LiteRT Next GPU delegate doesn't support 5D yet

[BmanClark]

I am getting the following when trying to run a model on LiteRT Next (2.0.0-alpha) on GPU:

Following operations are not supported by GPU delegate:
ADD: Tensor <code line> has bad input dims size: 5.
DIV: Tensor <code line> has bad input dims size: 5.
RESHAPE: Tensor <code line> has bad input dims size: 5.
RESHAPE: Tensor <code line> has bad input dims size: 5.
RESHAPE: Tensor <code line> has bad input dims size: 5.
...

From my understanding this feature is coming?

I was hoping there would be an issue I could track to find out when it gets supported, but it doesn't seem to exist yet (just one for CoreML), so I'm creating it so others can upvote and we can hopefully see when it's resolved/implemented 😃

Thanks!

I also get an error:
[third_party/odml/litert/litert/runtime/accelerators/dispatch/dispatch_accelerator.cc:113] Dispatch library directory is not set.
How do I set this? Is something missing in the packaging everything GPU into com.google.ai.edge.litert:2.0.0-alpha?

[gaikwadrahul8]

Hi, @BmanClark
Thank you for bringing this issue to our attention and it seems like at the moment LiteRT Next utilizes has limitations on the dimensionality of tensors it supports for various operations. The errors you're seeing like ADD: Tensor <code line> has bad input dims size: 5.) confirm that 5D tensors are not currently supported for those operations on the GPU delegate

"Dispatch library directory is not set" error when using the AAR and targeting GPU likely points to an internal configuration or packaging aspect of the com.google.ai.edge.litert:2.0.0-alpha library that needs to be addressed by us especially for an alpha version, I'll discuss this issue with our internal team and will update you.

Thank you for your cooperation and understanding.