Yu-won Lee comments

Results 230 comments of


                                            Yu-won Lee

Inference python scripts

From the latest prints: * **Before PEFT:** `get_input_embeddings().weight.shape == torch.Size([0])` and same for output — so the **base checkpoint already has empty embeddings**. * After loading the adapter, shapes remain...

Inference python scripts

The code was for `transformers==4.51.3` and I haven't tested with other versions. I'll make an update for it soon.

How can I integrate external models (e.g., SentenceTransformer) into reward_funcs.py without DeepSpeed conflicts?

I'm not really familiar with GRPO, so I'll check about this one. Sorry for the issue.

merge_lora.sh killed in Qwen2-7B-Instruct

1. I will found out what is the reason for the performance. 2. I've finetuned the model but it worked for me before, I will test this againg too. 3....

Got stuck while installing the Flash-attn package

Thanks for letting me know. When I was making this repo 2.5.8 was the stable one for me. If this works then I'll update the code and the env. Thanks!

Got stuck while installing the Flash-attn package

It looks like it's okay for using the latest flash-attn.

ValueError: Image features and image tokens do not match: tokens: 259, features 256

I think something is mismatched. Can you give a sample for your dataset that I could check for this one?

ValueError: Image features and image tokens do not match [in GROPO finetuning]

https://github.com/2U1/Qwen2-VL-Finetune/issues/152#issuecomment-2999215309 This may be the issue.

ValueError: Image features and image tokens do not match [in GROPO finetuning]

Or increase the ` --max_completion_length` options.

Training with text-only instruction data

Yes you could just erase the "image".