape[2] + t * uniforms.xShape[2];
                 for (var c = 0; c < ${o}; c += 1) {
-                    let xval = getX(b, t, c); // fetch from x
-                    let kval = getKernel(c, out_offset); // fetch from kernel
-                    sum += xval * kval;
+                    let xval = x[baseX + c];
+                    let kval = getKernel(c, out_offset);
+                    sum = fma(xval, kval, sum);
                 }
 
                 setOutputAtIndex(index, sum);
@@ -41,17 +42,20 @@ class m {
         `;
   }
 }
-function l(a) {
-  const { x: t, kernel: e } = a.inputs, { heads: o } = a.attrs, s = a.backend, r = t.shape[0], i = t.shape[1], u = t.shape[2], n = new m(r, o, i, u);
+function f(r) {
+  const { x: t, kernel: e } = r.inputs, { heads: o } = r.attrs, s = r.backend, n = t.shape[0], u = t.shape[1], a = t.shape[2];
+  if (m(e.shape, [a, 3 * a], "Error in QKV: "), a % o !== 0)
+    throw new Error(`Channel dimension ${a} must be divisible by number of heads ${o} in QKV.`);
+  const i = new l(n, o, u, a);
   return [
-    s.runWebGPUProgram(n, [t, e], "float32", [{ type: "int32", data: [0] }]),
-    s.runWebGPUProgram(n, [t, e], "float32", [{ type: "int32", data: [1] }]),
-    s.runWebGPUProgram(n, [t, e], "float32", [{ type: "int32", data: [2] }])
+    s.runWebGPUProgram(i, [t, e], "float32", [{ type: "int32", data: [0] }]),
+    s.runWebGPUProgram(i, [t, e], "float32", [{ type: "int32", data: [1] }]),
+    s.runWebGPUProgram(i, [t, e], "float32", [{ type: "int32", data: [2] }])
   ];
 }
-const f = {
+const b = {
   kernelName: "QKV",
   backendName: "webgpu",
-  kernelFunc: l
+  kernelFunc: f
 };
-p(f);
+p(b);

package/dist/ops/webgpu/rope.js

@@ -1,7 +1,7 @@
-import { g as h } from "../../webgpu_program-WaoMq-WD.js";
-import { f as l, c as m } from "../../webgpu_util-DhSeP4b6.js";
-import { f as g } from "../../index-UdZhlibC.js";
-class f {
+import { g as c } from "../../webgpu_program-PFzf1hAQ.js";
+import { f as m, c as l } from "../../webgpu_util-D____QpY.js";
+import { f as x, a2 as f } from "../../index-BoWRt-10.js";
+class S {
   variableNames = ["x", "sin", "cos"];
   outputShape;
   shaderKey = "Rope";
@@ -10,13 +10,13 @@ class f {
   workgroupSize = [64, 1, 1];
   size = !0;
   uniforms = "pastLen: i32";
-  constructor(t, s, n, o) {
-    this.shaderKey = `Rope_${o}`, this.outputShape = [t, s, n, o], this.dispatchLayout = l(this.outputShape), this.dispatch = m(this.dispatchLayout, this.outputShape, this.workgroupSize);
+  constructor(t, e, a, s) {
+    this.shaderKey = `Rope_${s}`, this.outputShape = [t, e, a, s], this.dispatchLayout = m(this.outputShape), this.dispatch = l(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     const t = this.outputShape[3];
     return `
-        ${h("index")} {
+        ${c("index")} {
             if (index < uniforms.size) {
                 let coords = getCoordsFromIndex(index); // [b, h, t, d]
                 let b = coords[0];
@@ -28,27 +28,31 @@ class f {
 
                 var outVal = 0.0;
 
+                let xIdx = b * uniforms.outShapeStrides[0] +
+                    h * uniforms.outShapeStrides[1] +
+                    t * uniforms.outShapeStrides[2] +
+                    d;
+
                 if (d < rotaryDim) {
-                    let pairIdx = d / 2;
-                    let cos = getCos(t + uniforms.pastLen, pairIdx, 0);
-                    let sin = getSin(t + uniforms.pastLen, pairIdx, 0);
+                    let idx = (t + uniforms.pastLen) * uniforms.cosShape[1] + d / 2;
+                    let cos = cos[idx];
+                    let sin = sin[idx];
 
-                    let ownX = getX(b, h, t, d) * cos;
+                    let ownX = x[xIdx] * cos;
+                    var evenOdd = 0.0;
 
                     if (d % 2 == 0) {
                         // even index
-                        let even = ownX;
-                        let odd = getX(b, h, t, d + 1);
-                        outVal = even - odd * sin;
+                        evenOdd = -x[xIdx + 1];
                     } else {
                         // odd index
-                        let even = getX(b, h, t, d - 1);
-                        let odd = ownX;
-                        outVal = even * sin + odd;
+                        evenOdd = x[xIdx - 1];
                     }
+
+                    outVal = fma(evenOdd, sin, ownX);
                 } else {
                     // pass through for non-rotary dims
-                    outVal = getX(b, h, t, d);
+                    outVal = x[xIdx];
                 }
 
                 setOutputAtIndex(index, outVal);
@@ -57,13 +61,22 @@ class f {
         `;
   }
 }
-function b(e) {
-  const { x: t, sin: s, cos: n } = e.inputs, { pastLen: o } = e.attrs, a = e.backend, r = t.shape[0], i = t.shape[1], d = t.shape[2], p = t.shape[3], u = new f(r, i, d, p), c = [{ type: "int32", data: [o] }];
-  return a.runWebGPUProgram(u, [t, s, n], "float32", c);
+function g(o) {
+  const { x: t, sin: e, cos: a } = o.inputs, { pastLen: s } = o.attrs, i = o.backend, p = t.shape[0], d = t.shape[1], n = t.shape[2], r = t.shape[3];
+  if (f(e.shape, a.shape, "Error in Rope: "), e.shape[0] < n + s)
+    throw new Error(
+      `Sin tensor shape ${e.shape} is not compatible with seqLength ${n} and pastLen ${s}.`
+    );
+  if (e.shape[1] * 2 < r)
+    throw new Error(`Sin tensor shape ${e.shape} is not compatible with feature dimension ${r}.`);
+  if (e.shape.length !== 3)
+    throw new Error(`Sin tensor must be 3-dimensional, but got shape ${e.shape}.`);
+  const h = new S(p, d, n, r), u = [{ type: "int32", data: [s] }];
+  return i.runWebGPUProgram(h, [t, e, a], "float32", u);
 }
-const x = {
+const b = {
   kernelName: "Rope",
   backendName: "webgpu",
-  kernelFunc: b
+  kernelFunc: g
 };
-g(x);
+x(b);

package/dist/ops/webgpu/scatterSub.js

@@ -1,7 +1,7 @@
-import { g as c } from "../../webgpu_program-WaoMq-WD.js";
-import { f as p, c as u } from "../../webgpu_util-DhSeP4b6.js";
-import { f as d } from "../../index-UdZhlibC.js";
-class h {
+import { g as p } from "../../webgpu_program-PFzf1hAQ.js";
+import { f as u, c as d } from "../../webgpu_util-D____QpY.js";
+import { f as h, a2 as o } from "../../index-BoWRt-10.js";
+class b {
   variableNames = ["labels", "softmaxProbs", "dy"];
   outputShape;
   shaderKey = "ScatterSub";
@@ -10,29 +10,31 @@ class h {
   workgroupSize = [64, 1, 1];
   size = !0;
   constructor(t, e) {
-    this.outputShape = [t, e], this.dispatchLayout = p(this.outputShape), this.dispatch = u(this.dispatchLayout, this.outputShape, this.workgroupSize);
+    this.outputShape = [t, e], this.dispatchLayout = u(this.outputShape), this.dispatch = d(this.dispatchLayout, this.outputShape, this.workgroupSize);
   }
   getUserCode() {
     return `
-            ${c("index")} {
+            ${p("index")} {
                 if (index < uniforms.size) {
                     let coords = getCoordsFromIndex(index); // [batch, depth]
-                    let idx = i32(getLabels(coords[0]));
-                    let prob = getSoftmaxProbsByOutputIndex(index);
-                    let dy = getDy(coords[0]);
+                    let idx = i32(labels[coords[0]]);
+                    let prob = softmaxProbs[index];
+                    let dy = dy[coords[0]];
                     setOutputAtIndex(index, select(prob, prob - 1.0, idx == coords[1]) * dy);
                 }
             }
             `;
   }
 }
-function b(o) {
-  const { logits: t, labels: e, dy: s } = o.inputs, r = o.backend, a = e.shape[0], i = t.shape[1], n = new h(a, i);
-  return r.runWebGPUProgram(n, [e, t, s], "float32");
+function f(a) {
+  const { logits: t, labels: e, dy: s } = a.inputs, c = a.backend, r = e.shape[0], i = t.shape[1];
+  o(s.shape, [r], "Error in EfficientScatterSub dy: "), o(t.shape, [r, i], "Error in EfficientScatterSub logits: "), o(e.shape, [r], "Error in EfficientScatterSub labels: ");
+  const n = new b(r, i);
+  return c.runWebGPUProgram(n, [e, t, s], "float32");
 }
 const l = {
   kernelName: "EfficientScatterSub",
   backendName: "webgpu",
-  kernelFunc: b
+  kernelFunc: f
 };
-d(l);
+h(l);

package/dist/ops/webgpu/utils/reductions.js

@@ -1,7 +1,7 @@
-import { p as l, j as d } from "../../../index-UdZhlibC.js";
-import { g as p } from "../../../webgpu_program-WaoMq-WD.js";
-import { r as f } from "../../../Reshape-Do18N3gO.js";
-import { c as x } from "../../../axis_util-5DTW2tFV.js";
+import { p as l, j as d } from "../../../index-BoWRt-10.js";
+import { g as p } from "../../../webgpu_program-PFzf1hAQ.js";
+import { r as f } from "../../../Reshape-DvudQDvJ.js";
+import { c as x } from "../../../axis_util-BzbKo31C.js";
 function I(e, r, t, s, u) {
   return `
            fn DIV_CEIL(a : u32, b : u32) -> u32 {

package/dist/{ops-aRTXR2Sr.js → ops-BFGCx8Ri.js}

@@ -1,19 +1,19 @@
-import { B as m, C as u, l, E as f, aP as gn, k as S, aQ as $n, aR as kn, aS as yn, p as En, aT as w, q as rn, x as _n, b as pn, aU as In, aV as Nn, N as An, al as Tn, I as v, aW as xn, aX as wn, aY as Sn, aZ as Mn, a_ as zn, a$ as qn, b0 as Bn, X as K, b1 as Gn, b2 as Kn, b3 as Rn, aG as Dn, $ as vn, Z as Vn, _ as Pn, b4 as jn, t as Fn, b5 as L, b6 as Wn, b7 as Yn, b8 as On, b9 as Cn, ba as Ln, bb as Xn, bc as Zn, bd as Jn, a as $, o as U, n as q, h as Qn, c as A, be as Hn, M as X, d as Un } from "./index-UdZhlibC.js";
+import { B as m, C as u, l, E as f, aP as gn, k as S, aQ as $n, aR as kn, aS as yn, p as En, aT as w, q as rn, x as _n, b as pn, aU as In, aV as Nn, N as An, al as Tn, I as v, aW as xn, aX as wn, aY as Sn, aZ as Mn, a_ as zn, a$ as qn, b0 as Bn, X as K, b1 as Gn, b2 as Kn, b3 as Rn, aG as Dn, $ as vn, Z as Vn, _ as Pn, b4 as jn, t as Fn, b5 as L, b6 as Wn, b7 as Yn, b8 as On, b9 as Cn, ba as Ln, bb as Xn, bc as Zn, bd as Jn, a as $, o as U, n as q, h as Qn, c as A, be as Hn, M as X, d as Un } from "./index-BoWRt-10.js";
 import { n as ne, a as ee, b as se } from "./non_max_suppression_impl-CsEgBuMA.js";
-import { r as k } from "./reshape-BR0eoLYN.js";
-import { s as mn } from "./split-CRU0PjVV.js";
-import { s as E } from "./sum-NLYbiDag.js";
-import { b as Z } from "./broadcast_to-Ba9h_8DO.js";
-import { s as _ } from "./slice-BHbDHObE.js";
-import { r as W } from "./range-D9CZhVlR.js";
-import { t as te } from "./tensor-Do9PKbIE.js";
-import { s as Y } from "./stack-ikk2Y8_P.js";
-import { c as re, z as ae } from "./zeros-DnPT2nD4.js";
-import { e as oe } from "./axis_util-5DTW2tFV.js";
-import { m as J } from "./max-ByjEGoFx.js";
-import { c as Q } from "./concat-CbXTetof.js";
-import { m as N } from "./mat_mul-dwmZz69e.js";
-import { t as H } from "./tensor2d-CWHxHpLh.js";
+import { r as k } from "./reshape-CdBq1WJ6.js";
+import { s as mn } from "./split-B-dikLRw.js";
+import { s as E } from "./sum-66ew2byf.js";
+import { b as Z } from "./broadcast_to-CdbwV-Dj.js";
+import { s as _ } from "./slice-BiOsknYS.js";
+import { r as W } from "./range-CRuAh-gd.js";
+import { t as te } from "./tensor-JwS7ZYY6.js";
+import { s as Y } from "./stack-B17UN2nn.js";
+import { c as re, z as ae } from "./zeros--BdLQ3oG.js";
+import { e as oe } from "./axis_util-BzbKo31C.js";
+import { m as J } from "./max-Ddnnb5xe.js";
+import { c as Q } from "./concat-CsxrgovM.js";
+import { m as N } from "./mat_mul-8m8pfdcx.js";
+import { t as H } from "./tensor2d-wxPAnDQy.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/{random_width-DbSpgl4o.js → random_width-sZORGo5k.js}

@@ -1,25 +1,25 @@
-import { B as T, C as I, E as O, bE as Oa, bF as Ma, bG as Ci, l as b, bH as Ii, X as L, bI as Di, bJ as $i, bK as Ti, bL as zi, h as Ei, bM as Li, bN as Fi, bO as Oi, bP as Mi, bQ as _a, bR as _i, bS as Ra, bT as Ri, bU as Ba, bV as Bi, I as Ge, k as kt, bq as Wa, bW as Wi, bX as Gi, bY as Pi, y as Pe, c as V, a as w, bZ as Ga, b_ as Ui, b$ as Vi, c0 as ji, p as ce, aU as pt, bv as Pa, c1 as Ki, c2 as Hi, c3 as qi, c4 as Ji, c5 as Zi, bx as Xi, c6 as Yi, c7 as Qi, F as Ua, al as Va, aq as ja, c8 as tr, c9 as Ka, o as z, ca as Ps, cb as Ha, cc as qa, j as pn, cd as Us, ce as Ja, cf as Za, cg as Xa, ch as Ya, ci as Qa, cj as tl, ck as el, bf as nl, cl as sl, q as he, b as et, n as U, cm as il, bm as rl, aw as ht, cn as ol, z as Q, co as al, cp as ll, cq as ul, cr as cl, cs as hl, ct as pl, cu as dl, cv as fl, J as ml, cw as gl, bj as bl, bp as yl, cx as wl, Q as kl, cy as xl, a8 as Nl, cz as vl, cA as Al, cB as Sl, as as Cl, cC as Il, ae as Dl, aV as $l, br as Tl, ak as zl, bs as El, G as Ll, aX as Fl, am as Ol, cD as Ml, cE as _l, cF as Rl, an as Bl, af as Wl, cG as Gl, cH as Pl, cI as Ul, U as Vl, bt as jl, cJ as Kl, cK as Hl, aS as ql, b1 as Jl, cL as Zl, cM as Xl, bu as Yl, a$ as Ql, P as tu, cN as eu, x as rs, ap as nu, bw as su, az as iu, V as ru, au as ou, at as au, K as lu, bc as uu, cO as cu, bd as hu, cP as pu, b3 as du, aQ as fu, ar as mu, cQ as gu, a9 as bu, W as yu, S as wu, L as ku, bz as xu, cR as Nu, bA as vu, av as Au, bC as Su, Y as Cu, cS as Iu, N as Du, b5 as $u, b4 as Tu, cT as Oe, cU as zu, i as Eu, aj as Vs, cV as Lu, t as x, aT as $e, cW as S, cX as He, cY as qe, ab as Vt, d as J, ac as Fu, cZ as js, w as Jt, M as Ou, T as Te, O as Mu, c_ as _u, m as Ks, c$ as Ru, d0 as Hs, d1 as Bu } from "./index-UdZhlibC.js";
-import { k as ke, c as Nt, o as ze, s as er, b as nr, d as Wu, m as sr, t as In, l as ir, v as os, a as Gu, S as Pu, p as Uu, w as as, x as rr, y as Je, z as Vu, A as ju } from "./selu_util-Ce6pu9IM.js";
-import { n as wt, w as ne, a as Ze, g as Xe, b as ls, t as K, c as Ce, d as Xt, e as Ku, u as dn, f as ye, h as fn, i as Hu, l as qu, s as us, m as or, j as qt, k as Ju } from "./ops-aRTXR2Sr.js";
-import { r as N } from "./reshape-BR0eoLYN.js";
-import { s as W } from "./sum-NLYbiDag.js";
-import { m as ct } from "./mat_mul-dwmZz69e.js";
-import { s as Qt } from "./split-CRU0PjVV.js";
-import { s as Zu, c as ar } from "./sin-CiBxrDqX.js";
-import { e as Hn, g as lr, h as cs, c as Xu } from "./axis_util-5DTW2tFV.js";
-import { a as se, e as ie, l as Yu } from "./log_sum_exp-BnmCkHWl.js";
-import { s as Dn } from "./stack-ikk2Y8_P.js";
-import { o as xe } from "./ones-C8Mfln6-.js";
-import { s as Dt } from "./slice-BHbDHObE.js";
-import { M as Qu, f as ur, r as tc, d as ec, a as $n } from "./dropout-DPfPgWWe.js";
-import { z as vt } from "./zeros-DnPT2nD4.js";
-import { c as pe } from "./concat-CbXTetof.js";
-import { g as cr } from "./gather-Bbh8DHhM.js";
-import { s as hr } from "./softmax-JMEIUo2J.js";
-import { m as Ee } from "./max-ByjEGoFx.js";
-import { t as nc } from "./tensor-Do9PKbIE.js";
-import { r as sc } from "./range-D9CZhVlR.js";
-import { v as ic } from "./variable-BTBkayv_.js";
+import { B as T, C as I, E as O, bE as Oa, bF as Ma, bG as Ci, l as b, bH as Ii, X as L, bI as Di, bJ as $i, bK as Ti, bL as zi, h as Ei, bM as Li, bN as Fi, bO as Oi, bP as Mi, bQ as _a, bR as _i, bS as Ra, bT as Ri, bU as Ba, bV as Bi, I as Ge, k as kt, bq as Wa, bW as Wi, bX as Gi, bY as Pi, y as Pe, c as V, a as w, bZ as Ga, b_ as Ui, b$ as Vi, c0 as ji, p as ce, aU as pt, bv as Pa, c1 as Ki, c2 as Hi, c3 as qi, c4 as Ji, c5 as Zi, bx as Xi, c6 as Yi, c7 as Qi, F as Ua, al as Va, aq as ja, c8 as tr, c9 as Ka, o as z, ca as Ps, cb as Ha, cc as qa, j as pn, cd as Us, ce as Ja, cf as Za, cg as Xa, ch as Ya, ci as Qa, cj as tl, ck as el, bf as nl, cl as sl, q as he, b as et, n as U, cm as il, bm as rl, aw as ht, cn as ol, z as Q, co as al, cp as ll, cq as ul, cr as cl, cs as hl, ct as pl, cu as dl, cv as fl, J as ml, cw as gl, bj as bl, bp as yl, cx as wl, Q as kl, cy as xl, a8 as Nl, cz as vl, cA as Al, cB as Sl, as as Cl, cC as Il, ae as Dl, aV as $l, br as Tl, ak as zl, bs as El, G as Ll, aX as Fl, am as Ol, cD as Ml, cE as _l, cF as Rl, an as Bl, af as Wl, cG as Gl, cH as Pl, cI as Ul, U as Vl, bt as jl, cJ as Kl, cK as Hl, aS as ql, b1 as Jl, cL as Zl, cM as Xl, bu as Yl, a$ as Ql, P as tu, cN as eu, x as rs, ap as nu, bw as su, az as iu, V as ru, au as ou, at as au, K as lu, bc as uu, cO as cu, bd as hu, cP as pu, b3 as du, aQ as fu, ar as mu, cQ as gu, a9 as bu, W as yu, S as wu, L as ku, bz as xu, cR as Nu, bA as vu, av as Au, bC as Su, Y as Cu, cS as Iu, N as Du, b5 as $u, b4 as Tu, cT as Oe, cU as zu, i as Eu, aj as Vs, cV as Lu, t as x, aT as $e, cW as S, cX as He, cY as qe, ab as Vt, d as J, ac as Fu, cZ as js, w as Jt, M as Ou, T as Te, O as Mu, c_ as _u, m as Ks, c$ as Ru, d0 as Hs, d1 as Bu } from "./index-BoWRt-10.js";
+import { k as ke, c as Nt, o as ze, s as er, b as nr, d as Wu, m as sr, t as In, l as ir, v as os, a as Gu, S as Pu, p as Uu, w as as, x as rr, y as Je, z as Vu, A as ju } from "./selu_util-BJEXVvjX.js";
+import { n as wt, w as ne, a as Ze, g as Xe, b as ls, t as K, c as Ce, d as Xt, e as Ku, u as dn, f as ye, h as fn, i as Hu, l as qu, s as us, m as or, j as qt, k as Ju } from "./ops-BFGCx8Ri.js";
+import { r as N } from "./reshape-CdBq1WJ6.js";
+import { s as W } from "./sum-66ew2byf.js";
+import { m as ct } from "./mat_mul-8m8pfdcx.js";
+import { s as Qt } from "./split-B-dikLRw.js";
+import { s as Zu, c as ar } from "./sin-BeA3tsEd.js";
+import { e as Hn, g as lr, h as cs, c as Xu } from "./axis_util-BzbKo31C.js";
+import { a as se, e as ie, l as Yu } from "./log_sum_exp-DbjkV734.js";
+import { s as Dn } from "./stack-B17UN2nn.js";
+import { o as xe } from "./ones-Dj0SDhHf.js";
+import { s as Dt } from "./slice-BiOsknYS.js";
+import { M as Qu, f as ur, r as tc, d as ec, a as $n } from "./dropout-DYs5QFGQ.js";
+import { z as vt } from "./zeros--BdLQ3oG.js";
+import { c as pe } from "./concat-CsxrgovM.js";
+import { g as cr } from "./gather-CMMy2KEG.js";
+import { s as hr } from "./softmax-Bv_6lyMX.js";
+import { m as Ee } from "./max-Ddnnb5xe.js";
+import { t as nc } from "./tensor-JwS7ZYY6.js";
+import { r as sc } from "./range-CRuAh-gd.js";
+import { v as ic } from "./variable-BuddVFLa.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.

package/dist/{range-D9CZhVlR.js → range-CRuAh-gd.js}

@@ -1,4 +1,4 @@
-import { E as e, R as f } from "./index-UdZhlibC.js";
+import { E as e, R as f } from "./index-BoWRt-10.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.

package/dist/{reciprocal-CGB48wZB.js → reciprocal-BvGAyKyu.js}

@@ -1,4 +1,4 @@
-import { B as c, C as n, E as e, V as s } from "./index-UdZhlibC.js";
+import { B as c, C as n, E as e, V as s } from "./index-BoWRt-10.js";
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.