Qwen Image incorrect device assignment during prompt encode

[vladmandic]

Describe the bug

In the new QwenImagePipeline method _get_qwen_prompt_embeds it does following:

        txt_tokens = self.tokenizer(
            txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt"
        ).to(self.device)

i assume this is a typo as it doesn't use actual device which is passed as function param and as such it breaks with any type of offloading with device mismatch since self.device defaults to entire model placement - which doesnt match.

simple solution is to change last part from .to(self.device) to .to(device)

Reproduction

N/A

Logs

RuntimeError: Expected all tensors to be on the same device, but got index is on cpu, different from other tensors on cuda:0 (when checking argument in method wrapper_CUDA__index_select)

System Info

diffusers==7ea065c5070a5278259e6f1effa9dccea232e62a

Who can help?

cc @naykun @yiyixuxu @sayakpaul @a-r-r-o-w

[vladmandic]

gentle ping? this is a blocker for qwen-image in sdnext and fix is trivial.