#ifndef SLEELA_AI_LLM_TRAINING_HPP #define SLEELA_AI_LLM_TRAINING_HPP #include "large_language_models_training.h" #include #include #include namespace sleela::ai { struct TrainingOptions { double learning_rate=0.0001; double gradient_clip_norm=1.0; std::uint32_t micro_batch_size=1; std::uint32_t accumulation_steps=1; std::uint64_t seed=0; }; struct TrainingMetrics { double loss=0; std::uint64_t tokens_processed=0; std::uint64_t optimizer_steps=0; }; class LanguageModelTrainer { public: explicit LanguageModelTrainer(const sllmt_backend& backend); ~LanguageModelTrainer(); LanguageModelTrainer(const LanguageModelTrainer&)=delete; LanguageModelTrainer& operator=(const LanguageModelTrainer&)=delete; LanguageModelTrainer(LanguageModelTrainer&&) noexcept; LanguageModelTrainer& operator=(LanguageModelTrainer&&) noexcept; void initialize(const std::string& model_id,const TrainingOptions& options={}); void shutdown() noexcept; bool is_ready() const noexcept; TrainingMetrics train_batch(const std::vector& input,const std::vector& target,const TrainingOptions& options={}); TrainingMetrics evaluate_batch(const std::vector& input,const std::vector& target); void save_checkpoint(const std::string& path); private: sllmt_trainer* handle_=nullptr; static void check(sllmt_status status); }; } #endif