echodict/llama.cpp

1#include "models.h"2 3 4 5llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :6    llm_graph_context(params) {7    const int64_t n_embd_head = hparams.n_embd_head_v();8 9    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());10 11    const float f_logit_scale = hparams.f_logit_scale;12 13    ggml_tensor * cur;14    ggml_tensor * inpL;15 16    inpL = build_inp_embd(model.tok_embd);17 18    // inp_pos - contains the positions19    ggml_tensor * inp_pos = build_inp_pos();20 21    auto * inp_attn = build_attn_inp_kv();22 23    ggml_tensor * inp_out_ids = build_inp_out_ids();24 25    for (int il = 0; il < n_layer; ++il) {26        // norm27        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);28        cb(cur, "attn_norm", il);29 30        ggml_tensor * ffn_inp = cur;31 32        // self-attention33        {34            // compute Q and K and RoPE them35            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,36                    n_embd_head, n_head, n_head_kv, il);37 38            if (model.layers[il].attn_q_norm) {39                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM, il);40                cb(Qcur, "Qcur", il);41            }42            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,43                                 ext_factor, attn_factor, beta_fast, beta_slow);44 45            if (model.layers[il].attn_k_norm) {46                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM, il);47                cb(Kcur, "Kcur", il);48            }49            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,50                                 ext_factor, attn_factor, beta_fast, beta_slow);51 52            cb(Qcur, "Qcur", il);53            cb(Kcur, "Kcur", il);54            cb(Vcur, "Vcur", il);55 56            cur = build_attn(inp_attn,57                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,58                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);59        }60        if (il == n_layer - 1 && inp_out_ids) {61            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);62            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);63            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);64        }65        ggml_tensor * attn_out = cur;66 67        // feed-forward network68        {69            cur = build_ffn(ffn_inp,70                    model.layers[il].ffn_up, NULL, NULL,71                    model.layers[il].ffn_gate, NULL, NULL,72                    model.layers[il].ffn_down, NULL, NULL,73                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);74            cb(cur, "ffn_out", il);75        }76        // add together residual + FFN + self-attention77        cur = ggml_add(ctx0, cur, inpL);78        cur = ggml_add(ctx0, cur, attn_out);79 80        cur = build_cvec(cur, il);81        cb(cur, "l_out", il);82 83        // input for next layer84        inpL = cur;85    }86    cur = inpL;87 88    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);89 90    cb(cur, "result_norm", -1);91    res->t_embd = cur;92 93    // lm_head94    cur = build_lora_mm(model.output, cur);95 96    if (f_logit_scale) {97        cur = ggml_scale(ctx0, cur, f_logit_scale);98    }99    cb(cur, "result_output", -1);100    res->t_logits = cur;101 102    ggml_build_forward_expand(gf, cur);103}104