1#include "models.h"
  2
  3llm_build_starcoder::llm_build_starcoder(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 * inpL;
 11
 12    inpL = build_inp_embd(model.tok_embd);
 13
 14    // inp_pos - contains the positions
 15    ggml_tensor * inp_pos = build_inp_pos();
 16
 17    auto * inp_attn = build_attn_inp_kv();
 18
 19    ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
 20    cb(pos, "pos_embd", -1);
 21
 22    inpL = ggml_add(ctx0, inpL, pos);
 23    cb(inpL, "inpL", -1);
 24
 25    ggml_tensor * inp_out_ids = build_inp_out_ids();
 26
 27    for (int il = 0; il < n_layer; ++il) {
 28        cur = build_norm(inpL,
 29                model.layers[il].attn_norm,
 30                model.layers[il].attn_norm_b,
 31                LLM_NORM, il);
 32        cb(cur, "attn_norm", il);
 33
 34        // self-attention
 35        {
 36            cur = build_lora_mm(model.layers[il].wqkv, cur);
 37            cb(cur, "wqkv", il);
 38
 39            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
 40            cb(cur, "bqkv", il);
 41
 42            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));
 43            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));
 44            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));
 45
 46            cb(Qcur, "Qcur", il);
 47            cb(Kcur, "Kcur", il);
 48            cb(Vcur, "Vcur", il);
 49
 50            cur = build_attn(inp_attn,
 51                    model.layers[il].wo, model.layers[il].bo,
 52                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 53        }
 54        if (il == n_layer - 1 && inp_out_ids) {
 55            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
 56            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
 57        }
 58        // add the input
 59        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
 60        cb(ffn_inp, "ffn_inp", il);
 61
 62        // FF
 63        {
 64            cur = build_norm(ffn_inp,
 65                    model.layers[il].ffn_norm,
 66                    model.layers[il].ffn_norm_b,
 67                    LLM_NORM, il);
 68            cb(cur, "ffn_norm", il);
 69
 70            cur = build_ffn(cur,
 71                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
 72                    NULL,                      NULL,                        NULL,
 73                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
 74                    NULL,
 75                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
 76            cb(cur, "ffn_out", il);
 77        }
 78        cur = ggml_add(ctx0, cur, ffn_inp);
 79
 80        cur = build_cvec(cur, il);
 81        cb(cur, "l_out", il);
 82
 83        // input for next layer
 84        inpL = cur;
 85    }
 86    cur = build_norm(inpL,
 87            model.output_norm,
 88            model.output_norm_b,
 89            LLM_NORM, -1);
 90
 91    cb(cur, "result_norm", -1);
 92    res->t_embd = cur;
 93
 94    cur = build_lora_mm(model.output, cur);
 95
 96    cb(cur, "result_output", -1);
 97    res->t_logits = cur;
 98
 99    ggml_build_forward_expand(gf, cur);
100}