Brunobkr/llama.cpp_AlgMor24_github

1#include "llama-model-saver.h"2 3#include "ggml.h"4#include "gguf.h"5 6#include "llama-arch.h"7#include "llama.h"8#include "llama-hparams.h"9#include "llama-model.h"10#include "llama-vocab.h"11 12#include <cstdint>13#include <string>14 15bool llama_model_saver_supports_arch(llm_arch arch) {16    switch (arch) {17        case LLM_ARCH_PLAMO3:18        case LLM_ARCH_GEMMA3:19        case LLM_ARCH_GEMMA3N:20        case LLM_ARCH_COHERE2:21        case LLM_ARCH_COHERE2MOE:22        case LLM_ARCH_OLMO2:23        case LLM_ARCH_BITNET:24        case LLM_ARCH_T5:25        case LLM_ARCH_EXAONE_MOE:26        case LLM_ARCH_AFMOE:27        case LLM_ARCH_APERTUS:28        case LLM_ARCH_MIMO2:29        case LLM_ARCH_STEP35:30        case LLM_ARCH_MUSE_GLIMMER:31        case LLM_ARCH_MELLUM:32        case LLM_ARCH_LAGUNA:33            return false;34        default:35            return true;36    }37}38 39llama_model_saver::llama_model_saver(const struct llama_model * model) :40        gguf_ctx(gguf_init_empty()), gguf_ctx_owned(true), model(model), llm_kv(model->arch) {41    GGML_ASSERT(llama_model_saver_supports_arch(model->arch));42}43 44llama_model_saver::llama_model_saver(enum llm_arch arch, struct gguf_context * gguf_ctx) :45        gguf_ctx(gguf_ctx == nullptr ? gguf_init_empty() : gguf_ctx), gguf_ctx_owned(gguf_ctx == nullptr), model(nullptr), llm_kv(arch) {}46 47llama_model_saver::~llama_model_saver() {48    if (gguf_ctx_owned) {49        gguf_free(gguf_ctx);50    }51}52 53void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) {54    gguf_set_val_u32(gguf_ctx, llm_kv(key).c_str(), value);55}56 57void llama_model_saver::add_kv(const enum llm_kv key, const int32_t value) {58    gguf_set_val_i32(gguf_ctx, llm_kv(key).c_str(), value);59}60 61void llama_model_saver::add_kv(const enum llm_kv key, const float value) {62    gguf_set_val_f32(gguf_ctx, llm_kv(key).c_str(), value);63}64 65void llama_model_saver::add_kv(const enum llm_kv key, const bool value) {66    gguf_set_val_bool(gguf_ctx, llm_kv(key).c_str(), value);67}68 69void llama_model_saver::add_kv(const enum llm_kv key, const char * value) {70    gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), value);71}72 73[[noreturn]]74void llama_model_saver::add_kv(const enum llm_kv key, const char value) {75    GGML_UNUSED(key);76    GGML_UNUSED(value);77    GGML_ABORT("fatal error"); // this should never be called, only needed to make the template below compile78}79 80template <typename Container>81void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, const bool per_layer) {82    GGML_ASSERT(model != nullptr || !per_layer);83    const size_t n_values = per_layer ? size_t(model->hparams.n_layer()) : value.size();84    GGML_ASSERT(n_values <= value.size());85 86    if (n_values == 0) {87        return;88    }89 90    if (per_layer) {91        bool all_values_the_same = true;92        for (size_t i = 1; i < n_values; ++i) {93            if (value[i] != value[0]) {94                all_values_the_same = false;95                break;96            }97        }98        if (all_values_the_same) {99            add_kv(key, value[0]);100            return;101        }102    }103 104    if (std::is_same<typename Container::value_type, uint8_t>::value) {105        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT8, value.data(), n_values);106    } else if (std::is_same<typename Container::value_type, int8_t>::value) {107        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT8, value.data(), n_values);108    } else if (std::is_same<typename Container::value_type, uint32_t>::value) {109        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT32, value.data(), n_values);110    } else if (std::is_same<typename Container::value_type, bool>::value) {111        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_BOOL, value.data(), n_values);112    } else if (std::is_same<typename Container::value_type, int32_t>::value) {113        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT32, value.data(), n_values);114    } else if (std::is_same<typename Container::value_type, float>::value) {115        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_FLOAT32, value.data(), n_values);116    } else if (std::is_same<Container, std::string>::value) {117        gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), reinterpret_cast<const char *>(value.data()));118    } else {119        GGML_ABORT("fatal error");120    }121}122// instantiate for external usage:123template void llama_model_saver::add_kv<std::vector<uint32_t>>(const enum llm_kv, const std::vector<uint32_t> &, const bool);124 125void llama_model_saver::add_kv(const enum llm_kv key, const std::vector<std::string> & value) {126    std::vector<const char *> tmp(value.size());127    for (size_t i = 0; i < value.size(); ++i) {128        tmp[i] = value[i].c_str();129    }130    gguf_set_arr_str(gguf_ctx, llm_kv(key).c_str(), tmp.data(), tmp.size());131}132 133void llama_model_saver::add_tensor(const struct ggml_tensor * tensor) {134    if (!tensor) {135        return;136    }137    if (gguf_find_tensor(gguf_ctx, tensor->name) >= 0) {138        const std::string tensor_name = tensor->name;139        GGML_ASSERT(140            tensor_name == "rope_freqs.weight" || tensor_name == "rope_factors_long.weight" ||141            tensor_name == "rope_factors_short.weight"); // FIXME142        return;143    }144    gguf_add_tensor(gguf_ctx, tensor);145}146 147void llama_model_saver::add_kv_from_model() {148    const llama_hparams & hparams = model->hparams;149    const llama_vocab   & vocab   = model->vocab;150 151    const int32_t n_vocab = vocab.n_tokens();152    std::vector<std::string> tokens(n_vocab);153    std::vector<float>       scores(n_vocab);154    std::vector<int32_t>     token_types(n_vocab);155 156    if (vocab.get_type() != LLAMA_VOCAB_TYPE_NONE) {157        for (int32_t id = 0; id < n_vocab; ++id) {158            const llama_vocab::token_data & token_data = vocab.get_token_data(id);159 160            tokens[id] = token_data.text;161            scores[id] = token_data.score;162 163            // FIXME should this be treated as flags?164            switch(token_data.attr) {165                case LLAMA_TOKEN_ATTR_UNKNOWN:      token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN;      break;166                case LLAMA_TOKEN_ATTR_UNUSED:       token_types[id] = LLAMA_TOKEN_TYPE_UNUSED;       break;167                case LLAMA_TOKEN_ATTR_NORMAL:       token_types[id] = LLAMA_TOKEN_TYPE_NORMAL;       break;168                case LLAMA_TOKEN_ATTR_CONTROL:      token_types[id] = LLAMA_TOKEN_TYPE_CONTROL;      break;169                case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;170                case LLAMA_TOKEN_ATTR_BYTE:         token_types[id] = LLAMA_TOKEN_TYPE_BYTE;         break;171                // case LLAMA_TOKEN_ATTR_NORMALIZED:   ???172                // case LLAMA_TOKEN_ATTR_LSTRIP:       ???173                // case LLAMA_TOKEN_ATTR_RSTRIP:       ???174                case LLAMA_TOKEN_ATTR_UNDEFINED:175                default:                            token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED;    break;176            }177        }178    }179 180    // add_kv(LLM_KV_GENERAL_TYPE,                      ???);181    add_kv(LLM_KV_GENERAL_ARCHITECTURE,              model->arch_name());182    // add_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION,      ???);183    // add_kv(LLM_KV_GENERAL_ALIGNMENT,                 ???);184    // add_kv(LLM_KV_GENERAL_FILE_TYPE,                 ???);185    // add_kv(LLM_KV_GENERAL_SAMPLING_SEQUENCE,         ???);186    // add_kv(LLM_KV_GENERAL_SAMPLING_TOP_K,            ???);187    // add_kv(LLM_KV_GENERAL_SAMPLING_TOP_P,            ???);188    // add_kv(LLM_KV_GENERAL_SAMPLING_MIN_P,            ???);189    // add_kv(LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY,  ???);190    // add_kv(LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD,    ???);191    // add_kv(LLM_KV_GENERAL_SAMPLING_TEMP,             ???);192    // add_kv(LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N,   ???);193    // add_kv(LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT,   ???);194    // add_kv(LLM_KV_GENERAL_SAMPLING_MIROSTAT,         ???);195    // add_kv(LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU,     ???);196    // add_kv(LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA,     ???);197    add_kv(LLM_KV_GENERAL_NAME,                      model->name);198    // add_kv(LLM_KV_GENERAL_AUTHOR,                    ???);199    // add_kv(LLM_KV_GENERAL_VERSION,                   ???);200    // add_kv(LLM_KV_GENERAL_URL,                       ???);201    // add_kv(LLM_KV_GENERAL_DESCRIPTION,               ???);202    // add_kv(LLM_KV_GENERAL_LICENSE,                   ???);203    // add_kv(LLM_KV_GENERAL_SOURCE_URL,                ???);204    // add_kv(LLM_KV_GENERAL_SOURCE_HF_REPO,            ???);205 206    add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());207    add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);208    add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);209    if (hparams.n_embd_out_impl > 0) {210        add_kv(LLM_KV_EMBEDDING_LENGTH_OUT,          hparams.n_embd_out_impl);211    }212    add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer_all);213    add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);214    add_kv(LLM_KV_FEED_FORWARD_LENGTH,               hparams.n_ff_arr, true);215    add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);216    add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp);217    add_kv(LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,  hparams.n_ff_chexp);218    add_kv(LLM_KV_SWIGLU_CLAMP_EXP,                  hparams.swiglu_clamp_exp);219    add_kv(LLM_KV_SWIGLU_CLAMP_SHEXP,                hparams.swiglu_clamp_shexp);220    add_kv(LLM_KV_USE_PARALLEL_RESIDUAL,             hparams.use_par_res);221    // add_kv(LLM_KV_TENSOR_DATA_LAYOUT,                ???);222    add_kv(LLM_KV_EXPERT_COUNT,                      hparams.n_expert);223    add_kv(LLM_KV_EXPERT_USED_COUNT,                 hparams.n_expert_used);224    add_kv(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);225    add_kv(LLM_KV_EXPERT_GROUP_COUNT,                hparams.n_expert_groups);226    add_kv(LLM_KV_EXPERT_GROUP_USED_COUNT,           hparams.n_group_used);227    add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale);228    add_kv(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm);229    add_kv(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);230    add_kv(LLM_KV_EXPERT_GROUP_SCALE,                hparams.expert_group_scale);231    add_kv(LLM_KV_EXPERTS_PER_GROUP,                 hparams.n_group_experts);232    add_kv(LLM_KV_MOE_EVERY_N_LAYERS,                hparams.moe_every_n_layers);233    add_kv(LLM_KV_NEXTN_PREDICT_LAYERS,              hparams.n_layer_nextn);234    add_kv(LLM_KV_NUM_DEEPSTACK_LAYERS,              hparams.n_deepstack_layers);235    add_kv(LLM_KV_DEEPSTACK_MAPPING,                 hparams.deepstack_mapping_arr);236    add_kv(LLM_KV_POOLING_TYPE,                      uint32_t(hparams.pooling_type));237    add_kv(LLM_KV_LOGIT_SCALE,                       hparams.f_logit_scale);238    add_kv(LLM_KV_DECODER_START_TOKEN_ID,            hparams.dec_start_token_id);239    add_kv(LLM_KV_DECODER_BLOCK_COUNT,               hparams.dec_n_layer);240    add_kv(LLM_KV_ATTN_LOGIT_SOFTCAPPING,            hparams.f_attn_logit_softcapping);241    add_kv(LLM_KV_ROUTER_LOGIT_SOFTCAPPING,          hparams.f_router_logit_softcapping);242    add_kv(LLM_KV_FINAL_LOGIT_SOFTCAPPING,           hparams.f_final_logit_softcapping);243    add_kv(LLM_KV_SWIN_NORM,                         hparams.swin_norm);244    add_kv(LLM_KV_RESCALE_EVERY_N_LAYERS,            hparams.rescale_every_n_layers);245    add_kv(LLM_KV_TIME_MIX_EXTRA_DIM,                hparams.time_mix_extra_dim);246    add_kv(LLM_KV_TIME_DECAY_EXTRA_DIM,              hparams.time_decay_extra_dim);247    add_kv(LLM_KV_RESIDUAL_SCALE,                    hparams.f_residual_scale);248    add_kv(LLM_KV_EMBEDDING_SCALE,                   hparams.f_embedding_scale);249    add_kv(LLM_KV_TOKEN_SHIFT_COUNT,                 hparams.token_shift_count);250    add_kv(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         hparams.n_moe_layer_step);251    // add_kv(LLM_KV_FULL_ATTENTION_INTERVAL,           ???); // saved as LLM_KV_ATTENTION_RECURRENT_LAYERS instead252 253    add_kv(LLM_KV_ATTENTION_HEAD_COUNT,              hparams.n_head_arr, true);254    add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);255    add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS,          hparams.f_max_alibi_bias);256    add_kv(LLM_KV_ATTENTION_CLAMP_KQV,               hparams.f_clamp_kqv);257    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k_full);258    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v_full);259    add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,           hparams.f_norm_eps);260    add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);261    add_kv(LLM_KV_ATTENTION_GROUPNORM_EPS,           hparams.f_norm_group_eps);262    add_kv(LLM_KV_ATTENTION_GROUPNORM_GROUPS,        hparams.n_norm_groups);263    add_kv(LLM_KV_ATTENTION_CAUSAL,                  hparams.causal_attn);264    add_kv(LLM_KV_ATTENTION_Q_LORA_RANK,             hparams.n_lora_q);265    add_kv(LLM_KV_ATTENTION_KV_LORA_RANK,            hparams.n_lora_kv);266    add_kv(LLM_KV_ATTENTION_DECAY_LORA_RANK,         hparams.n_lora_decay);267    add_kv(LLM_KV_ATTENTION_ICLR_LORA_RANK,          hparams.n_lora_iclr);268    add_kv(LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, hparams.n_lora_value_res_mix);269    add_kv(LLM_KV_ATTENTION_GATE_LORA_RANK,          hparams.n_lora_gate);270    add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,  hparams.n_rel_attn_bkts);271    add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);272    // add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,  ???);273    add_kv(LLM_KV_ATTENTION_SCALE,                   hparams.f_attention_scale);274    add_kv(LLM_KV_ATTENTION_OUTPUT_SCALE,            hparams.f_attn_out_scale);275    add_kv(LLM_KV_ATTENTION_VALUE_SCALE,             hparams.f_attn_value_scale);276    add_kv(LLM_KV_ATTENTION_TEMPERATURE_LENGTH,      hparams.attn_temp_length);277    add_kv(LLM_KV_ATTENTION_TEMPERATURE_SCALE,       hparams.f_attn_temp_scale);278    add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA,          hparams.n_embd_head_k_mla_impl);279    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_MLA,        hparams.n_embd_head_v_mla_impl);280    add_kv(LLM_KV_ATTENTION_KEY_LENGTH_SWA,          hparams.n_embd_head_k_swa);281    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,        hparams.n_embd_head_v_swa);282    add_kv(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,      hparams.indexer_n_head);283    add_kv(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,      hparams.indexer_head_size);284    add_kv(LLM_KV_ATTENTION_INDEXER_TOP_K,           hparams.indexer_top_k);285    add_kv(LLM_KV_ATTENTION_INDEXER_BLOCK_SIZE,      hparams.indexer_block_size);286    add_kv(LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS,    hparams.indexer_local_blocks);287    add_kv(LLM_KV_ATTENTION_INDEXER_TYPES,           hparams.is_indexer_full_impl, true);288    add_kv(LLM_KV_ATTENTION_RECURRENT_LAYERS,        hparams.is_recr_impl, true);289 290    const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;291 292    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot_full);293    add_kv(LLM_KV_ROPE_DIMENSION_COUNT_SWA,          hparams.n_rot_swa);294    add_kv(LLM_KV_ROPE_DIMENSION_SECTIONS,           hparams.rope_sections);295    add_kv(LLM_KV_ROPE_FREQ_BASE,                    hparams.rope_freq_base_train);296    add_kv(LLM_KV_ROPE_FREQ_BASE_SWA,                hparams.rope_freq_base_train_swa);297    // add_kv(LLM_KV_ROPE_SCALE_LINEAR,                 rope_scaling_factor); // old name298    add_kv(LLM_KV_ROPE_SCALING_TYPE,                 llama_rope_scaling_type_name(hparams.rope_scaling_type_train));299    add_kv(LLM_KV_ROPE_SCALING_FACTOR,               rope_scaling_factor);300    add_kv(LLM_KV_ROPE_SCALING_ATTN_FACTOR,          hparams.rope_attn_factor);301    add_kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,         hparams.n_ctx_orig_yarn);302    add_kv(LLM_KV_ROPE_SCALING_FINETUNED,            hparams.rope_finetuned);303    add_kv(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,         hparams.rope_yarn_log_mul);304    add_kv(LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,      hparams.yarn_ext_factor);305    add_kv(LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR,     hparams.yarn_attn_factor);306    add_kv(LLM_KV_ROPE_SCALING_YARN_BETA_FAST,       hparams.yarn_beta_fast);307    add_kv(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,       hparams.yarn_beta_slow);308 309    // TODO: implement split file support310    // add_kv(LLM_KV_SPLIT_NO,                          ???);311    // add_kv(LLM_KV_SPLIT_COUNT,                       ???);312    // add_kv(LLM_KV_SPLIT_TENSORS_COUNT,               ???);313 314    add_kv(LLM_KV_SSM_INNER_SIZE,                    hparams.ssm_d_inner);315    add_kv(LLM_KV_SSM_CONV_KERNEL,                   hparams.ssm_d_conv);316    add_kv(LLM_KV_SSM_STATE_SIZE,                    hparams.ssm_d_state);317    add_kv(LLM_KV_SSM_TIME_STEP_RANK,                hparams.ssm_dt_rank);318    add_kv(LLM_KV_SSM_GROUP_COUNT,                   hparams.ssm_n_group);319    add_kv(LLM_KV_SSM_DT_B_C_RMS,                    hparams.ssm_dt_b_c_rms);320 321    add_kv(LLM_KV_KDA_HEAD_DIM,                      hparams.n_embd_head_kda);322 323    add_kv(LLM_KV_WKV_HEAD_SIZE,                     hparams.wkv_head_size);324 325    add_kv(LLM_KV_TOKENIZER_MODEL,                   vocab.get_tokenizer_model());326    add_kv(LLM_KV_TOKENIZER_PRE,                     vocab.get_tokenizer_pre());327    add_kv(LLM_KV_TOKENIZER_LIST,                    tokens);328    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE,              token_types);329    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,        vocab.n_token_types());330    add_kv(LLM_KV_TOKENIZER_SCORES,                  scores);331    add_kv(LLM_KV_TOKENIZER_MERGES,                  vocab.get_bpe_merges());332    // FIXME llama_token is type i32 but when reading in a GGUF file u32 is expected, not an issue for writing though333    add_kv(LLM_KV_TOKENIZER_BOS_ID,                  uint32_t(vocab.token_bos()));334    add_kv(LLM_KV_TOKENIZER_EOS_ID,                  uint32_t(vocab.token_eos()));335    add_kv(LLM_KV_TOKENIZER_EOT_ID,                  uint32_t(vocab.token_eot()));336    add_kv(LLM_KV_TOKENIZER_EOM_ID,                  uint32_t(vocab.token_eom()));337    add_kv(LLM_KV_TOKENIZER_UNK_ID,                  uint32_t(vocab.token_unk()));338    add_kv(LLM_KV_TOKENIZER_SEP_ID,                  uint32_t(vocab.token_sep()));339    add_kv(LLM_KV_TOKENIZER_PAD_ID,                  uint32_t(vocab.token_pad()));340    // add_kv(LLM_KV_TOKENIZER_CLS_ID,                  uint32_t(vocab.token_bos())); // deprecated341    // add_kv(LLM_KV_TOKENIZER_MASK_ID,                 ???);342    add_kv(LLM_KV_TOKENIZER_ADD_BOS,                 vocab.get_add_bos());343    add_kv(LLM_KV_TOKENIZER_ADD_EOS,                 vocab.get_add_eos());344    add_kv(LLM_KV_TOKENIZER_ADD_SEP,                 vocab.get_add_sep());345    add_kv(LLM_KV_TOKENIZER_ADD_PREFIX,              vocab.get_add_space_prefix());346    add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,         vocab.get_remove_extra_whitespaces());347    add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,    vocab.get_precompiled_charsmap());348    // add_kv(LLM_KV_TOKENIZER_HF_JSON,                 ???);349    // add_kv(LLM_KV_TOKENIZER_RWKV,                    ???);350    add_kv(LLM_KV_TOKENIZER_FIM_PRE_ID,              uint32_t(vocab.token_fim_pre()));351    add_kv(LLM_KV_TOKENIZER_FIM_SUF_ID,              uint32_t(vocab.token_fim_suf()));352    add_kv(LLM_KV_TOKENIZER_FIM_MID_ID,              uint32_t(vocab.token_fim_mid()));353    add_kv(LLM_KV_TOKENIZER_FIM_PAD_ID,              uint32_t(vocab.token_fim_pad()));354    add_kv(LLM_KV_TOKENIZER_FIM_REP_ID,              uint32_t(vocab.token_fim_rep()));355    add_kv(LLM_KV_TOKENIZER_FIM_SEP_ID,              uint32_t(vocab.token_fim_sep()));356 357    // TODO: implement LoRA support358    // add_kv(LLM_KV_ADAPTER_TYPE,                      ???);359    // add_kv(LLM_KV_ADAPTER_LORA_ALPHA,                ???);360    // add_kv(LLM_KV_ADAPTER_LORA_TASK_NAME,            ???);361    // add_kv(LLM_KV_ADAPTER_LORA_PROMPT_PREFIX,        ???);362    // add_kv(LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS,   ???);363 364    add_kv(LLM_KV_POSNET_EMBEDDING_LENGTH,           hparams.posnet.n_embd);365    add_kv(LLM_KV_POSNET_BLOCK_COUNT,                hparams.posnet.n_layer);366 367    add_kv(LLM_KV_CONVNEXT_EMBEDDING_LENGTH,         hparams.convnext.n_embd);368    add_kv(LLM_KV_CONVNEXT_BLOCK_COUNT,              hparams.convnext.n_layer);369 370    add_kv(LLM_KV_CLASSIFIER_OUTPUT_LABELS,          model->classifier_labels);371 372    add_kv(LLM_KV_SHORTCONV_L_CACHE,                 hparams.n_shortconv_l_cache);373 374    add_kv(LLM_KV_XIELU_ALPHA_N,                     hparams.xielu_alpha_n);375    add_kv(LLM_KV_XIELU_ALPHA_P,                     hparams.xielu_alpha_p);376    add_kv(LLM_KV_XIELU_BETA,                        hparams.xielu_beta);377    add_kv(LLM_KV_XIELU_EPS,                         hparams.xielu_eps);378 379    // deprecated380    // add_kv(LLM_KV_TOKENIZER_PREFIX_ID,               ???);381    // add_kv(LLM_KV_TOKENIZER_SUFFIX_ID,               ???);382    // add_kv(LLM_KV_TOKENIZER_MIDDLE_ID,               ???);383 384    add_kv(LLM_KV_DENSE_2_FEAT_IN,                   hparams.dense_2_feat_in);385    add_kv(LLM_KV_DENSE_2_FEAT_OUT,                  hparams.dense_2_feat_out);386    add_kv(LLM_KV_DENSE_3_FEAT_IN,                   hparams.dense_3_feat_in);387    add_kv(LLM_KV_DENSE_3_FEAT_OUT,                  hparams.dense_3_feat_out);388}389 390void llama_model_saver::add_tensors_from_model() {391    if (model->output != nullptr &&392            std::string(model->output->name) != std::string(model->tok_embd->name)) {393        add_tensor(model->tok_embd); // some models use the same tensor for tok_embd and output394    }395    add_tensor(model->type_embd);396    add_tensor(model->pos_embd);397    add_tensor(model->tok_norm);398    add_tensor(model->tok_norm_b);399    add_tensor(model->output_norm);400    add_tensor(model->output_norm_b);401    add_tensor(model->output);402    add_tensor(model->output_b);403    add_tensor(model->output_norm_enc);404    add_tensor(model->output_s);405    add_tensor(model->output_in_s);406    add_tensor(model->cls);407    add_tensor(model->cls_b);408    add_tensor(model->cls_out);409    add_tensor(model->cls_out_b);410    add_tensor(model->cls_norm);411 412    for (const struct llama_layer & layer : model->layers) {413        for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {414            add_tensor(reinterpret_cast<const struct ggml_tensor * const *>(&layer)[i]);415        }416    }417}418 419void llama_model_saver::save(const std::string & path_model) {420    gguf_write_to_file(gguf_ctx, path_model.c_str(), false);421}422 423void llama_model_saver::save(FILE * file) {424    gguf_write_to_file_ptr(gguf_ctx, file, false);425}426