Update Quark to use latest Onnx version.

[Zacchae]

Following istallation instructions at: https://quark.docs.amd.com/latest/install.html I ensured I had build-essentials installed

apt list --installed | grep build-essentials
# prints "build-essential/jammy,now 12.9ubuntu3 amd64 [installed,automatic]"

and ran (I'm on Ubuntu 22.04)

python -m venv codes/quark
codes/quark/bin/pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
codes/quark/bin/pip install "cmake<4.0" amd-quark

Then, following instructions at https://quark.docs.amd.com/latest/supported_accelerators/ryzenai/tutorial_convert_fp32_or_fp16_to_bf16.html I ran:

codes/quark/bin/python -m quark.onnx.tools.convert_fp32_to_bf16 --input $FLOAT32_ONNX_MODEL_PATH --output $BFLOAT16_ONNX_MODEL_PATH --format with_cast

Which fails. Actually, it fails on import, so even the following fails:

codes/quark/bin/python -m quark.onnx.tools.convert_fp32_to_bf16
codes/quark/bin/python -c "import quark.onnx.tools.convert_fp32_to_bf16"

The traceback is:

Traceback (most recent call last):
  File "/usr/lib/python3.10/runpy.py", line 187, in _run_module_as_main
    mod_name, mod_spec, code = _get_module_details(mod_name, _Error)
  File "/usr/lib/python3.10/runpy.py", line 110, in _get_module_details
    __import__(pkg_name)
  File "/mnt/bpd_architecture/zacchae/codes/quark/lib/python3.10/site-packages/quark/onnx/__init__.py", line 8, in <module>
    from .calibrate import (create_calibrator_power_of_two, create_calibrator_float_scale)
  File "/mnt/bpd_architecture/zacchae/codes/quark/lib/python3.10/site-packages/quark/onnx/calibrate.py", line 27, in <module>
    from .quant_utils import (PowerOfTwoMethod, get_tensor_type_from_qType, quantize_data_pof2s, ExtendedQuantType)
  File "/mnt/bpd_architecture/zacchae/codes/quark/lib/python3.10/site-packages/quark/onnx/quant_utils.py", line 321, in <module>
    ONNX_INT_TYPE_RANGE = create_range_dict(dtype_ranges)
  File "/mnt/bpd_architecture/zacchae/codes/quark/lib/python3.10/site-packages/quark/onnx/quant_utils.py", line 286, in create_range_dict
    np_dtype = getattr(custom_element_types, dtype.lower())
NameError: name 'custom_element_types' is not defined

I have verified that I am on the latest quark version, i.e. the version that the linked docs (currently) point to

python -c "import quark; print(quark.__version__)"
# prints "0.9"

How can we proceed converting to bfloat16?

[Zacchae]

Temporary work-around: downgrade onnx to 1.18.0. Good luck fixing it on your end!