From 75698e60b3b0db7f443f8bf19d5d3a20ddc4ce0a Mon Sep 17 00:00:00 2001 From: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Date: Mon, 29 Jun 2026 18:45:53 -0400 Subject: [PATCH] [Bug] Fix sparse attention issue for GLM5.2 non-torch compile path (#47083) Signed-off-by: yewentao256 --- vllm/models/deepseek_v32/nvidia/attention.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/vllm/models/deepseek_v32/nvidia/attention.py b/vllm/models/deepseek_v32/nvidia/attention.py index 771a3d4f95..962eb6c57e 100644 --- a/vllm/models/deepseek_v32/nvidia/attention.py +++ b/vllm/models/deepseek_v32/nvidia/attention.py @@ -467,10 +467,6 @@ class DeepseekV32Attention(MLAAttention): index_rope_interleave=self._index_rope_interleave, ) - if attn_metadata is None: - output.zero_() - return - if self.indexer is not None: sparse_attn_indexer( q_c, @@ -492,6 +488,10 @@ class DeepseekV32Attention(MLAAttention): True, # skip_topk_buffer_clear (fused_norm_rope already did it) ) + if attn_metadata is None: + output.zero_() + return + num_actual = attn_metadata.num_actual_tokens # type: ignore[attr-defined] kv_cache = self.kv_cache if self._fp8_kv_needs_view: