vllm-project/vllm#1348
PagedAttention V2 added sequence-level parallelism to speed up small batches.
6 files·+764/-139·3 lanes