TLDR Tech / Data
Scaling LLM Post-Training at Netflix
Netflix developed an internal Post-Training Framework to scale LLM post-training, abstracting away infrastructure complexities with optimized data pipelines (on-the-fly sequence packing + document masking for skewed distributions), advanced parallelism (FSDP, tensor parallelism), and orchestration workflows using Ray and Verl, while integrating PyTorch, vLLM, and Hugging Face for flexibility and custom architectures.
Read the full story