All pull requests

Pull requests

• perf(jit): reduce repeated work in HBG kernel eager calls

  #3230·Leaf-Salix opened Oct 30, 2026Contributor·

  ·
  1

• fix(distributed): capture builtin collective dispatches in DFX

  #3229·georgebisbas opened Oct 30, 2026Contributor·

  ·
  14

• Add L4 cluster orchestration and persistent shared windows

  #3228·TaoZQY opened Oct 30, 2026·

  ·
  1

• feat(ir): add opt-in single-level local software pipelines

  #3227·Jacko-John opened Oct 30, 2026Contributor·

  ·
  16

• feat(codegen): execute A5 VF helpers through EmitC wrappers

  #3226·luohuan19 opened Oct 30, 2026Collaborator·

  ·
  37

• refactor(ir): materialize aligned A5 gather sources with textract

  #3225·Little-oil opened Oct 30, 2026Contributor·

  ·
  12

• perf(passes): fold A5 C2V BF16 epilogues into Fixpipe

  #3223·wangqin1723-max opened Oct 29, 2026Contributor·

  ·
  17

• fix: select Tile handles through native scf.if

  #3221·Little-oil opened Oct 29, 2026Contributor·

  ·
  34

• fix(codegen): initialize empty tile views and guard memory transfers

  #3220·luohuan19 opened Oct 29, 2026Collaborator·

  ·
  12

• fix(ir): rotate For/While Tile storage through scalar selectors

  #3216·Little-oil opened Oct 29, 2026Contributor·

  ·
  1

• Fix: preserve comparison validity and consume CPU simulator repairs

  bug
  #3214·chenshengxin2026 opened Oct 28, 2026Contributor·

  ·
  1

  63

• fix(ci): restore forked tests and GLM5 compilation

  #3211·tonibohnlein opened Oct 28, 2026Contributor·

  ·
  13

• feat(ir): add proven-safe borrowed GM tensor subviews

  #3208·yiyuan-123 opened Oct 28, 2026Contributor·

  ·
  38

• fix(collectives): reject InCore composite collectives in a multi-block SPMD launch

  #3206·georgebisbas opened Oct 28, 2026Contributor·

  ·
  24

• feat(ir): enable native packed FP4 MX tile matmul

  enhancement
  #3202·lwDavid opened Oct 28, 2026Contributor·

  ·
  12

  

• fix(ir): preserve AutoTile input snapshots and result storage

  #3199·tonibohnlein opened Oct 27, 2026Contributor·

  ·
  26

• perf(ir): index AutoTile reads once per function

  #3198·tonibohnlein opened Oct 27, 2026Contributor·

  ·
  16

• fix(ir): separate L0 search alignment from MAD work

  #3197·tonibohnlein opened Oct 27, 2026Contributor·

  ·
  6

• fix(ir): preserve valid extents in AutoTile Mat scratch

  #3196·tonibohnlein opened Oct 27, 2026Contributor·

  ·
  13

• feat(jit): route eager kernels through torch streams

  #3194·YunjiQin opened Oct 27, 2026Collaborator·

  ·
  41

• fix(ir): preserve accumulator windows and slice consumer types

  #3193·Little-oil opened Oct 27, 2026Contributor·

  ·
  1

  9

• feat(ir): add explicit-stride views and preserve GM load strides

  #3191·Little-oil opened Oct 27, 2026Contributor·

  ·
  1

  21

• fix(ci): allow image publication while main advances

  #3181·luohuan19 opened Oct 27, 2026Collaborator·

  ·
  2

• fix(toolchain): upgrade PTOAS to v0.70 and adapt division precision

  #3179·luohuan19 opened Oct 27, 2026Collaborator·

  ·
  2

• fix(jit): infer constant-width runtime slices

  #3177·lyfne123 opened Oct 27, 2026Collaborator·

  ·
  1

  29