Revert "Remove deprecated _aminmax operator (#125995)"

This reverts commit 0116ffa.

Reverted #125995 on behalf of https://github.com/huydhn due to Sorry for reverting your change but we need to reland this after I get rid of all usage of _aminmax internally in Meta ([comment](#125995 (comment)))

aten/src/ATen/native/ReduceAllOps.cpp

@@ -8,6 +8,7 @@
 #include <ATen/Functions.h>
 #include <ATen/NativeFunctions.h>
 #else
+#include <ATen/ops/_aminmax_native.h>
 #include <ATen/ops/aminmax.h>
 #include <ATen/ops/empty.h>
 #include <ATen/ops/max.h>
@@ -65,4 +66,11 @@ Tensor& max_unary_out(const Tensor &self, Tensor& out) {
   return out;
 }
 
+// DEPRECATED: Use at::aminmax instead
+std::tuple<Tensor, Tensor> _aminmax_all(const Tensor &self) {
+  TORCH_WARN_ONCE("_aminmax is deprecated as of PyTorch 1.11 and will be removed in a future release. Use aminmax instead."
+                  " This warning will only appear once per process.");
+  return at::aminmax(self);
+}
+
 } // namespace at::native

aten/src/ATen/native/TensorCompare.cpp

@@ -20,6 +20,7 @@
 #include <ATen/Functions.h>
 #include <ATen/NativeFunctions.h>
 #else
+#include <ATen/ops/_aminmax_native.h>
 #include <ATen/ops/_assert_async_native.h>
 #include <ATen/ops/_functional_assert_async_native.h>
 #include <ATen/ops/_print_native.h>
@@ -681,6 +682,13 @@ std::tuple<Tensor, Tensor> qmin(const Tensor& self, int64_t dim, bool keepdim) {
       at::_make_per_tensor_quantized_tensor(min, self.q_scale(), self.q_zero_point()), min_indices);
 }
 
+// DEPRECATED: Use at::aminmax instead
+std::tuple<Tensor, Tensor> _aminmax(const Tensor& self, int64_t dim, bool keepdim) {
+  TORCH_WARN_ONCE("_aminmax is deprecated as of PyTorch 1.11 and will be removed in a future release. Use aminmax instead."
+                  " This warning will only appear once per process.");
+  return at::aminmax(self, dim, keepdim);
+}
+
 TORCH_IMPL_FUNC(clamp_out)
 (
  const Tensor& /*self*/,

aten/src/ATen/native/native_functions.yaml

@@ -3762,6 +3762,18 @@
 # This function should be deprecated in favor of differential_analytic_matrix_function in FunctionsManual.cpp
 - func: matrix_exp_backward(Tensor self, Tensor grad) -> Tensor
 
+# DEPRECATED: Use torch.aminmax instead
+- func: _aminmax(Tensor self) -> (Tensor, Tensor)
+  dispatch:
+    CPU, CUDA: _aminmax_all
+  autogen: _aminmax.out
+
+# DEPRECATED: Use torch.aminmax instead
+- func: _aminmax.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor, Tensor)
+  dispatch:
+    CPU, CUDA: _aminmax
+  autogen: _aminmax.dim_out
+
 - func: aminmax(Tensor self, *, int? dim=None, bool keepdim=False) -> (Tensor min, Tensor max)
   device_check: NoCheck   # TensorIterator
   structured_delegate: aminmax.out

aten/src/ATen/native/quantized/cuda/FusedObsFakeQuant.cu

@@ -7,7 +7,7 @@
 #ifndef AT_PER_OPERATOR_HEADERS
 #include <ATen/Functions.h>
 #else
-#include <ATen/ops/aminmax.h>
+#include <ATen/ops/_aminmax.h>
 #include <ATen/ops/_fake_quantize_per_tensor_affine_cachemask_tensor_qparams.h>
 #include <ATen/ops/fake_quantize_per_channel_affine.h>
 #include <ATen/ops/fake_quantize_per_channel_affine_cachemask.h>
@@ -148,7 +148,7 @@ void _calculate_moving_average(
   cudaStream_t cuda_stream = at::cuda::getCurrentCUDAStream();
 
   if (per_row_fq) {
-    std::tie(x_min, x_max) = at::aminmax(x, 1);
+    std::tie(x_min, x_max) = at::_aminmax(x, 1);
     float* x_min_data = x_min.data_ptr<float>();
     float* x_max_data = x_max.data_ptr<float>();
     int num_threads = std::min(size, (int64_t)512);
@@ -165,7 +165,7 @@ void _calculate_moving_average(
         size);
     C10_CUDA_KERNEL_LAUNCH_CHECK();
   } else {
-    std::tie(x_min, x_max) = at::aminmax(x);
+    std::tie(x_min, x_max) = at::_aminmax(x);
     float* x_min_data = x_min.data_ptr<float>();
     float* x_max_data = x_max.data_ptr<float>();
     // Moving Average Min/Max observer for activations

test/allowlist_for_publicAPI.json

@@ -1109,6 +1109,7 @@
     "_add_relu",
     "_add_relu_",
     "_addmm_activation",
+    "_aminmax",
     "_amp_foreach_non_finite_check_and_unscale_",
     "_amp_update_scale_",
     "_assert_async",
@@ -1318,7 +1319,6 @@
     "_values_copy",
     "_weight_norm",
     "_weight_norm_interface",
-    "aminmax",
     "autocast",
     "broadcast_shapes",
     "candidate",

test/expect/HasDecompTest.test_has_decomposition.expect

@@ -18,6 +18,10 @@ aten::_add_relu.Tensor
 aten::_add_relu.out
 aten::_add_relu_.Scalar
 aten::_add_relu_.Tensor
+aten::_aminmax
+aten::_aminmax.dim
+aten::_aminmax.dim_out
+aten::_aminmax.out
 aten::_amp_foreach_non_finite_check_and_unscale
 aten::_amp_foreach_non_finite_check_and_unscale.out
 aten::_amp_foreach_non_finite_check_and_unscale_

test/forward_backward_compatibility/check_forward_backward_compatibility.py

@@ -9,9 +9,6 @@
 from torch._C import parse_schema
 
 
-# Run by backwards_compat CI job
-
-
 # How to run this test locally:
 # 1 Have two virtual environments (eg conda env), one without PyTorch installed (venv_nightly)
 #   one with your local changes (venv_yours).
@@ -143,10 +140,6 @@
     ("onednn::qconv2d_pointwise", datetime.date(2024, 12, 31)),
     ("onednn::qconv3d_pointwise", datetime.date(2024, 12, 31)),
     ("onednn::qconv2d_pointwise.binary", datetime.date(2024, 12, 31)),
-    ("aten::_aminmax", datetime.date(2024, 12, 31)),
-    ("aten::_aminmax.dim", datetime.date(2024, 12, 31)),
-    ("aten::_aminmax.out", datetime.date(2024, 12, 31)),
-    ("aten::_aminmax.dim_out", datetime.date(2024, 12, 31)),
 ]
 
 ALLOW_LIST_COMPILED = [

test/mobile/model_test/coverage.yaml

@@ -1048,6 +1048,7 @@ uncovered_ops:
   aten::__is__: 83
   aten::__isnot__: 81
   aten::__not__: 32
+  aten::_aminmax: 4
   aten::_convolution: 12
   aten::_convolution.deprecated: 3
   aten::_make_per_tensor_quantized_tensor: 2

test/mobile/model_test/model_ops.yaml

@@ -30,6 +30,7 @@ root_operators:
   aten::__range_length: 106
   aten::__rshift__.int: 2
   aten::__xor__.bool: 16
+  aten::_aminmax: 18
   aten::_convolution: 27
   aten::_convolution.deprecated: 3
   aten::_infer_size: 9

test/test_reductions.py

@@ -1220,10 +1220,18 @@ def test_amax(self, device, dtype):
     def test_aminmax(self, device, dtype):
 
         def _amin_wrapper(x, dim=None, keepdims=False):
-            return torch.aminmax(x, dim=dim, keepdim=keepdims)[0]
+            with self.assertWarnsOnceRegex(UserWarning, "_aminmax is deprecated"):
+                if dim is None:
+                    return torch._aminmax(x)[0]
+                else:
+                    return torch._aminmax(x, dim, keepdims)[0]
 
         def _amax_wrapper(x, dim=None, keepdims=False):
-            return torch.aminmax(x, dim=dim, keepdim=keepdims)[1]
+            with self.assertWarnsOnceRegex(UserWarning, "_aminmax is deprecated"):
+                if dim is None:
+                    return torch._aminmax(x)[1]
+                else:
+                    return torch._aminmax(x, dim, keepdims)[1]
 
         self._test_minmax_helper(_amin_wrapper, np.amin, device, dtype)
         self._test_minmax_helper(_amax_wrapper, np.amax, device, dtype)

torch/_dynamo/trace_rules.py

@@ -373,6 +373,7 @@
         "torch._add_relu_",
         "torch._add_relu",
         "torch._addmm_activation",
+        "torch._aminmax",
         "torch._amp_foreach_non_finite_check_and_unscale_",
         "torch._amp_update_scale_",
         "torch._assert_async",