Standard Kernels

Standard Kernels makes our high-performace GPU kernels usable.

Install and try GEGLU

First, install standardkernels:

python -m pip install standardkernels

Then import the custom module:

import torch
from standardkernels.nn import PreNormGEGLUFFN

block = PreNormGEGLUFFN().cuda().eval()
x = torch.randn(1024, 768, device="cuda", dtype=torch.bfloat16)

with torch.inference_mode():
    y = block(x)

The current GEGLU proof of concept is inference-only and requires its documented CUDA shape, dtype, and hardware contract. See the GEGLU guide for details.

Explore

• GEGLU guide
• Generated API reference