[fix]解决gpt oss nn moe权重加载出错

vllm/model_executor/models/gpt_oss.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import os
 from collections.abc import Iterable
 from typing import Optional
 
@@ -253,6 +254,7 @@ class GptOssModel(nn.Module):
             make_empty_intermediate_tensors_factory(
                 ["hidden_states", "residual"], self.config.hidden_size))
         self.aux_hidden_state_layers = tuple[int, ...]()
+        self.use_nn_moe = int(os.environ.get('MOE_NN', 1)) == 1
 
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.embedding(input_ids)
@@ -522,9 +524,11 @@ class GptOssModel(nn.Module):
                     narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
                 else:
                     narrow_weight = weight[:, :,
-                                           2 * tp_rank_start:2 * tp_rank_end]
+                                            2 * tp_rank_start:2 * tp_rank_end]
+                
+                if not self.use_nn_moe:
+                    narrow_weight = narrow_weight.permute(0, 2, 1).contiguous()
 
-                narrow_weight = narrow_weight.permute(0, 2, 1).contiguous()
                 param = params_dict[name]
 
                 param.copy_(narrow_weight)
@@ -536,9 +540,12 @@ class GptOssModel(nn.Module):
                     narrow_weight = weight[ep_rank_start:ep_rank_end, ...]
                 else:
                     narrow_weight = weight[:, tp_rank_start:tp_rank_end, :]
-                narrow_weight = narrow_weight.permute(0, 2, 1).contiguous()
-                param = params_dict[name]
 
+                if not self.use_nn_moe:
+                    narrow_weight = narrow_weight.permute(0, 2, 1).contiguous()
+
+                param = params_dict[name]
+                
                 param.copy_(narrow_weight)
                 loaded_params.add(name)
                 continue