Add support for fake distributed process groups.

Author: Victarry

megatron/training/arguments.py

@@ -1129,6 +1129,8 @@ def validate_args(args, defaults={}):
                 assert not args.distrib_optim_fully_reshardable_mem_efficient, \
                     '--distrib-optim-fully-reshardable-mem-efficient requires -enable-gloo-process-groups'
 
+    if args.fake_process_group:
+        assert not args.enable_gloo_process_groups, "Fake distributed group requires disabling Gloo process groups."
 
     # Checkpointing
     if args.ckpt_fully_parallel_save_deprecated and args.rank == 0:
@@ -2834,6 +2836,10 @@ def _add_distributed_args(parser):
                        "and must be consistent across all ranks.")
     group.add_argument('--replication-factor', default=2, type=int,
                        help="Number of machines storing the replica of a given rank's data.")
+    group.add_argument('--fake-process-group', action='store_true', default=False,
+                       help='If set, initialize with fake distributed process group and all distributed communication operations will be skipped. \
+                       This is quite useful for profiling memory usage of distributed training with just one GPU. \
+                       Setting WORLD_SIZE and RANK to the specific values for target distribtued scale.')
     return parser
 
 

megatron/training/initialize.py

@@ -346,6 +346,11 @@ def _initialize_distributed(get_embedding_ranks, get_position_embedding_ranks, s
             'rank': args.rank,
             'timeout': timedelta(minutes=args.distributed_timeout_minutes),
         }
+        if args.fake_process_group:
+            from torch.testing._internal.distributed.fake_pg import FakeStore
+            store = FakeStore()
+            init_process_group_kwargs['backend'] = 'fake'
+            init_process_group_kwargs['store'] = store
 
         torch.distributed.init_process_group(**init_process_group_kwargs)
         inprocess_restart.maybe_force_nccl_backend_init(device_id)

megatron/training/training.py

@@ -1773,7 +1773,9 @@ def training_log(
                 num_microbatches = get_num_microbatches()
                 report_theoretical_memory(args, num_microbatches=num_microbatches, verbose=True)
             report_memory(f'(after {iteration} iterations)')
-            report_memory_flag = False
+            if iteration > 1:
+                # Make sure the memory after the second iteration is reported to include optimizer state memory.
+                report_memory_flag = False
         # Write timers to wandb, don't reset the counts
         if args.log_timers_to_tensorboard:
             timers.write(timers_to_log, writer, iteration, normalizer=args.log_interval, reset=False)

megatron/training/utils.py

@@ -414,7 +414,7 @@ def is_last_rank():
 
 def print_rank_last(message):
     """If distributed is initialized, print only on last rank."""
-    if torch.distributed.is_initialized():
+    if torch.distributed.is_initialized() and torch.distributed.get_backend() != 'fake':
         if is_last_rank():
             print(message, flush=True)
     else: