From 86ce4d10e290e6179e6f1d5ee392f582814d5747 Mon Sep 17 00:00:00 2001 From: Jee Jee Li Date: Mon, 21 Sep 2026 10:59:11 +0800 Subject: [PATCH] [BUGFIX][HY4] Record indexer completion event for full CUDA graph capture (#57811) Signed-off-by: Jee Jee Li --- vllm/models/hy_v4/nvidia/attention.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/vllm/models/hy_v4/nvidia/attention.py b/vllm/models/hy_v4/nvidia/attention.py index 5e0a33f183..7f963cc707 100644 --- a/vllm/models/hy_v4/nvidia/attention.py +++ b/vllm/models/hy_v4/nvidia/attention.py @@ -758,6 +758,8 @@ class HYV4MLAAttention(nn.Module): """ if self.indexer is not None and self.is_sparse and not self.skip_topk: self.indexer(hidden_states, q_c, positions, self.indexer_rope_emb) + if self.is_sparse: + self.mla_attn.impl.record_logical_topk_ready() # type: ignore[attr-defined] out.copy_( self.mla_attn( q,