Add empty_like for NF4Tensor to support offloading (#881)

Driss, we can confirm semantics when you're back but I'm fairly confident this is okay

test/dtypes/test_nf4.py

@@ -20,6 +20,7 @@
     run_tests,
 )
 from torchao.dtypes.nf4tensor import (
+    NF4Tensor,
     linear_nf4,
     to_nf4,
     _INNER_TENSOR_NAMES_FOR_SHARDING,
@@ -270,6 +271,15 @@ def test_chunk_size_equivalence(self, dtype: torch.dtype, shape, chunk_size):
 
         torch.testing.assert_close(nf4_patched.quantized_data, nf4_base.quantized_data)
 
+    @unittest.skipIf(not torch.cuda.is_available(), "Need CUDA available")
+    @parametrize("input_size", [(512 * 512,), (512, 512)])
+    def test_empty_like(self, input_size: Union[Tuple[int], int]):
+        nf4_tensor = to_nf4(torch.rand(input_size, device="cuda"))
+        new_tensor = torch.empty_like(nf4_tensor, device="cpu")
+        self.assertTrue(isinstance(new_tensor, NF4Tensor))
+        self.assertEqual(new_tensor.get_device(), -1)  # that it's on CPU
+        self.assertEqual(new_tensor.size(), nf4_tensor.size())
+
 
 class TestFSDPOps(TestCase):
     @parametrize("input_size", [512 * 512, (512 * 512,), (512, 512)])

torchao/dtypes/nf4tensor.py

@@ -164,6 +164,20 @@ def nf4_detach(aten_op, args, kwargs=None):
     return NF4Tensor(*construct_nf4_args(nf4tensor, updated_attrs))
 
 
+@implements(
+    [
+        aten.empty_like.default,
+    ]
+)
+def nf4_empty_like(aten_op, args, kwargs=None):
+    nf4tensor = args[0]
+    updated_attrs = apply_to_inner_tensors(nf4tensor, aten_op, args[1:], kwargs)
+    if kwargs is not None and len(kwargs):
+        for key, value in kwargs.items():
+            updated_attrs[key] = value
+    return NF4Tensor(*construct_nf4_args(nf4tensor, updated_attrs))
+
+
 @implements(
     [
         aten.split.Tensor,