Remove _dequantize_per_tensor (#26681)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26681

att

Test Plan:
ci

Imported from OSS

Differential Revision: D17542833

fbshipit-source-id: 653e906b0e146763609c69ef0de7f9cf38621586

aten/src/ATen/core/OpsAlreadyMovedToC10.cpp

@@ -1153,7 +1153,6 @@ bool aten_op_is_not_moved_to_c10_yet(const c10::OperatorName& opName) {
     #endif
         {"aten::quantize_per_tensor", ""},
         {"aten::quantize_per_channel", ""},
-        {"aten::_dequantize_per_tensor", ""},
         {"aten::q_per_channel_axis", ""},
         {"aten::qscheme", ""},
         {"aten::to", "dtype_layout"},

aten/src/ATen/native/native_functions.yaml

@@ -3536,11 +3536,6 @@
   dispatch:
     QuantizedCPU: dequantize_quant
 
-- func: _dequantize_per_tensor(Tensor self, float scale, int zero_point, ScalarType dtype) -> Tensor
-  variants: function
-  dispatch:
-    CPU: dequantize_per_tensor_cpu
-
 - func: q_scale(Tensor self) -> float
   use_c10_dispatcher: full
   variants: function, method

aten/src/ATen/native/quantized/QTensor.cpp

@@ -30,30 +30,6 @@ Tensor dequantize_quant(const Tensor& self) {
   return get_qtensorimpl(self)->quantizer()->dequantize(self);
 }
 
-Tensor dequantize_per_tensor_cpu(
-    const Tensor& self,
-    double scale,
-    int64_t zero_point,
-    ScalarType dtype) {
-  TORCH_CHECK(
-      isQIntType(toQIntType(self.scalar_type())),
-      "Scalar type for quantized Tensor must have same underlying type as input.");
-  TORCH_CHECK(
-      dtype == toQIntType(self.scalar_type()),
-      "ScalarType argument must match the corresponding quantized scalar type of input integer Tensor");
-  // scalar type of output Tensor is hard-coded as float
-  Tensor f = at::empty(self.sizes(), self.options().dtype(at::kFloat));
-  AT_DISPATCH_QINT_TYPES(
-      toQIntType(self.scalar_type()), "dequantize_linear_cpu", [&]() {
-        underlying_t* qdata = self.data_ptr<underlying_t>();
-        auto* fdata = f.data_ptr<float>();
-        for (int i = 0; i < self.numel(); ++i) {
-          fdata[i] = (static_cast<float>(qdata[i]) - zero_point) * scale;
-        }
-      });
-  return f;
-}
-
 double q_scale_quant(const Tensor& self) {
   auto quantizer = get_qtensorimpl(self)->quantizer();
   TORCH_CHECK(quantizer->qscheme() == kPerTensorAffine);

test/test_quantized_tensor.py

@@ -129,15 +129,7 @@ class TestQuantizedTensor(TestCase):
         rqr = qr.dequantize()
         self.assertTrue(np.allclose(r.numpy(), rqr.numpy(), atol=2 / scale))
 
-    def test_qtensor_dequantize_per_tensor(self):
+    def test_qtensor_quantize_per_channel(self):
-        t = torch.arange(-10, 10, dtype=torch.int8)
-        scale = 3
-        zero_point = 2
-        qt = torch._dequantize_per_tensor(t, scale, zero_point, torch.qint8)
-        qt2 = torch._make_per_tensor_quantized_tensor(t, scale, zero_point)
-        self.assertEqual(qt, qt2.dequantize())
-
-    def test_qtensor_per_channel_affine(self):
         r = torch.rand(3, 2, dtype=torch.float) * 4 - 2
         scales = torch.tensor([0.2, 0.03], dtype=torch.double)
         zero_points = torch.tensor([5, 10], dtype=torch.long)