Sync master with upstream release b10435 - #621

Open

jan-service-account wants to merge 80 commits into

devfrom

update-dev-from-master-2026-08-15-00-21

Commits on Aug 10, 2026

ci : target ROCm 7.14 for build and release (#25775)
superm1
authored
0666ad2
opencl: transpose the K tile in local memory for FA prefill kernels (#26428)
wanghqc
authored
689e227
Address review comment of PR 25532 (#26852)
gaugarg-nv
authored
030ebb5

Commits on Aug 11, 2026

ggml-webgpu: fix CI errors from #25025 and #25262 (#26566)
yomaytk
authored
84f7129
common/peg : suppress incomplete escape sequences (#26780)
aldehir
authored
48d22e2
model : fix SWA not being enabled for EXAONE 4.5 (#26848)
junmo-kim
authored
14e78dd
tests : disable backend sampler hip multi output (#26878)

jimw567
and
Jim Wu
authored
4801e3c
tests : clean-up server test, use `tests.sh` in ci (#26886)
ggerganov
authored
b3df572
llama: add default load-mode auto, which avoids mmap on iGPUs (#26081)

authored
153d324
tests : fix running server tests on windows (#26889)
ggerganov
authored
9afff1b
model-conversion : use save_output_data for causual embeddings [no ci] (#26890)
danbev
authored
1138b85
ci: hip-quality-check: update vgpr spill ignore list (#26859)
IMbackK
authored
7044859
ui: fix context gauge for single-model usage (#25738)
intel00000
authored
8d274dd
mtmd: support pocket-tts (#26871)

ngxson
and
ServeurpersoCom
authored
6e62ba5
Dflash support for nemotron-3.5 (#26905)

lnigam
and
CISC
authored
cc078b4
convert : keep quantization scales for nemotron --mtp export (#26903)
ynankani
authored
5d16e81
requirements: use stable torch packages on s390x (#26864)
nikwen
authored
2468576
imatrix.cpp: Move finite check and only check touched experts (#26861)
bartowski1182
authored
38406d5
ci : add windows-rocm to check-release (#26897)
CISC
authored
70dfba5
chat : tighten bare function parsing for Qwen models (#26793)
aldehir
authored
ba360ef
spec : update speculative-simple (#26904)
ggerganov
authored
f785fc9
cuda : add warp-per-row wkv7 kernel for single-token decode (#26111)
123123213weqw
authored
5988633
CUDA: only disable CUDA graphs when mul_mat_id actually needs a stream sync (#26802)
grafail
authored
ebb546b
ci : add missing release check (#26923)
CISC
authored
7b13a84
chat : fix muse-glimmer detection of tool calls after EOM (#26879)
ruanslv
authored
0b1bad1

Commits on Aug 12, 2026

opencl: use flat mv q5_k when weight exceeds image1d_buffer_t limit (#26880)
lhez
authored
cb27fe9
convert : handle per_layer_config in Gemma4 (transformers 5.15) (#26882)

pluvium27
and
CISC
authored
6eff593
wavtokenizer-dec : bound posnet/convnext block_count against n_layer_all (#26892)

oakkaya
and
CISC
authored
55f453b
vulkan: add TQ2_0 (ternary) support (#25850)

authored
a7cd2f0
tests : update speculative params (#26925)
ggerganov
authored
a4a4c51
opencl: default FA c8 cluster width to 16 on X1E (#26433)
wanghqc
authored
89e0aa6
ui: add read_media tool (#25877)

authored
4dd1275
server : support slot save/restore with media inputs (#26640)
CHIPMUNK-T0T
authored
5d9e5ac
cmake : add config version support (ggml/1582)

danbev
authored and
ggerganov
committed
13fd0bb
sync : ggml
ggerganov
committed
af05a42
model : disallow integer dflash sliding_window_pattern (#26900)
CISC
authored
ece98b8
kleidiai: Add runtime feature detection mechanism for aarch64/kleidiai (#26076)
JonathanC-ARM
authored
132753b
gguf : harden loader against malformed tensor dims and metadata types (#25596)

harrison001
and
ggerganov
authored
d8a8bea
cmake : introduce semantic versioning (#26839)

danbev
and
ggerganov
authored
680a9ae
disable rocm cache (#26962)
ggerganov
authored
7a9ff95
ci : disable ubuntu-rocm (#26969)
ggerganov
authored
9558fa4
ci: fix thread sanitizer + remove ccache (#26927)

netrunnereve
and
CISC
authored
84e908c
common: add system-level config file (#26118)

jcmdln
and
ngxson
authored
8e7f22b

Commits on Aug 13, 2026

ui: Constants refactor (#26908)
ServeurpersoCom
authored
e21152d
ggml : fix arm builds, unused var (#26991)
ggerganov
authored
1f368f3
refactor: Clean up UI types (#26909)
allozaur
authored
a6040c9
ui: Stores architecture improvements (#26910)
allozaur
authored
094e53d
ui: Move `styles/` to `$lib` scope (#26950)
allozaur
authored
f2efd64
ui: Clean up contexts, remove prop drilling from Chat Form Actions (#26951)
allozaur
authored
d86c7d6
ggml-hip : remove -funsafe-math-optimizations (#26696)

jimw567
and
Jim Wu
authored
e79e4bf
server: refactor + correctness fixes for metrics (#26920)
ngxson
authored
decaf50
sycl : enhance concat to support Q4_0, Q4_1, Q5_0, Q5_1, Q8_0 (#26800)
arthw
authored
d415e65
sycl : Add DMMV ESIMD Q3_K kernel (#26251)
malsbat
authored
8efbf65
sycl: fuse UNARY(silu|sigmoid|softplus) + MUL (#26411)
Titaniumtown
authored
1ee1cd9
sycl: remove separate fp32 type promotion in gemm non-oneDNN path (#26372)
icfaust
authored
154d57a
ggml-cpu/ops: vectorize flash-attention V-cache F16 to F32 conversion (#26947)

jinzihao
and
jinzihao
authored
eeae28b
spec: enable backend sampling for both dflash & dspark (#26958)

ruixiang63
and
ggerganov
authored
0d0bfcd
common : auto-detect spec type from draft GGUF metadata (#26814)
aic0d3r
authored
f65e568
metal : add TQ2_0 support (#26980)
ggerganov
authored
4a84b0a
spec : auto-detect mtp draft model type (#27005)
CISC
authored
1d2869c
server : serve index.html with no-cache (#27006)
erusev
authored
981184e
chat : fix LFM2 tool call arg name prefix ambiguity (#26960)
aldehir
authored
2606220
[SYCL] Support host pinned mem to improve SYCL Host-to-Device Memory Access (#26789)
arthw
authored
a97123e
OpenVINO: Qwen3.5, memory optimization, and test-recurrent-state-rollback (#26952)

authored
aee56b3
ui: fix VITE_PUBLIC_SERVER variable reading (#24845)
brainrom
authored
9c5531e
refactor: Naming (#27001)
allozaur
authored
fa4ec45
ui: Refactor data-attrs constants, enum for bool strings (#27002)
ServeurpersoCom
authored
bdffafa
common: apply CPU parameters across tools (#27026)
nikwen
authored
a94d563

Commits on Aug 14, 2026

dflash : clarify output logging of target_layer_ids (#27013)
danbev
authored
2bacf9e
sycl: fuse the gated-delta-net state writeback cpy (#26643)
Titaniumtown
authored
3d93885
ggml: force single thread on wasi (#25686)
MendyBerger
authored
c6f6a92
sycl: fuse mul_mat(gate) + mul_mat(up) + GLU for q4_K dense FFN (#26779)
Titaniumtown
authored
6509138
tests : replace personal home directory paths with generic placeholders (#27043)

jimw567
and
Jim Wu
authored
885c5bb
server: allow accessing /metrics and /slots during llama_decode() (#27041)
ngxson
authored
77918ca
llama : allow virtual igpu devices (#26953)
ggerganov
authored
4c1a0af
ggml : recurrent state rollback for ggml_ssm_scan (#26623)

authored
1692f9e
ggml : bump version to 0.20.0 (ggml/1584)
ggerganov
committed
06ae232
sync : ggml
ggerganov
committed
9b05354
chat : pass reasoning_effort to template

sobakasu
and
aldehir
authored
7e4c0a9
jinja : fix quadratic cost in gather_string_parts (#27034)
Yunzez
authored
9e40df6