NVIDIAOfficial·17 Agent Skills Included

Model-Optimizer

Quantize, evaluate, and deploy LLMs faster on NVIDIA GPUs

Automates post-training quantization of LLMs to NVFP4, FP8, and INT4 formats using NVIDIA Model Optimizer. Handles Slurm job submission, accuracy evaluation, baseline comparison, and vLLM or SGLang deployment without manual scripting. Guides agents through recipe search, EAGLE3 speculative decoding, and day-0 release checks with enforced quality gates.

3,672|576

↗

npx skills add NVIDIA/Model-Optimizer --all -g -y

Available:

Tells the AI agent where the skills live, how to run tests and commits safely, and to never push code or open PRs without explicit user approval.

All Skills in This Repository (17)

Pure Emerald Level Indicators

📦 In Repo

NVIDIA

quant-recipe-search

Design and manage iterative optimization searches for quantization recipes.

monitor

Monitor submitted compute jobs and report their execution status changes.

eagle3-new-model

Configure EAGLE3 offline pipeline YAML files for new model checkpoints.

day0-release

Orchestrate quantized model release workflows with gated validation and publish decisions.

eagle3-triage

Diagnose failed EAGLE3 pipeline runs by identifying broken workflow stages.

compare-results

Compare baseline and candidate model evaluation results to determine quantization impact.

eagle3-validate

Validate EAGLE3 pipeline runs by checking completion, artifacts, and benchmark metrics.

eagle3-review-logs

Analyze EAGLE3 pipeline logs to classify failures and summarize task outcomes.

release-cherry-pick

Cherry-pick labeled merged pull requests into release branches via GitHub CLI.

monitor

Monitor PTQ, evaluation, and deployment jobs on SLURM clusters.

common

Provides shared environment setup, credentials, remote execution, and SLURM guidance for ModelOpt workflows.

deployment

Serve quantized or unquantized LLM checkpoints as OpenAI-compatible endpoints using vLLM, SGLang, or TRT-LLM.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install Model-Optimizer skills?▼

Run `npx skills add NVIDIA/Model-Optimizer --all -g -y` in your terminal to install all skills in this suite globally.

How to quantize a model to NVFP4 or FP8?▼

Ask your agent to quantize a model and the ptq skill runs ModelOpt's hf_ptq.py with the right format, calibration data, and validation checks for your GPU.

Can it deploy quantized models with vLLM or SGLang?▼

Yes. The deployment skill starts an OpenAI-compatible endpoint with vLLM, SGLang, or TensorRT-LLM and verifies health before reporting the URL.

Does it work on Slurm clusters?▼

Yes. The launcher and monitor skills submit quantization and evaluation jobs to Slurm, track their status, and report failures with root causes.

How do I know a quantized model is safe to publish?▼

The day0-release skill chains quantization, serving canary, baseline and candidate evaluations, and accuracy gates to return a clear publish decision.