Build Highs - for Pyhton highspy with hipo/blas

[Tooblippe]

Hi,

i was able to build Highs 1.3.1 from sources for:
a) hipo/blas and also
2) cuda and both worked.

I also want to compile highspy to run with pypsa but it gives error ERRORHIPO is only available in C++. Not yet supported in Python. - is it possible to build highspy yet for the new implementation.

I ran `pip install .' against the pyproject.toml with the Blas and Hipo Cmake flags set. It runs to almost completion and then the error - Process failed with very little further info.

thanks
Tobie

[galabovaa]

Hi Tobie,

We are currently working on HiGHS with HiPO in Python. There is some more work to do, as we want to keep HiPO optional. Please note, that using HIPO changes the license from MIT to Apache.

For the moment, you can use the developemnt version from conda-forge with

 conda install -c conda-forge/label/dev highspy

Or you could build it locally, if you check out branch py-l2 and run

 python -m pip install .  --config-settings=cmake.define.HIPO=ON  --config-settings=cmake.define.BUILD_OPENBLAS=ON

These build steps will change with our ongoing re-structure of the dependency management

[Tooblippe]

Thank you. @galabovaa - if you can point me in the right direction please. before i build it i just want to find out if the funciionality has not been baked in already.

[Tooblippe]

Just some feedback. I managed to build the version above for hipo/openblas and also after a lot of patching, drivers and binary path issues got the CUDA support compiled in on Windows using an RTX 5070.

I made detailed notes than might be helpful, or maybe just a note to myself so I can re-create on my PC.

• notes - https://gist.github.com/Tooblippe/c574810333ae5cbb5722f080e9046faf
• ChatGPT - discusssion - https://chatgpt.com/share/6a2a55d5-84cc-832e-a44c-0d6c012e72f2
• Summary of all mods/patches - https://gist.github.com/Tooblippe/b3692abbf8da1765a86f353374fe55d3

GPU workload

PYPSA solver output indicating Cuda and use of GPU

[jajhall]

Congratulations on building v1.13.1 in this way. However, it will be much easier once v1.14 is released later this month.

I see that you're using one of our PDLP solvers (cuPDLP-C) and it took a very long time. Indeed, the iteration speed is astonishingly slow.

Was it not possible to solve the problem with HiPO?

[Tooblippe]

It is a very very utility scale Pypsa model, it even gives cplex a hard time. I will be running HiPO a bit lalter.

I think our team will have to focus more on understanding the Linopy model, and what can be optimised in our network to improve performance, ensuring scaling and other issues that might be latently present in the model that increases solver time but does not increase accuracy if that makes sende. will give feedback on HiPO

[jajhall]

Did the PDLP solver get a solution eventually?

If Cplex struggled and this is not a large model for you, then you do need to reconsider what you think is reasonable to model, since there's no point in generating models that are prohibitively expensive to solve.

[galabovaa]

It would be great if you could try to build the "latest" branch as well:

To install locally without an existing OpenBLAS installation, run

python -m pip install ./highspy-extras
python -m pip install .

To install locally with an existing OpenBLAS installation, run

python -m pip install ./highspy-extras --config-settings=cmake.define.BUILD_OPENBLAS=OFF
python -m pip install .

If the OpenBLAS installation path is not in the default set, it could be provided with

    --config-settings=cmake.define.BLAS_LIBRARIES=/path/to/openblas/library

HiPO is relatively well-tested, CUDA we have yet to test and update!

[Tooblippe]

HI, Thank you. I will try to build as suggested and also see if I can compile CUDA support in. For us, CUDA solved the issue in half the time (600s) compared to Hipo (1200s) and Cplex (85 seconds). regards Tobie

…

On Fri, Jun 12, 2026 at 4:52 PM galabovaa ***@***.***> wrote: It would be great if you could try to build the "latest" branch as well: To install locally without an existing OpenBLAS installation, run python -m pip install ./highspy-extras python -m pip install . To install locally with an existing OpenBLAS installation, run python -m pip install ./highspy-extras --config-settings=cmake.define.BUILD_OPENBLAS=OFF python -m pip install . If the OpenBLAS installation path is not in the default set, it could be provided with --config-settings=cmake.define.BLAS_LIBRARIES=/path/to/openblas/library HiPO is relatively well-tested, CUDA we have yet to test and update! — Reply to this email directly, view it on GitHub <#2928?email_source=notifications&email_token=AAGETQMCV3WX5ZJJ5JDEV5D47QKMVA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZSHAYTGNJVUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17281355>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGETQPQVSTMZB7GMR5ACCL47QKMVAVCNFSNUABHKJSXA33TNF2G64TZHMYTENRZHE2DONJYHNCGS43DOVZXG2LPNY5TSNRZHE4DKNVBOYBA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

-- *Tobie Nortje* 079 897 9035

[galabovaa]

Cuda should now work locally: #3071