search

intel / intel-xpu-backend-for-triton

OpenAI Triton backend for Intel® GPUs
MIT License
131 stars 39 forks source link

[PyTorch Upstream] do_bench functionaliy error #1563

Closed etaf closed 2 months ago

etaf commented 3 months ago

Hi, we found two implementation issues in triton's do_bench when enabling max-autotune feature in Pytorch.

  1. do_bench should return the number of ms , but currently is ns. https://github.com/intel/intel-xpu-backend-for-triton/blob/1b2f15840e0d70eec50d84c7a0575cb835524def/python/triton/testing.py#L12-L21
  2. In do_bench , if USE_WALL_TIME, there should be a synchronize() between start/end time record, But currently is not. https://github.com/intel/intel-xpu-backend-for-triton/blob/1b2f15840e0d70eec50d84c7a0575cb835524def/python/triton/testing.py#L139-L148
etaf commented 3 months ago

@alexbaden can you help to check?

whitneywhtsang commented 3 months ago

@etaf Please verify. Do you also need this change to be backported to the release branch?

etaf commented 2 months ago

@whitneywhtsang Sorry for my late update, thanks for your work, and currently no need to backport to the release branch.