Current Approach
Use scikit-build-core and a CMakeLists file in the root directory to compile ggml as a shared library and install it relative to the ggml/ggml.py file so that it can be loaded by ctypes.CDLL.
Issues with Current Approach
#define's are not accessible from the shared library, this is an issue for values that are computed on the host system at compile time.- Platform specific functions such as those for CUDA, OpenCL, and Metal are also awkward to conditionally support with the ctypes approach
- Assuming the relative location of files in an installed package causes issues, especially when installing in ie. editable mode for local development.
- Some arguments such as ctypes Array's require awkward typing workarounds.
Alternative Binding Systems
Requirements
- Should be at least as fast as ctypes approach (large overhead is not acceptable for this application)
- Should be simple to maintain (ggml changes often, should be at least as simple as ctypes to add new function definitions / types / etc)
- Should be easy for others to extend or modify ggml
- Should use ggml's existing build system and flags when compiling, should not limit which platforms / optimizations can be used
Current Approach
Use scikit-build-core and a CMakeLists file in the root directory to compile ggml as a shared library and install it relative to the
ggml/ggml.pyfile so that it can be loaded byctypes.CDLL.Issues with Current Approach
#define's are not accessible from the shared library, this is an issue for values that are computed on the host system at compile time.Alternative Binding Systems
Requirements