biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Revert] Fix performance regression for GLM-4.7-GPTQ decode and MTP acceptance rate (#33771)

Browse Source 
Signed-off-by: aabbccddwasd <aabbccddwasd@qq.com>
This commit is contained in:
aabbccddwasd
2026-02-09 00:13:24 +08:00
committed by GitHub
parent c4df59ad43
commit 179ae7da8f
1 changed files with 1 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/v1/attention/backends/flashinfer.py
View File

@@ -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