wehub-resource-sync
315 lines
10 KiB
Python
315 lines
10 KiB
Python
# Copyright (c) 2025 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
|
|
|
|
import paddle
|
|
import paddle.distributed as dist
|
|
from paddle.distributed.fsdp.fully_shard import fully_shard
|
|
from paddle.io import DataLoader, Dataset, DistributedBatchSampler
|
|
|
|
|
|
class RandomDataset(Dataset):
|
|
def __init__(self, num_samples=100, input_dim=10):
|
|
self.num_samples = num_samples
|
|
self.input_dim = input_dim
|
|
|
|
def __getitem__(self, idx):
|
|
np.random.seed(2025)
|
|
data = np.random.randn(self.input_dim).astype('float32')
|
|
label = np.random.randn(self.input_dim).astype('float32')
|
|
return data, label
|
|
|
|
def __len__(self):
|
|
return self.num_samples
|
|
|
|
|
|
class StandardMLPExpert(paddle.nn.Layer):
|
|
def __init__(self, d_model, d_hidden):
|
|
super().__init__()
|
|
self.fc1 = paddle.nn.Linear(d_model, d_hidden)
|
|
self.fc2 = paddle.nn.Linear(d_hidden, d_model)
|
|
|
|
def forward(self, x):
|
|
x = self.fc1(x)
|
|
x = paddle.nn.functional.relu(x)
|
|
x = self.fc2(x)
|
|
return x
|
|
|
|
|
|
class TransformerLayer(paddle.nn.Layer):
|
|
def __init__(self, d_model, d_hidden, num_experts=2, top_k=1):
|
|
super().__init__()
|
|
self.num_experts = num_experts
|
|
self.top_k = top_k
|
|
self.d_model = d_model
|
|
|
|
self.gate = paddle.nn.Linear(d_model, num_experts, bias_attr=False)
|
|
|
|
self.experts = paddle.nn.LayerList(
|
|
[StandardMLPExpert(d_model, d_hidden) for _ in range(num_experts)]
|
|
)
|
|
|
|
def forward(self, x):
|
|
orig_shape = x.shape
|
|
x = x.reshape([-1, self.d_model])
|
|
|
|
gate_logits = self.gate(x)
|
|
gate_probs = paddle.nn.functional.softmax(gate_logits, axis=-1)
|
|
|
|
topk_probs, topk_indices = paddle.topk(gate_probs, self.top_k, axis=-1)
|
|
|
|
output = paddle.zeros_like(x)
|
|
for k in range(self.top_k):
|
|
expert_idx = topk_indices[:, k]
|
|
expert_weight = topk_probs[:, k].unsqueeze(-1)
|
|
for i in range(self.num_experts):
|
|
mask = (expert_idx == i).astype(x.dtype).unsqueeze(-1)
|
|
expert_out = self.experts[i](x)
|
|
output = output + mask * expert_weight * expert_out
|
|
|
|
output = output.reshape(orig_shape)
|
|
return output
|
|
|
|
|
|
class SimpleMoEModel(paddle.nn.Layer):
|
|
def __init__(self, d_model=10, d_hidden=20, num_experts=2, top_k=1):
|
|
super().__init__()
|
|
self.config = type('Config', (), {'num_experts_per_tok': top_k})()
|
|
self.embed = paddle.nn.Linear(d_model, d_model)
|
|
self.moe = TransformerLayer(d_model, d_hidden, num_experts, top_k)
|
|
self.output = paddle.nn.Linear(d_model, d_model)
|
|
|
|
def forward(self, x):
|
|
x = self.embed(x)
|
|
x = x.unsqueeze(1)
|
|
x = self.moe(x)
|
|
x = x.squeeze(1)
|
|
x = self.output(x)
|
|
return x
|
|
|
|
|
|
class TestSemiAutoParallelFSDP:
|
|
def __init__(self):
|
|
self._backend = os.getenv("backend")
|
|
self._seed = eval(os.getenv("seed"))
|
|
self._mesh = dist.ProcessMesh([0, 1], dim_names=["dp"])
|
|
dist.auto_parallel.set_mesh(self._mesh)
|
|
self.gradient_accumulation_steps = 2
|
|
|
|
def create_dist_loader(self, batch_size):
|
|
dataset = RandomDataset(num_samples=10, input_dim=10)
|
|
|
|
sampler = DistributedBatchSampler(
|
|
dataset=dataset,
|
|
batch_size=batch_size,
|
|
shuffle=True,
|
|
drop_last=False,
|
|
)
|
|
|
|
loader = DataLoader(
|
|
dataset=dataset,
|
|
batch_sampler=sampler,
|
|
)
|
|
|
|
return loader
|
|
|
|
def test_sharding_stage_3(self):
|
|
paddle.seed(self._seed)
|
|
model = paddle.nn.Linear(10, 10)
|
|
data_loader = self.create_dist_loader(1)
|
|
dist_loader = dist.shard_dataloader(
|
|
dataloader=data_loader,
|
|
meshes=[self._mesh],
|
|
shard_dims="dp",
|
|
)
|
|
opt = paddle.optimizer.AdamW(parameters=model.parameters())
|
|
|
|
# use sharding stage 3
|
|
opt = dist.shard_optimizer(opt, dist.ShardingStage3("dp", self._mesh))
|
|
|
|
stage_losses = []
|
|
tr_loss_add = float(0)
|
|
step = 0
|
|
for batch in dist_loader:
|
|
tr_loss = model(batch[0])
|
|
tr_loss.backward()
|
|
tr_loss_add += tr_loss
|
|
if (step + 1) % self.gradient_accumulation_steps == 0:
|
|
tr_loss_add /= self.gradient_accumulation_steps
|
|
tr_loss = tr_loss_add
|
|
opt.step()
|
|
opt.clear_grad()
|
|
step += 1
|
|
stage_losses.append(tr_loss._local_value()._md5sum())
|
|
return stage_losses
|
|
|
|
def test_fsdp(self):
|
|
paddle.seed(self._seed)
|
|
model = paddle.nn.Linear(10, 10)
|
|
data_loader = self.create_dist_loader(1)
|
|
dist_loader = dist.shard_dataloader(
|
|
dataloader=data_loader,
|
|
meshes=[self._mesh],
|
|
shard_dims="dp",
|
|
)
|
|
opt = paddle.optimizer.AdamW(parameters=model.parameters())
|
|
|
|
# use FSDP
|
|
model = fully_shard(
|
|
model, mesh=self._mesh, enable_tensor_fusion_and_overlap=False
|
|
)
|
|
|
|
stage_losses = []
|
|
tr_loss_add = float(0)
|
|
step = 0
|
|
for batch in dist_loader:
|
|
tr_loss = model(batch[0])
|
|
tr_loss.backward()
|
|
tr_loss_add += tr_loss
|
|
if (step + 1) % self.gradient_accumulation_steps == 0:
|
|
tr_loss_add /= self.gradient_accumulation_steps
|
|
tr_loss = tr_loss_add
|
|
opt.step()
|
|
opt.clear_grad()
|
|
step += 1
|
|
stage_losses.append(tr_loss._local_value()._md5sum())
|
|
return stage_losses
|
|
|
|
def test_fsdp_with_auto_dp(self):
|
|
dist.enable_auto_dp()
|
|
paddle.seed(self._seed)
|
|
model = paddle.nn.Linear(10, 10)
|
|
data_loader = self.create_dist_loader(2)
|
|
dist_loader = dist.shard_dataloader(
|
|
dataloader=data_loader,
|
|
meshes=[self._mesh],
|
|
)
|
|
opt = paddle.optimizer.AdamW(parameters=model.parameters())
|
|
|
|
# use FSDP
|
|
model = fully_shard(
|
|
model, mesh=self._mesh, enable_tensor_fusion_and_overlap=False
|
|
)
|
|
|
|
stage_losses = []
|
|
tr_loss_add = float(0)
|
|
step = 0
|
|
for batch in dist_loader:
|
|
tr_loss = model(batch[0])
|
|
tr_loss.backward()
|
|
tr_loss_add += tr_loss
|
|
if (step + 1) % self.gradient_accumulation_steps == 0:
|
|
tr_loss_add /= self.gradient_accumulation_steps
|
|
tr_loss = tr_loss_add
|
|
opt.step()
|
|
opt.clear_grad()
|
|
step += 1
|
|
stage_losses.append(tr_loss._local_value()._md5sum())
|
|
return stage_losses
|
|
|
|
def test_fsdp_with_tensor_fusion_and_overlap(self):
|
|
dist.init_parallel_env()
|
|
dist.enable_auto_dp()
|
|
paddle.seed(self._seed)
|
|
model = paddle.nn.Linear(10, 10)
|
|
data_loader = self.create_dist_loader(2)
|
|
dist_loader = dist.shard_dataloader(
|
|
dataloader=data_loader,
|
|
meshes=[self._mesh],
|
|
)
|
|
opt = paddle.optimizer.AdamW(parameters=model.parameters())
|
|
|
|
# use FSDP with tensor_fusion and overlap
|
|
model = fully_shard(
|
|
model, mesh=self._mesh, enable_tensor_fusion_and_overlap=True
|
|
)
|
|
|
|
stage_losses = []
|
|
tr_loss_add = float(0)
|
|
step = 0
|
|
for batch in dist_loader:
|
|
tr_loss = model(batch[0])
|
|
tr_loss.backward()
|
|
tr_loss_add += tr_loss
|
|
if (step + 1) % self.gradient_accumulation_steps == 0:
|
|
tr_loss_add /= self.gradient_accumulation_steps
|
|
tr_loss = tr_loss_add
|
|
opt.step()
|
|
opt.clear_grad()
|
|
step += 1
|
|
stage_losses.append(tr_loss._local_value()._md5sum())
|
|
return stage_losses
|
|
|
|
def test_fsdp_with_moe(self):
|
|
dist.init_parallel_env()
|
|
dist.enable_auto_dp()
|
|
paddle.seed(self._seed)
|
|
|
|
model = SimpleMoEModel(d_model=10, d_hidden=20, num_experts=2, top_k=1)
|
|
|
|
data_loader = self.create_dist_loader(2)
|
|
dist_loader = dist.shard_dataloader(
|
|
dataloader=data_loader,
|
|
meshes=[self._mesh],
|
|
)
|
|
opt = paddle.optimizer.AdamW(parameters=model.parameters())
|
|
|
|
model = fully_shard(
|
|
model, mesh=self._mesh, enable_tensor_fusion_and_overlap=True
|
|
)
|
|
|
|
stage_losses = []
|
|
tr_loss_add = float(0)
|
|
step = 0
|
|
for batch in dist_loader:
|
|
tr_loss = model(batch[0])
|
|
tr_loss.backward()
|
|
tr_loss_add += tr_loss
|
|
if (step + 1) % self.gradient_accumulation_steps == 0:
|
|
tr_loss_add /= self.gradient_accumulation_steps
|
|
tr_loss = tr_loss_add
|
|
opt.step()
|
|
opt.clear_grad()
|
|
step += 1
|
|
stage_losses.append(tr_loss._local_value()._md5sum())
|
|
return stage_losses
|
|
|
|
def run_test_case(self):
|
|
if self._backend == "cpu":
|
|
paddle.set_device("cpu")
|
|
elif self._backend == "gpu":
|
|
paddle.set_device("gpu:" + str(dist.get_rank()))
|
|
else:
|
|
raise ValueError("Only support cpu or gpu backend.")
|
|
|
|
losses_stage_3 = self.test_sharding_stage_3()
|
|
losses_fsdp = self.test_fsdp()
|
|
assert losses_stage_3 == losses_fsdp
|
|
|
|
losses_fsdp_with_auto_dp = self.test_fsdp_with_auto_dp()
|
|
assert losses_fsdp_with_auto_dp == losses_fsdp
|
|
|
|
losses_fsdp_fusion = self.test_fsdp_with_tensor_fusion_and_overlap()
|
|
assert losses_fsdp_fusion == losses_fsdp
|
|
|
|
losses_fsdp_moe = self.test_fsdp_with_moe()
|
|
assert len(losses_fsdp_moe) > 0, "MoE test should produce losses"
|
|
|
|
|
|
if __name__ == '__main__':
|
|
TestSemiAutoParallelFSDP().run_test_case()
|