GPU Debugging

Interactive GPU containers on Modal with SSH access. Use for debugging, profiling, prototyping, or any work where you need hands-on access to a GPU environment. Containers come pre-installed with PyTorch, Transformers, SGLang, and common ML tools. Your workspace is persisted to a Modal Volume and synced every 30 seconds.

For general Modal platform usage (app structure, function types, CLI, deployment), see the modal-basic-skills skill.

Quick Reference

Launch a Sandbox

# Launch with specific GPU (runs in background)
_SANDBOX_GPU=H100 python -m modal run modal-gpu-dev/tools/gpu_sandbox.py \
  --key-path ~/.ssh/id_ed25519.pub --gpu H100 --sandbox-id my-sandbox > /tmp/sandbox.log 2>&1 &