Address code review feedback: move imports to top, improve comments #2
Workflow file for this run
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| name: "Example: Batched Matrix Workflow" | ||
|
Check failure on line 1 in .github/workflows/example-batched-matrix.yml
|
||
| # This is an example workflow demonstrating how to use the batching feature | ||
| # to work around GitHub Actions' 256 job matrix limit | ||
| on: | ||
| workflow_dispatch: | ||
| inputs: | ||
| model-prefix: | ||
| description: "Model prefix to benchmark" | ||
| required: true | ||
| type: string | ||
| seq-lens: | ||
| description: "Sequence length config (e.g., 1k1k)" | ||
| required: true | ||
| type: string | ||
| jobs: | ||
| # Step 1: Determine how many batches are needed | ||
| get-batch-count: | ||
| runs-on: ubuntu-latest | ||
| outputs: | ||
| batch-count: ${{ steps.count.outputs.batch-count }} | ||
| steps: | ||
| - name: Checkout code | ||
| uses: actions/checkout@v4 | ||
| - id: count | ||
| run: | | ||
| pip install pydantic | ||
| BATCH_COUNT=$(python3 ${GITHUB_WORKSPACE}/utils/matrix-logic/generate_sweep_configs.py \ | ||
| full-sweep \ | ||
| --config-files ${GITHUB_WORKSPACE}/.github/configs/nvidia-master.yaml \ | ||
| ${GITHUB_WORKSPACE}/.github/configs/amd-master.yaml \ | ||
| --seq-lens ${{ inputs.seq-lens }} \ | ||
| --model-prefix ${{ inputs.model-prefix }} \ | ||
| --get-batch-count) | ||
| echo "batch-count=$BATCH_COUNT" >> $GITHUB_OUTPUT | ||
| echo "Total batches needed: $BATCH_COUNT" | ||
| # Step 2: Generate config for each batch | ||
| # This job runs once per batch (up to the batch-count) | ||
| get-batch-configs: | ||
| needs: get-batch-count | ||
| runs-on: ubuntu-latest | ||
| # Create a matrix with one entry per batch | ||
| strategy: | ||
| matrix: | ||
| # Generate array [0, 1, 2, ..., batch-count-1] | ||
| # This expression creates a comma-separated list of indices and wraps it in an array | ||
| # Example: if batch-count=3, this generates: [0, 1, 2] | ||
| # Note: This is a workaround for GitHub Actions' lack of native range array generation | ||
| batch-index: ${{ fromJson(format('[{0}]', join(range(0, fromJson(needs.get-batch-count.outputs.batch-count)), ','))) }} | ||
| outputs: | ||
| # Each batch gets its own output | ||
| configs-${{ matrix.batch-index }}: ${{ steps.get-configs.outputs.configs }} | ||
| steps: | ||
| - name: Checkout code | ||
| uses: actions/checkout@v4 | ||
| - id: get-configs | ||
| run: | | ||
| pip install pydantic | ||
| CONFIG_JSON=$(python3 ${GITHUB_WORKSPACE}/utils/matrix-logic/generate_sweep_configs.py \ | ||
| full-sweep \ | ||
| --config-files ${GITHUB_WORKSPACE}/.github/configs/nvidia-master.yaml \ | ||
| ${GITHUB_WORKSPACE}/.github/configs/amd-master.yaml \ | ||
| --seq-lens ${{ inputs.seq-lens }} \ | ||
| --model-prefix ${{ inputs.model-prefix }} \ | ||
| --batch-index ${{ matrix.batch-index }}) | ||
| echo "configs=$CONFIG_JSON" >> $GITHUB_OUTPUT | ||
| echo "Generated batch ${{ matrix.batch-index }}" | ||
| # Step 3: Run benchmarks for batch 0 | ||
| # You would create similar jobs for batch-1, batch-2, etc. if needed | ||
| benchmark-batch-0: | ||
| needs: get-batch-configs | ||
| # Only run if batch 0 exists | ||
| if: ${{ fromJson(needs.get-batch-count.outputs.batch-count) > 0 }} | ||
| uses: ./.github/workflows/benchmark-tmpl.yml | ||
| name: ${{ inputs.model-prefix }} ${{ inputs.seq-lens }} batch-0 / | ||
| strategy: | ||
| fail-fast: false | ||
| matrix: | ||
| config: ${{ fromJson(needs.get-batch-configs.outputs.configs-0) }} | ||
| secrets: inherit | ||
| with: | ||
| exp-name: "${{ inputs.model-prefix }}_${{ inputs.seq-lens }}_batch0" | ||
| isl: 1024 | ||
| osl: 1024 | ||
| max-model-len: 2048 | ||
| runner: ${{ matrix.config.runner }} | ||
| image: ${{ matrix.config.image }} | ||
| model: ${{ matrix.config.model }} | ||
| framework: ${{ matrix.config.framework }} | ||
| precision: ${{ matrix.config.precision }} | ||
| tp: ${{ matrix.config.tp }} | ||
| ep: ${{ matrix.config.ep }} | ||
| dp-attn: ${{ matrix.config.dp-attn }} | ||
| conc: ${{ matrix.config.conc }} | ||
| # Step 4 (optional): Collect results from all batches | ||
| collect-results: | ||
| needs: [get-batch-count, benchmark-batch-0] | ||
| if: ${{ always() }} | ||
| runs-on: ubuntu-latest | ||
| steps: | ||
| - name: Summary | ||
| run: | | ||
| echo "Processed ${{ needs.get-batch-count.outputs.batch-count }} batch(es)" | ||
| echo "Benchmark complete" | ||
| # Note: For production use with multiple batches, you would either: | ||
| # 1. Create multiple benchmark-batch-N jobs (one per possible batch) | ||
| # 2. Use a dynamic workflow generation approach | ||
| # 3. Use GitHub's reusable workflows with a loop construct (when available) | ||
| # | ||
| # The current InferenceMAX workflows split by model-prefix instead, | ||
| # which naturally keeps each job under the 256 limit. | ||
