All pull requests

Pull requests

• GLM-5.3 EXLR8: run the fragment check at load

  #809·harrisonfriia opened Oct 13, 2026Contributor·8 tasks done·

  ·

• cuda: EXL3 projections write fp32 too (qlinear decodeF32 and promptF32)

  #806·BHCC2025 opened Oct 13, 2026Contributor·

  ·
  1

• cuda: NVFP4 experts' prompt rows take the warp's row tile count as a constant

  #805·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  1

• glm: GLM-5.3-Flash (NVIDIA's NVFP4 checkpoint) as a CUDA family at TP over hosts

  #804·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  1

• XPU backend: Qwen3.8-27B on Intel Arc Pro B70 (Xe2) over Level Zero

  #802·Mirrdhyn opened Oct 13, 2026Contributor·7 of 8 tasks·

  ·

• native: each forwarded lane-backend entry calls the family's entry of the same name

  #800·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·

• cuda: sliding-window attention, sigmoid top-k routing and combine, dense rows for every family, checked against float64

  #798·jschmied opened Oct 13, 2026Contributor·8 tasks done·

  ·

• cuda: SwiGLU, output gates, Gated DeltaNet inputs, QK norm and RoPE for every family, checked against float64

  #797·jschmied opened Oct 13, 2026Contributor·8 tasks done·

  ·

• cuda/norms: one tree a width, IEEE 1/sqrt; Nemotron's add_rmsnorm on the shared kernel

  #796·jschmied opened Oct 13, 2026Contributor·8 tasks done·

  ·

• tools/zig/triton_aot_pack.py: the kernel set step for every family, from a manifest or a Recorder dump

  #795·zweck opened Oct 13, 2026·3 tasks done·

  ·
  1

• aot.zig: a launch with no exact captured variant takes one that assumes less of its arguments

  #794·zweck opened Oct 13, 2026·3 tasks done·

  ·
  1

• EXLR8: reduce barriers in Metal projections

  #789·mrv777 opened Oct 13, 2026Contributor·

  ·

• Qwen3.8-Flash-Next faster kernels in Zig engine

  #788·grantoverton opened Oct 13, 2026Contributor·8 tasks done·

  ·

• nemotron cuda: skip empty pocket layers and spread the pocket over many blocks

  #787·ThinkOffApp opened Oct 13, 2026Contributor·8 tasks done·

  ·

• GLM-5.3-Flash faster oQ kernels in Zig engine

  #786·grantoverton opened Oct 13, 2026Contributor·8 tasks done·

  ·
  1

• Qwen3.8-27B faster oQ kernels in Zig engine

  #784·grantoverton opened Oct 13, 2026Contributor·8 tasks done·

  ·

• Qwen3.8-Flash-Next support for oQ in Zig engine

  #774·grantoverton opened Oct 13, 2026Contributor·8 tasks done·

  ·
  2

• GLM-5.3-Flash support for oQ in Zig engine

  #773·grantoverton opened Oct 13, 2026Contributor·8 tasks done·

  ·
  2

• core: a checkpoint's pages out of the page cache once a loader is done with them

  #766·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  1

• glm cuda: agree settings across every rank and give each follower its own doorbell

  #765·Kurcide opened Oct 13, 2026·

  ·
  1

• cuda: the staged NVFP4 prompt entries leave an expert's items alone (skip), as the decode form does

  #764·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  1

• cuda: Nvfp4Experts.packGateUp packs gate and up in place, no scratch

  #763·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  1

• native: a CUDA family's prompt cache, weight share and open errors come from its own declarations

  #762·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  1

• server: the soft open-file limit is raised to the hard one when the server listens

  #760·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  1

• core: read ModelOpt NVFP4 checkpoints as stored

  #758·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  5