Repository navigation

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

Closed

bitalov wants to merge 724 commits into

ifm-ai:model/K2Horizonfrom

bitalov:k2-horizon-all-fixes-20260927

Closed

K2 Horizon: integrate all fixes on current upstream#5
bitalov wants to merge 724 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 23, 2026
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
Commits on Sep 28, 2026
Commits on Sep 29, 2026
Commits on Sep 30, 2026
- ggml-zdnn: impl buffer reset, fix memory leaks (#29637)
  taronaeo
  authored
  eae11d2
- vendor: update BoringSSL to 0.20260929.0 (#29669)
  cabelo
  authored
  931351e
- add GLM-5.3-Flash (GLM5-Next) support (#27773)
  
  authored
  649dcb1
- SYCL: reduce tensor allreduce sync with pinned host buffers (#29604)
  Captain-Tripps
  authored
  2a53ace
- ci : add models backend check (#29651)
  CISC
  authored
  72db1e0
- musa : define __CUDA_ARCH__ for device passes (#29508)
  yeahdongcn
  authored
  272aad8
- ci : fix Fusion / metal by adding glm5-next to MTL.csv (#29712)
  ServeurpersoCom
  authored
  db00347
- openvino: serve GET_ROWS on a weight view from the base Constant (#28381)
  ServeurpersoCom
  authored
  25747b0
- ui : type-safe API types, fetch helpers and download-ready models store plumbing (#29582)
  
  allozaur
  and
  ServeurpersoCom
  authored
  fa2bde5
- ui : model id grammar for sidecars, quants and capability parsing (#27946)
  allozaur
  authored
  f653250
- ui : Hugging Face Hub data layer (#27947)
  allozaur
  authored
  9b43336
- ui : model memory-fit estimation (#27957)
  allozaur
  authored
  4cfb6d1
- ui : model download pipeline (#27959)
  allozaur
  authored
  8664eae
- ui : shared model display primitives (#29644)
  allozaur
  authored
  4a096b8
- model-conversion : add --add-bos to run org model script (#29558)
  danbev
  authored
  8df332d
- renaming template fixture
  bitalov
  committed
  cbee4a2
- cpu: accept BF16 in src1 of mul_mat (#28937)
  ServeurpersoCom
  authored
  90c908d
- constants conflict fix
  bitalov
  committed
  757e62e
- ggml : add BF16 unary, GLU, binary and scale ops (CPU, CUDA) (#29675)
  am17an
  authored
  2090f60
- llama: llama_prefetch_rows (#29599)
  
  authored
  185103d
- ci : fix Models Backend Check by shortening the hrm_text fixture (#29744)
  ServeurpersoCom
  authored
  3b3d022
- ggml-et : remove useless alloca() (#29663)
  angt
  authored
  bdeb855
- jinja : support coerced array attributes (#29574)
  CISC
  authored
  ca2e203
- mimo : support dflash (convert + feature extraction) (#29650)
  
  ggerganov
  and
  CISC
  authored
  22bdcc4
- cli: exit on stdin EOF and drop the console wide Ctrl+C broadcast (#29722)
  ServeurpersoCom
  authored
  b046420
- codeowners : remove former ZenDNN owner (#29747)
  
  vishalMCE
  and
  Vishal Singh
  authored
  876c75b
- ggml/gguf : fix integer overflow (#29384)
  apach301
  authored
  2149c00
- glm5-next: give dead indexer slots unique scatter rows (#29745)
  ServeurpersoCom
  authored
  05af0d2
- batch: migrate the rest of examples to llama_batch_ext (#29601)
  ngxson
  authored
  60e9cf7
- llama: properly handle KV on training (#28520)
  ngxson
  authored
  81ff93e
- convert : fix LoRA conversion crash for Qwen3.5 V-head reorder (#28324)
  
  authored
  b016f46
- test-llama-archs : toggle causal_attn to catch graph shape changes (#29724)
  sihanyu03
  authored
  4f31296
- llama : preserve original batch order for speculative decoding layer inputs (#29019)
  hthadicherla
  authored
  4453b53
- args: fix cli download mmproj arg (#28977)
  pr3pony
  authored
  feb9a3d
- Hexagon: optimize ALLREDUCE with support for safe scatter mode (#29757)
  
  ebateni
  and
  max-krasnyansky
  authored
  a4d880f
- cuda: guard the iq4_nl dequantize row kernel against short rows (#29683)
  yeahdongcn
  authored
  f872b59
- ggml-opencl : replace alloca() with std::vector (#29765)
  angt
  authored
  f7b384c
Commits on Oct 1, 2026
- webgpu: fix SSM_SCAN binding aliasing (#29750)
  yomaytk
  authored
  0c1e570
- model : re-enable -sm tensor for qwen4exp (#28569)
  kh0pper
  authored
  10f340d
- docs : refresh CPU ops support matrix (#29666)
  CaramelizedCUDA
  authored
  66bcc27
- llama-bench : fix docs (#29464)
  mairp
  authored
  79625e0
- metal : use bf16 math for mxfp4 mul-mat (#29770)
  ggerganov
  authored
  2232bc8
- llama-bench : fix verbosity filter to show GGML_LOG_ERROR (#28229)
  kushalgarg101
  authored
  7dad6db
- vocab : honor BOS/EOS settings for PLaMo-2 and PLaMo-3 (#29734)
  
  tokinasin
  and
  CISC
  authored
  db33d3c
- opencl: mark vec subgroup bcast as supproted for Adreno E17 compiler (#29698)
  lhez
  authored
  3ec4df4
- common : add LLM-jp-4.1 Harmony dialect handler (#29681)
  e-mon
  authored
  b8f96c3
- BLAS : Document AOCL-BLAS build and label the device AOCL-BLAS (#29640)
  pradeeptrgit
  authored
  b0aca3c
- hex-workqueue: fix race condition in seqn getting out of sync with idx_read/write (#29785)
  max-krasnyansky
  authored
  3aa0ce9
- HIP: avoid treating CDNA as dgx spark for gqa_ratio 20 in fattn_mma dqk 576 (#29572)
  IMbackK
  authored
  f11d642
- llama-mmap : avoid a second full-size copy of each tensor with direct-io (#29749)
  
  praneshgo
  and
  Pranesh Gonegandla
  authored
  32dd62e
- jinja : skip copying loop scope unless a loop filter needs it (#29776)
  xarillian
  authored
  def4d40
- meta: clear inactive AllReduce shards with FILL, not SCALE (#29793)
  ggerganov
  authored
  5503b04
- mtmd: cap max_image to ubatch for non_causal models (#29773)
  ngxson
  authored
  552f18f
- adressing cisc comments
  bitalov
  committed
  fe498bd
- CUDA: Make CCCL configurable + pin it to 3.4.3 for CI jobs (#29792)
  ORippler
  authored
  7677678
- llama: fix qwen4exp (#29751)
  am17an
  authored
  66e0c17
- Qwen4Exp: add MTP (#29761)
  
  am17an
  and
  ggerganov
  authored
  c061df1
- CUDA: Handle compute type for NVFP4 on cublass path (#29173)
  
  ynankani
  and
  JohannesGaessler
  authored
  b56f34a
- llama : fix invalid assert in recurrent memory (#29799)
  ggerganov
  authored
  869034b
- webgpu: add bfloat16 support for MUL_MAT/MUL_MAT_ID/GET_ROWS- #29358 (#29358)
  yomaytk
  authored
  4b1622a
- metal : release temporary private transfer buffers (#29777)
  
  mvanlamz
  and
  nikwen
  authored
  13b4d71
- cuda : route sm70 to the Turing MMVQ nwarps table (#29753)
  
  tkittich
  and
  JohannesGaessler
  authored
  42d9581
- convert : write Gemma embedding scale for DFlash drafts (#29802)
  kabu1204
  authored
  2b36825
- server: return HTTP 400 for invalid embedding requests (#29060)
  SamMalayek
  authored
  d775ebf
- hexagon: shared strided DMA copy for CPY and CONCAT, any-dim CONCAT via DMA (#29685)
  
  njsyw1997
  and
  max-krasnyansky
  authored
  dcd387a
- llama : clamp kpool re-pool bound to existing pools (#29805)
  ggerganov
  authored
  81e39ad
- ci: fix Fusion / metal by updating the qwen4exp baseline (#29812)
  ServeurpersoCom
  authored
  e358d59
- skill: note about model-specific CLI arguments + testings (#29808)
  ngxson
  authored
  68e79bd
- common,rpc : fix cache dir creation through symlinks on buggy libstdc++ (#29816)
  angt
  authored
  f1cee99
- llama: refer to segment documentation [no ci] (#29074)
  JohannesGaessler
  authored
  78e2964
- CUDA: fix 2 broken Volta FA cases (#29803)
  JohannesGaessler
  authored
  ec7630a
- adressing cisc follows ups
  bitalov
  committed
  39ddb2b
- hexagon: add q2_k and q3_k quant type support (#29717)
  
  jhen0409
  and
  max-krasnyansky
  authored
  a868c3e
Commits on Oct 2, 2026
- qwen4exp: fix tests (#29819)
  am17an
  authored
  159c651
- hexagon: install rebuilt HTP skels (#29828)
  kurquhar
  authored
  5fc4f3c
- pyproject : add linux platform marker to uv torch source (#29177)
  Yezat
  authored
  207bdab
- vulkan: add logging to pipeline compile issues (#29794)
  0cc4m
  authored
  fb4b273
- ci : fix missing zdnn backend check (#29837)
  taronaeo
  authored
  254b177
- ggml : add `alloc_buffer_n` to buffer type interface (#23671)
  ggerganov
  authored
  631109b
- qwen4exp : optimize mask constructions (#29824)
  ggerganov
  authored
  4e2713c
- sycl : do not use slow oneDNN reference matmul and fattn (#28985)
  lslusarczyk
  authored
  c328acc
- sycl: large register file for D=512 FA vec kernels (#29062)
  Titaniumtown
  authored
  b933289
- vulkan: disable large matmul tile on Samsung GPUs with 32KB shared memory (#28531)
  jwsong98
  authored
  9e258a6
- SYCL: Q8_0 DMMV ESIMD and MMVQ wide load (#29186)
  
  cwriter
  and
  cwriter
  authored
  392ded6
- opencl: use sigmoid f16 for bf16 (#29787)
  lhez
  authored
  a8c9a4e
- llama : use GGML_ABORT instead of throw (#29840)
  angt
  authored
  6805ae3
- llama : silence unused-result warnings (#29839)
  angt
  authored
  8d81559
- common : remove fs_open_ifstream() by using u8path() (#29841)
  angt
  authored
  70849ee
- llama, server: add /v1/systemone API (models: laya, julia-1, lev, openjev, kev) (#29818)
  ngxson
  authored
  a4cb4c6
- metal : add tensor API flash attention kernel for F16 KV (#29570)
  ethanhq
  authored
  926862e
- readme : add cmd install commands (#29850)
  ggerganov
  authored
  d8fbd25
- model: support nimble decision model (#29844)
  ngxson
  authored
  46ca246
- ggml-cpu : fix soft_max_back wrong output when dst aliases src1 (#27096)
  
  devYRPauli
  and
  ggerganov
  authored
  dd4c286
- ggml-quants : avoid invalid rounding in qkx3 scale search (#29817)
  devYRPauli
  authored
  2923cf2
- ggml-cuda : fix cpy transposed path corrupting non-contiguous dst (#27663)
  devYRPauli
  authored
  134b2bb
- spec : add probabilistic sampling for simple draft and MTP (#27694)
  
  praneshgo
  and
  Pranesh Gonegandla
  authored
  1fb7ef3
- ci : use t4-medium for cuda jobs (#29842)
  CISC
  authored
  4ebdf2c
- CUDA: fuse shared experts into MMVQ (#29184)
  am17an
  authored
  bed0a85
- desloppify the parser / adress aldehir comments
  bitalov
  committed
  336cc79
- clean test-chat
  bitalov
  committed
  59db596
- remove fallback : model trained mostly on high anyway
  bitalov
  committed
  fbee594
Commits on Oct 3, 2026
- model: add support for clef decision model (text-only) (#29831)
  
  ngxson
  and
  CISC
  authored
  99b9548
- qwen4exp : halve the indexer score memory (#29825)
  
  ServeurpersoCom
  and
  0cc4m
  authored
  889edf4
- model : Add LFM2.5-Encoder-350M and LFM2.5-Encoder-230M (#29862)
  tdakhran
  authored
  cb7934c
- ggml-openvino: update to 2026.4.1, optimize performance, expand ops, improve device listing. (#29852)
  
  authored
  b92761a
- graph: gather the recurrent states once so the reserve covers every split (#29856)
  ServeurpersoCom
  authored
  436f6f8
- fix conflicts
  bitalov
  committed
  50abacf
Commits on Oct 4, 2026
- fix k2 attn_v_exp tn splitting and metal fusion baseline
  bitalov
  committed
  c9b78f7