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

286
vllm/model_executor/models/qwen2_moe.py
View File

@@ -24,6 +24,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Qwen2MoE model compatible with HuggingFace weights."""
from collections.abc import Iterable
from itertools import islice
from typing import Any, Optional, Union
@@ -41,29 +42,36 @@ 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
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
extract_layer_index,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
logger = init_logger(__name__)
class Qwen2MoeMLP(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -75,19 +83,24 @@ class Qwen2MoeMLP(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):
@@ -98,7 +111,6 @@ class Qwen2MoeMLP(nn.Module):
class Qwen2MoeSparseMoeBlock(nn.Module):
def __init__(
self,
config: Qwen2MoeConfig,
@@ -111,37 +123,39 @@ class Qwen2MoeSparseMoeBlock(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}."
)
self.experts = FusedMoE(num_experts=config.num_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
prefix=f"{prefix}.experts")
self.experts = FusedMoE(
num_experts=config.num_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
prefix=f"{prefix}.experts",
)
self.gate = ReplicatedLinear(config.hidden_size,
config.num_experts,
bias=False,
quant_config=None,
prefix=f"{prefix}.gate")
self.gate = ReplicatedLinear(
config.hidden_size,
config.num_experts,
bias=False,
quant_config=None,
prefix=f"{prefix}.gate",
)
if config.shared_expert_intermediate_size > 0:
self.shared_expert = Qwen2MoeMLP(
hidden_size=config.hidden_size,
intermediate_size=config.shared_expert_intermediate_size,
hidden_act=config.hidden_act,
quant_config=quant_config,
reduce_results=self.experts.must_reduce_shared_expert_outputs(
),
reduce_results=self.experts.must_reduce_shared_expert_outputs(),
prefix=f"{prefix}.shared_expert",
)
else:
self.shared_expert = None
self.shared_expert_gate = torch.nn.Linear(config.hidden_size,
1,
bias=False)
self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
# NOTE: hidden_states can have either 1D or 2D shape.
@@ -152,24 +166,26 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
if self.shared_expert is not None:
shared_output = self.shared_expert(hidden_states)
if self.shared_expert_gate is not None:
shared_output = F.sigmoid(
self.shared_expert_gate(hidden_states)) * shared_output
shared_output = (
F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_output
)
# 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 shared_output is not None:
final_hidden_states = final_hidden_states + shared_output
if self.tp_size > 1:
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel( # noqa E501
final_hidden_states)
final_hidden_states
)
return final_hidden_states.view(orig_shape)
class Qwen2MoeAttention(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -207,19 +223,23 @@ class Qwen2MoeAttention(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=True,
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=True,
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,
@@ -240,7 +260,10 @@ class Qwen2MoeAttention(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 {},
)
def forward(
self,
@@ -256,7 +279,6 @@ class Qwen2MoeAttention(nn.Module):
class Qwen2MoeDecoderLayer(nn.Module):
def __init__(
self,
config: Qwen2MoeConfig,
@@ -268,11 +290,10 @@ class Qwen2MoeDecoderLayer(nn.Module):
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
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
)
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
self.self_attn = Qwen2MoeAttention(
hidden_size=self.hidden_size,
num_heads=config.num_attention_heads,
@@ -289,24 +310,27 @@ class Qwen2MoeDecoderLayer(nn.Module):
# Note: Qwen/Qwen2-57B-A14B-Instruct does not have
# `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 = Qwen2MoeSparseMoeBlock(config=config,
quant_config=quant_config,
prefix=f"{prefix}.mlp")
config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
):
self.mlp = Qwen2MoeSparseMoeBlock(
config=config, quant_config=quant_config, prefix=f"{prefix}.mlp"
)
else:
self.mlp = Qwen2MoeMLP(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 = Qwen2MoeMLP(
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,
@@ -319,23 +343,20 @@ class Qwen2MoeDecoderLayer(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 Qwen2MoeModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -352,16 +373,18 @@ class Qwen2MoeModel(nn.Module):
)
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: Qwen2MoeDecoderLayer(config=config,
cache_config=cache_config,
quant_config=quant_config,
prefix=prefix),
lambda prefix: Qwen2MoeDecoderLayer(
config=config,
cache_config=cache_config,
quant_config=quant_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)
@@ -386,10 +409,9 @@ class Qwen2MoeModel(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
@@ -400,10 +422,10 @@ class Qwen2MoeModel(nn.Module):
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.num_experts)
num_experts=self.config.num_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"),
@@ -417,7 +439,7 @@ class Qwen2MoeModel(nn.Module):
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
@@ -431,8 +453,9 @@ class Qwen2MoeModel(nn.Module):
continue
name = name.replace(weight_name, param_name)
# Skip loading extra bias for GPTQ models.
if ((name.endswith(".bias") or name.endswith("_bias"))
and name not in params_dict):
if (
name.endswith(".bias") or name.endswith("_bias")
) and name not in params_dict:
continue
# Skip layers on other devices.
if is_pp_missing_parameter(name, self):
@@ -455,21 +478,25 @@ class Qwen2MoeModel(nn.Module):
if is_pp_missing_parameter(name, self):
continue
# Skip loading extra bias for GPTQ models.
if ((name.endswith(".bias") or name.endswith("_bias"))
and name not in params_dict):
if (
name.endswith(".bias") or name.endswith("_bias")
) and name not in params_dict:
continue
param = params_dict[name]
weight_loader = param.weight_loader
weight_loader(param,
loaded_weight,
name,
shard_id=shard_id,
expert_id=expert_id)
weight_loader(
param,
loaded_weight,
name,
shard_id=shard_id,
expert_id=expert_id,
)
break
else:
# Skip loading extra bias for GPTQ models.
if ((name.endswith(".bias") or name.endswith("_bias"))
and name not in params_dict):
if (
name.endswith(".bias") or name.endswith("_bias")
) and name not in params_dict:
continue
# Skip layers on other devices.
if is_pp_missing_parameter(name, self):
@@ -477,7 +504,8 @@ class Qwen2MoeModel(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
@@ -488,15 +516,15 @@ class Qwen2MoeModel(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 Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
fall_back_to_pt_during_load = False
packed_modules_mapping = {
"qkv_proj": [
@@ -516,17 +544,21 @@ class Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.model = Qwen2MoeModel(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 = Qwen2MoeModel(
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
)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)
@@ -538,8 +570,9 @@ class Qwen2MoeForCausalLM(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(
@@ -549,8 +582,7 @@ class Qwen2MoeForCausalLM(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)