nn.gelu (PRECISE, erf-based): x · (1 + erf(x/√2)) / 2 — composed exactly as mlx's python source (weak scalars promote to x's dtype). Used by the gemma-1 gated MLP and starcoder2's plain MLP; distinct from geluApprox (gelu_pytorch_tanh).
nn.gelu (PRECISE, erf-based): x · (1 + erf(x/√2)) / 2 — composed exactly as mlx's python source (weak scalars promote to x's dtype). Used by the gemma-1 gated MLP and starcoder2's plain MLP; distinct from geluApprox (gelu_pytorch_tanh).