#ifndef SLEELA_AI_LARGE_LANGUAGE_MODELS_HPP #define SLEELA_AI_LARGE_LANGUAGE_MODELS_HPP #include "large_language_models.h" #include #include #include #include #include #include namespace sleela::ai { struct GenerationOptions { std::uint32_t max_new_tokens = 256; double temperature = 0.7; double top_p = 0.95; std::uint64_t seed = 0; }; class LanguageModel { public: explicit LanguageModel(const sllm_backend& backend); ~LanguageModel(); LanguageModel(const LanguageModel&) = delete; LanguageModel& operator=(const LanguageModel&) = delete; LanguageModel(LanguageModel&& other) noexcept; LanguageModel& operator=(LanguageModel&& other) noexcept; void load(const std::string& model_id); void unload() noexcept; bool is_loaded() const noexcept; std::string backend_name() const; /* Return false from on_chunk to request cancellation. */ void generate(const std::string& prompt, const GenerationOptions& options, const std::function& on_chunk); // Borrowed native handle for interop with the C reasoning layer. Ownership // stays with this object; do not destroy the returned handle. sllm_engine* native_handle() const noexcept { return handle_; } private: sllm_engine* handle_ = nullptr; static void throw_if_error(sllm_status status); }; } // namespace sleela::ai #endif