Install Issue: "no suitable conversion function from "const at::DeprecatedTypeProperties" to "c10::ScalarType" exists"

[DavidRNickel]

I am currently trying to run pip install for this repo, but I am getting a total of 6 errors which read as follows:

difflogic/cuda/difflogic_kernel.cu(283): error: no suitable conversion function from "const at::DeprecatedTypeProperties" to "c10::ScalarType" exists
      [&] { const auto& the_type = x.type(); constexpr const char* at_dispatch_name = "logic_layer_cuda_forward"; at::ScalarType _st = ::detail::scalar_type(the_type); ; switch (_st) { case at::ScalarType::Double: { do { if constexpr (!at::should_include_kernel_dtype( at_dispatch_name, at::ScalarType::Double)) { if (!(false)) { ::c10::detail::torchCheckFail( __func__, "difflogic/cuda/difflogic_kernel.cu", static_cast<uint32_t>(283), (::c10::detail::torchCheckMsgImpl( "Expected " "false" " to be true, but got false.  " "(Could this error message be improved?  If so, " "please report an enhancement request to PyTorch.)", "dtype '", toString(at::ScalarType::Double), "' not selected for kernel tag ", at_dispatch_name))); }; } } while (0); using scalar_t [[maybe_unused]] = c10::impl::ScalarTypeToCPPTypeT<at::ScalarType::Double>; return ([&] { logic_layer_cuda_forward_kernel<scalar_t><<<blocks_per_grid, threads_per_block>>>( x.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>(), a.packed_accessor64<int64_t, 1, torch::RestrictPtrTraits>(), b.packed_accessor64<int64_t, 1, torch::RestrictPtrTraits>(), w.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>(), y.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>() ); })(); } case at::ScalarType::Float: { do { if constexpr (!at::should_include_kernel_dtype( at_dispatch_name, at::ScalarType::Float)) { if (!(false)) { ::c10::detail::torchCheckFail( __func__, "difflogic/cuda/difflogic_kernel.cu", static_cast<uint32_t>(283), (::c10::detail::torchCheckMsgImpl( "Expected " "false" " to be true, but got false.  " "(Could this error message be improved?  If so, " "please report an enhancement request to PyTorch.)", "dtype '", toString(at::ScalarType::Float), "' not selected for kernel tag ", at_dispatch_name))); }; } } while (0); using scalar_t [[maybe_unused]] = c10::impl::ScalarTypeToCPPTypeT<at::ScalarType::Float>; return ([&] { logic_layer_cuda_forward_kernel<scalar_t><<<blocks_per_grid, threads_per_block>>>( x.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>(), a.packed_accessor64<int64_t, 1, torch::RestrictPtrTraits>(), b.packed_accessor64<int64_t, 1, torch::RestrictPtrTraits>(), w.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>(), y.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>() ); })(); } case at::ScalarType::Half: { do { if constexpr (!at::should_include_kernel_dtype( at_dispatch_name, at::ScalarType::Half)) { if (!(false)) { ::c10::detail::torchCheckFail( __func__, "difflogic/cuda/difflogic_kernel.cu", static_cast<uint32_t>(283), (::c10::detail::torchCheckMsgImpl( "Expected " "false" " to be true, but got false.  " "(Could this error message be improved?  If so, " "please report an enhancement request to PyTorch.)", "dtype '", toString(at::ScalarType::Half), "' not selected for kernel tag ", at_dispatch_name))); }; } } while (0); using scalar_t [[maybe_unused]] = c10::impl::ScalarTypeToCPPTypeT<at::ScalarType::Half>; return ([&] { logic_layer_cuda_forward_kernel<scalar_t><<<blocks_per_grid, threads_per_block>>>( x.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>(), a.packed_accessor64<int64_t, 1, torch::RestrictPtrTraits>(), b.packed_accessor64<int64_t, 1, torch::RestrictPtrTraits>(), w.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>(), y.packed_accessor64<scalar_t, 2, torch::RestrictPtrTraits>() ); })(); } default: if (!(false)) { ::c10::detail::torchCheckFail( __func__, "difflogic/cuda/difflogic_kernel.cu", static_cast<uint32_t>(283), (::c10::detail::torchCheckMsgImpl( "Expected " "false" " to be true, but got false.  " "(Could this error message be improved?  If so, " "please report an enhancement request to PyTorch.)", '"', at_dispatch_name, "\" not implemented for '", toString(_st), "'"))); }; } }()

I'm wondering what a potential solution to this might be? I am unfamiliar with CPP/CUDA code, so my ability to read the stack trace here is quite limited. Is this an issue with version-control, or something upstream at the repository?

I have encountered these same errors on two different machines (one Ubuntu WSL2, the other Rocky Linux).

Machine specs:

• WSL: Cuda 12.9, Python 3.13.3, Torch 2.7.0+cu128
• Rocky: Cuda 12.7, Python 3.13.3, Torch 2.6.0+cu126

On the Rocky machine, I have to run python setup.py build because of an issue with pip not finding torch on Line 2 of setup.py. The output is roughly the same as the WSL machine running pip install difflogic or pip install -e . There are also some warnings thrown.

Thanks for any help you can afford me.

[Z-KN]

Hello, did you solve the problem?

[inikishev]

I have the same issue, except I am using python 3.12, the error being this: https://gist.github.com/inikishev/5ef8dcd15c91f5147759bfb5dcbe0d6e

[sayhanne]

I have the same issue with
``

Pytorch

torch                      2.9.0+cu130      pypi_0           pypi
torchvision                0.24.0+cu130     pypi_0           pypi

CUDA driver

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Wed_Aug_20_01:58:59_PM_PDT_2025
Cuda compilation tools, release 13.0, V13.0.88
Build cuda_13.0.r13.0/compiler.36424714_0

Cuda home path

/usr/local/cuda-13.0

I already tried with other versions of torch and cuda but the error persists.

[sayhanne]

After couple of trial and error, I have found a solution that works for me with newer version of torch (2.9.0+cu130) and NVIDIA CUDA Driver 13.0. Steps are below:

1. Instead of installing via pip, pull the difflogic from GitHub repo.
2. Modify the difflogic_kernel.cu under ./difflogic/cuda as follows (remove the deprecated function calls):

use x.is_cuda() instead of x.type().is_cuda() in line 13

#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor") // old
#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor") // new

replace ALL remaining .type() calls with .scalar_type()
e.g.,

AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.type(), ...  // old 
AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), ... // new

3. Then python setup.py sdist bdist_whell
4. If you want to install it like a pip package, then install it via .whl under the dist folder:
   pip install dist/*.whl

Now you can use difflogic package anywhere in your environment.