wolfram-boost

A transparent turbocharger for wolframclient. One import, zero API changes.

The official wolframclient gets sluggish with anything beyond tiny data—an 8 MB image takes 1.5 seconds, a short audio clip takes seconds. wolfram-boost swaps out the transport layer so the same data arrives in tens of milliseconds.

No API changes. No new concepts. Just a smoother experience.

The difference

Benchmarks: ARM64 Linux, Wolfram Engine 14.3, 20-round median.

Installation

pip install wolfram-boost

Requirements:

· Python ≥ 3.8 · NumPy ≥ 1.20 · wolframclient · Wolfram Engine or Mathematica installed and configured · Linux only (relies on /dev/shm)

Quick start

# Before
from wolframclient.evaluation import WolframLanguageSession

# After
from wolfram_boost import SHMWolframSession

session = SHMWolframSession(
    kernel="/opt/Wolfram/WolframEngine/14.3/Executables/WolframKernel"
)

# Everything else is the same
from wolframclient.language import wl
import numpy as np

img = np.random.rand(4096, 4096, 3).astype(np.float32)
result = session.evaluate(wl.ImageMeasurements(img, "MeanIntensity"))

audio = np.random.randint(-32768, 32767, size=(2, 44100), dtype=np.int16)
result = session.evaluate(wl.Mean(wl.Flatten(audio)))

small = np.random.rand(100, 2)
result = session.evaluate(wl.Mean(wl.Flatten(small)))

session.terminate()

How it works

You don't need to know. But if you're curious:

Your data
    │
    ├── small → normal ZMQ path (no overhead)
    │
    └── large → high-speed bypass → Wolfram Engine
                    │
              files cleaned up automatically

The decision is automatic. You can tweak the threshold if you want:

session = SHMWolframSession(kernel="...", threshold=10*1024)  # 10 KB

Supported data types

Everything you're likely to throw at Wolfram Engine:

float32 · float64 · float16 · int8 · int16 · int32 · int64 · uint8 · uint16 · uint32 · uint64 · complex64 · complex128 · bool

API

SHMWolframSession is a drop-in replacement for WolframLanguageSession. Same methods, same signatures. You can also use the alias BoostSession if you prefer:

from wolfram_boost import BoostSession

Limitations

· Linux only (uses /dev/shm) · Python and Wolfram Engine must run on the same machine · For multi-GB arrays, call np.ascontiguousarray() first

Tests

# Functional smoke test (all data types, cleanup verification)
python3 tests/smoke_test.py

# Benchmark (SHM vs ZMQ)
python3 tests/benchmark_shm_vs_zmq.py

# Extreme stress test (up to 512 MB+)
python3 tests/benchmark_hell.py

License

MIT