Hi team,
I am currently trying to perform Post-Training Quantization (PTQ) on the Qwen3-1.7B model for the 520 platform using GAI Deployment Toolkit v3.4.2 (with mtk_llm_sdk_v3.4.2).
When executing the quantization script ./qwen3-1.7B_02_ptq.sh, I hit a version check error indicating that the current system’s Converter version is 8.16, whereas the SDK explicitly requires >= 9.2.0.
1. Command & Environment Info
-
Toolkit Version: GAI Deployment Toolkit v3.4.2
-
SDK Version:
mtk_llm_sdk_v3.4.2(venv_3.4.2) -
Target Model: Qwen3-1.7B
-
Target SoC/Platform: 520
-
Execution Command:
./qwen3-1.7B_02_ptq.sh
2. Traceback Log
Plaintext
2026-10-12 16:33:03,289 [mtk_llm_sdk:INFO] precision_config.py:415: Precision Config Mode: manual
2026-10-12 16:33:03,289 [mtk_llm_sdk:INFO] precision_config.py:424: [PTQPrecisionConfig] max_llm_weight_bit=4
2026-10-12 16:33:03,289 [mtk_llm_sdk:ERROR] ptq.py:564: Found converter version 8.16 but >= 9.2.0 is required
Traceback (most recent call last):
File "/home/00/MJ/mt8189/mtk_llm_sdk_v3.4.2/venv_3.4.2/bin/mtk_ptq_llm", line 8, in <module>
sys.exit(main())
File "/home/00/MJ/mt8189/mtk_llm_sdk_v3.4.2/venv_3.4.2/lib/python3.11/site-packages/mtk_llm_sdk/utils/memory_profiler.py", line 402, in wrapper
retval = func(*args, **kwargs)
File "/home/00/MJ/mt8189/mtk_llm_sdk_v3.4.2/venv_3.4.2/lib/python3.11/site-packages/mtk_llm_sdk/ptq.py", line 1678, in main
args_sanity_checks(args)
File "/home/00/MJ/mt8189/mtk_llm_sdk_v3.4.2/venv_3.4.2/lib/python3.11/site-packages/mtk_llm_sdk/ptq.py", line 652, in args_sanity_checks
_converter_checks(args, config)
File "/home/00/MJ/mt8189/mtk_llm_sdk_v3.4.2/venv_3.4.2/lib/python3.11/site-packages/mtk_llm_sdk/ptq.py", line 564, in _converter_checks
logger.error(f'Found converter version {cvt_ver[0]}.{cvt_ver[1]} but >= 9.2.0 is required')
File "/home/00/MJ/mt8189/mtk_llm_sdk_v3.4.2/venv_3.4.2/lib/python3.11/site-packages/mtk_llm_sdk/utils/logger.py", line 234, in error
raise err
RuntimeError
3. Questions / Requests for Support
-
Does GAI Deployment Toolkit v3.4.2 / SDK v3.4.2 officially support the Qwen3 model series?
-
Regarding the hard constraint of
Converter >= 9.2.0, where can I download the corresponding NeuroPilot Converter v9.2.0 update package or release? -
If I cannot immediately update the Converter to v9.2.0 in my current environment, is there a recommended temporary workaround (e.g., lowering the version check threshold)? Or is version 8.16 fundamentally incompatible with Qwen3’s underlying architecture?
Thanks for your support!