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

342
vllm/model_executor/models/qwen3_moe.py
View File

@@ -22,6 +22,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Qwen3MoE model compatible with HuggingFace weights."""
import typing
from collections.abc import Callable, Iterable
from itertools import islice
@@ -33,38 +34,51 @@ from torch import nn
from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (get_ep_group, get_pp_group,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather)
from vllm.distributed import (
get_ep_group,
get_pp_group,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather,
)
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
QKVParallelLinear,
ReplicatedLinear,
RowParallelLinear)
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
QKVParallelLinear,
ReplicatedLinear,
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, maybe_remap_kv_scale_name)
default_weight_loader,
maybe_remap_kv_scale_name,
)
from vllm.model_executor.models.utils import sequence_parallel_chunk
from vllm.sequence import IntermediateTensors
from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, PPMissingLayer, extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
logger = init_logger(__name__)
class Qwen3MoeMLP(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -76,19 +90,24 @@ class Qwen3MoeMLP(nn.Module):
) -> None:
super().__init__()
self.gate_up_proj = MergedColumnParallelLinear(
hidden_size, [intermediate_size] * 2,
hidden_size,
[intermediate_size] * 2,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.gate_up_proj")
self.down_proj = RowParallelLinear(intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config,
reduce_results=reduce_results,
prefix=f"{prefix}.down_proj")
prefix=f"{prefix}.gate_up_proj",
)
self.down_proj = RowParallelLinear(
intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config,
reduce_results=reduce_results,
prefix=f"{prefix}.down_proj",
)
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):
@@ -99,7 +118,6 @@ class Qwen3MoeMLP(nn.Module):
class Qwen3MoeSparseMoeBlock(nn.Module):
def __init__(
self,
vllm_config: VllmConfig,
@@ -123,7 +141,8 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
if self.tp_size > config.num_experts:
raise ValueError(
f"Tensor parallel size {self.tp_size} is greater than "
f"the number of experts {config.num_experts}.")
f"the number of experts {config.num_experts}."
)
# Load balancing settings.
vllm_config = get_current_vllm_config()
@@ -132,36 +151,40 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
self.n_logical_experts = self.n_routed_experts
self.n_redundant_experts = eplb_config.num_redundant_experts
self.n_physical_experts = (self.n_logical_experts +
self.n_redundant_experts)
self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
self.n_local_physical_experts = self.n_physical_experts // self.ep_size
self.physical_expert_start = (self.ep_rank *
self.n_local_physical_experts)
self.physical_expert_end = (self.physical_expert_start +
self.n_local_physical_experts)
self.physical_expert_start = self.ep_rank * self.n_local_physical_experts
self.physical_expert_end = (
self.physical_expert_start + self.n_local_physical_experts
)
self.experts = FusedMoE(num_experts=self.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=True,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
prefix=f"{prefix}.experts",
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts,
is_sequence_parallel=self.is_sequence_parallel)
self.experts = FusedMoE(
num_experts=self.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=True,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
prefix=f"{prefix}.experts",
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts,
is_sequence_parallel=self.is_sequence_parallel,
)
self.gate = ReplicatedLinear(config.hidden_size,
config.num_experts,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.gate")
self.gate = ReplicatedLinear(
config.hidden_size,
config.num_experts,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.gate",
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
assert hidden_states.dim(
) <= 2, "Qwen3MoeSparseMoeBlock only supports 1D or 2D inputs"
assert hidden_states.dim() <= 2, (
"Qwen3MoeSparseMoeBlock only supports 1D or 2D inputs"
)
is_input_1d = hidden_states.dim() == 1
num_tokens, hidden_dim = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_dim)
@@ -171,21 +194,21 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
# router_logits: (num_tokens, n_experts)
router_logits, _ = self.gate(hidden_states)
final_hidden_states = self.experts(hidden_states=hidden_states,
router_logits=router_logits)
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits
)
if self.is_sequence_parallel:
final_hidden_states = tensor_model_parallel_all_gather(
final_hidden_states, 0)
final_hidden_states, 0
)
final_hidden_states = final_hidden_states[:num_tokens]
# return to 1d if input is 1d
return final_hidden_states.squeeze(0) if is_input_1d else \
final_hidden_states
return final_hidden_states.squeeze(0) if is_input_1d else final_hidden_states
class Qwen3MoeAttention(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -226,19 +249,23 @@ class Qwen3MoeAttention(nn.Module):
self.max_position_embeddings = max_position_embeddings
self.dual_chunk_attention_config = dual_chunk_attention_config
self.qkv_proj = QKVParallelLinear(hidden_size,
self.head_dim,
self.total_num_heads,
self.total_num_kv_heads,
bias=qkv_bias,
quant_config=quant_config,
prefix=f"{prefix}.qkv_proj")
self.qkv_proj = QKVParallelLinear(
hidden_size,
self.head_dim,
self.total_num_heads,
self.total_num_kv_heads,
bias=qkv_bias,
quant_config=quant_config,
prefix=f"{prefix}.qkv_proj",
)
self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
hidden_size,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.o_proj")
self.o_proj = RowParallelLinear(
self.total_num_heads * self.head_dim,
hidden_size,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.o_proj",
)
self.rotary_emb = get_rope(
self.head_dim,
@@ -259,7 +286,9 @@ class Qwen3MoeAttention(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)
@@ -273,13 +302,11 @@ class Qwen3MoeAttention(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)
@@ -289,7 +316,6 @@ class Qwen3MoeAttention(nn.Module):
class Qwen3MoeDecoderLayer(nn.Module):
def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None:
super().__init__()
@@ -300,11 +326,10 @@ class Qwen3MoeDecoderLayer(nn.Module):
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
max_position_embeddings = getattr(config, "max_position_embeddings",
8192)
dual_chunk_attention_config = getattr(config,
"dual_chunk_attention_config",
None)
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
dual_chunk_attention_config = getattr(
config, "dual_chunk_attention_config", None
)
self.self_attn = Qwen3MoeAttention(
hidden_size=self.hidden_size,
num_heads=config.num_attention_heads,
@@ -313,8 +338,8 @@ class Qwen3MoeDecoderLayer(nn.Module):
rope_scaling=rope_scaling,
max_position_embeddings=max_position_embeddings,
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,
prefix=f"{prefix}.self_attn",
@@ -323,23 +348,27 @@ class Qwen3MoeDecoderLayer(nn.Module):
# `mlp_only_layers` in the config.
layer_idx = extract_layer_index(prefix)
mlp_only_layers = ([] if not hasattr(config, "mlp_only_layers") else
config.mlp_only_layers)
mlp_only_layers = (
[] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
)
if (layer_idx not in mlp_only_layers) and (
config.num_experts > 0 and
(layer_idx + 1) % config.decoder_sparse_step == 0):
self.mlp = Qwen3MoeSparseMoeBlock(vllm_config=vllm_config,
prefix=f"{prefix}.mlp")
config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
):
self.mlp = Qwen3MoeSparseMoeBlock(
vllm_config=vllm_config, prefix=f"{prefix}.mlp"
)
else:
self.mlp = Qwen3MoeMLP(hidden_size=config.hidden_size,
intermediate_size=config.intermediate_size,
hidden_act=config.hidden_act,
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.mlp = Qwen3MoeMLP(
hidden_size=config.hidden_size,
intermediate_size=config.intermediate_size,
hidden_act=config.hidden_act,
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
)
def forward(
self,
@@ -352,23 +381,20 @@ class Qwen3MoeDecoderLayer(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
@support_torch_compile
class Qwen3MoeModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -385,17 +411,17 @@ class Qwen3MoeModel(nn.Module):
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.embed_tokens")
prefix=f"{prefix}.embed_tokens",
)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: Qwen3MoeDecoderLayer(vllm_config=vllm_config,
prefix=prefix),
lambda prefix: Qwen3MoeDecoderLayer(vllm_config=vllm_config, prefix=prefix),
prefix=f"{prefix}.layers",
)
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
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.embed_tokens(input_ids)
@@ -420,10 +446,9 @@ class Qwen3MoeModel(nn.Module):
for layer in islice(self.layers, self.start_layer, self.end_layer):
hidden_states, residual = layer(positions, hidden_states, 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.norm(hidden_states, residual)
return hidden_states
@@ -435,10 +460,10 @@ class Qwen3MoeModel(nn.Module):
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.num_experts,
num_redundant_experts=self.num_redundant_experts)
num_redundant_experts=self.num_redundant_experts,
)
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)
("qkv_proj", "q_proj", "q"),
@@ -449,15 +474,24 @@ class Qwen3MoeModel(nn.Module):
]
# Skip loading extra parameters for GPTQ/modelopt models.
ignore_suffixes = (".bias", "_bias", ".k_scale", "_k_scale",
".v_scale", "_v_scale", ".weight_scale",
"_weight_scale", ".input_scale", "_input_scale")
ignore_suffixes = (
".bias",
"_bias",
".k_scale",
"_k_scale",
".v_scale",
"_v_scale",
".weight_scale",
"_weight_scale",
".input_scale",
"_input_scale",
)
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
expert_params_mapping = self.get_expert_mapping()
for name, loaded_weight in weights:
for (param_name, weight_name, shard_id) in stacked_params_mapping:
for param_name, weight_name, shard_id in stacked_params_mapping:
# Skip non-stacked layers and experts (experts handled below).
if weight_name not in name:
continue
@@ -487,8 +521,7 @@ class Qwen3MoeModel(nn.Module):
continue
param = params_dict[name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader = getattr(param, "weight_loader", default_weight_loader)
if weight_loader == default_weight_loader:
weight_loader(param, loaded_weight)
else:
@@ -513,23 +546,27 @@ class Qwen3MoeModel(nn.Module):
continue
# Skip loading extra parameters for GPTQ/modelopt models.
if name_mapped.endswith(
ignore_suffixes
) and name_mapped not in params_dict:
if (
name_mapped.endswith(ignore_suffixes)
and name_mapped not in params_dict
):
continue
param = params_dict[name_mapped]
# We should ask the weight loader to return success or not
# here since otherwise we may skip experts with other
# available replicas.
weight_loader = typing.cast(Callable[..., bool],
param.weight_loader)
success = weight_loader(param,
loaded_weight,
name_mapped,
shard_id=shard_id,
expert_id=expert_id,
return_success=True)
weight_loader = typing.cast(
Callable[..., bool], param.weight_loader
)
success = weight_loader(
param,
loaded_weight,
name_mapped,
shard_id=shard_id,
expert_id=expert_id,
return_success=True,
)
if success:
name = name_mapped
break
@@ -541,8 +578,7 @@ class Qwen3MoeModel(nn.Module):
continue
# Skip loading extra parameters for GPTQ/modelopt models.
if name.endswith(
ignore_suffixes) and name not in params_dict:
if name.endswith(ignore_suffixes) and name not in params_dict:
continue
# Skip layers on other devices.
if is_pp_missing_parameter(name, self):
@@ -550,7 +586,8 @@ class Qwen3MoeModel(nn.Module):
# Remapping the name of FP8 kv-scale.
if name.endswith("kv_scale"):
remapped_kv_scale_name = name.replace(
".kv_scale", ".attn.kv_scale")
".kv_scale", ".attn.kv_scale"
)
if remapped_kv_scale_name not in params_dict:
logger.warning_once(
"Found kv scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv-scale is not loaded.", # noqa: E501
@@ -561,15 +598,15 @@ class Qwen3MoeModel(nn.Module):
else:
name = remapped_kv_scale_name
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
class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA,
MixtureOfExperts):
class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, MixtureOfExperts):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
@@ -590,17 +627,21 @@ class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA,
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.model = Qwen3MoeModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.lm_head = ParallelLMHead(config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"))
self.model = Qwen3MoeModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
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.model.embed_tokens.weight
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors)
self.model.make_empty_intermediate_tensors
)
# Set MoE hyperparameters
self.expert_weights = []
@@ -652,8 +693,7 @@ class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA,
assert self.num_local_physical_experts == num_local_physical_experts
self.num_physical_experts = num_physical_experts
self.num_local_physical_experts = num_local_physical_experts
self.num_redundant_experts = (num_physical_experts -
self.num_logical_experts)
self.num_redundant_experts = num_physical_experts - self.num_logical_experts
for layer in self.model.layers:
if isinstance(layer.mlp, Qwen3MoeSparseMoeBlock):
moe = layer.mlp
@@ -672,8 +712,9 @@ class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA,
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(
@@ -683,8 +724,7 @@ class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA,
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)
return loader.load_weights(weights)