• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Owner(s): ["module: dynamo"]
2
3import unittest
4
5import torch
6import torch._dynamo as torchdynamo
7from torch.testing._internal.common_utils import run_tests, TEST_CUDA, TestCase
8
9
10try:
11    import tabulate  # noqa: F401  # type: ignore[import]
12
13    from torch.utils.benchmark.utils.compile import bench_all
14
15    HAS_TABULATE = True
16except ImportError:
17    HAS_TABULATE = False
18
19
20@unittest.skipIf(not TEST_CUDA, "CUDA unavailable")
21@unittest.skipIf(not HAS_TABULATE, "tabulate not available")
22class TestCompileBenchmarkUtil(TestCase):
23    def test_training_and_inference(self):
24        class ToyModel(torch.nn.Module):
25            def __init__(self) -> None:
26                super().__init__()
27                self.weight = torch.nn.Parameter(torch.Tensor(2, 2))
28
29            def forward(self, x):
30                return x * self.weight
31
32        torchdynamo.reset()
33        model = ToyModel().cuda()
34
35        inference_table = bench_all(model, torch.ones(1024, 2, 2).cuda(), 5)
36        self.assertTrue(
37            "Inference" in inference_table
38            and "Eager" in inference_table
39            and "-" in inference_table
40        )
41
42        training_table = bench_all(
43            model,
44            torch.ones(1024, 2, 2).cuda(),
45            5,
46            optimizer=torch.optim.SGD(model.parameters(), lr=0.01),
47        )
48        self.assertTrue(
49            "Train" in training_table
50            and "Eager" in training_table
51            and "-" in training_table
52        )
53
54
55if __name__ == "__main__":
56    run_tests()
57