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

Standardise get_rope to use rope_parameters["partial_rotary_factor"], not rotary_dim (#30389)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-12-11 20:45:23 +00:00
committed by GitHub
parent 92fea56fd1
commit cf3eacfe58
83 changed files with 260 additions and 314 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/models/bamba.py
View File

@@ -178,14 +178,11 @@ class BambaAttentionDecoderLayer(nn.Module):
self.scaling = self.head_dim**-0.5
self.max_position_embeddings = max_position_embeddings
if hasattr(config, "attn_rotary_emb"):
rotary_dim = config.attn_rotary_emb # for backward compatibility
else:
rotary_dim = self.head_dim # default
rotary_dim = getattr(config, "attn_rotary_emb", self.head_dim)
config.rope_parameters["partial_rotary_factor"] = rotary_dim / self.head_dim
self.rotary_emb = get_rope(
head_size=self.head_dim,
rotary_dim=rotary_dim,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
is_neox_style=True,