Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fix(runtime): bundle Windows CUDA dependencies

LauraGPTc0b0174738b7
diff --git a/.github/workflows/build-llamacpp-binaries.yml b/.github/workflows/build-llamacpp-binaries.yml
index 9001b9317..925f0d018 100644
--- a/.github/workflows/build-llamacpp-binaries.yml
+++ b/.github/workflows/build-llamacpp-binaries.yml
@@ -135,6 +135,8 @@ 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
@@ -159,6 +161,8 @@ 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
@@ -250,12 +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 }}
diff --git a/runtime/llama.cpp/tests/test_release_workflow.py b/runtime/llama.cpp/tests/test_release_workflow.py
index d6542c462..d191311af 100644
--- a/runtime/llama.cpp/tests/test_release_workflow.py
+++ b/runtime/llama.cpp/tests/test_release_workflow.py
@@ -67,6 +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")