Fix MLA decode path returning unwritten (padded) rows (#19902)
This commit is contained in:
@@ -1704,7 +1704,8 @@ class AiterAttnBackend(AttentionBackend):
|
|||||||
num_kv_splits=num_kv_splits,
|
num_kv_splits=num_kv_splits,
|
||||||
)
|
)
|
||||||
|
|
||||||
return o[q_mask]
|
total_valid_q = int(qo_indptr[-1].item())
|
||||||
|
return o[:total_valid_q]
|
||||||
else:
|
else:
|
||||||
o = q.new_empty(
|
o = q.new_empty(
|
||||||
(q.shape[0], layer.tp_q_head_num, layer.v_head_dim),
|
(q.shape[0], layer.tp_q_head_num, layer.v_head_dim),
|
||||||
|
|||||||
Reference in New Issue
Block a user