infer_engine.hpp

#pragma once

#include "../models/llama/llama_config.hpp"
#include "distributed/distributed.hpp"
#include "infinicore/tensor.hpp"
#include "rank_worker.hpp"
#include "../models/infinilm_model.hpp"

#include <vector>

namespace infinilm::engine {

class InferEngine {
public:
    struct Input {
        infinicore::Tensor input_ids;

        infinicore::Tensor position_ids;

        infinicore::Tensor cache_positions;

        infinilm::InfinilmModel::Input to_model_input() const;
    };

    struct Output {
        infinicore::Tensor logits;
    };

    // Updated constructor: accept CacheConfig instead of CacheType
    InferEngine(
        const InfinilmModel::Config &config,
        const distributed::DistConfig &distributed_config = distributed::DistConfig(),
        infinicore::Device::Type device_type = infinicore::context::getDevice().getType(),
        const cache::CacheConfig *cache_config = nullptr);

    // Load a parameter to all workers (each can extract its shard inside RankWorker)
    void load_param(const std::string &name, const infinicore::Tensor &param);

    // return the parameters (i.e. weights and biases).
    std::vector<std::unordered_map<std::string, infinicore::nn::Parameter>> state_dict();

    // Run a single forward pass on all workers and return the outputs from all ranks
    Output forward(const Input &input);

    void reset_cache(const cache::CacheConfig *new_config);

    ~InferEngine();

    const distributed::DistConfig &get_dist_config() const;

    // Get current KV configuration
    const cache::CacheConfig *get_cache_config() const { return cache_config_.get(); }

protected:
    std::vector<std::unique_ptr<RankWorker>> workers_;
    distributed::CommunicationGroup communication_group_;
    const InfinilmModel::Config &model_config_;
    std::unique_ptr<cache::CacheConfig> cache_config_;
};

} // namespace infinilm::engine