[ci][onnx] Relax more test tolerances

[driazati]

Following on #11042, this changes tolerances to fix some other ONNX test failures that have come up in the past several days

test	link
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[llvm]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/main/3047/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[llvm]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/main/3044/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[cuda]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/main/3068/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[llvm]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-11032/2/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[cuda]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-10833/11/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[cuda]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-11035/1/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[llvm]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-11074/1/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[cuda]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-11068/1/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearsigmoid[cuda]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-11078/1/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[llvm]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-10867/10/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearsigmoid[cuda]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/ci-gpu-update/4/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearsigmoid[cuda]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/main/3132/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearsigmoid[cuda]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/main/3143/pipeline/
tests/python/frontend/onnx/test_forward.py::test_qlinearleakyrelu[llvm]	https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-11146/2/pipeline/

cc @areusch

[github-actions]

It has been a while since this PR was updated, @areusch please leave a review or address the outstanding comments. @driazati if this PR is still a work in progress, please convert it to a draft until it is ready for review.

[areusch]

cc @altanh can you take a look?

[github-actions]

It has been a while since this PR was updated, @altanh @areusch please leave a review or address the outstanding comments. @driazati if this PR is still a work in progress, please convert it to a draft until it is ready for review.

[altanh]

I'm suspicious of a single value being ~1e-2 off while the rest are below 1e-5... tricky. We could reduce tolerances to that but that's quite a reduction

[driazati]

Closing in favor of #11376, agreed that the tolerances are too high here (and not even enough for all the failures I've seen, some are off by atol=0.02) to merge and would silently indicate nothing is wrong. We don't really have any other choice but to disable these tests until someone can identify and fix the values

[masahi]

yeah, for quantized ops, getting accuracy aligned with frameworks is challenging, due to slight difference in how low-level numerics is done (fixed point vs fp32 etc). In PyTorch frontnend, we skip accuracy check entirely for quantized ops / models.