Felipe97/llama-cpp-compiled
01.1k
1#include "ggml.h"2#include "gguf.h"3 4#include <clocale>5#include <cstdio>6#include <string>7#include <sstream>8#include <vector>9 10#undef MIN11#undef MAX12#define MIN(a, b) ((a) < (b) ? (a) : (b))13#define MAX(a, b) ((a) > (b) ? (a) : (b))14 15template <typename T>16static std::string to_string(const T & val) {17 std::stringstream ss;18 ss << val;19 return ss.str();20}21 22static bool gguf_ex_write(const std::string & fname) {23 struct gguf_context * ctx = gguf_init_empty();24 25 gguf_set_val_u8 (ctx, "some.parameter.uint8", 0x12);26 gguf_set_val_i8 (ctx, "some.parameter.int8", -0x13);27 gguf_set_val_u16 (ctx, "some.parameter.uint16", 0x1234);28 gguf_set_val_i16 (ctx, "some.parameter.int16", -0x1235);29 gguf_set_val_u32 (ctx, "some.parameter.uint32", 0x12345678);30 gguf_set_val_i32 (ctx, "some.parameter.int32", -0x12345679);31 gguf_set_val_f32 (ctx, "some.parameter.float32", 0.123456789f);32 gguf_set_val_u64 (ctx, "some.parameter.uint64", 0x123456789abcdef0ull);33 gguf_set_val_i64 (ctx, "some.parameter.int64", -0x123456789abcdef1ll);34 gguf_set_val_f64 (ctx, "some.parameter.float64", 0.1234567890123456789);35 gguf_set_val_bool(ctx, "some.parameter.bool", true);36 gguf_set_val_str (ctx, "some.parameter.string", "hello world");37 38 gguf_set_arr_data(ctx, "some.parameter.arr.i16", GGUF_TYPE_INT16, std::vector<int16_t>{ 1, 2, 3, 4, }.data(), 4);39 gguf_set_arr_data(ctx, "some.parameter.arr.f32", GGUF_TYPE_FLOAT32, std::vector<float>{ 3.145f, 2.718f, 1.414f, }.data(), 3);40 gguf_set_arr_str (ctx, "some.parameter.arr.str", std::vector<const char *>{ "hello", "world", "!" }.data(), 3);41 42 struct ggml_init_params params = {43 /*.mem_size =*/ 128ull*1024ull*1024ull,44 /*.mem_buffer =*/ NULL,45 /*.no_alloc =*/ false,46 };47 48 struct ggml_context * ctx_data = ggml_init(params);49 50 const int n_tensors = 10;51 52 // tensor infos53 for (int i = 0; i < n_tensors; ++i) {54 const std::string name = "tensor_" + to_string(i);55 56 int64_t ne[GGML_MAX_DIMS] = { 1 };57 int32_t n_dims = rand() % GGML_MAX_DIMS + 1;58 59 for (int j = 0; j < n_dims; ++j) {60 ne[j] = rand() % 10 + 1;61 }62 63 struct ggml_tensor * cur = ggml_new_tensor(ctx_data, GGML_TYPE_F32, n_dims, ne);64 ggml_set_name(cur, name.c_str());65 66 {67 float * data = (float *) cur->data;68 for (int j = 0; j < ggml_nelements(cur); ++j) {69 data[j] = 100 + i;70 }71 }72 73 gguf_add_tensor(ctx, cur);74 }75 76 gguf_write_to_file(ctx, fname.c_str(), false);77 78 printf("%s: wrote file '%s;\n", __func__, fname.c_str());79 80 ggml_free(ctx_data);81 gguf_free(ctx);82 83 return true;84}85 86// just read tensor info87static bool gguf_ex_read_0(const std::string & fname) {88 struct gguf_init_params params = {89 /*.no_alloc = */ false,90 /*.ctx = */ NULL,91 };92 93 struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);94 95 if (!ctx) {96 fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname.c_str());97 return false;98 }99 100 printf("%s: version: %d\n", __func__, gguf_get_version(ctx));101 printf("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx));102 printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));103 104 // kv105 {106 const int n_kv = gguf_get_n_kv(ctx);107 108 printf("%s: n_kv: %d\n", __func__, n_kv);109 110 for (int i = 0; i < n_kv; ++i) {111 const char * key = gguf_get_key(ctx, i);112 113 printf("%s: kv[%d]: key = %s\n", __func__, i, key);114 }115 }116 117 // find kv string118 {119 const char * findkey = "some.parameter.string";120 121 const int keyidx = gguf_find_key(ctx, findkey);122 if (keyidx == -1) {123 printf("%s: find key: %s not found.\n", __func__, findkey);124 } else {125 const char * key_value = gguf_get_val_str(ctx, keyidx);126 printf("%s: find key: %s found, kv[%d] value = %s\n", __func__, findkey, keyidx, key_value);127 }128 }129 130 // tensor info131 {132 const int n_tensors = gguf_get_n_tensors(ctx);133 134 printf("%s: n_tensors: %d\n", __func__, n_tensors);135 136 for (int i = 0; i < n_tensors; ++i) {137 const char * name = gguf_get_tensor_name (ctx, i);138 const size_t size = gguf_get_tensor_size (ctx, i);139 const size_t offset = gguf_get_tensor_offset(ctx, i);140 141 printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);142 }143 }144 145 gguf_free(ctx);146 147 return true;148}149 150// read and create ggml_context containing the tensors and their data151static bool gguf_ex_read_1(const std::string & fname, bool check_data) {152 struct ggml_context * ctx_data = NULL;153 154 struct gguf_init_params params = {155 /*.no_alloc = */ false,156 /*.ctx = */ &ctx_data,157 };158 159 struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);160 161 printf("%s: version: %d\n", __func__, gguf_get_version(ctx));162 printf("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx));163 printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));164 165 // kv166 {167 const int n_kv = gguf_get_n_kv(ctx);168 169 printf("%s: n_kv: %d\n", __func__, n_kv);170 171 for (int i = 0; i < n_kv; ++i) {172 const char * key = gguf_get_key(ctx, i);173 174 printf("%s: kv[%d]: key = %s\n", __func__, i, key);175 }176 }177 178 // tensor info179 {180 const int n_tensors = gguf_get_n_tensors(ctx);181 182 printf("%s: n_tensors: %d\n", __func__, n_tensors);183 184 for (int i = 0; i < n_tensors; ++i) {185 const char * name = gguf_get_tensor_name (ctx, i);186 const size_t size = gguf_get_tensor_size (ctx, i);187 const size_t offset = gguf_get_tensor_offset(ctx, i);188 const auto type = gguf_get_tensor_type (ctx, i);189 190 const char * type_name = ggml_type_name(type);191 const size_t type_size = ggml_type_size(type);192 const size_t n_elements = size / type_size;193 194 printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu, type = %s, n_elts = %zu\n", __func__, i, name, size, offset, type_name, n_elements);195 }196 }197 198 // data199 {200 const int n_tensors = gguf_get_n_tensors(ctx);201 202 for (int i = 0; i < n_tensors; ++i) {203 printf("%s: reading tensor %d data\n", __func__, i);204 205 const char * name = gguf_get_tensor_name(ctx, i);206 207 struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);208 209 printf("%s: tensor[%d]: n_dims = %d, ne = (%d, %d, %d, %d), name = %s, data = %p\n",210 __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);211 212 // print first 10 elements213 const float * data = (const float *) cur->data;214 215 printf("%s data[:10] : ", name);216 for (int j = 0; j < MIN(10, ggml_nelements(cur)); ++j) {217 printf("%f ", data[j]);218 }219 printf("\n\n");220 221 // check data222 if (check_data) {223 const float * data = (const float *) cur->data;224 for (int j = 0; j < ggml_nelements(cur); ++j) {225 if (data[j] != 100 + i) {226 fprintf(stderr, "%s: tensor[%d], data[%d]: found %f, expected %f\n", __func__, i, j, data[j], float(100 + i));227 gguf_free(ctx);228 return false;229 }230 }231 }232 }233 }234 235 printf("%s: ctx_data size: %zu\n", __func__, ggml_get_mem_size(ctx_data));236 237 ggml_free(ctx_data);238 gguf_free(ctx);239 240 return true;241}242 243int main(int argc, char ** argv) {244 std::setlocale(LC_NUMERIC, "C");245 246 if (argc < 3) {247 printf("usage: %s data.gguf r|w [n]\n", argv[0]);248 printf("r: read data.gguf file\n");249 printf("w: write data.gguf file\n");250 printf("n: no check of tensor data\n");251 return -1;252 }253 bool check_data = true;254 if (argc == 4) {255 check_data = false;256 }257 258 srand(123456);259 260 const std::string fname(argv[1]);261 const std::string mode (argv[2]);262 263 GGML_ASSERT((mode == "r" || mode == "w") && "mode must be r or w");264 265 if (mode == "w") {266 GGML_ASSERT(gguf_ex_write(fname) && "failed to write gguf file");267 } else if (mode == "r") {268 GGML_ASSERT(gguf_ex_read_0(fname) && "failed to read gguf file");269 GGML_ASSERT(gguf_ex_read_1(fname, check_data) && "failed to read gguf file");270 }271 272 return 0;273}274