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

9
vllm/model_executor/models/minimax_text_01.py
View File

@@ -188,7 +188,7 @@ class MiniMaxText01Attention(nn.Module):
num_kv_heads: int,
rotary_dim: int,
max_position: int = 4096 * 32,
rope_theta: float = 10000,
rope_parameters: dict | None = None,
sliding_window: int | None = None,
quant_config: QuantizationConfig | None = None,
layer_idx: int = None,
@@ -214,7 +214,6 @@ class MiniMaxText01Attention(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.sliding_window = sliding_window
self.prefix = prefix
@@ -247,7 +246,7 @@ class MiniMaxText01Attention(nn.Module):
head_size=self.head_dim,
rotary_dim=rotary_dim,
max_position=max_position,
base=int(rope_theta),
rope_parameters=rope_parameters,
is_neox_style=True,
dtype=torch.float32,
)
@@ -287,8 +286,6 @@ class MiniMaxText01DecoderLayer(nn.Module):
self.hidden_size = config.hidden_size
self.expert_num = expert_num
rope_theta = getattr(config, "rope_theta", 10000)
head_dim = getattr(config, "head_dim", None)
if head_dim is None:
head_dim = config.hidden_size // config.num_attention_heads
@@ -328,7 +325,7 @@ class MiniMaxText01DecoderLayer(nn.Module):
else head_dim,
num_kv_heads=config.num_key_value_heads,
max_position=max_position_embeddings,
rope_theta=rope_theta,
rope_parameters=config.rope_parameters,
sliding_window=config.sliding_window,
quant_config=quant_config,
layer_idx=self._ilayer,