vime Documentation#
vime is an LLM post-training framework for RL scaling, providing two core capabilities:
High-Performance Training: Supports efficient training in various modes by connecting Megatron with vLLM;
Flexible Data Generation: Enables arbitrary training data generation workflows through custom data generation interfaces and server-based engines.
vime is built on slime, the RL framework behind GLM-4.7, GLM-4.6 and GLM-4.5. vime keeps slime’s training stack and data-generation design while using vLLM as the default rollout backend, and inherits broad model support from slime, including:
Qwen3 series (Qwen3Next, Qwen3MoE, Qwen3), Qwen2.5 series;
DeepSeek V3 series (DeepSeek V3, V3.1, DeepSeek R1);
Llama 3.
Start by Use Case#
New to vime: Quick Start
Configure training and rollout arguments: Usage Guide
Add custom generation, reward, or rollout functions: Customization Guide
Build agentic RL workflows: Agentic RL Training Roadmap
Configure production vLLM rollout topology: vLLM Config: Advanced Engine Deployment
Connect external rollout engines: External Rollout Engines Roadmap
Sync weights as byte-level deltas: Delta Weight Sync
Use PD disaggregation: PD Disaggregation
Use BF16 training with FP8 rollout or FP8 KV cache: Low Precision Training and Rollout
Understand CI and reliability coverage: CI (Continuous Integration)
Debug, trace, and profile long-running jobs: Debugging, Trace Viewer, Profiling
Get Started
MoE
Advanced Features
Other Usage
Developer Guide
Hardware Platforms