Skip to content

bf16xint16 fails on AMD ROCm with tensor comparison assertion error #797

@karthickai

Description

@karthickai

Describe the bug
The test compares a Helion bf16×int16 GEMM kernel against a pytorch reference implementation, but ROCm produces slightly different numerical results that exceed the comparison tolerance.

To Reproduce
Pull PR #794 and remove the @skipIfRocm decorator from test_bf16xint16 in test/test_examples.py. Run the test on ROCm environment.
Failed CI job: https://github.com/pytorch/helion/actions/runs/18213179448/job/51857480348

Expected behavior
The test should pass on ROCm since both implementations perform the same mathematical operation

  • helion kernel: converts int16 -> bf16 inside Triton kernel then performs GEMM
  • pytorch reference: converts int16 -> bf16 with PyTorch then performs torch.matmul

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions