cusolver native routines from cuda-python #651
-
|
I'd like to access the native cusolver routines from python, specifically potrf. The bindings I can find in the python ecosystem all manage the info parameter on a failed factor and they don't allow the input matrix to clobber in place (PyTorch solves copy the input tensor). Is there a way to access these routines so I can just pass info.data_ptr() to them straight from PyTorch? I'm trying to also avoid a build process with c++ code so if I ship a package I don't need wheels or a slow build process for the user. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
Hi @capybara-club, I believe nvmath-python exposes all the low-level bindings in the |
Beta Was this translation helpful? Give feedback.
-
|
Perfect! Thanks for being so helpful! |
Beta Was this translation helpful? Give feedback.
Hi @capybara-club, I believe nvmath-python exposes all the low-level bindings in the
nvmath.bindingsmodule:https://docs.nvidia.com/cuda/nvmath-python/latest/bindings/generated/nvmath.bindings.cusolverDn.xpotrf.html
Feel free to reach out to the nvmath-python team in the NVIDIA/nvmath-python repo!