Qwen3Omni lora微调后推理报错

使用qwen3-omni的megaton微调后，进行推理出现报错：

ValueError: Qwen3OmniMoeThinkerForConditionalGeneration does not support LoRA yet.

推理脚本：

CUDA_VISIBLE_DEVICES=0,1 \
MAX_PIXELS=1003520 \
swift infer \
    --model Qwen3-Omni/Qwen3-Omni-30B-A3B-Instruct \
    --adapters ckpt/v4-20251212-163234/checkpoint-3 \
    --merge_lora false \
    --stream true \
    --infer_backend vllm \
    --val_dataset ms-swift/data/train_test.jsonl \
    --vllm_gpu_memory_utilization 0.9 \
    --vllm_tensor_parallel_size 2 \
    --vllm_max_model_len 32768 \
    --max_new_tokens 2048 \
    --vllm_limit_mm_per_prompt '{'image': 3, 'video': 3, 'audio': 3}'

请问vllm版本有什么要求吗？我尝试了0.11和0.9.2都没效果

Start by reproducing the reported swift infer command with the Qwen3-Omni model, checkpoint, and vLLM backend. Compare the reported vLLM 0.11 and 0.9.2 environments while tracing the LoRA support check that raises the error. Done means identifying a supported configuration or confirming and documenting the limitation.