DPO Qwen2.5 Omni, inputs_embeds = base_model.thinker.model.embed_tokens(input_ids),RuntimeError: 'weight' must be 2-D

Describe the bug
DPO微调Qwen2.5 Omni-7B，设置流式读取

Your hardware and system info
训练脚本如下：

NPROC_PER_NODE=8 \
MAX_PIXELS=602112 \
VIDEO_MAX_PIXELS=602112 \
FPS_MAX_FRAMES=64 \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \


swift rlhf \
    --rlhf_type dpo \
    --model ./models/Qwen2.5-Omni-7B \
    --train_type full \
    --dataset xxx/dpo-v1-swift.json \
    --load_from_cache_file true \
    --split_dataset_ratio 0.01 \
    --torch_dtype bfloat16 \
    --num_train_epochs 1 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --learning_rate 1e-5 \
    --gradient_accumulation_steps 2 \
    --eval_steps 100 \
    --save_steps 100 \
    --save_total_limit 2 \
    --logging_steps 5 \
    --max_length 32000 \
    --output_dir output/qwen2_5_dpo \
    --warmup_ratio 0.05 \
    --save_only_model true \
    --dataloader_num_workers 4 \
    --dataset_num_proc 4 \
    --deepspeed zero3 \
    --attn_impl flash_attn \
    --rpo_alpha 0.1 \
    --padding_free true \
    --lora_rank 64 \
    --lora_alpha 32 \
    --freeze_vit true \
    --streaming true \
    --max_steps 1368

debug发现llm/template/template的qwen.py中780行
inputs_embeds = base_model.thinker.model.embed_tokens(input_ids)
中的base_model.thinker.model.embed_tokens.shape为[0]
Additional context
Add any other context about the problem here(在这里补充其他信息)

Start with llm/template/template/qwen.py around line 780 and trace how the DPO streaming path constructs base_model.thinker.model.embed_tokens. Reproduce the reported command with the listed Qwen2.5-Omni configuration, then inspect why the embedding weight has shape [0]. Done means identifying and correcting the failing path with a regression check for this training setup.