QwenVL2.5 bfloat16 Type Issue

[qirui-chen]

In the code, the tensor is forcibly cast to float32, while cos and sin may be inconsistent with the tensor due to the model type being bfloat16.

This problem occurred when using the flash attention 2 & bf16 Qwen2.5VL model.

[sorenmc]

Looks like they are fixing it in this pr huggingface/transformers#35837

[marwankefah]

Until they accept the pull request, you can modify modeling_qwen2_5_vl.py in the Transformers library by changing the following line:

tensor_ = tensor.float()

to

tensor_ = tensor