[Auto Parallel] Improve the APIs

python/paddle/distributed/auto_parallel/__init__.py

@@ -12,10 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .interface import shard_tensor  # noqa: F401
-from .interface import shard_op  # noqa: F401
+from .strategy import Strategy
 from .process_mesh import ProcessMesh
-from .reshard import Resharder  # noqa: F401
-from .cost_model import estimate_cost
+from .engine import Engine
+from .interface import shard_tensor
+from .interface import shard_op
+from .interface import recompute
+from .interface import fetch
 
 __all__ = []

python/paddle/distributed/auto_parallel/constants.py

@@ -0,0 +1,118 @@
+# 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 collections import defaultdict
+
+# _g_default_config[category][field] = default_value
+_g_default_config = defaultdict(dict)
+
+
+def get_category_default_config(category):
+    return _g_default_config[category]
+
+
+def set_category_default_config(category, default_value):
+    _g_default_config[category] = default_value
+
+
+def get_field_default_config(category, field):
+    return _g_default_config[category][field]
+
+
+def set_field_default_config(category, field, default_value):
+    _g_default_config[category][field] = default_value
+
+
+NOT_FOUND = "not_found"
+
+#########################################
+# base configuration
+#########################################
+BASE = "base"
+set_field_default_config(BASE, "auto_mode", "semi")
+set_field_default_config(BASE, "gradient_scale", True)
+set_field_default_config(BASE, "use_cache", True)
+set_field_default_config(BASE, "return_numpy", True)
+set_field_default_config(BASE, "all_ranks", False)
+set_field_default_config(BASE, "split_data", False)
+set_field_default_config(BASE, "seed", None)
+set_field_default_config(BASE, "reinit", False)  # Only for debug
+
+#########################################
+# recompute configuration
+#########################################
+RECOMPUTE = "recompute"
+set_field_default_config(RECOMPUTE, "enable", False)
+set_field_default_config(RECOMPUTE, "checkpoints", None)
+set_field_default_config(RECOMPUTE, "enable_tuning", False)
+
+#########################################
+# AMP configuration
+#########################################
+AMP = "amp"
+set_field_default_config(AMP, "enable", False)
+set_field_default_config(AMP, "init_loss_scaling", 32768.0)
+set_field_default_config(AMP, "incr_every_n_steps", 1000)
+set_field_default_config(AMP, "decr_every_n_nan_or_inf", 2)
+set_field_default_config(AMP, "incr_ratio", 2.0)
+set_field_default_config(AMP, "decr_ratio", 0.8)
+set_field_default_config(AMP, "use_dynamic_loss_scaling", True)
+set_field_default_config(AMP, "custom_white_list", [])
+set_field_default_config(AMP, "custom_black_list", [])
+set_field_default_config(AMP, "custom_black_varnames", [])
+set_field_default_config(AMP, "use_pure_fp16", False)
+set_field_default_config(AMP, "use_fp16_guard", True)
+set_field_default_config(AMP, "use_optimizer_fp16", False)
+
+#########################################
+# sharding configuration
+#########################################
+SHARDING = "sharding"
+set_field_default_config(SHARDING, "enable", False)
+set_field_default_config(SHARDING, "stage", 1)
+set_field_default_config(SHARDING, "sharding_degree", 8)
+set_field_default_config(SHARDING, "segment_broadcast_MB", 32.0)
+set_field_default_config(SHARDING, "enable_tuning", False)
+set_field_default_config(SHARDING, "tuning_range", [])
+
+#########################################
+# gradient merge configuration
+#########################################
+GRADIENT_MERGE = "gradient_merge"
+set_field_default_config(GRADIENT_MERGE, "enable", False)
+set_field_default_config(GRADIENT_MERGE, "k_steps", 1)
+set_field_default_config(GRADIENT_MERGE, "avg", True)
+
+#########################################
+# quantization configuration
+#########################################
+QAT = "qat"
+set_field_default_config(QAT, "enable", False)
+set_field_default_config(QAT, "channel_wise_abs_max", True)
+set_field_default_config(QAT, "weight_bits", 8)
+set_field_default_config(QAT, "activation_bits", 8)
+set_field_default_config(QAT, "not_quant_pattern", ['skip_quant'])
+set_field_default_config(QAT, "algo", None)
+
+# #########################################
+# auto tuning configuration
+# #########################################
+TUNING = "tuning"
+set_field_default_config(TUNING, "enable", False)
+set_field_default_config(TUNING, "batch_size", 1)
+set_field_default_config(TUNING, "dataset", None)
+set_field_default_config(TUNING, "profile_start_step", 1)
+set_field_default_config(TUNING, "profile_end_step", 1)
+set_field_default_config(TUNING, "run_after_tuning", True)
+set_field_default_config(TUNING, "verbose", True)

python/paddle/distributed/auto_parallel/converter.py

@@ -16,7 +16,7 @@
 import warnings
 import logging
 import numpy as np
-from ..utils import get_logger
+from .utils import get_logger
 
 
 class Converter(object):

python/paddle/distributed/auto_parallel/dist_attribute.py

@@ -173,6 +173,17 @@ def mark_annotated_as(self, dist_attr):
     def clear_annotated(self):
         self._is_annotated.clear()
 
+    def __eq__(self, other):
+        if not isinstance(other, TensorDistributedAttribute):
+            return False
+        if self.process_mesh != other.process_mesh:
+            return False
+        if self.dims_mapping != other.dims_mapping:
+            return False
+        if self._is_annotated != other._is_annotated:
+            return False
+        return True
+
     def __str__(self):
         str = "\n\ttensor_dist_attr = {"
         if self.is_annotated("process_mesh"):
@@ -486,6 +497,27 @@ def is_annotated_output_dims_mapping(self, name):
         else:
             return False
 
+    def __eq__(self, other):
+        if not isinstance(other, OperatorDistributedAttribute):
+            return False
+        if self.process_mesh != other.process_mesh:
+            return False
+        if self.op_type != other.op_type:
+            return False
+        if self.impl_type != other.impl_type:
+            return False
+        if self.impl_idx != other.impl_idx:
+            return False
+        if self._is_annotated != other._is_annotated:
+            return False
+        if self._is_recompute != other._is_recompute:
+            return False
+        if self.inputs_dist_attrs != other.inputs_dist_attrs:
+            return False
+        if self.outputs_dist_attrs != other.outputs_dist_attrs:
+            return False
+        return True
+
     def __str__(self):
         str = "\n\top_dist_attr = {"
         if self.is_annotated("process_mesh"):

python/paddle/distributed/auto_parallel/dist_context.py

@@ -126,9 +126,6 @@ def __init__(self,
         # A flag indicates whether the used parallelism is data parallel
         self._data_parallel = False
 
-        # flag whether using `to_static`
-        self._dygraph_mode = False
-
     @property
     def serial_main_program(self):
         return self._serial_main_program

python/paddle/distributed/auto_parallel/dist_op.py

@@ -23,6 +23,7 @@
 from .dist_attribute import append_op_output_suffix
 from .dist_attribute import get_tensor_dist_attr_field_keys
 from .dist_attribute import get_op_dist_attr_field_keys
+from .utils import convert_to_shard_spec, verify_shard_spec
 
 
 class DistributedOperator:
@@ -248,23 +249,106 @@ def __deepcopy__(self, memo):
         return result
 
 
-class DistributedModule:
+class DistributedOperatorHelper:
 
-    def __init__(self, serial_module, dist_attr=None):
-        self._serial_module = serial_module
-        self._dist_attr = dist_attr
+    def __init__(self, serial_op, process_mesh, in_dims_mappings,
+                 out_dims_mappings):
+        self._serial_op = serial_op
+        self._process_mesh = process_mesh
+        self._in_dims_mappings = in_dims_mappings
+        self._out_dims_mappings = out_dims_mappings
 
     def __call__(self, *args, **kwargs):
-        from .dist_context import get_default_distributed_context
+        tensor_to_dims_mapping = {}
+        index = 0
+        if self._in_dims_mappings:
+            assert len(args) + len(kwargs) == len(self._in_dims_mappings), \
+                "The length of dims_mapping {} does not matching the length output {}.".format(len(self._in_dims_mappings), len(args) + len(kwargs))
+        for arg in args:
+            if isinstance(arg, Variable) and self._in_dims_mappings:
+                tensor_to_dims_mapping[arg.name] = self._in_dims_mappings[index]
+            index += 1
+        for arg in kwargs.values() and self._in_dims_mappings:
+            if isinstance(arg, Variable):
+                tensor_to_dims_mapping[arg.name] = self._in_dims_mappings[index]
+            index += 1
+
         default_prog = paddle.fluid.default_main_program()
         cur_block = default_prog.current_block()
         op_size = len(cur_block.ops)
-        output = self._serial_module(*args, **kwargs)
+        output = self._serial_op(*args, **kwargs)
         new_op_size = len(cur_block.ops)
+
+        if isinstance(output, tuple) or isinstance(output, list):
+            new_output = list(output)
+        elif isinstance(output, Variable):
+            new_output = [output]
+        else:
+            raise ValueError("Unrecognized outpout.")
+
+        if self._out_dims_mappings:
+            assert len(new_output) == len(self._out_dims_mappings), \
+                "The length of dims_mapping {} does not matching the length output {}.".format(len(self._out_dims_mappings), len(new_output))
+        for i, item in enumerate(new_output):
+            if isinstance(item, Variable) and self._out_dims_mappings:
+                tensor_to_dims_mapping[item.name] = self._out_dims_mappings[i]
+
+        from .dist_context import get_default_distributed_context
         default_dist_ctx = get_default_distributed_context()
         for idx in range(op_size, new_op_size):
             op = cur_block.ops[idx]
-            dist_op = DistributedOperator(op, self._dist_attr)
-            dist_op.dist_attr.mark_annotated_as(self._dist_attr)
+            dist_op = DistributedOperator(op)
+            for name in dist_op.serial_op.input_arg_names:
+                if name in tensor_to_dims_mapping.keys():
+                    tensor = dist_op.get_serial_input(name)
+                    tensor_dist_attr = dist_op.dist_attr.get_input_dist_attr(
+                        name)
+                    dims_mapping = tensor_to_dims_mapping[name]
+                    if tensor is None:
+                        tensor_shape = []
+                    else:
+                        if tensor.type == core.VarDesc.VarType.READER \
+                            or tensor.type == core.VarDesc.VarType.LOD_TENSOR_ARRAY \
+                            or tensor.type == core.VarDesc.VarType.STEP_SCOPES:
+                            tensor_shape = []
+                        else:
+                            tensor_shape = tensor.shape
+                    if dims_mapping is not None:
+                        dims_mapping = tensor_to_dims_mapping[name]
+                        shard_spec = convert_to_shard_spec(
+                            dims_mapping, self._process_mesh)
+                        assert verify_shard_spec(shard_spec, tensor_shape, self._process_mesh), \
+                            "For tensor {}, shard_spec {} is invalid with tensor_shape {} and process_mesh {}.".format(
+                                name, shard_spec, tensor_shape, self._process_mesh)
+                        tensor_dist_attr.dims_mapping = dims_mapping
+                        tensor_dist_attr.mark_annotated("dims_mapping")
+            for name in dist_op.serial_op.output_arg_names:
+                if name in tensor_to_dims_mapping.keys():
+                    tensor = dist_op.get_serial_output(name)
+                    tensor_dist_attr = dist_op.dist_attr.get_output_dist_attr(
+                        name)
+                    dims_mapping = tensor_to_dims_mapping[name]
+                    if tensor is None:
+                        tensor_shape = []
+                    else:
+                        if tensor.type == core.VarDesc.VarType.READER \
+                            or tensor.type == core.VarDesc.VarType.LOD_TENSOR_ARRAY \
+                            or tensor.type == core.VarDesc.VarType.STEP_SCOPES:
+                            tensor_shape = []
+                        else:
+                            tensor_shape = tensor.shape
+                    if dims_mapping is not None:
+                        dims_mapping = tensor_to_dims_mapping[name]
+                        shard_spec = convert_to_shard_spec(
+                            dims_mapping, self._process_mesh)
+                        assert verify_shard_spec(shard_spec, tensor_shape, self._process_mesh), \
+                            "For tensor {}, shard_spec {} is invalid with tensor_shape {} and process_mesh {}.".format(
+                                name, shard_spec, tensor_shape, self._process_mesh)
+                        tensor_dist_attr.dims_mapping = dims_mapping
+                        tensor_dist_attr.mark_annotated("dims_mapping")
+            dist_op.dist_attr.process_mesh = self._process_mesh
+            if self._process_mesh is not None:
+                dist_op.dist_attr.mark_annotated("process_mesh")
             default_dist_ctx.add_dist_op_for_program(dist_op)
+
         return output