[BUGFIX][HY4] Record indexer completion event for full CUDA graph capture (#57811)
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
Jee Jee Li committed
Sep 21, 2026 at 10:59 UTC
86ce4d10e290e6179e6f1d5ee392f582814d5747
1 file changed
+2
vllm/models/hy_v4/nvidia/attention.py
+2
@@ -758,6 +758,8 @@ class HYV4MLAAttention(nn.Module):
758
"""
759
if self.indexer is not None and self.is_sparse and not self.skip_topk:
760
self.indexer(hidden_states, q_c, positions, self.indexer_rope_emb)
761
+ if self.is_sparse:
762
+ self.mla_attn.impl.record_logical_topk_ready() # type: ignore[attr-defined]
763
out.copy_(
764
self.mla_attn(
765
q,