Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_gemma2::load_arch_hparams(llama_model_loader & ml) {4    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;5    hparams.n_swa = 4096; // default value of gemma 26    load_swa_pattern(ml, 2);7    hparams.attn_soft_cap = true;8    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;9    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;10 11    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);12    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);13    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);14    ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING,      hparams.f_attn_logit_softcapping, false);15    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);16 17    switch (hparams.n_layer()) {18        case 26: type = LLM_TYPE_2B; break;19        case 42: type = LLM_TYPE_9B; break;20        case 46: type = LLM_TYPE_27B; break;21        default: type = LLM_TYPE_UNKNOWN;22   }23 24    // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L17325    hparams.f_attention_scale = type == LLM_TYPE_27B26        ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))27        : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));28}29 30void llama_model_gemma2::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_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading38 39    for (int i = 0; i < n_layer; ++i) {40        auto & layer = layers[i];41 42        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);43 44        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);45        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);46        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);47 48        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);49        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);50        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);51        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);52        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);53    }54}55 56std::unique_ptr<llm_graph_context> llama_model_gemma2::build_arch_graph(const llm_graph_params & params) const {57    return std::make_unique<graph>(*this, params);58}59 60llama_model_gemma2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {61    const int64_t n_embd_head = hparams.n_embd_head_k();62 63    ggml_tensor * cur;64    ggml_tensor * inpL;65 66    inpL = build_inp_embd(model.tok_embd);67 68    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));69    cb(inpL, "inp_scaled", -1);70 71    // inp_pos - contains the positions72    ggml_tensor * inp_pos = build_inp_pos();73 74    auto * inp_attn = build_attn_inp_kv_iswa();75 76    ggml_tensor * inp_out_ids = build_inp_out_ids();77 78    for (int il = 0; il < n_layer; ++il) {79        const float freq_base_l  = model.get_rope_freq_base (cparams, il);80        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);81 82        // norm83        cur = build_norm(inpL,84                model.layers[il].attn_norm, NULL,85                LLM_NORM_RMS, il);86        cb(cur, "attn_norm", il);87 88        // self-attention89        {90            // compute Q and K and RoPE them91            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,92                    n_embd_head, n_head, n_head_kv, il);93 94            Qcur = ggml_rope_ext(95                    ctx0, Qcur, inp_pos, nullptr,96                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,97                    ext_factor, attn_factor, beta_fast, beta_slow);98 99            Kcur = ggml_rope_ext(100                    ctx0, Kcur, inp_pos, nullptr,101                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,102                    ext_factor, attn_factor, beta_fast, beta_slow);103 104            cb(Qcur, "Qcur", il);105            cb(Kcur, "Kcur", il);106            cb(Vcur, "Vcur", il);107 108            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);109 110            cur = build_attn(inp_attn,111                    model.layers[il].wo, NULL, model.layers[il].wo_s,112                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);113        }114        if (il == n_layer - 1 && inp_out_ids) {115            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);116            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);117        }118        cur = build_norm(cur,119                model.layers[il].attn_post_norm, NULL,120                LLM_NORM_RMS, il);121        cb(cur, "attn_post_norm", il);122 123        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);124        cb(sa_out, "sa_out", il);125 126        cur = build_norm(sa_out,127                model.layers[il].ffn_norm, NULL,128                LLM_NORM_RMS, il);129        cb(cur, "ffn_norm", il);130 131        // feed-forward network132        {133            cur = build_ffn(cur,134                    model.layers[il].ffn_up,   NULL, NULL,135                    model.layers[il].ffn_gate, NULL, NULL,136                    model.layers[il].ffn_down, NULL, NULL,137                    NULL,138                    LLM_FFN_GELU, LLM_FFN_PAR, il);139            cb(cur, "ffn_out", il);140        }141        cur = build_norm(cur,142                model.layers[il].ffn_post_norm, NULL,143                LLM_NORM_RMS, -1);144        cb(cur, "ffn_post_norm", -1);145 146        cur = ggml_add(ctx0, cur, sa_out);147 148        cur = build_cvec(cur, il);149        cb(cur, "l_out", il);150 151        // input for next layer152        inpL = cur;153    }154    cur = inpL;155 156    cur = build_norm(cur,157            model.output_norm, NULL,158            LLM_NORM_RMS, -1);159 160    cb(cur, "result_norm", -1);161    res->t_embd = cur;162 163    // lm_head164    cur = build_lora_mm(model.output, cur, model.output_s);165 166    // final logit soft-capping167    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);168    cur = ggml_tanh(ctx0, cur);169    cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);170 171    cb(cur, "result_output", -1);172    res->t_logits = cur;173 174    ggml_build_forward_expand(gf, cur);175}176