Class ForwardGraph

src/models/omnivoice/generator.cpp:419–934

    virtual int64_t total_token_capacity() const noexcept = 0;
    virtual int64_t target_frame_capacity() const noexcept = 0;
    virtual double rebuild_clear_ms() const noexcept = 0;
    virtual double rebuild_build_ms() const noexcept = 0;
    virtual double rebuild_alloc_ms() const noexcept = 0;
    virtual double rebuild_init_ms() const noexcept = 0;
};

class ForwardGraph final : public GeneratorForwardGraph {
public:
    ForwardGraph(
        std::shared_ptr<WeightsRuntime> runtime,
        size_t graph_arena_bytes,
        int64_t total_token_capacity,
        int64_t target_frame_capacity,
        OmniVoiceGeneratorPerfMode perf_mode)
        : runtime_(std::move(runtime)),
          graph_arena_bytes_(graph_arena_bytes),
          perf_mode_(perf_mode) {
        rebuild(total_token_capacity, target_frame_capacity);
    }

    ~ForwardGraph() {
        clear_graph();
    }

    bool matches(
        const WeightsRuntime & runtime,
        int64_t total_tokens,
        int64_t target_frames) const override {
        return runtime_.get() == &runtime &&
            total_tokens_capacity_ >= total_tokens &&
            target_frame_capacity_ >= target_frames;
    }

    void rebuild(int64_t total_token_capacity, int64_t target_frame_capacity) override {
        if (total_token_capacity <= 0 || target_frame_capacity <= 0) {
            throw std::runtime_error("OmniVoice generator total token capacity is invalid");
        }
        const auto clear_start = Clock::now();
        clear_graph();
        const auto clear_end = Clock::now();
        rebuild_clear_ms_ = engine::debug::elapsed_ms(clear_start, clear_end);
        total_tokens_capacity_ = total_token_capacity;
        target_frame_capacity_ = target_frame_capacity;

        const auto build_start = Clock::now();
        ggml_init_params input_params{16ull * 1024ull * 1024ull, nullptr, true};
        input_ctx_.reset(ggml_init(input_params));
        if (input_ctx_ == nullptr) {
            throw std::runtime_error("failed to initialize OmniVoice generator input context");
        }
        ggml_init_params params{graph_arena_bytes_, nullptr, true};
        ctx_.reset(ggml_init(params));
        if (ctx_ == nullptr) {
            throw std::runtime_error("failed to initialize OmniVoice generator graph context");
        }

        const auto & config = runtime_->assets().config;
        const auto & weights = runtime_->weights();