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

Update rope_scaling to rope_parameters in preparation for Transformers v5 (#28542)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-11-19 18:06:36 +01:00
committed by GitHub
parent d44e9df7d4
commit a8b70304d6
104 changed files with 542 additions and 910 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -156,8 +156,6 @@ class BambaAttentionDecoderLayer(nn.Module):
prefix: str = "",
) -> None:
super().__init__()
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
self.hidden_size = config.hidden_size
tp_size = get_tensor_model_parallel_world_size()
@@ -178,7 +176,6 @@ class BambaAttentionDecoderLayer(nn.Module):
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5
self.rope_theta = rope_theta
self.max_position_embeddings = max_position_embeddings
if hasattr(config, "partial_rotary_factor"):
@@ -192,8 +189,7 @@ class BambaAttentionDecoderLayer(nn.Module):
head_size=self.head_dim,
rotary_dim=rotary_dim,
max_position=max_position_embeddings,
rope_scaling=rope_scaling,
base=rope_theta,
rope_parameters=config.rope_parameters,
is_neox_style=True,
dtype=torch.get_default_dtype(), # see impl of get_rope
)