mlp-block_kernel. Fused MLP block optimization using a custom CUDA kernel in CuPy. Implements in-place GELU fusion between two cuBLAS GEMMs to reduce memory traffic and kernel launches, achieving measurable speedups while preserving numerical correctness.