[Quant Tool] Prevent int32 quantized bias from clipping by adjusting …

…the weight's scale (microsoft#22020)

### Description
Fixes scenario in which a bias input quantized to int32 has a scale that
is too small. A bias with a scale that is smaller than a certain
threshold will overflow the range of an `int32` when quantized, which
significantly decreases accuracy.

Credit to @yihonglyu for finding out about this issue and the fix.

### Motivation and Context
Consider the following Convolution with very small weights and a
constant bias input of `[5, -4.5]`.

![image](https://github.com/user-attachments/assets/4bde2bd9-892f-4ae9-887b-61a6668779a1)

The QDQ quantizer first computes the following quantization scale for
`input_0` and `weight`:
- `input_0`: scale=0.5
- `weight`: scale=7.843e-10 **[really small]**

The QDQ quantizer then computes the bias input's scale as follows:
```
bias_scale = input_0_scale * weight_0_scale = 0.5 * 7.843e-10 = 3.9215686274509805e-11
```

This `bias_scale` is too small. Before this PR, the QDQ quantizer would
quantize the f32 bias with this `bias_scale`:
```
bias_quant = round(bias_f32 / bias_scale) =  round([5.0/bias_scale, -4.5/bias_scale]) = [127500000000, -114750000000]
```
These quantized bias values exceed the range of int32, and so are
clipped to [int32.min(), int32.max()], which is very inaccurate.

#### New approach
This PR increases the `weight_0_scale` by the necessary amount to ensure
that `bias_scale` (which equals `weight_0_scale * input_0_scale`) is
appropriate for the int32 quantization type.

The smallest valid bias scale is given by the normal scale formula: 
`bias_smallest_valid_scale = (bias_f32_max - bias_f32_min) / (int32_max
- int32_min)`

Then, we compute the candidate bias scale:
`bias_scale_candidate = input_0_scale * weight_0_scale`

If the candidate scale is smaller than the smallest valid scale, we
increase the `weight_0_scale` by the necessary ratio:
```python
if bias_scale_candidate < bias_smallest_valid_scale:
    ratio = bias_smallest_valid_scale / bias_scale_candidate
    weight_0_scale = ratio * weight_0_scale
```

Then, we recompute the final bias scale:
```python
bias_scale = input_0_scale * weight_0_scale
```

#### Impact on accuracy
Here's the above model's quantized output compared to the f32
(ground-truth) output.
- Before PR: 
  - f32 model output[0]: **5.0f**
  - qdq model output[0]: **0.075**
  - SNR: 0.1369 (higher is better)
- After PR:
  - f32 model output[0]: **5.0f**
  - qdq model output[0]: **4.992**
  - SNR: 55.656 (higher is better)

onnxruntime/python/tools/quantization/base_quantizer.py

@@ -21,7 +21,6 @@
 from .quant_utils import (
     ONNX_TYPE_TO_NP_TYPE,
     TENSOR_NAME_QUANT_SUFFIX,
-    QuantType,
     find_by_name,
     model_has_infer_metadata,
     normalize_axis,
@@ -40,18 +39,26 @@ def __init__(self, **data: Dict[str, Any]):
         for k, v in data.items():
             if not isinstance(k, str):
                 raise TypeError(f"Keys must be strings not {type(k)} for k={k!r}.")
-            if not isinstance(v, (int, str, np.ndarray)):
+            if k != "axis" and not isinstance(v, (int, str, np.ndarray)):
                 raise TypeError(f"Values must be numpy arrays, int, float, str not {type(v)} for k={k!r}.")
+            if k == "axis" and not isinstance(v, int) and v is not None:
+                raise TypeError(f"Axis value must be an int or None, not {type(v)}.")
             if k == "scale" and v.dtype not in (np.float32, np.float16):
                 raise ValueError(f"scale must a float32 or float16 numpy element but is {v.dtype} for k={k!r}")
             self.data[k] = v
 
+    def get(self, key, default_value=None):
+        return self.data.get(key, default_value)
+
     def __iter__(self):
         yield from self.data
 
     def __getitem__(self, key):
         return self.data[key]
 
+    def __setitem__(self, key, value):
+        self.data[key] = value
+
     def __len__(self):
         return len(self.data)
 
@@ -88,9 +95,10 @@ def __init__(
         self.force_quantize_no_input_check = (
             "ForceQuantizeNoInputCheck" in self.extra_options and self.extra_options["ForceQuantizeNoInputCheck"]
         )
-        self.is_weight_symmetric = self.extra_options.get(
-            "WeightSymmetric", weight_qType in (QuantType.QInt8, QuantType.QInt16, QuantType.QFLOAT8E4M3FN)
-        )
+
+        # If user does not explicitly set "WeightSymmetric", then the weight's quantization type determines
+        # the symmetry (i.e., signed integer types will use symmetric quantization). See `def is_weight_symmetric()`
+        self._is_weight_symmetric: bool | None = self.extra_options.get("WeightSymmetric", None)
         self.is_activation_symmetric = self.extra_options.get("ActivationSymmetric", False)
         self.min_real_range = self.extra_options.get("MinimumRealRange")
 
@@ -131,6 +139,16 @@ def __init__(
 
         self.tensor_quant_override_qtypes = self.tensor_quant_overrides.get_quant_types()
 
+    def is_weight_symmetric(self, weight_quant_type: onnx.TensorProto.DataType) -> bool:
+        if self._is_weight_symmetric is not None:
+            return self._is_weight_symmetric  # Return value explicitly set by user.
+        return weight_quant_type in (
+            onnx.TensorProto.INT4,
+            onnx.TensorProto.INT8,
+            onnx.TensorProto.INT16,
+            onnx.TensorProto.FLOAT8E4M3FN,
+        )
+
     def quantize_model(self):
         raise NotImplementedError
 
@@ -230,9 +248,19 @@ def quantize_bias_static_impl(self, bias_name, input_scale, weight_scale, beta=1
             # TODO: This formula should be explained including why the scale is not estimated for the bias as well.
             bias_scale = input_scale * weight_scale * beta
 
-            quantized_data = (np.asarray(bias_data) / bias_scale).round()
-            quantized_data = np.clip(quantized_data, np.iinfo(np.int32).min, np.iinfo(np.int32).max)
-            quantized_data = quantized_data.astype(np.int32)
+            # Quantize by dividing by bias_scale
+            quantized_data = np.asarray(bias_data, dtype=np.float64) / np.asarray(bias_scale, dtype=np.float64)
+            quantized_data = quantized_data.round()
+
+            # Clip quantized data to the range of a int32
+            int32_min = np.float64(np.iinfo(np.int32).min)
+            int32_max = np.float64(np.iinfo(np.int32).max)
+            if np.any(quantized_data < int32_min) or np.any(quantized_data > int32_max):
+                logging.warning(
+                    f"Quantized bias `{bias_name}` exceeds the range of a int32. The bias scale is too small."
+                )
+
+            quantized_data = np.clip(quantized_data, int32_min, int32_max).astype(np.int32)
 
             # update bias initializer
             bias_np_data = np.asarray(quantized_data, dtype=np.int32).reshape(bias_initializer.dims)
@@ -282,6 +310,7 @@ def quantize_initializer_impl(self, weight, qType, reduce_range=False, keep_floa
                                   If keep_float_weight is False, quantize the weight, or don't quantize the weight.
         :return: quantized weight name, zero point name, scale name
         """
+        # TODO(adrianlizarraga): This function is now only used by onnx_quantizer.py, so move it there.
         q_weight_name = weight.name + TENSOR_NAME_QUANT_SUFFIX
         zp_name = weight.name + "_zero_point"
         scale_name = weight.name + "_scale"
@@ -303,10 +332,11 @@ def quantize_initializer_impl(self, weight, qType, reduce_range=False, keep_floa
             assert isinstance(scale, np.ndarray), f"Unexpected type {type(scale)}"
 
         else:
-            _, _, zero_point, scale, q_weight_data = quantize_data(
+            symmetric = self.is_weight_symmetric(qType) if qType == self.weight_qType else self.is_activation_symmetric
+            zero_point, scale, q_weight_data = quantize_data(
                 weight_data.flatten(),
                 qType,
-                quant_overrides.get("symmetric", self.is_weight_symmetric),
+                quant_overrides.get("symmetric", symmetric),
                 reduce_range=quant_overrides.get("reduce_range", self.reduce_range and reduce_range),
                 min_real_range=self.min_real_range,
                 rmin_override=quant_overrides.get("rmin"),
@@ -371,6 +401,7 @@ def quantize_weight_per_channel_impl(
         reduce_range=True,
         keep_float_weight=False,
     ):
+        # TODO(adrianlizarraga): This function is now only used by onnx_quantizer.py, so move it there.
         initializer = find_by_name(weight_name, self.model.initializer())
         if initializer is None:
             raise ValueError("{} is not an initializer", weight_name)
@@ -409,13 +440,7 @@ def quantize_weight_per_channel_impl(
         if "quant_type" in quant_overrides_for_channels[0]:
             weight_qType = quant_overrides_for_channels[0]["quant_type"].tensor_type  # noqa: N806
 
-        symmetric = quant_overrides_for_channels[0].get(
-            "symmetric",
-            (
-                self.is_weight_symmetric
-                or weight_qType in (onnx.TensorProto.INT8, onnx.TensorProto.FLOAT8E4M3FN, onnx.TensorProto.INT4)
-            ),
-        )
+        symmetric = quant_overrides_for_channels[0].get("symmetric", self.is_weight_symmetric(weight_qType))
         reduce_range = quant_overrides_for_channels[0].get("reduce_range", self.reduce_range and reduce_range)
         zero_point_list = []
         scale_list = []
@@ -444,7 +469,7 @@ def quantize_weight_per_channel_impl(
                 ), f"Unexpected type {type(quantized_per_channel_data)}"
 
             else:
-                _, _, zero_point, scale, quantized_per_channel_data = quantize_data(
+                zero_point, scale, quantized_per_channel_data = quantize_data(
                     per_channel_data.flatten(),
                     weight_qType,
                     symmetric,

onnxruntime/python/tools/quantization/onnx_model.py

@@ -296,6 +296,26 @@ def get_largest_node_name_suffix(self, node_name_prefix):
 
         return suffix
 
+    def get_largest_initializer_name_suffix(self, initializer_name_prefix):
+        """
+        Gets the largest initializer name integer suffix for all initializer names that begin
+        with `initializer_name_prefix`. This can be used to create unique initializer names.
+
+        Example: for initializer names 'my_weight_0' and 'my_weight_3', this method returns 3 if
+                 `initializer_name_prefix` is 'my_weight_'.
+        """
+        suffix = -1
+
+        for initializer in self.model.graph.initializer:
+            if initializer.name.startswith(initializer_name_prefix):
+                try:
+                    index = int(initializer.name[len(initializer_name_prefix) :])
+                    suffix = max(index, suffix)
+                except ValueError:
+                    continue
+
+        return suffix
+
     def find_nodes_by_initializer(self, graph, initializer):
         """
         Find all nodes with given initializer as an input.