v0.8.0-rc0 Break in compatibility: DType int64 -> Go's int64 (it was Go's int) and many small improvements.

[janpfeifer]

This release fixes a long standing mistake: GoMLX previously made the assumption that Go int are 64 bits on the platforms that mattered -- it was made out of convenience in the early versions.

With this release we have the following conversion changes:

1. Slices and types converting from Go int will get a Tensor with DType Int32 or Int64 depending on the bit size of the int (it uses strconv.IntSize to check). If the bitsize is different (128 bits?) it panics and simply doesn't convert. This doesn't break compatibility.
2. Tensors of DType Int64 converting to Go will convert to (multi-dimension) slices of int64 and no longer int. This breaks compatibility, and may break your code!!. The fix is generally trivial, but still, it requires fixing.

The breaking is not done lightly ... but it seems worth the fix. Hopefully, there will be fewer of those in the future on the main APIs, and soon I can build a 1.0 release.

This release also includes many small improvements and fixes, listed below.

This is still an RC release, expect a few changes before the full v0.8.0 release.

• DType and Tensors:
  • Added support to Go's int64 -- breaks compatibility because DType Int64 when converted back to Go becomes int64
    and not int.
  • Renamed Local.Flat -> Local.FlatCopy : not to be mixed with LocalRef.Flat (which is not a copy).
• C++ code integrating with XLA:
  • Enable copy elision -- which makes std::move not necessary.
  • Temporarily copied xla/mlir/utils library to third-party/xla_mlir, since it is not available in all XLA distributions.
• Package context:
  • Added context.GetParamOr and context.GetGraphParamOr: it uses generics to cast to the desired type, and allowing a default value to be returned.
• Added recovery of some basic types (numeric and slices) when loading params from Json with checkpoints.
• Package train: Loop automatically sets LoopStep to context's "global_step" parameter.
• Package optimizer: Added GetGlobalStep.
• Package context: Added Context.EnumerateVariablesInScope() method.
• Package graph:
  • Added optional reduceAxes parameter to L2Norm and L1Norm.
  • Added L2NormSquare, L2Normalize and L2NormalizeWithEpsilon.
• Package nanlogger: added AttachToTrainer; improved docs.
• Package margaid:
  • automatic ending plot when loop finishes.
  • option to plot evaluation losses separately from training losses -- for when they include different terms.
• Example "Dogs vs Cats":
  • Added Byol (Bootstrap Your Own Latent) regularized models.
  • Added support for generating pairs of images for BYOL model.

---

This discussion was created from the release v0.8.0-rc0 Break in compatibility: DType int64 -> Go's int64 (it was Go's int) and many small improvements..