Loosen fp16 atol in ONNX export tests for onnxruntime 1.29 CPU numerics

[pggPL]

Description

L1_pytorch_onnx_unittest fails in nightly CI since 2026-08-18 on the fp16 (precision1) variants of test_export_linear_recipe, test_export_layernorm_linear_recipe and test_export_layernorm_mlp. The trigger is the onnxruntime 1.28.0 → 1.29.0 bump in the unpinned CI image: ORT executes the exported graph on the CPU EP, and 1.29 changed fp16 CPU numerics, so the ORT-vs-TE mismatch now reaches up to ~9e-3 (2–9 fp16 ULPs) against the hardcoded atol=1e-3. Verified by A/B on the same TE/torch: ORT 1.28 passes, ORT 1.29 fails; no TE commit is involved.

Since the comparison is half-precision GPU (TE) vs half-precision CPU (ORT), differences of a few ULPs from accumulation order are expected and do not indicate an incorrect ONNX representation. This PR makes the tolerance precision-aware: atol=2e-2 for fp16 (the value already used for FP8 CurrentScaling and FP8 MLP in the same file), keeping 1e-3 for fp32.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Add get_atol(precision) helper in tests/pytorch/test_onnx_export.py returning 2e-2 for fp16 and the default 1e-3 otherwise.
• Use it in the non-FP8 validation paths of _test_export_linear, _test_export_layernorm_linear, _test_export_layernorm_mlp and in the DelayedScaling branch of _test_export_layernorm_linear.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

The PR makes ONNX-export numerical validation precision-aware to accommodate ONNX Runtime 1.29 CPU fp16 differences while preserving the existing fp32 tolerance.

• Adds a shared helper returning 2e-2 for fp16 and 1e-3 otherwise.
• Applies it to Linear, LayerNormLinear, and LayerNormMLP validation paths, including the relevant DelayedScaling branch.

Confidence Score: 5/5

The PR appears safe to merge because the changed tolerance is limited to fp16 comparisons in the identified ONNX tests while fp32 remains unchanged.

The helper receives canonical torch dtype parameters, selects the intended fp16 tolerance at every changed call site, and preserves the previous tolerance for fp32; no concrete blocking or non-blocking defect remains.

Important Files Changed

Filename	Overview
tests/pytorch/test_onnx_export.py	Introduces and consistently applies a precision-aware absolute tolerance to the identified fp16 ONNX export comparisons without changing fp32 behavior.

_{Reviews (1): Last reviewed commit: "Loosen fp16 atol in ONNX export tests fo..." | Re-trigger Greptile}

[vthumbe1503]

/te-ci L1 pytorch

[fheinecke]

The CI failures are unrelated to this changeset and should be fixed in #3384