test/lazy/test_extract_compiled_graph.py

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

import unittest

from torch._lazy.ts_backend import init as init_ts_backend
init_ts_backend()
from torch._lazy import config
from torch._lazy.extract_compiled_graph import extract_compiled_graph
import torch
from torch import nn
import dis
import inspect
from torch import fx
import re
from contextlib import contextmanager
import copy

class ModuleConstScale(nn.Module):
    def forward(self, a):
        return a * 2

class ModuleSub(nn.Module):
    def forward(self, a, b):
        return a - b

class ModuleAddcmul(nn.Module):
    """
    addcmul function takes a at::Scalar which results in a special TSData containing a Scalar rather than a Tensor.
    """
    def forward(self, a, b, c):
        return torch.addcmul(a, b, c, value=5)

class ModuleReturnMulti(nn.Module):
    def forward(self, a, b):
        return (b + 1, a - 1)

# The default fx tracer will convert torch.randn to a constant.. We may need
# a custom tracer.
# class ModuleEagerTensor(nn.Module):
#     def __init__(self):
#         super().__init__()
#
#     def forward(self, a):
#         b = torch.randn(2, 3, device="cpu") # eager device
#         return a + b

# The module was planned to cover the case that a Fx graph return an eager
# tensor on the default device. It's harder than ModuleEagerTensor because
# we can not just override the device argument to Lazy since there is no
# explicit device argument.
#
# Unfortunately, the default fx tracer convert the return value of the forward
# method to a constant.. Comment out for now
# class ModuleReturnEagerTensorOnDefaultDevice(nn.Module):
#     def __init__(self):
#         super().__init__()
#
#     def forward(self):
#         return torch.tensor((2, 3), dtype=torch.float32)

class ModuleReturnDupTensor(nn.Module):
    """
    Handle the corner case that the same tensor appears multiple times in the
    returned tuple. torchbench like drq will hit this corner case when running
    thru torchdynamo..
    """
    def forward(self, a, b):
        c = a + b
        return a - b, c, a + 1, c

class ModuleInplaceUpdate(nn.Module):
    def forward(self, a, b):
        a.sub_(b)
        return b - 1, b + 1

@contextmanager
def force_fallback_ctx_mgr(fallback_op):
    oldconfig = config.get_force_fallback()
    config.set_force_fallback(fallback_op)
    try:
        yield None
    finally:
        config.set_force_fallback(oldconfig)

@contextmanager
def nop_ctx_mgr():
    try:
        yield None
    finally:
        pass

def gen_rand_args(mod):
    args = []
    for _ in range(len(inspect.signature(mod.forward).parameters)):
        args.append(torch.randn(2, 3))
    return args

def allclose(expected, actual):
    def unwrap(cont):
        if isinstance(cont, (list, tuple)) and len(cont) == 1:
            return cont[0]
        return cont
    expected = unwrap(expected)
    actual = unwrap(actual)

    if isinstance(expected, torch.Tensor) and isinstance(actual, torch.Tensor):
        return torch.allclose(expected, actual)
    elif isinstance(expected, (tuple, list)) and isinstance(actual, (tuple, list)):
        return len(expected) == len(actual) and all(torch.allclose(a, b) for a, b in zip(expected, actual))
    else:
        raise RuntimeError("Unexpected types")

def verify_reusing_compiled_graph(mod, exception_msg_pattern, ncase=10):
    args = gen_rand_args(mod)
    out = mod(*args)

    dis.dis(mod.forward)

    try:
        optimized_mod = extract_compiled_graph(fx.symbolic_trace(mod), args)
    except RuntimeError as e:
        if exception_msg_pattern is None:
            raise e  # reraise the exception
        exception_message = str(e)
        if not re.search(exception_msg_pattern, exception_message):
            raise RuntimeError(f"Exception message does not match the required pattern: {exception_message}") from e
        else:
            # We are done for the test case that expects an exception
            return

    if exception_msg_pattern is not None:
        raise RuntimeError(f"Expect an exception matching pattern {exception_msg_pattern}")
    print("return value of optimized_mod", optimized_mod(*args))

    # check correctness
    failed_index = []
    for i in range(ncase):
        rand_args = gen_rand_args(mod)
        rand_args_copy = copy.deepcopy(rand_args)
        expected = mod(*rand_args)
        actual = optimized_mod(*rand_args_copy)

        if not allclose(expected, actual):
            print(f"Incorrect results. expected {expected}, actual {actual}")
            failed_index.append(i)
            continue

        # make sure arguments match after calling the model forward method to handle inplace
        # updates.
        if not allclose(rand_args, rand_args_copy):
            print(f"Incorrect updated arguments. expected {rand_args}, actual {rand_args_copy}")
            failed_index.append(i)
            continue

    if len(failed_index) > 0:
        raise RuntimeError(f"Failed {len(failed_index)}/{ncase} cases")

def maketest(module_cls, exception_msg_pattern=None, ctxmgr=None):
    def wrapper(self):
        nonlocal ctxmgr
        if not ctxmgr:
            ctxmgr = nop_ctx_mgr()
        with ctxmgr:
            verify_reusing_compiled_graph(module_cls(), exception_msg_pattern)

    return wrapper

class OptimizeTest(unittest.TestCase):
    test_sub = maketest(ModuleSub)
    # Same as test_sub but force aten::sub to fallback
    # We expect an exception caught because of LTC fallabck.
    test_ltc_fallback = maketest(ModuleSub, exception_msg_pattern="fallback.*aten::sub", ctxmgr=force_fallback_ctx_mgr("aten::sub"))
    test_const_scale = maketest(ModuleConstScale)
    test_addcmul = maketest(ModuleAddcmul)
    test_return_multi = maketest(ModuleReturnMulti)
    test_return_dup_tensor = maketest(ModuleReturnDupTensor)
    test_inplace_update = maketest(ModuleInplaceUpdate)