Skip to content

Latest commit

 

History

History
8 lines (5 loc) · 646 Bytes

xformers.md

File metadata and controls

8 lines (5 loc) · 646 Bytes
Raw

Note for xformers installation

As bevfusion (or mmdet3d) works with Pytorch==1.10.2, we opt to keep the version of pytorch. However, xformers does not officially support Pytorch<2.0.0. Therefore, we made a customized version of xformers with Pytorch==1.10.2. We only change some searching path of Pytorch, it should be safe.

Only tested with cuda10.2 on V100. flash attention is not supported on V100, therefore our xformers will ignore flash attention on installation.

Note

Please DO NOT use triton or flash-attn with this xformers. If you have installed triton in your python environments, please uninstall it.