From 179ae7da8f48f76b0db4dc3e331153b24a36a96a Mon Sep 17 00:00:00 2001 From: aabbccddwasd <140953076+aabbccddwasd@users.noreply.github.com> Date: Mon, 9 Feb 2026 00:13:24 +0800 Subject: [PATCH] [Revert] Fix performance regression for GLM-4.7-GPTQ decode and MTP acceptance rate (#33771) Signed-off-by: aabbccddwasd --- vllm/v1/attention/backends/flashinfer.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/vllm/v1/attention/backends/flashinfer.py b/vllm/v1/attention/backends/flashinfer.py index 8e81c6fe9..26d372c11 100755 --- a/vllm/v1/attention/backends/flashinfer.py +++ b/vllm/v1/attention/backends/flashinfer.py @@ -919,9 +919,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]): # Guard access to seq_lens_cpu, which may not always be needed # and can be expensive to retrieve in async mode. needs_seq_lens_cpu = self.use_dcp or use_cascade or not is_only_trtllm_decode - seq_lens_cpu = ( - common_attn_metadata.seq_lens.cpu() if needs_seq_lens_cpu else None - ) + seq_lens_cpu = common_attn_metadata.seq_lens_cpu if needs_seq_lens_cpu else None seq_lens_np = seq_lens_cpu.numpy() if seq_lens_cpu is not None else None num_blocks_np = ( (seq_lens_np + (page_size - 1)) // page_size