All pull requests

Pull requests

• bugfix: infinite recursion caused by a tool named 'call'(#29967)

  documentationtesting
  #30088·Frost-54 opened Oct 7, 2026·

  ·
  1

  1

• ggml-cuda: assign two GDN state columns per warp

  CUDAggml
  #30087·SongXiaoXi opened Oct 7, 2026Contributor·

  ·
  4

  

• vendor : update cpp-httplib to 0.60.0

  vendor
  #30081·cabelo opened Oct 7, 2026Contributor·

  ·

• musa: use the tile lightning indexer kernel

  CUDAggml
  #30080·yeahdongcn opened Oct 7, 2026Collaborator·

  ·

• imatrix : include clocale for std::setlocale

  examples
  #30079·dazzywi opened Oct 7, 2026·

  ·

• CUDA: fix slow q4_0/q5_0 V dequantization on Blackwell with CUDA 12.8

  CUDAggml
  #30077·ilyas-mallah opened Oct 7, 2026·

  ·

• meta: mirror K/V tensors with empty shards

  #30076·neyuru opened Oct 7, 2026·

  ·

• Metal : stop overflow from happening in large matrix output offsets

  Apple Metalggmltesting
  #30072·feni6 opened Oct 6, 2026·

  ·

• server : reject RANK pooling on /embeddings and /v1/embeddings

  server
  #30068·Shreesh-Sree opened Oct 6, 2026·

  ·
  1

  1

• [Fix] Skip over -inf for CPU Flash Attention to prevent NaN during softmax

  ggmltesting
  #30066·AnkitNakhawa opened Oct 6, 2026·

  ·
  1

  1

• metal : few-row MMA mat-mul for the remaining src0 types

  Apple Metalggmlmerge readytesting
  #30065·pratiknarola-t opened Oct 6, 2026Contributor·

  ·
  

• mtmd: add cohere2 vision support

  conversionmtmd
  #30062·Terrencezzj opened Oct 6, 2026·

  ·
  7

• common : read ngram cache parts by reference instead of copying

  #30061·jadidbourbaki opened Oct 6, 2026·

  ·

• Option to copy GPU tensors instead of mmap host-ptr buffers

  #30060·lotar opened Oct 6, 2026·

  ·
  1

  

• LLama-wackMall-fusion.

  documentationggmlservertesting
  #30059·miltos22 opened Oct 6, 2026·

  ·
  5

• metal : few-row MMA kernels for MUL_MAT_ID and more src0 types

  Apple Metalggmltesting
  #30047·pratiknarola-t opened Oct 6, 2026Contributor·

  ·
  1

  

• ggml-cpu : add VNNI repack kernels for Q8_0 and Q4_0 on x86

  ggml
  #30036·njsyw1997 opened Oct 6, 2026Contributor·

  ·

• hexagon: optimize q2_k bpw and activation

  ggmlHexagon
  #30035·jhen0409 opened Oct 6, 2026Member·

  ·

• Red-teamed hardening and unit test coverage for /v1/systemone

  buildconversiondocumentationservertesting
  #30027·sometimesanotion opened Oct 5, 2026·

  ·
  2

• tests : consolidate context state tests into test-llama-context

  devopstesting
  #30023·ggerganov opened Oct 5, 2026Member·

  ·

• ggml-cuda: hip: Tune stream_k algo and mmq configs to use stream_k on AMD GCN

  CUDAggml
  #30022·thelittlefireman opened Oct 5, 2026Contributor·

  ·

• ggml-cuda: hip: tune GCN MMQ configs

  CUDAggml
  #30021·thelittlefireman opened Oct 5, 2026Contributor·

  ·

• server: refactor modalities handling

  documentationserver
  #30015·ngxson opened Oct 5, 2026Collaborator·

  ·
  8

• ggml: Change default visibility of symbols to hidden

  ggml
  #30014·ckastner opened Oct 5, 2026Collaborator·

  ·
  1

  

• server: fix sleep race that strands queued tasks and crashes on model unload

  draftservertesting
  #30012·vyalamar opened Oct 5, 2026·

  ·
  1

  1