|
| 1 | +from functools import wraps |
| 2 | +import os |
| 3 | +import random |
| 4 | + |
| 5 | +import torch |
| 6 | +import torch.distributed as dist |
| 7 | +import torch.distributed.autograd as dist_autograd |
| 8 | +from torch.distributed.optim import DistributedOptimizer |
| 9 | +import torch.distributed.rpc as rpc |
| 10 | +from torch.distributed.rpc import RRef |
| 11 | +from torch.distributed.rpc import ProcessGroupRpcBackendOptions |
| 12 | +import torch.multiprocessing as mp |
| 13 | +from torch.nn.parallel import DistributedDataParallel as DDP |
| 14 | +import torch.optim as optim |
| 15 | + |
| 16 | +NUM_EMBEDDINGS = 100 |
| 17 | +EMBEDDING_DIM = 16 |
| 18 | + |
| 19 | +class HybridModel(torch.nn.Module): |
| 20 | + r""" |
| 21 | + The model consists of a sparse part and a dense part. The dense part is an |
| 22 | + nn.Linear module that is replicated across all trainers using |
| 23 | + DistributedDataParallel. The sparse part is an nn.EmbeddingBag that is |
| 24 | + stored on the parameter server. |
| 25 | +
|
| 26 | + The model holds a Remote Reference to the embedding table on the parameter |
| 27 | + server. |
| 28 | + """ |
| 29 | + |
| 30 | + def __init__(self, emb_rref, device): |
| 31 | + super(HybridModel, self).__init__() |
| 32 | + self.emb_rref = emb_rref |
| 33 | + self.fc = DDP(torch.nn.Linear(16, 8).cuda(device), device_ids=[device]) |
| 34 | + self.device = device |
| 35 | + |
| 36 | + def forward(self, indices, offsets): |
| 37 | + emb_lookup = self.emb_rref.rpc_sync().forward(indices, offsets) |
| 38 | + return self.fc(emb_lookup.cuda(self.device)) |
| 39 | + |
| 40 | +def _retrieve_embedding_parameters(emb_rref): |
| 41 | + return [RRef(p) for p in emb_rref.local_value().parameters()] |
| 42 | + |
| 43 | + |
| 44 | +def _run_trainer(emb_rref, rank): |
| 45 | + r""" |
| 46 | + Each trainer runs a forward pass which involves an embedding lookup on the |
| 47 | + parameter server and running nn.Linear locally. During the backward pass, |
| 48 | + DDP is responsible for aggregating the gradients for the dense part |
| 49 | + (nn.Linear) and distributed autograd ensures gradients updates are |
| 50 | + propagated to the parameter server. |
| 51 | + """ |
| 52 | + |
| 53 | + # Setup the model. |
| 54 | + model = HybridModel(emb_rref, rank) |
| 55 | + |
| 56 | + # Retrieve all model parameters as rrefs for DistributedOptimizer. |
| 57 | + |
| 58 | + # Retrieve parameters for embedding table. |
| 59 | + model_parameter_rrefs = rpc.rpc_sync( |
| 60 | + "ps", _retrieve_embedding_parameters, args=(emb_rref,)) |
| 61 | + |
| 62 | + # model.parameters() only includes local parameters. |
| 63 | + for param in model.parameters(): |
| 64 | + model_parameter_rrefs.append(RRef(param)) |
| 65 | + |
| 66 | + # Setup distributed optimizer |
| 67 | + opt = DistributedOptimizer( |
| 68 | + optim.SGD, |
| 69 | + model_parameter_rrefs, |
| 70 | + lr=0.05, |
| 71 | + ) |
| 72 | + |
| 73 | + criterion = torch.nn.CrossEntropyLoss() |
| 74 | + |
| 75 | + def get_next_batch(rank): |
| 76 | + for _ in range(10): |
| 77 | + num_indices = random.randint(20, 50) |
| 78 | + indices = torch.LongTensor(num_indices).random_(0, NUM_EMBEDDINGS) |
| 79 | + |
| 80 | + # Generate offsets. |
| 81 | + offsets = [] |
| 82 | + start = 0 |
| 83 | + batch_size = 0 |
| 84 | + while start < num_indices: |
| 85 | + offsets.append(start) |
| 86 | + start += random.randint(1, 10) |
| 87 | + batch_size += 1 |
| 88 | + |
| 89 | + offsets_tensor = torch.LongTensor(offsets) |
| 90 | + target = torch.LongTensor(batch_size).random_(8).cuda(rank) |
| 91 | + yield indices, offsets_tensor, target |
| 92 | + |
| 93 | + # Train for 100 epochs |
| 94 | + for epoch in range(100): |
| 95 | + # create distributed autograd context |
| 96 | + for indices, offsets, target in get_next_batch(rank): |
| 97 | + with dist_autograd.context() as context_id: |
| 98 | + output = model(indices, offsets) |
| 99 | + loss = criterion(output, target) |
| 100 | + |
| 101 | + # Run distributed backward pass |
| 102 | + dist_autograd.backward(context_id, [loss]) |
| 103 | + |
| 104 | + # Tun distributed optimizer |
| 105 | + opt.step(context_id) |
| 106 | + |
| 107 | + # Not necessary to zero grads as each iteration creates a different |
| 108 | + # distributed autograd context which hosts different grads |
| 109 | + print("Training done for epoch {}".format(epoch)) |
| 110 | + |
| 111 | + |
| 112 | +def run_worker(rank, world_size): |
| 113 | + r""" |
| 114 | + A wrapper function that initializes RPC, calls the function, and shuts down |
| 115 | + RPC. |
| 116 | + """ |
| 117 | + |
| 118 | + # We need to use different port numbers in TCP init_method for init_rpc and |
| 119 | + # init_process_group to avoid port conflicts. |
| 120 | + rpc_backend_options = ProcessGroupRpcBackendOptions() |
| 121 | + rpc_backend_options.init_method='tcp://localhost:29501' |
| 122 | + |
| 123 | + # Rank 2 is master, 3 is ps and 0 and 1 are trainers. |
| 124 | + if rank == 2: |
| 125 | + rpc.init_rpc( |
| 126 | + "master", |
| 127 | + rank=rank, |
| 128 | + world_size=world_size, |
| 129 | + rpc_backend_options=rpc_backend_options) |
| 130 | + |
| 131 | + # Build the embedding table on the ps. |
| 132 | + emb_rref = rpc.remote( |
| 133 | + "ps", |
| 134 | + torch.nn.EmbeddingBag, |
| 135 | + args=(NUM_EMBEDDINGS, EMBEDDING_DIM), |
| 136 | + kwargs={"mode": "sum"}) |
| 137 | + |
| 138 | + # Run the training loop on trainers. |
| 139 | + futs = [] |
| 140 | + for trainer_rank in [0, 1]: |
| 141 | + trainer_name = "trainer{}".format(trainer_rank) |
| 142 | + fut = rpc.rpc_async( |
| 143 | + trainer_name, _run_trainer, args=(emb_rref, rank)) |
| 144 | + futs.append(fut) |
| 145 | + |
| 146 | + # Wait for all training to finish. |
| 147 | + for fut in futs: |
| 148 | + fut.wait() |
| 149 | + elif rank <= 1: |
| 150 | + # Initialize process group for Distributed DataParallel on trainers. |
| 151 | + dist.init_process_group( |
| 152 | + backend="gloo", rank=rank, world_size=2, |
| 153 | + init_method='tcp://localhost:29500') |
| 154 | + |
| 155 | + # Initialize RPC. |
| 156 | + trainer_name = "trainer{}".format(rank) |
| 157 | + rpc.init_rpc( |
| 158 | + trainer_name, |
| 159 | + rank=rank, |
| 160 | + world_size=world_size, |
| 161 | + rpc_backend_options=rpc_backend_options) |
| 162 | + |
| 163 | + # Trainer just waits for RPCs from master. |
| 164 | + else: |
| 165 | + rpc.init_rpc( |
| 166 | + "ps", |
| 167 | + rank=rank, |
| 168 | + world_size=world_size, |
| 169 | + rpc_backend_options=rpc_backend_options) |
| 170 | + # parameter server do nothing |
| 171 | + pass |
| 172 | + |
| 173 | + # block until all rpcs finish |
| 174 | + rpc.shutdown() |
| 175 | + |
| 176 | + |
| 177 | +if __name__=="__main__": |
| 178 | + # 2 trainers, 1 parameter server, 1 master. |
| 179 | + world_size = 4 |
| 180 | + mp.spawn(run_worker, args=(world_size, ), nprocs=world_size, join=True) |
0 commit comments