wehub-resource-sync
278 lines
8.8 KiB
Python
278 lines
8.8 KiB
Python
# Copyright (c) 2024 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 random
|
|
import unittest
|
|
|
|
import numpy as np
|
|
from test_to_static_pir_program import (
|
|
DemoNet,
|
|
create_data_loader,
|
|
)
|
|
|
|
import paddle
|
|
import paddle.distributed as dist
|
|
from paddle import nn
|
|
|
|
BATCH_SIZE = 4
|
|
BATCH_NUM = 40
|
|
IMAGE_SIZE = 16
|
|
CLASS_NUM = 8
|
|
np.random.seed(2024)
|
|
paddle.seed(2024)
|
|
|
|
|
|
class PPDemoNet(nn.Layer):
|
|
def __init__(self, mesh1, mesh2):
|
|
super().__init__()
|
|
self._mesh1 = mesh1
|
|
self._mesh2 = mesh2
|
|
self.linear_0 = nn.Linear(IMAGE_SIZE, IMAGE_SIZE, bias_attr=False)
|
|
self.linear_1 = nn.Linear(IMAGE_SIZE, CLASS_NUM, bias_attr=False)
|
|
self.relu_0 = nn.ReLU()
|
|
self.relu_1 = nn.ReLU()
|
|
self.relu_2 = nn.ReLU()
|
|
# shard the weights of this layer
|
|
self.linear_0.weight = dist.shard_tensor(
|
|
self.linear_0.weight,
|
|
self._mesh1,
|
|
[dist.Replicate()],
|
|
stop_gradient=False,
|
|
)
|
|
self.linear_1.weight = dist.shard_tensor(
|
|
self.linear_1.weight,
|
|
self._mesh2,
|
|
[dist.Replicate()],
|
|
stop_gradient=False,
|
|
)
|
|
|
|
def forward(self, x):
|
|
x.stop_gradient = False
|
|
out = self.relu_0(x) # trigger backward partial allreduce
|
|
out = self.linear_0(out)
|
|
out = self.relu_1(out)
|
|
out = dist.reshard(out, self._mesh2, [dist.Replicate()])
|
|
out = self.linear_1(out)
|
|
out = self.relu_2(out) # trigger forward partial allreduce
|
|
out = paddle.cast(out, 'float32')
|
|
return out
|
|
|
|
|
|
class DPDemoNet(nn.Layer):
|
|
def __init__(
|
|
self,
|
|
mesh,
|
|
):
|
|
super().__init__()
|
|
self._mesh = mesh
|
|
self.linear_0 = nn.Linear(IMAGE_SIZE, IMAGE_SIZE, bias_attr=False)
|
|
self.linear_1 = nn.Linear(IMAGE_SIZE, CLASS_NUM, bias_attr=False)
|
|
self.linear_0.weight = dist.shard_tensor(
|
|
self.linear_0.weight,
|
|
self._mesh,
|
|
[dist.Replicate()],
|
|
stop_gradient=False,
|
|
)
|
|
self.linear_1.weight = dist.shard_tensor(
|
|
self.linear_1.weight,
|
|
self._mesh,
|
|
[dist.Replicate()],
|
|
stop_gradient=False,
|
|
)
|
|
self.relu_0 = nn.ReLU()
|
|
self.relu_1 = nn.ReLU()
|
|
self.relu_2 = nn.ReLU()
|
|
|
|
def forward(self, x):
|
|
out = self.relu_0(x)
|
|
out = self.linear_0(out)
|
|
out = self.relu_1(out)
|
|
out = self.linear_1(out)
|
|
out = self.relu_2(out)
|
|
out = paddle.cast(out, 'float32')
|
|
return out
|
|
|
|
|
|
class TestMLPTensorParallel(unittest.TestCase):
|
|
def test_to_static_program(self):
|
|
paddle.base.set_flags({'FLAGS_enable_pir_api': 1})
|
|
mesh = dist.ProcessMesh([0, 1], dim_names=["x"])
|
|
mp_layer = DemoNet(mesh, True)
|
|
opt = paddle.optimizer.SGD(
|
|
learning_rate=0.1, parameters=mp_layer.parameters()
|
|
)
|
|
loss_fn = nn.MSELoss()
|
|
loader = create_data_loader()
|
|
dist_loader = dist.shard_dataloader(loader, meshes=[mesh])
|
|
dist_model = dist.to_static(mp_layer, dist_loader, loss_fn, opt)
|
|
|
|
dist_model.train()
|
|
for batch_id, (image, label) in enumerate(dist_loader()):
|
|
loss = dist_model(image, label)
|
|
|
|
|
|
class TestMLPReplicated(unittest.TestCase):
|
|
def test_to_static_program(self):
|
|
paddle.base.set_flags({'FLAGS_enable_pir_api': 1})
|
|
mesh = dist.ProcessMesh([0, 1], dim_names=["x"])
|
|
replicated_layer = DemoNet(mesh, False)
|
|
opt = paddle.optimizer.SGD(
|
|
learning_rate=0.1, parameters=replicated_layer.parameters()
|
|
)
|
|
loss_fn = nn.MSELoss()
|
|
loader = create_data_loader()
|
|
dist_loader = dist.shard_dataloader(loader, meshes=[mesh])
|
|
dist_model = dist.to_static(replicated_layer, dist_loader, loss_fn, opt)
|
|
|
|
dist_model.train()
|
|
for batch_id, (image, label) in enumerate(dist_loader()):
|
|
loss = dist_model(image, label)
|
|
|
|
|
|
class TestMLPPipelineParallel(unittest.TestCase):
|
|
def init_env(self):
|
|
paddle.seed(1024)
|
|
np.random.seed(1024)
|
|
random.seed(1024)
|
|
|
|
def test_to_static_program(self):
|
|
paddle.base.set_flags({'FLAGS_enable_pir_api': 1})
|
|
mesh1 = dist.ProcessMesh([0], dim_names=["x"])
|
|
mesh2 = dist.ProcessMesh([1], dim_names=["y"])
|
|
pp_layer = PPDemoNet(mesh1, mesh2)
|
|
opt = paddle.optimizer.SGD(
|
|
learning_rate=0.1, parameters=pp_layer.parameters()
|
|
)
|
|
loss_fn = nn.MSELoss()
|
|
loader = create_data_loader()
|
|
dist_loader = dist.shard_dataloader(loader, meshes=[mesh1, mesh2])
|
|
dist_model = dist.to_static(pp_layer, dist_loader, loss_fn, opt)
|
|
dist_model.train()
|
|
mode = "train"
|
|
|
|
for batch_id, (image, label) in enumerate(dist_loader()):
|
|
loss = dist_model(image, label)
|
|
|
|
def _pipeline_schedule(
|
|
self,
|
|
enable_schedule=False,
|
|
schedule_mode="FThenB",
|
|
accumulate_steps=1,
|
|
grad_merge=False,
|
|
enable_amp=True,
|
|
):
|
|
self.init_env()
|
|
paddle.set_flags({'FLAGS_enable_pir_api': 1})
|
|
mesh1 = dist.ProcessMesh([0], dim_names=["x"])
|
|
mesh2 = dist.ProcessMesh([1], dim_names=["x"])
|
|
pp_layer = PPDemoNet(mesh1, mesh2)
|
|
opt = paddle.optimizer.SGD(
|
|
learning_rate=0.1, parameters=pp_layer.parameters()
|
|
)
|
|
loss_fn = nn.MSELoss()
|
|
loader = create_data_loader(
|
|
BATCH_SIZE, BATCH_NUM, IMAGE_SIZE, CLASS_NUM
|
|
)
|
|
strategy = dist.Strategy()
|
|
strategy.pipeline.enable = enable_schedule
|
|
strategy.pipeline.schedule_mode = schedule_mode
|
|
strategy.pipeline.accumulate_steps = accumulate_steps
|
|
|
|
if enable_amp:
|
|
amp = strategy.amp
|
|
amp.enable = True
|
|
amp.dtype = 'float16'
|
|
amp.level = 'O2'
|
|
amp.use_master_weight = True
|
|
amp.use_master_grad = True
|
|
amp.use_promote = True
|
|
amp.init_loss_scaling = 1024.0
|
|
|
|
if grad_merge:
|
|
gradient_merge = strategy.gradient_merge
|
|
gradient_merge.enable = True
|
|
gradient_merge.k_steps = accumulate_steps
|
|
gradient_merge.avg = True
|
|
|
|
dist_loader = dist.shard_dataloader(loader, meshes=[mesh1, mesh2])
|
|
dist_model = dist.to_static(
|
|
pp_layer, dist_loader, loss_fn, opt, strategy
|
|
)
|
|
dist_model.train()
|
|
|
|
loss = None
|
|
for batch_id, (image, label) in enumerate(dist_loader()):
|
|
loss = dist_model(image, label)
|
|
|
|
if accumulate_steps > 1 and loss is not None:
|
|
loss = np.mean(loss)
|
|
return loss
|
|
|
|
def test_pp_pass(self):
|
|
ref_loss = self._pipeline_schedule()
|
|
# only split_program
|
|
loss_split_prog_acc1 = self._pipeline_schedule(
|
|
enable_schedule=False, schedule_mode="FThenB", accumulate_steps=1
|
|
)
|
|
self.assertEqual(ref_loss, loss_split_prog_acc1)
|
|
|
|
loss_split_prog_acc4 = self._pipeline_schedule(
|
|
enable_schedule=True,
|
|
schedule_mode="FThenB",
|
|
accumulate_steps=4,
|
|
grad_merge=True,
|
|
)
|
|
|
|
if ref_loss is None:
|
|
self.assertEqual(ref_loss, loss_split_prog_acc4)
|
|
else:
|
|
ret_1 = np.allclose(
|
|
loss_split_prog_acc4,
|
|
ref_loss,
|
|
rtol=1e-3,
|
|
atol=1e-2,
|
|
equal_nan=True,
|
|
)
|
|
self.assertEqual(ret_1, True)
|
|
|
|
def test_pp_pass_amp(self):
|
|
loss_split_prog_acc1 = self._pipeline_schedule(
|
|
enable_schedule=False,
|
|
schedule_mode="FThenB",
|
|
accumulate_steps=1,
|
|
enable_amp=True,
|
|
)
|
|
loss_split_prog_acc4 = self._pipeline_schedule(
|
|
enable_schedule=True,
|
|
schedule_mode="FThenB",
|
|
accumulate_steps=4,
|
|
grad_merge=True,
|
|
enable_amp=True,
|
|
)
|
|
|
|
cur_rank = paddle.distributed.get_rank()
|
|
if cur_rank == 1:
|
|
ret_1 = np.allclose(
|
|
loss_split_prog_acc4,
|
|
loss_split_prog_acc1,
|
|
rtol=1e-3,
|
|
atol=1e-2,
|
|
equal_nan=True,
|
|
)
|
|
self.assertEqual(ret_1, True)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|