Avoid computing lse in Ragged Prefill when there's no prefix. (#5476)
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
This commit is contained in:
co-authored by
Baizhou Zhang
parent
e465b08ddb
commit
bfa3922451
@@ -348,7 +348,7 @@ class FlashInferMLAAttnBackend(AttentionBackend):
|
||||
|
||||
if self.forward_metadata.use_ragged:
|
||||
# ragged prefill
|
||||
o, _ = self.prefill_wrapper_ragged.forward_return_lse(
|
||||
o = self.prefill_wrapper_ragged.forward(
|
||||
qall,
|
||||
k.view(-1, layer.tp_k_head_num, layer.head_dim),
|
||||
v.view(-1, layer.tp_k_head_num, layer.v_head_dim),
|
||||
|
||||
Reference in New Issue
Block a user