[BUG] Installation error : nvcc fatal : Unsupported gpu architecture 'compute_50'

[m44guy]

Bug summary

Installation from source following documentation

Failling step : 2.2.1.2. Install the DeePMD-kit’s python interface

Command : "pip install ."

*** Building project with Ninja...
[1/19] Building CUDA object lib/src/gpu/CMakeFiles/deepmd_op_cuda.dir/coord.cu.o
FAILED: [code=1] lib/src/gpu/CMakeFiles/deepmd_op_cuda.dir/coord.cu.o
/gnu/store/i3z3lxkyvakyq3cf3bj72if8fqbx77g8-profile/bin/nvcc ...
-c '/scratch/.../LAMMPS/guix _lammps_22Jul2025_avec_gpu/DeepMD/deepmd-kit/source/lib/src/gpu/coord.cu' -o lib/src/gpu/CMakeFiles/deepmd_op_cuda.dir/coord.cu.o
nvcc fatal : Unsupported gpu architecture 'compute_50'

DeePMD-kit Version

git tag > v3.1.3

Backend and its version

n/a

How did you download the software?

Built from source

Input Files, Running Commands, Error Log, etc.

see log file :
deepmd_error.txt

Steps to Reproduce

error during installation from source files. see log file above.

Further Information, Files, and Links

No response

[njzjz-bot]

Thanks for the report. This looks like a CUDA 13 / architecture-list issue rather than a problem in coord.cu itself.

From the attached log, CMake detects CUDA 13.0.88, then GPU autodetection fails and PyTorch's CMake logic falls back to its “common architectures” list:

-- Automatic GPU detection failed. Building for common architectures.
-- Autodetected CUDA architecture(s): 5.0;8.0;8.6;8.9;9.0;9.0a;10.0;10.0a;10.1a;12.0;12.0a
-- Added CUDA NVCC flags for: ... arch=compute_50,code=sm_50 ...
nvcc fatal   : Unsupported gpu architecture 'compute_50'

sm_50/compute_50 is no longer supported by CUDA 13, so nvcc aborts as soon as that fallback list is used.

Please try overriding the CUDA architecture list to match your actual GPU, for example:

# example for Ampere; adjust to your GPU
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=80"
export TORCH_CUDA_ARCH_LIST="8.0"
pip install .

Use the architecture corresponding to your GPU, e.g. 86 / 8.6 for many RTX 30xx cards, 89 / 8.9 for Ada, 90 / 9.0 for Hopper, etc. Setting both variables is useful here because the log also shows PyTorch warning that it ignores CMAKE_CUDA_ARCHITECTURES and wants TORCH_CUDA_ARCH_LIST.

If you don't need GPU support for this build, another workaround is to disable CUDA/GPU support in the build options.

— OpenClaw 2026.4.22 (model: gpt-5.5)

[m44guy]

Hi,
thanks for the update.

It goes further but there's another error later :
[34/95] Building CUDA object lib/src/gpu/CMakeFiles/deepmd_op_cuda.dir/neighbor_list.cu.o
FAILED: [code=2] lib/src/gpu/CMakeFiles/deepmd_op_cuda.dir/neighbor_list.cu.o
/gnu/store/i3z3lxkyvakyq3cf3bj72if8fqbx77g8-profile/bin/nvcc ...
...
-o lib/src/gpu/CMakeFiles/deepmd_op_cuda.dir/neighbor_list.cu.o
/gnu/store/i3z3lxkyvakyq3cf3bj72if8fqbx77g8-profile/bin/..//include/cccl/cuda/std/__cmath/fpclassify.h(104): error: the global scope has no "fpclassify"
{return ::fpclassify(__x);}
^
/gnu/store/i3z3lxkyvakyq3cf3bj72if8fqbx77g8-profile/bin/..//include/cccl/cuda/std/__cmath/fpclassify.h(117): error: the global scope has no "fpclassify"
{return ::fpclassify(__x);}
^

see attached file for more details.

deepmd_error_80.txt

[njzjz-bot]

Thanks for the updated log. This confirms that the first workaround did take effect:

-- Added CUDA NVCC flags for: -gencode;arch=compute_80,code=sm_80

so the new failure is a different one.

The error now comes from CUDA 13's CCCL/libcu++ headers while compiling files that include CUB, e.g. neighbor_list.cu and prod_env_mat.cu:

include/cccl/cuda/std/__cmath/fpclassify.h: error: the global scope has no "fpclassify"
include/cccl/cuda/std/__cmath/traits.h: error: the global scope has no "isgreater"

This is usually caused by the host C math declarations not being visible/usable to CUDA's <cuda/std/cmath> wrappers, and in your log the environment is quite new/nonstandard:

• CUDA 13.0.88
• GCC/G++ 15.2.0
• TensorFlow 2.21 / PyTorch in the same build
• Guix-provided CUDA/CCCL include paths

So I don't think this is a DeePMD kernel logic error in neighbor_list.cu; it looks more like a CUDA 13 + host compiler / packaging compatibility issue.

Suggested things to try:

1. Use a CUDA 12.x toolkit for now (this is likely the safest path for DeePMD-kit v3.1.x builds).
2. Or try a different supported host compiler via CUDACXX/CUDAHOSTCXX/CMAKE_CUDA_HOST_COMPILER (for example GCC 13/14 if available), and rebuild from a clean build directory.
3. If you must stay on CUDA 13 + Guix, please test whether a minimal CUB/CCCL program including <cub/block/block_scan.cuh> with the same nvcc -ccbin=/bin/g++ -std=c++17 fails similarly. If it does, the issue should probably be reported to the CUDA/Guix packaging side.

One smaller DeePMD-side improvement we may still want is documenting CUDA-13 architecture selection explicitly, because your first error was caused by PyTorch's fallback list still including compute_50, which CUDA 13 no longer accepts.

— OpenClaw 2026.4.22 (model: gpt-5.5)

[m44guy]

Thank you for your reply,
So I tried your suggestions : => downgrade cuda to 12.9 and downgrade gcc to 14.3 (compatibility cuda vs gcc)
I check my compiling

deepmd_error_cuda129_gcc143.txt

env :

$ g++ --version
g++ (GCC) 14.3.0
Copyright (C) 2024 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Tue_May_27_02:21:03_PDT_2025
Cuda compilation tools, release 12.9, V12.9.86
Build cuda_12.9.r12.9/compiler.36037853_0

I restart the installation in a new build directory ; but it still fails elsewhere (but earlier I think)

    -- Set GLIBCXX_USE_CXX_ABI=1
    -- Found CUDA: /gnu/store/kb7bgcvgl5ccx09cy8r9vw29ddx2xclb-profile (found version "12.9")
    CMake Error at /gnu/store/kb7bgcvgl5ccx09cy8r9vw29ddx2xclb-profile/share/cmake-4.1/Modules/CMakeDe    termineCompilerId.cmake:928 (message):
      Compiling the CUDA compiler identification source file
      "CMakeCUDACompilerId.cu" failed.

      Compiler: /gnu/store/kb7bgcvgl5ccx09cy8r9vw29ddx2xclb-profile/bin/nvcc

      Build flags:

      Id flags: --keep;--keep-dir;tmp;-ccbin=/bin/g++ -v

The whole output file is in attachment.