All pull requests

Pull requests

• ggml-cpu: fix RVV IQ4_NL and MXFP4 odd-block dot products

  #30643·inuEbisu opened Oct 16, 2026·

  ·
  1

• json-schema: accept non-schema additionalProperties values

  testing
  #30642·xingzihai opened Oct 16, 2026·

  ·
  1

• vulkan: Let q8/q4 KV use fa kernel optimized for Intel Xe

  ggmlVulkan
  #30641·elvinw-intel opened Oct 16, 2026·

  ·

• llama: create new fused op for QSA mask

  CUDAggmlmodeltesting
  #30640·am17an opened Oct 16, 2026Contributor·

  ·

• opencl: enable f32_f16 flash attention on Adreno a7x (E031)

  ggmlOpenCL
  #30639·EdgeBites opened Oct 16, 2026·

  ·
  1

• Replace MoE LRU cache with a 50/50 SLRU cache

  #30637·eapache opened Oct 16, 2026Contributor·

  ·
  2

• server: add realtime transcription API

  server
  #30636·ngxson opened Oct 16, 2026Collaborator·3 tasks·

  ·

• gguf-py : add numpy Q1_0 and Q2_0 de/quantization support

  conversion
  #30630·SomeNobody21112 opened Oct 15, 2026·

  ·
  1

  2

• hexagon: add support for q2_0 quant

  ggmlHexagon
  #30627·tboinovski1 opened Oct 15, 2026Contributor·

  ·

• server: bound resumable-stream session count to prevent memory exhaustion, fix #29498

  server
  #30625·Yunzez opened Oct 15, 2026Contributor·

  ·
  1

• scripts: Fix --strict mode for API check

  #30623·ckastner opened Oct 15, 2026Collaborator·

  ·
  

• ci: update device matrix for Snapdragon jobs by removing SM8750

  devops
  #30620·zhiyuan8 opened Oct 15, 2026Contributor·

  ·
  2

• ci: added ABI reports as github artifact

  devops
  #30619·sandrohanea opened Oct 15, 2026Contributor·

  ·
  

• tests : catch vanished rows in the fusion baseline

  documentationtesting
  #30618·ggerganov opened Oct 15, 2026Member·

  ·
  2

• model : MTP support for Mellum2.1

  conversionmodel
  #30612·Xarbirus opened Oct 15, 2026Contributor·

  ·
  2

• Fix failing RocM tests

  testing
  #30611·mailkv23 opened Oct 15, 2026Contributor·

  ·

• server : handle thinking disabled and output_config.effort in Anthrop…

  server
  #30608·homeofe opened Oct 15, 2026Contributor·

  ·
  1

• qwen3 : bidirectional attention and dense_2 output projection

  conversionmodel
  #30607·fzowl opened Oct 15, 2026·

  ·
  1

  1

• sycl: route BMG D=256 single-query decode to the FA vector kernel

  ggmlSYCLtesting
  #30606·anantshri opened Oct 15, 2026Contributor·

  ·

• Sycl iq3 crash fix

  ggmlSYCLtesting
  #30604·anantshri opened Oct 15, 2026Contributor·

  ·
  2

• cuda : fuse qwen4exp's SCALE -> SIGMOID -> SCALE -> hc_post chain

  CUDAggmlmerge ready
  #30602·am17an opened Oct 15, 2026Contributor·

  ·
  1

• llama: avoid signed overflow in KV seq_cp is_full predicate

  #30601·Jasmine-tim opened Oct 15, 2026Contributor·

  ·

• llama: avoid signed overflow in recurrent seq_add

  #30599·Jasmine-tim opened Oct 15, 2026Contributor·

  ·

• llama: avoid signed overflow in KV seq_add

  #30597·Jasmine-tim opened Oct 15, 2026Contributor·

  ·

• llama: clamp dry_allowed_length in DRY sampler init

  #30596·Jasmine-tim opened Oct 15, 2026Contributor·

  ·