chore: import upstream snapshot with attribution

python/paddle/distributed/transpiler/details/__init__.py

@@ -0,0 +1,25 @@
+# Copyright (c) 2018 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 .program_utils import (  # noqa: F401
+    delete_ops,
+    find_op_by_input_arg,
+    find_op_by_output_arg,
+)
+from .ufind import UnionFind  # noqa: F401
+from .vars_distributed import (  # noqa: F401
+    VarDistributed,
+    VarsDistributed,
+    VarStruct,
+)

python/paddle/distributed/transpiler/details/program_utils.py

@@ -0,0 +1,44 @@
+# Copyright (c) 2018 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.
+
+
+def delete_ops(block, ops):
+    for op in ops:
+        try:
+            idx = list(block.ops).index(op)
+            block._remove_op(idx)
+        except Exception as e:
+            print(e)
+
+
+def find_op_by_input_arg(block, arg_name):
+    for index, op in enumerate(block.ops):
+        if arg_name in op.input_arg_names:
+            return index
+    return -1
+
+
+def find_op_by_output_arg(block, arg_name, reverse=False):
+    if reverse:
+        pos = len(block.ops) - 1
+        while pos >= 0:
+            op = block.ops[pos]
+            if arg_name in op.output_arg_names:
+                return pos
+            pos -= 1
+    else:
+        for index, op in enumerate(block.ops):
+            if arg_name in op.output_arg_names:
+                return index
+    return -1

python/paddle/distributed/transpiler/details/ufind.py

@@ -0,0 +1,64 @@
+# Copyright (c) 2018 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.
+
+
+class UnionFind:
+    """Union-find data structure.
+
+    Union-find is a data structure that keeps track of a set of elements partitioned
+    into a number of disjoint (non-overlapping) subsets.
+
+    Reference:
+    https://en.wikipedia.org/wiki/Disjoint-set_data_structure
+
+    Args:
+      elements(list): The initialize element list.
+    """
+
+    def __init__(self, elements=None):
+        self._parents = []  # index -> parent index
+        self._index = {}  # element -> index
+        self._curr_idx = 0
+        if not elements:
+            elements = []
+        for ele in elements:
+            self._parents.append(self._curr_idx)
+            self._index.update({ele: self._curr_idx})
+            self._curr_idx += 1
+
+    def find(self, x):
+        # Find the root index of given element x,
+        # execute the path compress while finding the root index
+        if x not in self._index:
+            return -1
+        idx = self._index[x]
+        while idx != self._parents[idx]:
+            t = self._parents[idx]
+            self._parents[idx] = self._parents[t]
+            idx = t
+        return idx
+
+    def union(self, x, y):
+        # Union two given element
+        x_root = self.find(x)
+        y_root = self.find(y)
+
+        if x_root == y_root:
+            return
+        self._parents[x_root] = y_root
+
+    def is_connected(self, x, y):
+        # If two given elements have the same root index,
+        # then they are connected.
+        return self.find(x) == self.find(y)

python/paddle/distributed/transpiler/details/vars_distributed.py

@@ -0,0 +1,289 @@
+# Copyright (c) 2018 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 paddle.static import Variable
+
+
+class VarStruct:
+    """
+    record part properties of a Variable in python.
+    """
+
+    def __init__(self, name, shape, dtype, type, lod_level, persistable):
+        self.name = name
+        self.shape = shape
+        self.dtype = dtype
+        self.type = type
+        self.lod_level = lod_level
+        self.persistable = persistable
+
+
+class VarDistributed:
+    """
+    a class to record the var distributed on parameter servers.
+    the class will record the relationship between origin var and slice var.
+    the slice var's properties, such as type/shape/offset/endpoint.
+    """
+
+    def __init__(
+        self,
+        origin_var,
+        slice_var,
+        is_slice=None,
+        block_id=None,
+        offset=None,
+        vtype=None,
+        endpoint=None,
+    ):
+        """
+        Args:
+            origin_var(Variable|VarStruct): origin var properties
+            slice_var(Variable|VarStruct): slice var properties
+            is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard.
+            block_id(int|None): the number about the slice var.
+            offset(int|None): if the slice var is sliced, offset is the numel before the var.
+            vtype(str|None): a tag, such as Optimizer/Param/RemotePrefetch.
+            endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001"
+        """
+
+        if isinstance(origin_var, Variable):
+            self.origin = self.__create_var_struct(origin_var)
+        else:
+            self.origin = origin_var
+
+        if isinstance(slice_var, Variable):
+            self.slice = self.__create_var_struct(slice_var)
+        else:
+            self.slice = slice_var
+
+        if self.equal(self.origin, self.slice):
+            self.is_slice = False
+            self.block_id = 0
+            self.offset = 0
+        else:
+            self.is_slice = True
+            self.block_id = 0
+            self.offset = 0
+
+        if is_slice is not None:
+            self.is_slice = is_slice
+        if block_id is not None:
+            self.block_id = block_id
+        if offset is not None:
+            self.offset = offset
+
+        self.vtype = vtype
+        self.endpoint = endpoint
+
+    @staticmethod
+    def __create_var_struct(var):
+        return VarStruct(
+            var.name,
+            var.shape,
+            var.dtype,
+            var.type,
+            var.lod_level,
+            var.persistable,
+        )
+
+    @staticmethod
+    def equal(var1, var2):
+        """
+        the two var is equal or not.
+        Returns:
+            bool: equal will return True else False
+        """
+        assert isinstance(var1, VarStruct) and isinstance(var2, VarStruct)
+
+        return (
+            var1.name == var2.name
+            and var1.type == var2.type
+            and var1.shape == var2.shape
+            and var1.dtype == var2.dtype
+            and var1.lod_level == var2.lod_level
+            and var1.persistable == var2.persistable
+        )
+
+    def __str__(self):
+        origin_var_str = f"{self.origin.name} : base.{self.origin.type}.shape{self.origin.shape}.astype({self.origin.dtype})"
+
+        slice_var_str = (
+            f"{self.slice.name} : base.{self.slice.type}.shape{self.slice.shape}.astype({self.slice.dtype})"
+            f".slice({self.is_slice}).block({self.block_id}).offset({self.offset})"
+        )
+
+        return f"var owned: {self.vtype}, origin var: ( {origin_var_str} ), slice var: ( {slice_var_str} ), endpoint: {self.endpoint} "
+
+
+class VarsDistributed:
+    """
+    a gather about VarDistributed with many methods to find distributed vars.
+    through the class, we can get overview about the distributed parameters on parameter servers.
+    this class may centralized and convenient for developer to manage and get variable's distribute.
+    other module can also use this to find variables such io.py.
+    """
+
+    def __init__(self):
+        self.distributed_vars = []
+
+    def add_distributed_var(
+        self,
+        origin_var,
+        slice_var,
+        is_slice=None,
+        block_id=None,
+        offset=None,
+        vtype=None,
+        endpoint=None,
+    ):
+        """
+        add distributed var in this.
+
+        Args:
+            origin_var(Variable|VarStruct): origin var properties
+            slice_var(Variable|VarStruct): slice var properties
+            is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard.
+            block_id(int|None): the number about the slice var.
+            offset(int|None): if the slice var is sliced, offset is the numel before the var.
+            vtype(str|None): a tag, such as Optimizer/Param/RemotePrefetch.
+            endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001"
+        Returns:
+            None
+        """
+        self.distributed_vars.append(
+            VarDistributed(
+                origin_var,
+                slice_var,
+                is_slice,
+                block_id,
+                offset,
+                vtype,
+                endpoint,
+            )
+        )
+
+    def get_distributed_var_by_slice(self, var_name):
+        """
+        get distributed var by conditions.
+
+        Args:
+            var_name(str): slice var name, such as "w.trainer0.block1"
+        Returns:
+            VarDistributed: distributed var.
+        """
+        for dist_var in self.distributed_vars:
+            if dist_var.slice.name == var_name:
+                return dist_var
+        return None
+
+    @staticmethod
+    def equal(var1, var2):
+        """
+        the two var is equal or not.
+        Returns:
+            bool: equal will return True else False
+        """
+        return (
+            var1.name == var2.name
+            and var1.type == var2.type
+            and var1.shape == var2.shape
+            and var1.dtype == var2.dtype
+            and var1.lod_level == var2.lod_level
+            and var1.persistable == var2.persistable
+        )
+
+    def get_distributed_var_by_origin_and_ep(self, origin_var_name, endpoint):
+        """
+        get distributed var by conditions.
+
+        Args:
+            origin_var_name(str):
+            endpoint(str): the parameter endpoint, such as "127.0.0.1:1001"
+        Returns:
+            VarDistributed: distributed var.
+        """
+        for dist_var in self.distributed_vars:
+            if (
+                dist_var.origin.name == origin_var_name
+                and dist_var.endpoint == endpoint
+            ):
+                return dist_var
+        return None
+
+    def get_distributed_vars_by_vtypes(self, vtypes, groupby=False):
+        """
+        get distributed vars by conditions.
+
+        Args:
+            vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch"
+            groupby(bool|False): group by origin var or not.
+
+        Returns:
+            list: distributed var list.
+            dict: distributed var map when groupby=True
+        """
+        vtype_vars = []
+        for var in self.distributed_vars:
+            if var.vtype in vtypes:
+                vtype_vars.append(var)
+        if not groupby:
+            return vtype_vars
+
+        params_map = {}
+        for var in vtype_vars:
+            origin_var_name = var.origin.name
+
+            if origin_var_name in params_map.keys():
+                optimizers = params_map.get(origin_var_name)
+            else:
+                optimizers = []
+            optimizers.append(var)
+            params_map[origin_var_name] = optimizers
+        return params_map
+
+    def get_distributed_vars_by_ep(self, endpoint, vtype=None):
+        """
+        get distributed vars by conditions.
+
+        Args:
+            endpoint(str): the parameter server endpoint, such as "127.0.0.1:2001"
+            vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch"
+
+        Returns:
+            list: distributed var list.
+        """
+        endpoint_vars = []
+        for var in self.distributed_vars:
+            if var.endpoint == endpoint:
+                endpoint_vars.append(var)
+        if not vtype:
+            return endpoint_vars
+
+        vtype_vars = []
+        for var in endpoint_vars:
+            if var.vtype == vtype:
+                vtype_vars.append(var)
+        return vtype_vars
+
+    def overview(self):
+        """
+        get the overview string about all params on all parameter servers.
+
+        Returns:
+            Str: overview string.
+
+        """
+        vars_str = []
+        for var in self.distributed_vars:
+            vars_str.append(str(var))
+        return "\n".join(vars_str)