Configuring Dask to target GPU or CPU workers in a heterogeneous cluster

Distributed
1

I’m working with a manually deployed Dask cluster across multiple machines. Some of the worker nodes have GPUs, while others do not.

I’d like to understand how to group or label Dask workers such that:

  • Workloads that require GPU are scheduled only on GPU-enabled workers, and

  • Workloads that do not require GPU are scheduled only on CPU-only workers.

How can I configure this behavior at the worker and client level?
I’d really appreciate a concrete example, as I haven’t found a clear explanation for this setup in the official documentation.

Thanks in advance!

2

Hi @doughng,

Have you came across the Worker Resources page of the documentation?

There are several examples here on declaring them and using them in various contexts.

If not enough, could you describe your use case in more details so that we can better help?