ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-15 19:44:39 +00:00
Code Issues Packages Projects Releases Wiki Activity

update: remove restriction of group mode hd192 lse

Browse Source
This commit is contained in:
Jim
2025-05-27 20:51:15 +08:00
parent 128f5a1eab
commit b4170a2005
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View File

@@ -517,7 +517,7 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, optdim_list, mask_impl
continue
if hdim == 192 and tile.F_bn1 == 128:
# NOTE: this is used to speedup deepseek prefill case, we don't gen training
if pipeline.F_bias != 'no' or pipeline.F_lse == 't' or pipeline.F_dropout == 't':
if pipeline.F_bias != 'no' or pipeline.F_dropout == 't':
continue
# logits_soft_cap is only allowed if no bias
if not ((pipeline.F_logits == 't' and pipeline.F_bias == 'no') or pipeline.F_logits == 'f'):