chore: import upstream snapshot with attribution

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/__init__.py

@@ -0,0 +1,18 @@
+#   Copyright (c) 2021 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
+from .dygraph_sharding_optimizer import DygraphShardingOptimizer  # noqa: F401
+from .heter_parallel_optimizer import HeterParallelOptimizer  # noqa: F401
+from .hybrid_parallel_gradscaler import HybridParallelGradScaler  # noqa: F401
+from .hybrid_parallel_optimizer import HybridParallelOptimizer  # noqa: F401
+
+__all__ = []

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/heter_parallel_optimizer.py

@@ -0,0 +1,65 @@
+# Copyright (c) 2021 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 paddle.autograd as imperative_base
+from paddle import framework
+
+__all__ = []
+
+
+def _obtain_optimizer_parameters_list(optimizer):
+    if getattr(optimizer, '_param_groups', None) and isinstance(
+        optimizer._param_groups[0], dict
+    ):
+        parameters_list = []
+        for group in optimizer._param_groups:
+            for param in group['params']:
+                parameters_list.append(param)
+    else:
+        parameters_list = list(optimizer._parameter_list)
+
+    return parameters_list
+
+
+class HeterParallelOptimizer:
+    # adapter wrapper for optimizer
+    def __init__(self, optimizer, strategy):
+        self._inner_opt = optimizer
+        self._strategy = strategy
+
+        # NOTE(liubo48): In pure DataParallel mode,
+        # the gradient synchronization is achieved through reducer.
+
+    @imperative_base.no_grad()
+    @framework.dygraph_only
+    def step(self):
+        parameters_list = _obtain_optimizer_parameters_list(self._inner_opt)
+        self._inner_opt.step()
+
+    @imperative_base.no_grad()
+    def minimize(
+        self, loss, startup_program=None, parameters=None, no_grad_set=None
+    ):
+        # minimize does not support parameters in the form of param_group,
+        # so no need use _obtain_optimizer_parameters_list
+        parameter_list = (
+            parameters if parameters else self._inner_opt._parameter_list
+        )
+
+        return self._inner_opt.minimize(
+            loss, startup_program, parameter_list, no_grad_set
+        )
+
+    def __getattr__(self, item):
+        return getattr(self._inner_opt, item)

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/hybrid_parallel_gradscaler.py

@@ -0,0 +1,85 @@
+# Copyright (c) 2021 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 paddle
+import paddle.autograd as imperative_base
+from paddle import _legacy_C_ops
+
+from ...base.topology import ParallelMode
+
+__all__ = []
+
+
+class HybridParallelGradScaler:
+    def __init__(self, scaler, hcg):
+        self._scaler = scaler
+        self._hcg = hcg
+        self._use_dp_mode = (
+            self._hcg.get_parallel_mode() == ParallelMode.DATA_PARALLEL
+        )
+
+    def scale(self, var):
+        return self._scaler.scale(var)
+
+    def minimize(self, optimizer, *args, **kwargs):
+        if not self._enable:
+            return optimizer.minimize(*args, **kwargs)
+
+        #  unscale the grad
+        self._unscale(optimizer)
+
+        optimize_ops, params_grads = (None, None)
+
+        if hasattr(optimizer, "_set_auxiliary_var"):
+            optimizer._set_auxiliary_var('found_inf', self._found_inf)
+            optimize_ops, params_grads = optimizer.minimize(*args, **kwargs)
+            # TODO: Fix to _cache_found_inf after PaddleNLP update
+            self._cache_found_inf = optimizer._get_auxiliary_var('found_inf')
+        else:
+            if self._found_inf:
+                self._cache_found_inf = True
+            else:
+                optimize_ops, params_grads = optimizer.minimize(*args, **kwargs)
+                self._cache_found_inf = False
+
+        if self._use_dynamic_loss_scaling:
+            self._update()
+
+        return optimize_ops, params_grads
+
+    @imperative_base.no_grad()
+    def _unscale(self, optimizer):
+        if not self._enable:
+            return
+        param_grads = [
+            param._grad_ivar()
+            for param in optimizer._parameter_list
+            if param._grad_ivar() is not None
+        ]
+        _legacy_C_ops.check_finite_and_unscale(
+            param_grads, self._scale, param_grads, self._found_inf
+        )
+        # allreduce_max found_inf in check_group
+        if not self._use_dp_mode:
+            self._found_inf = paddle.cast(self._found_inf, dtype="int32")
+            # TODO(shenliang03) Since the minimize call in the optimizer is
+            # after the grad scaler, check_finite needs to synchronize global
+            # information. In the future, we should use check_group
+            paddle.distributed.all_reduce(
+                self._found_inf, op=paddle.distributed.ReduceOp.MAX, group=None
+            )
+            self._found_inf = paddle.cast(self._found_inf, dtype="bool")
+
+    def __getattr__(self, item):
+        return getattr(self._scaler, item)