All pull requests

Pull requests

• server : retry cached prompt restoration after clearing idle slots

  server
  #30997·NoahOksuz opened Oct 22, 2026Contributor·2 tasks done·

  ·
  1

  1

• vocab : keep \r\n together in the Laguna pre-tokenizer

  testing
  #30996·tkran2 opened Oct 22, 2026Contributor·

  ·

• hexagon: add conv2d support

  ggmlHexagon
  #30995·tboinovski1 opened Oct 22, 2026Contributor·

  ·

• server: realtime API: send complete after committed audio drains

  mtmdserver
  #30994·ngxson opened Oct 22, 2026Collaborator·

  ·
  1

  1

• model, mtmd: add chatterbox TTS support

  conversiondocumentationmtmd
  #30993·ngxson opened Oct 22, 2026Collaborator·

  ·

• server : add --prefill-budget-busy

  documentationserver
  #30992·enriqxxx opened Oct 22, 2026·

  ·
  1

• cpu : add AVX-VNNI tiled matmul

  ggml
  #30991·trivalik opened Oct 22, 2026Contributor·

  ·

• rpc : support multiple RDMA ports

  ggml
  #30990·ru-sh opened Oct 22, 2026·

  ·
  1

• ui: Models Manager as Route + UI improvements

  merge readyserver/ui
  #30988·allozaur opened Oct 22, 2026Contributor·

  ·
  2

• cuda : gather batched tensor reads into mapped host memory

  CUDAggmltesting
  #30986·bhargav191098 opened Oct 22, 2026·

  ·
  1

  

• metal : use mul_mm for 32 <= K < 64

  Apple Metalggmltesting
  #30982·vieduy opened Oct 22, 2026Contributor·

  ·

• [Model] Add MiniCPM-Jev

  conversionmtmd
  #30980·tc-mb opened Oct 22, 2026Contributor·

  ·
  1

• ui: Align hover card and context popup styling overall UI style

  merge readyserver/ui
  #30978·allozaur opened Oct 22, 2026Contributor·

  ·
  1

• ci : re-enable test-llama-context with FA on for WebGPU

  devops
  #30977·yomaytk opened Oct 22, 2026Member·

  ·
  2

• sycl : vectorize reordered-layout dequantization stores (Q4_0, K-quants, IQ3, MXFP4)

  ggmlSYCL
  #30976·anantshri opened Oct 22, 2026Contributor·

  ·
  1

• vocab : fix Nemotron-3.5-Embed tokenization

  conversionmerge ready
  #30974·tokinasin opened Oct 22, 2026Contributor·

  ·
  7

• llama : add adaptive single-graph MoE offload

  CUDAggmltesting
  #30971·pwilkin opened Oct 22, 2026Member·

  ·
  3

• server : share prompt prefix cells between slots (--slot-attach)

  documentationserver
  #30970·rochaudhari opened Oct 22, 2026·

  ·

• ggml : fix time overflow in ggml_time_us

  ggml
  #30969·frozenblade1224 opened Oct 22, 2026Contributor·

  ·
  4

• docs: add Intel backend selection guide, fix SYCL/OpenCL docs

  documentationOpenCLSYCL
  #30967·AshutoshSinghIntel opened Oct 22, 2026·

  ·

• llama_moe_cache: run missed experts on CPU, add tuning script

  documentationggmlserver
  #30966·am17an opened Oct 22, 2026Contributor·

  ·
  4

• qwen4exp : build the QSA selection without the kq_mask

  CUDAggmlmodeltesting
  #30965·sanmai opened Oct 22, 2026Contributor·

  ·
  7

• batched-bench : apply --lora adapters

  examples
  #30964·vieduy opened Oct 22, 2026Contributor·

  ·

• vocab : extract bpe_rank loading into function

  #30962·danbev opened Oct 22, 2026Member·

  ·
  2

• server : fix /v1/responses endpoint to accept extra_params

  documentationserver
  #30959·Dgarc359 opened Oct 22, 2026Contributor·

  ·