bert: bugfix for 1x1 training (#160)

Co-authored-by: zhouyu <zhouyu@baai.ac.cn>

training/benchmarks/bert/pytorch/dataloaders/dataset.py

@@ -1,14 +1,17 @@
-import random
-
 import h5py
 import numpy as np
 import os
+import sys
 
 import torch
 import torch.distributed as dist
 import torch.nn.functional as F
 from torch.utils.data import Dataset
 
+CURR_PATH = os.path.abspath(os.path.dirname(__file__))
+sys.path.append(os.path.abspath(os.path.join(CURR_PATH, "../../../")))
+from driver import dist_pytorch
+
 
 class PretrainingDataset(Dataset):
 
@@ -60,10 +63,14 @@ def exchange_padding_fast(device, max_batch_size, input_ids, segment_ids,
         input_mask = F.pad(input_mask, (0, 0, 0, pad_size))
         masked_lm_labels = F.pad(masked_lm_labels, (0, 0, 0, pad_size))
         next_sentence_labels = F.pad(next_sentence_labels, (0, pad_size))
-    ngpus = torch.distributed.get_world_size()
+
+    ngpus = 1
+    igpu = 0
+    if dist_pytorch.is_dist_avail_and_initialized():
+        ngpus = dist_pytorch.get_world_size()
+        igpu = dist_pytorch.get_rank()
     nseqs = input_mask.shape[0]
     ntokensperseq = input_mask.shape[1]
-    igpu = torch.distributed.get_rank()
 
     flattened_length_seq = nseqs * ntokensperseq
     flattened_length_nsp = nseqs
@@ -141,8 +148,11 @@ def decode_packet(flat_packet):
                            dtype=torch.float16)
     tensors_ = list(torch.split(tensors_, 1))
     # Address valueError: ProcessGroupGloo::allgather: invalid tensor size at index 0 (expected (2049), got (1, 2049))
-    torch.distributed.all_gather(tensors_,
-                                 tensors.view(torch.float16).unsqueeze(0))
+
+    if dist_pytorch.is_dist_avail_and_initialized():
+        dist.all_gather(tensors_, tensors.view(torch.float16).unsqueeze(0))
+    else:
+        tensors_ = tuple(tensors.view(torch.float16).unsqueeze(0))
 
     tensors_ = torch.stack(tensors_).view(torch.int16).long()
     input_ids_, segment_ids_, input_mask_, masked_lm_labels_, next_sentence_labels_ = decode_packet(

training/benchmarks/bert/pytorch/run_pretraining.py

@@ -42,6 +42,7 @@ def main():
     bert_driver = Driver(config, config.mutable_params)
     bert_driver.setup_config(argparse.ArgumentParser("Bert"))
     bert_driver.setup_modules(driver, globals(), locals())
+    config.distributed = dist_pytorch.get_world_size() > 1
 
     logger = bert_driver.logger