Cannot build for cuda 12.8

[atonderski]

Hi,

I am running into a bunch of issues trying to use your method on a rtx 5090, which requires cuda 12.8 for pytorch. Using the dockerfile i was able to isolate it to just changing
-ARG NVIDIA_CUDA_VERSION=12.6.3
+ARG NVIDIA_CUDA_VERSION=12.8.1
and
-ARG CUDA_ARCHITECTURES=native
+ARG CUDA_ARCHITECTURES=120

The build proceeds without errors, but running python reconstruction.py leads to a core dump:

*** Aborted at 1759417086 (unix time) try "date -d @1759417086" if you are using GNU date ***
PC: @                0x0 (unknown)
*** SIGABRT (@0xa) received by PID 10 (TID 0x7eea8671d000) from PID 10; stack trace: ***
    @     0x7eea31095046 (unknown)
    @     0x7eea86760520 (unknown)
    @     0x7eea867b49fc pthread_kill
    @     0x7eea86760476 raise
    @     0x7eea867467f3 abort
    @     0x7eea86221703 (unknown)
    @     0x7eea86237429 __gcc_personality_v0
    @     0x7eea838c2fe9 __libunwind_Unwind_Resume
    @     0x7eea3146cb74 (unknown)
    @     0x7eea316d0728 (unknown)
    @     0x7eea316c42da PyInit__core
    @     0x5c3e1f30797a (unknown)
    @     0x5c3e1f212419 (unknown)
    @     0x5c3e1f202723 _PyEval_EvalFrameDefault
    @     0x5c3e1f2121bc _PyFunction_Vectorcall
    @     0x5c3e1f201763 _PyEval_EvalFrameDefault
    @     0x5c3e1f2121bc _PyFunction_Vectorcall
    @     0x5c3e1f1fc967 _PyEval_EvalFrameDefault
    @     0x5c3e1f2121bc _PyFunction_Vectorcall
    @     0x5c3e1f1fc820 _PyEval_EvalFrameDefault
    @     0x5c3e1f2121bc _PyFunction_Vectorcall
    @     0x5c3e1f1fc820 _PyEval_EvalFrameDefault
    @     0x5c3e1f2121bc _PyFunction_Vectorcall
    @     0x5c3e1f1fc820 _PyEval_EvalFrameDefault
    @     0x5c3e1f2121bc _PyFunction_Vectorcall
    @     0x5c3e1f211619 (unknown)
    @     0x5c3e1f2e652f _PyObject_CallMethodIdObjArgs
    @     0x5c3e1f224f75 PyImport_ImportModuleLevelObject
    @     0x5c3e1f234e3c (unknown)
    @     0x5c3e1f211962 (unknown)
    @     0x5c3e1f21ff4b PyObject_Call
    @     0x5c3e1f202723 _PyEval_EvalFrameDefault
Aborted (core dumped)

And just trying to import pycolmap gives the following error:

>>> import pycolmap
Traceback (most recent call last):
  File "/usr/local/lib/python3.10/dist-packages/pycolmap/__init__.py", line 7, in <module>
    from . import _core
ImportError: generic_type: type "PositiveExponentialManifold" referenced unknown base type "ceres::Manifold"

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.10/dist-packages/pycolmap/__init__.py", line 9, in <module>
    raise RuntimeError(
RuntimeError: 
Cannot import the C++ backend pycolmap._core.
Make sure that you successfully install the package with
  $ python -m pip install pycolmap/

Any idea what could be going on? I understand this configuration is not tested, but I would very much like to get it to work and try out your method.

Thanks in advance :)

[Zador-Pataki]

Hi,

It is true that this is tricky, and I have had my fair share of trouble.

My la
st build on 12.8 I managed with

git clone git@github.com:Zador-Pataki/colmap.git
cd colmap
mkdir build
cd build
cmake .. -GNinja \
  -DCMAKE_CUDA_ARCHITECTURES=native \
  -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.8/bin/nvcc \
  -DCUDAToolkit_ROOT=/usr/local/cuda-12.8 \
  -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-12.8
ninja
sudo ninja install
cd ..
export CUDA_HOME=/usr/local/cuda-12.8
export CUDACXX=$CUDA_HOME/bin/nvcc
export pybind11_DIR=$(python -m pybind11 --cmakedir)

CMAKE_ARGS="-DCMAKE_CUDA_COMPILER=$CUDACXX \
-DCUDAToolkit_ROOT=$CUDA_HOME \
-DCUDA_TOOLKIT_ROOT_DIR=$CUDA_HOME \
-DCMAKE_CUDA_ARCHITECTURES=native \
-Dpybind11_DIR=$pybind11_DIR" \
pip install . --no-build-isolation --no-clean -v

Perhaps you could try it out.