Yu-won Lee comments

Results 230 comments of


                                            Yu-won Lee

Fail to merge lora of Qwen2.5-VL

@Helen1p If you didn't change the config for the model much you could just download the `config.json` from the model's repo and use it.

Dataset Format Support

I could make it but, lot of multi-modal dataset uses the format from llava so I'm using the same format. It's better for ask the gpt for the conversion code.

Script about qlora.

@R3xpook Thanks for the comment. @yeppp27 That's right. You can add bits too use qlora.

Script about qlora.

@yeppp27 Sorry, I forgot to tell you. You should use zero2 for QLoRA.

Script about qlora.

I haven't tested 72b model, that I have no resource for it. You could decrease the `pixel_nums` or unset `vision_lr` and `projector_lr`.

Missing reference picture, default is 65552 and mmco: unref short failure

That looks like some kind of video file type warning. I'm not sure with it.

Any pointers as to how I could integrate TRL GRPO into the mix?

`[rank0]: TypeError: GRPOTrainer.__init__() got an unexpected keyword argument 'data_collator'` Looks like this line shows that GRPO trainer dosen't accept data_collator. I haven't used GRPO so, I need to take a...

Any pointers as to how I could integrate TRL GRPO into the mix?

In the line, it says it does not use the data collator so, it may work with just crossing out the collator. ``` # Data collator def data_collator(features): # No...

Any pointers as to how I could integrate TRL GRPO into the mix?

Thanks for letting me know. I'll look in to it and implement some prefernce optimization techinques to this repo.

Any pointers as to how I could integrate TRL GRPO into the mix?

Now I've got some time to work on so I'll start trying on too!