wehub-resource-sync
133 lines
3.8 KiB
Python
133 lines
3.8 KiB
Python
# Copyright (c) 2026 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
|
|
from dygraph_to_static_utils import (
|
|
Dy2StTestBase,
|
|
)
|
|
|
|
import paddle
|
|
from paddle.distributed.fleet.utils import recompute
|
|
|
|
|
|
class ManualPyLayerRecompute(paddle.autograd.PyLayer):
|
|
@staticmethod
|
|
def forward(ctx, fn, x, y):
|
|
ctx.fn = fn
|
|
ctx.save_for_backward(x, y)
|
|
with paddle.no_grad():
|
|
out = fn(x, y)
|
|
return out
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_out):
|
|
x, y = ctx.saved_tensor()
|
|
with paddle.enable_grad():
|
|
x.stop_gradient = False
|
|
y.stop_gradient = False
|
|
out = ctx.fn(x, y)
|
|
grad_inputs = paddle.autograd.grad(out, [x, y], grad_out)
|
|
return (*grad_inputs,)
|
|
|
|
|
|
class TestManualPyLayerRecompute(Dy2StTestBase):
|
|
def test_recompute(self):
|
|
@paddle.jit.to_static()
|
|
def fn(x, y):
|
|
return x * y + paddle.sin(x)
|
|
|
|
x = paddle.randn([3, 3])
|
|
x.stop_gradient = False
|
|
y = paddle.randn([3, 3])
|
|
y.stop_gradient = False
|
|
|
|
out = ManualPyLayerRecompute.apply(fn, x, y)
|
|
out.backward()
|
|
grad_x1 = x.grad.numpy()
|
|
grad_y1 = y.grad.numpy()
|
|
|
|
x.clear_gradient()
|
|
y.clear_gradient()
|
|
|
|
out2 = fn(x, y)
|
|
out2.backward()
|
|
grad_x2 = x.grad.numpy()
|
|
grad_y2 = y.grad.numpy()
|
|
|
|
x.clear_gradient()
|
|
y.clear_gradient()
|
|
|
|
np.testing.assert_allclose(grad_x1, grad_x2, rtol=1e-05)
|
|
np.testing.assert_allclose(grad_y1, grad_y2, rtol=1e-05)
|
|
|
|
|
|
class SimpleRecomputeNetToStatic(paddle.nn.Layer):
|
|
def __init__(self, input_size=10):
|
|
super().__init__()
|
|
self.block = paddle.nn.Sequential(
|
|
paddle.nn.Linear(input_size, input_size, bias_attr=False),
|
|
paddle.nn.ReLU(),
|
|
)
|
|
self.block = paddle.jit.to_static(self.block)
|
|
|
|
def forward(self, inputs):
|
|
return recompute(self.block, inputs)
|
|
|
|
|
|
class SimpleRecomputeNet(paddle.nn.Layer):
|
|
def __init__(self, input_size=10):
|
|
super().__init__()
|
|
self.block = paddle.nn.Sequential(
|
|
paddle.nn.Linear(input_size, input_size, bias_attr=False),
|
|
paddle.nn.ReLU(),
|
|
)
|
|
|
|
def forward(self, inputs):
|
|
return self.block(inputs)
|
|
|
|
|
|
class TestDistributedPyLayerRecompute(Dy2StTestBase):
|
|
def test_recompute(self):
|
|
input_size = 10
|
|
inp = np.random.rand(2, input_size).astype("float32")
|
|
x = paddle.to_tensor(inp)
|
|
x.stop_gradient = False
|
|
y = paddle.to_tensor(inp)
|
|
y.stop_gradient = False
|
|
|
|
weight = np.random.rand(input_size, input_size).astype("float32")
|
|
model_x = SimpleRecomputeNet(input_size)
|
|
model_y = SimpleRecomputeNetToStatic(input_size)
|
|
|
|
with paddle.no_grad():
|
|
model_x.block[0].weight.set_value(weight)
|
|
model_y.block[0].weight.set_value(weight)
|
|
|
|
out_x = model_x(x)
|
|
loss_x = out_x.mean()
|
|
loss_x.backward()
|
|
|
|
out_y = model_y(y)
|
|
loss_y = out_y.mean()
|
|
loss_y.backward()
|
|
|
|
np.testing.assert_allclose(out_x.numpy(), out_y.numpy(), rtol=1e-05)
|
|
np.testing.assert_allclose(x.grad.numpy(), y.grad.numpy(), rtol=1e-05)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|