Sync master with upstream release v0.2.0 - #628

Open

jan-service-account wants to merge 211 commits into

devfrom

update-dev-from-master-2026-08-22-00-21

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
mtmd, common: various fixes (#27071)
ngxson
authored
6fed9f6
model : add support for MiniMaxText01ForCausalLM and MiniMaxM1ForCausalLM (#27018)

authored
16d222f
mtmd: fix Granite4 Vision image sequence assembly (#26653)

authored
9d57ce4

Commits on Aug 15, 2026

Fixed gating logic for problematic Intel driver version

rillomas
authored and
ggerganov
committed
7b38cb7
fixed indent

rillomas
authored and
ggerganov
committed
6b4344e
common: migrate the deprecated --mmap/--no-mmap to --load-mode (#26934)
fboudra
authored
0177dcc
vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat1 on Intel Xe (#25380)
fish-jiang
authored
9b0a2ce
fix: check gguf array type before reading (#27075)
ngxson
authored
27df919
common: support --models-dir loading MTP assistant models (#24431)

EZForever
and
ngxson
authored
5f754ea
vendor : update cpp-httplib to 0.53.1 (#27103)
cabelo
authored
77140d2
vendor: update BoringSSL to 0.20260813.0 (#27099)
cabelo
authored
adb55e5
server: re-design yield_to_queue thread model (#27133)
ngxson
authored
22b8e31
model: add Kimi-K3 text model (#26185)

authored
ad1de39
ui: read structuredContent from MCP tool result when content is empty (#26691)
Gautam0507
authored
0d9ceae
ui: mask API Key field in settings and error splash to stop browser a… (#26562)

crowmoed
and
ServeurpersoCom
authored
ece963f

Commits on Aug 16, 2026

Commits on Aug 17, 2026

[SYCL] support OP OPT_STEP_ADAMW, OPT_STEP_SGD (#25268)

arthw
and
NeoZhangJianyu
authored
37a215c
sycl: fix thread/block count in quantized cpy kernel launches (#27160)
Titaniumtown
authored
f275595
llama-bench: fix deprecation warnings missing trailing newline (#27179)
fboudra
authored
4695f00
ggml : bump version to 0.20.1 (ggml/1587)
ggerganov
committed
cea66f4
sync : ggml
ggerganov
committed
4197155
model : BailingMoE3 Support (#26608)

aetherbird
and
Dyluhn
authored
3733366
convert: add @ModelBase.example (#27208)
ngxson
authored
fa88ae9
ci : make release workflows use a deploy key (#27229)
ggerganov
authored
f9779dd
ci : allow make-release to target a specific commit (#27234)
ggerganov
authored
7c35571
mtmd: harden preprocessor_granite (#27235)
ngxson
authored
9f0d017
ci : restore release.yml check during make-release.yml (#27247)
ggerganov
authored
d83f72d
ui: add browser get_info tool (#27251)

ngxson
and
allozaur
authored
7077abb
model: support speculators-format checkpoints for DSpark (#26275)

authored
9cd719a
ci : reduce builds in build-xcframework.sh (#27252)
ggerganov
authored
805984d
ui: move get_datetime tool to frontend (#27255)

ngxson
and
ServeurpersoCom
authored
666f889
ci : push release tag explicitly in release.yml (#27261)
ggerganov
authored
34af94c
vendor: move hash to vendor (#27262)
ngxson
authored
39be55c
cuda : skip UMA override for HIP builds (#27083)
superm1
authored
60eeeb6
mtmd : skip thumbnail for non-tiled LFM2 images (#27246)

tdakhran
and
ngxson
authored
b75ecd1
vocab : support integer tokenizer scores (#27260)
CISC
authored
d8df12e
mtmd: use sha256 for input hashing (#27274)
ngxson
authored
ed1c3a2
server: save processed mtmd chunks as placeholder (#27278)
ngxson
authored
533b182
doc: document MCP stdio servers and CORS defaults in the server README [no release] [no ci] (#26847)
ServeurpersoCom
authored
087f94d
ci: more optimizations (#26983)
netrunnereve
authored
058df67
ui: Refactor Built-In Tools naming (Server/Browser) (#27271)
allozaur
authored
0021a77
ui: enforce alphabetical enum member ordering (#27272)
ngxson
authored
01818e4

Commits on Aug 18, 2026

Commits on Aug 19, 2026

gguf-py : add size guards to GGUFReader (#27188)

oakkaya
and
CISC
authored
0329fcd
tests: skip test-unicode build on win32/BUILD_SHARED_LIBS (#27336)
jeffbolznv
authored
79fe799
ci : add attestation for signed release artifacts (#25933)
quartzjer
authored
0adcc3b
ggml-webgpu: add mulmat with overlapping src0/src1 (e.g., for minimax-01) (#27321)
yomaytk
authored
5112b97
vulkan: tiled transpose for 0<->2 permuted CONT (#26585)

kh0pper
and
Kevin Hopper
authored
98d1e92
server: add dedup-cache-models preset option (#27346)
ngxson
authored
ee4c505
metal : dequantize q8_0 using packed types (#27370)
ggerganov
authored
8ef78e6
mtmd: add mtmd_bitmap_set_mergeable (#27348)
ngxson
authored
95c409c
ui: read persisted settings before the API key probe (#27365)
ServeurpersoCom
authored
77acca4
ggml: add ggml_rope_set_offset (+ metal support) (#27120)
ngxson
authored
fe8156f
ci : re-enable release dependency for sycl (#27385)
CISC
authored
645ca28
models : remove duplicate metadata load (#27378)
CISC
authored
2e92ecd
ci : add release attestation url (#27389)
CISC
authored
01ac3ad
sycl: report zero devices instead of aborting when the host has none (#27291)
Titaniumtown
authored
6cc504a
model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)

gabe-l-hart
and
CISC
authored
7221e24
Revert "common: share thread pools when `n_threads` differ (#27138)" (#27337)

nikwen
and
max-krasnyansky
authored
3e73446
vulkan: add null checks in ggml_vk_queue_command_pools_cleanup (#27353)
jetsontan
authored
b95502b
vulkan : dequant q8_0 KV once in coopmat1 (#25494)
Nathanw1014
authored
dc72703
server : make models endpoints private when authentication is enabled (#26347)
s0mecode
authored
ee0ea03
server: refactor sleep handling, allow access /metrics during sleep (#27376)
ngxson
authored
947fd9b
ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE (#26860)
ServeurpersoCom
authored
cd644c3
opencl: port fused ssm_scan kernel (Mamba-2, d_state in {128, 256}) to GPU (#26439)

wanghqc
and
lhez
authored
b062ba7
hexagon: fix FA HMX queue ordering and pack the rescale D matrices (#27042)
njsyw1997
authored
990e3bf
tensor-split meta backend fixes (#26502)
max-krasnyansky
authored
d59d455

Commits on Aug 20, 2026

Commits on Aug 21, 2026

docs: fix typos in ET.md (#27457)
itsvedantkumar
authored
a298422
TP: enable tensor split for LFM2/LFM2MOE (#26993)
cdanis
authored
b2e5e9b
sycl : fix load model with mlock issue (#27250)
arthw
authored
9e96cf7
sycl: fix multiple warnings in compiling sycl backend (#26713)
icfaust
authored
6602dd3
sycl: Update gate logic for Alchemist GPUs regarding OneDNN features. (#26635)

HumerousGorgon
and
arthw
authored
1cb3f5e
opencl: keep the vocab-scale K-quant lm_head on the CPU for Adreno A7X (compiler issue workaround) (#26440)

wanghqc
and
lhez
authored
cd26896
sycl : Add Q5_K ESIMD kernel (#26376)
malsbat
authored
9e89a19
test : make the FA V-is-view-of-K case a test case parameter (#27394)
ggerganov
authored
5fff128
sycl : add Q2_K reordered MMVQ and ESIMD kernels (#26336)
malsbat
authored
ff14356
kleidiai : add SME2 F32 GEMV kernel support (#26891)
chaxu01
authored
62b2269
ci : release clean-up (#27477)
ggerganov
authored
1719747
ci : add nightly-tag.txt to make-release (#27485)
ggerganov
authored
e467c2f
ui: Settings navigation cleanup (#27241)
allozaur
authored
5b6ddc9
Revert "sycl : add Q2_K reordered MMVQ and ESIMD kernels (#26336)" (#27486)
ggerganov
authored
7a0e42f
ci : remove duplicate flag (#27488)
ggerganov
authored
353b32d
ggml : bump version to 0.21.0 (ggml/1597)
ggerganov
committed
01ff204
sync : ggml
ggerganov
committed
5de25a7
scripts : add release.sh for release preparation (#27497)
ggerganov
authored
c4b0225
llama.cpp : bump version to 0.2.0 (#27498)
ggerganov
authored
bb4caa7