Yu-won Lee

@koookieee Yes you can use that, but you need to add a config.json file into the directory. I'm not sure why now the config.json file isn't save but I'll add...

@koookieee Yes the base model config. This repository isn’t backed by a team—it’s a personal side-project I tinker with in my free time. I started it because there wasn’t any...

@xqfJohn Most of the case the vocab isn't modified if yoiu haven't added or removed the token. If you haven't changed other things, you could just copy the `vocab.json` and...

Yes it basically input_ids == labels but, the user input should be masked to -100.

@natsunoshion @lucasjinreal Maybe it could be avoided with giviing option to processor I think. https://github.com/2U1/Qwen2-VL-Finetune/blob/1e8c37a8839c658d6f7ce4256ecab3ff0dd47943/src/train/data.py#L206 Adding `do_rescale=False` option in this line.

@natsunoshion Oh, I thought your data has already rescaled so, it causes the warninig every step.

Current version of llama.cpp has disalbed to run clip on gpu. However, I don't know why it still runs on cpu when I've installed the older version.

For people who need Qwen2-VL/Qwne2.5-VL for sequece classification, I'll share my code here. https://github.com/2U1/Qwen2-VL-Finetune?tab=readme-ov-file#classification-finetuning I've tailored the model for classification task and added some other loss fuction for it. Feedbacks...

Hi @Lsjay3, thanks for the interest in my work. vLLM’s paged‑attention mainly speeds up *decoding*, but this model is used for **classification only**, so there’s almost no win. Prefill may...

@Lsjay3 If you mean `torch.compile` and `flash-attention2`, it does support it. However `torch.compile` would has compile each time when the batch size changes.