All pull requests

Pull requests

• [WIP][AIROCMLIR-1307] Add SetInThreadTransposeReductionLayoutPass: Optimize warpsPerCTA generated by ITT

  #540·pabloantoniom opened Sep 24, 2026Collaborator·

  ·

• Expand conv bias onto the folded output-layout K axis

  #539·tuhinp-amd opened Sep 24, 2026Collaborator·

  ·
  8

• Support building on Windows with MSVC cl.exe

  #538·rachgupt93 opened Sep 24, 2026Collaborator·

  ·
  14

• Recombine fully masked split-KV attention partials on the host

  #535·umangyadav opened Sep 23, 2026Member·

  ·
  1

• Handle fully masked attention rows in kernel lowering

  #534·umangyadav opened Sep 23, 2026Member·

  ·
  18

  

• [AIROCMLIR-1240] rock dialect dynamic shapes verifier [WIP]

  #530·tuhinp-amd opened Sep 23, 2026Collaborator·

  ·
  8

• [AIROCMLIR-1306] Port missing rocMLIR test coverage into rocmlirTriton

  #527·dhernandez0 opened Sep 22, 2026Collaborator·

  ·
  54

  

• [AIROCMLIR-1163] [EXTERNAL] Port the upstream version of the cost-aware dot rematerialization patch

  #521·bogdan-petkovic opened Sep 22, 2026Collaborator·

  ·
  17

  

• [AIROCMLIR-1160] Fix splitKV pre-softmax fusion views and non-adjacent merge collapse

  #513·dhernandez0 opened Sep 21, 2026Collaborator·

  ·
  27

  

• [AIROCMLIR-1292] [EXTERNAL] Cherry-pick the upstream fix for the float-based 32-bit div/rem expansion

  #511·bogdan-petkovic opened Sep 21, 2026Collaborator·

  ·
  9

  

• [AIROCMLIR-1292] [EXTERNAL] Do not stagger CDNA4 padded layouts over rows that do not exist

  #508·bogdan-petkovic opened Sep 18, 2026Collaborator·

  ·
  11

  

• [AIROCMLIR-888] Implement per-problem narrowing in the quick-tuning DB

  #500·mirza-halilcevic opened Sep 17, 2026Collaborator·

  ·
  25

• [AIROCMLIR-1163] Enable the vendored Triton lit test suite in nightly CI

  #494·bogdan-petkovic opened Sep 16, 2026Collaborator·

  ·
  5

  

• [AIROCMLIR-1187] Select K-major LDS layout for pipelined WMMA and pre-gfx950 MFMA dots

  #471·justinrosner opened Sep 3, 2026Collaborator·

  ·
  17

• [WIP][AIROCMLIR-1097] Pull down LLVM fix and enable non-pow2 kPerBlock on gfx950

  #470·pabloantoniom opened Sep 3, 2026Collaborator·

  ·

• [AIROCMLIR-1141] Fix the tuning error for fp4 datatype

  #437·erizheng-amd opened Aug 20, 2026Collaborator·

  ·
  28

• [Windows] Enable the rocmlirTriton test suite on Windows

  #378·ppetrovi-amd opened Jul 28, 2026Contributor·

  ·

• [AIROCMLIR-899] Enable weekly CI stages (parameter sweeps + tuning)

  #361·bogdan-petkovic opened Jul 20, 2026Collaborator·

  ·
  26

  

• BLOCKED [AIROCMLIR-1025] Non-power-of-two tile decomposition to attention (gemm+gemm)

  #360·dhernandez0 opened Jul 20, 2026Collaborator·

  ·
  42

  

• [WIP][AIROCMLIR-1076] Support K not a multiple of kPerBlock via loop peeling

  #348·pabloantoniom opened Jul 15, 2026Collaborator·

  ·

• [AIROCMLIR-987] Widen im2col conv input loads by carrying a vectorization hint into Triton

  #329·erizheng-amd opened Jul 8, 2026Collaborator·

  ·
  73

• Lower tanh to gfx1250 hardware instructions

  #314·umangyadav opened Jun 26, 2026Member·

  ·
  56

  

• [WIP][AIROCMLIR-982] Simplify AffineMaps in TransformsToPointerArith

  #292·pabloantoniom opened Jun 16, 2026Collaborator·

  ·
  4

• [AIROCMLIR-911] Replace inf-clamping with fail-on-inf in allclose comparator

  #285·stefankoncarevic opened Jun 11, 2026Contributor·

  ·
  22

• [AIROCMLIR-899]Add weekly CI

  #275·djramic opened Jun 8, 2026Contributor·

  ·
  4