gemma3 in grpo doesn't work when using lmdeploy

[kakao-charlie-cs]

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)

I followed this example script for training gemma3-4b in GRPO

• https://github.com/modelscope/ms-swift/blob/main/examples/train/grpo/full_lmdeploy.sh

I used gemma3-4b, but no-vision model. So it is generally similar as gemma-3-1b-it.

• https://huggingface.co/gghfez/gemma-3-4b-novision

lmdeploy model which is not supported by turbomind doesn't have load_weights method.
Therefore, below lines raised Exception related the method.

ms-swift/swift/trainers/rlhf_trainer/grpo_trainer.py

Lines 558 to 565 in 9860d42

	if self.infer_rank >= 0:
	if self.args.async_generate:
	self._wait_queue()
	if self.args.use_vllm:
	llm_model = self.engine.inner_model
	else:
	llm_model = self.engine.engine.engine
	llm_model.load_weights(state_dict.items())

qwen2_5 based model works fine (since it was supported). But gemma3 not work properly.

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)

• single node H200-8ea
• deepspeed==0.14.5
• trl==0.16.1
• lmdeploy==0.7.2.post1
• torch==2.6.0
• CUDA==12.2
• ms-swift==3.3.0dev

Additional context
Add any other context about the problem here(在这里补充其他信息)

[RomanticGodVAN]

the same issue

[hjh0119]

The integration with LMDeploy currently only works with the turbomind backend, for non-TurboMind compatible models: plz use vLLM or pt backend