[P0] Off-chain / devshard implementation track #1220
Open for community contributors. Multiple parallel efforts in this direction are welcome to explore different approaches and accelerate progress.
💬 Comments (2)
hi @orvionx, thanks for you interest!
I'd say that these milestones are very broad to assess at this stage. there are some important differences between inference off-chain logic (devshards) and requirements for training.
specifically, because inference is a very well-defined flow and all devshards are doing exactly the same work all the time. decentralized training, on the other hand, is a research direction with a lot of open questions regarding particular low-communication training methods, validation mechanisms etc. and it should allow researchers to run experiments, test these approaches and iterate quickly.
the first iteration of how communication with main net will work is proposed here
I'd suggest to start with reviewing that issue and the corresponding PR with trainshard v0 plan and join the discussion from there.
🔄 Auto-synced from Issue #1220 every hour.
Hi, @tcharchian I’d like to work on this issue.
I reviewed the current
devshardstructure and would like to start with a scoped implementation pass rather than trying to cover the whole off-chain/devshard track at once.My initial plan is:
devshardpackages and current off-chain flowBefore I start the implementation, could you confirm whether there is a preferred first milestone or acceptance criteria for this track? If there is no strict preference, I can start by proposing a small PR around the current devshard flow and iterate from maintainer feedback.