This is the first release of tvm-ffi
Apache TVM FFI is an open ABI and FFI for ML systems. It is a minimal, framework-agnostic, yet flexible open convention with the following systems in mind:
- Kernel libraries: ship one wheel to support multiple frameworks, Python versions, and different languages.
- Kernel DSLs: reusable open ABI for JIT and AOT kernel exposure to PyTorch, JAX, and other ML runtimes.
- ML frameworks and runtimes: unified mechanism to connect libraries and DSLs that adopt the ABI convention.
- Coding agents: unified mechanism to package and ship generated code to production environments.
- ML infrastructure: cross-language support for Python, C++, and Rust, and DSLs.
It has the following technical features:
- DLPack-compatible Tensor data ABI to seamlessly support many frameworks such as PyTorch, JAX, CuPy and others that support DLPack convention.
- Compact value and function calling convention for common data types in machine learning.
- Stable, minimal, and flexible C ABI to support machine learning system use-cases.
- Out-of-the-box multi-language support for Python, C++, Rust, and future path for other languages.
Full Changelog: https://github.com/apache/tvm-ffi/commits/v0.1.0