Remove boilerplate missed by #46820 (#46956)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Jun 29, 2026 at 01:11 UTC
311ad689adcde0236d630ca202110f5b0fec85f8
6 files changed
+1
-35
vllm/model_executor/models/gemma4_unified.py
+1
-1
@@ -335,7 +335,6 @@ class Gemma4UnifiedForConditionalGeneration(Gemma4ForConditionalGeneration):
335
)
336
337
# --- MixtureOfExperts delegation to language_model ---
338
- self.expert_weights = self.language_model.expert_weights
338
self.moe_layers = self.language_model.moe_layers
339
self.num_moe_layers = self.language_model.num_moe_layers
340
self.num_logical_experts = self.language_model.num_logical_experts
@@ -345,6 +344,7 @@ class Gemma4UnifiedForConditionalGeneration(Gemma4ForConditionalGeneration):
344
self.num_expert_groups = self.language_model.num_expert_groups
345
self.num_shared_experts = self.language_model.num_shared_experts
346
self.num_redundant_experts = self.language_model.num_redundant_experts
347
+ self.set_eplb_state = self.language_model.set_eplb_state
348
349
gen_cfg = vllm_config.model_config.try_get_generation_config()
350
self._suppress_token_ids = gen_cfg.get("suppress_tokens") if gen_cfg else None
vllm/model_executor/models/mllama4.py
-11
@@ -812,17 +812,6 @@ class Llama4ForConditionalGeneration(
812
)
813
return self.language_model.get_eagle3_default_aux_hidden_state_layers()
814
815
- def set_eplb_state(
816
- self,
817
- expert_load_view: torch.Tensor,
818
- logical_to_physical_map: torch.Tensor,
819
- logical_replica_count: torch.Tensor,
820
- ):
821
- self.language_model.set_eplb_state(
822
- expert_load_view, logical_to_physical_map, logical_replica_count
823
- )
824
- self.expert_weights = self.language_model.expert_weights
825
-
815
def update_physical_experts_metadata(
816
self, num_physical_experts: int, num_local_physical_experts: int
817
):
vllm/model_executor/models/param2moe.py
-20
@@ -705,8 +705,6 @@ class Param2MoEModel(nn.Module):
705
class Param2MoEMixtureOfExperts(MixtureOfExperts):
706
"""Implements the vLLM MixtureOfExperts protocol for Param2MoE."""
707
708
- expert_weights: list[torch.Tensor]
709
-
708
def extract_moe_parameters(self, example_moe: Param2MoEMoEBlock | None) -> None:
709
if example_moe is None:
710
raise RuntimeError(
@@ -745,24 +743,6 @@ class Param2MoEMixtureOfExperts(MixtureOfExperts):
743
if hasattr(fused, "update_expert_map"):
744
fused.update_expert_map()
745
748
- def set_eplb_state(
749
- self,
750
- expert_load_view: torch.Tensor,
751
- logical_to_physical_map: torch.Tensor,
752
- logical_replica_count: torch.Tensor,
753
- ) -> None:
754
- self.expert_weights = []
755
- for layer_idx, layer in enumerate(self.moe_layers):
756
- if hasattr(layer, "get_expert_weights"):
757
- self.expert_weights.append(layer.get_expert_weights())
758
- if hasattr(layer, "set_eplb_state"):
759
- layer.set_eplb_state(
760
- moe_layer_idx=layer_idx,
761
- expert_load_view=expert_load_view,
762
- logical_to_physical_map=logical_to_physical_map,
763
- logical_replica_count=logical_replica_count,
764
- )
765
-
746
747
class Param2MoEForCausalLM(
748
nn.Module, SupportsPP, SupportsLoRA, Param2MoEMixtureOfExperts
vllm/model_executor/models/sarvam.py
-1
@@ -711,7 +711,6 @@ class SarvamMLAForCausalLM(nn.Module, SupportsPP, SupportsLoRA, SarvamMixtureOfE
711
self.model.make_empty_intermediate_tensors
712
)
713
714
- self.expert_weights = []
714
self.num_moe_layers = 0
715
716
self.moe_layers = []
vllm/models/deepseek_v32/nvidia/model.py
-1
@@ -336,7 +336,6 @@ class DeepseekV32ForCausalLM(DeepseekV2ForCausalLM):
336
def set_moe_parameters(self):
337
# Same as the base, but keyed on the MoE block type rather than the
338
# decoder-layer type (DeepseekV32DecoderLayer is a plain nn.Module).
339
- self.expert_weights = []
339
self.num_expert_groups = getattr(self.config, "n_group", 1)
340
self.moe_layers = []
341
self.moe_mlp_layers = []
vllm/models/deepseek_v32/nvidia/mtp.py
-1
@@ -175,7 +175,6 @@ class DeepseekV32MTP(nn.Module, DeepseekV2MixtureOfExperts):
175
self.set_moe_parameters()
176
177
def set_moe_parameters(self):
178
- self.expert_weights = []
178
self.num_moe_layers = self.config.num_nextn_predict_layers
179
self.num_expert_groups = self.config.n_group
180
self.moe_layers = []