/slime-user

SLIME Documentation Navigation Guide

Getting Started

Quick Start (Priority: High)

Path: docs/en/get_started/quick_start.md Purpose: Essential first read for setting up SLIME within one hour Contents:

• Environment setup (Docker recommended, hardware support)
• Model and dataset download
• Weight conversion (HF ↔ Megatron)
• Training script execution and parameter overview
• Key features: colocated actor/rollout, dynamic sampling, partial rollout, bf16 training with fp8 inference

Usage Guide (Priority: High)

Path: docs/en/get_started/usage.md Purpose: Comprehensive parameter reference and configuration guide Contents:

• Cluster resource allocation (actor/rollout GPUs, colocate mode)
• Training backend selection (Megatron vs FSDP)
• Megatron configuration (model params, parallelism, checkpoints)
• SGLang configuration (HF checkpoint, router setup)
• Data format (JSONL with prompt/label/metadata)
• RL hyperparameters (advantage estimators, loss calculation, TIS)
• Custom rollout/reward functions
• SGLang/Megatron integration details
• FSDP backend quick start

Customization (Priority: Medium)

Path: docs/en/get_started/customization.md Purpose: Advanced customization of SLIME workflows Use when: Building custom reward models, generation functions, or filters

Q&A (Priority: Low)

Path: docs/en/get_started/qa.md Purpose: Common issues and troubleshooting Use when: Debugging setup or runtime issues

Advanced Features

Fault Tolerance

Path: docs/en/advanced/fault-tolerance.md Purpose: Handling failures in distributed training Use when: Setting up multi-node training with automatic recovery

PD Disaggregation

Path: docs/en/advanced/pd-disaggregation.md Purpose: Prefill-Decode separation for improved efficiency Use when: Optimizing inference performance with separate prefill/decode servers

Speculative Decoding

Path: docs/en/advanced/speculative-decoding.md Purpose: Accelerating inference with draft models Use when: Need faster rollout generation

Architecture Support Beyond Megatron

Path: docs/en/advanced/arch-support-beyond-megatron.md Purpose: Using non-Megatron architectures Use when: Working with custom model architectures

Platform Support

AMD Tutorial

Path: docs/en/platform_support/amd_tutorial.md Purpose: Running SLIME on AMD GPUs Use when: Using AMD hardware instead of NVIDIA

Example Configurations

Single-Machine Examples

1. Qwen3-4B (docs/en/examples/qwen3-4B.md): 8xH100, basic setup
2. GLM4-9B (docs/en/examples/glm4-9B.md): 8xH100, includes dynamic sampling
3. Qwen3-4B Base (OpenHermes) (docs/en/examples/qwen3-4b-base-openhermes.md): FSDP backend example

Multi-Node Examples

1. Qwen3-30B-A3B (docs/en/examples/qwen3-30B-A3B.md): Multi-node MOE training
2. GLM-4.5-355B-A32B (docs/en/examples/glm4.5-355B-A32B.md): 64xH100, large-scale MOE
3. DeepSeek-R1 (docs/en/examples/deepseek-r1.md): 128xH100, massive scale

Developer Resources

Debug Guide

Path: docs/en/developer_guide/debug.md Purpose: Debugging techniques and tools Use when: Troubleshooting training issues

Blogs and Release Notes

Introducing SLIME

Path: docs/en/blogs/introducing_slime.md Purpose: Overview of SLIME framework and design philosophy

Release v0.1.0

Path: docs/en/blogs/release_v0.1.0.md Purpose: Initial release notes and features

Navigation Tips

1. For first-time users: Start with quick_start.md → usage.md
2. For basic training: Read quick_start.md and example docs (Qwen3-4B or GLM4-9B)
3. For advanced features: Check advanced/ directory after mastering basics
4. For custom workflows: Read customization.md and examples/ directory
5. For multi-node training: Study qwen3-30B-A3B.md or larger examples
6. For FSDP backend: See usage.md FSDP section and qwen3-4b-base-openhermes.md