LoRA node has new "quant" substring in onnx graph after quantization

[miaoqiz]

It seems that with quantization in the "auto-opt" command, it adds "quant" in LoRA related nodes in the graph

The command line is as follows:

---

olive auto-opt --model_name_or_path microsoft/Phi-3.5-mini-instruct --trust_remote_code --adapter_path path_adapter --output_path path_output --device cpu --provider CPUExecutionProvider --precision int4 --use_ort_genai --log_level 1"

---

Is there a way to reverse it to the original LoRA node name with no "quant" string?

Thanks very much!

[jambayk]

The .quant part was added since each adapter weight has 2/3 parameters (quantized weight, scale, and zero point) and I wanted to distinguish them from non quantized weights.

Olive/olive/passes/onnx/extract_adapters.py

Line 106 in d98186d

quantized_suffices = [".quant.weight", ".quant.scale", ".quant.zero_point"]

Do you have a use case where it cares about what the adapter weight names are?