Hi,
I’m working on deploying LLM inference on the NPU of the Genio 520-EVK (MT8371) and would like to confirm a few things about the workflow.
1. Pre-built models (public):
My understanding is that the pre-built LLM packages listed in the IoT AI Hub Model Zoo (e.g. llama3.2-1b, containing the compiled .dla files, tokenizer, and config) are publicly downloadable and can be run directly on the Genio 520 via the built-in llm_cmdline_tool in the Yocto image, with no special SDK access required. Is that correct?
2. Converting our own model:
For deploying our own LLM with NPU acceleration on the Genio 520, I’d like to understand the conversion path to produce the .dla files:
- Do I need NDA-gated access to the GAI Toolkit and/or the NeuroPilot SDK (Premium) to perform this conversion? If so, could you point me to how to request that access for our account?
- Which specific NeuroPilot version is required/matched for the Genio 520 (MT8371) LLM conversion?
- Is there any publicly available converter that can convert a Hugging Face / ONNX / TFLite LLM into the
.dlaformat for Genio 520 NPU inference, or is the GAI Toolkit the only supported path?
Thanks in advance for the clarification.
Best regards,