Are Eagle and LoRa compatible?

[dongyibo]

I have opened both Eagle and multiple LoRa paths at the same time, and I noticed a significant difference compared to when Eagle is not open.

[Fridge003]

They are not compatible by now.

[shuaills]

Hi, please check sgl-project/SpecForge#138.
We are releasing an EAGLE3-LoRA training pipeline and also its inference code. Stay tuned!

[Fridge003]

Hi, please check sgl-project/SpecForge#138. We are releasing an EAGLE3-LoRA training pipeline and also its inference code. Stay tuned!

Thank you @shuaills!
I have a small question: do we need to train different EAGLE3 weights for each combination of lora adaptor and base model?

[shuaills]

Hi, please check sgl-project/SpecForge#138. We are releasing an EAGLE3-LoRA training pipeline and also its inference code. Stay tuned!

Thank you @shuaills! I have a small question: do we need to train different EAGLE3 weights for each combination of lora adaptor and base model?

No, this is for multi-LoRa, single draft serving workload.

[Fridge003]

Hi, please check sgl-project/SpecForge#138. We are releasing an EAGLE3-LoRA training pipeline and also its inference code. Stay tuned!

Thank you @shuaills! I have a small question: do we need to train different EAGLE3 weights for each combination of lora adaptor and base model?

No, this is for multi-LoRa, single draft serving workload.

So the same base model will use the same draft model, no matter which lora adaptor it uses. Is that right?

[shuaills]

Hi, please check sgl-project/SpecForge#138. We are releasing an EAGLE3-LoRA training pipeline and also its inference code. Stay tuned!

Thank you @shuaills! I have a small question: do we need to train different EAGLE3 weights for each combination of lora adaptor and base model?

No, this is for multi-LoRa, single draft serving workload.

So the same base model will use the same draft model, no matter which lora adaptor it uses. Is that right?

YES!

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen it if needed.