// Muon-owned Moonshot-scaled orthogonal update with decoupled weight decay.
import storage;
import attributes;
struct PushConstants {
uint parameter_index;
uint orthogonal_index;
uint element_count;
float learning_rate;
float weight_decay;
float moonshot_scale;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("muon_apply")]
[domain("ml")]
[variant("generic")]
[dtype("f32")]
[status("experimental")]
[shader("compute")]
[numthreads(256, 1, 1)]
void main(uint3 dispatch_thread_id : SV_DispatchThreadID) {
uint index = dispatch_thread_id.x;
if (index >= push.element_count) {
return;
}
float parameter = load_f32(storage_buffers[push.parameter_index], index);
float orthogonal = load_f32(storage_buffers[push.orthogonal_index], index);
parameter = (1.0f - push.learning_rate * push.weight_decay) * parameter
- push.learning_rate * push.moonshot_scale * orthogonal;
store_f32(storage_buffers[push.parameter_index], index, parameter);
}