Preferences

The sglang and vllm numbers are with cuda graphs enabled.

Having said that, 1B model is an extreme example - hence the 1.5x speedup. For regular models and batch sizes this would probably buy you a few percent.


This item has no comments currently.