Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_gemma_embedding::load_arch_hparams(llama_model_loader & ml) {4    hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC;5    load_swa_pattern(ml, 6);6 7    hparams.causal_attn = false; // embeddings do not use causal attention8 9    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);10    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);11    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);12 13    //applied only if model converted with --sentence-transformers-dense-modules14    ml.get_key(LLM_KV_DENSE_2_FEAT_IN, hparams.dense_2_feat_in, false);15    ml.get_key(LLM_KV_DENSE_2_FEAT_OUT, hparams.dense_2_feat_out, false);16    ml.get_key(LLM_KV_DENSE_3_FEAT_IN, hparams.dense_3_feat_in, false);17    ml.get_key(LLM_KV_DENSE_3_FEAT_OUT, hparams.dense_3_feat_out, false);18 19    GGML_ASSERT((hparams.dense_2_feat_in == 0 || hparams.dense_2_feat_in == hparams.n_embd) && "dense_2_feat_in must be equal to n_embd");20    GGML_ASSERT((hparams.dense_3_feat_out == 0 || hparams.dense_3_feat_out == hparams.n_embd) && "dense_3_feat_out must be equal to n_embd");21 22    switch (hparams.n_layer()) {23        case 24: type = LLM_TYPE_0_3B; break;24        default: type = LLM_TYPE_UNKNOWN;25    }26    hparams.f_attention_scale = 1.0f / std::sqrt(float(hparams.n_embd_head_k()));27 28}29 30void llama_model_gemma_embedding::load_arch_tensors(llama_model_loader &) {31    LLAMA_LOAD_LOCALS;32 33    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);34 35    // output36    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);37    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);38 39    // if output is NULL, init from the input tok embed40    if (output == NULL) {41        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,   "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);42    }43 44    // Dense linear weights45    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.dense_2_feat_out}, TENSOR_NOT_REQUIRED);46    dense_3_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_3_OUT, "weight"), {hparams.dense_3_feat_in, n_embd}, TENSOR_NOT_REQUIRED);47 48 49    for (int i = 0; i < n_layer; ++i) {50        auto & layer = layers[i];51 52        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);53 54        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);55        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);56 57        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);58        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_k}, 0);59        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head_k}, 0);60 61        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);62        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);63        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);64        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);65        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);66    }67}68 69std::unique_ptr<llm_graph_context> llama_model_gemma_embedding::build_arch_graph(const llm_graph_params & params) const {70    return std::make_unique<graph>(*this, params);71}72 73llama_model_gemma_embedding::graph::graph(const llama_model & model, const llm_graph_params & params) :74    llm_graph_context(params) {75    const int64_t n_embd_head = hparams.n_embd_head_k();76 77    ggml_tensor * cur;78    ggml_tensor * inpL;79 80    inpL = build_inp_embd(model.tok_embd);81 82    // important: do not normalize weights for raw embeddings input (i.e. encoded image embeddings)83    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);84    cb(inpL, "inp_scaled", -1);85 86    // inp_pos - contains the positions87    ggml_tensor * inp_pos = build_inp_pos();88 89    auto * inp_attn = build_attn_inp_no_cache();90 91    ggml_tensor * inp_out_ids = build_inp_out_ids();92 93    for (int il = 0; il < n_layer; ++il) {94        const float freq_base_l  = model.get_rope_freq_base(cparams, il);95        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);96 97        // norm98        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);99        cb(cur, "attn_norm", il);100 101        // self-attention102        {103            // compute Q and K and RoPE them104            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,105                    n_embd_head, n_head, n_head_kv, il);106 107            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);108            cb(Qcur, "Qcur_normed", il);109 110            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,111                                 ext_factor, attn_factor, beta_fast, beta_slow);112 113            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);114            cb(Kcur, "Kcur_normed", il);115 116            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,117                                 ext_factor, attn_factor, beta_fast, beta_slow);118 119            cb(Qcur, "Qcur", il);120            cb(Kcur, "Kcur", il);121            cb(Vcur, "Vcur", il);122 123            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315124            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);125 126            cur =127                build_attn(inp_attn,128                    model.layers[il].wo, NULL, model.layers[il].wo_s,129                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);130        }131 132        if (il == n_layer - 1 && inp_out_ids) {133            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);134            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);135        }136 137        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);138        cb(cur, "attn_post_norm", il);139 140        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);141        cb(sa_out, "sa_out", il);142 143        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);144        cb(cur, "ffn_norm", il);145 146        // feed-forward network147        {148            cur = build_ffn(cur,149                model.layers[il].ffn_up, NULL, NULL,150                model.layers[il].ffn_gate, NULL, NULL,151                model.layers[il].ffn_down, NULL, NULL,152                NULL, LLM_FFN_GELU, LLM_FFN_PAR, il);153            cb(cur, "ffn_out", il);154        }155 156        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);157        cb(cur, "ffn_post_norm", -1);158 159        cur = ggml_add(ctx0, cur, sa_out);160 161        cur = build_cvec(cur, il);162        cb(cur, "l_out", il);163 164        // input for next layer165        inpL = cur;166    }167 168    cur = inpL;169 170    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);171 172    cb(cur, "result_norm", -1);173    res->t_embd = cur;174 175    ggml_build_forward_expand(gf, cur);176}177