Nvcc missing when installing cudatoolkit?

Question

I have installed cuda along pytorch with

conda install pytorch torchvision cudatoolkit=10.0 -c pytorch

However, it seems like nvcc was not installed along with it. If I want to use for example nvcc -V, I get the error that nvcc was not found, and that I should install it with sudo apt install nvidia-cuda-toolkit. Can I do this (I dont want to just try and then find out that it is not working/messes up the whole cuda setup). And is this a bug or expected behavior?

I am using Ubuntu 18.04 and have cuda 10.2

Solution 1

Met this question when installing cudatoolkit of 10.1 with PyTorch 1.4.

There is a conda-forge package https://anaconda.org/conda-forge/cudatoolkit-dev. After installing this, nvcc as well as other CUDA libraries will be then available at /home/li/anaconda3/envs/<env_name>/pkgs/cuda-toolkit in bin/ and lib/.

Solution 2

You can try

conda install -c conda-forge nvcc_linux-64

Currently this should get you 10.2. Nvidia has its own channel nvidia but the latest version there is 10.1.

Solution 3

cudatoolkit that is installed with pytorch is runtime only and does not come with the development compiler nvcc. To get nvcc you need to install cudatoolkit-dev which I believe is available from the conda-forge channel.

Solution 4

cudatoolkit-dev package from conda-forge did not work for me. I used the package from HCC with the latest pytorch (v1.9.0) https://anaconda.org/HCC/cudatoolkit.

The command to install cudatoolkit alongside pytorch and torchvision:

conda install pytorch torchvision cudatoolkit=10.2 -c pytorch -c hcc

After the installation, you can check

$ nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Wed_Oct_23_19:24:38_PDT_2019
Cuda compilation tools, release 10.2, V10.2.89

View more solutions

Comments

[Robert Crovella]

The cudatoolkit installed using conda install is not the same as the CUDA toolkit packaged up by NVIDIA. It is a subset, to provide the needed components for other packages installed by conda such as pytorch. It's likely that it is all you need if you only need to use pytorch. If you need/want the full CUDA toolkit for some other reason, you can install it using a variety of methods including some variant of the one you indicate, but it will install a separate copy, in another location from what is typically used by conda packages. This all looks like expected behavior to me.

[Robert Crovella]

NVIDIA publishes a CUDA linux install guide which may be useful. It doesn't have conda packages or methods in view, however.

[Robert Crovella]

You haven't mentioned if you have a GPU, or what GPU you have, or if you installed a GPU driver, or what GPU driver you installed, but that is all important. The simplest instruction for compatibility is to install the latest driver for your GPU, if you've not already done so. Installing a CUDA toolkit from NVIDIA may install a proper/sufficient driver for you, depending on what exactly you install. Generally, conda install ... does not install a GPU driver, in my experience.

[Luca Thiede]

An Faster-RCNN implementation I want to use needs nvcc. I do have a gpu, and Pytorch runs flawless otherwise. So the safest bet is to uninistall the conda cudatoolkit and install cuda manually? Do you know of any good installlation script that automates the installation?

[Robert Crovella]

I think you will discover that it is harder to get your conda install of pytorch to use a CUDA toolkit other than the one installed by conda. I don't know what the safest bet is; I regularly use a machine that has the cuda toolkit installed by conda and a separate install that I did using the instructions I already provided. That may or may not work for your use case. Versions do matter. Requests for tutorials, scripts, off-site resources are generally off-topic for stack overflow, from what I have seen.

[Bo Li]

Followed this answer but found out that this seems not solving the problems. The nvcc installed by nvcc_linux-64 is merely a shell script calling $(CUDA_HOME)/bin/nvcc

[tejasvi88]

Unfortunately, it currently has 10.1 and I need 10.2

[Bo Li]

@tejasvi88 try requesting this by github.com/conda-forge/cudatoolkit-dev-feedstock

[stason]

as it say in the description of the package: "A meta-package to enable the right nvcc." not really nvcc

[Wok]

Thank you so much! I was going crazy!

[Wok]

People, listen to the guys above my comment: this does NOT install nvcc! It is like a shortcut to nowhere! You will waste time figuring out what is wrong when you don't have nvcc in the first place!

[inverted_index]

Strange why it doesn't have version 10.2 :/

[Sterling]

Looks like this is also only for Linux

[Fazil]

@Sterling I haven't tested the commands on a Windows machine. So, probably it is only for Linux.

[Fazil]

Just checked cudatoolkit package on HCC repo. HCC provides cudatoolkit package only for Linux.

[Rafay Khan]

Gets stuck on "Executing transaction" and then eventually fails :/

[Rafay Khan]

Actually, worked for me. Thanks!

[Juergen]

From the Nvidia website ( link ), I see that CUDA packages can be installed selectively. Is it sufficient to only install nvcc?

[Deep Patel]

I haven't tried with selective install but seems okay since the nvcc installation will check for any dependencies beforehand