Commit 8f2bc575 authored by Ayush Dubey's avatar Ayush Dubey Committed by TensorFlower Gardener
Browse files

Reuse GPUs for multiple ranks in NcclManager test.

Since NCCL 2.1.2, NCCL uses cooperative group launch by default which
results in deadlocks in the NcclManager test if number of ranks > number of
GPUs.  Setting NCCL_LAUNCH_MODE to PARALLEL, which was the default before NCCL
2.1.2, fixes this.

PiperOrigin-RevId: 224531851
parent b56d14fb
Loading
Loading
Loading
Loading
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Please to comment