All pull requests

Pull requests

• [WIP] Add CPU offloading of row embedding token tables to tabular ICL models

  #1319·seunghwak opened Oct 30, 2026Collaborator·

  ·
  2

• WIP

  #1311·RBendias opened Oct 30, 2026Collaborator·

  ·
  2

• Add opt-in CUDA Graph replay for Kumo Tabular prediction

  #1290·GraysonWu opened Oct 28, 2026·

  ·
  6

  

• Compile categorical string lookup with native PyTorch operations

  #1275·aw471 opened Oct 27, 2026Collaborator·

  ·
  5

• Preserve nanosecond datetime precision with legacy microsecond defaults

  #1273·alan-kumo opened Oct 27, 2026Contributor·

  ·
  2

• Replace cuDF string sorting with native PyTorch on CUDA

  #1272·aw471 opened Oct 26, 2026Collaborator·

  ·
  7

• Add native variable-length entry permutation

  #1271·aw471 opened Oct 26, 2026Collaborator·

  ·
  5

• Preserve fitted recipe execution state across compiled conditionals

  #1258·aw471 opened Oct 23, 2026Collaborator·

  ·
  3

• Make single-member category shuffle fitting traceable

  #1243·aw471 opened Oct 22, 2026Collaborator·

  ·
  4

• [DO NOT REVIEW/WIP] Add KumoTabular context-sharded inference

  #1238·aw471 opened Oct 22, 2026Collaborator·

  ·
  2

• Add MatrixFormer matrix completion model

  #1203·jacobf18 opened Oct 20, 2026·

  ·
  1

• Fuse attention projection with per-head normalization

  #1200·autumnust opened Oct 20, 2026Collaborator·

  ·
  1

• Reduce redundant attention input preparation

  #1191·autumnust opened Oct 19, 2026Collaborator·4/4·

  ·
  1

• Reduce repeated table metadata preparation during inference

  #1170·autumnust opened Oct 17, 2026Collaborator·3/4·

  ·
  7

• Fuse shared-length query scaling during inference

  #1168·autumnust opened Oct 16, 2026Collaborator·2/4·

  ·
  6

  

• Make bounded KumoRelational task propagation compilable

  #1157·aw471 opened Oct 16, 2026Collaborator·

  ·
  1

• [FP8 5/5] Accelerate cached prediction with context splitting

  #1107·aw471 opened Oct 13, 2026Collaborator·

  ·
  2

  

• [Kumo-Anomaly 3/n] Support post-norm and dropout in TransformerBlock

  #1106·aw471 opened Oct 13, 2026Collaborator·

  ·
  3

• Support fullgraph compilation of TabICLv2 hierarchies

  #1104·puririshi98 opened Oct 12, 2026Collaborator·

  ·
  2

  

• [2/2] Reuse grouped recipe tables during estimator execution

  #1100·aw471 opened Oct 12, 2026Collaborator·

  ·
  9

• Add explicit target inversion to output recipes

  #1069·RBendias opened Oct 7, 2026Collaborator·

  ·
  2

• [1/2] Extract estimator batching helpers from model base

  #1047·aw471 opened Oct 5, 2026Collaborator·

  ·
  8

• Fix PowerTransform fitting for half-precision inputs

  #1039·aw471 opened Oct 2, 2026Collaborator·

  ·
  12

• Polars assisted sampling for KumoRelational

  #1026·alan-kumo opened Sep 29, 2026Contributor·

  ·
  1

• Fix QuantileTransform for 3D inputs

  #964·rusty1s opened Sep 24, 2026Collaborator·

  ·
  4