a CPU-only version of libdeepmd_c library

[chandagsin]

Dear developers

I would like to compile CP2K with deepmd-kit in a CPU-based system, which requires the libdeepmd_c library. Since the pre-compiled libdeepmd_c was built with CUDA, and thus does not work in a CPU system. Do you have a CPU-only libdeepmd_c, or, how can I build a CPU-only libdeepmd_c from the source code? Many thanks. CDS.

[njzjz]

built with CUDA, and thus does not work in a CPU system.

What error did you encounter? It should fall back to the CPU.

how can I build a CPU-only libdeepmd_c from the source code?

Please refer to this workflow https://github.com/deepmodeling/deepmd-kit/blob/master/.github/workflows/package_c.yml

[chandagsin]

Thanks. Here is the error raised:

DeePMD-kit: Cannot find libcudart.so.11.0
implib-gen: libcudart.so.11.0: failed to load library 'libcudart.so.11.0' via callback 'DP_cudart_dlopen'
implib-gen: libcudart.so.11.0: failed to resolve symbol '__cudaRegisterFatBinary', library failed to load
implib-gen: libcudart.so.11.0: failed to load library 'libcudart.so.11.0' via callback 'DP_cudart_dlopen'
implib-gen: libcudart.so.11.0: failed to resolve symbol '__cudaRegisterFunction', library failed to load
implib-gen: libcudart.so.11.0: failed to load library 'libcudart.so.11.0' via callback 'DP_cudart_dlopen'
implib-gen: libcudart.so.11.0: failed to resolve symbol '__cudaRegisterFatBinaryEnd', library failed to load
DeePMD-kit: Cannot find libcudart.so.11.0
implib-gen: libcudart.so.11.0: failed to load library 'libcudart.so.11.0' via callback 'DP_cudart_dlopen'
implib-gen: libcudart.so.11.0: failed to resolve symbol '__cudaRegisterFatBinary', library failed to load
implib-gen: libcudart.so.11.0: failed to load library 'libcudart.so.11.0' via callback 'DP_cudart_dlopen'
implib-gen: libcudart.so.11.0: failed to resolve symbol '__cudaRegisterFunction', library failed to load
implib-gen: libcudart.so.11.0: failed to load library 'libcudart.so.11.0' via callback 'DP_cudart_dlopen'
implib-gen: libcudart.so.11.0: failed to resolve symbol '__cudaRegisterFatBinaryEnd', library failed to load
DeePMD-kit WARNING: Environmental variable TF_INTRA_OP_PARALLELISM_THREADS is not set. Tune TF_INTRA_OP_PARALLELISM_THREADS for the best performance. See https://deepmd.rtfd.io/parallelism/ for more information.
DeePMD-kit WARNING: Environmental variable TF_INTRA_OP_PARALLELISM_THREADS is not set. Tune TF_INTRA_OP_PARALLELISM_THREADS for the best performance. See https://deepmd.rtfd.io/parallelism/ for more information.
Abort(1) on node 0 (rank 0 in comm 0): application called MPI_Abort(MPI_COMM_WORLD, 1) - process 0
Note: The following floating-point exceptions are signalling: IEEE_INVALID_FLAG IEEE_DIVIDE_BY_ZERO IEEE_OVERFLOW_FLAG IEEE_UNDERFLOW_FLAG IEEE_DENORMAL
STOP 1

I use CP2K 2025.1 with libdeepmd_c 2.2.7. this is the libdeepmd_c library that is by default downloaded and compiled when CP2K is installed.

[njzjz]

It has fallen back to the CPU. The error was not due to this.

[chandagsin]

Many thanks. the issue was that the path of graph.pb was too long, and CP2K truncated the path... :(