Forum Discussion
I built a Python runtime that loads precompiled MLIR artifacts from a closed-source compiler
I’ve been building Fluno, a closed-source compiler/runtime experiment for extracting selected hot regions from Python/PyTorch-style continuous inference loops and running them as precompiled native artifacts.
The public repo is not the compiler. It is the audit/runtime surface:
- a Python package ("fluno_runtime") that loads precompiled artifacts
- manifest/schema/hash/expiry validation before dynamic library loading
- a Windows x86_64 live artifact package
- benchmark docs and claim boundaries
- zero-compiler-internals public package structure
The current L-size continuous inference benchmark shows:
- PyTorch optimized repeated: 84.673 ms
- Fluno "hot_vector_repeated": 4.061 ms
- Fluno "hot_run_repeated": 7.245 ms
- max absolute error: 0.0 within the published 11-element "partial_summary_vector" scope
Important limitation: Fluno does not currently beat the handwritten Rust/C++ references on this row. The point of the current public release is not “faster than C++”; it is showing a Python-callable artifact runtime boundary with fail-closed validation and native-class latency.
4 Replies
If your compiler is closed‑source but the runtime is open, someone could extract the MLIR-to‑native part via debugging the runtime’s load behavior. That’s fine—you’re already accepting that by making the loader public.
By releasing the runtime and benchmarks first, you are establishing credibility on the hardest part of the problem: correctness and safety. The 0.0 max absolute error claim is more important than the 20x speedup. It proves the artifact is not hallucinating or drifting numerically.
Thank you for sharing this detailed overview of Fluno and its current capabilities. It sounds like a promising approach for integrating precompiled MLIR artifacts into a Python environment with strong validation and minimal runtime overhead.
Most Python acceleration projects (Numba, TorchScript, Cython) focus on compile-time transformation. You write Python, they turn it into something faster, and you trust that the generated code is correct.
