1#include "models.h"
  2
  3llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
  4    llm_graph_context(params) {
  5    const int64_t n_embd_head = hparams.n_embd_head_v;
  6
  7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
  8    GGML_ASSERT(n_embd_head == hparams.n_rot);
  9
 10    ggml_tensor * cur;
 11    ggml_tensor * inpL;
 12
 13    inpL = build_inp_embd(model.tok_embd);
 14
 15    // inp_pos - contains the positions
 16    ggml_tensor * inp_pos = build_inp_pos();
 17
 18    auto * inp_attn = build_attn_inp_kv();
 19
 20    ggml_tensor * inp_out_ids = build_inp_out_ids();
 21
 22    for (int il = 0; il < n_layer; ++il) {
 23        ggml_tensor * inpSA = inpL;
 24
 25        // norm
 26        {
 27            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 28            cb(cur, "attn_norm", il);
 29        }
 30        // self-attention
 31        {
 32            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
 33            cb(Qcur, "Qcur", il);
 34            if (model.layers[il].bq) {
 35                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
 36                cb(Qcur, "Qcur", il);
 37            }
 38            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 39            cb(Kcur, "Kcur", il);
 40            if (model.layers[il].bk) {
 41                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
 42                cb(Kcur, "Kcur", il);
 43            }
 44            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 45            cb(Vcur, "Vcur", il);
 46            if (model.layers[il].bv) {
 47                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
 48                cb(Vcur, "Vcur", il);
 49            }
 50            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 51            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 52            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 53
 54            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 55                                 ext_factor, attn_factor, beta_fast, beta_slow);
 56
 57            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 58                                 ext_factor, attn_factor, beta_fast, beta_slow);
 59
 60            cb(Qcur, "Qcur", il);
 61            cb(Kcur, "Kcur", il);
 62            cb(Vcur, "Vcur", il);
 63
 64            cur = build_attn(inp_attn,
 65                    model.layers[il].wo, NULL,
 66                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
 67        }
 68        if (il == n_layer - 1) {
 69            // skip computing output for unused tokens
 70            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
 71            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
 72        }
 73        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
 74        cb(ffn_inp, "ffn_inp", il);
 75
 76        // feed-forward network
 77        {
 78            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
 79            cb(cur, "ffn_norm", il);
 80
 81            cur = build_ffn(cur,
 82                    model.layers[il].ffn_up, NULL, NULL,
 83                    model.layers[il].ffn_gate, NULL, NULL,
 84                    model.layers[il].ffn_down, NULL, NULL,
 85                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
 86            cb(cur, "ffn_out", il);
 87        }
 88        cur = ggml_add(ctx0, cur, ffn_inp);
 89
 90        cur = build_cvec(cur, il);
 91        cb(cur, "l_out", il);
 92
 93        // input for next layer
 94        inpL = cur;
 95    }
 96    cur = inpL;
 97
 98    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 99
100    cb(cur, "result_norm", -1);
101    res->t_embd = cur;
102
103    // lm_head
104    cur = build_lora_mm(model.output, cur);
105
106    cb(cur, "result_output", -1);
107    res->t_logits = cur;
108
109    ggml_build_forward_expand(gf, cur);
110}