Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_bert::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);5 6    switch (hparams.n_layer()) {7        case 3:8            type = LLM_TYPE_17M; break; // bge-micro9        case 6:10            type = LLM_TYPE_22M; break; // MiniLM-L611        case 12:12            switch (hparams.n_embd) {13                case 384: type = LLM_TYPE_33M; break; // MiniLM-L12, bge-small14                case 768: type = LLM_TYPE_109M; break; // bge-base15                default: type = LLM_TYPE_UNKNOWN;16            } break;17        case 24:18            type = LLM_TYPE_335M; break; // bge-large19        default: type = LLM_TYPE_UNKNOWN;20    }21}22 23void llama_model_bert::load_arch_tensors(llama_model_loader &) {24    LLAMA_LOAD_LOCALS;25 26    if (n_token_types == 0) {27        throw std::runtime_error(arch_name() + " model needs to define token type count");28    }29    tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);30    type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);31 32    pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);33 34    cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);35    cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);36 37    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);38    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);39 40    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);41    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);42 43    for (int i = 0; i < n_layer; ++i) {44        auto & layer = layers[i];45 46        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);47 48        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);49        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);50 51        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);52        layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);53 54        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);55        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);56        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);57        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);58 59        layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);60        layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);61    }62}63 64std::unique_ptr<llm_graph_context> llama_model_bert::build_arch_graph(const llm_graph_params & params) const {65    return std::make_unique<graph>(*this, params);66}67 68llama_model_bert::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {69    const int64_t n_embd_head = hparams.n_embd_head_v();70 71    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());72 73    ggml_tensor * cur;74    ggml_tensor * inpL;75    ggml_tensor * inp_pos = nullptr;76 77    if (model.arch != LLM_ARCH_JINA_BERT_V2) {78        inp_pos = build_inp_pos();79    }80 81    // construct input embeddings (token, type, position)82    inpL = build_inp_embd(model.tok_embd);83 84    // token types are hardcoded to zero ("Sentence A")85    if (model.type_embd) {86        ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);87        inpL                    = ggml_add(ctx0, inpL, type_row0);88    }89    if (model.arch == LLM_ARCH_BERT) {90        inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);91    }92    cb(inpL, "inp_embd", -1);93 94    // embed layer norm95    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, 0);96    cb(inpL, "inp_norm", 0);97 98    auto * inp_attn = build_attn_inp_no_cache();99 100    ggml_tensor * inp_out_ids = build_inp_out_ids();101 102    for (int il = 0; il < n_layer; ++il) {103        ggml_tensor * cur = inpL;104 105        {106            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,107                    n_embd_head, n_head, n_head_kv, il);108 109            if (model.layers[il].attn_q_norm) {110                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);111 112                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);113 114                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);115            }116 117            if (model.layers[il].attn_k_norm) {118                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);119 120                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);121 122                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);123            }124 125            // RoPE126            if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||127                model.arch == LLM_ARCH_JINA_BERT_V3) {128                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,129                                     ext_factor, attn_factor, beta_fast, beta_slow);130 131                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,132                                     ext_factor, attn_factor, beta_fast, beta_slow);133            }134 135            cb(Qcur, "Qcur", il);136            cb(Kcur, "Kcur", il);137            cb(Vcur, "Vcur", il);138 139            cur = build_attn(inp_attn,140                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,141                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);142            cb(cur, "kqv_out", il);143        }144 145        if (il == n_layer - 1 && inp_out_ids) {146            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);147            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);148        }149 150        // re-add the layer input151        cur = ggml_add(ctx0, cur, inpL);152 153        // attention layer norm154        cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il);155 156        if (model.layers[il].attn_norm_2 != nullptr) {157            cur = ggml_add(ctx0, cur, inpL);  // re-add the layer input158            cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il);159        }160 161        ggml_tensor * ffn_inp = cur;162        cb(ffn_inp, "ffn_inp", il);163 164        // feed-forward network165        if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {166            // MoE branch167            cur = build_moe_ffn(cur,168                    model.layers[il].ffn_gate_inp,169                    model.layers[il].ffn_up_exps,170                    nullptr,171                    model.layers[il].ffn_down_exps,172                    nullptr,173                    hparams.n_expert, hparams.n_expert_used(),174                    LLM_FFN_GELU, false,175                    hparams.expert_weights_scale,176                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,177                    il);178            cb(cur, "ffn_moe_out", il);179        } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||180                   model.arch == LLM_ARCH_JINA_BERT_V3) {181            cur = build_ffn(cur,182                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,183                    NULL, NULL, NULL,184                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,185                    LLM_FFN_GELU, LLM_FFN_SEQ, il);186            cb(cur, "ffn_out", il);187        } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {188            const bool up_contains_gate = !model.layers[il].ffn_gate && model.layers[il].ffn_up->ne[1] != hparams.n_ff();189            auto type_op = up_contains_gate ? LLM_FFN_GEGLU : LLM_FFN_GELU;190            cur = build_ffn(cur,191                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,192                    model.layers[il].ffn_gate, NULL, NULL,193                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,194                    type_op, LLM_FFN_PAR, il);195            cb(cur, "ffn_out", il);196        } else {197            cur = build_ffn(cur,198                model.layers[il].ffn_up, NULL, NULL,199                model.layers[il].ffn_gate, NULL, NULL,200                model.layers[il].ffn_down, NULL, NULL,201                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);202            cb(cur, "ffn_out", il);203        }204 205        // attentions bypass the intermediate layer206        cur = ggml_add(ctx0, cur, ffn_inp);207 208        // output layer norm209        cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il);210 211        // input for next layer212        inpL = cur;213    }214 215    cur = inpL;216 217    cb(cur, "result_embd", -1);218    res->t_embd = cur;219 220    ggml_build_forward_expand(gf, cur);221}222