NeuronScope is pure Python with a small optional native layer. It runs on any environment that satisfies the requirements below. Anything outside the matrix is best-effort — no CI coverage, no release blocker.
Python
- CPython 3.10, 3.11, 3.12, 3.13. Older interpreters are refused at import time.
- PyPy is unsupported — the tensor path relies on CPython-specific bindings in the framework layer.
- Free-threaded 3.13 (PEP 703) is exercised in CI but marked experimental.
Operating systems
- Linux: glibc 2.28+ (Debian 10, RHEL 8, Ubuntu 20.04, and newer).
- macOS: 12.0 (Monterey) and newer, on both Intel and Apple silicon.
- Windows: Windows 10 21H2+ and Windows 11. WSL 2 is supported through the Linux path.
Accelerators
NeuronScope does not ship its own CUDA or ROCm wheels — it observes whichever framework you already have installed. The versions below are the ones covered by CI.
- NVIDIA: CUDA 12.1 and 12.4, driver 535+.
- AMD: ROCm 6.0+.
- Apple: Metal via
torch.mpsand JAX Metal. - CPU-only is a first-class configuration and is used for CI on every PR.
Memory and disk
The recorder streams observations to disk, so a recording's memory usage is bounded by the largest single tensor rather than by the artifact size. A working machine needs:
- ≥ 4 GB RAM free at record time for a medium-size model (BERT-base, GPT-2 small).
- ≥ 2 GB free disk for the artifact and its temporary spool.