Documentation

Hypertensor Documentation

This documentation is for Testnet v1.0 and this product is still undergoing development. Expect bugs.

Be aware that the blockchain will shut down and reset as we find and fix bugs and update the code.

The subnet repository is still undergoing development and is expected to be constantly undergoing updates that may require pulling new versions for anyone running an AI node.

Introduction

This documentation involves the application layer of the blockchain, decentralized AI model hosting (subnets), and how to interact with the protocol.

This documentation is an ongoing effort. If there are any inaccuracies, contact us through Discord.


Socials

Official socials

Any socials not listed here or on the website are not managed or endorsed by the Hypertensor team.


Whitepaper


Run Blockchain and AI Nodes

Note

During testnet, it's more important to run a model validator node hosting the machine-learning models than it is to run a Hypertensor blockchain node.


Outline


Intro

Hypertensor comprises two components, the blockchain, and the decentralized subnets, specifically the AI model hosting network.


Blockchain

The blockchain is responsible for forming consensus amongst peers hosting models, incentivizing those contributing to consensus and validating decentralized machine learning models, rewarding both blockchain validators and model validators, democratizing artificial intelligence, and being a payments & transactions infrastructure.


Decentralized Artificial Intelligence (Subnet)

The decentralized peer-to-peer machine learning model hosting network is responsible for hosting machine learning models, such as LLMs, image diffusers, or other categories of AI models. Models are hosted by peers with each peer hosting a subset of a model alongside other peers hosting other subsets of the model, together hosting an entire model.

Notes

  • To begin with, Hypertensor will be focused on decentralized LLM's. In the future Hypertensor and the community of developers will implement other AI categories, such as image diffusers and more.

  • Hypertensor can integrate and interface any decentralized AI hosting technology.


Peer Consensus Mechanism

The Peer Consensus Mechanism is the core of the blockchain. Each AI model-hosting peer interfaced with the Hypertensor Blockchain must submit consensus data on each peer hosting within the same model on each epoch. In between each epoch, validators form consensus and generate emissions.


Incentives

Users can generate rewards by validating the blockchain or hosting machine learning models as a model validator.

Blockchain validators generate rewards on each block. There is a split of the rewards on each block with a portion of rewards going to blockchain validators and the remaining going to model validators on each epoch.

Testnet v1.0 doesn't implement inflation but plans to transition into a NPoS mechanism for mainnet.


Peer Scores

Read more about consensus

Each peer is scored based on their contribution towards computing. In Petals Tensor, each peer is scored based on the following equation:

score=maxShareRatioxx+xscore = maxShareRatio * x * x + x

This equation incentivizes the following:

  • Run higher-performing compute with fewer peers than lower-performing compute with more peers.

    • e.g. One server with equal computing to 10 servers will garner higher rewards.

  • Run servers with similar or higher computing to others, increasing the total measurable computation over time to compete for rewards.

Theoretically, this can result in a left-skewed low distribution or negatively-skewed low distribution bell curve that is ever increasingly skewing left as technology for higher-compute improves.

Note

  • The scoring mechanism is expected to undergo multiple changes going into the future. The main priority is to incentivize users to host nodes with higher-performing GPUs to result in a low distribution of compute without undermining nodes with lower-performing GPUs.

  • It is up to each subnet to create a scoring mechanism.

Last updated