def old_uniform(kernel_size: int):
    max, min = 2.5, -2.5
    ksize_half = (kernel_size - 1) * 0.5
    kernel = torch.linspace(-ksize_half, ksize_half, steps=kernel_size)
    for i, j in enumerate(kernel):
        if min <= j <= max:
            kernel[i] = 1 / (max - min)
        else:
            kernel[i] = 0

    return kernel.unsqueeze(dim=0)

def new_uniform(kernel_size):
    kernel = torch.zeros(kernel_size)

    start_uniform_index = max(kernel_size // 2 - 2, 0)
    end_uniform_index = min(kernel_size // 2 + 3, kernel_size)

    min_, max_  = -2.5, 2.5
    kernel[start_uniform_index:end_uniform_index] = 1 / (max_ - min_)

    return kernel.unsqueeze(dim=0)

%timeit old_uniform(11)
>>> 354 µs ± 14.7 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

%timeit new_uniform(11)
>>> 13.6 µs ± 303 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

%timeit old_uniform(3)
>>> 123 µs ± 2.19 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

%timeit new_uniform(3)
>>> 11.6 µs ± 1.13 µs per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

for kernel_size in range(1, 101, 2):
    torch.testing.assert_close(old_uniform(kernel_size), new_uniform(kernel_size))