[MoE Refactor] Remove SharedFusedMoE class (#35782)
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
@@ -42,7 +42,7 @@ from vllm.distributed import (
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -163,7 +163,7 @@ class AXK1MoE(nn.Module):
|
||||
prefix=f"{prefix}.shared_experts",
|
||||
)
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
gate=self.gate,
|
||||
num_experts=config.n_routed_experts,
|
||||
@@ -916,7 +916,7 @@ class AXK1ForCausalLM(
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
@@ -950,7 +950,7 @@ class AXK1ForCausalLM(
|
||||
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -18,7 +18,7 @@ from vllm.distributed import (
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -124,8 +124,8 @@ class AfmoeMoE(nn.Module):
|
||||
prefix=f"{prefix}.shared_experts",
|
||||
)
|
||||
|
||||
# Routed experts using SharedFusedMoE
|
||||
self.experts = SharedFusedMoE(
|
||||
# Routed experts using FusedMoE
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=config.num_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
@@ -479,7 +479,7 @@ class AfmoeModel(nn.Module, EagleModelMixin):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
@@ -637,7 +637,7 @@ class AfmoeForCausalLM(nn.Module, SupportsPP, SupportsEagle3, SupportsLoRA):
|
||||
self.num_moe_layers = config.num_hidden_layers - config.num_dense_layers
|
||||
self.num_expert_groups = config.n_group
|
||||
|
||||
self.moe_layers: list[SharedFusedMoE] = []
|
||||
self.moe_layers: list[FusedMoE] = []
|
||||
example_moe = None
|
||||
for layer in self.model.layers:
|
||||
if isinstance(layer, PPMissingLayer):
|
||||
|
||||
@@ -14,7 +14,7 @@ from vllm.config.multimodal import BaseDummyOptions
|
||||
from vllm.distributed import get_tensor_model_parallel_rank
|
||||
from vllm.inputs import MultiModalDataDict
|
||||
from vllm.model_executor.layers.activation import get_act_fn
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.linear import ColumnParallelLinear, RowParallelLinear
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
@@ -214,7 +214,7 @@ class AriaProjector(nn.Module):
|
||||
return out
|
||||
|
||||
|
||||
class AriaFusedMoE(SharedFusedMoE):
|
||||
class AriaFusedMoE(FusedMoE):
|
||||
def weight_loader(
|
||||
self, param: nn.Parameter, loaded_weight: torch.Tensor, shard_id: str
|
||||
) -> None:
|
||||
|
||||
@@ -41,7 +41,7 @@ from vllm.distributed import (
|
||||
)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
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,
|
||||
@@ -285,7 +285,7 @@ class BailingMoE(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=self.num_experts,
|
||||
top_k=self.top_k,
|
||||
@@ -461,7 +461,7 @@ class BailingMoeModel(nn.Module):
|
||||
return hidden_states
|
||||
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -21,7 +21,7 @@ from vllm.model_executor.layers.fla.ops.layernorm_guard import (
|
||||
RMSNormGated,
|
||||
layernorm_fn,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE, SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -351,8 +351,8 @@ class BailingMoeV25(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
# Routed experts using SharedFusedMoE
|
||||
self.experts = SharedFusedMoE(
|
||||
# Routed experts using FusedMoE
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=self.num_experts,
|
||||
top_k=self.top_k,
|
||||
|
||||
@@ -11,7 +11,7 @@ from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.compilation.decorators import support_torch_compile
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
@@ -252,7 +252,7 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
|
||||
]
|
||||
stacked_params_mapping.extend(indexer_fused_mapping)
|
||||
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -48,9 +48,9 @@ from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoE,
|
||||
GateLinear,
|
||||
RoutingMethodType,
|
||||
SharedFusedMoE,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
@@ -311,7 +311,7 @@ class DeepseekV2MoE(nn.Module):
|
||||
prefix=f"{prefix}.shared_experts",
|
||||
)
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
gate=self.gate,
|
||||
num_experts=config.n_routed_experts,
|
||||
@@ -1432,7 +1432,7 @@ class DeepseekV2ForCausalLM(
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
@@ -1474,7 +1474,7 @@ class DeepseekV2ForCausalLM(
|
||||
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -40,7 +40,7 @@ from vllm.distributed import (
|
||||
)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
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,
|
||||
@@ -155,7 +155,7 @@ class Dots1MoE(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=config.n_routed_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
@@ -413,7 +413,7 @@ class Dots1Model(nn.Module):
|
||||
return hidden_states
|
||||
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -42,7 +42,7 @@ from vllm.distributed import (
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
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,
|
||||
@@ -188,7 +188,7 @@ class Ernie4_5_MoeMoE(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=config.moe_num_experts,
|
||||
top_k=config.moe_k,
|
||||
@@ -485,7 +485,7 @@ class Ernie4_5_MoeModel(nn.Module):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
@@ -667,7 +667,7 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, MixtureOfExpe
|
||||
self.num_moe_layers = len(moe_layers_indices)
|
||||
self.num_expert_groups = 1
|
||||
|
||||
self.moe_layers: list[SharedFusedMoE] = []
|
||||
self.moe_layers: list[FusedMoE] = []
|
||||
example_moe = None
|
||||
for layer in self.model.layers:
|
||||
if isinstance(layer, PPMissingLayer):
|
||||
|
||||
@@ -36,7 +36,7 @@ from vllm.config import CacheConfig, VllmConfig
|
||||
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
QKVParallelLinear,
|
||||
@@ -257,7 +257,7 @@ class Ernie4_5_VLMoeMoE(nn.Module):
|
||||
prefix=f"{prefix}.text_experts_gate",
|
||||
)
|
||||
|
||||
self.text_experts = SharedFusedMoE(
|
||||
self.text_experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=config.moe_num_experts[0],
|
||||
top_k=config.moe_k,
|
||||
@@ -294,7 +294,7 @@ class Ernie4_5_VLMoeMoE(nn.Module):
|
||||
prefix=f"{prefix}.vision_experts_gate",
|
||||
)
|
||||
|
||||
self.vision_experts = SharedFusedMoE(
|
||||
self.vision_experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=config.moe_num_experts[1],
|
||||
top_k=config.moe_k,
|
||||
@@ -649,7 +649,7 @@ class Ernie4_5_VLMoeForCausalLM(nn.Module, SupportsPP):
|
||||
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -31,7 +31,6 @@ from vllm.distributed import (
|
||||
get_tensor_model_parallel_world_size,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import ReplicatedLinear
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
@@ -130,7 +129,7 @@ class ExaoneMoe(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
gate=self.gate,
|
||||
num_experts=self.n_routed_experts,
|
||||
|
||||
@@ -42,7 +42,7 @@ from vllm.distributed import (
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
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,
|
||||
@@ -178,7 +178,7 @@ class Glm4MoE(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=config.n_routed_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
@@ -466,7 +466,7 @@ class Glm4MoeModel(nn.Module):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -41,7 +41,7 @@ from vllm.distributed import (
|
||||
get_pp_group,
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
@@ -308,7 +308,7 @@ class Glm4MoeLiteModel(nn.Module):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
@@ -334,7 +334,7 @@ class Glm4MoeLiteModel(nn.Module):
|
||||
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
@@ -616,7 +616,7 @@ class Glm4MoeLiteForCausalLM(
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -32,7 +32,7 @@ from transformers import PretrainedConfig
|
||||
|
||||
from vllm._aiter_ops import rocm_aiter_ops
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE, SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
@@ -260,7 +260,7 @@ class Glm4MoeLiteMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
|
||||
("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
|
||||
]
|
||||
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -42,7 +42,7 @@ from vllm.distributed import (
|
||||
)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -438,7 +438,7 @@ class HunYuanSparseMoeBlock(nn.Module):
|
||||
else:
|
||||
self.shared_mlp = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_mlp,
|
||||
num_experts=self.n_routed_experts,
|
||||
top_k=top_k,
|
||||
@@ -712,7 +712,7 @@ class HunYuanModel(nn.Module, EagleModelMixin):
|
||||
if _is_moe(self.config):
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -14,7 +14,7 @@ from vllm.distributed import (
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.kda import KimiDeltaAttention
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
@@ -144,7 +144,7 @@ class KimiMoE(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=num_experts,
|
||||
top_k=config.num_experts_per_token,
|
||||
@@ -476,7 +476,7 @@ class KimiLinearModel(nn.Module):
|
||||
if self.config.is_moe:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="w1",
|
||||
ckpt_down_proj_name="w2",
|
||||
|
||||
@@ -36,7 +36,7 @@ from vllm.model_executor.layers.attention import (
|
||||
Attention,
|
||||
ChunkedLocalAttention,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
QKVParallelLinear,
|
||||
@@ -127,7 +127,7 @@ class Llama4MoE(nn.Module):
|
||||
self.n_physical_experts = self.n_local_experts + self.n_redundant_experts
|
||||
self.n_local_physical_experts = self.n_physical_experts // self.ep_size
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_expert,
|
||||
num_experts=config.num_local_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
@@ -414,7 +414,7 @@ class Llama4Model(LlamaModel):
|
||||
params_dict: The dictionary of module parameters.
|
||||
loaded_params: The set of already loaded parameters.
|
||||
expert_params_mapping: The mapping of expert parameters. Must be
|
||||
generated by SharedFusedMoE.make_expert_params_mapping().
|
||||
generated by FusedMoE.make_expert_params_mapping().
|
||||
fused: Whether the expert weights are fused into a single weight
|
||||
tensor or are separate weight tensors for each expert.
|
||||
When fused is True, loaded_weight should have shape of:
|
||||
@@ -554,7 +554,7 @@ class Llama4Model(LlamaModel):
|
||||
fused_experts_params = False
|
||||
# Expert parameter mapping for the case where the expert weights are
|
||||
# not fused into a single weight tensor.
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
@@ -564,7 +564,7 @@ class Llama4Model(LlamaModel):
|
||||
)
|
||||
# Expert parameter mapping for the case where the expert weights are
|
||||
# fused into a single weight tensor.
|
||||
expert_params_mapping_fused = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping_fused = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_up_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -34,8 +34,8 @@ from vllm.distributed.parallel_state import get_pp_group
|
||||
from vllm.model_executor.layers.activation import ReLUSquaredActivation
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoE,
|
||||
GateLinear,
|
||||
SharedFusedMoE,
|
||||
activation_without_mul,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
@@ -210,7 +210,7 @@ class NemotronHMoE(nn.Module):
|
||||
self.fc1_latent_proj = None
|
||||
self.fc2_latent_proj = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=config.n_routed_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
@@ -652,7 +652,7 @@ class NemotronHModel(nn.Module):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
if self.has_moe:
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_params_mapping = FusedMoE.make_expert_params_mapping(
|
||||
# - FusedMoe.w1 (aka gate_proj) should be up_proj since that's
|
||||
# what the activation is applied to
|
||||
# - FusedMoe.w3 (aka up_proj) should be ignored since we're
|
||||
|
||||
@@ -44,7 +44,7 @@ from vllm.model_executor.layers.attention import (
|
||||
Attention,
|
||||
StaticSinkAttention,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -200,7 +200,7 @@ class OpenPanguMoE(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=config.n_routed_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
@@ -1149,7 +1149,7 @@ class OpenPanguModel(nn.Module):
|
||||
]
|
||||
has_experts = hasattr(self.config, "n_routed_experts")
|
||||
if has_experts:
|
||||
expert_merge_mapping = SharedFusedMoE.make_expert_params_mapping(
|
||||
expert_merge_mapping = FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -32,7 +32,7 @@ from vllm.distributed import (
|
||||
)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
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,
|
||||
@@ -353,7 +353,7 @@ class Param2MoEMoEBlock(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None # type: ignore[assignment]
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=self.num_experts,
|
||||
top_k=self.top_k,
|
||||
@@ -370,7 +370,7 @@ class Param2MoEMoEBlock(nn.Module):
|
||||
routed_scaling_factor=self.routed_scaling_factor,
|
||||
)
|
||||
|
||||
def maybe_get_fused_moe(self) -> SharedFusedMoE:
|
||||
def maybe_get_fused_moe(self) -> FusedMoE:
|
||||
return self.experts
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
@@ -690,7 +690,7 @@ class Param2MoEModel(nn.Module):
|
||||
return loaded_params
|
||||
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -40,7 +40,7 @@ from vllm.distributed import 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.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
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,
|
||||
@@ -164,7 +164,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
|
||||
else:
|
||||
self.shared_expert = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_expert,
|
||||
num_experts=config.num_experts,
|
||||
top_k=config.num_experts_per_tok,
|
||||
@@ -418,7 +418,7 @@ class Qwen2MoeModel(nn.Module):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -43,7 +43,7 @@ from vllm.distributed import (
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
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,
|
||||
@@ -205,7 +205,7 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
|
||||
self.shared_expert_gate = None
|
||||
self.shared_expert = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_expert,
|
||||
gate=self.gate,
|
||||
num_experts=self.n_routed_experts,
|
||||
@@ -508,7 +508,7 @@ class Qwen3MoeModel(nn.Module, EagleModelMixin):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -23,7 +23,7 @@ from vllm.distributed import (
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import (
|
||||
GemmaRMSNorm as Qwen3NextRMSNorm,
|
||||
)
|
||||
@@ -146,7 +146,7 @@ class Qwen3NextSparseMoeBlock(nn.Module):
|
||||
else:
|
||||
self.shared_expert = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_expert,
|
||||
gate=self.gate,
|
||||
num_experts=self.n_routed_experts,
|
||||
@@ -533,7 +533,7 @@ class Qwen3NextModel(nn.Module, EagleModelMixin):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -35,7 +35,7 @@ from vllm.distributed import (
|
||||
get_tensor_model_parallel_world_size,
|
||||
)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul
|
||||
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -335,7 +335,7 @@ class SarvamMLAMoE(nn.Module):
|
||||
else:
|
||||
self.shared_experts = None
|
||||
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.shared_experts,
|
||||
num_experts=self.num_experts,
|
||||
top_k=self.top_k,
|
||||
@@ -352,7 +352,7 @@ class SarvamMLAMoE(nn.Module):
|
||||
routed_scaling_factor=self.routed_scaling_factor,
|
||||
)
|
||||
|
||||
def maybe_get_fused_moe(self) -> SharedFusedMoE:
|
||||
def maybe_get_fused_moe(self) -> FusedMoE:
|
||||
return self.experts
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
@@ -529,7 +529,7 @@ class SarvamMLAModel(nn.Module):
|
||||
return hidden_states
|
||||
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
return SharedFusedMoE.make_expert_params_mapping(
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="gate_proj",
|
||||
ckpt_down_proj_name="down_proj",
|
||||
|
||||
@@ -24,7 +24,6 @@ from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import SiluAndMul, SwigluStepAndMul
|
||||
from vllm.model_executor.layers.attention import Attention
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
|
||||
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
ColumnParallelLinear,
|
||||
@@ -372,7 +371,7 @@ class FusedMoEBlock(nn.Module):
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.share_expert",
|
||||
)
|
||||
self.experts = SharedFusedMoE(
|
||||
self.experts = FusedMoE(
|
||||
shared_experts=self.share_expert,
|
||||
gate=self.gate,
|
||||
num_experts=config.moe_num_experts,
|
||||
|
||||
Reference in New Issue
Block a user