Add a basic shim and stable::Tensor is_contiguous API (#156228)

Add a limited is_contiguous in shim, stable::Tensor API with a test case
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156228
Approved by: https://github.com/desertfire

test/cpp_extensions/libtorch_agnostic_extension/libtorch_agnostic/csrc/kernel.cpp

@@ -237,3 +237,20 @@ STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
 STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
   m.impl("divide_neg_exp", &boxed_divide_neg_exp);
 }
+
+bool is_contiguous(Tensor t) {
+  return t.is_contiguous();
+}
+
+void boxed_is_contiguous(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
+  bool res = is_contiguous(to<Tensor>(stack[0]));
+  stack[0] = from(res);
+}
+
+STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
+  m.def("is_contiguous(Tensor t) -> bool");
+}
+
+STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
+  m.impl("is_contiguous", &boxed_is_contiguous);
+}

test/cpp_extensions/libtorch_agnostic_extension/libtorch_agnostic/ops.py

@@ -104,3 +104,15 @@ def divide_neg_exp(t) -> Tensor:
     Returns: divide(neg(t), exp(t))
     """
     return torch.ops.libtorch_agnostic.divide_neg_exp.default(t)
+
+
+def is_contiguous(t) -> bool:
+    """
+    Returns a bool indicating if the input tensor is contiguous
+
+    Args:
+        t: Tensor
+
+    Returns: is_contiguous(t)
+    """
+    return torch.ops.libtorch_agnostic.is_contiguous.default(t)

test/cpp_extensions/libtorch_agnostic_extension/test/test_libtorch_agnostic.py

@@ -142,6 +142,13 @@ if not IS_WINDOWS:
                     curr_mem = torch.cuda.memory_allocated(device)
                     self.assertEqual(curr_mem, init_mem)
 
+        def test_is_contiguous(self, device):
+            import libtorch_agnostic
+
+            t = torch.rand(2, 7, device=device)
+            self.assertTrue(libtorch_agnostic.ops.is_contiguous(t))
+            self.assertFalse(libtorch_agnostic.ops.is_contiguous(t.transpose(0, 1)))
+
         # TODO: Debug this:
         # torch._dynamo.exc.TorchRuntimeError: Dynamo failed to run FX node with fake tensors:
         # call_function libtorch_agnostic.my_ones_like.default(*(FakeTensor(..., size=(3, 1)), 'cpu'),