All pull requests

Pull requests

• convert : fix LoRA conversion with --outtype auto

  conversiondevopstesting
  #30433·teddygood opened Oct 13, 2026·

  ·

• ggml : add AArch64 NEON path for BF16 matrix operations (#29917)

  ggmltesting
  #30432·djeong20 opened Oct 13, 2026Contributor·

  ·
  1

• vulkan: Define the shader count for Intel Arc Pro B65

  ggmlVulkan
  #30431·TheBlueMatt opened Oct 13, 2026Contributor·

  ·

• sycl : fix get_rows on reordered quantized weights

  ggmlSYCLtesting
  #30430·skipzx-lab opened Oct 12, 2026·

  ·
  1

• ggml-cpu: faster AMX Q4_K decode (+13% tg) by walking each group across columns

  ggml
  #30429·ToukoUrsin opened Oct 12, 2026·

  ·

• fit-params: accept --spec-type, -md, -ngld arguments

  examplestesting
  #30427·claudiubelu opened Oct 12, 2026·

  ·

• server-common.cpp: add fallback for F_DUPFD_CLOEXEC when it is missing

  server
  #30425·barracuda156 opened Oct 12, 2026Contributor·

  ·
  1

• hexagon: add configurable HTP DCVS/HMX power mode

  documentationggmlHexagon
  #30423·pdhinaka opened Oct 12, 2026Contributor·

  ·
  1

• opencl: fix OOB tail access for adreno noshuffle gemv kernels

  ggmlOpenCL
  #30421·lhez opened Oct 12, 2026Contributor·

  ·

• cuda : warnings cleanup

  CUDAggml
  #30420·angt opened Oct 12, 2026Member·

  ·
  1

• opencl: use buf+offset instead of subbuffer for adreno moe expert sort

  ggmlmerge readyOpenCL
  #30418·lhez opened Oct 12, 2026Contributor·

  ·
  1

• common : add fs::path arg handler, fix --log-file on Windows

  #30417·angt opened Oct 12, 2026Member·

  ·

• ggml-rpc : fail gracefully on unreachable endpoints and failed handshakes

  examplesggmltesting
  #30415·figkcoci-debug opened Oct 12, 2026·

  ·
  2

• opencl: admit K = 256 weights to the Adreno kernels on the X2-90 and 830/840

  ggmlOpenCL
  #30414·wanghqc opened Oct 12, 2026Contributor·

  ·

• opencl: read each weight row once in the small-N f32/f16 GEMV

  ggmlOpenCL
  #30413·wanghqc opened Oct 12, 2026Contributor·

  ·

• opencl: run the tiled q4_K lm_head through a tiled GEMM at batch > 1 (bug fix)

  ggmlOpenCL
  #30412·wanghqc opened Oct 12, 2026Contributor·

  ·

• Clarify that the intel oneapi toolkit already has all necessary dependencies for sycl

  documentationSYCL
  #30407·alosslessdev opened Oct 12, 2026Contributor·

  ·

• llama : do not map foreign ranges into a backend buffer

  #30405·ggerganov opened Oct 12, 2026Member·

  ·

• mtmd : avoid padding-only Whisper audio chunks

  mtmdtesting
  #30403·hera2019 opened Oct 12, 2026·1 task done·

  ·

• metal : add IM2COL_3D

  Apple Metalggml
  #30400·luka-topi opened Oct 12, 2026·

  ·

• fix: bounds check in ggml-rpc rdma_recv

  ggml
  #30399·sohaibharraoui opened Oct 12, 2026·

  ·
  1

• Allow Qwen3.8 Flash Next to accept image embds for MTP

  model
  #30397·ORippler opened Oct 12, 2026Collaborator·

  ·
  1

  1

• models: support vllm decision 2 qwen 3.5 variant

  conversionmodel
  #30393·ngxson opened Oct 12, 2026Collaborator·

  ·

• server: include version info in the "initializing ..." log line

  server
  #30392·kucharskim opened Oct 12, 2026Contributor·

  ·
  1

• unicode : add opt-in NFC normalization for Unigram tokenizers

  conversionmerge readytesting
  #30390·sandrohanea opened Oct 12, 2026Contributor·

  ·
  9