Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_nomic_bert_moe::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);5    ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS,         hparams.moe_every_n_layers, 0);6 7    switch (hparams.n_layer()) {8        case 12:9            type = LLM_TYPE_475M; break;10        default: type = LLM_TYPE_UNKNOWN;11    }12}13 14void llama_model_nomic_bert_moe::load_arch_tensors(llama_model_loader &) {15    LLAMA_LOAD_LOCALS;16 17    if (n_token_types == 0) {18        throw std::runtime_error(arch_name() + " model needs to define token type count");19    }20    tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);21    type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);22 23    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);24    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);25 26    for (int i = 0; i < n_layer; ++i) {27        auto & layer = layers[i];28 29        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);30 31        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);32        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);33 34        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);35        layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);36 37        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {38            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);39            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);40            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);41        } else {42            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);43            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);44            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);45            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);46        }47 48        layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);49        layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);50    }51}52 53std::unique_ptr<llm_graph_context> llama_model_nomic_bert_moe::build_arch_graph(const llm_graph_params & params) const {54    return std::make_unique<graph>(*this, params);55}56 57