The Agent Skills Directory

[SAFE]: The skill uses nvidia-modelopt, an official NVIDIA library, to perform post-training quantization. External dependencies (e.g., transformers, mamba-ssm) are installed via standard package managers (pip, uv) and target well-known public registries.
[SAFE]: Remote execution and SLURM integration follow standard enterprise practices, requiring explicit user-provided SSH keys or existing cluster configurations. The launcher utility uses nemo_run (an NVIDIA internal tool) for job orchestration.
[SAFE]: The skill implements a mandatory validation gate in references/checkpoint-validation.md. This requires the agent to verify model size, layer-wise quantization coverage, and metadata consistency before reporting success or proceeding to deployment. This check mitigates risk from misconfigured quantization recipes.
[SAFE]: Instructions for handling 'unlisted models' emphasize local inspection and manual patching of the quantization plugin rather than downloading opaque remote payloads. It correctly handles trust_remote_code by prompting for manual dependency installation and code inspection.
[SAFE]: Calibration datasets are fetched from standard Hugging Face repositories. The fallback to cnn_dailymail is a standard practice for constrained environments.

ptq