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

145
vllm/model_executor/models/qwen.py
View File

@@ -6,6 +6,7 @@
# Copyright (c) Alibaba Cloud.
# LICENSE: https://huggingface.co/Qwen/Qwen-7B/blob/main/LICENSE
"""Inference-only QWen model compatible with HuggingFace weights."""
import json
from collections.abc import Iterable
from itertools import islice
@@ -21,21 +22,28 @@ from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
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
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead, VocabParallelEmbedding)
ParallelLMHead,
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
class QWenMLP(nn.Module):
@@ -51,16 +59,15 @@ class QWenMLP(nn.Module):
):
super().__init__()
self.gate_up_proj = MergedColumnParallelLinear(
hidden_size, [intermediate_size] * 2,
bias=False,
quant_config=quant_config)
self.c_proj = RowParallelLinear(intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config)
hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config
)
self.c_proj = RowParallelLinear(
intermediate_size, hidden_size, bias=False, quant_config=quant_config
)
if hidden_act != "silu":
raise ValueError(f"Unsupported activation: {hidden_act}. "
"Only silu is supported for now.")
raise ValueError(
f"Unsupported activation: {hidden_act}. Only silu is supported for now."
)
self.act_fn = SiluAndMul()
def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -71,7 +78,6 @@ class QWenMLP(nn.Module):
class QWenAttention(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -85,12 +91,10 @@ class QWenAttention(nn.Module):
):
super().__init__()
self.hidden_size = hidden_size
tensor_model_parallel_world_size = get_tensor_model_parallel_world_size(
)
tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
self.total_num_heads = num_heads
assert self.total_num_heads % tensor_model_parallel_world_size == 0
self.num_heads = (self.total_num_heads //
tensor_model_parallel_world_size)
self.num_heads = self.total_num_heads // tensor_model_parallel_world_size
self.head_dim = hidden_size // self.total_num_heads
self.c_attn = QKVParallelLinear(
hidden_size,
@@ -114,12 +118,14 @@ class QWenAttention(nn.Module):
base=rope_theta,
rope_scaling=rope_scaling,
)
self.attn = Attention(self.num_heads,
self.head_dim,
self.scaling,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn")
self.attn = Attention(
self.num_heads,
self.head_dim,
self.scaling,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
)
def forward(
self,
@@ -135,7 +141,6 @@ class QWenAttention(nn.Module):
class QWenBlock(nn.Module):
def __init__(
self,
config: PretrainedConfig,
@@ -148,20 +153,22 @@ class QWenBlock(nn.Module):
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
self.attn = QWenAttention(config.hidden_size,
config.num_attention_heads,
config.max_position_embeddings,
rope_theta=rope_theta,
rope_scaling=rope_scaling,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn")
self.attn = QWenAttention(
config.hidden_size,
config.num_attention_heads,
config.max_position_embeddings,
rope_theta=rope_theta,
rope_scaling=rope_scaling,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
)
self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
self.mlp = QWenMLP(config.hidden_size,
config.intermediate_size // 2,
quant_config=quant_config)
self.mlp = QWenMLP(
config.hidden_size, config.intermediate_size // 2, quant_config=quant_config
)
def forward(
self,
@@ -188,7 +195,6 @@ class QWenBlock(nn.Module):
@support_torch_compile
class QWenModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -205,13 +211,13 @@ class QWenModel(nn.Module):
)
self.start_layer, self.end_layer, self.h = make_layers(
config.num_hidden_layers,
lambda prefix: QWenBlock(
config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.h")
lambda prefix: QWenBlock(config, cache_config, quant_config, prefix=prefix),
prefix=f"{prefix}.h",
)
self.ln_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
self.make_empty_intermediate_tensors = (
make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size))
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
["hidden_states", "residual"], config.hidden_size
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.wte(input_ids)
@@ -241,16 +247,14 @@ class QWenModel(nn.Module):
residual,
)
if not get_pp_group().is_last_rank:
return IntermediateTensors({
"hidden_states": hidden_states,
"residual": residual
})
return IntermediateTensors(
{"hidden_states": hidden_states, "residual": residual}
)
hidden_states, _ = self.ln_f(hidden_states, residual)
return hidden_states
class QWenBaseModel(nn.Module):
def __init__(
self,
*,
@@ -265,18 +269,21 @@ class QWenBaseModel(nn.Module):
self.config = config
self.multimodal_config = multimodal_config
self.quant_config = quant_config
self.transformer = transformer_type(vllm_config=vllm_config,
prefix=maybe_prefix(
prefix, "transformer"))
self.lm_head = ParallelLMHead(config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"))
self.transformer = transformer_type(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer")
)
self.lm_head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
if self.config.tie_word_embeddings:
self.lm_head.weight = self.transformer.wte.weight
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.transformer.make_empty_intermediate_tensors)
self.transformer.make_empty_intermediate_tensors
)
def compute_logits(
self,
@@ -285,8 +292,7 @@ class QWenBaseModel(nn.Module):
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]:
stacked_params_mapping = [
# (param_name, shard_name, shard_id)
("gate_up_proj", "w2", 0),
@@ -297,7 +303,7 @@ class QWenBaseModel(nn.Module):
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
for (param_name, weight_name, shard_id) in stacked_params_mapping:
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in name:
continue
name = name.replace(weight_name, param_name)
@@ -319,8 +325,7 @@ class QWenBaseModel(nn.Module):
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader = getattr(param, "weight_loader", default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
@@ -338,14 +343,13 @@ class QWenLMHeadModel(QWenBaseModel, SupportsPP, SupportsLoRA):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
config = vllm_config.model_config.hf_config
if hasattr(config, "visual"):
hf_overrides = {
"architectures": ["QwenVLForConditionalGeneration"]
}
hf_overrides = {"architectures": ["QwenVLForConditionalGeneration"]}
raise RuntimeError(
"The configuration of this model indicates that it supports "
"vision inputs, but you instantiated the text-only version "
"of this model. Please use the vision model by setting "
f"`--hf-overrides '{json.dumps(hf_overrides)}'`")
f"`--hf-overrides '{json.dumps(hf_overrides)}'`"
)
super().__init__(vllm_config=vllm_config, prefix=prefix)
@@ -356,6 +360,7 @@ class QWenLMHeadModel(QWenBaseModel, SupportsPP, SupportsLoRA):
intermediate_tensors: Optional[IntermediateTensors] = None,
inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
hidden_states = self.transformer(input_ids, positions,
intermediate_tensors, inputs_embeds)
hidden_states = self.transformer(
input_ids, positions, intermediate_tensors, inputs_embeds
)
return hidden_states