// Copyright (c) 2020 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. syntax = "proto2"; package paddle.fleet; enum Mode { COLLECTIVE = 1; PS = 2; PIPELINE = 3; HETER = 4; // support XPU and GPU computing server } message RecomputeConfig { repeated string checkpoints = 1; } message AMPConfig { optional float init_loss_scaling = 1 [ default = 32768.0 ]; optional int32 incr_every_n_steps = 2 [ default = 1000 ]; optional int32 decr_every_n_nan_or_inf = 3 [ default = 2 ]; optional float incr_ratio = 4 [ default = 2.0 ]; optional float decr_ratio = 5 [ default = 0.8 ]; optional bool use_dynamic_loss_scaling = 6 [ default = true ]; repeated string custom_white_list = 7; repeated string custom_black_list = 8; repeated string custom_black_varnames = 9; } message LocalSGDConfig { optional int32 k_steps = 1 [ default = 1 ]; optional int32 begin_step = 2 [ default = 1 ]; } message AdaptiveLocalSGDConfig { optional int32 init_k_steps = 1 [ default = 1 ]; optional int32 begin_step = 2 [ default = 1 ]; } message GradientMergeConfig { optional int32 k_steps = 1 [ default = 1 ]; optional bool avg = 2 [ default = true ]; } message DGCConfig { optional int32 rampup_begin_step = 1 [ default = 0 ]; optional int32 rampup_step = 2 [ default = 1 ]; repeated float sparsity = 3; } message LarsConfig { optional float lars_coeff = 1 [ default = 0.001 ]; optional float lars_weight_decay = 2 [ default = 0.0005 ]; optional float epsilon = 3 [ default = 0.0 ]; repeated string exclude_from_weight_decay = 4; } message LambConfig { optional float lamb_weight_decay = 1 [ default = 0.01 ]; repeated string exclude_from_weight_decay = 2; } message BuildStrategy { optional bool enable_sequential_execution = 1 [ default = false ]; optional bool fuse_elewise_add_act_ops = 2 [ default = false ]; optional bool fuse_bn_act_ops = 3 [ default = false ]; optional bool fuse_relu_depthwise_conv = 4 [ default = false ]; optional bool fuse_broadcast_ops = 5 [ default = false ]; optional bool fuse_all_optimizer_ops = 6 [ default = false ]; optional bool enable_inplace = 7 [ default = false ]; optional bool enable_backward_optimizer_op_deps = 8 [ default = true ]; optional bool cache_runtime_context = 9 [ default = false ]; } message ExecutionStrategy { optional int32 num_threads = 1 [ default = 1 ]; optional int32 num_iteration_per_drop_scope = 2 [ default = 10 ]; optional int32 num_iteration_per_run = 3 [ default = 1 ]; optional bool use_thread_barrier = 4 [ default = false ]; } message AsyncConfig { optional int32 k_steps = 1 [ default = -1 ]; optional int32 max_merge_var_num = 2 [ default = 1 ]; optional int32 send_queue_size = 3 [ default = 16 ]; optional bool independent_recv_thread = 4 [ default = false ]; optional int32 min_send_grad_num_before_recv = 5 [ default = 1 ]; optional int32 thread_pool_size = 6 [ default = 1 ]; optional int32 send_wait_times = 7 [ default = 1 ]; optional bool runtime_split_send_recv = 8 [ default = false ]; } message PipelineConfig { optional int32 micro_batch = 1 [ default = 1 ]; } message DistributedStrategy { // bool options optional Mode mode = 1 [ default = COLLECTIVE ]; optional bool amp = 2 [ default = false ]; optional bool recompute = 3 [ default = false ]; optional bool localsgd = 4 [ default = false ]; optional bool dgc = 5 [ default = false ]; optional bool gradient_merge = 6 [ default = false ]; optional bool lars = 7 [ default = false ]; optional bool lamb = 8 [ default = false ]; optional bool pipeline = 9 [ default = false ]; optional bool elastic = 10 [ default = false ]; optional bool auto = 11 [ default = false ]; optional bool a_sync = 12 [ default = true ]; optional bool sync_nccl_allreduce = 13 [ default = true ]; optional int32 nccl_comm_num = 14 [ default = 1 ]; optional bool use_hierarchical_allreduce = 15 [ default = false ]; optional int32 hierarchical_allreduce_inter_nranks = 16 [ default = 1 ]; optional bool sync_batch_norm = 17 [ default = false ]; optional bool fuse_all_reduce_ops = 18 [ default = true ]; optional int32 fuse_grad_size_in_MB = 19 [ default = 32 ]; optional float fuse_grad_size_in_TFLOPS = 20 [ default = 50 ]; optional bool cudnn_exhaustive_search = 21 [ default = true ]; optional int32 conv_workspace_size_limit = 22 [ default = 4000 ]; optional bool cudnn_batchnorm_spatial_persistent = 23 [ default = true ]; optional bool adaptive_localsgd = 24 [ default = false ]; optional RecomputeConfig recompute_configs = 101; optional AMPConfig amp_configs = 102; optional LocalSGDConfig localsgd_configs = 103; optional GradientMergeConfig gradient_merge_configs = 104; optional DGCConfig dgc_configs = 105; optional PipelineConfig pipeline_configs = 106; optional AsyncConfig a_sync_configs = 107; optional LarsConfig lars_configs = 108; optional LambConfig lamb_configs = 109; optional AdaptiveLocalSGDConfig adaptive_localsgd_configs = 110; optional BuildStrategy build_strategy = 201; optional ExecutionStrategy execution_strategy = 202; } message DistributedJobInfo { optional int32 worker_num = 1; optional int32 server_num = 2; repeated string worker_ips = 3; repeated string server_endpoints = 4; optional string origin_startup = 5; optional string origin_main = 6; // without backpropagation and optimization optional string distributed_main = 7; // with backpropagation and optimization optional string optimizer_name = 8; // optimizer name optional DistributedStrategy strategy = 101; }