Open hjarnek opened 3 days ago
Hi @hjarnek The key part of that message is here:
CMake Error at /home/joel/.conda/envs/protax/share/cmake-3.30/Modules/Internal/CMakeCUDAFindToolkit.cmake:104 (message):
Failed to find nvcc.
Compiler requires the CUDA toolkit. Please set the CUDAToolkit_ROOT
variable.
It seems that it is looking for the CUDA toolkit even though you stated that you want to just use the CPU fallback.
Can you confirm that you ran pip install "jax[cpu]"
? If you can provide any more info about your environment setup, or earlier steps before cloning the PROTAX-GPU repo and installing requirements.txt
and pip install .
that would be helpful for us to see what is going on.
Yes, I ran pip install "jax[cpu]"
. If I run it again, I only get Requirement already satisfied
.
I just deleted the repo folder and conda environment and tried to install it from scratch again. The error seems to occur during installation of requirements.txt
, which yields:
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
pyppeteer 2.0.0 requires appdirs<2.0.0,>=1.4.3, which is not installed.
pyppeteer 2.0.0 requires importlib-metadata>=1.4, which is not installed.
Hi,
I think the issue may be on our end. Based on the output trying to search for CUDA toolkit, it looks like the problem is in CMakelists.txt (lines 12-24)
# if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64" AND CMAKE_SYSTEM_NAME MATCHES "Darwin")
if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
message(STATUS "Mac detected. Skipping CUDA configuration.")
set(CUDA_FOUND OFF)
else()
enable_language(CUDA)
find_package(CUDA)
if (CUDA_FOUND)
include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
else()
message(WARNING "CUDA not found or not supported. Skipping GPU ops.")
endif()
endif()
In short, we are calling enable_language(CUDA)
even though it isn't guaranteed that the cuda toolkit has been located. If I recall correctly, when we do this cmake attempts building a simple CUDA program with nvcc to test, which wouldn't be on your system. Can you try changing this to put it inside the CUDA_FOUND
block and see if this works for you? i.e:
# if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64" AND CMAKE_SYSTEM_NAME MATCHES "Darwin")
if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
message(STATUS "Mac detected. Skipping CUDA configuration.")
set(CUDA_FOUND OFF)
else()
find_package(CUDA)
if (CUDA_FOUND)
enable_language(CUDA) # Move to here
include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
else()
message(WARNING "CUDA not found or not supported. Skipping GPU ops.")
endif()
endif()
I tested this change on a colab CPU instance and was able to get it to build.
@Cleanliness – I tried and got the following wall of text error. I guess it has something to do with the pyppeteer error above.
pip install .
error
Hi,
Thanks for updating the usage instructions! I'm facing problems with the installation on Linux (Ubuntu 24.04). For a start, I just want to try PROTAX-CPU. So I have created the conda environment and installed the requirements except for CUDA. But when I run
pip install .
I get the following error. Do you have any idea what's going on?