whisper.cpp/whisper.cpp

#include "whisper.h"
#ifdef WHISPER_USE_COREML
#include "coreml/whisper-encoder.h"
#endif

#ifdef GGML_USE_METAL
#  include "ggml-metal.h"
#endif

#ifdef WHISPER_USE_OPENVINO
#include "openvino/whisper-openvino-encoder.h"
#endif

#include "ggml.h"
#include "ggml-alloc.h"

#include <algorithm>
#include <cassert>
#define _USE_MATH_DEFINES
#include <cmath>
#include <cstdio>
#include <cstdarg>
#include <cstring>
#include <fstream>
#include <map>
#include <set>
#include <string>
#include <thread>
#include <vector>
#include <regex>
#include <random>
#include <functional>

#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif

#if defined(GGML_BIG_ENDIAN)
#include <bit>

template<typename T>
static T byteswap(T value) {
    return std::byteswap(value);
}

template<>
float byteswap(float value) {
    return std::bit_cast<float>(byteswap(std::bit_cast<std::uint32_t>(value)));
}

template<typename T>
static void byteswap_tensor_data(ggml_tensor * tensor) {
    T * datum = reinterpret_cast<T *>(tensor->data);
    for (int i = 0; i < ggml_nelements(tensor); i++) {
        datum[i] = byteswap(datum[i]);
    }
}

static void byteswap_tensor(ggml_tensor * tensor) {
    switch (tensor->type) {
        case GGML_TYPE_I16: {
            byteswap_tensor_data<int16_t>(tensor);
            break;
        }
        case GGML_TYPE_F16: {
            byteswap_tensor_data<ggml_fp16_t>(tensor);
            break;
        }
        case GGML_TYPE_I32: {
            byteswap_tensor_data<int32_t>(tensor);
            break;
        }
        case GGML_TYPE_F32: {
            byteswap_tensor_data<float>(tensor);
            break;
        }
        default: { // GML_TYPE_I8
            break;
        }
    }
}

#define BYTESWAP_VALUE(d) d = byteswap(d)
#define BYTESWAP_FILTERS(f)            \
    do {                              \
        for (auto & datum : f.data) { \
            datum = byteswap(datum);  \
        }                             \
    } while (0)
#define BYTESWAP_TENSOR(t)       \
    do {                         \
        byteswap_tensor(t); \
    } while (0)
#else
#define BYTESWAP_VALUE(d) do {} while (0)
#define BYTESWAP_FILTERS(f) do {} while (0)
#define BYTESWAP_TENSOR(t) do {} while (0)
#endif

#define WHISPER_ASSERT(x) \
    do { \
        if (!(x)) { \
            log("WHISPER_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
            abort(); \
        } \
    } while (0)

// define this to enable verbose trace logging - useful for debugging purposes
//#define WHISPER_DEBUG

#if defined(WHISPER_DEBUG)
#define WHISPER_PRINT_DEBUG(...) \
    do { \
        fprintf(stderr, __VA_ARGS__); \
    } while (0)
#else
#define WHISPER_PRINT_DEBUG(...)
#endif

//#define WHISPER_USE_FLASH_ATTN
//#define WHISPER_USE_FLASH_FF
#define WHISPER_MAX_DECODERS 16

//
// ggml helpers
//

static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);

    if (plan.work_size > 0) {
        buf.resize(plan.work_size);
        plan.work_data = buf.data();
    }

    ggml_graph_compute(graph, &plan);
}

// faster matrix multiplications for tensors that do not have dimension 0 divisible by "pad"
// the idea is to represent the original matrix multiplication:
//
//   Z = X @ Y
//
// with the sum of two matrix multiplications:
//
//   Z = (X_0 @ Y_0) + (X_1 @ Y_1)
//
// here X_0 and Y_0 are views of X and Y that have dimension 0 divisible by "pad"
// and X_1 and Y_1 are the remaining views. X_1 and Y_1 end up being small matrices that can be processed with more
// general-purpose kernels
//
static struct ggml_tensor * ggml_mul_mat_pad(struct ggml_context * ctx, struct ggml_tensor * x, struct ggml_tensor * y, int pad = 32) {
    // use padding only if dimension 0 is at least 8 times larger than the padding
    // else we won't get much benefit from the optimization
    const int n_pad_req = 8;

    if (x->ne[0] % pad == 0 || x->ne[0] / pad < n_pad_req) {
        return ggml_mul_mat(ctx, x, y);
    }

    struct ggml_tensor * x_0 = ggml_view_3d(ctx, x, (x->ne[0]/pad)*pad, x->ne[1], x->ne[2], x->nb[1], x->nb[2], 0);
    struct ggml_tensor * x_1 = ggml_view_3d(ctx, x,  x->ne[0]%pad,      x->ne[1], x->ne[2], x->nb[1], x->nb[2], x_0->ne[0]*x_0->nb[0]);

    struct ggml_tensor * y_0 = ggml_view_3d(ctx, y, (y->ne[0]/pad)*pad, y->ne[1], y->ne[2], y->nb[1], y->nb[2], 0);
    struct ggml_tensor * y_1 = ggml_view_3d(ctx, y,  y->ne[0]%pad,      y->ne[1], y->ne[2], y->nb[1], y->nb[2], y_0->ne[0]*y_0->nb[0]);

    return ggml_add(ctx,
            ggml_mul_mat(ctx, x_0, y_0),
            ggml_mul_mat(ctx, x_1, y_1));
}

// TODO: check if other platforms can benefit from this optimization
#if defined(GGML_USE_METAL)
#define ggml_mul_mat ggml_mul_mat_pad
#endif

// available whisper models
enum e_model {
    MODEL_UNKNOWN,
    MODEL_TINY,
    MODEL_BASE,
    MODEL_SMALL,
    MODEL_MEDIUM,
    MODEL_LARGE,
};

static const std::map<std::string, std::pair<int, std::string>> g_lang = {
    { "en",  { 0,  "english",         } },
    { "zh",  { 1,  "chinese",         } },
    { "de",  { 2,  "german",          } },
    { "es",  { 3,  "spanish",         } },
    { "ru",  { 4,  "russian",         } },
    { "ko",  { 5,  "korean",          } },
    { "fr",  { 6,  "french",          } },
    { "ja",  { 7,  "japanese",        } },
    { "pt",  { 8,  "portuguese",      } },
    { "tr",  { 9,  "turkish",         } },
    { "pl",  { 10, "polish",          } },
    { "ca",  { 11,  "catalan",        } },
    { "nl",  { 12,  "dutch",          } },
    { "ar",  { 13,  "arabic",         } },
    { "sv",  { 14,  "swedish",        } },
    { "it",  { 15,  "italian",        } },
    { "id",  { 16,  "indonesian",     } },
    { "hi",  { 17,  "hindi",          } },
    { "fi",  { 18,  "finnish",        } },
    { "vi",  { 19,  "vietnamese",     } },
    { "he",  { 20,  "hebrew",         } },
    { "uk",  { 21,  "ukrainian",      } },
    { "el",  { 22,  "greek",          } },
    { "ms",  { 23,  "malay",          } },
    { "cs",  { 24,  "czech",          } },
    { "ro",  { 25,  "romanian",       } },
    { "da",  { 26,  "danish",         } },
    { "hu",  { 27,  "hungarian",      } },
    { "ta",  { 28,  "tamil",          } },
    { "no",  { 29,  "norwegian",      } },
    { "th",  { 30,  "thai",           } },
    { "ur",  { 31,  "urdu",           } },
    { "hr",  { 32,  "croatian",       } },
    { "bg",  { 33,  "bulgarian",      } },
    { "lt",  { 34,  "lithuanian",     } },
    { "la",  { 35,  "latin",          } },
    { "mi",  { 36,  "maori",          } },
    { "ml",  { 37,  "malayalam",      } },
    { "cy",  { 38,  "welsh",          } },
    { "sk",  { 39,  "slovak",         } },
    { "te",  { 40,  "telugu",         } },
    { "fa",  { 41,  "persian",        } },
    { "lv",  { 42,  "latvian",        } },
    { "bn",  { 43,  "bengali",        } },
    { "sr",  { 44,  "serbian",        } },
    { "az",  { 45,  "azerbaijani",    } },
    { "sl",  { 46,  "slovenian",      } },
    { "kn",  { 47,  "kannada",        } },
    { "et",  { 48,  "estonian",       } },
    { "mk",  { 49,  "macedonian",     } },
    { "br",  { 50,  "breton",         } },
    { "eu",  { 51,  "basque",         } },
    { "is",  { 52,  "icelandic",      } },
    { "hy",  { 53,  "armenian",       } },
    { "ne",  { 54,  "nepali",         } },
    { "mn",  { 55,  "mongolian",      } },
    { "bs",  { 56,  "bosnian",        } },
    { "kk",  { 57,  "kazakh",         } },
    { "sq",  { 58,  "albanian",       } },
    { "sw",  { 59,  "swahili",        } },
    { "gl",  { 60,  "galician",       } },
    { "mr",  { 61,  "marathi",        } },
    { "pa",  { 62,  "punjabi",        } },
    { "si",  { 63,  "sinhala",        } },
    { "km",  { 64,  "khmer",          } },
    { "sn",  { 65,  "shona",          } },
    { "yo",  { 66,  "yoruba",         } },
    { "so",  { 67,  "somali",         } },
    { "af",  { 68,  "afrikaans",      } },
    { "oc",  { 69,  "occitan",        } },
    { "ka",  { 70,  "georgian",       } },
    { "be",  { 71,  "belarusian",     } },
    { "tg",  { 72,  "tajik",          } },
    { "sd",  { 73,  "sindhi",         } },
    { "gu",  { 74,  "gujarati",       } },
    { "am",  { 75,  "amharic",        } },
    { "yi",  { 76,  "yiddish",        } },
    { "lo",  { 77,  "lao",            } },
    { "uz",  { 78,  "uzbek",          } },
    { "fo",  { 79,  "faroese",        } },
    { "ht",  { 80,  "haitian creole", } },
    { "ps",  { 81,  "pashto",         } },
    { "tk",  { 82,  "turkmen",        } },
    { "nn",  { 83,  "nynorsk",        } },
    { "mt",  { 84,  "maltese",        } },
    { "sa",  { 85,  "sanskrit",       } },
    { "lb",  { 86,  "luxembourgish",  } },
    { "my",  { 87,  "myanmar",        } },
    { "bo",  { 88,  "tibetan",        } },
    { "tl",  { 89,  "tagalog",        } },
    { "mg",  { 90,  "malagasy",       } },
    { "as",  { 91,  "assamese",       } },
    { "tt",  { 92,  "tatar",          } },
    { "haw", { 93,  "hawaiian",       } },
    { "ln",  { 94,  "lingala",        } },
    { "ha",  { 95,  "hausa",          } },
    { "ba",  { 96,  "bashkir",        } },
    { "jw",  { 97,  "javanese",       } },
    { "su",  { 98,  "sundanese",      } },
};

static const size_t MB = 1ull*1024*1024;

// TODO: avoid using GGUF
static const std::map<ggml_type, std::map<e_model, size_t>> MEM_REQ_MODEL = {
    { GGML_TYPE_F32,
        {
            { MODEL_TINY,     74ull*MB },
            { MODEL_BASE,    142ull*MB },
            { MODEL_SMALL,   466ull*MB },
            { MODEL_MEDIUM, 1464ull*MB },
            { MODEL_LARGE,  2952ull*MB },
        },
    },
    { GGML_TYPE_F16,
        {
            { MODEL_TINY,     74ull*MB },
            { MODEL_BASE,    142ull*MB },
            { MODEL_SMALL,   466ull*MB },
            { MODEL_MEDIUM, 1464ull*MB },
            { MODEL_LARGE,  2952ull*MB },
        },
    },
    { GGML_TYPE_Q4_0,
        {
            { MODEL_TINY,     26ull*MB },
            { MODEL_BASE,     50ull*MB },
            { MODEL_SMALL,   154ull*MB },
            { MODEL_MEDIUM,  470ull*MB },
            { MODEL_LARGE,   940ull*MB },
        },
    },
    { GGML_TYPE_Q4_1,
        {
            { MODEL_TINY,     32ull*MB },
            { MODEL_BASE,     58ull*MB },
            { MODEL_SMALL,   182ull*MB },
            { MODEL_MEDIUM,  562ull*MB },
            { MODEL_LARGE,  1124ull*MB },
        },
    },
    { GGML_TYPE_Q5_0,
        {
            { MODEL_TINY,     30ull*MB },
            { MODEL_BASE,     54ull*MB },
            { MODEL_SMALL,   170ull*MB },
            { MODEL_MEDIUM,  516ull*MB },
            { MODEL_LARGE,  1034ull*MB },
        },
    },
    { GGML_TYPE_Q5_1,
        {
            { MODEL_TINY,     32ull*MB },
            { MODEL_BASE,     58ull*MB },
            { MODEL_SMALL,   182ull*MB },
            { MODEL_MEDIUM,  562ull*MB },
            { MODEL_LARGE,  1124ull*MB },
        },
    },
    { GGML_TYPE_Q8_0,
        {
            { MODEL_TINY,     45ull*MB },
            { MODEL_BASE,     84ull*MB },
            { MODEL_SMALL,   268ull*MB },
            { MODEL_MEDIUM,  834ull*MB },
            { MODEL_LARGE,  1674ull*MB },
        },
    },
};

struct whisper_mel {
    int n_len;
    int n_len_org;
    int n_mel;

    std::vector<float> data;
};

struct whisper_filters {
    int32_t n_mel;
    int32_t n_fft;

    std::vector<float> data;
};

struct whisper_vocab {
    using id    = int32_t;
    using token = std::string;

    int n_vocab = 51864;

    std::map<token, id> token_to_id;
    std::map<id, token> id_to_token;

    // reference: https://github.com/openai/whisper/blob/248b6cb124225dd263bb9bd32d060b6517e067f8/whisper/tokenizer.py#L334-L349
    id token_eot        = 50256;
    id token_sot        = 50257;
    // task tokens (used only for multilingual models)
    id token_translate  = 50357;
    id token_transcribe = 50358;
    // other special tokens
    id token_solm       = 50359; // [TDRZ] used by tinydiarize models to indicate speaker turn
    id token_prev       = 50360;
    id token_nosp       = 50361;
    id token_not        = 50362; // no timestamps
    id token_beg        = 50363; // begin timestamps

    bool is_multilingual() const {
        return n_vocab == 51865;
    }
};

struct whisper_segment {
    int64_t t0;
    int64_t t1;

    std::string text;

    std::vector<whisper_token_data> tokens;

    bool speaker_turn_next;
};

// medium
// hparams: {
// 'n_mels': 80,
// 'n_vocab': 51864,
// 'n_audio_ctx': 1500,
// 'n_audio_state': 1024,
// 'n_audio_head': 16,
// 'n_audio_layer': 24,
// 'n_text_ctx': 448,
// 'n_text_state': 1024,
// 'n_text_head': 16,
// 'n_text_layer': 24
// }
//
// default hparams (Whisper tiny)
struct whisper_hparams {
    int32_t n_vocab       = 51864;
    int32_t n_audio_ctx   = 1500;
    int32_t n_audio_state = 384;
    int32_t n_audio_head  = 6;
    int32_t n_audio_layer = 4;
    int32_t n_text_ctx    = 448;
    int32_t n_text_state  = 384;
    int32_t n_text_head   = 6;
    int32_t n_text_layer  = 4;
    int32_t n_mels        = 80;
    int32_t ftype         = 1;
    float   eps           = 1e-5f;
};

// audio encoding layer
struct whisper_layer_encoder {
    // encoder.blocks.*.attn_ln
    struct ggml_tensor * attn_ln_0_w;
    struct ggml_tensor * attn_ln_0_b;

    // encoder.blocks.*.attn.out
    struct ggml_tensor * attn_ln_1_w;
    struct ggml_tensor * attn_ln_1_b;

    // encoder.blocks.*.attn.query
    struct ggml_tensor * attn_q_w;
    struct ggml_tensor * attn_q_b;

    // encoder.blocks.*.attn.key
    struct ggml_tensor * attn_k_w;

    // encoder.blocks.*.attn.value
    struct ggml_tensor * attn_v_w;
    struct ggml_tensor * attn_v_b;

    // encoder.blocks.*.mlp_ln
    struct ggml_tensor * mlp_ln_w;
    struct ggml_tensor * mlp_ln_b;

    // encoder.blocks.*.mlp.0
    struct ggml_tensor * mlp_0_w;
    struct ggml_tensor * mlp_0_b;

    // encoder.blocks.*.mlp.2
    struct ggml_tensor * mlp_1_w;
    struct ggml_tensor * mlp_1_b;
};

// token decoding layer
struct whisper_layer_decoder {
    // decoder.blocks.*.attn_ln
    struct ggml_tensor * attn_ln_0_w;
    struct ggml_tensor * attn_ln_0_b;

    // decoder.blocks.*.attn.out
    struct ggml_tensor * attn_ln_1_w;
    struct ggml_tensor * attn_ln_1_b;

    // decoder.blocks.*.attn.query
    struct ggml_tensor * attn_q_w;
    struct ggml_tensor * attn_q_b;

    // decoder.blocks.*.attn.key
    struct ggml_tensor * attn_k_w;

    // decoder.blocks.*.attn.value
    struct ggml_tensor * attn_v_w;
    struct ggml_tensor * attn_v_b;

    // decoder.blocks.*.cross_attn_ln
    struct ggml_tensor * cross_attn_ln_0_w;
    struct ggml_tensor * cross_attn_ln_0_b;

    // decoder.blocks.*.cross_attn.out
    struct ggml_tensor * cross_attn_ln_1_w;
    struct ggml_tensor * cross_attn_ln_1_b;

    // decoder.blocks.*.cross_attn.query
    struct ggml_tensor * cross_attn_q_w;
    struct ggml_tensor * cross_attn_q_b;

    // decoder.blocks.*.cross_attn.key
    struct ggml_tensor * cross_attn_k_w;

    // decoder.blocks.*.cross_attn.value
    struct ggml_tensor * cross_attn_v_w;
    struct ggml_tensor * cross_attn_v_b;

    // decoder.blocks.*.mlp_ln
    struct ggml_tensor * mlp_ln_w;
    struct ggml_tensor * mlp_ln_b;

    // decoder.blocks.*.mlp.0
    struct ggml_tensor * mlp_0_w;
    struct ggml_tensor * mlp_0_b;

    // decoder.blocks.*.mlp.2
    struct ggml_tensor * mlp_1_w;
    struct ggml_tensor * mlp_1_b;
};

struct whisper_kv_cache {
    struct ggml_tensor * k;
    struct ggml_tensor * v;

    struct ggml_context * ctx;

    // buf points to the memory allocated for both ggml_tensor 'k' and 'v' (see kv_cache_init)
    std::vector<uint8_t> buf;

    int n; // number of tokens currently in the cache
};

struct whisper_model {
    e_model type = MODEL_UNKNOWN;

    whisper_hparams hparams;
    whisper_filters filters;

    // encoder.positional_embedding
    struct ggml_tensor * e_pe;

    // encoder.conv1
    struct ggml_tensor * e_conv_1_w;
    struct ggml_tensor * e_conv_1_b;

    // encoder.conv2
    struct ggml_tensor * e_conv_2_w;
    struct ggml_tensor * e_conv_2_b;

    // encoder.ln_post
    struct ggml_tensor * e_ln_w;
    struct ggml_tensor * e_ln_b;

    // decoder.positional_embedding
    struct ggml_tensor * d_pe;

    // decoder.token_embedding
    struct ggml_tensor * d_te;

    // decoder.ln
    struct ggml_tensor * d_ln_w;
    struct ggml_tensor * d_ln_b;

    std::vector<whisper_layer_encoder> layers_encoder;
    std::vector<whisper_layer_decoder> layers_decoder;

    // context
    struct ggml_context * ctx;

    // the model memory buffer is read-only and can be shared between processors
    std::vector<uint8_t> * buf;

    // tensors
    int n_loaded;
    std::map<std::string, struct ggml_tensor *> tensors;
};

struct whisper_sequence {
    std::vector<whisper_token_data> tokens;

    // the accumulated transcription in the current iteration (used to truncate the tokens array)
    int result_len;

    double sum_logprobs_all; // the sum of the log probabilities of the tokens
    double sum_logprobs;     // the sum of the log probabilities of the tokens (first result_len tokens)
    double avg_logprobs;     // the average log probability of the tokens
    double entropy;          // the entropy of the tokens
    double score;            // likelihood rank score
};

// TAGS: WHISPER_DECODER_INIT
struct whisper_decoder {
    // each decoder keeps its own KV-cache
    whisper_kv_cache kv_self;

    // the currently generated sequence of tokens
    whisper_sequence sequence;

    int seek_delta; // the window shift found so far based on the decoded timestamp tokens

    bool failed;    // has the current segment failed to decode?
    bool completed; // has the decoder completed the current segment?
    bool has_ts;    // have we already sampled a non-beg timestamp token for the current segment?

    // new token probs, logits and logprobs after the last whisper_decode (1-dimensional array: [n_vocab])
    std::vector<float> probs;
    std::vector<float> logits;
    std::vector<float> logprobs;

    std::vector<whisper_token> tokens_tmp; // used for whisper_decode calls
};

// replace std::pair by using customized pair struct (reason: std::pair is very slow)
template<typename A, typename B>
struct whisper_pair {
    A first;
    B second;

    // Define a constructor that takes two arguments.
    whisper_pair(const A& a, const B& b) : first(a), second(b) {}
    // Define a constructor that takes no argument.
    whisper_pair() : first(A()), second(B()) {}
};

// beam-search helpers
struct kv_buf {
    std::vector<uint8_t> k;
    std::vector<uint8_t> v;
};

// ggml_allocr wrapper for whisper usage
struct whisper_allocr {
    ggml_allocr * alloc = nullptr;

    std::vector<uint8_t> meta;
    std::vector<uint8_t> data;
};

static size_t whisper_allocr_size(struct whisper_allocr & allocr) {
    return allocr.meta.size() + allocr.data.size();
}

// measure the memory usage of a graph and prepare the allocr's internal data buffer
static void whisper_allocr_graph_init(struct whisper_allocr & allocr, std::function<struct ggml_cgraph *()> && get_graph) {
    const int tensor_alignment = 32;

    auto & alloc = allocr.alloc;
    auto & meta  = allocr.meta;
    auto & data  = allocr.data;

    meta.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead());

    alloc = ggml_allocr_new_measure(tensor_alignment);

    const size_t alloc_size = ggml_allocr_alloc_graph(alloc, get_graph()) + tensor_alignment;

    ggml_allocr_free(alloc);

    data.resize(alloc_size);

    alloc = ggml_allocr_new(data.data(), data.size(), tensor_alignment);
}

static void whisper_allocr_free(struct whisper_allocr & allocr) {
    if (allocr.alloc) {
        ggml_allocr_free(allocr.alloc);
        allocr.alloc = nullptr;
    }
}

struct whisper_state {
    int64_t t_sample_us = 0;
    int64_t t_encode_us = 0;
    int64_t t_decode_us = 0;
    int64_t t_prompt_us = 0;
    int64_t t_mel_us = 0;

    int32_t n_sample = 0; // number of tokens sampled
    int32_t n_encode = 0; // number of encoder calls
    int32_t n_decode = 0; // number of decoder calls with n_tokens == 1 (text-generation)
    int32_t n_prompt = 0; // number of decoder calls with n_tokens >  1 (prompt encoding)
    int32_t n_fail_p = 0; // number of logprob threshold failures
    int32_t n_fail_h = 0; // number of entropy threshold failures

    // cross-attention KV cache for the decoders
    // shared between all decoders
    whisper_kv_cache kv_cross;
    whisper_mel mel;

    whisper_decoder decoders[WHISPER_MAX_DECODERS] = {};

    // buffer for swapping KV caches between decoders during beam-search
    std::vector<kv_buf> kv_swap_bufs;

    // reusable buffer for `struct ggml_graph_plan.work_data`
    std::vector<uint8_t> work_buffer;

    // ggml-alloc:
    // - stores meta info about the intermediate tensors into the `meta` buffers
    // - stores the actual tensor data into the `data` buffers
    whisper_allocr alloc_conv;
    whisper_allocr alloc_encode;
    whisper_allocr alloc_cross;
    whisper_allocr alloc_decode;

    // result of the encoder
    struct ggml_tensor * embd_conv = nullptr;
    struct ggml_tensor * embd_enc  = nullptr;

    // decode output (2-dimensional array: [n_tokens][n_vocab])
    std::vector<float> logits;

    std::vector<whisper_segment> result_all;
    std::vector<whisper_token>   prompt_past;

    // work container used to avoid memory allocations
    std::vector<whisper_pair<double, whisper_vocab::id>> logits_id;

    mutable std::mt19937 rng; // used for sampling at t > 0.0

    int lang_id = 0; // english by default

    std::string path_model; // populated by whisper_init_from_file()
#ifdef WHISPER_USE_COREML
    whisper_coreml_context * ctx_coreml = nullptr;
#endif

#ifdef GGML_USE_METAL
    ggml_metal_context * ctx_metal = nullptr;
#endif

#ifdef WHISPER_USE_OPENVINO
    whisper_openvino_context * ctx_openvino = nullptr;
#endif

    // [EXPERIMENTAL] token-level timestamps data
    int64_t t_beg = 0;
    int64_t t_last = 0;
    whisper_token tid_last;
    std::vector<float> energy; // PCM signal energy

    // [EXPERIMENTAL] speed-up techniques
    int32_t exp_n_audio_ctx = 0; // 0 - use default
};

struct whisper_context {
    int64_t t_load_us  = 0;
    int64_t t_start_us = 0;

    ggml_type wtype = ggml_type::GGML_TYPE_F16; // weight type (FP32 / FP16 / QX)
    ggml_type itype = ggml_type::GGML_TYPE_F16; // intermediate type (FP32 or FP16)

    whisper_model model;
    whisper_vocab vocab;
    whisper_state * state = nullptr;

    std::string path_model; // populated by whisper_init_from_file()
};

static void whisper_default_log(const char * text) {
    fprintf(stderr, "%s", text);
}

static whisper_log_callback whisper_log = whisper_default_log;

#ifdef __GNUC__
#ifdef __MINGW32__
__attribute__((gnu_format(printf, 1, 2)))
#else
__attribute__((format(printf, 1, 2)))
#endif
#endif
static void log(const char * fmt, ...) {
    if (!whisper_log) return;
    char buf[1024];
    va_list args;
    va_start(args, fmt);
    vsnprintf(buf, sizeof(buf), fmt, args);
    whisper_log(buf);
}

template<typename T>
static void read_safe(whisper_model_loader * loader, T & dest) {
    loader->read(loader->context, &dest, sizeof(T));
    BYTESWAP_VALUE(dest);
}

static bool kv_cache_init(
        const struct whisper_hparams & hparams,
             struct whisper_kv_cache & cache,
                           ggml_type   wtype,
                                 int   n_ctx) {
    const int64_t n_text_state = hparams.n_text_state;
    const int64_t n_text_layer = hparams.n_text_layer;

    const int64_t n_mem      = n_text_layer*n_ctx;
    const int64_t n_elements = n_text_state*n_mem;

    const size_t mem_bytes = 2*(ggml_type_size(wtype)*n_elements + ggml_tensor_overhead());

    cache.buf.resize(mem_bytes);

    struct ggml_init_params params = {
        /*.mem_size   =*/ cache.buf.size(),
        /*.mem_buffer =*/ cache.buf.data(),
        /*.no_alloc   =*/ false,
    };

    cache.ctx = ggml_init(params);

    if (!cache.ctx) {
        log("%s: failed to allocate memory for kv cache\n", __func__);
        return false;
    }

    cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
    cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);

    return true;
}

static bool kv_cache_reinit(struct whisper_kv_cache & cache) {
    WHISPER_ASSERT(cache.ctx);

    const int n_elements = ggml_nelements(cache.k);
    WHISPER_ASSERT(n_elements == ggml_nelements(cache.v));

    const ggml_type wtype = cache.k->type;
    WHISPER_ASSERT(wtype == cache.v->type);

    WHISPER_ASSERT(cache.buf.size() >= 2*n_elements*ggml_type_sizef(wtype));

    struct ggml_init_params params = {
        /*.mem_size   =*/ cache.buf.size(),
        /*.mem_buffer =*/ cache.buf.data(),
        /*.no_alloc   =*/ false,
    };

    cache.ctx = ggml_init(params);

    if (!cache.ctx) {
        log("%s: failed to allocate memory for kv cache\n", __func__);
        return false;
    }

    cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
    cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);

    return true;
}

static void kv_cache_free(struct whisper_kv_cache & cache) {
    if (cache.ctx) {
        ggml_free(cache.ctx);
        cache.ctx = nullptr;
    }
}

// load the model from a ggml file
//
// file format:
//
//   - hparams
//   - pre-computed mel filters
//   - vocab
//   - weights
//
// see the convert-pt-to-ggml.py script for details
//
static bool whisper_model_load(struct whisper_model_loader * loader, whisper_context & wctx) {
    log("%s: loading model\n", __func__);

    const int64_t t_start_us = ggml_time_us();

    wctx.t_start_us = t_start_us;

    auto & model = wctx.model;
    auto & vocab = wctx.vocab;

    // verify magic
    {
        uint32_t magic;
        read_safe(loader, magic);
        if (magic != GGML_FILE_MAGIC) {
            log("%s: invalid model data (bad magic)\n", __func__);
            return false;
        }
    }

    //load hparams
    {
        auto & hparams = model.hparams;

        read_safe(loader, hparams.n_vocab);
        read_safe(loader, hparams.n_audio_ctx);
        read_safe(loader, hparams.n_audio_state);
        read_safe(loader, hparams.n_audio_head);
        read_safe(loader, hparams.n_audio_layer);
        read_safe(loader, hparams.n_text_ctx);
        read_safe(loader, hparams.n_text_state);
        read_safe(loader, hparams.n_text_head);
        read_safe(loader, hparams.n_text_layer);
        read_safe(loader, hparams.n_mels);
        read_safe(loader, hparams.ftype);

        assert(hparams.n_text_state == hparams.n_audio_state);

        if (hparams.n_audio_layer == 4) {
            model.type = e_model::MODEL_TINY;
        }

        if (hparams.n_audio_layer == 6) {
            model.type = e_model::MODEL_BASE;
        }

        if (hparams.n_audio_layer == 12) {
            model.type = e_model::MODEL_SMALL;
        }

        if (hparams.n_audio_layer == 24) {
            model.type = e_model::MODEL_MEDIUM;
        }

        if (hparams.n_audio_layer == 32) {
            model.type = e_model::MODEL_LARGE;
        }

        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;

        hparams.ftype %= GGML_QNT_VERSION_FACTOR;

        // for the big tensors, we have the option to store the data in 16-bit floats or quantized
        // in order to save memory and also to speed up the computation
        wctx.wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
        if (wctx.wtype == GGML_TYPE_COUNT) {
            log("%s: invalid model (bad ftype value %d)\n", __func__, model.hparams.ftype);
            return false;
        }

        const size_t scale = model.hparams.ftype ? 1 : 2;

        log("%s: n_vocab       = %d\n", __func__, hparams.n_vocab);
        log("%s: n_audio_ctx   = %d\n", __func__, hparams.n_audio_ctx);
        log("%s: n_audio_state = %d\n", __func__, hparams.n_audio_state);
        log("%s: n_audio_head  = %d\n", __func__, hparams.n_audio_head);
        log("%s: n_audio_layer = %d\n", __func__, hparams.n_audio_layer);
        log("%s: n_text_ctx    = %d\n", __func__, hparams.n_text_ctx);
        log("%s: n_text_state  = %d\n", __func__, hparams.n_text_state);
        log("%s: n_text_head   = %d\n", __func__, hparams.n_text_head);
        log("%s: n_text_layer  = %d\n", __func__, hparams.n_text_layer);
        log("%s: n_mels        = %d\n", __func__, hparams.n_mels);
        log("%s: ftype         = %d\n", __func__, model.hparams.ftype);
        log("%s: qntvr         = %d\n", __func__, qntvr);
        log("%s: type          = %d\n", __func__, model.type);

        // print memory requirements
        {
            // TODO
            //log("%s: mem required  = %7.2f MB (+ %7.2f MB per decoder)\n", __func__,
            //        mem_required / 1024.0 / 1024.0, mem_required_decoder / 1024.0 / 1024.0);
        }

        // initialize all memory buffers
        // always have at least one decoder

        wctx.model.buf = new std::vector<uint8_t>();
        wctx.model.buf->resize(scale*MEM_REQ_MODEL.at(wctx.wtype).at(model.type));

        // we skip initialization of the state until it is needed
        // because it might be that state will always be provided externally.
    }

    // load mel filters
    {
        auto & filters = wctx.model.filters;

        read_safe(loader, filters.n_mel);
        read_safe(loader, filters.n_fft);

        filters.data.resize(filters.n_mel * filters.n_fft);
        loader->read(loader->context, filters.data.data(), filters.data.size() * sizeof(float));
        BYTESWAP_FILTERS(filters);
    }

    // load vocab
    {
        int32_t n_vocab = 0;
        read_safe(loader, n_vocab);

        //if (n_vocab != model.hparams.n_vocab) {
        //    log("%s: invalid model file '%s' (bad vocab size %d != %d)\n",
        //            __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
        //    return false;
        //}

        std::string word;
        std::vector<char> tmp;

        tmp.reserve(128);

        for (int i = 0; i < n_vocab; i++) {
            uint32_t len;
            read_safe(loader, len);

            if (len > 0) {
                tmp.resize(len);
                loader->read(loader->context, &tmp[0], tmp.size()); // read to buffer
                word.assign(&tmp[0], tmp.size());
            } else {
                // seems like we have an empty-string token in multi-language models (i = 50256)
                //log("%s: warning: empty-string token in vocab, i = %d\n", __func__, i);
                word = "";
            }

            vocab.token_to_id[word] = i;
            vocab.id_to_token[i] = word;

            //printf("%s: vocab[%d] = '%s'\n", __func__, i, word.c_str());
        }

        vocab.n_vocab = model.hparams.n_vocab;
        if (vocab.is_multilingual()) {
            vocab.token_eot++;
            vocab.token_sot++;
            vocab.token_translate++;
            vocab.token_transcribe++;
            vocab.token_solm++;
            vocab.token_prev++;
            vocab.token_nosp++;
            vocab.token_not++;
            vocab.token_beg++;
        }

        if (n_vocab < model.hparams.n_vocab) {
            log("%s: adding %d extra tokens\n", __func__, model.hparams.n_vocab - n_vocab);
            for (int i = n_vocab; i < model.hparams.n_vocab; i++) {
                if (i > vocab.token_beg) {
                    word = "[_TT_" + std::to_string(i - vocab.token_beg) + "]";
                } else if (i == vocab.token_eot) {
                    word = "[_EOT_]";
                } else if (i == vocab.token_sot) {
                    word = "[_SOT_]";
                } else if (i == vocab.token_solm) {
                    word = "[_SOLM_]";
                } else if (i == vocab.token_prev) {
                    word = "[_PREV_]";
                } else if (i == vocab.token_nosp) {
                    word = "[_NOSP_]";
                } else if (i == vocab.token_not) {
                    word = "[_NOT_]";
                } else if (i == vocab.token_beg) {
                    word = "[_BEG_]";
                } else {
                    word = "[_extra_token_" + std::to_string(i) + "]";
                }
                vocab.token_to_id[word] = i;
                vocab.id_to_token[i] = word;
            }
        }
    }

    size_t ctx_size = 0;

    const ggml_type wtype = wctx.wtype;
    const ggml_type vtype = wctx.wtype == GGML_TYPE_F32 ? GGML_TYPE_F32 : GGML_TYPE_F16; // conv type

    {
        const auto & hparams = model.hparams;

        const int n_vocab = hparams.n_vocab;

        const int n_audio_ctx   = hparams.n_audio_ctx;
        const int n_audio_state = hparams.n_audio_state;
        const int n_audio_layer = hparams.n_audio_layer;

        const int n_text_ctx   = hparams.n_text_ctx;
        const int n_text_state = hparams.n_text_state;
        const int n_text_layer = hparams.n_text_layer;

        const int n_mels = hparams.n_mels;

        // encoder
        {
            ctx_size += n_audio_ctx*n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_pe;

            ctx_size += 3*n_mels*n_audio_state*ggml_type_sizef(vtype);         // e_conv_1_w
            ctx_size +=          n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_conv_1_b

            ctx_size += 3*n_audio_state*n_audio_state*ggml_type_sizef(vtype);         // e_conv_2_w
            ctx_size +=                 n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_conv_2_b

            ctx_size += n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_ln_w;
            ctx_size += n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_ln_b;
        }

        // decoder
        {
            ctx_size += n_text_ctx*n_text_state*ggml_type_sizef(GGML_TYPE_F32); // d_pe;

            ctx_size += n_vocab*n_text_state*ggml_type_sizef(wtype); // d_te;

            ctx_size += n_text_state*ggml_type_sizef(GGML_TYPE_F32); // d_ln_w;
            ctx_size += n_text_state*ggml_type_sizef(GGML_TYPE_F32); // d_ln_b;
        }

        // encoder layers
        {
            ctx_size += n_audio_layer*(n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_ln_w
            ctx_size += n_audio_layer*(n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_ln_b

            ctx_size += n_audio_layer*(4*n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // mlp_0_w
            ctx_size += n_audio_layer*(              4*n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_0_b

            ctx_size += n_audio_layer*(4*n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // mlp_1_w
            ctx_size += n_audio_layer*(                n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_1_b

            ctx_size += n_audio_layer*(n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_0_w
            ctx_size += n_audio_layer*(n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_0_b

            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // attn_q_w
            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_q_b

            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_sizef(wtype)); // attn_k_w

            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // attn_v_w
            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_v_b

            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // attn_ln_1_w
            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_1_b
        }

        // decoder layers
        {
            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_ln_w
            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_ln_b

            ctx_size += n_text_layer*(4*n_text_state*n_text_state*ggml_type_sizef(wtype));         // mlp_0_w
            ctx_size += n_text_layer*(             4*n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_0_b

            ctx_size += n_text_layer*(4*n_text_state*n_text_state*ggml_type_sizef(wtype));         // mlp_1_w
            ctx_size += n_text_layer*(               n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_1_b

            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_0_w
            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_0_b

            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // attn_q_w
            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_q_b

            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype)); // attn_k_w

            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // attn_v_w
            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_v_b

            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // attn_ln_1_w
            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_1_b
                                                                                                //
            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_ln_0_w
            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_ln_0_b

            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // cross_attn_q_w
            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_q_b

            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype)); // cross_attn_k_w

            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // cross_attn_v_w
            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_v_b

            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // cross_attn_ln_1_w
            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_ln_1_b
        }

        ctx_size += (15 + 15*n_audio_layer + 24*n_text_layer)*512; // object overhead

        log("%s: model ctx     = %7.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
    }

    // create the ggml context
    {
        struct ggml_init_params params = {
            /*.mem_size   =*/ wctx.model.buf->size(),
            /*.mem_buffer =*/ wctx.model.buf->data(),
            /*.no_alloc   =*/ false,
        };

        model.ctx = ggml_init(params);
        if (!model.ctx) {
            log("%s: ggml_init() failed\n", __func__);
            return false;
        }
    }

    // prepare memory for the weights
    {
        auto & ctx = model.ctx;

        const auto & hparams = model.hparams;

        const int n_vocab = hparams.n_vocab;

        const int n_audio_ctx   = hparams.n_audio_ctx;
        const int n_audio_state = hparams.n_audio_state;
        const int n_audio_layer = hparams.n_audio_layer;

        const int n_text_ctx   = hparams.n_text_ctx;
        const int n_text_state = hparams.n_text_state;
        const int n_text_layer = hparams.n_text_layer;

        const int n_mels = hparams.n_mels;

        model.layers_encoder.resize(n_audio_layer);
        model.layers_decoder.resize(n_text_layer);

        // encoder
        {
            model.e_pe       = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_audio_state, n_audio_ctx);

            model.e_conv_1_w = ggml_new_tensor_3d(ctx, vtype,         3, n_mels, n_audio_state);
            model.e_conv_1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n_audio_state);

            model.e_conv_2_w = ggml_new_tensor_3d(ctx, vtype,         3, n_audio_state, n_audio_state);
            model.e_conv_2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n_audio_state);

            model.e_ln_w     = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
            model.e_ln_b     = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);

            // map by name
            model.tensors["encoder.positional_embedding"] = model.e_pe;

            model.tensors["encoder.conv1.weight"]         = model.e_conv_1_w;
            model.tensors["encoder.conv1.bias"]           = model.e_conv_1_b;

            model.tensors["encoder.conv2.weight"]         = model.e_conv_2_w;
            model.tensors["encoder.conv2.bias"]           = model.e_conv_2_b;

            model.tensors["encoder.ln_post.weight"]       = model.e_ln_w;
            model.tensors["encoder.ln_post.bias"]         = model.e_ln_b;

            for (int i = 0; i < n_audio_layer; ++i) {
                auto & layer = model.layers_encoder[i];

                layer.mlp_ln_w    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_audio_state);
                layer.mlp_ln_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_audio_state);

                layer.mlp_0_w     = ggml_new_tensor_2d(ctx, wtype,           n_audio_state, 4*n_audio_state);
                layer.mlp_0_b     = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_audio_state);

                layer.mlp_1_w     = ggml_new_tensor_2d(ctx, wtype,         4*n_audio_state, n_audio_state);
                layer.mlp_1_b     = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_audio_state);

                layer.attn_ln_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_audio_state);
                layer.attn_ln_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_audio_state);

                layer.attn_q_w    = ggml_new_tensor_2d(ctx, wtype,           n_audio_state, n_audio_state);
                layer.attn_q_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_audio_state);

                layer.attn_k_w    = ggml_new_tensor_2d(ctx, wtype,           n_audio_state, n_audio_state);

                layer.attn_v_w    = ggml_new_tensor_2d(ctx, wtype,           n_audio_state, n_audio_state);
                layer.attn_v_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_audio_state);

                layer.attn_ln_1_w = ggml_new_tensor_2d(ctx, wtype,           n_audio_state, n_audio_state);
                layer.attn_ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_audio_state);

                // map by name
                model.tensors["encoder.blocks." + std::to_string(i) + ".mlp_ln.weight"]     = layer.mlp_ln_w;
                model.tensors["encoder.blocks." + std::to_string(i) + ".mlp_ln.bias"]       = layer.mlp_ln_b;

                model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.0.weight"]      = layer.mlp_0_w;
                model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.0.bias"]        = layer.mlp_0_b;

                model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.2.weight"]      = layer.mlp_1_w;
                model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.2.bias"]        = layer.mlp_1_b;

                model.tensors["encoder.blocks." + std::to_string(i) + ".attn_ln.weight"]    = layer.attn_ln_0_w;
                model.tensors["encoder.blocks." + std::to_string(i) + ".attn_ln.bias"]      = layer.attn_ln_0_b;

                model.tensors["encoder.blocks." + std::to_string(i) + ".attn.query.weight"] = layer.attn_q_w;
                model.tensors["encoder.blocks." + std::to_string(i) + ".attn.query.bias"]   = layer.attn_q_b;

                model.tensors["encoder.blocks." + std::to_string(i) + ".attn.key.weight"]   = layer.attn_k_w;

                model.tensors["encoder.blocks." + std::to_string(i) + ".attn.value.weight"] = layer.attn_v_w;
                model.tensors["encoder.blocks." + std::to_string(i) + ".attn.value.bias"]   = layer.attn_v_b;

                model.tensors["encoder.blocks." + std::to_string(i) + ".attn.out.weight"]   = layer.attn_ln_1_w;
                model.tensors["encoder.blocks." + std::to_string(i) + ".attn.out.bias"]     = layer.attn_ln_1_b;
            }
        }

        // decoder
        {
            model.d_pe   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_text_state, n_text_ctx);

            model.d_te   = ggml_new_tensor_2d(ctx, wtype,         n_text_state, n_vocab);

            model.d_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
            model.d_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);

            // map by name
            model.tensors["decoder.positional_embedding"]   = model.d_pe;

            model.tensors["decoder.token_embedding.weight"] = model.d_te;

            model.tensors["decoder.ln.weight"]              = model.d_ln_w;
            model.tensors["decoder.ln.bias"]                = model.d_ln_b;

            for (int i = 0; i < n_text_layer; ++i) {
                auto & layer = model.layers_decoder[i];

                layer.mlp_ln_w          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);
                layer.mlp_ln_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.mlp_0_w           = ggml_new_tensor_2d(ctx, wtype,           n_text_state, 4*n_text_state);
                layer.mlp_0_b           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_text_state);

                layer.mlp_1_w           = ggml_new_tensor_2d(ctx, wtype,         4*n_text_state, n_text_state);
                layer.mlp_1_b           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.attn_ln_0_w       = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);
                layer.attn_ln_0_b       = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.attn_q_w          = ggml_new_tensor_2d(ctx, wtype,           n_text_state, n_text_state);
                layer.attn_q_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.attn_k_w          = ggml_new_tensor_2d(ctx, wtype,           n_text_state, n_text_state);

                layer.attn_v_w          = ggml_new_tensor_2d(ctx, wtype,           n_text_state, n_text_state);
                layer.attn_v_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.attn_ln_1_w       = ggml_new_tensor_2d(ctx, wtype,           n_text_state, n_text_state);
                layer.attn_ln_1_b       = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.cross_attn_ln_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);
                layer.cross_attn_ln_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.cross_attn_q_w    = ggml_new_tensor_2d(ctx, wtype,           n_text_state, n_text_state);
                layer.cross_attn_q_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.cross_attn_k_w    = ggml_new_tensor_2d(ctx, wtype,           n_text_state, n_text_state);

                layer.cross_attn_v_w    = ggml_new_tensor_2d(ctx, wtype,           n_text_state, n_text_state);
                layer.cross_attn_v_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                layer.cross_attn_ln_1_w = ggml_new_tensor_2d(ctx, wtype,           n_text_state, n_text_state);
                layer.cross_attn_ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_text_state);

                // map by name
                model.tensors["decoder.blocks." + std::to_string(i) + ".mlp_ln.weight"]           = layer.mlp_ln_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".mlp_ln.bias"]             = layer.mlp_ln_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.0.weight"]            = layer.mlp_0_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.0.bias"]              = layer.mlp_0_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.2.weight"]            = layer.mlp_1_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.2.bias"]              = layer.mlp_1_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".attn_ln.weight"]          = layer.attn_ln_0_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".attn_ln.bias"]            = layer.attn_ln_0_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".attn.query.weight"]       = layer.attn_q_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".attn.query.bias"]         = layer.attn_q_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".attn.key.weight"]         = layer.attn_k_w;

                model.tensors["decoder.blocks." + std::to_string(i) + ".attn.value.weight"]       = layer.attn_v_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".attn.value.bias"]         = layer.attn_v_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".attn.out.weight"]         = layer.attn_ln_1_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".attn.out.bias"]           = layer.attn_ln_1_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn_ln.weight"]    = layer.cross_attn_ln_0_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn_ln.bias"]      = layer.cross_attn_ln_0_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.query.weight"] = layer.cross_attn_q_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.query.bias"]   = layer.cross_attn_q_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.key.weight"]   = layer.cross_attn_k_w;

                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.value.weight"] = layer.cross_attn_v_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.value.bias"]   = layer.cross_attn_v_b;

                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.out.weight"]   = layer.cross_attn_ln_1_w;
                model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.out.bias"]     = layer.cross_attn_ln_1_b;
            }
        }
    }

    // load weights
    {
        size_t total_size = 0;

        model.n_loaded = 0;

        while (true) {
            int32_t n_dims;
            int32_t length;
            int32_t ttype;

            read_safe(loader, n_dims);
            read_safe(loader, length);
            read_safe(loader, ttype);

            if (loader->eof(loader->context)) {
                break;
            }

            int32_t nelements = 1;
            int32_t ne[4] = { 1, 1, 1, 1 };
            for (int i = 0; i < n_dims; ++i) {
                read_safe(loader, ne[i]);
                nelements *= ne[i];
            }

            std::string name;
            std::vector<char> tmp(length); // create a buffer
            loader->read(loader->context, &tmp[0], tmp.size()); // read to buffer
            name.assign(&tmp[0], tmp.size());

            if (model.tensors.find(name) == model.tensors.end()) {
                log("%s: unknown tensor '%s' in model file\n", __func__, name.data());
                return false;
            }

            auto tensor = model.tensors[name.data()];
            if (ggml_nelements(tensor) != nelements) {
                log("%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
                log("%s: shape: [%d, %d, %d], expected: [%d, %d, %d]\n",
                        __func__, ne[0], ne[1], ne[2], (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2]);
                return false;
            }

            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1] || tensor->ne[2] != ne[2]) {
                log("%s: tensor '%s' has wrong shape in model file: got [%d, %d, %d], expected [%d, %d, %d]\n",
                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], ne[0], ne[1], ne[2]);
                return false;
            }

            const size_t bpe = ggml_type_size(ggml_type(ttype));

            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
                log("%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
                return false;
            }

            loader->read(loader->context, tensor->data, ggml_nbytes(tensor));
            BYTESWAP_TENSOR(tensor);

            //printf("%48s - [%5d, %5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype), ggml_nbytes(tensor)/1024.0/1024.0);
            total_size += ggml_nbytes(tensor);
            model.n_loaded++;
        }

        log("%s: model size    = %7.2f MB\n", __func__, total_size/1024.0/1024.0);

        if (model.n_loaded == 0) {
            log("%s: WARN no tensors loaded from model file - assuming empty model for testing\n", __func__);
        } else if (model.n_loaded != (int) model.tensors.size()) {
            log("%s: ERROR not all tensors loaded from model file - expected %zu, got %d\n", __func__, model.tensors.size(), model.n_loaded);
            return false;
        }
    }

    wctx.t_load_us = ggml_time_us() - t_start_us;

    return true;
}

static bool whisper_encode_external(const whisper_state & wstate) {
    GGML_UNUSED(wstate);

#ifndef WHISPER_USE_COREML
    const bool use_coreml = false;
#else
    const bool use_coreml = wstate.ctx_coreml != nullptr;
#endif

#ifndef WHISPER_USE_OPENVINO
    const bool use_openvino = false;
#else
    const bool use_openvino = wstate.ctx_openvino != nullptr;
#endif

    return use_coreml || use_openvino;
}

static struct ggml_cgraph * whisper_build_graph_conv(
        whisper_context & wctx,
          whisper_state & wstate,
              const int   mel_offset) {
    const auto & model   = wctx.model;
    const auto & mel_inp = wstate.mel;
    const auto & hparams = model.hparams;

    const int n_ctx   = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;
    const int n_state = hparams.n_audio_state; GGML_UNUSED(n_state);

    const int n_mels = hparams.n_mels;

    struct ggml_init_params params = {
        /*.mem_size   =*/ wstate.alloc_conv.meta.size(),
        /*.mem_buffer =*/ wstate.alloc_conv.meta.data(),
        /*.no_alloc   =*/ true,
    };

    struct ggml_context * ctx0 = ggml_init(params);

    ggml_cgraph * gf = ggml_new_graph(ctx0);

    ggml_allocr * alloc = wstate.alloc_conv.alloc;

    struct ggml_tensor * mel = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 2*n_ctx, n_mels);
    ggml_allocr_alloc(alloc, mel);

    assert(mel->type == GGML_TYPE_F32);
    if (!ggml_allocr_is_measure(alloc)) {
        assert(mel_inp.n_mel == n_mels);

        float * dst = (float *) mel->data;
        memset(dst, 0, ggml_nbytes(mel));

        const int i0 = std::min(mel_offset, mel_inp.n_len);
        const int i1 = std::min(mel_offset + 2*n_ctx, mel_inp.n_len);

        for (int j = 0; j < mel_inp.n_mel; ++j) {
            for (int i = i0; i < i1; ++i) {
                dst[j*2*n_ctx + (i - i0)] = mel_inp.data[j*mel_inp.n_len + i];
            }
        }
    }

    struct ggml_tensor * cur = nullptr;

    if (!whisper_encode_external(wstate)) {
        // convolution + gelu
        {
            cur = ggml_conv_1d_ph(ctx0, model.e_conv_1_w, mel, 1, 1);
            cur = ggml_add(ctx0,
                    ggml_repeat(ctx0,
                        model.e_conv_1_b,
                        cur),
                    cur);

            cur = ggml_gelu(ctx0, cur);

            cur = ggml_conv_1d_ph(ctx0, model.e_conv_2_w, cur, 2, 1);
            cur = ggml_add(ctx0,
                    ggml_repeat(ctx0,
                        model.e_conv_2_b,
                        cur),
                    cur);

            cur = ggml_gelu(ctx0, cur);
        }

        wstate.embd_conv = cur;
    } else {
#ifdef WHISPER_USE_COREML
        cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx);
        ggml_allocr_alloc(alloc, cur);

        if (!ggml_allocr_is_measure(alloc)) {
            whisper_coreml_encode(wstate.ctx_coreml, (float *) mel->data, (float *) cur->data);
        }
#endif
#ifdef WHISPER_USE_OPENVINO
        cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx);
        ggml_allocr_alloc(alloc, cur);

        if (!ggml_allocr_is_measure(alloc)) {
            whisper_openvino_encode(wstate.ctx_openvino, mel, cur);
        }
#endif

        wstate.embd_enc = cur;
    }

    ggml_build_forward_expand(gf, cur);

    ggml_free(ctx0);

    return gf;
}

static struct ggml_cgraph * whisper_build_graph_encoder(
        whisper_context & wctx,
          whisper_state & wstate) {
    const auto & model   = wctx.model;
    const auto & hparams = model.hparams;

    const int n_ctx   = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;
    const int n_state = hparams.n_audio_state;
    const int n_head  = hparams.n_audio_head;
    const int n_layer = hparams.n_audio_layer;

    struct ggml_init_params params = {
        /*.mem_size   =*/ wstate.alloc_encode.meta.size(),
        /*.mem_buffer =*/ wstate.alloc_encode.meta.data(),
        /*.no_alloc   =*/ true,
    };

    struct ggml_context * ctx0 = ggml_init(params);

    ggml_cgraph * gf = ggml_new_graph(ctx0);

    ggml_allocr * alloc = wstate.alloc_encode.alloc;

    struct ggml_tensor * KQscale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
    ggml_allocr_alloc(alloc, KQscale);

    if (!ggml_allocr_is_measure(alloc)) {
        ggml_set_f32(KQscale, 1.0f/sqrt(float(n_state)/n_head));
    }

    struct ggml_tensor * cur = ggml_view_tensor(ctx0, wstate.embd_conv);

    // ===================================================================
    // NOTE: experimenting with partial evaluation of the encoder (ignore)
    //static int iter = -1;
    //const int n_iter = 1500/n_ctx;

    //iter = (iter + 1) % n_iter;

    //if (iter == 0) {
    //    memset(model.memory_cross_k->data, 0, ggml_nbytes(model.memory_cross_k));
    //    memset(model.memory_cross_v->data, 0, ggml_nbytes(model.memory_cross_v));
    //}

    static int iter = 0;

    const size_t e_pe_stride = model.e_pe->ne[0]*ggml_element_size(model.e_pe);
    const size_t e_pe_offset = model.e_pe->ne[0]*ggml_element_size(model.e_pe)*n_ctx*iter;

    struct ggml_tensor * e_pe = ggml_view_2d(ctx0, model.e_pe, model.e_pe->ne[0], n_ctx, e_pe_stride, e_pe_offset);

    cur = ggml_add(ctx0, e_pe, ggml_cont(ctx0, ggml_transpose(ctx0, cur)));

    // ===================================================================

    // original:
    //cur = ggml_add(ctx0, model.e_pe, ggml_transpose(ctx0, cur));

    struct ggml_tensor * inpL = cur;

    for (int il = 0; il < n_layer; ++il) {
        const auto & layer = model.layers_encoder[il];

        // norm
        {
            cur = ggml_norm(ctx0, inpL, hparams.eps);

            // cur = ln_0_w*cur + ln_0_b
            cur = ggml_add(ctx0,
                    ggml_mul(ctx0, cur, layer.attn_ln_0_w),
                    layer.attn_ln_0_b);
        }

        // self-attention
        {
            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                    layer.attn_q_w,
                    cur);

            Qcur = ggml_add(ctx0, Qcur, layer.attn_q_b);

            //Qcur = ggml_scale(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));

            // note: no bias for Key
            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0,
                    layer.attn_k_w,
                    cur);

            //Kcur = ggml_scale(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));

            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0,
                    layer.attn_v_w,
                    cur);

            Vcur = ggml_add(ctx0, Vcur, layer.attn_v_b);

            // ------

#ifdef WHISPER_USE_FLASH_ATTN
            struct ggml_tensor * Q =
                ggml_permute(ctx0,
                        ggml_cpy(ctx0,
                            Qcur,
                            ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
                        0, 2, 1, 3);

            struct ggml_tensor * K =
                ggml_permute(ctx0,
                        ggml_cpy(ctx0,
                            Kcur,
                            ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
                        0, 2, 1, 3);

            struct ggml_tensor * V =
                ggml_cpy(ctx0,
                        ggml_permute(ctx0,
                            ggml_reshape_3d(ctx0,
                                Vcur,
                                n_state/n_head, n_head, n_ctx),
                            1, 2, 0, 3),
                        ggml_new_tensor_3d(ctx0, wctx.itype, n_ctx, n_state/n_head, n_head));

            struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, false);
#else
            struct ggml_tensor * Q =
                ggml_permute(ctx0,
                        ggml_cpy(ctx0,
                            Qcur,
                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_state/n_head, n_head, n_ctx)),
                        0, 2, 1, 3);

            struct ggml_tensor * K =
                ggml_permute(ctx0,
                        ggml_cpy(ctx0,
                            Kcur,
                            ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
                        0, 2, 1, 3);

            // K * Q
            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);

            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQscale);

            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_scaled);

            struct ggml_tensor * V =
                ggml_cpy(ctx0,
                        ggml_permute(ctx0,
                            ggml_reshape_3d(ctx0,
                                Vcur,
                                n_state/n_head, n_head, n_ctx),
                            1, 2, 0, 3),
                        ggml_new_tensor_3d(ctx0, wctx.itype, n_ctx, n_state/n_head, n_head)
                        );

            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
#endif
            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);

            cur = ggml_cpy(ctx0,
                    KQV_merged,
                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx));
        }

        // projection
        {
            cur = ggml_mul_mat(ctx0,
                    layer.attn_ln_1_w,
                    cur);

            cur = ggml_add(ctx0, cur, layer.attn_ln_1_b);
        }

        // add the input
        cur = ggml_add(ctx0, cur, inpL);

        struct ggml_tensor * inpFF = cur;

        // feed-forward network
        {
            // norm
            {
                cur = ggml_norm(ctx0, inpFF, hparams.eps);

                // cur = mlp_ln_w*cur + mlp_ln_b
                cur = ggml_add(ctx0,
                        ggml_mul(ctx0, cur, layer.mlp_ln_w),
                        layer.mlp_ln_b);
            }

#ifdef WHISPER_USE_FLASH_FF
            cur = ggml_flash_ff(ctx0,
                    ggml_cpy(ctx0, cur, ggml_new_tensor_2d(ctx0, wstate.itype, n_state, n_ctx)),
                    layer.mlp_0_w, layer.mlp_0_b, layer.mlp_1_w, layer.mlp_1_b);
#else
            // fully connected
            cur = ggml_mul_mat(ctx0,
                    layer.mlp_0_w,
                    cur);

            cur = ggml_add(ctx0, cur, layer.mlp_0_b);

            // GELU activation
            cur = ggml_gelu(ctx0, cur);

            // projection
            cur = ggml_mul_mat(ctx0,
                    layer.mlp_1_w,
                    cur);

            cur = ggml_add(ctx0, cur, layer.mlp_1_b);
#endif
        }

        inpL = ggml_add(ctx0, cur, inpFF);
    }

    cur = inpL;

    // norm
    {
        cur = ggml_norm(ctx0, cur, hparams.eps);

        // cur = ln_f_g*cur + ln_f_b
        cur = ggml_add(ctx0,
                ggml_mul(ctx0, cur, model.e_ln_w),
                model.e_ln_b);
    }

    ggml_build_forward_expand(gf, cur);

    wstate.embd_enc = cur;

    //ggml_graph_print(gf);

    ////////////////////////////////////////////////////////////////////////////

    //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
    //        ggml_used_mem(ctx0)/1024.0/1024.0,
    //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
    //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
    //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
    //        wstate.get_buf_max_mem(3)/1024.0/1024.0);

    ggml_free(ctx0);

    return gf;
}

// pre-compute cross-attention memory
static struct ggml_cgraph * whisper_build_graph_cross(
        whisper_context & wctx,
          whisper_state & wstate) {
    const auto & model   = wctx.model;
    const auto & hparams = model.hparams;

    const int n_ctx   = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;
    const int n_state = hparams.n_audio_state;
    const int n_head  = hparams.n_audio_head;

    struct ggml_init_params params = {
        /*.mem_size   =*/ wstate.alloc_cross.meta.size(),
        /*.mem_buffer =*/ wstate.alloc_cross.meta.data(),
        /*.no_alloc   =*/ true,
    };

    struct ggml_context * ctx0 = ggml_init(params);

    ggml_cgraph * gf = ggml_new_graph(ctx0);

    ggml_allocr * alloc = wstate.alloc_cross.alloc;

    struct ggml_tensor * cur = ggml_view_tensor(ctx0, wstate.embd_enc);

    struct ggml_tensor * Kscale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
    ggml_allocr_alloc(alloc, Kscale);

    if (!ggml_allocr_is_measure(alloc)) {
        ggml_set_f32(Kscale, pow(float(n_state) / n_head, -0.25));
    }

    for (int il = 0; il < model.hparams.n_text_layer; ++il) {
        auto & layer = model.layers_decoder[il];

        struct ggml_tensor* Kcross = ggml_mul_mat(ctx0,
                layer.cross_attn_k_w,
                cur);

        Kcross = ggml_scale(ctx0, Kcross, Kscale);

        struct ggml_tensor* Vcross = ggml_mul_mat(ctx0,
                layer.cross_attn_v_w,
                cur);

        Vcross = ggml_add(ctx0,
                    Vcross,
                    layer.cross_attn_v_b);

        Vcross = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcross, n_state, n_ctx));

        struct ggml_tensor * k = ggml_view_1d(ctx0, wstate.kv_cross.k,
                n_state*n_ctx,
                (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));

        struct ggml_tensor * v = ggml_view_2d(ctx0, wstate.kv_cross.v, n_ctx, n_state,
                (   n_ctx)*ggml_element_size(wstate.kv_cross.v),
                (il*n_ctx)*ggml_element_size(wstate.kv_cross.v)*n_state);

        ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcross, k));
        ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcross, v));
    }

    //ggml_graph_print(gf);

    ggml_free(ctx0);

    return gf;
}

// evaluate the encoder with the given state
//
// given audio recording (more specifically, its log mel spectrogram), runs forward pass of the encoder
// part of the transformer model and returns the encoded features
//
//   - wctx:      the model
//   - wstate:     the state of the encoder
//   - n_threads:  number of threads to use
//   - mel_offset: offset in the mel spectrogram (i.e. audio offset)
//
static bool whisper_encode_internal(
        whisper_context & wctx,
          whisper_state & wstate,
              const int   mel_offset,
              const int   n_threads) {
    const int64_t t_start_us = ggml_time_us();

    // conv
    {
        auto & alloc = wstate.alloc_conv.alloc;

        ggml_allocr_reset(alloc);

        ggml_cgraph * gf = whisper_build_graph_conv(wctx, wstate, mel_offset);

        ggml_allocr_alloc_graph(alloc, gf);

        if (!whisper_encode_external(wstate)) {
            ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads);
        }
    }

    // encoder
    if (!whisper_encode_external(wstate)) {
        auto & alloc = wstate.alloc_encode.alloc;

        ggml_allocr_reset(alloc);

        ggml_cgraph * gf = whisper_build_graph_encoder(wctx, wstate);

        ggml_allocr_alloc_graph(alloc, gf);

#ifdef GGML_USE_METAL
        if (wstate.ctx_metal) {
            ggml_metal_set_n_cb     (wstate.ctx_metal, n_threads);
            ggml_metal_graph_compute(wstate.ctx_metal, gf);
        } else {
            ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads);
        }
#else
        ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads);
#endif
    }

    // cross
    {
        auto & alloc = wstate.alloc_cross.alloc;

        ggml_allocr_reset(alloc);

        ggml_cgraph * gf = whisper_build_graph_cross(wctx, wstate);

        ggml_allocr_alloc_graph(alloc, gf);

#ifdef GGML_USE_METAL
        if (wstate.ctx_metal) {
            ggml_metal_set_n_cb     (wstate.ctx_metal, n_threads);
            ggml_metal_graph_compute(wstate.ctx_metal, gf);
        } else {
            ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads);
        }
#else
        ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads);
#endif
    }

    // ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);

    wstate.t_encode_us += ggml_time_us() - t_start_us;
    wstate.n_encode++;

    return true;
}

static struct ggml_cgraph * whisper_build_graph_decoder(
         whisper_context & wctx,
         whisper_state   & wstate,
         whisper_decoder & decoder,
     const whisper_token * tokens,
                   int   n_tokens,
                   int   n_past) {
    const auto & model   = wctx.model;
    const auto & hparams = model.hparams;

    auto & kv_self = decoder.kv_self;

    WHISPER_ASSERT(!!kv_self.ctx);

    const int n_ctx   = hparams.n_text_ctx;
    const int n_state = hparams.n_text_state;
    const int n_head  = hparams.n_text_head;
    const int n_layer = hparams.n_text_layer;

    const int N = n_tokens;
    const int M = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;

    //WHISPER_PRINT_DEBUG("%s: n_past = %d, N = %d, M = %d, n_ctx = %d\n", __func__, n_past, N, M, n_ctx);

    struct ggml_init_params params = {
        /*.mem_size   =*/ wstate.alloc_decode.meta.size(),
        /*.mem_buffer =*/ wstate.alloc_decode.meta.data(),
        /*.no_alloc   =*/ true,
    };

    struct ggml_context * ctx0 = ggml_init(params);

    ggml_cgraph * gf = ggml_new_graph(ctx0);

    ggml_allocr * alloc = wstate.alloc_decode.alloc;

    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
    ggml_allocr_alloc(alloc, embd);

    if (!ggml_allocr_is_measure(alloc)) {
        memcpy(embd->data, tokens, N*ggml_element_size(embd));
    }

    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
    ggml_allocr_alloc(alloc, position);

    if (!ggml_allocr_is_measure(alloc)) {
        for (int i = 0; i < N; ++i) {
            ((int32_t *) position->data)[i] = n_past + i;
        }
    }

    struct ggml_tensor * KQscale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
    ggml_allocr_alloc(alloc, KQscale);

    if (!ggml_allocr_is_measure(alloc)) {
        ggml_set_f32(KQscale, pow(float(n_state)/n_head, -0.25));
    }

    // token encoding + position encoding
    struct ggml_tensor * cur =
        ggml_add(ctx0,
                ggml_get_rows(ctx0, model.d_te, embd),
                ggml_get_rows(ctx0, model.d_pe, position));

    struct ggml_tensor * inpL = cur;

    for (int il = 0; il < n_layer; ++il) {
        const auto & layer = model.layers_decoder[il];

        // norm
        {
            cur = ggml_norm(ctx0, inpL, hparams.eps);

            // cur = ln_0_w*cur + ln_0_b
            cur = ggml_add(ctx0,
                    ggml_mul(ctx0,
                        cur,
                        layer.attn_ln_0_w),
                    layer.attn_ln_0_b);
        }

        // self-attention
        {
            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                    layer.attn_q_w,
                    cur);

            Qcur = ggml_add(ctx0,
                        Qcur,
                        layer.attn_q_b);

            Qcur = ggml_scale(ctx0, Qcur, KQscale);

            // note: no bias for Key
            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0,
                    layer.attn_k_w,
                    cur);

            Kcur = ggml_scale(ctx0, Kcur, KQscale);

            // store key and value to memory
            {
                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0,
                        layer.attn_v_w,
                        cur);

                Vcur = ggml_add(ctx0,
                            Vcur,
                            layer.attn_v_b);

                Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_state, N));

                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_state, (ggml_element_size(kv_self.k)*n_state)*(il*n_ctx + n_past));
                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_state,
                        (   n_ctx)*ggml_element_size(kv_self.v),
                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_state + n_past*ggml_element_size(kv_self.v));

                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
            }

            // ------

            struct ggml_tensor * Q =
                ggml_permute(ctx0,
                        ggml_reshape_3d(ctx0, Qcur, n_state/n_head, n_head, N),
                        0, 2, 1, 3);

            struct ggml_tensor * K =
                ggml_view_3d(ctx0, kv_self.k,
                        n_state/n_head, n_past + N, n_head,
                        ggml_element_size(kv_self.k)*n_state,
                        ggml_element_size(kv_self.k)*n_state/n_head,
                        ggml_element_size(kv_self.k)*n_state*n_ctx*il);

            // K * Q
            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);

            //struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);

            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ, n_past);

            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);

            struct ggml_tensor * V =
                ggml_view_3d(ctx0, kv_self.v,
                        n_past + N, n_state/n_head, n_head,
                        n_ctx*ggml_element_size(kv_self.v),
                        n_ctx*ggml_element_size(kv_self.v)*n_state/n_head,
                        il*n_ctx*ggml_element_size(kv_self.v)*n_state);

            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);

            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);

            cur = ggml_cpy(ctx0,
                    KQV_merged,
                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, N));
        }

        // projection
        {
            cur = ggml_mul_mat(ctx0,
                    layer.attn_ln_1_w,
                    cur);

            cur = ggml_add(ctx0,
                    cur,
                    layer.attn_ln_1_b);
        }

        // add the input
        struct ggml_tensor * inpCA = ggml_add(ctx0, cur, inpL);

        // norm
        {
            cur = ggml_norm(ctx0, inpCA, hparams.eps); // note: we use inpCA here

            // cur = ln_0_w*cur + ln_0_b
            cur = ggml_add(ctx0,
                    ggml_mul(ctx0,
                        cur,
                        layer.cross_attn_ln_0_w),
                    layer.cross_attn_ln_0_b);
        }

        // cross-attention
        {
            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                    layer.cross_attn_q_w,
                    cur);

            Qcur = ggml_add(ctx0,
                        Qcur,
                        layer.cross_attn_q_b);

            Qcur = ggml_scale(ctx0, Qcur, KQscale);

            // Kcross is already scaled
            struct ggml_tensor * Kcross =
                ggml_view_3d(ctx0, wstate.kv_cross.k,
                        n_state/n_head, M, n_head,
                        ggml_element_size(wstate.kv_cross.k)*n_state,
                        ggml_element_size(wstate.kv_cross.k)*n_state/n_head,
                        ggml_element_size(wstate.kv_cross.k)*n_state*M*il);

            //struct ggml_tensor * Vcross =
            //    ggml_reshape_3d(ctx0,
            //            ggml_view_1d(ctx0, wstate.kv_cross.v, M*n_state, il*M*ggml_element_size(wstate.kv_cross.v)*n_state),
            //            n_state/n_head, n_head, M);

            //struct ggml_tensor * V_trans =
            //    ggml_cpy(ctx0,
            //            ggml_permute(ctx0, Vcross, 1, 2, 0, 3),
            //            ggml_new_tensor_3d(ctx0, Vcross->type, M, n_state/n_head, n_head));

            struct ggml_tensor * V =
                ggml_view_3d(ctx0, wstate.kv_cross.v,
                        M, n_state/n_head, n_head,
                        M*ggml_element_size(wstate.kv_cross.v),
                        M*ggml_element_size(wstate.kv_cross.v)*n_state/n_head,
                        il*M*ggml_element_size(wstate.kv_cross.v)*n_state);

            // ------

            struct ggml_tensor * Q =
                ggml_permute(ctx0,
                        ggml_reshape_3d(ctx0, Qcur, n_state/n_head, n_head, N),
                        0, 2, 1, 3);

            // K * Q
            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, Kcross, Q);

            //struct ggml_tensor * KQ_scaled =
            //    ggml_scale(ctx0,
            //            KQ,
            //            ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head))
            //            );

            // no masking for cross-attention
            //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);

            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ);

            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);

            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);

            // cur = KQV_merged.contiguous().view(n_state, N)
            cur = ggml_cpy(ctx0,
                    KQV_merged,
                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, N));
        }

        // projection
        {
            cur = ggml_mul_mat(ctx0,
                    layer.cross_attn_ln_1_w,
                    cur);

            cur = ggml_add(ctx0,
                    cur,
                    layer.cross_attn_ln_1_b);
        }

        // add the input
        cur = ggml_add(ctx0, cur, inpCA);

        struct ggml_tensor * inpFF = cur;

        // feed-forward network
        {
            // norm
            {
                cur = ggml_norm(ctx0, inpFF, hparams.eps);

                // cur = mlp_ln_w*cur + mlp_ln_b
                cur = ggml_add(ctx0,
                        ggml_mul(ctx0,
                            cur,
                            layer.mlp_ln_w),
                        layer.mlp_ln_b);
            }

            // fully connected
            cur = ggml_mul_mat(ctx0,
                    layer.mlp_0_w,
                    cur);

            cur = ggml_add(ctx0,
                    cur,
                    layer.mlp_0_b);

            // GELU activation
            cur = ggml_gelu(ctx0, cur);

            // projection
            cur = ggml_mul_mat(ctx0,
                    layer.mlp_1_w,
                    cur);

            cur = ggml_add(ctx0,
                    cur,
                    layer.mlp_1_b);
        }

        inpL = ggml_add(ctx0, cur, inpFF);
    }

    cur = inpL;

    // norm
    {
        cur = ggml_norm(ctx0, cur, hparams.eps);

        cur = ggml_add(ctx0,
                ggml_mul(ctx0,
                    cur,
                    model.d_ln_w),
                model.d_ln_b);
    }

    // compute logits only for the last token
    // comment this line to compute logits for all N tokens
    // might be useful in the future
    cur = ggml_view_2d(ctx0, cur, cur->ne[0], 1, cur->nb[1], (cur->ne[1] - 1)*cur->nb[1]);

    struct ggml_tensor * logits = ggml_mul_mat(ctx0, model.d_te, cur);

    ggml_build_forward_expand(gf, logits);

    ggml_free(ctx0);

    return gf;
}

// evaluate the decoder
//
// given text prompt + audio features -> computes the logits for the next token
//
//   - model:      the model
//   - n_threads:  number of threads to use
//   - tokens:     text prompt
//   - n_tokens:   number of tokens in the prompt
//   - n_past:     number of past tokens to prefix the prompt with
//
static bool whisper_decode_internal(
        whisper_context & wctx,
          whisper_state & wstate,
        whisper_decoder & decoder,
    const whisper_token * tokens,
              const int   n_tokens,
              const int   n_past,
              const int   n_threads) {
    const int64_t t_start_us = ggml_time_us();

    const auto & model   = wctx.model;
    const auto & hparams = model.hparams;

    const int n_vocab = hparams.n_vocab;

    auto & logits_out = wstate.logits;

    struct ggml_tensor * logits;

    // decoder
    {
        auto & alloc = wstate.alloc_decode.alloc;

        ggml_allocr_reset(alloc);

        ggml_cgraph * gf = whisper_build_graph_decoder(wctx, wstate, decoder, tokens, n_tokens, n_past);

        ggml_allocr_alloc_graph(alloc, gf);

        logits = gf->nodes[gf->n_nodes - 1];

#ifdef GGML_USE_METAL
        if (wstate.ctx_metal) {
            ggml_metal_set_n_cb     (wstate.ctx_metal, n_threads);
            ggml_metal_graph_compute(wstate.ctx_metal, gf);
        } else {
            ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads);
        }
#else
        ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads);
#endif
    }

    // extract logits for all N tokens
    //logits_out.resize(n_tokens*n_vocab);
    //memcpy(logits_out.data(), ggml_get_data(logits), sizeof(float)*n_tokens*n_vocab);

    // extract logits only for the last token
    logits_out.resize(n_vocab);
    memcpy(logits_out.data(), ggml_get_data(logits), sizeof(float)*n_vocab);

    if (n_tokens > 1) {
        //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
        //        ggml_used_mem(ctx0)/1024.0/1024.0,
        //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
        //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
        //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
        //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
    }

    if (n_tokens == 1) {
        wstate.t_decode_us += ggml_time_us() - t_start_us;
        wstate.n_decode++;
    } else {
        wstate.t_prompt_us += ggml_time_us() - t_start_us;
        wstate.n_prompt++;
    }

    return true;
}


//  500 -> 00:05.000
// 6000 -> 01:00.000
static std::string to_timestamp(int64_t t, bool comma = false) {
    int64_t msec = t * 10;
    int64_t hr = msec / (1000 * 60 * 60);
    msec = msec - hr * (1000 * 60 * 60);
    int64_t min = msec / (1000 * 60);
    msec = msec - min * (1000 * 60);
    int64_t sec = msec / 1000;
    msec = msec - sec * 1000;

    char buf[32];
    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);

    return std::string(buf);
}

#define SIN_COS_N_COUNT WHISPER_N_FFT
static float sin_vals[SIN_COS_N_COUNT];
static float cos_vals[SIN_COS_N_COUNT];

// In FFT, we frequently use sine and cosine operations with the same values.
// We can use precalculated values to speed up the process.
static void fill_sin_cos_table() {
    static bool is_filled = false;
    if (is_filled) return;
    for (int i = 0; i < SIN_COS_N_COUNT; i++) {
        double theta = (2*M_PI*i)/SIN_COS_N_COUNT;
        sin_vals[i] = sinf(theta);
        cos_vals[i] = cosf(theta);
    }
    is_filled = true;
}

// naive Discrete Fourier Transform
// input is real-valued
// output is complex-valued
static void dft(const std::vector<float> & in, std::vector<float> & out) {
    int N = in.size();

    out.resize(N*2);
    const int sin_cos_step = SIN_COS_N_COUNT / N;

    for (int k = 0; k < N; k++) {
        float re = 0;
        float im = 0;

        for (int n = 0; n < N; n++) {
            int idx = (k * n * sin_cos_step) % (SIN_COS_N_COUNT); // t = 2*M_PI*k*n/N
            re += in[n]*cos_vals[idx]; // cos(t)
            im -= in[n]*sin_vals[idx]; // sin(t)
        }

        out[k*2 + 0] = re;
        out[k*2 + 1] = im;
    }
}

// Cooley-Tukey FFT
// poor man's implementation - use something better
// input is real-valued
// output is complex-valued
static void fft(const std::vector<float> & in, std::vector<float> & out) {
    out.resize(in.size()*2);

    int N = in.size();

    if (N == 1) {
        out[0] = in[0];
        out[1] = 0;
        return;
    }

    if (N%2 == 1) {
        dft(in, out);
        return;
    }

    std::vector<float> even;
    std::vector<float> odd;

    even.reserve(N/2);
    odd.reserve(N/2);

    for (int i = 0; i < N; i++) {
        if (i % 2 == 0) {
            even.push_back(in[i]);
        } else {
            odd.push_back(in[i]);
        }
    }

    std::vector<float> even_fft;
    std::vector<float> odd_fft;

    fft(even, even_fft);
    fft(odd, odd_fft);

    const int sin_cos_step = SIN_COS_N_COUNT / N;
    for (int k = 0; k < N/2; k++) {
        int idx = k * sin_cos_step; // t = 2*M_PI*k/N
        float re = cos_vals[idx]; // cos(t)
        float im = -sin_vals[idx]; // sin(t)

        float re_odd = odd_fft[2*k + 0];
        float im_odd = odd_fft[2*k + 1];

        out[2*k + 0] = even_fft[2*k + 0] + re*re_odd - im*im_odd;
        out[2*k + 1] = even_fft[2*k + 1] + re*im_odd + im*re_odd;

        out[2*(k + N/2) + 0] = even_fft[2*k + 0] - re*re_odd + im*im_odd;
        out[2*(k + N/2) + 1] = even_fft[2*k + 1] - re*im_odd - im*re_odd;
    }
}

static bool hann_window(int length, bool periodic, std::vector<float> & output) {
    if (output.size() < static_cast<size_t>(length)) {
        output.resize(length);
    }
    int offset = -1;
    if (periodic) {
        offset = 0;
    }
    for (int i = 0; i < length; i++) {
        output[i] = 0.5*(1.0 - cosf((2.0*M_PI*i)/(length + offset)));
    }

    return true;
}

static void log_mel_spectrogram_worker_thread(int ith, const std::vector<float> & hann, const std::vector<float> & samples,
                                              int n_samples, int frame_size, int frame_step, int n_threads,
                                              const whisper_filters & filters, whisper_mel & mel) {
    std::vector<float> fft_in(frame_size, 0.0);
    std::vector<float> fft_out(2 * frame_step);
    // make sure n_fft == 1 + (WHISPER_N_FFT / 2), bin_0 to bin_nyquist
    int n_fft = 1 + (frame_size / 2);
    int i = ith;

    // calculate FFT only when fft_in are not all zero
    for (; i < std::min(n_samples / frame_step + 1, mel.n_len); i += n_threads) {
        const int offset = i * frame_step;

        // apply Hanning window (~10% faster)
        for (int j = 0; j < std::min(frame_size, n_samples - offset); j++) {
            fft_in[j] = hann[j] * samples[offset + j];
        }
        // fill the rest with zeros
        if (n_samples - offset < frame_size) {
            std::fill(fft_in.begin() + (n_samples - offset), fft_in.end(), 0.0);
        }

        // FFT
        fft(fft_in, fft_out);

        // Calculate modulus^2 of complex numbers
        // Use pow(fft_out[2 * j + 0], 2) + pow(fft_out[2 * j + 1], 2) causes inference quality problem? Interesting.
        for (int j = 0; j < frame_size; j++) {
            fft_out[j] = (fft_out[2 * j + 0] * fft_out[2 * j + 0] + fft_out[2 * j + 1] * fft_out[2 * j + 1]);
        }

        // mel spectrogram
        for (int j = 0; j < mel.n_mel; j++) {
            double sum = 0.0;

            // unroll loop (suggested by GH user @lunixbochs)
            int k = 0;
            for (k = 0; k < n_fft - 3; k += 4) {
                sum +=
                        fft_out[k + 0] * filters.data[j * n_fft + k + 0] +
                        fft_out[k + 1] * filters.data[j * n_fft + k + 1] +
                        fft_out[k + 2] * filters.data[j * n_fft + k + 2] +
                        fft_out[k + 3] * filters.data[j * n_fft + k + 3];
            }

            // handle n_fft remainder
            for (; k < n_fft; k++) {
                sum += fft_out[k] * filters.data[j * n_fft + k];
            }

            sum = log10(std::max(sum, 1e-10));

            mel.data[j * mel.n_len + i] = sum;
        }
    }

    // Otherwise fft_out are all zero
    double sum = log10(1e-10);
    for (; i < mel.n_len; i += n_threads) {
        for (int j = 0; j < mel.n_mel; j++) {
            mel.data[j * mel.n_len + i] = sum;
        }
    }
}

// ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L110-L157
static bool log_mel_spectrogram(
              whisper_state & wstate,
              const float * samples,
              const int   n_samples,
              const int   /*sample_rate*/,
              const int   frame_size,
              const int   frame_step,
              const int   n_mel,
              const int   n_threads,
              const whisper_filters & filters,
              const bool   debug,
              whisper_mel & mel) {
    const int64_t t_start_us = ggml_time_us();

    // Hanning window (Use cosf to eliminate difference)
    // ref: https://pytorch.org/docs/stable/generated/torch.hann_window.html
    // ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L147
    std::vector<float> hann;
    hann_window(frame_size, true, hann);


    // Calculate the length of padding
    int64_t stage_1_pad = WHISPER_SAMPLE_RATE * 30;
    int64_t stage_2_pad = frame_size / 2;

    // Initialize a vector and copy data from C array to it.
    std::vector<float> samples_padded;
    samples_padded.resize(n_samples + stage_1_pad + stage_2_pad * 2);
    std::copy(samples, samples + n_samples, samples_padded.begin() + stage_2_pad);

    // pad 30 seconds of zeros at the end of audio (480,000 samples) + reflective pad 200 samples at the end of audio
    std::fill(samples_padded.begin() + n_samples + stage_2_pad, samples_padded.begin() + n_samples + stage_1_pad + 2 * stage_2_pad, 0);

    // reflective pad 200 samples at the beginning of audio
    std::reverse_copy(samples + 1, samples + 1 + stage_2_pad, samples_padded.begin());

    mel.n_mel     = n_mel;
    // https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/SpectralOps.cpp#L936
    // Calculate number of frames + remove the last frame
    mel.n_len     = (samples_padded.size() - frame_size) / frame_step;
    // Calculate semi-padded sample length to ensure compatibility
    mel.n_len_org = 1 + (n_samples + stage_2_pad - frame_size) / frame_step;
    mel.data.resize(mel.n_mel * mel.n_len);


    {
        std::vector<std::thread> workers(n_threads - 1);
        for (int iw = 0; iw < n_threads - 1; ++iw) {
            workers[iw] = std::thread(
                    log_mel_spectrogram_worker_thread, iw + 1, std::cref(hann), samples_padded,
                    n_samples + stage_2_pad, frame_size, frame_step, n_threads,
                    std::cref(filters), std::ref(mel));
        }

        // main thread
        log_mel_spectrogram_worker_thread(0, hann, samples_padded, n_samples + stage_2_pad, frame_size, frame_step, n_threads, filters, mel);

        for (int iw = 0; iw < n_threads - 1; ++iw) {
            workers[iw].join();
        }
    }

    // clamping and normalization
    double mmax = -1e20;
    for (int i = 0; i < mel.n_mel*mel.n_len; i++) {
        if (mel.data[i] > mmax) {
            mmax = mel.data[i];
        }
    }

    mmax -= 8.0;

    for (int i = 0; i < mel.n_mel*mel.n_len; i++) {
        if (mel.data[i] < mmax) {
            mel.data[i] = mmax;
        }

        mel.data[i] = (mel.data[i] + 4.0)/4.0;
    }

    wstate.t_mel_us += ggml_time_us() - t_start_us;

    // Dump log_mel_spectrogram
    if (debug) {
        std::ofstream outFile("log_mel_spectrogram.json");
        outFile << "[";
        for (uint64_t i = 0; i < mel.data.size() - 1; i++) {
            outFile << mel.data[i] << ", ";
        }
        outFile << mel.data[mel.data.size() - 1] << "]";
        outFile.close();
    }

    return true;
}

// split text into tokens
//
// ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
//
// Regex (Python):
// r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
//
// Regex (C++):
// R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"
//
static std::vector<whisper_vocab::id> tokenize(const whisper_vocab & vocab, const std::string & text) {
    std::vector<std::string> words;

    // first split the text into words
    {
        std::string str = text;
        std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";

        std::regex re(pat);
        std::smatch m;

        while (std::regex_search(str, m, re)) {
            for (auto x : m) {
                words.push_back(x);
            }
            str = m.suffix();
        }
    }

    // find the longest tokens that form the words:
    std::vector<whisper_vocab::id> tokens;
    for (const auto & word : words) {
        if (word.empty()) continue;

        int i = 0;
        int n = word.size();
        while (i < n) {
            int j = n;
            bool found = false;
            while (j > i) {
                auto sub = word.substr(i, j-i);
                auto it = vocab.token_to_id.find(sub);
                if (it != vocab.token_to_id.end()) {
                    tokens.push_back(it->second);
                    i = j;
                    found = true;
                    break;
                }
                --j;
            }
            if (!found) {
                log("unknown token\n");
                ++i;
            }
        }
    }

    return tokens;
}

//
// interface implementation
//

#ifdef WHISPER_USE_COREML
// replace .bin with -encoder.mlmodelc
static std::string whisper_get_coreml_path_encoder(std::string path_bin) {
    auto pos = path_bin.rfind('.');
    if (pos != std::string::npos) {
        path_bin = path_bin.substr(0, pos);
    }

    // match "-qx_x"
    pos = path_bin.rfind('-');
    if (pos != std::string::npos) {
        auto sub = path_bin.substr(pos);
        if (sub.size() == 5 && sub[1] == 'q' && sub[3] == '_') {
            path_bin = path_bin.substr(0, pos);
        }
    }

    path_bin += "-encoder.mlmodelc";

    return path_bin;
}
#endif

#ifdef WHISPER_USE_OPENVINO
// replace .bin with-encoder-openvino.xml
static std::string whisper_openvino_get_path_encoder(std::string path_bin) {
    auto pos = path_bin.rfind('.');
    if (pos != std::string::npos) {
        path_bin = path_bin.substr(0, pos);
    }

    path_bin += "-encoder-openvino.xml";

    return path_bin;
}

static std::string whisper_openvino_get_path_cache(std::string path_bin) {
    auto pos = path_bin.rfind('.');
    if (pos != std::string::npos) {
        path_bin = path_bin.substr(0, pos);
    }

    path_bin += "-encoder-openvino-cache";

    return path_bin;
}
#endif

struct whisper_state * whisper_init_state(whisper_context * ctx) {
    fill_sin_cos_table();
    whisper_state * state = new whisper_state;

    if (!kv_cache_init(ctx->model.hparams, state->decoders[0].kv_self, ctx->itype, ctx->model.hparams.n_text_ctx)) {
        log("%s: kv_cache_init() failed for self-attention cache\n", __func__);
        delete state;
        return nullptr;
    }

    {
        const size_t memory_size = ggml_nbytes(state->decoders[0].kv_self.k) + ggml_nbytes(state->decoders[0].kv_self.v);
        log("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
    }

    if (!kv_cache_init(ctx->model.hparams, state->kv_cross, ctx->itype, ctx->model.hparams.n_audio_ctx)) {
        log("%s: kv_cache_init() failed for cross-attention cache\n", __func__);
        delete state;
        return nullptr;
    }

    {
        const size_t memory_size = ggml_nbytes(state->kv_cross.k) + ggml_nbytes(state->kv_cross.v);
        log("%s: kv cross size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
    }

#ifdef WHISPER_USE_COREML
    const auto path_coreml = whisper_get_coreml_path_encoder(ctx->path_model);

    log("%s: loading Core ML model from '%s'\n", __func__, path_coreml.c_str());
    log("%s: first run on a device may take a while ...\n", __func__);

    state->ctx_coreml = whisper_coreml_init(path_coreml.c_str());
    if (!state->ctx_coreml) {
        log("%s: failed to load Core ML model from '%s'\n", __func__, path_coreml.c_str());
#ifndef WHISPER_COREML_ALLOW_FALLBACK
        delete state;
        return nullptr;
#endif
    } else {
        log("%s: Core ML model loaded\n", __func__);
    }
#endif

    state->logits.reserve(ctx->vocab.n_vocab * ctx->model.hparams.n_text_ctx);

    state->logits_id.reserve(ctx->model.hparams.n_vocab);

    // TAGS: WHISPER_DECODER_INIT
    state->decoders[0].sequence.tokens.reserve(ctx->model.hparams.n_text_ctx);

    state->decoders[0].probs.reserve   (ctx->vocab.n_vocab);
    state->decoders[0].logits.reserve  (ctx->vocab.n_vocab);
    state->decoders[0].logprobs.reserve(ctx->vocab.n_vocab);

    // conv allocator
    {
        whisper_allocr_graph_init(state->alloc_conv,
                [&]() {
                    return whisper_build_graph_conv(*ctx, *state, 0);
                });

        log("%s: compute buffer (conv)   = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_conv) / 1024.0 / 1024.0);
    }

    // encoder allocator
    if (!whisper_encode_external(*state)) {
        whisper_allocr_graph_init(state->alloc_encode,
                [&]() {
                    return whisper_build_graph_encoder(*ctx, *state);
                });

        log("%s: compute buffer (encode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_encode) / 1024.0 / 1024.0);
    }

    // cross allocator
    {
        whisper_allocr_graph_init(state->alloc_cross,
                [&]() {
                    return whisper_build_graph_cross(*ctx, *state);
                });

        log("%s: compute buffer (cross)  = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_cross) / 1024.0 / 1024.0);
    }

    // decoder allocator
    {
        whisper_allocr_graph_init(state->alloc_decode,
                [&]() {
                    const auto & hparams = ctx->model.hparams;

                    // TODO: make sure this is the worst-case scenario
                    const int n_tokens = hparams.n_text_ctx;
                    const int n_past   = 0;

                    return whisper_build_graph_decoder(*ctx, *state, state->decoders[0], nullptr, n_tokens, n_past);
                });

        log("%s: compute buffer (decode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_decode) / 1024.0 / 1024.0);
    }

#ifdef GGML_USE_METAL
    state->ctx_metal = ggml_metal_init(1);
    if (!state->ctx_metal) {
        log("%s: ggml_metal_init() failed\n", __func__);
        delete state;
        return nullptr;
    }

    log("%s: Metal context initialized\n", __func__);

    // this allocates all Metal resources and memory buffers

    void * data_ptr  = NULL;
    size_t data_size = 0;

    // TODO: add mmap support
    //if (params.use_mmap) {
    //    data_ptr  = ctx->model.mapping->addr;
    //    data_size = ctx->model.mapping->size;
    //} else {
    //    data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
    //    data_size = ggml_get_mem_size  (ctx->model.ctx);
    //}

    data_ptr  = ggml_get_mem_buffer(ctx->model.ctx);
    data_size = ggml_get_mem_size  (ctx->model.ctx);

    const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);

    log("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);

#define WHISPER_METAL_CHECK_BUF(result)              \
    if (!(result)) {                                 \
        log("%s: failed to add metal buffer\n", __func__); \
        delete state;                                \
        return nullptr;                              \
    }

    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "data", data_ptr, data_size, max_size));

    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "meta_conv",   state->alloc_conv.meta.data(),   state->alloc_conv.meta.size(),   0));
    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "meta_encode", state->alloc_encode.meta.data(), state->alloc_encode.meta.size(), 0));
    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "meta_cross",  state->alloc_cross.meta.data(),  state->alloc_cross.meta.size(),  0));
    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "meta_decode", state->alloc_decode.meta.data(), state->alloc_decode.meta.size(), 0));

    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "data_conv",   state->alloc_conv.data.data(),   state->alloc_conv.data.size(),   0));
    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "data_encode", state->alloc_encode.data.data(), state->alloc_encode.data.size(), 0));
    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "data_cross",  state->alloc_cross.data.data(),  state->alloc_cross.data.size(),  0));
    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "data_decode", state->alloc_decode.data.data(), state->alloc_decode.data.size(), 0));

    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "kv_cross",  state->kv_cross.buf.data(), state->kv_cross.buf.size(), 0));

    WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, "kv_self_0", state->decoders[0].kv_self.buf.data(), state->decoders[0].kv_self.buf.size(), 0));
#undef WHISPER_METAL_CHECK_BUF
#endif

    state->rng = std::mt19937(0);

    return state;
}

int whisper_ctx_init_openvino_encoder(
        struct whisper_context * ctx,
                    const char * model_path,
                    const char * device,
                    const char * cache_dir) {
#ifndef WHISPER_USE_OPENVINO
    (void)(ctx);
    (void)(model_path);
    (void)(device);
    (void)(cache_dir);

    return 1;
#else
    if (!model_path && ctx->path_model.empty()) {
        log("%s: model_path is nullptr, and ctx has no model_path set.\n", __func__);
        return 1;
    }

    std::string path_encoder;
    if (!model_path) {
        //if model_path is not set, attempt to find it in the same directory as ggml-<model>.bin model
        path_encoder = whisper_openvino_get_path_encoder(ctx->path_model);
    } else {
        path_encoder = model_path;
    }

    std::string path_cache;
    if (!cache_dir) {
        //if cache_dir is not set, set it as a dir residing next to ggml-<model>.bin
        path_cache = whisper_openvino_get_path_cache(ctx->path_model);
    } else {
        path_cache = cache_dir;
    }

    log("%s: loading OpenVINO model from '%s'\n", __func__, path_encoder.c_str());
    log("%s: first run on a device may take a while ...\n", __func__);

    ctx->state->ctx_openvino = whisper_openvino_init(path_encoder.c_str(), device, path_cache.c_str());
    if (!ctx->state->ctx_openvino) {
        log("%s: failed to init OpenVINO encoder from '%s'\n", __func__, path_encoder.c_str());
        return 1;
    } else {
        log("%s: OpenVINO model loaded\n", __func__);
    }

    return 0;
#endif
}

struct whisper_context * whisper_init_from_file_no_state(const char * path_model) {
    log("%s: loading model from '%s'\n", __func__, path_model);

    auto fin = std::ifstream(path_model, std::ios::binary);
    if (!fin) {
        log("%s: failed to open '%s'\n", __func__, path_model);
        return nullptr;
    }

    whisper_model_loader loader = {};

    loader.context = &fin;

    loader.read = [](void * ctx, void * output, size_t read_size) {
        std::ifstream * fin = (std::ifstream*)ctx;
        fin->read((char *)output, read_size);
        return read_size;
    };

    loader.eof = [](void * ctx) {
        std::ifstream * fin = (std::ifstream*)ctx;
        return fin->eof();
    };

    loader.close = [](void * ctx) {
        std::ifstream * fin = (std::ifstream*)ctx;
        fin->close();
    };

    auto ctx = whisper_init_no_state(&loader);

    if (ctx) {
        ctx->path_model = path_model;
    }

    return ctx;
}

struct whisper_context * whisper_init_from_buffer_no_state(void * buffer, size_t buffer_size) {
    struct buf_context {
        uint8_t* buffer;
        size_t size;
        size_t current_offset;
    };

    buf_context ctx = { reinterpret_cast<uint8_t*>(buffer), buffer_size, 0 };

    log("%s: loading model from buffer\n", __func__);

    whisper_model_loader loader = {};

    loader.context = &ctx;

    loader.read = [](void * ctx, void * output, size_t read_size) {
        buf_context * buf = reinterpret_cast<buf_context *>(ctx);

        size_t size_to_copy = buf->current_offset + read_size < buf->size ? read_size : buf->size - buf->current_offset;

        memcpy(output, buf->buffer + buf->current_offset, size_to_copy);
        buf->current_offset += size_to_copy;

        return size_to_copy;
    };

    loader.eof = [](void * ctx) {
        buf_context * buf = reinterpret_cast<buf_context *>(ctx);

        return buf->current_offset >= buf->size;
    };

    loader.close = [](void * /*ctx*/) { };

    return whisper_init_no_state(&loader);
}

struct whisper_context * whisper_init_no_state(struct whisper_model_loader * loader) {
    ggml_time_init();

    whisper_context * ctx = new whisper_context;

    if (!whisper_model_load(loader, *ctx)) {
        loader->close(loader->context);
        log("%s: failed to load model\n", __func__);
        delete ctx;
        return nullptr;
    }

    loader->close(loader->context);

    return ctx;
}

struct whisper_context * whisper_init_from_file(const char * path_model) {
    whisper_context * ctx = whisper_init_from_file_no_state(path_model);
    if (!ctx) {
        return nullptr;
    }

    ctx->state = whisper_init_state(ctx);
    if (!ctx->state) {
        whisper_free(ctx);
        return nullptr;
    }

    return ctx;
}

struct whisper_context * whisper_init_from_buffer(void * buffer, size_t buffer_size) {
    whisper_context * ctx = whisper_init_from_buffer_no_state(buffer, buffer_size);
    if (!ctx) {
        return nullptr;
    }

    ctx->state = whisper_init_state(ctx);
    if (!ctx->state) {
        whisper_free(ctx);
        return nullptr;
    }

    return ctx;
}

struct whisper_context * whisper_init(struct whisper_model_loader * loader) {
    whisper_context * ctx = whisper_init_no_state(loader);
    if (!ctx) {
        return nullptr;
    }

    ctx->state = whisper_init_state(ctx);
    if (!ctx->state) {
        whisper_free(ctx);
        return nullptr;
    }

    return ctx;
}

void whisper_free_state(struct whisper_state * state)
{
    if (state) {
        kv_cache_free(state->kv_cross);

        for (int i = 0; i < WHISPER_MAX_DECODERS; ++i) {
            kv_cache_free(state->decoders[i].kv_self);
        }

#ifdef WHISPER_USE_COREML
        if (state->ctx_coreml != nullptr) {
            whisper_coreml_free(state->ctx_coreml);
            state->ctx_coreml = nullptr;
        }
#endif

#ifdef GGML_USE_METAL
        if (state->ctx_metal) {
            ggml_metal_free(state->ctx_metal);
            state->ctx_metal = nullptr;
        }
#endif

#ifdef WHISPER_USE_OPENVINO
        if (state->ctx_openvino != nullptr) {
            whisper_openvino_free(state->ctx_openvino);
            state->ctx_openvino = nullptr;
        }
#endif

        whisper_allocr_free(state->alloc_conv);
        whisper_allocr_free(state->alloc_decode);
        whisper_allocr_free(state->alloc_cross);
        whisper_allocr_free(state->alloc_encode);

        delete state;
    }
}

void whisper_free(struct whisper_context * ctx) {
    if (ctx) {
        if (ctx->model.ctx) {
            ggml_free(ctx->model.ctx);
        }
        if (ctx->model.buf) {
            delete ctx->model.buf;
        }

        whisper_free_state(ctx->state);

        delete ctx;
    }
}

void whisper_free_params(struct whisper_full_params * params) {
    if (params) {
        delete params;
    }
}

int whisper_pcm_to_mel_with_state(struct whisper_context * ctx, struct whisper_state * state, const float * samples, int n_samples, int n_threads) {
    if (!log_mel_spectrogram(*state, samples, n_samples, WHISPER_SAMPLE_RATE, WHISPER_N_FFT, WHISPER_HOP_LENGTH, WHISPER_N_MEL, n_threads, ctx->model.filters, false, state->mel)) {
        log("%s: failed to compute mel spectrogram\n", __func__);
        return -1;
    }

    return 0;
}

int whisper_pcm_to_mel(struct whisper_context * ctx, const float * samples, int n_samples, int n_threads) {
    return whisper_pcm_to_mel_with_state(ctx, ctx->state, samples, n_samples, n_threads);
}

// same as whisper_pcm_to_mel, but applies a Phase Vocoder to speed up the audio x2 (PV without phase lock is not good)
int whisper_pcm_to_mel_phase_vocoder_with_state(struct whisper_context * ctx, struct whisper_state * state, const float * samples, int n_samples, int n_threads) {
    if (!log_mel_spectrogram(*state, samples, n_samples, WHISPER_SAMPLE_RATE, 2 * WHISPER_N_FFT, 2 * WHISPER_HOP_LENGTH, WHISPER_N_MEL, n_threads, ctx->model.filters, false, state->mel)) {
        log("%s: failed to compute mel spectrogram\n", __func__);
        return -1;
    }

    return 0;
}

// same as whisper_pcm_to_mel, but applies a Phase Vocoder to speed up the audio x2 (PV without phase lock is not good)
int whisper_pcm_to_mel_phase_vocoder(struct whisper_context * ctx, const float * samples, int n_samples, int n_threads) {
    return whisper_pcm_to_mel_phase_vocoder_with_state(ctx, ctx->state, samples, n_samples, n_threads);
}

// same as whisper_pcm_to_mel, but applies WSOLA to speed up the audio x2
// TODO

// same as whisper_pcm_to_mel, but applies HPTSM to speed up the audio x2
// TODO

// same as whisper_pcm_to_mel, but applies PV (with phase lock) to speed up the audio x2
// TODO

int whisper_set_mel_with_state(
        struct whisper_context * /*ctx*/,
          struct whisper_state * state,
                   const float * data,
                           int   n_len,
                           int   n_mel) {
    if (n_mel != WHISPER_N_MEL) {
        log("%s: invalid number of mel bands: %d (expected %d)\n", __func__, n_mel, WHISPER_N_MEL);
        return -1;
    }

    state->mel.n_len     = n_len;
    state->mel.n_len_org = n_len;
    state->mel.n_mel     = n_mel;

    state->mel.data.resize(n_len*n_mel);
    memcpy(state->mel.data.data(), data, n_len*n_mel*sizeof(float));

    return 0;
}

int whisper_set_mel(
        struct whisper_context * ctx,
        const float * data,
        int n_len,
        int n_mel) {
    return whisper_set_mel_with_state(ctx, ctx->state, data, n_len, n_mel);
}

int whisper_encode_with_state(struct whisper_context * ctx, struct whisper_state * state, int offset, int n_threads) {
    if (!whisper_encode_internal(*ctx, *state, offset, n_threads)) {
        log("%s: failed to eval\n", __func__);
        return -1;
    }

    return 0;
}

int whisper_encode(struct whisper_context * ctx, int offset, int n_threads) {
    if (!whisper_encode_internal(*ctx, *ctx->state, offset, n_threads)) {
        log("%s: failed to eval\n", __func__);
        return -1;
    }

    return 0;
}

int whisper_decode_with_state(struct whisper_context * ctx, struct whisper_state * state, const whisper_token * tokens, int n_tokens, int n_past, int n_threads) {
    const int selected_decoder_id = 0;

    if (!whisper_decode_internal(*ctx, *state, state->decoders[selected_decoder_id], tokens, n_tokens, n_past, n_threads)) {
        log("%s: failed to eval\n", __func__);
        return 1;
    }

    return 0;
}

int whisper_decode(struct whisper_context * ctx, const whisper_token * tokens, int n_tokens, int n_past, int n_threads) {
    // TODO: add selected_decoder_id to state
    const int selected_decoder_id = 0;

    if (ctx->state == nullptr) {
        log("%s: ERROR state was not loaded.\n", __func__);
        return false;
    }

    if (!whisper_decode_internal(*ctx, *ctx->state, ctx->state->decoders[selected_decoder_id], tokens, n_tokens, n_past, n_threads)) {
        log("%s: failed to eval\n", __func__);
        return 1;
    }

    return 0;
}

int whisper_tokenize(struct whisper_context * ctx, const char * text, whisper_token * tokens, int n_max_tokens) {
    const auto res = tokenize(ctx->vocab, text);

    if (n_max_tokens < (int) res.size()) {
        log("%s: too many resulting tokens: %d (max %d)\n", __func__, (int) res.size(), n_max_tokens);
        return -1;
    }

    for (int i = 0; i < (int) res.size(); i++) {
        tokens[i] = res[i];
    }

    return res.size();
}

int whisper_lang_max_id() {
    auto max_id = 0;
    for (const auto & kv : g_lang) {
        max_id = std::max(max_id, kv.second.first);
    }

    return max_id;
}

int whisper_lang_id(const char * lang) {
    if (!g_lang.count(lang)) {
        for (const auto & kv : g_lang) {
            if (kv.second.second == lang) {
                return kv.second.first;
            }
        }

        log("%s: unknown language '%s'\n", __func__, lang);
        return -1;
    }
    return g_lang.at(lang).first;
}

const char * whisper_lang_str(int id) {
    for (const auto & kv : g_lang) {
        if (kv.second.first == id) {
            return kv.first.c_str();
        }
    }

    log("%s: unknown language id %d\n", __func__, id);
    return nullptr;
}

int whisper_lang_auto_detect_with_state(
        struct whisper_context * ctx,
          struct whisper_state * state,
                           int   offset_ms,
                           int   n_threads,
                         float * lang_probs) {
    const int seek = offset_ms/10;

    if (seek < 0) {
        log("%s: offset %dms is before the start of the audio\n", __func__, offset_ms);
        return -1;
    }

    if (seek >= state->mel.n_len_org) {
        log("%s: offset %dms is past the end of the audio (%dms)\n", __func__, offset_ms, state->mel.n_len_org*10);
        return -2;
    }

    // run the encoder
    if (whisper_encode_with_state(ctx, state, seek, n_threads) != 0) {
        log("%s: failed to encode\n", __func__);
        return -6;
    }

    const std::vector<whisper_token> prompt = { whisper_token_sot(ctx) };

    if (whisper_decode_with_state(ctx, state, prompt.data(), prompt.size(), 0, n_threads) != 0) {
        log("%s: failed to decode\n", __func__);
        return -7;
    }

    auto & logits_id = state->logits_id;
    logits_id.clear();

    for (const auto & kv : g_lang) {
        const auto token_lang = whisper_token_lang(ctx, kv.second.first);
        logits_id.emplace_back(state->logits[token_lang], kv.second.first);
    }

    // sort descending
    {
        using pair_type = std::remove_reference<decltype(logits_id)>::type::value_type;
        std::sort(logits_id.begin(), logits_id.end(), [](const pair_type & a, const pair_type & b) {
            return a.first > b.first;
        });
    }

    // softmax
    {
        const auto max = logits_id[0].first;

        double sum = 0.0f;
        for (auto & kv : logits_id) {
            kv.first = exp(kv.first - max);
            sum += kv.first;
        }

        for (auto & kv : logits_id) {
            kv.first /= sum;
        }
    }

    {
        for (const auto & prob : logits_id) {
            if (lang_probs) {
                lang_probs[prob.second] = prob.first;
            }

            //printf("%s: lang %2d (%3s): %f\n", __func__, prob.second, whisper_lang_str(prob.second), prob.first);
        }
    }

    return logits_id[0].second;
}

int whisper_lang_auto_detect(
        struct whisper_context * ctx,
                           int   offset_ms,
                           int   n_threads,
                         float * lang_probs) {
    return whisper_lang_auto_detect_with_state(ctx, ctx->state, offset_ms, n_threads, lang_probs);
}

int whisper_model_n_vocab(struct whisper_context * ctx) {
    return ctx->model.hparams.n_vocab;
}

int whisper_model_n_audio_ctx(struct whisper_context * ctx) {
    return ctx->model.hparams.n_audio_ctx;
}

int whisper_model_n_audio_state(struct whisper_context * ctx) {
    return ctx->model.hparams.n_audio_state;
}

int whisper_model_n_audio_head(struct whisper_context * ctx) {
    return ctx->model.hparams.n_audio_head;
}

int whisper_model_n_audio_layer(struct whisper_context * ctx) {
    return ctx->model.hparams.n_audio_layer;
}

int whisper_model_n_text_ctx(struct whisper_context * ctx) {
    return ctx->model.hparams.n_text_ctx;
}

int whisper_model_n_text_state(struct whisper_context * ctx) {
    return ctx->model.hparams.n_text_state;
}

int whisper_model_n_text_head(struct whisper_context * ctx) {
    return ctx->model.hparams.n_text_head;
}

int whisper_model_n_text_layer(struct whisper_context * ctx) {
    return ctx->model.hparams.n_text_layer;
}

int whisper_model_n_mels(struct whisper_context * ctx) {
    return ctx->model.hparams.n_mels;
}

int whisper_model_ftype(struct whisper_context * ctx) {
    return ctx->model.hparams.ftype;
}

int whisper_model_type(struct whisper_context * ctx) {
    return ctx->model.type;
}

const char *whisper_model_type_readable(struct whisper_context * ctx) {
    switch (ctx->model.type) {
    case e_model::MODEL_TINY:
        return "tiny";
    case e_model::MODEL_BASE:
        return "base";
    case e_model::MODEL_SMALL:
        return "small";
    case e_model::MODEL_MEDIUM:
        return "medium";
    case e_model::MODEL_LARGE:
        return "large";
    default:
        return "unknown";
    }
}

int whisper_n_len_from_state(struct whisper_state * state) {
    return state->mel.n_len_org;
}

int whisper_n_len(struct whisper_context * ctx) {
    return ctx->state->mel.n_len_org;
}

int whisper_n_vocab(struct whisper_context * ctx) {
    return ctx->vocab.n_vocab;
}

int whisper_n_text_ctx(struct whisper_context * ctx) {
    return ctx->model.hparams.n_text_ctx;
}

int whisper_n_audio_ctx(struct whisper_context * ctx) {
    return ctx->model.hparams.n_audio_ctx;
}

int whisper_is_multilingual(struct whisper_context * ctx) {
    return ctx->vocab.is_multilingual() ? 1 : 0;
}

float * whisper_get_logits(struct whisper_context * ctx) {
    return ctx->state->logits.data();
}

float * whisper_get_logits_from_state(struct whisper_state * state) {
    return state->logits.data();
}

const char * whisper_token_to_str(struct whisper_context * ctx, whisper_token token) {
    return ctx->vocab.id_to_token.at(token).c_str();
}

whisper_token whisper_token_eot(struct whisper_context * ctx) {
    return ctx->vocab.token_eot;
}

whisper_token whisper_token_sot(struct whisper_context * ctx) {
    return ctx->vocab.token_sot;
}

whisper_token whisper_token_solm(struct whisper_context * ctx) {
    return ctx->vocab.token_solm;
}

whisper_token whisper_token_prev(struct whisper_context * ctx) {
    return ctx->vocab.token_prev;
}

whisper_token whisper_token_nosp(struct whisper_context * ctx) {
    return ctx->vocab.token_nosp;
}

whisper_token whisper_token_not(struct whisper_context * ctx) {
    return ctx->vocab.token_not;
}

whisper_token whisper_token_beg(struct whisper_context * ctx) {
    return ctx->vocab.token_beg;
}

whisper_token whisper_token_lang(struct whisper_context * ctx, int lang_id) {
    return whisper_token_sot(ctx) + 1 + lang_id;
}

whisper_token whisper_token_translate(struct whisper_context * ctx) {
    return ctx->vocab.token_translate;
}

whisper_token whisper_token_transcribe(struct whisper_context * ctx) {
    return ctx->vocab.token_transcribe;
}

void whisper_print_timings(struct whisper_context * ctx) {
    const int64_t t_end_us = ggml_time_us();

    log("\n");
    log("%s:     load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0f);
    if (ctx->state != nullptr) {

        const int32_t n_sample = std::max(1, ctx->state->n_sample);
        const int32_t n_encode = std::max(1, ctx->state->n_encode);
        const int32_t n_decode = std::max(1, ctx->state->n_decode);
        const int32_t n_prompt = std::max(1, ctx->state->n_prompt);

        log("%s:     fallbacks = %3d p / %3d h\n", __func__, ctx->state->n_fail_p, ctx->state->n_fail_h);
        log("%s:      mel time = %8.2f ms\n", __func__, ctx->state->t_mel_us / 1000.0f);
        log("%s:   sample time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3f * ctx->state->t_sample_us, n_sample, 1e-3f * ctx->state->t_sample_us / n_sample);
        log("%s:   encode time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3f * ctx->state->t_encode_us, n_encode, 1e-3f * ctx->state->t_encode_us / n_encode);
        log("%s:   decode time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3f * ctx->state->t_decode_us, n_decode, 1e-3f * ctx->state->t_decode_us / n_decode);
        log("%s:   prompt time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3f * ctx->state->t_prompt_us, n_prompt, 1e-3f * ctx->state->t_prompt_us / n_prompt);
    }
    log("%s:    total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0f);
}

void whisper_reset_timings(struct whisper_context * ctx) {
    if (ctx->state != nullptr) {
        ctx->state->t_sample_us = 0;
        ctx->state->t_encode_us = 0;
        ctx->state->t_decode_us = 0;
        ctx->state->t_prompt_us = 0;
        ctx->state->n_sample = 0;
        ctx->state->n_encode = 0;
        ctx->state->n_decode = 0;
        ctx->state->n_prompt = 0;
    }
}

static int whisper_has_coreml(void) {
#ifdef WHISPER_USE_COREML
    return 1;
#else
    return 0;
#endif
}

static int whisper_has_openvino(void) {
#ifdef WHISPER_USE_OPENVINO
    return 1;
#else
    return 0;
#endif
}

const char * whisper_print_system_info(void) {
    static std::string s;

    s  = "";
    s += "AVX = "       + std::to_string(ggml_cpu_has_avx())       + " | ";
    s += "AVX2 = "      + std::to_string(ggml_cpu_has_avx2())      + " | ";
    s += "AVX512 = "    + std::to_string(ggml_cpu_has_avx512())    + " | ";
    s += "FMA = "       + std::to_string(ggml_cpu_has_fma())       + " | ";
    s += "NEON = "      + std::to_string(ggml_cpu_has_neon())      + " | ";
    s += "ARM_FMA = "   + std::to_string(ggml_cpu_has_arm_fma())   + " | ";
    s += "METAL = "     + std::to_string(ggml_cpu_has_metal())     + " | ";
    s += "F16C = "      + std::to_string(ggml_cpu_has_f16c())      + " | ";
    s += "FP16_VA = "   + std::to_string(ggml_cpu_has_fp16_va())   + " | ";
    s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";
    s += "BLAS = "      + std::to_string(ggml_cpu_has_blas())      + " | ";
    s += "SSE3 = "      + std::to_string(ggml_cpu_has_sse3())      + " | ";
    s += "SSSE3 = "     + std::to_string(ggml_cpu_has_ssse3())     + " | ";
    s += "VSX = "       + std::to_string(ggml_cpu_has_vsx())       + " | ";
    s += "COREML = "    + std::to_string(whisper_has_coreml())     + " | ";
    s += "OPENVINO = "  + std::to_string(whisper_has_openvino())   + " | ";

    return s.c_str();
}

////////////////////////////////////////////////////////////////////////////

struct whisper_full_params * whisper_full_default_params_by_ref(enum whisper_sampling_strategy strategy) {
    struct whisper_full_params params = whisper_full_default_params(strategy);

    struct whisper_full_params* result = new whisper_full_params();
    *result = params;
    return result;
}

struct whisper_full_params whisper_full_default_params(enum whisper_sampling_strategy strategy) {
    struct whisper_full_params result = {
        /*.strategy          =*/ strategy,

        /*.n_threads         =*/ std::min(4, (int32_t) std::thread::hardware_concurrency()),
        /*.n_max_text_ctx    =*/ 16384,
        /*.offset_ms         =*/ 0,
        /*.duration_ms       =*/ 0,

        /*.translate         =*/ false,
        /*.no_context        =*/ true,
        /*.single_segment    =*/ false,
        /*.print_special     =*/ false,
        /*.print_progress    =*/ true,
        /*.print_realtime    =*/ false,
        /*.print_timestamps  =*/ true,

        /*.token_timestamps  =*/ false,
        /*.thold_pt          =*/ 0.01f,
        /*.thold_ptsum       =*/ 0.01f,
        /*.max_len           =*/ 0,
        /*.split_on_word     =*/ false,
        /*.max_tokens        =*/ 0,

        /*.speed_up          =*/ false,
        /*.debug_mode        =*/ false,
        /*.audio_ctx         =*/ 0,

        /*.tdrz_enable       =*/ false,

        /*.initial_prompt    =*/ nullptr,
        /*.prompt_tokens     =*/ nullptr,
        /*.prompt_n_tokens   =*/ 0,

        /*.language          =*/ "en",
        /*.detect_language   =*/ false,

        /*.suppress_blank    =*/ true,
        /*.suppress_non_speech_tokens =*/ false,

        /*.temperature       =*/  0.0f,
        /*.max_initial_ts    =*/  1.0f,
        /*.length_penalty    =*/ -1.0f,

        /*.temperature_inc   =*/  0.4f,
        /*.entropy_thold     =*/  2.4f,
        /*.logprob_thold     =*/ -1.0f,
        /*.no_speech_thold   =*/  0.6f,

        /*.greedy            =*/ {
            /*.best_of   =*/ -1,
        },

        /*.beam_search      =*/ {
            /*.beam_size =*/ -1,

            /*.patience  =*/ -1.0f,
        },

        /*.new_segment_callback           =*/ nullptr,
        /*.new_segment_callback_user_data =*/ nullptr,

        /*.progress_callback           =*/ nullptr,
        /*.progress_callback_user_data =*/ nullptr,

        /*.encoder_begin_callback           =*/ nullptr,
        /*.encoder_begin_callback_user_data =*/ nullptr,

        /*.logits_filter_callback           =*/ nullptr,
        /*.logits_filter_callback_user_data =*/ nullptr,
    };

    switch (strategy) {
        case WHISPER_SAMPLING_GREEDY:
            {
                result.greedy = {
                    /*.best_of   =*/ 2, // TODO: increase to 5 when we speed-up batch decoding
                };
            } break;
        case WHISPER_SAMPLING_BEAM_SEARCH:
            {
                result.beam_search = {
                    /*.beam_size =*/ 2, // TODO: increase to 5 when we speed-up batch decoding

                    /*.patience  =*/ -1.0f,
                };
            } break;
    }

    return result;
}

// forward declarations
static std::vector<float> get_signal_energy(const float * signal, int n_samples, int n_samples_per_half_window);
static void whisper_exp_compute_token_level_timestamps(
        struct whisper_context & ctx,
          struct whisper_state & state,
                           int   i_segment,
                         float   thold_pt,
                         float   thold_ptsum);

static inline bool should_split_on_word(const char * txt, bool split_on_word) {
    if (!split_on_word) return true;

    return txt[0] == ' ';
}

// wrap the last segment to max_len characters
// returns the number of new segments
static int whisper_wrap_segment(struct whisper_context & ctx, struct whisper_state & state, int max_len, bool split_on_word) {
    auto segment = state.result_all.back();

    int res = 1;
    int acc = 0;

    std::string text;

    for (int i = 0; i < (int) segment.tokens.size(); i++) {
        const auto & token = segment.tokens[i];
        if (token.id >= whisper_token_eot(&ctx)) {
            continue;
        }

        const auto txt = whisper_token_to_str(&ctx, token.id);
        const int cur = strlen(txt);

        if (acc + cur > max_len && i > 0 && should_split_on_word(txt, split_on_word)) {
            state.result_all.back().text = std::move(text);
            state.result_all.back().t1 = token.t0;
            state.result_all.back().tokens.resize(i);
            state.result_all.back().speaker_turn_next = false;

            state.result_all.push_back({});
            state.result_all.back().t0 = token.t0;
            state.result_all.back().t1 = segment.t1;

            // add tokens [i, end] to the new segment
            state.result_all.back().tokens.insert(
                state.result_all.back().tokens.end(),
                    segment.tokens.begin() + i,
                    segment.tokens.end());

            state.result_all.back().speaker_turn_next = segment.speaker_turn_next;

            acc = 0;
            text = "";

            segment = state.result_all.back();
            i = -1;

            res++;
        } else {
            acc += cur;
            text += txt;
        }
    }

    state.result_all.back().text = std::move(text);

    return res;
}

static const std::vector<std::string> non_speech_tokens = {
    "\"", "#", "(", ")", "*", "+", "/", ":", ";", "<", "=", ">", "@", "[", "\\", "]", "^",
    "_", "`", "{", "|", "}", "~", "「", "」", "『", "』", "<<", ">>", "<<<", ">>>", "--",
    "---", "-(", "-[", "('", "(\"", "((", "))", "(((", ")))", "[[", "]]", "{{", "}}", "♪♪",
    "♪♪♪","♩", "♪", "♫", "♬", "♭", "♮", "♯"
};

// process the logits for the selected decoder
// - applies logit filters
// - computes logprobs and probs
static void whisper_process_logits(
              struct whisper_context & ctx,
               struct whisper_state  & state,
    const struct whisper_full_params   params,
              struct whisper_decoder & decoder,
                               float   temperature) {
    const auto & vocab      = ctx.vocab;
    const auto & tokens_cur = decoder.sequence.tokens;

    const bool is_initial = tokens_cur.size() == 0;
    const int  n_logits   = vocab.id_to_token.size();

    WHISPER_ASSERT(n_logits == ctx.vocab.n_vocab);

    // extract the logits for the last token
    // we will be mutating, and therefore we don't want to use the ctx.logits buffer directly
    auto & probs    = decoder.probs;
    auto & logits   = decoder.logits;
    auto & logprobs = decoder.logprobs;
    {
        logits.resize(n_logits);
        memcpy(logits.data(), state.logits.data() + (state.logits.size() - n_logits), n_logits*sizeof(float));

        if (temperature > 0.0f) {
            for (int i = 0; i < n_logits; i++) {
                logits[i] /= temperature;
            }
        }

        // will be populated a bit later
        probs.resize(n_logits);
        logprobs.resize(n_logits);
    }

    // apply logit filters here
    // ref: https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L480-L493
    {
        // suppress blank
        // https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L388-L390
        if (params.suppress_blank) {
            if (is_initial) {
                logits[vocab.token_eot]           = -INFINITY;
                logits[vocab.token_to_id.at(" ")] = -INFINITY;
            }
        }

        // suppress <|notimestamps|> token
        // ref: https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L410-L412
        logits[vocab.token_not] = -INFINITY;

        // suppress sot and nosp tokens
        logits[vocab.token_sot]  = -INFINITY;
        logits[vocab.token_nosp] = -INFINITY; // TODO: ignore this token for now

        // [TDRZ] when tinydiarize is disabled, suppress solm token
        if (params.tdrz_enable == false) {
            logits[vocab.token_solm] = -INFINITY;
        }

        // suppress task tokens
        logits[vocab.token_translate]  = -INFINITY;
        logits[vocab.token_transcribe] = -INFINITY;

        if (params.logits_filter_callback) {
            params.logits_filter_callback(&ctx, &state, tokens_cur.data(), tokens_cur.size(), logits.data(), params.logits_filter_callback_user_data);
        }

        // suppress non-speech tokens
        // ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
        if (params.suppress_non_speech_tokens) {
            for (const std::string & token : non_speech_tokens) {
                const std::string suppress_tokens[] = {token, " " + token};
                for (const std::string & suppress_token : suppress_tokens) {
                    if (vocab.token_to_id.find(suppress_token) != vocab.token_to_id.end()) {
                        logits[vocab.token_to_id.at(suppress_token)] = -INFINITY;
                    }
                }
            }

            // allow hyphens "-" and single quotes "'" between words, but not at the beginning of a word
            if (vocab.token_to_id.find(" -") != vocab.token_to_id.end()) {
                logits[vocab.token_to_id.at(" -")] = -INFINITY;
            }
            if (vocab.token_to_id.find(" '") != vocab.token_to_id.end()) {
                logits[vocab.token_to_id.at(" '")] = -INFINITY;
            }
        }

        // timestamps have to appear in pairs, except directly before EOT; mask logits accordingly
        // https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L414-L424
        {
            const bool last_was_timestamp        = tokens_cur.size() > 0 && tokens_cur.back().id >= vocab.token_beg;
            const bool penultimate_was_timestamp = tokens_cur.size() < 2 || tokens_cur[tokens_cur.size() - 2].id >= vocab.token_beg;

            //log("last_was_timestamp=%d penultimate_was_timestamp=%d\n", last_was_timestamp, penultimate_was_timestamp);

            if (last_was_timestamp) {
                if (penultimate_was_timestamp) {
                    for (int i = vocab.token_beg; i < n_logits; ++i) {
                        logits[i] = -INFINITY;
                    }
                } else {
                    for (int i = 0; i < vocab.token_eot; ++i) {
                        logits[i] = -INFINITY;
                    }
                }
            }
        }

        // the initial timestamp cannot be larger than max_initial_ts
        // ref: https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L426-L429
        if (is_initial && params.max_initial_ts > 0.0f) {
            const float precision = float(WHISPER_CHUNK_SIZE)/ctx.model.hparams.n_audio_ctx;
            const int   tid0      = std::round(params.max_initial_ts/precision);

            for (int i = vocab.token_beg + tid0 + 1; i < n_logits; ++i) {
                logits[i] = -INFINITY;
            }
        }

        // condition timestamp tokens to be increasing
        // ref: https://github.com/openai/whisper/pull/831#issuecomment-1385910556
        if (decoder.has_ts) {
            const int tid0 = decoder.seek_delta/2;

            for (int i = vocab.token_beg; i < vocab.token_beg + tid0; ++i) {
                logits[i] = -INFINITY;
            }
        }

        // populate the logprobs array (log_softmax)
        {
            const float logit_max = *std::max_element(logits.begin(), logits.end());
            float logsumexp = 0.0f;
            for (int i = 0; i < n_logits; ++i) {
                if (logits[i] > -INFINITY) {
                    logsumexp += expf(logits[i] - logit_max);
                }
            }
            logsumexp = logf(logsumexp) + logit_max;

            for (int i = 0; i < n_logits; ++i) {
                if (logits[i] > -INFINITY) {
                    logprobs[i] = logits[i] - logsumexp;
                } else {
                    logprobs[i] = -INFINITY;
                }
            }
        }

        // if sum of probability over timestamps is above any other token, sample timestamp
        // ref: https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L431-L437
        {
            // logsumexp over timestamps
            float timestamp_logprob = -INFINITY;
            {
                float logsumexp = 0.0f;
                const float logprob_max = *std::max_element(logprobs.begin() + vocab.token_beg, logprobs.end());
                for (int i = vocab.token_beg; i < n_logits; ++i) {
                    if (logprobs[i] > -INFINITY) {
                        logsumexp += expf(logprobs[i] - logprob_max);
                    }
                }
                if (logsumexp > 0.0f) {
                    timestamp_logprob = logf(logsumexp) + logprob_max;
                }
            }

            const float max_text_token_logprob = *std::max_element(logprobs.begin(), logprobs.begin() + vocab.token_beg);

            //log("timestamp_logprob=%f max_text_token_logprob=%f\n", timestamp_logprob, max_text_token_logprob);

            if (timestamp_logprob > max_text_token_logprob) {
                for (int i = 0; i < vocab.token_beg; ++i) {
                    logits[i]   = -INFINITY;
                    logprobs[i] = -INFINITY;
                }
            }
        }
    }

    // compute probs
    {
        for (int i = 0; i < n_logits; ++i) {
            if (logits[i] == -INFINITY) {
                probs[i] = 0.0f;
            } else {
                probs[i] = expf(logprobs[i]);
            }
        }
    }

#if 0
    // print first 100 logits - token string : logit
    for (int i = 0; i < 100; i++) {
        const auto token   = vocab.id_to_token.at(i);
        const auto prob    = probs[i];
        const auto logit   = logits[i];
        const auto logprob = logprobs[i];
        printf("%s : prob=%9.5f logit=%9.5f logprob=%9.5f\n", token.c_str(), prob, logit, logprob);
    }

    // "And", "and", " And", " and"
    printf("logits[\"and\"]  = %f\n", logits[vocab.token_to_id.at("and")]);
    printf("logits[\"And\"]  = %f\n", logits[vocab.token_to_id.at("And")]);
    printf("logits[\" and\"] = %f\n", logits[vocab.token_to_id.at(" and")]);
    printf("logits[\" And\"] = %f\n", logits[vocab.token_to_id.at(" And")]);
    printf("logits[\" so\"]  = %f\n", logits[vocab.token_to_id.at(" so")]);

    printf("logprobs[\"and\"]  = %f\n", logprobs[vocab.token_to_id.at("and")]);
    printf("logprobs[\"And\"]  = %f\n", logprobs[vocab.token_to_id.at("And")]);
    printf("logprobs[\" and\"] = %f\n", logprobs[vocab.token_to_id.at(" and")]);
    printf("logprobs[\" And\"] = %f\n", logprobs[vocab.token_to_id.at(" And")]);
    printf("logprobs[\" so\"]  = %f\n", logprobs[vocab.token_to_id.at(" so")]);

    printf("probs[\"and\"]  = %f\n", probs[vocab.token_to_id.at("and")]);
    printf("probs[\"And\"]  = %f\n", probs[vocab.token_to_id.at("And")]);
    printf("probs[\" and\"] = %f\n", probs[vocab.token_to_id.at(" and")]);
    printf("probs[\" And\"] = %f\n", probs[vocab.token_to_id.at(" And")]);
    printf("probs[\" so\"]  = %f\n", probs[vocab.token_to_id.at(" so")]);
#endif
}

static whisper_token_data whisper_sample_token(
            whisper_context & ctx,
              whisper_state & state,
      const whisper_decoder & decoder,
                       bool   best) {
    whisper_token_data result = {
        0, 0, 0.0f, 0.0f, 0.0f, 0.0f, -1, -1, 0.0f,
    };

    const auto & vocab = ctx.vocab;

    const auto & probs    = decoder.probs;
    const auto & logprobs = decoder.logprobs;

    const int n_logits = vocab.n_vocab;

    {
        double sum_ts = 0.0;
        double max_ts = 0.0;

        for (int i = vocab.token_beg; i < n_logits; i++) {
            if (probs[i] == -INFINITY) {
                continue;
            }

            sum_ts += probs[i];
            if (max_ts < probs[i]) {
                max_ts = probs[i];
                result.tid = i;
            }
        }

        result.pt    = max_ts/(sum_ts + 1e-10);
        result.ptsum = sum_ts;
    }

    if (best) {
        for (int i = 0; i < n_logits; ++i) {
            if (result.p < probs[i]) {
                result.id   = i;
                result.p    = probs[i];
                result.plog = logprobs[i];
            }
        }
    } else {
        std::discrete_distribution<> dist(probs.begin(), probs.end());

        result.id   = dist(state.rng);
        result.p    = probs[result.id];
        result.plog = logprobs[result.id];
    }

    if (result.id >= vocab.token_beg) {
        result.tid = result.id;
        result.pt  = result.p;
    }

    state.n_sample++;

    return result;
}

static std::vector<whisper_token_data> whisper_sample_token_topk(
            whisper_context & ctx,
              whisper_state & state,
      const whisper_decoder & decoder,
                        int   k) {
    const auto & vocab = ctx.vocab;

    const auto & probs    = decoder.probs;
    const auto & logits   = decoder.logits;
    const auto & logprobs = decoder.logprobs;

    const int n_logits = vocab.n_vocab;

    auto & logits_id = state.logits_id;

    logits_id.resize(n_logits);
    for (int i = 0; i < n_logits; ++i) {
        logits_id[i].first = logits[i];
        logits_id[i].second = i;
    }

    {
        using pair_type = std::remove_reference<decltype(logits_id)>::type::value_type;
        std::partial_sort(
                logits_id.begin(),
                logits_id.begin() + k, logits_id.end(),
                [](const pair_type & a, const pair_type & b) {
            return a.first > b.first;
        });
    }

    std::vector<whisper_token_data> result;
    result.reserve(k);

    whisper_token tid = vocab.token_beg;

    float pt    = 0.0;
    float ptsum = 0.0;

    {
        double sum_ts = 0.0;
        double max_ts = 0.0;

        for (int i = vocab.token_beg; i < n_logits; i++) {
            if (probs[i] == -INFINITY) {
                continue;
            }

            sum_ts += probs[i];
            if (max_ts < probs[i]) {
                max_ts = probs[i];
                tid = i;
            }
        }

        pt    = max_ts/(sum_ts + 1e-10);
        ptsum = sum_ts;
    }

    for (int i = 0; i < k; ++i) {
        const auto id = logits_id[i].second;

        result.push_back({ id, tid, probs[id], logprobs[id], pt, ptsum, -1, -1, 0.0f, });

        if (result[i].id >= vocab.token_beg) {
            result[i].tid = result[i].id;
            result[i].pt  = result[i].p;
        }
    }

    state.n_sample++;

    return result;
}

// ref: https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L178-L192
static void whisper_sequence_score(
        const struct whisper_full_params & params,
                        whisper_sequence & sequence) {
    if (sequence.result_len == 0) {
        return;
    }

    double result = 0.0f;

    for (int i = 0; i < sequence.result_len; ++i) {
        result += sequence.tokens[i].plog;
    }

    sequence.sum_logprobs = result;
    sequence.avg_logprobs = result/sequence.result_len;

    double penalty = sequence.result_len;

    if (params.length_penalty > 0.0f) {
        penalty = pow((5.0 + penalty)/6.0, params.length_penalty);
    }

    sequence.score = result/penalty;

    // compute the entropy of the sequence of the last 32 tokens
    {
        const int n = 32;

        int cnt = 0;
        double entropy = 0.0f;

        std::map<whisper_token, int> token_counts;
        for (int i = std::max(0, sequence.result_len - n); i < sequence.result_len; ++i) {
            token_counts[sequence.tokens[i].id]++;
            cnt++;
        }

        for (const auto & kv : token_counts) {
            const auto p = kv.second/(double)cnt;
            entropy -= p*log(p);

            //WHISPER_PRINT_DEBUG("entropy: %d %f %f, count %d\n", kv.first, p, log(p), kv.second);
        }

        sequence.entropy = entropy;
    }
}

static bool whisper_kv_swap_fast(
                   std::vector<int> & view,
                    whisper_decoder   src[],
                std::vector<kv_buf> & kv_swap_bufs,
                          const int & n_decoders) {
    WHISPER_PRINT_DEBUG("%s: n_decoders %d\n", __func__, n_decoders);

    // (decoder->buffer->decoder or decoder->buffer + decoder->decoder)
    std::set<int> two_copy; // decoder indices require two copies to safely modify KV caches

    // (buffer->decoder or decoder->decoder)
    std::set<int> one_copy; // decoder indices require one copy to safely modify KV caches

    // (decoder<->decoder)
    std::set<int> p_swap_set; // decoder indices able to swap KV-cache pointers
    std::vector<whisper_pair<int, int>> p_swap_vec;
    p_swap_vec.reserve(n_decoders);

    // see https://github.com/ggerganov/whisper.cpp/wiki
    for (int i = 0; i < n_decoders; i++) {
        // zero-copy (no modification)
        if (i == view[i] || view[i] < 0) {
            continue;
        }

        bool is_one_copy = true;
        // since we modify data sequentially, we only consider decoder indices after current index
        for (int j = i + 1; j < n_decoders; j++) {
            if (i == view[j]) {
                // detect symmetric diagram
                if (j == view[i]) {
                    p_swap_set.insert(i);
                    p_swap_set.insert(j);
                    p_swap_vec.emplace_back(i, j);
                } else {
                    two_copy.insert(i);
                    is_one_copy = false;
                }
                break;
            }
        }
        if (is_one_copy) {
            one_copy.insert(i);
        }
    }

    kv_swap_bufs.resize(n_decoders);

    for (int i = 0; i < n_decoders; i++) {
        kv_swap_bufs[i].k.resize(ggml_nbytes(src[i].kv_self.k));
        kv_swap_bufs[i].v.resize(ggml_nbytes(src[i].kv_self.v));
    }

    for (auto & i : two_copy) {
        // make a copy of KV caches
        WHISPER_PRINT_DEBUG("%s: store KV cache into swap: idx %d\n", __func__, i);
        memcpy(kv_swap_bufs[i].k.data(), src[i].kv_self.k->data, kv_swap_bufs[i].k.size());
        memcpy(kv_swap_bufs[i].v.data(), src[i].kv_self.v->data, kv_swap_bufs[i].v.size());
    }

    // since two-copy decoder KV caches are protected by kv_swap_bufs, modify them first
    for (auto & i : two_copy) {
        // skip the decoder indices that require pointer swapping
        if (p_swap_set.find(i) != p_swap_set.end()) {
            continue;
        }

        if (two_copy.find(view[i]) != two_copy.end()) {
            // modify KV caches of decoder using data from kv_swap_bufs
            WHISPER_PRINT_DEBUG("%s: two-copy decoder using   swap buffers: swap[%d] -> %d\n", __func__, view[i], i);
            memcpy(src[i].kv_self.k->data, kv_swap_bufs[view[i]].k.data(), kv_swap_bufs[view[i]].k.size());
            memcpy(src[i].kv_self.v->data, kv_swap_bufs[view[i]].v.data(), kv_swap_bufs[view[i]].v.size());
        } else {
            // modify KV caches of decoder using data from correspond decoder KV caches directly
            WHISPER_PRINT_DEBUG("%s: two-copy decoder without swap buffers:      %d  -> %d\n", __func__, view[i], i);
            memcpy(src[i].kv_self.k->data, src[view[i]].kv_self.k->data, ggml_nbytes(src[view[i]].kv_self.k));
            memcpy(src[i].kv_self.v->data, src[view[i]].kv_self.v->data, ggml_nbytes(src[view[i]].kv_self.v));
        }
    }

    // then modify one-copy decoder KV caches
    for (auto & i : one_copy) {
        // skip the decoder indices that require pointer swapping
        if (p_swap_set.find(i) != p_swap_set.end()) {
            continue;
        }

        if (two_copy.find(view[i]) != two_copy.end()) {
            // modify KV caches of decoder using data from kv_swap_bufs
            WHISPER_PRINT_DEBUG("%s: one-copy decoder using   swap buffers: swap[%d] -> %d\n", __func__, view[i], i);
            memcpy(src[i].kv_self.k->data, kv_swap_bufs[view[i]].k.data(), kv_swap_bufs[view[i]].k.size());
            memcpy(src[i].kv_self.v->data, kv_swap_bufs[view[i]].v.data(), kv_swap_bufs[view[i]].v.size());
        } else {
            // modify KV caches of decoder using data from correspond decoder KV caches directly
            WHISPER_PRINT_DEBUG("%s: one-copy decoder without swap buffers:      %d  -> %d\n", __func__, view[i], i);
            memcpy(src[i].kv_self.k->data, src[view[i]].kv_self.k->data, ggml_nbytes(src[view[i]].kv_self.k));
            memcpy(src[i].kv_self.v->data, src[view[i]].kv_self.v->data, ggml_nbytes(src[view[i]].kv_self.v));
        }
    }

    // swap the pointers
    for (auto & i : p_swap_vec) {
        WHISPER_PRINT_DEBUG("%s: swap pointers: %d <-> %d\n", __func__, i.first, i.second);
        std::swap(src[i.first].kv_self, src[i.second].kv_self);
    }

    return true;
}

int whisper_full_with_state(
        struct whisper_context * ctx,
          struct whisper_state * state,
    struct whisper_full_params   params,
                   const float * samples,
                           int   n_samples) {
    // clear old results
    auto & result_all = state->result_all;

    result_all.clear();

    if (n_samples > 0) {
        // compute log mel spectrogram
        if (params.speed_up) {
            // TODO: Replace PV with more advanced algorithm
            log("%s: failed to compute log mel spectrogram\n", __func__);
            return -1;
        } else {
            if (whisper_pcm_to_mel_with_state(ctx, state, samples, n_samples, params.n_threads) != 0) {
                log("%s: failed to compute log mel spectrogram\n", __func__);
                return -2;
            }
        }
    }

    // auto-detect language if not specified
    if (params.language == nullptr || strlen(params.language) == 0 || strcmp(params.language, "auto") == 0 || params.detect_language) {
        std::vector<float> probs(whisper_lang_max_id() + 1, 0.0f);

        const auto lang_id = whisper_lang_auto_detect_with_state(ctx, state, 0, params.n_threads, probs.data());
        if (lang_id < 0) {
            log("%s: failed to auto-detect language\n", __func__);
            return -3;
        }
        state->lang_id = lang_id;
        params.language = whisper_lang_str(lang_id);

        log("%s: auto-detected language: %s (p = %f)\n", __func__, params.language, probs[whisper_lang_id(params.language)]);
        if (params.detect_language) {
            return 0;
        }
    }

    if (params.token_timestamps) {
        state->t_beg    = 0;
        state->t_last   = 0;
        state->tid_last = 0;
        if (n_samples > 0) {
            state->energy = get_signal_energy(samples, n_samples, 32);
        }
    }

    const int seek_start = params.offset_ms/10;
    const int seek_end = params.duration_ms == 0 ? whisper_n_len_from_state(state) : seek_start + params.duration_ms/10;

    // if length of spectrogram is less than 1.0s (100 frames), then return
    // basically don't process anything that is less than 1.0s
    // see issue #39: https://github.com/ggerganov/whisper.cpp/issues/39
    if (seek_end < seek_start + (params.speed_up ? 50 : 100)) {
        return 0;
    }

    // a set of temperatures to use
    // [ t0, t0 + delta, t0 + 2*delta, ..., < 1.0f + 1e-6f ]
    std::vector<float> temperatures;
    if (params.temperature_inc > 0.0f) {
        for (float t = params.temperature; t < 1.0f + 1e-6f; t += params.temperature_inc) {
            temperatures.push_back(t);
        }
    } else {
        temperatures.push_back(params.temperature);
    }

    // initialize the decoders
    int n_decoders = 1;

    switch (params.strategy) {
        case WHISPER_SAMPLING_GREEDY:
            {
                n_decoders = params.greedy.best_of;
            } break;
        case WHISPER_SAMPLING_BEAM_SEARCH:
            {
                n_decoders = std::max(params.greedy.best_of, params.beam_search.beam_size);
            } break;
    };

    n_decoders = std::max(1, n_decoders);

    // TAGS: WHISPER_DECODER_INIT
    for (int j = 1; j < n_decoders; j++) {
        auto & decoder = state->decoders[j];

        if (decoder.kv_self.ctx == nullptr) {
            decoder.kv_self = state->decoders[0].kv_self;
            if (!kv_cache_reinit(decoder.kv_self)) {
                log("%s: kv_cache_reinit() failed for self-attention, decoder %d\n", __func__, j);
                return -4;
            }

            WHISPER_PRINT_DEBUG("%s: initialized self-attention kv cache, decoder %d\n", __func__, j);

            decoder.sequence.tokens.reserve(state->decoders[0].sequence.tokens.capacity());

            decoder.probs.resize   (ctx->vocab.n_vocab);
            decoder.logits.resize  (ctx->vocab.n_vocab);
            decoder.logprobs.resize(ctx->vocab.n_vocab);

            // TODO: not very clean - look for a better way and potentially merging with the init of decoder 0
#ifdef GGML_USE_METAL
#define WHISPER_METAL_CHECK_BUF(result)              \
            if (!(result)) {                                 \
                log("%s: failed to add metal buffer\n", __func__); \
                return 0;                              \
            }

            const std::string kv_name = "kv_self_" + std::to_string(j);
            auto & kv_self = decoder.kv_self;

            WHISPER_METAL_CHECK_BUF(ggml_metal_add_buffer(state->ctx_metal, kv_name.c_str(), kv_self.buf.data(), kv_self.buf.size(), 0));
#undef WHISPER_METAL_CHECK_BUF
#endif
        }
    }

    // the accumulated text context so far
    auto & prompt_past = state->prompt_past;
    if (params.no_context) {
        prompt_past.clear();
    }

    // prepare prompt
    {
        std::vector<whisper_token> prompt_tokens;

        // initial prompt
        if (!params.prompt_tokens && params.initial_prompt) {
            prompt_tokens.resize(2048);
            prompt_tokens.resize(whisper_tokenize(ctx, params.initial_prompt, prompt_tokens.data(), prompt_tokens.size()));
            params.prompt_tokens   = prompt_tokens.data();
            params.prompt_n_tokens = prompt_tokens.size();
        }

        // prepend the prompt tokens to the prompt_past
        if (params.prompt_tokens && params.prompt_n_tokens > 0) {
            // parse tokens from the pointer
            for (int i = 0; i < params.prompt_n_tokens; i++) {
                prompt_past.push_back(params.prompt_tokens[i]);
            }
            std::rotate(prompt_past.begin(), prompt_past.end() - params.prompt_n_tokens, prompt_past.end());
        }
    }

    // overwrite audio_ctx, max allowed is hparams.n_audio_ctx
    if (params.audio_ctx > whisper_n_audio_ctx(ctx)) {
        log("%s: audio_ctx is larger than the maximum allowed (%d > %d)\n", __func__, params.audio_ctx, whisper_n_audio_ctx(ctx));
        return -5;
    }
    state->exp_n_audio_ctx = params.audio_ctx;

    // these tokens determine the task that will be performed
    std::vector<whisper_token> prompt_init = { whisper_token_sot(ctx) };
    if (whisper_is_multilingual(ctx)) {
        const int lang_id = whisper_lang_id(params.language);
        state->lang_id = lang_id;
        prompt_init.push_back(whisper_token_lang(ctx, lang_id));
        if (params.translate) {
            prompt_init.push_back(whisper_token_translate(ctx));
        } else {
            prompt_init.push_back(whisper_token_transcribe(ctx));
        }
    }

    int seek = seek_start;

    std::vector<whisper_token> prompt;
    prompt.reserve(whisper_n_text_ctx(ctx));

    struct beam_candidate {
        int decoder_idx;
        int seek_delta;

        bool has_ts;

        whisper_sequence sequence;
    };

    std::vector<beam_candidate> beam_candidates;

    // main loop
    while (true) {
        if (params.progress_callback) {
            const int progress_cur = (100*(seek - seek_start))/(seek_end - seek_start);

            params.progress_callback(
                ctx, ctx->state, progress_cur, params.progress_callback_user_data);
        }

        // of only 1 second left, then stop
        if (seek + 100 >= seek_end) {
            break;
        }

        if (params.encoder_begin_callback) {
            if (params.encoder_begin_callback(ctx, state, params.encoder_begin_callback_user_data) == false) {
                log("%s: encoder_begin_callback returned false - aborting\n", __func__);
                break;
            }
        }

        // encode audio features starting at offset seek
        if (!whisper_encode_internal(*ctx, *state, seek, params.n_threads)) {
            log("%s: failed to encode\n", __func__);
            return -6;
        }

        // if there is a very short audio segment left to process, we remove any past prompt since it tends
        // to confuse the decoder and often make it repeat or hallucinate stuff
        if (seek > seek_start && seek + 500 >= seek_end) {
            prompt_past.clear();
        }

        int best_decoder_id = 0;

        for (int it = 0; it < (int) temperatures.size(); ++it) {
            const float t_cur = temperatures[it];

            int n_decoders_cur = 1;

            switch (params.strategy) {
                case whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY:
                    {
                        if (t_cur > 0.0f) {
                            n_decoders_cur = params.greedy.best_of;
                        }
                    } break;
                case whisper_sampling_strategy::WHISPER_SAMPLING_BEAM_SEARCH:
                    {
                        if (t_cur > 0.0f) {
                            n_decoders_cur = params.greedy.best_of;
                        } else {
                            n_decoders_cur = params.beam_search.beam_size;
                        }
                    } break;
            };

            n_decoders_cur = std::max(1, n_decoders_cur);

            WHISPER_PRINT_DEBUG("\n%s: decoding with %d decoders, temperature = %.2f\n", __func__, n_decoders_cur, t_cur);

            // TAGS: WHISPER_DECODER_INIT
            for (int j = 0; j < n_decoders_cur; ++j) {
                auto & decoder = state->decoders[j];

                decoder.kv_self.n = 0;

                decoder.sequence.tokens.clear();
                decoder.sequence.result_len       = 0;
                decoder.sequence.sum_logprobs_all = 0.0;
                decoder.sequence.sum_logprobs     = -INFINITY;
                decoder.sequence.avg_logprobs     = -INFINITY;
                decoder.sequence.entropy          = 0.0;
                decoder.sequence.score            = -INFINITY;

                decoder.seek_delta = 100*WHISPER_CHUNK_SIZE;

                decoder.failed    = false;
                decoder.completed = false;
                decoder.has_ts    = false;
            }

            // init prompt and kv cache for the current iteration
            // run whisper_decoder() only for decoder 0 and copy the results for the other decoders
            {
                prompt.clear();

                // if we have already generated some text, use it as a prompt to condition the next generation
                if (!prompt_past.empty() && t_cur < 0.5f && params.n_max_text_ctx > 0) {
                    int n_take = std::min(std::min(params.n_max_text_ctx, whisper_n_text_ctx(ctx)/2), int(prompt_past.size()));

                    prompt = { whisper_token_prev(ctx) };
                    prompt.insert(prompt.begin() + 1, prompt_past.end() - n_take, prompt_past.end());
                }

                // init new transcription with sot, language (opt) and task tokens
                prompt.insert(prompt.end(), prompt_init.begin(), prompt_init.end());

                // print the prompt
                WHISPER_PRINT_DEBUG("\n\n");
                for (int i = 0; i < (int) prompt.size(); i++) {
                    WHISPER_PRINT_DEBUG("%s: prompt[%d] = %s\n", __func__, i, ctx->vocab.id_to_token.at(prompt[i]).c_str());
                }
                WHISPER_PRINT_DEBUG("\n\n");

                if (!whisper_decode_internal(*ctx, *state, state->decoders[0], prompt.data(), prompt.size(), 0, params.n_threads)) {
                    log("%s: failed to decode\n", __func__);
                    return -7;
                }

                {
                    const int64_t t_start_sample_us = ggml_time_us();

                    whisper_process_logits(*ctx, *state, params, state->decoders[0], t_cur);

                    state->decoders[0].kv_self.n += prompt.size();

                    for (int j = 1; j < n_decoders_cur; ++j) {
                        auto & decoder = state->decoders[j];

                        memcpy(decoder.kv_self.k->data, state->decoders[0].kv_self.k->data, ggml_nbytes(decoder.kv_self.k));
                        memcpy(decoder.kv_self.v->data, state->decoders[0].kv_self.v->data, ggml_nbytes(decoder.kv_self.v));

                        decoder.kv_self.n += prompt.size();

                        memcpy(decoder.probs.data(),    state->decoders[0].probs.data(),    decoder.probs.size()*sizeof(decoder.probs[0]));
                        memcpy(decoder.logits.data(),   state->decoders[0].logits.data(),   decoder.logits.size()*sizeof(decoder.logits[0]));
                        memcpy(decoder.logprobs.data(), state->decoders[0].logprobs.data(), decoder.logprobs.size()*sizeof(decoder.logprobs[0]));
                    }

                    state->t_sample_us += ggml_time_us() - t_start_sample_us;
                }
            }

            for (int i = 0, n_max = whisper_n_text_ctx(ctx)/2 - 4; i < n_max; ++i) {
                const int64_t t_start_sample_us = ggml_time_us();

                if (params.strategy == whisper_sampling_strategy::WHISPER_SAMPLING_BEAM_SEARCH) {
                    beam_candidates.clear();
                }

                // generate new sequence candidates for each decoder
                for (int j = 0; j < n_decoders_cur; ++j) {
                    auto & decoder = state->decoders[j];

                    if (decoder.completed || decoder.failed) {
                        continue;
                    }

                    switch (params.strategy) {
                        case whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY:
                            {
                                if (t_cur < 1e-6f) {
                                    decoder.sequence.tokens.push_back(whisper_sample_token(*ctx, *state, decoder, true));
                                } else {
                                    decoder.sequence.tokens.push_back(whisper_sample_token(*ctx, *state, decoder, false));
                                }

                                decoder.sequence.sum_logprobs_all += decoder.sequence.tokens.back().plog;
                            } break;
                        case whisper_sampling_strategy::WHISPER_SAMPLING_BEAM_SEARCH:
                            {
                                const auto tokens_new = whisper_sample_token_topk(*ctx, *state, decoder, params.beam_search.beam_size);

                                for (const auto & token : tokens_new) {
                                    beam_candidates.push_back({ j, decoder.seek_delta, decoder.has_ts, decoder.sequence });
                                    beam_candidates.back().sequence.tokens.push_back(token);
                                    beam_candidates.back().sequence.sum_logprobs_all += token.plog;

                                    //WHISPER_PRINT_DEBUG("%s: beam candidate: %s (%f, %f)\n", __func__, ctx->vocab.id_to_token.at(token.id).c_str(), token.plog, beam_candidates.back().sequence.sum_logprobs_all);
                                }
                            } break;
                    };
                }

                // for beam-search, choose the top candidates and update the KV caches
                if (params.strategy == whisper_sampling_strategy::WHISPER_SAMPLING_BEAM_SEARCH) {
                    std::sort(
                            beam_candidates.begin(),
                            beam_candidates.end(),
                            [](const beam_candidate & a, const beam_candidate & b) {
                        return a.sequence.sum_logprobs_all > b.sequence.sum_logprobs_all;
                    });

                    uint32_t cur_c = 0;
                    std::vector<int> decoder_idx(n_decoders_cur, -1);

                    for (int j = 0; j < n_decoders_cur; ++j) {
                        auto & decoder = state->decoders[j];

                        if (decoder.completed || decoder.failed) {
                            continue;
                        }

                        auto & cur = beam_candidates[cur_c++];

                        while (beam_candidates.size() > cur_c && beam_candidates[cur_c].sequence.sum_logprobs_all == cur.sequence.sum_logprobs_all && i > 0) {
                            ++cur_c;
                        }

                        decoder.sequence   = cur.sequence;
                        decoder.seek_delta = cur.seek_delta;
                        decoder.has_ts     = cur.has_ts;

                        decoder_idx[j] = cur.decoder_idx;
                        WHISPER_PRINT_DEBUG("%s: beam search: decoder %d: from decoder %d: token = %10s, plog = %8.5f, sum_logprobs = %8.5f\n",
                                __func__, j, cur.decoder_idx, ctx->vocab.id_to_token.at(decoder.sequence.tokens.back().id).c_str(), decoder.sequence.tokens.back().plog, decoder.sequence.sum_logprobs_all);
                    }

                    // update KV caches
                    whisper_kv_swap_fast(decoder_idx, state->decoders, state->kv_swap_bufs, n_decoders_cur);
                }

                // update the decoder state
                // - check if the sequence is completed
                // - check if the sequence is failed
                // - update sliding window based on timestamp tokens
                for (int j = 0; j < n_decoders_cur; ++j) {
                    auto & decoder = state->decoders[j];

                    if (decoder.completed || decoder.failed) {
                        continue;
                    }

                    auto & has_ts     = decoder.has_ts;
                    auto & failed     = decoder.failed;
                    auto & completed  = decoder.completed;
                    auto & seek_delta = decoder.seek_delta;
                    auto & result_len = decoder.sequence.result_len;

                    {
                        const auto & token = decoder.sequence.tokens.back();

                        // timestamp token - update sliding window
                        if (token.id > whisper_token_beg(ctx)) {
                            const int seek_delta_new = 2*(token.id - whisper_token_beg(ctx));

                            // do not allow to go back in time
                            if (has_ts && seek_delta > seek_delta_new && result_len < i) {
                                failed = true; // TODO: maybe this is not a failure ?
                                continue;
                            }

                            seek_delta = seek_delta_new;
                            result_len = i + 1;
                            has_ts = true;
                        }

#ifdef WHISPER_DEBUG
                        {
                            const auto tt = token.pt > 0.10 ? ctx->vocab.id_to_token.at(token.tid) : "[?]";
                            WHISPER_PRINT_DEBUG("%s: id = %3d, decoder = %d, token = %6d, p = %6.3f, ts = %10s, %6.3f, result_len = %4d '%s'\n",
                                    __func__, i, j, token.id, token.p, tt.c_str(), token.pt, result_len, ctx->vocab.id_to_token.at(token.id).c_str());
                        }
#endif

                        // end of segment
                        if (token.id == whisper_token_eot(ctx) ||               // end of text token
                           (params.max_tokens > 0 && i >= params.max_tokens) || // max tokens per segment reached
                           (has_ts && seek + seek_delta + 100 >= seek_end)      // end of audio reached
                           ) {
                            if (result_len == 0) {
                                if (seek + seek_delta + 100 >= seek_end) {
                                    result_len = i + 1;
                                } else {
                                    failed = true;
                                    continue;
                                }
                            }

                            if (params.single_segment) {
                                result_len = i + 1;
                                seek_delta = 100*WHISPER_CHUNK_SIZE;
                            }

                            completed = true;
                            continue;
                        }

                        // TESTS: if no tensors are loaded, it means we are running tests
                        if (ctx->model.n_loaded == 0) {
                            seek_delta = 100*WHISPER_CHUNK_SIZE;
                            completed = true;
                            continue;
                        }
                    }

                    // sometimes, the decoding can get stuck in a repetition loop
                    // this is an attempt to mitigate such cases - we flag the decoding as failed and use a fallback strategy
                    if (i == n_max - 1 && (result_len == 0 || seek_delta < 100*WHISPER_CHUNK_SIZE/2)) {
                        failed = true;
                        continue;
                    }
                }

                // check if all decoders have finished (i.e. completed or failed)
                {
                    bool completed_all = true;

                    for (int j = 0; j < n_decoders_cur; ++j) {
                        auto & decoder = state->decoders[j];

                        if (decoder.completed || decoder.failed) {
                            continue;
                        }

                        completed_all = false;
                    }

                    if (completed_all) {
                        break;
                    }
                }

                state->t_sample_us += ggml_time_us() - t_start_sample_us;

                // obtain logits for the next token
                for (int j = 0; j < n_decoders_cur; ++j) {
                    auto & decoder = state->decoders[j];

                    if (decoder.failed || decoder.completed) {
                        continue;
                    }

                    decoder.tokens_tmp.resize(1);
                    decoder.tokens_tmp[0] = decoder.sequence.tokens.back().id;

                    //WHISPER_PRINT_DEBUG("%s: decoder %d: token %d, kv_self.n %d, seek_delta %d\n", __func__, j, decoder.tokens_tmp[0], decoder.kv_self.n, decoder.seek_delta);

                    if (!whisper_decode_internal(*ctx, *state, decoder, decoder.tokens_tmp.data(), decoder.tokens_tmp.size(), decoder.kv_self.n, params.n_threads)) {
                        log("%s: failed to decode\n", __func__);
                        return -8;
                    }

                    {
                        const int64_t t_start_sample_us = ggml_time_us();

                        whisper_process_logits(*ctx, *state, params, decoder, t_cur);

                        ++decoder.kv_self.n;

                        state->t_sample_us += ggml_time_us() - t_start_sample_us;
                    }
                }
            }

            // rank the resulting sequences and select the best one
            {
                double best_score = -INFINITY;

                for (int j = 0; j < n_decoders_cur; ++j) {
                    auto & decoder = state->decoders[j];

                    if (decoder.failed) {
                        continue;
                    }

                    decoder.sequence.tokens.resize(decoder.sequence.result_len);
                    whisper_sequence_score(params, decoder.sequence);

                    WHISPER_PRINT_DEBUG("%s: decoder %2d: score = %8.5f, result_len = %3d, avg_logprobs = %8.5f, entropy = %8.5f\n",
                            __func__, j, decoder.sequence.score, decoder.sequence.result_len, decoder.sequence.avg_logprobs, decoder.sequence.entropy);

                    if (decoder.sequence.result_len > 32 && decoder.sequence.entropy < params.entropy_thold) {
                        WHISPER_PRINT_DEBUG("%s: decoder %2d: failed due to entropy %8.5f < %8.5f\n",
                                __func__, j, decoder.sequence.entropy, params.entropy_thold);

                        decoder.failed = true;
                        state->n_fail_h++;

                        continue;
                    }

                    if (best_score < decoder.sequence.score) {
                        best_score = decoder.sequence.score;
                        best_decoder_id = j;
                    }
                }

                WHISPER_PRINT_DEBUG("%s: best decoder = %d\n", __func__, best_decoder_id);
            }

            // was the decoding successful for the current temperature?
            // do fallback only if:
            // - we are not at the last temperature
            // - we are not at the end of the audio (3 sec)
            if (it != (int) temperatures.size() - 1 &&
                seek_end - seek > 10*WHISPER_CHUNK_SIZE) {
                bool success = true;

                const auto & decoder = state->decoders[best_decoder_id];

                if (decoder.failed || decoder.sequence.avg_logprobs < params.logprob_thold) {
                    success = false;
                    state->n_fail_p++;
                }

                if (success) {
                    //for (auto & token : ctx->decoders[best_decoder_id].sequence.tokens) {
                    //    WHISPER_PRINT_DEBUG("%s: token = %d, p = %6.3f, pt = %6.3f, ts = %s, str = %s\n", __func__, token.id, token.p, token.pt, ctx->vocab.id_to_token.at(token.tid).c_str(), ctx->vocab.id_to_token.at(token.id).c_str());
                    //}

                    break;
                }
            }

            WHISPER_PRINT_DEBUG("\n%s: failed to decode with temperature = %.2f\n", __func__, t_cur);
        }

        // output results through a user-provided callback
        {
            const auto & best_decoder = state->decoders[best_decoder_id];

            const auto seek_delta = best_decoder.seek_delta;
            const auto result_len = best_decoder.sequence.result_len;

            const auto & tokens_cur = best_decoder.sequence.tokens;

            //WHISPER_PRINT_DEBUG("prompt_init.size() = %d, prompt.size() = %d, result_len = %d, seek_delta = %d\n", prompt_init.size(), prompt.size(), result_len, seek_delta);

            // update prompt_past
            prompt_past.clear();
            if (prompt.front() == whisper_token_prev(ctx)) {
                prompt_past.insert(prompt_past.end(), prompt.begin() + 1, prompt.end() - prompt_init.size());
            }

            for (int i = 0; i < result_len; ++i) {
                prompt_past.push_back(tokens_cur[i].id);
            }

            if (!tokens_cur.empty() && ctx->model.n_loaded > 0) {
                int  i0 = 0;
                auto t0 = seek + 2*(tokens_cur.front().tid - whisper_token_beg(ctx));

                std::string text;
                bool speaker_turn_next = false;

                for (int i = 0; i < (int) tokens_cur.size(); i++) {
                    //printf("%s: %18s %6.3f %18s %6.3f\n", __func__,
                    //        ctx->vocab.id_to_token[tokens_cur[i].id].c_str(), tokens_cur[i].p,
                    //        ctx->vocab.id_to_token[tokens_cur[i].tid].c_str(), tokens_cur[i].pt);

                    if (params.print_special || tokens_cur[i].id < whisper_token_eot(ctx)) {
                        text += whisper_token_to_str(ctx, tokens_cur[i].id);
                    }

                    // [TDRZ] record if speaker turn was predicted after current segment
                    if (params.tdrz_enable && tokens_cur[i].id == whisper_token_solm(ctx)) {
                        speaker_turn_next = true;
                    }

                    if (tokens_cur[i].id > whisper_token_beg(ctx) && !params.single_segment) {
                        const auto t1 = seek + 2*(tokens_cur[i].tid - whisper_token_beg(ctx));

                        if (!text.empty()) {
                            const auto tt0 = params.speed_up ? 2*t0 : t0;
                            const auto tt1 = params.speed_up ? 2*t1 : t1;

                            if (params.print_realtime) {
                                if (params.print_timestamps) {
                                    printf("[%s --> %s]  %s\n", to_timestamp(tt0).c_str(), to_timestamp(tt1).c_str(), text.c_str());
                                } else {
                                    printf("%s", text.c_str());
                                    fflush(stdout);
                                }
                            }

                            //printf("tt0 = %d, tt1 = %d, text = %s, token = %s, token_id = %d, tid = %d\n", tt0, tt1, text.c_str(), ctx->vocab.id_to_token[tokens_cur[i].id].c_str(), tokens_cur[i].id, tokens_cur[i].tid);

                            result_all.push_back({ tt0, tt1, text, {}, speaker_turn_next });
                            for (int j = i0; j <= i; j++) {
                                result_all.back().tokens.push_back(tokens_cur[j]);
                            }

                            int n_new = 1;

                            if (params.token_timestamps) {
                                whisper_exp_compute_token_level_timestamps(
                                        *ctx, *state, result_all.size() - 1, params.thold_pt, params.thold_ptsum);

                                if (params.max_len > 0) {
                                    n_new = whisper_wrap_segment(*ctx, *state, params.max_len, params.split_on_word);
                                }
                            }
                            if (params.new_segment_callback) {
                                params.new_segment_callback(ctx, state, n_new, params.new_segment_callback_user_data);
                            }
                        }
                        text = "";
                        while (i < (int) tokens_cur.size() && tokens_cur[i].id > whisper_token_beg(ctx)) {
                            i++;
                        }
                        i--;
                        t0 = t1;
                        i0 = i + 1;
                        speaker_turn_next = false;
                    }
                }

                if (!text.empty()) {
                    const auto t1 = seek + seek_delta;

                    const auto tt0 = params.speed_up ? 2*t0 : t0;
                    const auto tt1 = params.speed_up ? 2*t1 : t1;

                    if (params.print_realtime) {
                        if (params.print_timestamps) {
                            printf("[%s --> %s]  %s\n", to_timestamp(tt0).c_str(), to_timestamp(tt1).c_str(), text.c_str());
                        } else {
                            printf("%s", text.c_str());
                            fflush(stdout);
                        }
                    }

                    result_all.push_back({ tt0, tt1, text, {} , speaker_turn_next });
                    for (int j = i0; j < (int) tokens_cur.size(); j++) {
                        result_all.back().tokens.push_back(tokens_cur[j]);
                    }

                    int n_new = 1;

                    if (params.token_timestamps) {
                        whisper_exp_compute_token_level_timestamps(
                                *ctx, *state, result_all.size() - 1, params.thold_pt, params.thold_ptsum);

                        if (params.max_len > 0) {
                            n_new = whisper_wrap_segment(*ctx, *state, params.max_len, params.split_on_word);
                        }
                    }
                    if (params.new_segment_callback) {
                        params.new_segment_callback(ctx, state, n_new, params.new_segment_callback_user_data);
                    }
                }
            }

            // update audio window
            seek += seek_delta;

            WHISPER_PRINT_DEBUG("seek = %d, seek_delta = %d\n", seek, seek_delta);
        }
    }

    return 0;
}

int whisper_full(
        struct whisper_context * ctx,
    struct whisper_full_params   params,
                   const float * samples,
                           int   n_samples) {
    return whisper_full_with_state(ctx, ctx->state, params, samples, n_samples);
}

int whisper_full_parallel(
        struct whisper_context * ctx,
        struct whisper_full_params params,
        const float * samples,
        int n_samples,
        int n_processors) {
    if (n_processors == 1) {
        return whisper_full(ctx, params, samples, n_samples);
    }
    int ret = 0;

    // prepare separate states for each thread
    std::vector<whisper_state*> states;

    const int offset_samples = (WHISPER_SAMPLE_RATE*params.offset_ms)/1000;
    const int n_samples_per_processor = (n_samples - offset_samples)/n_processors;

    // the calling thread will process the first chunk
    // while the other threads will process the remaining chunks

    std::vector<std::thread> workers(n_processors - 1);
    for (int i = 0; i < n_processors - 1; ++i) {
        // create a new state for each thread
        states.push_back(whisper_init_state(ctx));

        const int start_samples = offset_samples + (i + 1)*n_samples_per_processor;
        const int n_samples_cur = (i == n_processors - 2) ? n_samples - start_samples : n_samples_per_processor;

        auto params_cur = params;

        params_cur.offset_ms = 0;
        params_cur.print_progress = false;
        params_cur.print_realtime = false;

        params_cur.new_segment_callback = nullptr;
        params_cur.new_segment_callback_user_data = nullptr;

        params_cur.progress_callback = nullptr;
        params_cur.progress_callback_user_data = nullptr;

        workers[i] = std::thread(whisper_full_with_state, ctx, states[i], std::move(params_cur), samples + start_samples, n_samples_cur);
    }

    {
        auto params_cur = params;

        // We need to disable the print real-time for this one as well, otherwise it will show only for the first chunk.
        params_cur.print_realtime = false;

        // Run the first transformation using default state but only for the first chunk.
        ret = whisper_full_with_state(ctx, ctx->state, std::move(params_cur), samples, offset_samples + n_samples_per_processor);
    }

    for (int i = 0; i < n_processors - 1; ++i) {
        workers[i].join();
    }

    const int64_t offset_t = (int64_t) params.offset_ms/10.0;

    // combine results into result_state->result_all from all other states
    for (int i = 0; i < n_processors - 1; ++i) {
        auto& results_i = states[i]->result_all;

        for (auto& result : results_i) {
            // correct the segment timestamp taking into account the offset
            result.t0 += 100 * ((i + 1) * n_samples_per_processor) / WHISPER_SAMPLE_RATE + offset_t;
            result.t1 += 100 * ((i + 1) * n_samples_per_processor) / WHISPER_SAMPLE_RATE + offset_t;

            // make sure that segments are not overlapping
            if (!ctx->state->result_all.empty()) {
                result.t0 = std::max(result.t0, ctx->state->result_all.back().t1);
            }

            ctx->state->result_all.push_back(std::move(result));

            // call the new_segment_callback for each segment
            if (params.new_segment_callback) {
                params.new_segment_callback(ctx, ctx->state, 1, params.new_segment_callback_user_data);
            }
        }

        ctx->state->t_mel_us += states[i]->t_mel_us;

        ctx->state->t_sample_us += states[i]->t_sample_us;
        ctx->state->t_encode_us += states[i]->t_encode_us;
        ctx->state->t_decode_us += states[i]->t_decode_us;
        ctx->state->t_prompt_us += states[i]->t_prompt_us;

        ctx->state->n_sample += states[i]->n_sample;
        ctx->state->n_encode += states[i]->n_encode;
        ctx->state->n_decode += states[i]->n_decode;
        ctx->state->n_prompt += states[i]->n_prompt;

        whisper_free_state(states[i]);
    }

    // average the timings
    ctx->state->t_mel_us    /= n_processors;
    ctx->state->t_sample_us /= n_processors;
    ctx->state->t_encode_us /= n_processors;
    ctx->state->t_decode_us /= n_processors;

    // print information about the audio boundaries
    log("\n");
    log("%s: the audio has been split into %d chunks at the following times:\n", __func__, n_processors);
    for (int i = 0; i < n_processors - 1; ++i) {
        log("%s: split %d - %s\n", __func__, (i + 1), to_timestamp(100*((i + 1)*n_samples_per_processor)/WHISPER_SAMPLE_RATE + offset_t).c_str());
    }
    log("%s: the transcription quality may be degraded near these boundaries\n", __func__);

    return ret;
}

int whisper_full_n_segments_from_state(struct whisper_state * state) {
    return state->result_all.size();
}

int whisper_full_n_segments(struct whisper_context * ctx) {
    return ctx->state->result_all.size();
}

int whisper_full_lang_id_from_state(struct whisper_state * state) {
    return state->lang_id;
}

int whisper_full_lang_id(struct whisper_context * ctx) {
    return ctx->state->lang_id;
}

int64_t whisper_full_get_segment_t0_from_state(struct whisper_state * state, int i_segment) {
    return state->result_all[i_segment].t0;
}

int64_t whisper_full_get_segment_t0(struct whisper_context * ctx, int i_segment) {
    return ctx->state->result_all[i_segment].t0;
}

int64_t whisper_full_get_segment_t1_from_state(struct whisper_state * state, int i_segment) {
    return state->result_all[i_segment].t1;
}

int64_t whisper_full_get_segment_t1(struct whisper_context * ctx, int i_segment) {
    return ctx->state->result_all[i_segment].t1;
}

bool whisper_full_get_segment_speaker_turn_next_from_state(struct whisper_state * state, int i_segment) {
    return state->result_all[i_segment].speaker_turn_next;
}

bool whisper_full_get_segment_speaker_turn_next(struct whisper_context * ctx, int i_segment) {
    return ctx->state->result_all[i_segment].speaker_turn_next;
}

const char * whisper_full_get_segment_text_from_state(struct whisper_state * state, int i_segment) {
    return state->result_all[i_segment].text.c_str();
}

const char * whisper_full_get_segment_text(struct whisper_context * ctx, int i_segment) {
    return ctx->state->result_all[i_segment].text.c_str();
}

int whisper_full_n_tokens_from_state(struct whisper_state * state, int i_segment) {
    return state->result_all[i_segment].tokens.size();
}

int whisper_full_n_tokens(struct whisper_context * ctx, int i_segment) {
    return ctx->state->result_all[i_segment].tokens.size();
}

const char * whisper_full_get_token_text_from_state(struct whisper_context * ctx, struct whisper_state * state, int i_segment, int i_token) {
    return ctx->vocab.id_to_token[state->result_all[i_segment].tokens[i_token].id].c_str();
}

const char* whisper_full_get_token_text(struct whisper_context * ctx, int i_segment, int i_token) {
    return ctx->vocab.id_to_token[ctx->state->result_all[i_segment].tokens[i_token].id].c_str();
}

whisper_token whisper_full_get_token_id_from_state(struct whisper_state * state, int i_segment, int i_token) {
    return state->result_all[i_segment].tokens[i_token].id;
}

whisper_token whisper_full_get_token_id(struct whisper_context * ctx, int i_segment, int i_token) {
    return ctx->state->result_all[i_segment].tokens[i_token].id;
}

struct whisper_token_data whisper_full_get_token_data_from_state(struct whisper_state * state, int i_segment, int i_token) {
    return state->result_all[i_segment].tokens[i_token];
}

struct whisper_token_data whisper_full_get_token_data(struct whisper_context * ctx, int i_segment, int i_token) {
    return ctx->state->result_all[i_segment].tokens[i_token];
}

float whisper_full_get_token_p_from_state(struct whisper_state * state, int i_segment, int i_token) {
    return state->result_all[i_segment].tokens[i_token].p;
}

float whisper_full_get_token_p(struct whisper_context * ctx, int i_segment, int i_token) {
    return ctx->state->result_all[i_segment].tokens[i_token].p;
}

// =================================================================================================

//
// Temporary interface needed for exposing ggml interface
// Will be removed in the future when ggml becomes a separate library
//

WHISPER_API int whisper_bench_memcpy(int n_threads) {
    fputs(whisper_bench_memcpy_str(n_threads), stderr);
    return 0;
}

WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
    static std::string s;
    s = "";
    char strbuf[256];

    ggml_time_init();

    size_t n    = 20;
    size_t arr  = n_threads > 0 ? 1024llu : n_threads; // trick to avoid compiler optimizations

    // 1GB MB array
    const size_t size = arr*1024llu*1024llu;

    // single-thread
    {
        char * src = (char *) malloc(size);
        char * dst = (char *) malloc(size);

        for (size_t i = 0; i < size; i++) src[i] = i;

        memcpy(dst, src, size); // heat-up

        double tsum = 0.0;
        double sum  = 0.0;

        for (size_t i = 0; i < n; i++) {
            const int64_t t0 = ggml_time_us();

            memcpy(dst, src, size);

            const int64_t t1 = ggml_time_us();

            tsum += (t1 - t0)*1e-6;

            src[rand() % size] = rand() % 256;
        }

        snprintf(strbuf, sizeof(strbuf), "memcpy: %.2f GB/s (1 thread)\n", (double) (n*size)/(tsum*1024llu*1024llu*1024llu));
        s += strbuf;

        // needed to prevent the compiler from optimizing the memcpy away
        {
            for (size_t i = 0; i < size; i++) sum += dst[i];

            snprintf(strbuf, sizeof(strbuf), "sum:    %f\n", sum);
            s += strbuf;
        }

        free(src);
        free(dst);
    }

    return s.c_str();
}

WHISPER_API int whisper_bench_ggml_mul_mat(int n_threads) {
    fputs(whisper_bench_ggml_mul_mat_str(n_threads), stderr);
    return 0;
}

WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
    static std::string s;
    s = "";
    char strbuf[256];

    ggml_time_init();

    const int n_max = 128;

    const std::vector<size_t> sizes = {
        64, 128, 256, 512, 1024, 2048, 4096,
    };

    const size_t N_max = sizes.back();

    // a: N*N*sizeof(float)
    // b: N*N*sizeof(float)
    // c: N*N*sizeof(float)
    // when F16 is used, there is an extra work buffer of size N*N*sizeof(float)
    std::vector<uint8_t> buf(3llu*N_max*N_max*sizeof(float) + 3*ggml_tensor_overhead());
    std::vector<uint8_t> work;

    // put a bunch of random data in the buffer
    for (size_t i = 0; i < buf.size(); i++) buf[i] = i;

    for (int j = 0; j < (int) sizes.size(); j++) {
        int n_q4_0 = 0;
        int n_q4_1 = 0;
        int n_q5_0 = 0;
        int n_q5_1 = 0;
        int n_q8_0 = 0;
        int n_fp16 = 0;
        int n_fp32 = 0;

        // GFLOPS/s
        double s_q4_0 = 0.0;
        double s_q4_1 = 0.0;
        double s_q5_0 = 0.0;
        double s_q5_1 = 0.0;
        double s_q8_0 = 0.0;
        double s_fp16 = 0.0;
        double s_fp32 = 0.0;

        const size_t N = sizes[j];

        for (int k = 0; k < 7; ++k) {
            const ggml_type wtype =
                k == 0 ? GGML_TYPE_Q4_0 :
                k == 1 ? GGML_TYPE_Q4_1 :
                k == 2 ? GGML_TYPE_Q5_0 :
                k == 3 ? GGML_TYPE_Q5_1 :
                k == 4 ? GGML_TYPE_Q8_0 :
                k == 5 ? GGML_TYPE_F16  : GGML_TYPE_F32;

            double & s = k == 0 ? s_q4_0 : k == 1 ? s_q4_1 : k == 2 ? s_q5_0 : k == 3 ? s_q5_1 : k == 4 ? s_q8_0 : k == 5 ? s_fp16 : /*k == 6*/ s_fp32;
            int    & n = k == 0 ? n_q4_0 : k == 1 ? n_q4_1 : k == 2 ? n_q5_0 : k == 3 ? n_q5_1 : k == 4 ? n_q8_0 : k == 5 ? n_fp16 : /*k == 6*/ n_fp32;

            struct ggml_init_params gparams = {
                /*.mem_size   =*/ buf.size(),
                /*.mem_buffer =*/ buf.data(),
                /*.no_alloc   =*/ false,
            };

            struct ggml_context * ctx0 = ggml_init(gparams);

            struct ggml_tensor * a = ggml_new_tensor_2d(ctx0, wtype,         N, N);
            struct ggml_tensor * b = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, N, N);

            struct ggml_tensor * c = ggml_mul_mat(ctx0, a, b);

            struct ggml_cgraph gf = ggml_build_forward(c);

            double tsum = 0.0;

            // heat-up
            ggml_graph_compute_helper(work, &gf, n_threads);

            for (int i = 0; i < n_max; ++i) {
                const int64_t t0 = ggml_time_us();

                ggml_graph_compute_helper(work, &gf, n_threads);

                const int64_t t1 = ggml_time_us();

                tsum += (t1 - t0)*1e-6;
                n++;

                if (tsum > 1.0 && n >= 3) {
                    break;
                }
            }

            ggml_free(ctx0);

            s = ((2.0*N*N*N*n)/tsum)*1e-9;
        }

        // Q4_0 | Q4_1
        snprintf(strbuf, sizeof(strbuf), "%4zu x %4zu: Q4_0 %7.1f GFLOPS (%3d runs) | Q4_1 %7.1f GFLOPS (%3d runs)\n",
                N, N, s_q4_0, n_q4_0, s_q4_1, n_q4_1);
        s += strbuf;

        // Q5_0 | Q5_1 | Q8_0
        snprintf(strbuf, sizeof(strbuf), "%4zu x %4zu: Q5_0 %7.1f GFLOPS (%3d runs) | Q5_1 %7.1f GFLOPS (%3d runs) | Q8_0 %7.1f GFLOPS (%3d runs)\n",
                N, N, s_q5_0, n_q5_0, s_q5_1, n_q5_1, s_q8_0, n_q8_0);
        s += strbuf;

        // F16 | F32
        snprintf(strbuf, sizeof(strbuf), "%4zu x %4zu: F16  %7.1f GFLOPS (%3d runs) | F32  %7.1f GFLOPS (%3d runs)\n",
                N, N, s_fp16, n_fp16, s_fp32, n_fp32);
        s += strbuf;
    }

    return s.c_str();
}

// =================================================================================================

// =================================================================================================

//
// Experimental stuff below
//
// Not sure if these should be part of the library at all, because the quality of the results is not
// guaranteed. Might get removed at some point unless a robust algorithm implementation is found
//

// =================================================================================================

//
// token-level timestamps
//

static int timestamp_to_sample(int64_t t, int n_samples) {
    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
}

static int64_t sample_to_timestamp(int i_sample) {
    return (100ll*i_sample)/WHISPER_SAMPLE_RATE;
}

// a cost-function / heuristic that is high for text that takes longer to pronounce
// obviously, can be improved
static float voice_length(const std::string & text) {
    float res = 0.0f;

    for (char c : text) {
        if (c == ' ') {
            res += 0.01f;
        } else if (c == ',') {
            res += 2.00f;
        } else if (c == '.') {
            res += 3.00f;
        } else if (c == '!') {
            res += 3.00f;
        } else if (c == '?') {
            res += 3.00f;
        } else if (c >= '0' && c <= '9') {
            res += 3.00f;
        } else {
            res += 1.00f;
        }
    }

    return res;
}

// average the fabs of the signal
static std::vector<float> get_signal_energy(const float * signal, int n_samples, int n_samples_per_half_window) {
    const int hw = n_samples_per_half_window;

    std::vector<float> result(n_samples);

    for (int i = 0; i < n_samples; i++) {
        float sum = 0;
        for (int j = -hw; j <= hw; j++) {
            if (i + j >= 0 && i + j < n_samples) {
                sum += fabs(signal[i + j]);
            }
        }
        result[i] = sum/(2*hw + 1);
    }

    return result;
}

static void whisper_exp_compute_token_level_timestamps(
        struct whisper_context & ctx,
          struct whisper_state & state,
                           int   i_segment,
                         float   thold_pt,
                         float   thold_ptsum) {
    auto & segment = state.result_all[i_segment];
    auto & tokens  = segment.tokens;

    const int n_samples = state.energy.size();

    if (n_samples == 0) {
        log("%s: no signal data available\n", __func__);
        return;
    }

    const int64_t t0 = segment.t0;
    const int64_t t1 = segment.t1;

    const int n = tokens.size();

    if (n == 0) {
        return;
    }

    if (n == 1) {
        tokens[0].t0 = t0;
        tokens[0].t1 = t1;

        return;
    }

    auto & t_beg    = state.t_beg;
    auto & t_last   = state.t_last;
    auto & tid_last = state.tid_last;

    for (int j = 0; j < n; ++j) {
        auto & token = tokens[j];

        if (j == 0) {
            if (token.id == whisper_token_beg(&ctx)) {
                tokens[j    ].t0 = t0;
                tokens[j    ].t1 = t0;
                tokens[j + 1].t0 = t0;

                t_beg    = t0;
                t_last   = t0;
                tid_last = whisper_token_beg(&ctx);
            } else {
                tokens[j    ].t0 = t_last;
            }
        }

        const int64_t tt = t_beg + 2*(token.tid - whisper_token_beg(&ctx));

        tokens[j].id    = token.id;
        tokens[j].tid   = token.tid;
        tokens[j].p     = token.p;
        tokens[j].pt    = token.pt;
        tokens[j].ptsum = token.ptsum;

        tokens[j].vlen = voice_length(whisper_token_to_str(&ctx, token.id));

        if (token.pt > thold_pt && token.ptsum > thold_ptsum && token.tid > tid_last && tt <= t1) {
            if (j > 0) {
                tokens[j - 1].t1 = tt;
            }
            tokens[j].t0 = tt;
            tid_last = token.tid;
        }
    }

    tokens[n - 2].t1 = t1;
    tokens[n - 1].t0 = t1;
    tokens[n - 1].t1 = t1;

    t_last = t1;

    // find intervals of tokens with unknown timestamps
    // fill the timestamps by proportionally splitting the interval based on the token voice lengths
    {
        int p0 = 0;
        int p1 = 0;

        while (true) {
            while (p1 < n && tokens[p1].t1 < 0) {
                p1++;
            }

            if (p1 >= n) {
                p1--;
            }

            //printf("p0=%d p1=%d t0=%lld t1=%lld\n", p0, p1, tokens[p0].t0, tokens[p1].t1);

            if (p1 > p0) {
                double psum = 0.0;
                for (int j = p0; j <= p1; j++) {
                    psum += tokens[j].vlen;
                }

                //printf("analyzing %d - %d, psum = %f\n", p0, p1, psum);

                const double dt = tokens[p1].t1 - tokens[p0].t0;

                // split the time proportionally to the voice length
                for (int j = p0 + 1; j <= p1; j++) {
                    const double ct = tokens[j - 1].t0 + dt*tokens[j - 1].vlen/psum;

                    tokens[j - 1].t1 = ct;
                    tokens[j    ].t0 = ct;
                }
            }

            p1++;
            p0 = p1;
            if (p1 >= n) {
                break;
            }
        }
    }

    // fix up (just in case)
    for (int j = 0; j < n - 1; j++) {
        if (tokens[j].t1 < 0) {
            tokens[j + 1].t0 = tokens[j].t1;
        }

        if (j > 0) {
            if (tokens[j - 1].t1 > tokens[j].t0) {
                tokens[j].t0 = tokens[j - 1].t1;
                tokens[j].t1 = std::max(tokens[j].t0, tokens[j].t1);
            }
        }
    }

    // VAD
    // expand or contract tokens based on voice activity
    {
        const int hw = WHISPER_SAMPLE_RATE/8;

        for (int j = 0; j < n; j++) {
            if (tokens[j].id >= whisper_token_eot(&ctx)) {
                continue;
            }

            int s0 = timestamp_to_sample(tokens[j].t0, n_samples);
            int s1 = timestamp_to_sample(tokens[j].t1, n_samples);

            const int ss0 = std::max(s0 - hw, 0);
            const int ss1 = std::min(s1 + hw, n_samples);

            const int ns = ss1 - ss0;

            float sum = 0.0f;

            for (int k = ss0; k < ss1; k++) {
                sum += state.energy[k];
            }

            const float thold = 0.5*sum/ns;

            {
                int k = s0;
                if (state.energy[k] > thold && j > 0) {
                    while (k > 0 && state.energy[k] > thold) {
                        k--;
                    }
                    tokens[j].t0 = sample_to_timestamp(k);
                    if (tokens[j].t0 < tokens[j - 1].t1) {
                        tokens[j].t0 = tokens[j - 1].t1;
                    } else {
                        s0 = k;
                    }
                } else {
                    while (state.energy[k] < thold && k < s1) {
                        k++;
                    }
                    s0 = k;
                    tokens[j].t0 = sample_to_timestamp(k);
                }
            }

            {
                int k = s1;
                if (state.energy[k] > thold) {
                    while (k < n_samples - 1 && state.energy[k] > thold) {
                        k++;
                    }
                    tokens[j].t1 = sample_to_timestamp(k);
                    if (j < ns - 1 && tokens[j].t1 > tokens[j + 1].t0) {
                        tokens[j].t1 = tokens[j + 1].t0;
                    } else {
                        s1 = k;
                    }
                } else {
                    while (state.energy[k] < thold && k > s0) {
                        k--;
                    }
                    s1 = k;
                    tokens[j].t1 = sample_to_timestamp(k);
                }
            }
        }
    }

    // fixed token expand (optional)
    //{
    //    const int t_expand = 0;

    //    for (int j = 0; j < n; j++) {
    //        if (j > 0) {
    //            tokens[j].t0 = std::max(0, (int) (tokens[j].t0 - t_expand));
    //        }
    //        if (j < n - 1) {
    //            tokens[j].t1 = tokens[j].t1 + t_expand;
    //        }
    //    }
    //}

    // debug info
    //for (int j = 0; j < n; ++j) {
    //    const auto & token = tokens[j];
    //    const auto tt = token.pt > thold_pt && token.ptsum > 0.01 ? whisper_token_to_str(&ctx, token.tid) : "[?]";
    //    printf("%s: %10s %6.3f %6.3f %6.3f %6.3f %5d %5d '%s'\n", __func__,
    //            tt, token.p, token.pt, token.ptsum, token.vlen, (int) token.t0, (int) token.t1, whisper_token_to_str(&ctx, token.id));

    //    if (tokens[j].id >= whisper_token_eot(&ctx)) {
    //        continue;
    //    }
    //}
}

void whisper_set_log_callback(whisper_log_callback callback) {
    whisper_log = callback;
}