The first framework that can post train DeepSeek V4-pro on a single-node?
7 comments
This is so good. Is it kinda similar in concept to QLoRA training?
Hi, thank you so much :)
Yes, you are very sharp, to make truely it scalable, we have to finetune quantized LLMs directly, similar ideas as QLoRA. The algorithmic innovation is based on our research about orthogonal finetuning OFT and QOFT, which has demonstrated better training stability when finetuning quantized base models in such a scale.
Thank you for the information, creating this and releasing the research behind it. Truly appreciate.
We just opensourced a project called Orbit, which can RL post train trillion scale LLMs like deepseek v4. We found it pretty cool! Looking forward to some comments!
- Tim
Blog post:
https://spherelab.ai/orbit/