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.