[zero] fixed the activation offload (#647)

* fixed the activation offload

* polish

colossalai/utils/activation_checkpoint.py

@@ -7,6 +7,19 @@ from torch.utils.checkpoint import check_backward_validity, detach_variable
 from colossalai.context.random import get_states, get_current_mode, set_seed_states, set_mode, sync_states
 from .cuda import get_current_device
 
+
+def copy_to_device(obj, device):
+    if torch.is_tensor(obj):
+        return obj.to(device)
+    elif isinstance(obj, list):
+        return [copy_to_device(i, device) for i in obj]
+    elif isinstance(obj, tuple):
+        return tuple([copy_to_device(v, device) for v in obj])
+    elif isinstance(obj, dict):
+        return {k: copy_to_device(v, device) for k, v in obj.items()}
+    else:
+        return obj
+
 class CheckpointFunction(torch.autograd.Function):
 
     @staticmethod
@@ -27,6 +40,13 @@ class CheckpointFunction(torch.autograd.Function):
         else:
             ctx.had_autocast_in_fwd = False
         
+        if activation_offload:
+            inputs_cuda = copy_to_device(args, ctx.device)
+        else:
+            inputs_cuda = args
+
+        with torch.no_grad():
+            outputs = run_function(*inputs_cuda)
         # Save non-tensor inputs in ctx, keep a placeholder None for tensors
         # to be filled out during the backward.
         ctx.inputs = []
@@ -34,10 +54,8 @@ class CheckpointFunction(torch.autograd.Function):
         tensor_inputs = []
         for i, arg in enumerate(args):
             if torch.is_tensor(arg):
-                if ctx.activation_offload:
+                if activation_offload:
-                    tmp = arg.detach().cpu()
+                    tensor_inputs.append(copy_to_device(arg, 'cpu'))
-                    tmp.requires_grad = arg.requires_grad
-                    tensor_inputs.append(tmp)
                 else:
                     tensor_inputs.append(arg)
                 ctx.tensor_indices.append(i)
@@ -46,18 +64,15 @@ class CheckpointFunction(torch.autograd.Function):
                 ctx.inputs.append(arg)
 
         ctx.save_for_backward(*tensor_inputs)
-
-        with torch.no_grad():
-            outputs = run_function(*args)
         return outputs
 
     @staticmethod
     def backward(ctx, *args):
         if not torch.autograd._is_checkpoint_valid():
             raise RuntimeError(
-                "Checkpointing is not compatible with .grad() or when an `inputs` parameter"
+                "Checkpointing is not compatible with .grad() or when an `inputs` parameter is "
-                " is passed to .backward(). Please use .backward() and do not pass its `inputs`"
+                "passed to .backward(). Please use .backward() and do not pass its `inputs` argument."
-                " argument.")
+                )
         # Copy the list to avoid modifying original list.
         inputs = list(ctx.inputs)
         tensor_indices = ctx.tensor_indices
@@ -74,12 +89,12 @@ class CheckpointFunction(torch.autograd.Function):
         for parallel_mode, state in ctx.fwd_seed_states.items():
             set_seed_states(parallel_mode, state)
         set_mode(ctx.fwd_current_mode)
+        if ctx.activation_offload:
+            tensors = copy_to_device(tensors, ctx.device)
 
         # Fill in inputs with appropriate saved tensors.
         for i, idx in enumerate(tensor_indices):
-            tmp = tensors[i].detach().to(ctx.device)
+            inputs[idx] = tensors[i]
-            tmp.requires_grad = tensors[i].requires_grad
-            inputs[idx] = tmp
         detached_inputs = detach_variable(tuple(inputs))
         if ctx.had_autocast_in_fwd:
             with torch.enable_grad(), torch.cuda.amp.autocast():