Prerequisites

  • Architecture: x86_64
  • Chip models: NVIDIA H20
  • Host driver: 610.43.02
  • Container toolkit (optional) : nvidia-container-toolkit

Image contents

Built on

harbor.baai.ac.cn/flagos-base/flagos-base-nvidia-cuda13.3:2.1.2 open_in_new

Python

3.12

Major Python packages

  • flag_gems==5.3.4
  • flagtree==0.6.1
  • numpy==2.3.5
  • torch==2.11.0+cu130
  • torchaudio==2.11.0+cu130
  • torchvision==0.26.0+cu130
  • triton==3.6.0

Switch compiler

This image includes both FlagTree (default) and Triton. To switch, run compiler triton inside the container. Use compiler flagtree to switch back, or compiler to check the active compiler.

Launch

With the container toolkit (optional):

  docker run --rm -it \
  --gpus all \
  harbor.baai.ac.cn/flagos-runtime/flagos-runtime-nvidia-cuda13.3:2.1.2 bash

Without a toolkit — plain docker / podman:

  docker run --rm -it \
  --device /dev/nvidia0 \
  --device /dev/nvidiactl \
  --device /dev/nvidia-uvm \
  -v /usr/bin/nvidia-smi:/usr/bin/nvidia-smi:ro \
  -v /usr/lib/x86_64-linux-gnu/libnvidia-ml.so.1:/usr/lib/x86_64-linux-gnu/libnvidia-ml.so.1:ro \
  -v /usr/lib/x86_64-linux-gnu/libcuda.so.1:/usr/lib/x86_64-linux-gnu/libcuda.so.1:ro \
  harbor.baai.ac.cn/flagos-runtime/flagos-runtime-nvidia-cuda13.3:2.1.2 bash

Verify

Inside the container, confirm the accelerator is visible:

  nvidia-smi
Edit this page

Last updated 15 Aug 2026, 12:51 +0800. history