A Systems Design Comparison of LLM RL Post-training Frameworks
A deep systems comparison of verl, SLIME, AReaL, and ms-swift across rollout architecture, async control, partial rollout, weight sync, and platform trade-offs.
Blog
/
Why Prefill and Decode Should Not Share a GPU
A systems-oriented introduction to Prefill-Decode disaggregation in LLM inference, from KV cache physics and roofline intuition to vLLM and SGLang implementation trade-offs.