frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[bug fix] Fix llama4 spec decoding (#22691)

Browse Source 
Signed-off-by: qizixi <qizixi@meta.com>
Co-authored-by: Lu Fang <30275821+houseroad@users.noreply.github.com>
This commit is contained in:
qizixi
2025-08-19 17:53:24 +09:00
committed by GitHub
parent 31fd3265c8
commit 5bfe0dea7a
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/models/llama4.py
View File

@ -195,7 +195,9 @@ class Llama4Attention(nn.Module):
is_neox_style=is_neox_style,
) if not self.nope else None
attn_cls = Attention if self.nope else ChunkedLocalAttention
use_chunked_local_attn = not self.nope and config.attention_chunk_size
attn_cls = (ChunkedLocalAttention
if use_chunked_local_attn else Attention)
self.attn = attn_cls(
self.num_heads,
self.head_dim,
@ -206,7 +208,7 @@ class Llama4Attention(nn.Module):
prefix=f"{prefix}.attn",
**({
"attention_chunk_size": config.attention_chunk_size
} if not self.nope else {}))
} if use_chunked_local_attn else {}))
def _get_attn_scale(self, positions: torch.Tensor) -> torch.Tensor:
floor = torch.floor((positions + 1.0) / self.floor_scale)