Scaling LLM Post-Training at Netflix

This article describes Netflix's internal Post-Training Framework for adapting LLMs to production recommendation, personalization, and search use cases.

• •Loss masking ensures only assistant tokens contribute to training loss, preventing quality degradation from prompt tokens
• •Sequence packing with document masks reduces padding waste and keeps shapes consistent across FSDP workers
• •Model setup covers FSDP/TP sharding, LoRA integration, and memory optimizations for large vocabularies (>128k tokens)
• •Ray orchestrates distributed workflows supporting SFT, DPO, RL (GRPO), and Knowledge Distillation
• •Architecture is organized around four pillars: Data, Model, Compute, and Workflow, built on Netflix's Mako GPU platform with PyTorch, Ray, and vLLM