← Home

ONNX-MLIR

2026 · Open source · Merged contributor

ONNX-MLIR is an open-source compiler built on LLVM and MLIR that turns machine learning models in the ONNX format into native code with minimal runtime support.

Contributions

  • Fixed PadV2 decomposition so padding constants match the input tensor's element type, resolving compilation failures for F16 and F64 inputs.
  • Reused the existing float conversion helper to convert the F32 attribute with the destination type's floating-point semantics.
  • Added FileCheck regression cases for F16 and F64 alongside the existing F32 coverage.

Outcomes

  • Merged upstream into onnx/onnx-mlir as PR #3645.
  • Resolved issue #3408 covering PadV2 compilation with F16 and F64 inputs.