K2 Horizon: integrate all fixes on current upstream - #5

Draft

bitalov wants to merge 560 commits into

ifm-ai:model/K2Horizonfrom

bitalov:k2-horizon-all-fixes-20260927

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

Commits on Sep 15, 2026

ci: Bump CUDA Windows x64 builds to 13.4.1 (#28930)
shivamkumard-ctrl
authored
0ecb159
metal : add FA kernels for HSK=96, HSV=64 (MiniCPM3) (#28599)
wyanzhao
authored
1e7bcf3
ci: bump kleidiai runners from 22.04 to 24.04 (#28885)
taronaeo
authored
1af6c65
opencl: add generic ssm_scan (#28881)
lhez
authored
6ec1a7e
OpenVINO: optimize stateful decode and GPU MoE inference (#28638)

authored
77d554b
vulkan: support sparse Flash Attention (#28105)
0cc4m
authored
fc82583
models : move build_arch_graph() after graph() template specialization (#28934)
cpeterso
authored
9e71716
cuda: support row-contiguous SUM_ROWS (#26308)
MohamedElashri
authored
5431581
rpc : hash-cache only weights (#28789)

authored
7609846
docs: Rule of thumb for AI review time [no ci] (#28945)
JohannesGaessler
authored
6011c34
ci: build MUSA for only 1 arch (#28944)
am17an
authored
d1d3c33
opencl: choose the MoE expert matmul by batch size for speculative decoding/MTP (#27637)

wanghqc
and
lhez
authored
9f31776
HIP: Enable AllReduce for ROCm (#27825)
Stastez
authored
38a5b42
hex-cpy: use dma if src and dst are contiguous (#28906)
trivikram-reddy1
authored
72b590d
hexagon: add back missing contiguous fast-path and hvx_copy_uu for each run (#28886)
jhen0409
authored
930e2fa

Commits on Sep 16, 2026

Commits on Sep 17, 2026

Commits on Sep 18, 2026

Commits on Sep 19, 2026

cuda : fix CUB argsort corruption caused by in-place keys (#28389)

authored
b23701f
metal : support qwen4exp hc ops (#29000)
ggerganov
authored
59fc5a1
test-llama-archs : generate dummy test vocab (#29084)
ggerganov
authored
efa28e9
metal : fix FA support checks (#29122)
ggerganov
authored
60b06ab
metal : add MoE and SSM_CONV fusion optimizations (#28948)
ggerganov
authored
5b59b83
json-schema : accept escaped hyphen in regex patterns (#29127)
ChihebBENCHEIKH1
authored
eb1e1f4
server : improve startup log messages (#29125)
ggerganov
authored
1af554f
hexagon: enable support for TOP_K op (#29113)
aparmp-quic
authored
7d4b92b
hexagon: add support for GEGLU_QUICK (#29114)
aparmp-quic
authored
851cb34
hexagon: enable I32 GET_ROWS (#29116)
aparmp-quic
authored
e613ef2
chat : add dedicated Ling 3.0 (Bailing V3) parser (#28682)

aetherbird
and
aldehir
authored
59657a6
chat : fix gemma4 required tool grammar (#29115)
aldehir
authored
f072b10

Commits on Sep 20, 2026

metal: add F16 input to the FWHT (#29094)
bri-prism
authored
9a9f939
fix(mamba) : make time-step projection input contiguous (#28832)
abetlen
authored
4260903
ui: Fix mobile breakpoint + content overflow issues (#29108)

allozaur
and
ServeurpersoCom
authored
b23efaa
CUDA: enable sparse fa for qwen4 (#28770)
am17an
authored
3cf0325
common/peg : handle invalid utf-8 sequences in the AST (#29161)
aldehir
authored
3d82ef6
metal : support arbitrary hc in dsv4_hc_pre (#29169)
ggerganov
authored
a894dae
CUDA: tune FA for Gemma 4 on Ampere or newer (#29152)
JohannesGaessler
authored
ce8caa6

Commits on Sep 21, 2026

Commits on Sep 22, 2026

Commits on Sep 23, 2026

sampler: reduce the size of the probe (#29285)
max-krasnyansky
authored
441df11
vulkan: hide internal symbols to prevent duplicate-dlopen state destruction (#29139)
ewintr
authored
08b1d2a
sycl : support op get_rows_back, only support fp32/fp16 (#25266)
arthw
authored
4d7d770
sycl: extend MMVQ GLU fusion, add rms_norm+scale and ssm_conv+silu fusions (#28931)
anantshri
authored
5e48b31
sycl : support new UT case for mul_mat_hadamard fp16 (#29218)
arthw
authored
384a534
cuda: top-k MoE should always fire (#28432)
am17an
authored
1a67982
ggml-meta: resolve multi buffer views (#29266)
0cc4m
authored
9425611
vulkan: add IQ4_XS MMQ/MMV matmul kernels (#28415)
pwilkin
authored
b1ff4ca
sycl : fix compile warnings
ggerganov
committed
e97545d
ggml : bump version to 0.25.0 (ggml/1635)
ggerganov
committed
503549c
sync : ggml
ggerganov
committed
45062d4
make-release : update summary prompt
ggerganov
committed
183d2a0
ci : run python (jinja) test (#29302)
CISC
authored
86b2daa
model : support Gemma4 DSpark draft backbone (#29226)
hthadicherla
authored
633733d
model-conversion : add causal-compare-logits recipe (#29305)
danbev
authored
18f9f7b
ci : fix build-cmake runner target (#29299)
CISC
authored
26758d3
server: Dedup the draft HF model via dedup-cache-models (#27934)
DreamingWater
authored
bcbc936
server: accept OpenAI video_url content type and data: video URIs (#27921)
calebrio02
authored
057494f
metal : key the fa-vec tuned table by family instead of SKU (#29075)
forforever73
authored
ee3ecce
jinja : parse unary +/- before variables (#29244)
cs-fisha
authored
4e416ee
server: fix token counting API crash on sleep (#29309)
willweimike
authored
42916d8
CUDA: enable sparse-fa for dsv4 prefill (again) (#29298)

am17an
and
ServeurpersoCom
authored
dc9879c
metal: add the missing f32 x bf16 mul_mv variants (#28741)
ServeurpersoCom
authored
9575389
common : keep HF cache dir as path, expose UTF-8 only for logs (#29320)
angt
authored
bddf826
CUDA: add a reserve to avoid spurious warning on older GCC builds (#29317)
am17an
authored
66fba63
ggml : bump version to 0.25.1 (ggml/1637)
ggerganov
committed
e4e2f62
sync : ggml
ggerganov
committed
177cd8c
llama.cpp : bump version to 0.5.0 (#29333)
ggerganov
authored
7fe450e
opencl: add A8 Q6_K non-MoE dp4a binary kernel (#29057)
shaofeiqi
authored
fee39dd
ci : use hf-jobs-cpu-xl runner in server sanitize workflow (#29297)
ggerganov
authored
6e60f35
tests: add `-b/--backend` option to test-llama-archs for testing a specific backend (#27372)

yomaytk
and
JohannesGaessler
authored
d2e5458
server: allow preset to set log file (#29334)
ngxson
authored
b9ae43a
convert : allow vision target for DFlash/Dspark (#29339)
tdakhran
authored
bd4f514

Commits on Sep 24, 2026

Commits on Sep 25, 2026

Commits on Sep 26, 2026

hexagon: find software divide calls using binary inspection tool (#29449)
trivikram-reddy1
authored
4e74811
opencl: add A8 Q8_0 non-MoE dp4a binary kernel (#29439)
shaofeiqi
authored
9f70b2c
ggml-cpu: tiled mul_mat for k-quants (#27851)

jbooth
and
ggerganov
authored
d834d44
polished Readme and llama-bench (#28968)

truecoder34
and
plotnikov.v10
authored
965f897
jinja : support noncall test statements with arg (#29443)
CISC
authored
a1de614
llama : fix K/V and recurrent state cleanup after failed restores (#27530)
CHIPMUNK-T0T
authored
08618ff
jinja : fix compile error (#29468)
CISC
authored
86a24a1
jinja : implement sameas test (#29448)
CISC
authored
81bc6b8
Revert "Change max context length for auto-fitting with unified KV (#28849)" (#29437)
gaugarg-nv
authored
2145525
server : fix wake_fd warning on Windows (#29479)
angt
authored
fcb3074
cuda: add F16 input to the FWHT (#29096)
bri-prism
authored
6f856c7
musa: build the docker images from the PH1 MUSA SDK image (#29481)
yeahdongcn
authored
694ec23
cuda: support Nemotron 3 Puzzle state size 96 for ssm scan (#28717)
anavp-nvidia
authored
9588757

Commits on Sep 27, 2026

hexagon: support for backend sampler (#29502)
max-krasnyansky
authored
2b129cc
hexagon: support tiled Q4_0 and Q8_0 GET_ROWS (#29511)

kurquhar
and
max-krasnyansky
authored
7ac59a6
hrm : fix layer placement of `z_l_init` weight (#29512)
ggerganov
authored
85ca3b5
llama-bench : fix OOB access of hf_file (#29515)
angt
authored
187664b
ci : enable GGML_SCHED_DEBUG_REALLOC=1 for ctest workflows (#29514)
ggerganov
authored
7fb2b08
RPC: use RDMA completion channel to not spin (#29440)
am17an
authored
d7fb90e
common : throw instead of abort on grammar without llguidance (#29516)
angt
authored
da6c28e
vulkan: fix argsort kernel selection for Adreno (#29469)
0cc4m
authored
cea7462
HIP: Enable fattn-mma kernel on cdna for dkq > 256 for large batch sizes (#28907)
IMbackK
authored
2ebd9ae
CUDA: tune fp16 tile FlashAttention configs for head sizes 40-112 (#26289)
animeshsri14
authored
36d7b08
sycl: FWHT kernels for block widths above 512 (#29243)
bri-prism
authored
c829670
opencl: refine bin kernel loading condition (#29503)
lhez
authored
c9064dd
jinja : add support for dict builtin (#29477)
CISC
authored
33c923d
ci : bump ty to 0.0.84 (#29529)
CISC
authored
6fd50a4
convert : export YaRN scaling parameters for PLaMo-3 (#29528)
tokinasin
authored
9adc7f4
jinja : support sequence indices in selectattr and rejectattr
bitalov
committed
ecf9741
model : add K2 Horizon dense and MoVA support
bitalov
committed
63dded7
chat : support K2 Horizon reasoning and tool calls
bitalov
committed
93819d9
Merge the original K2 Horizon branch into the validated integration
bitalov
committed
d70e235
conversion: remove obsolete K2 Aurora alias
bitalov
committed
f13798d

Commits on Sep 28, 2026

k2-horizon: enforce response schemas and load YaRN betas
bitalov
committed
540e938