compile still not working yet

Author: jerryzh168

tutorials/developer_api_guide/tensor_parallel.py

@@ -177,17 +177,16 @@ def rowwise_shard(m: torch.nn.Module, mesh: DeviceMesh) -> torch.nn.Module:
 
     y_colwise = d_up(input_dtensor)
     print("y_colwise:", y_colwise.shape)
-    # doesn't work, see BUG in rowwise_shard()
     print("result:", d_dn(y_colwise))
     print("Distributed works!")
 
     # doesn't work
     # [rank0]: torch._dynamo.exc.TorchRuntimeError: Failed running call_function <built-in function linear>(*(DTensor(local_tensor=FakeTensor(..., device='cuda:0', size=(128, 1024)), device_mesh=DeviceMesh('cuda', [0, 1,
     # 2, 3]), placements=(Replicate(),)), DTensor(local_tensor=MyDTypeTensorTP(data=FakeTensor(..., device='cuda:0', size=(128, 1024)), shape=torch.Size([1024, 1024]), device=cuda:0, dtype=torch.float32, requires_grad=False), device_mesh=DeviceMesh('cuda', [0, 1, 2, 3]), placements=(Shard(dim=0),)), None), **{}):
     # [rank0]: a and b must have same reduction dim, but got [128, 1024] X [128, 1024].
-    c_up = torch.compile(d_up)
-    c_dn = torch.compile(d_dn)
-    print("compiled result:", c_dn(c_up(input_dtensor)))
-    print("torch.compile works!")
+    # c_up = torch.compile(d_up)
+    # c_dn = torch.compile(d_dn)
+    # print("compiled result:", c_dn(c_up(input_dtensor)))
+    # print("torch.compile works!")
 
     dist.destroy_process_group()