pytorch/test/lazy/test_ts_opinfo.py

# Owner(s): ["oncall: jit"]

import functools
import itertools
import os
from collections.abc import Sequence
from pathlib import Path
from unittest import skip

import yaml

import torch
import torch._lazy
import torch._lazy.config
import torch._lazy.ir_cache
import torch._lazy.metrics
import torch._lazy.ts_backend
from torch.testing._internal.common_device_type import (
    instantiate_device_type_tests,
    ops,
)
from torch.testing._internal.common_methods_invocations import op_db
from torch.testing._internal.common_utils import run_tests, TestCase
from torch.testing._internal.jit_utils import JitTestCase


torch._lazy.ts_backend.init()


def get_test_device():
    return "cuda" if "LTC_TS_CUDA" in os.environ else "cpu"


def remove_suffixes(l):
    return [x.split(".")[0] for x in l]


def init_lists():
    path_to_script = Path(os.path.abspath(os.path.dirname(__file__)))
    TS_NATIVE_FUNCTIONS_PATH = (
        path_to_script.parent.parent / "aten/src/ATen/native/ts_native_functions.yaml"
    )
    with open(TS_NATIVE_FUNCTIONS_PATH) as f:
        yaml_ts = yaml.load(f, yaml.SafeLoader)
    LAZY_OPS_LIST = set(
        remove_suffixes(
            itertools.chain(
                yaml_ts["full_codegen"], yaml_ts["supported"], yaml_ts["autograd"]
            )
        )
    )
    HAS_SYMINT_SUFFIX = yaml_ts["symint"]
    FALLBACK_LIST = {"clamp"}
    SKIP_RUNTIME_ERROR_LIST = {
        "index_select",  # Empty output_sizes is not supported
        "clone",  # is clone decomposed?
        # General ASAN Failure due to related to generating bool values.
        # https://github.com/pytorch/pytorch/issues/74519
        # https://github.com/pytorch/pytorch/issues/63034
        "nonzero",  # ASAN failure (paste: P501906539)
        "all",  # ASAN failure
        "any",  # ASAN failure
        "logdet",  # ASAN failure
    }
    SKIP_INCORRECT_RESULTS_LIST = {
        "squeeze",  # Value out of range
        "t",  # Value out of range
        "transpose",  # Value out of range
        "bernoulli",  # incorrect results
        "pow",  # incorrect results
        "addcdiv",  # incorrect results (on CI not locally?)
    }
    # The following ops all show up directly in ts_native_functions.yaml,
    # but run functionalized versions of the composite kernels in core.
    # This means that we don't expect the ops to show directly in the LTC metrics.
    FUNCTIONAL_DECOMPOSE_LIST = {
        "diag_embed",
        "block_diag",
        "new_empty_strided",
        "narrow_copy",
        "pixel_shuffle",
        "pixel_unshuffle",
        "select_backward",
        "_trilinear",
        "linalg_inv_ex",
        "linalg_pinv.atol_rtol_tensor",
        "logsumexp",
    }
    # For some ops, we don't support all variants. Here we use formatted_name
    # to uniquely identify the variant.
    SKIP_VARIANT_LIST = {"norm_nuc", "min_reduction_with_dim"}

    return (
        LAZY_OPS_LIST,
        FALLBACK_LIST,
        SKIP_RUNTIME_ERROR_LIST,
        SKIP_INCORRECT_RESULTS_LIST,
        FUNCTIONAL_DECOMPOSE_LIST,
        HAS_SYMINT_SUFFIX,
        SKIP_VARIANT_LIST,
    )


(
    LAZY_OPS_LIST,
    FALLBACK_LIST,
    SKIP_RUNTIME_ERROR_LIST,
    SKIP_INCORRECT_RESULTS_LIST,
    FUNCTIONAL_DECOMPOSE_LIST,
    HAS_SYMINT_SUFFIX,
    SKIP_VARIANT_LIST,
) = init_lists()

torch.manual_seed(42)


def clone_move(t):
    dev = "lazy"
    copy_t = t.detach().clone().requires_grad_(True).to(device=dev)
    return copy_t


class TestLazyTensor(JitTestCase):
    @skip("Disable until autograd supports symints")
    def testConvolutionBackward(self):
        test_device = get_test_device()
        inp = torch.rand(1, 3, 128, 128, device=test_device, requires_grad=True)
        inp_copy = clone_move(inp)
        grad = torch.rand(1, 32, 121, 121, device=test_device)  # no requires_grad
        grad_copy = clone_move(grad)
        weight = torch.rand(32, 3, 8, 8, device=test_device, requires_grad=True)
        weight_copy = clone_move(weight)
        bias = torch.rand(32, device=test_device, requires_grad=True)
        bias_copy = clone_move(bias)

        # run eager
        conv_out = torch.nn.functional.conv2d(inp, weight, bias)
        (inp_grad, weight_grad, bias_grad) = torch.autograd.grad(
            [conv_out], [inp, weight, bias], [grad]
        )

        # run lazy
        conv_copy_out = torch.nn.functional.conv2d(inp_copy, weight_copy, bias_copy)
        (inp_copy_grad, weight_copy_grad, bias_copy_grad) = torch.autograd.grad(
            [conv_copy_out], [inp_copy, weight_copy, bias_copy], [grad_copy]
        )

        # check numerics
        torch.testing.assert_close(bias_copy_grad.cpu(), bias_grad.cpu())

        torch.testing.assert_close(weight_copy_grad.cpu(), weight_grad.cpu())
        torch.testing.assert_close(inp_copy_grad.cpu(), inp_grad.cpu())

    def test_view_mark_step_preserved(self):
        test_device = get_test_device()
        inp = torch.rand(4, device=test_device)
        inp_lazy = clone_move(inp)

        def foo(x, *, mark_step):
            y = x.view(2, 2)
            y.add_(1)
            z = x + x  # noqa: F841

            if mark_step:
                torch._lazy.mark_step()

            # y and x should contiue to be aliased after the mark_step call.
            y.add_(1)
            return x

        out_ref = foo(inp, mark_step=False)
        out = foo(inp_lazy, mark_step=True)
        # out will have some pending mutations, which will be synced by the .cpu() call.
        torch.testing.assert_close(out_ref.cpu(), out.cpu())

    def test_tensor_ctr(self):
        test_device = get_test_device()
        inp = torch.tensor([[1, 2, 3, 4, 5]], device=test_device)
        inp_lazy = torch.tensor([[1, 2, 3, 4, 5]], device="lazy")

        def foo(x):
            # Calling a view op to ensure that functionalization wrapping occurs.
            return x.view(-1)

        out_ref = foo(inp)
        out = foo(inp_lazy)
        torch.testing.assert_close(out_ref.cpu(), out.cpu())


class TestLazyOpInfo(TestCase):
    @ops(
        [
            op
            for op in op_db
            if op.name in LAZY_OPS_LIST
            and op.name not in SKIP_RUNTIME_ERROR_LIST
            and op.name not in FUNCTIONAL_DECOMPOSE_LIST
            and op.formatted_name not in SKIP_VARIANT_LIST
        ],
        allowed_dtypes=(torch.float,),
    )
    def test_dispatched_to_lazy(self, device, dtype, op):
        def get_name(op):  # noqa: F841
            l = [op.name]
            if op.variant_test_name != "":
                l.append(op.variant_test_name)
            return ".".join(l)

        global HAS_SYMINT_SUFFIX, FALLBACK_LIST
        samples = op.sample_inputs("lazy", dtype, requires_grad=False)
        sample = next(iter(samples))
        args = [sample.input] + list(sample.args)
        kwargs = sample.kwargs
        torch._lazy.mark_step()
        torch._lazy.wait_device_ops()
        torch._lazy.metrics.reset()

        op(*args, **kwargs)
        torch._lazy.mark_step()
        torch._lazy.wait_device_ops()
        prefix = "aten" if op.name in FALLBACK_LIST else "lazy"
        symint_suffix = "_symint" if op.name in HAS_SYMINT_SUFFIX else ""
        found = f"{prefix}::{op.name}{symint_suffix}" in remove_suffixes(
            torch._lazy.metrics.counter_names()
        )
        # check aliases
        if not found:
            for alias in op.aliases:
                alias_found = (
                    f"{prefix}::{alias.name}{symint_suffix}"
                    in remove_suffixes(torch._lazy.metrics.counter_names())
                )
                found = found or alias_found
                if found:
                    break
        self.assertTrue(found)

    @ops(
        [
            op
            for op in op_db
            if op.name in LAZY_OPS_LIST
            and op.name not in SKIP_RUNTIME_ERROR_LIST | SKIP_INCORRECT_RESULTS_LIST
        ],
        allowed_dtypes=(torch.float,),
    )  # noqa: B950
    def test_correctness(self, device, dtype, op):
        test_device = get_test_device()

        def clone_to_device(input, dev):
            if isinstance(input, torch.Tensor):
                return input.detach().clone().to(device=dev)
            if isinstance(input, Sequence) and not isinstance(input, str):
                return tuple(map(functools.partial(clone_to_device, dev=dev), input))
            return input

        def assert_allclose_rec(t):
            a, b = t
            self.assertEqual(type(a), type(b))
            if isinstance(a, torch.Tensor):
                self.assertTrue(
                    torch.allclose(clone_to_device(a, test_device), b, atol=1e-4)
                )

            if isinstance(a, Sequence):
                map(assert_allclose_rec, zip(a, b))

        samples = op.sample_inputs("lazy", dtype, requires_grad=False)
        for sample in samples:
            # Need to run mark step so that all random ops are computed in the right order
            torch._lazy.mark_step()

            args = [sample.input] + list(sample.args)
            kwargs = sample.kwargs
            copy_args = clone_to_device(args, test_device)

            r_exp = op(*copy_args, **kwargs)
            r_actual = op(*args, **kwargs)

            torch._lazy.mark_step()
            assert_allclose_rec((r_actual, r_exp))

    @ops(
        [
            op
            for op in op_db
            if op.name in LAZY_OPS_LIST
            and op.name not in SKIP_RUNTIME_ERROR_LIST | SKIP_INCORRECT_RESULTS_LIST
        ],
        allowed_dtypes=(torch.float,),
    )  # noqa: B950
    def test_correctness_with_reusing_ir(self, device, dtype, op):
        torch._lazy.config.set_reuse_ir(True)
        test_device = get_test_device()

        def clone_to_device(input, dev):
            if isinstance(input, torch.Tensor):
                return input.detach().clone().to(device=dev)
            if isinstance(input, Sequence) and not isinstance(input, str):
                return tuple(map(functools.partial(clone_to_device, dev=dev), input))
            return input

        def assert_allclose_rec(t):
            a, b = t
            self.assertEqual(type(a), type(b))
            if isinstance(a, torch.Tensor):
                self.assertTrue(
                    torch.allclose(clone_to_device(a, test_device), b, atol=1e-4)
                )

            if isinstance(a, Sequence):
                map(assert_allclose_rec, zip(a, b))

        samples = op.sample_inputs("lazy", dtype, requires_grad=False)
        for sample in samples:
            # Need to run mark step so that all random ops are computed in the right order
            torch._lazy.mark_step()

            args = [sample.input] + list(sample.args)
            kwargs = sample.kwargs
            copy_args = clone_to_device(args, test_device)

            r_exp = op(*copy_args, **kwargs)
            r_actual = op(*args, **kwargs)

            torch._lazy.mark_step()
            assert_allclose_rec((r_actual, r_exp))

        torch._lazy.ir_cache.reset()
        torch._lazy.config.set_reuse_ir(False)


# TODO: after we move to master, add Lazy as a new Device here:
# https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/common_device_type.py#L532
instantiate_device_type_tests(TestLazyOpInfo, globals(), only_for="cpu")


class TestLazyDynamicOps(TestCase):
    @classmethod
    def setUpClass(cls) -> None:
        # Setup the dynamic shape mode
        cls.old_ssa_mode = torch._C._lazy._get_symbolic_shape_mode()
        torch._C._lazy._set_symbolic_shape_mode(True)
        return super().setUpClass()

    @classmethod
    def tearDownClass(cls) -> None:
        torch._C._lazy._set_symbolic_shape_mode(cls.old_ssa_mode)
        return super().tearDownClass()

    def test_nonzero_dynamic(self):
        # Test that nonzero gives upper bounds sizes when symbolic shape mode is enabled
        test_device = get_test_device()
        x1 = torch.tensor(
            [[0, 1.0, 2.0], [3.0, 0, 0]], device=test_device, requires_grad=True
        )
        x1_lazy = clone_move(x1)
        x2_lazy = torch.nonzero(x1_lazy)

        # FIXME: Add bindings to get upper bounds
        # self.assertEqual(tuple(x2_lazy.size()), (6, 2))

        # We should still be able to instantiate it and get the actual result
        x2_eager = x2_lazy.cpu()
        self.assertEqual(tuple(x2_eager.size()), (3, 2))

    def test_adaptiveavgpool3d_dynamic(self):
        # Test that adaptive_avg_pool3d gives correct shapes with lazy backend
        img_cpu = torch.zeros([2, 3, 4, 5, 6], device="cpu")
        out_cpu = torch.nn.AdaptiveAvgPool3d(2).to(device="cpu")(img_cpu)

        test_device = get_test_device()
        img_lazy = torch.zeros([2, 3, 4, 5, 6], device=test_device)
        out_lazy = torch.nn.AdaptiveAvgPool3d(2).to(test_device)(img_lazy)

        self.assertEqual(out_cpu.shape, out_lazy.shape)


if __name__ == "__main__":
    run_tests()