diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3433480..c13c2a2 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -16,6 +16,22 @@ if(CMAKE_CUDA_COMPILER)
   set(CMAKE_CUDA_STANDARD 17)
   set(CMAKE_CUDA_STANDARD_REQUIRED ON)
 
+  if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12)
+    message(STATUS "PyTorch NVTX headers workaround: Yes")
+    # only do this if nvToolsExt is not defined and CUDA::nvtx3 exists
+    if(NOT TARGET CUDA::nvToolsExt AND TARGET CUDA::nvtx3)
+      add_library(CUDA::nvToolsExt INTERFACE IMPORTED)
+      # The TORCH_CUDA_USE_NVTX3 compile definition tells PyTorch to use the NVTX3 headers
+      # instead of the legacy NVTX headers. This is necessary when building with CUDA 12+
+      # and the CUDA::nvtx3 target is available, but CUDA::nvToolsExt is not defined.
+      # Without this definition, PyTorch may fail to find the correct NVTX headers.
+      target_compile_definitions(CUDA::nvToolsExt INTERFACE TORCH_CUDA_USE_NVTX3)
+      target_link_libraries(CUDA::nvToolsExt INTERFACE CUDA::nvtx3)
+    endif()
+  else()
+    message(STATUS "PyTorch NVTX headers workaround: No")
+  endif()
+
   message(STATUS "INSTALLING EXTENSIONS WITH CUDA!")
   string(REGEX REPLACE ".[0-9][0-9]|\\." "" CUDA_V ${CMAKE_CUDA_COMPILER_VERSION})
   message(STATUS "CMAKE_CUDA_COMPILER = ${CMAKE_CUDA_COMPILER}")
diff --git a/pyproject.toml b/pyproject.toml
index 4a80e71..11e0694 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -36,7 +36,7 @@ license-files = ["LICENSE"]
 exclude = ["**/.mypy_cache/**", "**/build/**", "**/.vscode/**"]
 
 [build-system]
-requires = ["scikit-build-core>=0.10", "pybind11>=2.10", "cmake", "ninja"]
+requires = ["scikit-build-core>=0.10", "pybind11>=2.10", "cmake", "ninja", "torch"]
 build-backend = "scikit_build_core.build"
 
 [tool.isort]