Question: Any timeline for open-sourcing the training recipe?

[01lin]

Question

The README mentions:

"we will also open-source the training recipe soon, so you can train your own DFlash draft model to accelerate any LLM."

I'm very interested in training custom DFlash draft models. Could you share more details?

1. Is there a rough timeline for releasing the training recipe?
2. What will the training recipe include? (e.g., data format, training scripts, hyperparameter configs, hardware requirements)
3. Will it support training draft models for arbitrary base LLMs, or only specific architectures?

Being able to train custom draft models would significantly expand DFlash's applicability. Thanks for the great work!

Context

• Use case: Want to train a DFlash draft model for a custom fine-tuned LLM
• Backend: vLLM / vLLM-ascend

[NotMainstream]

I came here to ask about the same thing. There are so many people posting daily different custom distills with improved performance and it is impossible for you guys to track them all and create draft models for them. It would help so much the community if you release the recipe sooner and other developers can try to improve on it if possible.

[HyokaChen]

同样期待一波源码，有一些微调模型想要自定义训练一下 dflash，用微调模型的基座模型，试下来命中率只有个位数～～

[dongyibo]

Looking forward to it