-
58cb9138e4
vulkan : check for null vkEnumerateInstanceVersion (#29872)
Ian McKellar
2026-10-06 06:28:54 -07:00
-
4f54067615
HIP: use -O0 for host code in debug builds (#29795)
Johannes Gäßler
2026-10-06 13:53:00 +02:00
-
f0c41e0168
models : consolidate nextn row cropping into shared helpers (#30017)
Georgi Gerganov
2026-10-06 14:08:54 +03:00
-
c07242a989
qwen35: pass transaction state view
xrip
2026-10-06 12:46:38 +03:00
-
d7a695ef67
scripts : limit apiabi checks to libllama and libmtmd (#30038)
Daniel Bevenius
2026-10-06 11:42:59 +02:00
-
c73f1401e5
build: fix stale graph and batch API uses
xrip
2026-10-06 09:38:18 +00:00
-
7d6656d788
Merge branch 'ggml-org:master' into master
Ilya Maslennikov
2026-10-06 12:09:06 +03:00
-
-
6c73b3e12d
convert : add text_config as fallback [transformers 5.18] (#30040)
Kartik GuliaandSigbjørn Skjæret
2026-10-06 14:34:26 +05:30
-
411bf3c608
Merge upstream master
xrip
2026-10-06 12:03:55 +03:00
-
-
1a3011cc0c
llama : re-reserve the sched when the nextn extraction flags change (#30020)
Piotr Wilkin (ilintar)
2026-10-06 10:41:23 +02:00
-
6753a033f0
ggml: refactor selective expert copying to user code (#29943)
Aman GuptaandGeorgi Gerganov
2026-10-06 13:29:45 +05:30
-
cbb7d52ecb
test-llama-archs : initialize backends before generating models (#30034)
Ihar Hrachyshka
2026-10-06 03:49:52 -04:00
-
63bef2728d
vendor : update LibreSSL to 4.3.3 [no ci] (#30019)
Alessandro de Oliveira Faria (A.K.A.CABELO)
2026-10-06 04:14:45 -03:00
-
b9a5a00b86
ggml-openvino: fix CI tests; fix GPU regressions. (#30037)
Ravi PanchumarthyandMustafa Cavus
2026-10-05 23:57:04 -07:00
-
43fe9c6428
llama: fix k-pool scatter data race on shared sequences (#29994)
Pascal
2026-10-06 07:41:57 +02:00
-
5e03bdd870
hexagon: ssm-conv updates (#29971)
Todor BoinovskiandMax Krasnyansky
2026-10-05 17:42:40 -07:00
-
50569eb87d
hexagon: add pool op support (#29995)
Aparna M PandMax Krasnyansky
2026-10-06 04:04:32 +05:30
-
7049ff0cbe
ci : add 1accel label [no ci] (#30016)
Sigbjørn Skjæret
2026-10-05 21:52:20 +02:00
-
c250304960
ci : skip container re-tagging when Require Docker is disabled (#30008)
Georgi Gerganov
2026-10-05 21:22:08 +03:00
-
8345f33395
hexagon: matmul and flash-atten scalability updates (#29974)
2026-10-05 08:55:21 -07:00
-
d812350493
llama.cpp : bump version to 0.6.0 (#29997)
Georgi Gerganov
2026-10-05 18:13:51 +03:00
-
4d60b4d087
common, server : report model input/output modalities in GET /models (#29987)
Adrien Gallouët
2026-10-05 17:08:29 +02:00
-
c06f84160a
sync : ggml
Georgi Gerganov
2026-10-05 17:34:35 +03:00
-
f05c8b2780
ggml : bump version to 0.26.0 (ggml/1652)
Georgi Gerganov
2026-10-05 17:26:51 +03:00
-
e117148a41
CUDA: make the alloc_deps check batch independent (#29986)
Aman Gupta
2026-10-05 19:14:23 +05:30
-
6c59c40076
vulkan: fix Flash Attention shmem write out of bounds (#29988)
Ruben Ortlam
2026-10-05 15:15:00 +02:00
-
3c9e747f7e
vulkan: revert mul_mat_id tile selection PR #29182 (#29936)
virajwad
2026-10-05 06:11:21 -07:00
-
b809b886d9
cuda: use the vector lightning indexer kernel on MUSA (#29990)
Pascal
2026-10-05 15:09:53 +02:00
-
994e8f2222
ci : add "Require Docker" flag to make-release workflow (#29989)
Georgi Gerganov
2026-10-05 16:04:47 +03:00
-
9d853bb36a
webui: Use toLocaleString() format consistently across chat message statistics (#27990)
Giovanni Rivera
2026-10-05 05:57:41 -07:00
-
8f9ae20c86
ci : disable failing test on virtual Metal device (#29993)
Georgi Gerganov
2026-10-05 15:10:16 +03:00
-
9871df5911
server: support vision input for Clef (#29969)
Xuan-Son Nguyen
2026-10-05 14:02:40 +02:00
-
8b2fbaf32c
CUDA: Optimize accumulation in mmq for NVFP4 type (#29857)
Konrad Moren
2026-10-05 13:43:07 +02:00
-
2ed93db472
ci : disable unused qemu in docker build (#29984)
Sigbjørn Skjæret
2026-10-05 12:33:10 +02:00
-
8e1642198d
server: reject partial media truncation (#24076)
Yufeng HeandPascal
2026-10-05 17:22:23 +08:00
-
806eee9841
vulkan: fix stale prealloc_y reuse across flash attention and soft_max (#29591)
François-Xavier Gsell
2026-10-05 16:38:26 +08:00
-
b3daa077a5
vulkan: sparse flash attention for quantized K/V (#29639)
François-Xavier Gsell
2026-10-05 16:37:54 +08:00
-
c173a53bdf
llama : fix unexpected graph reallocation in the k-pool models (#29958)
Georgi GerganovandPascal
2026-10-05 11:36:25 +03:00
-
210791069b
kv-cache: fix restoring mismatched KV cache rotation by saving exact rotation metadata (#28498)
Evan HuusandGeorgi Gerganov
2026-10-05 04:27:30 -04:00
-
e5983d6704
ci : winget urls must be separate strings (#29978)
Sigbjørn Skjæret
2026-10-05 09:36:57 +02:00
-
4ca6b76f0b
ci : fix docker workflow permissions (#29979)
Sigbjørn Skjæret
2026-10-05 09:36:03 +02:00
-
9f12cd4a4c
ggml-cpu : add Q8_0 IME1 matrix kernel for SpacemiT X60 (#28479)
Alan Tseng
2026-10-05 15:34:15 +08:00
-
ebe18bee5a
vulkan : Fix undeclared identifiers when -DGGML_VULKAN_RUN_TESTS=ON (#29912)
Ed Addario
2026-10-05 08:18:55 +01:00
-
8216c84623
webgpu: add MMVQ support for Q1_0/Q5_0/Q5_1/Q3_K/Q5_K/Q6_K/MXFP4 (#29483)
Masashi Yoshimura
2026-10-05 15:56:45 +09:00
-
1b43d31169
cuda: tile the lightning indexer over keys and tokens for 4 heads (#29901)
Pascal
2026-10-05 08:39:47 +02:00
-
a3a1c4747f
metal : few-row MMA mat-mul (#29869)
pratiknarola-tandGeorgi Gerganov
2026-10-05 10:59:13 +05:30
-
9d3aba6b5e
CUDA: use MMVF for thin f16/bf16 mul_mat at small batch size (#29633)
ynankaniandJohannes Gäßler
2026-10-05 04:56:50 +00:00
-
d89651a7b2
CUDA: prefer whole-tile FlashAttention scheduling for efficient two-stage kernels (#29435)
anujj
2026-10-05 08:32:07 +05:30
-
a7fb71fab8
log, server: self contained colors, split child commands from logs in router mode (#29895)
PascalandXuan-Son Nguyen
2026-10-05 01:59:37 +02:00
-
0bb496dbd3
llama: support both embd + raw tokens in batch (#29622)
Xuan-Son Nguyen
2026-10-05 01:35:49 +02:00
-
2ca15f5404
CUDA: refactor swizzling code (#29612)
Johannes Gäßler
2026-10-04 22:48:30 +02:00
-
a7b94df2c6
ggml-cpu: support BF16/FP16/FP32 K tails in tinyBLAS on x86 (#29806)
SXX
2026-10-05 03:22:17 +08:00
-
0eb6d9a813
cuda : move neu_padded to where it is used (#29940)
Adrien Gallouët
2026-10-04 21:21:31 +02:00
-
2e7c58c547
ci : windows llvm build requires ninja multi-config (#29959)
Sigbjørn Skjæret
2026-10-04 19:28:24 +02:00
-
7f2dd88b0a
ci : add windows arm64 vulkan release (#29954)
Sigbjørn Skjæret
2026-10-04 18:46:15 +02:00
-
bf79dbbcd0
AGENTS.md : revamp (#29656)
Aman Gupta
2026-10-04 21:38:19 +05:30
-
dbe4c3ed42
chat-peg-parser : clear current_tool when pending_tool_call is reset (#29942)
Anas
2026-10-04 17:59:21 +02:00
-
46847e6158
ci : set default permissions (#29945)
Sigbjørn Skjæret
2026-10-04 16:13:26 +02:00
-
2bc5635734
cuda : move blocks_per_col to where it is used (#29939)
Adrien Gallouët
2026-10-04 15:47:21 +02:00
-
dd266785c2
CUDA: fix MMQ memory fault if n_expert >> n_ubatch (#29941)
Johannes Gäßler
2026-10-04 14:10:43 +02:00
-
16c163d561
vulkan: fix rdna4 mat_vec tuning (#29934)
Ruben Ortlam
2026-10-04 14:05:21 +02:00
-
0504396140
imatrix: calculate activation-based statistics for new format (GGUF) imatrices (#14891)
2026-10-04 10:24:53 +01:00
-
8330e96967
spec : fix n-gram drafts rejected at temp > 0 after truncation (#29924)
Pranesh GonegandlaandPranesh Gonegandla
2026-10-04 14:15:22 +05:30
-
6716df694b
common : prepare load_from_models_dir() for path conversion (#29674)
Adrien Gallouët
2026-10-04 10:35:33 +02:00
-
bf9a0ccce7
server : fix dead LLAMA_ARG_HF_REPO_FILE key in preset allow-list (#29938)
Adrien Gallouët
2026-10-04 10:34:43 +02:00
-
0faee50042
ci : pushing tag needs deploy key (#29937)
Sigbjørn Skjæret
2026-10-04 09:45:49 +02:00
-
f98b31c67e
ci : improve release flow (#29913)
Sigbjørn Skjæret
2026-10-04 09:31:24 +02:00
-
11fe02151f
webgpu: add f16 support to fill/set_rows (#29897)
Masashi Yoshimura
2026-10-04 09:07:47 +09:00
-
836d57176d
mtmd : fix deprecated strdup warning on Windows (#29863)
Adrien Gallouët
2026-10-03 19:47:54 +02:00
-
eec18f5d32
vendor : update cpp-httplib to 0.59.0 (#29886)
Alessandro de Oliveira Faria (A.K.A.CABELO)
2026-10-03 14:08:47 -03:00
-
1537a0a8b2
server : fix laya abort by limiting n_batch to n_ubatch (#29903)
Nik Bogatyrev
2026-10-03 18:19:13 +03:00
-
edd6e2bbda
common : add common_is_tty() helper and fix deprecated warnings on Windows (#29860)
Adrien Gallouët
2026-10-03 15:48:24 +02:00
-
9bf55f4a36
chat : honor json_schema in Ling 3.0 parser (#29813)
Yash Raj Pandey
2026-10-03 09:40:50 -04:00
-
a55e952b85
ci: fix flaky ADD_ADD f16 by using the fused ADD tolerance (#29904)
Pascal
2026-10-03 14:56:07 +02:00
-
436f6f89e1
graph: gather the recurrent states once so the reserve covers every split (#29856)
Pascal
2026-10-03 14:02:05 +02:00
-
b92761a515
ggml-openvino: update to 2026.4.1, optimize performance, expand ops, improve device listing. (#29852)
2026-10-03 14:29:25 +05:30
-
cb7934c52c
model : Add LFM2.5-Encoder-350M and LFM2.5-Encoder-230M (#29862)
Tarek Dakhran
2026-10-03 08:44:45 +02:00
-
889edf43dd
qwen4exp : halve the indexer score memory (#29825)
PascalandRuben Ortlam
2026-10-03 07:19:00 +02:00
-
99b95488ca
model: add support for clef decision model (text-only) (#29831)
Xuan-Son NguyenandSigbjørn Skjæret
2026-10-03 02:50:48 +02:00
-
bed0a85660
CUDA: fuse shared experts into MMVQ (#29184)
Aman Gupta
2026-10-03 02:26:27 +08:00
-
4ebdf2c74a
ci : use t4-medium for cuda jobs (#29842)
Sigbjørn Skjæret
2026-10-02 17:31:19 +02:00
-
1fb7ef3e33
spec : add probabilistic sampling for simple draft and MTP (#27694)
2026-10-02 20:22:54 +05:30
-
134b2bb756
ggml-cuda : fix cpy transposed path corrupting non-contiguous dst (#27663)
Yash Raj Pandey
2026-10-02 10:47:30 -04:00
-
2923cf2862
ggml-quants : avoid invalid rounding in qkx3 scale search (#29817)
Yash Raj Pandey
2026-10-02 10:37:35 -04:00
-
dd4c286f38
ggml-cpu : fix soft_max_back wrong output when dst aliases src1 (#27096)
Yash Raj PandeyandGeorgi Gerganov
2026-10-02 10:37:11 -04:00
-
46ca246de9
model: support nimble decision model (#29844)
Xuan-Son Nguyen
2026-10-02 14:56:53 +02:00
-
d8fbd2583a
readme : add cmd install commands (#29850)
Georgi Gerganov
2026-10-02 15:39:11 +03:00
-
926862e574
metal : add tensor API flash attention kernel for F16 KV (#29570)
Ethan Guo
2026-10-02 18:19:21 +08:00
-
a4cb4c61fd
llama, server: add /v1/systemone API (models: laya, julia-1, lev, openjev, kev) (#29818)
Xuan-Son Nguyen
2026-10-02 11:56:04 +02:00
-
70849ee82c
common : remove fs_open_ifstream() by using u8path() (#29841)
Adrien Gallouët
2026-10-02 11:43:45 +02:00
-
8d81559fa7
llama : silence unused-result warnings (#29839)
Adrien Gallouët
2026-10-02 11:07:06 +02:00
-
6805ae35df
llama : use GGML_ABORT instead of throw (#29840)
Adrien Gallouët
2026-10-02 10:44:44 +02:00
-
a8c9a4e7cc
opencl: use sigmoid f16 for bf16 (#29787)
lhez
2026-10-02 01:18:51 -07:00
-
392ded6546
SYCL: Q8_0 DMMV ESIMD and MMVQ wide load (#29186)
cwriterandcwriter
2026-10-02 10:14:35 +02:00
-
9e258a6e0a
vulkan: disable large matmul tile on Samsung GPUs with 32KB shared memory (#28531)
Jiwoong Song
2026-10-02 17:13:20 +09:00
-
b933289545
sycl: large register file for D=512 FA vec kernels (#29062)
Titaniumtown
2026-10-02 01:12:48 -07:00
-
c328acc91d
sycl : do not use slow oneDNN reference matmul and fattn (#28985)
Łukasz Ślusarczyk
2026-10-02 10:11:03 +02:00
-
4e2713c162
qwen4exp : optimize mask constructions (#29824)
Georgi Gerganov
2026-10-02 11:08:50 +03:00
-
631109b34d
ggml : add
alloc_buffer_n to buffer type interface (#23671)
Georgi Gerganov
2026-10-02 11:08:08 +03:00
-
254b177307
ci : fix missing zdnn backend check (#29837)
Aaron Teo
2026-10-02 15:10:45 +08:00