Kyle Daruwalla

Agreed. Perhaps Flux docs should start pointing users to these libraries too? The most stable way to save a Flux model right now is to save the named tuple representation...

One thing we can do on the Flux side is to make clear that there are two options for storing models: 1. The code prior to loading will define the...

Taking a quick glance at the test suite indicates that we never attempt to read an intermediate slice outside of the first chunk in a shard. Any access beyond the...

Yeah that's what the commit does. The test failure of a segfault is puzzling, but I believe it is genuine. ~~I'm using my PR branch in another codebase, and I'm...

Yes the loading logic already supports that, and the plan was to move to that once we have multiple weight options.

When you say "same switch" do you mean defining something like `Flux.throw_sanity_check_errors()` which flips the requested precision switch in the OP as well as turning off `CUDA.allowscalar`? Or do you...

One option is to build the default `functor` on ConstructionBase (which will recurse arrays). Then adopt #41 with a `Numeric` tag that applies to numbers and numeric arrays.

Jax uses [`tree_structure`](https://jax.readthedocs.io/en/latest/_autosummary/jax.tree_util.tree_structure.html#jax.tree_util.tree_structure) to return the `treedef`. I think it could be confusing to use `*structure` both in verb and noun form. Could we do `structural_def(x)` or `structural(x)` to imply...

What about `tree_def(x)` or `graph_def(x)`? So "give me the graph definition that underlies these nested `struct`ures."

I think any ancestral tree/graph analogies beyond referring to the parent/children of a single node is confusing. I am okay with something like `strip` or `simplify` (or the noun versions).