add support qwen1.5-based DPO for Flash Attention 2

[changingivan]

What does this PR do?

Fixes bug: when use qwen1.5 models and Flash Attention 2 on stage DPO, the defalut padding_side of the qwen1.5 tokenizer is 'right'. However , as qwen1.5 define, tokenizer.padding_side = 'left' should be call before tokenizing the input , otherwise unexpected behaviour may happened and it will raise a ValueError.

192.168.1.38:   File "/usr/local/lib/python3.10/dist-packages/transformers/models/qwen2/modeling_qwen2.py", line 1173, in forward
192.168.1.38:     outputs = self.model(
192.168.1.38:   File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1502, in _wrapped_call_impl
192.168.1.38:     return self._call_impl(*args, **kwargs)
192.168.1.38:   File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1548, in _call_impl
192.168.1.38:     result = forward_call(*args, **kwargs)
192.168.1.38:   File "/usr/local/lib/python3.10/dist-packages/transformers/models/qwen2/modeling_qwen2.py", line 1008, in forward
192.168.1.38:     raise ValueError(
192.168.1.38: ValueError: You are attempting to perform batched generation with padding_side='right' this may lead to unexpected behaviour for Flash Attention version of Qwen2. Make sure to  call `tokenizer.padding_side  = 'left'` before tokenizing the input.

Before submitting

• [✔] Did you read the contributor guideline?

[hiyouga]

Llama Factory does not support left padding in training, we have posted a patch to fix this problem in 4a6ca62, thank you for pointing it out.