DeepSeek-v4-flash-0731 sft

Checklist / 检查清单

• I have searched existing issues, and this is a new question or discussion topic. / 我已经搜索过现有的 issues，确认这是一个新的问题与讨论。

Question Description / 问题描述

example from https://swift.readthedocs.io/zh-cn/v4.4/BestPractices/deepseek-v4.html

PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
NPROC_PER_NODE=4 \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
megatron sft \
    --model /code/hewangsh/models/DeepSeek-V4-Flash-0731 \
    --save_safetensors true \
    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#1000' \
    --model_author swift \
    --model_name swift-robot \
    --merge_lora true \
    --load_from_cache_file true \
    --add_non_thinking_prefix true \
    --loss_scale ignore_empty_think \
    --split_dataset_ratio 0.01 \
    --tuner_type lora \
    --lora_rank 16 \
    --lora_alpha 32 \
    --tensor_model_parallel_size 1 \
    --expert_model_parallel_size 4 \
    --micro_batch_size 1 \
    --global_batch_size 32 \
    --padding_free false \
    --group_by_length true \
    --recompute_granularity full \
    --recompute_method uniform \
    --recompute_num_layers 1 \
    --moe_permute_fusion true \
    --moe_grouped_gemm true \
    --moe_shared_expert_overlap true \
    --moe_aux_loss_coeff 1e-3 \
    --num_train_epochs 1 \
    --finetune true \
    --cross_entropy_loss_fusion true \
    --lr 1e-4 \
    --lr_warmup_fraction 0.05 \
    --min_lr 1e-5 \
    --output_dir megatron_output/DeepSeek-V4-Flash \
    --eval_steps 200 \
    --save_steps 200 \
    --max_length 4096 \
    --dataloader_num_workers 8 \
    --dataset_num_proc 8 \
    --no_save_optim true \
    --no_save_rng true \
    --sequence_parallel true \
    --mtp_num_layers 1 \
    --attention_backend flash

Error:

[rank3]: Traceback (most recent call last):
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/cli/_megatron/sft.py", line 7, in <module>
[rank3]:     megatron_sft_main()
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/megatron/pipelines/train/sft.py", line 99, in megatron_sft_main
[rank3]:     return MegatronSft(args).main()
[rank3]:            ^^^^^^^^^^^^^^^^^^^^^^^^
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/pipelines/base.py", line 52, in main
[rank3]:     result = self.run()
[rank3]:              ^^^^^^^^^^
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/megatron/pipelines/train/sft.py", line 72, in run
[rank3]:     trainer = self.prepare_trainer()
[rank3]:               ^^^^^^^^^^^^^^^^^^^^^^
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/megatron/pipelines/train/sft.py", line 38, in prepare_trainer
[rank3]:     return MegatronTrainer(self.args, self.template)
[rank3]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/megatron/trainers/base.py", line 68, in __init__
[rank3]:     self.prepare_model()
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/megatron/trainers/base.py", line 188, in prepare_model
[rank3]:     self.unwrapped_models = get_mcore_model(args, self.template.config)
[rank3]:                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/megatron/model/utils.py", line 82, in get_mcore_model
[rank3]:     config = get_mcore_model_config(args, hf_config)
[rank3]:              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank3]:   File "/usr/local/lib/python3.12/site-packages/swift/megatron/model/utils.py", line 73, in get_mcore_model_config
[rank3]:     config = ModelConfig(**kwargs)
[rank3]:              ^^^^^^^^^^^^^^^^^^^^^
[rank3]:   File "<string>", line 305, in __init__
[rank3]:   File "/usr/local/lib/python3.12/site-packages/mcore_bridge/config/model_config.py", line 342, in __post_init__
[rank3]:     super().__post_init__()
[rank3]:   File "/usr/local/lib/python3.12/site-packages/megatron/core/transformer/transformer_config.py", line 1836, in __post_init__
[rank3]:     raise ValueError(
[rank3]: ValueError: Expert bias for aux-loss-free routing only supports sigmoid score function.Please set --moe-router-score-function sigmoid for sigmoid score function.

Start at swift/megatron/model/utils.py, especially get_mcore_model_config, and trace how MegatronSft.prepare_model constructs ModelConfig. Reproduce the supplied four-GPU command and compare its generated configuration with the validation in mcore_bridge/config/model_config.py and megatron/core/transformer/transformer_config.py. Done when the example's configuration and the reported validation error are resolved.