What problem does it solve? Setting up reinforcement learning post-training for large language models requires coordinating distributed training (Megatron-LM) with high-throughput rollout generation (SGLang), which involves complex configuration, data buffering, and failure handling that is difficult to get right manually. ## Core Features & Use Cases - GRPO and Async Training Workflows: Step-by-step launch commands for standard GRPO training, asynchronous rollout-training overlap, and multi-turn agentic training with custom generate functions. - Configuration Reference: Covers the three argument categories (Megatron, SGLang-prefixed, and slime-specific) plus the batch-size constraint formula and pre-configured model scripts for GLM, Qwen3, DeepSeek V3, and Llama 3. - Troubleshooting Guidance: Diagnoses SGLang engine crashes, weight sync timeouts, CUDA OOM, reward collapse, and data loading bottlenecks with concrete flag-level fixes. - Use Case: A researcher fine-tuning Qwen3-4B with GRPO on math reasoning data can follow the workflow to prepare JSONL prompts, source the model script, launch training across 8 GPUs, and monitor reward curves in TensorBoard. ## Quick Start Ask the agent to help you launch a GRPO training run with slime for your model and JSONL prompt dataset, including the correct parallelism and batch-size flags.
