Feature Request: Support EAGLE3 models for draft model / speculative decoding use cases

[roykim98]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

I've been reading about some success regarding Eagle-3 for speculative decoding. Model architectures like LlamaForCausalLMEagle3 are not supported for converting to GGUF to be used as draft models.

This ask is for llama.cpp to support converting these model architectures to GGUF and have them be able to be used as draft models.

Motivation

https://huggingface.co/baseten-admin/EAGLE3-gpt-oss-120b-bf16 has an EAGLE-3 model for GPT-OSS (seems to be 885M parameters to gpt-oss-120b's ~5B) that could be used a draft model. GPT-OSS has demonstrated itself to be fairly capable for agentic use-cases with recent updates, but the prompt processing / inference speeds for localized compute are just shy of fast.

Possible Implementation

No response

[whatever1983]

yeah, definitely. Llama.cpp is so behind SGlang and TensorRT in those areas, it is crazy. eagle3 is definitely better than the "Qwen2.5 0.5B" transplant-vocab draft models it is not even funny.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.

[nullnuller]

Just bumping it for interest

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.