numfast 0.0.1

GPU-first numerical computing framework — Runtime, Driver, Compute, Agent SDK

NumFast

GPU-first numerical computing framework — Runtime, Driver, Compute, High-Level API and Agent SDK.

pip install numfast

Architecture

User (human or AI)              ─┐
    │                            │ Agent SDK
    ▼                            │ AI-agnostic
Agent SDK                        ─┘
    │
    ▼
Session — owns Runtime
    │
    ├──► Runtime — kernel table, compiler, executor
    ├──► Driver — CPU (reference) · WebGPU · CUDA (stub)
    ├──► Operations — scan(), matmul(), fft(), sort(), histogram()
    └──► Compute — 6 GPU algorithms with CPU==GPU conformance

Quick start

from numfast import scan, sort, matmul
import numpy as np

# Prefix sum
x = np.array([1.0, 2.0, 3.0, 4.0, 5.0])
y = scan(x)
print(y)  # [1. 3. 6. 10. 15.]

# Bitonic sort
x = np.random.random(8)
s = sort(x)

# Matrix multiplication
A = np.ones((4, 4), dtype=np.float64)
B = np.ones((4, 4), dtype=np.float64)
C = matmul(A.flatten(), B.flatten(), M=4, N=4, K=4)

Agent SDK

from numfast import AgentSDK

sdk = AgentSDK()

with sdk.session() as session:
    # Run a script
    result = session.run("x = 2 + 2")

    # Compare two arrays
    result = session.diff(
        np.array([1.0, 2.0, 3.0]),
        np.array([1.0, 2.0, 3.0]),
    )

    # Compare CPU vs GPU
    cpu_result = ...
    gpu_result = ...
    session.compare(cpu_result, gpu_result)

    # Run benchmarks
    session.benchmark("matmul_4x4", "matmul(A, B, M=4, N=4, K=4)")

Tools

Tool	Description
scan()	Inclusive prefix sum (CPU/GPU)
matmul()	Tiled matrix multiplication
fft()	Fast Fourier Transform
sort()	Bitonic sort (power of 2)
histogram()	Binned histogram with atomics

Backends

• CPU — reference implementation, float64, always available
• WebGPU — GPU via wgpu-py (Windows, Linux, macOS)
• CUDA, OpenCL, Metal — driver interface ready (stubs)

Every GPU kernel has CPU == GPU conformance tests.

Requirements

• Python >= 3.11
• NumPy (CPU)
• wgpu-py (optional, for GPU)

Development

# Run all tests
pytest

# Run conformance suite
python -m tests.conformance.test_compute

License

AGPL-3.0-only

Status

Pre-alpha. API is stable (frozen per ADR-005). Active development.