ScaleLLM icon indicating copy to clipboard operation
ScaleLLM copied to clipboard

cuda graph capture may occasionally become stuck with multiple gpus.

Open guocuimi opened this issue 1 year ago • 0 comments

It is a known issue that CUDA graph capture may occasionally become stuck when multiple workers are in use. further investigation is needed.

Disabled cuda graph for multiple gpus by default for now.

guocuimi avatar Apr 18 '24 21:04 guocuimi