Shared launch intake for any TAO workflow or action. Use when the user wants to run TAO AutoML, train, evaluate, infer, export, generate TensorRT engines, or launch DEFT/workflow jobs on an execution platform.
TAO Launch Workflow is a shared, model-agnostic intake skill that runs before launching any NVIDIA TAO workflow or model action — AutoML, train, evaluate, infer, export, TensorRT engine generation, or DEFT/application jobs — on an execution platform. It solves the problem of ad-hoc, unsafe launches by enforcing a non-negotiable preflight gate: no runner scripts, launch scripts, shims, workspaces, state files, logs, or dependency installs are created until every preflight condition passes.
The preflight requires the execution platform to be selected from a packaged helper (rather than by scanning docs or folders), platform and model-specific credentials to be satisfied, the default container image to be resolved from packaged metadata and confirmed or explicitly overridden, a successful platform access check, dataset inputs mapped to concrete spec keys, required compute-shape fields known, required local tools present, and a final launch review — including, for AutoML, explicit recommendation count/budget, concurrency, algorithm, metric, direction, and searched ranges — shown and confirmed by the user. It also defines monitoring behavior: whether to keep polling job logs in-chat and how often to post status, with clear rules about when a final response ends monitoring, and it tracks the original request so it resumes toward the goal after clearing a blocker. A bundled benchmark reports it passed NVSkills-Eval with 100% on the security dimension.
Target users are developers and ML engineers launching TAO training and related jobs on platforms such as SLURM, Brev, Kubernetes, or local Docker who want a consistent, credential-safe, user-confirmed launch process.
Before launching any TAO workflow or model action — AutoML, train, evaluate, infer, export, TensorRT engine generation, or DEFT/application jobs — as a shared intake and preflight step on the chosen execution platform.
A nine-point preflight: platform selected from the packaged helper, platform and model credentials satisfied, the default image resolved and confirmed, a successful platform access check, datasets mapped to spec keys, compute-shape fields known, required local tools present, and a final user-confirmed launch review.
Supported execution platforms per the skill card include SLURM, Brev, Kubernetes, and local Docker. The choices are enumerated by a packaged list_tao_platforms helper rather than by scanning configuration folders.
You choose whether the agent keeps polling backend job logs in chat and how often to post status (default every 5 minutes). A final response ends chat-side monitoring; while a launched job is non-terminal the agent continues polling and sends in-progress updates.
The packaged TAO skill bank helper scripts must be available (via TAO_SKILL_BANK_PATH), along with credentials for the selected platform and model.
Quick Setup:
.claude/skills/Repository
nvidia/skills