Hello, I noticed that the accelerator configuration is improved in Flyte 2. We had to do something similar in our Flyte 1 fork. One thing that I think might still be missing is a way to configure the container runtime class name depending on the accelerator being used. For nvidia gpu tasks we typically need to configure
nvidia-cdi
on the pods.
f
freezing-airport-6809
09/09/2026, 5:23 AM
we can add support for this too, do you want to propose. but also dont make it necessary
c
clean-glass-36808
09/09/2026, 5:33 AM
Yeah I will put something up
f
freezing-airport-6809
09/09/2026, 5:38 AM
the only thing is make it optional
c
clean-glass-36808
09/09/2026, 5:47 AM
Yeah of course.
clean-glass-36808
09/09/2026, 6:59 AM
Looks like we may be able to express this with the accelerator device class pod template so I'll try that out first