Lukasz Stafiniak comments

Results 114 comments of


                                            Lukasz Stafiniak

Introduce a new log level `Exclude_prefixed`, or `Explicit_excluding_prefixed`

Obsolete now that we plan to get rid of prefix filtering.

Output hyperlinks for the PrintBox text backend once newest version of printbox-text is released

But be careful because test diffing strips away and otherwise messes up ANSI-wrapped text! I wasted at least half a day on this.

Instead of introducing category theory by itself, discuss the `Preface` library

[GitHub page](https://github.com/xvw/preface)

Implement "parameter punning" for the code notation `%cd`

Let's have it also populate the label with the labels collected from the right-hand-side of the punning line. Then we could actually use it in the `Train.sgd_one` code!

Move the `%cd` syntax to the arrayjit library / package

We should move shape inference to arrayjit, and add to arrayjit a "tensor light" variant of `Tensor` that does rootness tracking and shape inference handling, without backpropagation.

Move the `%cd` syntax to the arrayjit library / package

Don't agree anymore with the previous comment about rootness tracking. Regarding shape inference, it's tricky. Regarding `%cd`, it wouldn't work without shape inference. Altogether this needs to be driven by...

Implement an xLSTM LLM

https://github.com/NX-AI/xlstm

Study and incorporate Andrej Karpathy's `llm.c` lessons

The "study" part is certainly aiming at versions 0.6.x, but many solutions will wait till 0.9.x.

Add an LLVM / clang backend

It could perhaps eventually also be used for Nvidia: https://www.cse.ust.hk/~weiwa/papers/gass-ppopp22-poster.pdf "POSTER: An LLVM-based Open-Source Compiler for NVIDIA GPUs" Da Yan, Wei Wang, Xiaowen Chu

Concatenate along a new or an existing axis: `einsum` extension

Bumping this to high priority.