Lukasz Stafiniak comments

Results 114 comments of


                                            Lukasz Stafiniak

Any lessons from Imbue for training-in-the-large?

Also, from llm.c: https://github.com/karpathy/llm.c/discussions/677

Implement local let-bindings in the `%nn_cd` syntax

This is needless complexity.

Implement local let-bindings in the `%nn_cd` syntax

This does make sense. The environment needed will hold a kind (e.g. non-diff tensor, code, array etc.) and slot (LHS, RHS1 etc.) of the bindings.

Implement a fork-based backend as stepping stone to distributing across machines

Not helpful, `Domain`-based devices already implemented.

Implement a fork-based backend as stepping stone to distributing across machines

Restoring this as it provides more angles on the design space. (And even possibility for efficiency gain due to not using domain-based GC, but that should be irrelevant to the...

Deep dives into Petalisp and Caten

The Discord of Caten is very informative. Although my initial reaction to the code was I'm overwhelmed, this looks increasingly promising.

Rename (virtual) `device` -> `stream`, and `physical_device` -> `device`

Moving this to 0.5 to have a semblance of stability (it's trivial but breaks all prior usage).

Rename (virtual) `device` -> `stream`, and `physical_device` -> `device`

Moving this back to 0.4.2 -- it's a deeply breaking change but fits better with the scope of 0.4.2.

Rename (virtual) `device` -> `stream`, and `physical_device` -> `device`

This is fixed with the most recent commit https://github.com/ahrefs/ocannl/commit/502ec4e5a85dca6e3e3788bdabff0ccc6e3ee018

ppx_op: Include `config.label` in the label of the primary (returned) tensor, when available, and not incorporating a label from a function argument

This feature might interact badly with #210 leading to labels that are too bloated. It would still make sense if there is no function (tensor) argument, but such cases would...