Failing to build mmcv from source with cuda 12.1 #2860
Comments
|
I also an getting a similar error in trying to build |
|
Thanks for the notification. This is caused by nightly torch https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/ATen.h Torch has update their c++ compatible since 2.1.0+. We will update the compatible of MMCV ASAP. |
|
Hello! Any updates on this? I cannot install mmcv, hence I cannot use any of the other tools like mmdetection or mmyolo (I want to try out RTMDet). I ran the following: mim install "mmcv>=2.0.0" And I get the following error: C:\anaconda3\envs\openmmlab\lib\site-packages\torch\include\c10/util/C++17.h(27): fatal error C1189: #error: You need C++17 to compile PyTorch I assume this is related to the issue posted by isobelm, hence I'm making a comment here. If you have a solution to this problem that I can try, that'd be great! Thanks for your hard work on this! |
|
Any updates? |
|
Hello! Would appreciate an update on this. Thanks! |
|
I haven’t yet since I thought that’s for building from source. I’m installing mmcv over mim, as directed in the mmdetection repo… I’m trying to use RTMDet. Is there a way to use mmcv from source with the changes you mentioned (without mim install)? Thanks! |
|
I was using mmPose, which also recommended installing via mim, I installed from source to get around the issues I was having installing from mim (probably the same ones you're having). As far as I know, you shouldn't have any issue using RTMDet if you install mmcv from source with the changes described in this thread, although I'm no expert. |
|
Oh awesome, will try that out! This is a very dumb question, but how do you connect mmcv to other mm repos after installing from source…? |
|
There are docs that describe how to build from source, just follow the instructions included and verify the installation and it'll work, there's no extra step to connect them: https://mmcv.readthedocs.io/en/latest/get_started/build.html#build-mmcv |
|
Awesome! Thanks! I’ll give this a shot :)) |
|
I'm trying to build using this Dockerfike: FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04 AS mmcv-builder
ENV FORCE_CUDA 1
ENV MMCV_WITH_OPS 1
RUN apt update && apt install -y python3 python3-pip git
ARG PYTORCH_VERSION=2.1.0.dev20230831+cu121
ARG TORCHVISION_VERSION=v0.16.0.dev20230831
RUN pip install --pre torch==$PYTORCH_VERSION torchvision==$TORCHVISION_VERSION --index-url https://download.pytorch.org/whl/nightly/cu121
RUN pip install openmim
RUN git clone --depth 1 --branch fix-cpp17 https://github.com/grimoire/mmcv
WORKDIR mmcv
RUN mim install .But getting this error: |
|
Found the solution. |
when will the update completed? Very unconvenient for pytorch 2.1.0 users. |
|
What can I do to get past this error on Ubuntu? I'm trying to install ComfyUI Impact Pack on Ubuntu 23, and there's no way to enable MMCV because of this error. I believe it's related? |
Hi @juWuBabaaaa , we will provide the pre-built packages for torch2.1 soon. |
Ended up installing 2.0.0 from source with the instructions in the GiHub, and replaced all "c++14" to "c++17" in setup.py before running install. |
|
I added statements for compiler selection in Cmakelist.txt and successfully compiled them, I hope it will be helpful to you |
Prerequisite
Environment
Reproduces the problem - code sample
N/A
Reproduces the problem - command or script
Reproduces the problem - error message
Additional information
I have been unable to build mmcv from source in my current environment. This may be due to my Cuda version being to recent, but that is difficult to change as I do not have admin powers on the computer I'm using.
Should this be possible? Am I doing something wrong? Any help would be appreciated.
The text was updated successfully, but these errors were encountered: