pytorch icon indicating copy to clipboard operation
pytorch copied to clipboard
verified publisher icon ROCmSoftwarePlatform

[ROCm] Set vector size to 4 for in-place vectorized elementwise operations for MI300

Open jerrymannil opened this issue 9 months ago • 1 comments

  • In-place vectorized elementwise operations are slower on MI300 with bigger vector size

jerrymannil avatar Apr 09 '25 23:04 jerrymannil

Jenkins build for d07bcc9ebc6dad4abcec0b6b6ee1e4b57f412c4b commit is in progress Links: Blue Ocean view / Build artifacts

rocm-repo-management-api[bot] avatar Apr 09 '25 23:04 rocm-repo-management-api[bot]