Move CUDA tensors to CPU before moving to XLA. (#6060)

test/test_core_aten_ops.py

@@ -5,10 +5,24 @@
 import torch
 import torch._export
 
+import os
 import tempfile
 import unittest
 
 
+def onlyIfTorchSupportsCUDA(fn):
+  return unittest.skipIf(
+      not torch.cuda.is_available(), reason="requires PyTorch CUDA support")(
+          fn)
+
+
+def onlyIfPJRTDeviceIsCUDA(fn):
+  return unittest.skipIf(
+      os.environ.get("PJRT_DEVICE") not in ("GPU", "CUDA"),
+      reason="requires CUDA as PJRT_DEVICE")(
+          fn)
+
+
 def diff_output(testcase, output1, output2, rtol, atol):
   if isinstance(output1, torch.Tensor):
     testcase.assertIsInstance(output2, torch.Tensor)
@@ -4664,6 +4678,18 @@ def test_aten_where_self_0(self):
     kwargs = dict()
     run_export_and_compare(self, torch.ops.aten.where.self, args, kwargs)
 
+  @onlyIfTorchSupportsCUDA
+  @onlyIfPJRTDeviceIsCUDA
+  def test_aten_move_cuda_to_xla(self):
+    # Assumes CPU-XLA data movement works.
+    t_cpu = torch.arange(5)
+    t_cuda = t_cpu.to("cuda")
+    # Move tensor CUDA -> XLA.
+    t_xla = t_cuda.to(xm.xla_device())
+    # Move the XLA tensor back to CPU, and check that it is the same as
+    # the original CPU tensor.
+    self.assertTrue(torch.equal(t_cpu, t_xla.cpu()))
+
 
 if __name__ == '__main__':
   unittest.main()

torch_xla/csrc/runtime/tensor_source.h

@@ -54,8 +54,13 @@ class AtenSource : public TensorSource {
     if (target_torch_type != tensor.type().scalarType()) {
       TORCH_LAZY_COUNTER("AtenSourceDowncasts", 1);
     }
-    tensor_ = std::move(tensor.to(target_torch_type, /*non_blocking=*/false,
-                                  /*copy=*/true, at::MemoryFormat::Contiguous));
+    // TODO(ysiraichi): check, first, if tensor lives in a device that the
+    // current PjRt client has access. If so, we don't need to go through the
+    // CPU.
+    tensor_ = std::move(
+        tensor.to(at::TensorOptions().device(at::kCPU).dtype(target_torch_type),
+                  /*non_blocking=*/false,
+                  /*copy=*/true, at::MemoryFormat::Contiguous));
   }
 
   const void* data() const override { return tensor_.const_data_ptr(); }