All pull requests

Pull requests

• Fold single-element widening converts in propagate_constant (regression from #5138)

  #5495·klin2024 opened Nov 9, 2026Collaborator·

  ·
  4

• Direct FP32 3x3 stride-2 convolution kernel for gfx12 (conv3x3s2)

  #5493·kazhang2 opened Nov 9, 2026Collaborator·

  ·
  1

• Fix Selu lowering returning 0 or NaN for large positive inputs

  #5492·akshat-lakhera opened Nov 8, 2026·

  ·
  1

• Port the CPU target to oneDNN 3

  #5489·adityas-amd opened Nov 7, 2026Member·

  ·
  1

• Sink layout copies into concat buffers and fold layouts into copying reshapes

  #5488·pfultz2 opened Nov 7, 2026Collaborator·

  ·
  11

• Onnxruntime Weekly Sync 2026-11-06

  automateddependenciesonnxruntimeskip bot checks
  #5486·github-actions[bot] opened Nov 6, 2026Bot·

  ·
  

• Time benchmark candidates as HIP graphs

  #5485·kazhang2 opened Nov 6, 2026Collaborator·

  ·
  3

• [AIMIGRAPHX-1371] Keep program outputs standard in eliminate_concat

  #5484·anupsing-amd opened Nov 6, 2026Contributor·

  ·
  8

• Clamp layernorm variance to avoid NaN from cancellation

  #5483·VishalX opened Nov 6, 2026Contributor·1 task done·

  ·
  19

• Fuse the chunked linear attention recurrence into one kernel

  #5482·pfultz2 opened Nov 5, 2026Collaborator·

  ·
  14

• gpu: use hip::load_scalar for if/loop host scalars

  #5481·ivarusic-amd opened Nov 5, 2026Contributor·

  ·
  4

• Fold axis-linear operations on a dot input into its constant operand

  #5479·pfultz2 opened Nov 4, 2026Collaborator·

  ·
  16

• Binary cache: store only winner if problem cache is used

  Windows
  #5478·pnikolic-amd opened Nov 4, 2026Collaborator·

  ·
  

• Binary cache read only option

  Windows
  #5477·pnikolic-amd opened Nov 4, 2026Collaborator·

  ·
  2

  

• Support per-channel quantized models with expanded QDQ parameters

  #5475·royzhang-amd opened Nov 4, 2026Contributor·

  ·

• Prevent FP16 GroupNorm overflow with float intermediates

  #5474·royzhang-amd opened Nov 4, 2026Contributor·

  ·
  3

• Check driver input names against the program parameters

  #5469·pfultz2 opened Nov 3, 2026Collaborator·

  ·
  6

• Support stride and dilation in the channelwise convolution kernel

  #5468·pfultz2 opened Nov 2, 2026Collaborator·

  ·
  14

• Fuse pointwise into the MLIR conv/dot that no sibling input depends on

  #5467·pfultz2 opened Nov 2, 2026Collaborator·

  ·
  4

• Add MXR recompiling: keep kernel recipes and recompile compiled programs

  #5466·anilmartha opened Nov 2, 2026·

  ·
  68

• Onnxruntime Weekly Sync 2026-10-30

  automateddependenciesonnxruntimeskip bot checks
  #5463·github-actions[bot] opened Oct 30, 2026Bot·

  ·
  

• Bump RocMlirTriton with triton_amd_codegen library

  #5462·ilango100 opened Oct 30, 2026·

  ·
  6

• Avoid stack overflow for deep recursion in the fix combinator

  #5459·pfultz2 opened Oct 30, 2026Collaborator·

  ·
  7

• Fix GPU codegen layout pinning and deep-graph stack overflows

  #5458·cliffcai opened Oct 30, 2026Contributor·6 of 7 tasks·

  ·
  10

• [AIMIGRAPHX-595] NVFP4 E4M3-scaled GEMM on gfx1250

  #5457·anupsing-amd opened Oct 29, 2026Contributor·4 of 5 tasks·

  ·
  81