[Installation] <title>Can this be applied to 5090?

[lh06-x]

Is there an existing issue for this?

• I have searched the existing issues

Have you followed all the steps in the FAQ?

• I have tried the steps in the FAQ.

Current Behavior

I tried to apply Torchsparse to 5090, in cuda12.8 and PyTorch Nightly. But there is always the problem of illegal memory access. I wanna know how to solve this.

Error Line

File "/root/autodl-tmp/.conda-stuff/envs/pointcept/lib/python3.10/site-packages/torchsparse-2.1.0-py3.10-linux-x86_64.egg/torchsparse/nn/functional/conv/kmap/func/hashmap_on_the_fly.py", line 185, in build_kmap_Fetch_on_Demand_hashmap_on_the_fly
results = torch.t(kmap["out_in_map"]).contiguous()
torch.AcceleratorError: CUDA error: an illegal memory access was encountered
Search for cudaErrorIllegalAddress' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. Compile with TORCH_USE_CUDA_DSA` to enable device-side assertions.

Environment

- GCC:13.2
- NVCC:13.2
- PyTorch:2.91
- PyTorch CUDA:12.8

Full Error Log

Error Log

[PUT YOUR ERROR LOG HERE]

[prakashknaikade]

hi, did you manage to solve this? @lh06-x

[lh06-x]

Oh right. I have received all of your messages. I haven't find a way to solve this problem. I just changed my gpu from 5090 to 4090. The reason behind the problem is that the Torchsparse hasn't update for cuda 12.8 and above. But 5090 requires the cuda12.8 or above. So if you can find a way to make the Torchsparse compile well in cuda12.8 or if you can find a way to use lower-version cuda(such as cuda12.4) on 5090, maybe this problem can be solved. If you have any new findings, please let me know. Thank'u very much. @prakashknaikade

[MinChoi0129]

Thanks to claude code, I fixed it. Now, I can run on Blackwell GPU! (RTX5090)

Python: 3.9
torch: 2.8.0

pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128

---

(I'm trying to use torchsparse==1.4.0 for specific reason)

git clone --branch v1.4.0 --depth 1 https://github.com/mit-han-lab/torchsparse.git
cd torchsparse

Patch: .type() → .scalar_type() (PyTorch 2.x compatibility)

sed -i 's/.type(), "/.scalar_type(), "/g' torchsparse/backend/convolution/convolution_cuda.cu torchsparse/backend/devoxelize/devoxelize_cuda.cu torchsparse/backend/voxelize/voxelize_cuda.cu

pip install .

---

python -c "import torchsparse; print('torchsparse', torchsparse.version)"
[Out] torchsparse 1.4.0

[prakashknaikade]

I also managed to install but it does not really work>
File "/home/miniconda3/envs/scenescript/lib/python3.9/site-packages/torchsparse/nn/modules/norm.py", line 24, in forward batch_size = torch.max(coords[:, 0]).item() + 1 torch.AcceleratorError: CUDA error: an illegal memory access was encountered CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile withTORCH_USE_CUDA_DSAto enable device-side assertions.

But have you managed to perform any operation using it? @MinChoi0129

[MinChoi0129]

I'm using torchsparse to run the MF-MOS network as described in the SCNU-RISLAB GitHub (https://github.com/SCNU-RISLAB/MF-MOS). I’ve tested it in inference mode, and it works perfectly—no errors or warnings so far. Unfortunately, I'm not an expert in fixing complex CUDA errors, so if you're trying to figure out what's happening under the hood, I’d suggest using the Claude VS Code extension. It might help you identify any potential issues quite faster.

[prakashknaikade]

which gcc, python and nvcc version did use? @MinChoi0129

[MinChoi0129]

Here it is.

(mfmos) root@7a3b8b5ae17e:/workspace# nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Wed_Jan_15_19:20:09_PST_2025
Cuda compilation tools, release 12.8, V12.8.61
Build cuda_12.8.r12.8/compiler.35404655_0

(mfmos) root@7a3b8b5ae17e:/workspace# gcc --version
gcc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Copyright (C) 2021 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

(mfmos) root@7a3b8b5ae17e:/workspace# python --version
Python 3.9.23

[lh06-x]

@MinChoi0129 Thanks for your sharing, I'll try to use cluade code to fix this.