All pull requests

Pull requests

• ci : enable GGML_SCHED_DEBUG_REALLOC=1 for ctest workflows

  devops
  #29514·ggerganov opened Sep 27, 2026Member·

  ·

• llama + CUDA: flash_attn_ext_rows (fix unified kv for multi-seq)

  Apple MetalAscend NPUCUDAggmlHexagonmodelOpenCLOpenVINOSYCLtestingVulkanWebGPU
  #29510·am17an opened Sep 27, 2026Contributor·

  ·

• server: don't store the draft KV in context checkpoints

  server
  #29509·myml opened Sep 27, 2026·

  ·
  1

• musa: define __CUDA_ARCH__ and fix the q8_0 and iq4_nl dequantization kernels

  CUDAggml
  #29508·yeahdongcn opened Sep 27, 2026Collaborator·

  ·
  1

• sycl: remove duplicate block-size defines from op headers

  ggmlSYCL
  #29507·Titaniumtown opened Sep 27, 2026Contributor·

  ·

• sycl: use ExternalProject to let this backend be built with SYCL compiler while everything else could use another

  ggmlSYCL
  #29506·a1batross opened Sep 26, 2026·

  ·
  2

• Fix c++ odr by properly using GGML_COMMON_DECL_CPP

  ggml
  #29504·XZiar opened Sep 26, 2026·

  ·
  1

• opencl: refine a8x bin kernel loading condition

  ggmlmerge readyOpenCL
  #29503·lhez opened Sep 26, 2026Contributor·

  ·

• sycl: add IQ3_S multi-column MMVQ

  ggmlSYCL
  #29500·clemenswasser opened Sep 26, 2026·

  ·

• bug: fix grammar builder crash on invalid minItems/maxItems (#29462)

  Apple MetalbuildconversionCUDAdevopsdocumentationggmlHexagonjinja parsermodelmtmdnixOpenCLOpenVINOserver/uiserverSYCLtestingvendorVulkanWebGPU
  #29497·Cid-oe opened Sep 26, 2026·

  ·
  1

  1

• server : report usage across all choices when n > 1

  server
  #29496·bhaskargurram-ai opened Sep 26, 2026·

  ·
  1

  1

• cli: fix exit code 0 when media loading fails

  devopsdocumentationserver
  #29492·Yu-Haikuo opened Sep 26, 2026·

  ·

• server : fix usage.completion_tokens undercount with n > 1

  server
  #29486·unbug opened Sep 26, 2026·

  ·
  1

  2

• webgpu: add MMVQ support for Q1_0/Q5_0/Q5_1/Q3_K/Q5_K/Q6_K/MXFP4

  ggmlWebGPU
  #29483·yomaytk opened Sep 26, 2026Member·

  ·
  2

• server: emit tool results before user text in Anthropic conversion

  server
  #29482·gputier opened Sep 26, 2026·

  ·
  1

• kv-cache : fix iSWA seq_pos_max answering from the SWA cache only

  #29480·charan-rathore opened Sep 26, 2026·

  ·
  1

  1

• ggml-cuda: HIP: optimize non-saturating packed byte subtraction (__vsubss4)

  CUDAggml
  #29478·thelittlefireman opened Sep 26, 2026Contributor·

  ·
  

• jinja : add support for dict builtin

  jinja parsertesting
  #29477·CISC opened Sep 26, 2026Member·

  ·

• vulkan: Tune GDN kernel, fix Intel performance

  ggmlVulkan
  #29476·0cc4m opened Sep 26, 2026Contributor·

  ·

• common : fix HF cache paths on Windows

  #29475·angt opened Sep 26, 2026Member·

  ·

• Fix: Handle unaligned writes in ggml_backend_webgpu_buffer_set_tensor

  ggmlWebGPU
  #29471·jbooth opened Sep 26, 2026Contributor·

  ·

• server: roll back batch state after slot exceptions

  server
  #29470·lilong555 opened Sep 26, 2026·

  ·
  3

• vulkan: fix argsort kernel selection for Adreno

  ggmlmerge readyVulkan
  #29469·0cc4m opened Sep 26, 2026Contributor·

  ·
  1

• server : rollback failed request's partial prompt batch

  server
  #29467·SamJSui opened Sep 26, 2026·

  ·
  2

• OoD documentation for llama-bench

  documentationexamples
  #29464·mairp opened Sep 26, 2026·

  ·
  2