Luca Grementieri

A simple solution should be to just remove the parameter `dtype` inside the call to `np.asanyarray`. https://github.com/mikedh/trimesh/blob/7c90a94a427e7af1152bdc2dfb3e6ab6cfd2826c/trimesh/grouping.py#L604

The export of `return_11_almost_zeros ` generates the textual IR ``` ENTRY main.5 { Arg_0.1 = f32[2,2]{1,0} parameter(0) Arg_1.2 = f32[2,2]{1,0} parameter(1) constant.3 = f32[11]{0} constant({...}) ROOT tuple.4 = (f32[11]{0}) tuple(constant.3)...

Yes, I think that https://github.com/google/jax/blob/main/jax/tools/jax_to_ir.py#L133 should be modified to print large constants returning a fully-functional textual IR.

`_MetaAbstractArray` is the base class of all `jaxtyping` types and annotations so the check `isinstance(type_hint, _MetaAbstractArray)` is the best way to identify `jaxtyping` annotations. Surely there are workarounds to not...

Surely the first point should be verified. For the second point, I did some experiments and I think that `typing.get_type_hints` is sufficient to access shape annotations. I'll come back to...

From my experiments I think it's possible to support all the features using `jaxtyping` with no manual parsing! I think my pull request should be ready for next week.

I finally had time to push my PR https://github.com/brentyi/jax_dataclasses/pull/6 ! Sorry for the delay!