Lukasz Stafiniak

https://github.com/pytorch/ao/tree/main/torchao/prototype/low_bit_optim https://x.com/marksaroufim/status/1809398186198593566

I have a feeling it might be related to `Utils.waiter`. https://github.com/ahrefs/ocannl/actions/runs/9488230357/job/26146684090

https://arxiv.org/abs/1804.10694 > Tiramisu uses a four-level intermediate representation that allows full separation between the algorithms, loop transformations, data layouts, and communication.

https://github.com/bytecodealliance/wasmtime/tree/main/cranelift

https://github.com/huggingface/candle The goals of Candle are a bit less aligned with OCANNL than those of Luminal, e.g. Candle has hand-written kernels. But Candle is minimalistic and already very efficient.

https://arxiv.org/abs/2405.10480

This would simplify things and preemptively solve some potential user problems.

Delimited scopes are less surprising. Do we need a good name for extended scopes e.g. `%debug_below_sexp`, `%track_below_sexp`, or can we get rid of them?

It is significant complexity in the implementation, too cumbersome to write, too distracting to read, and doesn't integrate well with projects as a logging mechanism. Instead of prefix-based, we will...

This would be great to quickly get in-depth tracing for large parts of code.