KBaba7/llama.cpp

1#include "ggml.h"2#include "gguf.h"3 4#include <cstdio>5#include <string>6#include <sstream>7#include <vector>8 9#undef MIN10#undef MAX11#define MIN(a, b) ((a) < (b) ? (a) : (b))12#define MAX(a, b) ((a) > (b) ? (a) : (b))13 14template <typename T>15static std::string to_string(const T & val) {16    std::stringstream ss;17    ss << val;18    return ss.str();19}20 21static bool gguf_ex_write(const std::string & fname) {22    struct gguf_context * ctx = gguf_init_empty();23 24    gguf_set_val_u8  (ctx, "some.parameter.uint8",    0x12);25    gguf_set_val_i8  (ctx, "some.parameter.int8",    -0x13);26    gguf_set_val_u16 (ctx, "some.parameter.uint16",   0x1234);27    gguf_set_val_i16 (ctx, "some.parameter.int16",   -0x1235);28    gguf_set_val_u32 (ctx, "some.parameter.uint32",   0x12345678);29    gguf_set_val_i32 (ctx, "some.parameter.int32",   -0x12345679);30    gguf_set_val_f32 (ctx, "some.parameter.float32",  0.123456789f);31    gguf_set_val_u64 (ctx, "some.parameter.uint64",   0x123456789abcdef0ull);32    gguf_set_val_i64 (ctx, "some.parameter.int64",   -0x123456789abcdef1ll);33    gguf_set_val_f64 (ctx, "some.parameter.float64",  0.1234567890123456789);34    gguf_set_val_bool(ctx, "some.parameter.bool",     true);35    gguf_set_val_str (ctx, "some.parameter.string",   "hello world");36 37    gguf_set_arr_data(ctx, "some.parameter.arr.i16", GGUF_TYPE_INT16,   std::vector<int16_t>{ 1, 2, 3, 4, }.data(), 4);38    gguf_set_arr_data(ctx, "some.parameter.arr.f32", GGUF_TYPE_FLOAT32, std::vector<float>{ 3.145f, 2.718f, 1.414f, }.data(), 3);39    gguf_set_arr_str (ctx, "some.parameter.arr.str",                    std::vector<const char *>{ "hello", "world", "!" }.data(), 3);40 41    struct ggml_init_params params = {42        /*.mem_size   =*/ 128ull*1024ull*1024ull,43        /*.mem_buffer =*/ NULL,44        /*.no_alloc   =*/ false,45    };46 47    struct ggml_context * ctx_data = ggml_init(params);48 49    const int n_tensors = 10;50 51    // tensor infos52    for (int i = 0; i < n_tensors; ++i) {53        const std::string name = "tensor_" + to_string(i);54 55        int64_t ne[GGML_MAX_DIMS] = { 1 };56        int32_t n_dims = rand() % GGML_MAX_DIMS + 1;57 58        for (int j = 0; j < n_dims; ++j) {59            ne[j] = rand() % 10 + 1;60        }61 62        struct ggml_tensor * cur = ggml_new_tensor(ctx_data, GGML_TYPE_F32, n_dims, ne);63        ggml_set_name(cur, name.c_str());64 65        {66            float * data = (float *) cur->data;67            for (int j = 0; j < ggml_nelements(cur); ++j) {68                data[j] = 100 + i;69            }70        }71 72        gguf_add_tensor(ctx, cur);73    }74 75    gguf_write_to_file(ctx, fname.c_str(), false);76 77    printf("%s: wrote file '%s;\n", __func__, fname.c_str());78 79    ggml_free(ctx_data);80    gguf_free(ctx);81 82    return true;83}84 85// just read tensor info86static bool gguf_ex_read_0(const std::string & fname) {87    struct gguf_init_params params = {88        /*.no_alloc = */ false,89        /*.ctx      = */ NULL,90    };91 92    struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);93 94    if (!ctx) {95        fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname.c_str());96        return false;97    }98 99    printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));100    printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));101    printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));102 103    // kv104    {105        const int n_kv = gguf_get_n_kv(ctx);106 107        printf("%s: n_kv: %d\n", __func__, n_kv);108 109        for (int i = 0; i < n_kv; ++i) {110            const char * key = gguf_get_key(ctx, i);111 112            printf("%s: kv[%d]: key = %s\n", __func__, i, key);113        }114    }115 116    // find kv string117    {118        const char * findkey = "some.parameter.string";119 120        const int keyidx = gguf_find_key(ctx, findkey);121        if (keyidx == -1) {122            printf("%s: find key: %s not found.\n", __func__, findkey);123        } else {124            const char * key_value = gguf_get_val_str(ctx, keyidx);125            printf("%s: find key: %s found, kv[%d] value = %s\n", __func__, findkey, keyidx, key_value);126        }127    }128 129    // tensor info130    {131        const int n_tensors = gguf_get_n_tensors(ctx);132 133        printf("%s: n_tensors: %d\n", __func__, n_tensors);134 135        for (int i = 0; i < n_tensors; ++i) {136            const char * name   = gguf_get_tensor_name  (ctx, i);137            const size_t size   = gguf_get_tensor_size  (ctx, i);138            const size_t offset = gguf_get_tensor_offset(ctx, i);139 140            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);141        }142    }143 144    gguf_free(ctx);145 146    return true;147}148 149// read and create ggml_context containing the tensors and their data150static bool gguf_ex_read_1(const std::string & fname, bool check_data) {151    struct ggml_context * ctx_data = NULL;152 153    struct gguf_init_params params = {154        /*.no_alloc = */ false,155        /*.ctx      = */ &ctx_data,156    };157 158    struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);159 160    printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));161    printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));162    printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));163 164    // kv165    {166        const int n_kv = gguf_get_n_kv(ctx);167 168        printf("%s: n_kv: %d\n", __func__, n_kv);169 170        for (int i = 0; i < n_kv; ++i) {171            const char * key = gguf_get_key(ctx, i);172 173            printf("%s: kv[%d]: key = %s\n", __func__, i, key);174        }175    }176 177    // tensor info178    {179        const int n_tensors = gguf_get_n_tensors(ctx);180 181        printf("%s: n_tensors: %d\n", __func__, n_tensors);182 183        for (int i = 0; i < n_tensors; ++i) {184            const char * name   = gguf_get_tensor_name  (ctx, i);185            const size_t size   = gguf_get_tensor_size  (ctx, i);186            const size_t offset = gguf_get_tensor_offset(ctx, i);187 188            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);189        }190    }191 192    // data193    {194        const int n_tensors = gguf_get_n_tensors(ctx);195 196        for (int i = 0; i < n_tensors; ++i) {197            printf("%s: reading tensor %d data\n", __func__, i);198 199            const char * name = gguf_get_tensor_name(ctx, i);200 201            struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);202 203            printf("%s: tensor[%d]: n_dims = %d, ne = (%d, %d, %d, %d), name = %s, data = %p\n",204                __func__, i, ggml_n_dims(cur), int(cur->ne[0]), int(cur->ne[1]), int(cur->ne[2]), int(cur->ne[3]), cur->name, cur->data);205 206            // print first 10 elements207            const float * data = (const float *) cur->data;208 209            printf("%s data[:10] : ", name);210            for (int j = 0; j < MIN(10, ggml_nelements(cur)); ++j) {211                printf("%f ", data[j]);212            }213            printf("\n\n");214 215            // check data216            if (check_data) {217                const float * data = (const float *) cur->data;218                for (int j = 0; j < ggml_nelements(cur); ++j) {219                    if (data[j] != 100 + i) {220                        fprintf(stderr, "%s: tensor[%d], data[%d]: found %f, expected %f\n", __func__, i, j, data[j], float(100 + i));221                        gguf_free(ctx);222                        return false;223                    }224                }225            }226        }227    }228 229    printf("%s: ctx_data size: %zu\n", __func__, ggml_get_mem_size(ctx_data));230 231    ggml_free(ctx_data);232    gguf_free(ctx);233 234    return true;235}236 237int main(int argc, char ** argv) {238    if (argc < 3) {239        printf("usage: %s data.gguf r|w [n]\n", argv[0]);240        printf("r: read data.gguf file\n");241        printf("w: write data.gguf file\n");242        printf("n: no check of tensor data\n");243        return -1;244    }245    bool check_data = true;246    if (argc == 4) {247        check_data = false;248    }249 250    srand(123456);251 252    const std::string fname(argv[1]);253    const std::string mode (argv[2]);254 255    GGML_ASSERT((mode == "r" || mode == "w") && "mode must be r or w");256 257    if (mode == "w") {258        GGML_ASSERT(gguf_ex_write(fname) && "failed to write gguf file");259    } else if (mode == "r") {260        GGML_ASSERT(gguf_ex_read_0(fname) && "failed to read gguf file");261        GGML_ASSERT(gguf_ex_read_1(fname, check_data) && "failed to read gguf file");262    }263 264    return 0;265}266