modal-projects

Use when adding, debugging, validating, or productionizing new model support in modal-training-gym, especially Slime recipes and model configs.

This skill should be used when the user asks to "check the progress of a run", "what step is run X on", "is my training run stuck", "what is run <id> doing", "grab the app logs for a run", "how far along is the rollout", or otherwise wants to know the live status, current step, current phase, or blocker of a modal-training-gym training run given a run id (or partial id / wandb id / Modal app id).

Runbook for launching, monitoring, and debugging training jobs on Modal: entrypoint discovery, detached app workflow, log filtering, container inspection, volume state, image build behavior, and debugging strategy. Use when running modal run, checking app status, reading logs, or troubleshooting a Modal training job.

One-stop reference for the modal-training-gym repo: package layout, core abstractions (ModelConfig, DatasetConfig, framework two-class split, build_app, cloudpickle caller resolution), the tutorial generator system, model catalog, tools/ directory, and common gotchas. Use when working with tutorials, models, frameworks, or any code in this repository.

Tiered validation pipeline for checking whether tutorials and examples still run as documented: Tier 0 (local compile), Tier 1 (cheap drift checks), Tier 2 (scheduled smoke on Modal), Tier 3 (full multi-node validation). Covers discovery, pass criteria, and failure recording. Use when validating, testing, or checking tutorials and examples for drift.

Foundational Modal platform knowledge for auto-research agents. Covers app structure, function types, CLI usage, deployment patterns, and development workflow. This skill auto-triggers when code imports `modal` and provides the base layer that the GPU-specific research skills build on.

Interactive GPU sandboxes on Modal with SSH access for debugging, profiling, and prototyping. Launch a container with any GPU (T4 to B200), SSH in, iterate on code with pre-installed ML tools. Use when you need hands-on GPU access for debugging, profiling with nsys/ncu, prototyping training scripts, or testing model serving.

Training apps, experiment patterns, and persistent storage on Modal. Write single-GPU and multi-node training jobs with volumes for data and checkpoints, secrets for W&B/HuggingFace, and fault tolerance via retries and checkpoint auto-resume. Use when you need to run training, fine-tuning, or batch experiments on Modal GPUs.

Parallel Claude Code agent orchestration. Spawn multiple autonomous agents, each with its own GPU or compute, to divide work in parallel. Each agent works independently and reports findings via structured reports; the parent monitors progress and steers agents. Use when you need to run multiple debugging sessions, experiments, or research tasks in parallel across separate GPUs.

This skill provides guidance for building and deploying applications on the Modal cloud platform. Use this skill whenever the user mentions Modal or has code that imports the `modal` SDK. This skill should also trigger when the user needs to run Python code with horizontal scalability (e.g. batch jobs), needs access to GPUs (e.g. AI/ML workloads including training and inference) or needs to execute processes in a sandboxed environment.