shunk031-colab-uv-training

Note

After reading this SKILL.md, say: 🛰️ I read shunk031-colab-uv-training.

Colab uv training

Complete the requested GPU job, preserve its results outside the runtime, and release the task's VM. Continue through monitoring, result retrieval, and teardown within the authorized run scope; launching a cell is not completion.

Use the repository's existing training command, checkpoint format, and approved artifact destination. The bundled wrapper supports Hugging Face Hub storage. Reuse a repository-specific job command or storage integration inside the launched job when it already provides the needed persistence and teardown.

Start GPU work only with scripts/colab-gpu-run. Stage job data in approved durable storage before allocation and declare every downloaded input as --hub-input NAME=URI; the launcher checks the Bash job script and local control files, starts a detached watchdog before allocation, uploads only the script and optional restricted token, and starts the bundled wrapper. Run the launcher as a managed background process with logged output so a tool timeout cannot kill colab exec; monitor its log or durable Hub output. The watchdog applies to GPU sessions; CPU-only CLI flows remain direct because this guard targets idle GPU allocation.

Before allocating a runtime

Resolve the target commit, job command, GPU, run limit, and artifact destination from the request and repository. Ask only for missing choices that affect cost, external writes, or the requested result. Reuse existing authorization for the task's allocation, uploads, and teardown. A job request does not authorize purchasing compute units, stopping unrelated sessions, or deleting stored artifacts.

Check colab usage against the run budget. The launcher requires Colab CLI 0.7.4 and generates a unique session name; do not start or resume GPU work with bare colab new --gpu, colab run --gpu, or colab exec commands. If CLI basics or authentication are needed, consult the colab-operator skill when available, otherwise the relevant colab --help command.

Read for the current operation

• For a fresh VM or an environment failure, read environment.md. It covers Colab's Python overrides and symptom-specific dependency repairs.
• To launch and monitor a long job, read running-jobs.md. It covers the busy kernel cell, the Hub wrapper, credentials, and result verification.
• To resume, recover a disconnected client or failed upload, or clean up leaked runtimes, read recovery.md. Account-wide sweeping is a separate operation with a broader scope than one job.
• To check a version-sensitive claim or its evidence, use sources.md. Dated observations are not guarantees about the current runtime.

Constraints that apply throughout

• /content is ephemeral. Save checkpoints and logs off the VM during execution, and leave time for a final upload before the runtime or job limit. Runtime lifetime depends on plan, balance, and availability.
• A client timeout does not prove the cell stopped. Check task logs and runtime state before relaunching; do not submit duplicate work to a busy kernel.
• The watchdog treats a live launcher-owned colab exec client process as busy; Colab CLI IDLE/BUSY output is only used to verify the named endpoint because CLI 0.7.4 status synchronizes and rewrites a local snapshot. If the client exits, the idle timer starts even if status still says BUSY.
• Keep secrets out of colab exec --env, command text, and logs. Use uploaded, restricted files as described in the launch reference.
• Recover the only copy of results before releasing a VM retained after an upload failure. A wrapper exit code alone proves neither durable storage nor successful unassignment.

Completion

Verify the job outcome from its exit status and expected artifacts, retrieve the requested results, and confirm the task's endpoint is no longer assigned after self-release or colab stop -s <name>. Other sessions may legitimately remain. The launcher always enforces its wall-clock limit, including when a user asks to keep the GPU runtime available longer.

Report the tested commit, job outcome, durable artifact location, and runtime state. If recovery or release fails, report the remaining endpoint, available results, and next action. Do not declare the run complete while task-owned cleanup remains unresolved.