blackwell cu128

[13918763630]

Is spconv support the latest cuda12.8 and the blackwell? Can i use this package on RTX 5090? Thank you so much!

[linya205]

+1

[Cupcee]

For people wondering, it seems like it does not work currently on Cuda 12.8. I tried compiling from source since the wheels don't exist, and I get segfault in pytorch backward pass when trying to train models.

[boxiao-wv]

Crying for the support

[linya205]

“pip install spconv-cu120” works

[Wlll7Ai]

+1

[Phytered]

+1

[SindreYang]

I found it！ https://github.com/MrForExample/Comfy3D_Pre_Builds/tree/main/_Build_Wheels/_Wheels_win_py312_torch2.7.0_cu128

[adithyamurali]

+1 Has anyone been able to make this work on CUDA 12.8?

[kenomo]

Building and basic tests work, when building spconv and cumm from sources:

docker run --rm -it --gpus all \
	-e CUMM_CUDA_VERSION='128' \
	-e CUMM_DISABLE_JIT=1 \
	-e SPCONV_DISABLE_JIT=1 \
	-e CUMM_CUDA_ARCH_LIST='12.0' \
	pytorch/pytorch:2.7.1-cuda12.8-cudnn9-devel \
bash -c "
apt update && apt -y install git build-essential && \
pip install pccm wheel && \
git clone https://github.com/FindDefinition/cumm.git /opt/cumm && \
cd /opt/cumm && pip install . && \
git clone https://github.com/kenomo/spconv.git /opt/spconv && \
cd /opt/spconv && pip install . && \
python /opt/spconv/test/benchmark.py && \
python /opt/spconv/test/fake_train.py
"

Using cumm-cu128 fails!

[fatrybl]

@kenomo I admit that your method works perfect out-of-the-box. Thank you. Appreciate it a lot!

Here is the same copy-paste command (tested on RTX 5070Ti with cuda 12.9 on host machine):

docker run --rm -it --gpus all
-e CUMM_CUDA_VERSION='128'
-e CUMM_DISABLE_JIT=1
-e SPCONV_DISABLE_JIT=1
-e CUMM_CUDA_ARCH_LIST='12.0'
pytorch/pytorch:2.7.1-cuda12.8-cudnn9-devel
bash -c "apt update && apt -y install git build-essential && pip install pccm wheel && git clone https://github.com/FindDefinition/cumm.git /opt/cumm && cd /opt/cumm && pip install . && git clone https://github.com/kenomo/spconv.git /opt/spconv && cd /opt/spconv && pip install . && python /opt/spconv/test/benchmark.py && python /opt/spconv/test/fake_train.py"

[Arthav24]

I am able to build libspconv.so on SM_120 Cuda13 & 12.8 but executing example ./main ../benchmark-pc.jarr throws this error

Hello libspconv!!!
0 20 796 788
0 49 650 1056
0 48 1112 1162
0 25 669 477
0 20 825 995
num voxels [125562, 3] [200000, 3]
[3, 3, 3] [80, 1600, 1600] [80, 1600, 1600]
native example
terminate called after throwing an instance of 'std::runtime_error'
  what():  /opt/spconv/example/libspconv/spconv/src/spconvlib/spconv/csrc/sparse/convops/gemmops/GemmTunerSimple/GemmTunerSimple_tune_and_cache.cc(91)
!all_profile_res.empty() assert faild. can't find suitable algorithm
Aborted (core dumped)

Running on RTX 5070Ti Any help @traveller59 @kenomo

tried different combination of branches too CUMM - SPCONV - CUDA version - STATUS master - master - 12.8 - Same error v0.8.2 - v2.3.8 - 13.0 - Same error