RLHF Infrastructure: Miles Server Arguments Documentation
Source: https://x.com (tweet) Author: unknown (Prime Intellect ecosystem) Date: 2025-2026
Summary
Full documentation dropped for Miles server arguments covering Ray resource scheduling and R3 (Rollout Routing Replay) for MoE models. Includes tips for Blackwell hardware.
Key Claims
- RLHF infra is described as “a beast”
- R3 = Rollout Routing Replay — a system for routing RL rollouts in MoE architectures
- Blackwell-specific tip: colocate specific resources for optimal performance
- Ray is the orchestration layer for distributed RLHF compute
Concepts
- RL Infrastructure — production RLHF infrastructure documentation