Method gaussian_topk

subprojects/llama.cpp/src/models/gemma3n-iswa.cpp:312–318

input x shape: [n_embd, n_tokens] output shape: [n_embd, n_tokens]

// input x shape: [n_embd, n_tokens]
// output  shape: [n_embd, n_tokens]
ggml_tensor * llm_build_gemma3n_iswa::gaussian_topk(ggml_tensor * x) {
    ggml_tensor * mean = ggml_mean(ctx0, x);
    ggml_tensor * std  = ggml_sqrt(ctx0, ggml_scale(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))),
                                                    1.0f / (float) (x->ne[0] - 1)));
    ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul));
    return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x));
}

//
// altup functions