Function build_coco_quantizer_ids

src/models/vevo2/components.cpp:898–933

    return engine::core::reshape_tensor(
        ctx,
        ids,
        engine::core::TensorShape::from_dims({1, projected_btd.shape.dims[1]}));
}

}  // namespace

struct Vevo2CocoTokenizerGraph {
    Vevo2CocoTokenizerGraph(
        ggml_backend_t backend,
        engine::core::BackendType backend_type,
        size_t graph_context_bytes,
        const Vevo2CocoTokenizerConfig & config,
        const Vevo2CocoTokenizerWeights & weights,
        int64_t feature_frames)
        : backend(backend),
          uses_whisper(config.coco_type == "content_style"),
          feature_frames(feature_frames),
          output_frames(coco_output_frames(feature_frames, config.downsample_rate)) {
        if (backend == nullptr) {
            throw std::runtime_error("Vevo2 Coco tokenizer graph backend is not initialized");
        }
        if (feature_frames <= 0) {
            throw std::runtime_error("Vevo2 Coco tokenizer graph requires positive feature frames");
        }
        if (uses_whisper && !weights.whisper_input.has_value()) {
            throw std::runtime_error("Vevo2 content-style tokenizer graph requires Whisper input weights");
        }
        if (!weights.chromagram_input.has_value()) {
            throw std::runtime_error("Vevo2 Coco tokenizer graph requires chromagram input weights");
        }

        ggml_init_params params{graph_context_bytes, nullptr, true};
        ctx.reset(ggml_init(params));
        if (ctx == nullptr) {
            throw std::runtime_error("failed to initialize Vevo2 Coco tokenizer graph context");
        }

        engine::core::ModuleBuildContext build_ctx{ctx.get(), "vevo2.coco_tokenizer", backend_type};
        chromagram_input = engine::core::make_tensor(