ERGO-Code
Metal Acceleration for cuPDLP-C
FEATURE REQUEST
Would it be possible to get a Metal accelerated version of cuPDLP-C anytime sooner? I believe without creating a direct dependency on High cost Nvidia GPUs, it would give a way to accelerate the algorithms using GPU on Mac devices
We will be writing our own version of PDLP, and will look to enable it to run on Macs
We will be writing our own version of PDLP, and will look to enable it to run on Macs
And this sounds glorious. Is there an ETA? Is this in the pdlp-raphael branch? I observed it is quite fast on some problems.
Work has started on the new PDLP solver but it's not yet part of the repo. We've got external funding to develop it that will grease the wheels and will focus our minds on a delivery date of November. The pdlp-raphael branch is just a sandbox for an intern to develop his own PDLP solver. Any speed in that branch is the same as latest.
Work has started on the new PDLP solver but it's not yet part of the repo. We've got external funding to develop it that will grease the wheels and will focus our minds on a delivery date of November. The
pdlp-raphaelbranch is just a sandbox for an intern to develop his own PDLP solver. Any speed in that branch is the same aslatest.
By the way, how do I build the highspy package so that I can use "latest" in my python code?
I thought we had this documented somewhere, but no...
I've done it many times, and have the following note (for Linux)
- Let . be the root of HiGHS, as checked out from GitHub, and ./build be your build folder
- Decide on your install folder ("install") for HiGHS and set CMAKE_INSTALL_PREFIX to its location by running
cd build; cmake .. -DCMAKE_INSTALL_PREFIX=install
- Install HiGHS by running (from ./build)
make install
- Go back to . and run
sudo pip install . --break-system-packages
- You may need to have run
sudo apt install python3-pip
If you're on Windows I don't know how to do it