colo init context add device attr. (#866)

colossalai/tensor/colo_tensor.py

@@ -58,6 +58,10 @@ class ColoTensor(object):
     def shape(self):
         return torch.Size(self._size)
 
+    @property
+    def device(self):
+        return self._torch_tensor.device
+
     def size(self, dim=None):
         if dim is None:
             return self.shape
@@ -105,14 +109,14 @@ class ColoTensor(object):
                                              device=self._device)
         return self._torch_tensor
 
-    def set_spec(self, spec: str, lazy_shard: bool=False) -> None:
+    def set_spec(self, spec: str, lazy_shard: bool = False) -> None:
         self._shard_spec = spec
         if lazy_shard == False:
             self._shard()
 
     def _shard(self):
         assert self._shard_spec is not None, 'You should call set_spec() before _shard() ColoTensor.'
-        if self._shard_spec == "1Drow": # TODO It actually represents the sharding layout for Linear-1Drow-weight, but we make it simpler now.
+        if self._shard_spec == "1Drow":    # TODO It actually represents the sharding layout for Linear-1Drow-weight, but we make it simpler now.
             num_partition = gpc.get_world_size(ParallelMode.TENSOR)
             local_rank = gpc.get_local_rank(ParallelMode.TENSOR)
             dim = -1
@@ -121,11 +125,11 @@ class ColoTensor(object):
             # Reshape to get shard for this rank and we don't want autograd
             # recording here for the narrow op and 'local_shard' should be a
             # leaf variable in the autograd graph.
-            self._torch_tensor = self._torch_tensor.narrow(dim, 
-                    local_rank * chunk_size, chunk_size).detach().contiguous() # TODO Shall we clone() here since detach() will point to the old tensor?
+            self._torch_tensor = self._torch_tensor.narrow(dim, local_rank * chunk_size, chunk_size).detach(
+            ).contiguous()    # TODO Shall we clone() here since detach() will point to the old tensor?
             self._torch_tensor.requires_grad = self._requires_grad
             self._size = self._torch_tensor.size()
-            self._device = device # TODO A `fake` device now because torch_tensor.device always = cpu
+            self._device = device    # TODO A `fake` device now because torch_tensor.device always = cpu
 
     @classmethod
     def __torch_function__(cls, func, types, args=(), kwargs=None):

colossalai/utils/model/colo_init_context.py

+from colossalai.utils.cuda import get_current_device
 from .utils import InsertPostInitMethodToModuleSubClasses
 import torch
 # from colossalai.logging import get_dist_logger
@@ -8,9 +9,15 @@ from colossalai.tensor import ColoTensor
 
 class ColoInitContext(InsertPostInitMethodToModuleSubClasses):
 
-    def __init__(self, lazy_memory_allocate=False):
+    def __init__(self, lazy_memory_allocate: bool = False, device: torch.device = torch.device('cpu')):
+        """
+        Args:
+            lazy_memory_allocate (bool, optional): whether to allocate memory for the parameter tensors. Defaults to False.
+            device (torch.device, optional): the device parameters initialized are resident on. Defaults to torch.device('cpu').
+        """
         super().__init__()
         self._lazy_memory_allocate = lazy_memory_allocate
+        self._device = device
 
     def _post_init_method(self, module: torch.nn.Module, *args, **kwargs):
         """
@@ -26,4 +33,5 @@ class ColoInitContext(InsertPostInitMethodToModuleSubClasses):
         save_torch_payload = True if not self._lazy_memory_allocate else False
         for name, param in name_list:
             delattr(module, name)
-            setattr(module, name, ColoTensor.init_from_torch_tensor(tensor=param, save_payload=save_torch_payload))
+            setattr(module, name,
+                    ColoTensor.init_from_torch_tensor(tensor=param.to(self._device), save_payload=save_torch_payload))

tests/test_tensor/test_context.py

@@ -5,17 +5,16 @@ import torch
 from colossalai.tensor import ColoTensor
 from copy import deepcopy
 
+from colossalai.utils.cuda import get_current_device
 
-def test_linear():
+
+def test_lazy_init():
     in_dim = 4
     out_dim = 5
 
     with ColoInitContext(lazy_memory_allocate=True) as ctx:
         fc = torch.nn.Linear(in_dim, out_dim, bias=True)
 
-    print(fc.weight.numel())
-    print(fc.bias.numel())
-
     # lazy_memory_allocate=True, no payload is maintained
     assert fc.weight._torch_tensor.numel() == 0
 
@@ -23,5 +22,18 @@ def test_linear():
     assert fc.weight._torch_tensor.numel() == in_dim * out_dim
 
 
+def test_device():
+    in_dim = 4
+    out_dim = 5
+
+    with ColoInitContext(lazy_memory_allocate=True, device=get_current_device()) as ctx:
+        fc = torch.nn.Linear(in_dim, out_dim, bias=True)
+
+    # eval an lazy parameter
+    fc.weight.torch_tensor()
+    assert fc.weight.device == get_current_device()
+
+
 if __name__ == '__main__':
-    test_linear()
+    test_lazy_init()
+    test_device()