frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

relax tol for flaky nansum_out_dtype_cuda_float32 test (#121550)

Browse Source 
TestReductionsCUDA.test_nansum_out_dtype_cuda_float32 would fail or pass depending on the random inputs. Observed by ROCm internal QA testing.  But same problematic random inputs breaks the test for CUDA, verified on V100.

There is precedent in another test within the same file to relax tolerance.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121550
Approved by: https://github.com/albanD
This commit is contained in:
Jeff Daily
2024-03-14 15:28:45 +00:00
committed by PyTorch MergeBot
parent 7e13b5ba29
commit be0bdf111c
1 changed files with 18 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
test/test_reductions.py
View File

@ -1713,6 +1713,20 @@ class TestReductions(TestCase):
self._test_reduction_function_with_numpy(torch.count_nonzero, np.count_nonzero, device, dtype)
self._test_reduction_function_with_numpy(torch.count_nonzero, np.count_nonzero, device, dtype, True)
# TODO: Investigate why the output is not close to numpy.
def _get_relaxed_tolerances_for(self, dtype):
if dtype == torch.float16:
atol = 0.4
rtol = 1e-2
elif dtype == torch.float32:
atol = 7e-05
rtol = 3e-06
else:
# Default values
atol = None
rtol = None
return atol, rtol
def _test_sum_reduction_vs_numpy(self, torch_fn, np_fn, device, dtype, with_keepdim=False, with_extremal=False):
def is_integral(dtype):
return dtype in integral_types()
@ -1731,16 +1745,7 @@ class TestReductions(TestCase):
exact_dtype = False
# TODO: Investigate why the output is not close to numpy.
if dtype == torch.float16:
atol = 0.4
rtol = 1e-2
elif dtype == torch.float32:
atol = 7e-05
rtol = 3e-06
else:
# Default values
atol = None
rtol = None
atol, rtol = self._get_relaxed_tolerances_for(dtype)
self._test_reduction_function_with_numpy(torch_fn, np_fn, device, dtype,
atol=atol, rtol=rtol, exact_dtype=exact_dtype,
with_keepdim=with_keepdim, with_extremal=with_extremal)
@ -1771,12 +1776,14 @@ class TestReductions(TestCase):
out_dtype = dtype
inp_dtypes = all_types_and(torch.half) if out_dtype.is_floating_point else integral_types()
for inp_dtype in inp_dtypes:
# TODO: Investigate why the output is not close to numpy.
atol, rtol = self._get_relaxed_tolerances_for(dtype)
shape = _rand_shape(random.randint(2, 5), min_size=5, max_size=10)
x = _generate_input(shape, inp_dtype, device, with_extremal=False)
torch_fn = partial(torch.nansum, dtype=out_dtype)
np_out_dtype = torch_to_numpy_dtype_dict[out_dtype]
np_fn = partial(np.nansum, dtype=np_out_dtype)
self.compare_with_numpy(torch_fn, np_fn, x, device=None, dtype=None)
self.compare_with_numpy(torch_fn, np_fn, x, device=None, dtype=None, atol=atol, rtol=rtol)
@dtypes(*all_types_and(torch.half))
def test_argminmax_multiple(self, device, dtype):