PyTorch Error: Calling linear solver with sparse tensors requires compiling PyTorch with CUDA cuDSS and is not supported in ROCm build

Question

I'm trying to use torch.sparse.spsolve to solve a linear system of equations as follows:

    A_sparse = torch.sparse_coo_tensor(indices, values, size=(eq_counter, self.num_regions))
    A_sparse_csr = A_sparse.to_sparse_csr()  
    A_sparse_csr = A_sparse_csr.cuda()

    # Create dense vector b
    b = torch.tensor(b_values, dtype=slopes.dtype, device=slopes.device)
    b = b.cuda()


    # Solve the linear system A c = b using torch.sparse.spsolve
    intercepts = torch.sparse.spsolve(A_sparse_csr, b)  # Shape: (num_regions,)

However, I get the following strange error --> RuntimeError: Calling linear solver with sparse tensors requires compiling PyTorch with CUDA cuDSS and is not supported in ROCm build.

I have an Nvidia RTX 40 series graphics card, so I don't understand why the ROCm build (related to AMD GPUs) is even relevant/showing up in the error? I tried to diagnose the error with the following code snippet:

print(f"PyTorch Version: {torch.__version__}")
print(f"CUDA Version: {torch.version.cuda}")
print(f"Is CUDA Available: {torch.cuda.is_available()}")

# Check the current CUDA device
if torch.cuda.is_available():
   print(f"Current CUDA Device: {torch.cuda.current_device()}")
   print(f"Device Name: {torch.cuda.get_device_name(torch.cuda.current_device())}")

The output was as expected:

PyTorch Version: 2.5.1+cu124
CUDA Version: 12.4
Is CUDA Available: True
Current CUDA Device: 0
Device Name: NVIDIA GeForce RTX 4090

So I really don't know what the issue is and any help in resolving this would be much appreciated! Thanks in advance

binbin chan · Accepted Answer · 2024-12-08 11:02:21Z

0

RuntimeError: Calling linear solver with sparse tensors requires compiling PyTorch with CUDA cuDSS and is not supported in ROCm build.

It means to use torch.sparse.spsolve, you have to

compile Pytorch using cuDSS (which is a C++ lib), rather than using binaries of pytorch from pip install torch.
use Nvidia GPU (is not supported in ROCm build of AMD GPU).

You have made the second point as you use RTX 4090, so you only need to install cuDSS and compile pytorch from the source, refer to https://github.com/pytorch/pytorch?tab=readme-ov-file#from-source. Note you should export USE_CUDSS=1 before compile.

Using the compiled pytorch with cuDSS, the following demo codes run OK. demo codes

answered Dec 8, 2024 at 11:02

binbin chan

1

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

PyTorch Error: Calling linear solver with sparse tensors requires compiling PyTorch with CUDA cuDSS and is not supported in ROCm build

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related