Frequently Asked Questions (FAQ)
How to control the GPU devices used by Alpa?
Method 1: Use CUDA_VISIBLE_DEVICES
CUDA_VISIBLE_DEVICES
works for alpa, but there are some caveats.
If you use Ray cluster, you should not put CUDA_VISIBLE_DEVICES
before the python script you run.
You should apply this environment variable to ray start --head
. For example, CUDA_VISIBLE_DEVICES=0,1 ray start --head
.
Method 2: Use arguments in alpa.init
You can use the arguments of alpa.init
to configure the number of devices to use.
See the docstring
Method 3: Use other Ray features
If you are familiar with Ray, you can use advanced Ray features like placement group.