llama : add Gemma4 MTP #23398
am17an
merged
30 commits into
ggml-org:master
from
am17an:gemma4-mtp
garloff
commented
on 2026-06-02
llama: Gemma 4 MTP
f268966d
fix multi-seq
9af0434d
add assert that draft + shared kv should be on same device
7b87cd35
add Q rot when cache is quantized
27461cd8
add temp hack to not use fit with gemma4, rm later
777af6af
add exception in test-llama-archs
c0da00af
move assistant to separate file
dd97604f
ggerganov
commented
on 2026-06-05
add unified assistant
4eaa3cee
ggerganov
commented
on 2026-06-05
Merge branch 'master' into pr/23398
5954f196
cont : adjust to hparams changes
d78a3864
cont : avoid computations on the CPU
f0438b1b
Merge branch 'master' into pr/23398
65eef954
Merge remote-tracking branch 'origin/master' into pr/23398
2e6d351a
cont : clean-up
cb40363f
cont : clean-up
d2483639
cont : fix handling of unused tensors
88172738
cont : fix undefined
a85f38c5
fix typo
efd651a8
cont : enable gemma4 graph reuse
9bc26704
cont : fix assert
fd06e52f
cont : fix quantized cache
a2986416
cont : fix names
ca707bfa
cont : fix names
6b7e371d
cont : add reference for draft positions
5cf5efb1
ServeurpersoCom
approved these changes
on 2026-06-07
cont : fix multi-modality
22a56247
cont : add comment about ctx_src
dbd34b22
cont : clean-up server fit logic
b76904a8
cont : clean-up llama_context
6efaa0aa
py : fix names
aff138f8
cont : rename ctx_src -> ctx_other
6bc7f5af
ggerganov
approved these changes
on 2026-06-07
pwilkin
approved these changes
on 2026-06-07
am17an
commented
on 2026-06-07
Reviewers
Assignees
Labels
model
testing
examples
python
server
Milestone
No milestone
Login to write a write a comment.
Login via GitHub