xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-15 10:13:37 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Disable FlexAttention direct block mask building for encoder-only models (#27344)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py 2025-10-23 00:39:08 +08:00 committed by GitHub
parent c9461e05a4
commit 084a9dae80
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/v1/attention/backends/flex_attention.py
View File

@ -658,7 +658,10 @@ class FlexAttentionMetadataBuilder(AttentionMetadataBuilder[FlexAttentionMetadat
total_cache_tokens=total_cache_tokens, total_cache_tokens=total_cache_tokens,
decode_offset=offset_tensor, decode_offset=offset_tensor,
num_blocks_per_seq=num_blocks_per_seq, num_blocks_per_seq=num_blocks_per_seq,
direct_build=self.direct_build, # FIXME(Isotr0py): direct build has issue to build bidirectional
# attention block mask for encoder-only models, disable it temporarily.
# see: https://github.com/vllm-project/vllm/pull/27329#issuecomment-3431484053
direct_build=(self.direct_build and common_attn_metadata.causal),
q_block_size=self.q_block_size, q_block_size=self.q_block_size,
kv_block_size=self.kv_block_size, kv_block_size=self.kv_block_size,
) )