echodict/llama.cpp

1#include "models.h"2 3llm_build_refact::llm_build_refact(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {4    const int64_t n_embd_head = hparams.n_embd_head_v();5 6    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());7 8    ggml_tensor * cur;9    ggml_tensor * inpL;10 11    inpL = build_inp_embd(model.tok_embd);12 13    auto * inp_attn = build_attn_inp_kv();14 15    ggml_tensor * inp_out_ids = build_inp_out_ids();16 17    for (int il = 0; il < n_layer; ++il) {18        ggml_tensor * inpSA = inpL;19 20        cur = build_norm(inpL,21                model.layers[il].attn_norm, NULL,22                LLM_NORM_RMS, il);23        cb(cur, "attn_norm", il);24 25        // self-attention26        {27            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,28                    n_embd_head, n_head, n_head_kv, il);29 30            cb(Qcur, "Qcur", il);31            cb(Kcur, "Kcur", il);32            cb(Vcur, "Vcur", il);33 34            cur = build_attn(inp_attn,35                    model.layers[il].wo, NULL, model.layers[il].wo_s,36                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);37        }38        if (il == n_layer - 1 && inp_out_ids) {39            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);40            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);41        }42        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);43        cb(ffn_inp, "ffn_inp", il);44 45        // feed-forward network46        {47            cur = build_norm(ffn_inp,48                    model.layers[il].ffn_norm, NULL,49                    LLM_NORM_RMS, il);50            cb(cur, "ffn_norm", il);51 52            cur = build_ffn(cur,53                    model.layers[il].ffn_up,   NULL, NULL,54                    model.layers[il].ffn_gate, NULL, NULL,55                    model.layers[il].ffn_down, NULL, NULL,56                    NULL,57                    LLM_FFN_SILU, LLM_FFN_PAR, il);58            cb(cur, "ffn_out", il);59        }60        cur = ggml_add(ctx0, cur, ffn_inp);61 62        cur = build_cvec(cur, il);63        cb(cur, "l_out", il);64 65        // input for next layer66        inpL = cur;67    }68    cur = inpL;69 70    cur = build_norm(cur,71            model.output_norm, NULL,72            LLM_NORM_RMS, -1);73 74    cb(cur, "result_norm", -1);75    res->t_embd = cur;76 77    // lm_head78    cur = build_lora_mm(model.output, cur);79 80    cb(cur, "result_output", -1);81    res->t_logits = cur;82 83    ggml_build_forward_expand(gf, cur);84}85