Blis BLAS-like Linear Algebra Library

1.3.3 · active · verified Sun Mar 29

Blis is a Python library that provides a high-performance, self-contained C-extension for BLAS-like dense linear algebra operations. Developed by Explosion (the creators of spaCy), it focuses on delivering fast matrix multiplication and other linear algebra routines without requiring external system dependencies. The library is actively maintained, with a current version of 1.3.3, and releases often include support for newer Python versions and updates to underlying dependencies like NumPy.

Warnings

breaking Blis versions 1.0.0 and later introduce a breaking change due to dependency on NumPy 2.0. Wheels built against NumPy 1.x are not compatible with NumPy 2.0. If you distribute binaries, they must be built against NumPy 2.x to be compatible with both NumPy 1.x and 2.x runtimes.
Fix: Ensure your project's `blis` installation and any dependent packages are built against NumPy 2.x for maximum compatibility. If you are building from source, this will typically happen automatically when NumPy 2.x is present. Downstream projects using `blis` should verify their build processes with NumPy 2.x.
gotcha For optimal performance on non-x86_64 or osx/arm64 architectures, you may need to compile `blis` from source with specific architecture flags. The provided wheels are optimized for common architectures, but for other CPUs, a generic (slower) build might be used unless `BLIS_ARCH` environment variable is set during installation.
Fix: If experiencing poor performance on specific hardware, refer to the `blis` GitHub repository documentation on 'Building BLIS for alternative architectures' and install using `BLIS_ARCH="your_arch" pip install blis --no-binary blis`.
gotcha Blis is designed for single-threaded execution for its intended workloads (ML inference). While it is re-entrant and safe to use concurrently from multiple threads with immutable data, concurrent mutation of input NumPy arrays from different threads while Blis is running can lead to data races or segmentation faults.
Fix: Avoid concurrent mutation of NumPy arrays used as input to Blis operations across multiple threads. Ensure data is immutable during Blis computations or manage threading carefully at a higher level.
gotcha Earlier versions around `v1.0.0` experienced memory access errors and instability on Windows. This issue was addressed in `v1.2.0` by reverting to an older, more stable version of the vendored Blis library.
Fix: Users on Windows encountering instability should upgrade to `blis` version `1.2.0` or newer.

Install

pip install blis Install stable version

Imports

einsum
```
from blis.py import einsum
```
Provides a high-level Python API for Einstein summation, similar to NumPy's einsum but optimized for Blis routines.
cy
```
cimport blis.cy
```
Used for direct Cython access, offering fused-type, nogil bindings to the underlying Blis library.

Quickstart

This quickstart demonstrates how to use the high-level `einsum` function from `blis.py` for efficient matrix multiplication. It uses NumPy arrays as input and output, showcasing how Blis integrates with the NumPy ecosystem. The `einsum` function supports various tensor operations with specific restrictions that allow direct mapping to optimized Blis routines. Ensure NumPy is installed as a dependency.

import numpy as np
from blis.py import einsum

dim_a, dim_b, dim_c = 500, 128, 300

arr1 = np.random.rand(dim_a, dim_b).astype('float32')
arr2 = np.random.rand(dim_b, dim_c).astype('float32')
out = np.zeros((dim_a, dim_c), dtype='float32')

# Perform matrix multiplication using einsum
einsum('ab,bc->ac', arr1, arr2, out=out)

print(f"Output shape: {out.shape}")
# Example of returning a new array with transposed output
out_transposed = einsum('ab,bc->ca', arr1, arr2)
print(f"Transposed output shape: {out_transposed.shape}")

view raw JSON →