openeuler/areal repository overview

⁠Quick reference

• The official AReaL Ascend docker images

• Maintained by: openEuler CloudNative SIG⁠

• Where to get help: openEuler CloudNative SIG⁠, openEuler⁠

⁠AReaL | openEuler

Current AReaL docker images are built on the openEuler⁠. This repository is free to use and exempted from per-user rate limits.

AReaL⁠ is an open-source, large-scale asynchronous reinforcement learning system for LLM post-training. It decouples rollout generation from training, supports RLHF/RLVR and agentic RL workflows, and scales from a single node to large clusters with fully/partially/asynchronous rollout modes.

This image ships AReaL with torch_npu⁠ so that post-training runs seamlessly on the Ascend NPU. The software stack is validated on Ascend 910B4 cards with CANN 9.1.0, torch==2.10.0 and torch_npu==2.10.0.post6 on Python 3.12, following the Ascend examples guard of the workflows repo (project scripts⁠).

Read more about Ascend at hiascend.com⁠ and explore the AReaL documentation at areal-project.github.io/AReaL⁠.

⁠Supported tags and respective Dockerfile links

The tag of each AReaL docker image consists of the AReaL version and the version of the software stack. The details are as follows

⁠Usage

⁠Quick start 1: supported devices

• Atlas A2 Training series (Atlas 800T A2, Atlas 900 A2 PoD, Atlas 200T A2 Box16, Atlas 300T A2)

• Atlas 800I A2 Inference series (Atlas 800I A2)

⁠Quick start 2: setup environment using container

# Update DEVICE according to your device (/dev/davinci[0-7])
export DEVICE=/dev/davinci0
# Update the areal image
export IMAGE=openeuler/areal:2.1.1-torch_npu2.10.0-cann9.1.0-python3.12-oe2403sp4
docker run --rm \
--name my-areal \
--device $DEVICE \
--device /dev/davinci_manager \
--device /dev/devmm_svm \
--device /dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /root/.cache:/root/.cache \
-it $IMAGE bash

⁠Quick start 3: validate the training environment on Ascend NPU

Validate the environment inside the container:

python3 -c "import torch, torch_npu; print('torch=', torch.__version__); print('torch_npu=', torch_npu.__version__); print('is_available:', torch.npu.is_available()); print('count:', torch.npu.device_count())"
python3 -c "from importlib.metadata import version; print('areal', version('areal'))"
areal --help

The AReaL source tree (with examples/) is installed at /opt/AReaL. Run a GSM8K SFT smoke test the same way the Ascend CI does (single NPU, Qwen/Qwen3-1.7B as the actor, FSDP training backend; the model and dataset are downloaded at runtime, use the HF mirror in China):

cd /opt/AReaL
export HF_ENDPOINT=https://hf-mirror.com
export SWANLAB_MODE=disabled
python3 examples/math/gsm8k_sft.py --config examples/math/gsm8k_sft.yaml \
    scheduler.type=local \
    cluster.n_gpus_per_node=1 \
    actor.backend=fsdp:d1p1t1 \
    actor.path=Qwen/Qwen3-1.7B \
    experiment_name=areal-sft-smoke \
    train_dataset.batch_size=2 \
    valid_dataset.batch_size=2 \
    +total_train_steps=1 \
    +train_dataset.setup_timeout=600 \
    +valid_dataset.setup_timeout=600 \
    stats_logger.wandb.mode=disabled

The vLLM rollout engine and the Megatron training backend are not bundled in this image; install the matched vLLM-Ascend stack and follow the upstream NPU guides before running rollout-backed examples.

⁠Question and answering

If you have any questions or want to use some special features, please submit an issue or a pull request on openeuler-docker-images⁠.