NanoKV: High-Ratio KV Cache Compression for Autoregressive Video Generation
NanoKV codec and NanoKV-Fast attention readers for HY-WorldPlay, LingBot-World 2.0, and Matrix-Game 3.5.
Project page · Code · Inference guide
Checkpoints
| Backbone | Directory | Codec | Fast readers | Total |
|---|---|---|---|---|
| HY-WorldPlay | hy-worldplay/ |
1 | 41 | 281 MiB |
| LingBot-World 2.0 | lingbot-world-2.0/ |
1 | 40 | 1,260 MiB |
| Matrix-Game 3.5 | matrix-game-3.5/ |
1 | 30 | 452 MiB |
Each directory contains codec.pt, readers/, inference profiles, a portable Fast manifest template, checksums, and a model-specific guide. NanoKV and Fast share the same codec; Fast also uses the readers.
Download
pip install -U huggingface_hub
hf download jiacheng1234/NanoKV --local-dir ./checkpoints
To download one backbone:
hf download jiacheng1234/NanoKV --include "hy-worldplay/*" --local-dir ./checkpoints
Prepare and run
Install the base model following the setup guide. From the NanoKV code checkout, set ARTIFACTS to the chosen backbone directory:
export ARTIFACTS="$PWD/checkpoints/hy-worldplay"
(cd "$ARTIFACTS" && sha256sum -c SHA256SUMS)
python -m nanokv.prepare_artifacts --model hy \
--artifacts "$ARTIFACTS" --output "$ARTIFACTS/manifest.json"
| Backbone | --model |
ARTIFACTS subdirectory |
|---|---|---|
| HY-WorldPlay | hy |
hy-worldplay |
| LingBot-World 2.0 | lingbot |
lingbot-world-2.0 |
| Matrix-Game 3.5 | mg |
matrix-game-3.5 |
The local manifest binds the Fast readers to your download location. Use the matching nanokv.json or fast.json profile with the inference commands.
Licenses
Each backbone's files retain their model-specific license. See LICENSES.md and the license files inside each directory.