import torch
from kernels import get_kernel
# Download optimized kernels from the Hugging Face hub
ops = get_kernel("mohitsha/aiter_aiter_operator")
# Random tensor
x = torch.randn((10, 10), dtype=torch.bfloat16, device="cuda")
y = torch.randn((10, 10), dtype=torch.bfloat16, device="cuda")
out = ops.div(y, x)
print(out)