Covers LLM reinforcement learning training systems: the train-inference split between trainer and rollout engines, the training step loop, and optimizations like KV/prefix caching, continuous and speculative decoding, activation checkpointing, and weight sync.

Save to open this file