[Tracking Issue][ONNX] Complete missing and limited operators in ONNX frontend

[tlopex]

The Relax ONNX frontend (python/tvm/relax/frontend/onnx/onnx_frontend.py) currently supports 157 operators. However, 7 operators are commented out in the converter map, and 11 have known limitations. This issue tracks completing support for these operators.

Scope

Each missing operator needs:

1. Converter implementation in onnx_frontend.py
2. Tests in tests/python/relax/test_frontend_from_onnx.py

Each limited operator needs the specific restriction lifted and corresponding test coverage added.

Missing operators (commented out in converter map)

Tier 1 — High value

• If — ONNX control flow (conditional branching). Needed for models with runtime-dependent execution paths.
• MatMulInteger — INT8 matrix multiplication. Required for quantized model inference (e.g. ONNX QDQ models).

Tier 2 — Lower priority

• MaxRoiPool — Max ROI pooling.
• MatMulInteger16 — INT16 matrix multiplication. Rarely used.
• Optional / OptionalHasElement / OptionalGetElement — ONNX Optional type handling. Used in some models with optional outputs.

Operators with known limitations

Shape/dynamic restrictions

• Unsqueeze — Does not support symbolic axes or dynamic axes. Only constant axes work.
• Squeeze — Does not support symbolic axes.
• Slice — Only supports constant parameters. Dynamic/symbolic start, end, step not supported.
• Reshape — Special case for shape [-1] not fully handled.

Op-specific restrictions

• ConvTranspose — Does not support 3D (ConvTranspose3d, ndim=5).
• Resize — Does not support dynamic ROI.
• ArgMax / ArgMin — Does not support select_last_index attribute.
• SequenceInsert — Does not support inserting with new axis.
• SplitToSequence — Only supports keepdims=1.

Implementation reference

• Converter base class: OnnxOpConverter in onnx_frontend.py
• Converter map: _get_convert_map() (line 3937)
• Existing tests: tests/python/relax/test_frontend_from_onnx.py

The shape/dynamic restriction fixes (Unsqueeze, Squeeze, Slice) are good first issues — the converter class already exists, just needs to handle additional cases.

If you're interested in contributing, please comment below to claim a task before starting work.

cc @KJlaccHoeUM9l

[OmarAzizi]

Hi @tlopex, I'd like to take MatMulInteger and If (Tier 1). I will work on both and open two separate pull requests

[Aharrypotter]

Hi @tlopex , I'd like to claim the Tier 2 operator: MatMulInteger16, and the Optional suite (Optional, OptionalHasElement, OptionalGetElement). I will open a PR with the implementations and tests soon. Thank you!

[Dayuxiaoshui]

Hi @tlopex, I'd like to claim ConvTranspose 3D support (ConvTranspose3d, ndim=5) and Resize with dynamic ROI, with implementations in onnx_frontend.py and tests in tests/python/relax/test_frontend_from_onnx.py; I can submit one PR or split into two if that’s easier to review—please let me know if this overlaps with anyone else’s work. Thank you!

[LudovicoYIN]

Hi @tlopex, I'd like to claim the implementation of MaxRoiPool.

[Aharrypotter]

Hi @tlopex, I'd like to claim the implementation of shape/dynamic restriction fixes (Unsqueeze, Squeeze, Slice) .