autozyme 0.3.0

Drop-in accelerators for scientific Python packages

autozyme

Drop-in accelerators for scientific Python packages. Each accelerator is a validated, monkey-patched fast replacement for a hot function in an upstream library (scanpy, scvelo, etc.). Install once; existing scripts run faster automatically. It bundles no datasets; it accelerates your own data in place.

Install

pip install autozyme

autozyme itself has no hard dependencies on upstream packages. A patch activates only when its upstream is importable in your environment.

Use

import autozyme
# autozyme 0.3.0: 16 patches available, 4 subsets — autozyme.activate(<name>) to enable

autozyme.activate("scvelo")
import scvelo as scv
scv.tl.recover_dynamics(adata)   # uses fast version transparently

Supported parameter scope

Each accelerator is validated for a specific parameter envelope. Supported parameters (n_comps, resolution, the data itself, …) can be set freely; a handful of parameters per patch are not on the fast path and fall back to the upstream implementation automatically (correct result, just not accelerated). A few documented approximations are validated only at their stated configuration. Each patch's supported scope and out-of-scope behavior are documented in a SCOPE.md next to its source in the repository, at autozyme_py/src/autozyme/<patch>/SCOPE.md (github.com/ElliotXie/autozyme).

A few patches accelerate version-fragile upstreams that bind one release's internal API (e.g. cell2location, pinned to 0.1.5). Those carry a tested install recipe under "Tested environment" in that SCOPE.md and an opt-in extra, e.g. pip install "autozyme[cell2location]".

Opt-out levels

# per-call (only namespace-fn patches; class-method patches don't expose this)
fn(..., zyme=False)

# per-block (works for ALL patches incl. class-method — hard kill switch)
with autozyme.disabled():
    fn(...)        # uses captured original
    other_fn(...)  # also original

# per-package
autozyme.restore("scvelo")
autozyme.activate("scvelo")

# session-wide
autozyme.restore_all()

# never activate (set before import; `activate()` returns False without binding)
import os; os.environ["AUTOZYME_DISABLED"] = "1"

# suppress per-activation stderr marker (keeps logs clean; banner unaffected)
import os; os.environ["AUTOZYME_QUIET"] = "1"

Multiprocessing workers

activate() binds patches into the upstream namespace of the current process. Fork-mode workers (default on Linux for multiprocessing) inherit the bindings via copy-on-write memory. Spawn-mode workers (default on macOS / Windows, and what loky, joblib, concurrent.futures.ProcessPoolExecutor(mp_context="spawn") use) start fresh — they import autozyme but no patch is active.

For spawn-mode pools, activate inside each worker via the pool's initializer:

from multiprocessing import Pool

def _init():
    import autozyme
    autozyme.activate("cell2location")

with Pool(4, initializer=_init) as pool:
    pool.map(work_fn, items)

joblib.Parallel(backend="loky") uses the same pattern via the initializer= arg on its underlying executor.

Threading

autozyme.set_threads(8)

Status and introspection

python -m autozyme              # one-screen dashboard of patches + upstream availability

autozyme.list_patches()                   # ['cell2location', 'obspy', 'prody', ...]
autozyme.list_patches(installed=True)     # only patches whose upstream is importable
autozyme.status()                         # {'cell2location': 'inactive', 'prody': 'active', ...}
autozyme.inspect("prody")                 # detailed binding info for one patch
autozyme.env_snapshot()                   # structured dict for provenance / Methods capture

Citation

autozyme releases on GitHub get a Zenodo DOI. Cite the version + the patch name(s) you activated. The activation marker logged to stderr is the canonical provenance record, e.g.

[autozyme] activated cell2location -> 2 target(s) in cell2location 0.1.5

In Methods: "We used autozyme v0.3.0 with the cell2location patch (tested against cell2location 0.1.5)."

License

MIT