| # Copyright (c) Facebook, Inc. and its affiliates. |
| # All rights reserved. |
| # |
| # This source code is licensed under the BSD-style license found in the |
| # LICENSE file in the root directory of this source tree. |
| |
| from torch.testing._internal.common_utils import TestCase, run_tests, is_iterable_of_tensors |
| import torch |
| import torch.nn.functional as F |
| from torch import Tensor |
| import functools |
| import itertools |
| import copy |
| import warnings |
| import unittest |
| from torch.testing._internal.common_device_type import instantiate_device_type_tests, \ |
| skipCUDAIfNoMagma |
| from torch.testing._internal.common_device_type import ops, onlyCPU |
| from torch.testing._internal.common_dtype import floating_types_and, integral_types |
| from functorch_lagging_op_db import functorch_lagging_op_db |
| from functorch_additional_op_db import additional_op_db |
| from common_utils import ( |
| get_fallback_and_vmap_exhaustive, |
| get_exhaustive_batched_inputs, |
| opinfo_in_dict, |
| xfail, |
| skip, |
| skipOps, |
| check_vmap_fallback, |
| IS_FBCODE, |
| ) |
| import types |
| from torch.utils._pytree import tree_flatten, tree_unflatten, tree_map |
| from functorch import grad, vjp, vmap |
| import torch.autograd.forward_ad as fwAD |
| from functorch._src.eager_transforms import _as_tuple, jvp |
| from functorch.compile import decomposition_table |
| aten = torch.ops.aten |
| |
| # Version of autograd.grad that handles outputs that don't depend on inputs |
| def _autograd_grad(outputs, inputs, grad_outputs=None, retain_graph=False, create_graph=True): |
| inputs, inputs_spec = tree_flatten(inputs) |
| result = [torch.zeros_like(inp) for inp in inputs] |
| diff_argnums = tuple(i for i, inp in enumerate(inputs) if inp.requires_grad) |
| inputs = tuple(inputs[i] for i in diff_argnums) |
| if grad_outputs is None: |
| diff_outputs = tuple(out for out in outputs if out.requires_grad) |
| else: |
| something = [(out, go) for out, go in zip(outputs, grad_outputs) |
| if out.requires_grad] |
| if len(something) == 0: |
| diff_outputs, grad_outputs = (), () |
| else: |
| diff_outputs, grad_outputs = zip(*something) |
| if len(diff_outputs) == 0: |
| return tuple(torch.zeros_like(inp) for inp in inputs) |
| grad_inputs = torch.autograd.grad(diff_outputs, inputs, grad_outputs, |
| retain_graph=retain_graph, |
| create_graph=create_graph, |
| allow_unused=True) |
| grad_inputs = tuple(torch.zeros_like(inp) if gi is None else gi |
| for gi, inp in zip(grad_inputs, inputs)) |
| for idx, grad_inp in zip(diff_argnums, grad_inputs): |
| result[idx] = grad_inp |
| return tree_unflatten(result, inputs_spec) |
| |
| |
| def diff_arg(arg, requires_grad=True): |
| def is_differentiable_arg(arg): |
| if requires_grad: |
| return arg.requires_grad |
| else: |
| return arg.is_floating_point() or arg.is_complex() |
| if is_iterable_of_tensors(arg): |
| if all([is_differentiable_arg(a) for a in arg]): |
| return True |
| if all([not is_differentiable_arg(a) for a in arg]): |
| return False |
| raise RuntimeError("NYI: The test runner can't handle this") |
| return isinstance(arg, Tensor) and is_differentiable_arg(arg) |
| |
| |
| # Given f, returns an f' such that: |
| # - f' takes only positional arguments |
| # - All arguments to f' are floating-point Tensors |
| # - All outputs of f' are floating-point Tensors |
| def normalize_op_input_output2(f, args, kwargs, output_process_fn_grad=None, requires_grad=True): |
| flat_args, args_spec = tree_flatten(args) |
| diff_argnums = tuple(i for i, arg in enumerate(flat_args) if diff_arg(arg, requires_grad=requires_grad)) |
| |
| assert len(diff_argnums) > 0 |
| primals = tuple(flat_args[i] for i in diff_argnums) |
| |
| @functools.wraps(f) |
| def wrapped(*primals): |
| _args = list(flat_args) |
| for num, arg in zip(diff_argnums, primals): |
| _args[num] = arg |
| _args = tree_unflatten(_args, args_spec) |
| result = f(*_args, **kwargs) |
| if output_process_fn_grad is not None: |
| result = output_process_fn_grad(result) |
| if isinstance(result, tuple): |
| # TODO: Remove the following hack for namedtuples |
| result = tuple(result) |
| result = tuple(r for r in result if torch.is_floating_point(r)) |
| assert len(result) > 0 |
| return result |
| return wrapped, primals |
| |
| def normalize_op_input_output(f, sample, requires_grad=True): |
| args = tuple([sample.input] + list(sample.args)) |
| return normalize_op_input_output2(f, args, sample.kwargs, sample.output_process_fn_grad, requires_grad=requires_grad) |
| |
| def ref_vjp(f, *primals): |
| result = f(*primals) |
| |
| def wrapped(cotangents): |
| return _autograd_grad(_as_tuple(result), primals, _as_tuple(cotangents)) |
| |
| return result, wrapped |
| |
| def ref_jvp(f, primals, tangents): |
| with fwAD.dual_level(): |
| duals = tuple(fwAD.make_dual(p, t) for p, t in zip(primals, tangents)) |
| result_duals = f(*duals) |
| result_duals, spec = tree_flatten(result_duals) |
| primals_out, tangents_out = zip(*(fwAD.unpack_dual(d) for d in result_duals)) |
| return tree_unflatten(primals_out, spec), tree_unflatten(tangents_out, spec) |
| |
| # Returns a new function g(*args, *cotangents) that computes vjps and |
| # sample (*args, *cotangents) |
| def get_vjpfull_variant(f, sample): |
| fn, primals = normalize_op_input_output(f, sample) |
| result = fn(*primals) |
| cotangents = _as_tuple( |
| tree_map(lambda x: torch.randn_like(x, requires_grad=True), result)) |
| num_primals = len(primals) |
| args = (*primals, *cotangents) |
| |
| @functools.wraps(f) |
| def wrapped(*args): |
| primals = args[:num_primals] |
| cotangents = args[num_primals:] |
| result, vjp_fn = vjp(fn, *primals) |
| if isinstance(result, torch.Tensor): |
| assert len(cotangents) == 1 |
| cotangents = cotangents[0] |
| return vjp_fn(cotangents) |
| |
| return wrapped, args |
| |
| def get_jvp_variant(f, sample): |
| # We want this higher-order variant of jvp, so that it can |
| # be used to wrap vmap |
| fn, primals = normalize_op_input_output(f, sample, requires_grad=False) |
| tangents = _as_tuple( |
| tree_map(lambda x: torch.randn_like(x), primals)) |
| |
| @functools.wraps(f) |
| def wrapped(*args): |
| tangents = args |
| primals_out, tangents_out = jvp(fn, primals, tangents) |
| |
| if isinstance(primals_out, torch.Tensor): |
| return (primals_out, tangents_out) |
| else: |
| flat_primals_out, _ = tree_flatten(primals_out) |
| flat_tangents_out, _ = tree_flatten(tangents_out) |
| return tuple(flat_primals_out + flat_tangents_out) |
| |
| return wrapped, tangents |
| |
| |
| def is_inplace(op, variant): |
| if hasattr(variant, "__wrapped__"): |
| return variant.__wrapped__ is op.get_inplace() |
| return variant is op.get_inplace() |
| |
| |
| vjp_fail = { |
| xfail('nn.functional.dropout'), # randomness testing artifact |
| xfail('nn.functional.rrelu'), # randomness testing artifact |
| xfail('linalg.cholesky'), |
| xfail('linalg.inv'), |
| xfail('linalg.matrix_power'), |
| xfail('tensor_split'), |
| xfail('to_sparse'), |
| xfail('nn.functional.ctc_loss'), |
| xfail('nn.functional.fractional_max_pool3d'), |
| xfail('nn.functional.fractional_max_pool2d'), |
| } |
| |
| class TestOperators(TestCase): |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=(torch.float,)) |
| @skipOps('TestOperators', 'test_grad', vjp_fail) |
| def test_grad(self, device, dtype, op): |
| if op.name in vjp_fail: |
| self.skipTest("Skipped; Expected failures") |
| return |
| |
| if not op.supports_autograd: |
| self.skipTest("Skipped! Autograd not supported.") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=True) |
| |
| # TODO: test in-place |
| if is_inplace(op, op.get_op()): |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| for sample in samples: |
| args = [sample.input] + list(sample.args) |
| kwargs = sample.kwargs |
| |
| diff_argnums = tuple(i for i, arg in enumerate(args) if diff_arg(arg)) |
| assert len(diff_argnums) > 0 |
| diff_args = tuple(args[i] for i in diff_argnums) |
| |
| def wrapped_fn(*args, **kwargs): |
| result = op(*args, **kwargs) |
| if sample.output_process_fn_grad is not None: |
| result = sample.output_process_fn_grad(result) |
| |
| # Reduce into single value for grad |
| if isinstance(result, torch.Tensor): |
| return result.sum() |
| result = sum([res.sum() for res in result]) |
| return result |
| |
| result = grad(wrapped_fn, diff_argnums)(*args, **kwargs) |
| expected = _autograd_grad(_as_tuple(wrapped_fn(*args, **kwargs)), diff_args) |
| |
| self.assertEqual(result, expected) |
| |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=(torch.float,)) |
| @skipOps('TestOperators', 'test_jvp', set({ |
| xfail('nn.functional.dropout'), # randomness testing artifact; not actually a problem |
| xfail('nn.functional.rrelu'), # randomness testing artifact; not actually a problem |
| |
| # See https://github.com/pytorch/pytorch/issues/69034 |
| # RuntimeError: expected scalar type double but found float |
| xfail('minimum'), |
| xfail('min', 'binary'), |
| xfail('maximum'), |
| xfail('max', 'binary'), |
| |
| # The following don't have a forward-mode AD formula in PyTorch core |
| # (check derivatives.yaml). |
| xfail('var_mean'), |
| xfail('std_mean'), |
| # https://gist.github.com/zou3519/f62a167fb46cda01d7f238f61dd9ccf9 |
| xfail('linalg.eigvalsh'), |
| # https://gist.github.com/zou3519/b86616d01ca375a4bd17403277f49225 |
| xfail('nn.functional.dropout', device_type='cuda'), |
| |
| # ============================================= |
| # NB: The above failures also fail using PyTorch core's |
| # forward-mode AD and vmap. |
| # The failures below are functorch-specific issues |
| # ============================================= |
| |
| # Composite ops that do bad things. Need to be fixed in PyTorch core. |
| # RuntimeError: Cannot access data pointer of Tensor that doesn't have storage |
| xfail('linalg.inv'), |
| xfail('linalg.matrix_power'), |
| xfail('linalg.cholesky'), |
| xfail('tensor_split'), |
| |
| # Causing a CUDA assert, needs investigation |
| skip('div', 'floor_rounding', device_type='cuda'), |
| skip('div', 'no_rounding_mode', device_type='cuda'), |
| skip('div', 'trunc_rounding', device_type='cuda'), |
| skip('true_divide', device_type='cuda'), |
| })) |
| def test_jvp(self, device, dtype, op): |
| # TODO: when we change supports_autograd to supports_backward_ad, also change in this file |
| if not op.supports_forward_ad: |
| self.skipTest("Skipped! Forward AD not supported.") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=True) |
| # TODO: test in-place |
| if is_inplace(op, op.get_op()): |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| for sample in samples: |
| fn, primals = normalize_op_input_output(op, sample, requires_grad=False) |
| tangents = tree_map(lambda x: torch.randn_like(x), primals) |
| primal_outs, tangent_outs = jvp(fn, primals, tangents) |
| expected_primal_outs, expected_tangent_outs = ref_jvp(fn, primals, tangents) |
| self.assertEqual(primal_outs, expected_primal_outs) |
| self.assertEqual(tangent_outs, expected_tangent_outs) |
| |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=(torch.float,)) |
| @skipOps('TestOperators', 'test_vjp', vjp_fail.union({ |
| skip('nn.functional.conv_transpose3d', device_type='cuda'), # numerical precision |
| })) |
| def test_vjp(self, device, dtype, op): |
| if not op.supports_autograd: |
| self.skipTest("Skipped! Autograd not supported.") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=True) |
| |
| # TODO: test in-place |
| if is_inplace(op, op.get_op()): |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| def _test(_op): |
| for sample in samples: |
| fn, primals = normalize_op_input_output(_op, sample) |
| result = fn(*primals) |
| cotangents = tree_map(lambda x: torch.randn_like(x), result) |
| |
| out, vjp_fn = vjp(fn, *primals) |
| self.assertEqual(out, result) |
| result_vjps = vjp_fn(cotangents) |
| |
| _, vjp_fn = ref_vjp(fn, *primals) |
| expected_vjps = vjp_fn(cotangents) |
| |
| self.assertEqual(result_vjps, expected_vjps) |
| |
| _test(op) |
| for a_op in op.aliases: |
| _test(a_op) |
| |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=(torch.float,)) |
| @skipOps('TestOperators', 'test_vjpvjp', vjp_fail.union({ |
| skip('nn.functional.conv_transpose3d'), # numerical precision problem |
| })) |
| def test_vjpvjp(self, device, dtype, op): |
| if not op.supports_autograd: |
| self.skipTest("Skipped! Autograd not supported.") |
| return |
| if not op.supports_gradgrad: |
| self.skipTest("Skipped! Operation does not support gradgrad") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=True) |
| |
| # TODO: test in-place |
| if is_inplace(op, op.get_op()): |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| for sample in samples: |
| fn, args = get_vjpfull_variant(op, sample) |
| result = fn(*args) |
| cotangents = tree_map(lambda x: torch.randn_like(x), result) |
| |
| # Compute vjp of vjp |
| _, vjp_fn = vjp(fn, *args) |
| result_vjps = vjp_fn(cotangents) |
| |
| # Compute ref_vjp of vjp. We could have done ref_vjp of ref_vjp, |
| # but since we're confident that vjp works by itself, this is |
| # an equivalent way to test that. |
| _, vjp_fn = ref_vjp(fn, *args) |
| expected_vjps = vjp_fn(cotangents) |
| |
| self.assertEqual(result_vjps, expected_vjps) |
| |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=(torch.float,)) |
| def test_vmapvjpvjp(self, device, dtype, op): |
| self.skipTest("Skipped; these tests take too long") |
| op_skip = set({ |
| }) |
| op_skip = op_skip.union(vjp_fail) |
| if op.name in op_skip: |
| self.skipTest("Skipped; Expected failures") |
| return |
| |
| if not op.supports_autograd: |
| self.skipTest("Skipped! Autograd not supported.") |
| return |
| if not op.supports_gradgrad: |
| self.skipTest("Skipped! Operation does not support gradgrad") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=True) |
| |
| # TODO: test in-place |
| if is_inplace(op, op.get_op()): |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| for sample in samples: |
| fn, args = get_vjpfull_variant(op, sample) |
| result = fn(*args) |
| cotangents = tree_map(lambda x: torch.randn_like(x), result) |
| cotangents, _ = tree_flatten(cotangents) |
| num_args = len(args) |
| |
| args_and_cotangents = tuple(args) + tuple(cotangents) |
| |
| def vjp_of_vjp(*args_and_cotangents): |
| args = args_and_cotangents[:num_args] |
| cotangents = args_and_cotangents[num_args:] |
| result, vjp_fn = vjp(fn, *args) |
| result_vjps = vjp_fn(cotangents) |
| result, _ = tree_flatten(result) |
| result_vjps, _ = tree_flatten(result_vjps) |
| return (*result, *result_vjps) |
| |
| for loop_out, batched_out in \ |
| get_fallback_and_vmap_exhaustive(vjp_of_vjp, args_and_cotangents, {}): |
| self.assertEqual(loop_out, batched_out, atol=1e-4, rtol=1e-4) |
| vmapvjp_fail = vjp_fail.union({ |
| # All of the following are bugs and need to be fixed |
| xfail('diag_embed'), |
| xfail('eig'), |
| xfail('view_as_complex'), |
| xfail('fft.ihfft'), |
| xfail('fft.ihfft'), |
| xfail('fft.rfft'), |
| xfail('fft.rfft'), |
| xfail('fft.rfftn'), |
| xfail('cdist'), |
| xfail('fmax'), |
| xfail('fmin'), |
| xfail('index_add'), |
| xfail('index_copy'), |
| xfail('index_fill'), |
| xfail('linalg.det', ''), |
| xfail('linalg.eigh'), |
| xfail('linalg.householder_product'), |
| xfail('linalg.matrix_norm'), |
| xfail('linalg.norm'), |
| xfail('linalg.slogdet'), |
| xfail('logdet'), |
| xfail('lu_unpack'), |
| xfail('masked_fill'), |
| xfail('masked_scatter'), |
| xfail('matrix_exp'), |
| xfail('nanquantile'), |
| xfail('norm', 'fro'), |
| xfail('norm', 'nuc'), |
| xfail('prod'), |
| xfail('put'), |
| xfail('quantile'), |
| xfail('symeig'), |
| xfail('take'), |
| xfail('linalg.tensorinv'), |
| xfail('nn.functional.conv_transpose2d', device_type='cuda'), |
| xfail('nanmean'), |
| xfail('block_diag'), |
| xfail('nn.functional.dropout'), |
| xfail('fft.ihfft2'), |
| xfail('fft.ihfftn'), |
| xfail('double', 'channels_last'), |
| xfail('nn.functional.gaussian_nll_loss'), |
| xfail('nn.functional.poisson_nll_loss'), |
| xfail('nn.functional.conv1d', device_type='cuda'), |
| xfail('fft.rfft2'), |
| xfail('lu'), |
| skip('qr'), # Nondetermistic |
| xfail('_masked.prod'), # calls aten::item |
| xfail('stft'), |
| xfail('nn.functional.glu'), |
| |
| xfail('nn.functional.fractional_max_pool3d'), |
| xfail('as_strided'), |
| xfail('nn.functional.fractional_max_pool2d'), |
| |
| # PyTorch changed its convolution recently. |
| # Maybe it is responsible for all of the following changes. |
| xfail('nn.functional.conv_transpose1d'), |
| xfail('nn.functional.conv_transpose2d'), |
| xfail('nn.functional.conv_transpose3d'), |
| |
| }) |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=(torch.float,)) |
| @skipOps('TestOperators', 'test_vmapvjp', vmapvjp_fail) |
| def test_vmapvjp(self, device, dtype, op): |
| # These are too annoying to put into the list above |
| if op.name in {'nn.functional.linear'}: |
| self.skipTest("Skipped! ExpectedF failures") |
| if not op.supports_autograd: |
| self.skipTest("Skipped! Autograd not supported.") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=True) |
| |
| # TODO: test in-place |
| if is_inplace(op, op.get_op()): |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| for sample in samples: |
| fn, args = get_vjpfull_variant(op, sample) |
| for loop_out, batched_out in get_fallback_and_vmap_exhaustive(fn, args, {}): |
| self.assertEqual(loop_out, batched_out, atol=1e-4, rtol=1e-4) |
| |
| # There are several variations we care about |
| # 1) primal batched (TODO) |
| # 2) tangent batched (batched grads) <-- |
| # 3) both batched (TODO) |
| # The below tests (2) only. |
| @ops(functorch_lagging_op_db, allowed_dtypes=(torch.float,)) |
| @skipOps('TestOperators', 'test_vmapjvp', { |
| xfail('nn.functional.dropout'), # randomness |
| |
| # TODO: fails in core due to in-place batched nto non-batched |
| # but fails here for a different reason |
| xfail('linalg.householder_product'), |
| |
| # Try to in-place batched tensor into non-batched tensor |
| xfail('matrix_exp'), |
| xfail('lu'), |
| xfail('fill_'), |
| xfail('block_diag'), # TODO: We expect this to fail in core, but it doesn't |
| xfail('index_add'), |
| xfail('index_copy'), |
| xfail('index_put'), |
| xfail('index_fill'), |
| xfail('masked_fill'), |
| xfail('masked_scatter'), |
| |
| # https://gist.github.com/zou3519/c42d032c0111c6b65235583d391bf7a3 |
| xfail('nn.functional.linear'), |
| |
| # These are issues that should be fixed in core. See repro in core: |
| # https://github.com/pytorch/functorch/pull/232#discussion_r751405155 |
| # RuntimeError: expected scalar type double but found float |
| xfail('minimum'), |
| xfail('min', 'binary'), |
| xfail('maximum'), |
| xfail('max', 'binary'), |
| |
| # Apprently these support forward AD, but we get "Trying to use forward AD..." |
| # These are cases where OpInfo has supports_forward_ad=True, but disables the test |
| xfail('var_mean'), |
| xfail('std_mean'), |
| xfail('linalg.eigvalsh'), |
| |
| # functorch doesn't support channels_last |
| # PyTorch core's vmap doesn't have a batching rule for `double`, if it |
| # did it would also not support channels last, so I'm including this |
| # xfail "above the line". |
| xfail('double', 'channels_last'), |
| |
| # See https://github.com/pytorch/pytorch/issues/66357 |
| xfail('nn.functional.pad', 'circular'), |
| |
| # RuntimeError: expand: the number of sizes provided (1) must be greater or equal to the number of dimensions in the tensor (2) |
| xfail('nanquantile'), |
| xfail('quantile'), |
| |
| # RuntimeError: vmap: inplace arithmetic(self, *extra_args) |
| xfail('nn.functional.gelu'), |
| |
| # Not implemented |
| xfail('scatter'), |
| |
| # ============================================= |
| # NB: The above failures also fail in PyTorch core. |
| # The failures below only fail in functorch |
| # ============================================= |
| |
| # Composite ops that do bad things. Need to be fixed in PyTorch core. |
| # RuntimeError: Cannot access data pointer of Tensor that doesn't have storage |
| xfail('tensor_split'), |
| xfail('linalg.inv'), |
| xfail('linalg.matrix_power'), |
| xfail('linalg.cholesky'), |
| |
| # Causing a CUDA assert, needs investigation |
| skip('div', 'floor_rounding', device_type='cuda'), |
| skip('div', 'no_rounding_mode', device_type='cuda'), |
| skip('div', 'trunc_rounding', device_type='cuda'), |
| skip('true_divide', device_type='cuda'), |
| }) |
| def test_vmapjvp(self, device, dtype, op): |
| if is_inplace(op, op.get_op()): |
| # TODO: test in-place |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=False) |
| |
| if not op.supports_forward_ad: |
| self.skipTest("Skipped! Forward AD not supported.") |
| return |
| |
| for sample in samples: |
| arg_values = [sample.input] + list(sample.args) |
| kwarg_values = sample.kwargs |
| args = tuple([*arg_values, *kwarg_values]) |
| fn, args = get_jvp_variant(op, sample) |
| for loop_out, batched_out in get_fallback_and_vmap_exhaustive(fn, args, {}, bdims=(0,)): |
| self.assertEqual(loop_out, batched_out, atol=1e-4, rtol=1e-4) |
| |
| |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=(torch.float,)) |
| @skipOps('TestOperators', 'test_vmapvjp_has_batch_rule', vmapvjp_fail.union({ |
| xfail('view_as_complex'), |
| xfail('__getitem__'), |
| xfail('cdist'), |
| xfail('cholesky'), |
| xfail('clamp', 'scalar'), |
| xfail('complex'), |
| xfail('copysign'), |
| xfail('corrcoef'), |
| xfail('cummax'), |
| xfail('cummin'), |
| xfail('cumprod'), |
| xfail('diag_embed'), |
| xfail('eig'), |
| xfail('fft.ihfft'), |
| xfail('fft.rfft'), |
| xfail('fft.rfftn'), |
| xfail('cdist'), |
| xfail('fill_'), |
| xfail('fmax'), |
| xfail('fmin'), |
| xfail('index_add'), |
| xfail('index_copy'), |
| xfail('index_fill'), |
| xfail('index_select'), |
| xfail('linalg.cholesky'), |
| xfail('linalg.cholesky_ex'), |
| xfail('linalg.det'), |
| xfail('linalg.eig'), |
| xfail('linalg.eigh'), |
| xfail('linalg.eigvals'), |
| xfail('linalg.householder_product'), |
| xfail('linalg.lstsq'), |
| xfail('linalg.inv'), |
| xfail('linalg.matrix_norm'), |
| xfail('linalg.matrix_power'), |
| xfail('linalg.norm'), |
| xfail('linalg.pinv'), |
| xfail('linalg.qr'), |
| xfail('linalg.pinv', 'hermitian'), |
| xfail('linalg.slogdet'), |
| xfail('linalg.solve'), |
| xfail('linalg.tensorinv'), |
| xfail('linalg.vector_norm'), |
| xfail('logdet'), |
| xfail('lu'), |
| xfail('lu_solve'), |
| xfail('lu_unpack'), |
| xfail('masked_fill'), |
| xfail('masked_scatter'), |
| xfail('masked_select'), |
| xfail('matrix_exp'), |
| xfail('max', 'reduction_no_dim'), |
| xfail('median'), |
| xfail('min', 'reduction_no_dim'), |
| xfail('nanmedian'), |
| xfail('nanquantile'), |
| xfail('nn.functional.conv_transpose2d'), |
| xfail('nn.functional.gelu'), |
| xfail('nn.functional.pad', 'circular'), |
| xfail('norm', 'fro'), |
| xfail('norm', 'inf'), |
| xfail('norm', 'nuc'), |
| xfail('pinverse'), |
| xfail('prod'), |
| xfail('put'), |
| xfail('quantile'), |
| xfail('renorm'), |
| xfail('repeat_interleave'), |
| xfail('solve'), |
| xfail('symeig'), |
| xfail('take'), |
| xfail('tensor_split'), |
| xfail('to_sparse'), |
| xfail('trace'), |
| xfail('unfold'), |
| xfail('vdot'), |
| xfail('nanmean'), |
| xfail('block_diag'), |
| xfail('nn.functional.dropout'), |
| xfail('nn.functional.batch_norm'), |
| xfail('_masked.prod'), |
| xfail('fft.ihfft2'), |
| xfail('fft.ihfftn'), |
| xfail('fft.rfft2'), |
| xfail('nn.functional.embedding'), |
| xfail('cross'), |
| xfail('double', 'channels_last'), |
| xfail('linalg.cross'), |
| xfail('nn.functional.conv1d'), |
| xfail('nn.functional.gaussian_nll_loss'), |
| xfail('nn.functional.hardsigmoid'), |
| xfail('nn.functional.huber_loss'), |
| xfail('nn.functional.instance_norm'), |
| xfail('nn.functional.poisson_nll_loss'), |
| xfail('nn.functional.conv_transpose3d'), |
| xfail('_masked.norm'), |
| xfail('_masked.normalize'), |
| xfail('nn.functional.bilinear'), |
| xfail('nn.functional.prelu'), |
| xfail('nn.functional.glu'), |
| xfail('nn.functional.fractional_max_pool3d'), |
| xfail('as_strided'), |
| xfail('linalg.solve_triangular'), |
| xfail('stft'), |
| xfail('nn.functional.rrelu'), |
| xfail('nn.functional.embedding_bag'), |
| xfail('nn.functional.softshrink'), |
| xfail('nn.functional.conv_transpose1d'), |
| xfail('nn.functional.max_pool3d'), |
| xfail('istft'), |
| xfail('nn.functional.fractional_max_pool2d'), |
| xfail('linalg.tensorsolve'), |
| })) |
| def test_vmapvjp_has_batch_rule(self, device, dtype, op): |
| # These are too annoying to put into the list above |
| if op.name in {'nn.functional.linear', 'nn.functional.conv2d'}: |
| self.skipTest("Skipped! ExpectedF failures") |
| if not op.supports_autograd: |
| self.skipTest("Skipped! Autograd not supported.") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=True) |
| |
| # TODO: test in-place |
| if is_inplace(op, op.get_op()): |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| def test(): |
| for sample in samples: |
| fn, args = get_vjpfull_variant(op, sample) |
| for _ in get_fallback_and_vmap_exhaustive(fn, args, {}, compute_loop_out=False): |
| pass |
| for a_op in op.aliases: |
| fn, args = get_vjpfull_variant(a_op, sample) |
| for _ in get_fallback_and_vmap_exhaustive(fn, args, {}, compute_loop_out=False): |
| pass |
| check_vmap_fallback(self, test, op, dry_run=False) |
| |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=(torch.float,)) |
| @skipOps('TestOperators', 'test_vjpvmap', vjp_fail.union({ |
| # fallback path doesn't work |
| xfail('H'), |
| # All of the following are bugs and need to be fixed |
| xfail('__getitem__'), |
| xfail('clamp', ''), |
| xfail('dsplit'), |
| xfail('fill_'), |
| xfail('gradient'), |
| xfail('hsplit'), |
| xfail('vsplit'), |
| xfail('dstack'), |
| xfail('hstack'), |
| xfail('index_put'), |
| xfail('linalg.multi_dot'), |
| xfail('vstack'), |
| xfail('nn.functional.batch_norm'), |
| xfail('cdist'), |
| xfail('lu_solve'), |
| xfail('lu_unpack'), |
| xfail('matrix_exp'), |
| xfail('view_as_complex'), |
| xfail('nn.functional.gaussian_nll_loss'), |
| xfail('double', 'channels_last'), |
| xfail('masked_select'), |
| xfail('nn.functional.fractional_max_pool3d'), |
| xfail('nn.functional.glu'), |
| xfail('as_strided'), |
| xfail('nn.functional.fractional_max_pool2d'), |
| })) |
| def test_vjpvmap(self, device, dtype, op): |
| # NB: there is no vjpvmap_has_batch_rule test because that is almost |
| # certainly redundant with the vmap_has_batch_rule test in test_vmap.py |
| |
| # one-off skip |
| if op.name == 'nn.functional.dropout': |
| self.skipTest("Skipped!") |
| |
| if not op.supports_autograd: |
| # If the op doesn't support autograd, vmap(op) won't either |
| self.skipTest("Skipped! Autograd not supported.") |
| return |
| |
| # TODO: test in-place |
| if is_inplace(op, op.get_op()): |
| self.skipTest("Skipped! NYI: inplace-testing not supported.") |
| return |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=True) |
| |
| for sample in samples: |
| args = [sample.input] + list(sample.args) |
| kwargs = sample.kwargs |
| |
| for batched_args, in_dims, kwargs in get_exhaustive_batched_inputs(args, kwargs): |
| vmapped_op = vmap(op, in_dims) |
| fn, primals = normalize_op_input_output2(vmapped_op, batched_args, kwargs, |
| sample.output_process_fn_grad) |
| result = fn(*primals) |
| cotangents = tree_map(lambda x: torch.randn_like(x), result) |
| |
| _, vjp_fn = vjp(fn, *primals) |
| result_vjps = vjp_fn(cotangents) |
| |
| _, vjp_fn = ref_vjp(fn, *primals) |
| expected_vjps = vjp_fn(cotangents) |
| |
| self.assertEqual(result_vjps, expected_vjps) |
| |
| class InplaceError(Exception): |
| def __repr__(self): |
| return "Decomposition Tensor with no elem was created (probably due to an in-place op)" |
| |
| |
| |
| def ref_vjp_no_create(f, *primals): |
| result = f(*primals) |
| |
| def wrapped(cotangents): |
| return _autograd_grad(_as_tuple(result), primals, _as_tuple(cotangents), create_graph=False) |
| |
| return result, wrapped |
| |
| run_decompositions = set() |
| run_ops = set() |
| class TestDecompositionOpInfo(TestCase): |
| |
| @unittest.skipIf(IS_FBCODE, "__torch_dispatch__ is buggy") |
| @ops(functorch_lagging_op_db + additional_op_db, allowed_dtypes=[torch.float32, torch.float64, torch.float16, torch.bfloat16] + [*integral_types()] ) |
| # entries in here need don't work and need to be fixed. |
| # Each one of these is a bug (or needs to be investigated) |
| @skipOps('TestDecompositionOpInfo', 'test_decomposition', { |
| skip('view_as_complex'), |
| xfail('linalg.cholesky'), |
| xfail('linalg.inv'), |
| xfail('linalg.matrix_power'), |
| xfail('to_sparse'), |
| skip('tensor_split'), |
| skip('mvlgamma'), |
| skip('tanh', device_type='cuda'), # cuda bfloat16 failure |
| skip('eig'), |
| skip('nn.functional.dropout'), |
| skip('_masked.softmin'), |
| skip('_masked.log_softmax'), |
| skip('stft'), |
| skip('_masked.softmax'), |
| skip('_masked.normalize'), |
| # Some weird matmul stuff with int64 matmuls |
| # inplace op |
| skip('resize_'), |
| }) |
| def test_decomposition(self, device, dtype, op): |
| # copied from common_utils.py |
| dtype_precisions = { |
| torch.float16 : (0.001, 1e-5), |
| torch.bfloat16 : (0.016, 1e-5), |
| torch.float32 : (1.3e-6, 1e-5), |
| torch.float64 : (1e-7, 1e-7), |
| torch.complex32 : (0.001, 1e-5), |
| torch.complex64 : (1.3e-6, 1e-5), |
| torch.complex128 : (1e-7, 1e-7), |
| } |
| # Returns the "default" rtol and atol for comparing scalars or |
| # tensors of the given dtypes. |
| def _getDefaultRtolAndAtol(dtype0, dtype1): |
| rtol = max(dtype_precisions.get(dtype0, (0, 0))[0], |
| dtype_precisions.get(dtype1, (0, 0))[0]) |
| atol = max(dtype_precisions.get(dtype0, (0, 0))[1], |
| dtype_precisions.get(dtype1, (0, 0))[1]) |
| return rtol, atol |
| def op_assert_equal(op, a, b): |
| assert a.dtype == b.dtype |
| # Some ops, like those involving reductions, are fundamentally non-decomposable with precision guarantees |
| tol_table = { |
| (torch.bfloat16, aten._softmax_backward_data): (0.016, 1e-2), # aggghhhhhhhhhh I hate reductions and floating point |
| (torch.bfloat16, aten._log_softmax_backward_data): (0.016, 1e-2), |
| # (torch.float16, aten.im2col_backward): (0.016, 1e-2), |
| } |
| msg = f"{op} decomposition failed" |
| if (b.dtype, op) in tol_table: |
| rtol, atol = tol_table[(b.dtype, op)] |
| else: |
| rtol, atol = _getDefaultRtolAndAtol(a.dtype, b.dtype) |
| assert torch.allclose(a, b, rtol=rtol, atol=atol), msg |
| |
| # We check the correctness of each decomposition right after running it. |
| # So, when we encounter a decomposition, we run the function normally, and then run the decomposition, and ensure they're identical. |
| # The way this is implemented, there could .... technically be an exponential blow up, but it's probably fine for now. |
| class DecompositionTensor(torch.Tensor): |
| elem: torch.Tensor |
| |
| __slots__ = ['elem'] |
| |
| @staticmethod |
| def __new__(cls, elem): |
| r = torch.Tensor._make_wrapper_subclass( |
| cls, elem.size(), |
| strides=elem.stride(), storage_offset=elem.storage_offset(), |
| dtype=elem.dtype, layout=elem.layout, |
| device=elem.device, requires_grad=elem.requires_grad |
| ) |
| |
| r.elem = elem |
| return r |
| |
| def __repr__(self): |
| return f"DecompositionTensor(elem={self.elem})" |
| |
| @classmethod |
| def __torch_dispatch__(cls, func, types, args=(), kwargs=None): |
| global run_ops |
| run_ops.add(func) |
| def unwrap_tensor(e): |
| if isinstance(e, DecompositionTensor): |
| if not hasattr(e, 'elem'): |
| raise InplaceError() |
| return e.elem |
| return e |
| |
| |
| real_out = func(*tree_map(unwrap_tensor, args), **tree_map(unwrap_tensor, kwargs)) |
| |
| |
| if func in decomposition_table and func != torch.ops.aten.detach: |
| upcast_table = set([ |
| aten.tanh_backward, |
| ]) |
| decomposition = decomposition_table[func] |
| global run_decompositions |
| run_decompositions.add(func) |
| input_dtypes = set() |
| |
| def upcast_tensor(x): |
| if isinstance(x, Tensor) and (x.dtype == torch.bfloat16 or x.dtype == torch.float16): |
| input_dtypes.add(x.dtype) |
| x = x.to(dtype=torch.float32) |
| return x |
| # Theoretically, most PyTorch ops compute intermediates as fp32. But this breaks some ops... |
| if func in upcast_table: |
| args = tree_map(upcast_tensor, args) |
| kwargs = tree_map(upcast_tensor, kwargs) |
| decomp_out = decomposition(*args, **kwargs) |
| real_out_flat = tree_flatten(real_out)[0] |
| decomp_out_flat = tree_flatten(decomp_out)[0] |
| assert(len(real_out_flat) == len(decomp_out_flat)) |
| assert(len(input_dtypes) <= 1) |
| input_dtypes = list(input_dtypes) |
| for a, b in zip(real_out_flat, decomp_out_flat): |
| if len(input_dtypes) > 0: |
| b = b.to(dtype=input_dtypes[0]) |
| op_assert_equal(func, a, b) |
| |
| def wrap_tensor(e): |
| if e is None: |
| return DecompositionTensor(torch.empty(())) |
| return DecompositionTensor(e) if type(e) == torch.Tensor else e |
| wrapped_out = tree_map(wrap_tensor, real_out) |
| return wrapped_out |
| |
| if dtype not in op.supported_dtypes(dtype): |
| self.skipTest("Dtype not in op's supported dtypes") |
| return |
| if is_inplace(op, op.get_op()): |
| self.skipTest("op is inplace") |
| return |
| _requires_grad = op.supports_autograd and dtype.is_floating_point |
| |
| samples = op.sample_inputs(device, dtype, requires_grad=_requires_grad) |
| |
| # Acquires variants to test |
| def wrap_tensor(x): |
| if type(x) == torch.Tensor: |
| return DecompositionTensor(x) |
| return x |
| |
| try: |
| func = op.get_op() |
| for sample_input in samples: |
| if _requires_grad: |
| fn, primals = normalize_op_input_output(func, sample_input) |
| primals = tree_map (lambda x: x.abs() if isinstance(x, torch.Tensor) else x, primals) |
| |
| decomp_out, decomp_vjp_fn = ref_vjp_no_create(fn, *tree_map(wrap_tensor, primals)) |
| cotangents = tree_map(lambda x: torch.randn_like(x), decomp_out) |
| |
| decomp_grads = decomp_vjp_fn(cotangents) |
| |
| else: |
| args = [sample_input.input] + list(sample_input.args) |
| kwargs = sample_input.kwargs |
| orig_out = func(*args, **kwargs) |
| |
| args = tree_map(wrap_tensor, args) |
| kwargs = tree_map(wrap_tensor, kwargs) |
| decomp_out = func(*args, **kwargs) |
| |
| |
| except InplaceError: |
| self.skipTest("op is inplace") |
| return |
| except RuntimeError as e: |
| if "not implemented for" in str(e): |
| self.skipTest(str(e)) |
| return |
| if "Mismatch in shape: grad_output" in str(e): |
| self.skipTest("Some weird issue with autograd engine and tensor subclasses") |
| return |
| raise e |
| |
| @unittest.skipIf(IS_FBCODE, "__torch_dispatch__ is buggy") |
| def test_placeholder(self): |
| global run_ops, run_decompositions |
| with open('op_analysis/run_ops.txt', 'w') as f: |
| def get_names(l): |
| return sorted([x.__name__ for x in l]) |
| for op in get_names(run_ops): |
| f.write(f'{op}\n') |
| with open('op_analysis/run_decompositions.txt', 'w') as f: |
| for op in get_names(run_decompositions): |
| f.write(f'{op}\n') |
| |
| only_for = ("cpu", "cuda") |
| instantiate_device_type_tests(TestOperators, globals(), only_for=only_for) |
| instantiate_device_type_tests(TestDecompositionOpInfo, globals(), only_for=only_for) |
| |
| if __name__ == '__main__': |
| run_tests() |
