virtaava's picture
Upload README.md with huggingface_hub
6c80790 verified
# Qualcomm AI Runtime — V75 Android arm64 (v2.46.0.260424)
Mirror of the Qualcomm AI Runtime SDK Community Edition runtime libraries needed
to execute precompiled-QNN-ONNX bundles on **Snapdragon 8 Gen 3 (Hexagon V75)**
Android devices.
This mirror exists so the Sona Forge app has a single, stable download endpoint
under `huggingface.co/sona-forge/` for every binary it needs at runtime. Sona
Forge is a local-first Android image-transformation app and never phones home
to any provider.
## Files
| File | Side | Bytes | Purpose |
|---|---|---:|---|
| `libQnnHtp.so` | App / arm64-v8a | 3,495,096 | Generic HTP backend dispatcher |
| `libQnnSystem.so` | App / arm64-v8a | 3,864,784 | QNN system support |
| `libQnnHtpV75Stub.so` | App / arm64-v8a | 736,288 | V75 dispatcher stub (CPU side) |
| `libQnnHtpPrepare.so` | App / arm64-v8a | 90,431,072 | Graph compiler / preparation pass |
| `libQnnHtpV75.so` | DSP / hexagon-v75 | 11,388,152 | V75 backend (loaded onto Hexagon) |
| `libQnnHtpV75Skel.so` | DSP / hexagon-v75 | 11,085,052 | V75 FastRPC skel (loaded onto Hexagon) |
The four `App` libs are linked against Android **bionic** and load on the
application processor via `dlopen` / `System.load`. The two `DSP` libs are
Hexagon DSP6 binaries loaded onto the Hexagon NPU itself by FastRPC.
## Source
Extracted from the official Qualcomm AI Runtime SDK Community Edition v2.46.0,
specifically the artefacts under:
- `qairt/2.46.0.260424/lib/aarch64-android/`
- `qairt/2.46.0.260424/lib/hexagon-v75/unsigned/`
## Licence
Distributed under the **Qualcomm AI Engine Direct SDK License Agreement** that
covers QAIRT Community Edition. See [LICENSE](LICENSE) for the redistribution
terms. Sona Forge end-users acquire the runtime on their device when the app
fetches this pack; the download itself is the redistribution event.
## Compatibility
QAIRT runtime is back-compat across the 2.x minor series. v2.46 runtime loads
context binaries compiled against v2.42 (and later) Workbench releases.
## See also
- [`sona-forge/sd15-controlnet-canny-qnn-8gen3`](https://huggingface.co/sona-forge/sd15-controlnet-canny-qnn-8gen3) — the SD 1.5 + ControlNet-Canny precompiled-QNN-ONNX bundle this runtime accelerates.