| 42 | |
| 43 | |
| 44 | class DistributedDataParallel(Module): |
| 45 | |
| 46 | def __init__(self, module): |
| 47 | super(DistributedDataParallel, self).__init__() |
| 48 | self.warn_on_half = True if dist._backend == dist.dist_backend.GLOO else False |
| 49 | |
| 50 | self.module = module |
| 51 | self.data_parallel_group = mpu.get_data_parallel_group() |
| 52 | src_rank = mpu.get_tensor_model_parallel_rank() |
| 53 | for p in self.module.parameters(): |
| 54 | if torch.is_tensor(p): |
| 55 | dist.broadcast(p, src_rank, group=self.data_parallel_group) |
| 56 | |
| 57 | def allreduce_params(reduce_after=True, |
| 58 | no_scale=False, |
| 59 | fp32_allreduce=False): |
| 60 | if (self.needs_reduction): |
| 61 | self.needs_reduction = False |
| 62 | buckets = {} |
| 63 | for name, param in self.module.named_parameters(): |
| 64 | if param.requires_grad and param.grad is not None: |
| 65 | tp = (param.data.type()) |
| 66 | if tp not in buckets: |
| 67 | buckets[tp] = [] |
| 68 | buckets[tp].append(param) |
| 69 | if self.warn_on_half: |
| 70 | if torch.cuda.HalfTensor in buckets: |
| 71 | print( |
| 72 | 'WARNING: gloo dist backend for half parameters may be extremely slow.', |
| 73 | 'It is recommended to use the NCCL backend in this case.' |
| 74 | ) |
| 75 | self.warn_on_half = False |
| 76 | for tp in buckets: |
| 77 | bucket = buckets[tp] |
| 78 | grads = [param.grad.data for param in bucket] |
| 79 | coalesced = _flatten_dense_tensors(grads) |
| 80 | if fp32_allreduce: |
| 81 | coalesced = coalesced.float() |
| 82 | if not no_scale and not reduce_after: |
| 83 | coalesced /= dist.get_world_size( |
| 84 | group=self.data_parallel_group) |
| 85 | dist.all_reduce(coalesced, group=self.data_parallel_group) |
| 86 | torch.cuda.synchronize() |
| 87 | if not no_scale and reduce_after: |
| 88 | coalesced /= dist.get_world_size( |
| 89 | group=self.data_parallel_group) |
| 90 | for buf, synced in zip( |
| 91 | grads, _unflatten_dense_tensors(coalesced, grads)): |
| 92 | buf.copy_(synced) |
| 93 | |
| 94 | self.hook_handles = [] |
| 95 | self.hooks = [] |
| 96 | for param in list(self.module.parameters()): |
| 97 | |
| 98 | def allreduce_hook(*unused): |
| 99 | Variable._execution_engine.queue_callback(allreduce_params) |
| 100 | |
| 101 | self.allreduce_params = allreduce_params |