Repository navigation

Sync master with upstream (2026-11-02) - #172

Merged

jimw567 merged 424 commits into

jimwu.master-sync-upstream-1102

Nov 3, 2026

Merged

Sync master with upstream (2026-11-02)#172
jimw567 merged 424 commits into
masterfrom
jimwu.master-sync-upstream-1102

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

Commits on Oct 22, 2026
Commits on Oct 23, 2026
- sycl: Add graph record and replay (#28725)
  malsbat
  authored
  7377bad
- sycl: Replace host waits in fattn-mkl with explicit dependencies (#30684)
  malsbat
  authored
  0a9ef00
- hexagon: Op Fusion refactoring, optimizations and new fusions like `BINARY_NX` (back-to-back ADDs, etc) (#30843)
  max-krasnyansky
  authored
  643b0cd
- vocab : fix Nemotron-3.5-Embed tokenization (#30974)
  
  authored
  f9599d0
- Extend probabilistic sampling for DFlash2 drafts (#30663)
  
  praneshgo
  and
  Pranesh Gonegandla
  authored
  1ab09d6
- sycl : pass device into mul_mat_uses_onednn (#31014)
  Titaniumtown
  authored
  f64a498
- CANN: missing `const` qualifiers added for consistency (#30933)
  rauletorresc
  authored
  f5027fb
- ci : add behavior check workflow (#31036)
  CISC
  authored
  f00c908
- ggml : fix time overflow in ggml_time_us (#30969)
  frozenblade1224
  authored
  4231b23
- ui : bump packages to clear npm audit (#31038)
  allozaur
  authored
  e1e3b2c
- hexagon: add CEIL op (#31006)
  max-krasnyansky
  authored
  36e36e6
- CUDA: chunk graphs to amortize graph launch (#30804)
  am17an
  authored
  280141e
- vendor : update LibreSSL to 4.4.0 [no ci] (#31049)
  angt
  authored
  da2ee1e
- qwen4exp : build the QSA selection without the kq_mask (#30965)
  sanmai
  authored
  5c31142
- allows the usage of 4-10 gpus for stepfun and laguna in tensor parallel (#24554)
  krampenschiesser
  authored
  06230cb
- CUDA: fix build after chunked CUDA graphs (#30804) (#31053)
  
  sreeram-11
  and
  JohannesGaessler
  authored
  9ff0d7e
- ggm-cuda: Enforce DP4A MMQ tile height multiple of warp size (#30837)
  thelittlefireman
  authored
  1a24df1
- hip: fix bf16 WMMA operand vector type for gfx11/gfx12 (#30885)
  waltercool
  authored
  9145ddb
- common : allow concurrent downloads across processes (#29312)
  angt
  authored
  21ec307
- hexagon: add SGN op (#30346)
  
  kurquhar
  and
  max-krasnyansky
  authored
  0ae0e19
- mtmd : add Soprano TTS support (#28607)
  hans00
  authored
  9ae73e6
Commits on Oct 24, 2026
Commits on Oct 25, 2026
- model: DSV4.1 flash (#31086)
  
  am17an
  and
  danielhanchen
  authored
  74bca39
- chat : reject JSON nested more than 128 levels before walking it (#30242)
  
  ykhrustalev
  and
  ngxson
  authored
  b3fa699
- ci : update CPU job assignments (#31136)
  ggerganov
  authored
  344be0a
- opencl: add initial q2_k support (#30999)
  
  lhez
  and
  wanghqc
  authored
  b083c38
- opencl: add topk and lightning indexer (#31026)
  lhez
  authored
  9a4580b
- hexagon: generic dims collapse/reshape for unary, binary and matmul ops (#31071)
  max-krasnyansky
  authored
  f0fb598
- vulkan : one warp tile per subgroup in the small matmul tiles (#30106)
  ianloic
  authored
  459a1ba
- vulkan: Add a larger tile size for conv3d (coopmat2 only) (#29920)
  jeffbolznv
  authored
  e321bda
- ggml-cuda: fuse qwen MRoPE + set-rows (#31025)
  am17an
  authored
  a55c16e
- CUDA: fuse FFN gate/up matmuls and GLU in MMQ with interleaved gate/up rows (#29948)
  anujj
  authored
  c49edf0
- tools : reject truncated GGUF tensor data in merge (#31044)
  
  shreya-11733
  and
  Shreya Agrawal
  authored
  7790a92
- llama : respect no_alloc in all memory caches (#31087)
  ggerganov
  authored
  fa3645f
- grammar: limit recursion depth to prevent stack overflow (#29495) (#29755)
  Hrushikesh-ramilla
  authored
  f1168ac
- llama-quant : fix 4D tensor quantization using ggml_nrows() and name-based expert detection (#30820)
  anthony-wss
  authored
  5c965fd
- llama: fix exhausting CPU and MEM caused by dry sequence breaker (#31035)
  trunk-trick
  authored
  63412d8
- ci : disable test-llama-context with fa off in the webgpu (#30977)
  yomaytk
  authored
  1a0b56c
- qwen: merge tensors on convert (#31028)
  am17an
  authored
  2280f0f
- tests : adjust default threads for test-llama-context (#31157)
  ggerganov
  authored
  6050b46
- common, grammar: add recursive depth limit, make it a requirement (#31153)
  
  authored
  48c9f08
- cuda : use std::string instead of VLA for archName (#30463)
  angt
  authored
  d854d98
Commits on Oct 26, 2026
Commits on Oct 27, 2026
Commits on Oct 28, 2026
Commits on Oct 29, 2026
Commits on Oct 30, 2026
- common : end reasoning budget forcing on an end tag from the prompt (#31374)
  tdakhran
  authored
  3c809f4
- opencl: add initial q2_0 support (#31340)
  
  shaofeiqi
  and
  lhez
  authored
  e0f142d
- memory: refuse rollbacks past the snapshots of the last ubatch (#31368)
  ServeurpersoCom
  authored
  ab3768c
- kv-cache : keep rotation constants on the GPU (#31073)
  
  anton-iermolenko
  and
  ServeurpersoCom
  authored
  fb59972
- gemma3n, gemma4 : create token_embd before the tied output (#31394)
  m4zar
  authored
  344846b
- ci : extract most apple jobs into daily workflow (#31408)
  danbev
  authored
  e216f4e
- cmake : embed an app manifest in all Windows exes (#31329)
  angt
  authored
  d63fbfe
- ci : add missing ccache-clear to windows behavior check (#31413)
  CISC
  authored
  27b8464
- vendor : update cpp-httplib to 0.64.0 (#31400)
  cabelo
  authored
  1ab23c3
- quant : allow quantization for LFM3-Audio codec decoder (#31386)
  tdakhran
  authored
  889f2ec
- common : drop the Windows argv reconstruction (#31417)
  angt
  authored
  aa588a3
- common : delete fs_path_to_utf8() (#31423)
  angt
  authored
  a34de86
- opencl: add initial iq1 support (#31341)
  
  shaofeiqi
  and
  lhez
  authored
  792fcf4
- rpc : delete fs_path_to_utf8() (#31424)
  angt
  authored
  5047748
Commits on Oct 31, 2026
- sycl: fix f32 to q2_0/q4_0/q5_0 rounding in CPY (#31120)
  Torchit1
  authored
  8956c76
- chat : strip the think block in the qwen3-coder with reasoning_format none (#31380)
  tdakhran
  authored
  1052f7e
- opencl: add fwht (#31434)
  
  shaofeiqi
  and
  lhez
  authored
  0ec4027
- mtmd : resize Gemma 4 images without letterbox padding (#30922)
  
  boshjerns
  and
  ngxson
  authored
  7bd1808
- server : fix /v1/responses endpoint to accept extra_params (#30959)
  Dgarc359
  authored
  2871678
- server: cap shared decision prefixes at the first output (#30710)
  aiqubits
  authored
  f3586e6
- reshaping so mm on 2d instead of 3d tensor for lfm sc (#30831)
  anthonymaltsev
  authored
  ddf48f8
- llama : fix rotation of attention tensors in layers with shared KV (#31406)
  nikwen
  authored
  0138937
- vendor : update miniaudio to 0.11.26 (#31459)
  cabelo
  authored
  f6312ad
- llama : enable tensor split for LFM3 (#31431)
  tdakhran
  authored
  bc428e4
- tests : enable Gemma 4 in `test-llama-archs` (#31401)
  nikwen
  authored
  79bce0c
- mtmd : fix InternVL CLS token order (#30886)
  boshjerns
  authored
  d5ef408
- hexagon : fix DMA transfers with strides > 24 bits (#31442)
  
  marktodorovich
  and
  max-krasnyansky
  authored
  a91eab6
Commits on Nov 1, 2026
- opencl: add initial iq3 support (#31338)
  
  shaofeiqi
  and
  lhez
  authored
  060b2fe
- sycl: Try PkgConfig for LevelZero includes / libs (#31076)
  mctylr-gh
  authored
  dde879b
- requirements : fix numpy dependencies (#31508)
  CISC
  authored
  785322d
Commits on Nov 2, 2026
- webgpu : fix CI by keeping scratch alloc sizes monotonic in batch length (#31422)
  ServeurpersoCom
  authored
  1c5ed70
- cmake : add explicit thread linkage (#31414)
  o7si
  authored
  4e910a8
- requirements : minor adjustments for python 3.14 wheels (#31534)
  CISC
  authored
  fcfe4b5
- docs: fix Huawei ascend brand (#31463)
  licy666
  authored
  f72dc30
- ggml : fix clang-cl build with AVX512 BF16 (#31276)
  YagUber
  authored
  a7a490a
- jinja : implement map('filter') (#24033)
  MuoDoo
  authored
  5969566
- tests : merge identical if branches (#31523)
  nikwen
  authored
  00ef136
- docs : fix Vulkan example command in build.md (#31530)
  Septa2112
  authored
  aac6da8
- ui: don't reload a loading model on select, prefer it in pre-selection (#31326)
  allozaur
  authored
  b12b405
- ui: show interrupted state for tool calls that never returned (#31314)
  allozaur
  authored
  64be95b
- ui: keep queued steering messages in their own conversation (#31304)
  allozaur
  authored
  90080ce
- ui: align alert dialog styling with other overlay surfaces (#31312)
  allozaur
  authored
  23426d7
- ui: disable model deletion the server would refuse (#31305)
  allozaur
  authored
  5c2ea1a
- sycl: dmmv code cleanup (#31325)
  Titaniumtown
  authored
  5e18eb0
- sycl : fast ggml_sycl_comm_exchange() on Intel Arc Pro B70s (#30956)
  92es
  authored
  2c93f0b
- vulkan: avoid redundant scans when zeroing skipped rows (#31266)
  SG-Amadeus
  authored
  cb02250
- vulkan: improve small-n split-K selection on RDNA3 and RDNA4 (#30146)
  ravel7524
  authored
  07ec7c8
- vulkan: add tensor-parallel allreduce function (#25051)
  
  authored
  dcbb091
- opencl: add topk_mask (#31441)
  lhez
  authored
  84d9ce0
- ggml : use `vm_allocate` instead of `posix_memalign` on all Apple platforms (#31437)
  ykhrustalev
  authored
  ea24b95
- ggml-zdnn: share buffers with host (#31481)
  taronaeo
  authored
  6ed8120
- docs: CANN backend supports BF16 on 910B, document GGML_CANN_NO_PINNED (#31249)
  li-lizhe
  authored
  458fcfc
- readme : fix links in README.md for maintainer PRs (#31487)
  maternion
  authored
  6b00d56
- context : simplify setters (#29730)
  nikwen
  authored
  cd8427d
- imatrix : fix --nextn crash (#31522)
  sanmai
  authored
  3f8af15
- llama : flash_attn_ext_rows - more efficient unified kv for multi-seq (#29510)
  
  am17an
  and
  ggerganov
  authored
  1b6935b
- model: add Clover support (#31306)
  
  Maphist0
  and
  CISC
  authored
  e4b6c8c
- mtmd: Pocket tts - support for voices without voice cloning (#31429)
  
  authored
  d02b4c0
- docs: update s390x build docs (#31546)
  taronaeo
  authored
  29c720e
- Merge upstream/master into fork master
  
  Jim Wu
  and
  claude
  committed
  dc8bede