chore: import upstream snapshot with attribution

python/paddle/distributed/launch/plugins/__init__.py

@@ -0,0 +1,89 @@
+# Copyright (c) 2022 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
+
+__all__ = []
+
+
+# print configuration after args are well filled in controller init
+def log(ctx):
+    ctx.logger.info("-----------  Configuration  ----------------------")
+    for arg, value in sorted(vars(ctx.args).items()):
+        ctx.logger.info(f"{arg}: {value}")
+    ctx.logger.info("--------------------------------------------------")
+
+
+def process_args(ctx):
+    # reset device by args
+    # argdev = ctx.args.gpus or ctx.args.xpus or ctx.args.npus
+    argdev = ctx.args.devices
+    if argdev:
+        for d in argdev.split(','):
+            if d not in ctx.node.device.labels:
+                ctx.logger.error(
+                    f'Device not found {d} from {argdev} for setting {ctx.node.device.labels}'
+                )
+
+    if ctx.args.ips:
+        ips = ctx.args.ips.split(',')
+        if '127.0.0.1' in ips and len(ips) != 1:
+            raise ValueError("127.0.0.1 in ips is not allowed in multi-nodes.")
+
+
+def collective_compatible(ctx):
+    force_use_args = int(os.getenv("PADDLE_LAUNCH_WITH_ARGS", "0"))
+    if 'PADDLE_TRAINER_ENDPOINTS' in ctx.envs:
+        eps = ctx.envs['PADDLE_TRAINER_ENDPOINTS'].split(',')
+        hosts = {h.split(':')[0] for h in eps}
+        if force_use_args:
+            ctx.args.master = None
+        else:
+            ctx.args.master = eps[0] if ':' in eps[0] else f'{eps[0]}:6768'
+        ctx.args.nnodes = len(hosts)
+        ctx.logger.info(f'args reset by env PADDLE_TRAINER_ENDPOINTS\n{eps}')
+
+    if 'DISTRIBUTED_TRAINER_ENDPOINTS' in ctx.envs:
+        eps = ctx.envs['DISTRIBUTED_TRAINER_ENDPOINTS'].split(',')
+        hosts = {h.split(':')[0] for h in eps}
+        if force_use_args:
+            ctx.args.master = None
+        else:
+            ctx.args.master = eps[0]
+        ctx.args.nnodes = len(hosts)
+        ctx.logger.info(
+            f'args reset by env DISTRIBUTED_TRAINER_ENDPOINTS\n{eps}'
+        )
+
+
+def rewrite_host_ip(ctx):
+    if ctx.args.host is not None and "." in ctx.args.host:
+        ctx.logger.warning(f'Host ip reset to {ctx.args.host}')
+        ctx.node.ip = ctx.args.host
+
+
+def test_mode(ctx):
+    if ctx.args.training_script == 'run_check':
+        ctx.logger.info('Paddle Distributed Test begin...')
+        if int(ctx.args.nnodes) < 2:
+            ctx.args.nnodes = 2
+        ctx.args.training_script = f'{os.path.dirname(__file__)}/test.py'
+
+
+enabled_plugins = [
+    test_mode,
+    collective_compatible,
+    rewrite_host_ip,
+    process_args,
+]

python/paddle/distributed/launch/plugins/test.py

@@ -0,0 +1,105 @@
+# Copyright (c) 2022 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 numpy as np
+
+import paddle
+from paddle.distributed import fleet
+from paddle.io import DataLoader, Dataset
+from paddle.vision.models import ResNet
+from paddle.vision.models.resnet import BottleneckBlock
+
+base_lr = 0.1
+momentum_rate = 0.9
+l2_decay = 1e-4
+
+epoch = 3
+batch_num = 1
+batch_size = 1
+class_dim = 102
+
+
+# define a random dataset
+class RandomDataset(Dataset):
+    def __init__(self, num_samples):
+        self.num_samples = num_samples
+
+    def __getitem__(self, idx):
+        image = np.random.random([3, 224, 224]).astype('float32')
+        label = np.random.randint(0, class_dim - 1, (1,)).astype('int64')
+        return image, label
+
+    def __len__(self):
+        return self.num_samples
+
+
+def optimizer_setting(parameter_list=None):
+    optimizer = paddle.optimizer.Momentum(
+        learning_rate=base_lr,
+        momentum=momentum_rate,
+        weight_decay=paddle.regularizer.L2Decay(l2_decay),
+        parameters=parameter_list,
+    )
+    return optimizer
+
+
+def train_resnet():
+    fleet.init(is_collective=True)
+
+    resnet = ResNet(BottleneckBlock, 18, num_classes=class_dim)
+    optimizer = optimizer_setting(parameter_list=resnet.parameters())
+    optimizer = fleet.distributed_optimizer(optimizer)
+    resnet = fleet.distributed_model(resnet)
+
+    dataset = RandomDataset(batch_num * batch_size)
+    train_loader = DataLoader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=True,
+        drop_last=True,
+        num_workers=2,
+    )
+
+    print("Distributed training start...")
+    for eop in range(epoch):
+        resnet.train()
+
+        for batch_id, data in enumerate(train_loader()):
+            img, label = data
+            label.stop_gradient = True
+
+            out = resnet(img)
+            loss = paddle.nn.functional.cross_entropy(input=out, label=label)
+            avg_loss = paddle.mean(x=loss)
+            acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
+            acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
+
+            avg_loss.backward()
+            optimizer.step()
+            resnet.clear_gradients()
+
+            print(
+                f"[Epoch {eop}, batch {batch_id}] loss: {avg_loss:.5f}, acc1: {acc_top1:.5f}, acc5: {acc_top5:.5f}"
+            )
+
+    print("Distributed training completed")
+
+
+if __name__ == '__main__':
+    import os
+
+    nnodes = os.getenv('PADDLE_NNODES')
+    cn = os.getenv('PADDLE_LOCAL_SIZE')
+    print(f"Prepare distributed training with {nnodes} nodes {cn} cards")
+    train_resnet()