Function build_coco_convnext_block

src/models/vevo2/components.cpp:812–844

    }).build(ctx, hidden, weights.pwconv2);
    hidden = scale_last_dim(ctx, hidden, weights.gamma);
    hidden = engine::modules::TransposeModule({{0, 2, 1, 3}, 3}).build(ctx, hidden);
    return engine::modules::AddModule{}.build(ctx, input_bct, hidden);
}

engine::core::TensorValue build_coco_vocos_encoder(
    engine::core::ModuleBuildContext & ctx,
    const engine::core::TensorValue & input_bct,
    const Vevo2CocoTokenizerWeights & weights,
    const Vevo2CocoTokenizerConfig & config) {
    auto hidden = engine::modules::Conv1dModule({
        config.hidden_size,
        config.vocos_dim,
        7,
        1,
        3,
        1,
        weights.encoder.embed.bias.has_value(),
    }).build(ctx, input_bct, weights.encoder.embed);
    hidden = engine::modules::TransposeModule({{0, 2, 1, 3}, 3}).build(ctx, hidden);
    hidden = engine::modules::LayerNormModule({config.vocos_dim, 1.0e-6F, true, true})
                 .build(ctx, hidden, weights.encoder.norm);
    hidden = engine::modules::TransposeModule({{0, 2, 1, 3}, 3}).build(ctx, hidden);
    for (const auto & block : weights.encoder.convnext) {
        hidden = build_coco_convnext_block(ctx, hidden, block, config);
    }
    hidden = engine::modules::TransposeModule({{0, 2, 1, 3}, 3}).build(ctx, hidden);
    hidden = engine::modules::LayerNormModule({config.vocos_dim, 1.0e-6F, true, true})
                 .build(ctx, hidden, weights.encoder.final_norm);
    return engine::modules::LinearModule({
        config.vocos_dim,
        config.hidden_size,
        true,
        GGML_PREC_F32,
    }).build(ctx, hidden, weights.encoder.output);
}