Difficulty of Running BEV Former on Blackwell Architecture GPU RTX5060, Request for Update & solution.

[ThuraTunScibotics]

Hello, can someone help me how can I setup BEVformer on RTX5060 since the device only support Pytorch >=2.7.1 / sm120 kernel CUDA12.8?

However, I tried to install BEVformer on the thsi RTX5060 device, but can't work with the torch version of BEVformer. Although I tried with torch2.7.1 and upgrading mmdet3d, mmcv, mmseg, the code base of BEVformer, as well as mmdet3d model builder doesn't support in mmdet3d>1.1.0.

So, is there anyone has already tried for this incompatibility issue of running BEVformer on RTX5060 device? Please, can anyone provide for using BEV former on high Blackwell architecture GPU?

Thanks.

[Arthav24]

Hey @ThuraTunScibotics
were you able to make any progress on this ?

[QZJGeorge]

Hello, we are the authors of nuCarla. We have successfully adapted BEVFormer and several other perception models to run on PyTorch 2.7 with CUDA 12.8. We welcome you to take a look!

[Arthav24]

Hey @QZJGeorge , That sounds great. I’ll definitely look at it. I’ve been working on building an end-to-end transformer-based pipeline for multi-object tracking (MOT) with LiDAR–camera sensor fusion.

[nowander]

Hey @QZJGeorge , That sounds great. I’ll definitely look at it. I’ve been working on building an end-to-end transformer-based pipeline for multi-object tracking (MOT) with LiDAR–camera sensor fusion.

i tried,but failed.Did you run it successfully?

[QZJGeorge]

Could you open an issue at my repo and attach the error information?

[nowander]

Could you open an issue at my repo and attach the error information?

i opened #2 issue yesterday