Skip to content

Re-enable dynamo tests with TODOs #4746

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Mar 10, 2023
Merged

Re-enable dynamo tests with TODOs #4746

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
13f55ef
Re-enable dynamo tests with TODOs
wonjoo-wj Mar 9, 2023
3ea7413
Run linter
wonjoo-wj Mar 9, 2023
514507b
Set manual seed for dynamo unit tests
wonjoo-wj Mar 9, 2023
42dfaa9
Update random seed generation function to use the existing function
wonjoo-wj Mar 9, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 24 additions & 5 deletions test/dynamo/test_dynamo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,9 +11,19 @@
import torchvision
import unittest

# Setup import folders.
xla_test_folder = os.path.dirname(os.path.dirname(os.path.abspath(sys.argv[0])))
sys.path.append(xla_test_folder)

import test_utils


class DynamoInferenceBasicTest(unittest.TestCase):

@classmethod
def setUpClass(self):
test_utils._set_rng_seed(42)

def fn_simple(self, x, y):
a = torch.cos(x)
b = torch.sin(y)
Expand Down Expand Up @@ -79,6 +89,10 @@ def test_resnet18(self):

class DynamoTrainingBasicTest(unittest.TestCase):

@classmethod
def setUpClass(self):
test_utils._set_rng_seed(42)

def fn_simple(self, input):
loss_fn = torch.nn.CrossEntropyLoss()
target = torch.tensor([1, 2, 3], dtype=torch.long).to(input.device)
Expand Down Expand Up @@ -123,7 +137,6 @@ def test_simple_model(self):
self.assertTrue(torch.allclose(res_cpu_3, res_xla_dynamo_3.cpu()))
self.assertTrue(torch.allclose(input.grad, xla_input.grad.cpu()))

@unittest.skip("Broke by functionalization, #4680")
def test_resnet18(self):
torch._dynamo.reset()
met.clear_counters()
Expand Down Expand Up @@ -164,7 +177,8 @@ def test_resnet18(self):
# Graph 2: backward
# Graph 3: sync input for backward
# Graph 4: sync input for backward (TODO(JackCaoG) understand why there are two graphs)
self.assertEqual(met.metric_data('CompileTime')[0], 4)
# TODO @wonjoo CompileTime has increased to 5 after enabling functionalization (#4680)
self.assertEqual(met.metric_data('CompileTime')[0], 5)
# We execute 3 grphs per step.
self.assertEqual(met.metric_data('ExecuteTime')[0], sample_count * 3)
# one for each forward and one for each backward
Expand All @@ -176,6 +190,10 @@ def test_resnet18(self):

class DynamoTrainingOptimizerTest(unittest.TestCase):

@classmethod
def setUpClass(self):
test_utils._set_rng_seed(42)

def fn_simple(self, input, optimizer):
loss_fn = torch.nn.CrossEntropyLoss()
optimizer.zero_grad(True)
Expand Down Expand Up @@ -219,7 +237,6 @@ def test_simple_model(self):
assert torch.allclose(input.grad, xla_input.grad.cpu())
assert torch.allclose(input, xla_input.cpu())

@unittest.skip("Broke by functionalization, #4680")
def test_resnet18(self):
torch._dynamo.reset()
met.clear_counters()
Expand Down Expand Up @@ -267,9 +284,11 @@ def test_resnet18(self):
# Graph 3: optimizer
# Graph 4: sync input for backward
# Graph 5: sync input for backward (TODO(JackCaoG) understand why there are two graphs)
self.assertEqual(met.metric_data('CompileTime')[0], 5)
# TODO @wonjoo CompileTime has increased 5->7 after enabling functionalization (#4680)
self.assertEqual(met.metric_data('CompileTime')[0], 7)
# We execute 4 grphs per step when optimizer is enabled.
self.assertEqual(met.metric_data('ExecuteTime')[0], sample_count * 4)
# TODO @wonjoo ExecuteTime has increased by 1 after enabling functionalization (#4680)
self.assertEqual(met.metric_data('ExecuteTime')[0], (sample_count * 4) + 1)
Copy link
Collaborator Author

@wonjoo-wj wonjoo-wj Mar 9, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@JackCaoG, I just saw that this was failing as well -- the ExecuteTime is offset by 1 regardless of the sample size.

# one for each forward, backward and optimizer
self.assertEqual(
met.metric_data('RunCachedGraphInputData')[0], sample_count * 3)
Expand Down