Library / SDK
Inference & ServingFine-tuning & Training
xFormers
1Memory 2efficient 3transformer 4components 5from 6Meta
the six Ws · specification
W1
Who
Developed and maintained by Meta's FAIR research team.
W2
What
xFormers offers modular, memory efficient transformer components including fused attention kernels used across research and production models.
W3
Where
Hosted on GitHub at facebookresearch/xformers and distributed via PyPI.
W4
When
Released in 2021 and continuously updated alongside PyTorch.
W5
Why
It gives researchers interchangeable, highly optimized building blocks so they can assemble fast transformer variants without writing custom kernels.
W6
With
Built on PyTorch and often paired with FlashAttention kernels.
W7
Watch
BSD style license, open source, maintained by Meta, roughly 10.5k GitHub stars.
alternatives
works with