wehub-resource-sync
202 lines
6.6 KiB
Python
202 lines
6.6 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 os
|
|
|
|
import numpy as np
|
|
from semi_auto_parallel_simple_net import (
|
|
DemoNet,
|
|
TestSimpleNetForSemiAutoParallel,
|
|
)
|
|
|
|
import paddle
|
|
import paddle.distributed as dist
|
|
from paddle import nn
|
|
|
|
|
|
class TestSimpleNetWithAmpForSemiAutoParallel(TestSimpleNetForSemiAutoParallel):
|
|
def __init__(self):
|
|
self._dtype = os.getenv("dtype")
|
|
self._backend = os.getenv("backend")
|
|
self._seed = eval(os.getenv("seed"))
|
|
self._use_adam = eval(os.getenv("use_adam"))
|
|
self._use_master_grad = bool(eval(os.getenv("use_master_grad")))
|
|
self._mesh = dist.ProcessMesh([0, 1], dim_names=["x"])
|
|
|
|
paddle.set_device(self._backend)
|
|
self.init_single_card_net_result()
|
|
|
|
def check_tensor_eq(self, tensor_a, tensor_b):
|
|
super().check_tensor_eq(tensor_a, tensor_b, rtol=1e-5, atol=1e-7)
|
|
|
|
def run_dynamic_amp(
|
|
self, layer, level='O1', shard_input=False, run_dist=False
|
|
):
|
|
# create loss
|
|
loss_fn = nn.MSELoss()
|
|
if self._use_adam:
|
|
opt = paddle.optimizer.Adam(
|
|
learning_rate=0.001, parameters=layer.parameters()
|
|
)
|
|
else:
|
|
opt = paddle.optimizer.AdamW(
|
|
learning_rate=0.001, parameters=layer.parameters()
|
|
)
|
|
|
|
if level == 'O2':
|
|
layer, opt = paddle.amp.decorate(
|
|
models=layer,
|
|
level='O2',
|
|
master_grad=self._use_master_grad,
|
|
optimizers=opt,
|
|
dtype=self._dtype,
|
|
)
|
|
|
|
scaler = paddle.amp.GradScaler(init_loss_scaling=1024)
|
|
if run_dist:
|
|
scaler = dist.shard_scaler(scaler)
|
|
# run forward and backward
|
|
for _ in range(2):
|
|
image, label = self.init_input_data()
|
|
if shard_input:
|
|
image = dist.shard_tensor(image, self._mesh, [dist.Shard(0)])
|
|
|
|
with paddle.amp.auto_cast(level=level, dtype=self._dtype):
|
|
out = layer(image)
|
|
loss = loss_fn(out, label)
|
|
|
|
scaled = scaler.scale(loss)
|
|
scaled.backward()
|
|
scaler.step(opt)
|
|
scaler.update()
|
|
opt.clear_grad()
|
|
return loss, layer.parameters()
|
|
|
|
def init_single_card_net_result(self):
|
|
self.set_random_seed(self._seed)
|
|
|
|
(
|
|
self.base_loss_o1,
|
|
self.base_parameters_o1,
|
|
) = self.run_dynamic_amp(DemoNet('demo_weight_O1'), 'O1')
|
|
|
|
self.set_random_seed(self._seed)
|
|
(
|
|
self.base_loss_o2,
|
|
self.base_parameters_o2,
|
|
) = self.run_dynamic_amp(DemoNet('demo_weight_O2'), 'O2')
|
|
|
|
def test_dp_demo_net(self):
|
|
tol = 0.005
|
|
self.set_random_seed(self._seed)
|
|
(
|
|
self.dp_loss_o1,
|
|
self.dp_parameters_o1,
|
|
) = self.run_dynamic_amp(
|
|
DemoNet('dp_demo_weight_O1'), 'O1', shard_input=True, run_dist=True
|
|
)
|
|
np.testing.assert_allclose(
|
|
self.dp_loss_o1.numpy(),
|
|
self.base_loss_o1.numpy(),
|
|
rtol=tol,
|
|
atol=tol,
|
|
)
|
|
for param, param_base in zip(
|
|
self.dp_parameters_o1, self.base_parameters_o1
|
|
):
|
|
np.testing.assert_allclose(
|
|
param.numpy(), param_base.numpy(), rtol=tol, atol=tol
|
|
)
|
|
np.testing.assert_allclose(
|
|
param.grad.numpy(), param_base.grad.numpy(), rtol=tol, atol=tol
|
|
)
|
|
|
|
self.set_random_seed(self._seed)
|
|
(
|
|
self.dp_loss_o2,
|
|
self.dp_parameters_o2,
|
|
) = self.run_dynamic_amp(DemoNet('dp_demo_weight_O2'), 'O2')
|
|
np.testing.assert_allclose(
|
|
self.dp_loss_o2.numpy(),
|
|
self.base_loss_o2.numpy(),
|
|
rtol=tol,
|
|
atol=tol,
|
|
)
|
|
for param, param_base in zip(
|
|
self.dp_parameters_o2, self.base_parameters_o2
|
|
):
|
|
np.testing.assert_allclose(
|
|
param.numpy(), param_base.numpy(), rtol=tol, atol=tol
|
|
)
|
|
np.testing.assert_allclose(
|
|
param.grad.numpy(), param_base.grad.numpy(), rtol=tol, atol=tol
|
|
)
|
|
|
|
def test_mp_demo_net(self):
|
|
tol = 0.005
|
|
self.set_random_seed(self._seed)
|
|
mp_layer_o1 = dist.shard_layer(
|
|
DemoNet("mp_demo_weight_O1"), self._mesh, self.shard_fn
|
|
)
|
|
(
|
|
self.mp_loss_o1,
|
|
self.mp_parameters_o1,
|
|
) = self.run_dynamic_amp(mp_layer_o1, 'O1', run_dist=True)
|
|
np.testing.assert_allclose(
|
|
self.mp_loss_o1.numpy(),
|
|
self.base_loss_o1.numpy(),
|
|
rtol=tol,
|
|
atol=tol,
|
|
)
|
|
for param, param_base in zip(
|
|
self.mp_parameters_o1, self.base_parameters_o1
|
|
):
|
|
np.testing.assert_allclose(
|
|
param.numpy(), param_base.numpy(), rtol=tol, atol=tol
|
|
)
|
|
np.testing.assert_allclose(
|
|
param.grad.numpy(), param_base.grad.numpy(), rtol=tol, atol=tol
|
|
)
|
|
|
|
self.set_random_seed(self._seed)
|
|
mp_layer_o2 = dist.shard_layer(
|
|
DemoNet("mp_demo_weight_O2"), self._mesh, self.shard_fn
|
|
)
|
|
(
|
|
self.mp_loss_o2,
|
|
self.mp_parameters_o2,
|
|
) = self.run_dynamic_amp(mp_layer_o2, 'O2', run_dist=True)
|
|
np.testing.assert_allclose(
|
|
self.mp_loss_o2.numpy(), self.base_loss_o2.numpy(), rtol=tol
|
|
)
|
|
for param, param_base in zip(
|
|
self.mp_parameters_o2, self.base_parameters_o2
|
|
):
|
|
np.testing.assert_allclose(
|
|
param.numpy(), param_base.numpy(), rtol=tol, atol=tol
|
|
)
|
|
np.testing.assert_allclose(
|
|
param.grad.numpy(), param_base.grad.numpy(), rtol=tol, atol=tol
|
|
)
|
|
|
|
def run_test_case(self):
|
|
if self._dtype == "bfloat16" and not paddle.amp.is_bfloat16_supported():
|
|
return
|
|
self.test_dp_demo_net()
|
|
self.test_mp_demo_net()
|
|
|
|
|
|
if __name__ == '__main__':
|
|
TestSimpleNetWithAmpForSemiAutoParallel().run_test_case()
|