Marin - CoreWeave GPU Canary Ferry #96
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: Marin - CoreWeave GPU Canary Ferry | |
| on: | |
| schedule: | |
| - cron: '0 10 * * *' # Daily at 10 AM UTC | |
| workflow_dispatch: | |
| inputs: | |
| target_tokens: | |
| description: 'Override training token budget' | |
| type: number | |
| required: false | |
| keep_nodepool: | |
| description: 'Keep CW node pool alive after the run (for faster re-runs)' | |
| type: boolean | |
| default: false | |
| permissions: | |
| contents: read # actions/checkout | |
| packages: write # docker login ghcr.io for iris cluster start | |
| issues: write # claude triage files issues | |
| id-token: write # claude-code-action OIDC | |
| jobs: | |
| canary-ferry-cw: | |
| runs-on: ubuntu-latest | |
| timeout-minutes: 180 | |
| concurrency: | |
| group: canary-ferry-cw-iris-canary | |
| cancel-in-progress: true | |
| env: | |
| RUN_ID: canary-gpu-${{ github.run_id }}-${{ github.run_attempt }} | |
| CANARY_ACCELERATOR: gpu | |
| CANARY_BATCH_SIZE: "32" | |
| CANARY_TARGET_TOKENS: "6553600" | |
| CANARY_MIN_STEPS: "40" | |
| CANARY_MAX_LOSS: "8.0" | |
| CANARY_MAX_WALL_CLOCK: "7200" | |
| WANDB_ENTITY: marin-community | |
| WANDB_PROJECT: marin | |
| IRIS_CONFIG: lib/iris/examples/coreweave-canary.yaml | |
| # Must match the label_prefix and namespace in IRIS_CONFIG so teardown | |
| # targets only this cluster's resources. | |
| IRIS_LABEL_PREFIX: iris-canary | |
| IRIS_NAMESPACE: iris-canary | |
| steps: | |
| - name: Checkout code | |
| uses: actions/checkout@v4 | |
| - name: Set up Python 3.12 | |
| uses: actions/setup-python@v5 | |
| with: | |
| python-version: "3.12" | |
| - name: Install uv | |
| uses: astral-sh/setup-uv@v7 | |
| with: | |
| enable-cache: true | |
| - name: Install dependencies | |
| run: uv sync --all-packages --extra=cpu --no-default-groups | |
| - name: Write CoreWeave kubeconfig | |
| run: | | |
| mkdir -p ~/.kube | |
| echo "${{ secrets.CW_KUBECONFIG }}" > ~/.kube/coreweave-iris | |
| chmod 600 ~/.kube/coreweave-iris | |
| - name: Set up Docker Buildx | |
| uses: docker/setup-buildx-action@v3 | |
| - name: Log in to GitHub Container Registry | |
| uses: docker/login-action@v3 | |
| with: | |
| registry: ghcr.io | |
| username: ${{ github.actor }} | |
| password: ${{ secrets.GITHUB_TOKEN }} | |
| - name: Start CoreWeave cluster | |
| run: .venv/bin/iris -v --config=${{ env.IRIS_CONFIG }} cluster start | |
| env: | |
| BUILDKIT_PROGRESS: plain | |
| R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }} | |
| R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }} | |
| - name: Submit canary ferry | |
| id: submit | |
| shell: bash -l {0} | |
| run: | | |
| JOB_ID=$(.venv/bin/iris --config=${{ env.IRIS_CONFIG }} \ | |
| job run --no-wait \ | |
| --memory=16G --disk=16G --cpu=1 --extra=cpu \ | |
| -e MARIN_PREFIX s3://marin-na/marin/ \ | |
| -e RUN_ID "$RUN_ID" \ | |
| -e CANARY_ACCELERATOR "$CANARY_ACCELERATOR" \ | |
| -e CANARY_BATCH_SIZE "$CANARY_BATCH_SIZE" \ | |
| -e CANARY_TARGET_TOKENS "$CANARY_TARGET_TOKENS" \ | |
| -e WANDB_ENTITY "$WANDB_ENTITY" \ | |
| -e WANDB_PROJECT "$WANDB_PROJECT" \ | |
| -e WANDB_API_KEY "$WANDB_API_KEY" \ | |
| -e HF_TOKEN "$HF_TOKEN" \ | |
| -e AWS_ACCESS_KEY_ID "$R2_ACCESS_KEY_ID" \ | |
| -e AWS_SECRET_ACCESS_KEY "$R2_SECRET_ACCESS_KEY" \ | |
| -e AWS_ENDPOINT_URL "https://74981a43be0de7712369306c7b19133d.r2.cloudflarestorage.com" \ | |
| -- python -m experiments.ferries.canary_ferry) | |
| echo "job_id=$JOB_ID" >> "$GITHUB_OUTPUT" | |
| echo "Submitted job: $JOB_ID" | |
| env: | |
| WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }} | |
| HF_TOKEN: ${{ secrets.HF_TOKEN }} | |
| R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }} | |
| R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }} | |
| CANARY_TARGET_TOKENS: ${{ inputs.target_tokens || env.CANARY_TARGET_TOKENS }} | |
| - name: Wait for canary ferry | |
| shell: bash -l {0} | |
| run: | | |
| JOB_ID="${{ steps.submit.outputs.job_id }}" | |
| echo "Polling job status: $JOB_ID" | |
| while true; do | |
| STATE=$(.venv/bin/iris --config=${{ env.IRIS_CONFIG }} \ | |
| job list --json --prefix "$JOB_ID" \ | |
| | jq -r --arg id "$JOB_ID" '[.[] | select(.job_id == $id)][0].state // empty') | |
| case "$STATE" in | |
| JOB_STATE_SUCCEEDED) | |
| echo "Job succeeded" | |
| exit 0 | |
| ;; | |
| JOB_STATE_PENDING|JOB_STATE_BUILDING|JOB_STATE_RUNNING) | |
| echo "$(date -u +%H:%M:%S) Job state: $STATE" | |
| sleep 30 | |
| ;; | |
| "") | |
| echo "Job not found: $JOB_ID" | |
| exit 1 | |
| ;; | |
| *) | |
| echo "Job finished with state: $STATE" | |
| .venv/bin/iris --config=${{ env.IRIS_CONFIG }} \ | |
| job list --json --prefix "$JOB_ID" \ | |
| | jq --arg id "$JOB_ID" '.[] | {job_id, state, error}' || true | |
| exit 1 | |
| ;; | |
| esac | |
| done | |
| - name: Validate canary metrics | |
| continue-on-error: ${{ inputs.target_tokens != '' }} | |
| shell: bash -l {0} | |
| run: .venv/bin/python scripts/canary/validate_canary_metrics.py | |
| env: | |
| MARIN_PREFIX: s3://marin-na/marin/ | |
| AWS_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }} | |
| AWS_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }} | |
| AWS_ENDPOINT_URL: https://74981a43be0de7712369306c7b19133d.r2.cloudflarestorage.com | |
| - name: Summarize GPU canary profile | |
| continue-on-error: true | |
| shell: bash -l {0} | |
| run: | | |
| .venv/bin/python -m marin.profiling.cli summarize \ | |
| --run-target "$RUN_ID" \ | |
| --entity "$WANDB_ENTITY" \ | |
| --project "$WANDB_PROJECT" | |
| env: | |
| WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }} | |
| - name: Capture failure diagnostics | |
| if: failure() | |
| run: | | |
| echo "=== Controller logs ===" | |
| kubectl --kubeconfig ~/.kube/coreweave-iris -n ${{ env.IRIS_NAMESPACE }} \ | |
| logs -l app=iris-controller --tail=500 || true | |
| echo "=== Controller pod describe ===" | |
| kubectl --kubeconfig ~/.kube/coreweave-iris -n ${{ env.IRIS_NAMESPACE }} \ | |
| describe pod -l app=iris-controller || true | |
| echo "=== Task pod logs ===" | |
| kubectl --kubeconfig ~/.kube/coreweave-iris -n ${{ env.IRIS_NAMESPACE }} \ | |
| logs -l iris.managed=true --tail=200 || true | |
| echo "=== Task pod describe ===" | |
| kubectl --kubeconfig ~/.kube/coreweave-iris -n ${{ env.IRIS_NAMESPACE }} \ | |
| describe pod -l iris.managed=true || true | |
| echo "=== Warning events ===" | |
| kubectl --kubeconfig ~/.kube/coreweave-iris -n ${{ env.IRIS_NAMESPACE }} \ | |
| get events --sort-by='.lastTimestamp' --field-selector type!=Normal || true | |
| - name: Claude triage | |
| id: claude_triage | |
| if: failure() && github.event_name == 'schedule' | |
| uses: anthropics/claude-code-action@v1 | |
| timeout-minutes: 30 | |
| with: | |
| claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN || secrets.CLAUDE_MAX_OAUTH_TOKEN }} | |
| prompt: | | |
| Read .agents/skills/canary-triage/SKILL.md and follow it. | |
| claude_args: | | |
| --model opus | |
| --max-turns 50 | |
| --allowedTools "Bash(kubectl:*),Bash(gh:*),Bash(.venv/bin/iris:*),Bash(.venv/bin/python:*),Bash(cat:*),Bash(jq:*),Bash(head:*),Bash(tail:*),Bash(grep:*)" | |
| env: | |
| CANARY_LANE: gpu | |
| CANARY_JOB_ID: ${{ steps.submit.outputs.job_id }} | |
| CANARY_RUN_ID: ${{ env.RUN_ID }} | |
| IRIS_CONFIG: ${{ env.IRIS_CONFIG }} | |
| IRIS_NAMESPACE: ${{ env.IRIS_NAMESPACE }} | |
| WANDB_ENTITY: ${{ env.WANDB_ENTITY }} | |
| WANDB_PROJECT: ${{ env.WANDB_PROJECT }} | |
| WANDB_API_KEY: ${{ secrets.WANDB_API_KEY }} | |
| GHA_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }} | |
| # `cluster stop` only deletes Pods; NodePools survive and rely on the | |
| # CW autoscaler to scale down. Delete them explicitly to avoid lingering | |
| # H100 costs. | |
| - name: Tear down CoreWeave cluster | |
| if: always() | |
| run: | | |
| .venv/bin/iris -v --config=${{ env.IRIS_CONFIG }} cluster stop || true | |
| if [ "${{ inputs.keep_nodepool }}" != "true" ]; then | |
| kubectl --kubeconfig ~/.kube/coreweave-iris \ | |
| delete nodepool -l iris-${{ env.IRIS_LABEL_PREFIX }}-managed=true | |
| else | |
| echo "Keeping node pool alive (keep_nodepool=true)" | |
| fi | |
| - name: Notify Slack on failure | |
| if: failure() && github.event_name == 'schedule' | |
| env: | |
| SLACK_WEBHOOK_URL: ${{ secrets.SLACK_WEBHOOK_URL }} | |
| FALLBACK_TEXT: ":red_circle: *GPU Canary failed*\nRun: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}" | |
| run: | | |
| if [ -f slack_message.md ]; then | |
| TEXT=$(cat slack_message.md) | |
| else | |
| TEXT="$FALLBACK_TEXT" | |
| fi | |
| PAYLOAD=$(python3 -c "import sys,json; print(json.dumps({'text': sys.stdin.read()}))" <<< "$TEXT") | |
| curl -sf -X POST -H 'Content-Type: application/json' -d "$PAYLOAD" "$SLACK_WEBHOOK_URL" |