All pull requests

Pull requests

• core: nearest upsample copies rows for every datum type

  #3250·EvanCarroll opened Nov 1, 2026Contributor·

  ·

• onnx: MultiHeadAttention takes attention_bias as Sdpa's additive mask

  #3249·EvanCarroll opened Nov 1, 2026Contributor·

  ·
  1

• core: Cast and GroupNorm convert f16 across the executor's threads

  #3248·EvanCarroll opened Nov 1, 2026Contributor·

  ·

• suite-unit: cover RotateHalf

  #3244·kali opened Oct 31, 2026Collaborator·

  ·
  1

• linalg: run mmm inline on a one-thread pool instead of the global pool

  #3234·dundysm opened Oct 31, 2026·

  ·
  1

  2

• pulse: blockify sliding-window masks and pulses finer than the mask c…

  #3227·kali opened Oct 30, 2026Collaborator·

  ·
  1

• core: a deploy-runtime pool serving concurrent callers through a fixe…

  #3216·kali opened Oct 30, 2026Collaborator·

  ·
  1

• Fix/layer norm keep last axis

  #3211·kali opened Oct 29, 2026Collaborator·

  ·
  7

• cuda: run f32 cuDNN convolutions in FP32 unless TRACT_CUDA_TF32 opts in

  #3201·kali opened Oct 28, 2026Collaborator·

  ·
  1

• Fix/gpu lazy host tensor writes

  #3198·kali opened Oct 28, 2026Collaborator·

  ·
  1

• api: make an input constant from the public API

  #3184·czoli1976 opened Oct 27, 2026Contributor·

  ·
  2

• Feat/qwen3 asr example

  #3178·kali opened Oct 27, 2026Collaborator·

  ·
  3

  

• examples: add Pixel Lab, live background replacement on WebGPU

  #3176·czoli1976 opened Oct 27, 2026Contributor·

  ·
  5

  

• core,cuda: reap per-session caches of sessions dropped on another thread

  #3162·czoli1976 opened Oct 26, 2026Contributor·

  ·
  2

• linalg, nnef: add the ggml IQ2_S block format

  #3080·czoli1976 opened Oct 20, 2026Contributor·

  ·

• linalg, nnef: add the ggml Q6_K block format

  #3077·czoli1976 opened Oct 20, 2026Contributor·

  ·

• linalg, nnef: add the ggml Q5_K block format

  #3076·czoli1976 opened Oct 20, 2026Contributor·

  ·

• linalg, nnef: add the ggml Q4_K block format

  #3075·czoli1976 opened Oct 20, 2026Contributor·

  ·
  6

  

• cuda: Q3_K, IQ4_XS, and IQ2_S matmul fallback by expanding constants to floats

  #3063·czoli1976 opened Oct 18, 2026Contributor·

  ·

• cuda: Q4_K matmul fallback through a per-run device dequant

  #3062·czoli1976 opened Oct 18, 2026Contributor·

  ·

• cuda: Q5_K matmul fallback through a per-run device dequant

  #3061·czoli1976 opened Oct 18, 2026Contributor·

  ·

• cuda: Q6_K matmul fallback through a per-run device dequant

  #3060·czoli1976 opened Oct 18, 2026Contributor·

  ·

• cuda: Q8_0 matmul fallback through a per-run device dequant

  #3059·czoli1976 opened Oct 18, 2026Contributor·

  ·

• metal: Q3_K, IQ4_XS, and IQ2_S matmul fallback by expanding constants to floats

  #3058·czoli1976 opened Oct 17, 2026Contributor·

  ·

• metal: Q4_K matmul fallback through a per-run device dequant

  #3057·czoli1976 opened Oct 17, 2026Contributor·

  ·