1#include "models.h"
  2
  3llm_build_gpt2::llm_build_gpt2(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    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
  6
  7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
  8
  9    ggml_tensor * cur;
 10    ggml_tensor * pos;
 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    pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
 21    cb(pos, "pos_embd", -1);
 22
 23    inpL = ggml_add(ctx0, inpL, pos);
 24    cb(inpL, "inpL", -1);
 25
 26    ggml_tensor * inp_out_ids = build_inp_out_ids();
 27
 28    for (int il = 0; il < n_layer; ++il) {
 29        cur = build_norm(inpL,
 30                model.layers[il].attn_norm,
 31                model.layers[il].attn_norm_b,
 32                LLM_NORM, il);
 33        cb(cur, "attn_norm", il);
 34
 35        // self-attention
 36        {
 37            cur = build_lora_mm(model.layers[il].wqkv, cur);
 38            cb(cur, "wqkv", il);
 39
 40            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
 41            cb(cur, "bqkv", il);
 42
 43            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
 44            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
 45            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 46
 47            cb(Qcur, "Qcur", il);
 48            cb(Kcur, "Kcur", il);
 49            cb(Vcur, "Vcur", il);
 50
 51            cur = build_attn(inp_attn,
 52                    model.layers[il].wo, model.layers[il].bo,
 53                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 54        }
 55
 56        if (il == n_layer - 1 && inp_out_ids) {
 57            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
 58            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
 59        }
 60
 61        // add the input
 62        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
 63        cb(ffn_inp, "ffn_inp", il);
 64
 65        // FF
 66        {
 67            cur = build_norm(ffn_inp,
 68                    model.layers[il].ffn_norm,
 69                    model.layers[il].ffn_norm_b,
 70                    LLM_NORM, il);
 71            cb(cur, "ffn_norm", il);
 72
 73            cur = build_ffn(cur,
 74                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
 75                    NULL,                      NULL,                        NULL,
 76                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
 77                    NULL,
 78                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
 79            cb(cur, "ffn_out", il);
 80        }
 81
 82        cur = ggml_add(ctx0, cur, ffn_inp);
 83
 84        cur = build_cvec(cur, il);
 85        cb(cur, "l_out", il);
 86
 87        // input for next layer
 88        inpL = cur;
 89    }
 90
 91    cur = build_norm(inpL,
 92            model.output_norm,
 93            model.output_norm_b,
 94            LLM_NORM, -1);
 95
 96    cb(cur, "result_norm", -1);
 97    res->t_embd = cur;
 98
 99    cur = build_lora_mm(model.output, cur);
100
101    cb(cur, "result_output", -1);
102    res->t_logits = cur;
103
104    ggml_build_forward_expand(gf, cur);
105}