Gemma-4 PLE/MoE VLM support - #495

Open

arielge wants to merge 12 commits into

mainfrom

gemma4_vlm_ple_moe

Commits on Sep 10, 2026

initial Gemma-4 PLE/MoE VLM support (with CPU tower)

arielge
and
GPT 5.6
committed
1bf391f
add vision tower adapter

arielge
and
GPT 5.6
committed
92f3627
correctness fixes

arielge
and
GPT 5.6
committed
0c3f549
refactor Gemma 4 models to reduce compilations

arielge
and
GPT 5.6
committed
0a2d1a4
remove redundant stored variables

arielge
and
GPT 5.6
committed
9939db6
various fixes

arielge
and
GPT 5.6
committed
8a5fb91
remove unused forward methods + unified state logic

arielge
and
GPT 5.6
committed
06616a4
fix image slots issue

arielge
and
GPT 5.6
committed
f031e44
unify block classes

arielge
and
GPT 5.6
committed
07935fa
remove double cpu transfer
arielge
committed
bdf43b1
reduce duplication in token-to-pad substitution

arielge
and
GPT 5.6
committed
baf5d8a
add output clamp assert

arielge
and
GPT 5.6
committed
930f21a