Rather than re-convert the model, we're writing a custom Rust implementation for Hydra, as ONNX performed poorly for JTP-3 and GPU acceleration was inaccessible.
This repo will be kept up for posterity!
An ONNX-converted version of JTP-3, for use with Akasha, a desktop-native image gallery for data hoarders.
Support is currently experimental.
Kimi K2.7 converted the weights using export_onnx_v2.py. All functions exposed in RedRocket's inference code, to my knowledge, is ported to Akasha.
The only thing we don't currently have is extensions; this is pending a potential rewrite of ONNX handling in general (considering a move to Burn) as it's the most involved part by far.
To use, download jtp-3-hydra-fpXX.onnx and its matching .onnx.data, as well as tags.csv, tags.txt and pos_embed.safetensors, and aim an Akasha [[models]] table at it.
Currently JTP support is only available on Akasha's dev branch, and isn't currently compiled into nightly binaries.
