module

ML::GGUF::QuantMatmul

Constants

QK_K = 256

Class methods

matmul_add(x : Array(Float32), rows : Int32, in_dim : Int32, w_raw : Bytes, w_type : TensorType, out_dim : Int32, bias : Array(Float32)) : Array(Float32)

Fused matmul: result[o] = Σ_j x[j] * dequant(W_raw[o, j]) + bias[o] W_raw is quantized weight data as raw bytes, row-major [out_dim rows, in_dim cols]. Each row is a sequence of quantized blocks covering in_dim elements.

Source