Repository navigation
Conversation
There was a problem hiding this comment.
🟡 Changes recommended
Critical float16 precision errors in normalization leave small L2 inputs unchanged and collapse large L1/L2 inputs to zeros instead of producing normalized outputs.
1 open issue
What changed in this PR
Adds ONNX LpNormalization support to the Relax frontend, allowing imported models to normalize tensors along a selected axis using L1 or L2 norms.
Changes:
- Implements and registers the converter with attribute validation and zero-norm handling.
- Adds structural and numerical tests covering axes, zero slices, float16 inputs, and invalid attributes.
- Enables the default ONNX backend conformance test.
| File | Description |
|---|---|
tests/python/relax/test_frontend_onnx.py |
Tests L1/L2 lowering, numerical results, and attribute validation. |
tests/python/relax/test_frontend_onnx_backend.py |
Enables the default LpNormalization backend test. |
python/tvm/relax/frontend/onnx/onnx_frontend.py |
Implements and registers the LpNormalization converter. |
Review effort: Max
There was a problem hiding this comment.
Warning
Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.
Copilot review overview
3 open issues
Review effort: Lite
|
Hi, I noticed this overlaps with #20818, which I opened on October 25 for the same operator. I've updated that PR against current main and fixed its reference-runtime compatibility: TVM still runs the full opset 13/22 matrix, while the ONNX Runtime reference uses the equivalent opset 13 kernel supported by CI's pinned runtime. The tests cover defaults, axes, static/symbolic shapes, zero-norm slices, and float16/bfloat16 accumulation in float32. Happy to collaborate or consolidate the implementations/tests. Could the maintainers advise which PR they would prefer to continue with so we can avoid duplicating work? |
|
closing as #20818 merged |


This PR adds support for the ONNX LpNormalization operator in the Relax ONNX frontend.
LpNormalization normalizes the input along
axisby its L1 or L2 norm:Changes
LpNormalizationconverter (_impl_v1) inpython/tvm/relax/frontend/onnx/onnx_frontend.py, supporting theaxis(default-1) andp(default2) attributes, and registered it in the converter map.p=1andp=2are accepted, as defined by the ONNX spec (onnxruntime also rejects other values); other values raise aValueError.axisis normalized and range-checked against the input rank.where(norm == 0, 1, norm)rather than clamping with an epsilon, so non-zero norms are never altered and float16 inputs don't underflow to0 / 0.test_lp_normintests/python/relax/test_frontend_onnx.py:p in {1, 2}across positive and negative axes,pand out-of-rangeaxis.test_lpnormalization_default) by addinglpnormalizationto_INCLUDE_OPSintests/python/relax/test_frontend_onnx_backend.py.