github/ROCm
1
1
Fork 0
mirror of https://github.com/ROCm/ROCm.git synced 2026-01-06 21:33:57 -05:00
Code Issues Packages Projects Releases Wiki Activity

711 hipBLASLT performance decline known issue added (#5730)

Browse Source 
* hipBLASLT performance decline known issue added

* Update RELEASE.md

Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>

* GitHub Issue added

* Ram's feedback incorporated

* GitHub Issue added

* Update RELEASE.md

Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>

---------

Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
This commit is contained in:
Pratik Basyal
2025-12-03 08:50:25 -05:00
committed by GitHub
parent b4031ef23c
commit e8fdc34b71
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
RELEASE.md
View File

@@ -874,7 +874,11 @@ Due to partial data corruption in the Electrically Erasable Programmable Read-On
### Incorrect results in gemm_ex operations for rocBLAS and hipBLAS ### Incorrect results in gemm_ex operations for rocBLAS and hipBLAS
Some `gemm_ex` operations with 8-bit input datatypes (`int8`, `float8`, `bfloat8`) for specific matrix dimensions (K = 1 and number of workgroup > 1) might yield in incorrect results. The issue results from incorrect tailloop code that fails to consider workgroup index when calculating valid element size. The issue will be fixed in a future ROCm release. Some `gemm_ex` operations with 8-bit input data types (`int8`, `float8`, `bfloat8`) for specific matrix dimensions (K = 1 and number of workgroups > 1) might yield incorrect results. The issue results from incorrect tailloop code that fails to consider workgroup index when calculating valid element size. The issue will be fixed in a future ROCm release. See [GitHub issue #5722](https://github.com/ROCm/ROCm/issues/5722).
### hipBLASLt performance variation for a particular FP8 GEMM operation on AMD Instinct MI325X GPUs
If you’re using hipBLASLt on AMD Instinct MI325X GPUs for large FP8 GEMM operations (such as 9728x8192x65536), you might observe a noticeable performance variation. The issue is currently under investigation and will be fixed in a future ROCm release. See [GitHub issue #5734](https://github.com/ROCm/ROCm/issues/5734).
## ROCm resolved issues ## ROCm resolved issues