Skip to content

Compatibility matrix

Task: W14.T01 / W14.T02 · Audit: 2026-07-28 (portable SIMD + portability CI)

OS Arch Python Torch Native kernel Notes
Windows 10/11 x64 3.11–3.13 pinned in constraints.txt MSVC OpenMP DLL; runtime AVX-512→AVX2→scalar Preferred lab path; CI soft compile
Linux x64 3.11–3.13 same GCC .so CI hard gate; same runtime dispatch linux-native + linux-py-matrix
Linux arm64 3.12 same Clang/GCC .so + NEON portability (ubuntu-24.04-arm); err=0
macOS arm64 3.11–3.13 same Clang .so + NEON portability (macos-latest); see MACOS_NOTES.md
macOS x86_64 3.11–3.13 same Clang .so; AVX2/scalar portability (macos-15-intel)
Anything else 3.11+ same NumPy fallback Correctness first; no fake speedups

Requires-python: >=3.11 (pyproject.toml).

Extras: [dev], [hf] (transformers), [all].

CI: Windows (3.12) + Linux native hard (3.12) + Linux Python matrix 3.11/3.12/3.13 + portability (linux-arm64, macos-arm64, macos-x86_64).

Policy: floats need not be bit-identical across machines; same conclusions vs tests/golden_floors.json. Runtime ISA details: 41_PORTABLE_SIMD_KERNEL.md.

Torch upper bound: TORCH_PIN_POLICY.md (W14.T03) — torch>=2.1,<2.13. Optional HF / torchao probes: OPTIONAL_EXTRAS_MATRIX.md (W14.T06).