The problem GPUFlex solves

The Problem It Solves

Training AI models is expensive and resource-intensive, while millions of GPUs sit idle around the world. GPUFlex unlocks this untapped power, making model training cheaper, faster, and decentralized.

It removes the complexity of distributed training by handling model sharding, GPU tracking, and result verification — all while securely rewarding contributors through a staking system.

What It’s Good For
• Lowering the cost of training AI models
• Monetizing idle GPU time
• Scaling ML experiments across a global network
• Ensuring fair and verified contributions in a trustless system

Challenges we ran into

The current Yellow protocol lacked a data field in the AppSession, which several teams needed to support use cases involving data messaging. The proposed solution was to add a SessionData field to the protocol.

During the hackathon, we explored the database schema and identified two implementation options: json or VARCHAR. Relevant migrations were created. We also modified the repository code, adding SessionData to app_session.go and noting the need for inclusion in handlers.go. Test cases were written, though some failed due to incomplete setup.

Adding this field opens the door to more complex applications like RWA, IoT, and AI — all of which typically require structured session data. However, the current project structure lacks clarity, and a proper test framework is still needed. Overall, the changes mark a notable improvement over the 2024 version, and we’re grateful to the Yellow team for their support.

Two possible realisations - json or VARCHAR
https://github.com/Kyrylog/nitrolite/blob/main/clearnode/config/migrations/postgres/20250614152411_add_initial_schema.sql

Added SessionData to app_session.go https://github.com/erc7824/nitrolite/commit/4ca37fdf203c43aa40c7bb5879e9f8d027430d4d

Need to add SessionData to handlers.go
https://github.com/erc7824/nitrolite/commit/3741d0e5189f6db6a6d8166e5e47ff0df7b9f91b