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

15
vllm/model_executor/models/qwen3.py
View File

@@ -42,6 +42,7 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.config import set_default_rope_theta
from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
from .qwen2 import Qwen2MLP as Qwen3MLP
@@ -57,14 +58,13 @@ class Qwen3Attention(nn.Module):
hidden_size: int,
num_heads: int,
num_kv_heads: int,
rope_parameters: dict,
max_position: int = 4096 * 32,
head_dim: int | None = None,
rms_norm_eps: float = 1e-06,
qkv_bias: bool = False,
rope_theta: float = 10000,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
rope_scaling: tuple | None = None,
prefix: str = "",
attn_type: str = AttentionType.DECODER,
dual_chunk_attention_config: dict[str, Any] | None = None,
@@ -89,7 +89,6 @@ class Qwen3Attention(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.dual_chunk_attention_config = dual_chunk_attention_config
self.qkv_proj = QKVParallelLinear(
@@ -113,8 +112,7 @@ class Qwen3Attention(nn.Module):
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position,
base=self.rope_theta,
rope_scaling=rope_scaling,
rope_parameters=rope_parameters,
dual_chunk_attention_config=dual_chunk_attention_config,
)
self.attn = Attention(
@@ -166,9 +164,7 @@ class Qwen3DecoderLayer(nn.Module):
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
# Requires transformers > 4.32.0
rope_theta = getattr(config, "rope_theta", 1000000)
rope_scaling = getattr(config, "rope_scaling", None)
set_default_rope_theta(config, default_theta=1000000)
dual_chunk_attention_config = getattr(
config, "dual_chunk_attention_config", None
)
@@ -187,13 +183,12 @@ class Qwen3DecoderLayer(nn.Module):
num_heads=config.num_attention_heads,
max_position=config.max_position_embeddings,
num_kv_heads=config.num_key_value_heads,
rope_theta=rope_theta,
rms_norm_eps=config.rms_norm_eps,
qkv_bias=getattr(config, "attention_bias", False),
head_dim=getattr(config, "head_dim", None),
cache_config=cache_config,
quant_config=quant_config,
rope_scaling=rope_scaling,
rope_parameters=config.rope_parameters,
prefix=f"{prefix}.self_attn",
attn_type=attn_type,
dual_chunk_attention_config=dual_chunk_attention_config,