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

[BugFix] Support EP/DP + EPLB with MTP (#25311)

Browse Source 
Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Sage Moore <sage@neuralmagic.com>
Co-authored-by: Sage Moore <sage@neuralmagic.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
This commit is contained in:
Ilya Markov
2025-11-05 16:22:17 +01:00
committed by GitHub
parent 5d16d0fa62
commit e50c454672
27 changed files with 957 additions and 529 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
vllm/model_executor/models/glm4_moe.py
View File

@@ -62,7 +62,7 @@ from vllm.model_executor.model_loader.weight_utils import (
)
from vllm.sequence import IntermediateTensors
from .interfaces import SupportsLoRA, SupportsPP
from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
@@ -127,7 +127,7 @@ class Glm4MoE(nn.Module):
self.routed_scaling_factor = config.routed_scaling_factor
self.ep_group = get_ep_group().device_group
self.ep_rank = self.ep_group.rank()
self.ep_rank = get_ep_group().rank_in_group
self.ep_size = self.ep_group.size()
self.n_routed_experts: int = config.n_routed_experts
self.n_shared_experts: int = config.n_shared_experts
@@ -616,7 +616,35 @@ class Glm4MoeModel(nn.Module):
return loaded_params
class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
class Glm4MixtureOfExperts(MixtureOfExperts):
def extract_moe_parameters(self, example_moe: Glm4MoE | None) -> None:
if example_moe is None:
raise RuntimeError("No Glm4MoE layer found in model.layers.")
else:
self.num_logical_experts = example_moe.n_logical_experts
self.num_physical_experts = example_moe.n_physical_experts
self.num_local_physical_experts = example_moe.n_local_physical_experts
self.num_routed_experts = example_moe.n_routed_experts
self.num_shared_experts = example_moe.n_shared_experts
self.num_redundant_experts = example_moe.n_redundant_experts
def update_physical_experts_metadata(
self,
num_physical_experts: int,
num_local_physical_experts: int,
) -> None:
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
for moe in self.moe_mlp_layers:
moe.n_local_physical_experts = num_local_physical_experts
moe.n_physical_experts = num_physical_experts
moe.n_redundant_experts = self.num_redundant_experts
moe.experts.update_expert_map()
class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, Glm4MixtureOfExperts):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
@@ -659,7 +687,9 @@ class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace
self.num_expert_groups = config.n_group
self.moe_layers: list[SharedFusedMoE] = []
self.moe_layers = []
self.moe_mlp_layers: list[Glm4MoE] = []
example_moe = None
for layer in self.model.layers:
if isinstance(layer, PPMissingLayer):
@@ -669,33 +699,10 @@ class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
if isinstance(layer.mlp, Glm4MoE):
# Pick last one layer since the first ones may be dense layers.
example_moe = layer.mlp
self.moe_mlp_layers.append(layer.mlp)
self.moe_layers.append(layer.mlp.experts)
if example_moe is None:
raise RuntimeError("No Glm4MoE layer found in model.layers.")
self.num_logical_experts = example_moe.n_logical_experts
self.num_physical_experts = example_moe.n_physical_experts
self.num_local_physical_experts = example_moe.n_local_physical_experts
self.num_routed_experts = example_moe.n_routed_experts
self.num_shared_experts = example_moe.n_shared_experts
self.num_redundant_experts = example_moe.n_redundant_experts
def set_eplb_state(
self,
expert_load_view: torch.Tensor,
logical_to_physical_map: torch.Tensor,
logical_replica_count: torch.Tensor,
) -> None:
for layer_idx, layer in enumerate(self.moe_layers):
# Register the expert weights.
self.expert_weights.append(layer.get_expert_weights())
layer.set_eplb_state(
moe_layer_idx=layer_idx,
expert_load_view=expert_load_view,
logical_to_physical_map=logical_to_physical_map,
logical_replica_count=logical_replica_count,
)
self.extract_moe_parameters(example_moe)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.get_input_embeddings(input_ids)