Unable to install pytorch bindings - lcuda not found

[akshay-krishnan]

I am trying to build the pytorch bindings, but I am unable to because lcuda was not found on my system.
ld: cannot find -lcuda
I have cuda installed in /usr/local/cuda-11.3 and added /usr/local/cuda-11.3/bin to my $PATH and /usr/local/cuda-11.3/lib64/ to my $LD_LIBRARY_PATH.

I am able to build the library itself with cmake, but no success with the torch bindings.
The g++ command fails since it is unable to find lcuda. I also cannot find a libcuda.so in my /usr/local/cuda-11.3/lib64/. I found one in /usr/local/cuda-11.3/lib64/stubs/, and added this path to my LD_LIBRARY_PATH, but that did not help.

I appreciate any help on this.

[bdfbdf321]

have you solved this problem？

[akshay-krishnan]

Yes, it could have been specific to my environment setup (or to my cuda version 11.3). I had to manually link to the libcuda.so in my conda environment lib.

[bdfbdf321]

@akshay-krishnan How to link libcuda.so ? could you say more details about this solution. I still have this problem.

[ry85]

@akshay-krishnan How did you manually linked libcuda.so?

[akshay-krishnan]

Sorry for the late response folks.

There was a libcuda.so in this path on my system /lib/x86_64-linux-gnu/libcuda.so, but I think that path is not included in LD's library path.
I saw that the linked library path in my command is the /lib folder in my conda environment (miniconda3/envs/<env_name>/lib/libcuda.so). So I linked the libcuda.so from /lib/x86_64-linux-gnu to this folder with the following command:

ln -s /usr/lib/x86_64-linux-gnu/libcuda.so miniconda3/envs/<env_path>/lib/libcuda.so

I don't know if that is the right libcuda.so to be linked, but installation completed and I was able to import the tinycudann module.

[akshay-krishnan]

So I tried replicating this in a new environment, and the fix does not work for me anymore. Its not able to find lcuda even though I have a symlink to libcuda.so in the lib folder that it is searching in.

@Tom94 any help here would be great!

[akshay-krishnan]

I also tried with cuda 11.7, but I face the same problem, so this is probably not version-specific.

[olestole]

I was able to resolve my issue regarding -lcuda which couldn't be found by following @akshay-krishnan's symlink-solution

[yenchenlin]

Add the following line to your .bashrc or .zshrc (depending on your shell) should solve the problem:

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

[ry85]

The answer from @akshay-krishnan worked for me also,

ln -s /usr/lib/x86_64-linux-gnu/libcuda.so anaconda3/envs/nerfstudio/lib/libcuda.so

and not,

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

It seems there is no clear-cut solution to this problem, right now.

[KamiCalcium]

Add the following line to your .bashrc or .zshrc (depending on your shell) should solve the problem:

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

@yenchenlin I'm able to install

pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

After setting the path to /usr/local/cuda/lib64/stubs, thank you!

However, when I tried to run instant-NGP under nerfstudio I have this error:

FAILED: nerfacc_cuda.so 
c++ pybind.cuda.o render_weight.cuda.o pack.cuda.o cdf.cuda.o render_transmittance.cuda.o intersection.cuda.o render_transmittance_cub.cuda.o contraction.cuda.o ray_marching.cuda.o -shared -L/home/user/anaconda3/envs/nerfstudio/lib/python3.8/site-packages/torch/lib -lc10 -lc10_cuda -ltorch_cpu -ltorch_cuda_cu -ltorch_cuda_cpp -ltorch -ltorch_python -L/home/user/anaconda3/envs/nerfstudio/lib64 -lcudart -o nerfacc_cuda.so
/usr/bin/ld: cannot find -lcudart: No such file or directory
collect2: error: ld returned 1 exit status
ninja: build stopped: subcommand failed.

I tried

export LIBRARY_PATH="/usr/local/cuda/lib64:$LIBRARY_PATH"

since I do found libcudart.so under /usr/local/cuda/lib64 but it does not resolve the issue. Any suggestions? Thank you very much!

[Tom94]

@KamiCalcium I'm afraid that's an issue on NerfAcc's side, not tcnn.

That said, if you would like to reproduce instant-ngp's results, I'd warmly encourage using the official implementation as none of the third party re-implementations I've tried match the performance vs. quality tradeoff.

[MilkClouds]

I've solved lcuda not found issue, on WSL2

by executing ln -s /usr/lib/wsl/lib/libcuda.so ~/anaconda3/envs/(your env)/lib/libcuda.so

may be helpful to someone.

[xiechun0318]

I had the same problem installing it in a conda env on WSL. Here is how I solved it.

the compiler tries to find libcuda.so in ~/miniconda3/envs/(envname)/lib, but for some reason CUDA installed it to ~/miniconda3/envs/(envname)/lib/stubs

so I just created a link for libcuda.so outside the /stubs folder by ln -s ~/miniconda3/envs/(envname/lib/stubs/libcuda.so ~/miniconda3/envs/xrnerf/lib/libcuda.so

Then it compiled with no error.

I hope it helps.

[Eric-Ho-Matrix]

Add the following line to your .bashrc or .zshrc (depending on your shell) should solve the problem:

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

Just to record, this method also works for WSL2 (Ubuntu 20.02 LTS)

[whoiszzj]

I meet the same problem and solve it with
ln -s /usr/lib/x86_64-linux-gnu/libcuda.so miniconda3/envs/<env_path>/lib/libcuda.so
Note that the libcuda.so must from the Nvidia Driver Lib, not Cuda Lib
Nvidia Driver Lib:

Cuda lib: <cuda_home>/lib64/stubs:

Good luck!

[lenismerino]

Add the following line to your .bashrc or .zshrc (depending on your shell) should solve the problem:

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

Just to record, this method also works for WSL2 (Ubuntu 20.02 LTS)

This also worked on WSL2 Ubuntu 22.04.02, cuda 11.8 with cudnn, Python 3.10 via pip

[Zazexy]

Add the following line to your .bashrc or .zshrc (depending on your shell) should solve the problem:

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

It works perfectly! thx!

[samsartor]

The LIBRARY_PATH solution did not work for me, but export LDFLAGS=-L/usr/local/cuda/lib64/stub did.

[qgli]

I tried with @Zazexy and @samsartor's method:

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"
export LDFLAGS=-L/usr/local/cuda/lib64/stub

and it works for me, thank you!
specs: RTX4070ti, windows11, cuda11.8, gcc 11 and pytorch 2.0.1

[kwand]

I can confirm the above solution works, though there seems to be a typo (second line should be export LDFLAGS=-L/usr/local/cuda/lib64/stubs, with an 's' for stubs).

On CUDA 11.8, PyTorch 2.0, GCC11 (all within an conda env).

Many thanks, and I hope this is fixed soon.

EDIT: Unfortunately, it is still broken. When I try to run it, I get an undefined symbol: _ZN3c104cuda9SetDeviceEi error.

[jeff999955]

The LIBRARY_PATH solution did not work for me, but export LDFLAGS=-L/usr/local/cuda/lib64/stub did.

This works for me, thanks!

[wy-luke]

I had the same problem installing it in a conda env on WSL. Here is how I solved it.

the compiler tries to find libcuda.so in ~/miniconda3/envs/(envname)/lib, but for some reason CUDA installed it to ~/miniconda3/envs/(envname)/lib/stubs

so I just created a link for libcuda.so outside the /stubs folder by ln -s ~/miniconda3/envs/(envname/lib/stubs/libcuda.so ~/miniconda3/envs/xrnerf/lib/libcuda.so

Then it compiled with no error.

I hope it helps.

Great! It's the problem.

You could also run export LIBRARY_PATH=".../conda/envs/env_name/lib/stubs/:$LIBRARY_PATH".

Replace the path to yours

It works for me. Thanks.

[Dynaomit]

Add the following line to your .bashrc or .zshrc (depending on your shell) should solve the problem:

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

@yenchenlin I'm able to install

pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

After setting the path to /usr/local/cuda/lib64/stubs, thank you!

However, when I tried to run instant-NGP under nerfstudio I have this error:

FAILED: nerfacc_cuda.so 
c++ pybind.cuda.o render_weight.cuda.o pack.cuda.o cdf.cuda.o render_transmittance.cuda.o intersection.cuda.o render_transmittance_cub.cuda.o contraction.cuda.o ray_marching.cuda.o -shared -L/home/user/anaconda3/envs/nerfstudio/lib/python3.8/site-packages/torch/lib -lc10 -lc10_cuda -ltorch_cpu -ltorch_cuda_cu -ltorch_cuda_cpp -ltorch -ltorch_python -L/home/user/anaconda3/envs/nerfstudio/lib64 -lcudart -o nerfacc_cuda.so
/usr/bin/ld: cannot find -lcudart: No such file or directory
collect2: error: ld returned 1 exit status
ninja: build stopped: subcommand failed.

I tried

export LIBRARY_PATH="/usr/local/cuda/lib64:$LIBRARY_PATH"

since I do found libcudart.so under /usr/local/cuda/lib64 but it does not resolve the issue. Any suggestions? Thank you very much!

Did you find a solution to this Nerfacc related problem?

[vule20]

Add the following line to your .bashrc or .zshrc (depending on your shell) should solve the problem:

export LIBRARY_PATH="/usr/local/cuda/lib64/stubs:$LIBRARY_PATH"

Thanks so much. This works and solved my problem.

[gary-young]

The LIBRARY_PATH solution did not work for me, but export LDFLAGS=-L/usr/local/cuda/lib64/stub did.

This works for me, thanks!

LDFLAGS works for me!! Thanks!!

[pengc02]

Sorry for the late response folks.

There was a libcuda.so in this path on my system /lib/x86_64-linux-gnu/libcuda.so, but I think that path is not included in LD's library path. I saw that the linked library path in my command is the /lib folder in my conda environment (miniconda3/envs/<env_name>/lib/libcuda.so). So I linked the libcuda.so from /lib/x86_64-linux-gnu to this folder with the following command:

ln -s /usr/lib/x86_64-linux-gnu/libcuda.so miniconda3/envs/<env_path>/lib/libcuda.so

I don't know if that is the right libcuda.so to be linked, but installation completed and I was able to import the tinycudann module.

this works, and if there's no libcuda.so in /usr/lib/x86_64-linux-gnu/, you can symlink one using libcuda.so.1 or something like that by

ln -s libcuda.so.1 libcuda.so

[Invisiphantom]

export LIBRARY_PATH=$LIBRARY_PATH:/usr/lib/wsl/lib
This works for me.