[refactor] remove unused op_mlp (#26673)

python/sglang/srt/models/deepseek_v2.py

@@ -2114,19 +2114,6 @@ class DeepseekV2DecoderLayer(nn.Module):
             )
         )
 
-    def op_mlp(self, state):
-        hidden_states = state.pop("hidden_states_mlp_input")
-        if not (
-            enable_moe_dense_fully_dp()
-            and (not self.is_layer_sparse)
-            and hidden_states.shape[0] == 0
-        ):
-            state.hidden_states_mlp_output = self.mlp(
-                hidden_states, state.forward_batch
-            )
-        else:
-            state.hidden_states_mlp_output = hidden_states
-
     def op_comm_postprocess_layer(self, state):
         hidden_states, residual = self.layer_communicator.postprocess_layer(
             state.pop("hidden_states_mlp_output"),

python/sglang/srt/models/glm4_moe.py

@@ -1017,19 +1017,6 @@ class Glm4MoeDecoderLayer(nn.Module):
             )
         )
 
-    def op_mlp(self, state):
-        hidden_states = state.pop("hidden_states_mlp_input")
-        if not (
-            enable_moe_dense_fully_dp()
-            and (not self.is_layer_sparse)
-            and hidden_states.shape[0] == 0
-        ):
-            state.hidden_states_mlp_output = self.mlp(
-                hidden_states, state.forward_batch
-            )
-        else:
-            state.hidden_states_mlp_output = hidden_states
-
     def op_comm_postprocess_layer(self, state):
         hidden_states, residual = self.layer_communicator.postprocess_layer(
             state.pop("hidden_states_mlp_output"),

python/sglang/srt/models/glm4_moe_lite.py

@@ -737,19 +737,6 @@ class Glm4MoeLiteDecoderLayer(nn.Module):
             )
         )
 
-    def op_mlp(self, state):
-        hidden_states = state.pop("hidden_states_mlp_input")
-        if not (
-            enable_moe_dense_fully_dp()
-            and (not self.is_layer_sparse)
-            and hidden_states.shape[0] == 0
-        ):
-            state.hidden_states_mlp_output = self.mlp(
-                hidden_states, state.forward_batch
-            )
-        else:
-            state.hidden_states_mlp_output = hidden_states
-
     def op_comm_postprocess_layer(self, state):
         hidden_states, residual = self.layer_communicator.postprocess_layer(
             state.pop("hidden_states_mlp_output"),

python/sglang/srt/models/mimo_v2.py

@@ -808,10 +808,6 @@ class MiMoV2DecoderLayer(nn.Module):
             )
         )
 
-    def op_mlp(self, state):
-        hidden_states = state.pop("hidden_states_mlp_input")
-        state.hidden_states_mlp_output = self.mlp(hidden_states, state.forward_batch)
-
     def op_comm_postprocess_layer(self, state):
         hidden_states, residual = self.layer_communicator.postprocess_layer(
             state.pop("hidden_states_mlp_output"),

python/sglang/srt/models/minimax_m2.py

@@ -1069,12 +1069,6 @@ class MiniMaxM2DecoderLayer(nn.Module):
             )
         )
 
-    def op_mlp(self, state):
-        hidden_states = state.pop("hidden_states_mlp_input")
-        state.hidden_states_mlp_output = self.block_sparse_moe(
-            hidden_states, state.forward_batch
-        )
-
     def op_comm_postprocess_layer(self, state):
         """Communication postprocess for layer - TBO operation"""
         hidden_states, residual = self.layer_communicator.postprocess_layer(

python/sglang/srt/models/qwen3_moe.py

@@ -891,10 +891,6 @@ class Qwen3MoeDecoderLayer(nn.Module):
             )
         )
 
-    def op_mlp(self, state):
-        hidden_states = state.pop("hidden_states_mlp_input")
-        state.hidden_states_mlp_output = self.mlp(hidden_states, state.forward_batch)
-
     def op_comm_postprocess_layer(self, state):
         hidden_states, residual = self.layer_communicator.postprocess_layer(
             state.pop("hidden_states_mlp_output"),