Vime
An LLM post-training framework for reinforcement-learning scaling, built on slime's training and data-generation design, that uses vLLM with a router as its default rollout backend.
Vime provides high-performance training by connecting Megatron with vLLM for efficient training in various modes, and flexible data generation through custom interfaces and server-based engines, covering arbitrary training-data-generation workflows. It inherits slime's broad model support, including the Qwen, DeepSeek V3 and Llama 3 families.
Its training module (Megatron) reads from a data buffer and synchronizes parameters to the rollout module after training; its rollout module (vLLM plus a router) launches vLLM inference engines and routes generation requests, and custom generate functions can wrap generation with multi-turn loops, tool calls, environment or sandbox interaction, and verifier-based rewards. The data buffer bridges prompt initialization, custom data, and rollout generation, including agentic workflows that produce samples through the same interface.
The project positions itself alongside other vLLM-ecosystem post-training frameworks such as NeMo RL, OpenRLHF, prime-rl, SkyRL and verl, aiming to bring slime's training paradigm into the vLLM ecosystem on aligned release cycles.