AVRID refactors verl orchestration for LLM post-training

AVRID is a surgical strike on the complexity of distributed RL frameworks, prioritizing transparency and developer ergonomics over feature bloat.

• –Replaces verl's indirection with a "single-controller" architecture to simplify debugging in complex distributed environments
• –Implements token-aware dispatch to maximize GPU utilization by accounting for sequence length variation across shards
• –Leverages Ray placement groups to ensure tight co-location for high-performance generation and training rollouts
• –Fills a critical gap in MLOps education by documenting the architectural "why" of LLM post-training infrastructure