|
| 1 | +name: Test Metal Backend |
| 2 | + |
| 3 | +on: |
| 4 | + pull_request: |
| 5 | + push: |
| 6 | + branches: |
| 7 | + - main |
| 8 | + - release/* |
| 9 | + |
| 10 | +concurrency: |
| 11 | + group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }} |
| 12 | + cancel-in-progress: false |
| 13 | + |
| 14 | +jobs: |
| 15 | + test-metal-builds: |
| 16 | + name: test-executorch-metal-build |
| 17 | + uses: pytorch/test-infra/.github/workflows/macos_job.yml@main |
| 18 | + with: |
| 19 | + runner: macos-m2-stable |
| 20 | + python-version: '3.11' |
| 21 | + submodules: 'recursive' |
| 22 | + ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }} |
| 23 | + timeout: 90 |
| 24 | + script: | |
| 25 | + set -eux |
| 26 | +
|
| 27 | + echo "::group::Test ExecuTorch Metal build" |
| 28 | + PYTHON_EXECUTABLE=python CMAKE_ARGS="-DEXECUTORCH_BUILD_METAL=ON" ${CONDA_RUN} --no-capture-output ./install_executorch.sh |
| 29 | + echo "::endgroup::" |
| 30 | +
|
| 31 | + export-voxtral-metal-artifact: |
| 32 | + name: export-voxtral-metal-artifact |
| 33 | + uses: pytorch/test-infra/.github/workflows/macos_job.yml@main |
| 34 | + secrets: inherit |
| 35 | + with: |
| 36 | + runner: macos-m2-stable |
| 37 | + python-version: '3.11' |
| 38 | + submodules: 'recursive' |
| 39 | + ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }} |
| 40 | + timeout: 90 |
| 41 | + secrets-env: EXECUTORCH_HF_TOKEN |
| 42 | + upload-artifact: voxtral-metal-export |
| 43 | + script: | |
| 44 | + set -eux |
| 45 | +
|
| 46 | + echo "::group::Setup Huggingface" |
| 47 | + ${CONDA_RUN} pip install -U "huggingface_hub[cli]" accelerate |
| 48 | + ${CONDA_RUN} huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN |
| 49 | + echo "::endgroup::" |
| 50 | +
|
| 51 | + echo "::group::Setup Optimum-ExecuTorch" |
| 52 | + OPTIMUM_ET_VERSION=$(cat .ci/docker/ci_commit_pins/optimum-executorch.txt) |
| 53 | + echo "Using optimum-executorch version: ${OPTIMUM_ET_VERSION}" |
| 54 | + ${CONDA_RUN} pip install git+https://github.com/huggingface/optimum-executorch.git@${OPTIMUM_ET_VERSION} |
| 55 | + ${CONDA_RUN} pip install mistral-common librosa |
| 56 | + echo "::endgroup::" |
| 57 | +
|
| 58 | + echo "::group::Setup ExecuTorch" |
| 59 | + PYTHON_EXECUTABLE=python ${CONDA_RUN} ./install_executorch.sh |
| 60 | + echo "::endgroup::" |
| 61 | +
|
| 62 | + echo "::group::Pip List" |
| 63 | + ${CONDA_RUN} pip list |
| 64 | + echo "::endgroup::" |
| 65 | +
|
| 66 | + echo "::group::Export Voxtral" |
| 67 | + ${CONDA_RUN} optimum-cli export executorch \ |
| 68 | + --model "mistralai/Voxtral-Mini-3B-2507" \ |
| 69 | + --task "multimodal-text-to-text" \ |
| 70 | + --recipe "metal" \ |
| 71 | + --dtype bfloat16 \ |
| 72 | + --max_seq_len 1024 \ |
| 73 | + --output_dir ./ |
| 74 | + ${CONDA_RUN} python -m executorch.extension.audio.mel_spectrogram \ |
| 75 | + --feature_size 128 \ |
| 76 | + --stack_output \ |
| 77 | + --max_audio_len 300 \ |
| 78 | + --output_file voxtral_preprocessor.pte |
| 79 | +
|
| 80 | + test -f model.pte |
| 81 | + test -f aoti_metal_blob.ptd |
| 82 | + test -f voxtral_preprocessor.pte |
| 83 | + echo "::endgroup::" |
| 84 | +
|
| 85 | + echo "::group::Store Voxtral Artifacts" |
| 86 | + mkdir -p "${RUNNER_ARTIFACT_DIR}" |
| 87 | + cp model.pte "${RUNNER_ARTIFACT_DIR}/" |
| 88 | + cp aoti_metal_blob.ptd "${RUNNER_ARTIFACT_DIR}/" |
| 89 | + cp voxtral_preprocessor.pte "${RUNNER_ARTIFACT_DIR}/" |
| 90 | + ls -al "${RUNNER_ARTIFACT_DIR}" |
| 91 | + echo "::endgroup::" |
| 92 | +
|
| 93 | + test-voxtral-metal-e2e: |
| 94 | + name: test-voxtral-metal-e2e |
| 95 | + needs: export-voxtral-metal-artifact |
| 96 | + uses: pytorch/test-infra/.github/workflows/macos_job.yml@main |
| 97 | + with: |
| 98 | + runner: macos-m2-stable |
| 99 | + python-version: '3.11' |
| 100 | + submodules: 'recursive' |
| 101 | + ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }} |
| 102 | + timeout: 90 |
| 103 | + download-artifact: voxtral-metal-export |
| 104 | + script: | |
| 105 | + set -eux |
| 106 | +
|
| 107 | + echo "::group::Print machine info" |
| 108 | + uname -a |
| 109 | + if [ $(uname -s) == Darwin ]; then |
| 110 | + sw_vers |
| 111 | + # Print RAM in GB |
| 112 | + RAM_BYTES=$(sysctl -n hw.memsize) |
| 113 | + RAM_GB=$(echo "scale=2; $RAM_BYTES/1024/1024/1024" | bc) |
| 114 | + echo "Available RAM (GB): $RAM_GB" |
| 115 | + sysctl machdep.cpu.brand_string |
| 116 | + sysctl machdep.cpu.core_count |
| 117 | + # Print number of GPU cores (Apple Silicon) |
| 118 | + if command -v system_profiler &> /dev/null; then |
| 119 | + GPU_CORES=$(system_profiler SPDisplaysDataType | awk '/Total Number of Cores/ {print $5; exit}') |
| 120 | + if [ -z "$GPU_CORES" ]; then |
| 121 | + # Fallback: try to parse "Core Count" from Apple GPU section |
| 122 | + GPU_CORES=$(system_profiler SPDisplaysDataType | awk '/Core Count/ {print $3; exit}') |
| 123 | + fi |
| 124 | + echo "GPU Cores: ${GPU_CORES:-Unknown}" |
| 125 | + else |
| 126 | + echo "system_profiler not available, cannot determine GPU cores." |
| 127 | + fi |
| 128 | + fi |
| 129 | + echo "::endgroup::" |
| 130 | +
|
| 131 | + echo "::group::Setup ExecuTorch Requirements" |
| 132 | + CMAKE_ARGS="-DEXECUTORCH_BUILD_METAL=ON" ${CONDA_RUN} --no-capture-output ./install_requirements.sh |
| 133 | + echo "::endgroup::" |
| 134 | +
|
| 135 | + echo "::group::Pip List" |
| 136 | + ${CONDA_RUN} pip list |
| 137 | + echo "::endgroup::" |
| 138 | +
|
| 139 | + echo "::group::Prepare Voxtral Artifacts" |
| 140 | + cp "${RUNNER_ARTIFACT_DIR}/model.pte" . |
| 141 | + cp "${RUNNER_ARTIFACT_DIR}/aoti_metal_blob.ptd" . |
| 142 | + cp "${RUNNER_ARTIFACT_DIR}/voxtral_preprocessor.pte" . |
| 143 | + TOKENIZER_URL="https://huggingface.co/mistralai/Voxtral-Mini-3B-2507/resolve/main/tekken.json" |
| 144 | + curl -L $TOKENIZER_URL -o tekken.json |
| 145 | + ls -al model.pte aoti_metal_blob.ptd voxtral_preprocessor.pte tekken.json |
| 146 | + echo "::endgroup::" |
| 147 | +
|
| 148 | + echo "::group::Create Test Audio File" |
| 149 | + say -o call_samantha_hall.aiff "Call Samantha Hall" |
| 150 | + afconvert -f WAVE -d LEI16 call_samantha_hall.aiff call_samantha_hall.wav |
| 151 | + echo "::endgroup::" |
| 152 | +
|
| 153 | + echo "::group::Build Voxtral Runner" |
| 154 | + ${CONDA_RUN} cmake --preset llm \ |
| 155 | + -DEXECUTORCH_BUILD_METAL=ON \ |
| 156 | + -DCMAKE_INSTALL_PREFIX=cmake-out \ |
| 157 | + -DCMAKE_BUILD_TYPE=Release \ |
| 158 | + -Bcmake-out -S. |
| 159 | + ${CONDA_RUN} cmake --build cmake-out -j$(( $(sysctl -n hw.ncpu) - 1 )) --target install --config Release |
| 160 | +
|
| 161 | + ${CONDA_RUN} cmake -DEXECUTORCH_BUILD_METAL=ON \ |
| 162 | + -DCMAKE_BUILD_TYPE=Release \ |
| 163 | + -Sexamples/models/voxtral \ |
| 164 | + -Bcmake-out/examples/models/voxtral/ |
| 165 | + ${CONDA_RUN} cmake --build cmake-out/examples/models/voxtral --target voxtral_runner --config Release |
| 166 | + echo "::endgroup::" |
| 167 | +
|
| 168 | + echo "::group::Run Voxtral Runner" |
| 169 | + set +e |
| 170 | + OUTPUT=$(cmake-out/examples/models/voxtral/voxtral_runner \ |
| 171 | + --model_path model.pte \ |
| 172 | + --data_path aoti_metal_blob.ptd \ |
| 173 | + --tokenizer_path tekken.json \ |
| 174 | + --audio_path call_samantha_hall.wav \ |
| 175 | + --processor_path voxtral_preprocessor.pte \ |
| 176 | + --temperature 0 2>&1) |
| 177 | + EXIT_CODE=$? |
| 178 | + set -e |
| 179 | +
|
| 180 | + echo "$OUTPUT" |
| 181 | +
|
| 182 | + if ! echo "$OUTPUT" | grep -iq "Samantha"; then |
| 183 | + echo "Expected output 'Samantha' not found in output" |
| 184 | + exit 1 |
| 185 | + fi |
| 186 | +
|
| 187 | + if [ $EXIT_CODE -ne 0 ]; then |
| 188 | + echo "Unexpected exit code: $EXIT_CODE" |
| 189 | + exit $EXIT_CODE |
| 190 | + fi |
| 191 | + echo "::endgroup::" |
0 commit comments