Function rmsprop_step 

pub fn rmsprop_step<T: Triton, const BLOCK_SIZE: i32>(
    params_ptr: T::Pointer<f32>,
    grad_ptr: T::Pointer<f32>,
    square_avg_ptr: T::Pointer<f32>,
    n_elements: i32,
    lr: f32,
    alpha: f32,
    eps: f32,
    weight_decay: f32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<f32>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<f32>>>,

RMSprop step (no momentum, not centred).

square_avg = alpha * square_avg + (1 - alpha) * g²
p          = p - lr * g / (sqrt(square_avg) + eps)

Grid: [ceil(n_elements / BLOCK_SIZE), 1, 1].