vllm

Serving, batching, memory management, or performance work that specifically uses vLLM in training-adjacent or inference systems.