Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_falcon::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 32: type = LLM_TYPE_7B; break;8        case 60: type = LLM_TYPE_40B; break;9        default: type = LLM_TYPE_UNKNOWN;10    }11}12 13void llama_model_falcon::load_arch_tensors(llama_model_loader &) {14    LLAMA_LOAD_LOCALS;15 16    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);17 18    // output19    {20        output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);21        output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);22 23        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);24        if (!output) {25            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // needs to be on GPU26        }27    }28 29    for (int i = 0; i < n_layer; ++i) {30        auto & layer = layers[i];31 32        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);33        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);34 35        layer.attn_norm_2   = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);36        layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);37 38        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);39        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);40 41        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);42        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);43    }44}45 46std::unique_ptr<llm_graph_context> llama_model_falcon::build_arch_graph(const llm_graph_params & params) const {47    return std::make_unique<graph>(*this, params);48}49 50llama_model_falcon::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {51    const int64_t n_embd_head = hparams.n_embd_head_v();52 53    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());54    GGML_ASSERT(n_embd_head == n_rot);55 56    ggml_tensor * cur;57    ggml_tensor * inpL;58 59    inpL = build_inp_embd(model.tok_embd);60 61    // inp_pos - contains the positions62    ggml_tensor * inp_pos = build_inp_pos();63 64    auto * inp_attn = build_attn_inp_kv();65 66    ggml_tensor * inp_out_ids = build_inp_out_ids();67 68    for (int il = 0; il < n_layer; ++il) {69        ggml_tensor * attn_norm;70 71        attn_norm = build_norm(inpL,72                model.layers[il].attn_norm,73                model.layers[il].attn_norm_b,74                LLM_NORM, il);75        cb(attn_norm, "attn_norm", il);76 77        // self-attention78        {79            if (model.layers[il].attn_norm_2) {80                // Falcon-40B81                cur = build_norm(inpL,82                        model.layers[il].attn_norm_2,83                        model.layers[il].attn_norm_2_b,84                        LLM_NORM, il);85                cb(cur, "attn_norm_2", il);86            } else {87                cur = attn_norm;88            }89 90            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,91                    n_embd_head, n_head, n_head_kv, il);92 93            // using mode = 2 for neox mode94            Qcur = ggml_rope_ext(95                    ctx0, Qcur, inp_pos, nullptr,96                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,97                    ext_factor, attn_factor, beta_fast, beta_slow98                    );99 100            Kcur = ggml_rope_ext(101                    ctx0, Kcur, inp_pos, nullptr,102                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,103                    ext_factor, attn_factor, beta_fast, beta_slow104                    );105 106            cb(Qcur, "Qcur", il);107            cb(Kcur, "Kcur", il);108            cb(Vcur, "Vcur", il);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/sqrtf(float(n_embd_head)), il);113        }114 115        if (il == n_layer - 1 && inp_out_ids) {116            cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);117            inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);118            attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);119        }120 121        ggml_tensor * ffn_inp = cur;122 123        // feed forward124        {125            cur = build_ffn(attn_norm, // !! use the attn norm, not the result126                    model.layers[il].ffn_up,   NULL, NULL,127                    NULL,                      NULL, NULL,128                    model.layers[il].ffn_down, NULL, NULL,129                    NULL,130                    LLM_FFN_GELU, LLM_FFN_SEQ, il);131            cb(cur, "ffn_out", il);132        }133 134        cur = ggml_add(ctx0, cur, ffn_inp);135        cur = ggml_add(ctx0, cur, inpL);136 137        cur = build_cvec(cur, il);138        cb(cur, "l_out", il);139 140        // input for next layer141        inpL = cur;142    }143 144    cur = inpL;145 146    // norm147    cur = build_norm(cur,148            model.output_norm,149            model.output_norm_b,150            LLM_NORM, -1);151 152    cb(cur, "result_norm", -1);153    res->t_embd = cur;154 155    cur = build_lora_mm(model.output, cur, model.output_s);156 157    cb(cur, "result_output", -1);158    res->t_logits = cur;159 160    ggml_build_forward_expand(gf, cur);161}162