Fix MLA decode path returning unwritten (padded) rows (#19902)

This commit is contained in:
Clint
2026-03-06 02:54:29 -06:00
committed by GitHub
parent 0252ca8255
commit 27053aa5ed

View File

@@ -1704,7 +1704,8 @@ class AiterAttnBackend(AttentionBackend):
num_kv_splits=num_kv_splits,
)
return o[q_mask]
total_valid_q = int(qo_indptr[-1].item())
return o[:total_valid_q]
else:
o = q.new_empty(
(q.shape[0], layer.tp_q_head_num, layer.v_head_dim),