vLLM 称在 DeepSeek V4.1 Flash 解码上追平官方 DeepInfer
vLLM 团队称在 batch 8–256 的解码配置下与 DeepInfer 持平，靠引入 DeepSeek 开源 kernel 并调优 prefill/decode。
How much speed do you give up by not running a lab's own inference engine? For DeepSeek-V4.1-Flash, maybe none. The vLLM team says vLLM now matches DeepInfer, DeepSeek's in-house engine, on DeepSeek's decode configuration across batch sizes 8 to 256. They got there by pulling in DeepSeek's open-source kernels and doing a lot of prefill and decode tuning, and they ran it at DEP32 where DeepInfer used EP64DP64. Fun detail: DeepInfer was originally built on vLLM, so open source came full circle here. If you serve DeepSeek models, rerun your own benchmark on the latest vLLM with your real prompt lengths before you take on a second serving stack. Link in Comments