You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
compatibility with CUDA.jl v6 / latest CUDA_Runtime_jll #20
If I try to add CUDA@6 in an environment where CuPDLPx.jl is installed, I get the following resolver error:
ERROR: Unsatisfiable requirements detected for package CUDA_Runtime_jll [76a88914]:
CUDA_Runtime_jll [76a88914] log:
├─possible versions are: 0.1.0 - 0.23.0 or uninstalled
├─restricted by compatibility requirements with CuPDLPx [bcd6524d] to versions: 0.19.2
│ └─CuPDLPx [bcd6524d] log:
│ ├─possible versions are: 0.2.3 - 0.2.7 or uninstalled
│ └─restricted to versions * by project [47468d1f], leaving only versions: 0.2.3 - 0.2.7
│ └─project [47468d1f] log:
│ ├─possible versions are: 0.0.0 or uninstalled
│ └─project [47468d1f] is fixed to version 0.0.0
└─restricted by compatibility requirements with cuFFT [533571aa] to versions: [0.21.0, 0.23.0] — no versions left
└─cuFFT [533571aa] log:
├─possible versions are: 6.0.0 - 6.2.0 or uninstalled
└─required (without additional version restrictions) by CUDA [052768ef]
└─CUDA [052768ef] log:
├─possible versions are: 0.1.0 - 6.2.0 or uninstalled
└─restricted to versions 6 by an explicit requirement, leaving only versions: 6.0.0 - 6.2.0
could the CUDA_Runtime_jll compatibility be updated? Not sure if that requires further changes to this package
Thanks for the report! The pin came from CuPDLPx declaring CUDA_Runtime_jll as a direct dependency even though it's only used transitively via the cuPDLPx_jll binary (which has no version bound on it). I've removed that unused dependency in #21, which lets CUDA_Runtime_jll float to whatever the rest of the environment needs.
Locally, I confirmed that dev-ing the branch and then Pkg.add(name="CUDA", version="6") resolves and installs without the error. Once #21 is merged and v0.2.8 is registered, it'll be available via the registry, but if you'd like to verify sooner, you can test against the branch directly:
Thanks @simeonschaub, and thanks for the afiro.mps repro!
This isn't a bindings issue — the crash is inside libcusparse.so.12 in the native cuPDLPx_jll binary. Since v0.2.8, the CUDA 13 build defaults to the experimental cusparseSpMVOp API, whereas the CUDA 12 build uses the classic cusparseSpMV. That SpMVOp API changed between CUDA 13.1 and 13.3, so when CUDA_Runtime_jll 0.23 loads the 13.3 cuSPARSE, its ABI no longer matches what the binary was built against → segfault in cusparseSpMVOp_bufferSize.
I'll rebuild the CUDA-13 binaries and cut a new release.
In the meantime you can avoid it by either pinning the CUDA-12 runtime:
using CUDA
CUDA.set_runtime_version!(v"12.9") # then restart Julia
or using cuPDLPx_jll v0.2.7, whose CUDA-13 build still uses cusparseSpMV.
Activity
ZedongPeng commentedon Jun 21, 2026
Thanks for the report! The pin came from
CuPDLPxdeclaringCUDA_Runtime_jllas a direct dependency even though it's only used transitively via thecuPDLPx_jllbinary (which has no version bound on it). I've removed that unused dependency in #21, which letsCUDA_Runtime_jllfloat to whatever the rest of the environment needs.Locally, I confirmed that
dev-ing the branch and thenPkg.add(name="CUDA", version="6")resolves and installs without the error. Once #21 is merged and v0.2.8 is registered, it'll be available via the registry, but if you'd like to verify sooner, you can test against the branch directly:Let me know if it works on your end!
ZedongPeng commentedon Jun 23, 2026
v0.2.8has now been registered, so you should be able to install it directly with:pkg> add CuPDLPxsimeonschaub commentedon Jun 24, 2026
Hmm, I am now getting segfaults with CuPDLPx in combination with CUDA_Runtime_jll 0.23. Running the tests outputs the following:
It looks like there might be some incompatibility with cuSPARSE?
I uploaded afiro.mps to https://gist.github.com/simeonschaub/eeef6af648c5d98a9dc0282855585f3a in case you can't reproduce
ZedongPeng commentedon Jul 1, 2026
Thanks @simeonschaub, and thanks for the
afiro.mpsrepro!This isn't a bindings issue — the crash is inside
libcusparse.so.12in the nativecuPDLPx_jllbinary. Since v0.2.8, the CUDA 13 build defaults to the experimentalcusparseSpMVOpAPI, whereas the CUDA 12 build uses the classiccusparseSpMV. That SpMVOp API changed between CUDA 13.1 and 13.3, so whenCUDA_Runtime_jll0.23 loads the 13.3 cuSPARSE, its ABI no longer matches what the binary was built against → segfault incusparseSpMVOp_bufferSize.I'll rebuild the CUDA-13 binaries and cut a new release.
In the meantime you can avoid it by either pinning the CUDA-12 runtime:
or using
cuPDLPx_jllv0.2.7, whose CUDA-13 build still usescusparseSpMV.ZedongPeng commentedon Jul 10, 2026
Hi @simeonschaub . CuPDLPx.jl v0.3.0 has been released. Give it a try and see if it works for you.
simeonschaub commentedon Jul 13, 2026
Can confirm CuPDLPx.jl v0.3.0 works without issues