Keep infinities in clamp with one bound

[Muhammad Usman Mateen (usmanmateen)]

aten_clamp (with only min or only max), aten_clamp_min, aten_clamp_max and the 0-d tensor branches of aten_clamp_min_tensor / aten_clamp_max_tensor leave the other Clip bound empty. The Clip spec says the bounds default to numeric_limits::lowest() and numeric_limits::max(), and ONNX Runtime does exactly that. So for floating-point inputs an infinity is clipped to the largest finite value:

import math

import torch


class Model(torch.nn.Module):
    def forward(self, x):
        return x.clamp(min=1e-6), x.clamp(max=0.0)


x = torch.tensor([math.inf, -math.inf, 2.0, -3.0])
onnx_program = torch.onnx.export(Model(), (x,), dynamo=True)
print(Model()(x))
# (tensor([   inf, 1.0000e-06, 2.0000e+00, 1.0000e-06]), tensor([ 0., -inf,  0., -3.]))
print(onnx_program(x))
# (tensor([3.4028e+38, 1.0000e-06, 2.0000e+00, 1.0000e-06]), tensor([ 0.0000e+00, -3.4028e+38,  0.0000e+00, -3.0000e+00]))

After that, torch.isinf no longer sees the value, and something like 1 / x gives 2.9e-39 instead of 0.

This adds a small _clip helper that passes -inf / inf for the missing bound when the input is floating point, and uses it in those five places. Integer inputs keep the omitted bound, since there the type's limits are the right default. NaN still propagates, as before.

Tests

test_clamp_with_one_bound_keeps_infinities in tests/function_libs/torch_lib/e2e_ops_tests.py exports clamp with only min and only max, clamp_min and clamp_max with a scalar, and clamp_min and clamp_max with a 0-d tensor bound. The input contains inf, -inf and nan, and the test checks the outputs against eager. It fails on main (inf vs 3.4028e+38 in the first output) and passes with this change.

• python -m pytest tests/function_libs/torch_lib/e2e_ops_tests.py -k "clamp or clip or hardtanh or relu6": 2 passed
• python -m pytest tests/function_libs/torch_lib/ops_test.py -k "clamp or hardtanh or relu6 or clip": 21 passed, 14 skipped (124 subtests passed)
• ruff check and ruff format --check (0.15.8) on the changed files

Tested with torch 2.14.1+cpu and onnxruntime 1.31.0.