The official slime Ascend docker images
Maintained by: openEuler CloudNative SIG
Where to get help: openEuler CloudNative SIG, openEuler
Current slime docker images are built on the openEuler. This repository is free to use and exempted from per-user rate limits.
slime is an open-source RL training framework from THUDM designed for large-scale post-training of large language models. It supports synchronous/asynchronous agentic RL and RLHF workflows on top of SGLang inference engines and Megatron training backends.
This image ships the Ascend-adapted slime tree from the Ascend/slime-ascend fork (upstream slime has no Ascend support) together with torch_npu, so 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 (quick start runbook and setup scripts).
Read more about Ascend at hiascend.com and explore the slime repository at github.com/THUDM/slime.
The tag of each slime docker image consists of the slime version and the version of the software stack. The details are as follows
| Tags | Currently | Architectures |
|---|---|---|
| 0.4.0-torch_npu2.10.0-cann9.1.0-python3.12-oe2403sp4 | slime 0.4.0, Ascend-adapted fork at c69676f (torch_npu 2.10.0.post6, CANN 9.1.0) on openEuler 24.03-LTS-SP4 | amd64 |
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)
# Update DEVICE according to your device (/dev/davinci[0-7])
export DEVICE=/dev/davinci0
# Update the slime image
export IMAGE=openeuler/slime:0.4.0-torch_npu2.10.0-cann9.1.0-python3.12-oe2403sp4
docker run --rm \
--name my-slime \
--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 $IMAGE bash
Validate the environment inside the container:
python3 -c "import torch, torch_npu; print('torch=', torch.__version__); print('torch_npu=', torch_npu.__version__); print('is_available:', torch.npu.is_available()); print('count:', torch.npu.device_count())"
python3 -c "import slime; print('slime:', slime.__file__)"
The slime source tree (with the examples/ directory of the Ascend fork) is installed at /opt/slime-ascend. Run a minimal asynchronous RL recipe the same way the Ascend CI does (four NPUs, Qwen/Qwen2.5-0.5B-Instruct from ModelScope, two rollouts):
cd /opt/slime-ascend
export PYTHONPATH=/opt/slime-ascend:$PYTHONPATH
bash examples/fully_async/run-qwen2.5-0.5B-fully_async.sh
The example script launches the Ray cluster and the SGLang rollout engines on the NPU and finishes after the configured rollouts plus optimizer steps. The SGLang and Megatron backends required by the rollout engines are not bundled in this image; follow the Ascend fork's scripts/ascend_script/quick_install.sh and the Ascend quick start runbook to install the matched stack.
If you have any questions or want to use some special features, please submit an issue or a pull request on openeuler-docker-images.
Content type
Image
Digest
sha256:acf983347…
Size
5.2 GB
Last updated
7 days ago
docker pull openeuler/slimePulls:
32
Last week