diff --git a/vllm/model_executor/models/gemma4_unified.py b/vllm/model_executor/models/gemma4_unified.py index 9cc0710c4d..64c084b616 100644 --- a/vllm/model_executor/models/gemma4_unified.py +++ b/vllm/model_executor/models/gemma4_unified.py @@ -335,7 +335,6 @@ class Gemma4UnifiedForConditionalGeneration(Gemma4ForConditionalGeneration): ) # --- MixtureOfExperts delegation to language_model --- - self.expert_weights = self.language_model.expert_weights self.moe_layers = self.language_model.moe_layers self.num_moe_layers = self.language_model.num_moe_layers self.num_logical_experts = self.language_model.num_logical_experts @@ -345,6 +344,7 @@ class Gemma4UnifiedForConditionalGeneration(Gemma4ForConditionalGeneration): self.num_expert_groups = self.language_model.num_expert_groups self.num_shared_experts = self.language_model.num_shared_experts self.num_redundant_experts = self.language_model.num_redundant_experts + self.set_eplb_state = self.language_model.set_eplb_state gen_cfg = vllm_config.model_config.try_get_generation_config() self._suppress_token_ids = gen_cfg.get("suppress_tokens") if gen_cfg else None diff --git a/vllm/model_executor/models/mllama4.py b/vllm/model_executor/models/mllama4.py index 855fe5a47a..178dae506c 100644 --- a/vllm/model_executor/models/mllama4.py +++ b/vllm/model_executor/models/mllama4.py @@ -812,17 +812,6 @@ class Llama4ForConditionalGeneration( ) return self.language_model.get_eagle3_default_aux_hidden_state_layers() - def set_eplb_state( - self, - expert_load_view: torch.Tensor, - logical_to_physical_map: torch.Tensor, - logical_replica_count: torch.Tensor, - ): - self.language_model.set_eplb_state( - expert_load_view, logical_to_physical_map, logical_replica_count - ) - self.expert_weights = self.language_model.expert_weights - def update_physical_experts_metadata( self, num_physical_experts: int, num_local_physical_experts: int ): diff --git a/vllm/model_executor/models/param2moe.py b/vllm/model_executor/models/param2moe.py index 3386f9545f..ff56cf505f 100644 --- a/vllm/model_executor/models/param2moe.py +++ b/vllm/model_executor/models/param2moe.py @@ -705,8 +705,6 @@ class Param2MoEModel(nn.Module): class Param2MoEMixtureOfExperts(MixtureOfExperts): """Implements the vLLM MixtureOfExperts protocol for Param2MoE.""" - expert_weights: list[torch.Tensor] - def extract_moe_parameters(self, example_moe: Param2MoEMoEBlock | None) -> None: if example_moe is None: raise RuntimeError( @@ -745,24 +743,6 @@ class Param2MoEMixtureOfExperts(MixtureOfExperts): if hasattr(fused, "update_expert_map"): fused.update_expert_map() - def set_eplb_state( - self, - expert_load_view: torch.Tensor, - logical_to_physical_map: torch.Tensor, - logical_replica_count: torch.Tensor, - ) -> None: - self.expert_weights = [] - for layer_idx, layer in enumerate(self.moe_layers): - if hasattr(layer, "get_expert_weights"): - self.expert_weights.append(layer.get_expert_weights()) - if hasattr(layer, "set_eplb_state"): - layer.set_eplb_state( - moe_layer_idx=layer_idx, - expert_load_view=expert_load_view, - logical_to_physical_map=logical_to_physical_map, - logical_replica_count=logical_replica_count, - ) - class Param2MoEForCausalLM( nn.Module, SupportsPP, SupportsLoRA, Param2MoEMixtureOfExperts diff --git a/vllm/model_executor/models/sarvam.py b/vllm/model_executor/models/sarvam.py index fd28e3b391..f59579b1bc 100644 --- a/vllm/model_executor/models/sarvam.py +++ b/vllm/model_executor/models/sarvam.py @@ -711,7 +711,6 @@ class SarvamMLAForCausalLM(nn.Module, SupportsPP, SupportsLoRA, SarvamMixtureOfE self.model.make_empty_intermediate_tensors ) - self.expert_weights = [] self.num_moe_layers = 0 self.moe_layers = [] diff --git a/vllm/models/deepseek_v32/nvidia/model.py b/vllm/models/deepseek_v32/nvidia/model.py index 0cd7fb02ed..b139b8ba23 100644 --- a/vllm/models/deepseek_v32/nvidia/model.py +++ b/vllm/models/deepseek_v32/nvidia/model.py @@ -336,7 +336,6 @@ class DeepseekV32ForCausalLM(DeepseekV2ForCausalLM): def set_moe_parameters(self): # Same as the base, but keyed on the MoE block type rather than the # decoder-layer type (DeepseekV32DecoderLayer is a plain nn.Module). - self.expert_weights = [] self.num_expert_groups = getattr(self.config, "n_group", 1) self.moe_layers = [] self.moe_mlp_layers = [] diff --git a/vllm/models/deepseek_v32/nvidia/mtp.py b/vllm/models/deepseek_v32/nvidia/mtp.py index 7a04d4f3d3..0efa1ac7a7 100644 --- a/vllm/models/deepseek_v32/nvidia/mtp.py +++ b/vllm/models/deepseek_v32/nvidia/mtp.py @@ -175,7 +175,6 @@ class DeepseekV32MTP(nn.Module, DeepseekV2MixtureOfExperts): self.set_moe_parameters() def set_moe_parameters(self): - self.expert_weights = [] self.num_moe_layers = self.config.num_nextn_predict_layers self.num_expert_groups = self.config.n_group self.moe_layers = []