All pull requests

Pull requests

• Add max allocation size reporting and split MoE cache groups

  #31139·sasa7812 opened Oct 25, 2026·

  ·
  1

  4

• sycl: make oneDNN FA cache thread-safe and memoize runtime failures

  ggmlSYCL
  #31138·Titaniumtown opened Oct 25, 2026Contributor·

  ·

• deepseekv4.1: follow-up review fixes

  Apple MetalconversionggmlHexagonmodelOpenCLOpenVINOVulkanWebGPU
  #31137·am17an opened Oct 25, 2026Contributor·

  ·
  5

• model : Add support for Metis-1.6 models

  conversionmodeltesting
  #31134·lernex opened Oct 25, 2026·

  ·
  5

• sycl: cache device-id

  ggmlSYCL
  #31126·Titaniumtown opened Oct 25, 2026Contributor·

  ·

• sycl : add DPAS flash attention kernel for quantized KV cache

  documentationggmlSYCLtesting
  #31125·Torchit1 opened Oct 25, 2026·

  ·
  9

• sycl : fix f32 → q2_0/q4_0/q5_0 rounding so CPY stops failing

  ggmlSYCL
  #31120·Torchit1 opened Oct 24, 2026·

  ·
  4

• ui: Sidebar resizing improvements

  merge readyserver/ui
  #31119·allozaur opened Oct 24, 2026Contributor·

  ·
  3

• vulkan: skip maxStorageBufferRange pre-check for non-compute tensors (fixes MoE cache crash on devices without shader_64b_indexing)

  ggmlVulkan
  #31114·FamousMonsterr opened Oct 24, 2026·

  ·
  1

  4

• sycl: work-group argmax for short TOP_K rows

  ggmlSYCL
  #31110·newjordan opened Oct 24, 2026Contributor·

  ·
  1

• sycl: bound CONV_TRANSPOSE_1D input iteration

  ggmlSYCLtesting
  #31109·newjordan opened Oct 24, 2026Contributor·

  ·

• metal : add bf16_bf16 and f32_bf16 mul_mm kernels

  Apple Metalggmltesting
  #31105·rashid5925 opened Oct 24, 2026·

  ·
  1

• common : add longPathAware manifest to Windows exes (fixes #30935)

  server
  #31102·origin-labs-ai opened Oct 24, 2026Contributor·

  ·
  1

  1

• server : cap dry_sequence_breakers count at 128 (fixes #30833)

  server
  #31101·origin-labs-ai opened Oct 24, 2026Contributor·

  ·
  1

  1

• jinja : cap from_json nesting depth (fixes #30834)

  jinja parser
  #31099·origin-labs-ai opened Oct 24, 2026Contributor·

  ·
  1

  3

• server : validate grammar trigger type from JSON (fixes #30835)

  server
  #31098·origin-labs-ai opened Oct 24, 2026Contributor·

  ·
  1

  2

• moe-cache: split tensor if its size reached backend limit

  #31095·rhjdvsgsgks opened Oct 24, 2026Contributor·

  ·
  1

• ggml : possible fix for the radix TOP_K thread timing bug

  CUDAggmltestingVulkan
  #31084·sanmai opened Oct 24, 2026Contributor·

  ·

• ggml-cpu : gate RVV intrinsics on __riscv_vector (#22159)

  ggml
  #31082·Xeonacid opened Oct 24, 2026·

  ·
  1

• ggml : link Threads::Threads as PUBLIC on ggml-base (#31007)

  ggml
  #31080·Fogyvishnu opened Oct 24, 2026·

  ·
  1

• vulkan : fix out-of-bounds NVFP4 UE4M3 scale lookup

  ggmlVulkan
  #31079·flyingdoog opened Oct 24, 2026·

  ·
  1

• sycl : Try PkgConfig for LevelZero includes / libs

  ggmlSYCL
  #31076·mctylr-gh opened Oct 24, 2026Contributor·

  ·

• metal : faster MXFP4 mat-vec, full threadgroups for unary/binary/GLU

  Apple Metalggml
  #31075·am17an opened Oct 24, 2026Contributor·

  ·

• kv-cache : keep rotation constants on the GPU

  testing
  #31073·anton-iermolenko opened Oct 24, 2026·

  ·
  2

  

• unicode : add continuation check lambda

  #31070·danbev opened Oct 24, 2026Member·

  ·