Migrating to MMCV version 2.x & Nvidia Blackwell B200 GPUs

[deepak-vij]

Hi @hli2020 @eltociear @faikit @Yihanhu @wljungbergh , after jumping through hoops, I finally got the the entire UniAD2.0 training/eval environment (BevFormer, Stage1, Stage2) working on Nvidia Blackwell B200 environment. However, I ended up using MVCC libraries version 1.6.1 compatible libraries. I tried switching to MVCC 2.x codebase. However, due to major code refactoring, I gave up and ended up staying with version 1.6.1.

The question I have is, has anyone in the community tried the migration effort from MVCC version 1.x towards version 2.x. It looks like substantial amount of migration work due to major code refactoring. Although, benefits of doing so seem very tempting (optimized cuda kernels, support for tocrh.compile, kernel fusion etc.).

Please let me know your thoughts before I start thinking through this migration effort. Appreciate it. Thanks.

[niouoklul]

Hi @deepak-vij and UniAD maintainers ,

Thanks for starting this discussion about migrating UniAD 2.0 to MMCV 2.x on GPUs like the Blackwell B200 — I’m encountering a similar situation and would love some detailed guidance.

Specifically, would it be possible for you or anyone in the community to provide:

1. The exact set of installation steps you used to get UniAD training / evaluation working on the Blackwell B200 (including:

   • PyTorch + CUDA versions
   • exact mmcv/mmcv‑full version and how it was installed
   • mmdet, mmdet3d, mmseg versions that worked
   • any required patch files / configuration changes
   • whether openmim was used or manual wheel installs

2. If you attempted migration to MMCV 2.x, could you share:

   • where the migration breaks (specific files / APIs)
   • which parts of UniAD code depend on MMCV 1.x internals
   • pointers on how to update those parts to MMCV 2.x (example diffs or a patch)

3. Any working environment configuration (environment.yaml or requirements list) that others can reproduce easily.

I’m open to testing combinations and providing logs if that helps — any concrete steps or best practices would be much appreciated!

Thanks!