class

MicroGPT::WeightStore

Constructors

new(mats : Array(Mat), adam_mats : Array(Mat) | Nil = nil)

weight_mats: all trainable parameters adam_mats: corresponding Adam m/v Mats (2 per weight mat: m, v) Order: [w0_m, w0_v, w1_m, w1_v, ...]

adam_m_offset

Adam state lives at known offsets in the same buffer

adam_m_ptr(mat_idx : Int32) : Pointer(Float32)

Get Adam M pointer for a specific mat index

adam_v_offset

adam_v_ptr(mat_idx : Int32) : Pointer(Float32)

Get Adam V pointer for a specific mat index

bulk_adam_step(grad_mats : Array(Mat), lr : Float64, t : Int32)

Bulk Adam: gather gradients, then one kernel launch for all parameters

download_all

Download all weights back to CPU

finalize

mat_offsets

mats

store_ptr

total_bytes

total_floats

weights_bytes