[release/2.11] Add gfx1100, gfx1101, gfx1151 to hipBLASLt preferred archs#3317
[release/2.11] Add gfx1100, gfx1101, gfx1151 to hipBLASLt preferred archs#3317nemanjaudovic wants to merge 1 commit into
Conversation
|
Jenkins build for af33880defa4615ef5ed71f1fbe752d2c7482608 commit finished as FAILURE |
|
@nemanjaudovic Is there any context you can provide to substantiate why we can now include these additional archs in the hipblasLt preferred archs? Either hipblasLT documentation or testing results from your end? |
|
@jithunnair-amd In ROCm 7.12/7.13 timeframe we landed multiple PRs with tuned hipblaslt kernels for these architectures: ROCm/rocm-libraries#7581, ROCm/rocm-libraries#7468, ROCm/rocm-libraries#5946, ROCm/rocm-libraries#3913, ROCm/rocm-libraries#3879, ROCm/rocm-libraries#5130 In rocm/pytorch, these targets have been in prefer_hipblaslt list in release/2.10 and release/2.9: https://github.com/ROCm/pytorch/blob/release/2.10/aten/src/ATen/Context.cpp#L449, https://github.com/ROCm/pytorch/blame/2e97395fce125ce26a03425e7ecd3d82b3a45c1c/aten/src/ATen/Context.cpp#L474 This PR is aligning 2.11 with this and addresses QA reported regression when using nightly builds from main (where this change is not present): ROCM-25429 We also submitted a PR to main to change this so we don't have to keep changing our release branches going forward: pytorch#185375 |
Add gfx1100, gfx1101, gfx1151 to hipBLASLt preferred archs