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

29
vllm/model_executor/models/chameleon.py
View File

@@ -265,8 +265,7 @@ class ChameleonAttention(nn.Module):
hidden_size: int,
num_heads: int,
num_kv_heads: int,
rope_theta: float = 10000,
rope_scaling: dict[str, Any] | None = None,
rope_parameters: dict[str, Any],
max_position_embeddings: int = 4096,
quant_config: QuantizationConfig | None = None,
bias: bool = False,
@@ -293,7 +292,6 @@ class ChameleonAttention(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
self.qkv_proj = QKVParallelLinear(
@@ -318,8 +316,7 @@ class ChameleonAttention(nn.Module):
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
base=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=rope_parameters,
)
self.attn = Attention(
@@ -369,14 +366,6 @@ class ChameleonDecoderLayer(nn.Module):
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
if rope_scaling is not None and getattr(
config, "original_max_position_embeddings", None
):
rope_scaling["original_max_position_embeddings"] = (
config.original_max_position_embeddings
)
max_position_embeddings = getattr(config, "max_position_embeddings", 4096)
self.self_attn = ChameleonAttention(
@@ -385,8 +374,7 @@ class ChameleonDecoderLayer(nn.Module):
num_kv_heads=getattr(
config, "num_key_value_heads", config.num_attention_heads
),
rope_theta=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=config.rope_parameters,
max_position_embeddings=max_position_embeddings,
quant_config=quant_config,
bias=False,
@@ -439,14 +427,6 @@ class ChameleonSwinDecoderLayer(nn.Module):
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
if rope_scaling is not None and getattr(
config, "original_max_position_embeddings", None
):
rope_scaling["original_max_position_embeddings"] = (
config.original_max_position_embeddings
)
max_position_embeddings = getattr(config, "max_position_embeddings", 4096)
self.self_attn = ChameleonAttention(
@@ -455,8 +435,7 @@ class ChameleonSwinDecoderLayer(nn.Module):
num_kv_heads=getattr(
config, "num_key_value_heads", config.num_attention_heads
),
rope_theta=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=config.rope_parameters,
max_position_embeddings=max_position_embeddings,
quant_config=quant_config,
bias=False,