All pull requests

Pull requests

• server : pass in --json-schema to chat template

  server
  #31862·aldehir opened Nov 7, 2026Contributor·

  ·
  1

• common : sanitize invalid UTF-8 in PEG tool call arguments

  testing
  #31861·apuravmanhas opened Nov 7, 2026·

  ·

• workaround: gpt-oss gguf conversion for p-e-w/gpt-oss-20b-heretic-ara-v4

  conversion
  #31860·linellazatin opened Nov 7, 2026·

  ·

• server : downgrade per-request proxy logs to trace

  server
  #31859·ggerganov opened Nov 7, 2026Member·

  ·

• hexagon: n_hvx and n_threads cleanup

  documentationggmlHexagonmerge ready
  #31857·max-krasnyansky opened Nov 7, 2026Member·

  ·
  1

• CUDA: use block reductions for whole-row pool_1d

  CUDAggmltesting
  #31855·topppg opened Nov 7, 2026·

  ·

• ggml : switch to F32 src1 in the CPU backend

  ggmltesting
  #31854·ggerganov opened Nov 7, 2026Member·

  ·
  4

• llama, mtmd: follow the reference precision of ggml_prec_set_acc calls

  CUDAggmlHexagonmodelmtmdOpenCLtestingVulkanWebGPU
  #31851·ServeurpersoCom opened Nov 7, 2026Contributor·

  ·
  3

• llama : upload expert copies on a separate stream with an event-lifetime fix

  CUDAggmltesting
  #31850·bahooo22 opened Nov 7, 2026·

  ·
  2

• ggml-zdnn: make zTensor memory usage visible for fitting

  ggmlIBM zDNN
  #31849·taronaeo opened Nov 7, 2026Member·

  ·

• gguf : retain zero-sized dimensions while trimming trailing ones

  ggmltesting
  #31845·ppAlexa opened Nov 7, 2026Contributor·

  ·

• cuda: respect src1 GGML_PREC_F32 in F32/F16 mul_mat

  Apple MetalCUDAggmltesting
  #31844·ServeurpersoCom opened Nov 7, 2026Contributor·

  ·
  1

• rpc : avoid stack exhaustion in create_node with deep graphs

  ggml
  #31843·Pyolar opened Nov 7, 2026·1 task done·

  ·
  1

  1

• llama : reuse scheduler buffers across requests

  testing
  #31841·wsj040303 opened Nov 7, 2026·

  ·
  1

  1

  

• ROCm/HIP: fixed BLAS workspace instead of rocBLAS allocating one per call

  CUDAggml
  #31838·fox5611 opened Nov 7, 2026·

  ·
  1

  3

• ui: show audio in Discover for models whose mmproj has an audio encoder

  server/ui
  #31837·allozaur opened Nov 7, 2026Contributor·

  ·

• Gate copying split inputs with the flag GGML_TENSOR_FLAG_COMPUTE of consuming graph nodes

  ggml
  #31835·zyang-dev opened Nov 7, 2026Contributor·

  ·
  2

• common/json-schema : fix allOf with plain types (e.g. integer) generating an empty object grammar

  testing
  #31834·chinmaykamjith001 opened Nov 6, 2026·

  ·
  1

  1

• ggml-cpu: reject MUL_MAT in supports_op when vec_dot kernel is missing

  ggml
  #31832·xPuffyPuffx opened Nov 6, 2026·1 task done·

  ·

• opencl: improve q8_0 and MoE small-n verification for spec

  ggmlOpenCL
  #31831·shaofeiqi opened Nov 6, 2026Contributor·

  ·

• hexagon: decode-gated collective buffer isolation for TG speedup

  ggml
  #31827·ebateni opened Nov 6, 2026Contributor·

  ·

• meta : split tied LM head and size compute arena per device

  ggml
  #31826·ebateni opened Nov 6, 2026Contributor·

  ·
  1

• ggml-zdnn: accept non-zDNN tensors, invalidate compute tensors after every operation

  ggmlIBM zDNN
  #31824·taronaeo opened Nov 6, 2026Member·

  ·
  1

• discussion : ggml_prec updates and formalization

  Apple Metalggmltesting
  #31821·ngxson opened Nov 6, 2026Collaborator·

  ·
  9

• Hexagon : gated fusion for gate/up matmul + GLU

  ggmlHexagon
  #31820·ebateni opened Nov 6, 2026Contributor·

  ·