[Gemma3-1b-it] Make calibration dataset failed

NeuroPilot NeuroPilot - GenAI
,
1

An error occurs when creating the calibration dataset using the latest version of the MTK LLM SDK V3.4.2. According to the MTK LLM SDK documentation, this version already supports Gemma 3. Will an example for Gemma 3 be released in the GAI Toolkit, or how can this error be resolved?

Traceback (most recent call last):
File “/home/viauser/miniconda3/envs/np8_py38/lib/python3.8/site-packages/mtk_llm_sdk/tokenizers/tokenization_utils_base.py”, line 2197, in _from_pretrained
tokenizer = cls(*init_inputs, **init_kwargs)
File “/home/viauser/miniconda3/envs/np8_py38/lib/python3.8/site-packages/mtk_llm_sdk/tokenizers/tokenization_gemma_fast.py”, line 100, in init
super().init(
File “/home/viauser/miniconda3/envs/np8_py38/lib/python3.8/site-packages/mtk_llm_sdk/tokenizers/tokenization_utils_fast.py”, line 107, in init
fast_tokenizer = TokenizerFast.from_file(fast_tokenizer_file)
Exception: data did not match any variant of untagged enum ModelWrapper at line 2379610 column 3

2

Hi Lori,

Thanks for the question.

For now, Gemma 3 is yet to plan to be supported with NP8 platform(Genio-520, Genio-720).

It’s to besupported with MTK LLM SDK V3.4.2 + NP9 platform, and there will have a GAI tutorial for Gemma3 soon in the begining of 2026.

Important Notice

  • Library upgrades on NP8 may enable limited experimentation but are not supported for production.
  • Do not upgrade Neuron libraries on production devices to force compatibility. Neuron libraries bind tightly to hardware (MDLA versions) and mixed configurations may cause instability or accuracy issues.
  • For stable deployment and predictable behavior, use the documented platform and SDK combinations.