bf16

BF16 is a Python package for converting floating-point numbers to BFloat16 format using optimized Windows x86-64 assembly.

Live Data

Pepy.tech data

Github.com data

Pypi.org data

Socket.dev data

Virustotal Scan Results

VirusTotal

More Badges

Installation

To install bf16 the 2 easiest options are either run this command:

pip install bf16

or click this button to download a zip containing only necessary files to download and also extract it:

Usage

import bf16

x = bf16.bfloat16(3.14)
print(x)  # Outputs the BFloat16 approximation as a float and for the record it also actually output 3.125 because its a real brain-float16
print(f"Size: {len(x.data)} bytes")  # Always 2 bytes

Requirements

• Windows AMD x86-64
• Python 3.6+
• cffi The package includes a compiled assembly executable for fast conversion.

Acknowledgments

• Thank you to Tomasz Grysztar for his work on FASM which made it possible to optimize bf16.dll and make it smaller
• Jeremy Gordon thank you for making golink which helped with earlier versions of bf16

Changelog

1.0.1

bf16 has completely changed with the last update here is the main stuff

• added nearly all operators that function that exist within float32 to bfloat16
• instead of bf16.obj and bf16.exe it is now just a bf16.dll
• optimized to not have to open and exit bf16.dll everytime
• added Changelog
• added acknowlegement to golink I forgot

1.0.2

bf16 has gained a way for people to know they are downloading a safe package

• corrected README.md old changelog for 1.0.1
• added repo and its link

1.0.3

bf16 now has more verified details

• added more verified details
• made SECURITY.md workflow and CODEOWNERS

1.0.4

bf16 now has finished documentation.

• Finished community Standards
• Added downloading Zip file allowing you to only download only necessary files saving memory including README.md, LICENSE, bf16/__init__.py, and bf16/bf16.dll NOTE: go to [[CHANGELOG.md]] for more specific information on the changelog

1.0.5

bf16 is now much faster!

• optimized bf16.dll by using vectorization merged functions and merged sections
• switched from GoLink&NASM to FASM for increased speed and smaller dll
• at the cost of adding a dependency uses cffi to make __init__.py much faster