Ondřej Čertík

Allow to initialize nested structs using "Transactions(transactions=[], accounts=[])"

4

With the following definitions: ```python from lpython import dataclass, i64 @dataclass class TransactionItem: account_idx: i64 commodity: str amount: i64 @dataclass class Transaction: date: str name: str cleared: bool comments: str...

idea: `pytest` style `assert` based unit testing

1

We should expose the Assert ASR node to the frontend, either by introducing an assert statement, and/or just using existing syntax `call assert(x == 5)`, with the appropriate `assert` subroutine...

Use LFortran to compile any Fortran to WASM

7

LFortran (as of today) can compile 6 out of 18 Fortran packages in SciPy: https://github.com/lfortran/lfortran/pull/2743, and we keep working on compiling the rest. At the same time, I think it's...

Implement other sampling methods

2

Currently only the "greedy" sampling is implemented (the token with the highest probability is selected). Implement other sampling methods, some options are: * top-p * top-k * temperature (here is...

TODO: * [ ] Finish it to work correctly for any MPI rank * [ ] Write a test for the result, ensuring it is correct * [ ] Add...

Benchmark against PyTorch with Accelerate enabled

3

So far we only benchmarked against PyTorch+OpenBLAS. We should also benchmark against PyTorch+Accelerate. Here are a few ways how to do it: * https://towardsdatascience.com/installing-pytorch-on-apple-m1-chip-with-gpu-acceleration-3351dc44d67c * https://twitter.com/isuru_f/status/1636013111592329216 (`conda install "libblas=*=*_accelerate"`)

This approximates the erf() function directly. To use it, apply the following patch: ```diff --- a/gpt2.f90 +++ b/gpt2.f90 @@ -109,7 +109,7 @@ real(sp), intent(in) :: x(:,:), fc_w(:,:), fc_b(:), proj_w(:,:), proj_b(:)...

Batching means that multiple input streams are being computed at the same time, which can vectorize better and thus speedup the inference per token.