1
7
1
DeepSeek
K100-AI跑Deepseek-V4实测
2026-04-24 20:04:40
1550
https://modelscope.cn/models/FlagRelease/DeepSeek-V4-Flash-hygon-FlagOS/summary docker pull harbor.baai.ac.cn/flagrelease-public/flagrelease-hygon-release-model\_hygon-new-gems-tree-tree\_0.5.0\_hcu3.0-gems\_5.0.1rc0-cx\_none-python\_3.10.12-torch\_2.9.0\_das.opt1.dtk2604.20260206.g275d08c2-gpu\_hygon001-arc\_amd64-driver\_6.3.31-v1.5.0a:2026042410131 export NCCL\_SOCKET\_IFNAME=eno1 export NCCL\_IB\_DISABLE=1 export NCCL\_DEBUG=INFO export USE\_FLAGGEMS=1 export NCCL\_DEBUG=WARN export GEMS\_VENDOR=hygon docker run \\ --name flagos \\ --network=host \\ --ipc=host \\ --device=/dev/kfd \\ --device=/dev/mkfd \\ --device=/dev/dri \\ -v /opt/hyhal:/opt/hyhal \\ -v /baai:/baai \\ --group-add video \\ --cap-add=SYS\_PTRACE \\ --security-opt seccomp=unconfined \\ -itd \\ harbor.baai.ac.cn/flagrelease-public/flagrelease-hygon-release-model\_hygon-new-gems-tree-tree\_0.5.0\_hcu3.0-gems\_5.0.1rc0-cx\_none-python\_3.10.12-torch\_2.9.0\_das.opt1.dtk2604.20260206.g275d08c2-gpu\_hygon001-arc\_amd64-driver\_6.3.31-v1.5.0a:2026042410131 cd /workspace bash run\_node0.sh bash run\_node1.sh # The reasoning answers are DCD
全部评论(6)
只看楼主
C
CrystalRays
5楼
不是服务化部署。。。
6/10
杨
杨涌全
4楼
智力和吞吐量咋样, 如果和k100 的 minimax 2.5比
5/9
H
huangyy
回复 @杨涌全
同问
5/11
A
aq
楼主
3楼
用新的镜像继续测试
镜像拉取:
docker run \
  --name flagos \
  --network=host \
  --ipc=host \
  --device=/dev/kfd \
  --device=/dev/mkfd \
  --device=/dev/dri \
  -v /opt/hyhal:/opt/hyhal \
  -v /baai:/baai \
  --group-add video \
  --cap-add=SYS_PTRACE \
  --security-opt seccomp=unconfined \
  -itd \
  harbor.baai.ac.cn/flagrelease-public/flagrelease-hygon-release-model_hygon-new-gems-tree-tree_0.5.0_hcu3.0-gems_5.0.1rc0-cx_none-python_3.10.12-torch_2.9.0_das.opt1.dtk2604.20260206.g275d08c2-gpu_hygon001-arc_amd64-driver_6.3.31-v1.5.0a:202604242039

docker exec -it flagos /bin/bash

主节点配置:
cat /home/models/DeepSeek-V4-FlagOS-main/run_node0.sh
#!/bin/bash
# TCP 控制面
export NCCL_SOCKET_IFNAME=ens1f0np0
# RDMA 数据面
export NCCL_IB_DISABLE=0
export NCCL_IB_HCA=mlx5_0
export NCCL_IB_GID_INDEX=3
export NCCL_DEBUG=WARN # 避免过多日志,但保留警告
export USE_FLAGGEMS=1
export GEMS_VENDOR=hygon
CMD="torchrun --nproc-per-node=8 --nnodes=2 --node-rank=0 --master-addr=172.16.88.1 --master-port=29500 generate_dir.py --ckpt-path /home/models/DeepSeek-V4-Flash-hygon-FlagOS/ --config config_flash_v4.json --interactive"
exec ${CMD}
echo "主节点已启动"

从节点配置:
cat /home/models/DeepSeek-V4-FlagOS-main/run_node1.sh
#!/bin/bash
# TCP 控制面
export NCCL_SOCKET_IFNAME=ens1f0np0
# RDMA 数据面
export NCCL_IB_DISABLE=0
export NCCL_IB_HCA=mlx5_0
export NCCL_IB_GID_INDEX=3
export NCCL_DEBUG=WARN # 避免过多日志,但保留警告
export USE_FLAGGEMS=1
export GEMS_VENDOR=hygon
CMD="torchrun --nproc-per-node=8 --nnodes=2 --node-rank=1 --master-addr=172.16.88.1 --master-port=29500 generate_dir.py --ckpt-path /home/models/DeepSeek-V4-Flash-hygon-FlagOS/ --config config_flash_v4.json --interactive"
exec ${CMD}
echo "从节点已启动"

4/26
小
小洛克
2楼
666
4/26
K
kaiwong_mnt0vxc6icgq
1楼
:qiang:emoji:qiang:emoji:qiang:emoji 谢谢大神! 魔搭哪边已经有更新镜像。
4/24
没有更多内容了
--
楼主
Level.4
0 帖子
楼主热帖
暂无数据