feat(sync): upstream master sync (42 commits) + Gemma4 MTP via PR #23398 vendor#93

Merged

marksverdhei merged 72 commits into

htfrom

feat/gemma4-mtp-vendor

Jun 7, 2026

Commits on Jun 4, 2026

build : use umbrella Headers directory for XCFramework module map (#23974)
gmarzjr
authored
webui: fix tool selector toggle/counter, key tools by stable identity (#24065)
ServeurpersoCom
authored
agents: refactor, include more guidelines (#24111)
ngxson
authored
server: avoid unnecessary checkpoint restore when new tokens are present (#24110)

Abioy
and
ggerganov
authored
ggml: vectorize ggml_vec_dot_q4_1_q8_1 with WASM SIMD128 (#22209)
sirohikartik
authored
convert: Fix Gemma 4 Unified conversion (#24118)
pcuenca
authored
return filter to save memory (#24125)

forforever73
and
lvyichen
authored
ui: added single line reasoning preview (#23601)

gugugiyu
and
allozaur
authored
ui: Fixed packages (#24119)
allozaur
authored
Move duplicated imatrix code into single common imatrix-loader.cpp (#22445)
bartowski1182
authored
webui: [a11y] fix keyboard navigation issues in chat interface and sidebar (#23132)
vignesh191
authored
arg: fix double mtp downloads (#24128)
ngxson
authored
server : disable on-device spec checkpoints (#24108)
ggerganov
authored

Commits on Jun 5, 2026

Commits on Jun 6, 2026

Commits on Jun 7, 2026

docker : bump cuda13 to 13.3.0 (#24228)
CISC
authored
convert : fix Gemma4 with no audio encoder (#24242)
CISC
authored
arg: Skip mmproj download when user supplied mmproj (#24239)
konradmb
authored
chore(sync): merge upstream master into ht (42 commits, 006640408..465b1f0e7)
marksverdhei
committed
chore(sync): adapt DFlash to hparams.n_layer() method post-#24060
marksverdhei
committed
llama: Gemma 4 MTP

am17an
authored and
marksverdhei
committed
fix multi-seq

am17an
authored and
marksverdhei
committed
add assert that draft + shared kv should be on same device

am17an
authored and
marksverdhei
committed
add Q rot when cache is quantized

am17an
authored and
marksverdhei
committed
add temp hack to not use fit with gemma4, rm later

am17an
authored and
marksverdhei
committed
add exception in test-llama-archs

am17an
authored and
marksverdhei
committed
move assistant to separate file

am17an
authored and
marksverdhei
committed
add unified assistant

am17an
authored and
marksverdhei
committed
cont : adjust to hparams changes

ggerganov
authored and
marksverdhei
committed
cont : avoid computations on the CPU

ggerganov
authored and
marksverdhei
committed
cont : clean-up

ggerganov
authored and
marksverdhei
committed
cont : clean-up

ggerganov
authored and
marksverdhei
committed
cont : fix handling of unused tensors

ggerganov
authored and
marksverdhei
committed
cont : fix undefined

ggerganov
authored and
marksverdhei
committed
fix typo

CISC
authored and
marksverdhei
committed
cont : enable gemma4 graph reuse

ggerganov
authored and
marksverdhei
committed
cont : fix assert

ggerganov
authored and
marksverdhei
committed
cont : fix quantized cache

ggerganov
authored and
marksverdhei
committed
cont : fix names

ggerganov
authored and
marksverdhei
committed
cont : fix names

ggerganov
authored and
marksverdhei
committed
cont : add reference for draft positions

ggerganov
authored and
marksverdhei
committed
cont : fix multi-modality

ggerganov
authored and
marksverdhei
committed
cont : add comment about ctx_src

ggerganov
authored and
marksverdhei
committed
cont : clean-up server fit logic

ggerganov
authored and
marksverdhei
committed
cont : clean-up llama_context

ggerganov
authored and
marksverdhei
committed
py : fix names

ggerganov
authored and
marksverdhei
committed
cont : rename ctx_src -> ctx_other

ggerganov
authored and
marksverdhei
committed
chore(sync): drop intermediate llama_set_mtp_source call
marksverdhei
committed