SYSTEM Cited by 1 source
Torchtune¶
Torchtune is PyTorch's reference library of post-training recipes (SFT, DPO, LoRA, QLoRA, and related fine-tuning workflows). Netflix credits it as one of three OSS projects whose reference implementations and design patterns informed its internal Post-Training Framework. First canonical wiki reference: sources/2026-02-13-netflix-scaling-llm-post-training-at-netflix.
Role¶
- Canonical reference for PyTorch post-training recipes: SFT, DPO, LoRA.
- Informed Netflix's standardised recipe concept — the framework's user-facing unit (SFT/DPO/RL/KD) is config-file-selected, conceptually aligned with torchtune's recipe pattern.
Relationship to Netflix's framework¶
Not a direct dependency — Netflix owns its internal optimised model definitions and tokenizer-compat layer — but torchtune's recipe ergonomics and post-training utilities shaped the Netflix framework's Data/Model/Compute surface.