#pragma once

#include "llama-arch.h"

#include "llama-batch.h"

#include "llama-hparams.h"

#include "llama-adapter.h"

#include <cstdint>

#include <vector>

#include <memory>

#include <set>

#include <functional>

#include <map>

struct ggml_cgraph;

struct ggml_context;

struct ggml_tensor;

struct llama_cparams;

struct llama_layer;

struct llama_memory_context_i;

class llama_kv_cache_context;

class llama_kv_cache_dsa_context;

class llama_kv_cache_dsv4_raw_context;

class llama_kv_cache_dsv4_context;

class llama_kv_cache_iswa_context;

class llama_memory_recurrent_context;

class llama_memory_hybrid_context;

class llama_memory_hybrid_iswa_context;

// certain models (typically multi-modal) can produce different types of graphs

enum llm_graph_type {

    LLM_GRAPH_TYPE_DEFAULT,

    LLM_GRAPH_TYPE_ENCODER,

    LLM_GRAPH_TYPE_DECODER,

    LLM_GRAPH_TYPE_DECODER_MTP,

};

enum llm_fused_op {

    LLM_FUSED_OP_FLASH_ATTN,

    LLM_FUSED_OP_GDN_AR,

    LLM_FUSED_OP_GDN_CH,

    LLM_FUSED_OP_LIGHTNING_INDEXER,

};

enum llm_ffn_op_type : int {

    LLM_FFN_NONE = 0,           // sentinel: unset; archs must assign before use

    LLM_FFN_SILU,

    LLM_FFN_GELU,

    LLM_FFN_RELU,

    LLM_FFN_RELU_SQR,

    LLM_FFN_SWIGLU,

    LLM_FFN_GEGLU,

    LLM_FFN_REGLU,

    LLM_FFN_SWIGLU_OAI_MOE,

};

enum llm_ffn_gate_type {

    LLM_FFN_SEQ,

    LLM_FFN_PAR, // ffn_gate is parallel to ffn_up

};

enum llm_norm_type {

    LLM_NORM,

    LLM_NORM_RMS,

    LLM_NORM_GROUP,

};

// TODO: tmp - need something better to pass the data from the encoder to the decoder

struct llama_cross {

    // the output embeddings from the encoder as a ggml tensor

    // TODO: this needs more work to be correct, for now copy the embeddings data to host memory

    //       ref: https://github.com/ggml-org/llama.cpp/pull/11213#discussion_r1969892524

    //ggml_tensor * t_embd = nullptr;

    int64_t n_embd = 0;

    int64_t n_enc  = 0;

    // embeddings data copied to host memory (tmp)

    std::vector<float> v_embd;

    // needed to construct the cross-attention mask in the decoder

    std::vector<std::set<llama_seq_id>> seq_ids_enc;

};

struct llm_graph_params;

//

// llm_graph_input

//

class llm_graph_input_i {

public:

    llm_graph_input_i() {

        const char * LLAMA_GRAPH_INPUT_DEBUG = getenv("LLAMA_GRAPH_INPUT_DEBUG");

        debug = LLAMA_GRAPH_INPUT_DEBUG ? atoi(LLAMA_GRAPH_INPUT_DEBUG) : 0;

    }

    virtual ~llm_graph_input_i() = default;

    virtual void set_input(const llama_ubatch * ubatch) = 0;

    // return true if the resulting input tensors using the provided graph parameters would be

    //   the same as the previous input tensors that we have currently stored in the object

    virtual bool can_reuse(const llm_graph_params & params) {

        // returning false here by default will prevent from reusing the graph if the check

        //   for the input type has not been implemented yet

        GGML_UNUSED(params);

        return false;

    }

protected:

    // env: LLAMA_GRAPH_INPUT_DEBUG

    int debug = 0;

};

using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;

class llm_graph_input_embd : public llm_graph_input_i {

public:

    llm_graph_input_embd(int64_t n_embd) : n_embd(n_embd) {}

    virtual ~llm_graph_input_embd() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * tokens = nullptr; // I32 [n_batch]

    ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]

    const int64_t n_embd = 0;

};

// similar to llm_graph_input_embd but with an additional hidden state input

class llm_graph_input_embd_h : public llm_graph_input_i {

public:

    llm_graph_input_embd_h(int64_t n_embd) : n_embd(n_embd) {}

    virtual ~llm_graph_input_embd_h() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * tokens = nullptr; // I32 [n_batch]

    ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]

    ggml_tensor * h      = nullptr; // F32 [n_embd, n_batch]

    const int64_t n_embd = 0;

};

class llm_graph_input_pos : public llm_graph_input_i {

public:

    llm_graph_input_pos(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}

    virtual ~llm_graph_input_pos() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * pos = nullptr; // I32 [n_batch]

    const uint32_t n_pos_per_embd = 1;

};

// temperature tuning, used by llama4

class llm_graph_input_attn_temp : public llm_graph_input_i {

public:

    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale, float f_attn_temp_offset)

        : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale), f_attn_temp_offset(f_attn_temp_offset) {}

    virtual ~llm_graph_input_attn_temp() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * attn_scale = nullptr; // F32 [n_batch]

    const uint32_t n_attn_temp_floor_scale;

    const float    f_attn_temp_scale;

    const float    f_attn_temp_offset;

};

class llm_graph_input_pos_bucket : public llm_graph_input_i {

public:

    llm_graph_input_pos_bucket(const llama_hparams & hparams) : hparams(hparams) {}

    virtual ~llm_graph_input_pos_bucket() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch]

    const llama_hparams hparams;

};

class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {

public:

    llm_graph_input_pos_bucket_kv(

            const llama_hparams & hparams,

            const llama_kv_cache_context * mctx) : hparams(hparams), mctx(mctx) {}

    virtual ~llm_graph_input_pos_bucket_kv() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]

    const llama_hparams hparams;

    const llama_kv_cache_context * mctx;

};

class llm_graph_input_out_ids : public llm_graph_input_i {

public:

    llm_graph_input_out_ids(

            const llama_hparams & hparams,

            const llama_cparams & cparams,

            uint32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}

    virtual ~llm_graph_input_out_ids() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * out_ids; // I32 [n_outputs]

    const llama_hparams hparams;

    const llama_cparams cparams;

    const uint32_t n_outputs;

};

class llm_graph_input_mean : public llm_graph_input_i {

public:

    llm_graph_input_mean(const llama_cparams & cparams) : cparams(cparams) {}

    virtual ~llm_graph_input_mean() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * mean; // F32 [n_batch, n_batch]

    const llama_cparams cparams;

};

class llm_graph_input_cls : public llm_graph_input_i {

public:

    llm_graph_input_cls(const llama_cparams & cparams, const llm_arch arch) : cparams(cparams), arch(arch) {}

    virtual ~llm_graph_input_cls() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * cls; // I32 [n_batch]

    const llama_cparams cparams;

    const llm_arch arch;

};

class llm_graph_input_rs : public llm_graph_input_i {

public:

    llm_graph_input_rs(const llama_memory_recurrent_context * mctx) : mctx(mctx) {}

    virtual ~llm_graph_input_rs() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * s_copy;  // I32 [n_rs]

    // views of s_copy, computed once per graph

    // and shared across layers which use build_rs

    ggml_tensor * s_copy_main;   // I32 [n_seqs]

    ggml_tensor * s_copy_extra;  // I32 [n_rs - n_seqs]

    const llama_memory_recurrent_context * mctx;

    // used in view offsets, need to match for valid graph reuse

    uint32_t head;

    int32_t rs_z;

};

class llm_graph_input_cross_embd : public llm_graph_input_i {

public:

    llm_graph_input_cross_embd(

            const llama_cross * cross) : cross(cross) {}

    virtual ~llm_graph_input_cross_embd() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * cross_embd; // F32 [n_embd, n_outputs_enc]

    const llama_cross * cross;

};

class llm_graph_input_attn_no_cache : public llm_graph_input_i {

public:

    llm_graph_input_attn_no_cache(const llama_hparams & hparams, const llama_cparams & cparams) :

        hparams(hparams),

        cparams(cparams) {

    }

    ~llm_graph_input_attn_no_cache() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }

    ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }

    // n_tokens == n_batch

    ggml_tensor * self_kq_mask         = nullptr; // F32/F16 [n_tokens, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_cnv     = nullptr; //         [n_tokens, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_swa     = nullptr; // F32/F16 [n_tokens, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //         [n_tokens, n_batch/n_stream, 1, n_stream]

    const llama_hparams hparams;

    const llama_cparams cparams;

};

class llm_graph_input_attn_kv : public llm_graph_input_i {

public:

    llm_graph_input_attn_kv(

            const llama_hparams & hparams,

            const llama_cparams & cparams,

            const llama_kv_cache_context * mctx) :

        hparams(hparams),

        cparams(cparams),

        mctx(mctx) {

    }

    ~llm_graph_input_attn_kv() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * get_k_idxs() const { return self_k_idxs; }

    ggml_tensor * get_v_idxs() const { return self_v_idxs; }

    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }

    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]

    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]

    ggml_tensor * self_kq_mask     = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]

    // note: assumes v_rot^2 == I

    ggml_tensor * self_k_rot = nullptr;

    ggml_tensor * self_v_rot = nullptr;

    // note: these have to be copies because in order to be able to reuse a graph, its inputs

    //       need to carry these parameters with them. otherwise, they can point to freed

    //       llm_graph_params from a previous batch, causing stack-use-after-return

    const llama_hparams hparams;

    const llama_cparams cparams;

    const llama_kv_cache_context * mctx;

};

// V-less input for the KV cache

// ref: https://github.com/ggml-org/llama.cpp/pull/19067

class llm_graph_input_attn_k : public llm_graph_input_i {

public:

    llm_graph_input_attn_k(

            const llama_hparams & hparams,

            const llama_cparams & cparams,

            const llama_kv_cache_context * mctx) :

        hparams(hparams),

        cparams(cparams),

        mctx(mctx) {

    }

    ~llm_graph_input_attn_k() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * get_k_idxs() const { return self_k_idxs; }

    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }

    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]

    ggml_tensor * self_kq_mask     = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]

    const llama_hparams hparams;

    const llama_cparams cparams;

    const llama_kv_cache_context * mctx;

};

class llm_graph_input_attn_k_dsa : public llm_graph_input_i {

public:

    llm_graph_input_attn_k_dsa(

            const llama_hparams & hparams,

            const llama_cparams & cparams,

            const llama_kv_cache_dsa_context * mctx) :

        hparams(hparams),

        cparams(cparams),

        mctx(mctx) {

    }

    ~llm_graph_input_attn_k_dsa() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * get_k_idxs_mla() const { return self_k_idxs_mla; }

    ggml_tensor * get_k_idxs_lid() const { return self_k_idxs_lid; }

    ggml_tensor * get_kq_mask_mla() const { return self_kq_mask_mla_cnv; }

    ggml_tensor * get_kq_mask_lid() const { return self_kq_mask_lid; }

    ggml_tensor * self_k_idxs_mla = nullptr; // I64 [n_batch]

    ggml_tensor * self_k_idxs_lid = nullptr; // I64 [n_batch]

    ggml_tensor * self_kq_mask_mla     = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_mla_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_lid     = nullptr; // F32     [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_lid_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_k_rot_lid = nullptr;

    const llama_hparams hparams;

    const llama_cparams cparams;

    const llama_kv_cache_dsa_context * mctx;

};

class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {

public:

    llm_graph_input_attn_kv_iswa(

            const llama_hparams & hparams,

            const llama_cparams & cparams,

            const llama_kv_cache_iswa_context * mctx) :

        hparams(hparams),

        cparams(cparams),

        mctx(mctx) {

    }

    ~llm_graph_input_attn_kv_iswa() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    ggml_tensor * get_k_idxs()     const { return self_k_idxs; }

    ggml_tensor * get_v_idxs()     const { return self_v_idxs; }

    ggml_tensor * get_k_idxs_swa() const { return self_k_idxs_swa; }

    ggml_tensor * get_v_idxs_swa() const { return self_v_idxs_swa; }

    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }

    ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }

    ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]

    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]

    ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]

    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]

    ggml_tensor * self_kq_mask         = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_cnv     = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_swa     = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_k_rot = nullptr;

    ggml_tensor * self_v_rot = nullptr;

    ggml_tensor * self_k_rot_swa = nullptr;

    ggml_tensor * self_v_rot_swa = nullptr;

    const llama_hparams hparams;

    const llama_cparams cparams;

    const llama_kv_cache_iswa_context * mctx;

};

// DSV4 raw graph inputs are SWA-only, but their mask may be stream-shaped

// so raw K can be concatenated with DSV4 compressed K in one attention op.

class llm_graph_input_dsv4_raw {

public:

    llm_graph_input_dsv4_raw(

            const llama_cparams & cparams,

            const llama_kv_cache_dsv4_raw_context * mctx) :

        cparams(cparams),

        mctx(mctx) {

    }

    void set_input(const llama_ubatch * ubatch);

    ggml_tensor * get_k_idxs() const { return self_k_idxs; }

    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }

    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]

    ggml_tensor * self_kq_mask     = nullptr; // F32/F16 [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_kq_mask_cnv = nullptr; //         [n_kv, n_batch/n_stream, 1, n_stream]

    ggml_tensor * self_k_rot = nullptr;

    const llama_cparams cparams;

    const llama_kv_cache_dsv4_raw_context * mctx;

};

class llm_graph_input_dsv4 : public llm_graph_input_i {

public:

    struct comp_input {

        ggml_tensor * state_pos        = nullptr; // I32 [n_state]

        ggml_tensor * state_persist_src_idxs = nullptr; // I32 [n_state_persist]

        ggml_tensor * state_persist_dst_idxs = nullptr; // I32 [n_state_persist]

        ggml_tensor * state_read_idxs  = nullptr; // I32 [ratio*n_state_write]

        ggml_tensor * state_write_idxs = nullptr; // I64 [n_state_write]

        ggml_tensor * state_write_pos  = nullptr; // I32 [n_state_write]

        ggml_tensor * kq_mask    = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]

        ggml_tensor * k_rot      = nullptr;

    };

    llm_graph_input_dsv4(

            const llama_cparams & cparams,

            std::unique_ptr<llm_graph_input_dsv4_raw> inp_raw,

            const llama_kv_cache_dsv4_context * mctx) :

        inp_raw(std::move(inp_raw)),

        cparams(cparams),

        mctx(mctx) {

    }

    ~llm_graph_input_dsv4() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    llm_graph_input_dsv4_raw * get_raw() const { return inp_raw.get(); }

    const comp_input & get_csa() const { return inp_csa; }

    const comp_input & get_hca() const { return inp_hca; }

    const comp_input & get_lid() const { return inp_lid; }

    std::unique_ptr<llm_graph_input_dsv4_raw> inp_raw;

    comp_input inp_csa;

    comp_input inp_hca;

    comp_input inp_lid;

    const llama_cparams cparams;

    const llama_kv_cache_dsv4_context * mctx;

};

class llm_graph_input_attn_cross : public llm_graph_input_i {

public:

    llm_graph_input_attn_cross(const llama_cross * cross) : cross(cross) {}

    ~llm_graph_input_attn_cross() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; }

    ggml_tensor * cross_kq_mask     = nullptr; // F32/F16 [n_outputs_enc, n_batch, 1, 1]

    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32/F16 [n_outputs_enc, n_batch, 1, 1]

    const llama_cross * cross = nullptr;

};

class llm_graph_input_mem_hybrid : public llm_graph_input_i {

public:

    llm_graph_input_mem_hybrid(

            const llama_cparams & cparams,

            std::unique_ptr<llm_graph_input_attn_kv> inp_attn,

            std::unique_ptr<llm_graph_input_rs>      inp_rs,

            const llama_memory_hybrid_context *      mctx) :

        inp_attn(std::move(inp_attn)),

        inp_rs(std::move(inp_rs)),

        cparams(cparams),

        mctx(mctx) { }

    virtual ~llm_graph_input_mem_hybrid() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    std::unique_ptr<llm_graph_input_attn_kv> inp_attn;

    std::unique_ptr<llm_graph_input_rs>      inp_rs;

    llm_graph_input_attn_kv * get_attn() const { return inp_attn.get(); }

    llm_graph_input_rs      * get_recr() const { return inp_rs.get(); }

    const llama_cparams cparams;

    const llama_memory_hybrid_context * mctx;

};

class llm_graph_input_mem_hybrid_k : public llm_graph_input_i {

public:

    llm_graph_input_mem_hybrid_k(

            const llama_cparams & cparams,

            std::unique_ptr<llm_graph_input_attn_k> inp_attn,

            std::unique_ptr<llm_graph_input_rs>      inp_rs,

            const llama_memory_hybrid_context *      mctx) :

        inp_attn(std::move(inp_attn)),

        inp_rs(std::move(inp_rs)),

        cparams(cparams),

        mctx(mctx) { }

    virtual ~llm_graph_input_mem_hybrid_k() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    std::unique_ptr<llm_graph_input_attn_k> inp_attn;

    std::unique_ptr<llm_graph_input_rs>      inp_rs;

    llm_graph_input_attn_k * get_attn() const { return inp_attn.get(); }

    llm_graph_input_rs      * get_recr() const { return inp_rs.get(); }

    const llama_cparams cparams;

    const llama_memory_hybrid_context * mctx;

};

class llm_graph_input_mem_hybrid_iswa : public llm_graph_input_i {

public:

    llm_graph_input_mem_hybrid_iswa(

            const llama_cparams & cparams,

            std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn,

            std::unique_ptr<llm_graph_input_rs>          inp_rs,

            const llama_memory_hybrid_iswa_context *     mctx) :

        inp_attn(std::move(inp_attn)),

        inp_rs(std::move(inp_rs)),

        cparams(cparams),

        mctx(mctx) { }

    virtual ~llm_graph_input_mem_hybrid_iswa() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn;

    std::unique_ptr<llm_graph_input_rs>          inp_rs;

    llm_graph_input_attn_kv_iswa * get_attn() const { return inp_attn.get(); }

    llm_graph_input_rs           * get_recr() const { return inp_rs.get(); }

    const llama_cparams cparams;

    const llama_memory_hybrid_iswa_context * mctx;

};

class llm_graph_input_sampling : public llm_graph_input_i {

public:

    llm_graph_input_sampling(std::map<llama_seq_id, llama_sampler *> samplers) :

        samplers(std::move(samplers)) { }

    virtual ~llm_graph_input_sampling() = default;

    void set_input(const llama_ubatch * ubatch) override;

    bool can_reuse(const llm_graph_params & params) override;

    std::map<llama_seq_id, llama_sampler *> samplers;

};

//

// llm_graph_result

//

// these objects deliver the result from the graph build process back to the llama_context

// note that the input tensors created for the graph are referenced here - the goal is to be able to populate their

//   specific data, by calling the set_inputs() method

// along with the input tensors, the object also provides commonly used outputs tensors, such as logits, embeddings, etc.

//   these are used by the llama_context to extact the relevant data, based on the compute parameters

// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)

using llm_graph_cb = std::function<void(const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il)>;

class llm_graph_result;

struct llm_graph_params {

    llm_arch arch = LLM_ARCH_UNKNOWN;

    llama_hparams hparams;

    llama_cparams cparams;

    llama_ubatch ubatch; // note: intentionally make a copy

    llm_graph_type gtype;

    ggml_backend_sched_t sched;

    ggml_backend_t backend_cpu;

    const llama_adapter_cvec     * cvec;

    const llama_adapter_loras    * loras;

    const llama_memory_context_i * mctx;

    const llama_cross            * cross;

    std::map<llama_seq_id, llama_sampler *> samplers;

    static bool samplers_equal(

          const std::map<llama_seq_id, llama_sampler *> & lhs,

          const std::map<llama_seq_id, llama_sampler *> & rhs) {

        if (lhs.size() != rhs.size()) {

            return false;

        }

        for (const auto & [seq_id, sampler] : lhs) {

            auto it = rhs.find(seq_id);

            if (it == rhs.end() || it->second != sampler) {

                return false;

            }

        }

        return true;

    }

    uint32_t n_outputs;

    llm_graph_cb cb;

    llm_graph_result * res;

    // return true if the "other" params would result in a graph with the same topology as with the current params

    //   having the same topology allows us to reuse the graph in some cases

    bool allow_reuse(const llm_graph_params & other) const {

        // first check the ubatch

        bool can_reuse_ubatch =

            ubatch.equal_seqs() == other.ubatch.equal_seqs() &&

            ubatch.n_tokens     == other.ubatch.n_tokens &&

            ubatch.n_seq_tokens == other.ubatch.n_seq_tokens &&

            ubatch.n_seqs       == other.ubatch.n_seqs &&

            ubatch.n_seqs_unq   == other.ubatch.n_seqs_unq &&

            (

                (!ubatch.token && !other.ubatch.token) ||

                (!ubatch.embd  && !other.ubatch.embd)  ||

                (ubatch.token && other.ubatch.token && ubatch.embd && other.ubatch.embd)

            );

        // when we split the batch using "equal_seqs" we have to verify that the participating sequences are the same

        //   the reason is because the set of attention streams would be different for different sequences

        if (can_reuse_ubatch && ubatch.equal_seqs()) {

            if (!ubatch.data) {

                // if the old ubatch does not own it's data, then we cannot guarantee that it is still alive, and

                //   therefore we cannot perform the sequence id check. normally should never happen

                can_reuse_ubatch = false;

            } else {

                for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {

                    can_reuse_ubatch &= ubatch.seq_id_unq[s] == other.ubatch.seq_id_unq[s];

                }

            }

        }

        if (!can_reuse_ubatch) {

            return false;

        }

        if (n_outputs != other.n_outputs) {

            return false;

        }

        if (!samplers_equal(samplers, other.samplers)) {

            return false;

        }

        if (samplers.size() > 0) {

            if (!ubatch.data || !other.ubatch.data) {

                return false;

            }

            // check that the outputs are the same for all samplers

            for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {

                if (ubatch.output[i]    != other.ubatch.output[i] ||

                    ubatch.seq_id[i][0] != other.ubatch.seq_id[i][0]) {

                    return false;

                }

            }

        }

        // TODO: https://github.com/ggml-org/llama.cpp/pull/24340#discussion_r3448035248

        if (cparams.nextn_layer_offset != other.cparams.nextn_layer_offset) {

            return false;

        }

        return

            cparams.embeddings              == other.cparams.embeddings              &&

            cparams.embeddings_nextn        == other.cparams.embeddings_nextn        &&

            cparams.embeddings_nextn_masked == other.cparams.embeddings_nextn_masked &&

            cparams.causal_attn             == other.cparams.causal_attn             &&

            arch  == other.arch  &&

            gtype == other.gtype &&

            cvec  == other.cvec  &&

            loras == other.loras &&

            cross == other.cross;

    }

};

struct llm_graph_fused_node {

    llm_fused_op op;

    ggml_tensor * tensor;

    int il;

};

class llm_graph_result {

public:

    llm_graph_result(int64_t max_nodes);

    virtual ~llm_graph_result() = default;

    ggml_tensor * get_inp_tokens()  const { return t_inp_tokens; }

    ggml_tensor * get_logits()      const { return t_logits; }

    ggml_tensor * get_embd()        const { return t_embd; }

    ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }

    ggml_tensor * get_h_nextn()     const { return t_h_nextn; }

    ggml_tensor * get_layer_inp(int il) const { return t_layer_inp[il]; }

    ggml_cgraph  * get_gf()  const { return gf; }

    ggml_context * get_ctx() const { return ctx_compute.get(); }

    int64_t get_max_nodes() const;

    void reset();

    void set_inputs(const llama_ubatch * ubatch);

    void set_outputs(const llm_graph_params & params);

    // try to update the existing graph result using the new graph parameters in order to reuse it

    // this can only be done if we determine that the resulting graph using the new graph parameters

    //   would be identical to the existing graph. in that case, we simply have to update the memory

    //   contexts of the input tensors of the graph and we can reuse it for another computation

    // return true if the graph was updated and can be reused

    bool can_reuse(const llm_graph_params & params);

    llm_graph_input_i * add_input(llm_graph_input_ptr input);

    void add_fused_node(llm_graph_fused_node result);

    const std::vector<llm_graph_fused_node> & get_fused_nodes() const { return fused_nodes; }

    void set_params(const llm_graph_params & params);

    // important graph nodes

    ggml_tensor * t_inp_tokens  = nullptr;

    ggml_tensor * t_inp_embd    = nullptr; // [n_embd_inp, n_tokens]

    ggml_tensor * t_logits      = nullptr;

    ggml_tensor * t_embd        = nullptr;

    ggml_tensor * t_embd_pooled = nullptr;

    ggml_tensor * t_h_nextn     = nullptr; // [n_embd, n_outputs] hidden state before final output norm

    std::vector<ggml_tensor *> t_layer_inp;

    std::map<llama_seq_id, ggml_tensor *> t_sampled_logits;

    std::map<llama_seq_id, ggml_tensor *> t_candidates;

    std::map<llama_seq_id, ggml_tensor *> t_sampled;

    std::map<llama_seq_id, ggml_tensor *> t_sampled_probs;

    std::vector<llm_graph_input_ptr> inputs;

    std::vector<llm_graph_fused_node> fused_nodes;

    ggml_context_ptr ctx_compute;

    // memory buffers used to evaluate the model

    std::vector<uint8_t> buf_compute_meta;

    ggml_cgraph * gf;

    int64_t max_nodes;

private:

    // keep a copy of the previous graph parameters

    // we will use this to determine whether the graph can be reused by comparing them with the new parameters

    // note: these are updated after constructing the new graph

    llm_graph_params params;

    // env: LLAMA_GRAPH_RESULT_DEBUG

    int debug = 0;

};

using llm_graph_result_ptr = std::unique_ptr<llm_graph_result>;

//

// llm_graph_context

//

// used in build_rs to properly order writes and avoid unnecessary copies

using llm_graph_get_rows_fn = std::function<ggml_tensor * (ggml_context *, ggml_tensor * states, ggml_tensor * ids)>;

struct llm_graph_qkv {

    ggml_tensor * q; // [n_embd_head, n_head,    n_tokens]

    ggml_tensor * k; // [n_embd_head, n_head_kv, n_tokens]

    ggml_tensor * v; // [n_embd_head, n_head_kv, n_tokens]

};

struct llm_graph_context {

    const llm_arch arch;

    const llama_hparams & hparams;

    const llama_cparams & cparams;

    const llama_ubatch  & ubatch;

    const int64_t n_embd;

    const int64_t n_layer;

    const int64_t n_layer_nextn;

    const int64_t n_rot;

    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)

    const int64_t n_head;

    const int64_t n_head_kv;

    const int64_t n_embd_head_k;

    const int64_t n_embd_k_gqa;

    const int64_t n_embd_head_v;

    const int64_t n_embd_v_gqa;

    const int64_t n_expert;

    const int64_t n_expert_used;

    const float freq_base;

    const float freq_scale;

    const float ext_factor;

    const float attn_factor;

    const float beta_fast;

    const float beta_slow;

    const float norm_eps;

    const float norm_rms_eps;

    const int64_t n_tokens;

    const int64_t n_outputs;

    const int32_t n_ctx_orig; // yarn

    const enum llama_pooling_type pooling_type;

    const enum llama_rope_type    rope_type;

    ggml_backend_sched_t sched;

    ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?

    const llama_adapter_cvec     * cvec;

    const llama_adapter_loras    * loras;

    const llama_memory_context_i * mctx;

    const llama_cross            * cross;

    std::map<llama_seq_id, llama_sampler *> samplers;

    const llm_graph_cb & cb_func;

    llm_graph_result * res;

    ggml_context * ctx0 = nullptr;

    ggml_cgraph  * gf   = nullptr;

    llm_graph_context(const llm_graph_params & params);

    virtual ~llm_graph_context() = default;

    void cb(ggml_tensor * cur, const char * name, int il) const;

    //

    // common

    //

    ggml_tensor * build_cvec(

             ggml_tensor * cur,

                     int   il) const;

    // do mat_mul, while optionally apply lora and per-tensor scale

    ggml_tensor * build_lora_mm(

              ggml_tensor * w,

              ggml_tensor * cur,

              ggml_tensor * w_s = nullptr) const;

    // do mat_mul_id, while optionally apply lora and per-expert scale

    ggml_tensor * build_lora_mm_id(

              ggml_tensor * w,   // ggml_tensor * as

              ggml_tensor * cur, // ggml_tensor * b

              ggml_tensor * ids,

              ggml_tensor * w_s = nullptr) const;

    ggml_tensor * build_norm(

             ggml_tensor * cur,

             ggml_tensor * mw,

             ggml_tensor * mb,

           llm_norm_type   type,

                     int   il) const;

    // compute Q, K, V projections with optional bias and reshape

    // supports both fused wqkv and separate wq/wk/wv paths

    llm_graph_qkv build_qkv(

        const llama_layer & layer,

              ggml_tensor * cur,

                  int64_t   n_embd_head,

                  int64_t   n_head,

                  int64_t   n_head_kv,

                      int   il) const;

    ggml_tensor * build_ffn(

             ggml_tensor * cur,

             ggml_tensor * up,

             ggml_tensor * up_b,

             ggml_tensor * up_s,

             ggml_tensor * gate,

             ggml_tensor * gate_b,

             ggml_tensor * gate_s,

             ggml_tensor * down,

             ggml_tensor * down_b,

             ggml_tensor * down_s,

             ggml_tensor * act_scales,

         llm_ffn_op_type   type_op,

       llm_ffn_gate_type   type_gate,

                     int   il) const;

    // build MoE FFN without bias tensors

    ggml_tensor * build_moe_ffn(

             ggml_tensor * cur,

             ggml_tensor * gate_inp,

             ggml_tensor * up_exps,

             ggml_tensor * gate_exps,

             ggml_tensor * down_exps,

             ggml_tensor * exp_probs_b,

                 int64_t   n_expert,

                 int64_t   n_expert_used,

         llm_ffn_op_type   type_op,

                    bool   norm_w,

                   float   w_scale,

            llama_expert_gating_func_type gating_op,

                     int   il,

             ggml_tensor * probs_in = nullptr,

             ggml_tensor * gate_up_exps = nullptr,

             ggml_tensor * up_exps_s = nullptr,

             ggml_tensor * gate_exps_s = nullptr,

             ggml_tensor * down_exps_s = nullptr,

             ggml_tensor * selected_experts_in = nullptr) const;

    ggml_tensor * build_moe_ffn(

             ggml_tensor * cur,

             ggml_tensor * gate_inp,

             ggml_tensor * gate_inp_b,

             ggml_tensor * up_exps,

             ggml_tensor * up_exps_b,

             ggml_tensor * gate_exps,

             ggml_tensor * gate_exps_b,

             ggml_tensor * down_exps,

             ggml_tensor * down_exps_b,

             ggml_tensor * exp_probs_b,

                 int64_t   n_expert,

                 int64_t   n_expert_used,

         llm_ffn_op_type   type_op,

                    bool   norm_w,

                   float   w_scale,

            llama_expert_gating_func_type gating_op,

                     int   il,

             ggml_tensor * probs_in = nullptr,

             ggml_tensor * gate_up_exps = nullptr,

             ggml_tensor * gate_up_exps_b = nullptr,

             ggml_tensor * up_exps_s = nullptr,

             ggml_tensor * gate_exps_s = nullptr,

             ggml_tensor * down_exps_s = nullptr,

             ggml_tensor * selected_experts_in = nullptr) const;

    //

    // inputs

    //

    ggml_tensor * build_inp_embd(ggml_tensor * tok_embd) const;

    ggml_tensor * build_inp_pos() const;

    ggml_tensor * build_inp_attn_scale() const;

    ggml_tensor * build_inp_out_ids() const;

    ggml_tensor * build_inp_mean() const;

    ggml_tensor * build_inp_cls() const;

    ggml_tensor * build_inp_cross_embd() const;

    ggml_tensor * build_inp_pos_bucket_enc() const;

    ggml_tensor * build_inp_pos_bucket_dec() const;

    ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const;

    //

    // attention

    //

    ggml_tensor * build_attn_mha(

            ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]

            ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]

            ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)