NVIDIA
MinkowskiEngine does not support RTX 5090 (sm_120 architecture)
MinkowskiEngine does not support RTX 5090 (sm_120 architecture)
Issue Description
I recently attempted to run code using MinkowskiEngine on a system with an RTX 5090 GPU but encountered errors related to CUDA architecture compatibility. The RTX 5090 uses the sm_120 architecture, while the current version of MinkowskiEngine (0.5.4) appears to only support up to sm_90.
Environment Information
- GPU: NVIDIA RTX 5090
- CUDA Version: CUDA 12.8 installed on the system (required for RTX 5090)
- CUDA Environment: Using conda environment configured with CUDA 12.1 packages
- PyTorch Version: 2.5.1
- MinkowskiEngine Version: 0.5.4
- Operating System: Ubuntu
- Python Version: 3.10
Error Message
When attempting to run MinkowskiEngine operations, I encounter the following error:
CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
The detailed CUDA error indicates this is due to the sm_120 architecture not being supported.
Analysis
The RTX 5090 uses NVIDIA's latest Blackwell architecture with compute capability sm_120. After examining MinkowskiEngine's installation process, I found that the CUDA architectures specified during compilation do not include sm_120, preventing it from running on the RTX 5090.
There may also be compatibility issues between the system-level CUDA 12.8 (required for RTX 5090) and the conda environment configured with CUDA 12.1 packages. While CUDA 12.8 is backward compatible, this combination might contribute to the problem.
Request
- Please update MinkowskiEngine to support NVIDIA Blackwell architecture (sm_120)
- Temporary solution: Could you provide specific guidance on how to add sm_120 support when compiling from source?
Attempted Solutions
- Tried running on CPU, but many critical operations (such as furthest_point_sampling in PointNet++) show "CPU not supported" errors
- Tried running with CUDA_LAUNCH_BLOCKING=1, but the same error persists
- Attempted to use different CUDA environments but encountered similar issues
Thank you for your attention and assistance!
Follow the new #621 and I successfully installed it. You need to install the newest PyTorch through
Follow the new #621 and I successfully installed it. You need to install the newest PyTorch through
I tried but failed. Is it suitable for CUDA 12.8?
Follow
same here, I tried the cuda12 installation branch, still show 47 errors compile errors blablbalbal
raise RuntimeError(message) from e RuntimeError: Error compiling objects for extension
Follow the new #621 and I successfully installed it. You need to install the newest PyTorch through
I tried but failed. Is it suitable for CUDA 12.8?
Yes, I use the cuda 12.8. You could try pulling a cuda docker with has all the cuda cudnn setup to avoid problem with drivers.
Follow the new #621 and I successfully installed it. You need to install the newest PyTorch through
I tried but failed. Is it suitable for CUDA 12.8?
Yes, I use the cuda 12.8. You could try pulling a cuda docker with has all the cuda cudnn setup to avoid problem with drivers.
Thank u! I will try docker. May I ask do you also use docker to install it successfully?
Follow the new #621 and I successfully installed it. You need to install the newest PyTorch through
I tried but failed. Is it suitable for CUDA 12.8?
Yes, I use the cuda 12.8. You could try pulling a cuda docker with has all the cuda cudnn setup to avoid problem with drivers.
Thank u! I will try docker. May I ask do you also use docker to install it successfully?
Yes, I successfully install in both docker and the host machine.
Here is the docker build file with the necessary fix
ARG PYTORCH="2.7.0"
ARG CUDA="12.8"
ARG CUDNN="9"
FROM pytorch/pytorch:${PYTORCH}-cuda${CUDA}-cudnn${CUDNN}-devel
ENV TORCH_NVCC_FLAGS="-Xfatbin -compress-all"
ARG CUDA_SM_LIST="6.0 6.1 6.2 7.0 7.2 7.5 8.0 8.6 8.9 9.0 12.0"
# 2) Export it for CMake (>=3.18) and PyTorch builds:
ENV CMAKE_CUDA_ARCHITECTURES=${CUDA_SM_LIST} \
TORCH_CUDA_ARCH_LIST=${CUDA_SM_LIST}
# Install dependencies
RUN apt-get update
RUN apt-get install -y git ninja-build cmake build-essential libopenblas-dev \
xterm xauth openssh-server tmux wget mate-desktop-environment-core
WORKDIR /workspace
RUN git clone https://github.com/CiSong10/MinkowskiEngine.git && \
cd MinkowskiEngine && \
git checkout cuda12-installation
WORKDIR ./MinkowskiEngine
# Apply the fix for error: NVTX3 header collisions (https://github.com/NVIDIA/MinkowskiEngine/issues/614#issuecomment-2886009673)
RUN sed -i '/sources=.*BIND_FILES.*/a\ define_macros=[('"'"'NVTX_DISABLE'"'"', None)],' setup.py
# Apply the fix for error: Compilation: compiler finds both std::to_address and cuda::std::to_address (https://github.com/NVIDIA/MinkowskiEngine/issues/596)
RUN sed -i 's/\bauto __raw = __to_address(__r.get());/auto __raw = std::__to_address(__r.get());/' /usr/include/c++/11/bits/shared_ptr_base.h
RUN python setup.py install --force_cuda --blas=openblas
Hi, I would like to share my way to solve this problem.
First, you must install an old version setuptools(mine is 67.6.1),and you must install(
sudo apt-get install -y git ninja-build cmake build-essential libopenblas-dev).
after that, do
git checkout cuda12-installation
then
sed -i '/sources=.*BIND_FILES.*/a\ define_macros=[('"'"'NVTX_DISABLE'"'"', None)],' setup.py
sudo sed -i 's/\bauto __raw = __to_address(__r.get());/auto __raw = std::__to_address(__r.get());/' /usr/include/c++/11/bits/shared_ptr_base.h
By the way, I face the error:
MinkowskiEngine/setup.py", line 315 define_macros=[('NVTX_DISABLE', None)], ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
what i do is delete one because there are two,
define_macros=[('NVTX_DISABLE', None)]
then
python setup.py install --force_cuda --blas=openblas
and it indicated:Using /home/user/anaconda3/envs/seg/lib/python3.10/site-packages
Finished processing dependencies for MinkowskiEngine==0.5.4
I hope my answer can help all of us who are using fxxking 5090
