-
fb4b2737a8
vulkan: add logging to pipeline compile issues (#29794)
Ruben Ortlam
2026-10-02 08:27:56 +02:00
-
207bdab950
pyproject : add linux platform marker to uv torch source (#29177)
Kasimir Tanner
2026-10-02 07:32:13 +02:00
-
5fc4f3c8c7
hexagon: install rebuilt HTP skels (#29828)
kurquhar
2026-10-01 19:07:38 -07:00
-
159c651f57
qwen4exp: fix tests (#29819)
Aman Gupta
2026-10-02 09:12:25 +08:00
-
a868c3e3c5
hexagon: add q2_k and q3_k quant type support (#29717)
Jhen-Jie HongandMax Krasnyansky
2026-10-02 05:12:25 +08:00
-
ec7630a640
CUDA: fix 2 broken Volta FA cases (#29803)
Johannes Gäßler
2026-10-01 21:55:41 +02:00
-
78e2964c23
llama: refer to segment documentation [no ci] (#29074)
Johannes Gäßler
2026-10-01 21:52:40 +02:00
-
f1cee9941b
common,rpc : fix cache dir creation through symlinks on buggy libstdc++ (#29816)
Adrien Gallouët
2026-10-01 20:15:02 +02:00
-
68e79bd8cd
skill: note about model-specific CLI arguments + testings (#29808)
Xuan-Son Nguyen
2026-10-01 19:53:23 +02:00
-
e358d59178
ci: fix Fusion / metal by updating the qwen4exp baseline (#29812)
Pascal
2026-10-01 18:57:34 +02:00
-
81e39ad343
llama : clamp kpool re-pool bound to existing pools (#29805)
Georgi Gerganov
2026-10-01 19:55:34 +03:00
-
dcd387a412
hexagon: shared strided DMA copy for CPY and CONCAT, any-dim CONCAT via DMA (#29685)
Yiwei ShaoandMax Krasnyansky
2026-10-01 08:38:37 -07:00
-
d775ebf363
server: return HTTP 400 for invalid embedding requests (#29060)
Sam Malayek
2026-10-01 10:52:43 -04:00
-
2b36825cbc
convert : write Gemma embedding scale for DFlash drafts (#29802)
Yu Chengye
2026-10-01 22:34:32 +08:00
-
42d958167a
cuda : route sm70 to the Turing MMVQ nwarps table (#29753)
2026-10-01 20:21:26 +07:00
-
13b4d7135a
metal : release temporary private transfer buffers (#29777)
Mike van LammerenandNiklas Wenzel
2026-10-01 09:17:23 -04:00
-
4b1622afb7
webgpu: add bfloat16 support for MUL_MAT/MUL_MAT_ID/GET_ROWS- #29358 (#29358)
Masashi Yoshimura
2026-10-01 22:09:07 +09:00
-
869034b4bb
llama : fix invalid assert in recurrent memory (#29799)
Georgi Gerganov
2026-10-01 14:49:03 +03:00
-
b56f34ab13
CUDA: Handle compute type for NVFP4 on cublass path (#29173)
ynankaniandJohannes Gäßler
2026-10-01 11:23:52 +00:00
-
c061df1983
Qwen4Exp: add MTP (#29761)
Aman GuptaandGeorgi Gerganov
2026-10-01 19:13:27 +08:00
-
66e0c17ee1
llama: fix qwen4exp (#29751)
Aman Gupta
2026-10-01 19:13:27 +08:00
-
7677678503
CUDA: Make CCCL configurable + pin it to 3.4.3 for CI jobs (#29792)
Oliver Simons
2026-10-01 13:05:29 +02:00
-
552f18f912
mtmd: cap max_image to ubatch for non_causal models (#29773)
Xuan-Son Nguyen
2026-10-01 11:55:11 +02:00
-
5503b04b05
meta: clear inactive AllReduce shards with FILL, not SCALE (#29793)
Georgi Gerganov
2026-10-01 12:43:57 +03:00
-
def4d406ae
jinja : skip copying loop scope unless a loop filter needs it (#29776)
a u s t i n
2026-10-01 02:10:57 -06:00
-
32dd62ee6d
llama-mmap : avoid a second full-size copy of each tensor with direct-io (#29749)
Pranesh GonegandlaandPranesh Gonegandla
2026-10-01 13:14:31 +05:30
-
f11d642a27
HIP: avoid treating CDNA as dgx spark for gqa_ratio 20 in fattn_mma dqk 576 (#29572)
uvos
2026-10-01 07:51:38 +02:00
-
3aa0ce9bca
hex-workqueue: fix race condition in seqn getting out of sync with idx_read/write (#29785)
Max Krasnyansky
2026-09-30 22:35:52 -07:00
-
b0aca3c653
BLAS : Document AOCL-BLAS build and label the device AOCL-BLAS (#29640)
Pradeep Rao
2026-10-01 11:05:01 +05:30
-
b8f96c3e82
common : add LLM-jp-4.1 Harmony dialect handler (#29681)
e-mon
2026-10-01 14:33:55 +09:00
-
3ec4df42d9
opencl: mark vec subgroup bcast as supproted for Adreno E17 compiler (#29698)
lhez
2026-09-30 22:33:23 -07:00
-
db33d3cb89
vocab : honor BOS/EOS settings for PLaMo-2 and PLaMo-3 (#29734)
Toki NasinandSigbjørn Skjæret
2026-10-01 14:31:42 +09:00
-
7dad6db858
llama-bench : fix verbosity filter to show GGML_LOG_ERROR (#28229)
Kushal Garg
2026-10-01 10:50:00 +05:30
-
2232bc8b5f
metal : use bf16 math for mxfp4 mul-mat (#29770)
Georgi Gerganov
2026-10-01 08:19:24 +03:00
-
79625e056e
llama-bench : fix docs (#29464)
Marlon Paz
2026-10-01 09:18:46 +04:00
-
66bcc27706
docs : refresh CPU ops support matrix (#29666)
CaramelizedCUDA
2026-10-01 14:17:51 +09:00
-
10f340d1a2
model : re-enable -sm tensor for qwen4exp (#28569)
Kevin Hopper
2026-10-01 00:16:49 -05:00
-
0c1e57098b
webgpu: fix SSM_SCAN binding aliasing (#29750)
Masashi Yoshimura
2026-10-01 11:11:48 +09:00
-
f7b384c1e5
ggml-opencl : replace alloca() with std::vector (#29765)
Adrien Gallouët
2026-09-30 23:59:46 +02:00
-
f872b59112
cuda: guard the iq4_nl dequantize row kernel against short rows (#29683)
R0CKSTAR
2026-10-01 04:33:22 +08:00
-
a4d880fd5c
Hexagon: optimize ALLREDUCE with support for safe scatter mode (#29757)
Ehsan BateniandMax Krasnyansky
2026-09-30 15:56:44 -04:00
-
feb9a3d6de
args: fix cli download mmproj arg (#28977)
pr3pony
2026-10-01 02:45:39 +08:00
-
4453b535fd
llama : preserve original batch order for speculative decoding layer inputs (#29019)
Hrishith Thadicherla
2026-09-30 11:17:40 -07:00
-
4f31296a90
test-llama-archs : toggle causal_attn to catch graph shape changes (#29724)
Sihan Yu
2026-09-30 20:39:53 +03:00
-
b016f461be
convert : fix LoRA conversion crash for Qwen3.5 V-head reorder (#28324)
2026-09-30 20:30:21 +03:00
-
81ff93ea1d
llama: properly handle KV on training (#28520)
Xuan-Son Nguyen
2026-09-30 18:09:08 +02:00
-
60e9cf7a7b
batch: migrate the rest of examples to llama_batch_ext (#29601)
Xuan-Son Nguyen
2026-09-30 18:08:43 +02:00
-
05af0d2b13
glm5-next: give dead indexer slots unique scatter rows (#29745)
Pascal
2026-09-30 17:10:48 +02:00
-
2149c00f44
ggml/gguf : fix integer overflow (#29384)
Daniel Kuts
2026-09-30 17:59:00 +03:00
-
876c75b1f6
codeowners : remove former ZenDNN owner (#29747)
Vishal SinghandVishal Singh
2026-09-30 20:08:21 +05:30
-
b04642061d
cli: exit on stdin EOF and drop the console wide Ctrl+C broadcast (#29722)
Pascal
2026-09-30 16:24:46 +02:00
-
22bdcc4cdd
mimo : support dflash (convert + feature extraction) (#29650)
Georgi GerganovandSigbjørn Skjæret
2026-09-30 17:04:35 +03:00
-
ca2e2037b6
jinja : support coerced array attributes (#29574)
Sigbjørn Skjæret
2026-09-30 15:33:26 +02:00
-
bdeb855b30
ggml-et : remove useless alloca() (#29663)
Adrien Gallouët
2026-09-30 15:14:45 +02:00
-
3b3d022b82
ci : fix Models Backend Check by shortening the hrm_text fixture (#29744)
Pascal
2026-09-30 15:03:38 +02:00
-
185103dcf5
llama: llama_prefetch_rows (#29599)
2026-09-30 20:27:18 +08:00
-
2090f60f0b
ggml : add BF16 unary, GLU, binary and scale ops (CPU, CUDA) (#29675)
Aman Gupta
2026-09-30 20:24:59 +08:00
-
90c908d06d
cpu: accept BF16 in src1 of mul_mat (#28937)
Pascal
2026-09-30 14:14:45 +02:00
-
8df332de1b
model-conversion : add --add-bos to run org model script (#29558)
Daniel Bevenius
2026-09-30 12:52:02 +02:00
-
4a096b8ff6
ui : shared model display primitives (#29644)
Aleksander Grygier
2026-09-30 12:42:18 +02:00
-
8664eaea30
ui : model download pipeline (#27959)
Aleksander Grygier
2026-09-30 12:42:17 +02:00
-
4cfb6d1c75
ui : model memory-fit estimation (#27957)
Aleksander Grygier
2026-09-30 12:42:17 +02:00
-
9b43336114
ui : Hugging Face Hub data layer (#27947)
Aleksander Grygier
2026-09-30 12:42:16 +02:00
-
f653250407
ui : model id grammar for sidecars, quants and capability parsing (#27946)
Aleksander Grygier
2026-09-30 12:42:16 +02:00
-
fa2bde5543
ui : type-safe API types, fetch helpers and download-ready models store plumbing (#29582)
Aleksander GrygierandPascal
2026-09-30 12:42:15 +02:00
-
25747b08e7
openvino: serve GET_ROWS on a weight view from the base Constant (#28381)
Pascal
2026-09-30 11:17:02 +02:00
-
db00347a4b
ci : fix Fusion / metal by adding glm5-next to MTL.csv (#29712)
Pascal
2026-09-30 09:50:23 +02:00
-
272aad8b98
musa : define __CUDA_ARCH__ for device passes (#29508)
R0CKSTAR
2026-09-30 15:14:37 +08:00
-
72db1e02ff
ci : add models backend check (#29651)
Sigbjørn Skjæret
2026-09-30 09:06:24 +02:00
-
2a53ace3be
SYCL: reduce tensor allreduce sync with pinned host buffers (#29604)
Captain-Tripps
2026-09-30 01:25:13 -05:00
-
649dcb1036
add GLM-5.3-Flash (GLM5-Next) support (#27773)
2026-09-30 08:20:32 +02:00
-
931351ea50
vendor: update BoringSSL to 0.20260929.0 (#29669)
Alessandro de Oliveira Faria (A.K.A.CABELO)
2026-09-30 01:44:32 -03:00
-
eae11d2217
ggml-zdnn: impl buffer reset, fix memory leaks (#29637)
Aaron Teo
2026-09-30 12:21:41 +08:00
-
19e28a2770
Hexagon f16 activation ops (#29209)
cqderekandMax Krasnyansky
2026-09-30 05:59:13 +08:00
-
a6ea155d3d
gguf : reject tensor size that wraps after padding (#26979)
Xiang Chen
2026-09-30 04:46:20 +08:00
-
d3954b9324
ggml : check row bounds in get_rows_back (#29575)
Adrien Gallouët
2026-09-29 22:44:41 +02:00
-
48de2a1bcb
model : support classifier_pooling for rerankers (#29627)
bosh
2026-09-30 03:33:05 +07:00
-
7fee178464
hexagon: optimize concat op (#29673)
Trivikram Reddy
2026-09-29 14:52:10 -05:00
-
6a2743f028
CUDA: bitonic argsort handles rows wider than one block (#28957)
Pascal
2026-09-29 20:09:10 +02:00
-
748d4225b9
ggml-cuda: HIP: optimize packed byte subtraction (
__vsubss4 -> __vsub4) (#29478)
thelittlefiremanandCarl Philipp Klemm
2026-09-29 19:59:19 +02:00
-
cee37ffea0
ci: add zdnn backend build but not test (#29541)
Aaron Teo
2026-09-30 01:45:14 +08:00
-
6dbbac4429
opencl: fix get_tensor for q5_K adreno gemm_nonshuffle kernel (#29555)
lhez
2026-09-29 10:44:42 -07:00
-
5c200e0c8d
vulkan: Tune GDN kernel, fix Intel performance (#29476)
Ruben Ortlam
2026-09-29 19:44:20 +02:00
-
83dd71f869
vulkan : Load F32 A matrix 2 at a time when its 2-aligned (#29254)
Matt Corallo
2026-09-29 17:39:36 +00:00
-
94a0ae3e72
vulkan: MOE aware mat_mul_id tile selection (#29182)
Ankit Khandelwal
2026-09-29 23:08:59 +05:30
-
da89bb3ccc
ggml : fix c++ odr by properly using GGML_COMMON_DECL_CPP (#29504)
2026-09-29 10:31:31 -07:00
-
a3f84faf49
vocab : keep </s> NORMAL in PLaMo-2 and PLaMo-3 (#29580)
Toki Nasin
2026-09-30 02:30:40 +09:00
-
284153e069
ggml : accumulate f16 dot products in f32 on AVX512-FP16 (#29545)
Adrien Gallouët
2026-09-29 19:29:12 +02:00
-
b5cf8ce02a
ggml : require input tensors to be GGML_OP_NONE (#29647)
Georgi Gerganov
2026-09-29 20:23:33 +03:00
-
e904318a2d
hexagon: add FP32 GELU_ERF and GEGLU_ERF support (#29631)
Aparna M P
2026-09-29 22:05:47 +05:30
-
d280808f5d
common : stop accepting draft tokens at EOG (#29638)
Ethan Guo
2026-09-30 00:25:41 +08:00
-
ba0ba54d93
server : remove the built-in UI's service worker when the UI is not served (#29565)
Emanuil Rusev
2026-09-29 18:48:17 +03:00
-
00af63567a
common : use fs::path for config dir (#29649)
Adrien Gallouët
2026-09-29 17:22:49 +02:00
-
c85b92c69c
tests : adjust server string regex to also match m2 utlra results (#29648)
Georgi Gerganov
2026-09-29 15:32:35 +03:00
-
31385c9ceb
common : add fs_write_atomic() (#29642)
Adrien Gallouët
2026-09-29 13:46:53 +02:00
-
8019dc563b
ggml : collect all input tensors into graph_inputs (#29634)
Georgi Gerganov
2026-09-29 13:37:50 +03:00
-
86ea01d05e
ggml-zdnn: fix 0-row tensor crash (#29636)
Aaron Teo
2026-09-29 16:51:10 +08:00
-
18b74ff68e
musa: build the docker image and CI container from the MUSA SDK images (#29624)
R0CKSTARandyeahdongcn
2026-09-29 16:42:20 +08:00
-
c13e04e1dd
ggml : speed up model loading (#29598)
Adrien GallouëtandJohannes Gäßler
2026-09-29 10:35:08 +02:00
-
c8cda8b4fe
ci: remove gpu-rocm keyed directory logs (#28940)
Aaron Teo
2026-09-29 16:21:54 +08:00