Skip to content
This repository has been archived by the owner on Jul 22, 2024. It is now read-only.

horovod in powerai 1.6.2, error occurs with hvd.init() #264

Open
wacoder opened this issue Jun 11, 2020 · 2 comments
Open

horovod in powerai 1.6.2, error occurs with hvd.init() #264

wacoder opened this issue Jun 11, 2020 · 2 comments

Comments

@wacoder
Copy link

wacoder commented Jun 11, 2020

Hi,
Try to install Horovod with PyTorch based on powerai image following official tutorial, error occurs when try to hvd.init(). Is there any material on how to set up DDL_OPTIONS to use ddl? Thanks

  • code

import horovod.torch as hvd
hvd.init()
[2020-06-11 03:01:22.297540: W horovod/common/ops/ddl_operations.cc:105] DDL backend has been deprecated. Please, start using the NCCL backend by building Horovod with 'HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_GPU_BROADCAST=NCCL'.
terminate called after throwing an instance of 'std::logic_error'
what(): DDL_OPTIONS env variable needs to be set to use DDL.
Aborted (core dumped)`


I also tried to set up env variable for DDL_OPTIONS with export DDL_OPTIONS=DDL it produce another error

`

import horovod.torch as hvd
hvd.init()
[2020-06-11 03:14:33.814715: W horovod/common/ops/ddl_operations.cc:105] DDL backend has been deprecated. Please, start using the NCCL backend by building Horovod with 'HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_GPU_BROADCAST=NCCL'.
E 03:14:33.814 634 652 DDL:86 ] ill-formated option :[DDL]
terminate called after throwing an instance of 'std::system_error'
what(): Resource deadlock avoided
Aborted (core dumped)
`

@bnemanich
Copy link

You will need to launch the job using the ddlrun tool. This will setup the appropriate environment variables. For more information on ddlrun, see https://www.ibm.com/support/knowledgecenter/SS5SF7_1.6.2/navigation/wmlce_ddlrun.html.

@wacoder
Copy link
Author

wacoder commented Jun 11, 2020

@bnemanich thanks very much. Does that mean I can use horovod in script, and use ddlrun to run?

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants