# Copyright (c) 2023 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 collections import logging from ..auto_parallel.static.utils import ( get_logger, ) from .pass_base import PassBase, register_pass from .pass_utils import AutoParallelStreamType, split_matmul_grad_to_matmul logger = get_logger(logging.INFO) # For allreduce pattern in the backward phase of column parallel linear: # dX, dY = matmul_grad(X, Y, dOut) # dX = all_reduce_sum(dX) # Split matmul_grad to 2 matmul: # dX = matmul(dOut, Y^T) # dX = all_reduce_sum(dX) # dY = matmul(X^T, dOut) # # Then the all_reduce sum can overlap with the compute of dY. @register_pass("allreduce_matmul_grad_overlapping") class AllreduceMatmulGradOverlappingPass(PassBase): def __init__(self): super().__init__() self.op_namescope = "/auto_parallel/allreduce_matmul_grad_overlapping" self.set_attr("dist_context", None) def _check_self(self): if self.get_attr("dist_context") 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") block = main_program.global_block() matmul_grad_id_to_allreduce_id = ( self._get_all_matmul_grad_and_allreduce_pairs(block) ) logger.info( f"overlap matmul_grad and allreduce: {matmul_grad_id_to_allreduce_id}" ) self._split_matmul_grad_and_multi_streaming_allreduce( block, matmul_grad_id_to_allreduce_id ) def _get_all_matmul_grad_and_allreduce_pairs(self, block): ops = block.ops op_num = len(ops) matmul_grad_id_to_allreduce_id = collections.OrderedDict() for i, op_i in enumerate(ops): if ( op_i.type == 'matmul_v2_grad' and op_i.attr("trans_x") is False and op_i.attr("trans_y") is False ): x_grad = op_i.output("X@GRAD") for j in range(i + 1, op_num): op_j = ops[j] if ( op_j.type == 'c_allreduce_sum' and op_j.input("X") == x_grad ): matmul_grad_id_to_allreduce_id[i] = j return matmul_grad_id_to_allreduce_id def _split_matmul_grad_and_multi_streaming_allreduce( self, block, matmul_grad_id_to_allreduce_id ): ops = block.ops for matmul_grad_id, allreduce_id in reversed( matmul_grad_id_to_allreduce_id.items() ): matmul_grad_op = ops[matmul_grad_id] allreduce_op = ops[allreduce_id] # NOTE(Sonder): When there are ops between matmul_grad and allreduce, we should check whether # these ops rely on the output of the intermediate ops. If so, we should not split the matmul_grad. # Otherwise, the output of the intermediate ops will get wrong results. skip_overlapping = False moved_ops_output = [] matmul_grad_output = matmul_grad_op.output('Y@GRAD')[0] for idx in range(matmul_grad_id + 1, allreduce_id): if matmul_grad_output in ops[idx].desc.input_arg_names(): moved_ops_output.extend(ops[idx].desc.output_arg_names()) else: for input_name in ops[idx].desc.input_arg_names(): if input_name in moved_ops_output: skip_overlapping = True if skip_overlapping: continue # matmul_grad_op => matmul_v2 + reshape + reshape + matmul_v2 + reshape split_matmul_grad_to_matmul( block, matmul_grad_id, self.dist_context, self.op_namescope ) # NOTE(Ruibiao): Required OP scheduling order: matmul(dOut, Y^T) -> all_reduce_sum(dX) -> matmul(X^T, dOut). # all_reduce_sum(dX) and matmul(X^T, dOut) cannot be swapped. Otherwise, after buffer_shared_inplace_pass # adding share_buffer OP before all_reduce_sum, all_reduce_sum will synchronous with comp-stream, and then # the matmul op before it cannot be overlapped. allreduce_op_dist_attr = ( self.dist_context.get_op_dist_attr_for_program(allreduce_op) ) allreduce_op_dist_attr.execution_stream = ( AutoParallelStreamType.MP_STREAM.value ) allreduce_op_inputs = allreduce_op.desc.input_names() allreduce_op_outputs = allreduce_op.desc.output_names() allreduce_op_inputs = { name: allreduce_op.input(name) for name in allreduce_op_inputs } allreduce_op_outputs = { name: allreduce_op.output(name) for name in allreduce_op_outputs } # matmul_v2 + reshape + reshape + matmul_v2 + reshape + ... + original all_reduce_sum # => # matmul_v2 + new all_reduce_sum + reshape + reshape + matmul_v2 + reshape + ... + original all_reduce_sum # # NOTE(liym27): new all_reduce_sum must be inserted to "the next of the first matmul_v2", otherwise another # pass fused_linear_param_grad_add will not work. allreduce_op = block._insert_op_without_sync( index=matmul_grad_id + 1, type=allreduce_op.type, inputs=allreduce_op_inputs, outputs=allreduce_op_outputs, attrs=allreduce_op.all_attrs(), ) self.dist_context.set_op_dist_attr_for_program( allreduce_op, allreduce_op_dist_attr ) # Remove the original allreduce op block._remove_op(allreduce_id + 5, sync=False) block._sync_with_cpp()