Opt-AI/llama_v3_2_1b_instruct-genie-w4a16-qualcomm_snapdragon_8_elite-optai

Llama 3.2 1B Instruct Genie W4A16 for Snapdragon 8 Elite

This repository contains Opt-AI packaged runtime artifacts for meta-llama/Llama-3.2-1B-Instruct, targeting Qualcomm Snapdragon 8 Elite.

The files in this repository were imported from Qualcomm AI Hub artifacts and uploaded by Opt-AI for Hugging Face distribution.

What is included

• Quantized Genie binary shards (*.bin)
• genie_config.json
• htp_backend_ext_config.json
• config.json
• tokenizer.json
• tokenizer_config.json
• metadata.yaml
• sample_prompt.txt

Runtime and packaging

• Runtime: Qualcomm Genie
• Backend: QNN HTP (QnnHtp)
• Target device family: Qualcomm Snapdragon 8 Elite
• Target code: qualcomm_snapdragon_8_elite
• Context length: 4096
• Vocabulary size: 128256
• Packaging: W4A16 Genie bundle
• QAIRT version: 2.42.0.251225135753_193295

Base model

• Base model: meta-llama/Llama-3.2-1B-Instruct
• Architecture: LlamaForCausalLM
• Hidden size: 2048
• Number of layers: 16
• Attention heads: 32
• KV heads: 8
• Torch dtype in config: bfloat16

Usage notes

These files are packaged for the Qualcomm Genie / QNN runtime. They are not standard Hugging Face Transformers weight files for direct from_pretrained() loading.

Keep the .bin files and JSON config files together in the same directory layout when deploying.

Example prompt

What do llamas eat? Keep the answer under ten words.

Provenance and licensing

• Imported from Qualcomm AI Hub artifacts
• Derived from meta-llama/Llama-3.2-1B-Instruct
• Review the upstream Meta Llama license terms and any Qualcomm AI Hub redistribution requirements before production use