RadixArk open-sources weight teleportation, extending Miles delta weight sync
Overview
RadixArk says it has open-sourced weight teleportation, a method for faster weight updates from trainers to samplers in large-scale reinforcement learning.
The company says the method builds on delta weight sync in Miles, takes it further, and contributes the changes back to Miles.
RadixArk reports a sampler pause time under 0.23 seconds and about 3 seconds of non-blocking compression and preparation. These figures come from the company's own post on X.
Written by AI from the articles below · updated Nov 2, 6:55 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
RadixArk@radixarkOfficialRadixArk open-sources weight teleportation for faster RL weight syncAIRadixArk says it has open-sourced weight teleportation, a method for faster weight updates from trainers to samplers in large-scale reinforcement learning. The company reports a sampler pause time under 0.23 seconds and about 3 seconds of non-blocking compression and preparation, and says it was built on delta weight sync in Miles and contributed back to Miles.
Heat Trend
Not enough continuous observations to show a trend yet.