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

311
vllm/model_executor/models/glm4_moe.py
View File

@@ -22,6 +22,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only GLM-4.5, GLM-4.6 model compatible with HuggingFace weights."""
import typing
from collections.abc import Callable, Iterable
from itertools import islice
@@ -34,35 +35,48 @@ from transformers.models.glm4_moe import Glm4MoeConfig
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)
from vllm.distributed import (
get_ep_group,
get_pp_group,
get_tensor_model_parallel_world_size,
)
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,
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.shared_fused_moe import SharedFusedMoE
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.sequence import IntermediateTensors
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
make_empty_intermediate_tensors_factory, make_layers,
maybe_prefix)
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
is_pp_missing_parameter,
make_empty_intermediate_tensors_factory,
make_layers,
maybe_prefix,
)
logger = init_logger(__name__)
class Glm4MoeMLP(nn.Module):
def __init__(
self,
hidden_size: int,
@@ -74,19 +88,24 @@ class Glm4MoeMLP(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):
@@ -97,7 +116,6 @@ class Glm4MoeMLP(nn.Module):
class Glm4MoE(nn.Module):
def __init__(
self,
config: Glm4MoeConfig,
@@ -116,8 +134,10 @@ class Glm4MoE(nn.Module):
self.n_shared_experts: int = config.n_shared_experts
if config.hidden_act != "silu":
raise ValueError(f"Unsupported activation: {config.hidden_act}. "
"Only silu is supported for now.")
raise ValueError(
f"Unsupported activation: {config.hidden_act}. "
"Only silu is supported for now."
)
# NOTE In the transformers implementation, the gate isn't an nn.Linear,
# so we cannot use ReplicatedLinear here.
# See: https://github.com/huggingface/transformers/blob/v4.55.1/src/transformers/models/glm4_moe/modeling_glm4_moe.py#L260
@@ -128,7 +148,8 @@ class Glm4MoE(nn.Module):
dtype=torch.float32,
)
self.gate.e_score_correction_bias = nn.Parameter(
torch.empty(config.n_routed_experts, dtype=torch.float32))
torch.empty(config.n_routed_experts, dtype=torch.float32)
)
# Load balancing settings.
vllm_config = get_current_vllm_config()
@@ -137,18 +158,16 @@ class Glm4MoE(nn.Module):
self.n_redundant_experts = eplb_config.num_redundant_experts
self.n_logical_experts = self.n_routed_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
)
if config.n_shared_experts is not None:
intermediate_size = (config.moe_intermediate_size *
config.n_shared_experts)
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
self.shared_experts = Glm4MoeMLP(
hidden_size=config.hidden_size,
intermediate_size=intermediate_size,
@@ -195,7 +214,8 @@ class Glm4MoE(nn.Module):
routed_scaling_factor=1.0,
e_score_correction_bias=self.gate.e_score_correction_bias,
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts)
num_redundant_experts=self.n_redundant_experts,
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
num_tokens, hidden_dim = hidden_states.shape
@@ -204,27 +224,27 @@ class Glm4MoE(nn.Module):
# router_logits: (num_tokens, n_experts)
router_logits = self.gate(hidden_states.to(dtype=torch.float32))
fused_moe_out = self.experts(hidden_states=hidden_states,
router_logits=router_logits)
fused_moe_out = self.experts(
hidden_states=hidden_states, router_logits=router_logits
)
if self.shared_experts is not None:
shared_output, final_hidden_states = fused_moe_out
assert shared_output is not None
final_hidden_states = \
final_hidden_states * self.routed_scaling_factor\
+ shared_output
final_hidden_states = (
final_hidden_states * self.routed_scaling_factor + shared_output
)
else:
final_hidden_states = fused_moe_out * self.routed_scaling_factor
if self.tp_size > 1:
final_hidden_states = (
self.experts.maybe_all_reduce_tensor_model_parallel(
final_hidden_states))
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
final_hidden_states
)
return final_hidden_states.view(num_tokens, hidden_dim)
class Glm4MoeAttention(nn.Module):
def __init__(
self,
config: Glm4MoeConfig,
@@ -266,19 +286,23 @@ class Glm4MoeAttention(nn.Module):
self.max_position_embeddings = max_position_embeddings
self.use_qk_norm = use_qk_norm
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",
)
partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
self.rotary_emb = get_rope(
@@ -311,10 +335,12 @@ class Glm4MoeAttention(nn.Module):
qkv, _ = self.qkv_proj(hidden_states)
q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
if self.use_qk_norm:
q = self.q_norm(q.reshape(-1, self.num_heads,
self.head_dim)).reshape(q.shape)
k = self.k_norm(k.reshape(-1, self.num_kv_heads,
self.head_dim)).reshape(k.shape)
q = self.q_norm(q.reshape(-1, self.num_heads, self.head_dim)).reshape(
q.shape
)
k = self.k_norm(k.reshape(-1, self.num_kv_heads, self.head_dim)).reshape(
k.shape
)
q, k = self.rotary_emb(positions, q, k)
attn_output = self.attn(q, k, v)
@@ -323,7 +349,6 @@ class Glm4MoeAttention(nn.Module):
class Glm4MoeDecoderLayer(nn.Module):
def __init__(
self,
config: Glm4MoeConfig,
@@ -336,11 +361,10 @@ class Glm4MoeDecoderLayer(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",
131072)
max_position_embeddings = getattr(config, "max_position_embeddings", 131072)
# DecoderLayers are created with `make_layers` which passes the prefix
# with the layer's index.
layer_idx = int(prefix.split(sep='.')[-1])
layer_idx = int(prefix.split(sep=".")[-1])
self.layer_idx = layer_idx
self.self_attn = Glm4MoeAttention(
@@ -360,8 +384,10 @@ class Glm4MoeDecoderLayer(nn.Module):
use_qk_norm=config.use_qk_norm,
)
if (config.n_routed_experts is not None
and layer_idx >= config.first_k_dense_replace):
if (
config.n_routed_experts is not None
and layer_idx >= config.first_k_dense_replace
):
self.mlp = Glm4MoE(
config=config,
quant_config=quant_config,
@@ -369,16 +395,18 @@ class Glm4MoeDecoderLayer(nn.Module):
enable_eplb=enable_eplb,
)
else:
self.mlp = Glm4MoeMLP(hidden_size=config.hidden_size,
intermediate_size=config.intermediate_size,
hidden_act=config.hidden_act,
quant_config=quant_config,
prefix=f"{prefix}.mlp")
self.mlp = Glm4MoeMLP(
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.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.routed_scaling_factor = config.routed_scaling_factor
def forward(
@@ -391,12 +419,9 @@ class Glm4MoeDecoderLayer(nn.Module):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
else:
hidden_states, residual = self.input_layernorm(
hidden_states, residual)
hidden_states = self.self_attn(positions=positions,
hidden_states=hidden_states)
hidden_states, residual = self.post_attention_layernorm(
hidden_states, residual)
hidden_states, residual = self.input_layernorm(hidden_states, residual)
hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states)
hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
hidden_states = self.mlp(hidden_states)
return hidden_states, residual
@@ -407,9 +432,9 @@ class Glm4MoeDecoderLayer(nn.Module):
"positions": -1,
"intermediate_tensors": 0,
"inputs_embeds": 0,
})
}
)
class Glm4MoeModel(nn.Module):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -423,9 +448,8 @@ class Glm4MoeModel(nn.Module):
if get_pp_group().is_first_rank:
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
prefix=f"{prefix}.embed_tokens")
config.vocab_size, config.hidden_size, prefix=f"{prefix}.embed_tokens"
)
else:
self.embed_tokens = PPMissingLayer()
@@ -438,15 +462,16 @@ class Glm4MoeModel(nn.Module):
prefix=prefix,
enable_eplb=enable_eplb,
),
prefix=f"{prefix}.layers")
prefix=f"{prefix}.layers",
)
if get_pp_group().is_last_rank:
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
else:
self.norm = PPMissingLayer()
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)
@@ -473,27 +498,26 @@ class Glm4MoeModel(nn.Module):
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
def make_empty_intermediate_tensors(
self, batch_size: int, dtype: torch.dtype,
device: torch.device) -> IntermediateTensors:
return IntermediateTensors({
"hidden_states":
torch.zeros((batch_size, self.config.hidden_size),
dtype=dtype,
device=device),
"residual":
torch.zeros((batch_size, self.config.hidden_size),
dtype=dtype,
device=device),
})
self, batch_size: int, dtype: torch.dtype, device: torch.device
) -> IntermediateTensors:
return IntermediateTensors(
{
"hidden_states": torch.zeros(
(batch_size, self.config.hidden_size), dtype=dtype, device=device
),
"residual": torch.zeros(
(batch_size, self.config.hidden_size), dtype=dtype, device=device
),
}
)
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
# Params for weights, fp8 weight scales, fp8 activation scales
@@ -502,10 +526,10 @@ class Glm4MoeModel(nn.Module):
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts)
num_experts=self.config.n_routed_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"),
@@ -522,7 +546,7 @@ class Glm4MoeModel(nn.Module):
spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
if spec_layer is not None:
continue
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
@@ -532,7 +556,7 @@ class Glm4MoeModel(nn.Module):
# name will be updated to mlp.experts[0].gate_up_proj, which
# will then be updated below in expert_params_mapping
# for mlp.experts[0].gate_gate_up_proj, which breaks load.
if (("mlp.experts." in name) and name not in params_dict):
if ("mlp.experts." in name) and name not in params_dict:
continue
name = name.replace(weight_name, param_name)
# Skip loading extra bias for GPTQ models.
@@ -567,14 +591,17 @@ class Glm4MoeModel(nn.Module):
# 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
@@ -598,8 +625,9 @@ class Glm4MoeModel(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
)
weight_loader(param, loaded_weight)
loaded_params.add(name)
@@ -627,24 +655,26 @@ class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.model = Glm4MoeModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model"))
self.model = Glm4MoeModel(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
if get_pp_group().is_last_rank:
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
)
self.expert_weights = []
# Set MoE hyperparameters
self.num_moe_layers = (config.num_hidden_layers -
config.first_k_dense_replace)
self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace
self.num_expert_groups = config.n_group
self.moe_layers: list[FusedMoE] = []
@@ -695,8 +725,9 @@ class Glm4MoeForCausalLM(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(
@@ -706,8 +737,7 @@ class Glm4MoeForCausalLM(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)
@@ -715,13 +745,14 @@ class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
return self.model.get_expert_mapping()
def get_spec_layer_idx_from_weight_name(config: Glm4MoeConfig,
weight_name: str) -> Optional[int]:
if hasattr(config,
"num_nextn_predict_layers") and (config.num_nextn_predict_layers
> 0):
def get_spec_layer_idx_from_weight_name(
config: Glm4MoeConfig, weight_name: str
) -> Optional[int]:
if hasattr(config, "num_nextn_predict_layers") and (
config.num_nextn_predict_layers > 0
):
layer_idx = config.num_hidden_layers
for i in range(config.num_nextn_predict_layers):
if f"layers.{layer_idx+i}." in weight_name:
if f"layers.{layer_idx + i}." in weight_name:
return layer_idx + i
return None