Zum Inhalt springen
Kitbay
Skill

pytorch-fsdp

MITAktualisiert am 11. Nov. 2026262.606GitHub-Sterne

An optional Hermes Agent skill, credited to Orchestra Research, for writing and debugging PyTorch Fully Sharded Data Parallel code to train large models across devices.

Was der Skill vermittelt

An optional Hermes Agent skill, credited to Orchestra Research, for writing and debugging PyTorch Fully Sharded Data Parallel code to train large models across devices. Its tags cover sharding strategies, mixed precision, CPU offloading and FSDP2, and it lists torch 2.0 or later and transformers as dependencies. The SKILL.md itself is short: it says it was generated from the official documentation and sends the agent to a reference file of runnable snippets for wrapping, sharding, checkpointing and mixed precision instead of writing FSDP code from memory. Its scripts and assets folders are placeholders with nothing in them, and its reference list names a different file from the one the quick reference cites.

Funktionen

  • Look up FSDP wrapping and sharding patterns
  • Set up mixed precision and CPU offloading
  • Debug FSDP training code from reference snippets