微调deepseek ocr 报错 fsdp为空

[jcyfootballer]

Describe the bug
What the bug is, and how to reproduce, better with screenshots(描述bug以及复现过程，最好有截图)

Your hardware and system info
Write your system info like CUDA version/system/GPU/torch version here(在这里给出硬件信息和系统信息，如CUDA版本，系统，GPU型号和torch版本等)

Additional context
Add any other context about the problem here(在这里补充其他信息)
ms-swift/.venv/lib/python3.11/site-packages/transformers/training_args.py", line 1887, in post_init
elif FSDPOption.FULL_SHARD in self.fsdp and FSDPOption.SHARD_GRAD_OP in self.fsdp:
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: argument of type 'NoneType' is not iterable

[hjh0119]

plz provide reproduce script and env infos

[Dong237]

I am using this command to start training and got the exactly same error

#!/bin/bash

# Multi-node distributed training configuration
# Set these based on your cluster setup:
export NNODES=${NNODES:-1}                    # Total number of nodes
export NODE_RANK=${NODE_RANK:-0}              # Current node rank (0 for master, 1,2,3... for workers)
export MASTER_ADDR=${MASTER_ADDR:-"127.0.0.1"} # Master node IP address
export MASTER_PORT=${MASTER_PORT:-29500}      # Master node port for communication

# Per-node configuration
export NPROC_PER_NODE=8                       # Number of processes per node (GPUs per node)
export WORLD_SIZE=$((NNODES * NPROC_PER_NODE)) # Total number of processes across all nodes

# CUDA configuration
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True'

# Run name for logging
run_name="qwen3_1point7b_instruct"
output_dir="/mnt/bn/youxiang-lf/models/$run_name"

# Create output directory
mkdir -p $output_dir

# Use torchrun for proper distributed training
torchrun \
    --nnodes=$NNODES \
    --nproc_per_node=$NPROC_PER_NODE \
    --node_rank=$NODE_RANK \
    --master_addr=$MASTER_ADDR \
    --master_port=$MASTER_PORT \
    --max_restarts=3 \
    --rdzv_id=$run_name \
    --rdzv_backend=c10d \
    --rdzv_endpoint=$MASTER_ADDR:$MASTER_PORT \
    -m swift.cli.rlhf \
    --rlhf_type gkd \
    --model Qwen3-1.7B \
    --teacher_model Qwen3-1.7B \
    --train_type full \
    --custom_dataset_info dataset_info_ms_swift.json \
    --dataset xxx \
    --split_dataset_ratio 0.01 \
    --seq_kd false \
    --lmbda 1 \
    --beta 0.5 \
    --torch_dtype bfloat16 \
    --num_train_epochs 2 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --learning_rate 1e-5 \
    --gradient_accumulation_steps 1 \
    --eval_steps 5 \
    --save_steps 5 \
    --save_total_limit 5 \
    --logging_steps 5 \
    --max_length 4096 \
    --max_completion_length 1024 \
    --output_dir $output_dir \
    --warmup_ratio 0.05 \
    --save_only_model true \
    --dataloader_num_workers 16 \
    --dataset_num_proc 16 \
    --deepspeed zero2 \
    --attn_impl flash_attn \
    --teacher_deepspeed zero3_offload \
    --use_vllm true \
    --vllm_mode colocate \
    --vllm_gpu_memory_utilization 0.3 \
    --sleep_level 1 2>&1 | tee -a $output_dir/training.log

plz provide reproduce script and env infos

[hjh0119]

run normally on my end

swift rlhf \
    --rlhf_type gkd \
    --model Qwen/Qwen3-1.7B \
    --teacher_model Qwen/Qwen3-1.7B \
    --train_type full \
    --dataset AI-ModelScope/alpaca-gpt4-data-en#2000 \
    --split_dataset_ratio 0.01 \
    --seq_kd false \
    --lmbda 1 \
    --beta 0.5 \
    --torch_dtype bfloat16 \
    --num_train_epochs 2 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --learning_rate 1e-5 \
    --gradient_accumulation_steps 1 \
    --eval_steps 5 \
    --save_steps 5 \
    --save_total_limit 5 \
    --logging_steps 5 \
    --max_length 4096 \
    --max_completion_length 1024 \
    --output_dir $output_dir \
    --warmup_ratio 0.05 \
    --save_only_model true \
    --dataloader_num_workers 16 \
    --dataset_num_proc 16 \
    --deepspeed zero2 \
    --attn_impl flash_attn \
    --teacher_deepspeed zero3_offload \
    --use_vllm true \
    --vllm_mode colocate \
    --vllm_gpu_memory_utilization 0.3 \
    --sleep_level 1

[hjh0119]

plz provide env infos

[Ryann-Ran]

Passing --fsdp "" explicitly solves the issue temporarily.

[hjh0119]

fixed in #7235