All pull requests

Pull requests

• fix(common): explicit value_type init of spec->lookup entry

  #176·Buddychek opened Oct 20, 2026·

  ·

• cuda : emit routed MMQ down inputs from paired GLU

  #175·GenerelSchwerz opened Oct 16, 2026Owner·

  ·

• cuda : vectorize Q8_0 BF16 conversion

  #174·GenerelSchwerz opened Oct 13, 2026Owner·

  ·

• cuda : fuse dilated convolution, history and retained outputs

  #173·GenerelSchwerz opened Oct 12, 2026Owner·

  ·

• cuda : elide copies before gated activations

  #172·GenerelSchwerz opened Oct 12, 2026Owner·

  ·

• cuda: fuse signed square-root row gates

  #171·GenerelSchwerz opened Oct 12, 2026Owner·

  ·

• cuda : compact sparse routed MMQ scheduling

  #169·GenerelSchwerz opened Oct 12, 2026Owner·

  ·

• cuda : simplify contiguous Q8_1 input quantization addressing

  #168·GenerelSchwerz opened Oct 12, 2026Owner·

  ·

• cuda : fuse small-grid multi-column MMVQ gate/up projections

  #167·GenerelSchwerz opened Oct 11, 2026Owner·

  ·

• cuda : pair floating projections with compatible compute policies

  #166·GenerelSchwerz opened Oct 11, 2026Owner·

  ·

• cuda : pack short floating MMVF and HC-up reductions

  #164·GenerelSchwerz opened Oct 11, 2026Owner·

  ·

• cuda : specialize two-token gated delta net

  #163·GenerelSchwerz opened Oct 10, 2026Owner·

  ·

• cuda : skip empty MMQ stream-k fixup

  #162·GenerelSchwerz opened Oct 10, 2026Owner·

  ·

• cuda : emit prepared images from gated RMS outputs

  #161·GenerelSchwerz opened Oct 10, 2026Owner·

  ·

• cuda : emit typed and quantized activation images from HC outputs

  #160·GenerelSchwerz opened Oct 9, 2026Owner·

  ·

• cuda : specialize unbatched MMVQ indexing

  #159·GenerelSchwerz opened Oct 9, 2026Owner·

  ·

• cuda : pair ordinary floating projections and scalar epilogues

  #158·GenerelSchwerz opened Oct 9, 2026Owner·

  ·

• cuda: preserve weighted RMS reduction order with vector loads

  #157·GenerelSchwerz opened Oct 9, 2026Owner·

  ·

• cuda : combine routed reduction and shared expert gates

  CUDAggmltesting
  #156·GenerelSchwerz opened Oct 9, 2026Owner·

  ·

• cuda : distribute low-grid MMF work and prefetch weights

  CUDAggml
  #155·GenerelSchwerz opened Oct 8, 2026Owner·

  ·

• cuda: emit native MXFP4 images from HC normalization

  CUDAggmltesting
  #153·GenerelSchwerz opened Oct 8, 2026Owner·

  ·

• cuda: emit banked HC MMQ activation images

  CUDAggmltesting
  #152·GenerelSchwerz opened Oct 8, 2026Owner·

  ·

• cuda: bypass dead repeats for compatible matrix fanout

  CUDAggmltesting
  #150·GenerelSchwerz opened Oct 7, 2026Owner·

  ·

• cuda: emit banked RMS and HC activation images

  CUDAggmltesting
  #149·GenerelSchwerz opened Oct 7, 2026Owner·

  ·

• cuda: share banked activation images

  CUDAggmltesting
  #148·GenerelSchwerz opened Oct 7, 2026Owner·

  ·