星桥贯双境，光算入原生 · Complete native WebGPU integration

.github/workflows/core.yml

@@ -24,15 +24,19 @@ jobs:
           - id: macos-arm64
             runner: macos-15
             executable_suffix: ""
+            binary_dir: build-ci/bin
           - id: macos-x64
             runner: macos-15-intel
             executable_suffix: ""
+            binary_dir: build-ci/bin
           - id: windows-x64
             runner: windows-2022
             executable_suffix: ".exe"
+            binary_dir: build-ci/bin/Release
           - id: linux-x64
             runner: ubuntu-24.04
             executable_suffix: ""
+            binary_dir: build-ci/bin
     runs-on: ${{ matrix.runner }}
     steps:
       - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
@@ -59,7 +63,7 @@ jobs:
         shell: bash
         run: >-
           python tools/run_offline_check.py
-          --validate build-ci/bin/light_ocr_validate${{ matrix.executable_suffix }}
+          --validate ${{ matrix.binary_dir }}/light_ocr_validate${{ matrix.executable_suffix }}
           --bundle models/generated/ppocrv6-small-onnx-20260714.2
           --fixture corpus/fixtures/generated-hello-123/fixture.json
       - name: Runtime check with the network namespace disabled
@@ -75,7 +79,7 @@ jobs:
       - name: Repeated lifecycle RSS gate
         shell: bash
         run: >-
-          ./build-ci/bin/light_ocr_leak_check${{ matrix.executable_suffix }}
+          ./${{ matrix.binary_dir }}/light_ocr_leak_check${{ matrix.executable_suffix }}
           --bundle models/generated/ppocrv6-small-onnx-20260714.2
           --pixels corpus/fixtures/generated-hello-123/pixels.bin
           --width 800 --height 180 --stride 2400 --format bgr8
@@ -88,9 +92,9 @@ jobs:
         with:
           name: light-ocr-core-${{ matrix.id }}
           path: |
-            build-ci/bin/light_ocr_*
-            build-ci/bin/onnxruntime*
-            build-ci/bin/libonnxruntime*
+            ${{ matrix.binary_dir }}/light_ocr_*
+            ${{ matrix.binary_dir }}/onnxruntime*
+            ${{ matrix.binary_dir }}/libonnxruntime*
             build-ci/**/light_ocr_core.lib
             build-ci/**/liblight_ocr_core.a
             reports/release/${{ matrix.id }}

.github/workflows/npm-release.yml

@@ -165,6 +165,7 @@ jobs:
           python tools/npm_release.py stage-native \
             --platform-id "${{ matrix.id }}" \
             --build-dir build-npm \
+            --configuration Release \
             --metadata-dir "reports/npm/${{ matrix.id }}" \
             --output-dir "dist/native-input/${{ matrix.id }}"
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
@@ -183,6 +184,31 @@ jobs:
           if-no-files-found: error
           retention-days: 90
 
+  reproduce-webgpu-models:
+    needs: validate
+    runs-on: ubuntu-24.04
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065
+        with:
+          python-version: "3.11"
+      - name: Install the hash-locked WebGPU model toolchain
+        run: python -m pip install --require-hashes -r tools/webgpu/requirements.lock
+      - name: Reproduce and verify the deterministic WebGPU models
+        shell: bash
+        run: |
+          python tools/bootstrap_models.py --cache-dir .cache/models
+          python tools/package_model_bundle.py
+          python tools/webgpu/convert_models.py \
+            --output .cache/reproduced/webgpu-fp16-20260719.1
+          cmp models/derived/webgpu-fp16-20260719.1/det/inference.onnx \
+            .cache/reproduced/webgpu-fp16-20260719.1/det/inference.onnx
+          cmp models/derived/webgpu-fp16-20260719.1/rec/inference.onnx \
+            .cache/reproduced/webgpu-fp16-20260719.1/rec/inference.onnx
+          cmp models/derived/webgpu-fp16-20260719.1/provenance.json \
+            .cache/reproduced/webgpu-fp16-20260719.1/provenance.json
+
   derive-apple-models:
     needs: validate
     runs-on: macos-15
@@ -208,7 +234,7 @@ jobs:
           retention-days: 30
 
   assemble:
-    needs: [build-native, derive-apple-models]
+    needs: [build-native, reproduce-webgpu-models, derive-apple-models]
     runs-on: ubuntu-24.04
     env:
       RELEASE_VERSION: ${{ inputs.version }}
@@ -228,19 +254,20 @@ jobs:
         with:
           name: npm-apple-fp16-models
           path: models/generated/apple-fp16-20260715.1
-      - name: Bootstrap and package the self-contained Apple model bundle
+      - name: Bootstrap and package the self-contained native model bundle
         shell: bash
         run: |
           python tools/bootstrap_models.py --cache-dir .cache/models
           python tools/package_model_bundle.py
+          python tools/webgpu/package_bundle.py
           python tools/apple/package_bundle.py \
             --qualification-report contracts/apple-provider-baselines.json
       - name: Assemble and deterministically pack six packages
         shell: bash
         run: |
           python tools/npm_release.py assemble \
             --version "$RELEASE_VERSION" \
-            --bundle models/generated/ppocrv6-small-apple-20260715.1 \
+            --bundle models/generated/ppocrv6-small-native-20260719.1 \
             --native-root dist/native-input \
             --output-dir dist/npm/staging
           python tools/npm_release.py pack \

.github/workflows/tiled-qualification.yml

@@ -27,15 +27,23 @@ jobs:
           - id: macos-arm64
             runner: macos-15
             executable_suffix: ""
+            binary_dir: build-qualification/bin
+            node_runtime_dir: build-qualification/node-runtime
           - id: macos-x64
             runner: macos-15-intel
             executable_suffix: ""
+            binary_dir: build-qualification/bin
+            node_runtime_dir: build-qualification/node-runtime
           - id: linux-x64
             runner: ubuntu-24.04
             executable_suffix: ""
+            binary_dir: build-qualification/bin
+            node_runtime_dir: build-qualification/node-runtime
           - id: windows-x64
             runner: windows-2022
             executable_suffix: ".exe"
+            binary_dir: build-qualification/bin/Release
+            node_runtime_dir: build-qualification/node-runtime/Release
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 180
     steps:
@@ -123,8 +131,9 @@ jobs:
           bundle="$PWD/models/generated/ppocrv6-small-onnx-20260714.2"
           root="$PWD/reports/qualification/${{ matrix.id }}"
           executable="${{ matrix.executable_suffix }}"
+          binary_dir="${{ matrix.binary_dir }}"
           fixture_root="$PWD/corpus/tiled-v1/fixtures"
-          ./build-qualification/bin/light_ocr_leak_check${executable} \
+          ./${binary_dir}/light_ocr_leak_check${executable} \
             --bundle "$bundle" \
             --pixels "$fixture_root/tiled-four-way-intersection-2048/pixels.bin" \
             --width 2048 --height 2048 --stride 6144 --format bgr8 \
@@ -136,14 +145,14 @@ jobs:
             tiled-four-way-intersection-2048
           do
             python oracle/run_tiled_benchmark.py \
-              --native-benchmark "build-qualification/bin/light_ocr_benchmark${executable}" \
+              --native-benchmark "${binary_dir}/light_ocr_benchmark${executable}" \
               --bundle "$bundle" \
               --fixture "$fixture_root/$fixture/fixture.json" \
               --platform-id "${{ matrix.id }}" \
               --build-metadata "$root/build-manifest.json" \
               --report "$root/core/$fixture.json"
             for diagnostics in on off; do
-              ./build-qualification/bin/light_ocr_memory_gate${executable} \
+              ./${binary_dir}/light_ocr_memory_gate${executable} \
                 --bundle "$bundle" \
                 --pixels "$fixture_root/$fixture/pixels.bin" \
                 --width 2048 --height 2048 --stride 6144 --format bgr8 \
@@ -156,7 +165,8 @@ jobs:
       - name: Capture Node 22 latency and peak reports
         shell: bash
         env:
-          LIGHT_OCR_NODE_BINARY: ${{ github.workspace }}/build-qualification/bin/light_ocr_node.node
+          LIGHT_OCR_NODE_BINARY: ${{ github.workspace }}/${{ matrix.node_runtime_dir }}/native/light_ocr_node.node
+          LIGHT_OCR_RUNTIME_DESCRIPTOR: ${{ github.workspace }}/${{ matrix.node_runtime_dir }}/native/runtime-descriptor.json
           LIGHT_OCR_MODEL_BUNDLE: ${{ github.workspace }}/models/generated/ppocrv6-small-onnx-20260714.2
         run: |
           set -euo pipefail
@@ -182,7 +192,8 @@ jobs:
       - name: Capture Node 24 latency and peak reports
         shell: bash
         env:
-          LIGHT_OCR_NODE_BINARY: ${{ github.workspace }}/build-qualification/bin/light_ocr_node.node
+          LIGHT_OCR_NODE_BINARY: ${{ github.workspace }}/${{ matrix.node_runtime_dir }}/native/light_ocr_node.node
+          LIGHT_OCR_RUNTIME_DESCRIPTOR: ${{ github.workspace }}/${{ matrix.node_runtime_dir }}/native/runtime-descriptor.json
           LIGHT_OCR_MODEL_BUNDLE: ${{ github.workspace }}/models/generated/ppocrv6-small-onnx-20260714.2
         run: |
           set -euo pipefail

.github/workflows/webgpu-native.yml

@@ -0,0 +1,185 @@
+name: native webgpu
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: native-webgpu-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  contract:
+    name: ${{ matrix.id }} contract
+    timeout-minutes: 90
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - id: linux-x64
+            runner: ubuntu-24.04
+            node_runtime: build-webgpu/node-runtime/native
+          - id: windows-x64
+            runner: windows-2022
+            node_runtime: build-webgpu/node-runtime/Release/native
+    runs-on: ${{ matrix.runner }}
+    steps:
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065
+        with:
+          python-version: "3.11"
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020
+        with:
+          node-version: "22"
+      - name: Validate contract unit tests
+        run: >-
+          python -m unittest
+          tests.python.test_webgpu_runtime
+          tests.python.test_webgpu_models
+          tests.python.test_npm_release
+          tests.python.test_webgpu_qualification
+          tests.python.test_webgpu_report_review
+      - name: Validate the checked-in real-device report pair
+        if: matrix.id == 'linux-x64'
+        shell: bash
+        run: >-
+          python tools/webgpu/review_reports.py
+          --reports-root reports/webgpu-qualification
+          --output "$RUNNER_TEMP/webgpu-review-candidate.json"
+      - name: Bootstrap common dependencies and model
+        shell: bash
+        run: |
+          python tools/bootstrap_dependencies.py --cache-dir .cache/dependencies
+          python tools/bootstrap_dependencies.py --cache-dir .cache/dependencies --offline
+          python tools/bootstrap_models.py --cache-dir .cache/models
+          python tools/package_model_bundle.py
+          python tools/webgpu/package_bundle.py
+      - name: Reproduce the locked internal FP16 derivations (Linux)
+        if: runner.os != 'Windows'
+        shell: bash
+        run: |
+          python -m pip install --require-hashes -r tools/webgpu/requirements.lock
+          python tools/webgpu/convert_models.py \
+            --output .cache/reproduced/webgpu-fp16-20260719.1
+          cmp models/derived/webgpu-fp16-20260719.1/det/inference.onnx \
+            .cache/reproduced/webgpu-fp16-20260719.1/det/inference.onnx
+          cmp models/derived/webgpu-fp16-20260719.1/rec/inference.onnx \
+            .cache/reproduced/webgpu-fp16-20260719.1/rec/inference.onnx
+          cmp models/derived/webgpu-fp16-20260719.1/provenance.json \
+            .cache/reproduced/webgpu-fp16-20260719.1/provenance.json
+      - name: Assemble and revalidate the locked SDK
+        shell: bash
+        run: |
+          python tools/webgpu/build_runtime.py \
+            --platform "${{ matrix.id }}" \
+            --package-cache .cache/webgpu-packages \
+            --output-dir "dist/webgpu-sdk/${{ matrix.id }}"
+          python tools/webgpu/build_runtime.py \
+            --validate-sdk "dist/webgpu-sdk/${{ matrix.id }}"
+          python tools/webgpu/build_runtime.py \
+            --platform "${{ matrix.id }}" \
+            --offline \
+            --package-cache .cache/webgpu-packages \
+            --output-dir "dist/webgpu-sdk-offline/${{ matrix.id }}"
+      - name: Install Node development files (Linux)
+        if: runner.os != 'Windows'
+        shell: bash
+        run: |
+          node_version="$(node -p process.versions.node)"
+          npx --yes node-gyp@11.4.2 install "$node_version" --devdir "$PWD/.cache/node-gyp"
+          node_root="$PWD/.cache/node-gyp/$node_version"
+          test -f "$node_root/include/node/node_api.h"
+          echo "NODE_INCLUDE_DIR=$node_root/include/node" >> "$GITHUB_ENV"
+      - name: Install Node development files (Windows)
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: |
+          $nodeVersion = node -p process.versions.node
+          $devDir = Join-Path $env:GITHUB_WORKSPACE ".cache/node-gyp"
+          npx --yes node-gyp@11.4.2 install $nodeVersion --devdir $devDir
+          $nodeRoot = Join-Path $devDir $nodeVersion
+          $headers = Join-Path $nodeRoot "include/node"
+          $library = Get-ChildItem $nodeRoot -Recurse -Filter node.lib | Select-Object -First 1
+          if (-not (Test-Path (Join-Path $headers "node_api.h")) -or $null -eq $library) {
+            throw "Verified Node development files are incomplete"
+          }
+          "NODE_INCLUDE_DIR=$headers" | Out-File -FilePath $env:GITHUB_ENV -Append
+          "NODE_LIBRARY=$($library.FullName)" | Out-File -FilePath $env:GITHUB_ENV -Append
+      - name: Configure WebGPU qualification build (Linux)
+        if: runner.os != 'Windows'
+        shell: bash
+        run: >-
+          cmake -S . -B build-webgpu -G Ninja
+          -DCMAKE_BUILD_TYPE=Release
+          -DLIGHT_OCR_DEPENDENCY_CACHE_DIR="$PWD/.cache/dependencies"
+          -DLIGHT_OCR_ONNXRUNTIME_FLAVOR=webgpu
+          -DLIGHT_OCR_WEBGPU_SDK_DIR="$PWD/dist/webgpu-sdk/${{ matrix.id }}"
+          -DLIGHT_OCR_WEBGPU_QUALIFICATION_BUILD=ON
+          -DLIGHT_OCR_BUILD_NODE=ON
+          -DLIGHT_OCR_BUILD_TESTS=ON
+          -DLIGHT_OCR_BUILD_TOOLS=ON
+          -DLIGHT_OCR_NODE_INCLUDE_DIR="$NODE_INCLUDE_DIR"
+          -DLIGHT_OCR_NODE_EXECUTABLE="$(command -v node)"
+      - name: Configure WebGPU qualification build (Windows)
+        if: runner.os == 'Windows'
+        shell: pwsh
+        run: >-
+          cmake -S . -B build-webgpu -G "Visual Studio 17 2022" -A x64
+          "-DLIGHT_OCR_DEPENDENCY_CACHE_DIR=$env:GITHUB_WORKSPACE/.cache/dependencies"
+          -DLIGHT_OCR_ONNXRUNTIME_FLAVOR=webgpu
+          "-DLIGHT_OCR_WEBGPU_SDK_DIR=$env:GITHUB_WORKSPACE/dist/webgpu-sdk/${{ matrix.id }}"
+          -DLIGHT_OCR_WEBGPU_QUALIFICATION_BUILD=ON
+          -DLIGHT_OCR_BUILD_NODE=ON
+          -DLIGHT_OCR_BUILD_TESTS=ON
+          -DLIGHT_OCR_BUILD_TOOLS=ON
+          "-DLIGHT_OCR_NODE_INCLUDE_DIR=$env:NODE_INCLUDE_DIR"
+          "-DLIGHT_OCR_NODE_LIBRARY=$env:NODE_LIBRARY"
+          "-DLIGHT_OCR_NODE_EXECUTABLE=$((Get-Command node).Source)"
+      - name: Compile, link, and run hardware-independent tests
+        shell: bash
+        run: |
+          cmake --build build-webgpu --config Release --parallel
+          ctest --test-dir build-webgpu -C Release --output-on-failure -LE node
+      - name: Verify the staged addon and descriptor from a sterile directory
+        shell: bash
+        env:
+          LIGHT_OCR_NODE_BINARY: ${{ github.workspace }}/${{ matrix.node_runtime }}/light_ocr_node.node
+          LIGHT_OCR_RUNTIME_DESCRIPTOR: ${{ github.workspace }}/${{ matrix.node_runtime }}/runtime-descriptor.json
+        run: |
+          sterile="$(mktemp -d)"
+          cd "$sterile"
+          node -e "const x=require(process.env.GITHUB_WORKSPACE+'/bindings/node/js/load-native.cjs').loadNative(); console.log(JSON.stringify(x.binding.runtimeContract))"
+      - name: Generate license, SBOM, and native package input
+        shell: bash
+        run: |
+          python tools/generate_release_metadata.py \
+            --build-dir build-webgpu \
+            --configuration Release \
+            --platform-id "${{ matrix.id }}" \
+            --output-dir "reports/webgpu/${{ matrix.id }}"
+          python tools/npm_release.py stage-native \
+            --platform-id "${{ matrix.id }}" \
+            --build-dir build-webgpu \
+            --configuration Release \
+            --metadata-dir "reports/webgpu/${{ matrix.id }}" \
+            --output-dir "dist/webgpu-native/${{ matrix.id }}" \
+            --runtime-flavor webgpu \
+            --webgpu-artifact-manifest "dist/webgpu-sdk/${{ matrix.id }}/artifact-manifest.json" \
+            --qualification-build
+          LIGHT_OCR_NODE_BINARY="$PWD/dist/webgpu-native/${{ matrix.id }}/native/light_ocr_node.node" \
+          LIGHT_OCR_RUNTIME_DESCRIPTOR="$PWD/dist/webgpu-native/${{ matrix.id }}/native/runtime-descriptor.json" \
+          node -e "const x=require('./bindings/node/js/load-native.cjs').loadNative(); console.log(JSON.stringify(x.runtimePolicy))"
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
+        with:
+          name: native-webgpu-${{ matrix.id }}
+          path: |
+            dist/webgpu-sdk/${{ matrix.id }}
+            dist/webgpu-native/${{ matrix.id }}
+            reports/webgpu/${{ matrix.id }}
+          if-no-files-found: error
+          retention-days: 30