You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
How to reproduce the benchmark results for SGLang v0.3.0 compared to vLLM v0.6.0
In short, with multi step enabled, in online scenarios that we benchmarked, the Median TTFT of vLLM is 3 times that of SGLang, and the Median ITL is 10 times that of SGLang. Lower Median TTFT and ITL are better. vLLM's multi-step optimization did not improve throughput while ensuring lower Median TTFT and ITL. Also, under maximum throughput benchmark, if vLLM does not set gpu util to 0.95 separately and uses the default configuration instead, its maximum throughput is lower than that of SGLang.
We referred to the reproduction method in vllm-project/vllm#8176, and added the --num-scheduler-steps 10 parameter when starting the vLLM server. The gpu_memory_utilization of vLLM is by default 0.9 at both TP 1 and TP 4, while SGLang's mem_frac is 0.88 at TP 1 and 0.85 at TP 4, so we manually set it to 0.88 at TP 4.