[util] fixed activation checkpointing on torch 1.9 (#719)

colossalai/utils/activation_checkpoint.py

@@ -68,6 +68,9 @@ class CheckpointFunction(torch.autograd.Function):
             else:
                 ctx.inputs.append(arg)
 
+        if activation_offload:
+            ctx.tensor_inputs = tensor_inputs
+        else:
             ctx.save_for_backward(*tensor_inputs)
         return outputs
 
@@ -79,6 +82,10 @@ class CheckpointFunction(torch.autograd.Function):
         # Copy the list to avoid modifying original list.
         inputs = list(ctx.inputs)
         tensor_indices = ctx.tensor_indices
+
+        if ctx.activation_offload:
+            tensors = ctx.tensor_inputs
+        else:
             tensors = ctx.saved_tensors
 
         # store the current states