Closed aittalam closed 7 months ago
A note on this can be added to the documentation in the bottom of the README##Ray job submission
section.
Just got this issue as well replicating a run from scratch, will open a PR for it
Resolved with the above PR ^
Jobs which do not rely on
ray.remote
, such aslm-harness
, run without GPUs by default.If we want to add GPUs to our jobs, we need to manually add gpus to
ray.submit
(see here), for instance when it is called in the example notebooks.This is ok, I just wanted to raise this to make sure we describe this behavior in the documentation. The reason is that
lm-harness
jobs break if e.g. we run an evaluation on a quantized model without using a GPU.