[AutoParallel] dist_scale (#48295)

python/paddle/distributed/auto_parallel/operators/__init__.py

@@ -35,3 +35,4 @@ from . import dist_fused_attention
 from . import dist_reduce_sum_p
 from . import dist_shape
 from . import dist_assign
+from . import dist_scale

python/paddle/distributed/auto_parallel/operators/dist_scale.py

+# 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.
+
+from .common import DistributedOperatorImplContainer
+from .common import DistributedOperatorImpl
+from .common import register_distributed_operator_impl_container
+from .common import register_distributed_operator_impl
+from .dist_default import DistributedDefaultImpl0
+from ..utils import compute_compatible_and_update_dim_mapping
+
+
+class DistributedScale(DistributedOperatorImplContainer):
+    def __init__(self, op_type):
+        super().__init__(op_type)
+
+
+register_distributed_operator_impl_container(DistributedScale("scale"))
+
+
+class DistributedScaleImpl(DistributedOperatorImpl):
+    def __init__(self, name):
+        super().__init__(name)
+        self._forward_implemented = True
+        self._backward_implemented = True
+
+    def is_input_compatible(self, dist_op):
+        return True
+
+    def is_output_compatible(self, dist_op):
+        return True
+
+    def is_auto_compatible(self, dist_op):
+        if (not self.is_input_compatible(dist_op)) or (
+            not self.is_output_compatible(dist_op)
+        ):
+            return False
+
+        op_desc = dist_op.serial_op.desc
+        op_dist_attr = dist_op.dist_attr
+        x_name = op_desc.input('X')[0]
+        out_name = op_desc.output('Out')[0]
+        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
+        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
+
+        if x_dims_mapping != out_dims_mapping:
+            return False
+
+        return True
+
+    def update_dims_mapping(self, dist_op):
+        changed = False
+        op_desc = dist_op.serial_op.desc
+        op_dist_attr = dist_op.dist_attr
+        x_name = op_desc.input('X')[0]
+        out_name = op_desc.output('Out')[0]
+        x_dims_mapping = op_dist_attr.get_input_dims_mapping(x_name)
+        out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name)
+
+        for i in range(len(x_dims_mapping)):
+            dim_changed = compute_compatible_and_update_dim_mapping(
+                [x_dims_mapping, out_dims_mapping], [i, i]
+            )
+            if dim_changed:
+                changed = True
+
+        return changed
+
+    @staticmethod
+    def forward(ctx, *args, **kwargs):
+        DistributedDefaultImpl0.forward(ctx, *args, **kwargs)
+
+    @staticmethod
+    def backward(ctx, *args, **kwargs):
+        DistributedDefaultImpl0.backward(ctx, *args, **kwargs)
+
+
+register_distributed_operator_impl("scale", DistributedScaleImpl("scale"))

python/paddle/fluid/tests/unittests/auto_parallel/test_dist_scale.py

+# 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 unittest
+import paddle
+from paddle.distributed.fleet import auto
+
+paddle.enable_static()
+
+
+def make_program():
+    main_program = paddle.fluid.Program()
+    start_program = paddle.fluid.Program()
+    with paddle.static.program_guard(main_program, start_program):
+        x = paddle.static.data(name='x', shape=[4, 4, 8], dtype='float32')
+        x.stop_gradient = False
+        auto.shard_tensor(
+            x, auto.ProcessMesh([0, 1], dim_names=["x"]), [None, "x", None]
+        )
+        res = paddle.scale(x, scale=2.0, bias=1.0)
+    return main_program, start_program
+
+
+def parallelizer(program_func, rank):
+    from paddle.distributed.auto_parallel.completion import Completer
+    from paddle.distributed.auto_parallel.partitioner import Partitioner
+    from paddle.distributed.auto_parallel.dist_context import DistributedContext
+
+    main_program, start_program = program_func()
+
+    dist_context = DistributedContext()
+    completer = Completer(dist_context)
+    completer.complete_forward_annotation(main_program)
+    dist_context.block_state.parse_forward_blocks(main_program)
+
+    partitioner = Partitioner(dist_context, rank)
+    dist_main_prog, _, _ = partitioner.partition(
+        main_program, start_program, []
+    )
+
+    return dist_main_prog, dist_context
+
+
+class TestDistScale(unittest.TestCase):
+    def test_dist_scale(self):
+
+        dist_main_prog, dist_context = parallelizer(make_program, 0)
+        ops = dist_main_prog.global_block().ops
+        scale_op = ops[0]
+        dist_op = dist_context.get_dist_op_for_program(scale_op)
+        dist_op.dist_attr.impl_type == "scale"
+        dist_op.dist_attr.impl_idx == 0
+
+        in_name = scale_op.input_arg_names[0]
+        out_name = scale_op.output_arg_names[0]
+        in_dims_mapping = dist_op.dist_attr.get_input_dims_mapping(in_name)
+        out_dims_mapping = dist_op.dist_attr.get_output_dims_mapping(out_name)
+
+        assert in_dims_mapping == out_dims_mapping
+
+
+if __name__ == "__main__":
+    unittest.main()