💡This repository provides a PyPI simple index of prebuilt wheels for BEVFusionx.
⚡By using this index, you can conveniently install via pip install, saving time on local builds and configuration.
To install, use the --extra-index-url flag with pip install:
pip install {package} --extra-index-url https://rathaumons.github.io/bevfusionx-index/{suffix}/
💡See the Summary Table below for more details on {package} and {suffix}.
Examples:
# torchpack - any (for both CPU/GPU)
pip install torchpack==0.3.2 --extra-index-url https://rathaumons.github.io/bevfusionx-index/any/
# flash-attention - cu130 (CUDA 13.0 for Consumer + Workstation + Jetson)
pip install flash-attn==1.2.1 --extra-index-url https://rathaumons.github.io/bevfusionx-index/cu130/
# mmcv - cu130d (CUDA 13.0 for Data Center)
pip install mmcv-full==1.7.4 --extra-index-url https://rathaumons.github.io/bevfusionx-index/cu130d/
| Package</br>Name | Latest</br>Version | Index</br>Suffix | Release</br>Tag | Built with PyTorch</br>torch/torchvision |
|---|---|---|---|---|
cumm-cu130 |
v0.9.1 |
cu130d |
v0.9.1-cumm-cu130d |
None/None |
flash-attn |
v1.2.1 |
cu113 |
v1.2.1-flash-attention-cu113 |
1.12.1+cu113/0.13.1+cu113 |
flash-attn |
v1.2.1 |
cu121 |
v1.2.1-flash-attention-cu121 |
2.5.1+cu121/0.20.1+cu121 |
flash-attn |
v1.2.1 |
cu126 |
v1.2.1-flash-attention-cu126 |
2.11.0+cu126/0.26.0+cu126 |
flash-attn |
v1.2.1 |
cu128 |
v1.2.1-flash-attention-cu128 |
2.11.0+cu128/0.26.0+cu128 |
flash-attn |
v1.2.1 |
cu130d |
v1.2.1-flash-attention-cu130d |
2.11.0+cu130/0.26.0+cu130 |
flash-attn |
v1.2.1 |
cu130 |
v1.2.1-flash-attention-cu130 |
2.11.0+cu130/0.26.0+cu130 |
mmcv-full |
v1.7.4 |
cu113 |
v1.7.4-mmcv-cu113 |
1.12.1+cu113/0.13.1+cu113 |
mmcv-full |
v1.7.4 |
cu121 |
v1.7.4-mmcv-cu121 |
2.5.1+cu121/0.20.1+cu121 |
mmcv-full |
v1.7.4 |
cu126 |
v1.7.4-mmcv-cu126 |
2.11.0+cu126/0.26.0+cu126 |
mmcv-full |
v1.7.4 |
cu128 |
v1.7.4-mmcv-cu128 |
2.11.0+cu128/0.26.0+cu128 |
mmcv-full |
v1.7.4 |
cu130d |
v1.7.4-mmcv-cu130d |
2.11.0+cu130/0.26.0+cu130 |
mmcv-full |
v1.7.4 |
cu130 |
v1.7.4-mmcv-cu130 |
2.11.0+cu130/0.26.0+cu130 |
spconv-cu130 |
v2.4.1 |
cu130d |
v2.4.1-spconv-cu130d |
None/None |
torchpack |
v0.3.2 |
any |
v0.3.2-torchpack-any |
None/None |
THESE WHEELS WERE BUILT SPECIFICALLY FOR BEVFUSION𝕏 AND MAY NOT WORK IN YOUR ENVIROMENT.
The targeted NVIDIA GPU architectures for these builds were determined based on the logic of nvidia-arch using gpu_type="cons+jets", min_sm=60, except for:
cu130d builds, which used gpu_type="dcen".flash-attn builds, which used min_sm=75 since it does not support older architectures.| Index Suffix | Included NVIDIA Arches for NVCC |
|---|---|
cu130 |
7.5;8.6;8.7;8.9;11.0;12.0;12.1+PTX |
cu130d |
7.5;8.0;8.6;8.9;9.0;10.0;10.3;12.0+PTX |
cu128 |
6.0;6.1;6.2;7.0;7.2;7.5;8.6;8.7;8.9;12.0+PTX |
cu126 |
6.0;6.1;6.2;7.0;7.2;7.5;8.6;8.7;8.9+PTX |
cu121 |
6.0;6.1;6.2;7.0;7.2;7.5;8.6;8.7;8.9+PTX |
cu113 |
6.0;6.1;6.2;7.0;7.2;7.5;8.6+PTX |
The base Docker images are sourced from ratharog/manylinux_2_28_x86_64.