[BUG]: `libmathdx.so` not found, even with `nvidia-libmathdx-cu13` installed.

[murdock-grewar]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Runtime Error

Component

device

Describe the bug

libmathdx.so cannot be located.

How to Reproduce

Attempting to run the example examples/device/cufftdx_convolution_performance.py fails.

Command:
uv run --with nvidia-libmathdx-cu13 --with ./nvmath-python[cu13-dx] --with cupy python ./nvmath-python/examples/device/cufftdx_convolution_performance.py

An error is produced:
cuda.pathfinder._dynamic_libs.load_dl_common.DynamicLibNotFoundError: Failure finding "libmathdx.so": No such file: libmathdx.so*, No such file: libmathdx.so*, No such file: libmathdx.so*

Expected behavior

cuda.pathfinder should find libmathdx.so.

Operating System

No response

nvidia-smi output

No response

[carterbox]

I assume you are using Linux on x86 because you have not stated otherwise. I am not able to reproduce the issue with the provided command. In fact, the command fails earlier (during install) when uv says it cannot install both nvmath-python[cu13] and nvmath-python[cu12] simultaneously.

I am able to run the example if I create an environment and install the package in the conventional way as described in our documentation.

uv venv
source .venv/bin/activate
uv pip install nvmath-python[cu13-dx] cupy-cuda13x
python ./examples/device/cufftdx_convolution_performance.py

I suggest that there is a bug in uv run or in the way that you are using uv run. However, I do not know enough about uv to debug.

[murdock-grewar]

I assume you are using Linux on x86 because you have not stated otherwise.

Yes, Linux x86_64.

I am not able to reproduce the issue with the provided command. In fact, the command fails earlier (during install) when uv says it cannot install both nvmath-python[cu13] and nvmath-python[cu12] simultaneously.

Yes, I had the same issue when my working directory was set to the root of this project. I thought to report it, but then I did cd .. and this seemed to resolve the issue. I figured I might have just been doing something wrong.

I am able to run the example if I create an environment and install the package in the conventional way as described in our documentation.

I think the conventional way to set up the environment for a package like this is uv sync. In this case, I tried to use uv sync --only-group cu13-dx but I encountered the error you mentioned above.

uv venv
source .venv/bin/activate
uv pip install nvmath-python[cu13-dx] cupy-cuda13x
python ./examples/device/cufftdx_convolution_performance.py

Won't this grab nvmath-python[cu13-dx] from pypi or some other public repository? I want to install the package from source. I used .[cu13-dx] instead and that worked.

I suggest that there is a bug in uv run or in the way that you are using uv run. However, I do not know enough about uv to debug.

uv sync also fails. I guess the maintainers don't use uv, or they follow some nonstandard process for setting up the environment, like the one you described.

[carterbox]

We don't support uv specific commands: add, sync, lock, tool, etc. because having a single lock file for this project is not currently possible. We only support the pip-like installers.

[carterbox]

I'm going to close this issue because it is not a bug in nvmath-python.