All pull requests

Pull requests

• server : fix server_tokens media chunk handling in push_back and keep_first

  server
  #29956·sanjeevafk opened Oct 4, 2026·

  ·
  1

  2

• cuda : add BF16 support for XIELU

  CUDAdocumentationggmltesting
  #29955·Qiao12-pixel opened Oct 4, 2026·

  ·
  1

• CUDA: fix MMQ ncols rounding direction for alloc

  CUDAggml
  #29953·JohannesGaessler opened Oct 4, 2026Contributor·

  ·
  1

• CUDA: fuse FFN gate/up matmuls and GLU in MMQ with interleaved gate/up rows

  CUDAggmltesting
  #29948·anujj opened Oct 4, 2026Contributor·

  ·
  1

• common, server : add model type to GET /models

  documentationserver
  #29944·angt opened Oct 4, 2026Member·

  ·

• ggml: refactor selective expert copying to user code

  ggmltesting
  #29943·am17an opened Oct 4, 2026Contributor·

  ·
  6

  

• cuda : move neu_padded to where it is used

  CUDAggmlmerge ready
  #29940·angt opened Oct 4, 2026Member·

  ·

• vulkan: Fix Intel prefill regression on MoE models

  ggmlVulkan
  #29936·virajwad opened Oct 4, 2026Contributor·

  ·
  6

• vulkan: Support split_k for conv3d

  ggmltestingVulkan
  #29929·jeffbolznv opened Oct 4, 2026Contributor·

  ·

• feat: add GLM5Next MTP, optimize

  model
  #29928·pwilkin opened Oct 3, 2026Member·

  ·
  17

• ggml-cuda: HIP: use __builtin_amdgcn_perm for q1_0 unpacking

  CUDAggml
  #29927·thelittlefireman opened Oct 3, 2026Contributor·

  ·
  1

  

• ci: retain the anchor when testing recurrent rollback

  testing
  #29923·eapache opened Oct 3, 2026Contributor·

  ·
  

• vulkan: Add a larger tile size for conv3d (coopmat2 only)

  ggmltestingVulkan
  #29920·jeffbolznv opened Oct 3, 2026Contributor·

  ·

• server : add --cache-reuse-hybrid for recurrent/hybrid models

  server
  #29918·bcmyguest opened Oct 3, 2026·

  ·
  1

• ggml-rpc: validate PAD_REFLECT_1D parameters

  examplesggmltesting
  #29915·ZZ2266 opened Oct 3, 2026·2 of 3 tasks·

  ·
  1

• Fix undeclared identifiers when compiling with -DGGML_VULKAN_RUN_TESTS=ON

  ggmlVulkan
  #29912·EAddario opened Oct 3, 2026Contributor·

  ·
  1

• cuda: do not reject MUL_MAT_ID with GGML_PREC_F32 for quantized weights

  CUDAggmlVulkan
  #29911·AoiYori opened Oct 3, 2026·

  ·
  4

• ggml-cuda: fix a mountain of VGPR spills on Q2_K

  CUDAggml
  #29910·thelittlefireman opened Oct 3, 2026Contributor·

  ·

• cuda: tile the lightning indexer over keys and tokens for 4 heads

  CUDAggmltesting
  #29901·ServeurpersoCom opened Oct 3, 2026Contributor·

  ·
  13

• Improve qwen3.8 27b performance on a 3090

  CUDAggmlserver
  #29900·comtom opened Oct 3, 2026·

  ·
  1

• ui: fix text attachment preview overflow

  server/ui
  #29898·shamnadxyz opened Oct 3, 2026·

  ·

• ggml: add GGML_TYPE_BF16X - near-lossless bfloat16 recompression (11.5 bpw)

  Apple MetalconversionCUDAexamplesggmltesting
  #29896·dfytensor opened Oct 3, 2026·

  ·
  2

• log, server: self contained colors, split child commands from logs in router mode

  server
  #29895·ServeurpersoCom opened Oct 3, 2026Contributor·

  ·
  1

  19

• [SYCL]: fix memory errors in mul_mat, split buffer, host pool

  ggmlSYCL
  #29889·mctylr-gh opened Oct 3, 2026Contributor·

  ·
  5

• llama : add a GPU cache for MoE experts kept in host memory

  ggml
  #29887·am17an opened Oct 3, 2026Contributor·

  ·
  14