#pragma once

#include "engine/community_models/minimax_music3/ar_runtime.h"
#include "engine/community_models/minimax_music3/condition_encoder.h"
#include "engine/community_models/minimax_music3/flow_sampler.h"
#include "engine/community_models/minimax_music3/vocoder.h"
#include "engine/framework/core/execution_context.h"
#include "engine/framework/runtime/model.h"

#include <memory>
#include <vector>

namespace engine::models::minimax_music3 {

namespace detail {

void append_cropped_interleaved_audio(
    runtime::AudioBuffer & destination,
    const runtime::AudioBuffer & chunk,
    int64_t left_frames,
    int64_t right_frames);

}  // namespace detail

class MiniMaxMusic3PipelineRuntime {
public:
    MiniMaxMusic3PipelineRuntime(
        core::ExecutionContext & execution,
        std::shared_ptr<const MiniMaxMusic3Assets> assets,
        size_t graph_arena_bytes,
        size_t weight_context_bytes,
        assets::TensorStorageType storage_type,
        bool memory_saver,
        bool pipeline_overlap = false,
        MiniMaxMusic3PerfOptions perf = {});
    ~MiniMaxMusic3PipelineRuntime();

    runtime::AudioBuffer generate(const MiniMaxMusic3Request & request);
    // K independent takes of the same prompt sharing one batched AR pass;
    // take i uses seed take_seeds[i] end to end (AR, depth, flow noise).
    std::vector<runtime::AudioBuffer> generate_ensemble(
        const MiniMaxMusic3Request & request,
        const std::vector<uint64_t> & take_seeds);
    void release_runtime_graphs();

private:
    struct Impl;
    std::unique_ptr<Impl> impl_;
};

}  // namespace engine::models::minimax_music3