[NPU] Use use_dsa to dispatch Ascend DSA attention (#28436)
This commit is contained in:
@@ -45,12 +45,12 @@ def handle_attention_ascend(attn, forward_batch):
|
|||||||
and not forward_batch.forward_mode.is_target_verify()
|
and not forward_batch.forward_mode.is_target_verify()
|
||||||
and not forward_batch.forward_mode.is_draft_extend_v2()
|
and not forward_batch.forward_mode.is_draft_extend_v2()
|
||||||
):
|
):
|
||||||
if hasattr(attn, "indexer"):
|
if hasattr(attn, "use_dsa") and attn.use_dsa:
|
||||||
return AttnForwardMethod.DSA_NPU
|
return AttnForwardMethod.DSA_NPU
|
||||||
else:
|
else:
|
||||||
return AttnForwardMethod.MHA_NPU
|
return AttnForwardMethod.MHA_NPU
|
||||||
else:
|
else:
|
||||||
if hasattr(attn, "indexer"):
|
if hasattr(attn, "use_dsa") and attn.use_dsa:
|
||||||
return AttnForwardMethod.DSA_NPU
|
return AttnForwardMethod.DSA_NPU
|
||||||
else:
|
else:
|
||||||
return AttnForwardMethod.MLA_NPU
|
return AttnForwardMethod.MLA_NPU
|
||||||
|
|||||||
Reference in New Issue
Block a user