Leo Fang

icon indicating a website link https://leofang.github.io/about icon indicating an email [email protected]

icon company @NVIDIA icon location Greater NYC area icon location Python CUDA tech lead @NVIDIA. Open source contributor on my spare time.

Results 1175 comments of Leo Fang

Make `cupyx.scipy` submodule imports lazy

/test mini

Make `cupyx.scipy` submodule imports lazy

/test mini

Make `cupyx.scipy` submodule imports lazy

@ev-br it seems the CI failures are real issues, could you take a look? 🙂

Can't use `cupy.ndarray`'s `__setitem__` when the `value` is an array-like object that implements `__cuda_array_interface__` or `__cupy_get_ndarray__` but also implements `__eq__`.

https://github.com/cupy/cupy/issues/8352 is related (it's about `__getitem__`).

Cusolver potrs/potrs_batched

Hi @stavoltafunzia currently the non-batched POTRS is also used in `cupyx.lapack.posv()` routine when the matrix A is only 2D: https://github.com/cupy/cupy/blob/1e8ade16f7f0af01e748e706e8c8302ffd2fc873/cupyx/lapack.py#L301-L320 Does it not work?

Cusolver potrs/potrs_batched

I think it makes sense. How about adding a `is_factorized=False` bool flag to `posv` and use it to skip `potrf`?

Cusolver potrs/potrs_batched

> Though, in my opinion, it would be cleaner to have some kind of standalone `potrs` function exposed, then `posv` can be a simple call to `potrf` and `potrs`. It...

Cusolver potrs/potrs_batched

Closed by #9116.

need window support for cupy plz

Hi @xalteropsx could you be more specific? What "window support" are you referring to?

mpi4py>=4 very long build and very high memory consumption

FWIW we also observed Cython 3.1.0 leading to at least 2x slower build time: https://github.com/NVIDIA/cuda-python/issues/642. We haven't had time to dig into why this happens, though.