gopalsarda

@rkooo567 Thanks for the great work, looking forward to this PR getting merged. I wanted to try using EAGLE for target models with tp > 1 (mixtral 8x7b). I was...

> This is in progress with @WoosukKwon! We are planning to have a bigger refactoring that makes all arch SPMD based! @WoosukKwon That is great! If possible, can you please...

> the acceptance rate changes based on the number of spec token, which is not expected. Acceptance rate should not be affected by the propose length Maybe I am missing...

Thanks for looking into this. Updating `attention_backend` to `fa3` in [ServerArgs](https://github.com/sgl-project/SpecForge/blob/main/specforge/modeling/target/eagle3_target_model.py#L191) is still giving OOM. ``` [rank4]: Traceback (most recent call last): [rank4]: File "/mnt/git/SpecForge/scripts/train_eagle3.py", line 775, in [rank4]: main()...

Atleast for gpt-oss-120b, using `--target-model-backend hf` did not work. But it can maybe work for a smaller model.

@zyksir Sorry for tagging directly. Just wondering if you noticed anything like this during the development of the feature in https://github.com/sgl-project/SpecForge/pull/239