We're thrilled to announce the release of Targon v2, a significant rewrite of our core infrastructure that sets the stage for exciting developments on our roadmap. This update brings substantial improvements in efficiency and flexibility for Subnet 4 and the entire Bittensor ecosystem.
We've dramatically reduced our codebase from over 3,500 lines of code in v1 to just a little over 1,800 lines of code in v2. This leaner architecture positions us to more nimbly implement future features and improvements.
We've addressed the issue of miners buffering tokens on SN4, which was artificially inflating Tokens Per Second (TPS) metrics. In v1, miners were taking in a validator request, and then pausing responses for multiple seconds before sending a large response all at once.
Targon v2 employs Jaro-Winkler Similarity 1 checks to ensure response similarity to ground truth and grades miners on absolute time, moving us closer to a mechanism that rewards only miners who are doing honest and valuable work.
Miners now have the flexibility to use the OpenAI Query Schema, opening up alternatives to TGI such as VLLM or other preferred systems. This change promotes greater diversity and innovation in our mining ecosystem. As a miner, you can now choose any inference engine that is compatible with the OpenAI Query Schema. This gives miners freedom to experiment with different infrastructures until they land on one that produces the most accurate results and the quickest speeds.
As we build on the foundation of Targon v2, here's a glimpse of what's coming:
We're excited about the possibilities Targon v2 unlocks and look forward to building these new features alongside the Bittensor community.
Jaro-Winkler Similarity is a metric used to measure the similarity between two strings. Introduced in 1990 by William E. Winkler, Jaro-Winkler builds on the original Jaro Similarity by adding a prefix scale that gives more favorable ratings to strings that match from the beginning. ↩