gpucomm-bot

GPU-aware automation and CI engine for the gpucomm organization.

Features

• GPU / CUDA validation in PRs
• CI checks for compute pipelines
• Benchmark triggers before release
• Issue and PR automation

Commands

• /gpu-check → run GPU validation
• /benchmark → run performance tests
• /pause → pause automation
• /analyze → manual validation

Philosophy

Minimal, GPU-focused, automation-first.

Deployment

Local Development

1. Copy .env.example to .env and fill in values
2. Install dependencies: npm install
3. Start the bot: npm start
4. For public webhook access, use ngrok:

   ngrok http 3000

   Use the ngrok HTTPS URL + /webhook as your GitHub App webhook URL.

Production (Render)

We tried Railway first but the trial expired. Render was used instead with free tier.

1. Create a Render account and connect your GitHub repo
2. Create a Web Service with:
   • Build Command: npm install
   • Start Command: npm start
   • Plan: Free
3. Add environment variables:

   PORT=3000
   HOST=0.0.0.0
   WEBHOOK_SECRET=<your-secret>
   GITHUB_APP_ID=<your-app-id>
   GITHUB_INSTALLATION_ID=<your-installation-id>
   GITHUB_PRIVATE_KEY_B64=<base64-encoded-pem>
   GPUCOMM_GPU_LABEL=gpu-required
   

GitHub App Setup

1. Create: Go to https://github.com/settings/apps/new and create an app named gpucomm-bot
2. Install: Install the app on your account/org at https://github.com/apps/gpucomm-bot/installations/new
3. Webhook: In GitHub App settings → Webhook, enter https://<your-domain>/webhook
4. Permissions: Set to: Pull requests (read/write), Issues (read/write), Contents (read), Metadata (read)
5. Events: Subscribe to: Pull request, Issues, Release
6. Private key: Generate a private key, then convert to base64:

   base64 -i private-key.pem | tr -d '\n'

Bot in Action

PR with gpu-required label applied:

Issue with bot comment:

Real GPU CI

GPU CI (.github/workflows/gpu.yml) is configured for a self-hosted runner with NVIDIA drivers + CUDA toolkit.

• Runner labels: self-hosted, linux, x64, gpu
• Workflow gate: runs only if PR has label gpu-required or the PR title contains gpu/cuda
• CUDA smoke test: bash scripts/gpu-ci.sh (compiles scripts/cuda_smoke_test.cu with nvcc)
• Optional: set GPUCOMM_ENFORCE_CUDA_VERSION=true to enforce config/gpu.json.allowed_cuda_versions
• PyTorch smoke test: set repo/org variable PYTORCH_INDEX_URL (example: https://download.pytorch.org/whl/cu121) then run bash scripts/gpu-ci.sh pytorch

GitHub App auth

To let the bot comment/label on PRs, configure these environment variables for the running service:

• GITHUB_APP_ID: GitHub App ID
• GITHUB_INSTALLATION_ID: installation ID for the target org/repo
• GITHUB_PRIVATE_KEY (PEM string) or GITHUB_PRIVATE_KEY_B64 (base64 PEM)
• Optional: GITHUB_API_URL (defaults to https://api.github.com, set this for GHES)
• Optional: GPUCOMM_GPU_LABEL (label name to apply for GPU-related PRs; label should already exist)