-
e9616d415f
cuda : gate DT3 MMQ by batch size, cuBLAS keeps the large-batch prefill
dt3
Millaguie
2026-08-11 08:51:29 +02:00
-
981f439ff3
tests : judge the batched DT3 path strictly when it is MMQ
Millaguie
2026-08-11 08:33:06 +02:00
-
f46e8072d8
cuda : add MMQ kernel for DT3
Millaguie
2026-08-11 08:33:06 +02:00
-
9622c56b0e
tests : accept either accumulator precision in the DT3 GEMM bit-identity gate
Millaguie
2026-08-11 03:31:42 +02:00
-
ad6dd747d4
vulkan : fp32 accumulators for the DT3 dequant matmul fallback
Millaguie
2026-08-11 03:01:09 +02:00
-
8ba4db150f
vulkan : add DT3 dequant, get_rows and scalar mul_mat_vec
Millaguie
2026-08-11 03:00:47 +02:00
-
c01c26b56e
tests : skip test-dt3-gpu on backends that do not implement DT3
Millaguie
2026-08-10 23:09:23 +02:00
-
10e1fe3d3c
cuda : decode DT3 bytes once in the MMVQ vec_dot
Millaguie
2026-08-10 21:51:59 +02:00
-
0f33afbe56
tests : declare the generic DT3 vec_dot weak in the parity test
Millaguie
2026-08-10 18:48:19 +02:00
-
4e109bc7e6
tests : check the arch DT3 vec_dot is bit-identical to the generic
Millaguie
2026-08-10 18:28:21 +02:00
-
bf4eca0eb6
ggml : add AVX-512 DT3 vec_dot
Millaguie
2026-08-10 18:28:15 +02:00
-
b285eb8a4f
tests : gate the DT3 GEMM fallback by bit-identity with an F16 GEMM
Millaguie
2026-08-10 15:20:34 +02:00
-
e5c6656dbf
tests : judge DT3 mul_mat on norm error, add fp16 reference and controls
Millaguie
2026-08-10 15:17:49 +02:00
-
659b1ace9e
tests : add DT3 GPU vs CPU parity test
Millaguie
2026-08-10 15:01:38 +02:00
-
0cc5e310c1
cuda : add DT3 MMVQ kernel
Millaguie
2026-08-10 15:01:38 +02:00
-
1a1f869f93
cuda : add DT3 dequantization
Millaguie
2026-08-10 15:01:28 +02:00
-
a277f4c6f1
tests : check DT3 byte positions against hand-computed literals
Millaguie
2026-08-10 13:46:21 +02:00
-
6d6552c862
llama : warn when quantizing to DT3
Millaguie
2026-08-10 13:46:21 +02:00
-
4efd061d43
ggml : harden DT3 validation and reference quantizer
Millaguie
2026-08-10 13:46:21 +02:00
-
b300ba053d
tests : add bit-level DT3 tests and Rust parity driver
Millaguie
2026-08-10 13:21:52 +02:00
-
985b0ecba2
gguf-py : add DT3
Millaguie
2026-08-10 13:17:53 +02:00
-
693eb7d719
llama : register the DT3 file type
Millaguie
2026-08-10 13:16:36 +02:00
-
698f37f40b
ggml-cpu : add DT3 generic vec_dot and type traits
Millaguie
2026-08-10 13:13:44 +02:00
-
160ea6c428
ggml : add DT3 reference quantization and dequantization
Millaguie
2026-08-10 13:11:50 +02:00
-
5c175d940f
ggml: add block_dt3, the dual-plane ternary block
Millaguie
2026-08-10 12:58:49 +02:00
-
030ebb558a
Address review comment of PR 25532 (#26852)
Gaurav Garg
2026-08-11 00:02:25 +05:30
-
689e227db4
opencl: transpose the K tile in local memory for FA prefill kernels (#26428)
Hongqiang Wang
2026-08-10 11:09:19 -07:00
-
0666ad2b2b
ci : target ROCm 7.14 for build and release (#25775)
Mario Limonciello
2026-08-10 12:53:12 -05:00
-
dd1ea52433
llama : support multi-output backend sampling (#25532)

Gaurav GargandGeorgi Gerganov
2026-08-10 19:28:56 +05:30
-
d2f83055d6
ggml-cpu : fix CPU affinity mask being ignored on Android (#26838)
Hitesh Chopra
2026-08-10 17:43:40 +05:30
-
f8def7fe16
ggml : require contiguous src for ROLL on CUDA and Metal (#25928)
Yash Raj Pandey
2026-08-10 08:01:44 -04:00
-
4dee52f82d
ui: UI/chat form follow ups (#26743)
Pascal
2026-08-10 13:32:51 +02:00
-
e5275f6f77
ci : don't specify python version in server-sanitize for broader runner compatibility (#26840)
Sigbjørn Skjæret
2026-08-10 13:32:22 +02:00
-
4ae84dea27
server: add more tool isolation support (ssh remote + podman rootless) (#26774)

PascalandXuan Son Nguyen
2026-08-10 13:31:09 +02:00
-
62bf73d25c
model: Muse Glimmer Support (#26841)






2026-08-10 13:07:27 +02:00
-
a52077c4ca
chat : Align Laguna-S-2.1 chat template to huggingface (#26232)
master
Guido Imperiale
2026-08-10 11:20:59 +01:00
-
4c6766fd7e
vendor: sync subprocess.h and drop local patches (#26808)
Pascal
2026-08-10 11:59:08 +02:00
-
86c298fb8a
llama: Restore quantization of mmprojs (#26818)

Pedro CuencaandXuan Son Nguyen
2026-08-10 11:58:32 +02:00
-
2e2d99cfd2
ci: Add support for CUDA 13.4 ARM64 builds for Windows (#26650)


2026-08-10 14:16:44 +05:30
-
7a20b417f4
model: add MTP support for Nemotron model (#26725)
Ruixiang Wang
2026-08-10 10:25:24 +02:00
-
e23e9440eb
vendor : update cpp-httplib to 0.53.0 (#26821)
Alessandro de Oliveira Faria (A.K.A.CABELO)
2026-08-10 04:57:45 -03:00
-
157b81fe6d
model : Granite-Switch Architecture (#25107)
Bar Haim
2026-08-10 00:53:46 -07:00
-
6ad4ab0ea0
readme : remove dev branches (#26832)
Georgi Gerganov
2026-08-10 09:53:26 +03:00
-
92d1bb0c99
ui: Linting & Formatting scripts (#26819)
Aleksander Grygier
2026-08-10 08:38:37 +02:00
-
1e396e72a8
server: gate the docker tools runtime tests on a real container run (#26826)
Pascal
2026-08-10 08:32:58 +02:00
-
0377426cef
model-saver : fix expert shared/chunk FFN length key clobber (#26693)
Caleb DeLeeuw
2026-08-09 23:32:01 -07:00
-
aea252fb4a
ci: fix the ctest sanitize runs (#26593)
Eve
2026-08-10 06:31:28 +00:00
-
f401bb1390
ggml-webgpu : refactor several wgsl files and simplify flash_attn wgsl. (#26134)
Masashi Yoshimura
2026-08-10 15:29:41 +09:00
-
74ce15741b
ui: degrade the working directory picker when file search is off (#26811)
Pascal
2026-08-09 21:20:23 +02:00
-
936918514c
ci: add pr-draft-label (#26801)
Xuan-Son Nguyen
2026-08-09 16:51:21 +02:00
-
08659901c4
ggml-cpu : fix missing Q5_0 dispatch in SpaceMiT backend (#26792)
Hao-Chen2337
2026-08-09 18:16:53 +08:00
-
61141f1487
ci: rm
GGML_HIP_ROCWMMA_FATTN (#26760)
Aaron Teo
2026-08-09 18:15:28 +08:00
-
7ba604f1cb
server: report the isolate working directory from get_info (#26773)

PascalandXuan-Son Nguyen
2026-08-09 00:42:50 +02:00
-
687e778927
CUDA: fuse rms_norm + mul + rope (+ view + set_rows) (#26767)
Rafail Giavrimis
2026-08-08 17:32:37 +01:00
-
18f7ad7fc9
server, ui: only offer a working directory when a tool reads it (#26762)
Pascal
2026-08-08 16:36:21 +02:00
-
dd2c7c4471
server: add initial tool isolation support (via docker) (#26507)

Xuan-Son NguyenandPascal
2026-08-08 16:35:53 +02:00
-
69bf643791
CUDA: fix thread/block count in quantized cpy kernel launches (#26731)
Rafail Giavrimis
2026-08-08 05:40:04 +01:00
-
3653e6d6d5
tts: account for the vocoder pass in the timings line (#26733)
Pascal
2026-08-07 22:35:52 +02:00
-
fc6545d322
allozaur/feat/chat form contenteditable (#26717)
Aleksander Grygier
2026-08-07 20:40:10 +02:00
-
1621a3d388
tests : speed-up server test suite 3x (#26734)
Georgi Gerganov
2026-08-07 21:38:32 +03:00
-
6de1b63473
allozaur/feat/chat slash commands (#26716)
Aleksander Grygier
2026-08-07 20:20:01 +02:00
-
f8e30266d2
sycl: coalesce the ssm_conv window loads (#26612)
Titaniumtown
2026-08-07 11:09:32 -07:00
-
a194a75b7e
metal : fix NORM/RMS_NORM for row lengths that leave a partial simdgroup (#26708)
robertomeroni
2026-08-07 20:09:07 +02:00
-
23634783c5
ui: Filesystem
@mentions for Chat Form (#26715)
Aleksander Grygier
2026-08-07 18:45:54 +02:00
-
4cb22cd537
mtmd: fix longest_edge ignoring min/max pixels (#26638)
Xuan-Son Nguyen
2026-08-07 18:05:15 +02:00
-
4cf5cab65d
sync : ggml
Georgi Gerganov
2026-08-07 17:10:39 +03:00
-
933f46f3cb
ggml : bump version to 0.19.0 (ggml/1581)
Georgi Gerganov
2026-08-07 17:10:01 +03:00
-
9ba73fd1f5
server : clarify comment in eval_llama_cmpl_schema [no ci] [no release] (#26720)
Daniel Bevenius
2026-08-07 15:39:33 +02:00
-
f4f7758cae
webui: load the model selected via ?model= when ?load=true (#26707)
Emanuil Rusev
2026-08-07 16:31:40 +03:00
-
34e9ee57f5
ui: set npm
min-release-age to protect against supply-chain attacks (#26711)
Niklas Wenzel
2026-08-07 14:53:51 +02:00
-
dff15d4ac9
server: (router) add LRU scheduler (#26572)
Xuan-Son Nguyen
2026-08-07 14:46:53 +02:00
-
e1470ee6a2
server: (router) do not evict busy models (#26567)
Xuan-Son Nguyen
2026-08-07 14:39:59 +02:00
-
217df17ac3
mtmd: stop feeding the text stream again during Qwen3-TTS generation (#26706)
Pascal
2026-08-07 13:32:52 +02:00
-
cb26014d96
ggml : add aarch64 HWCAP fallbacks and fix fp16 variant detection (#25554)
Kilian Hu
2026-08-07 13:07:10 +02:00
-
82bb48500a
ui: read model modalities from the router model list (#26709)
Pascal
2026-08-07 12:07:58 +02:00
-
42e98813e4
Mitigate crashing issue on Windows MSYS2 UCRT64 environment (GCC 16.1.0) (#26555)
Masato Nakasaka
2026-08-07 02:17:16 -07:00
-
fc3f10b389
sycl: fix UE4M3 parsing (#25608)
Chris Lee
2026-08-06 23:28:53 -06:00
-
6b5c2efb4e
sycl: *glu flat path (#26354)
Titaniumtown
2026-08-06 22:24:40 -07:00
-
31558dbb76
sycl : Support DSv4 OPs: LIGHTNING_INDEXER,DSV4_HC_COMB,DSV4_HC_POST,DSV4_HC_PRE (#26568)
Neo Zhang
2026-08-07 13:22:23 +08:00
-
c1f4109898
sycl : update guide Q&A and script for device setting (#26442)
Neo Zhang
2026-08-07 13:18:47 +08:00
-
eef5f3e343
sycl : fix error Error OP FLASH_ATTN_EXT on arc770 (#26441)
Neo Zhang
2026-08-07 13:17:56 +08:00
-
c074cb3f76
sycl : enhance OP set_rows to support all missed data types (#26515)
Neo Zhang
2026-08-07 12:52:52 +08:00
-
5b87ed30f8
cuda: fix warnings for unused variable/function (#26688)
David Friehs
2026-08-07 06:51:56 +02:00
-
d8d9887228
ci: abort if build requirements are missing (#26368)
Niklas Wenzel
2026-08-07 06:50:48 +02:00
-
e40bf88642
metal : avoid
threadgroup matrix array instantiation in kernel_lightning_indexer (#26646)
JamePeng
2026-08-07 12:49:14 +08:00
-
15586e2d71
mtmd: add chunk save/load function (#26645)
Xuan-Son Nguyen
2026-08-06 19:46:40 +02:00
-
6a32c29a74
server: fix empty response for /cors-proxy (#26656)
Xuan-Son Nguyen
2026-08-06 15:07:22 +02:00
-
eb5667a169
convert : fix DeepseekV4 rope parameters with transformers 5.x (#26673)
Sigbjørn Skjæret
2026-08-06 15:06:52 +02:00
-
3db4ff877d
model-loader : fix quantized reshaped tensor strides (#26672)
Georgi Gerganov
2026-08-06 15:21:44 +03:00
-
e700bfb37f
convert : accept "ExaoneMoeForCausalLM" arch spelling (#26660)
Csaba Kecskemeti
2026-08-06 03:56:04 -07:00
-
a1f96d4fc2
ci : onboard AMD ROCm CI with gfx1151 fixes (#26544)




2026-08-06 01:43:26 -07:00
-
9de0fcf2b3
model-conversion : add --model-name to conversion scripts (#26665)
Daniel Bevenius
2026-08-06 09:38:06 +02:00
-
803b7fcae8
vulkan: fix submission batching size, add debug tools for diagnosing causes of DeviceLost drivers errors (#26371)
Ruben Ortlam
2026-08-06 09:24:13 +02:00
-
c8e03ce812
mtmd/ggml: add ggml_build_forward_order (#26649)
Pascal
2026-08-06 00:47:59 +02:00
-
f9e832c10e
server: harden the file_glob_search directory walk (#26626)
Pascal
2026-08-05 21:31:54 +02:00
-
360e1349f0
tests: re-enable MiniMax M3 in
test-llama-archs (#26633)
Niklas Wenzel
2026-08-05 17:58:34 +02:00
-
b06aa774c0
mtmd: Unlimited-OCR fix max_tiles, setting in converter (#25614)
Saba Fallah
2026-08-05 15:30:14 +02:00
-
cd0fa6051a
grammar : degrade max repetition >= 2000 to unbounded (#26613)
Aldehir Rojas
2026-08-05 07:39:10 -05:00
-
717dad5c8e
mtmd: support multi-row batching for deepseek-ocr (#26154)

Xuan-Son NguyenandSaba Fallah
2026-08-05 13:34:52 +02:00
-
9a688e51e6
fit: Fix memory allocation for MTP layers (#26605)
Sergey Malinin
2026-08-05 14:29:45 +03:00