Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
hpc100 opened this issue · comments
Hello,
Thank's for your amazing work.
I install Point Transformer V3 on A6000 GPU and when I run CUDA_LAUNCH_BLOCKING=1 CUDA_VISIBLE_DEVICES=1 sh scripts/train.sh -g 1 -d s3dis -c semseg-pt-v3m1-1-rpe -n semseg-pt-v3m1-1-rpe
I got this error
RuntimeError: Trainer: CUDA error: the provided PTX was compiled with an unsupported toolchain. Compile with
TORCH_USE_CUDA_DSA to enable device-side assertions.
I install point_opts and pointgroup_ops with TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
Any solutions ?
Thank's in advance.
Hi, no need to set TORCH_CUDA_ARCH_LIST if you build the lib and run the lib on the same GPU arch.
I got the same., with installing without TORCH_CUDA_ARCH_LIST and simply run python setup.py install
Yes, I used it to make docker environments.