pub fn muon_update<T: Triton, const BLOCK_SIZE: i32>(
params_ptr: T::Pointer<f32>,
grad_ptr: T::Pointer<f32>,
n_elements: i32,
lr: 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>>>,Expand description
Parameter update: W -= lr · G_orth.
Grid: [ceil(n_elements / BLOCK_SIZE), 1, 1].