[BUG] <title>CUDA error: invalid configuration argument when using stride=2 in sparse convolutions (RTX 40-series GPUs / Ada Lovelace Architecture)

[shr19976]

Is there an existing issue for this?

• I have searched the existing issues

Current Behavior

Problem Description：

When performing sparse convolution operations (especially with stride=2), a RuntimeError: CUDA error: invalid configuration argument occurs.
The error happens during kernel map generation in torchsparse, specifically at:
torchsparse/nn/functional/conv/hash/query.py:48 when calling torch.full().

Key Observations ：

1. ​stride=1 works, stride=2 fails​
2. ​GPU Architecture Specificity​
   Reproduced only on Ada Lovelace (Compute Capability 8.9) GPUs. Untested on other architectures (e.g., Ampere).
   ​3. Asynchronous Error Reporting​
   Error message mentions possible asynchronous reporting, but CUDA_LAUNCH_BLOCKING=1 does not resolve the issue.

Error Log
Full Stack Trace :
RuntimeError: CUDA error: invalid configuration argument
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 with TORCH_USE_CUDA_DSA to enable device-side assertions.

Traceback (most recent call last):
File "minimal_repro.py", line 16, in
y = conv(x)
...
File ".../torchsparse/nn/functional/conv/hash/query.py", line 48, in convert_transposed_out_in_map
out_in_map_t = torch.full(
^^^^^^^^^^^
RuntimeError: CUDA error: invalid configuration argument

​Additional Information
​Attempted Fixes ：

Recompiled torchsparse with explicit GPU architecture:
export TORCH_CUDA_ARCH_LIST="8.9"
pip install --force-reinstall git+https://github.com/mit-han-lab/torchsparse.git

Set CUDA_LAUNCH_BLOCKING=1 and TORCH_USE_CUDA_DSA=1, but no resolution.

Expected Behavior

Questions for Developers
​1. Architecture Compatibility​
Is TorchSparse officially supported on Ada Lovelace (Compute Capability 8.9)?

2. ​Stride Configuration Limitations​
   Are there known issues with stride=2 in sparse convolutions? Any special parameter requirements?

3. ​Debugging Suggestions​
   How to further diagnose the torch.full() CUDA configuration error?

Environment

- GCC: 11.4.0
- NVCC:11.8.89
- PyTorch:2.2.0
- PyTorch CUDA:11.8
- TorchSparse:2.1.0
- GPU: NVIDIA GeForce RTX 4070 Ti SUPER (Compute Capability 8.9)
- Python: 3.11.11
- OS: WSL-Ubuntu 22.04

Anything else?

Possible Causes

1. Mismatch Between TorchSparse and PyTorch/CUDA Versions

2. GPU Compute Capability Not Properly Supported

3. Lack of Recompilation from Source with Proper Architecture Flags

[shr19976]

I have installed TorchSparse by running the following command to attempt compatibility between my CUDA and TorchSparse:
** MAX_JOBS=2 TORCH_CUDA_ARCH_LIST="8.9" pip install --upgrade git+https://github.com/mit-han-lab/torchsparse.git **