Fix accuracy drop of dsv3 run in dp enablement (#8677)

Co-authored-by: wunhuang <wunhuang@amd.com>
This commit is contained in:
kk
2025-09-05 07:51:16 +08:00
committed by GitHub
parent e96973742c
commit 918e3d4c27
2 changed files with 99 additions and 68 deletions

View File

@@ -1085,7 +1085,13 @@ class DeepseekV2AttentionMLA(nn.Module):
and not forward_batch.forward_mode.is_target_verify()
and not forward_batch.forward_mode.is_draft_extend()
):
return AttnForwardMethod.MHA
if is_dp_attention_enabled():
if sum(forward_batch.extend_prefix_lens_cpu) == 0:
return AttnForwardMethod.MHA
else:
return AttnForwardMethod.MLA
else:
return AttnForwardMethod.MHA
else:
return AttnForwardMethod.MLA
else: