This walkthrough is planned but not published yet. The transcript below covers the same material if you would rather read it now.
Use .options() to ask for GPUs on the calls that need them, and only those.
Resource requirements belong on the call, not the definition. The options method takes num_cpus and num_gpus and returns a configured version of your distributed function, which you call with .remote(). The scheduler places that call on a node that can satisfy the request. Requesting a GPU for every task in a mixed pipeline wastes capacity and makes tasks queue behind GPU availability they do not need.