The GitHub Actions job "Lint" on tvm.git/main has succeeded.
Run started by GitHub user tlopex (triggered by tlopex).

Head commit for run:
9d5253e2ea59897664b92c0e8b5508fd83475fdd / yanght27 <[email protected]>
[Fix][Relax] Skip ReorderPermuteDimsAfterConcat for unknown-rank inputs (#20216)

Fixes #20204.

`ReorderPermuteDimsAfterConcat` derives the permutation for
`permute_dims(..., axes=None)` from the input rank. When an input tensor
has unknown rank, its `ndim` is `-1`, but the pass previously asserted
that `ndim >= 0`. As a result, valid Relax IR could raise an
`InternalError`.

This PR changes the axis derivation helper to return
`ffi::Optional<ffi::Array<int64_t>>`. When the axes are implicit and the
input rank is unknown, a concrete permutation cannot be derived, so the
pass conservatively leaves the original expression unchanged. Existing
behavior is preserved for known-rank inputs and explicit axes.

A regression test covers unknown-rank inputs followed by `concat` and
verifies that the pass neither rewrites the expression nor raises an
error.

Testing:

- Built the modified C++ target successfully.
- `test_transform_reorder_permute_dims_after_concat.py`: 7 passed.
- Pre-commit hooks passed on both modified files.
- The original issue reproducer completes without an `InternalError`.

Report URL: https://github.com/apache/tvm/actions/runs/33233704592

With regards,
GitHub Actions via GitBox


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to