[BE]: Replace clone detach with detach clone to be more efficient (#144469)

Skylion007 · pytorchmergebot · commit bbec35f028aa · 2025-01-09T18:28:39.000Z
Follow up to #144270 and fix some vulkan code Pull Request resolved: #144469 Approved by: https://github.com/awgu
diff --git a/aten/src/ATen/native/vulkan/ops/Random.cpp b/aten/src/ATen/native/vulkan/ops/Random.cpp
@@ -70,7 +70,7 @@ static Tensor rand_like(
   // numbers from a uniform distribution on the interval [0,1). To match the CPU
   // implementation, we simplify the range to [0,1] and tolerate the small
   // chance of 1 being sampled.
-  return input_arg.clone().detach().uniform_(0.0, 1.0);
+  return input_arg.detach().clone().uniform_(0.0, 1.0);
 }
 
 static Tensor& normal_(
@@ -129,7 +129,7 @@ static Tensor randn_like(
     const std::optional<c10::MemoryFormat> /* not implemented */) {
   // Returns a tensor with the same size as input that is filled with random
   // numbers from a normal distribution with mean 0 and standard deviation 1.
-  return input_arg.clone().detach().normal_(0.0, 1.0);
+  return input_arg.detach().clone().normal_(0.0, 1.0);
 }
 
 TORCH_LIBRARY_IMPL(aten, Vulkan, m) {
diff --git a/test/cpp/lazy/test_lazy_ops_util.cpp b/test/cpp/lazy/test_lazy_ops_util.cpp
@@ -140,7 +140,7 @@ void TestBackward(
     const torch::Tensor& input = inputs[i];
     if (input.defined()) {
       torch::Tensor oinput =
-          input.clone().detach().set_requires_grad(input.requires_grad());
+          input.detach().clone().set_requires_grad(input.requires_grad());
       input_vars.push_back(oinput);
 
       torch::Tensor xinput = CopyToDevice(input, device)
diff --git a/torch/csrc/utils/tensor_new.cpp b/torch/csrc/utils/tensor_new.cpp
@@ -1524,8 +1524,8 @@ Tensor new_tensor(
     if (THPVariable_Check(data)) {
       auto ret = PyErr_WarnEx(
           PyExc_UserWarning,
-          "To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() "
-          "or sourceTensor.clone().detach().requires_grad_(True), rather than tensor.new_tensor(sourceTensor).",
+          "To copy construct from a tensor, it is recommended to use sourceTensor.detach().clone() "
+          "or sourceTensor.detach().clone().requires_grad_(True), rather than tensor.new_tensor(sourceTensor).",
           1);
       if (ret != 0)
         throw python_error();
