HopeJW
I have uploaded the libspconv that supports SM87. Please give the new one a try. Thank you!
No, this is a bug caused by the wrong configuration of LD_LIBRARY_PATH.
Unfortunately, we can not run the CUDA-BEVFusion project on a Pascal architecture GPU. Because the libspconv is designed for Ampere MMA instructions. So, currently only SM87(Ampere Orin), SM80(Ampere A100), and...
For the CUDA-BEVFusion project, currently only SM87 (Ampere, Orin), SM80 (Ampere, A100), and SM101 (Blackwell) are supported. Unfortunately, Xavier (Volta) does not support running this project as the primary limitation...
This is normal. We evaluate the difference between the torch and tensor by using mAP, not absolute difference.
This is because of the nondeterministic implementation in the BEVFusion pipeline, such as Voxelization and Spconv.
The difference will be expected if you are running on fp16 (trtexec --onnx=model.onnx --fp16). Because fp16 has a lower representation precision compared to fp32. If you are running on fp32,...
Hi, libspconv with SM90 has been supported. ```bash $> cuobjdump --list-elf Lidar_AI_Solution/libraries/3DSparseConvolution/libspconv/lib/x86_64_cuda12.6/libspconv.so ELF file 1: libspconv.1.sm_80.cubin ELF file 2: libspconv.2.sm_86.cubin ELF file 3: libspconv.3.sm_90.cubin ELF file 4: libspconv.4.sm_80.cubin ELF file...
SparseConv2d = SparseConv3d with Z=1.
Currently, libspconv only supports CUDA 11.4 and 12.8 on platforms with SM_80 or higher.