cpp/model_manager.h

#pragma once

#include <mutex>
#include <optional>
#include <string>
#include <utility>

#include "llm_manager_types.h"

class ModelManager {
public:
    explicit ModelManager(const ManagerConfig &config);

    bool initialize_default(std::string &error);
    bool switch_model(const std::string &model, std::string &error);
    bool restart_active(std::string &error);
    std::optional<WorkerInfo> active_worker();
    json models_view();

private:
    std::mutex mu_;
    std::optional<WorkerInfo> active_;
    bool switch_in_progress_ = false;

    std::string default_model_;
    std::string llama_server_bin_;
    std::string worker_host_;
    std::string worker_bind_host_;
    int base_port_;
    int switch_timeout_sec_;
    int n_ctx_;
    int n_threads_;
    int n_gpu_layers_;
    int n_batch_;
    int n_ubatch_;
    int next_port_;

    int allocate_port();
    void finish_switch(bool ok);
    pid_t spawn_worker(const std::string &model, int port);
    bool wait_until_ready(pid_t pid, int port, int timeout_sec);
    std::pair<int, std::string> http_get(int port, const std::string &target);
};

bool is_alive(pid_t pid);
void shutdown_worker(pid_t pid, int wait_seconds = 15);