TAO Platform Execution SDK
Submit, monitor, and manage NVIDIA TAO GPU training jobs across supported platforms.
The documentation limits credential handling to environment-variable presence checks, scopes credentials by platform, and explains S3/HF/NGC data flow, output locations, cancellation, and instance cleanup. However, it auto-installs an unpinned nvidia-tao-sdk[all] by default, creating network and supply-chain side effects, and lacks an explicit final confirmation immediately before launching a GPU job, so points are deducted.
Core APIs, platform differences, polling, error patterns, and failure feedback are documented reasonably well. However, dependency versions are not pinned, key helpers, skill_info.yaml, scripts, and SDK implementations are outside the selected skill evidence, and no key-path test suite is included; the score is therefore conservative under static-review limits.
The audience, trigger phrases, platform selection, monitoring preferences, scope boundaries, and local-Docker multi-node limitation are fairly clear. Chinese-language guidance and mainland-China reachability for package indexes, registries, S3, HF, and NGC are not addressed, limiting environment-fit evidence.
The documentation is reasonably layered and includes preflight guidance, references, error patterns, scope boundaries, examples, Apache-2.0 licensing, version 0.1.0, and NVIDIA ownership. It lacks the recommended Instructions and Examples sections, a changelog or explicit update path, and exposes some dependency assumptions, so points are deducted.
The skill could directly support complex TAO workflows involving submission, monitoring, logs, S3 I/O, multi-node execution, and multiple platforms. Static files do not verify real job outputs; the benchmark contains only one plan-description task, and platform prerequisites plus automatic installation can be costly, so the score remains below the static ceiling.
Committed evals.json, BENCHMARK.md, error mappings, and concrete code examples provide some auditability. However, there is only one evaluation task, no negative task or committed key-path tests, and the benchmark report is not independently reproducible from the supplied files, so coverage is thin.
- Before execution, manually confirm the platform, image, credential scope, data URIs, output persistence, GPU/cloud cost, and whether automatic dependency installation is permitted.
- Pin and review nvidia-tao-sdk[all] and its platform dependencies, and verify reachability of required package indexes, container registries, S3, HF Hub, or NGC.
- Do not pass agent-host local paths directly to SLURM or Kubernetes; after launch, verify that outputs are persisted rather than relying on container-ephemeral storage.
- The benchmark covers only one positive description task and does not establish real TAO-job correctness, reliability, or security.
What does this skill do, and when should you use it?
This skill guides use of the NVIDIA TAO Execution SDK on Brev, SLURM, local Docker, and Kubernetes. It enables job handles for status polling, log streaming, cancellation, failure analysis, and result retrieval. The SDK also wraps S3 input downloads and output uploads and supports multi-node and platform-specific execution features. It is best suited to TAO workflows that need durable tracking or platform integration beyond a simple docker run.
Checks for and installs nvidia-tao-sdk[all]; verifies the presence of environment variables without reading credential values; obtains the credential requirements for the selected platform; builds container commands with build_entrypoint; creates, monitors, cancels, and inspects jobs through the SDK; downloads declared inputs from S3, HF Hub, or NGC and uploads declared outputs; resolves result destinations using mounted storage or S3; and supports Brev instance listing and deletion.
- A TAO user needs to launch training on a Brev GPU instance and monitor it over time.
- A team needs to submit multi-node distributed training through SLURM and Lustre queues.
- An engineer wants S3 inputs and outputs handled automatically around a TAO container run.
- A developer needs to debug TAO jobs through Kubernetes Jobs or local Docker.
- A user wants to chain multiple TAO jobs while preserving job IDs and run state across context breaks.
What are this skill's strengths and limitations?
- One execution pattern covers Brev, SLURM, local Docker, and Kubernetes.
- Provides job handles, status polling, logs, cancellation, failure analysis, and result lookup.
- Includes S3-aware input and output handling.
- Persists job IDs and state paths so background jobs can be reattached later.
- Scopes credential requests by platform and does not read credential values.
- Requires Python 3.10+, nvidia-tao-sdk[all], and a working pip configuration.
- Platform, S3, and container execution may require additional environment variables or login credentials.
- Without a persistent mount or S3 bucket, outputs remain in an ephemeral container directory.
- Long-running monitoring requires the chat or runtime to remain available, otherwise status must be refreshed manually.
- The supplied source provides no skill-specific test suite or independent evidence for every platform.
How do you install this skill?
Install the skill with the NVIDIA skills CLI:
npx skills add nvidia/skills --skill tao-run-platform --yes
It becomes available when the agent next loads skills and encounters a relevant task. Before using it, install the SDK dependency:
python -m pip install "nvidia-tao-sdk[all]"
The README does not state a separate price for this skill.
How do you use this skill?
Ask the agent to “use the TAO SDK” or perform a “TAO platform run,” and describe the TAO workflow. Before launch, select Brev, SLURM, local Docker, or Kubernetes; choose whether long-running monitoring should remain enabled and set the update interval, which defaults to five minutes. Confirm the resolved container image or provide image=<override>, then supply only the environment variables required by the selected platform and data URIs. The agent creates the job and continues polling until completion, failure, cancellation, or an explicit request to detach.
How does this skill compare with similar options?
Compared with running TAO directly through docker run, this skill is aimed at workflows needing job handles, persistent monitoring, S3 I/O wrapping, scheduler integration, or multi-job orchestration. Most simple TAO runs do not need the SDK.