We just released flyte-sdk version v2.0.0b25 - Its a massive release and effort by a lot of folks! Top features: • Triggers - cron and rate schedules • Blazing fast data downloads. 5GB in 5s on a 10Gbps • Dist training support (torch elastic) • Support for

tranium, gpus, and tpus

• Sync/async apis for file and dir • flyteidl2 usage • Faster flyte imports and cli • Lots of examples for HoF, UDFs • Context passing in tasks • Push/pull private images • Provides externally built images by name -

Image.from_ref_name("my_image")

. --

flyte run --image my_image="<http://docker.io/|docker.io/>...."

Thank you to everyone for all the feedback so far on the Flyte 2.0 beta 🙏 flyte https://github.com/flyteorg/flyte-sdk/releases/tag/v2.0.0b25

@powerful-horse-58724 One quick question: Does it also support “mps” apple mac silicon. It will be helpful for local development and now even pytorch supports “mps” I think this can be added as a

device

type

Using mps for local runs should work. I might try to make an example later when I'm back at computer, but you can could have something like: if run==local: device="mps" and pass that value to the workflow.

It should automatically detect the device type, if not explicitly provided: For example if cuda: then cuda; elif mps the mps else cpu

You should be able to do that too If you can do it in python it can be part of your workflow!

Ya that is true. I was thinking to make it part of the library itself.