What does this skill do, and when should you use it?
This skill guides NeMo AutoModel launches on the current node, Slurm HPC clusters, and SkyPilot cloud environments. It covers torchrun, multi-node rendezvous, containers, mounts, environment variables, and Nsight Systems profiling. The launchers configure distributed runtime variables automatically for multi-node jobs. It is intended for launch YAML authoring and submission troubleshooting, not model onboarding or state-dict adapter implementation.
Generates or adjusts YAML for the requested launch mode; explains automodel and torchrun commands for interactive runs; configures Slurm job names, nodes, tasks per node, time limits, accounts, partitions, container images, mounts, cache paths, environment variables, and master_port; configures SkyPilot cloud, accelerators, node count, spot usage, disk size, region, and setup commands; explains WORLD_SIZE, MASTER_ADDR, MASTER_PORT, and NCCL environment variables; enables Slurm nsys profile and identifies the .nsys-rep output; provides checkpointing and restore guidance for spot preemption.
- A NeMo AutoModel developer needs a single-node development or debugging run using all local GPUs and wants an interactive torchrun launch.
- An HPC user needs to submit a multi-node training job and configure Slurm resources, containers, mounts, and distributed communication.
- A cloud user needs to submit a NeMo AutoModel job through SkyPilot on AWS, GCP, Azure, Lambda, or Kubernetes.
- A SkyPilot spot user wants to limit lost work after preemption by configuring short checkpoint intervals and restore_from.
- A Slurm user needs a short diagnostic profiling run that produces an Nsight Systems report.
- A job owner is troubleshooting port collisions, missing mount paths, environment-variable expansion, or insufficient wall time.
How do you install this skill?
- Manually confirm permissions for Slurm, containers, mounts, cloud resources, and environment variables before execution; do not place real credentials in configuration or logs.
- Add versions and prerequisites for NeMo AutoModel, torchrun, Slurm, SkyPilot, and Nsight Systems, together with failure diagnosis, recovery, and rollback guidance.
- Reachability from mainland-China networks for cloud regions, image registries, and Hugging Face dependencies is unverified; spot preemption and checkpoint restoration may still interrupt jobs or lose work.
- The benchmark covers only three positive tasks and supplies no raw execution evidence, so it cannot replace validation in the target environment.
- Shell / CLI
- Local filesystem
torchrunSlurmSkyPilotEnroot/PyxisNsight Systems
Install the specific skill through the skills CLI documented in the repository README:
npx skills add nvidia/skills --skill nemo-automodel-launcher-config --yes
To target Codex explicitly:
npx skills add nvidia/skills --skill nemo-automodel-launcher-config --agent codex --yes
The supplied material does not document a standalone package, version requirement, or exact manual-copy destination.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/nemo-automodel-launcher-config" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Generate a Slurm configuration for NeMo AutoModel with 2 nodes and 8 tasks per node
- Configure a SkyPilot spot run with checkpoint recovery.
After installation, ask for a concrete launch configuration, such as: “Generate a Slurm configuration for NeMo AutoModel with 2 nodes and 8 tasks per node” or “Configure a SkyPilot spot run with checkpoint recovery.” Interactive mode requires no extra YAML section; when neither slurm: nor skypilot: is present, the CLI routes to torchrun automatically. Include the relevant platform and profiling requirements in the request.
What are this skill's strengths and limitations?
- Covers interactive, Slurm, and SkyPilot launch paths in one skill.
- Identifies the important fields for multi-node communication, containers, mounts, and environment variables.
- Includes practical warnings for port collisions, spot preemption, asynchronous checkpoints, and profiling overhead.
- The skill metadata names NVIDIA as author and specifies Apache-2.0 licensing.
- Its scope stops at launch mechanics; it does not implement model architectures, Hugging Face state-dict adapters, or model files.
- Successful Slurm and SkyPilot execution depends on the user's existing cluster, cloud account, and runtime environment.
- The supplied material gives no independent test results, support matrix, or specific version compatibility information.
- Nsight profiling adds overhead and creates large report artifacts, so it is unsuitable for normal production training.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| NeMo AutoModel Launcher Configuration this page ✓ NVIDIA · Official | 54 · Use with care | ★ 3.8k | today | Apache-2.0 |
| Megatron-LM on SLURM ✓ NVIDIA · Official | 56 · Use with care | ★ 3.8k | today | Apache-2.0 |
| NeMo MBridge Multi-Node Slurm ✓ NVIDIA · Official | 44 · Blocked | ★ 3.8k | today | Apache-2.0 |
| TAO Jobs on Kubernetes ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO on SLURM ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
Interactive launch is for development and debugging on the current node; Slurm is for HPC batch scheduling and multi-node execution; SkyPilot is for cloud-agnostic and spot-based jobs. These are the three launch alternatives explicitly covered by the skill.
How did FollowSkills review this skill?
The skill provides configuration guidance and command examples without apparent malware, credential theft, or destructive defaults; it references HF_TOKEN, WANDB_API_KEY, cloud submission, container mounts, and environment variables, and warns about spot preemption and profiling overhead. Deductions apply because it does not require execution-time confirmation or sufficiently disclose credential handling, permission scope, external effects, or rollback procedures.
The main interactive, Slurm, SkyPilot, and profiling paths are documented coherently with fields and common pitfalls. Deductions apply because prerequisites, version compatibility, failure diagnosis, and recovery are thin; the multi-node environment-variable statement mentions nodes times ntasks_per_node while the SkyPilot example does not define ntasks_per_node, creating possible ambiguity. Static calibration caps this at 10.
The name, trigger conditions, three launch modes, and explicit routing boundary are clear, and launcher configuration is separated from model onboarding. Deductions apply because non-fit cases, input validation, and environment requirements are incomplete; Chinese-language support and mainland-China reachability are not established. Interactive and local Slurm use are not entirely dependent on overseas services.
The SKILL.md has clear metadata, templates, field descriptions, examples, pitfalls, and code anchors; the skill card adds Apache 2.0 licensing, ownership, version information, and references. Deductions apply because Purpose, formal Examples, prerequisites, limitations, troubleshooting, changelog, and a clear maintenance/update path are missing; BENCHMARK.md records the same static-validation gaps.
The skill produces directly reusable YAML, bash, and runtime-behavior guidance, and its three evaluation tasks align closely with the documented content; the benchmark reports strong correctness and effectiveness. Deductions apply because evaluation covers only three positive scenarios with no negative-trigger cases, the report is not independently reproducible here, and real-environment correctness and runnability still require human validation. Static calibration caps this at 7.
The fixed revision, evaluation tasks, ground truths, result summary, code anchors, and named references provide some auditability. Deductions apply because there are no raw execution logs, CI tests, or committed tests covering this skill's key paths; the benchmark is primarily a declarative summary and cannot be independently reproduced from the supplied files. Static calibration caps this at 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →