Last updated: 06/16/2026.
This document describes how to use verl for reinforcement learning training on Cambricon MLU (MLU370 / MLU590).
Here we list all MLU related files for reference, we will continue to add new features.
verl_hardware_plugin/
├── engines
├── cncl_checkpoint_engine.py # support checkpoint engine
├── fsdp_mlu.py # fsdp related model support
├── megatron_mlu.py # megatron related model support
└── platforms
└── platform_mlu.py # basic platform settings
And docs to start.
user_guide_flagos/
├── README.md # This file
├── install_guidance.md # Installation guide
└── quick_start.md # Quick start
- Installation Guide — Docker setup, component installation
- Quick Start — Run your first GRPO training job
| Item | Description |
|---|---|
| Device type | mlu |
| Vendor identifier | cambricon |
| Communication backend | cncl |
| Device visibility env var | MLU_VISIBLE_DEVICES |
| Ray resource name | GPU |
| IPC support | Yes |