A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 16:30 UTC
[Revert] Fix performance regression for GLM-4.7-GPTQ decode and MTP acceptance rate (#33771)
Signed-off-by: aabbccddwasd <aabbccddwasd@qq.com>
aabbccddwasd committed
Feb 9, 2026 at 00:13 UTC
179ae7da8f48f76b0db4dc3e331153b24a36a96a
1 file changed
+1
-3
vllm/v1/attention/backends/flashinfer.py
+1
-3
@@ -919,9 +919,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
919
# Guard access to seq_lens_cpu, which may not always be needed
920
# and can be expensive to retrieve in async mode.
921
needs_seq_lens_cpu = self.use_dcp or use_cascade or not is_only_trtllm_decode
922
- seq_lens_cpu = (
923
- common_attn_metadata.seq_lens.cpu() if needs_seq_lens_cpu else None
924
- )
922
+ seq_lens_cpu = common_attn_metadata.seq_lens_cpu if needs_seq_lens_cpu else None
923
seq_lens_np = seq_lens_cpu.numpy() if seq_lens_cpu is not None else None
924
num_blocks_np = (
925
(seq_lens_np + (page_size - 1)) // page_size