# 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 logging from functools import reduce import paddle from paddle.distributed.auto_parallel.operators.common import ( ParallelMode, is_data_parallel_reduce_op, is_parameter_related, ) from paddle.distributed.auto_parallel.process_group import new_process_group from paddle.distributed.auto_parallel.utils import ( _get_comm_group, get_logger, get_var_numel, insert_dependencies_for_vars, is_backward_op, is_dep_skip_op, is_loss_grad_op, is_optimize_op, naive_set_dist_op_attr_for_program_by_mesh_and_mapping, set_var_dist_attr, ) from paddle.distributed.fleet.meta_optimizers.sharding.utils import get_var_size from paddle.fluid.executor import _is_enable_standalone_executor from paddle.framework import core from paddle.static import default_main_program, default_startup_program from paddle.utils import unique_name from .pass_base import PassBase, register_pass OpRole = core.op_proto_and_checker_maker.OpRole OP_ROLE_KEY = core.op_proto_and_checker_maker.kOpRoleAttrName() _skip_ops = [ 'create_py_reader', 'create_double_buffer_reader', 'read', 'slice', 'split', 'assign', "send_v2", ] # update here to support new optimizers _supported_optimizer_type = [ "adam", "adamax", "adamw", "decayed_adagrad", "momentum", "dgc_momentum", "lars_momentum", "merged_momentum", "lamb", "sgd", ] _logger = get_logger(logging.INFO) def _is_reshard_op(op): return op.desc.has_attr( "op_namescope" ) and "/auto_parallel/reshard" in op.desc.attr('op_namescope') # NOTE we add the "auto_parallel" prefix to the pass in order to # indicate that this pass should obey some constrains by auto_parallel # for example all ops and vars should has dist attr before and after pass # should use dist op instead of custom comm op @register_pass("auto_parallel_sharding") class ShardingPass(PassBase): def __init__(self): super().__init__() self.set_attr("dist_context", None) self.set_attr("stage", None) self.set_attr("sharding_degree", None) # for parallelizer self.set_attr("degree", None) # for parallelizer_v2 self.set_attr("enable_overlap", None) self.set_attr("param_comm_stream_num", None) self.set_attr("grad_comm_stream_num", None) self.set_attr("param_bucket_size_numel", None) self.set_attr("grad_bucket_size_numel", None) self.set_attr("partition_algor", None) self.set_attr("enable_hierarchical_comm", None) self.set_attr("params_grads", []) self.set_attr("global_rank", -1) self.dp_groups = set() self.sharding_infos = [] self.varname_to_sharding_info = {} self.sharding_hybrid_dp = False self.outer_dp_group = None self.shared_params_grads = [] def _check_self(self): if self.get_attr("dist_context") is None: return False if self.get_attr("stage") not in [1, 2, 3]: return False if self.get_attr("sharding_degree") is not None: if ( not isinstance(self.get_attr("sharding_degree"), int) ) or self.get_attr("sharding_degree") <= 1: return False elif self.get_attr("degree") is not None: if (not isinstance(self.get_attr("degree"), int)) or self.get_attr( "degree" ) <= 1: return False else: return False if len(self.get_attr("params_grads")) <= 0: return False if (not isinstance(self.get_attr("global_rank"), int)) or self.get_attr( "global_rank" ) < 0: return False if self.get_attr("enable_overlap") is None: return False if self.get_attr("param_comm_stream_num") is None: return False if self.get_attr("grad_comm_stream_num") is None: return False if self.get_attr("param_bucket_size_numel") is None: return False if self.get_attr("grad_bucket_size_numel") is None: return False if self.get_attr("partition_algor") is None: return False if self.get_attr("enable_hierarchical_comm") is None: return False return True def _check_conflict(self, other_pass): return True def _apply_single_impl(self, main_program, startup_program, context): self._dist_context = self.get_attr("dist_context") self.sharding_world_size = int( self.get_attr("sharding_degree") or self.get_attr("degree") ) self.stage = int(self.get_attr("stage")) self.global_rank = int(self.get_attr("global_rank")) self.enable_overlap = self.get_attr("enable_overlap") self.param_comm_stream_num = int(self.get_attr("param_comm_stream_num")) self.grad_comm_stream_num = int(self.get_attr("grad_comm_stream_num")) self.enable_hierarchical_comm = self.get_attr( "enable_hierarchical_comm" ) if self.param_comm_stream_num > 1 or self.grad_comm_stream_num > 1: assert ( self.enable_overlap ), "multiple comm stream need enable_overlap to be True" self.param_bucket_size_numel = int( self.get_attr("param_bucket_size_numel") ) self.grad_bucket_size_numel = int( self.get_attr("grad_bucket_size_numel") ) self.partition_algor = self.get_attr("partition_algor") params_grads = self.get_attr("params_grads") main_block, startup_block = ( main_program.global_block(), startup_program.global_block(), ) # NOTE Multi / Sub-Block Support # we assume that only parameter are present and partitioned in main_block, # there is NO new param in sub_block, and all params in sub_block follows the same # partition as main_block. the above contraint fullfill the 3 most common use-cases in Paddle sub_block: # 1. subblock for lr scheduler # 2. sub-block uses the same or partial network of main-block, e.g. GPT3 generation model # 3. sub-block used for double backward self._build_sharding_groups(main_block, params_grads) for block in main_program.blocks: self._shard_optimizer(block, startup_block, params_grads, context) self._shard_gradient_synchronization(block) self._shard_parameter(block, startup_block) context.set_attr("params_grads", self.shared_params_grads) self._optimization_pass(main_program, startup_program) def _build_sharding_groups(self, main_block, params_grads): self._collective_data_parallel_groups(main_block) self._build_sharding_infos(main_block, params_grads) def _collective_data_parallel_groups(self, main_block): for op in main_block.ops: if not _is_forward_op(op) or op.type in _skip_ops: continue # NOTE: there aren't dist_attr in the ops which reshard insert, # and should be skip in sharding. if _is_reshard_op(op): continue group = _inference_data_parallel_group_for_operator( self.global_rank, op, self._dist_context ) if group is not None: self.dp_groups.add(group) # TODO(JZ-LIANG) allow more than one dp groups in network, support more general distribution # genetated by auto search if len(self.dp_groups) != 1: raise NotImplementedError( "So far Only and Exactly one data parallel group in network are supported, but got [{}] different data parallel groups".format( len(self.dp_groups) ) ) def _build_sharding_infos(self, main_block, params_grads): # order params params_grads = re_order_program( main_block, params_grads, self._dist_context ) # partition for dp_group in self.dp_groups: assert ( dp_group.nranks >= self.sharding_world_size ), "sharding world size [{}] should not larger than dp world size [{}]".format( self.sharding_world_size, dp_group.nranks ) assert ( dp_group.nranks % self.sharding_world_size == 0 ), "sharding world size [{}] should be divisible by dp world size [{}]".format( self.sharding_world_size, dp_group.nranks ) assert ( self.global_rank in dp_group.ranks ), "current ranks [{}] does NOT belong to the data parallel group [{}]".format( self.global_rank, dp_group.ranks ) assert ( len(params_grads) >= self.sharding_world_size ), "number of parameters [{}] is not enough to be shard among [{}] ranks".format( len(params_grads), self.sharding_world_size ) # sharding hybrid data parallel: partial sharding param within if dp_group.nranks > self.sharding_world_size: self.sharding_hybrid_dp = True assert self.param_comm_stream_num < 2 assert self.grad_comm_stream_num < 2 assert ( len(self.dp_groups) == 1 ), "hybrid sharding and data parallelism are supported only when there is excatly one data parallel group in the network" outer_dp_group, sharding_group = _get_dp_and_sharding_groups( dp_group.ranks, self.sharding_world_size, self.global_rank ) sharding_group = new_process_group(sharding_group) self.outer_dp_group = new_process_group(outer_dp_group) else: sharding_group = dp_group self._dist_context._sharding_group = sharding_group # TODO(JZ-LIANG) when support multiple dp groups in future, should group param and bind them to corresponding dp group sharding_info = ShardingInfo( sharding_group, self.global_rank, params_grads, self.partition_algor, ) self.sharding_infos.append(sharding_info) for param in sharding_info.params: self.varname_to_sharding_info[param.name] = sharding_info def _shard_optimizer( self, main_block, startup_block, params_grads, pass_context ): """ sharding all optimizer related ops and vars, include: gradient clip ops & vars weight decay ops & vars optimizer ops and states """ self._shard_amp_related_op_and_vars(main_block, pass_context) self._shard_weight_decay(main_block) # self._shard_gradient_clip(main_block) self._shard_optimizer_ops_and_states(main_block, startup_block) self._insert_optimizer_broadcasts(main_block, startup_block) def _shard_amp_related_op_and_vars(self, main_block, pass_context): if self.stage < 2: return for idx, op in reversed(list(enumerate(main_block.ops))): # shard amp related param_grad cast if _is_param_grad_fp32_cast_op(main_block, op): output_name = op.output_arg_names[0] param_name = output_name[: output_name.find("@")] if not self._is_parameter_in_local_shard(param_name): main_block._remove_op(idx, sync=False) main_block._remove_var(output_name, sync=False) # shard check nan inf elif op.type in ["check_finite_and_unscale", "update_loss_scaling"]: reversed_x = [] for input_name in op.desc.input('X'): param_name = input_name[: input_name.find("@")] if self._is_parameter_in_local_shard(param_name): reversed_x.append(input_name) # NOTE: When `reversed_x` is [], check_finite_and_unscale will be replaced by `fill_constant` op. # The output of check_finite_and_unscale is be set False if reversed_x: op.desc.set_input('X', reversed_x) op.desc.set_output('Out', reversed_x) else: if op.type == "check_finite_and_unscale": op_role = op.attr('op_role') out_name = op.output_arg_names[0] out_var = main_block.vars[out_name] main_block._remove_op(idx, sync=False) main_block._insert_op_without_sync( idx, type="fill_constant", outputs={"Out": out_var}, attrs={ "shape": out_var.shape, "dtype": out_var.dtype, "value": 0, OP_ROLE_KEY: op_role, }, ) else: main_block._remove_op(idx, sync=False) main_block._sync_with_cpp() def _shard_gradient_clip(self, main_block): if self.stage < 2: return # TODO (JZ-LIANG) support calculate global norm with tensor parallelism removed_op_type = ['elementwise_mul', 'squared_l2_norm', 'clip_by_norm'] removed_op_idx = set() removed_tmp_var = set() for idx, op in list(enumerate(main_block.ops)): if not _is_gradient_clip_op(op): continue if op.type in removed_op_type: input_name = op.input("X")[0] param_name = input_name[: input_name.find("@GRAD")] if not self._is_parameter_in_local_shard(param_name): removed_op_idx.add(idx) if op.type in ['squared_l2_norm', 'clip_by_norm']: for output_name in op.output_arg_names: removed_tmp_var.add(output_name) for idx, op in reversed(list(enumerate(main_block.ops))): if not _is_gradient_clip_op(op): continue if idx in removed_op_idx: main_block._remove_op(idx, sync=False) for varname in removed_tmp_var: main_block._remove_var(varname, sync=False) for idx, op in list(enumerate(main_block.ops)): if not _is_gradient_clip_op(op): continue if op.type == 'sum': reserved_vars = [] for input_name in op.input_arg_names: if input_name not in removed_tmp_var: reserved_vars.append(input_name) op.desc.set_input("X", reserved_vars) sum_op_output = op.output_arg_names[0] for i, sharding_info in enumerate(self.sharding_infos): new_op = main_block._insert_op( idx + i + 1, type='c_allreduce_sum', inputs={'X': [sum_op_output]}, outputs={'Out': [sum_op_output]}, attrs={ 'ring_id': sharding_info.group.id, 'op_namescope': "/gradient_clip_model_parallelism", 'use_calc_stream': True, OP_ROLE_KEY: OpRole.Optimize, }, ) dist_attr = ( self._dist_context.get_tensor_dist_attr_for_program( main_block.var(sum_op_output) ) ) # assert dist_attr is not None # naive_set_dist_op_attr_for_program_by_mesh_and_mapping( # new_op, dist_attr.process_mesh, dist_attr.dims_mapping, # self._dist_context) break main_block._sync_with_cpp() def _shard_weight_decay(self, main_block): if self.stage < 2: return for idx, op in reversed(list(enumerate(main_block.ops))): if not _is_weight_decay_op(op): continue else: raise NotImplementedError( "weight decay is NOT supported by now" ) main_block._sync_with_cpp() def _shard_optimizer_ops_and_states(self, main_block, startup_block): should_removed_optimizer_states = [] for idx, op in reversed(list(enumerate(main_block.ops))): if not is_optimize_op(op): break if op.type in _supported_optimizer_type: assert "Param" in op.input_names assert len(op.input("Param")) == 1 param_name = op.input("Param")[0] if not self._is_parameter_in_local_shard(param_name): should_removed_optimizer_states.extend( [ varname for varname in op.output_arg_names if varname != param_name ] ) main_block._remove_op(idx, sync=False) else: self.shared_params_grads.append( self._get_param_grad(param_name) ) for idx, op in reversed(list(enumerate(startup_block.ops))): if ( len(op.output_arg_names) == 1 and op.output_arg_names[0] in should_removed_optimizer_states ): startup_block._remove_op(idx, sync=False) for varname in should_removed_optimizer_states: if main_block.has_var(varname): main_block._remove_var(varname, sync=False) if startup_block.has_var(varname): startup_block._remove_var(varname, sync=False) main_block._sync_with_cpp() startup_block._sync_with_cpp() def _insert_optimizer_broadcasts(self, main_block, startup_block): if self.stage > 2 or self.param_bucket_size_numel > 1: return for sharding_info in self.sharding_infos: for param in sharding_info.params: assert main_block.has_var(param.name) assert startup_block.has_var(param.name) new_op = main_block.append_op( type='c_broadcast', inputs={'X': param}, outputs={'Out': param}, attrs={ 'ring_id': sharding_info.group.id, 'root': sharding_info.get_var_rank(param.name), 'use_calc_stream': True, OP_ROLE_KEY: OpRole.Optimize, }, ) new_op._set_attr( 'op_namescope', str('/') + ParallelMode.DataParallel ) param_dist_attr = ( self._dist_context.get_tensor_dist_attr_for_program(param) ) assert param_dist_attr is not None naive_set_dist_op_attr_for_program_by_mesh_and_mapping( new_op, param_dist_attr.process_mesh, param_dist_attr.dims_mapping, self._dist_context, ) main_block._sync_with_cpp() def _is_parameter_in_local_shard(self, param_name): assert param_name in self.varname_to_sharding_info sharding_info = self.varname_to_sharding_info[param_name] return sharding_info.is_in_local_shard(param_name) def _get_param_grad(self, param_name): assert param_name in self.varname_to_sharding_info sharding_info = self.varname_to_sharding_info[param_name] p_g = sharding_info.get_param_grad(param_name) assert p_g is not None return p_g def _shard_gradient_synchronization(self, main_block): if self.stage < 2: return dp_ring_ids = [group.id for group in self.dp_groups] for idx, op in reversed(list(enumerate(main_block.ops))): if _is_param_grad_allreduce_op(op, main_block): input_name = op.input_arg_names[0] base_name = _get_base_name_from_grad_name(input_name) sharding_info = self.varname_to_sharding_info[base_name] reduce_op = _insert_reduce_op( main_block, idx, input_name, sharding_info.group.id, sharding_info.get_var_rank(base_name), self._dist_context, ) if ( not self.sharding_hybrid_dp or not sharding_info.is_in_local_shard(base_name) ): main_block._remove_op(idx + 1, sync=False) else: op._set_attr("ring_id", self.outer_dp_group.id) op._set_attr( 'op_namescope', str('/') + ParallelMode.DataParallel ) # NOTE: # var@GRAD = sum(var@GRAD@RENAME@0, var@GRAD@RENAME@1) # If the var is not in local rank and it is output of many ops, or the var is renamed in another words, # the sum op should be removed. if _is_param_grad_sum_op(op, main_block): out_name = op.output_arg_names[0] base_name = _get_base_name_from_grad_name(out_name) sharding_info = self.varname_to_sharding_info[base_name] if not sharding_info.is_in_local_shard(base_name): main_block._remove_op(idx, sync=False) main_block._sync_with_cpp() def _shard_parameter(self, main_block, startup_block): if self.stage < 3: return dp_ring_ids = [group.id for group in self.dp_groups] for sharding_info in self.sharding_infos: ( need_broadcast_vars, param_usage, ) = sharding_info.get_broadcast_vars_and_param_usage(main_block) not_used_param_nane = [] for param_name in param_usage: if ( param_usage[param_name] == 0 and sharding_info.get_var_rank(param_name) != sharding_info.local_rank ): not_used_param_nane.append(param_name) for idx, op in reversed(list(enumerate(main_block.ops))): if is_optimize_op(op): continue for input_name in op.input_arg_names: # NOTE hack for embedding op when AMP 02-3 # paddle amp force embedding (lookup table) to be run on fp32 if _is_param_fp16_cast_op( main_block, op, sharding_info.param_names ): continue if input_name not in need_broadcast_vars: continue root_rank = sharding_info.get_var_rank(input_name) if root_rank == sharding_info.local_rank: broadcast_varname = input_name else: broadcast_varname = unique_name.generate( input_name + "@BroadCast" ) input_var = main_block.var(input_name) new_var = main_block.create_var( name=broadcast_varname, shape=input_var.shape, dtype=input_var.dtype, persistable=False, ) ref_dist_attr = ( self._dist_context.get_tensor_dist_attr_for_program( input_var ) ) out_var_dist_attr = set_var_dist_attr( self._dist_context, new_var, ref_dist_attr.dims_mapping, ref_dist_attr.process_mesh, ) op._rename_input(input_name, broadcast_varname) _insert_init_and_broadcast_op( main_block, idx, broadcast_varname, sharding_info.local_rank, root_rank, sharding_info.group.id, op.attr('op_role'), self._dist_context, ) for idx, op in reversed(list(enumerate(main_block.ops))): if op.type != "cast": continue input_name = op.input_arg_names[0] output_name = op.output_arg_names[0] if input_name in not_used_param_nane: main_block._remove_op(idx, sync=False) main_block._remove_var(output_name, sync=False) for idx, op in reversed(list(enumerate(startup_block.ops))): assert len(op.output_arg_names) == 1 output_name = op.output_arg_names[0] if ( op.type == "c_broadcast" and op.attr("ring_id") in dp_ring_ids ): if ( self.outer_dp_group and sharding_info.get_var_rank(output_name) == sharding_info.local_rank ): op._set_attr("ring_id", self.outer_dp_group.id) else: startup_block._remove_op(idx, sync=False) continue if ( op.type != "c_broadcast" and output_name in param_usage and sharding_info.get_var_rank(output_name) != sharding_info.local_rank ): startup_block._remove_op(idx, sync=False) for param_name in param_usage: if ( sharding_info.get_var_rank(param_name) != sharding_info.local_rank ): main_block._remove_var(param_name, sync=False) startup_block._remove_var(param_name, sync=False) main_block._sync_with_cpp() startup_block._sync_with_cpp() def _optimization_pass(self, main_program, startup_program): if self.stage <= 1: return self.grad_coalesce_prefix = 'sharding_coalesce_grad_' self.param_coalesce_prefix = 'sharding_coalesce_param_' # NOTE PR#49275 for detail self.comm_op_scheduling_priority = -1 # TODO support multiple sub_blocks assert ( len(self.sharding_infos) == 1 ), "gradient synchronization optimization only support one sharding group right now, but got [{}].".format( len(self.sharding_infos) ) sharding_info = self.sharding_infos[0] with paddle.static.program_guard(main_program, startup_program): self._gradient_sync_optimization(sharding_info) # TODO independent the logic of fuse and overlap # support overlap when no fuse if self.param_bucket_size_numel > 1: if self.stage == 2: self._fuse_overlap_parameter_comm_stage_two(sharding_info) elif self.stage == 3: self._fuse_overlap_parameter_comm_stage_three(sharding_info) def _gradient_sync_optimization(self, sharding_info): if self.grad_bucket_size_numel <= 1 and (not self.enable_overlap): return main_block = default_main_program().global_block() startup_block = default_startup_program().global_block() coalesce_to_group_map, grad_name_to_group_map = self._group_grads( main_block, sharding_info, ) self._overlap_grad_comm( main_block, sharding_info, coalesce_to_group_map, grad_name_to_group_map, ) def _fuse_overlap_parameter_comm_stage_two(self, sharding_info): main_block = default_main_program().global_block() startup_block = default_startup_program().global_block() group_to_param_map, param_to_group_map = group_param( sharding_info, self.param_bucket_size_numel ) _logger.info("Sharding Stage2 Optimization:") _logger.info( "Param Bucket size is [{}], [{}] Parameters are fused into [{}] Buckets".format( self.param_bucket_size_numel, len(param_to_group_map.keys()), len(group_to_param_map.keys()), ) ) broadcast_var_to_group_map = {} if self.enable_overlap: # if the communication is cross node, comm will be slow and calc will therefore # wait for comm. enable multi-comm-stream # TODO revise me in future # 1. manager the comm and corresponding stream # 2. allow more than two streams and open to be config self.param_comm_group_stream_pairs = [] ranks = sharding_info.group.ranks for i in range(self.param_comm_stream_num): if i == 0: group = sharding_info.group else: group = new_process_group(ranks, force_new_group=True) # NOTE here stream is just a presentation with different name, # it is up to executor to create the exact streams given the name. stream = "sharding_param_comm_stream{}".format(i) self.param_comm_group_stream_pairs.append( { "comm_group": group, "comm_stream": stream, } ) _logger.info( "Parameter Communication would use [{}] streams.".format( self.param_comm_stream_num ) ) self.op_to_stream_idx = {} for i, param_group in enumerate(group_to_param_map.keys()): assert len(param_group) >= 1 if len(param_group) > 1: coalesce_var_name = unique_name.generate( self.param_coalesce_prefix + str(i) ) startup_block.create_var( name=coalesce_var_name, dtype=param_group.dtype, persistable=True, stop_gradient=True, ) param_group.coalesce_var = main_block.create_var( name=coalesce_var_name, dtype=param_group.dtype, persistable=True, stop_gradient=True, ) startup_block.append_op( type="coalesce_tensor", inputs={"Input": param_group.vars}, outputs={ "Output": param_group.vars, "FusedOutput": param_group.coalesce_var, }, attrs={ "copy_data": True, "use_align": True, "dtype": param_group.dtype, OP_ROLE_KEY: OpRole.Forward, }, ) else: param_group.coalesce_var = param_group.vars[0] _logger.info( "Bucket[{}] size [{}]MB.".format( i, sum([get_var_size(p) for p in param_group.vars]), ) ) _logger.debug( "Bucket[{}] parameters: {}.".format( i, [p.name for p in param_group.vars], ) ) broadcast_var_to_group_map[ param_group.coalesce_var.name ] = param_group # TODO revise me to manager stream and comm comm_stream_idx = i % self.param_comm_stream_num comm_group = self.param_comm_group_stream_pairs[comm_stream_idx][ 'comm_group' ] comm_stream = self.param_comm_group_stream_pairs[comm_stream_idx][ 'comm_stream' ] new_op = main_block.append_op( type='c_broadcast', inputs={'X': param_group.coalesce_var}, outputs={'Out': param_group.coalesce_var}, attrs={ 'ring_id': comm_group.id, 'root': param_group.rank, 'use_calc_stream': True, OP_ROLE_KEY: OpRole.Optimize, }, ) self.op_to_stream_idx[new_op] = comm_stream_idx new_op._set_attr( 'op_namescope', str('/') + ParallelMode.DataParallel ) if self.enable_overlap: new_op.dist_attr.execution_stream = comm_stream new_op.dist_attr.scheduling_priority = ( self.comm_op_scheduling_priority ) # NOTE the current dist context lack the presentation for bucket tensor which # composes many tensor with different dims_mapping. we DO NOT assign dist attr # for it currently. # add dependencies: # 1. all broadcast depend on its pre collective # 2. coalesce broadcast add nop to resolute data flow dependencies dep_map = {} for i, op in enumerate(main_block.ops): if is_sharding_param_broadcast_op(op): broadcast_varname = op.output("Out")[0] broadcast_var = main_block.vars[broadcast_varname] param_group = broadcast_var_to_group_map[broadcast_varname] comm_stream = None if self.enable_overlap: comm_stream = op.dist_attr.execution_stream # FIXME remove me when upgrade to multi-comm version if len(dep_map.keys()) < self.param_comm_stream_num: op = _get_broadcast_first_depend_op(main_block) prior_var = main_block.vars[op.output("ParamOut")[0]] else: pre_op = main_block.ops[i - self.param_comm_stream_num] assert is_sharding_param_broadcast_op( pre_op ), "Unexpected: sharding broadcast pre op should be broadcast." prior_var = main_block.vars[pre_op.output("Out")[0]] # broadcast order dependencies dep_map[i] = [(i, [prior_var], [broadcast_var], comm_stream)] if len(param_group.vars) > 1: # in shard coalesce depend to optimizer if param_group.is_in_local_shard: last_grad = param_group.vars[-1] dep_map[i].append( (i, [last_grad], [broadcast_var], comm_stream) ) # coalesce resolution post deps dep_map[i].append( (i + 1, [broadcast_var], param_group.vars, comm_stream) ) # insert deps indice = sorted(list(dep_map.keys()), reverse=True) for i in indice: for idx, prior_vars, post_vars, comm_stream in dep_map[i][::-1]: depend_op = insert_dependencies_for_vars( main_block, idx, prior_vars, post_vars, self._dist_context, OpRole.Optimize, process_mesh=[ -1 ], # hack to avoid initialize the dist attr for coalesce var is_recompute=False, sync=False, op_namescope="sharding_stage2_broadcast_dep", ) if self.enable_overlap: depend_op.dist_attr.execution_stream = comm_stream depend_op.dist_attr.scheduling_priority = ( self.comm_op_scheduling_priority ) main_block._sync_with_cpp() def _fuse_overlap_parameter_comm_stage_three(self, sharding_info): pass def _group_grads( self, block, sharding_info, ): """ conditions for gradients to be grouped: 1. group size < grad_bucket_size_numel 2. same dp group (TODO) 3. same src rank 4. same dtype 5. dependency: grad would NOT be used by other ops within group segment main logic: 1. record coalesce group 2. record all dp allreduce/reduce op idx 3. insert coalesce op 4. insert coalesce dependency (avoid allocate memory too early) 5. modify and remove allreduce/reduce op 6. ensure sharding-dp hybrid parallel logic gradients inside same group would be fuse into one coalesce tensor """ ops = block.ops if self.grad_bucket_size_numel < 1: # numel for transformer layer # h = 4096 + 1 # ffn_numel = 2 * (4 * h) * h # mha_numel = 3 * h * h + h * h # max_fuse_numel = ffn_numel + mha_numel self.grad_bucket_size_numel = 1 first_backward_op = None for op in ops: if is_loss_grad_op(op): first_backward_op = op # not backward op, sharding for inference if first_backward_op is None: return first_backward_varname = first_backward_op.output_arg_names[0] cur_group = VarGroup(self.grad_bucket_size_numel) grad_groups = [] grouped_grad_names = set() def op_depend_on_group(op, group): vars_ = set(op.input_arg_names + op.output_arg_names) var_names = set([var.name for var in group.vars]) return len(vars_.intersection(var_names)) > 0 # analyze groups i = 0 while i < len(ops): op = ops[i] if is_data_parallel_reduce_op(op): assert ( op.type == "c_reduce_sum" ), "Sharding should reduce grad first and than allreduce if Hybrid Sharding with Data-Parallel" grad_name = op.output_arg_names[0] param_name = _get_base_name_from_grad_name(grad_name) rank = sharding_info.get_var_rank(param_name) grad_var = block.var(grad_name) if cur_group.acceptable(grad_var, rank): assert grad_name not in grouped_grad_names cur_group.collect(grad_var, rank) else: grad_groups.append(cur_group) cur_group = VarGroup(self.grad_bucket_size_numel) cur_group.collect(grad_var, rank) if len(cur_group.vars) == 1: cur_group.coalesce_op_idx = i - 1 # NOTE coalesce dependency: control when allocate memory for gradients # too early would increase the peak memory requirement, too later would hurt the performance j = 2 while is_dep_skip_op(ops[i - j]): j += 1 dep_op = ops[i - j] dep_varname = dep_op.output_arg_names[0] cur_group.coalesce_dep_varname = dep_varname grouped_grad_names.add(grad_name) cur_group.reduce_op_indices.append(i) if self.sharding_hybrid_dp and sharding_info.is_in_local_shard( param_name ): cur_group.is_in_local_shard = True assert ( ops[i + 1].type == "c_allreduce_sum" ), "Sharding should reduce grad first and than allreduce if Hybrid Sharding with Data-Parallel" assert ( ops[i + 1].output_arg_names[0] == grad_name ), "Hybrid Sharding with Data-Parallel should sync same gradient var" cur_group.allreduce_op_indices.append(i + 1) i += 1 elif op_depend_on_group(op, cur_group): grad_groups.append(cur_group) cur_group = VarGroup(self.grad_bucket_size_numel) i += 1 # some grad not in this rank may not be used after dp reduced if len(cur_group.vars) >= 1: grad_groups.append(cur_group) _logger.info("Sharding Gradient Communication Optimization:") _logger.info( "Gradient Bucket size is [{}], [{}] Gradients are fused into [{}] Buckets.".format( self.grad_bucket_size_numel, len(grouped_grad_names), len(grad_groups), ) ) # create coalesce tesnor and record op idx grad_name_to_group_map = {} coalesce_to_group_map = {} modify_reduce_op_map = {} coalesce_op_map = {} remove_reduce_op_indices = [] for i, group in enumerate(grad_groups): if len(group.vars) > 1: group.coalesce_var = block.create_var( name=unique_name.generate( self.grad_coalesce_prefix + str(i) ), dtype=group.dtype, persistable=False, stop_gradient=True, ) coalesce_op_map[group.coalesce_op_idx] = group last_reduce_op_idx = group.reduce_op_indices.pop() modify_reduce_op_map[last_reduce_op_idx] = group remove_reduce_op_indices.extend(group.reduce_op_indices) if group.is_in_local_shard: last_allreduce_op_idx = group.allreduce_op_indices.pop() modify_reduce_op_map[last_allreduce_op_idx] = group remove_reduce_op_indices.extend(group.allreduce_op_indices) else: group.coalesce_var = group.vars[0] for grad in group.vars: grad_name_to_group_map[grad.name] = group coalesce_to_group_map[group.coalesce_var.name] = group coalesce_op_set = set(coalesce_op_map.keys()) modify_op_set = set(modify_reduce_op_map.keys()) remove_op_set = set(remove_reduce_op_indices) confilct = coalesce_op_set.intersection(modify_op_set) assert len(confilct) == 0 confilct = coalesce_op_set.intersection(remove_op_set) assert len(confilct) == 0 confilct = modify_op_set.intersection(remove_op_set) assert len(confilct) == 0 # update block for idx, op in reversed(list(enumerate(block.ops))): if idx in modify_reduce_op_map: group = modify_reduce_op_map[idx] grad_name = op.output_arg_names[0] assert ( grad_name == group.vars[-1].name ), "Unexpected: it is supposed to sync [{}] but got [{}]".format( group.vars[-1].name, grad_name ) op._rename_input(grad_name, group.coalesce_var.name) op._rename_output(grad_name, group.coalesce_var.name) if idx in remove_reduce_op_indices: block._remove_op(idx, sync=False) if idx in coalesce_op_map: group = coalesce_op_map[idx] first_grad_name = group.vars[0].name assert ( first_grad_name in op.output_arg_names ), "Unexpected: op is supposed to generate grad [{}] but got [{}]".format( first_grad_name, str(op) ) grad_names = [grad.name for grad in group.vars] concated_shapes = [] concated_ranks = [] for grad_ in group.vars: shape = grad_.shape concated_shapes.extend(shape) concated_ranks.append(len(shape)) coalesce_op = block._insert_op_without_sync( idx, type="coalesce_tensor", inputs={"Input": grad_names}, outputs={ "Output": grad_names, "FusedOutput": group.coalesce_var, }, attrs={ "copy_data": False, "use_align": True, "dtype": group.dtype, "concated_shapes": concated_shapes, "concated_ranks": concated_ranks, OP_ROLE_KEY: OpRole.Backward, }, ) depend_op = insert_dependencies_for_vars( block, idx, block.var(group.coalesce_dep_varname), group.coalesce_var, self._dist_context, OpRole.Backward, process_mesh=[ -1 ], # hack to avoid initialize the dist attr for coalesce var is_recompute=False, sync=False, op_namescope="sharding_grad_coalesce_dep", ) block._sync_with_cpp() return coalesce_to_group_map, grad_name_to_group_map def _overlap_grad_comm( self, block, sharding_info, coalesce_to_group_map, grad_name_to_group_map, ): """ overlap gradient communication with backward & optimizer computation. 1. assign gradient communications to grad comm stream 2. for coalesce gradient communication: 2.1 insert before communication dependencies 2.2 insert after communication dependencies only when need 3. there is not need to add explicit dependencies for non-coalesce gradient communication P.S. this overlap pass is ONLY adapted for standalone executor (graph based) and stream awared allocator. """ if not _is_enable_standalone_executor() or (not self.enable_overlap): return self.grad_comm_group_stream_pairs = [] ranks = sharding_info.group.ranks # NOTE since the gradient synchronization has calculation, there would be computation # competition between backward calculation. therefore should limit the number of stream used. for i in range(self.grad_comm_stream_num): if i == 0: group = sharding_info.group else: group = new_process_group(ranks, force_new_group=True) # NOTE here stream is just a presentation with different name, # it is up to executor to create the exact streams given the name. stream = "sharding_grad_comm_stream{}".format(i) self.grad_comm_group_stream_pairs.append( { "comm_group": group, "comm_stream": stream, } ) ops = block.ops # analyze dependencies dep_map = {} reduce_op_count = 0 grad_comm_op_to_stream_idx = {} for idx, op in enumerate(ops): if is_data_parallel_reduce_op(op): if op.type == "c_allreduce_sum": continue stream_idx = reduce_op_count % self.grad_comm_stream_num grad_comm_op_to_stream_idx[op] = stream_idx comm_group = self.grad_comm_group_stream_pairs[stream_idx][ "comm_group" ] comm_stream = self.grad_comm_group_stream_pairs[stream_idx][ "comm_stream" ] reduce_varname = op.output("Out")[0] grad_group = coalesce_to_group_map[reduce_varname] assert grad_group.coalesce_var.name == reduce_varname # coalesce deps if len(grad_group.vars) > 1: # NOTE should prior vars to be all grads ? # when the grad_ops' order is random # prior dep dep_map[idx] = [ ( idx, grad_group.vars[-1], grad_group.coalesce_var, comm_stream, ) ] # post dep post_idx = idx + 1 if self.sharding_hybrid_dp and grad_group.is_in_local_shard: post_idx += 1 dep_map[idx].append( ( post_idx, grad_group.coalesce_var, grad_group.vars, comm_stream, ) ) # assign stream op.dist_attr.execution_stream = comm_stream op.dist_attr.scheduling_priority = ( self.comm_op_scheduling_priority ) op._set_attr("ring_id", comm_group.id) if self.sharding_hybrid_dp and grad_group.is_in_local_shard: next_op = ops[idx + 1] assert next_op.type == "c_allreduce_sum" assert next_op.output("Out")[0] == reduce_varname # FIXME hybrid sharding-dp support multi comm & stream in feature # next_op._set_attr("ring_id", comm_group.id) next_op.dist_attr.execution_stream = comm_stream next_op.dist_attr.scheduling_priority = ( self.comm_op_scheduling_priority ) idx += 1 reduce_op_count += 1 idx += 1 # insert deps indice = sorted(list(dep_map.keys()), reverse=True) for i in indice: for idx, prior_vars, post_vars, comm_stream in dep_map[i][::-1]: depend_op = insert_dependencies_for_vars( block, idx, prior_vars, post_vars, self._dist_context, OpRole.Backward, process_mesh=[ -1 ], # hack to avoid initialize the dist attr for coalesce var is_recompute=False, sync=False, op_namescope="sharding_grad_comm_dep", ) depend_op.dist_attr.execution_stream = comm_stream depend_op.dist_attr.scheduling_priority = ( self.comm_op_scheduling_priority ) # hierarchical grad comm if self.enable_hierarchical_comm: # NOTE so far we only support Isomorphic cluster with 8 ranks per node # TODO unifiy here create communicators # create communicators nranks_per_node = 8 assert self.sharding_world_size % nranks_per_node == 0 global_group = sharding_info.group global_ranks = global_group.ranks relative_idx_in_node = self.global_rank % nranks_per_node node_idx = self.global_rank // nranks_per_node inter_node_ranks = [ rank for rank in global_ranks if rank % nranks_per_node == relative_idx_in_node ] _logger.info( "Sharding Gradient Hierarchical Communication Optimization." ) _logger.info( "current global rank idx: {}.".format(self.global_rank) ) _logger.info( "local inter node ranks idx: {}.".format(inter_node_ranks) ) assert ( len(inter_node_ranks) == self.sharding_world_size // nranks_per_node ) intra_node_ranks = [ rank for rank in global_ranks if rank // nranks_per_node == node_idx ] assert len(intra_node_ranks) == nranks_per_node _logger.info( "local intra node ranks idx: {}.".format(intra_node_ranks) ) inter_node_groups = [] intra_node_groups = [] for _ in range(self.grad_comm_stream_num): # TODO re-use one origin communicator inter_node_groups.append( new_process_group(inter_node_ranks, force_new_group=True) ) intra_node_groups.append( new_process_group(intra_node_ranks, force_new_group=True) ) # update program for idx, op in reversed(list(enumerate(block.ops))): if is_data_parallel_reduce_op(op): assert op.type == "c_reduce_sum" grad_comm_stream_idx = grad_comm_op_to_stream_idx[op] inter_node_group = inter_node_groups[grad_comm_stream_idx] intra_node_group = intra_node_groups[grad_comm_stream_idx] reduce_varname = op.output("Out")[0] if self.enable_overlap: comm_stream = op.dist_attr.execution_stream dst_rank = int(op.attr("root_id")) in_peer = False if dst_rank % nranks_per_node == relative_idx_in_node: in_peer = True intra_node_dst = dst_rank % nranks_per_node op._set_attr('ring_id', intra_node_group.id) op._set_attr('root_id', intra_node_dst) if in_peer: inter_node_dst = dst_rank // nranks_per_node new_op = block._insert_op_without_sync( idx + 1, type='c_reduce_sum', inputs={"X": reduce_varname}, outputs={ "Out": reduce_varname, }, attrs={ 'ring_id': inter_node_group.id, 'root_id': inter_node_dst, 'use_calc_stream': True, OP_ROLE_KEY: OpRole.Backward, }, ) new_op._set_attr( 'op_namescope', str('/') + ParallelMode.DataParallel ) if self.enable_overlap: new_op.dist_attr.execution_stream = comm_stream new_op.dist_attr.scheduling_priority = ( self.comm_op_scheduling_priority ) block._sync_with_cpp() def _get_broadcast_first_depend_op(block): for op in block.ops: if op.type in _supported_optimizer_type: return op raise Exception("Could not find optimizer op.") def _insert_init_and_broadcast_op( block, insert_idx, varname, local_rank, root_rank, ring_id, op_role, dist_context, ): """ empty op for initialization """ broadcast_var = block.var(varname) broadcast_var_dist_attr = dist_context.get_tensor_dist_attr_for_program( broadcast_var ) new_op = block._insert_op_without_sync( insert_idx, type='c_broadcast', inputs={'X': varname}, outputs={'Out': varname}, attrs={ 'ring_id': ring_id, 'root': root_rank, 'use_calc_stream': True, OP_ROLE_KEY: op_role, }, ) new_op._set_attr('op_namescope', str('/') + ParallelMode.DataParallel) naive_set_dist_op_attr_for_program_by_mesh_and_mapping( new_op, broadcast_var_dist_attr.process_mesh, broadcast_var_dist_attr.dims_mapping, dist_context, ) if local_rank != root_rank: new_op = block._insert_op_without_sync( insert_idx, type="empty", outputs={"Out": broadcast_var.name}, attrs={ "shape": broadcast_var.shape, "dtype": broadcast_var.dtype, OP_ROLE_KEY: op_role, }, ) naive_set_dist_op_attr_for_program_by_mesh_and_mapping( new_op, broadcast_var_dist_attr.process_mesh, broadcast_var_dist_attr.dims_mapping, dist_context, ) return def _insert_reduce_op( block, insert_idx, reduce_var, ring_id, root_id, dist_context, op_role=OpRole.Backward, use_calc_stream=True, ): assert ( root_id >= 0 ), "root id should be a positive int, but now root id is {}".format(root_id) new_op = block._insert_op_without_sync( insert_idx, type='c_reduce_sum', inputs={'X': [reduce_var]}, outputs={'Out': [reduce_var]}, attrs={ 'ring_id': ring_id, 'root_id': root_id, 'use_calc_stream': use_calc_stream, OP_ROLE_KEY: op_role, }, ) dist_attr = dist_context.get_tensor_dist_attr_for_program( block.var(reduce_var) ) naive_set_dist_op_attr_for_program_by_mesh_and_mapping( new_op, dist_attr.process_mesh, dist_attr.dims_mapping, dist_context ) new_op._set_attr('op_namescope', str('/') + ParallelMode.DataParallel) return new_op def _get_dp_and_sharding_groups(origin_group, sharding_group_size, rank): dp_axis = 0 sharding_axis = 1 shape = [len(origin_group) // sharding_group_size, sharding_group_size] dp_group = _get_comm_group(origin_group, shape, dp_axis, rank) sharding_group = _get_comm_group(origin_group, shape, sharding_axis, rank) return dp_group, sharding_group def _is_gradient_clip_op(op): return op.desc.has_attr("op_namescope") and op.desc.attr( "op_namescope" ).startswith("/gradient_clip") def _is_weight_decay_op(op): return op.desc.has_attr("op_namescope") and op.desc.attr( "op_namescope" ).startswith("/regularization") def _is_param_grad_fp32_cast_op(block, op): if not is_backward_op(op): return False if not _is_desired_cast_op( block, op, core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32 ): return False output_name = op.output_arg_names[0] base_name = output_name[: output_name.find("@")] if not block.has_var(base_name): return False return block.var(base_name).is_parameter def _is_param_fp16_cast_op(block, op, params): if is_optimize_op(op): return False if not _is_desired_cast_op(block, op): return False input_name = op.input_arg_names[0] if input_name not in params: return False return True def _is_desired_cast_op( block, op, src_var_type=core.VarDesc.VarType.FP32, dst_var_type=core.VarDesc.VarType.FP16, ): if op.type != "cast": return False assert len(op.input_arg_names) == 1 assert len(op.output_arg_names) == 1 input_var = block.var(op.input_arg_names[0]) output_var = block.var(op.output_arg_names[0]) if input_var.dtype != src_var_type or output_var.dtype != dst_var_type: return False return True def _get_base_name_from_grad_name(grad_name): base_name = None if ".cast_fp16@GRAD" in grad_name: base_name = grad_name[: grad_name.find(".cast_fp16@GRAD")] elif "@GRAD" in grad_name: base_name = grad_name[: grad_name.find("@GRAD")] return base_name def _is_param_grad_allreduce_op(op, block): if not is_data_parallel_reduce_op(op): return False output_name = op.output_arg_names[0] base_name = _get_base_name_from_grad_name(output_name) if not block.has_var(base_name): return False return block.var(base_name).is_parameter def _is_param_grad_sum_op(op, block): if not is_backward_op(op): return False if op.type != "sum": return False output_name = op.output_arg_names[0] base_name = _get_base_name_from_grad_name(output_name) if not block.has_var(base_name): return False return block.var(base_name).is_parameter def _is_forward_op(op): return op.attr("op_role") == 0 def is_sharding_param_broadcast_op(op): return ( op.type == "c_broadcast" and op.desc.has_attr("op_namescope") and ParallelMode.DataParallel in op.desc.attr("op_namescope") ) def _inference_data_parallel_group_for_operator(rank_id, op, dist_context): dp_group = None for input_name in op.input_arg_names: if not is_parameter_related(input_name, op.block): dist_attr = dist_context.get_op_dist_attr_for_program(op) process_mesh = dist_attr.process_mesh input_dim_mapping = dist_attr.get_input_dims_mapping(input_name) mesh_shape = process_mesh.shape # TODO(JZ-LIANG) replace with specific batch size dimension batch_size_axis = input_dim_mapping[0] if batch_size_axis > -1 and mesh_shape[batch_size_axis] > 1: group_ranks = _get_comm_group( process_mesh.process_ids, process_mesh.shape, batch_size_axis, rank_id, ) dp_group = new_process_group(group_ranks) break return dp_group def partition_by_use_order(params, group_size): """ shard the continouse param into same rank and divide the forward&backward computation into segement, which will favor the fuse pass in later. we assume that the params is already sorted by utilization order. """ mapping = {} total_param_mem = 0.0 param2mem = [] for param in params: mem = get_var_size(param) total_param_mem += mem param2mem.append((param, mem)) mapping = {x: [] for x in range(group_size)} cur_rank = 0 mem_accu = 0.0 for param, mem in param2mem: if mem_accu > total_param_mem * 1.0 * (cur_rank + 1) / group_size: cur_rank += 1 mapping[cur_rank].append(param) mem_accu += mem return mapping def partition_by_greedy_even(params, group_size): """ use greedy alogrithm to partition parameter as even as possible. """ mapping = {} for rank_ in range(group_size): mapping[rank_] = [] sizes = [0] * group_size for param in params: rank = sizes.index(min(sizes)) mapping[rank].append(param) numel = reduce(lambda x, y: x * y, param.shape) assert ( numel > 0 ), "param [{}] should larger than 0, but it is [{}]".format( param.name, numel ) sizes[rank] += numel return mapping def partition_parameters(params, group_size, algor="greedy_even"): if algor == "greedy_even": rank_to_params = partition_by_greedy_even(params, group_size) else: rank_to_params = partition_by_use_order(params, group_size) _logger.info("Sharding Parameter Partition:") for k, v in rank_to_params.items(): _logger.info( "Rank:{}, Parameter Size:{} MB.".format( k, sum([get_var_size(var) for var in v]) ) ) _logger.info("Params in this rank: {}.".format([var.name for var in v])) return rank_to_params def re_order_program(block, param_grads, dist_context): # record order pname_to_pg_pairs = {} for p, g in param_grads: pname_to_pg_pairs[p.name] = (p, g) use_order = [] for op in block.ops: for input_name in op.input_arg_names: if (input_name in pname_to_pg_pairs) and ( input_name not in use_order ): use_order.append(input_name) if len(use_order) == len(pname_to_pg_pairs): break # reorder optimzier last_op = block.ops[-1] pname_to_op = {} num_ops = len(block.ops) remove_op_indices = [] # TODO support case when optimizer is not the last op if is_optimize_op(last_op) and last_op.type in _supported_optimizer_type: # record optimizer for idx, op in reversed(list(enumerate(block.ops))): if op.type not in _supported_optimizer_type: break assert len(op.input("Param")) == 1 pname_to_op[op.input("Param")[0]] = op remove_op_indices.append(idx) assert len(use_order) == len(pname_to_op) # append new opts for pname in use_order: new_op = block.append_op(type='nop') new_op.desc.copy_from(pname_to_op[pname].desc) dist_context.set_op_dist_attr_for_program( new_op, dist_context.get_op_dist_attr_for_program(pname_to_op[pname]), ) # remove old opts for idx in remove_op_indices: block._remove_op(idx, sync=False) block._sync_with_cpp() assert len(block.ops) == num_ops # TODO reorder gradient clip order _logger.info( "Sharding the Order of param being used: {}.".format(use_order) ) return [pname_to_pg_pairs[p] for p in use_order] def group_param(sharding_info, fuse_size): """ param are group by: rank id fuse_size dtype """ group_to_param_map = {} param_to_group_map = {} bucket = [] cur_group = VarGroup(fuse_size) for param in sharding_info.params: rank = sharding_info.get_var_rank(param.name) if cur_group.acceptable(param, rank): cur_group.collect(param, rank) else: cur_group = VarGroup(fuse_size) cur_group.collect(param, rank) cur_group.is_in_local_shard = sharding_info.is_in_local_shard( param.name ) if cur_group in group_to_param_map: group_to_param_map[cur_group].append(param.name) else: group_to_param_map[cur_group] = [param.name] param_to_group_map[param.name] = cur_group return group_to_param_map, param_to_group_map class ShardingInfo: def __init__(self, group, rank, params_grads, partition_algor): self.group = group self.params_grads = dict([(p.name, (p, g)) for p, g in params_grads]) assert len(self.params_grads) == len( set(self.params_grads) ), "found duplicated param in params_grads" self.params = [p for p, _ in params_grads] self.param_names = [p.name for p in self.params] self.group_size = group.nranks self.global_rank = rank self.local_rank = group.ranks.index(self.global_rank) self.partition_algor = partition_algor # rank in below mapping are local rank in this sharding group self.rank_to_params = partition_parameters( self.params, self.group_size, self.partition_algor ) # include fp32 and fp16 param self.param_to_rank = dict() self._map_param_to_rank() def _map_param_to_rank(self): """ mapping parameters to the rank which holds it. """ for rank, params in self.rank_to_params.items(): for param in params: self.param_to_rank[param.name] = rank def get_var_rank(self, varname): if varname in self.param_to_rank: return self.param_to_rank[varname] return -1 # determine fp32 and fp16 (cast) param def is_in_local_shard(self, param_name): return self.get_var_rank(param_name) == self.local_rank # NOTE the follwo logic is designed for supporting AMP O1 when # the param would be cast to fp16 before used for caculation. # and sharding should only broadcast the casted fp16 param # instead of the origin fp32 version param. def get_broadcast_vars_and_param_usage(self, block): broadcast_vars = set([]) fp16_params = set([]) fp16_to_fp32 = {} param_usage = {x: 0 for x in self.param_names} for op in block.ops: if is_optimize_op(op): continue for input_name in op.input_arg_names: if input_name in self.param_names: param_usage[input_name] += 1 for op in block.ops: if not _is_param_fp16_cast_op(block, op, self.param_names): continue input_name = op.input_arg_names[0] output_name = op.output_arg_names[0] broadcast_vars.add(output_name) fp16_params.add(output_name) fp16_to_fp32[output_name] = input_name param_usage[input_name] -= 1 self.param_to_rank[output_name] = self.param_to_rank[input_name] for param, usage in param_usage.items(): if usage > 0: broadcast_vars.add(param) return broadcast_vars, param_usage def get_param_grad(self, param_name): if not self.is_in_local_shard(param_name): raise ValueError( "param[{}] not in current rank.".format(param_name) ) if param_name not in self.params_grads: raise ValueError('param[{}] not in params_grads'.format(param_name)) return self.params_grads.get(param_name, None) class VarGroup: def __init__(self, max_size): self.max_siez = max_size self.dtype = None self.rank = -1 self.numel = 0 self.vars = [] self.coalesce_var = None self.coalesce_dep_varname = None self.coalesce_op_idx = None self.reduce_op_indices = [] self.allreduce_op_indices = [] self.is_in_local_shard = False def acceptable(self, param, rank): if self.numel == 0: return True else: if param.dtype != self.dtype: return False if rank != self.rank: return False if self.numel + get_var_numel(param) > self.max_siez: return False return True def collect(self, param, rank): self.dtype = param.dtype self.rank = rank self.numel += get_var_numel(param) self.vars.append(param) def __len__(self): return len(self.vars)