[Docs] Build from source with Python 3.12 + PyTorch 2.8 + CUDA 12.9 is easy and works on Linux

[JohannesTheo]

📚 The doc issue

Hey there,

to anyone looking for a setup with a more recent PyTorch version, I just wanted to document that the following versions did work for me on Linux. I did not exhaustively test the install but a simple training with mmdet 3 ran just fine for a few iterations :)

# env
conda create --name my_env python=3.12
conda activate my_env
pip install torch==2.8.0+cu129 torchvision==0.23.0+cu129 --index-url https://download.pytorch.org/whl/cu129
conda install -c nvidia cuda==12.9

# mmcv from source: https://mmcv.readthedocs.io/en/latest/get_started/build.html
# gcc (Debian 12.2.0-14+deb12u1) 12.2.0
git clone https://github.com/open-mmlab/mmcv.git
cd mmcv
pip install -r requirements/optional.txt
pip install -e . -v

Hope this helps, will update when properly verified.

Suggest a potential alternative/fix

No response

[jaquilio]

I am trying to do the same on my Windows 64 machine with PyTorch 2.8.0 + CUDA 12.9. But. when I run "pip install -e . -v", I get the following error.

(train) c:\qlabRoot\projects\VtTrain\mmcv>pip install -e . -v
Using pip 25.2 from C:\Users\jaqu\anaconda3\envs\train\Lib\site-packages\pip (python 3.13)
Obtaining file:///C:/qlabRoot/projects/VtTrain/mmcv
  Running command python setup.py egg_info
  C:\qlabRoot\projects\VtTrain\mmcv\setup.py:5: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
    from pkg_resources import DistributionNotFound, get_distribution, parse_version
  Traceback (most recent call last):
    File "<string>", line 2, in <module>
      exec(compile('''
      ~~~~^^^^^^^^^^^^
      # This is <pip-setuptools-caller> -- a caller that pip uses to run setup.py
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
      ...<32 lines>...
      exec(compile(setup_py_code, filename, "exec"))
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
      ''' % ('C:\\qlabRoot\\projects\\VtTrain\\mmcv\\setup.py',), "<pip-setuptools-caller>", "exec"))
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "<pip-setuptools-caller>", line 35, in <module>
    File "C:\qlabRoot\projects\VtTrain\mmcv\setup.py", line 485, in <module>
      version=get_version(),
              ~~~~~~~~~~~^^
    File "C:\qlabRoot\projects\VtTrain\mmcv\setup.py", line 47, in get_version
      return locals()['__version__']
             ~~~~~~~~^^^^^^^^^^^^^^^
  KeyError: '__version__'
  error: subprocess-exited-with-error

[JohannesTheo]

Unfortunately, I have no experience with mmcv on Windows 😅

I used Python 3.12 but that should not be the problem. I suspect something else.

Have you checked the things mentioned here?

https://mmcv.readthedocs.io/en/latest/get_started/build.html#build-on-windows

In particular "6. Set up MSVC compiler"?

They mention that only utf-8 is recognized and you should set your System to English.

Just a wild guess 😄 Hope it helps, good luck!

[darklynx-ctrl]

will it work with mmdetection especially for training mask-rcnn
ImportError: /home/darklynx/miniconda3/envs/mmdet_env/lib/python3.10/site-packages/mmcv/_ext.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c105ErrorC2ENS_14SourceLocationESs
just no idea how to fix that

[JohannesTheo]

Yes, it will work but you have to apply this change in mmdet to install it: #3096 (comment)

However, before you install mmdet, you should check your mmcv installation. Did you run python .dev_scripts/check_installation.py as described here (step 6)? Looks like your compilation did not work correctly.

https://mmcv.readthedocs.io/en/latest/get_started/build.html#build-on-linux

[jaquilio]

After digging further into the KeyError: '__version__' error I reported above, I found out that it is related to the changes in Python 3.13. Specifically, it pertains to PEP 667 changes made in Python 3.13 as discussed here.

In case it helps others running into this issue: Modifying the get_version function in mmcv/setup.py as follows fixed it for me:

def get_version():
    version_file = 'mmcv/version.py'
    d = {}
    with open(version_file, encoding='utf-8') as f:
        exec(compile(f.read(), version_file, 'exec'), globals(), d)
    return d['__version__']

P.S. I wanted to submit a pull request for this change, but, ran into an issue elsewhere unrelated to this.

[ympc08]

Have you try python .dev_scripts/check_installation.py ?

[azztt]

I followed this (for this exact example and with cuda 12.8, python 3.10 as well). To run mmdet3d, should I also be installing mmdet before mmdet3d as per instructions from mmdet3d docs? Because when I do that, mmdet installation fails because it needs mmcv <2.2.0 and same for mmdet3d.

[JohannesTheo]

Hey @azztt, to install (or use) mmdet you have to modify,

• mmdetection/mmdet/__init__.py and manually change
• mmcv_maximum_version = '2.2.0' to mmcv_maximum_version = '2.2.1'

because they have a strict < assert check in place. I have no experience with mmdet3d, but you might have to do something similar there if you run into errors. Hope this helps :)

Unfortunately, all the OpenMMLab projects seem to be abandoned. I don't have much time right now but I'm considering a more general solution to all of these problems in the near future.

[azztt]

Hey, thanks for reply. Does this mean I'll need to build both of them from source as well? If yes, I'll try that and let post the results here.

Thanks

[JohannesTheo]

Assuming you have sucessfully installed mmcv 2.2.0, you should be able to simply:

• git clone https://github.com/open-mmlab/mmdetection.git
• apply the change from above in mmdetection/mmdet/__init__.py
• pip install ./mmdetection or pip install -e ./mmdetection (depending on your need)
• now install mmdet3d

But as I mentioned, I have no experience with mmdet3d and it might well be that you have to fix more things.

[yw155]

I successfully installed the MMENGINE and MMCV from the source. Then I validate the installation of MMCV and meet the following errors:

File "/home/mmlab/mmcv/.dev_scripts/check_installation.py", line 4, in
from mmcv.ops import box_iou_rotated
File "/home/mmlab/mmcv/mmcv/ops/init.py", line 3, in
from .active_rotated_filter import active_rotated_filter
File "/home/mmlab/mmcv/mmcv/ops/active_rotated_filter.py", line 10, in
ext_module = ext_loader.load_ext(
^^^^^^^^^^^^^^^^^^^^
File "/home/mmlab/mmcv/mmcv/utils/ext_loader.py", line 13, in load_ext
ext = importlib.import_module('mmcv.' + name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/anaconda3/envs/mmlab/lib/python3.12/importlib/init.py", line 90, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ModuleNotFoundError: No module named 'mmcv._ext'

My installation environment is: Python 3.12, Pytorch 2.6.0, CUDA 12.4.

What can I do?

[JohannesTheo]

Hm, I have not tested these versions. But you can check if there is a file like _ext.cpython-312-x86_64-linux-gnu.so (might have a different name on you system) in the mmcv/mmcv folder (same level as mmcv/mmcv/utils folder etc.).

If this is not the case, maybe try different versions or carefully re-read the docs I linked above and re-compile mmcv.

Other people have the same problem btw (and a potential solution): #3310

[jaquilio]

Honestly, I get the impression that the original maintainers of OpenMMLab projects aren't active any more and the project setups haven't been updated for a long time and hence keep failing while trying to build with newer versions of PyTorch, Cuda, Python etc. I found the following where people seem to be more interested in maintaining them.
https://github.com/VBTI-development