From 3576691c426ea7565789da98b8440600f8ca488e Mon Sep 17 00:00:00 2001 From: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Date: Sat, 26 Sep 2026 11:35:28 -0400 Subject: [PATCH] [GLM5.3 Bug] Fix sparse indexer attn topk backend selection (#58594) Signed-off-by: yewentao256 --- vllm/models/deepseek_v32/attention.py | 1 + 1 file changed, 1 insertion(+) diff --git a/vllm/models/deepseek_v32/attention.py b/vllm/models/deepseek_v32/attention.py index e6fb00434a..f7d0e76e54 100644 --- a/vllm/models/deepseek_v32/attention.py +++ b/vllm/models/deepseek_v32/attention.py @@ -495,6 +495,7 @@ class DeepseekV32Attention(MLAAttention): self._vllm_config.parallel_config.cp_kv_cache_interleave_size ), skip_topk_buffer_clear=True, + topk_backend=self.indexer.indexer_op.topk_backend, ) self.impl.record_logical_topk_ready() # type: ignore[attr-defined]