Reading
5 things I recommend: essays and papers worth your time. Each one links to the original.
Agentic RL Infra 主流技术路线综述
A synthesis of eleven 2025–2026 public reports on agentic RL infrastructure, organised around the throughput / off-policy trade-off rather than model names. Colocated vs disaggregated serving, staleness control, train–inference mismatch correction, and credit assignment are compared side by side — the rare survey that still points at unresolved engineering problems.
GLM-5.2 RL weight transfer in 4 seconds using NIXL and ModelExpress
A concrete answer to the weight-sync bottleneck in large-scale RL: why NCCL’s static process groups block elastic inference, and how a NIXL pull path plus traced vLLM load_layouts gets GLM-5.2 from 86s to under 4s. Worth reading for the lower-bound bandwidth math and the pause-consensus detail alone.
从会说话到会试错: 一篇读懂 LLM Reinforcement Learning
A Chinese-language introduction to post-training that spends its words on the loop rather than on a catalogue of algorithm names: sample a rollout, score it, update the policy. PPO and GRPO appear as answers to a question instead of as headings, which is the right order to meet them in. About 35 minutes, and worth reading before the papers it summarises.
Orca: A Distributed Serving System for Transformer-Based Generative Models
Where iteration-level scheduling — what everyone now calls continuous batching — was first described. Worth reading alongside a modern scheduler to see how much of the design survived and how much was reinvented under a new name.
vLLM: Easy, Fast, and Cheap LLM Serving with PagedAttention
The post that introduced paged KV cache to most people. Read it for the memory accounting in the middle section — it explains why continuous batching alone does not get you far once sequences get long, which is the part most reimplementations skip.