v1.2.67 - Train portable Minimax H3 LoRAs locally, smaller downloads & bugfixes

[imprsnst]

What's Changed

MiniMax H3 LoRAs now work in other apps e.g. ComfyUI

A LoRA trained here did nothing anywhere else. Several people reported the same thing: train an H3 LoRA, load it in another app, and it has no effect even at maximum strength. It was not a weak LoRA, it was a naming problem. H3 stores attention as one combined tensor, our loader splits it into three while reading the file, and the adapter was saved against those split names. Nothing else recognised them, so every layer was skipped in silence.

LoRAs are now saved using the names in the published checkpoint, so one file works both here and elsewhere. H3 LoRAs trained in other apps load here too.

You will need to retrain. LoRAs made before this release are in the old format. They still work inside Inline Studio, but not outside it.

The new file is about 40 percent larger. That is unavoidable: the three attention parts each carry their own data, and the combined format has to keep all of it.

Smaller MiniMax H3 downloads

66.3GB was the only option. The community pruned and pruned_fp8_scaled builds now load, and the fp8 one is 21.0GB for the same model. It appears in the model popup as an optional download.

Two things to know. It saves download and disk, not memory: every build takes the same room once loaded. And training still needs the bf16 file, because the smaller builds leave out a piece the trainer needs. The trainer says so clearly instead of failing part way in.

The int8_convrot files still do not load. Their weights are stored rotated, and that cannot be undone outside the app that did it.

Fixes

MiniMax H3 showed "loading model" for the whole render. It only reported at the start and the end, so a render that takes minutes looked stuck on the loading step. It now shows the step count as it goes, and stopping mid render works properly.

The queue vanished after a page refresh. Generation kept running in the background but the app forgot about it, so you saw an empty queue against a busy GPU. It now asks what is still running when the page loads.

H3 rendered slower than it needed to. Part of the model was kept at higher precision than the rest, which quietly pushed everything around it to higher precision too and lost a faster code path.

LoRA strength was applied wrongly if you changed it. The saved file left out the alpha value, so any setting other than the default was ignored when the LoRA was applied. This affected all models.

Upgrading

If you installed with git:

cd Inline-Studio && git pull
cd core && ./webui.sh --install --extra all
uv pip install --python .venv/bin/python -U inline-studio-frontend

That last line matters this time: the web UI ships as a separate package and the installer
does not force it to update, so without it you get the new engine and the old interface.

If you installed from PyPI:

pip install -U -r requirements.txt

Full Changelog: v1.2.66...v1.2.67

---

This discussion was created from the release v1.2.67 - Train portable Minimax H3 LoRAs locally, smaller downloads & bugfixes.