Method altup_correct

subprojects/llama.cpp/src/models/gemma3n-iswa.cpp:364–384

input predictions shape: [n_embd, n_tokens, n_altup] input activated shape: [n_embd, n_tokens] output shape: [n_embd, n_tokens, n_altup]

// input activated         shape: [n_embd, n_tokens]
// output                  shape: [n_embd, n_tokens, n_altup]
ggml_tensor * llm_build_gemma3n_iswa::altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
    ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
    cb(modalities, "modalities", il);

    ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);
    ggml_tensor * innovation        = ggml_sub(ctx0, activated, active_prediction);  // [n_embd, n_tokens]
    cb(innovation, "innovation", il);

    ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities);  // [n_altup, n_tokens]
    all_coefs               = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f);                    // + 1.0
    cb(all_coefs, "all_coefs", il);
    all_coefs = ggml_transpose(ctx0, all_coefs);                                               // [n_tokens, n_altup]
    all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup);                           // [1, n_tokens, n_altup]

    innovation              = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1);
    ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs);   // [n_embd, n_tokens, n_altup]
    corrected               = ggml_add(ctx0, corrected, predictions);  // [n_embd, n_tokens, n_altup]
    cb(corrected, "corrected", il);

    return corrected;
}