README.md

metadata

title: MicroMixer-2
emoji: 🫪
colorFrom: purple
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
license: apache-2.0
short_description: Play with the four attention-free MicroMixer-2 checkpoints.

MicroMixer-2 Discord-dialogue Playground

Interactive Gradio demo for the four attention-free, MLP-only language models in llaa33219/micromixer-2.

Variant	Max context	Params
100K	64	~125K
300K	128	~431K
500K	128	~779K
1M	4096	~1.02M

All checkpoints are byte-level (vocabulary = 256) and were trained on mookiezi/Discord-Dialogues, so prompts in User: / Assistant: format tend to work best.

How it works

• The first time you pick a model, app.py downloads its model.pt from the Hub via huggingface_hub.hf_hub_download and caches it.
• The architecture (src/model.py) and the byte-level tokenizer (src/tokenizer.py) are vendored from llaa33219/MicroMixer-2.
• MicroMixerConfig is rebuilt from the exact hyperparameters listed on each model card, so load_state_dict matches the published checkpoints.

Why Docker SDK

The HF Spaces "gradio" base image pins gradio==4.44.0, which has two Python-3.13-specific problems out of the box:

• it imports audioop transitively (via pydub), but the module was removed from the stdlib in 3.13;
• its oauth.py still does from huggingface_hub import HfFolder, which was removed in huggingface_hub 1.0.

Pinning old versions to dodge those would mean a fragile requirements.txt. Instead, this Space uses the docker SDK with a minimal Dockerfile that installs the latest torch / gradio / huggingface_hub / audioop-lts from scratch. No version pins anywhere - whatever the ecosystem ships today is what runs.

Files

.
├── app.py              # Gradio entry point
├── Dockerfile          # python:3.13-slim + deps + `python app.py`
├── requirements.txt    # no version pins, latest of everything
├── src/
│   ├── __init__.py
│   ├── model.py        # MicroMixer-2 V4 architecture
│   └── tokenizer.py    # ByteTokenizer (vocab=256)
└── README.md

License

Apache 2.0, in line with the upstream project.