All pull requests

Pull requests

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

  Apple Metalggmltesting
  #30680·ggerganov opened Oct 16, 2026Member·

  ·

• llama : optimize vocabulary loading

  #30679·mykhailen opened Oct 16, 2026·

  ·
  

• skills: add slice-model

  documentation
  #30677·ngxson opened Oct 16, 2026Collaborator·

  ·

• model: Granite switch decision

  conversiondocumentationserver
  #30676·gabe-l-hart opened Oct 16, 2026Collaborator·

  ·
  12

• tests: give the random model generator a per-tensor filler table

  testing
  #30674·ServeurpersoCom opened Oct 16, 2026Contributor·

  ·
  

• model : add Kolibri-1 (Aleph-Alpha) support

  conversionmerge readymodeltesting
  #30673·riteme opened Oct 16, 2026·

  ·
  1

• [CUDA] : Add SKU-aware FP16/BF16 accumulation precision gating for NVIDIA GPUs

  CUDAggml
  #30672·mmalpe-nv opened Oct 16, 2026·

  ·
  1

• model: add Beam (Reflection)

  conversionjinja parsermodeltesting
  #30670·danielhanchen opened Oct 16, 2026Contributor·

  ·
  2

• server : fix swa checkpoint invalidation logic

  server
  #30666·tdakhran opened Oct 16, 2026Contributor·

  ·
  

• vulkan: add scratch buffer system, use for dequant_kv

  ggmlVulkan
  #30665·0cc4m opened Oct 16, 2026Contributor·

  ·
  2

• Extend probabilistic sampling for DFlash2 drafts

  #30663·praneshgo opened Oct 16, 2026Contributor·

  ·
  2

• cuda : prevent premature eviction of graphs about to be used

  CUDAggml
  #30662·frank-2077 opened Oct 16, 2026·

  ·
  4

• perf(sampling): Reduce greedy logits without temporary candidates

  testing
  #30659·achembarpu opened Oct 16, 2026·

  ·

• opencl: add q4_0 o4 GEMV for long-vocab lm_head (a7x)

  ggmlOpenCL
  #30658·EdgeBites opened Oct 16, 2026·

  ·
  1

• model : add dspark GIDD log-SNR conditioning for DFlash draft models

  conversiondocumentationggmlmodelOpenCL
  #30657·ggarembroo opened Oct 16, 2026·

  ·
  2

• vulkan: ggml src1 precision support

  ggmltestingVulkan
  #30656·0cc4m opened Oct 16, 2026Contributor·

  ·

• hexagon: fix VTCM size query on images that need layout out-params

  ggmlHexagon
  #30655·tdakhran opened Oct 16, 2026Contributor·

  ·
  1

• ci: update ROCm 10.0.0 -> 10.1.0

  devops
  #30652·fboudra opened Oct 16, 2026Contributor·

  ·
  1

• rpc : do not leave or serve truncated cache files

  ggmltesting
  #30648·TonyWu2333 opened Oct 16, 2026·

  ·
  1

• opencl: use Adreno trans-weight GEMM for narrow q4_0 weights (a7x)

  ggmlOpenCL
  #30646·EdgeBites opened Oct 16, 2026·

  ·
  1

• metal : add tensor API flash attention kernel for small batches with GQA

  Apple Metalggmltesting
  #30645·ethanhq opened Oct 16, 2026Contributor·

  ·
  

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

  ggmltesting
  #30643·inuEbisu opened Oct 16, 2026·

  ·
  1

  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·

  ·