Merge pull request #3570 from modelscope/codex/llamacpp-blackwell-windows-3296

feat(runtime): publish Windows CUDA asset for Blackwell

.github/workflows/build-llamacpp-binaries.yml

@@ -135,6 +135,34 @@ jobs:
               -DGGML_CUDA_FA=OFF
               -DGGML_CUDA_NCCL=OFF
               -DGGML_CUDA_GRAPHS=OFF
+              -DGGML_OPENMP=OFF
+              -DCMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded
+              -DGGML_AVX=OFF
+              -DGGML_AVX2=OFF
+              -DGGML_AVX_VNNI=OFF
+              -DGGML_AVX512=OFF
+              -DGGML_AVX512_VBMI=OFF
+              -DGGML_AVX512_VNNI=OFF
+              -DGGML_AVX512_BF16=OFF
+              -DGGML_FMA=OFF
+              -DGGML_F16C=OFF
+              -DGGML_BMI2=OFF
+          - os: windows-latest
+            name: windows-x64-cuda-blackwell
+            cuda: true
+            cuda_version: '13.2.0'
+            cuda_architectures: '120'
+            build_target: llama-funasr-sensevoice
+            timeout_minutes: 90
+            cmake_flags: >-
+              -DGGML_NATIVE=OFF
+              -DGGML_CUDA=ON
+              -DGGML_CUDA_FORCE_CUBLAS=ON
+              -DGGML_CUDA_FA=OFF
+              -DGGML_CUDA_NCCL=OFF
+              -DGGML_CUDA_GRAPHS=OFF
+              -DGGML_OPENMP=OFF
+              -DCMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded
               -DGGML_AVX=OFF
               -DGGML_AVX2=OFF
               -DGGML_AVX_VNNI=OFF
@@ -226,12 +254,47 @@ jobs:
           cp build/bin/llama-funasr-* pkg/ 2>/dev/null || true
           cp build/bin/Release/llama-funasr-* pkg/ 2>/dev/null || true
           cp README.md download-funasr-model.sh pkg/ 2>/dev/null || true
+          if [ "${{ matrix.cuda }}" = "true" ]; then
+            cuda_root="$(cygpath -u "$CUDA_PATH")"
+            for dll in cublas64_13.dll cublasLt64_13.dll; do
+              source_path="$(find "$cuda_root" -type f -iname "$dll" -print -quit)"
+              if [ -z "$source_path" ]; then
+                echo "Missing CUDA redistributable: $dll" >&2
+                exit 1
+              fi
+              cp "$source_path" pkg/
+            done
+            license_path="$(find "$cuda_root" -maxdepth 3 -type f \
+              \( -iname 'EULA.txt' -o -iname 'LICENSE' \) -print -quit)"
+            if [ -z "$license_path" ]; then
+              echo "Missing CUDA redistribution license" >&2
+              exit 1
+            fi
+            cp "$license_path" pkg/NVIDIA-CUDA-LICENSE.txt
+          fi
           echo "--- packaged binaries ---"; ls -la pkg
           if [ "${{ runner.os }}" = "Windows" ]; then
             7z a "funasr-llamacpp-${{ matrix.name }}.zip" ./pkg/*
           else
             tar czf "funasr-llamacpp-${{ matrix.name }}.tar.gz" -C pkg .
           fi
+      - name: Audit Windows CUDA package dependencies
+        if: matrix.cuda
+        run: |
+          exe="pkg/llama-funasr-sensevoice.exe"
+          test -f "$exe"
+          test -f pkg/cublas64_13.dll
+          test -f pkg/cublasLt64_13.dll
+          test -f pkg/NVIDIA-CUDA-LICENSE.txt
+          dependencies="$(objdump -p "$exe" | awk '/DLL Name:/ { print $3 }')"
+          printf '%s\n' "$dependencies"
+          grep -Fxi 'cublas64_13.dll' <<<"$dependencies"
+          for dll in MSVCP140.dll VCOMP140.DLL VCRUNTIME140.dll VCRUNTIME140_1.dll; do
+            if grep -Fxi "$dll" <<<"$dependencies"; then
+              echo "Unexpected dynamic runtime dependency: $dll" >&2
+              exit 1
+            fi
+          done
       - uses: actions/upload-artifact@v4
         with:
           name: funasr-llamacpp-${{ matrix.name }}
@@ -257,4 +320,4 @@ jobs:
           gh release create "$tag" dist/* \
             --repo "${{ github.repository }}" \
             --title "FunASR llama.cpp runtime $version" \
-            --notes "Prebuilt self-contained binaries for the FunASR llama.cpp / GGUF runtime: SenseVoice, Paraformer and Fun-ASR-Nano with built-in FSMN-VAD. Download the default quantized model with \`bash download-funasr-model.sh <sensevoice|paraformer|nano>\` (the helper requires the Hugging Face CLI: \`pip install -U huggingface_hub\`), then run \`llama-funasr-cli\` / \`llama-funasr-sensevoice\` / \`llama-funasr-paraformer\`. Use the default x64 asset for maximum CPU compatibility; use the x64-avx2 asset on CPUs with AVX2/FMA/F16C/BMI2 for higher throughput. The Vulkan assets are \`linux-x64-vulkan\` and \`windows-x64-vulkan\`; they require a working Vulkan driver/ICD and enable SenseVoiceSmall graph execution with \`llama-funasr-sensevoice ... --backend vulkan\`. Build from source with \`-DGGML_VULKAN=ON\` to validate platform-specific GPU stacks. The Windows CUDA asset is \`windows-x64-cuda\`; it requires an NVIDIA driver compatible with the CUDA Toolkit version configured by the release workflow, targets CUDA architecture 86, and enables SenseVoiceSmall graph execution with \`llama-funasr-sensevoice ... --backend cuda\`. Build from source for other GPU architectures. No Python ASR runtime or local build is required. Docs: $docs"
+            --notes "Prebuilt self-contained binaries for the FunASR llama.cpp / GGUF runtime: SenseVoice, Paraformer and Fun-ASR-Nano with built-in FSMN-VAD. Download the default quantized model with \`bash download-funasr-model.sh <sensevoice|paraformer|nano>\` (the helper requires the Hugging Face CLI: \`pip install -U huggingface_hub\`), then run \`llama-funasr-cli\` / \`llama-funasr-sensevoice\` / \`llama-funasr-paraformer\`. Use the default x64 asset for maximum CPU compatibility; use the x64-avx2 asset on CPUs with AVX2/FMA/F16C/BMI2 for higher throughput. The Vulkan assets are \`linux-x64-vulkan\` and \`windows-x64-vulkan\`; they require a working Vulkan driver/ICD and enable SenseVoiceSmall graph execution with \`llama-funasr-sensevoice ... --backend vulkan\`. Build from source with \`-DGGML_VULKAN=ON\` to validate platform-specific GPU stacks. The Windows CUDA assets are \`windows-x64-cuda\` for CUDA architecture 86 and \`windows-x64-cuda-blackwell\` for CUDA architecture 120 (RTX 50 / Blackwell); both require a compatible NVIDIA driver and enable SenseVoiceSmall graph execution with \`llama-funasr-sensevoice ... --backend cuda\`. Build from source for other GPU architectures. A successful CI build verifies code generation and archive integrity, not execution on physical Blackwell hardware. No Python ASR runtime or local build is required. Docs: $docs"

runtime/llama.cpp/README.md

@@ -93,11 +93,11 @@ FSMN-VAD without Python.
 
 ### Optional Windows CUDA backend for SenseVoiceSmall
 
-The CPU release ZIPs are portable packages. Tagged releases also publish
-`funasr-llamacpp-windows-x64-cuda.zip` for SenseVoiceSmall graph execution on
-NVIDIA GPUs that match CUDA architecture 86. Download the CUDA ZIP from
-[runtime-llamacpp-v0.1.9](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.9),
-then select the backend at runtime:
+The CPU release ZIPs are portable packages. Tagged releases publish separate
+SenseVoiceSmall CUDA packages: `funasr-llamacpp-windows-x64-cuda.zip` targets
+CUDA architecture 86, while `funasr-llamacpp-windows-x64-cuda-blackwell.zip`
+targets CUDA architecture 120 (`sm_120`) for RTX 50 / Blackwell GPUs. Select the
+matching archive, then enable the backend at runtime:
 
 ```bash
 # From the extracted windows-x64-cuda package:
@@ -105,7 +105,9 @@ then select the backend at runtime:
   -m sensevoice-small-q8.gguf --vad fsmn-vad.gguf -a sample.wav --backend cuda
 ```
 
-Build from source to target other GPU architectures:
+The Blackwell package uses the same command from its extracted directory. Build
+from source to target other GPU architectures or to reproduce the architecture
+120 build locally:
 
 ```bash
 cmake -B build-cuda -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON \
@@ -115,10 +117,11 @@ cmake --build build-cuda -j --target llama-funasr-sensevoice
   -m sensevoice-small-f16.gguf -a sample.wav --backend cuda
 ```
 
-Use the matching `CMAKE_CUDA_ARCHITECTURES` value for your GPU. RTX 50 /
-Blackwell cards report compute capability 12.0 (`sm_120`), so the current
-`windows-x64-cuda` prebuilt package for architecture 86 will not cover those
-cards.
+Use the matching `CMAKE_CUDA_ARCHITECTURES` value for your GPU. A successful
+release workflow build verifies architecture 120 code generation and ZIP
+integrity; it does not prove execution on physical Blackwell hardware. Keep
+hardware-specific reports open until the matching archive is retested on the
+reported GPU.
 
 `--backend cpu` remains the default and is what the portable cross-platform
 prebuilt binaries use. The CUDA package requires an NVIDIA driver compatible

runtime/llama.cpp/tests/test_release_workflow.py

@@ -16,6 +16,19 @@ def test_windows_cuda_release_asset_is_in_matrix():
     assert "timeout_minutes: 90" in workflow
 
 
+def test_windows_blackwell_cuda_release_asset_is_in_matrix():
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+    entry = workflow.split("name: windows-x64-cuda-blackwell", maxsplit=1)[1].split(
+        "          - os:", maxsplit=1
+    )[0]
+
+    assert "cuda: true" in entry
+    assert "cuda_version: '13.2.0'" in entry
+    assert "cuda_architectures: '120'" in entry
+    assert "build_target: llama-funasr-sensevoice" in entry
+    assert "timeout_minutes: 90" in entry
+
+
 def test_linux_vulkan_release_asset_is_in_matrix():
     workflow = WORKFLOW.read_text(encoding="utf-8")
 
@@ -54,6 +67,20 @@ def test_windows_cuda_build_uses_cuda_toolkit_and_flags():
     assert "--target \"${{ matrix.build_target }}\"" in workflow
 
 
+def test_windows_cuda_archives_bundle_runtime_dependencies():
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+
+    assert "-DCMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded" in workflow
+    assert "-DGGML_OPENMP=OFF" in workflow
+    assert "cublas64_13.dll" in workflow
+    assert "cublasLt64_13.dll" in workflow
+    assert "NVIDIA-CUDA-LICENSE.txt" in workflow
+    assert "Audit Windows CUDA package dependencies" in workflow
+    assert "objdump -p" in workflow
+    assert "MSVCP140.dll" in workflow
+    assert "VCOMP140.DLL" in workflow
+
+
 def test_linux_vulkan_build_uses_vulkan_sdk_and_flags():
     workflow = WORKFLOW.read_text(encoding="utf-8")
 
@@ -88,10 +115,13 @@ def test_release_notes_explain_cpu_and_cuda_windows_assets():
     assert "--backend cuda" in readme
     assert "Windows CUDA" in readme
     assert "CUDA architecture 86" in readme
+    assert "windows-x64-cuda-blackwell" in readme
+    assert "CUDA architecture 120" in readme
     assert "Build from source" in readme
     assert "other GPU architectures" in readme
     assert "CMAKE_CUDA_ARCHITECTURES=120" in readme
     assert "sm_120" in readme
+    assert "does not prove execution on physical Blackwell hardware" in readme
 
 
 def test_release_notes_explain_vulkan_linux_asset():
@@ -134,6 +164,16 @@ def test_github_release_notes_mention_vulkan_asset():
     assert "--backend vulkan" in release_notes
     assert "GGML_VULKAN=ON" in release_notes
 
+
+def test_github_release_notes_distinguish_cuda_architecture_assets():
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+    release_notes = workflow.split('--notes "', maxsplit=1)[1]
+
+    assert "windows-x64-cuda-blackwell" in release_notes
+    assert "CUDA architecture 86" in release_notes
+    assert "CUDA architecture 120" in release_notes
+
+
 def test_readme_documents_lightweight_http_server():
     readme = (ROOT / "runtime" / "llama.cpp" / "README.md").read_text(encoding="utf-8")