SGLang claims 873 tokens per second for DeepSeek V4.1 Flash; a user questions its benchmarks
SGLang says it reached that speed at batch size 1 on four GB300s. A user challenges a 781-tokens-per-second-per-user figure in its dsv4-flash cookbook, saying the reproduction script uses random inputs that lead to repetitive output.
TLDR
SGLang says it reached 873 tokens per second with DeepSeek V4.1 Flash at batch size 1 on four GB300s within 24 hours of launch. The project says its new engineering guide documents the optimizations and includes instructions to reproduce the results. A user questions a 781-tokens-per-second-per-user figure advertised in SGLang’s dsv4-flash cookbook, also at batch size 1. The user says the reproduction script uses random token IDs that lead to repetitive output, and claims actual performance is only 200–300 tokens per second.
Combined views
356
1 Source, first seen 19d ago