Prerequisites

  • Architecture: x86_64
  • Chip models: Hygon BW1000
  • Host driver: 6.3.30-V1.4.1a
  • Container toolkit (optional) : dcu-container-toolkit >= 1.3.0

Image contents

Built on

harbor.baai.ac.cn/flagos-runtime/flagos-runtime-hygon-dtk26.04:2.2.0 open_in_new

Python

3.10

Application package

megatron-core[training]==0.18.2

Launch

Published: harbor.baai.ac.cn/flagos-app/megatron_training0.18.2-hygon-dtk26.04:2.2.0-0.3.0

The image name is long — assign it to a variable first:

  IMG=harbor.baai.ac.cn/flagos-app/megatron_training0.18.2-hygon-dtk26.04:2.2.0-0.3.0

The two approaches below are alternatives — pick the one that matches how your host runs containers:

With the container toolkit

Start an interactive shell:

  docker run --rm -it \
  -e DCU_VISIBLE_DEVICES=all \
  $IMG bash

Start the app with its default settings:

  docker run --rm -it \
  -e DCU_VISIBLE_DEVICES=all \
  $IMG

Pass arguments to the launcher:

  docker run --rm -it \
  -e DCU_VISIBLE_DEVICES=all \
  $IMG megatron-train --model-type GPT

Without a toolkit — plain docker / podman

Start an interactive shell:

  docker run --rm -it \
  --device /dev/kfd \
  --device /dev/mkfd \
  --device /dev/dri \
  --group-add video \
  -v /opt/hyhal:/opt/hyhal \
  --security-opt seccomp=unconfined \
  $IMG bash

Start the app with its default settings:

  docker run --rm -it \
  --device /dev/kfd \
  --device /dev/mkfd \
  --device /dev/dri \
  --group-add video \
  -v /opt/hyhal:/opt/hyhal \
  --security-opt seccomp=unconfined \
  $IMG

Pass arguments to the launcher:

  docker run --rm -it \
  --device /dev/kfd \
  --device /dev/mkfd \
  --device /dev/dri \
  --group-add video \
  -v /opt/hyhal:/opt/hyhal \
  --security-opt seccomp=unconfined \
  $IMG megatron-train --model-type GPT
Edit this page

Last updated 26 Sep 2026, 12:57 +0800. history