训练minimax- h3的nf4模型的lora，使用adapter训练时融合lora会报错

[rank0]: Traceback (most recent call last):
[rank0]: File "/data4/hjhuang/code/DiffSynth-Studio-2.1.4/examples/minimax_h3/model_training/train.py", line 190, in
[rank0]: model = MiniMaxH3TrainingModule(
[rank0]: File "/data4/hjhuang/code/DiffSynth-Studio-2.1.4/examples/minimax_h3/model_training/train.py", line 50, in init
[rank0]: self.switch_pipe_to_training_mode(
[rank0]: File "/data4/hjhuang/code/DiffSynth-Studio-2.1.4/diffsynth/diffusion/training_module.py", line 322, in switch_pipe_to_training_mode
[rank0]: pipe.load_lora(getattr(pipe, preset_lora_model), preset_lora_path)
[rank0]: File "/data4/hjhuang/code/DiffSynth-Studio-2.1.4/diffsynth/diffusion/base_pipeline.py", line 280, in load_lora
[rank0]: lora_loader.fuse_lora_to_base_model(module, lora, alpha=alpha)
[rank0]: File "/data4/hjhuang/code/DiffSynth-Studio-2.1.4/diffsynth/utils/lora/general.py", line 67, in fuse_lora_to_base_model
[rank0]: state_dict_base["weight"] = state_dict_base["weight"].to(device=self.device, dtype=self.torch_dtype) + weight_lora
[rank0]: RuntimeError: The size of tensor a (57802752) must match the size of tensor b (21504) at non-singleton dimension 0

Start with examples/minimax_h3/model_training/train.py, then trace switch_pipe_to_training_mode in diffsynth/diffusion/training_module.py through load_lora in diffsynth/diffusion/base_pipeline.py. Inspect diffsynth/utils/lora/general.py around line 67 and compare the base and LoRA tensor shapes for the NF4 adapter. Done means the MiniMax H3 NF4 training example can fuse the adapter without the reported size-mismatch error.