llama.cpp
[Speculative decoding] feat: add EAGLE3 speculative decoding support
#18039

Merged

[Speculative decoding] feat: add EAGLE3 speculative decoding support #18039

ggerganov merged 28 commits into ggml-org:master from ruixiang63:eagle3-adapt-new-arch

github-actions added model

github-actions added examples

github-actions added python

github-actions added ggml

ggerganov commented on 2025-12-15

ggerganov commented on 2025-12-17

pwilkin added hot

ngxson commented on 2026-01-15

github-actions added server

ruixiang63 force pushed from 4567954a 179 days ago

ruixiang63 force pushed to 4567954a 179 days ago

ruixiang63 force pushed from 4567954a to 5eafd448 179 days ago

ruixiang63 force pushed from 5eafd448 177 days ago

ruixiang63 force pushed to e2ad5426 175 days ago

ruixiang63 force pushed 170 days ago

ruixiang63 force pushed 164 days ago

ggerganov assigned

ggerganov 162 days ago

ggerganov commented on 2026-06-05

ruixiang63 force pushed 161 days ago

ggerganov commented on 2026-06-08

llama : enable layer input extraction

0806a964

spec: support eagle3

800494f8

eagle3: fix params bug

16e65554

eagle3: support Gemma4 eagle3 from RedHatAI

752bf233

eagle3: set sync when get features from target

b32d9ebe

eagle3 : fix ubatch handling in embd_layer_inp extraction and encoder

7c5f428d

eagle3: adapt to upstream changes

91b9cfc7

eagle3: fix rebase issues and adapt to upstream changes

4ca8087b

ruixiang63 force pushed to 4ca8087b 158 days ago

eagle3:exclude the eagle3 arch from test-llama-archs

413c16da

eagle3: fix editorconfig check failures

6c212225

github-actions added testing

ggerganov commented on 2026-06-08

eagle3: fix multi-seq issue in d2t vocab mapping

ac7e2b2f

cont : minor style / clean-up

544aaa2f

Merge branch 'master' into pr/18039

5738c9a4

ggerganov commented on 2026-06-10

spec : remove `common_speculative_setup_draft_model()`

b9f41d18

llama : clean-up unused API

f3fbbedf

eagle3: set d2t vocab mapping in decode graph

8002c4ce

cont : assert layer inputs are configured

33b02df4

hparams : use n_embd_inp instead of n_embd_target_features

78572218

eagle3: make output.weight optional and inherit from target model whe…

9b2543d0

haparams : generic norm-before-residual param

1d55316e

llama-ext : consistent names

2de116b2

cont : fix

f4088797

hparams : remove target_hidden_size

d3732331

cparams : rename output_layer_inp -> embeddings_layer_inp

5caedbcd

arch : reuse ATTN_NORM_2 instead of adding new hidden norm

0274f0fc

ggerganov marked this pull request as ready for review 155 days ago

ggerganov requested a review from

JohannesGaessler 155 days ago

ggerganov requested a review from

CISC 155 days ago

ggerganov requested a review 155 days ago

ggerganov marked this pull request as draft 155 days ago

llama : clean-up names

9baa68be

cont : add assert + comment

0bd54498

ggerganov marked this pull request as ready for review 155 days ago

bartowski1182 commented on 2026-06-11

pwilkin requested changes on 2026-06-11

CISC commented on 2026-06-11

pwilkin approved these changes on 2026-06-11

CISC approved these changes on 2026-06-11

Update conversion/llama.py

7c42aff5

CISC approved these changes on 2026-06-11

ggerganov approved these changes on 2026-06-12

ggerganov merged 88a39274 into master 154 days ago

Reviewers

ggerganov

CISC

pwilkin

bartowski1182

ngxson

jeffbolznv

JohannesGaessler

Assignees

ggerganov

Labels

model testing examples python server ggml hot

Milestone

No milestone

llama.cpp [Speculative decoding] feat: add EAGLE3 speculative decoding support #18039 Merged

[Speculative decoding] feat: add EAGLE3 speculative decoding support #18039

llama.cpp
[Speculative decoding] feat: add EAGLE3 speculative decoding support
#18039

Merged