Skip to content

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

ci: publish AVX2 llama.cpp runtime assets

zhifu gao1c22668808b6
diff --git a/.github/workflows/build-llamacpp-binaries.yml b/.github/workflows/build-llamacpp-binaries.yml
index fc28880ab..56d6b4d7c 100644
--- a/.github/workflows/build-llamacpp-binaries.yml
+++ b/.github/workflows/build-llamacpp-binaries.yml
@@ -20,10 +20,68 @@ jobs:
fail-fast: false
matrix:
include:
- - { os: ubuntu-latest, name: linux-x64 }
- - { os: ubuntu-24.04-arm, name: linux-arm64 }
- - { os: macos-latest, name: macos-arm64 }
- - { os: windows-latest, name: windows-x64 }
+ - os: ubuntu-latest
+ name: linux-x64
+ cmake_flags: >-
+ -DGGML_NATIVE=OFF
+ -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: ubuntu-latest
+ name: linux-x64-avx2
+ cmake_flags: >-
+ -DGGML_NATIVE=OFF
+ -DGGML_AVX=ON
+ -DGGML_AVX2=ON
+ -DGGML_AVX_VNNI=OFF
+ -DGGML_AVX512=OFF
+ -DGGML_AVX512_VBMI=OFF
+ -DGGML_AVX512_VNNI=OFF
+ -DGGML_AVX512_BF16=OFF
+ -DGGML_FMA=ON
+ -DGGML_F16C=ON
+ -DGGML_BMI2=ON
+ - os: ubuntu-24.04-arm
+ name: linux-arm64
+ cmake_flags: -DGGML_NATIVE=OFF
+ - os: macos-latest
+ name: macos-arm64
+ cmake_flags: -DGGML_NATIVE=OFF
+ - os: windows-latest
+ name: windows-x64
+ cmake_flags: >-
+ -DGGML_NATIVE=OFF
+ -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-avx2
+ cmake_flags: >-
+ -DGGML_NATIVE=OFF
+ -DGGML_AVX=ON
+ -DGGML_AVX2=ON
+ -DGGML_AVX_VNNI=OFF
+ -DGGML_AVX512=OFF
+ -DGGML_AVX512_VBMI=OFF
+ -DGGML_AVX512_VNNI=OFF
+ -DGGML_AVX512_BF16=OFF
+ -DGGML_FMA=ON
+ -DGGML_F16C=ON
+ -DGGML_BMI2=ON
runs-on: ${{ matrix.os }}
defaults:
run:
@@ -36,20 +94,9 @@ jobs:
# Release binaries must be portable across user machines, not tuned to
# the ephemeral CI runner CPU. A native ggml build may emit AVX512 or
# AVX-VNNI instructions and then crash with SIGILL on ordinary CPUs.
- # Keep the prebuilt x64 package conservative; optimized AVX2/AVX512
- # builds can be published as separate assets later.
- cmake -B build -DCMAKE_BUILD_TYPE=Release \
- -DGGML_NATIVE=OFF \
- -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
+ # Keep the default x64 package conservative, and publish explicit
+ # x64-avx2 assets for machines that support AVX2/FMA/F16C/BMI2.
+ cmake -B build -DCMAKE_BUILD_TYPE=Release ${{ matrix.cmake_flags }}
cmake --build build --config Release -j 2
- name: Package
run: |
@@ -85,4 +132,4 @@ jobs:
gh release create "${{ github.ref_name }}" dist/* \
--repo "${{ github.repository }}" \
--title "FunASR llama.cpp runtime ${{ github.ref_name }}" \
- --notes "Prebuilt self-contained binaries for the FunASR llama.cpp / GGUF runtime — SenseVoice, Paraformer and Fun-ASR-Nano with built-in FSMN-VAD (a whisper.cpp-style on-device ASR, strong on Chinese). Get a model with \`bash download-funasr-model.sh <sensevoice|paraformer|nano>\`, then run \`llama-funasr-cli\` / \`llama-funasr-sensevoice\` / \`llama-funasr-paraformer\`. No Python, no build. Docs: runtime/llama.cpp/README.md"
+ --notes "Prebuilt self-contained binaries for the FunASR llama.cpp / GGUF runtime — SenseVoice, Paraformer and Fun-ASR-Nano with built-in FSMN-VAD (a whisper.cpp-style on-device ASR, strong on Chinese). Get a model with \`bash download-funasr-model.sh <sensevoice|paraformer|nano>\`, 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. No Python, no build. Docs: runtime/llama.cpp/README.md"