|
| 1 | +# Owner(s): ["oncall: profiler"] |
| 2 | +import functools |
| 3 | +from typing import Iterator, Optional |
| 4 | + |
| 5 | +import torch |
| 6 | +from torch._C._profiler import _EventType |
| 7 | +from torch.profiler import _memory_profiler, _utils |
| 8 | +from torch.testing._internal.common_utils import run_tests, TestCase |
| 9 | + |
| 10 | + |
| 11 | +profile = functools.partial( |
| 12 | + torch.profiler.profile, record_shapes=True, profile_memory=True, with_stack=True |
| 13 | +) |
| 14 | + |
| 15 | + |
| 16 | +class ScaleLayer(torch.nn.Module): |
| 17 | + def __init__(self) -> None: |
| 18 | + super().__init__() |
| 19 | + self.scale = torch.nn.Parameter(torch.rand(()), requires_grad=True) |
| 20 | + |
| 21 | + def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 22 | + return x * self.scale |
| 23 | + |
| 24 | + |
| 25 | +class TestIdentifyGradients(TestCase): |
| 26 | + def gradient_detected( |
| 27 | + self, |
| 28 | + prof: torch.profiler.profile, |
| 29 | + ctx: _EventType, |
| 30 | + grad_tensor: torch.Tensor, |
| 31 | + parameter: Optional[torch.Tensor] = None, |
| 32 | + ) -> None: |
| 33 | + |
| 34 | + # This is not an exhaustive check, but for the purpose of unit testing |
| 35 | + # it is sufficient. |
| 36 | + def key_matches_tensor(key, tensor) -> bool: |
| 37 | + # Vacuous case. |
| 38 | + if tensor is None: |
| 39 | + return True |
| 40 | + |
| 41 | + if key is None: |
| 42 | + return False |
| 43 | + |
| 44 | + return tensor.storage().data_ptr() == key.storage_ptr |
| 45 | + |
| 46 | + tree = prof.profiler.kineto_results.experimental_event_tree() |
| 47 | + for node in _utils.traverse_dfs(tree): |
| 48 | + for p_key, p_grad_key in _memory_profiler.extract_gradients(node): |
| 49 | + if node.tag == ctx and key_matches_tensor(p_grad_key, grad_tensor): |
| 50 | + if parameter is None: |
| 51 | + return True # Don't need to check parameter; we're done. |
| 52 | + |
| 53 | + elif p_key is not None: |
| 54 | + # For a complex workflow a gradient could correspond to |
| 55 | + # different parameters at different points in a trace. |
| 56 | + # However this will not happen in the relatively simple |
| 57 | + # cases tested here, so if `extract_gradients` identifies |
| 58 | + # the parameter corresponding to a particular gradient it |
| 59 | + # must be the one we expect. |
| 60 | + self.assertTrue(key_matches_tensor(p_key, parameter)) |
| 61 | + return True |
| 62 | + |
| 63 | + return False |
| 64 | + |
| 65 | + def assertGradientDetected(self, name: str, *args, **kwargs) -> None: |
| 66 | + self.assertTrue( |
| 67 | + self.gradient_detected(*args, **kwargs), |
| 68 | + f"Failed to identify gradient `{name}` from profile.", |
| 69 | + ) |
| 70 | + |
| 71 | + def assertOnlyGradients( |
| 72 | + self, prof: torch.profiler.profile, tensors: Iterator[torch.Tensor] |
| 73 | + ) -> None: |
| 74 | + allowed_set = {t.storage().data_ptr() for t in tensors} |
| 75 | + |
| 76 | + tree = prof.profiler.kineto_results.experimental_event_tree() |
| 77 | + for node in _utils.traverse_dfs(tree): |
| 78 | + for _, p_grad_key in _memory_profiler.extract_gradients(node): |
| 79 | + self.assertTrue( |
| 80 | + p_grad_key.storage_ptr in allowed_set, |
| 81 | + f"Tensor wrongly marked as gradient: {node.name}: {p_grad_key}", |
| 82 | + ) |
| 83 | + |
| 84 | + def test_extract_gradients_low_level(self) -> None: |
| 85 | + x = torch.ones((1,)) |
| 86 | + w0 = torch.ones((1,), requires_grad=True) |
| 87 | + w1 = torch.ones((1,), requires_grad=True) |
| 88 | + |
| 89 | + def check(cold_start: bool): |
| 90 | + self.assertEqual(w0.grad is None, cold_start) |
| 91 | + self.assertEqual(w1.grad is None, cold_start) |
| 92 | + with profile() as prof: |
| 93 | + z = x.expand(4) * w0 |
| 94 | + (z * w1).sum().backward() |
| 95 | + |
| 96 | + # Gradient detection through op inspection does not provide a |
| 97 | + # reference to the parameter corresponding to the gradient. |
| 98 | + self.assertGradientDetected("w0", prof, _EventType.TorchOp, w0.grad) |
| 99 | + self.assertGradientDetected("w1", prof, _EventType.TorchOp, w1.grad) |
| 100 | + self.assertOnlyGradients(prof, (w0.grad, w1.grad)) |
| 101 | + |
| 102 | + check(cold_start=True) |
| 103 | + check(cold_start=False) |
| 104 | + |
| 105 | + def test_extract_gradients_from_module(self) -> None: |
| 106 | + model = torch.nn.Sequential(torch.nn.Linear(2, 1), ScaleLayer()) |
| 107 | + named_parameters = {name: p for name, p in model.named_parameters()} |
| 108 | + self.assertEqual(len(named_parameters), 3) |
| 109 | + |
| 110 | + def assert_only_gradients(prof: torch.profiler.profile): |
| 111 | + gradients = tuple(i.grad for i in named_parameters.values()) |
| 112 | + self.assertFalse(any(i is None for i in gradients)) |
| 113 | + self.assertOnlyGradients(prof, gradients) |
| 114 | + |
| 115 | + def check(cold_start: bool): |
| 116 | + x = torch.ones((2, 2)) |
| 117 | + with profile() as prof: |
| 118 | + model(x).sum().backward() |
| 119 | + |
| 120 | + for name, p in named_parameters.items(): |
| 121 | + # The first time we run a module none of the `.grad` fields |
| 122 | + # have been initialized. This is fine; in that case we can |
| 123 | + # detect everything we need in the profiled section. |
| 124 | + self.assertNotEqual( |
| 125 | + self.gradient_detected(prof, _EventType.PyCall, p.grad, p), |
| 126 | + cold_start, |
| 127 | + name, |
| 128 | + ) |
| 129 | + |
| 130 | + # Op based detection should still identify the gradients. |
| 131 | + self.assertGradientDetected(name, prof, _EventType.TorchOp, p.grad) |
| 132 | + assert_only_gradients(prof) |
| 133 | + |
| 134 | + # We can detect gradients even when `.backward()` is not called. |
| 135 | + with profile() as prof: |
| 136 | + model(torch.ones((2, 2))) |
| 137 | + |
| 138 | + for name, p in named_parameters.items(): |
| 139 | + self.assertGradientDetected(name, prof, _EventType.PyCall, p.grad, p) |
| 140 | + self.assertFalse( |
| 141 | + self.gradient_detected(prof, _EventType.TorchOp, p.grad), name |
| 142 | + ) |
| 143 | + assert_only_gradients(prof) |
| 144 | + |
| 145 | + check(cold_start=True) |
| 146 | + check(cold_start=False) |
| 147 | + |
| 148 | + def _test_extract_gradients_from_optimizer(self, set_to_none: bool) -> None: |
| 149 | + |
| 150 | + x = torch.ones((1,)) |
| 151 | + w0 = torch.ones((1,), requires_grad=True) |
| 152 | + w1 = torch.ones((1,), requires_grad=True) |
| 153 | + optimizer = torch.optim.SGD((w0, w1), lr=0.1, momentum=0.9) |
| 154 | + |
| 155 | + def check(cold_start: bool): |
| 156 | + self.assertEqual(w0.grad is None, cold_start) |
| 157 | + self.assertEqual(w1.grad is None, cold_start) |
| 158 | + with profile() as prof: |
| 159 | + optimizer.zero_grad(set_to_none=set_to_none) |
| 160 | + z = x.expand(4) * w0 |
| 161 | + (z * w1).sum().backward() |
| 162 | + optimizer.step() |
| 163 | + |
| 164 | + # Optimizer instrumentation runs late in the step, so we can detect |
| 165 | + # gradients for both cold and warm start. |
| 166 | + self.assertGradientDetected("w0", prof, _EventType.PyCall, w0.grad, w0) |
| 167 | + self.assertGradientDetected("w1", prof, _EventType.PyCall, w1.grad, w1) |
| 168 | + |
| 169 | + self.assertGradientDetected("w0", prof, _EventType.TorchOp, w0.grad) |
| 170 | + self.assertGradientDetected("w1", prof, _EventType.TorchOp, w1.grad) |
| 171 | + self.assertOnlyGradients(prof, (w0.grad, w1.grad)) |
| 172 | + |
| 173 | + with profile() as prof: |
| 174 | + for _ in range(2): |
| 175 | + optimizer.zero_grad(set_to_none=set_to_none) |
| 176 | + z = x.expand(4) * w0 |
| 177 | + (z * w1).sum().backward() |
| 178 | + optimizer.step() |
| 179 | + |
| 180 | + # Inspected state is cached, so if we replace gradients (as is the |
| 181 | + # case for `set_to_none=True`) our python instrumentation will not |
| 182 | + # see them. |
| 183 | + # TODO(robieta): Should `.step()` be excluded from caching? |
| 184 | + self.assertNotEqual( |
| 185 | + self.gradient_detected(prof, _EventType.PyCall, w0.grad, w0), |
| 186 | + set_to_none, |
| 187 | + ) |
| 188 | + |
| 189 | + self.assertNotEqual( |
| 190 | + self.gradient_detected(prof, _EventType.PyCall, w1.grad, w1), |
| 191 | + set_to_none, |
| 192 | + ) |
| 193 | + |
| 194 | + if set_to_none: |
| 195 | + with self.assertRaisesRegex(AssertionError, "Tensor wrongly marked"): |
| 196 | + self.assertOnlyGradients(prof, (w0.grad, w1.grad)) |
| 197 | + |
| 198 | + check(cold_start=True) |
| 199 | + check(cold_start=False) |
| 200 | + |
| 201 | + def test_extract_gradients_from_optimizer(self) -> None: |
| 202 | + self._test_extract_gradients_from_optimizer(set_to_none=False) |
| 203 | + |
| 204 | + def test_extract_gradients_from_optimizer_set_to_none(self) -> None: |
| 205 | + self._test_extract_gradients_from_optimizer(set_to_none=True) |
| 206 | + |
| 207 | + def test_extract_gradients_from_module_and_optimizer(self) -> None: |
| 208 | + # Module and optimizer are thoroughly tested individually and should be |
| 209 | + # additive. Thus we can manage with a lightweight check that they don't |
| 210 | + # interact adversely. |
| 211 | + model = torch.nn.Sequential(torch.nn.Linear(2, 1), ScaleLayer()) |
| 212 | + optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9) |
| 213 | + with profile() as prof: |
| 214 | + model(torch.ones((2, 2))).sum().backward() |
| 215 | + optimizer.step() |
| 216 | + |
| 217 | + self.assertGradientDetected( |
| 218 | + "weight", prof, _EventType.PyCall, model[0].weight.grad, model[0].weight |
| 219 | + ) |
| 220 | + |
| 221 | + |
| 222 | +if __name__ == "__main__": |
| 223 | + run_tests() |
0 commit comments