348 lines
11 KiB
Python
348 lines
11 KiB
Python
# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
import unittest
|
|
|
|
import numpy as np
|
|
|
|
import paddle
|
|
from paddle.distributed.fleet.recompute.recompute_hybrid import recompute_hybrid
|
|
from paddle.distributed.fleet.utils import recompute
|
|
|
|
|
|
class Layer(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.linear1 = paddle.nn.Linear(10, 10)
|
|
self.linear2 = paddle.nn.Linear(10, 10)
|
|
self.linear3 = paddle.nn.Linear(10, 10)
|
|
self.silu1 = paddle.nn.Silu()
|
|
self.silu2 = paddle.nn.Silu()
|
|
self.silu3 = paddle.nn.Silu()
|
|
|
|
def forward(self, x, y):
|
|
assert type(x) is tuple
|
|
assert len(x) == 2
|
|
o1 = self.silu1(self.linear1(x[0]))
|
|
o2 = self.silu2(self.linear2(x[1]))
|
|
o3 = self.silu3(self.linear3(y))
|
|
o = o1 + o2 + o3
|
|
return o
|
|
|
|
|
|
class TestPyLayer(unittest.TestCase):
|
|
def test_tuple_input(self):
|
|
layer = Layer()
|
|
x1 = paddle.rand(shape=[10, 10])
|
|
x1.stop_gradient = False
|
|
x2 = paddle.rand(shape=[10, 10])
|
|
x2.stop_gradient = False
|
|
y = paddle.rand(shape=[10, 10])
|
|
y.stop_gradient = False
|
|
o = recompute(layer, (x1, x2), y)
|
|
loss = paddle.mean(o, keepdim=True)
|
|
loss.backward()
|
|
|
|
def test_tuple_input_with_non_tensor(self):
|
|
layer = Layer()
|
|
x1 = paddle.rand(shape=[10, 10])
|
|
x1.stop_gradient = False
|
|
y = paddle.rand(shape=[10, 10])
|
|
y.stop_gradient = False
|
|
try:
|
|
o = recompute(layer, (x1, True), y)
|
|
except ValueError:
|
|
pass
|
|
|
|
def test_tuple_input_with_different_stop_gradient(self):
|
|
layer = Layer()
|
|
x1 = paddle.rand(shape=[10, 10])
|
|
x1.stop_gradient = False
|
|
x2 = paddle.rand(shape=[10, 10])
|
|
y = paddle.rand(shape=[10, 10])
|
|
y.stop_gradient = False
|
|
try:
|
|
o = recompute(layer, (x1, True), y)
|
|
except ValueError:
|
|
pass
|
|
|
|
def test_tuple_input_all_no_gradient(self):
|
|
layer = Layer()
|
|
x1 = paddle.rand(shape=[10, 10])
|
|
x2 = paddle.rand(shape=[10, 10])
|
|
y = paddle.rand(shape=[10, 10])
|
|
y.stop_gradient = False
|
|
o = recompute(layer, (x1, x2), y)
|
|
loss = paddle.mean(o, keepdim=True)
|
|
loss.backward()
|
|
|
|
|
|
class _MockMpGroup:
|
|
"""Minimal mock of a model-parallel group for single-GPU unit tests.
|
|
Only used when partition=False, so nranks/rank are never actually accessed.
|
|
"""
|
|
|
|
nranks = 1
|
|
rank = 0
|
|
|
|
|
|
class TestRecomputeHybridProtectTensors(unittest.TestCase):
|
|
"""Tests for recompute_hybrid() with various tensor inputs.
|
|
Uses a MockMpGroup so no distributed init is needed,
|
|
and partition=False so mp_group is never actually invoked.
|
|
"""
|
|
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
if not paddle.is_compiled_with_cuda():
|
|
raise unittest.SkipTest("Requires GPU")
|
|
|
|
def test_forward_backward_plain_tensor(self):
|
|
"""recompute_hybrid with a plain tensor input completes correctly."""
|
|
linear = paddle.nn.Linear(8, 8)
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
ctx = {'mp_group': _MockMpGroup(), 'offload': False, 'partition': False}
|
|
out = recompute_hybrid(ctx, linear, x)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
def test_forward_backward_multiple_tensors(self):
|
|
"""recompute_hybrid with multiple tensor inputs completes correctly."""
|
|
|
|
class TwoInputLayer(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.linear = paddle.nn.Linear(8, 8)
|
|
|
|
def forward(self, x, y):
|
|
return self.linear(x) + self.linear(y)
|
|
|
|
layer = TwoInputLayer()
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
y = paddle.rand([4, 8])
|
|
y.stop_gradient = False
|
|
ctx = {'mp_group': _MockMpGroup(), 'offload': False, 'partition': False}
|
|
out = recompute_hybrid(ctx, layer, x, y)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
self.assertIsNotNone(y.grad)
|
|
|
|
|
|
class TestRecomputeWithClosureTensors(unittest.TestCase):
|
|
"""End-to-end tests for recompute() with closure-captured tensors."""
|
|
|
|
# ------------------------------------------------------------------
|
|
# basic: plain function with a closure tensor, no release
|
|
# ------------------------------------------------------------------
|
|
|
|
def test_plain_function_with_closure_tensor(self):
|
|
"""recompute() on a plain function that captures a tensor in its
|
|
closure must complete forward and backward correctly."""
|
|
grid = paddle.rand([4, 8])
|
|
|
|
def fn(x):
|
|
return x * grid
|
|
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
out = recompute(fn, x)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
# ------------------------------------------------------------------
|
|
# basic: no closure at all
|
|
# ------------------------------------------------------------------
|
|
|
|
def test_function_without_closure(self):
|
|
"""recompute() on a function that has no closure must not raise."""
|
|
|
|
def simple_fn(x):
|
|
return x * 2.0
|
|
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
out = recompute(simple_fn, x)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
# ------------------------------------------------------------------
|
|
# basic: closure with only non-tensor values
|
|
# ------------------------------------------------------------------
|
|
|
|
def test_function_with_non_tensor_closure(self):
|
|
"""Closure holding only non-tensor values must be handled safely."""
|
|
scale = 3.0
|
|
|
|
def fn(x):
|
|
return x * scale
|
|
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
out = recompute(fn, x)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
# ------------------------------------------------------------------
|
|
# basic: nn.Layer (uses run_function.forward.__closure__)
|
|
# ------------------------------------------------------------------
|
|
|
|
def test_layer_with_closure_in_forward(self):
|
|
"""recompute() on an nn.Layer that captures a tensor in forward's
|
|
closure must complete forward and backward correctly."""
|
|
|
|
class ClosureLayer(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.linear = paddle.nn.Linear(8, 8)
|
|
mask = paddle.ones([4, 8])
|
|
|
|
# Define forward as a closure over `mask`
|
|
def _forward_impl(x):
|
|
return self.linear(x) * mask
|
|
|
|
self._forward_impl = _forward_impl
|
|
|
|
def forward(self, x):
|
|
return self._forward_impl(x)
|
|
|
|
layer = ClosureLayer()
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
out = recompute(layer, x)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
# ------------------------------------------------------------------
|
|
# gradient correctness with closure tensor
|
|
# ------------------------------------------------------------------
|
|
|
|
def test_gradient_correctness_with_closure_tensor(self):
|
|
"""Gradients computed via recompute (with closure tensor) must match
|
|
those computed without recompute."""
|
|
paddle.seed(42)
|
|
grid = paddle.rand([4, 8])
|
|
linear = paddle.nn.Linear(8, 8)
|
|
|
|
def fn(x):
|
|
return linear(x) + grid
|
|
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
|
|
# reference: no recompute
|
|
out_ref = fn(x)
|
|
loss_ref = out_ref.mean()
|
|
loss_ref.backward()
|
|
grad_ref = x.grad.numpy().copy()
|
|
x.clear_gradient()
|
|
|
|
# with recompute
|
|
out_rc = recompute(fn, x)
|
|
loss_rc = out_rc.mean()
|
|
loss_rc.backward()
|
|
np.testing.assert_allclose(x.grad.numpy(), grad_ref, rtol=1e-5)
|
|
|
|
def test_non_reentrant_with_closure_tensor(self):
|
|
"""use_reentrant=False path with a closure-captured tensor must
|
|
complete forward and backward correctly."""
|
|
grid = paddle.rand([4, 8])
|
|
|
|
def fn(x):
|
|
return x * grid
|
|
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
out = recompute(fn, x, use_reentrant=False)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
def test_closure_tensor_preserve_rng_state_false(self):
|
|
"""recompute() with preserve_rng_state=False and a closure tensor must
|
|
execute the else-branch in backward (L441), completing forward and
|
|
backward correctly."""
|
|
grid = paddle.rand([4, 8])
|
|
|
|
def fn(x):
|
|
return x * grid
|
|
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
out = recompute(fn, x, preserve_rng_state=False)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
|
|
class TestRecomputeRegression(unittest.TestCase):
|
|
"""Regression tests for recompute() basic scenarios.
|
|
|
|
These cover the plain-tensor, tuple-only-input, and keyword-argument
|
|
code paths and serve as a safety net against regressions regardless of
|
|
the internal protection mechanism used.
|
|
"""
|
|
|
|
def test_plain_tensor(self):
|
|
"""recompute() with a plain tensor arg completes forward/backward."""
|
|
linear = paddle.nn.Linear(8, 8)
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
out = recompute(linear, x)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
def test_tuple_only_input(self):
|
|
"""recompute() with a tuple-only tensor arg completes forward/backward."""
|
|
|
|
class TupleInputLayer(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.linear = paddle.nn.Linear(8, 8)
|
|
|
|
def forward(self, xy):
|
|
x, y = xy
|
|
return self.linear(x) + self.linear(y)
|
|
|
|
layer = TupleInputLayer()
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
y = paddle.rand([4, 8])
|
|
y.stop_gradient = False
|
|
out = recompute(layer, (x, y))
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
self.assertIsNotNone(y.grad)
|
|
|
|
def test_with_kwargs(self):
|
|
"""recompute() with extra keyword arguments completes forward/backward."""
|
|
|
|
class ScaledLinear(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.linear = paddle.nn.Linear(8, 8)
|
|
|
|
def forward(self, x, scale=1.0):
|
|
return self.linear(x) * scale
|
|
|
|
layer = ScaledLinear()
|
|
x = paddle.rand([4, 8])
|
|
x.stop_gradient = False
|
|
out = recompute(layer, x, scale=2.0)
|
|
out.mean().backward()
|
|
self.assertIsNotNone(x.grad)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|