Prince Canuma comments

Results 572 comments of


                                            Prince Canuma

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

Btw, since you are doing a overhaul. I think we need to get closer to the PEFT package API because more multimodal models are using task loras such as Phi-4-mm:...

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

Thanks @Goekdeniz-Guelmez, you rock and ship crazy fast! I'm testing it and will give you feedback. Feel free to ping me if you don't hear anything in the next 24h.

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

Hey @Goekdeniz-Guelmez This is what I got: ![Screenshot 2025-03-24 at 11 53 25 PM](https://github.com/user-attachments/assets/c170a539-1fab-450b-b534-b110c1370e6d) I noticed the loss went up and ran out of memory (96GB M3 max), did this...

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

Hey @Goekdeniz-Guelmez Just took it for a spin again and it works really well! I have been running it for good solid 30 min, the loss has converged to arround...

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

@Goekdeniz-Guelmez I got this error towards the end. The error is because a particular sample is too small. Can we add a skip such samples if they show up instead...

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

@Goekdeniz-Guelmez awesome, this looks good to me. Thank you very very much, for the amazing work and speed!❤️ What's missing? and what's next here? (GRPO maybe?)

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

Yes, absolutely! All of those are much needed and welcome ❤️

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

Of course, just move it to ready and we can get started with the review.

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

Hey Goekdeniz Hope you got your deserved rest, Any updates?

overhaul and adding full weight fine-tuning (Quantized training, GRPO, DPO, ORPO)

Awesome, enjoy the holiday and time with family this can wait 💪🏾 I wish I was in your shoes