# 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 from .common import infer_shape from .common import DistributedOperatorImplContainer from .common import DistributedOperatorImpl from .common import register_distributed_operator_impl_container from .common import gradient_synchronization from .common import naive_copy_op_dist_attr_for_program, register_distributed_operator_impl, set_comm_op_dist_attr_for_program from ..utils import is_dim_shard from ..utils import is_dim_replicate from ..utils import compute_compatible_and_update_dim_mapping from ..dist_attribute import OperatorDistributedAttribute from paddle.fluid import core, unique_name from paddle.fluid.data_feeder import check_variable_and_dtype, check_dtype from paddle.distributed.fleet.meta_optimizers.common import OP_ROLE_KEY, OpRole from ..process_group import new_process_group from ..utils import _get_comm_group, _get_idx_in_axis, _get_corresponding_rank, set_var_dist_attr from ..cost import build_comp_desc_from_dist_op, build_comm_desc_from_dist_op from ..cost import build_comm_costs_from_descs, build_comp_costs_from_descs, build_dp_costs from ..cost import EmbeddingOpCost, EmbeddingGradOpCost from paddle.distributed.auto_parallel.cost.comm_op_cost import AllreduceSumOpCost, IdentityOpCost class DistributedEmbedding(DistributedOperatorImplContainer): def __init__(self, op_type): super(DistributedEmbedding, self).__init__(op_type) register_distributed_operator_impl_container( DistributedEmbedding("lookup_table_v2")) register_distributed_operator_impl_container( DistributedEmbedding("c_embedding")) register_distributed_operator_impl_container( DistributedEmbedding("lookup_table")) def adopt_lookup_table_v1(ctx, main_block, src_op, Ids_var): assert len( Ids_var.shape ) == 3, "input Ids to lookup_table should have 3 dimensions but got [{}] with shape [{}]".format( Ids_var.name, Ids_var.shape) if not Ids_var.stop_gradient: raise NotImplementedError( 'Requiring the gradient of Ids of lookup_table(v1）dist op is not currently supported. Please open an issue with details on your use case so that we can prioritize adding this (for instance, adversarial training for language model).' ) target_shape = list(Ids_var.shape[:-1]) intermediate_var_0 = main_block.create_var( name=unique_name.generate_with_ignorable_key(".".join( ["dist_reshape", 'tmp'])), dtype=Ids_var.dtype, shape=target_shape, type=core.VarDesc.VarType.LOD_TENSOR, persistable=False, stop_gradient=True) target_shape = [0] + list(Ids_var.shape[:-1]) xshape_var = main_block.create_var( name=unique_name.generate_with_ignorable_key(".".join( ["dist_Xshape", 'tmp'])), dtype=Ids_var.dtype, shape=target_shape, type=core.VarDesc.VarType.LOD_TENSOR, persistable=False, stop_gradient=True) # TODO use inplace reshape for memory saving reshape_op = main_block.append_op(type='reshape2', inputs={'X': [Ids_var]}, outputs={ 'Out': [intermediate_var_0], 'XShape': [xshape_var] }, attrs={ "shape": [0, -1], }) # set dist attr op_dist_attr = ctx.get_op_dist_attr_for_program(src_op) Ids_var_dist_attr = op_dist_attr.get_input_dist_attr(Ids_var.name) assert Ids_var_dist_attr is not None intermediate_var_0_dist_attr = set_var_dist_attr( ctx, intermediate_var_0, Ids_var_dist_attr.dims_mapping, Ids_var_dist_attr.process_mesh) set_var_dist_attr(ctx, xshape_var, [-1] + list(Ids_var_dist_attr.dims_mapping), Ids_var_dist_attr.process_mesh) op_dist_attr.del_input_dist_attr(Ids_var.name) op_dist_attr.set_input_dist_attr(intermediate_var_0.name, intermediate_var_0_dist_attr) new_op_dist_attr = OperatorDistributedAttribute() new_op_dist_attr.process_mesh = Ids_var_dist_attr.process_mesh new_op_dist_attr.impl_type = "default" new_op_dist_attr.impl_idx = 0 new_op_dist_attr.set_input_dims_mapping(Ids_var.name, Ids_var_dist_attr.dims_mapping) new_op_dist_attr.set_output_dims_mapping(intermediate_var_0.name, Ids_var_dist_attr.dims_mapping) new_op_dist_attr.set_output_dims_mapping( xshape_var.name, [-1] + list(Ids_var_dist_attr.dims_mapping)) ctx.set_op_dist_attr_for_program(reshape_op, new_op_dist_attr) return intermediate_var_0 # RowParallel class DistributedEmbeddingImpl(DistributedOperatorImpl): def __init__(self, name): super(DistributedEmbeddingImpl, self).__init__(name) self._forward_implemented = True self._backward_implemented = True def calc_cost(self, op_role, dist_op, ctx, cluster): """Calculate the cost by the op role.""" cost = None if int(op_role) == int(OpRole.Forward): cost = self.calc_fwd_cost(dist_op, ctx, cluster) elif int(op_role) == int(OpRole.Backward): cost = self.calc_bwd_cost(dist_op, ctx, cluster) assert cost is not None return cost def calc_fwd_cost(self, dist_op, ctx, cluster): # calc comp op cost desc_mapping = build_comp_desc_from_dist_op(dist_op=dist_op, dist_context=ctx) processes = dist_op.dist_attr.process_mesh.processes # embedding need start_index cost_mapping = build_comp_costs_from_descs(EmbeddingOpCost, ctx, processes, desc_mapping, cluster) serial_op = dist_op.serial_op parallel_axis = dist_op.dist_attr.get_input_dims_mapping( serial_op.input("W")[0])[0] attrs = {"use_calc_stream": True, "use_model_parallel": True} var_names = serial_op.output("Out") c_allreduce_sum_desc_mapping = build_comm_desc_from_dist_op( "c_allreduce_sum", dist_op, ctx, var_names, attrs=attrs, parallel_axis=parallel_axis) comm_op_cost_list = build_comm_costs_from_descs( AllreduceSumOpCost, ctx, processes, c_allreduce_sum_desc_mapping, cluster) res_cost = [cost_mapping, comm_op_cost_list] return res_cost def calc_bwd_cost(self, dist_op, ctx, cluster): # by now the backward function only insert the gradient allreduce for dist op itself res = [] backward_op = dist_op.serial_op main_block = backward_op.block dist_attr = dist_op.dist_attr embedding_row_dim_mapping = dist_attr.get_input_dims_mapping( backward_op.input("W")[0])[0] parallel_axis = embedding_row_dim_mapping attrs = {"use_calc_stream": True, "use_model_parallel": True} var_names = [backward_op.input("Out@GRAD")[0]] c_identity_desc_mapping = build_comm_desc_from_dist_op( "c_identity", dist_op, ctx, var_names, attrs=attrs, parallel_axis=parallel_axis) process_mesh = dist_attr.process_mesh processes = process_mesh.processes comm_op_cost_list = build_comm_costs_from_descs( IdentityOpCost, ctx, processes, c_identity_desc_mapping, cluster) res.append(comm_op_cost_list) # calc comp op cost desc_mapping = build_comp_desc_from_dist_op(dist_op=dist_op, dist_context=ctx) cost_mapping = build_comp_costs_from_descs(EmbeddingGradOpCost, ctx, processes, desc_mapping, cluster) res.append(cost_mapping) # need gradient allreduce var_dim_mapping = dist_attr.get_input_dims_mapping( backward_op.input("Ids")[0]) mesh_shape = process_mesh.topology batch_size_axis = var_dim_mapping[0] if batch_size_axis > -1 and mesh_shape[batch_size_axis] > 1: parallel_axis = batch_size_axis attrs = {"use_calc_stream": True} var_names = [backward_op.output('W@GRAD')[0]] build_dp_costs(res, dist_op, ctx, var_names, attrs, parallel_axis, cluster) return res def is_input_compatible(self, dist_op): op_desc = dist_op.serial_op.desc op_dist_attr = dist_op.dist_attr ids_name = op_desc.input('Ids')[0] w_name = op_desc.input('W')[0] ids_dims_mapping = op_dist_attr.get_input_dims_mapping(ids_name) w_dims_mapping = op_dist_attr.get_input_dims_mapping(w_name) if is_dim_replicate(w_dims_mapping[-2]) or is_dim_shard( w_dims_mapping[-1]): return False # Other dimensions must be replicate except the batch dimension for mapping in ids_dims_mapping[1:]: if is_dim_shard(mapping): return False return True def is_output_compatible(self, dist_op): op_desc = dist_op.serial_op.desc op_dist_attr = dist_op.dist_attr out_name = op_desc.output('Out')[0] out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name) # Other dimensions must be replicate except the batch dimension for mapping in out_dims_mapping[1:]: if is_dim_shard(mapping): return False 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 ids_name = op_desc.input('Ids')[0] w_name = op_desc.input('W')[0] out_name = op_desc.output('Out')[0] out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name) ids_dims_mapping = op_dist_attr.get_input_dims_mapping(ids_name) w_dims_mapping = op_dist_attr.get_input_dims_mapping(w_name) if ids_dims_mapping != out_dims_mapping[:len(ids_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 ids_name = op_desc.input('Ids')[0] w_name = op_desc.input('W')[0] out_name = op_desc.output('Out')[0] ids_dims_mapping = op_dist_attr.get_input_dims_mapping(ids_name) w_dims_mapping = op_dist_attr.get_input_dims_mapping(w_name) out_dims_mapping = op_dist_attr.get_output_dims_mapping(out_name) for i in range(len(ids_dims_mapping)): dim_changed = compute_compatible_and_update_dim_mapping( [ids_dims_mapping, out_dims_mapping], [i, i]) if dim_changed: changed = True dim_changed = compute_compatible_and_update_dim_mapping( [w_dims_mapping, out_dims_mapping], [-1, -1]) if dim_changed: changed = True return changed @staticmethod def forward(ctx, *args, **kwargs): """ kwargs: inputname_mapping & outputname_mapping """ dist_op_context = ctx.dist_op_context main_block = dist_op_context.work_block startup_block = dist_op_context.startup_block src_op = dist_op_context.cur_src_op rank_id = dist_op_context.rank_id op_dist_attr = ctx.get_op_dist_attr_for_program(src_op) assert op_dist_attr is not None, "backward op [{}] don't have dist attribute !".format( str(src_op)) # check validation of inputs / outputs assert 'Ids' in kwargs, "input [{}] is not given".format('Ids') assert 'W' in kwargs, "input [{}] is not given".format('W') assert 'Out' in kwargs, "output [{}] is not given".format('Out') assert len( kwargs['Ids'] ) == 1, "row_parallel_embedding input Ids take 1 variable but got {}".format( kwargs['Ids']) assert len( kwargs['W'] ) == 1, "row_parallel_embedding input W take 1 variable but got {}".format( kwargs['W']) assert len( kwargs['Out'] ) == 1, "row_parallel_embedding output Out take 1 variable but got {}".format( kwargs['Out']) Ids_var = main_block.var(kwargs['Ids'][0]) Weight_var = main_block._var_recursive(kwargs['W'][0]) Out_var = main_block.var(kwargs['Out'][0]) # support lookup_table_v1 if src_op.type == 'lookup_table': Ids_var = adopt_lookup_table_v1(ctx, main_block, src_op, Ids_var) # got dist attribute info embedding_row_dim_mapping = op_dist_attr.get_input_dims_mapping( Weight_var.name)[0] assert embedding_row_dim_mapping >= 0, "row_parallel_embedding's row should be divided by a specific mesh axis, but got [{}]".format( embedding_row_dim_mapping) process_mesh_shape = op_dist_attr.process_mesh.topology process_mesh_group = op_dist_attr.process_mesh.processes # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism if rank_id not in process_mesh_group: rank_id = _get_corresponding_rank(ctx, op_dist_attr.process_mesh, rank_id) # A generalized method to caculate embedding offset using cartisian product relative_idx = _get_idx_in_axis(process_mesh_group, process_mesh_shape, embedding_row_dim_mapping, rank_id) per_part_size = Weight_var.shape[0] relative_idx = relative_idx * per_part_size # TODO caculate ring id parallel_axis = embedding_row_dim_mapping group_ranks = _get_comm_group(process_mesh_group, process_mesh_shape, parallel_axis, rank_id) group = new_process_group(group_ranks) # append op check_variable_and_dtype(Ids_var, 'input', ['int32', 'int64'], 'c_embedding') # infer new var shape with op dist attr out_tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(Out_var) assert out_tensor_dist_attr is not None out_var_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name) assert out_var_dist_attr is not None ref_shape = infer_shape(main_block, Out_var, out_tensor_dist_attr, out_var_dist_attr) intermediate_var_0 = main_block.create_var( name=unique_name.generate_with_ignorable_key(".".join( ["c_embedding", 'tmp'])), dtype=Weight_var.dtype, shape=Out_var.shape, type=core.VarDesc.VarType.LOD_TENSOR, persistable=False, stop_gradient=Out_var.stop_gradient) # set intermediate_var_0's dist_attr with Out_var's dist_attr ctx.set_tensor_dist_attr_for_program(intermediate_var_0, out_var_dist_attr) check_variable_and_dtype( Out_var, 'tensor', ['float16', 'float32', 'float64', 'int32', 'int64'], 'c_allreduce_sum') c_embedding_op = main_block.append_op( type='c_embedding', inputs={ 'Ids': [Ids_var], 'W': [Weight_var] }, outputs={'Out': [intermediate_var_0]}, attrs={ "start_index": relative_idx, OP_ROLE_KEY: src_op.attr('op_role') }) if intermediate_var_0.shape != ref_shape: intermediate_var_0.desc.set_shape(ref_shape) # use_model_parallel c_allreduce_sum_op = main_block.append_op( type='c_allreduce_sum', inputs={'X': [intermediate_var_0]}, outputs={'Out': [Out_var]}, attrs={ 'ring_id': group.id, 'use_calc_stream': True, 'use_model_parallel': True, OP_ROLE_KEY: src_op.attr('op_role') }) if Out_var.shape != ref_shape: Out_var.desc.set_shape(ref_shape) # set dist op's dist_attr with serial op's dist_attr # matmulv2 embedding_op_dist_attr = OperatorDistributedAttribute() embedding_op_dist_attr.process_mesh = op_dist_attr.process_mesh embedding_op_dist_attr.impl_type = op_dist_attr.impl_type embedding_op_dist_attr.impl_idx = op_dist_attr.impl_idx for input_varname in c_embedding_op.desc.input_arg_names(): input_dist_attr = op_dist_attr.get_input_dist_attr(input_varname) assert input_dist_attr is not None, "dist_attr is {}".format( op_dist_attr) embedding_op_dist_attr.set_input_dist_attr(input_varname, input_dist_attr) output_varname = c_embedding_op.desc.output_arg_names()[0] output_dist_attr = op_dist_attr.get_output_dist_attr(Out_var.name) assert output_dist_attr is not None, "dist_attr is {}".format( op_dist_attr) embedding_op_dist_attr.set_output_dist_attr(output_varname, output_dist_attr) ctx.set_op_dist_attr_for_program(c_embedding_op, embedding_op_dist_attr) # allreduce allreduce_op_dist_attr = OperatorDistributedAttribute() allreduce_op_dist_attr.process_mesh = op_dist_attr.process_mesh allreduce_op_dist_attr.impl_type = op_dist_attr.impl_type allreduce_op_dist_attr.impl_idx = op_dist_attr.impl_idx for input_varname in c_allreduce_sum_op.desc.input_arg_names(): input_var = main_block.var(input_varname) tensor_dist_attr = ctx.get_tensor_dist_attr_for_program(input_var) assert tensor_dist_attr is not None allreduce_op_dist_attr.set_input_dist_attr(input_varname, tensor_dist_attr) for output_varname in c_allreduce_sum_op.desc.output_arg_names(): output_dist_attr = op_dist_attr.get_output_dist_attr(output_varname) assert output_dist_attr is not None, "dist_attr is {}".format( op_dist_attr) allreduce_op_dist_attr.set_output_dist_attr(output_varname, output_dist_attr) ctx.set_op_dist_attr_for_program(c_allreduce_sum_op, allreduce_op_dist_attr) # param initialization sync if Weight_var.is_parameter and not op_dist_attr.is_recompute: if Weight_var.name in dist_op_context.already_init_sync_vars: return dist_op_context.already_init_sync_vars.add(Weight_var.name) param = startup_block.var(Weight_var.name) param_dist_attr = ctx.get_tensor_dist_attr_for_program(param) process_mesh = param_dist_attr.process_mesh dim_mapping = param_dist_attr.dims_mapping # NOTE all not splited axis should be presented in mesh for axis, size in enumerate(process_mesh.topology): if size <= 1 or axis in dim_mapping: pass else: group_ranks = _get_comm_group(process_mesh.processes, process_mesh.topology, axis, rank_id) sync_group = new_process_group(group_ranks) startup_block.append_op(type='c_broadcast', inputs={'X': param}, outputs={'Out': param}, attrs={ 'ring_id': sync_group.id, 'root': 0, 'use_calc_stream': True, OP_ROLE_KEY: OpRole.Forward }) @staticmethod def backward(ctx, *args, **kwargs): # by now the backward function only insert the gradient allreduce for dist op itself dist_op_context = ctx.dist_op_context main_block = dist_op_context.work_block backward_op = dist_op_context.cur_src_op rank_id = dist_op_context.rank_id dist_attr = ctx.get_op_dist_attr_for_program(backward_op) assert dist_attr is not None, "backward op [{}] don't have dist attribute !".format( str(backward_op)) # FIXME (JZ-LIANG) Remove this hack to support any op mesh group for Pipeline Parallelism if rank_id not in dist_attr.process_mesh.processes: rank_id = _get_corresponding_rank(ctx, dist_attr.process_mesh, rank_id) assert 'Ids' in kwargs, "input [{}] is not given".format('Ids') assert 'W' in kwargs, "input [{}] is not given".format('W') assert 'Out@GRAD' in kwargs, "input [{}] is not given".format('Out') assert 'W@GRAD' in kwargs, "output [{}] is not given".format('W@GRAD') assert len( kwargs['Ids'] ) == 1, "row_parallel_embedding input Ids take 1 variable but got {}".format( kwargs['Ids']) assert len( kwargs['W'] ) == 1, "row_parallel_embedding input Ids take 1 variable but got {}".format( kwargs['W']) assert len( kwargs['Out@GRAD'] ) == 1, "row_parallel_embedding input Ids take 1 variable but got {}".format( kwargs['Out']) assert len( kwargs['W@GRAD'] ) == 1, "row_parallel_embedding output Ids take 1 variable but got {}".format( kwargs['W@GRAD']) Ids_var = main_block.var(kwargs['Ids'][0]) Weight_var = main_block.var(kwargs['W'][0]) Out_grad = main_block.var(kwargs['Out@GRAD'][0]) Weight_grad = main_block.var(kwargs['W@GRAD'][0]) embedding_row_dim_mapping = dist_attr.get_input_dims_mapping( Weight_var.name)[0] assert embedding_row_dim_mapping >= 0, "row_parallel_embedding's row should be divided by a specific mesh axis, but got [{}]".format( embedding_row_dim_mapping) process_mesh_shape = dist_attr.process_mesh.topology process_mesh_group = dist_attr.process_mesh.processes # A generalized method to caculate embedding offset using cartisian product relative_idx = _get_idx_in_axis(process_mesh_group, process_mesh_shape, embedding_row_dim_mapping, rank_id) per_part_size = Weight_var.shape[0] relative_idx = relative_idx * per_part_size check_variable_and_dtype( Out_grad, 'tensor', ['float16', 'float32', 'float64', 'int32', 'int64'], '_c_identity') intermediate_var_0 = main_block.create_var( name=unique_name.generate_with_ignorable_key(".".join( ["c_embedding", '@tmp_0@GRAD'])), dtype=Out_grad.dtype, shape=Out_grad.shape, type=core.VarDesc.VarType.LOD_TENSOR, persistable=False, stop_gradient=Out_grad.stop_gradient) # copy X_var's dist_attr to intermediate_var_0's dist_attr out_grad_dist_attr = dist_attr.get_input_dist_attr(Out_grad.name) assert out_grad_dist_attr is not None ctx.set_tensor_dist_attr_for_program(intermediate_var_0, out_grad_dist_attr) group_ranks = _get_comm_group(process_mesh_group, process_mesh_shape, embedding_row_dim_mapping, rank_id) group = new_process_group(group_ranks) c_identity_op = main_block.append_op( type='c_identity', inputs={'X': [Out_grad]}, outputs={'Out': intermediate_var_0}, attrs={ 'ring_id': group.id, 'use_calc_stream': True, 'use_model_parallel': True, OP_ROLE_KEY: OpRole.Backward, }) check_variable_and_dtype(intermediate_var_0, 'x', ['float16', 'float32', 'float64'], 'linear') check_dtype(intermediate_var_0.dtype, 'dtype', ['float16', 'float32', 'float64'], 'linear') set_comm_op_dist_attr_for_program(c_identity_op, dist_attr.process_mesh, out_grad_dist_attr, ctx) c_embedding_grad_op_desc = main_block.append_op(type='nop').desc c_embedding_grad_op_desc.set_type("c_embedding_grad") c_embedding_grad_op_desc.set_input('Ids', [Ids_var.name]) c_embedding_grad_op_desc.set_input('W', [Weight_var.name]) c_embedding_grad_op_desc.set_input('Out@GRAD', [intermediate_var_0.name]) c_embedding_grad_op_desc.set_output('W@GRAD', [Weight_grad.name]) c_embedding_grad_op_desc._set_attr('start_index', relative_idx) c_embedding_grad_op_desc._set_attr(OP_ROLE_KEY, OpRole.Backward) c_embedding_grad_op = main_block.ops[-1] assert c_embedding_grad_op.type == "c_embedding_grad" naive_copy_op_dist_attr_for_program(c_embedding_grad_op, backward_op, ctx) # data parallel gradient synchronization act_grad_names = [Ids_var.name] out_grad_names = [kwargs['W@GRAD'][0]] gradient_synchronization(ctx, backward_op, act_grad_names, out_grad_names, rank_id) register_distributed_operator_impl("lookup_table_v2", DistributedEmbeddingImpl("row_parallel")) register_distributed_operator_impl("c_embedding", DistributedEmbeddingImpl("row_parallel")) register_distributed_operator_impl("lookup_table", DistributedEmbeddingImpl("row_parallel"))