Repository navigation

Merge upstream master into prism (2026-10-11) - #352

Open

bri-prism wants to merge 945 commits into

prismfrom

merge/upstream-20261011

Open

Merge upstream master into prism (2026-10-11)#352
bri-prism wants to merge 945 commits into
prismfrom
merge/upstream-20261011

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

Commits on Oct 1, 2026
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
Commits on Oct 3, 2026
Commits on Oct 4, 2026
- webgpu: add f16 support to fill/set_rows (#29897)
  yomaytk
  authored
  11fe021
- ci : improve release flow (#29913)
  CISC
  authored
  f98b31c
- ci : pushing tag needs deploy key (#29937)
  CISC
  authored
  0faee50
- server : fix dead LLAMA_ARG_HF_REPO_FILE key in preset allow-list (#29938)
  angt
  authored
  bf9a0cc
- common : prepare load_from_models_dir() for path conversion (#29674)
  angt
  authored
  6716df6
- spec : fix n-gram drafts rejected at temp > 0 after truncation (#29924)
  
  praneshgo
  and
  Pranesh Gonegandla
  authored
  8330e96
- imatrix: calculate activation-based statistics for new format (GGUF) imatrices (#14891)
  
  authored
  0504396
- vulkan: fix rdna4 mat_vec tuning (#29934)
  0cc4m
  authored
  16c163d
- CUDA: fix MMQ memory fault if n_expert >> n_ubatch (#29941)
  JohannesGaessler
  authored
  dd26678
- cuda : move blocks_per_col to where it is used (#29939)
  angt
  authored
  2bc5635
- ci : set default permissions (#29945)
  CISC
  authored
  46847e6
- chat-peg-parser : clear current_tool when pending_tool_call is reset (#29942)
  T-Anas
  authored
  dbe4c3e
- AGENTS.md : revamp (#29656)
  am17an
  authored
  bf79dbb
- ci : add windows arm64 vulkan release (#29954)
  CISC
  authored
  7f2dd88
- ci : windows llvm build requires ninja multi-config (#29959)
  CISC
  authored
  2e7c58c
- cuda : move neu_padded to where it is used (#29940)
  angt
  authored
  0eb6d9a
- ggml-cpu: support BF16/FP16/FP32 K tails in tinyBLAS on x86 (#29806)
  SongXiaoXi
  authored
  a7b94df
- CUDA: refactor swizzling code (#29612)
  JohannesGaessler
  authored
  2ca15f5
- llama: support both embd + raw tokens in batch (#29622)
  ngxson
  authored
  0bb496d
- log, server: self contained colors, split child commands from logs in router mode (#29895)
  
  ServeurpersoCom
  and
  ngxson
  authored
  a7fb71f
Commits on Oct 5, 2026
Commits on Oct 6, 2026
Commits on Oct 7, 2026
- ggml-webgpu: no dawn native features on wasi (#27069)
  MendyBerger
  authored
  f498f86
- sycl: add IQ3_S multi-column MMVQ (#29500)
  clemenswasser
  authored
  78651c4
- vulkan: fix amd iGPU slow checkpoint read (#30049)
  0cc4m
  authored
  4d756bc
- sycl : fattn_kv_buffers cleanup (#27689)
  Pratyush-gg
  authored
  5e5b628
- sycl: accelerate GLM MLA prefill with MKL flash attention (#29171)
  anantshri
  authored
  d2a79e6
- [SYCL] fix the issue in mixed different model GPUs in FA (#29071)
  
  arthw
  and
  ggerganov
  authored
  005a1e1
- tests : retain the anchor when testing recurrent rollback (#29923)
  eapache
  authored
  fa3c2fa
- ggml-webgpu: fix flash_attn supports_op check for overlapping KV (#28205)
  yomaytk
  authored
  36a7391
- imatrix : include clocale for std::setlocale (#30079)
  
  dazzywi
  and
  yarik
  authored
  2690873
- vendor : update cpp-httplib to 0.60.0 (#30081)
  cabelo
  authored
  b7dafa0
- musa: use the tile lightning indexer kernel (#30080)
  yeahdongcn
  authored
  ad21565
- convert : fix token configuration for PLaMo-3 (#29843)
  tokinasin
  authored
  7481354
- vocab : add plamo fim tokens (#30090)
  CISC
  authored
  42b021b
- sampling : use greedy selection for eligible temperature-zero chains (#29797)
  
  hthadicherla
  and
  ggerganov
  authored
  d0b490f
- qwen3tts : guard speaker_encoder_config patch for CustomVoice variant (#29179)
  SIDDARTHAREDDY8
  authored
  48499d2
- feat: add GLM5Next MTP, optimize (#29928)
  
  authored
  b9acf13
- metal : fix MUL_MAT+ADD fusion when the residual is itself a MUL_MAT (#30100)
  
  nyo16
  and
  ggerganov
  authored
  7e8324f
- metal : few-row MMA mat-mul for the remaining src0 types (#30065)
  pratiknarola-t
  authored
  9881906
- llama: share the nextn tensor flags between models (#30097)
  ServeurpersoCom
  authored
  448147d
- server : accumulate generated text and tokens as parse input (#29876)
  aldehir
  authored
  18b5f8b
- cuda: update uncoalesced memory reads in pool2d (#29425)
  
  shenron0101
  and
  ggerganov
  authored
  42c787e
- llama : add a GPU cache for MoE experts kept in host memory (#29887)
  am17an
  authored
  d6cf9ac
- mtmd: add cohere2 vision support (#30062)
  Terrencezzj
  authored
  50a6c5c
- cuda: FWHT kernels for block widths above 512 (#29100)
  bri-prism
  authored
  b86d2f0
- model : add LiquidAI/d1-3B decision model (#30110)
  tdakhran
  authored
  88dcc46
- chat : name tool and argument parser rules by index (#30088)
  
  Frost-54
  and
  aldehir
  authored
  5de7334
- chat: fix jinja parser for TranslateGemma (#30096)
  nbud
  authored
  bd4eeaa
- hexagon: improved GELU accuracy (#30104)
  kurquhar
  authored
  7081510
Commits on Oct 8, 2026
- hexagon: support tiled Q4_K and Q6_K GET_ROWS (#30115)
  kurquhar
  authored
  aa5e009
- model : add LiquidAI/d1-omni-600M decision model (#30114)
  tdakhran
  authored
  a657f7e
- hexagon: Q6_K weight dequant speedup (#30121)
  kurquhar
  authored
  06cad0b
- hexagon: enable alloc_buffer_n (#30126)
  
  max-krasnyansky
  and
  jhen0409
  authored
  9c2e0e4
- sycl: remove duplicate block-size defines from op headers (#29507)
  Titaniumtown
  authored
  9b4ed0c
- convert : support Qwen3.5 embedding models (#27920)
  SamMalayek
  authored
  75118a3
- sycl: add grouped MoE XMX GEMM (#29245)
  
  cwriter
  and
  cwriter
  authored
  847f447
- ggml-cuda: assign four GDN state columns per warp (#30087)
  SongXiaoXi
  authored
  24e4183
- model : support classifier_activation for rerankers (#29692)
  boshjerns
  authored
  37ac634
- sycl: stage bulk uploads (model loading) through a pinned ring buffer (#29608)
  
  cwriter
  and
  cwriter
  authored
  000bee5
- sycl: fuse the delta-net alpha gate (add + unary + mul) (#29687)
  Titaniumtown
  authored
  ff30363
- CUDA: fix norm family kernels when ne[2]/ne[3] exceed grid dim limits (#28175)
  edenfunf
  authored
  fda1866
- cuda : Use byte strides for roll to allow non-contiguous ROLL operations (#29547)
  bertaye
  authored
  d888016
- hex-mmadd: do not assume aligned read/write when bias-add is fused (#30133)
  max-krasnyansky
  authored
  097f5b5
- sycl: FWHT optimizations (#29605)
  bri-prism
  authored
  46baf1f
- cuda : support arbitrary striding for unary ops on f16, f32, and bf16 (#29781)
  
  saady789
  and
  JohannesGaessler
  authored
  08246a2
- vendor : update cpp-httplib to 0.60.1 (#30134)
  angt
  authored
  03aa006
- vulkan: extend sparse FA support to coopmat2 (#30003)
  jeffbolznv
  authored
  dac3087
- vulkan : fix TOP_K for +inf/NaN inputs and k = 1 on negative values (#30107)
  gianni-cor
  authored
  ff5888f
- server : preserve context checkpoints across slot save/restore (#26004)
  
  Tough-Respawn
  and
  okigan
  authored
  033df86
- llama: support MoE cache over multiple GPUs (#30112)
  am17an
  authored
  c811cb8
- ui: apply ui_settings on first visit in router mode (#29668)
  simonether
  authored
  4f92965
- CUDA: fix CCCL version guard breaking on major version rollover (#29453)
  
  hey-gm
  and
  ORippler
  authored
  1167d3f
- CUDA : looped PAD kernel for more than 65535 rows or slices (#30147)
  pskrunner14
  authored
  c35b667
- CUDA: fix MMQ out-of-bounds reads (#29953)
  JohannesGaessler
  authored
  fc9ce6b
- model : fix DFlash output head sharing (#30111)
  hthadicherla
  authored
  a11f57b
- CUDA: improve top-k algorithm selection (#28713)
  
  praneshgo
  and
  Pranesh Gonegandla
  authored
  71ad059
- Musa FWHT fix (#30167)
  bri-prism
  authored
  de7fa0a
Commits on Oct 9, 2026
Commits on Oct 10, 2026
Commits on Oct 11, 2026
Commits on Oct 12, 2026
- dspark: fix flake8 errors in the DSpark conversion scripts
  bri-prism
  committed
  20d603c
- dspark: log through logging instead of print in the DSpark/DFly scripts
  bri-prism
  committed
  31eb081