Mindspore动态组网方式无法使用PyNative MODE
收藏回复举报
Mindspore动态组网方式无法使用PyNative MODE
t('forum.solved') 已解决
发表于2024-05-21 19:27:46
0 查看

我想用动态组网方式启动单机多卡并行,我的脚本如下:

echo "=========================================="
echo "Please run the script as: "
echo "bash run.sh"
echo "It is better to use the absolute path."
echo "==========================================="

rm -rf device
mkdir device
cp ./parallel_test.py ./device
cd ./device
echo "start training"

i=0
export MS_WORKER_NUM=1
export MS_SCHED_HOST=127.0.0.1  # Scheduler IP address
export MS_SCHED_PORT=1400           # Scheduler port
export MS_ROLE=MS_WORKER
export MS_NODE_ID=$i
python ./parallel_test.py > worker_$i.log 2>&1 &


# Launch 1 scheduler.
export MS_WORKER_NUM=1
export MS_SCHED_HOST=127.0.0.1  # Scheduler IP address
export MS_SCHED_PORT=1400        # Scheduler port
export MS_ROLE=MS_SCHED
python ./parallel_test.py > scheduler.log 2>&1 &

代码如下

# parallel_test.py
ms.set_context()
ms.set_auto_parallel_context(parallel_mode=ms.ParallelMode.DATA_PARALLEL)
init()

报错如下

Traceback (most recent call last):
  File "/parallel_test/device/./parallel_test.py", line 22, in <module>
    init()
  File "/lib/python3.9/site-packages/mindspore/communication/management.py", line 180, in init
    init_cluster()
RuntimeError: PyNative mode is not supported in dynamic networking for now.

----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore/ccsrc/distributed/cluster/cluster_context.cc:217 InitNodeRole

运行机器:GPU RTX2080ti 希望尽快支持Pynative模式,GRAPH MODE模式不支持AnnAssign。两者都不能使用动态组网方式运行我的代码。希望尽快适配。

我要发帖子