Qwen3-Next-80b-a3b-instruct-fp8不支持GRPO训练

Traceback (most recent call last):
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/cli/_megatron/rlhf.py", line 7, in
[rank3]: megatron_rlhf_main()
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/train/rlhf.py", line 71, in megatron_rlhf_main
[rank3]: return MegatronRLHF(args).main()
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/llm/base.py", line 49, in main
[rank3]: result = self.run()
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/train/sft.py", line 77, in run
[rank3]: self.trainer.train(train_dataset, val_dataset, data_collator)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/grpo_trainer.py", line 70, in train
[rank3]: super().train(train_dataset, val_dataset, data_collator)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/base.py", line 1016, in train
[rank3]: pretrain(
[rank3]: File "/root/.cache/modelscope/hub/_github/Megatron-LM/megatron/training/training.py", line 737, in pretrain
[rank3]: iteration, num_floating_point_operations_so_far = train(
[rank3]: File "/root/.cache/modelscope/hub/_github/Megatron-LM/megatron/training/training.py", line 2298, in train
[rank3]: ) = train_step(
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/base.py", line 516, in train_step
[rank3]: new_data_iterator = self._replace_data_iterator(data_iterator, model)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/grpo_trainer.py", line 503, in _replace_data_iterator
[rank3]: micro_batch_data = self._generate_and_score_completions(rollout_batch)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/grpo_trainer.py", line 528, in _generate_and_score_completions
[rank3]: rollout_batch = self._generate_completions(rollout_batch)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/utils.py", line 154, in wrapper
[rank3]: return func(self, *args, **kwargs)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/grpo_trainer.py", line 691, in _generate_completions
[rank3]: self._move_model_to_vllm()
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/utils.py", line 154, in wrapper
[rank3]: return func(self, *args, **kwargs)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/grpo_trainer.py", line 230, in _move_model_to_vllm
[rank3]: self._export_and_load_weights()
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/swift/megatron/trainers/grpo_trainer.py", line 266, in _export_and_load_weights
[rank3]: llm_model.load_weights(weight_iterator)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/vllm/model_executor/models/qwen3_next.py", line 1279, in load_weights
[rank3]: return loader.load_weights(weights)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/vllm/model_executor/model_loader/online_quantization.py", line 173, in patched_model_load_weights
[rank3]: return original_load_weights(auto_weight_loader, weights, mapper=mapper)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/vllm/model_executor/models/utils.py", line 335, in load_weights
[rank3]: autoloaded_weights = set(self._load_module("", self.module, weights))
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/vllm/model_executor/models/utils.py", line 288, in _load_module
[rank3]: yield from self._load_module(
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/vllm/model_executor/models/utils.py", line 261, in _load_module
[rank3]: loaded_params = module_load_weights(weights)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/vllm/model_executor/models/qwen3_next.py", line 1122, in load_weights
[rank3]: weight_loader(param, loaded_weight)
[rank3]: File "/opt/conda/envs/zzb_megatron/lib/python3.10/site-packages/vllm/model_executor/layers/linear.py", line 555, in weight_loader_v2
[rank3]: param.load_column_parallel_weight(loaded_weight=loaded_weight)
[rank3]: AttributeError: 'Parameter' object has no attribute 'load_column_parallel_weight'

Reproduce the GRPO run for Qwen3-Next-80b-a3b-instruct-fp8 and start by reading swift/megatron/trainers/grpo_trainer.py around _export_and_load_weights and vllm/model_executor/models/qwen3_next.py around load_weights. Trace the weight loader in vllm/model_executor/layers/linear.py; done means the model can proceed through GRPO weight loading without the reported AttributeError.