How to run dpgen model_devi on RTX A6000 (CUDA 11+) when lammps-dp conda packages are CUDA 10 only?

[Collapsar126]

Hi,

I’m trying to run the model_devi stage of dpgen with GPU DeepMD on a
workstation with RTX A6000 (compute 8.6, CUDA 11+) and ran into a
compatibility issue.

• Training uses a conda env with deepmd-kit CUDA 11.6 GPU build and works fine.
• For model_devi, I created another env with
  deepmd-kit + lammps-dp CUDA 10.1 GPU builds from conda.deepmodeling.com.

When I run LAMMPS with pair_style deepmd, TensorFlow/DeepMD prints an error like:

CUDA runtime implicit initialization on GPU:0 failed.
Status: device kernel image is invalid

nvidia-smi shows an RTX A6000, so I suspect the CUDA 10.1 build of lammps-dp
is not compatible with this GPU.

mamba search shows CUDA 11 builds for deepmd-kit, but I can’t find any
CUDA 11 lammps-dp packages on the deepmodeling channel.

Questions:

1. What is the recommended way to run dpgen’s model_devi on RTX A6000?

   • Use CPU-only lammps-dp for model_devi? That will take much more time
   • Or manually compile LAMMPS + DeePMD against CUDA 11?

2. Are there plans to provide CUDA 11+ lammps-dp conda packages for newer GPUs?

Thanks a lot for any suggestions!

[njzjz-bot]

The lammps-dp packages from the old DeepModeling channel are legacy packages and should not be mixed with a newer CUDA DeePMD environment. The official channel itself has been deprecated since DeePMD-kit 3.0.

The current recommended installation is a consistent environment from conda-forge:

conda create -n deepmd deepmd-kit lammps -c conda-forge
conda activate deepmd
lmp -h

Alternatively, current DeePMD-kit provides the LAMMPS extra and official CUDA installers; the current installation documentation is here. Verify that lmp -h lists the DEEPMD package/pair style and that the selected CUDA runtime is compatible with the installed NVIDIA driver.

For an RTX A6000, CPU LAMMPS is only a fallback. A source build is also valid when a cluster requires a particular CUDA/MPI/LAMMPS stack, but it is no longer necessary merely because the obsolete lammps-dp channel lacks CUDA 11 packages. Use the lmp executable from the same modern environment in DP-GEN's model_devi.command.

Coding agent: Codex
Codex version: codex-cli 0.149.0
Model: gpt-5.6-sol
Reasoning effort: xhigh