slime-rl-training

1# slime: LLM Post-Training Framework for RL Scaling
2 
3slime is an LLM post-training framework from Tsinghua's THUDM team, powering GLM-4.5, GLM-4.6, and GLM-4.7. It connects Megatron-LM for training with SGLang for high-throughput rollout generation.
4 
5## When to Use slime
6 
7**Choose slime when you need:**
8- Megatron-LM native training with SGLang inference
9- Custom data generation workflows with flexible data buffers
10- Training GLM, Qwen3, DeepSeek V3, or Llama 3 models
11- Research-grade framework with production backing (Z.ai)
12 
13**Consider alternatives when:**
14- You need enterprise-grade stability features → use **miles**
15- You want flexible backend swapping → use **verl**
16- You need PyTorch-native abstractions → use **torchforge**
17 
18## Key Features
19 
20- **Training**: Megatron-LM with full parallelism support (TP, PP, DP, SP)
21- **Rollout**: SGLang-based high-throughput generation with router
22- **Data Buffer**: Flexible prompt management and sample storage
23- **Models**: GLM-4.x, Qwen3, DeepSeek V3/R1, Llama 3
24 
25## Architecture Overview
26 
27```
28┌─────────────────────────────────────────────────────────┐
29│                    Data Buffer                          │
30│ - Prompt initialization and management                  │
31│ - Custom data generation and filtering                  │
32│ - Rollout sample storage                                │
33└─────────────┬───────────────────────────┬───────────────┘
34              │                           │
35…