Prerequisites

  • Architecture: x86_64
  • Chip models: Hygon BW1000
  • Host driver: 6.3.30-V1.4.1a

Image contents

Python

3.10

Application package

sglang==0.5.11

sglang-fl==0.2.0

Component versions

ComponentVersion
FlagGems5.4.0 (6ed2f390)
FlagTree0.7.0+triton3.6 (3bc8649b)
FlagCX0.13.0
triton3.6.0 (backend: hcu)
torch2.10.0

Environment

  • LD_LIBRARY_PATH must additionally contain /usr/local/lib/python3.10/dist-packages/torch/lib:/opt/dtk/cuda/cuda-12/lib64, otherwise FlagCX fails to load with libcudart.so.12: cannot open shared object file
  • FLAGCX_PATH=/sgl-workspace/FlagCX
  • SGLANG_FL_DIST_BACKEND=flagcx
  • SGLANG_FL_FLAGOS_BLACKLIST=cumsum,layer_norm
  • SGLANG_FL_PREFER=vendor
  • SGLANG_FL_STRICT=1

Launch

Published: harbor.baai.ac.cn/flagos-app/sglang0.5.11-hygon-dtk26.04:2.2.0-0.2.0

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

  IMG=harbor.baai.ac.cn/flagos-app/sglang0.5.11-hygon-dtk26.04:2.2.0-0.2.0

Without a toolkit — plain docker / podman

This image ships no default entrypoint command, so the launch command must be given explicitly.

Start an interactive shell:

  docker run --rm -it \
  --network host \
  --ipc host \
  --privileged \
  --group-add video \
  --cap-add SYS_PTRACE \
  --security-opt seccomp=unconfined \
  --device /dev/kfd \
  --device /dev/mkfd \
  --device /dev/dri \
  -v /opt/hyhal:/opt/hyhal:ro \
  -e HIP_VISIBLE_DEVICES=0,1 \
  -e CUDA_VISIBLE_DEVICES=0,1 \
  $IMG bash

⚠️ Device selection must be pinned at docker run -e time (the HIP_VISIBLE_DEVICES / CUDA_VISIBLE_DEVICES above); exporting them later inside the container makes device enumeration hang.

Then, inside the container, start the server with tp=2:

  source /opt/dtk/env.sh
export LD_LIBRARY_PATH=/usr/local/lib/python3.10/dist-packages/torch/lib:/opt/dtk/cuda/cuda-12/lib64:$LD_LIBRARY_PATH
export FLAGCX_PATH=/sgl-workspace/FlagCX
export SGLANG_FL_DIST_BACKEND=flagcx
export SGLANG_FL_FLAGOS_BLACKLIST=cumsum,layer_norm
export SGLANG_FL_PREFER=vendor
export SGLANG_FL_STRICT=1
python3 -m sglang.launch_server --model-path <path> --host 0.0.0.0 --port 30000 --tp 2 --page-size 64 --disable-radix-cache --context-length 262144 --mem-fraction-static 0.80 --chunked-prefill-size 4096 --max-running-requests 64 --trust-remote-code --disable-piecewise-cuda-graph --reasoning-parser qwen3

For a thinking model (e.g. the Qwen3 family), --reasoning-parser qwen3 is required — otherwise the chain of thought is not separated into reasoning_content.

Edit this page

Last updated 24 Sep 2026, 15:33 +0800. history