ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 13:11:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

[rocm-libraries] ROCm/rocm-libraries#5504 (commit 47f86c7)

Browse Source 
[CK Tile] Add sink token gradient support in FMHA backward
 pass (#5504)

## Motivation

Adds sink token support to the FMHA backward kernel (dot_do_o pipeline):

## Technical Details

- Extend BlockFmhaBwdOGradDotOPipelineProblem with LSEDataType
- Add sink_ptr/d_sink_ptr/lse_ptr/nhead to FmhaBwdOGradDotOCommonKargs
- Compute per-head sink gradient via atomic accumulation in the pipeline
- Update example runner with reference validation for sink gradient

## Test Plan

Add new test case

## Test Result

WIP

## Submission Checklist

- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
This commit is contained in:
Linjun-AMD
2026-04-02 03:17:45 +00:00
committed by assistant-librarian[bot]
parent c1127a36f5
commit 08792e0b31
12 changed files with 380 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py
View File

@@ -533,6 +533,7 @@ using fmha_bwd_dot_do_o_pipeline_problem_{F_idx} = ck_tile::BlockFmhaBwdOGradDot
typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::OGradDataType,
typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::DDataType,
typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
/* BlockSize = M0 = */ {F_bm0},
{F_hdim},
{F_mode},