Merge branch 'free-pipe-output-t5-fix' into 'main'

added flag to control deallocation of pipeline outputs

See merge request ADLR/megatron-lm!377

megatron/arguments.py

@@ -681,6 +681,9 @@ def _add_distributed_args(parser):
                        help='Call torch.cuda.empty_cache() each iteration '
                        '(training and eval), to reduce fragmentation.'
                        '0=off, 1=moderate, 2=aggressive.')
+    group.add_argument('--deallocate-pipeline-outputs', action='store_true',
+                       default=False, help='If set, pipeline output tensors '
+                       'are deallocated during the forward pass.')
     return parser
 
 

megatron/schedules.py

@@ -42,14 +42,14 @@ def get_forward_backward_func():
         forward_backward_func = forward_backward_no_pipelining
     return forward_backward_func
 
-def free_output_tensor(output_tensors):
+def free_output_tensor(output_tensors, deallocate_pipeline_outputs):
     '''Pseudo-free (i.e., set to scalar) the output tensor's '.data' field.
 
     This method should be called right after the output tensor has been
     sent to the next pipeline stage. At this point, the output tensor is
     only useful for its '.grad_fn' field, and not its '.data'.
     '''
-    if output_tensors is None:
+    if not deallocate_pipeline_outputs or output_tensors is None:
         return
     if isinstance(output_tensors, torch.Tensor):
         output_tensors = [output_tensors]
@@ -164,7 +164,11 @@ def backward_step(optimizer, input_tensor, output_tensor, output_tensor_grad):
     # Backward pass.
     if output_tensor_grad[0] is None:
         output_tensor = optimizer.scale_loss(output_tensor[0])
-    custom_backward(output_tensor[0], output_tensor_grad[0])
+    if args.deallocate_pipeline_outputs:
+        custom_backward(output_tensor[0], output_tensor_grad[0])
+    else:
+        torch.autograd.backward(output_tensor[0],
+                                grad_tensors=output_tensor_grad[0])
 
     # Collect the grad of the input_tensor.
     input_tensor_grad = [None]
@@ -372,7 +376,7 @@ def forward_backward_pipelining_with_interleaving(forward_step_func, data_iterat
                     output_tensor, recv_prev=recv_prev,
                     tensor_shape=tensor_shape,
                     timers=timers)
-        free_output_tensor(output_tensor)
+        free_output_tensor(output_tensor, args.deallocate_pipeline_outputs)
         input_tensors[next_forward_model_chunk_id].append(input_tensor)
 
     # Run 1F1B in steady state.
@@ -437,7 +441,7 @@ def forward_backward_pipelining_with_interleaving(forward_step_func, data_iterat
                     output_tensor, input_tensor_grad,
                     recv_prev=recv_prev, recv_next=recv_next,
                     tensor_shape=tensor_shape, timers=timers)
-        free_output_tensor(output_tensor)
+        free_output_tensor(output_tensor, args.deallocate_pipeline_outputs)
 
         # Put input_tensor and output_tensor_grad in data structures in the
         # right location.
@@ -571,6 +575,7 @@ def forward_backward_pipelining_without_interleaving(forward_step_func, data_ite
     stages.
 
     Returns dictionary with losses if the last stage, empty dict otherwise."""
+    args = get_args()
     timers = get_timers()
 
     assert len(model) == 1
@@ -612,7 +617,7 @@ def forward_backward_pipelining_without_interleaving(forward_step_func, data_ite
         if not forward_only:
             input_tensors.append(input_tensor)
             output_tensors.append(output_tensor)
-            free_output_tensor(output_tensor)
+            free_output_tensor(output_tensor, args.deallocate_pipeline_outputs)
 
     # Before running 1F1B, need to receive first forward tensor.
     # If all microbatches are run in warmup / cooldown phase, then no need to
@@ -641,7 +646,7 @@ def forward_backward_pipelining_without_interleaving(forward_step_func, data_ite
             # Add input_tensor and output_tensor to end of list.
             input_tensors.append(input_tensor)
             output_tensors.append(output_tensor)
-            free_output_tensor(output_tensor)
+            free_output_tensor(output_tensor, args.deallocate_pipeline_outputs)
 
             # Pop input_tensor and output_tensor from the start of the list for
             # the backward pass.