shuhand0

Fp64 is not supported in jax-metal backend and is unknown when the support will be there as of now. We will post the update here if the situation changes.

Will you be able to reproduce the issue with a smaller module?

We are aware of the issue and will integrate the solution in next release.

Looking into add the conversion of the op.

This is a case of the dimension_numbers we didn't support. We will look into expanding our conversion patterns for convolution op.

The issue is lacking of conv1d support. Will look into adding a conversion sequence to shape it to conv2d.

We are looking into adding the conversion for the op.

The backend kernel doesn't support rank > 4 for reduce op. Is it possible for the app to work around the issue by reshaping the tensor, e.g., ```a = jnp.zeros(...

The reduction dimension is not limited to 1 nor all axes from stablehlo spec, so the pattern(reshape-reduce-reshape) will not resolve all the cases. We will look into whether a more...

Thanks for requesting the feature, and we will track it and update here when we have a plan.