Eric Meier

``` components = { "unet": UNet2DConditionModel.from_pretrained( checkpoint, subfolder="unet", use_safetensors=True, torch_dtype=torch.bfloat16 ), } components['unet'].to(device) # device is xla, it fails here ``` Just loading a unet from diffusers, not sure what...

Yeah tried with your versions, and it seems to load the model fine, let's hope everything else goes well as well 🙏

Seems to work fine, thanks again for your help @JackCaoG 🙏 TLDR for anyone who stumbles upon this: Use `openxla` as the backend, not `openxla_eval`. Use the nightly builds mentioned...

Fixes https://github.com/lancedb/lance/issues/2615 as well

Added some untested tests, not sure how to run integration tests locally to see if they work

Can we get a release with this fix? 🙏