Builder output with quantization enabled yields incorrect negative scales.

[aendk]

#Describe the bug
Quantization scales are defined to always be positive in the onnx documentation.
Creating a qdq-enabled onnx-representation using the builder.py lead to plenty of negative scales, contrary to what is stated in the documentation.
It also causes exceptions in various execution providers.

Steps to reproduce the behavior:

1. Download a model from hugginface, in our case Phi-3.5-mini-instruct.
2. Create a quantized onnx-represenation from it, with a call similar to this:

python builder.py -m /path/to/downloaded/model -e cuda -p int4 -o /desired/path/to/result --extra_options use_qdq=1

3. Open this new onnx-representation using netron
4. Click on a random DequantizeLinear node.
5. In the opened sidebar, find Inputs, and there x_scale. Click on the associated +-sign on the right.
6. See that there are many negative values.

Expected behavior
In Step 6, only 0 or positive values should be visible.

Desktop (please complete the following information):
ubuntu

Screenshots

[kunal-vaishnavi]

When quantizing to INT4 precision, the model builder assumes that the scales are symmetric. When is_symmetric = True, the scales can be positive or negative. If is_symmetric = False, the scales will be non-negative.

onnxruntime-genai/src/python/py/models/builder.py

Lines 463 to 474 in 2f2686f

	def to_int4(self, model):
	quant = MatMul4BitsQuantizer(
	model=model,
	block_size=self.quant_attrs["int4"]["block_size"],
	is_symmetric=True,
	accuracy_level=self.quant_attrs["int4"]["accuracy_level"],
	nodes_to_exclude=[],
	quant_format=QuantFormat.QDQ if self.quant_attrs["use_qdq"] else QuantFormat.QOperator,
	op_types_to_quantize=self.quant_attrs["int4"]["op_types_to_quantize"],
	)
	quant.process()
	return quant.model.model

This can be fixed by adding the ability to set is_symmetric using the --extra_options. Here's the PR to enable this.