Nvrtc error with CUDA

1

I am using PyTorch 2.5.1 (CUDA12.4) with Python 3.10.15 on Windows 11. I am performing operations on complex numbers as follows:

a = torch.randn(2, 10, 5, 6)  # dtype -> torch.float32
c = (1j * a).exp()  # expected dtype -> torch.complex64

I am getting the following error:

nvrtc: error: failed to open nvrtc-builtins64_124.dll.
Make sure that nvrtc-builtins64_124.dll is installed correctly.

Executing only 1j * a does not raise an error. The error occurs in the exponentiation. The complete error message is very long to paste here, but the first few lines are:

RuntimeError: 
  #ifdef __HIPCC__
  #define ERROR_UNSUPPORTED_CAST ;
  // corresponds to aten/src/ATen/native/cuda/thread_constants.h
  #define CUDA_OR_ROCM_NUM_THREADS 256
  // corresponds to aten/src/ATen/cuda/detail/OffsetCalculator.cuh
  #define MAX_DIMS 16
  #ifndef __forceinline__
  #define __forceinline__ inline __attribute__((always_inline))
  #endif
  #else

It mentions something about unsupported cast error, but I am not sure why this occurs and how to fix it. I have tried changing the Python versions (3.11, 3.12) and PyTorch versions (2.5.1cu121, 2.5.1cu118, 2.4.0cu124) but nothing seems to work. I have the CudaToolkit and cuDNN installed from NVIDIA’s website. Thanks in advance for your help.

2

As an update, this happens for abs() and log() operations, but not for sum(). It is not clear why some of the operations are not executed on CUDA. I need to resolve this to run the rest of the code on GPU. I would appreciate any insights on this.

Thanks in advance for your help.

3

The error is pointing to a missing nvrtc library in your build, which is used for runtime compilation of some kernels. I don’t have a Windows system to reproduce the issue and cannot reproduce it on Linux.
Ad a workaround you could try to install nvrtc manually.
CC @malfet in case you have seen this issue before.