All pull requests

Pull requests

• opencl: make extra_q8_0 layout explicit

  ggmlOpenCL
  #30899·kurquhar opened Oct 20, 2026Contributor·

  ·

• webgpu: throttle in-flight batches during prefill only

  ggmlWebGPU
  #30898·thomas-0816 opened Oct 20, 2026·

  ·

• opencl: add F32-to-I32 ARGMAX

  ggmlOpenCL
  #30897·kurquhar opened Oct 20, 2026Contributor·

  ·

• ggml-opencl : add IM2COL_3D support

  ggmlOpenCL
  #30895·kurquhar opened Oct 20, 2026Contributor·

  ·

• opencl: fix rope program build error with newer Clang

  ggmlOpenCL
  #30893·bashbaug opened Oct 20, 2026Contributor·

  ·

• model-saver: fix per-layer array lengths for MTP models

  testing
  #30891·ravel7524 opened Oct 20, 2026Contributor·

  ·

• metal : fix memory leaks and cleanup logic in GPU capture path

  Apple Metalggml
  #30890·nikwen opened Oct 20, 2026Member·

  ·

• vendor: update BoringSSL to 0.20261020.0

  vendor
  #30889·cabelo opened Oct 20, 2026Contributor·

  ·
  1

• mtmd : fix InternVL CLS token order

  mtmd
  #30886·boshjerns opened Oct 20, 2026Contributor·

  ·

• hip: fix bf16 WMMA operand vector type for gfx11/gfx12

  CUDAggml
  #30885·waltercool opened Oct 20, 2026·

  ·

• ggml : add batched tensor copy API

  Apple MetalAscend NPUCUDAggmlHexagonIBM zDNNOpenCLOpenVINOSYCLtestingVulkanWebGPU
  #30881·ggerganov opened Oct 20, 2026Member·

  ·
  3

• llama : add llama_arch_* query API and refactor test-llama-archs to use it

  testing
  #30880·ggerganov opened Oct 20, 2026Member·

  ·

• gguf : add string length accessor

  ggmltesting
  #30876·ggerganov opened Oct 20, 2026Member·

  ·
  1

• opencl:fix issues:29798, opencl set_tensor drops writes to q8_0 views…

  ggmlOpenCL
  #30872·GaoXiangYa opened Oct 20, 2026·

  ·
  1

• ggml-cpu: add SVE-optimized q5_0 and q5_1 dot-product kernels

  ggml
  #30867·t-kusanagi opened Oct 20, 2026·

  ·
  1

• Pin empty slot

  server
  #30866·Dimi8146 opened Oct 20, 2026·

  ·
  1

• Substitute rerank placeholder from the original template only.

  servertesting
  #30864·Pengkun-ZHU opened Oct 20, 2026·

  ·

• llama : scale moe down projection per expert token

  ggmlmerge ready
  #30861·ggerganov opened Oct 20, 2026Member·

  ·
  5

• CUDA: Enable small_k (regardless of K) for MMVQ on Jetson Orin when bs=1

  CUDAggml
  #30858·meitmt opened Oct 20, 2026·1 task done·

  ·

• sycl : add f16 support for SWIGLU_OAI

  ggmlSYCL
  #30857·mctylr-gh opened Oct 20, 2026Contributor·

  ·

• server : forward llama.cpp sampling params on /v1/messages

  documentationserver
  #30856·Dgarc359 opened Oct 20, 2026·

  ·
  15

• ggml-cpu : add AVX512BW/VL path for iq3_xxs dot product

  ggml
  #30853·ray-x opened Oct 20, 2026·

  ·

• cpu : accept F16 in src1 of mul_mat

  ggml
  #30851·AdvayMonga opened Oct 20, 2026·

  ·

• Handle small moe caches better

  #30848·eapache opened Oct 20, 2026Contributor·

  ·

• Improvements to the Memory Access Pattern in ggml_compute_forward_mul_mat_one_chunk

  ggml
  #30846·kazutoshi-ak opened Oct 20, 2026·

  ·
  2