mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 12:54:11 +08:00
e3c5c369ba
Part of my effort to move everything to pytest and decrease the number of testrunner frameworks in ci Gives xmls but they might look a weird b/c module level tests vs tests in classes. Doesn't give skip/disable test infra because those are tied to classes. (for future ref, could either put tests in classes or move the check_if_enable stuff into a pytest hook) Tested in CI and checked that the same number of tests are run Pull Request resolved: https://github.com/pytorch/pytorch/pull/95659 Approved by: https://github.com/huydhn
79 lines
2.4 KiB
Python
79 lines
2.4 KiB
Python
# Owner(s): ["oncall: distributed"]
|
|
|
|
# Copyright 2019 Kakao Brain
|
|
#
|
|
# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.
|
|
#
|
|
# This source code is licensed under the BSD license found in the
|
|
# LICENSE file in the root directory of this source tree.
|
|
import pytest
|
|
import torch
|
|
from torch import nn
|
|
|
|
from torch.distributed.pipeline.sync import Pipe
|
|
from torch.testing._internal.common_utils import run_tests
|
|
|
|
|
|
def test_inplace_on_requires_grad(setup_rpc):
|
|
model = nn.Sequential(nn.Linear(1, 1), nn.ReLU(inplace=True))
|
|
model = Pipe(model, checkpoint="always")
|
|
|
|
x = torch.rand(1)
|
|
y = model(x).local_value()
|
|
|
|
message = r"a leaf Variable that requires grad .* used in an in-place operation."
|
|
with pytest.raises(RuntimeError, match=message):
|
|
y.backward()
|
|
|
|
|
|
@pytest.mark.xfail(strict=True)
|
|
def test_inplace_on_not_requires_grad(setup_rpc):
|
|
# In-place operation on a tensor not requiring grad doesn't cause a
|
|
# RuntimeError. Currently, we cannot detect this case.
|
|
model = nn.Sequential(nn.ReLU(inplace=True))
|
|
model = Pipe(model, [1], devices=["cpu"], checkpoint="always")
|
|
|
|
x = torch.rand(1)
|
|
y = model(x).local_value()
|
|
del model
|
|
|
|
message = r"a leaf Variable that requires grad .* used in an in-place operation."
|
|
with pytest.raises(RuntimeError, match=message):
|
|
y.backward()
|
|
|
|
|
|
@pytest.mark.xfail(strict=True)
|
|
def test_inplace_incorrect_grad(setup_rpc):
|
|
class M(nn.Module):
|
|
def forward(self, foo_bar):
|
|
# 'foo' requires grad but 'bar' does not. In-place operation on
|
|
# 'bar' won't cause a RuntimeError.
|
|
foo, bar = foo_bar
|
|
|
|
# add_(1) is not idempotent, in contrast to relu_(). If it is
|
|
# executed multiple times, it will accumulates each difference onto
|
|
# 'bar'.
|
|
bar.add_(1)
|
|
|
|
# 'bar' is still captured by checkpointing. 'foo' will get
|
|
# incorrect grad.
|
|
return foo * bar
|
|
|
|
model = nn.Sequential(M())
|
|
model = Pipe(model, [1], devices=["cpu"], checkpoint="always")
|
|
|
|
foo = torch.tensor([1.0], requires_grad=True)
|
|
bar = torch.tensor([1.0])
|
|
|
|
output = model((foo, bar)).local_value()
|
|
del model
|
|
output.backward()
|
|
|
|
# The gradient of 'foo' should be 2, but it is 3 actually because
|
|
# bar.add_(1) was executed twice due to checkpointing.
|
|
assert foo.grad.item() == 2.0
|
|
|
|
|
|
if __name__ == "__main__":
|
|
run_tests()
|