All pull requests

Pull requests

• hexagon: enable alloc_buffer_n

  ggmlHexagonmerge ready
  #30126·max-krasnyansky opened Oct 7, 2026Member·

  ·
  1

• server: flush metrics more frequently

  server
  #30125·hugendudel opened Oct 7, 2026·2 tasks done·

  ·

• sampling : skip zero-probability candidates in typical entropy

  testing
  #30124·akanyaani opened Oct 7, 2026·

  ·
  1

• hexagon: add support for Q1_0

  ggmlHexagon
  #30122·tboinovski1 opened Oct 7, 2026Contributor·

  ·

• vulkan: Optimize pool_2d when the pooling window is large

  ggmltestingVulkan
  #30119·jeffbolznv opened Oct 7, 2026Contributor·

  ·

• grammar : add support for multi-token matching

  documentationtesting
  #30117·aldehir opened Oct 7, 2026Contributor·

  ·
  3

• server: use a secure RNG for generated IDs

  server
  #30116·benthecarman opened Oct 7, 2026·

  ·

• server : add --decision-type CLI option and fallback detection for /v1/systemone

  server
  #30113·jimdawdy-hub opened Oct 7, 2026·

  ·
  1

• llama: support MoE cache over multiple GPUs

  testing
  #30112·am17an opened Oct 7, 2026Contributor·

  ·
  2

• llama : fix DFlash output head sharing

  conversionmodel
  #30111·hthadicherla opened Oct 7, 2026Contributor·

  ·
  2

• model : use exact GELU for ModernBERT encoders

  model
  #30108·boshjerns opened Oct 7, 2026Contributor·

  ·

• vulkan : fix TOP_K for +inf/NaN inputs and k = 1 on negative values

  ggmltestingVulkan
  #30107·gianni-cor opened Oct 7, 2026·

  ·
  1

• vulkan : one warp tile per subgroup in the small matmul tiles

  ggmlVulkan
  #30106·ianloic opened Oct 7, 2026Contributor·

  ·
  1

• vulkan : honor FP32 source precision in MUL_MAT_ID

  #30105·linuxid10t opened Oct 7, 2026·

  ·
  3

• CUDA/HIP: fix race in flash_attn_ext_f16_process_tile when nbatch_combine != DKQ/2

  CUDAggml
  #30103·IMbackK opened Oct 7, 2026Contributor·

  ·
  1

  4

• CUDA : fix PAD for more than 65535 rows or slices

  CUDAggmltesting
  #30102·pskrunner14 opened Oct 7, 2026Contributor·

  ·

• WIP: Persist checkpoints when saving hybrid model slots to disk

  server
  #30101·rain-sk opened Oct 7, 2026·

  ·
  2

• CUDA: add SM103 SASS to non-native builds

  CUDAggml
  #30099·ynankani opened Oct 7, 2026Contributor·

  ·
  5

• examples : fix draft sampling and rejection handling

  examples
  #30095·danbev opened Oct 7, 2026Member·

  ·
  1

• AIX support for executable path resolution and change quad name

  ggmltesting
  #30094·mehendarkarprajwal opened Oct 7, 2026Contributor·

  ·
  7

• vulkan: warn (not silent STATUS) when glslc lacks a shader extension and a GPU path gets disabled

  ggmlVulkan
  #30092·chaosmonkies opened Oct 7, 2026·

  ·
  1

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

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

  ·
  7

  

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

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

  ·

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

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

  ·
  3

  

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

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

  ·
  1

  1