cuequivariance-ops-torch (CUDA 13)

0.10.0 verified Fri May 01 auth: no python

GPU-accelerated PyTorch extensions for equivariant neural network primitives, specifically compiled for CUDA 13. Current version 0.10.0, requires Python >=3.10. Part of the cuequivariance ecosystem for efficient group-equivariant operations.

pip install cuequivariance-ops-torch-cu13

Common errors

error ImportError: cannot import name '...' from 'cuequivariance_ops_torch_cu13' ↓

cause Trying to import directly from the pip package name instead of the importable module name.

fix

Change import to 'import cuequivariance_ops_torch' (without the CUDA suffix).

error RuntimeError: CUDA error: no kernel image is available for execution on the device ↓

cause The installed binary was compiled for a different GPU architecture (e.g., sm_80) and is not compatible with your GPU.

fix

Use a version compiled for your GPU's compute capability, or build from source. Alternatively, set environment variable 'CUDA_LAUNCH_BLOCKING=1' for debugging.

error ModuleNotFoundError: No module named 'cuequivariance' ↓

cause The core 'cuequivariance' library is not installed.

fix

Run 'pip install cuequivariance' to install the base dependency.

error AssertionError: Torch not compiled with CUDA enabled ↓

cause PyTorch is installed without CUDA support (CPU-only), but the operations require GPU.

fix

Install a CUDA-enabled version of PyTorch, e.g., 'pip install torch==2.x+cu118' (adjust to your CUDA version).

Warnings

gotcha Import the package as 'cuequivariance_ops_torch', not with the CUDA suffix ('cuequivariance_ops_torch_cu13'). The suffix is only for pip installation to select the correct CUDA version. ↓

fix Use 'import cuequivariance_ops_torch as coe_ops' in code.

breaking CUDA compatibility: The cu13 wheel only works with CUDA 13.x. Installing on a system with a different CUDA version will cause runtime errors (e.g., 'CUDA driver version is insufficient'). ↓

fix Match the package suffix (e.g., cu12, cu13) to your installed CUDA toolkit version. Use 'pip install cuequivariance-ops-torch-cu12' for CUDA 12.

gotcha The package may not be compatible with all CUDA compute capabilities; check the wheel's target architectures (e.g., sm_80, sm_90). Running on an unsupported GPU will raise a 'no kernel image' error. ↓

fix Ensure your GPU is supported (e.g., Ampere or newer). If not, you may need to build from source or use a CPU fallback.

Imports

all relevant ops
wrong
```
from cuequivariance_ops_torch_cu13 import ...
```
correct
```
import cuequivariance_ops_torch as coe_ops
```
The package name for import does not include the CUDA suffix; import using 'cuequivariance_ops_torch'.

Quickstart

Basic usage: import the package (without CUDA suffix) and call an operation on a CUDA tensor.

import torch
import cuequivariance_ops_torch as coe_ops

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
x = torch.randn(10, 10, device=device)
y = coe_ops.some_op(x)  # Replace with actual operation
print(y.shape)