Function edgetam_build_repvit_graph

sam3.cpp:4796–4830

Build the full RepViT backbone graph. Input: [W, H, 3, 1] image Returns 4 feature maps as stage_outs[0..3], each [C_stage, W_stage, H_stage, 1]

// Input: [W, H, 3, 1] image
// Returns 4 feature maps as stage_outs[0..3], each [C_stage, W_stage, H_stage, 1]
static void edgetam_build_repvit_graph(struct ggml_context* ctx,
                                        struct ggml_tensor* input,
                                        const sam3_model& model,
                                        struct ggml_tensor* stage_outs[4]) {
    const auto& repvit = model.repvit;
    const auto& hp = model.hparams;

    // Stem: conv1 (3→24, k=3, s=2, p=1) + GELU, conv2 (24→48, k=3, s=2, p=1), NO GELU
    // Use ggml_conv_2d_direct to avoid im2col + cont overhead (single Metal kernel)
    auto* x = ggml_conv_2d(ctx, repvit.stem_conv1_w, input, 2, 2, 1, 1, 1, 1);
    x = edgetam_conv2d_bias(ctx, x, repvit.stem_conv1_b);
    x = ggml_gelu(ctx, x);
    x = ggml_conv_2d(ctx, repvit.stem_conv2_w, x, 2, 2, 1, 1, 1, 1);
    x = edgetam_conv2d_bias(ctx, x, repvit.stem_conv2_b);
    // No GELU after stem conv2

    // 4 stages
    for (int s = 0; s < hp.repvit_num_stages; ++s) {
        auto& stage = repvit.stages[s];

        // Downsample at start of stages 1, 2, 3
        if (stage.has_downsample) {
            x = edgetam_repvit_downsample_forward(ctx, x, stage.downsample);
        }

        // Process all blocks in this stage
        for (int b = 0; b < (int)stage.blocks.size(); ++b) {
            x = edgetam_repvit_block_forward(ctx, x, stage.blocks[b]);
        }

        // Store stage output.
        // Keep as [W, H, C, 1] — the EdgeTAM FPN handles this layout directly
        stage_outs[s] = x;
    }
}

// Full EdgeTAM image encoding: preprocess → RepViT → FPN → state
static bool edgetam_encode_image(sam3_state& state,