| 1 | // This file contains functionality related to "GGUF" files, the binary file format used by ggml. |
| 2 | // GGUF files have the following structure: |
| 3 | // |
| 4 | // 1. File magic "GGUF" (4 bytes). |
| 5 | // 2. File version (uint32_t). |
| 6 | // 3. Number of ggml tensors in file (int64_t). |
| 7 | // 4. Number of key-value-pairs in file (int64_t). |
| 8 | // 5. For each KV pair: |
| 9 | // 1. The key (string). |
| 10 | // 2. The value type (gguf_type). |
| 11 | // 3a. If the value type is GGUF_TYPE_ARRAY: |
| 12 | // 1. The type of the array (gguf_type). |
| 13 | // 2. The number of elements in the array (uint64_t). |
| 14 | // 3. The binary representation of each element in the array. |
| 15 | // 3b. Otherwise: |
| 16 | // 1. The binary representation of the value. |
| 17 | // 6. For each ggml tensor: |
| 18 | // 1. The tensor name (string). |
| 19 | // 2. The number of dimensions of the tensor (uint32_t). |
| 20 | // 3. For each dimension: |
| 21 | // 1. The size of the tensor in the dimension (int64_t). |
| 22 | // 4. The tensor data type (ggml_type). |
| 23 | // 5. The tensor data offset in the tensor data binary blob (uint64_t). |
| 24 | // 7. The tensor data binary blob (optional, aligned). |
| 25 | // |
| 26 | // Strings are serialized as the string length (uint64_t) followed by the C string without the null terminator. |
| 27 | // All enums are stored as int32_t. |
| 28 | // All bool values are stored as int8_t. |
| 29 | // If the special key "general.alignment" (uint32_t) is defined it is used for alignment, |
| 30 | // otherwise GGUF_DEFAULT_ALIGNMENT is used. |
| 31 | // |
| 32 | // Module maintainer: Johannes Gäßler (@JohannesGaessler, johannesg@5d6.de) |
| 33 | |
| 34 | #pragma once |
| 35 | |
| 36 | #include "ggml.h" |
| 37 | |
| 38 | #include <stdbool.h> |
| 39 | #include <stdint.h> |
| 40 | |
| 41 | #define GGUF_MAGIC "GGUF" |
| 42 | #define GGUF_VERSION 3 |
| 43 | |
| 44 | #define GGUF_KEY_GENERAL_ALIGNMENT "general.alignment" |
| 45 | |
| 46 | #define GGUF_DEFAULT_ALIGNMENT 32 |
| 47 | |
| 48 | #ifdef __cplusplus |
| 49 | extern "C" { |
| 50 | #endif |
| 51 | |
| 52 | // types that can be stored as GGUF KV data |
| 53 | enum gguf_type { |
| 54 | GGUF_TYPE_UINT8 = 0, |
| 55 | GGUF_TYPE_INT8 = 1, |
| 56 | GGUF_TYPE_UINT16 = 2, |
| 57 | GGUF_TYPE_INT16 = 3, |
| 58 | GGUF_TYPE_UINT32 = 4, |
| 59 | GGUF_TYPE_INT32 = 5, |
| 60 | GGUF_TYPE_FLOAT32 = 6, |
| 61 | GGUF_TYPE_BOOL = 7, |
| 62 | GGUF_TYPE_STRING = 8, |
| 63 | GGUF_TYPE_ARRAY = 9, |
| 64 | GGUF_TYPE_UINT64 = 10, |
| 65 | GGUF_TYPE_INT64 = 11, |
| 66 | GGUF_TYPE_FLOAT64 = 12, |
| 67 | GGUF_TYPE_COUNT, // marks the end of the enum |
| 68 | }; |
| 69 | |
| 70 | struct gguf_context; |
| 71 | |
| 72 | struct gguf_init_params { |
| 73 | bool no_alloc; |
| 74 | |
| 75 | // if not NULL, create a ggml_context and allocate the tensor data in it |
| 76 | struct ggml_context ** ctx; |
| 77 | }; |
| 78 | |
| 79 | GGML_API struct gguf_context * gguf_init_empty(void); |
| 80 | GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params); |
| 81 | //GGML_API struct gguf_context * gguf_init_from_buffer(..); |
| 82 | |
| 83 | GGML_API void gguf_free(struct gguf_context * ctx); |
| 84 | |
| 85 | GGML_API const char * gguf_type_name(enum gguf_type type); |
| 86 | |
| 87 | GGML_API uint32_t gguf_get_version (const struct gguf_context * ctx); |
| 88 | GGML_API size_t gguf_get_alignment (const struct gguf_context * ctx); |
| 89 | GGML_API size_t gguf_get_data_offset(const struct gguf_context * ctx); |
| 90 | |
| 91 | GGML_API int64_t gguf_get_n_kv(const struct gguf_context * ctx); |
| 92 | GGML_API int64_t gguf_find_key(const struct gguf_context * ctx, const char * key); // returns -1 if key is not found |
| 93 | GGML_API const char * gguf_get_key (const struct gguf_context * ctx, int64_t key_id); |
| 94 | |
| 95 | GGML_API enum gguf_type gguf_get_kv_type (const struct gguf_context * ctx, int64_t key_id); |
| 96 | GGML_API enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int64_t key_id); |
| 97 | |
| 98 | // will abort if the wrong type is used for the key |
| 99 | GGML_API uint8_t gguf_get_val_u8 (const struct gguf_context * ctx, int64_t key_id); |
| 100 | GGML_API int8_t gguf_get_val_i8 (const struct gguf_context * ctx, int64_t key_id); |
| 101 | GGML_API uint16_t gguf_get_val_u16 (const struct gguf_context * ctx, int64_t key_id); |
| 102 | GGML_API int16_t gguf_get_val_i16 (const struct gguf_context * ctx, int64_t key_id); |
| 103 | GGML_API uint32_t gguf_get_val_u32 (const struct gguf_context * ctx, int64_t key_id); |
| 104 | GGML_API int32_t gguf_get_val_i32 (const struct gguf_context * ctx, int64_t key_id); |
| 105 | GGML_API float gguf_get_val_f32 (const struct gguf_context * ctx, int64_t key_id); |
| 106 | GGML_API uint64_t gguf_get_val_u64 (const struct gguf_context * ctx, int64_t key_id); |
| 107 | GGML_API int64_t gguf_get_val_i64 (const struct gguf_context * ctx, int64_t key_id); |
| 108 | GGML_API double gguf_get_val_f64 (const struct gguf_context * ctx, int64_t key_id); |
| 109 | GGML_API bool gguf_get_val_bool(const struct gguf_context * ctx, int64_t key_id); |
| 110 | GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int64_t key_id); |
| 111 | GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int64_t key_id); |
| 112 | GGML_API size_t gguf_get_arr_n (const struct gguf_context * ctx, int64_t key_id); |
| 113 | |
| 114 | // get raw pointer to the first element of the array with the given key_id |
| 115 | // for bool arrays, note that they are always stored as int8 on all platforms (usually this makes no difference) |
| 116 | GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int64_t key_id); |
| 117 | |
| 118 | // get ith C string from array with given key_id |
| 119 | GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int64_t key_id, size_t i); |
| 120 | |
| 121 | GGML_API int64_t gguf_get_n_tensors (const struct gguf_context * ctx); |
| 122 | GGML_API int64_t gguf_find_tensor (const struct gguf_context * ctx, const char * name); // returns -1 if the tensor is not found |
| 123 | GGML_API size_t gguf_get_tensor_offset(const struct gguf_context * ctx, int64_t tensor_id); |
| 124 | GGML_API const char * gguf_get_tensor_name (const struct gguf_context * ctx, int64_t tensor_id); |
| 125 | GGML_API enum ggml_type gguf_get_tensor_type (const struct gguf_context * ctx, int64_t tensor_id); |
| 126 | GGML_API size_t gguf_get_tensor_size (const struct gguf_context * ctx, int64_t tensor_id); |
| 127 | |
| 128 | // removes key if it exists, returns id that the key had prior to removal (-1 if it didn't exist) |
| 129 | GGML_API int64_t gguf_remove_key(struct gguf_context * ctx, const char * key); |
| 130 | |
| 131 | // overrides an existing KV pair or adds a new one, the new KV pair is always at the back |
| 132 | GGML_API void gguf_set_val_u8 (struct gguf_context * ctx, const char * key, uint8_t val); |
| 133 | GGML_API void gguf_set_val_i8 (struct gguf_context * ctx, const char * key, int8_t val); |
| 134 | GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val); |
| 135 | GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t val); |
| 136 | GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val); |
| 137 | GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t val); |
| 138 | GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float val); |
| 139 | GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val); |
| 140 | GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t val); |
| 141 | GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double val); |
| 142 | GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val); |
| 143 | GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val); |
| 144 | |
| 145 | // creates a new array with n elements of the given type and copies the corresponding number of bytes from data |
| 146 | GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, size_t n); |
| 147 | |
| 148 | // creates a new array with n strings and copies the corresponding strings from data |
| 149 | GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, size_t n); |
| 150 | |
| 151 | // set or add KV pairs from another context |
| 152 | GGML_API void gguf_set_kv(struct gguf_context * ctx, const struct gguf_context * src); |
| 153 | |
| 154 | // add tensor to GGUF context, tensor name must be unique |
| 155 | GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor); |
| 156 | |
| 157 | // after changing a tensor's type, the offsets of all tensors with higher indices are immediately recalculated |
| 158 | // in such a way that the tensor data remains as one contiguous block (except for padding) |
| 159 | GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type); |
| 160 | |
| 161 | // assumes that at least gguf_get_tensor_size bytes can be read from data |
| 162 | GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data); |
| 163 | |
| 164 | // writing gguf files can be done in 3 ways: |
| 165 | // |
| 166 | // - write the entire gguf_context to a binary file in a single pass: |
| 167 | // |
| 168 | // gguf_write_to_file(ctx, fname, /*only_meta =*/ false); |
| 169 | // |
| 170 | // - write only the meta data to a file, then re-open the file and append the tensor data: |
| 171 | // |
| 172 | // gguf_write_to_file(ctx, fname, /*only_meta =*/ true); |
| 173 | // FILE * f = fopen(fname, "ab"); |
| 174 | // fwrite(f, ...); // write tensor data |
| 175 | // fclose(f); |
| 176 | // |
| 177 | // - first prepare a file with a placeholder for the meta data, write the tensor data, then write the meta data: |
| 178 | // |
| 179 | // FILE * f = fopen(fname, "wb"); |
| 180 | // const size_t size_meta = gguf_get_meta_size(ctx); |
| 181 | // fseek(f, size_meta, SEEK_SET); |
| 182 | // fwrite(f, ...); // write tensor data |
| 183 | // void * data = malloc(size_meta); |
| 184 | // gguf_get_meta_data(ctx, data); |
| 185 | // rewind(f); |
| 186 | // fwrite(data, 1, data, f); |
| 187 | // free(data); |
| 188 | // fclose(f); |
| 189 | // |
| 190 | |
| 191 | // write the entire context to a binary file |
| 192 | GGML_API bool gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta); |
| 193 | |
| 194 | // get the size in bytes of the meta data (header, kv pairs, tensor info) including padding |
| 195 | GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx); |
| 196 | |
| 197 | // writes the meta data to pointer "data" |
| 198 | GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data); |
| 199 | |
| 200 | #ifdef __cplusplus |
| 201 | } |
| 202 | #endif |
| 203 | |