The Agent Skills Directory

[SAFE]: The skill provides technical instructions and code snippets for distributed training that follow industry best practices for performance engineering and model reproducibility.
[EXTERNAL_DOWNLOADS]: Mentions downloading common datasets (e.g., FineWeb-edu) and using reputable frameworks like PyTorch's torchtitan, Hugging Face's nanotron, and Karpathy's nanochat. All links provided in the sources point to official repositories or academic platforms like arXiv.
[COMMAND_EXECUTION]: Includes standard CLI examples for systems profiling (nsys, torch.profiler), monitoring (iostat, iotop), and cloud storage synchronization (rclone, aws-cli). these are contextually appropriate for the skill's stated purpose of managing large-scale compute runs.
[PRIVILEGE_ESCALATION]: Mentions administrative tuning commands like ulimit and blockdev for I/O optimization, which are standard in high-performance computing (HPC) environments and do not represent a security risk here.
[DYNAMIC_EXECUTION]: Discusses torch.compile and kernel fusion, which are standard performance features of the PyTorch ecosystem intended to improve training efficiency (MFU).

ai-distributed-training