训练LongCat-Flash-Omni sft报错

[Mryangkaitong]

如题，训练LongCat-Flash-Omni sft报错。

model_path=model/LongCat-Flash-Omni

megatron sft \ --model ${model_path} \ --load_safetensors true \ --save_safetensors true \ --dataset 'data/base_user_profile.json' \ 'data/online.json' \ --val_dataset 'paras2s_1217.json' \ --tensor_model_parallel_size 4 \ --pipeline_model_parallel_size 2 \ --expert_model_parallel_size 4 \ --context_parallel_size 1 \ --moe_grouped_gemm true \ --moe_shared_expert_overlap true \ --moe_aux_loss_coeff 0 \ --micro_batch_size 1 \ --global_batch_size 16 \ --freeze_llm false \ --freeze_vit true \ --freeze_aligner true \ --recompute_granularity full \ --recompute_method uniform \ --recompute_num_layers 1 \ --max_epochs 2 \ --finetune true \ --rpo_alpha 0.1 \ --cross_entropy_loss_fusion true \ --lr 1e-6 \ --lr_warmup_fraction 0.00 \ --min_lr 1e-7 \ --save ${output_path} \ --eval_interval 100 \ --save_interval 100 \ --max_length 8192 \ --packing true \

报错：

^^^^^^^^^^^^^^^ File "<string>", line 313, in __init__ File "ms_swift_1109/swift/megatron/argument/train_args.py", line 42, in __post_ init__ assert self.megatron_model_meta is not None, f'Model: {self.model} is not supported.' ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AssertionError: Model: model/LongCat-Flash-Omni is not supported. super().__post_init__() File "ms_swift_1109/swift/megatron/argument/megatron_base_args.py", line 22, in __post_init__ assert self.megatron_model_meta is not None, f'Model: {self.model} is not supported.' ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AssertionError: Model: model/LongCat-Flash-Omni is not supported.

LongCat-Flash-Omni 模型：https://github.com/meituan-longcat/LongCat-Flash-Omni