From 1c2c1eb8b9fdd2e67c45afb6123ccc07c0177555 Mon Sep 17 00:00:00 2001 From: bnellnm <49004751+bnellnm@users.noreply.github.com> Date: Thu, 23 Apr 2026 11:22:34 -0400 Subject: [PATCH] [MoE Refactor] Rename FusedMoE.make_expert_params_mapping to fused_moe_make_expert_params_mapping (#40671) Signed-off-by: Bill Nell --- .../layers/fused_moe/__init__.py | 2 ++ vllm/model_executor/layers/fused_moe/layer.py | 19 +++++++++++++++++++ vllm/model_executor/models/AXK1.py | 9 ++++++--- vllm/model_executor/models/afmoe.py | 7 +++++-- vllm/model_executor/models/arctic.py | 5 ++++- vllm/model_executor/models/aria.py | 4 +++- vllm/model_executor/models/bailing_moe.py | 7 +++++-- .../models/bailing_moe_linear.py | 7 +++++-- vllm/model_executor/models/dbrx.py | 4 +++- vllm/model_executor/models/deepseek_eagle.py | 6 ++++-- vllm/model_executor/models/deepseek_mtp.py | 6 ++++-- vllm/model_executor/models/deepseek_v2.py | 5 +++-- vllm/model_executor/models/dots1.py | 7 +++++-- vllm/model_executor/models/ernie45_moe.py | 7 +++++-- vllm/model_executor/models/ernie45_vl_moe.py | 7 +++++-- vllm/model_executor/models/exaone_moe.py | 7 +++++-- vllm/model_executor/models/flex_olmo.py | 4 +++- vllm/model_executor/models/gemma4.py | 5 ++++- vllm/model_executor/models/glm4_moe.py | 7 +++++-- vllm/model_executor/models/glm4_moe_lite.py | 10 ++++++---- .../models/glm4_moe_lite_mtp.py | 7 +++++-- vllm/model_executor/models/glm4_moe_mtp.py | 7 +++++-- vllm/model_executor/models/gpt_oss.py | 7 +++++-- vllm/model_executor/models/granitemoe.py | 7 +++++-- vllm/model_executor/models/grok1.py | 7 +++++-- vllm/model_executor/models/hunyuan_v1.py | 7 +++++-- vllm/model_executor/models/interns1_pro.py | 4 +++- vllm/model_executor/models/jamba.py | 7 +++++-- vllm/model_executor/models/kimi_linear.py | 7 +++++-- vllm/model_executor/models/lfm2_moe.py | 7 +++++-- vllm/model_executor/models/llama4.py | 11 +++++++---- vllm/model_executor/models/longcat_flash.py | 7 +++++-- vllm/model_executor/models/mimo_v2_flash.py | 7 +++++-- vllm/model_executor/models/minimax_m2.py | 7 +++++-- vllm/model_executor/models/minimax_text_01.py | 4 +++- vllm/model_executor/models/mixtral.py | 7 +++++-- vllm/model_executor/models/mllama4.py | 6 ++++-- vllm/model_executor/models/nemotron_h.py | 3 ++- vllm/model_executor/models/nemotron_h_mtp.py | 6 ++++-- vllm/model_executor/models/olmoe.py | 7 +++++-- vllm/model_executor/models/openpangu.py | 7 +++++-- vllm/model_executor/models/openpangu_mtp.py | 6 ++++-- vllm/model_executor/models/param2moe.py | 7 +++++-- vllm/model_executor/models/phimoe.py | 7 +++++-- vllm/model_executor/models/qwen2_moe.py | 7 +++++-- vllm/model_executor/models/qwen3_5_mtp.py | 6 ++++-- vllm/model_executor/models/qwen3_moe.py | 7 +++++-- vllm/model_executor/models/qwen3_next.py | 7 +++++-- vllm/model_executor/models/qwen3_next_mtp.py | 6 ++++-- vllm/model_executor/models/sarvam.py | 7 +++++-- vllm/model_executor/models/step3_text.py | 4 +++- vllm/model_executor/models/step3p5.py | 7 +++++-- .../model_executor/models/transformers/moe.py | 7 +++++-- 53 files changed, 254 insertions(+), 98 deletions(-) diff --git a/vllm/model_executor/layers/fused_moe/__init__.py b/vllm/model_executor/layers/fused_moe/__init__.py index 1b2ce61f7c..a154ede547 100644 --- a/vllm/model_executor/layers/fused_moe/__init__.py +++ b/vllm/model_executor/layers/fused_moe/__init__.py @@ -19,6 +19,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe_method_base import ( from vllm.model_executor.layers.fused_moe.layer import ( FusedMoE, FusedMoeWeightScaleSupported, + fused_moe_make_expert_params_mapping, ) from vllm.model_executor.layers.fused_moe.modular_kernel import ( FusedMoEActivationFormat, @@ -65,6 +66,7 @@ __all__ = [ "RoutingMethodType", "activation_without_mul", "apply_moe_activation", + "fused_moe_make_expert_params_mapping", "override_config", "get_config", ] diff --git a/vllm/model_executor/layers/fused_moe/layer.py b/vllm/model_executor/layers/fused_moe/layer.py index 7adac0374c..012c732850 100644 --- a/vllm/model_executor/layers/fused_moe/layer.py +++ b/vllm/model_executor/layers/fused_moe/layer.py @@ -1618,6 +1618,25 @@ class FusedMoE(PluggableLayer): return s +# This is a temporary forwarding method which will be removed/modified layer. +def fused_moe_make_expert_params_mapping( + model: torch.nn.Module, + ckpt_gate_proj_name: str, + ckpt_down_proj_name: str, + ckpt_up_proj_name: str, + num_experts: int, + num_redundant_experts: int = 0, +) -> list[tuple[str, str, int, str]]: + return FusedMoE.make_expert_params_mapping( + model, + ckpt_gate_proj_name, + ckpt_down_proj_name, + ckpt_up_proj_name, + num_experts, + num_redundant_experts, + ) + + # Mark the FusedMoE weight_loader as supporting MoE-specific parameters # to avoid expensive runtime reflection in model loading code FusedMoE.weight_loader.supports_moe_loading = True # type: ignore[attr-defined] diff --git a/vllm/model_executor/models/AXK1.py b/vllm/model_executor/models/AXK1.py index c33d5b9737..c8f56ca97f 100644 --- a/vllm/model_executor/models/AXK1.py +++ b/vllm/model_executor/models/AXK1.py @@ -42,7 +42,10 @@ 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 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, @@ -916,7 +919,7 @@ 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 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", @@ -950,7 +953,7 @@ class AXK1ForCausalLM( # 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", diff --git a/vllm/model_executor/models/afmoe.py b/vllm/model_executor/models/afmoe.py index e34a418c98..2216e4948b 100644 --- a/vllm/model_executor/models/afmoe.py +++ b/vllm/model_executor/models/afmoe.py @@ -18,7 +18,10 @@ 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 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, @@ -479,7 +482,7 @@ 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 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", diff --git a/vllm/model_executor/models/arctic.py b/vllm/model_executor/models/arctic.py index 0c9267994b..6ab55a4b1b 100644 --- a/vllm/model_executor/models/arctic.py +++ b/vllm/model_executor/models/arctic.py @@ -18,7 +18,10 @@ 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 fused_experts, fused_topk +from vllm.model_executor.layers.fused_moe import ( + fused_experts, + fused_topk, +) from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import ( MergedColumnParallelLinear, diff --git a/vllm/model_executor/models/aria.py b/vllm/model_executor/models/aria.py index 9696dec6d8..48c8d9a441 100644 --- a/vllm/model_executor/models/aria.py +++ b/vllm/model_executor/models/aria.py @@ -14,7 +14,9 @@ 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 FusedMoE +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 diff --git a/vllm/model_executor/models/bailing_moe.py b/vllm/model_executor/models/bailing_moe.py index ef4f66614a..56e119207d 100644 --- a/vllm/model_executor/models/bailing_moe.py +++ b/vllm/model_executor/models/bailing_moe.py @@ -41,7 +41,10 @@ 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, @@ -461,7 +464,7 @@ class BailingMoeModel(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="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/bailing_moe_linear.py b/vllm/model_executor/models/bailing_moe_linear.py index df36659b10..e26adc1743 100644 --- a/vllm/model_executor/models/bailing_moe_linear.py +++ b/vllm/model_executor/models/bailing_moe_linear.py @@ -21,7 +21,10 @@ from vllm.model_executor.layers.fla.ops.layernorm_guard import ( RMSNormGated, layernorm_fn, ) -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, @@ -990,7 +993,7 @@ class BailingMoeV25Model(nn.Module): def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: """Get expert parameter mapping for MoE layers.""" - 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", diff --git a/vllm/model_executor/models/dbrx.py b/vllm/model_executor/models/dbrx.py index a72f4e4871..6c798bf2f3 100644 --- a/vllm/model_executor/models/dbrx.py +++ b/vllm/model_executor/models/dbrx.py @@ -15,7 +15,9 @@ 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, +) from vllm.model_executor.layers.linear import ( QKVParallelLinear, ReplicatedLinear, diff --git a/vllm/model_executor/models/deepseek_eagle.py b/vllm/model_executor/models/deepseek_eagle.py index 5c439cdf48..f975b32adc 100644 --- a/vllm/model_executor/models/deepseek_eagle.py +++ b/vllm/model_executor/models/deepseek_eagle.py @@ -8,7 +8,9 @@ import torch.nn as nn from vllm.compilation.decorators import support_torch_compile from vllm.config import VllmConfig -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.logits_processor import LogitsProcessor from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -105,7 +107,7 @@ class DeepseekV2Model(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="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/deepseek_mtp.py b/vllm/model_executor/models/deepseek_mtp.py index 898e433340..37f94c687a 100644 --- a/vllm/model_executor/models/deepseek_mtp.py +++ b/vllm/model_executor/models/deepseek_mtp.py @@ -11,7 +11,9 @@ 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 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.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig @@ -252,7 +254,7 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts): ] stacked_params_mapping.extend(indexer_fused_mapping) - 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", diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py index 53bcf87c6c..3d0b1c4245 100644 --- a/vllm/model_executor/models/deepseek_v2.py +++ b/vllm/model_executor/models/deepseek_v2.py @@ -51,6 +51,7 @@ from vllm.model_executor.layers.fused_moe import ( FusedMoE, GateLinear, RoutingMethodType, + fused_moe_make_expert_params_mapping, ) from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm from vllm.model_executor.layers.linear import ( @@ -1432,7 +1433,7 @@ 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 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", @@ -1474,7 +1475,7 @@ class DeepseekV2ForCausalLM( # 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", diff --git a/vllm/model_executor/models/dots1.py b/vllm/model_executor/models/dots1.py index 181bd598e8..f58fc4da92 100644 --- a/vllm/model_executor/models/dots1.py +++ b/vllm/model_executor/models/dots1.py @@ -40,7 +40,10 @@ 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, @@ -413,7 +416,7 @@ class Dots1Model(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="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/ernie45_moe.py b/vllm/model_executor/models/ernie45_moe.py index 58dd61e9d9..a2b0eccde6 100644 --- a/vllm/model_executor/models/ernie45_moe.py +++ b/vllm/model_executor/models/ernie45_moe.py @@ -42,7 +42,10 @@ 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, @@ -485,7 +488,7 @@ 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 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", diff --git a/vllm/model_executor/models/ernie45_vl_moe.py b/vllm/model_executor/models/ernie45_vl_moe.py index b4e7af9304..38ed756ba4 100644 --- a/vllm/model_executor/models/ernie45_vl_moe.py +++ b/vllm/model_executor/models/ernie45_vl_moe.py @@ -36,7 +36,10 @@ 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 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, @@ -649,7 +652,7 @@ 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 = 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", diff --git a/vllm/model_executor/models/exaone_moe.py b/vllm/model_executor/models/exaone_moe.py index dd91a18962..80b7e0957e 100644 --- a/vllm/model_executor/models/exaone_moe.py +++ b/vllm/model_executor/models/exaone_moe.py @@ -30,7 +30,10 @@ from vllm.distributed import ( get_pp_group, get_tensor_model_parallel_world_size, ) -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 ReplicatedLinear from vllm.model_executor.layers.logits_processor import LogitsProcessor @@ -326,7 +329,7 @@ class ExaoneMoeModel(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", diff --git a/vllm/model_executor/models/flex_olmo.py b/vllm/model_executor/models/flex_olmo.py index 1b2047eb23..2ff9d86056 100644 --- a/vllm/model_executor/models/flex_olmo.py +++ b/vllm/model_executor/models/flex_olmo.py @@ -20,7 +20,9 @@ from torch import nn from vllm.config import VllmConfig from vllm.distributed import get_tensor_model_parallel_world_size from vllm.logger import init_logger -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 ReplicatedLinear from vllm.model_executor.models.olmoe import OlmoeAttention, OlmoeForCausalLM diff --git a/vllm/model_executor/models/gemma4.py b/vllm/model_executor/models/gemma4.py index 42762e36f8..bb91fd601e 100644 --- a/vllm/model_executor/models/gemma4.py +++ b/vllm/model_executor/models/gemma4.py @@ -37,7 +37,10 @@ from vllm.forward_context import get_forward_context 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, GateLinear +from vllm.model_executor.layers.fused_moe import ( + FusedMoE, + GateLinear, +) from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import ( ColumnParallelLinear, diff --git a/vllm/model_executor/models/glm4_moe.py b/vllm/model_executor/models/glm4_moe.py index 680e746099..aeec6fefa2 100644 --- a/vllm/model_executor/models/glm4_moe.py +++ b/vllm/model_executor/models/glm4_moe.py @@ -42,7 +42,10 @@ 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, @@ -466,7 +469,7 @@ 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 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", diff --git a/vllm/model_executor/models/glm4_moe_lite.py b/vllm/model_executor/models/glm4_moe_lite.py index 5dc33ec18b..77aaa179aa 100644 --- a/vllm/model_executor/models/glm4_moe_lite.py +++ b/vllm/model_executor/models/glm4_moe_lite.py @@ -41,7 +41,9 @@ from vllm.distributed import ( get_pp_group, ) from vllm.logger import init_logger -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.logits_processor import LogitsProcessor from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -308,7 +310,7 @@ 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 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", @@ -334,7 +336,7 @@ class Glm4MoeLiteModel(nn.Module): # Params for weights, fp8 weight scales, fp8 activation scales # (param_name, weight_name, expert_id, shard_id) - expert_params_mapping = 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", @@ -616,7 +618,7 @@ 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 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", diff --git a/vllm/model_executor/models/glm4_moe_lite_mtp.py b/vllm/model_executor/models/glm4_moe_lite_mtp.py index e00476abac..596cb48fac 100644 --- a/vllm/model_executor/models/glm4_moe_lite_mtp.py +++ b/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", diff --git a/vllm/model_executor/models/glm4_moe_mtp.py b/vllm/model_executor/models/glm4_moe_mtp.py index cde94673e5..791ecabebe 100644 --- a/vllm/model_executor/models/glm4_moe_mtp.py +++ b/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", diff --git a/vllm/model_executor/models/gpt_oss.py b/vllm/model_executor/models/gpt_oss.py index b6edc34430..d12db96c5d 100644 --- a/vllm/model_executor/models/gpt_oss.py +++ b/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", diff --git a/vllm/model_executor/models/granitemoe.py b/vllm/model_executor/models/granitemoe.py index f57a8c942b..e3585a6dd7 100644 --- a/vllm/model_executor/models/granitemoe.py +++ b/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", diff --git a/vllm/model_executor/models/grok1.py b/vllm/model_executor/models/grok1.py index c9aa3d2068..f06122a7fd 100644 --- a/vllm/model_executor/models/grok1.py +++ b/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, diff --git a/vllm/model_executor/models/hunyuan_v1.py b/vllm/model_executor/models/hunyuan_v1.py index 9d3ebe4ed9..fca801b748 100644 --- a/vllm/model_executor/models/hunyuan_v1.py +++ b/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", diff --git a/vllm/model_executor/models/interns1_pro.py b/vllm/model_executor/models/interns1_pro.py index 9612ea57b2..36f669179c 100644 --- a/vllm/model_executor/models/interns1_pro.py +++ b/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, diff --git a/vllm/model_executor/models/jamba.py b/vllm/model_executor/models/jamba.py index b4b3b6873d..84e96def6c 100644 --- a/vllm/model_executor/models/jamba.py +++ b/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", diff --git a/vllm/model_executor/models/kimi_linear.py b/vllm/model_executor/models/kimi_linear.py index 21940fb2e1..29d827d196 100644 --- a/vllm/model_executor/models/kimi_linear.py +++ b/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", diff --git a/vllm/model_executor/models/lfm2_moe.py b/vllm/model_executor/models/lfm2_moe.py index 4b49430c1f..55b00d2b9e 100644 --- a/vllm/model_executor/models/lfm2_moe.py +++ b/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", diff --git a/vllm/model_executor/models/llama4.py b/vllm/model_executor/models/llama4.py index c9495a743b..bfcb72a6a7 100644 --- a/vllm/model_executor/models/llama4.py +++ b/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", diff --git a/vllm/model_executor/models/longcat_flash.py b/vllm/model_executor/models/longcat_flash.py index 945fcb6150..d81df6f337 100644 --- a/vllm/model_executor/models/longcat_flash.py +++ b/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", diff --git a/vllm/model_executor/models/mimo_v2_flash.py b/vllm/model_executor/models/mimo_v2_flash.py index 0b466f1660..0fe31c129e 100644 --- a/vllm/model_executor/models/mimo_v2_flash.py +++ b/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", diff --git a/vllm/model_executor/models/minimax_m2.py b/vllm/model_executor/models/minimax_m2.py index 84d8dda533..c4a00f4101 100644 --- a/vllm/model_executor/models/minimax_m2.py +++ b/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", diff --git a/vllm/model_executor/models/minimax_text_01.py b/vllm/model_executor/models/minimax_text_01.py index 67d7cb2d8b..c73fbf7009 100644 --- a/vllm/model_executor/models/minimax_text_01.py +++ b/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, diff --git a/vllm/model_executor/models/mixtral.py b/vllm/model_executor/models/mixtral.py index c182444f66..cbfc254dda 100644 --- a/vllm/model_executor/models/mixtral.py +++ b/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", diff --git a/vllm/model_executor/models/mllama4.py b/vllm/model_executor/models/mllama4.py index 227ef2fa66..8fe1be721c 100644 --- a/vllm/model_executor/models/mllama4.py +++ b/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", diff --git a/vllm/model_executor/models/nemotron_h.py b/vllm/model_executor/models/nemotron_h.py index 9b8ed68560..537e19afbc 100644 --- a/vllm/model_executor/models/nemotron_h.py +++ b/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 diff --git a/vllm/model_executor/models/nemotron_h_mtp.py b/vllm/model_executor/models/nemotron_h_mtp.py index 12551d4254..fe737438c3 100644 --- a/vllm/model_executor/models/nemotron_h_mtp.py +++ b/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", diff --git a/vllm/model_executor/models/olmoe.py b/vllm/model_executor/models/olmoe.py index fcde2e41af..1f342ad173 100644 --- a/vllm/model_executor/models/olmoe.py +++ b/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", diff --git a/vllm/model_executor/models/openpangu.py b/vllm/model_executor/models/openpangu.py index 96b837e42a..68ab4a9ae4 100644 --- a/vllm/model_executor/models/openpangu.py +++ b/vllm/model_executor/models/openpangu.py @@ -44,7 +44,10 @@ from vllm.model_executor.layers.attention import ( Attention, StaticSinkAttention, ) -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, @@ -1149,7 +1152,7 @@ class OpenPanguModel(nn.Module): ] has_experts = hasattr(self.config, "n_routed_experts") if has_experts: - expert_merge_mapping = FusedMoE.make_expert_params_mapping( + expert_merge_mapping = fused_moe_make_expert_params_mapping( self, ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/openpangu_mtp.py b/vllm/model_executor/models/openpangu_mtp.py index 91b454a4bc..3a04ccdff5 100644 --- a/vllm/model_executor/models/openpangu_mtp.py +++ b/vllm/model_executor/models/openpangu_mtp.py @@ -28,7 +28,9 @@ from vllm.config import VllmConfig # SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project -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.logits_processor import LogitsProcessor from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -147,7 +149,7 @@ class OpenPanguMTP(nn.Module): ("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", diff --git a/vllm/model_executor/models/param2moe.py b/vllm/model_executor/models/param2moe.py index 4d1b3ff1b9..e8ea2dbc0e 100644 --- a/vllm/model_executor/models/param2moe.py +++ b/vllm/model_executor/models/param2moe.py @@ -32,7 +32,10 @@ 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, @@ -690,7 +693,7 @@ class Param2MoEModel(nn.Module): return loaded_params 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="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/phimoe.py b/vllm/model_executor/models/phimoe.py index 7d6083f202..5770420ce5 100644 --- a/vllm/model_executor/models/phimoe.py +++ b/vllm/model_executor/models/phimoe.py @@ -35,7 +35,10 @@ from vllm.compilation.decorators import support_torch_compile from vllm.config import CacheConfig, VllmConfig from vllm.distributed import get_pp_group, 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.linear import ( QKVParallelLinear, ReplicatedLinear, @@ -514,7 +517,7 @@ class PhiMoEModel(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", diff --git a/vllm/model_executor/models/qwen2_moe.py b/vllm/model_executor/models/qwen2_moe.py index 7fc3c6a7dd..77eea390ed 100644 --- a/vllm/model_executor/models/qwen2_moe.py +++ b/vllm/model_executor/models/qwen2_moe.py @@ -40,7 +40,10 @@ 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 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, @@ -418,7 +421,7 @@ 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 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", diff --git a/vllm/model_executor/models/qwen3_5_mtp.py b/vllm/model_executor/models/qwen3_5_mtp.py index bbb296d28c..e86b205b9f 100644 --- a/vllm/model_executor/models/qwen3_5_mtp.py +++ b/vllm/model_executor/models/qwen3_5_mtp.py @@ -12,7 +12,9 @@ from vllm.compilation.decorators import support_torch_compile from vllm.config import VllmConfig from vllm.distributed.parallel_state import get_pp_group from vllm.logger import init_logger -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 from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -194,7 +196,7 @@ class Qwen3_5MultiTokenPredictor(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="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/qwen3_moe.py b/vllm/model_executor/models/qwen3_moe.py index 520126718f..4ec1be3367 100644 --- a/vllm/model_executor/models/qwen3_moe.py +++ b/vllm/model_executor/models/qwen3_moe.py @@ -43,7 +43,10 @@ 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, @@ -516,7 +519,7 @@ 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 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", diff --git a/vllm/model_executor/models/qwen3_next.py b/vllm/model_executor/models/qwen3_next.py index 2a4021be6e..96d7e9c713 100644 --- a/vllm/model_executor/models/qwen3_next.py +++ b/vllm/model_executor/models/qwen3_next.py @@ -23,7 +23,10 @@ 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 FusedMoE +from vllm.model_executor.layers.fused_moe import ( + FusedMoE, + fused_moe_make_expert_params_mapping, +) from vllm.model_executor.layers.layernorm import ( GemmaRMSNorm as Qwen3NextRMSNorm, ) @@ -533,7 +536,7 @@ 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 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", diff --git a/vllm/model_executor/models/qwen3_next_mtp.py b/vllm/model_executor/models/qwen3_next_mtp.py index 751d7c23eb..2f411c48a6 100644 --- a/vllm/model_executor/models/qwen3_next_mtp.py +++ b/vllm/model_executor/models/qwen3_next_mtp.py @@ -11,7 +11,9 @@ from vllm.compilation.decorators import support_torch_compile from vllm.config import VllmConfig from vllm.distributed.parallel_state import get_pp_group from vllm.logger import init_logger -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 from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.vocab_parallel_embedding import ( @@ -145,7 +147,7 @@ class Qwen3NextMultiTokenPredictor(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="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/sarvam.py b/vllm/model_executor/models/sarvam.py index c770e20320..a0ab6c0ce2 100644 --- a/vllm/model_executor/models/sarvam.py +++ b/vllm/model_executor/models/sarvam.py @@ -35,7 +35,10 @@ 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 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, @@ -529,7 +532,7 @@ class SarvamMLAModel(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="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/step3_text.py b/vllm/model_executor/models/step3_text.py index 912a1b0754..8f08f6c607 100644 --- a/vllm/model_executor/models/step3_text.py +++ b/vllm/model_executor/models/step3_text.py @@ -18,7 +18,9 @@ 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 ( ColumnParallelLinear, diff --git a/vllm/model_executor/models/step3p5.py b/vllm/model_executor/models/step3p5.py index a0bc1211bf..df051fb873 100644 --- a/vllm/model_executor/models/step3p5.py +++ b/vllm/model_executor/models/step3p5.py @@ -23,7 +23,10 @@ from vllm.distributed import ( 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 import ( + FusedMoE, + fused_moe_make_expert_params_mapping, +) from vllm.model_executor.layers.layernorm import GemmaRMSNorm from vllm.model_executor.layers.linear import ( ColumnParallelLinear, @@ -637,7 +640,7 @@ class Step3p5Model(nn.Module): ] # New per-expert format: .moe.experts.E.gate_proj.weight_packed [out, in] - per_expert_mapping = FusedMoE.make_expert_params_mapping( + per_expert_mapping = fused_moe_make_expert_params_mapping( self, ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", diff --git a/vllm/model_executor/models/transformers/moe.py b/vllm/model_executor/models/transformers/moe.py index cf13958ef7..51a51799ff 100644 --- a/vllm/model_executor/models/transformers/moe.py +++ b/vllm/model_executor/models/transformers/moe.py @@ -25,7 +25,10 @@ from vllm.config.utils import getattr_iter from vllm.distributed import get_dp_group, get_ep_group from vllm.forward_context import ForwardContext, get_forward_context from vllm.model_executor.custom_op import PluggableLayer -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.models.interfaces import MixtureOfExperts from vllm.model_executor.models.utils import maybe_prefix from vllm.platforms import current_platform @@ -179,7 +182,7 @@ class MoEMixin(MixtureOfExperts): num_redundant_experts = self.parallel_config.eplb_config.num_redundant_experts for gate_proj, down_proj, up_proj in ckpt_names: expert_mapping.extend( - FusedMoE.make_expert_params_mapping( + fused_moe_make_expert_params_mapping( self, ckpt_gate_proj_name=gate_proj, ckpt_down_proj_name=down_proj,