// CHECK-DAG:  %[[CST_1:.*]] = "tfl.pseudo_const"() <{value = dense<0.000000e+00> : tensor<4x4xf16>}> : () -> tensor<4x4xf16>
  // CHECK-DAG:  %[[DQ_0:.*]] = "tfl.dequantize"(%[[CST_0]]) : (tensor<4x4xf16>) -> tensor<4x4xf32>
  // CHECK-DAG:  %[[DQ_1:.*]] = "tfl.dequantize"(%[[CST_1]]) : (tensor<4x4xf16>) -> tensor<4x4xf32>

    "Bitcast",
    "BroadcastArgs",
    "BroadcastTo",
    "CheckNumerics",
    "ConcatV2",
    "ConjugateTranspose",
    "DebugGradientIdentity",
    "DeepCopy",
    "DepthToSpace",
    "Dequantize",
    "EditDistance",
    "Empty",
    "EnsureShape",
    "ExpandDims",
    "ExtractImagePatches",
    "ExtractVolumePatches",
    "FakeQuantWithMinMaxArgs",
    "FakeQuantWithMinMaxArgsGradient",

// CHECK-NO-UNPACK: %[[DEQUANTIZE:.+]] = stablehlo.uniform_dequantize %[[QUANTIZE_1]] : (tensor<1x3x!quant.uniform<i8:f32, {{.*}}>>) -> tensor<1x3xf32>
// CHECK-NO-UNPACK: return %[[DEQUANTIZE]] : tensor<1x3xf32>

// -----

// Tests that a simple dot_general without CustomAggregators is not quantized.

    METHOD_UNSPECIFIED = 0;  // go/do-include-enum-unspecified

    // Apply default weight-only quantization. Weights are quantized during
    // conversion, then dequantized during inference.
    // Activation: f32, Weight: qi8, Bias: f32
    WEIGHT_ONLY = 1;

    // Apply default dynamic range quantization. Quantized tensor value's
    // ranges are determined during graph runtime.

// Returns element type from attribute Type 'type_attr'.
mlir::Type GetShapeStrippedType(mlir::TypeAttr type_attr);

// Returns true if 'val' is not from Quantize op or
// from Quantize Op with same quant type as 'qtype_attr'
bool NotFromQuantOpOrSameQuantType(mlir::Value val, mlir::TypeAttr qtype_attr);

}  // namespace tflite

A. Unique TensorFlower <******@****.***> 1713119208 -0700

                  const tensorflow::quantization::PyFunctionLibrary*
                      quantization_py_function_library = nullptr);

// Quantize the model with calibration data. Throw errors if `fully_quantize`
// is specified by the calibration data are not sufficient to quantize the
// model.
PyObject* MlirQuantizeModel(PyObject* data, bool disable_per_channel,
                            bool fully_quantize, int inference_type,

// RUN: stablehlo-quant-opt %s -split-input-file -stablehlo-prepare-quantize=bit-width=4 -verify-diagnostics | FileCheck %s

// CHECK-LABEL: func @dot_int4
// CHECK-SAME: (%[[ARG_0:.*]]: tensor<?x3xf32>) -> tensor<?x2xf32>
func.func @dot_int4(%arg0: tensor<?x3xf32>) -> tensor<?x2xf32> {
  // CHECK: %[[cst:.*]] = stablehlo.constant
  // CHECK: %[[q1:.*]] = "quantfork.qcast"(%[[cst]])
  // CHECK-SAME: quant.uniform<i8:f32, 0.0040316890267764818:127>

// RUN: tf-quant-opt %s -split-input-file -quant-prepare-quantize='post-training-quantize=true enable-per-channel-quantization=true' | FileCheck %s

module {
  func.func private @conv_with_bias_and_relu(%arg0: tensor<1x3x4x3xf32>) -> tensor<*xf32> {
    %cst = "tf.Const"() {device = "", value = dense<[7.11401462, 7.05456924]> : tensor<2xf32>} : () -> tensor<2xf32>

// CHECK-LABEL: remove_quantized_io
func.func @remove_quantized_io(
  %arg0: tensor<1x10x!quant.uniform<i8:f32, 0.1:-128>>,
  %arg1: tensor<1x5xf32>) -> (tensor<1x10x!quant.uniform<i8:f32, 0.2:42>>, tensor<1x5xf32>) {
  %0 = "tf.MyRequantize"(%arg0) : (tensor<1x10x!quant.uniform<i8:f32, 0.1:-128>>) -> tensor<1x10x!quant.uniform<i8:f32, 0.2:42>>
  %1 = "tf.Intermediate"(%arg1) : (tensor<1x5xf32>) -> tensor<1x5xf32>