Giving 👀 to Bittensor with Decentralized subnet inference at scale.
Subnet 19 is a Bittensor subnetwork, focused around inference at scale, with an initial focus on image generation & manipulation models.
- A ton of models supported - with arbitrary models to come 🙏🏻
- DPO technique integrated into scoring - miners compete head to head! 💀
- Custom trained XGBoost reward model to ensure miners can't game the system whilst allowing a variety of GPU types 📈
- Validators can sell their organic traffic in a completely decentralized way 💰
We have SDXL Turbo, Kandinsky, Inpainting, Scribble, Controlnets, Upscaling, Image-To-Image, Uncrop support, the list goes on!
GPU: 1x80 GB or 3x24gb or 2x48gb vram
You might be able to use 2x24gb vram or 1x48gb vram, but as models change this is subject to change
Storage: 200Gb
RAM: 32Gb
GPU: 1x80 GB or 1x24GB or 1x48GB
Storage: 200Gb
Ram: 64Gb (Recommended)
Setups tested for mining so far:
Name | CUDA Version | Ubuntu Version | Python Version | Works |
---|---|---|---|---|
H1OO | 12.0 | 22.04 | 3.10.12 | ✅ |
A100 | 11.8 | 22.04 | 3.10.12 | ✅ / ❌ (hit and miss) |
RTX 4090 | 11.8 | 22.04 | 3.10.12 | ✅ |
A6000* | 12.0 | 22.04 | 3.10.12 | ✅ |
A40 | 12.0 | 22.04 | 3.10.12 | ✅ |
L40 | 12.0 | 22.04 | 3.10.12 | ❌ |
A100 SXM | 11.8 | 22.04 | 3.10.12 | ❌ |
Note: That's not to say you can't use other GPU's!
See here for our response to key metrics used for assigning validator emissions