sync 91d990483a17

Browse files

Files changed (5) hide show

README.md +17 -14
build/webgpu/bench.json +0 -1
build/webgpu/manifest.json +21 -102
build/webgpu/metadata.json +6 -6
build/webgpu/test.json +5 -6

README.md CHANGED Viewed

@@ -18,19 +18,19 @@ See the [standard ONNX `Loop` spec](https://onnx.ai/onnx/operators/onnx__Loop.ht
 ## Inputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
-| `M` | `m` | `I` | — | — | Required uint32 trip-count limit encoded as either a scalar tensor or a one-element rank-1 tensor. The lowering executes at most `min(M, scan_output.shape[0])` iterations. | required |
-| `cond` | `cond` | `B` | — | — | Required initial condition encoded as either a scalar tensor or a one-element rank-1 tensor. A false value executes zero iterations; a true value permits all iterations selected by `M`. This fixed lowering does not update the condition inside the loop. | required |
-| `v_initial` | `v_initial` | `T` | `1` | — | Initial rank-1 loop-carried state of shape `[dim]`. | required |
-| `step` | `step` | `T` | `1` | — | Implementation-specific rank-1 increment of shape `[dim]`, added elementwise to the state on every executed iteration. This is not a standard ONNX `Loop` input. | required |
 ## Outputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- |
-| `v_final` | `v_final` | `T` | `1` | same as `v_initial` | Final rank-1 state of shape `[dim]` after the executed additions. | required |
-| `scan_output` | `scan_output` | `T` | `2` | — | Dense tensor of shape `[scan_steps, dim]`. Each executed row contains the updated state for that iteration; rows beyond the executed trip count are zero-filled. | required |
 ## Type constraints
@@ -42,7 +42,7 @@ See the [standard ONNX `Loop` spec](https://onnx.ai/onnx/operators/onnx__Loop.ht
 ## Files
-- [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, provenance)
 - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth)
 - [`test.json`](build/webgpu/test.json) — correctness cases
 - [`bench.json`](build/webgpu/bench.json) — benchmark + tuning cases
@@ -50,15 +50,18 @@ See the [standard ONNX `Loop` spec](https://onnx.ai/onnx/operators/onnx__Loop.ht
 ## Use with `@huggingface/kernels`
-The loader automatically allocates outputs whose metadata it can derive from the manifest contract and this call.
-The explicit `outputs` entries provide shape and logical dtype metadata for the results listed below:
 - `scan_output`
-Each entry either requests an optional result or supplies metadata that cannot be inferred from the inputs.
 The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version.
 Replace each `*Data` placeholder with a typed array containing the corresponding input data.

 ## Inputs
+| Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
+| `m` | `M` | `I` | — | — | Required uint32 trip-count limit encoded as either a scalar tensor or a one-element rank-1 tensor. The lowering executes at most `min(M, scan_output.shape[0])` iterations. | required |
+| `cond` | — | `B` | — | — | Required initial condition encoded as either a scalar tensor or a one-element rank-1 tensor. A false value executes zero iterations; a true value permits all iterations selected by `M`. This fixed lowering does not update the condition inside the loop. | required |
+| `v_initial` | — | `T` | `1` | — | Initial rank-1 loop-carried state of shape `[dim]`. | required |
+| `step` | — | `T` | `1` | — | Implementation-specific rank-1 increment of shape `[dim]`, added elementwise to the state on every executed iteration. This is not a standard ONNX `Loop` input. | required |
 ## Outputs
+| Name | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- |
+| `v_final` | `T` | `1` | same as `v_initial` | Final rank-1 state of shape `[dim]` after the executed additions. | required |
+| `scan_output` | `T` | `2` | — | Dense tensor of shape `[scan_steps, dim]`. Each executed row contains the updated state for that iteration; rows beyond the executed trip count are zero-filled. | required |
 ## Type constraints
 ## Files
+- [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, provenance)
 - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth)
 - [`test.json`](build/webgpu/test.json) — correctness cases
 - [`bench.json`](build/webgpu/bench.json) — benchmark + tuning cases
 ## Use with `@huggingface/kernels`
+```sh
+npm install --save-exact @huggingface/kernels@0.0.1-preview.2
+```
+Outputs with inferable metadata are allocated automatically. Explicit `outputs` entries request optional results or provide metadata that cannot be inferred from the supplied inputs and attributes.
+This example supplies explicit metadata for:
 - `scan_output`
 The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version.
+It follows the `v1` branch as fixes land. To pin exact artifact bytes, pass a 40-character commit `revision` instead of `version`.
 Replace each `*Data` placeholder with a typed array containing the corresponding input data.

build/webgpu/bench.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "ai.onnx.Loop",
   "cases": [
     {
       "name": "lowered_HEALTHY_baseline_dim128_steps128",

 {
   "cases": [
     {
       "name": "lowered_HEALTHY_baseline_dim128_steps128",

build/webgpu/manifest.json CHANGED Viewed

@@ -3,59 +3,21 @@
   "name": "Loop",
   "conformance": "internal-lowering",
   "sinceVersion": 25,
-  "description": "Support status: the standard ONNX `Loop` control-flow operator is not implemented because standalone kernel packages cannot carry or execute its body graph. This internal lowering performs one fixed recurrence—adding `step` to a rank-1 loop-carried tensor for up to `M` iterations while writing a dense scan tensor; `step` is not an ONNX `Loop` input, and this lowering must not be treated as ONNX `Loop`.",
-  "inputs": [
-    {
-      "role": "M",
-      "dtype": "I",
-      "description": "Required uint32 trip-count limit encoded as either a scalar tensor or a one-element rank-1 tensor. The lowering executes at most `min(M, scan_output.shape[0])` iterations."
-    },
-    {
-      "role": "cond",
-      "dtype": "B",
-      "description": "Required initial condition encoded as either a scalar tensor or a one-element rank-1 tensor. A false value executes zero iterations; a true value permits all iterations selected by `M`. This fixed lowering does not update the condition inside the loop."
-    },
-    {
-      "role": "v_initial",
-      "dtype": "T",
-      "rank": 1,
-      "description": "Initial rank-1 loop-carried state of shape `[dim]`."
-    },
-    {
-      "role": "step",
-      "dtype": "T",
-      "rank": 1,
-      "description": "Implementation-specific rank-1 increment of shape `[dim]`, added elementwise to the state on every executed iteration. This is not a standard ONNX `Loop` input."
-    }
-  ],
-  "outputs": [
-    {
-      "role": "v_final",
-      "dtype": "T",
-      "rank": 1,
-      "description": "Final rank-1 state of shape `[dim]` after the executed additions.",
-      "shape": "shapes.v_initial"
-    },
-    {
-      "role": "scan_output",
-      "dtype": "T",
-      "rank": 2,
-      "description": "Dense tensor of shape `[scan_steps, dim]`. Each executed row contains the updated state for that iteration; rows beyond the executed trip count are zero-filled."
-    }
-  ],
-  "typeConstraints": { "T": ["float32"], "I": ["uint32"], "B": ["uint32", "bool"] },
-  "args": {
-    "m": { "kind": "tensor", "semantic": "M", "role": "input" },
-    "cond": { "kind": "tensor", "semantic": "cond", "role": "input" },
-    "v_initial": { "kind": "tensor", "semantic": "v_initial", "role": "input" },
-    "step": { "kind": "tensor", "semantic": "step", "role": "input" },
-    "v_final": { "kind": "tensor", "semantic": "v_final", "role": "output" },
-    "scan_output": { "kind": "tensor", "semantic": "scan_output", "role": "output" }
   },
   "variants": [
     {
       "id": "lowered_add_step",
-      "when": ["ranks.M <= 1", "numel(shapes.M) == 1", "ranks.cond <= 1", "numel(shapes.cond) == 1", "ranks.v_initial == 1", "ranks.step == 1", "ranks.v_final == 1", "ranks.scan_output == 2", "dim(shapes.step, 0) == dim(shapes.v_initial, 0)", "dim(shapes.v_final, 0) == dim(shapes.v_initial, 0)", "dim(shapes.scan_output, 1) == dim(shapes.v_initial, 0)"],
       "derive": { "dimBlocks": "ceil(dim(shapes.v_initial, 0) / 256)" },
       "passes": [
         {
@@ -63,61 +25,18 @@
           "name": "Loop",
           "shader": "loop-add-step.wgsl.jinja",
           "bindings": [
-            {
-              "name": "m",
-              "arg": "m",
-              "semantic": "M",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "u32",
-              "length": 1
-            },
-            {
-              "name": "cond",
-              "arg": "cond",
-              "semantic": "cond",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "u32",
-              "length": 1
-            },
-            {
-              "name": "v_initial",
-              "arg": "v_initial",
-              "semantic": "v_initial",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "f32"
-            },
-            {
-              "name": "step",
-              "arg": "step",
-              "semantic": "step",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "f32"
-            },
-            {
-              "name": "v_final",
-              "arg": "v_final",
-              "semantic": "v_final",
-              "buffer": { "type": "storage" },
-              "elementType": "f32"
-            },
-            {
-              "name": "scan_output",
-              "arg": "scan_output",
-              "semantic": "scan_output",
-              "buffer": { "type": "storage" },
-              "elementType": "f32"
-            },
             {
               "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": {
-                "name": "Params",
-                "fields": [
-                  { "name": "dim", "type": "u32", "value": "dim(shapes.v_initial, 0)" },
-                  { "name": "scanSteps", "type": "u32", "value": "dim(shapes.scan_output, 0)" }
-                ]
-              }
             }
           ],
           "dispatch": { "x": "dimBlocks" }

   "name": "Loop",
   "conformance": "internal-lowering",
   "sinceVersion": 25,
+  "inputs": {
+    "m": { "onnx": "M", "dtype": "I" },
+    "cond": { "dtype": "B" },
+    "v_initial": { "dtype": "T", "rank": 1 },
+    "step": { "dtype": "T", "rank": 1 }
   },
+  "outputs": {
+    "v_final": { "dtype": "T", "rank": 1, "shape": "shapes.v_initial" },
+    "scan_output": { "dtype": "T", "rank": 2 }
+  },
+  "typeConstraints": { "T": ["float32"], "I": ["uint32"], "B": ["uint32", "bool"] },
   "variants": [
     {
       "id": "lowered_add_step",
+      "when": ["ranks.m <= 1", "numel(shapes.m) == 1", "ranks.cond <= 1", "numel(shapes.cond) == 1", "ranks.v_initial == 1", "ranks.step == 1", "ranks.v_final == 1", "ranks.scan_output == 2", "dim(shapes.step, 0) == dim(shapes.v_initial, 0)", "dim(shapes.v_final, 0) == dim(shapes.v_initial, 0)", "dim(shapes.scan_output, 1) == dim(shapes.v_initial, 0)"],
       "derive": { "dimBlocks": "ceil(dim(shapes.v_initial, 0) / 256)" },
       "passes": [
         {
           "name": "Loop",
           "shader": "loop-add-step.wgsl.jinja",
           "bindings": [
+            { "arg": "m", "elementType": "u32", "length": 1 },
+            { "arg": "cond", "elementType": "u32", "length": 1 },
+            { "arg": "v_initial", "elementType": "f32" },
+            { "arg": "step", "elementType": "f32" },
+            { "arg": "v_final", "elementType": "f32" },
+            { "arg": "scan_output", "elementType": "f32" },
             {
               "name": "params",
+              "struct": [
+                { "name": "dim", "type": "u32", "value": "dim(shapes.v_initial, 0)" },
+                { "name": "scanSteps", "type": "u32", "value": "dim(shapes.scan_output, 0)" }
+              ]
             }
           ],
           "dispatch": { "x": "dimBlocks" }

build/webgpu/metadata.json CHANGED Viewed

@@ -1,18 +1,18 @@
 {
   "name": "ai.onnx.Loop",
-  "id": "_ai_onnx_loop_webgpu_fb2e7a5",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "bench.json": "xq5d0FI/A9ESZ8tyzz5qAxP39qxO5jGxCNji2YREKxI=",
       "loop-add-step.wgsl.jinja": "Sk4601cYmNLKCMIFX47m3mKo2M5oh/fekvOwLOUN1D8=",
-      "manifest.json": "C5Kc57/iLS4XSFrNuO6zRpUHfXkgOI8Hm+LlxTfs/eg=",
-      "test.json": "UG6yKjMh8XfBVE9AZSF2FfMZjBNwXajveageqJpYjWY="
     }
   },
-  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
-  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Loop" }
 }

 {
   "name": "ai.onnx.Loop",
+  "id": "_ai_onnx_loop_webgpu_801da86",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "bench.json": "ZSbqJEih0tA29HPs9gNfjrVWJeylslE0PXPC0uKM82s=",
       "loop-add-step.wgsl.jinja": "Sk4601cYmNLKCMIFX47m3mKo2M5oh/fekvOwLOUN1D8=",
+      "manifest.json": "H3mdjtpFSSHh+Osunjn3M82kXBVm5ga/DK+8Ha0Wwoc=",
+      "test.json": "Rp36C1l3Leiipq280ww8mCj4exFK53UlSD0thVu/2P0="
     }
   },
+  "provenance": { "kernel": { "sha": "91d990483a174128daf7673f3f37a7c890493ae1", "dirty": false } },
+  "webgpu": { "manifestSpec": "2.0", "variants": { "lowered_add_step": ["loop-add-step.wgsl.jinja"] } }
 }

build/webgpu/test.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "ai.onnx.Loop",
   "cases": [
     {
       "name": "lowered_add_step_scan",
@@ -32,7 +31,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.Opset11WithNoVariadicInputsAndOutputs",
-        "notes": "Projection onto the framework's lowered add-step loop: one true iteration produces one scan output value."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [1] } },
@@ -50,7 +49,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.ZeroIterations",
-        "notes": "Projection onto the framework's lowered add-step loop: trip count zero must leave carried state unchanged and produce no scan elements."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [0] } },
@@ -68,7 +67,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.ExitDueToMaxIterations",
-        "notes": "Projection onto the framework's lowered add-step loop: two iterations advance the carried state and produce two scan values."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [2] } },
@@ -138,7 +137,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.ExitDueToMaxIterations",
-        "notes": "ONNX Loop trip count and initial condition are scalar values. The project-lowered add-step form should accept scalar logical uint32 metadata as well as [1] tensors."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
@@ -192,7 +191,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.ExitDueToMaxIterations",
-        "notes": "ONNX only requires the loop condition to be a bool; runtimes that store it as an integer may emit any nonzero value. The lowered kernel must treat any nonzero cond (here 2) as true via cond[0] != 0u, matching the reference's !== 0 test, not a literal == 1 comparison."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [3] } },

 {
   "cases": [
     {
       "name": "lowered_add_step_scan",
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.Opset11WithNoVariadicInputsAndOutputs",
+        "notes": "The internal lowered add-step loop runs one true iteration and produces one scan output value."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [1] } },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.ZeroIterations",
+        "notes": "With a zero trip count, the internal lowered add-step loop leaves carried state unchanged and produces no scan elements."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [0] } },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.ExitDueToMaxIterations",
+        "notes": "The internal lowered add-step loop runs two iterations, advances carried state twice, and produces two scan values."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [2] } },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.ExitDueToMaxIterations",
+        "notes": "The fixed add-step lowering accepts scalar uint32 trip-count and condition metadata, applies three iterations, and emits one scan row per iteration."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [], "data": { "kind": "values", "values": [3] } },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/controlflow/loop_test.cc",
         "test": "Loop.ExitDueToMaxIterations",
+        "notes": "The lowered integer-backed condition must treat any nonzero value, including 2, as logical true rather than comparing only with 1."
       },
       "inputs": {
         "m": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [3] } },