-
Notifications
You must be signed in to change notification settings - Fork 75
[release/2.8] Add gfx1150/gfx1151 to hipBLASLt supported architecture list #2582
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[release/2.8] Add gfx1150/gfx1151 to hipBLASLt supported architecture list #2582
Conversation
|
Jenkins build for b8bf96307449423ce014ec608fc1a0a1348f8033 commit finished as FAILURE |
|
@mstankov-amd gfx115x does not seem supported in hipBLASLt CC. @jagadish-amd |
|
hipBLASLt support for gfx1150/gfx1151 has been added since ROCm 6.4.1: https://github.com/ROCm/hipBLASLt/blob/release/rocm-rel-6.4/tensilelite/Tensile/Source/lib/include/Tensile/AMDGPU.hpp#L74 |
|
you're right @slojosic-amd, it is supported (according to great help from @jagadish-amd who figured it out, it is enabled in 6.4.2). Given this PR also updates the preferred backend, have you benchmarked hipBLASLt versus rocBLAS for perf? |
|
probably needs ROCm 6.4.2 condition, (reference ROCm/hipBLASLt#2067) |
hipBLASLt kernel yaml files for gfx1150 and gfx1151 are based on Navi33 yaml files which are again only copy of Navi31 yaml files. We have done some internal testing on Navi31 and it depends from model to model but overall conclusion is that we should change prefered BLAS backend for all RDNA 3/3.5 targets to be hipBLASLt. Please check this PR also: https://github.com/pytorch/pytorch/pull/164744/files |
Done: 39aaf53 |
|
Jenkins build for a62940d9960c0905ac7bb1904b7ba29b7963b544 commit finished as NOT_BUILT Detected error during Pytorch building: |
|
Jenkins build for a62940d9960c0905ac7bb1904b7ba29b7963b544 commit finished as FAILURE |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
lgtm, approved.
No description provided.