fix dap init in benchmark (#24)

benchmark/perf.py

@@ -35,19 +35,6 @@ def main():
 
     args = parser.parse_args()
 
-    args.distributed = False
-    if 'WORLD_SIZE' in os.environ:
-        args.distributed = int(os.environ['WORLD_SIZE']) > 1
-
-    args.local_rank = int(os.environ['LOCAL_RANK'])
-
-    torch.cuda.set_device(args.local_rank)
-    torch.distributed.init_process_group(backend='nccl', init_method='env://')
-    args.world_size = torch.distributed.get_world_size()
-    args.global_rank = torch.distributed.get_rank()
-    print(
-        'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
-        % (args.global_rank, args.world_size))
     init_dap(args.dap_size)
 
     precision = torch.bfloat16

fastfold/distributed/core.py

@@ -8,7 +8,9 @@ def ensure_divisibility(numerator, denominator):
     """Ensure that numerator is divisible by the denominator."""
     assert numerator % denominator == 0, '{} is not divisible by {}'.format(numerator, denominator)
 
-def set_distributed_environ(key, value):
+
+def set_missing_distributed_environ(key, value):
+    if key not in os.environ:
         os.environ[str(key)] = str(value)
 
 
@@ -21,19 +23,18 @@ def init_dap(tensor_model_parallel_size_=None):
         else:
             tensor_model_parallel_size_ = 1
 
-    if torch.torch.distributed.is_initialized():
+    if torch.distributed.is_initialized():
         _logger = colossalai.logging.get_dist_logger()
         _logger.error(
             "use fastfold.distributed.init_dap instead of torch.distributed.init_process_group!")
         exit(-1)
 
     # set distributed environ for single device launch
-    if 'RANK' not in os.environ:
-        set_distributed_environ('WORLD_SIZE', 1)
-        set_distributed_environ('RANK', 0)
-        set_distributed_environ('LOCAL_RANK', 0)
-        set_distributed_environ('MASTER_ADDR', "localhost")
-        set_distributed_environ('MASTER_PORT', 10045)
+    set_missing_distributed_environ('WORLD_SIZE', 1)
+    set_missing_distributed_environ('RANK', 0)
+    set_missing_distributed_environ('LOCAL_RANK', 0)
+    set_missing_distributed_environ('MASTER_ADDR', "localhost")
+    set_missing_distributed_environ('MASTER_PORT', -1)
 
     colossalai.launch_from_torch(
         config={"parallel": dict(tensor=dict(size=tensor_model_parallel_size_))})