miles-rl-training

You can execute RL training for large MoE models using a Miles-based workflow that applies FP8 and INT4 optimization alongside Megatron-style parallelism. This setup provides enterprise-grade guidance for production-scale training and rollout routing replay.

Train-inference alignment in speculative RL workflows ensures consistency between training and inference states during MoE model optimization. The Miles workflow provides guidance to maintain this alignment, which is critical for robust rollout routing replay in production environments.

Yes, you can use SGLang with Megatron-style parallelism for MoE model training within the Miles workflow. This combination satisfies production requirements for integrating rollout routing replay and executing robust, large-scale RL training.

Yes, you need Ray and PyTorch version 2.0.0 or higher, along with Transformers 4.40.0 or higher and SGLang-router 0.2.3 or higher. These dependencies are required to support the production-ready Miles RL training workflow for large MoE models.

The best way to handle rollout routing replay is through the Miles workflow, which integrates SGLang to manage routing replay robustly. This approach ensures train-inference alignment and supports speculative RL workflows for large MoE models.

MoE model training requires specialized RL workflows to address the complexity and scale challenges of production environments. The Miles workflow provides enterprise-grade guidance for applying FP8 and INT4 optimization, train-inference alignment, and routing replay to manage these challenges robustly.