Add QInt8, QUInt8, QInt32 quantized scalar types with full quantizati… (#1531)

* Add QInt8, QUInt8, QInt32 quantized scalar types with full quantization support

- Uncomment and fix ScalarType enum entries (QInt8=12, QUInt8=13, QInt32=14)
- Fix incorrect QUInt32 name to QInt32 to match PyTorch
- Add torch.qint8, torch.quint8, torch.qint32 dtype aliases
- Add torch.is_quantized() and Tensor.is_quantized() methods
- Add IsQuantized() extension method on ScalarType

Native quantization bindings:
- Add THSTensor_quantize_per_tensor, THSTensor_quantize_per_channel (C++ and P/Invoke)
- Add THSTensor_dequantize (C++ and P/Invoke)
- Add THSTensor_q_scale, THSTensor_q_zero_point (C++ and P/Invoke)
- Add THSTensor_int_repr (C++ and P/Invoke)
- Add THSTensor_q_per_channel_scales, THSTensor_q_per_channel_zero_points, THSTensor_q_per_channel_axis (C++ and P/Invoke)

Managed API:
- Add torch.quantize_per_tensor() and torch.quantize_per_channel() static methods
- Add torch.dequantize() static method
- Add Tensor.dequantize(), Tensor.q_scale(), Tensor.q_zero_point() instance methods
- Add Tensor.int_repr() instance method
- Add Tensor.q_per_channel_scales(), Tensor.q_per_channel_zero_points(), Tensor.q_per_channel_axis() instance methods

Unit tests for all new functionality (13 tests)

* Address PR review: add quantized type ElementSize + fix tensor dispose leak

- Add QInt8 (1), QUInt8 (1), QInt32 (4) to ElementSize() so it no
  longer throws NotImplementedException for quantized types
- Split chained tensor().reshape() in QuantizePerChannel test to
  properly dispose the intermediate 1D tensor via 'using var'

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: alinpahontu2912

src/Native/LibTorchSharp/THSTensor.cpp

@@ -2306,3 +2306,48 @@ Tensor THSTensor_unflatten_names(Tensor tensor, const char** names, const int64_
 
     return nullptr;
 }
+
+Tensor THSTensor_quantize_per_tensor(const Tensor tensor, double scale, int64_t zero_point, int8_t scalar_type)
+{
+    CATCH_TENSOR(torch::quantize_per_tensor(*tensor, scale, zero_point, at::ScalarType(scalar_type)));
+}
+
+Tensor THSTensor_quantize_per_channel(const Tensor tensor, const Tensor scales, const Tensor zero_points, int64_t axis, int8_t scalar_type)
+{
+    CATCH_TENSOR(torch::quantize_per_channel(*tensor, *scales, *zero_points, axis, at::ScalarType(scalar_type)));
+}
+
+Tensor THSTensor_dequantize(const Tensor tensor)
+{
+    CATCH_TENSOR(tensor->dequantize());
+}
+
+double THSTensor_q_scale(const Tensor tensor)
+{
+    CATCH_RETURN(double, 0.0, tensor->q_scale());
+}
+
+int64_t THSTensor_q_zero_point(const Tensor tensor)
+{
+    CATCH_RETURN(int64_t, 0, tensor->q_zero_point());
+}
+
+Tensor THSTensor_int_repr(const Tensor tensor)
+{
+    CATCH_TENSOR(tensor->int_repr());
+}
+
+Tensor THSTensor_q_per_channel_scales(const Tensor tensor)
+{
+    CATCH_TENSOR(tensor->q_per_channel_scales());
+}
+
+Tensor THSTensor_q_per_channel_zero_points(const Tensor tensor)
+{
+    CATCH_TENSOR(tensor->q_per_channel_zero_points());
+}
+
+int64_t THSTensor_q_per_channel_axis(const Tensor tensor)
+{
+    CATCH_RETURN(int64_t, 0, tensor->q_per_channel_axis());
+}

src/Native/LibTorchSharp/THSTensor.h

@@ -1808,3 +1808,15 @@ EXPORT_API(Tensor) THSTensor_kaiser_window(const int64_t len, bool periodic, dou
 
 EXPORT_API(Tensor) THSTensor_stft(const Tensor x, int64_t n_fft, int64_t hop_length, int64_t win_length, const Tensor window, bool normalized, int64_t onesided, bool return_complex);
 EXPORT_API(Tensor) THSTensor_istft(const Tensor x, int64_t n_fft, int64_t hop_length, int64_t win_length, const Tensor window, bool center, bool normalized, int64_t onesided, int64_t length, bool return_complex);
+
+// Quantization Ops
+
+EXPORT_API(Tensor) THSTensor_quantize_per_tensor(const Tensor tensor, double scale, int64_t zero_point, int8_t scalar_type);
+EXPORT_API(Tensor) THSTensor_quantize_per_channel(const Tensor tensor, const Tensor scales, const Tensor zero_points, int64_t axis, int8_t scalar_type);
+EXPORT_API(Tensor) THSTensor_dequantize(const Tensor tensor);
+EXPORT_API(double) THSTensor_q_scale(const Tensor tensor);
+EXPORT_API(int64_t) THSTensor_q_zero_point(const Tensor tensor);
+EXPORT_API(Tensor) THSTensor_int_repr(const Tensor tensor);
+EXPORT_API(Tensor) THSTensor_q_per_channel_scales(const Tensor tensor);
+EXPORT_API(Tensor) THSTensor_q_per_channel_zero_points(const Tensor tensor);
+EXPORT_API(int64_t) THSTensor_q_per_channel_axis(const Tensor tensor);

src/TorchSharp/PInvoke/LibTorchSharp.THSTensor.cs

@@ -2206,6 +2206,33 @@ internal static extern IntPtr THSTensor_upsample_nearest3d(IntPtr input,
         internal static extern IntPtr THSTensor_histogram_out_t(IntPtr input, IntPtr bins, IntPtr weight, bool density, out IntPtr hist, out IntPtr bin_edges, out IntPtr r_bin_edges);
         [DllImport("LibTorchSharp")]
         internal static extern IntPtr THSTensor_histogram_out_i(IntPtr input, long bins, IntPtr range, int length, IntPtr weight, bool density, out IntPtr hist, out IntPtr bin_edges, out IntPtr r_bin_edges);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern IntPtr THSTensor_quantize_per_tensor(IntPtr tensor, double scale, long zero_point, sbyte scalar_type);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern IntPtr THSTensor_quantize_per_channel(IntPtr tensor, IntPtr scales, IntPtr zero_points, long axis, sbyte scalar_type);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern IntPtr THSTensor_dequantize(IntPtr tensor);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern double THSTensor_q_scale(IntPtr tensor);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern long THSTensor_q_zero_point(IntPtr tensor);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern IntPtr THSTensor_int_repr(IntPtr tensor);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern IntPtr THSTensor_q_per_channel_scales(IntPtr tensor);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern IntPtr THSTensor_q_per_channel_zero_points(IntPtr tensor);
+
+        [DllImport("LibTorchSharp")]
+        internal static extern long THSTensor_q_per_channel_axis(IntPtr tensor);
     }
 #pragma warning restore CA2101
 }

src/TorchSharp/Tensor/Tensor.cs

@@ -271,6 +271,95 @@ internal IntPtr MoveHandle()
             /// </summary>
             public bool is_complex() => torch.is_complex(dtype);
 
+            /// <summary>
+            /// Returns True if the data type of input is a quantized data type i.e., one of torch.qint8, torch.quint8, and torch.qint32.
+            /// </summary>
+            public bool is_quantized() => torch.is_quantized(dtype);
+
+            /// <summary>
+            /// Given a quantized Tensor, returns a dequantized (float) Tensor.
+            /// </summary>
+            public Tensor dequantize()
+            {
+                var res = NativeMethods.THSTensor_dequantize(Handle);
+                if (res == IntPtr.Zero) { CheckForErrors(); }
+                return new Tensor(res);
+            }
+
+            /// <summary>
+            /// Given a quantized Tensor, returns the scale of the quantization as a double.
+            /// </summary>
+            public double q_scale()
+            {
+                var res = NativeMethods.THSTensor_q_scale(Handle);
+                CheckForErrors();
+                return res;
+            }
+
+            /// <summary>
+            /// Given a quantized Tensor, returns the zero_point of the quantization as a long.
+            /// </summary>
+            public long q_zero_point()
+            {
+                var res = NativeMethods.THSTensor_q_zero_point(Handle);
+                CheckForErrors();
+                return res;
+            }
+
+            /// <summary>
+            /// Given a quantized Tensor, returns a Tensor of the underlying integer representation.
+            /// </summary>
+            public Tensor int_repr()
+            {
+                var res = NativeMethods.THSTensor_int_repr(Handle);
+                if (res == IntPtr.Zero) { CheckForErrors(); }
+                return new Tensor(res);
+            }
+
+            /// <summary>
+            /// Given a quantized Tensor quantized per channel, returns a Tensor of the scales of the quantization for each channel.
+            /// </summary>
+            public Tensor q_per_channel_scales()
+            {
+                var res = NativeMethods.THSTensor_q_per_channel_scales(Handle);
+                if (res == IntPtr.Zero) { CheckForErrors(); }
+                return new Tensor(res);
+            }
+
+            /// <summary>
+            /// Given a quantized Tensor quantized per channel, returns a Tensor of the zero points of the quantization for each channel.
+            /// </summary>
+            public Tensor q_per_channel_zero_points()
+            {
+                var res = NativeMethods.THSTensor_q_per_channel_zero_points(Handle);
+                if (res == IntPtr.Zero) { CheckForErrors(); }
+                return new Tensor(res);
+            }
+
+            /// <summary>
+            /// Given a quantized Tensor quantized per channel, returns the axis along which per channel quantization is applied.
+            /// </summary>
+            public long q_per_channel_axis()
+            {
+                var res = NativeMethods.THSTensor_q_per_channel_axis(Handle);
+                CheckForErrors();
+                return res;
+            }
+
+            internal Tensor _quantize_per_tensor(double scale, long zero_point, ScalarType dtype)
+            {
+                var res = NativeMethods.THSTensor_quantize_per_tensor(Handle, scale, zero_point, (sbyte)dtype);
+                if (res == IntPtr.Zero) { CheckForErrors(); }
+                return new Tensor(res);
+            }
+
+            internal Tensor _quantize_per_channel(Tensor scales, Tensor zero_points, long axis, ScalarType dtype)
+            {
+                var res = NativeMethods.THSTensor_quantize_per_channel(Handle, scales.Handle, zero_points.Handle, axis, (sbyte)dtype);
+                if (res == IntPtr.Zero) { CheckForErrors(); }
+                return new Tensor(res);
+            }
+
             /// <summary>
             /// Returns True if the input is a single element tensor which is not equal to zero after type conversions,
             /// i.e. not equal to torch.tensor([0.]) or torch.tensor([0]) or torch.tensor([False]).
@@ -7359,9 +7448,9 @@ public enum ScalarType : sbyte
             ComplexFloat32 = 9,
             ComplexFloat64 = 10,
             Bool = 11,
-            //QInt8 = 12,
-            //QUInt8 = 13,
-            //QUInt32 = 14,
+            QInt8 = 12,
+            QUInt8 = 13,
+            QInt32 = 14,
             BFloat16 = 15
         }
 
@@ -7493,6 +7582,18 @@ public static bool is_complex(ScalarType type)
             }
         }
 
+        public static bool is_quantized(ScalarType type)
+        {
+            switch (type) {
+            case ScalarType.QInt8:
+            case ScalarType.QUInt8:
+            case ScalarType.QInt32:
+                return true;
+            default:
+                return false;
+            }
+        }
+
         public static long max_int_value(ScalarType type)
         {
             switch (type) {
@@ -7543,6 +7644,10 @@ public static long max_int_value(ScalarType type)
         public static ScalarType cfloat = ScalarType.ComplexFloat32;
         public static ScalarType cdouble = ScalarType.ComplexFloat64;
 
+        public static ScalarType qint8 = ScalarType.QInt8;
+        public static ScalarType quint8 = ScalarType.QUInt8;
+        public static ScalarType qint32 = ScalarType.QInt32;
+
         /// <summary>
         /// Creates a new dispose scope for the current thread. Any tensor created within the dispose scope will
         /// be automatically disposed once the dispose scope is disposed.

src/TorchSharp/Tensor/TensorExtensionMethods.cs

@@ -303,6 +303,9 @@ public static int ElementSize(this ScalarType type)
                 ScalarType.ComplexFloat32 => 8,
                 ScalarType.ComplexFloat64 => 16,
                 ScalarType.Bool => 1,
+                ScalarType.QInt8 => 1,
+                ScalarType.QUInt8 => 1,
+                ScalarType.QInt32 => 4,
                 ScalarType.BFloat16 => 2,
                 _ => throw new NotImplementedException()
             };
@@ -368,6 +371,23 @@ internal static bool IsComplex(this ScalarType type)
             }
         }
 
+        /// <summary>
+        /// Indicates whether a given element type is quantized.
+        /// </summary>
+        /// <param name="type">The input type.</param>
+        /// <returns></returns>
+        internal static bool IsQuantized(this ScalarType type)
+        {
+            switch (type) {
+            case ScalarType.QInt8:
+            case ScalarType.QUInt8:
+            case ScalarType.QInt32:
+                return true;
+            default:
+                return false;
+            }
+        }
+
         /// <summary>
         /// Save the tensor in a .NET-specific format.
         /// </summary>

src/TorchSharp/Tensor/torch.PointwiseOps.cs

@@ -779,6 +779,47 @@ public static Tensor fake_quantize_per_channel_affine(Tensor input, Tensor scale
         public static Tensor fake_quantize_per_tensor_affine(Tensor input, Tensor scale, Tensor zero_point, long quant_min, long quant_max)
             => throw new NotImplementedException();
 
+        // https://pytorch.org/docs/stable/generated/torch.quantize_per_tensor
+        /// <summary>
+        /// Converts a float tensor to a quantized tensor with given scale and zero point.
+        /// </summary>
+        /// <param name="input">Float tensor to quantize</param>
+        /// <param name="scale">Scale to apply in quantization formula</param>
+        /// <param name="zero_point">Offset in integer value that maps to float zero</param>
+        /// <param name="dtype">The desired data type of returned tensor. Must be a quantized type (torch.qint8, torch.quint8, or torch.qint32).</param>
+        /// <returns>A newly quantized tensor</returns>
+        public static Tensor quantize_per_tensor(Tensor input, double scale, long zero_point, ScalarType dtype)
+        {
+            if (!is_quantized(dtype))
+                throw new ArgumentException("dtype must be a quantized type (QInt8, QUInt8, or QInt32)", nameof(dtype));
+            return input._quantize_per_tensor(scale, zero_point, dtype);
+        }
+
+        // https://pytorch.org/docs/stable/generated/torch.quantize_per_channel
+        /// <summary>
+        /// Converts a float tensor to a per-channel quantized tensor with given scales and zero points.
+        /// </summary>
+        /// <param name="input">Float tensor to quantize</param>
+        /// <param name="scales">Float 1D tensor of scales to use, size should match input.size(axis)</param>
+        /// <param name="zero_points">Integer 1D tensor of offsets to use, size should match input.size(axis)</param>
+        /// <param name="axis">Dimension on which to apply per-channel quantization</param>
+        /// <param name="dtype">The desired data type of returned tensor. Must be a quantized type (torch.qint8, torch.quint8, or torch.qint32).</param>
+        /// <returns>A newly quantized tensor</returns>
+        public static Tensor quantize_per_channel(Tensor input, Tensor scales, Tensor zero_points, long axis, ScalarType dtype)
+        {
+            if (!is_quantized(dtype))
+                throw new ArgumentException("dtype must be a quantized type (QInt8, QUInt8, or QInt32)", nameof(dtype));
+            return input._quantize_per_channel(scales, zero_points, axis, dtype);
+        }
+
+        // https://pytorch.org/docs/stable/generated/torch.dequantize
+        /// <summary>
+        /// Returns an fp32 Tensor by dequantizing a quantized Tensor.
+        /// </summary>
+        /// <param name="input">A quantized tensor</param>
+        /// <returns>A dequantized (float) tensor</returns>
+        public static Tensor dequantize(Tensor input) => input.dequantize();
+
         // https://pytorch.org/docs/stable/generated/torch.fix
         /// <summary>
         /// Returns a new tensor with the truncated integer values of the elements of input.