Build fails on Ubuntu 24.04 with CUDA 12.6: missing launch.cuh and _Float* type errors

[nurartun]

markdown

System Information

• OS: Ubuntu 24.04.4 LTS
• GCC: 13.3.0 (Ubuntu 13.3.0-6ubuntu2~24.04.1)
• CUDA: 12.6.85
• GPU: NVIDIA GeForce RTX 5060 Ti (Compute Capability 12.0 - Blackwell)

Build Attempts & Errors

1. Standard CMake + Ninja

• Command: cmake -G Ninja .. && ninja
• Error: CMake fails during compiler detection phase with _Float64, _Float128 type errors in glibc headers.

2. Direct nvcc compilation (all .cu and .cpp files)

• Command:

  nvcc -std=c++17 -arch=sm_89 -I src/ src/cuda/launch.cu src/cuda/cuda_kernels.cu src/bindings.cpp src/insertion_points.cpp src/run_mcswell.cpp -o gci -lz -lpthread

Error:

text
src/cuda/launch.cu:22:10: fatal error: cuda/launch.cuh: No such file or directory
CMake with forced C++17 and STDC_WANT_IEC_60559_TYPES_EXT

Command:

bash
cmake -G Ninja
-DCMAKE_CXX_STANDARD=17
-DCMAKE_CUDA_STANDARD=17
-DCMAKE_CXX_FLAGS="-D__STDC_WANT_IEC_60559_TYPES_EXT__"
-DCMAKE_CUDA_FLAGS="-D__STDC_WANT_IEC_60559_TYPES_EXT__ -arch=sm_89"
-DCMAKE_CUDA_ARCHITECTURES="89"
-DCMAKE_CXX_COMPILER_WORKS=ON
-DCMAKE_CUDA_COMPILER_WORKS=ON
..
Error: Still fails with iscanonical and issignaling redeclaration errors due to glibc C23 float types.Full Error Logs (Relevant parts)
text
src/cuda/launch.cu:22:10: fatal error: cuda/launch.cuh: No such file or directory
22 | #include "cuda/launch.cuh"
| ^~~~~~~~~~~~~~~~~
text
/usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h:24:8: error: ‘_Float64’ does not name a type; did you mean ‘_Float16’?
text
/usr/include/x86_64-linux-gnu/bits/iscanonical.h:51:24: error: ‘_Float128’ was not declared in this scope; did you mean ‘__float128’?
text
/usr/include/math.h:1082:24: error: ‘int issignaling’ redeclared as different kind of entity
Request
I am using this tool for a molecular glue design project (Glioblastoma research). I need to run MCSwell to perform hydration mapping on a protein-protein interface.

Could you please:

Provide the missing launch.cuh file (or update the repository with it), and/or

Point me to a precompiled binary for Linux x86_64 with CUDA 12 support, or

Advise on a workaround for the glibc (C23) _Float* compatibility issue on modern Ubuntu systems.

I am happy to test any fixes or provide further logs if needed. Thank you for your time and this great tool

[nbruciaferri]

Hi @nurartun, I will look into this issue this week and hope to get back to you with a fix ASAP.

Thanks for trying MCSwell! :)

[nbruciaferri]

Hi @nurartun, I dug into this and reproduced it.

I installed CUDA 12.6 nvcc (v12.6.20) via NVIDIA's conda channel on a machine that already runs Ubuntu 24.04.4 / glibc 2.39 / GCC 13.3.0 — the same OS/GCC/glibc as your report. Results:

• System nvcc 12.0.140 on PATH: clean configure, no errors.
• Conda nvcc 12.6.20, explicitly pointed to and paired with a forced system host compiler: clean configure, no errors.
• Conda nvcc 12.6.20 resolved "naturally" via PATH (an activated conda env with cuda-nvcc shadowing the system toolkit, exactly how a plain cmake .. behaves once that env is active): reproduces your exact errors, verbatim:

  error: 'int issignaling' redeclared as different kind of entity
  error: '_Float128' was not declared in this scope; did you mean '__float128'?
  

So this is the exact pitfall the README's installation step 4 ("Remove conda CUDA packages") already warns about: OpenMM's conda dependencies pull in their own cuda-nvcc package, and if that shadows the system CUDA Toolkit on PATH (or that step was skipped/didn't fully take), CMake's own compiler-detection picks up the conda-provided nvcc and its bundled headers conflict with glibc 2.39's newer _Float64/_Float128 types on Ubuntu 24.04.

Could you check:

which nvcc
nvcc --version

If nvcc resolves into a conda/mamba/miniconda3/anaconda3 path rather than /usr/bin/nvcc or /usr/local/cuda*/bin/nvcc, that's almost certainly the cause. Try either:

• Following step 4 in the README (conda remove --force cuda-nvcc cuda-cudart cuda-cudart-dev cuda-driver-dev cuda-nvrtc cuda-nvrtc-dev cuda-profiler-api, then unset NVCC_PREPEND_FLAGS CXX CC), or
• Explicitly setting CUDACXX=/usr/bin/nvcc (or wherever your system toolkit lives) before configuring.

Separately, re: the "missing launch.cuh" part of your report — that file is present (include/cuda/launch.cuh); your manual nvcc command used -I src/ only, but the CUDA headers live under include/cuda/, so it needs -I include/ too. Not a repo bug, just the manual command missing an include path.

I've opened #6, which adds a configure-time warning when nvcc resolves into a conda/mamba environment (pointing back at step 4) so this fails with a clear message instead of a cryptic glibc header error, plus broader GPU-architecture support (including forward-compatible support for newer GPUs like your Blackwell RTX 5060 Ti via PTX fallback, and an explicit "CUDA >= 12.0" check with a clear error message).

Let us know if fixing the nvcc resolution gets you unblocked!