# Owner(s): ["module: dynamo"] import copy import re import unittest from textwrap import dedent from unittest.mock import patch import torch import torch._dynamo import torch._dynamo.test_case import torch.fx.traceback as fx_traceback import torch.utils._pytree as pytree from torch._dynamo.testing import CompileCounter, expectedFailureDynamic, rand_strided from torch._functorch.aot_autograd import _aot_export_function, create_functional_call from torch._subclasses.fake_tensor import FakeTensorMode from torch.fx.experimental.proxy_tensor import make_fx from torch.profiler import profile from torch.testing import FileCheck from torch.testing._internal.common_utils import compare_equal_outs_and_grads def maybe_dupe_op(x): y = x + 1 z = x + 2 if x.numel() < 5: return y, y else: return y, z def is_dynamic_shape_test(test_name): return test_name.endswith("_dynamic_shapes") aten = torch.ops.aten lib = torch.library.Library("custom", "DEF") # noqa: TOR901 lib.define("maybe_dupe_op(Tensor a) -> (Tensor, Tensor)") lib.impl("maybe_dupe_op", maybe_dupe_op, "CPU") lib.impl("maybe_dupe_op", maybe_dupe_op, "Meta") class AotAutogradFallbackTests(torch._dynamo.test_case.TestCase): def test_LSTM(self): # https://github.com/pytorch/torchdynamo/issues/1147 class Repro(torch.nn.Module): def __init__(self) -> None: super().__init__() self.self_mod_model_lstm_lstm = torch.nn.LSTM( 64, 64, num_layers=2, bidirectional=True ) def forward(self, permute: torch.Tensor): self_mod_model_lstm_lstm = self.self_mod_model_lstm_lstm(permute) return (self_mod_model_lstm_lstm,) mod = Repro() aot_mod = torch._dynamo.optimize("aot_eager")(mod) args = [((92, 4, 64), (1, 5888, 92), torch.float32, "cpu", False)] args = [ rand_strided(sh, st, dt, dev).requires_grad_(rg) for (sh, st, dt, dev, rg) in args ] eager_result = mod(*args) aot_result = aot_mod(*args) self.assertTrue(torch._dynamo.testing.same(eager_result, aot_result)) def test_mutation(self): # https://github.com/pytorch/torchdynamo/issues/1301 def fn(param, y): prev_grad = torch.is_grad_enabled() try: torch.set_grad_enabled(False) param.add_(y) finally: torch.set_grad_enabled(prev_grad) return y y = torch.randn(4) x = torch.nn.Parameter(torch.randn(4)) aot_fn = torch._dynamo.optimize("aot_eager")(fn) # This should not error: we mutated an autograd leaf under no_grad mode. aot_fn(x, y) def test_mutation1(self): def fn(_stack0: torch.Tensor, diagonal_chunked_attention_scores: torch.Tensor): getitem = diagonal_chunked_attention_scores[ ( slice(None, None, None), slice(None, None, None), slice(None, 256, None), slice(None, 257, None), ) ] _stack0[ ( slice(None, None, None), slice(None, -1, None), slice(None, None, None), slice(256, None, None), ) ] = getitem view = _stack0.view(1, 12, 1024, 513) return (view,) x = torch.randn(torch.Size([12, 4, 256, 513])) y = torch.randn(torch.Size([12, 3, 512, 513])) aot_fn = torch._dynamo.optimize("aot_eager")(fn) aot_fn(x, y) def test_negative_testing_mutation(self): def fn(_stack0: torch.Tensor, diagonal_chunked_attention_scores: torch.Tensor): getitem = diagonal_chunked_attention_scores[ ( slice(None, None, None), slice(None, None, None), slice(None, 256, None), slice(None, 257, None), ) ] _stack0 = torch.sin(_stack0) _stack0[ ( slice(None, None, None), slice(None, -1, None), slice(None, None, None), slice(256, None, None), ) ] = getitem view = _stack0.view(1, 12, 1024, 513) return (view,) x = torch.randn(torch.Size([12, 4, 256, 513])) y = torch.randn(torch.Size([12, 3, 512, 513])) aot_fn = torch._dynamo.optimize("aot_eager")(fn) aot_fn(x, y) def test_negative_testing(self): def fn(x, y): return torch.sin(x).add_(y) y = torch.randn(4) x = torch.randn(4) aot_fn = torch._dynamo.optimize("aot_eager")(fn) aot_fn(x, y) def test_call_fn_with_non_const_inputs_aot_safe(self): class ModuleSpecialFwd(torch.nn.Module): def __init__(self) -> None: super().__init__() self.conv = torch.nn.Conv2d( in_channels=3, out_channels=20, kernel_size=(5, 5) ) def _conv_forward(self, x): return self.conv._conv_forward(x, self.conv.weight, self.conv.bias) def forward(self, x): return self._conv_forward(x) # Init mod mod = ModuleSpecialFwd() rx = torch.randn([3, 10, 10]) # Run it for real real = mod(rx) # Run it in export graph, _ = torch._dynamo.export(mod)(rx) # Run exported graph with AOT self.assertTrue(torch._dynamo.testing.same(real, graph(rx))) aot_fn = torch._dynamo.optimize("aot_eager")(graph) aot_fn(rx) def test_call_fn_with_non_const_inputs_aot_unsafe(self): class ModuleSpecialFwd(torch.nn.Module): def _some_bad_fwd(self, param, y): prev_grad = torch.is_grad_enabled() try: torch.set_grad_enabled(False) param.add_(y) finally: torch.set_grad_enabled(prev_grad) return y def forward(self, x, y): return self._some_bad_fwd(x, y) # Init mod mod = ModuleSpecialFwd() x = torch.nn.Parameter(torch.randn(4)) y = torch.randn([4]) # Run it for real real = mod(x, y) # Run it in export graph, _ = torch._dynamo.export(mod)(x, y) # Assert equal self.assertTrue(torch._dynamo.testing.same(real, graph(x, y))) # Run exported graph with AOT aot_fn = torch._dynamo.optimize("aot_eager")(graph) # This should not error: we mutated an autograd leaf under no_grad mode. aot_fn(x, y) def test_call_fn_with_non_const_inputs_aot_unsafe_control_flow(self): class ModuleSpecialFwd(torch.nn.Module): def _some_bad_fwd(self, param, y): if y[0][0] < 3: return y + param return param * y def forward(self, x, y): a = x * y a = self._some_bad_fwd(a, a) b = x + y return a * b # Init mod mod = ModuleSpecialFwd() x = torch.nn.Parameter(torch.randn([2, 2])) y = torch.randn([2, 2]) # Run it for real real = mod(x, y) # Run it through optimize, with our capturing fn gms = [] counter = CompileCounter() def capturing_fn(gm, inputs): nonlocal gms gms.append(gm) return counter(gm, inputs) optimized_mod = torch._dynamo.optimize(capturing_fn)(mod) # Assert equal self.assertTrue(torch._dynamo.testing.same(real, optimized_mod(x, y))) # Uncomment to reproduce commented out graphs below. # for gm in gms: # print("GM CODE", gm.code) self.assertEqual(counter.frame_count, 4) self.assertEqual(counter.op_count, 7) # Graph 1 # def forward(self, x : torch.nn.parameter.Parameter, y : torch.Tensor): # mul = x * y; x = y = None # return (mul,) # BREAK # Graph 2 # def forward(self, y : torch.Tensor): # getitem = y[0]; y = None # getitem_1 = getitem[0]; getitem = None # lt = getitem_1 < 3; getitem_1 = None # return (lt,) # BREAK # Graph 3 # def forward(self, param : torch.Tensor, y : torch.Tensor): # add = y + param; y = param = None # return (add,) # BREAK # Graph 4 # def forward(self, _stack0 : torch.Tensor, x : torch.nn.parameter.Parameter, y : torch.Tensor): # add = x + y; x = y = None # mul = _stack0 * add; _stack0 = add = None # return (mul,) # Run fn with AOT torch._dynamo.reset() aot_fn = torch._dynamo.optimize("aot_eager")(optimized_mod) aot_fn(x, y) # Note: Dynamo recompilation guarding invalid grad # # This test is a spiritual equivalent to test_invalid_requires_grad_fake in test_autodispatch.py # The point of this test is to invoke aot_autograd in a way that would normally trigger an assertion # (This is what test_invalid_requires_grad_fake) does. However, the point of this test is to prove # that we do not hit this assertion, as dynamo recompiles correctly and protects this condition. # # Subnote: The reason for us having test_invalid_requires_grad_fake utilizing fake tensors # is because dynamo sends fake tensors down to aot_autograd. @patch("torch._functorch.config.debug_assert", True) def test_requires_grad_fake_via_dynamo_recompiles(self): class F(torch.nn.Module): def forward(self, x, y): return (x + y,) x = torch.randn(3, 3, requires_grad=True) y = torch.randn(3, 3, requires_grad=True) z = torch.randn(3, 3, requires_grad=False) cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") failure_reason = None def guard_fail_fn(failure): nonlocal failure_reason failure_reason = failure[0] fxy = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) compare_equal_outs_and_grads(self, F(), fxy, (x, y)) compare_equal_outs_and_grads(self, F(), fxy, (x, z)) self.assertIn( """tensor 'L['y']' requires_grad mismatch. expected requires_grad=1""", failure_reason, ) # Reset failure reason failure_reason = None self.assertEqual(cc.frame_count, 2) torch._dynamo.reset() # for new backend cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") fxz = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) compare_equal_outs_and_grads(self, F(), fxz, (x, z)) compare_equal_outs_and_grads(self, F(), fxz, (x, z)) self.assertEqual(cc.frame_count, 1) self.assertTrue(failure_reason is None) def test_double_backward_errors(self): # Remove this test after we get double backward to actually work for grad_output in (torch.tensor(1.0, requires_grad=True), None): x = torch.tensor(1.0, requires_grad=True) err = "torch.compile with aot_autograd does not currently support double backward" # The following cases should be equivalent: # (1) double backward entirely inside compiled function def f1(x): y = x.sin().exp() (gx,) = torch.autograd.grad( y, x, create_graph=True, grad_outputs=grad_output ) torch.autograd.grad(gx, x) return gx compiled_f1 = torch.compile(backend="aot_eager")(f1) f1(x) with self.assertRaisesRegex(RuntimeError, err): compiled_f1(x) # (2) the second half of double backward outside compiled function def f2(x): y = x.sin().exp() (gx,) = torch.autograd.grad( y, x, create_graph=True, grad_outputs=grad_output ) return gx compiled_f2 = torch.compile(backend="aot_eager")(f2) gx = compiled_f2(x) with self.assertRaisesRegex(RuntimeError, err): torch.autograd.grad(gx, x) # (3) double backward entirely outside compiled function def f3(x): y = x.sin().exp() return y compiled_f3 = torch.compile(backend="aot_eager")(f3) y = compiled_f3(x) (gx,) = torch.autograd.grad( y, x, create_graph=True, grad_outputs=grad_output ) with self.assertRaisesRegex(RuntimeError, err): torch.autograd.grad(gx, x) # create_graph=False def f4(x): y = x.sin().exp() return y compiled_f4 = torch.compile(backend="aot_eager")(f4) x = torch.tensor(1.0, requires_grad=True) y = compiled_f4(x) (gx,) = torch.autograd.grad(y, x, create_graph=False, grad_outputs=grad_output) @patch("torch._functorch.config.debug_assert", True) def test_arg_dupe_via_dynamo_recompiles(self): class F(torch.nn.Module): def forward(self, x, y): x = x.trunc_() y = y.trunc_() return (x + y,) x = torch.randn(3, 3, requires_grad=True) x1, x2, x3, x4 = x.clone(), x.clone(), x.clone(), x.clone() y = torch.randn(3, 3, requires_grad=True) y1, y2, y4 = y.clone(), y.clone(), y.clone() cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") failure_reason = None def guard_fail_fn(failure): nonlocal failure_reason failure_reason = failure[0] fxy = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) # Note: to prevent a recompilation between the two calls, # we need to clone x and y on each use. # fxy mutates the input's metadata, so otherwise dynamo will end up recompiling. fxy(x1, y1) fxy(x2, y2) self.assertTrue(failure_reason is None) # Reset failure reason failure_reason = None self.assertEqual(cc.frame_count, 1) torch._dynamo.reset() # for new backend cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") fxx = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) fxx(x3, x3) fxx(x4, y4) self.assertEqual(cc.frame_count, 2) self.assertIn("""L['x'] is L['y']""", failure_reason) @patch("torch._functorch.config.debug_assert", True) def test_arg_dupe_via_dynamo_recompiles_many_args_param_non_tensor_arg(self): class F(torch.nn.Module): def __init__(self) -> None: super().__init__() self.mean = torch.nn.Parameter(torch.randn(3, 3)) def forward(self, a, b, e, f): a.trunc_() b.trunc_() return (a + b + self.mean) * e * f a = torch.randn(3, 3, requires_grad=True) b = torch.randn(3, 3, requires_grad=True) a1, a2 = a.clone(), a.clone() b1, b2 = b.clone(), b.clone() failure_reason = None def guard_fail_fn(failure): nonlocal failure_reason failure_reason = failure[0] self.assertTrue(failure_reason is None) cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f(a1, a1, 2, 2) f(a2, b2, 2, 2) self.assertEqual(cc.frame_count, 2) self.assertIn( """L['a'] is L['b']""", failure_reason, ) torch._dynamo.reset() cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") c = torch.randn(3, 3, requires_grad=True) d = torch.randn(3, 3, requires_grad=True) c3, c4 = c.clone(), c.clone() d3, d4 = d.clone(), d.clone() f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f(c3, c3, 3, 3) f(c4, d4, 3, 3) self.assertEqual(cc.frame_count, 2) self.assertIn("""L['a'] is L['b']""", failure_reason) @patch("torch._functorch.config.debug_assert", True) def test_arg_dupe_via_dynamo_recompiles_many_with_global(self): z = None class F(torch.nn.Module): def __init__(self) -> None: super().__init__() self.mean = torch.nn.Parameter(torch.randn(3, 3)) def forward(self, a, b, e, f): a.trunc_() b.trunc_() return (a + b + z + self.mean) * e * f a = torch.randn(3, 3, requires_grad=True) b = torch.randn(3, 3, requires_grad=True) z = a a1, a2 = a.clone(), a.clone() b1, b2 = b.clone(), b.clone() failure_reason = None def guard_fail_fn(failure): nonlocal failure_reason failure_reason = failure[0] self.assertTrue(failure_reason is None) cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f(a1, a1, 2, 2) f(a2, b2, 2, 2) self.assertEqual(cc.frame_count, 2) self.assertIn( """L['a'] is L['b']""", failure_reason, ) @patch("torch._functorch.config.debug_assert", True) def test_arg_dupe_via_dynamo_recompiles_many_args_param_non_tensor_arg_list(self): class F(torch.nn.Module): def __init__(self) -> None: super().__init__() self.mean = torch.nn.Parameter(torch.randn(3, 3)) def forward(self, e, f, a, b): a.trunc_() b.trunc_() return (a + b + self.mean) * e[0] * f[0] a = torch.randn(3, 3, requires_grad=True) b = torch.randn(3, 3, requires_grad=True) a1, a2 = a.clone(), a.clone() b1, b2 = b.clone(), b.clone() failure_reason = None def guard_fail_fn(failure): nonlocal failure_reason failure_reason = failure[0] self.assertTrue(failure_reason is None) cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f([3, 2, 1], [4, 5, 6], a1, a1) f([3, 2, 1], [4, 5, 6], a2, b2) self.assertEqual(cc.frame_count, 2) self.assertIn( """L['a'] is L['b']""", failure_reason, ) torch._dynamo.reset() cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") c = torch.randn(3, 3, requires_grad=True) d = torch.randn(3, 3, requires_grad=True) c3, c4 = c.clone(), c.clone() d3, d4 = d.clone(), d.clone() f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f([3, 2, 1], [4, 5, 6], c3, c3) f([3, 2, 1], [4, 5, 6], c4, d4) self.assertEqual(cc.frame_count, 2) @patch("torch._functorch.config.debug_assert", True) def test_arg_dupe_via_dynamo_recompiles_many_args_param(self): class F(torch.nn.Module): def __init__(self) -> None: super().__init__() self.mean = torch.nn.Parameter(torch.randn(3, 3)) def forward(self, a, b): a.trunc_() b.trunc_() return a + b + self.mean a = torch.randn(3, 3, requires_grad=True) b = torch.randn(3, 3, requires_grad=True) a1, a2 = a.clone(), a.clone() b1, b2 = b.clone(), b.clone() failure_reason = None def guard_fail_fn(failure): nonlocal failure_reason failure_reason = failure[0] self.assertTrue(failure_reason is None) cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f(a1, a1) f(a2, b2) self.assertEqual(cc.frame_count, 2) self.assertIn( """L['a'] is L['b']""", failure_reason, ) torch._dynamo.reset() cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") c = torch.randn(3, 3, requires_grad=True) d = torch.randn(3, 3, requires_grad=True) c3, c4 = c.clone(), c.clone() d3, d4 = d.clone(), d.clone() f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f(c3, c3) f(c4, d4) self.assertEqual(cc.frame_count, 2) self.assertIn("""L['a'] is L['b']""", failure_reason) @patch("torch._functorch.config.debug_assert", True) def test_arg_dupe_via_dynamo_recompiles_many_args(self): class F(torch.nn.Module): def forward(self, a, b, c, d): a.trunc_() b.trunc_() c.trunc_() d.trunc_() return (a + b + c + d,) a = torch.randn(3, 3, requires_grad=True) b = torch.randn(3, 3, requires_grad=True) a1, a2, a3, a4 = a.clone(), a.clone(), a.clone(), a.clone() b1, b2, b3, b4 = b.clone(), b.clone(), b.clone(), b.clone() failure_reason = None def guard_fail_fn(failure): nonlocal failure_reason failure_reason = failure[0] self.assertTrue(failure_reason is None) cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f(a1, a1, a1, a1) f(a2, b2, b2, b2) self.assertEqual(cc.frame_count, 2) self.assertIn( """L['a'] is L['b']""", failure_reason, ) torch._dynamo.reset() cc = torch._dynamo.testing.CompileCounterWithBackend("aot_eager") c = torch.randn(3, 3, requires_grad=True) d = torch.randn(3, 3, requires_grad=True) c3, c4 = c.clone(), c.clone() d3, d4 = d.clone(), d.clone() f = torch._dynamo.optimize(cc, guard_fail_fn=guard_fail_fn)(F()) f(a3, b3, c3, c3) f(a4, b4, c4, d4) self.assertEqual(cc.frame_count, 2) self.assertIn("""L['c'] is L['d']""", failure_reason) def test_alias_inputs(self): def fn(): a = torch.tensor([1]) a = a[0:1] b = a.squeeze() a[0] = 0 if a[0] < 1e5: pass a[0] = 2 return b ref_output = fn() aot_fn = torch._dynamo.optimize("aot_eager")(fn) actual_output = aot_fn() self.assertEqual(ref_output, actual_output) def test_grad_inputs_alias_inputs(self): class Test(torch.autograd.Function): @staticmethod def forward(ctx, x, y): ctx.save_for_backward(x) return y @staticmethod def backward(ctx, grad): (x,) = ctx.saved_tensors return x, grad def fn(x, y): return Test.apply(x, y) x = torch.ones(1, requires_grad=True) y = torch.ones(1, requires_grad=True) compiled_fn = torch.compile(fn, backend="aot_eager") out = compiled_fn(x, y) out.sum().backward() @expectedFailureDynamic # https://github.com/pytorch/pytorch/issues/103539 @torch._dynamo.config.patch(automatic_dynamic_shapes=False) @patch("torch._functorch.config.debug_assert", True) def test_multiple_aot_autograd_calls_dupe_args(self): # this is just dealing with the fact that # aot_module_simplified expects submods to always return tuples/lists class WrapperModule(torch.nn.Module): def __init__(self, mod): super().__init__() self.mod = mod def forward(self, *args): out = self.mod(*args) if isinstance(out, (list, tuple)): return out return (out,) def compile_submod(input_mod, args): from functorch.compile import nop from torch._functorch.aot_autograd import aot_module_simplified class WrapperModule(torch.nn.Module): def __init__(self) -> None: super().__init__() self.original = input_mod self.submod = aot_module_simplified(input_mod, args, nop) def forward(self, *args): return self.submod(*args) return WrapperModule() def test_compile(fx_g, example_inps): split_gm = torch.fx.passes.split_module.split_module( fx_g, None, lambda node: 1 if "mul" in str(node) else 0 ) submod_1_inps = split_gm.submod_0(*example_inps) split_gm.submod_0 = compile_submod( WrapperModule(split_gm.submod_0), example_inps ) split_gm.submod_1 = compile_submod( WrapperModule(split_gm.submod_1), submod_1_inps ) return split_gm @torch._dynamo.optimize(test_compile) def f(a): b, c = torch.ops.custom.maybe_dupe_op(a) return (b.mul_(c),) f(torch.ones(4)) f(torch.ones(6)) def test_nn_parameter_construction(self): # https://github.com/pytorch/pytorch/issues/99569 def fn(x): y = x.sin() z = torch.nn.Parameter(torch.ones(1)) return y + z x = torch.rand((4, 4)) opt_fn = torch._dynamo.optimize("aot_eager")(fn) self.assertTrue(torch._dynamo.testing.same(fn(x), opt_fn(x))) def test_aot_sequence_nr(self): class Model(torch.nn.Module): def __init__(self) -> None: super().__init__() self.conv1 = torch.nn.Conv2d( in_channels=16, out_channels=16, kernel_size=(1, 1), stride=1, padding="same", bias=True, ) self.bn1 = torch.nn.BatchNorm2d(num_features=16) self.relu1 = torch.nn.ReLU() self.fc1 = torch.nn.Linear(in_features=1638400, out_features=1) self.loss_fn = torch.nn.L1Loss() def forward(self, x, target): y = x x = self.conv1(x) x = self.bn1(x) x = self.relu1(x) x = x + y x = torch.flatten(x) x = self.fc1(x) output = self.loss_fn(x, target) return (output,) mod = Model() mod.train() x = torch.rand(100, 16, 32, 32, requires_grad=True) target = torch.rand(1) # Use dynamo export to get the fx graph module g_mod, _ = torch._dynamo.export(mod, x, target) def _prepare_model_args(): named_parameters = dict(g_mod.named_parameters(remove_duplicate=False)) named_buffers = dict(g_mod.named_buffers(remove_duplicate=False)) params_and_buffers = { **dict(named_parameters), **dict(named_buffers), } params_and_buffers_flat, params_spec = pytree.tree_flatten( params_and_buffers ) params_len = len(params_and_buffers_flat) functional_call = create_functional_call(g_mod, params_spec, params_len) return params_and_buffers_flat, functional_call full_args, fn_to_trace = _prepare_model_args() param_and_buf_len = len(full_args) full_args.extend([x, target]) # aot_export requires a graph mod input of fwd graph # returns the full fwd/bwd graph in graph mod format with torch.enable_grad(), fx_traceback.preserve_node_meta(): fx_g, _, _, _ = _aot_export_function( fn_to_trace, full_args, decompositions=None, num_params_buffers=param_and_buf_len, no_tangents=True, ) # Walk all the nodes in fx graph. # Write the resulting ops to a table min_seq_nr = -1 seq_table = "SeqNr|OrigAten|SrcFn|FwdSrcFn\n" for node in fx_g.graph.nodes: if "call_" in node.op and "getitem" not in str(node.target): seq_nr = node.meta.get("seq_nr", -1) if seq_nr < 0: continue if min_seq_nr < 0: min_seq_nr = seq_nr source_fn_stack = node.meta.get("source_fn_stack", []) orig_aten = node.meta.get("original_aten", "") mod_name = "" if len(source_fn_stack) > 0: mod_name = source_fn_stack[-1][0] # Make all seq_nr relative so it starts at 0 seq_nr = seq_nr - min_seq_nr # For backward nodes, also test that metadata from the corresponding # forward node is copied over. fwd_source_fn_stack = node.meta.get("fwd_source_fn_stack", []) fwd_mod_name = "" if len(fwd_source_fn_stack): fwd_mod_name = fwd_source_fn_stack[-1][0] seq_table = ( seq_table + f"{seq_nr}|{orig_aten}|{mod_name}|{fwd_mod_name}\n" ) self.maxDiff = None self.assertExpectedInline( seq_table, dedent( """\ SeqNr|OrigAten|SrcFn|FwdSrcFn 0|aten.convolution.default|l__self___conv1| 0|aten.add.Tensor|l__self___bn1| 1|aten._native_batch_norm_legit_functional.default|l__self___bn1| 2|aten.relu.default|l__self___relu1| 2|aten.detach.default|l__self___relu1| 2|aten.detach.default|l__self___relu1| 3|aten.add.Tensor|add| 4|aten.view.default|flatten| 5|aten.view.default|l__self___fc1| 6|aten.t.default|l__self___fc1| 7|aten.addmm.default|l__self___fc1| 8|aten.view.default|l__self___fc1| 9|aten.sub.Tensor|l__self___loss_fn| 10|aten.abs.default|l__self___loss_fn| 11|aten.mean.default|l__self___loss_fn| 11|aten.ones_like.default||l__self___loss_fn 11|aten.expand.default||l__self___loss_fn 11|aten.div.Scalar||l__self___loss_fn 10|aten.sgn.default||l__self___loss_fn 10|aten.mul.Tensor||l__self___loss_fn 8|aten.view.default||l__self___fc1 7|aten.t.default||l__self___fc1 7|aten.mm.default||l__self___fc1 7|aten.t.default||l__self___fc1 7|aten.mm.default||l__self___fc1 7|aten.t.default||l__self___fc1 7|aten.sum.dim_IntList||l__self___fc1 7|aten.view.default||l__self___fc1 6|aten.t.default||l__self___fc1 5|aten.view.default||l__self___fc1 4|aten.view.default|| 2|aten.detach.default||l__self___relu1 2|aten.detach.default||l__self___relu1 2|aten.threshold_backward.default||l__self___relu1 1|aten.native_batch_norm_backward.default||l__self___bn1 0|aten.convolution_backward.default||l__self___conv1 11|aten.add.Tensor||l__self___loss_fn """ ), ) def test_split_with_sizes_aot_autograd_cleans_up_traceback_meta(self): from torch._functorch.aot_autograd import setup_stacktrace_preservation_hooks def fn(result, split_sizes): rs = torch.ops.aten.split_with_sizes(result, split_sizes.tolist()) return rs example_inputs = ( torch.randn(32, requires_grad=True), torch.tensor((7, 16, 9)), ) outs = fn(*example_inputs) setup_stacktrace_preservation_hooks([out.grad_fn for out in outs]) with fx_traceback.preserve_node_meta(): (outs[0].sum() + outs[1].sum() + outs[2].sum()).backward() self.assertNotIn("grad_fn_seq_nr", fx_traceback.current_meta) self.assertNotIn("in_grad_fn", fx_traceback.current_meta) # https://github.com/pytorch/pytorch/issues/110121 def test_aot_export_joint_simple_repro(self): class Mod(torch.nn.Module): def __init__(self, *args, **kwargs) -> None: super().__init__(*args, **kwargs) self.linear = torch.nn.Linear(5, 7) def forward(self, x): return self.linear(x) def mini_backend(gm, sample_inputs): from torch._functorch.aot_autograd import aot_export_joint_simple fake_mode = torch._dynamo.utils.detect_fake_mode(sample_inputs) with patch.object(fake_mode, "allow_non_fake_inputs", True), fake_mode: return aot_export_joint_simple(gm, sample_inputs, trace_joint=False) sample_inputs = [torch.rand((3, 4, 5))] model = Mod() m_compiled = torch.compile(model, backend=mini_backend) out_ref = model(*sample_inputs) out_test = m_compiled(*sample_inputs) self.assertEqual(out_ref, out_test) def test_eager_sequence_nr(self): class Model(torch.nn.Module): def __init__(self) -> None: super().__init__() self.conv1 = torch.nn.Conv2d( in_channels=16, out_channels=16, kernel_size=(1, 1), stride=1, padding="same", bias=True, ) self.bn1 = torch.nn.BatchNorm2d(num_features=16) self.relu1 = torch.nn.ReLU() self.fc1 = torch.nn.Linear(in_features=1638400, out_features=1) self.loss_fn = torch.nn.L1Loss() def forward(self, x, target): y = x x = self.conv1(x) x = self.bn1(x) x = self.relu1(x) x = x + y x = torch.flatten(x) x = self.fc1(x) output = self.loss_fn(x, target) return (output,) def grad_with_create_graph(mod, x, target): y = mod(x, target) # Set create_graph=True to ensure that the sequence_nr # for backward ops continues to count down. (gx,) = torch.autograd.grad( y[0], x, create_graph=True, grad_outputs=grad_output ) return gx x = torch.rand(100, 16, 32, 32, requires_grad=True) target = torch.rand(1) mod = Model() args = [mod, x, target] grad_output = torch.tensor(1.0, requires_grad=True) compiled_f1 = torch.compile(backend="aot_eager")(grad_with_create_graph) model_instance = compiled_f1 with profile( activities=[torch.profiler.ProfilerActivity.CPU], record_shapes=True, ) as kineto_prof: res = model_instance(*args) bwd_set = set() prof_str = "SeqNr|Thread|FwdThread|Name\n" for event in kineto_prof.events(): if event.sequence_nr >= 0: prof_str = ( prof_str + f"{event.sequence_nr}|{event.thread}" f"|{event.fwd_thread}|{event.name}|\n" ) if re.search(r"Backward[01]", event.name): bwd_set.add(event.sequence_nr) self.assertTrue(len(bwd_set), 13) def test_aot_grad_mode_mutation(self): for compiler in ["aot_eager", "inductor"]: def f(x): y = x * x torch.set_grad_enabled(False) return y.clone(), y f_compiled = torch.compile(f, backend=compiler, fullgraph=True) torch.set_grad_enabled(True) x = torch.ones(3, requires_grad=True) * 3 y_ref = f(x) self.assertEqual(torch.is_grad_enabled(), False) torch.set_grad_enabled(True) y = f_compiled(x) self.assertEqual(torch.is_grad_enabled(), False) torch.set_grad_enabled(True) self.assertEqual(y_ref, y) self.assertIsNone(y_ref[0].grad_fn) self.assertIsNone(y[0].grad_fn) self.assertIsNotNone(y_ref[1].grad_fn) self.assertIsNotNone(y[1].grad_fn) # Check that the grad computed for the inputs, given the input, is the same # The tangent to `y[0]`, which has grad_required=False, is irrelevant self.assertEqual( sum(y_ref[1].grad_fn(torch.tensor([-1.0, 2.0, 0.0]))), sum( x for x in y[1].grad_fn.apply(None, torch.tensor([-1.0, 2.0, 0.0])) if x is not None ), ) def test_aot_autograd_raises_invalid_leaf_set(self): @torch.compile def f(x): x.set_(torch.ones(2)) # We still want to make sure that this raises x = torch.ones(2, requires_grad=True) with self.assertRaisesRegex( RuntimeError, "is being used in an in-place operation" ): f(x) def test_aot_autograd_expand_mutation_functionalizes(self): def fn(x): y = x.expand(3, *x.shape) y[0, 0].add_(5) return y opt_fn = torch.compile(fn, backend="aot_eager") x = torch.arange(6) x_opt = x.clone().detach() self.assertEqual(fn(x), opt_fn(x_opt)) self.assertEqual(x, x_opt) def test_aot_autograd_expand_mutation_backwards(self): def fn(x, z): y = x.expand(3, *x.shape) y[1, 1].mul_(5) ret = y * z return ret opt_fn = torch.compile(fn, backend="aot_eager") x = torch.arange(6, dtype=torch.float) z = x.clone().detach() x_opt = x.clone().detach() z_opt = x.clone().detach() z.requires_grad = True z_opt.requires_grad = True res = fn(x, z) opt_res = opt_fn(x_opt, z_opt) self.assertEqual(res, opt_res) res.sum().backward() opt_res.sum().backward() self.assertEqual(x, x_opt) self.assertEqual(z.grad, z_opt.grad) def test_data_ptr_access_copy(self): import torch._functorch.config as _config with _config.patch(fake_tensor_allow_unsafe_data_ptr_access=False): with FakeTensorMode(): x = torch.randn(3) y = copy.copy(x) self.assertEqual(y.shape, x.shape) def test_data_ptr_access_fails_in_forward(self): with torch.library._scoped_library("mylib", "FRAGMENT") as lib: torch.library.define("mylib::foo", "(Tensor x) -> Tensor", lib=lib) @torch.library.impl("mylib::foo", "CompositeImplicitAutograd", lib=lib) def _(x): x.data_ptr() return x.clone() x = torch.randn(3) def data_ptr_graph_input(x): r0 = torch.ops.mylib.foo(x) return r0 def data_ptr_graph_intermediate(x): y = x.clone() r0 = torch.ops.mylib.foo(y) return r0 tests = [data_ptr_graph_input, data_ptr_graph_intermediate] def ctx(): return self.assertRaisesRegex( RuntimeError, "Cannot access data pointer" ) for f in tests: with ctx(): make_fx(f, tracing_mode="fake")(x) with ctx(): make_fx(f, tracing_mode="symbolic")(x) with ctx(): torch.compile(f, backend="eager", fullgraph=True)(x) def test_data_ptr_access_fails_in_backward(self): with torch.library._scoped_library("mylib", "FRAGMENT") as lib: torch.library.define("mylib::foo", "(Tensor x) -> Tensor", lib=lib) backward_called = False class Foo(torch.autograd.Function): @staticmethod def forward(ctx, x): return x.clone() @staticmethod def backward(ctx, grad): nonlocal backward_called backward_called = True grad.data_ptr() return grad.clone() @torch.library.impl("mylib::foo", "CompositeImplicitAutograd", lib=lib) def _(x): return Foo.apply(x) def f(x): return torch.ops.mylib.foo(x) x = torch.randn(3, requires_grad=True) with self.assertRaisesRegex(RuntimeError, "Cannot access data pointer"): y = torch.compile(f, backend="aot_eager", fullgraph=True)(x) self.assertTrue(backward_called) # We don't know how to catch multiple mutations to the same memory location @unittest.expectedFailure def test_aot_autograd_expand_mutation_error(self): def fn(x): y = x.expand(3, *x.shape) y[0:3, 0].add_(5) return y opt_fn = torch.compile(fn, backend="aot_eager") x = torch.arange(6) x_opt = x.clone().detach() with self.assertRaises(Exception): fn(x) with self.assertRaises(Exception): opt_fn(x_opt) @torch._functorch.config.patch(donated_buffer=True) def test_donated_buffer1(self): logger_name = "torch._functorch._aot_autograd.jit_compile_runtime_wrappers" @torch.compile() def relu(x): return torch.nn.functional.relu(x) with self.assertLogs(logger_name, level="INFO") as captured: relu(torch.rand([3, 3], requires_grad=True)).sum().backward() if is_dynamic_shape_test(self._testMethodName): # an extra symint exists expected_msg = "bw_donated_idxs=[1]" else: expected_msg = "bw_donated_idxs=[0]" # le is a donated buffer from relu FileCheck().check(expected_msg).run("\n".join(captured.output)) @torch._functorch.config.patch("donated_buffer", True) def test_donated_buffer2(self): logger_name = "torch._functorch._aot_autograd.jit_compile_runtime_wrappers" # we will re-use the graph for g across f1 and f2 @torch.compile() def g(activation, param2): return torch.matmul(activation, param2) def f(inp, param1, param2): activation = inp + param1 return g(activation, param2) inp = torch.ones(4, 4) param1 = torch.ones(4, 4, requires_grad=True) param2 = torch.ones(4, 4, requires_grad=True) with self.assertLogs(logger_name, level="INFO") as captured: f(inp, param1, param2).sum().backward() FileCheck().check("bw_donated_idxs=[]").run("\n".join(captured.output)) @torch._functorch.config.patch("donated_buffer", True) def test_donated_buffer3(self): logger_name = "torch._functorch._aot_autograd.jit_compile_runtime_wrappers" # we will re-use the graph for g across f1 and f2 @torch.compile() def g(activation, param2): return torch.matmul(activation, param2) def f(inp, param1, param2): # exp saves it output (the activation) for bw activation = torch.exp(inp + param1) return g(activation, param2) inp = torch.ones(4, 4) param1 = torch.ones(4, 4, requires_grad=True) param2 = torch.ones(4, 4, requires_grad=True) with self.assertLogs(logger_name, level="INFO") as captured: f(inp, param1, param2).sum().backward() FileCheck().check("bw_donated_idxs=[]").run("\n".join(captured.output)) @torch._functorch.config.patch("donated_buffer", True) def test_donated_buffer4(self): logger_name = "torch._functorch._aot_autograd.jit_compile_runtime_wrappers" class Mod(torch.nn.Module): def __init__(self) -> None: super().__init__() self.param = torch.nn.Parameter(torch.zeros([2, 2])) def forward(self, x: torch.Tensor) -> torch.Tensor: return torch.nn.functional.relu(x) + self.param mod = Mod() mod = torch.compile(mod) inp = torch.ones([2, 2], requires_grad=True) with self.assertLogs(logger_name, level="INFO") as captured: mod(inp).sum().backward() # Forward graph: # %primals_1 : [num_users=1] = placeholder[target=primals_1] # %primals_2 : [num_users=1] = placeholder[target=primals_2] # %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%primals_2,), kwargs = {}) # %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%relu, %primals_1), kwargs = {}) # %le : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%relu, 0), kwargs = {}) # return [add, le] # # `le` is a donated buffer FileCheck().check("bw_donated_idxs=[0]").run("\n".join(captured.output)) @torch._functorch.config.patch("donated_buffer", True) def test_donated_buffer5(self): logger_name = "torch._functorch._aot_autograd.jit_compile_runtime_wrappers" @torch.compile() def f(x, z): y = x.view(2, 3) z = torch.nn.functional.relu(z) return torch.mm(y, x) + z inp = [ torch.rand([3, 2], requires_grad=True), torch.rand([2, 2], requires_grad=True), ] with self.assertLogs(logger_name, level="INFO") as captured: f(*inp).sum().backward() # Forward graph: # %primals_1 : [num_users=3] = placeholder[target=primals_1] # %primals_2 : [num_users=1] = placeholder[target=primals_2] # %view : [num_users=1] = call_function[target=torch.ops.aten.view.default](args = (%primals_1, [2, 3]), kwargs = {}) # %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%primals_2,), kwargs = {}) # %mm : [num_users=1] = call_function[target=torch.ops.aten.mm.default](args = (%view, %primals_1), kwargs = {}) # %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mm, %relu), kwargs = {}) # %le : [num_users=1] = call_function[target=torch.ops.aten.le.Scalar](args = (%relu, 0), kwargs = {}) # return [add, primals_1, le] # # `le` is a donated buffer but primals_1 is not. FileCheck().check("bw_donated_idxs=[1]").run("\n".join(captured.output)) @torch._functorch.config.patch("donated_buffer", True) def test_donated_buffer_with_retain_or_create_graph1(self): # Gives non-empty bw_donated_idxs class Mod(torch.nn.Module): def __init__(self) -> None: super().__init__() self.param = torch.nn.Parameter(torch.zeros([3, 3])) def forward(self, x): return torch.nn.functional.relu(x) + self.param inp = torch.randn(3, 3, requires_grad=True) mod = torch.compile(Mod()) for _ in range(5): mod(inp).sum().backward() @torch._functorch.config.patch("donated_buffer", True) def test_donated_buffer_with_retain_or_create_graph2(self): # Gives non-empty bw_donated_idxs class Mod(torch.nn.Module): def __init__(self) -> None: super().__init__() self.param = torch.nn.Parameter(torch.zeros([3, 3])) def forward(self, x): return torch.nn.functional.relu(x) + self.param inp = torch.randn(3, 3, requires_grad=True) mod = torch.compile(Mod()) out = mod(inp).sum() for _ in range(5): out.backward(retain_graph=True) out.backward() @torch._functorch.config.patch("donated_buffer", True) def test_donated_buffer_with_retain_or_create_graph3(self): # Gives non-empty bw_donated_idxs class Mod(torch.nn.Module): def __init__(self) -> None: super().__init__() self.param = torch.nn.Parameter(torch.zeros([3, 3])) def forward(self, x): return torch.nn.functional.relu(x) + self.param inp = torch.randn(3, 3, requires_grad=True) mod = torch.compile(Mod()) mod(inp).sum().backward(create_graph=True) out = mod(inp).sum() for _ in range(5): out.backward(retain_graph=True) out.backward() @torch._functorch.config.patch("donated_buffer", True) def test_donated_buffer_with_retain_or_create_graph4(self): # Gives non-empty bw_donated_idxs class Mod(torch.nn.Module): def __init__(self) -> None: super().__init__() self.param = torch.nn.Parameter(torch.zeros([3, 3])) def forward(self, x): return torch.nn.functional.relu(x) + self.param inp = torch.randn(3, 3, requires_grad=True) mod = torch.compile(Mod()) mod(inp).sum().backward() out = mod(inp).sum() with self.assertRaisesRegex( RuntimeError, r"This backward function was compiled with non-empty donated " r"buffers which requires create_graph=False and retain_graph=False. " r"Please keep backward\(create_graph=False, retain_graph=False\) " r"across all backward\(\) function calls, or set " r"torch._functorch.config.donated_buffer=False to disable " r"donated buffer.", ): out.backward(retain_graph=True) if __name__ == "__main__": from torch._dynamo.test_case import run_tests run_tests()