On Oct. 24, 2026, SGLang said its tests matched DeepInfer on both prefill and decode. It cited up to 1.04× throughput on four GB300 and four B300 GPUs for prefill. For a test on 64 GB300 GPUs, it described its decode result as “up to 7.4% throughput,” without defining what the percentage measures. SGLang also said it provides an end-to-end reproduction guide and a full serving stack for production use.
Explaining the prefill comparison, @BanghuaZ said the team reran DeepInfer on B300 and GB300 hardware and used its fastest tested configuration at each point as a baseline. Both engines ran on the same hardware with matching DP/EP configurations, the author said. @BanghuaZ also said DeepInfer achieved higher prefill throughput on SGLang’s B300 hardware at several configurations than in its published results. The author acknowledged remaining methodological differences and said the team planned to submit its results to AgentX.
Separately, vLLM said it matched DeepInfer on DeepSeek’s decode configuration for DeepSeek-V4.1-Flash across batch sizes 8 to 256, using DEP32 against DeepInfer’s EP64DP64. Its post did not claim prefill parity. In a quote-post of vLLM’s announcement, @woosuk_k said “our results” had been verified in the @SemiAnalysis_ AgentX benchmark.
