NExT-GPT Chapter 6 inference resources
FP16 inference resources used by the course demo, matching the official NExT-GPT 7b_tiva_v0 architecture and adaptation checkpoint.
The bundle contains restored Vicuna-7B-v0, ImageBind huge, the original NExT-GPT adaptation checkpoint, Stable Diffusion v1.5, AudioLDM l-full, and Zeroscope v2 576w. model_manifest.json records their fixed upstream sources and resource paths. Each component remains subject to its original model license and usage conditions; the code license of the demo does not replace those terms.
Use the accompanying NExT-GPT demo code to load this multi-component bundle. It is not a single Transformers AutoModel checkpoint.
The resources retain the FP16 inference values of the reproduced model. Re-loading them produced identical token IDs and PNG/WAV/MP4 bytes for the three stored roundtrip cases. This checks resource conversion, not the semantic quality of the outputs. The model can omit requested modalities or hallucinate visual details.
- Downloads last month
- -