[MoE Refactor] Rename FusedMoE.make_expert_params_mapping to...

[MoE Refactor] Rename FusedMoE.make_expert_params_mapping to fused_moe_make_expert_params_mapping (#40671)
Signed-off-by: Bill Nell <bnell@redhat.com>

vllm/model_executor/models/glm4_moe_lite_mtp.py

@@ -32,7 +32,10 @@ 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
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 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 +263,7 @@ class Glm4MoeLiteMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
             ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
         ]
 
-        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+        expert_params_mapping = fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",

vllm/model_executor/models/glm4_moe_mtp.py

@@ -31,7 +31,10 @@ import torch.nn as nn
 from transformers import PretrainedConfig
 
 from vllm.config import CacheConfig, ParallelConfig, VllmConfig
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 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
@@ -247,7 +250,7 @@ class Glm4MoeMTP(nn.Module, Glm4MixtureOfExperts):
 
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
-        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+        expert_params_mapping = fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",

vllm/model_executor/models/gpt_oss.py

@@ -20,7 +20,10 @@ from vllm.distributed import (
     tensor_model_parallel_all_gather,
 )
 from vllm.model_executor.layers.attention import Attention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.fused_moe.config import FusedMoEParallelConfig
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
@@ -331,7 +334,7 @@ class GptOssModel(nn.Module, EagleModelMixin):
         # Params for weights, weight scales, activation scales
         # (param_name, weight_name, expert_id, shard_id)
         # NOTE: this is only used for quark.
-        return FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="w1",
             ckpt_down_proj_name="w2",

vllm/model_executor/models/granitemoe.py

@@ -39,7 +39,10 @@ from vllm.distributed import (
     tensor_model_parallel_all_gather,
 )
 from vllm.model_executor.layers.attention import Attention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     QKVParallelLinear,
@@ -351,7 +354,7 @@ class GraniteMoeModel(nn.Module):
 
         # Params for weights, fp8 weight scales, fp8 activation scales
         # (param_name, weight_name, expert_id, shard_id)
-        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+        expert_params_mapping = fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="w1",
             ckpt_down_proj_name="w2",

vllm/model_executor/models/grok1.py

@@ -38,7 +38,10 @@ 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 GeluAndMul
 from vllm.model_executor.layers.attention import Attention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
@@ -519,7 +522,7 @@ class Grok1Model(nn.Module):
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
         # Map expert parameter names to standard names
         num_experts = _get_num_experts(self.config)
-        return FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name=self.ckpt_gate_proj_name,
             ckpt_down_proj_name=self.ckpt_down_proj_name,

vllm/model_executor/models/hunyuan_v1.py

@@ -42,7 +42,10 @@ 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 FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     ColumnParallelLinear,
@@ -712,7 +715,7 @@ 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 FusedMoE.make_expert_params_mapping(
+            return fused_moe_make_expert_params_mapping(
                 self,
                 ckpt_gate_proj_name="gate_proj",
                 ckpt_down_proj_name="down_proj",

vllm/model_executor/models/interns1_pro.py

@@ -41,7 +41,9 @@ 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 FusedMoE
+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,

vllm/model_executor/models/jamba.py

@@ -14,7 +14,10 @@ from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.distributed.parallel_state import get_pp_group
 from vllm.model_executor.layers.attention import Attention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     QKVParallelLinear,
@@ -378,7 +381,7 @@ class JambaModel(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 FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",

vllm/model_executor/models/kimi_linear.py

@@ -14,7 +14,10 @@ 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 FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.kda import KimiDeltaAttention
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
@@ -476,7 +479,7 @@ 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 = FusedMoE.make_expert_params_mapping(
+            expert_params_mapping = fused_moe_make_expert_params_mapping(
                 self,
                 ckpt_gate_proj_name="w1",
                 ckpt_down_proj_name="w2",

vllm/model_executor/models/lfm2_moe.py

@@ -15,7 +15,10 @@ 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 FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
@@ -482,7 +485,7 @@ class Lfm2MoeModel(nn.Module):
         return hidden_states
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
-        return FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="w1",
             ckpt_down_proj_name="w2",

vllm/model_executor/models/llama4.py

@@ -36,7 +36,10 @@ from vllm.model_executor.layers.attention import (
     Attention,
     ChunkedLocalAttention,
 )
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     QKVParallelLinear,
@@ -414,7 +417,7 @@ 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 FusedMoE.make_expert_params_mapping().
+                generated by fused_moe_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 +557,7 @@ 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 = FusedMoE.make_expert_params_mapping(
+        expert_params_mapping = fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
@@ -564,7 +567,7 @@ class Llama4Model(LlamaModel):
         )
         # Expert parameter mapping for the case where the expert weights are
         # fused into a single weight tensor.
-        expert_params_mapping_fused = FusedMoE.make_expert_params_mapping(
+        expert_params_mapping_fused = fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_up_proj",
             ckpt_down_proj_name="down_proj",

vllm/model_executor/models/longcat_flash.py

@@ -46,7 +46,10 @@ from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group
 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.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
@@ -622,7 +625,7 @@ class LongcatFlashForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
     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 FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",

vllm/model_executor/models/mimo_v2_flash.py

@@ -22,7 +22,10 @@ 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 FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
@@ -511,7 +514,7 @@ class MiMoV2Model(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 FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",

vllm/model_executor/models/minimax_m2.py

@@ -38,7 +38,10 @@ from vllm.distributed import (
     get_tensor_model_parallel_world_size,
 )
 from vllm.model_executor.layers.attention import Attention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     QKVParallelLinear,
@@ -393,7 +396,7 @@ class MiniMaxM2Model(nn.Module, EagleModelMixin):
         return hidden_states
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
-        return FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="w1",
             ckpt_down_proj_name="w2",

vllm/model_executor/models/minimax_text_01.py

@@ -24,7 +24,9 @@ from vllm.distributed.parallel_state import (
 from vllm.forward_context import get_forward_context
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.attention import Attention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+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,

vllm/model_executor/models/mixtral.py

@@ -40,7 +40,10 @@ from vllm.distributed import (
     get_tensor_model_parallel_world_size,
 )
 from vllm.model_executor.layers.attention import Attention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     QKVParallelLinear,
@@ -364,7 +367,7 @@ class MixtralModel(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 FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="w1",
             ckpt_down_proj_name="w2",

vllm/model_executor/models/mllama4.py

@@ -40,7 +40,9 @@ from vllm.config.multimodal import BaseDummyOptions
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.inputs import MultiModalDataDict
 from vllm.model_executor.layers.attention import MMEncoderAttention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.linear import (
     ColumnParallelLinear,
     QKVParallelLinear,
@@ -1072,7 +1074,7 @@ class Llama4ForConditionalGeneration(
     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 FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",

vllm/model_executor/models/nemotron_h.py

@@ -37,6 +37,7 @@ from vllm.model_executor.layers.fused_moe import (
     FusedMoE,
     GateLinear,
     activation_without_mul,
+    fused_moe_make_expert_params_mapping,
 )
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
@@ -652,7 +653,7 @@ 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 = FusedMoE.make_expert_params_mapping(
+            expert_params_mapping = fused_moe_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

vllm/model_executor/models/nemotron_h_mtp.py

@@ -11,7 +11,9 @@ import torch.nn as nn
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.config.parallel import ParallelConfig
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import ColumnParallelLinear
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
@@ -399,7 +401,7 @@ class NemotronHMTP(nn.Module, SupportsPP):
         if getattr(self.config, "model_type", None) == "nemotron_h_puzzle":
             num_experts = self.config.mtp_n_routed_experts
         if num_experts is not None:
-            expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            expert_params_mapping = fused_moe_make_expert_params_mapping(
                 self,
                 ckpt_gate_proj_name="up_proj",
                 ckpt_down_proj_name="down_proj",

vllm/model_executor/models/olmoe.py

@@ -32,7 +32,10 @@ from vllm.distributed import (
 from vllm.distributed.utils import split_tensor_along_last_dim
 from vllm.logger import init_logger
 from vllm.model_executor.layers.attention import Attention
-from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.fused_moe import (
+    FusedMoE,
+    fused_moe_make_expert_params_mapping,
+)
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     QKVParallelLinear,
@@ -336,7 +339,7 @@ class OlmoeModel(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 FusedMoE.make_expert_params_mapping(
+        return fused_moe_make_expert_params_mapping(
             self,
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",