Function build_coco_vocos_encoder

src/models/vevo2/components.cpp:846–876

        GGML_PREC_F32,
    }).build(ctx, hidden, weights.encoder.output);
}

engine::core::TensorValue build_l2_normalized_rows(
    engine::core::ModuleBuildContext & ctx,
    const engine::core::TensorValue & input) {
    auto contiguous = engine::core::ensure_backend_addressable_layout(ctx, input);
    auto squared = engine::modules::MulModule{}.build(ctx, contiguous, contiguous);
    auto summed = engine::modules::ReduceSumModule({1}).build(ctx, squared);
    auto norm = engine::core::wrap_tensor(
        ggml_sqrt(ctx.ggml, ggml_scale_bias(ctx.ggml, summed.tensor, 1.0F, 1.0e-20F)),
        summed.shape,
        GGML_TYPE_F32);
    auto repeated_norm = engine::modules::RepeatModule({contiguous.shape}).build(ctx, norm);
    return engine::core::wrap_tensor(
        ggml_div(ctx.ggml, contiguous.tensor, repeated_norm.tensor),
        contiguous.shape,
        GGML_TYPE_F32);
}

engine::core::TensorValue build_coco_quantizer_ids(
    engine::core::ModuleBuildContext & ctx,
    const engine::core::TensorValue & encoder_output_btd,
    const Vevo2CocoTokenizerWeights & weights,
    const Vevo2CocoTokenizerConfig & config) {
    auto hidden_bdt = engine::modules::TransposeModule({{0, 2, 1, 3}, 3}).build(ctx, encoder_output_btd);
    auto projected_bdt = engine::modules::Conv1dModule({
        config.hidden_size,
        config.codebook_dim,
        1,
        1,
        0,
        1,
        weights.quantizer.in_project.bias.has_value(),
    }).build(ctx, hidden_bdt, weights.quantizer.in_project);