轻芯贯卷，双路同归 · Complete lightweight AMD recognition backend

.github/workflows/npu-native.yml

@@ -4,7 +4,7 @@ on:
   workflow_dispatch:
     inputs:
       amd_sdk_run_id:
-        description: Optional run containing amdnpu-candidate-sdk imported from Ryzen AI 1.8
+        description: Optional run containing an amdnpu-candidate-sdk; otherwise build lightweight IREE support
         required: false
         default: ""
         type: string
@@ -20,19 +20,25 @@ permissions:
 
 jobs:
   build:
-    # Draft release assets are visible only to tokens with push access.
-    permissions:
-      contents: write
-      actions: read
     runs-on: ubuntu-24.04
+    timeout-minutes: 180
     steps:
       - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
       - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020
         with:
           node-version: "22"
       - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065
         with:
-          python-version: "3.11"
+          python-version: "3.12"
+      - name: Install AMD build dependencies
+        if: inputs.amd_sdk_run_id == ''
+        run: sudo apt-get update && sudo apt-get install -y clang lld ninja-build uuid-dev
+      - name: Cache pinned AMD build tools and model artifacts
+        if: inputs.amd_sdk_run_id == ''
+        uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830
+        with:
+          path: .cache/amd-aie
+          key: amd-aie-${{ runner.os }}-${{ hashFiles('tools/npu/iree-source.lock.json', 'tools/npu/iree-frontend.requirements.txt', 'tools/npu/*iree*.py', 'tools/npu/*amdaie*.py', 'tools/npu/runtime/*', 'src/inference/amdnpu/runtime_api.h') }}
       - name: Assemble pinned Intel native SDK
         run: python tools/npu/build_openvino.py --package-cache .cache/npu --output-dir dist/npu-sdk/openvino
       - name: Import acquired AMD candidate
@@ -43,11 +49,9 @@ jobs:
           path: dist/npu-sdk/amdnpu
           run-id: ${{ inputs.amd_sdk_run_id }}
           github-token: ${{ github.token }}
-      - name: Fetch locked AMD deployment runtime and compiled recognition models
+      - name: Build lightweight AMD runtime and all recognition buckets
         if: inputs.amd_sdk_run_id == ''
-        env:
-          GH_TOKEN: ${{ github.token }}
-        run: python tools/npu/fetch_amdnpu.py --package-cache .cache/npu --output-dir dist/npu-sdk/amdnpu
+        run: python tools/npu/build_amdaie.py --work-dir .cache/amd-aie --output-dir dist/npu-sdk/amdnpu --jobs 2
       - name: Build the native addon with packaged NPU SDKs
         shell: bash
         run: |

.github/workflows/npu-support-release.yml

@@ -19,10 +19,8 @@ permissions:
 
 jobs:
   build:
-    # The AMD deployment archive remains a draft asset until v0.5.9 ships.
-    permissions:
-      contents: write
     runs-on: ubuntu-24.04
+    timeout-minutes: 180
     strategy:
       fail-fast: false
       matrix:
@@ -34,10 +32,18 @@ jobs:
           node-version: '22'
       - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065
         with:
-          python-version: '3.11'
+          python-version: '3.12'
+      - name: Install AMD build dependencies
+        if: matrix.provider == 'amdnpu'
+        run: sudo apt-get update && sudo apt-get install -y clang lld ninja-build uuid-dev
+      - name: Cache pinned AMD build tools and model artifacts
+        if: matrix.provider == 'amdnpu'
+        uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830
+        with:
+          path: .cache/amd-aie
+          key: amd-aie-${{ runner.os }}-${{ hashFiles('tools/npu/iree-source.lock.json', 'tools/npu/iree-frontend.requirements.txt', 'tools/npu/*iree*.py', 'tools/npu/*amdaie*.py', 'tools/npu/runtime/*', 'src/inference/amdnpu/runtime_api.h') }}
       - name: Prepare the selected provider only
         env:
-          GH_TOKEN: ${{ github.token }}
           PROVIDER: ${{ matrix.provider }}
           RELEASE_VERSION: ${{ inputs.version }}
         run: |
@@ -46,7 +52,7 @@ jobs:
           if [[ "$PROVIDER" == openvino ]]; then
             python tools/npu/build_openvino.py --package-cache .cache/npu --output-dir dist/sdk
           else
-            python tools/npu/fetch_amdnpu.py --package-cache .cache/npu --output-dir dist/sdk
+            python tools/npu/build_amdaie.py --work-dir .cache/amd-aie --output-dir dist/sdk --jobs 2
           fi
       - name: Build a separate CPU and NPU addon
         env:
@@ -73,6 +79,17 @@ jobs:
             --sdk-dir dist/sdk --output-dir dist/package
           mkdir -p dist/tarballs
           npm pack ./dist/package --ignore-scripts --json --pack-destination dist/tarballs > dist/tarballs/pack-manifest.json
+          python tools/npu/record_support_pack.py --package-dir dist/package \
+            --pack-manifest dist/tarballs/pack-manifest.json --output-file dist/tarballs/support-size-report.json
+      - name: Test AMD package without NPU hardware
+        if: matrix.provider == 'amdnpu'
+        env:
+          LIGHT_OCR_AMD_TEST_PACKAGE: ${{ github.workspace }}/dist/package
+          LIGHT_OCR_AMD_TEST_BUNDLE: ${{ github.workspace }}/models/generated/ppocrv6-small-onnx-20260714.2
+          LIGHT_OCR_AMD_TEST_ARCHIVE: ${{ github.workspace }}/dist/tarballs/arcships-light-ocr-amdnpu-linux-x64-gnu-${{ inputs.version }}.tgz
+        run: |
+          node --test packages/runtime/test/*.test.cjs
+          python -m unittest tests.python.test_amdaie_runtime
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: npu-support-${{ matrix.provider }}-${{ inputs.version }}

CHANGELOG.md

@@ -21,14 +21,21 @@ Release preparation; npm publication and GitHub Release are pending.
   pinned OpenVINO 2026.4.0 libraries; default platform packages do not. See
   [Intel NPU 加速技术方案](docs/intel-npu-acceleration.md).
 - Added a separately installed AMD NPU backend (`provider: "amdnpu"`) for Linux x64
-  glibc and Ryzen AI STX/KRK devices. It isolates the vendor ORT runtime,
-  loads 20 immutable BF16 recognition contexts, and runs detection on CPU.
+  glibc and Ryzen AI STX/KRK devices. The default deployment now uses a small
+  IREE C ABI runtime, 20 CPU/NPU recognition modules and one shared weight archive.
+  It offloads 19 dense pointwise convolutions, splits wide matrices to fit DMA
+  limits, keeps other recognition operators on IREE CPU, and runs detection on ORT CPU.
   Added model compilation and SDK import tools, descriptor validation,
   license/SBOM staging, and candidate/release workflows. AMD is excluded
   from Auto and enabled only through `execution.provider: "amdnpu"` (or
   `--provider amdnpu`). NPU device reports are optional for release;
-  `deviceValidated` remains false. The optional AMD support package includes Ryzen AI 1.8
-  deployment libraries and real compiled contexts bound to Small 0.3.4.
+  `deviceValidated` remains false. BF16 operands use BFP16ebs8 microkernels;
+  numerical parity and hardware execution remain unverified. Deployment artifacts
+  are bound to Small 0.3.4, with no vendor SDK, compiler, Python or XRT in the AMD payload.
+  Legacy vendor deployments retain their separate ORT API/namespace path.
+  Added hardware-free package/C ABI regression checks and sampled CPU transform
+  comparisons; fixed null deployment JSON leaking an unclassified TypeError.
+  These checks do not qualify the AMD BFP16ebs8 kernel or device execution.
   Neither support package is automatically installed or selected by Auto.
   The existing Small model bundle remains unchanged. See [AMD NPU 加速技术方案](docs/amd-npu-acceleration.md)
   and [NPU SDK 构建与发布](docs/npu-runtime-release.md).

CMakeLists.txt

@@ -39,7 +39,7 @@ set(LIGHT_OCR_OPENVINO_SDK_DIR "" CACHE PATH
 set(LIGHT_OCR_OPENVINO_C_LIBRARY "" CACHE FILEPATH
   "OpenVINO C runtime loaded by direct C++ callers; defaults to <SDK>/lib/libopenvino_c.so")
 set(LIGHT_OCR_AMDNPU_SDK_DIR "" CACHE PATH
-  "Verified Ryzen AI runtime SDK with precompiled BF16 recognition buckets (Linux x64 glibc)")
+  "Verified AMD runtime deployment with precompiled recognition buckets (Linux x64 glibc)")
 
 if(LIGHT_OCR_ENABLE_SANITIZERS AND LIGHT_OCR_ENABLE_THREAD_SANITIZER)
   message(FATAL_ERROR "Address/Undefined sanitizers and ThreadSanitizer require separate builds")

docs/amd-aie-compilation-results.json

@@ -0,0 +1,197 @@
+{
+  "schemaVersion": "1.0",
+  "stage": "development",
+  "date": "2026-10-07",
+  "sourceModelSha256": "5435fd747c9e0efe15a96d0b378d5bd157e9492ed8fd80edf08f30d02fa24634",
+  "sourceLockSha256": "a58136fc60e8eb5023798c3e8d3a7014abbeb2f3107f6a727c5b8f1f0cd0bf4a",
+  "compiler": {
+    "sourceCommits": {
+      "amdAie": "fddfec1be6ceefbdb890079d957947dfa1fe0848",
+      "third_party/aie-rt": "a6196ebd905ceaf92cabc2e594017fa11b9e4323",
+      "third_party/bootgen": "92e09bf37ea17d7b1f0e102a2548f27fb768651c",
+      "third_party/iree": "26083307b8753a11870b8caf590afea15c9429fa",
+      "third_party/mlir-air": "ca2448548d66d7e8e05bacb9e81e7df48a9ef729",
+      "third_party/openssl": "0e04e101507df752500093ff4b04cce5a898abaa",
+      "iree/third_party/benchmark": "192ef10025eb2c4cdd392bc502f0c852196baa48",
+      "iree/third_party/llvm-project": "8be29edc20dfb19245ab8cf49669c80125ece4f5",
+      "iree/third_party/flatcc": "9362cd00f0007d8cbee7bff86e90fb4b6b227ff3",
+      "iree/third_party/musl": "3f701faace7addc75d16dea8a6cd769fa5b3f260"
+    },
+    "localPatches": [
+      {
+        "path": "patches/check-matmul-map-ranks.patch",
+        "sha256": "5a0a9b3efe57cc4d257b8fd9d08f59a4c158a07c71a7242d21b812314ee9fb05",
+        "reason": "Reject scalar/rank-one generic maps before matmul dimension access"
+      }
+    ],
+    "localSourceDiffSha256": "b62c2c5415db8739e8799ca8f1f28658c1e27229ad0897fafce505150d6cb6f9",
+    "binaries": [
+      {
+        "path": "tools/iree-compile",
+        "bytes": 6256,
+        "sha256": "a751017528d0e96e3331e119cfa87c9705e802a3fcc27c68b4a74f8afbe7e3f3"
+      },
+      {
+        "path": "tools/iree-opt",
+        "bytes": 6248,
+        "sha256": "5b7123611e66efbbc527c0dfb0ac225a1597193cf4bbdb69022fdc4c169366b4"
+      },
+      {
+        "path": "lib/libIREECompiler.so",
+        "bytes": 271528912,
+        "sha256": "352f1d6e4c7f3945f7fbbdf09255598cad153ec3afbac52f353241559cbc4408"
+      }
+    ]
+  },
+  "runtimeBinaries": [
+    {
+      "path": "tools/iree-run-module",
+      "bytes": 1843488,
+      "sha256": "b269f29d707f00a426f09a2fb962806b61e7c3a4edabe7da4b13d9057f8ee3f5"
+    },
+    {
+      "path": "tools/iree-dump-module",
+      "bytes": 462336,
+      "sha256": "83eb65201baa9071dd14af67e265b70b0b3239b76bcf6bb7835dd5ddb4e8d34b"
+    }
+  ],
+  "target": "npu4",
+  "deviceHAL": "amdxdna",
+  "coreRows": 2,
+  "coreCols": 2,
+  "deviceValidated": false,
+  "numericsValidated": false,
+  "fullRecognitionSupported": false,
+  "reports": {
+    "matrixTargets": {
+      "compileStage": "targets",
+      "status": "failed",
+      "kernelMode": "peano-bf16",
+      "kernelPrecision": "bf16 converted to bfp16ebs8 inside Peano matmul microkernel",
+      "results": [
+        {
+          "variant": "projection-matmul",
+          "status": "compiled",
+          "returnCode": 0,
+          "artifact": {
+            "path": "projection-matmul.mlir",
+            "bytes": 1248515,
+            "sha256": "e25ab9d86cb7e14fcf56e09f8a27bf915fe31e074526a2825650c18419c15a4d"
+          },
+          "log": {
+            "path": "projection-matmul.log",
+            "bytes": 0,
+            "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+          }
+        },
+        {
+          "variant": "projection-matmul-padded",
+          "status": "compiled",
+          "returnCode": 0,
+          "artifact": {
+            "path": "projection-matmul-padded.mlir",
+            "bytes": 1248480,
+            "sha256": "be7d04f2b328b244ac74aa475abbb5b59e0e8059395fcd99714b8ead5a62a256"
+          },
+          "log": {
+            "path": "projection-matmul-padded.log",
+            "bytes": 0,
+            "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+          }
+        },
+        {
+          "variant": "projection",
+          "status": "failed",
+          "returnCode": 1,
+          "log": {
+            "path": "projection.log",
+            "bytes": 5614,
+            "sha256": "96adc90f3fe65b0a682cfdb697640621073dced26f9b590f4d88289ddc6ff593"
+          }
+        },
+        {
+          "variant": "mlp-residual",
+          "status": "failed",
+          "returnCode": 1,
+          "log": {
+            "path": "mlp-residual.log",
+            "bytes": 8236,
+            "sha256": "00db9558a947f4b6c22fbc4b5e28e1998ac0b7bfd93f4365b336c09cbb11f4f2"
+          }
+        }
+      ]
+    },
+    "matrixBinary": {
+      "compileStage": "binary",
+      "status": "compiled",
+      "kernelMode": "peano-bf16",
+      "kernelPrecision": "bf16 converted to bfp16ebs8 inside Peano matmul microkernel",
+      "results": [
+        {
+          "variant": "projection-matmul-padded",
+          "status": "compiled",
+          "returnCode": 0,
+          "artifact": {
+            "path": "projection-matmul-padded.vmfb",
+            "bytes": 612008,
+            "sha256": "c1d76da61b859a4aade426fa562eb997f32c73b40242cc359a07cf7d1f3613de"
+          },
+          "log": {
+            "path": "projection-matmul-padded.log",
+            "bytes": 0,
+            "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+          }
+        }
+      ]
+    },
+    "graphSources": {
+      "compileStage": "sources",
+      "status": "compiled",
+      "kernelMode": "vector",
+      "kernelPrecision": "source IR types",
+      "results": [
+        {
+          "variant": "recognition",
+          "status": "compiled",
+          "returnCode": 0,
+          "artifact": {
+            "path": "recognition.mlir",
+            "bytes": 42484239,
+            "sha256": "2bb0403b4a48ad345e4837e37515b633a9b38d044d8aaa19f75d2d58c008b5dc"
+          },
+          "log": {
+            "path": "recognition.log",
+            "bytes": 0,
+            "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+          }
+        }
+      ]
+    },
+    "graphTargets": {
+      "compileStage": "targets",
+      "status": "failed",
+      "kernelMode": "vector",
+      "kernelPrecision": "source IR types",
+      "results": [
+        {
+          "variant": "recognition",
+          "status": "crashed",
+          "returnCode": -11,
+          "log": {
+            "path": "recognition.log",
+            "bytes": 97,
+            "sha256": "beeae0292954fe1578e7f14762b2437daa2c88904dfa257b5362f332a124faf2"
+          },
+          "signal": 11
+        }
+      ]
+    }
+  },
+  "remainingWork": [
+    "layout and precision conversion dispatch lowering or coarse CPU/NPU partition",
+    "reduce full graph PackOp/packTranspose crash",
+    "evaluate BFP16 precision changes",
+    "share weights across widths",
+    "native backend and optional package integration"
+  ]
+}