Add Modal GPU CI workflow

.github/workflows/gpu-benchmark-modal.yml

@@ -0,0 +1,115 @@
+name: GPU Benchmark (Modal)
+
+on:
+  workflow_dispatch:
+    inputs:
+      gpu:
+        description: 'GPU type (L4, A10G, A100, H100)'
+        default: 'L4'
+        type: string
+      epochs:
+        description: 'Number of training epochs'
+        default: '5'
+        type: string
+      channels:
+        description: 'Number of model channels (8=tiny, 32=prod, 64=large)'
+        default: '32'
+        type: string
+      residual_blocks:
+        description: 'Number of residual blocks (2=tiny, 16=prod)'
+        default: '16'
+        type: string
+      samples:
+        description: 'Number of samples (0=all eligible)'
+        default: '50'
+        type: string
+      dataset:
+        description: 'Dataset: s3 (205 samples, matches EC2) or dataset_4 (2885 samples)'
+        default: 's3'
+        type: string
+      wandb_project:
+        description: 'WandB project name'
+        default: 'elf-net-ci'
+        type: string
+
+jobs:
+  benchmark:
+    runs-on: ubuntu-latest
+    env:
+      MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
+      MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v5
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install Modal
+        run: uv pip install --system modal
+
+      - name: Run benchmark
+        id: bench
+        env:
+          GPU: ${{ inputs.gpu || 'L4' }}
+          EPOCHS: ${{ inputs.epochs || '5' }}
+          CHANNELS: ${{ inputs.channels || '32' }}
+          RESIDUAL_BLOCKS: ${{ inputs.residual_blocks || '16' }}
+          SAMPLES: ${{ inputs.samples || '50' }}
+          DATASET: ${{ inputs.dataset || 's3' }}
+          WANDB_PROJECT: ${{ inputs.wandb_project || 'elf-net-ci' }}
+        run: |
+          OUTPUT=$(modal run modal/benchmark.py \
+            --gpu "$GPU" \
+            --epochs "$EPOCHS" \
+            --channels "$CHANNELS" \
+            --residual-blocks "$RESIDUAL_BLOCKS" \
+            --samples "$SAMPLES" \
+            --dataset "$DATASET" \
+            --wandb-project "$WANDB_PROJECT" 2>&1 | tee /dev/stderr)
+
+          # Parse metrics from output
+          for key in VAL_LOSS TRAIN_LOSS WALLCLOCK GPU DATASET SAMPLES WANDB_URL; do
+            val=$(echo "$OUTPUT" | grep -oP "BENCHMARK_${key}=\K.*" || true)
+            if [ -n "$val" ]; then
+              echo "BENCHMARK_${key}=${val}" >> "$GITHUB_OUTPUT"
+            fi
+          done
+
+      - name: Summary
+        env:
+          EPOCHS: ${{ inputs.epochs || '5' }}
+          CHANNELS: ${{ inputs.channels || '32' }}
+          RESIDUAL_BLOCKS: ${{ inputs.residual_blocks || '16' }}
+          VAL_LOSS: ${{ steps.bench.outputs.BENCHMARK_VAL_LOSS }}
+          TRAIN_LOSS: ${{ steps.bench.outputs.BENCHMARK_TRAIN_LOSS }}
+          WALLCLOCK: ${{ steps.bench.outputs.BENCHMARK_WALLCLOCK }}
+          GPU: ${{ steps.bench.outputs.BENCHMARK_GPU }}
+          DATASET: ${{ steps.bench.outputs.BENCHMARK_DATASET }}
+          SAMPLES: ${{ steps.bench.outputs.BENCHMARK_SAMPLES }}
+          WANDB_URL: ${{ steps.bench.outputs.BENCHMARK_WANDB_URL }}
+          WANDB_PROJECT: ${{ inputs.wandb_project || 'elf-net-ci' }}
+        run: |
+          echo "## GPU Benchmark Results (Modal)" >> "$GITHUB_STEP_SUMMARY"
+          echo "" >> "$GITHUB_STEP_SUMMARY"
+          echo "### Configuration" >> "$GITHUB_STEP_SUMMARY"
+          echo "| Parameter | Value |" >> "$GITHUB_STEP_SUMMARY"
+          echo "|-----------|-------|" >> "$GITHUB_STEP_SUMMARY"
+          echo "| Platform | Modal ($GPU) |" >> "$GITHUB_STEP_SUMMARY"
+          echo "| Epochs | $EPOCHS |" >> "$GITHUB_STEP_SUMMARY"
+          echo "| Channels | $CHANNELS |" >> "$GITHUB_STEP_SUMMARY"
+          echo "| Residual Blocks | $RESIDUAL_BLOCKS |" >> "$GITHUB_STEP_SUMMARY"
+          echo "| Dataset | $DATASET ($SAMPLES samples) |" >> "$GITHUB_STEP_SUMMARY"
+          COMMIT_LINK="[\`${GITHUB_SHA::7}\`](${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/commit/${GITHUB_SHA})"
+          echo "| Commit | $COMMIT_LINK |" >> "$GITHUB_STEP_SUMMARY"
+          if [ -n "$WANDB_URL" ]; then
+            WANDB_RUN_ID=$(echo "$WANDB_URL" | grep -oP '[^/]+$')
+            echo "| WandB | [$WANDB_PROJECT](https://wandb.ai/PrinceOA/$WANDB_PROJECT) run [$WANDB_RUN_ID]($WANDB_URL) |" >> "$GITHUB_STEP_SUMMARY"
+          fi
+          echo "" >> "$GITHUB_STEP_SUMMARY"
+          echo "### Results" >> "$GITHUB_STEP_SUMMARY"
+          echo "| Metric | Value |" >> "$GITHUB_STEP_SUMMARY"
+          echo "|--------|-------|" >> "$GITHUB_STEP_SUMMARY"
+          echo "| val_loss | $VAL_LOSS |" >> "$GITHUB_STEP_SUMMARY"
+          echo "| train_loss | $TRAIN_LOSS |" >> "$GITHUB_STEP_SUMMARY"
+          echo "| Wallclock | ${WALLCLOCK}s |" >> "$GITHUB_STEP_SUMMARY"

.github/workflows/gpu-e2e-modal.yml

@@ -0,0 +1,32 @@
+name: GPU E2E (Modal)
+
+on:
+  pull_request:
+    branches: [main]
+  workflow_dispatch:
+    inputs:
+      epochs:
+        description: 'Number of training epochs'
+        default: '5'
+        type: string
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    env:
+      MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
+      MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v5
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install Modal
+        run: uv pip install --system modal
+
+      - name: Run e2e test on Modal GPU
+        env:
+          EPOCHS: ${{ inputs.epochs || '5' }}
+        run: modal run modal/ci.py --epochs "$EPOCHS"