Method build

subprojects/llama.cpp/tools/mtmd/models/mobilenetv5.cpp:248–451

}

ggml_cgraph * clip_graph_mobilenetv5::build() {
    ggml_tensor * inp = build_inp_raw();

    // 1. Stem - Conv2dSame(3, 64, kernel_size=(3, 3), stride=(2, 2))
    ggml_tensor * cur = pad_same_2d(inp, 3, 3, 2, 2);  // Apply SAME padding

    cur = ggml_conv_2d_direct(ctx0, model.mobilenet_stem_conv_w, cur, 2, 2, 0, 0, 1, 1);  // padding=0
    if (model.mobilenet_stem_conv_b) {
        cur = ggml_add(ctx0, cur, model.mobilenet_stem_conv_b);
    }
    if (model.mobilenet_stem_norm_w) cur = rms_norm_2d(cur, model.mobilenet_stem_norm_w);
    cur = ggml_gelu(ctx0, cur);


    // 2. Blocks
    std::vector<ggml_tensor*> intermediate_features;
    const int total_blocks = model.mobilenet_blocks.size();

    auto is_stage_start = [&](int i) {
        if (i == 0) return true;
        for (int end_idx : model.mobilenet_stage_ends) {
            if (i == end_idx + 1) return true;
        }
        return false;
    };

    auto is_fusion_point = [&](int i) {
        if (model.mobilenet_stage_ends.size() >= 4) {
                if (i == model.mobilenet_stage_ends[2]) return true; // End of Stage 2
                if (i == model.mobilenet_stage_ends[3]) return true; // End of Stage 3
        } else {
            if (i == total_blocks - 1) return true;
        }
        return false;
    };

    for (int i = 0; i < total_blocks; i++) {
        const auto & block = model.mobilenet_blocks[i];
        int stride = is_stage_start(i) ? 2 : 1;

        if (block.s0_conv_exp_w)      cur = build_edge_residual(cur, block, stride);
        else if (block.attn_q_w)      cur = build_mobilenet_attn(cur, block);
        else                          cur = build_inverted_residual(cur, block, stride);

        if (is_fusion_point(i)) {

            intermediate_features.push_back(cur);
        }
    }

    // 3. Multi-Scale Fusion Adapter (MSFA)
    if (!intermediate_features.empty()) {

        // A. Reference Resolution: PyTorch implementation uses inputs[0]
        // We assume intermediate_features[0] is the "High Resolution" target.
        // In MobileNet designs, this is typically the feature map with the smallest stride (e.g. 32x32).
        ggml_tensor* target_feat = intermediate_features[0];
        int high_res_w = target_feat->ne[0];