[RESEARCH] finding libnvvm.so

[rwgk]

Creating this as a place for myself to collect information related to finding libnvvm.so

Note that libnvvm.so is unusual: while most CTK .so files live under <loc>/lib or <loc>/lib64, it lives under <loc>/nvvm/lib64.

[rwgk]

$ ./cuda_12.8.0_570.86.10_linux.run --toolkit --toolkitpath=$HOME/ctk_downloads/scratch

$ cd scratch
$ find . -name '*.so' | grep -v -e '^\./extras/' -e '^\./nsight-' -e '^\./compute-sanitizer/' -e '/eclipse_[0-9]*\.so$'
./nvvm/lib64/libnvvm.so
./targets/x86_64-linux/lib/libnvblas.so
./targets/x86_64-linux/lib/libcusolver.so
./targets/x86_64-linux/lib/libnppim.so
./targets/x86_64-linux/lib/libnpps.so
./targets/x86_64-linux/lib/libcudart.so
./targets/x86_64-linux/lib/libcufft.so
./targets/x86_64-linux/lib/libnvJitLink.so
./targets/x86_64-linux/lib/libcusolverMg.so
./targets/x86_64-linux/lib/libcuinj64.so
./targets/x86_64-linux/lib/libnvfatbin.so
./targets/x86_64-linux/lib/stubs/libcusolver.so
./targets/x86_64-linux/lib/stubs/libnvidia-ml.so
./targets/x86_64-linux/lib/stubs/libnppim.so
./targets/x86_64-linux/lib/stubs/libnpps.so
./targets/x86_64-linux/lib/stubs/libcufft.so
./targets/x86_64-linux/lib/stubs/libnvJitLink.so
./targets/x86_64-linux/lib/stubs/libcusolverMg.so
./targets/x86_64-linux/lib/stubs/libnvfatbin.so
./targets/x86_64-linux/lib/stubs/libnppif.so
./targets/x86_64-linux/lib/stubs/libcublas.so
./targets/x86_64-linux/lib/stubs/libnppial.so
./targets/x86_64-linux/lib/stubs/libcurand.so
./targets/x86_64-linux/lib/stubs/libcublasLt.so
./targets/x86_64-linux/lib/stubs/libnppidei.so
./targets/x86_64-linux/lib/stubs/libnvjpeg.so
./targets/x86_64-linux/lib/stubs/libcuda.so
./targets/x86_64-linux/lib/stubs/libnvrtc.alt.so
./targets/x86_64-linux/lib/stubs/libnppig.so
./targets/x86_64-linux/lib/stubs/libcusparse.so
./targets/x86_64-linux/lib/stubs/libnppisu.so
./targets/x86_64-linux/lib/stubs/libnppicc.so
./targets/x86_64-linux/lib/stubs/libnppitc.so
./targets/x86_64-linux/lib/stubs/libnppist.so
./targets/x86_64-linux/lib/stubs/libcufftw.so
./targets/x86_64-linux/lib/stubs/libnvrtc.so
./targets/x86_64-linux/lib/stubs/libnppc.so
./targets/x86_64-linux/lib/libnppif.so
./targets/x86_64-linux/lib/libcublas.so
./targets/x86_64-linux/lib/libcufile_rdma.so
./targets/x86_64-linux/lib/libcufile.so
./targets/x86_64-linux/lib/libnppial.so
./targets/x86_64-linux/lib/libaccinj64.so
./targets/x86_64-linux/lib/libcurand.so
./targets/x86_64-linux/lib/libcublasLt.so
./targets/x86_64-linux/lib/libnppidei.so
./targets/x86_64-linux/lib/libnvrtc-builtins.alt.so
./targets/x86_64-linux/lib/libnvToolsExt.so
./targets/x86_64-linux/lib/libOpenCL.so
./targets/x86_64-linux/lib/libnvjpeg.so
./targets/x86_64-linux/lib/libnvrtc.alt.so
./targets/x86_64-linux/lib/libnppig.so
./targets/x86_64-linux/lib/libcusparse.so
./targets/x86_64-linux/lib/libnvrtc-builtins.so
./targets/x86_64-linux/lib/libnppisu.so
./targets/x86_64-linux/lib/libnppicc.so
./targets/x86_64-linux/lib/libnppitc.so
./targets/x86_64-linux/lib/libnppist.so
./targets/x86_64-linux/lib/libcufftw.so
./targets/x86_64-linux/lib/libnvrtc.so
./targets/x86_64-linux/lib/libnppc.so

[rwgk]

Inside a cccl [Dev Container: cuda12.8-gcc13 @ ... ]

$ cd /usr/local/cuda
$ find . -name '*.so' | grep -v -e '^\./compute-sanitizer/'
./nvvm/lib64/libnvvm.so
./targets/x86_64-linux/lib/libnvperf_target.so
./targets/x86_64-linux/lib/libcupti.so
./targets/x86_64-linux/lib/libcudart.so
./targets/x86_64-linux/lib/libnvJitLink.so
./targets/x86_64-linux/lib/libcuinj64.so
./targets/x86_64-linux/lib/stubs/libnvidia-ml.so
./targets/x86_64-linux/lib/stubs/libnvJitLink.so
./targets/x86_64-linux/lib/stubs/libcurand.so
./targets/x86_64-linux/lib/stubs/libcuda.so
./targets/x86_64-linux/lib/stubs/libnvrtc.alt.so
./targets/x86_64-linux/lib/stubs/libnvrtc.so
./targets/x86_64-linux/lib/libpcsamplingutil.so
./targets/x86_64-linux/lib/libaccinj64.so
./targets/x86_64-linux/lib/libcurand.so
./targets/x86_64-linux/lib/libnvperf_host.so
./targets/x86_64-linux/lib/libnvrtc-builtins.alt.so
./targets/x86_64-linux/lib/libnvToolsExt.so
./targets/x86_64-linux/lib/libnvrtc.alt.so
./targets/x86_64-linux/lib/libnvrtc-builtins.so
./targets/x86_64-linux/lib/libcheckpoint.so
./targets/x86_64-linux/lib/libnvrtc.so

[rwgk]

Ubuntu 24.04 workstation with CTK 12.6.1 installed into /usr/local

$ cd /usr/local/cuda
$ find . -name '*.so' | grep -v -e '^\./compute-sanitizer/' -e '/eclipse_[0-9]*\.so$'
./nvvm/lib64/libnvvm.so
./targets/x86_64-linux/lib/libnvperf_target.so
./targets/x86_64-linux/lib/libcupti.so
./targets/x86_64-linux/lib/libnvblas.so
./targets/x86_64-linux/lib/libcusolver.so
./targets/x86_64-linux/lib/libnppim.so
./targets/x86_64-linux/lib/libnpps.so
./targets/x86_64-linux/lib/libcudart.so
./targets/x86_64-linux/lib/libcufft.so
./targets/x86_64-linux/lib/libnvJitLink.so
./targets/x86_64-linux/lib/libcusolverMg.so
./targets/x86_64-linux/lib/libcuinj64.so
./targets/x86_64-linux/lib/libnvfatbin.so
./targets/x86_64-linux/lib/stubs/libcusolver.so
./targets/x86_64-linux/lib/stubs/libnvidia-ml.so
./targets/x86_64-linux/lib/stubs/libnppim.so
./targets/x86_64-linux/lib/stubs/libnpps.so
./targets/x86_64-linux/lib/stubs/libcufft.so
./targets/x86_64-linux/lib/stubs/libnvJitLink.so
./targets/x86_64-linux/lib/stubs/libcusolverMg.so
./targets/x86_64-linux/lib/stubs/libnvfatbin.so
./targets/x86_64-linux/lib/stubs/libnppif.so
./targets/x86_64-linux/lib/stubs/libcublas.so
./targets/x86_64-linux/lib/stubs/libnppial.so
./targets/x86_64-linux/lib/stubs/libcurand.so
./targets/x86_64-linux/lib/stubs/libcublasLt.so
./targets/x86_64-linux/lib/stubs/libnppidei.so
./targets/x86_64-linux/lib/stubs/libnvjpeg.so
./targets/x86_64-linux/lib/stubs/libcuda.so
./targets/x86_64-linux/lib/stubs/libnppig.so
./targets/x86_64-linux/lib/stubs/libcusparse.so
./targets/x86_64-linux/lib/stubs/libnppisu.so
./targets/x86_64-linux/lib/stubs/libnppicc.so
./targets/x86_64-linux/lib/stubs/libnppitc.so
./targets/x86_64-linux/lib/stubs/libnppist.so
./targets/x86_64-linux/lib/stubs/libcufftw.so
./targets/x86_64-linux/lib/stubs/libnvrtc.so
./targets/x86_64-linux/lib/stubs/libnppc.so
./targets/x86_64-linux/lib/libnppif.so
./targets/x86_64-linux/lib/libcublas.so
./targets/x86_64-linux/lib/libcufile_rdma.so
./targets/x86_64-linux/lib/libcufile.so
./targets/x86_64-linux/lib/libnppial.so
./targets/x86_64-linux/lib/libpcsamplingutil.so
./targets/x86_64-linux/lib/libaccinj64.so
./targets/x86_64-linux/lib/libcurand.so
./targets/x86_64-linux/lib/libcublasLt.so
./targets/x86_64-linux/lib/libnvperf_host.so
./targets/x86_64-linux/lib/libnppidei.so
./targets/x86_64-linux/lib/libnvToolsExt.so
./targets/x86_64-linux/lib/libOpenCL.so
./targets/x86_64-linux/lib/libnvjpeg.so
./targets/x86_64-linux/lib/libnppig.so
./targets/x86_64-linux/lib/libcusparse.so
./targets/x86_64-linux/lib/libnvrtc-builtins.so
./targets/x86_64-linux/lib/libnppisu.so
./targets/x86_64-linux/lib/libnppicc.so
./targets/x86_64-linux/lib/libnppitc.so
./targets/x86_64-linux/lib/libnppist.so
./targets/x86_64-linux/lib/libcufftw.so
./targets/x86_64-linux/lib/libcheckpoint.so
./targets/x86_64-linux/lib/libnvrtc.so
./targets/x86_64-linux/lib/libnppc.so

[rwgk]

Using a venv (named scratchenv here):

$ pip install cuda-bindings

Collecting cuda-bindings
  Using cached cuda_bindings-12.8.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (13 kB)
Using cached cuda_bindings-12.8.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (11.2 MB)
Installing collected packages: cuda-bindings
Successfully installed cuda-bindings-12.8.0

$ pip install nvidia-cuda-nvcc-cu12

Collecting nvidia-cuda-nvcc-cu12
  Using cached nvidia_cuda_nvcc_cu12-12.8.61-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl.metadata (1.7 kB)
Using cached nvidia_cuda_nvcc_cu12-12.8.61-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl (40.1 MB)
Installing collected packages: nvidia-cuda-nvcc-cu12
Successfully installed nvidia-cuda-nvcc-cu12-12.8.61

$ find scratchenv -name '*.so*'

scratchenv/lib/python3.12/site-packages/cuda/ccudart.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/cudart.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/nvrtc.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/cnvrtc.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/cuda.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/_internal/nvjitlink.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/_internal/utils.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/cydriver.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/nvrtc.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/driver.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/cyruntime.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/nvjitlink.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/cynvjitlink.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/_bindings/cydriver.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/_bindings/cynvrtc.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/_lib/cyruntime/cyruntime.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/_lib/cyruntime/utils.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/_lib/utils.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/cynvrtc.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/bindings/runtime.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/cuda/ccuda.cpython-312-x86_64-linux-gnu.so
scratchenv/lib/python3.12/site-packages/nvidia/cuda_nvcc/nvvm/lib64/libnvvm.so

[rwgk]

When using conda (miniforge), this appears to always exist, after installing cuda-nvcc:

$CONDA_PREFIX/nvvm/lib64/

With these files:

     libnvvm.so -> libnvvm.so.4.0.0
     libnvvm.so.4 -> libnvvm.so.4.0.0
62M  libnvvm.so.4.0.0

I tried:

• conda install cuda-version=12.8 cuda-nvcc into the base environment.

• conda create --yes -n scratchenv python=3.12 cuda-version=12.8 cuda-nvcc

Additionally, $CONDA_PREFIX/lib/python3.12/site-packages appears to always be included in sys.path.

[rwgk]

For the record, I also experimented with this script:

https://github.com/rwgk/stuff/blob/2398710e6be12f65c044b91c8cb105093c720f43/cuda-python/cuda_bindings_nvvm_proc_self_maps.py

Example:

Into a scratchvenv:

pip install cuda_bindings-12.8.0-cp312-cp312-linux_x86_64.whl

Example output WITHOUT setting LD_LIBRARY_PATH:

label: after import cuda
NO DIFF


label: after import cuda.bindings
NO DIFF


label: after import cuda.bindings.nvvm
--- before
+++ after
@@ -1,8 +1,15 @@
+/home/rgrossekunst/scratchenv/lib/python3.12/site-packages/cuda/bindings/_internal/nvvm.cpython-312-x86_64-linux-gnu.so
+/home/rgrossekunst/scratchenv/lib/python3.12/site-packages/cuda/bindings/_internal/utils.cpython-312-x86_64-linux-gnu.so
+/home/rgrossekunst/scratchenv/lib/python3.12/site-packages/cuda/bindings/cynvvm.cpython-312-x86_64-linux-gnu.so
+/home/rgrossekunst/scratchenv/lib/python3.12/site-packages/cuda/bindings/nvvm.cpython-312-x86_64-linux-gnu.so
 /usr/bin/python3.12
 /usr/lib/locale/locale-archive
 /usr/lib/x86_64-linux-gnu/gconv/gconv-modules.cache
 /usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2
 /usr/lib/x86_64-linux-gnu/libc.so.6
 /usr/lib/x86_64-linux-gnu/libexpat.so.1.9.1
+/usr/lib/x86_64-linux-gnu/libgcc_s.so.1
 /usr/lib/x86_64-linux-gnu/libm.so.6
+/usr/lib/x86_64-linux-gnu/libpthread.so.0
+/usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.33
 /usr/lib/x86_64-linux-gnu/libz.so.1.3


label: after cuda.bindings.nvvm.version() failure
--- before
+++ after
@@ -7,9 +7,12 @@
 /usr/lib/x86_64-linux-gnu/gconv/gconv-modules.cache
 /usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2
 /usr/lib/x86_64-linux-gnu/libc.so.6
+/usr/lib/x86_64-linux-gnu/libcuda.so.560.35.03
+/usr/lib/x86_64-linux-gnu/libdl.so.2
 /usr/lib/x86_64-linux-gnu/libexpat.so.1.9.1
 /usr/lib/x86_64-linux-gnu/libgcc_s.so.1
 /usr/lib/x86_64-linux-gnu/libm.so.6
 /usr/lib/x86_64-linux-gnu/libpthread.so.0
+/usr/lib/x86_64-linux-gnu/librt.so.1
 /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.33
 /usr/lib/x86_64-linux-gnu/libz.so.1.3

After export LD_LIBRARY_PATH=/usr/local/cuda/nvvm/lib64 the output changes:

--- /home/rgrossekunst/z/vfailure       2025-02-11 15:53:15.693518916 -0800
+++ /home/rgrossekunst/z/vsuccess       2025-02-11 15:53:05.281172240 -0800
@@ -29,10 +29,10 @@
  /usr/lib/x86_64-linux-gnu/libz.so.1.3
 
 
-label: after cuda.bindings.nvvm.version() failure
+label: after cuda.bindings.nvvm.version() success
 --- before
 +++ after
-@@ -7,9 +7,12 @@
+@@ -7,9 +7,13 @@
  /usr/lib/x86_64-linux-gnu/gconv/gconv-modules.cache
  /usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2
  /usr/lib/x86_64-linux-gnu/libc.so.6
@@ -45,5 +45,6 @@
 +/usr/lib/x86_64-linux-gnu/librt.so.1
  /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.33
  /usr/lib/x86_64-linux-gnu/libz.so.1.3
++/usr/local/cuda-12.6/nvvm/lib64/libnvvm.so.4.0.0

[rwgk]

Possible approaches for locating libnvvm.so:

• Rely on the existing $ORIGIN/../../../nvidia/cuda_nvcc/nvvm/lib64 rpath addition to discover site-packages/nvidia/cuda_nvcc/nvvm/lib64.

• Traverse sys.path, look for lib subdirectories, look for nvvm/lib64 in their parent directories. — This will work for conda-based installations with separate venv. — We wouldn't need the $ORIGIN/../../../../../../nvvm/lib64 rpath addition anymore, although it would be fine to keep. — Note that this is similar to looking for $CONDA_PREFIX/nvvm/lib64, but without requiring that environment variable.

• Look for $CUDA_HOME/nvvm/lib64 or $CUDA_PATH/nvvm/lib64 as the last resort, e.g. to find /usr/local/cuda/nvvm/lib64.

In theory we could look for /usr/local/cuda/nvvm/lib64 as the very last resort, but is that wise?

[kkraus14]

@rwgk could probably also learn from how Numba finds nvvm today as well: https://github.com/NVIDIA/numba-cuda/blob/bf487d78a40eea87f009d636882a5000a7524c95/numba_cuda/numba/cuda/cudadrv/libs.py#L54

[rwgk]

Some ChatGPT queries, for more datapoints:

• https://chatgpt.com/share/67d03b9a-07c4-8008-bad7-5fcc777cf37e

For easy reference, copy-pasting summaries:

General:

• Use cuda-toolkit-config if available (it's direct and robust).
• Fallback to which nvcc or CUDA_HOME for general compatibility.
• For CMake-based projects, rely on FindCUDA.
• Handle Conda-based installations separately.

Python-based:

• numba.cuda.runtime.get_libdevice() – if numba is installed, this directly returns the path to libnvvm.
• cupy.get_cuda_path() – cleanest if you're already using CuPy.
• torch.utils.cpp_extension.CUDA_HOME – if PyTorch is in use.
• Fallback to os.getenv() – safest for broad compatibility.

site-packages/nvidia:

• For direct lookup → use importlib or pkg_resources
• For flexible discovery → use sys.path + glob
• For environment-based lookup → use CUDA_PATH

[kkraus14]

• Use cuda-toolkit-config if available (it's direct and robust).

I assume this is referring to the apt package cuda-toolkit-config-common and isn't an actually usable tool as far as I'm aware. There's too many different OS package managers where interacting directly with them is probably a bad idea.

• Fallback to which nvcc or CUDA_HOME for general compatibility.

CUDA_HOME isn't an officially supported environment variable and is just something that multiple OSS projects have adopted as a "failsafe" for when they can't find things in expected locations or if a user has a non-standard install location. which nvcc will not work in situations where the user doesn't have nvcc installed, i.e. if they're using wheel or conda packages.

• For CMake-based projects, rely on FindCUDA.

Can't rely on having CMake installed and FindCUDA is deprecated where FindCUDAToolkit is the more updated version, but it still doesn't support finding libraries in wheel packages and doesn't work if an nvcc package isn't installed.

• Handle Conda-based installations separately.

This is an option, but we need to handle more than Conda. We need to handle OS package managers putting things in standard locations found by the loader / linker, as well as the situation where users have installed the CUDA toolkit into a manual location and have set things like LD_LIBRARY_PATH and PATH appropriately.

Python-based:

• numba.cuda.runtime.get_libdevice() – if numba is installed, this directly returns the path to libnvvm.

We intend to move numba.cuda to depend on cuda.core for its internal usage of CUDA libraries, so this is a non-starter.

• cupy.get_cuda_path() – cleanest if you're already using CuPy.
• torch.utils.cpp_extension.CUDA_HOME – if PyTorch is in use.

We hope to have cupy and pytorch adopt and depend on cuda.core in the future so these are non-starters.

• Fallback to os.getenv() – safest for broad compatibility.

There isn't any officially supported environment variable to indicate the CUDA toolkit location and different package managers can have different layouts for the various libraries of the CUDA toolkit.

site-packages/nvidia:

• For direct lookup → use importlib or pkg_resources

There isn't really a python package within the current CUDA wheels. In theory there's a site-packages/nvidia folder that you can import as a no-op and then you can check nvidia.__dir__() and check if there's a cuda_nvcc entry in the list (the cuda-nvcc package is what contains the nvvm DSO).

• For flexible discovery → use sys.path + glob

If a user is using wheels I don't think this helps us versus just importing the nvidia namespace package / directory as described above.

• For environment-based lookup → use CUDA_PATH

CUDA_PATH isn't an officially supported environment variable similar to CUDA_HOME.

[rwgk]

Thanks @kkraus14, that's very useful, for me to better understand the environment.

We intend to move numba.cuda to depend on cuda.core for its internal usage of CUDA libraries, so this is a non-starter.

We hope to have cupy and pytorch adopt and depend on cuda.core in the future so these are non-starters.

These are interesting to me, though, and I already looked pretty closely at the numba code, after you pointed me to it a month ago (your comment here from Feb 12, my #447 experiment). I figure, roughly, we need to provide a functional union of what numba.cuda, cupy, and pytorch do for discovering CUDA libraries and headers. I want to spend a small/modest amount of time to also look at the relevant code in cupy and pytorch.