Difficulty of Running mmdetection2.x (for bevformer) on Blackwell Architecture GPU RTX5070, Request for Update & solution.

[long2double]

Hello, can someone help me how can I setup BEVformer on RTX5070 since the device only support Pytorch >=2.7.1 / sm120 kernel CUDA12.8?

However, I tried to install BEVformer on the thsi RTX5070 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 RTX5070 device? Please, can anyone provide for using BEVFormer on high Blackwell architecture GPU?

Thanks.

[rehan243]

Are you running the latest CUDA 12.x toolkit on your RTX 5070? We've seen similar issues with mmdetection and newer CUDA versions, turns out it's often a PyTorch version mismatch. I'd suggest trying to build mmcv from source with the specific CUDA version you're on - we did this for our production environment with CUDA 12.2 and PyTorch 2.1, and it worked fine. The thing is, mmdet3d>1.1.0 has some breaking changes, so you might need to tweak the BEVformer codebase a bit. We hit this exact issue with our RAG pipeline at Reallytics.ai when upgrading our inference cluster to newer GPUs.