All pull requests

Pull requests

• vulkan: Add workaround for device free memory > total memory on Intel Windows

  ggmlVulkan
  #29835·rillomas opened Oct 2, 2026Contributor·

  ·
  1

• server : reject cross-origin and non-JSON requests to /tools

  server
  #29834·Snakinya opened Oct 2, 2026·1 task done·

  ·
  1

• server: model-agnostic impl of systemone/decision API

  server
  #29832·ngxson opened Oct 2, 2026Collaborator·

  ·

• model: add support for clef decision model

  conversiondocumentationmodelservertesting
  #29831·ngxson opened Oct 1, 2026Collaborator·3 tasks·

  ·

• qwen4exp : halve the indexer score memory

  model
  #29825·ServeurpersoCom opened Oct 1, 2026Contributor·

  ·
  6

• qwen4exp : optimize mask constructions

  Apple Metalggmlmodel
  #29824·ggerganov opened Oct 1, 2026Member·

  ·
  1

• model : add support for Ouro (ByteDance looped LM)

  conversionmodel
  #29823·BrandeisPatrick opened Oct 1, 2026·

  ·

• vulkan: compute every row of an expert in mul_mm_id when ids repeat

  ggmltestingVulkan
  #29822·Anjielon opened Oct 1, 2026Contributor·

  ·
  1

• llama, server: add /v1/systemone API (models: laya, julia-1, lev, openjev, kev)

  conversiondocumentationmodelserver
  #29818·ngxson opened Oct 1, 2026Collaborator·3 tasks done·

  ·
  2

  

• ggml-quants : avoid invalid rounding in qkx3 scale search

  ggmltesting
  #29817·devYRPauli opened Oct 1, 2026Contributor·

  ·
  1

• metal : page-align the destination buffer in ggml_metal_get_tensor_async

  Apple Metalggmltesting
  #29814·Ethan0456 opened Oct 1, 2026·

  ·
  1

• chat : honor json_schema in Ling 3.0 parser

  testing
  #29813·devYRPauli opened Oct 1, 2026Contributor·

  ·
  1

• cpu: offset the flash attention KQ max to keep F16 sums in range

  ggml
  #29810·ServeurpersoCom opened Oct 1, 2026Contributor·

  ·
  1

• sycl : accelerate MXFP4 MoE with arithmetic decoding and weight reordering

  ggmlSYCL
  #29809·Captain-Tripps opened Oct 1, 2026·

  ·

• CUDA: Remove redundant CUDA copies after SSM_SCAN

  CUDAggmltesting
  #29807·anavp-nvidia opened Oct 1, 2026Contributor·

  ·

• ggml-cpu: support BF16 K tails in llamafile tinyBLAS

  ggml
  #29806·SongXiaoXi opened Oct 1, 2026Contributor·

  ·

• ggml : prune deprecated API

  Ascend NPUCUDAggmlOpenVINOSYCLtestingVulkan
  #29801·ggerganov opened Oct 1, 2026Member·

  ·
  1

• metal : add GGML_OP_CROSS_ENTROPY_LOSS

  Apple Metalggmltesting
  #29800·wrafique007 opened Oct 1, 2026·

  ·

• sampling : use greedy selection for eligible temperature-zero chains

  merge readytesting
  #29797·hthadicherla opened Oct 1, 2026Contributor·

  ·
  2

• llama : avoid redundant synchronization in output getters

  CUDAggmlSYCLtesting
  #29796·hthadicherla opened Oct 1, 2026Contributor·

  ·
  9

• HIP: use -O0 for host code in debug builds

  ggml
  #29795·JohannesGaessler opened Oct 1, 2026Contributor·

  ·

• vulkan: add logging to pipeline compile issues

  ggmlVulkan
  #29794·0cc4m opened Oct 1, 2026Contributor·

  ·

• vulkan: add fp8 and scaled matmul support

  ggmlVulkan
  #29791·0cc4m opened Oct 1, 2026Contributor·5/5·

  ·

• opencl: use sigmoid f16 for bf16

  ggmlmerge readyOpenCL
  #29787·lhez opened Oct 1, 2026Contributor·

  ·

• common : sum all full-throughput perf levels for the Apple thread default

  testing
  #29784·JamesFranc opened Oct 1, 2026·

  ·
  1