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

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
vllm/model_executor/models/qwen3.py
View File

@@ -22,6 +22,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Qwen3 model compatible with HuggingFace weights."""
from collections.abc import Iterable
from typing import Any, Optional, Union
@@ -35,8 +36,7 @@ from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
@@ -46,14 +46,12 @@ from vllm.sequence import IntermediateTensors
from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
from .qwen2 import Qwen2MLP as Qwen3MLP
from .qwen2 import Qwen2Model
from .utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
maybe_prefix)
from .utils import AutoWeightsLoader, PPMissingLayer, extract_layer_index, maybe_prefix
logger = init_logger(__name__)
class Qwen3Attention(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -131,7 +129,9 @@ class Qwen3Attention(nn.Module):
**{
"layer_idx": extract_layer_index(prefix),
"dual_chunk_attention_config": dual_chunk_attention_config,
} if dual_chunk_attention_config else {},
}
if dual_chunk_attention_config
else {},
)
self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
@@ -144,12 +144,10 @@ class Qwen3Attention(nn.Module):
qkv, _ = self.qkv_proj(hidden_states)
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
# Add qk-norm
q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim,
self.head_dim)
q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim)
q_by_head = self.q_norm(q_by_head)
q = q_by_head.view(q.shape)
k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim,
self.head_dim)
k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim)
k_by_head = self.k_norm(k_by_head)
k = k_by_head.view(k.shape)
q, k = self.rotary_emb(positions, q, k)
@@ -159,7 +157,6 @@ class Qwen3Attention(nn.Module):
class Qwen3DecoderLayer(nn.Module):
def __init__(
self,
config: Qwen3Config,
@@ -172,9 +169,9 @@ class Qwen3DecoderLayer(nn.Module):
# Requires transformers > 4.32.0
rope_theta = getattr(config, "rope_theta", 1000000)
rope_scaling = getattr(config, "rope_scaling", None)
dual_chunk_attention_config = getattr(config,
"dual_chunk_attention_config",
None)
dual_chunk_attention_config = getattr(
config, "dual_chunk_attention_config", None
)
# By default, Qwen3 uses causal attention as it is a decoder-only model.
# You can override the HF config with `is_causal=False` to enable
@@ -192,8 +189,8 @@ class Qwen3DecoderLayer(nn.Module):
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),
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,
@@ -208,10 +205,10 @@ class Qwen3DecoderLayer(nn.Module):
quant_config=quant_config,
prefix=f"{prefix}.mlp",
)
self.input_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.post_attention_layernorm = RMSNorm(config.hidden_size,
eps=config.rms_norm_eps)
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.post_attention_layernorm = RMSNorm(
config.hidden_size, eps=config.rms_norm_eps
)
def forward(
self,
@@ -224,16 +221,14 @@ class Qwen3DecoderLayer(nn.Module):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
else:
hidden_states, residual = self.input_layernorm(
hidden_states, residual)
hidden_states, residual = self.input_layernorm(hidden_states, residual)
hidden_states = self.self_attn(
positions=positions,
hidden_states=hidden_states,
)
# Fully Connected
hidden_states, residual = self.post_attention_layernorm(
hidden_states, residual)
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
hidden_states = self.mlp(hidden_states)
return hidden_states, residual
@@ -251,13 +246,13 @@ ALL_DECODER_LAYER_TYPES = {
"positions": -1,
"intermediate_tensors": 0,
"inputs_embeds": 0,
})
}
)
class Qwen3Model(Qwen2Model):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__(vllm_config=vllm_config,
prefix=prefix,
decoder_layer_type=Qwen3DecoderLayer)
super().__init__(
vllm_config=vllm_config, prefix=prefix, decoder_layer_type=Qwen3DecoderLayer
)
class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
@@ -283,25 +278,28 @@ class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
self.lora_config = lora_config
self.quant_config = quant_config
self.model = Qwen3Model(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.model = Qwen3Model(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
if get_pp_group().is_last_rank:
if config.tie_word_embeddings:
self.lm_head = self.model.embed_tokens
else:
self.lm_head = ParallelLMHead(config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(
prefix, "lm_head"))
self.lm_head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
else:
self.lm_head = PPMissingLayer()
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
self.model.make_empty_intermediate_tensors
)
def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
self.model.aux_hidden_state_layers = layers
@@ -320,8 +318,9 @@ class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.model(input_ids, positions, intermediate_tensors,
inputs_embeds)
hidden_states = self.model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return hidden_states
def compute_logits(
@@ -331,11 +330,9 @@ class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
logits = self.logits_processor(self.lm_head, hidden_states)
return logits
def load_weights(self, weights: Iterable[tuple[str,
torch.Tensor]]) -> set[str]:
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(
self,
skip_prefixes=(["lm_head."]
if self.config.tie_word_embeddings else None),
skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
)
return loader.load_weights(weights)