Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_cohere2::load_arch_hparams(llama_model_loader & ml) {4    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;5    load_swa_pattern(ml, 4);6 7    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;8    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;9 10    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,       hparams.rope_freq_base_train_swa, false);11    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);12    ml.get_key(LLM_KV_LOGIT_SCALE,              hparams.f_logit_scale);13    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);14 15    switch (hparams.n_layer()) {16        case 32: type = LLM_TYPE_8B; break;17        default: type = LLM_TYPE_UNKNOWN;18    }19}20 21void llama_model_cohere2::load_arch_tensors(llama_model_loader &) {22    LLAMA_LOAD_LOCALS;23 24    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);25 26    // output27    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);28    // init output from the input tok embed29    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab },30                                      TENSOR_DUPLICATED);31 32    for (int i = 0; i < n_layer; ++i) {33        auto & layer = layers[i];34 35        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);36 37        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);38        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }, 0);39 40        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, 0);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_cohere2::build_arch_graph(const llm_graph_params & params) const {47    return std::make_unique<graph>(*this, params);48}49 50llama_model_cohere2::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 55    const float f_logit_scale = hparams.f_logit_scale;56 57    ggml_tensor * cur;58    ggml_tensor * inpL;59 60    inpL = build_inp_embd(model.tok_embd);61 62    // inp_pos - contains the positions63    ggml_tensor * inp_pos = build_inp_pos();64 65    auto * inp_attn = build_attn_inp_kv_iswa();66 67    ggml_tensor * inp_out_ids = build_inp_out_ids();68 69    for (int il = 0; il < n_layer; ++il) {70        const bool is_swa = hparams.is_swa(il);71        // UNUSED:72        // const float freq_base_l  = model.get_rope_freq_base (cparams, il);73        // const float freq_scale_l = model.get_rope_freq_scale(cparams, il);74 75        // norm76        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);77        cb(cur, "attn_norm", il);78        ggml_tensor * ffn_inp = cur;79 80        // self-attention81        {82            // rope freq factors for 128k context83            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);84 85            // compute Q and K and RoPE them86            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,87                    n_embd_head, n_head, n_head_kv, il);88 89            if (is_swa) {90                Qcur = ggml_rope_ext(91                        ctx0, Qcur, inp_pos, rope_factors,92                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,93                        ext_factor, attn_factor, beta_fast, beta_slow94                        );95 96                Kcur = ggml_rope_ext(97                        ctx0, Kcur, inp_pos, rope_factors,98                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,99                        ext_factor, attn_factor, beta_fast, beta_slow100                        );101            }102 103            cb(Qcur, "Qcur", il);104            cb(Kcur, "Kcur", il);105            cb(Vcur, "Vcur", il);106 107            cur = build_attn(inp_attn,108                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,109                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);110        }111 112        if (il == n_layer - 1 && inp_out_ids) {113            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);114            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);115            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);116        }117 118        ggml_tensor * attn_out = cur;119 120        // feed-forward network121        {122            cur = build_ffn(ffn_inp,123                    model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,124                    model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,125                    model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,126                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);127            cb(cur, "ffn_out", il);128        }129 130        // add together residual + FFN + self-attention131        cur = ggml_add(ctx0, cur, inpL);132        cur = ggml_add(ctx0, cur, attn_out);133 134        cur = build_cvec(cur, il);135        cb(cur, "l_out", il);136 137        // input for next layer138        inpL = cur;139    }140 141    cur = inpL;142 143    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);144 145    cb(cur, "result_norm", -1);146    res->t_embd = cur;147 148    // lm_head149    cur = build_lora_mm(model.output, cur, model.output_s);150 151    if (f_logit_scale) {152        cur = ggml_scale(ctx0, cur, f_logit_scale);153    }154 155    cb(cur, "result_output", -1);156    res->t_logits = cur;157 158    ggml_build_forward_expand(gf, cur);159}160