// generate.h — генерация текста с KV-кэшем #pragma once #include "checkpoint.hpp" #include #include namespace xt { struct GenConfig { int max_new_tokens = 200; float temperature = 0.8f; int top_k = 40; float top_p = 0.95f; int seed = -1; bool stream = true; int repeat_penalty = 1; }; std::string generate(const Params& p, const Tokenizer& tok, const std::string& prompt, const GenConfig& gc, std::vector* out_ids = nullptr, bool fast = false); std::vector generate_batch(const Params& p, const Tokenizer& tok, const std::vector& prompts, const GenConfig& gc); }