Sync fork master with upstream (2026-09-17) - #109

Merged

jimw567 merged 606 commits into

jimwu.sync-upstream-20260917

Sep 21, 2026

This pull request is big! We're only showing the most recent 250 commits

Commits on Sep 3, 2026

sycl : enhance the api to support peer-to-peer copy (#27550)
arthw
authored
0df974d
vulkan: fix FA dequant path engagement (#28190)
Nathanw1014
authored
c7bda03
CUDA: Allow concurrent streams per split for multi-GPU (#28198)

tannerbruhn
and
am17an
authored
0ba6499
mtmd : add `const` in various places (#28307)
madsmtm
authored
f45576a
metal : fix glu dispatch with ne00 = 1 (#28306)
ggerganov
authored
0df017d
metal : add sparse FA (#28098)
ggerganov
authored
7bb0fc1
mtmd: propagate const to preproc class (#28310)
ngxson
authored
de8656b
server : use pytest-xdist for server tests (#28298)
danbev
authored
42f0225
ops: add Hexagon to ops.md and update main README.md (#28263)
tboinovski1
authored
e107984
snapdragon: ci updates to use new run script (#28293)
max-krasnyansky
authored
d30500b
tune MMVQ to MMQ crossover for SM87 (#28285)
kbenkhaled
authored
8c1a251
opencl: quant lm_head / decode GEMV and medium-batch GEMM optimizations (speculative decoding/MTP) (#26477)
wanghqc
authored
95ef7fc
common/json-schema : fix GBNF grammar generation for empty object schemas (#28279)
SergeySklyarov
authored
c5a5535
llama: fix whole source code rebuilt on each new commit (#28278)
ngxson
authored
d230ddd

Commits on Sep 4, 2026

Commits on Sep 6, 2026

metal : add remaining fa-vec tunings for M2 Max (#28458)
nikwen
authored
971595d
ui : embed assets directly with CMake (#28445)
angt
authored
c457e3b
common: add --log-jsonl (#28437)
ngxson
authored
7620399
ui: Improve Chat Messages rendering performance (#28460)
allozaur
authored
0afb805
grammar : fix max repetition threshold (#28469)
aldehir
authored
9e0e220
cuda: fixes races in mmid and mmf (#28475)
am17an
authored
73a43d1
opencl: properly choose weights pack for q4_K, q5_K mul_mat (#28402)
lhez
authored
d03efa5
[Model] Support for Spark2_5ForCausalLM implementation (#27868)

authored
3ad1ba7
models : fix GDN normalization from `max` to `rsqrt` (#28068)

danielhanchen
and
ggerganov
authored
5fdfa62
convert : add `--fuse-qkv` flag to fuse Q/K/V into QKV during HF-to-GGUF conversion (#22780)
JoursBleu
authored
465e49b

Commits on Sep 7, 2026

Commits on Sep 8, 2026

Commits on Sep 9, 2026

Add IQ type handling for MoE (#28476)

cwriter
and
cwriter
authored
304665f
tests : use less threads for data initialization (#28325)
ggerganov
authored
30b6a75
vulkan: add f16 B-type matmul pipelines and warp tile size tuning for Intel coopmat1 (#27471)
fish-jiang
authored
1945e09
vulkan: add dedicated iq4_xs mat-vec shader (#28426)
WakeUpMorty
authored
df750f7
jinja: treat a null left operand of in as a plain lookup (#28620)
ServeurpersoCom
authored
b31b71f
mtmd: propagate video ID to bitmap (#28601)
ngxson
authored
6de9cdb
model: fix granite3 moe unknown parameter count (#28632)
taronaeo
authored
e2d2c0d
args: officially deprecate --mmap|mlock|dio (#28334)
taronaeo
authored
14a9d09
CUDA: replace GGML_FA_ALL_QUANTS with GGML_FA_QUANTS, more control over what is compiled (#28079)

pwilkin
and
JohannesGaessler
authored
5a4d0fe
CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#28552)

authored
d4abd57
llama : use int32_t for llama_sampler_chain_n return type (#28631)

linsen458-spec
and
linsen
authored
4850c77
py : bump numpy to 2.4.6 (#28649)
CISC
authored
9cf3bf2
py : lower numpy to 2.2.6 (#28654)
CISC
authored
4b98ab8
vulkan: Convert FILL to distribute workgroups in 2D to avoid exceeding maxComputeWorkGroupCount (#28592)
rillomas
authored
22397c3
hexagon: rope updates (#28628)

tboinovski1
and
max-krasnyansky
authored
6d9c82e
vulkan: use spec constant for matrix matrix multiplication A-type (#25773)
0cc4m
authored
91f6a6c
ci: fix sanitizer tests (#28583)
netrunnereve
authored
434ddbb

Commits on Sep 10, 2026

Commits on Sep 11, 2026

Commits on Sep 12, 2026

Commits on Sep 13, 2026

Commits on Sep 14, 2026

common : move llama_n_rs_seq to before llama_decode (#28749)
danbev
authored
093a2f8
sycl : fix oneDNN scratchpad breaking the pool free order (#28704)
lslusarczyk
authored
661643e
ci : remove gg_sum summary logic (#28857)
ggerganov
authored
15d8f2d
ci : trigger self-hosted CI on changes to ci/run.sh (#28859)
ggerganov
authored
89fe242
ggml-cpu : disable PCH and fix CACHE_LINE_SIZE ambiguity to fix heap corruption (#28882)
ggerganov
authored
2f53959
sycl: rfc: Use radix select for top_k (#28670)

cwriter
and
cwriter
authored
21f6b0d
llama: add Maple 20B-A1B ternary MoE architecture (CPU) (#27000)
AlexGabbia
authored
3d10bcd
tests(s390x): add non-vxe build to tests (#28776)
taronaeo
authored
be2c6d7
models : fix mimo2 swa pattern load (#28865)
CISC
authored
1aca1f9
models : fix incorrect uses of get_key_or_arr (#28868)
CISC
authored
97e4ca7
tests : add fusion baseline README and broaden fusion CI triggers (#28893)
ggerganov
authored
bbdd9f2
ggml : bump version to 0.24.0 (ggml/1627)
ggerganov
committed
eeea731
sync : ggml
ggerganov
committed
d9e03f1
llama.cpp : bump version to 0.4.1 (#28900)
ggerganov
authored
b29c606
scripts: Add script to verify API/ABI compatibility (#28579)
ckastner
authored
dfe4516
cmake : remove precompiled headers (#28892)
danbev
authored
f3a184b
release : added gfx1103 to ubuntu rocm build (#28423)
apparle
authored
b4fa47d
qwen4exp: enable rms_norm + mul fusion (#28896)
am17an
authored
41abbfd
ci : add ubuntu-cuda builds to release (#28186)

ORippler
and
CISC
authored
391fac1
HIP: fattn-mma: use fp32 accumulation on MFMA devices (#28576)
IMbackK
authored
bfdc321
CI: hip-quality-check: ignore spill added in bfdc32183d57f1e35bacf35c47d6311e2028bbbc (#28909)
IMbackK
authored
96ffdc4
ci : reuse build tag name when used instead of safe one (#28911)
CISC
authored
7cf1c54
webui: stop re-probing disabled /tools endpoint on every message (#28646)
geckguy
authored
1bc7a5a

Commits on Sep 15, 2026

Commits on Sep 16, 2026

Commits on Sep 17, 2026

Commits on Sep 19, 2026

Merge upstream/master into fork master

Jim Wu
and
claude
committed
6ea0bbe