Question about performance

[SwibonX]

Could you please provide the relevant code for performance testing? Because during my testing process, the performance seems to be worse than that of vLLM.

[interestingLSY]

I disabled CUDA Graph (--enforce-eager) and multi-step scheduling when benchmarking vLLM. Could you please test this configuration?