search

conda-forge / pytorch-cpu-feedstock

A conda-smithy repository for pytorch-cpu.
BSD 3-Clause "New" or "Revised" License
17 stars 45 forks source link

Make Magma optional for cuda builds? #275

Open zklaus opened 2 days ago

zklaus commented 2 days ago

Comment:

A conda-forge environment with nothing but pytorch for cuda in it currently ways in at 7.2GB, which is perceived to be rather on the heavy side of things.

Looking at potential for slimming things down, libmagma at ~2GB looks like a good candidate.

The pytorch docs seem to suggest that libmagma is used as an alternative for cusolver, which is included anyway, at a much more modest 150 MB.

In the past, magma was significantly faster than cusolver, as demonstrated by [1]. However, a recent 2024 paper by the magma authors [2] shows that cusolver has made progress and is now faster for some of the most important problems. Magma still offers significant performance benefits for certain workloads, but given that pytorch has the ability to switch between the available libraries, we could make magma an optional dependency, i.e. merely include it in run_constrained and leave it up to the user to choose space or performance optimization based on their use-case.

Is this feasible or am I missing something about the use of magma in pytorch?

Do you think this is desirable?

[1] S. Abdelfattah, A. Haidar, S. Tomov and J. Dongarra, "Analysis and Design Techniques towards High-Performance and Energy-Efficient Dense Linear Solvers on GPUs," in IEEE Transactions on Parallel and Distributed Systems, vol. 29, no. 12, pp. 2700-2712, 1 Dec. 2018, doi: 10.1109/TPDS.2018.2842785. keywords: {Graphics processing units;Energy efficiency;Task analysis;Multicore processing;Dense linear solvers;GPU computing;energy efficiency},

[2] Abdelfattah A, Beams N, Carson R, et al. MAGMA: Enabling exascale performance with accelerated BLAS and LAPACK for diverse GPU architectures. The International Journal of High Performance Computing Applications. 2024;38(5):468-490. doi:10.1177/10943420241261960

isuruf commented 2 days ago

we could make magma an optional dependency, i.e. merely include it in run_constrained and leave it up to the user to choose space or performance optimization based on their use-case.

If you build with magma, you need it at runtime even if you don't use magma.