1#include "models.h"
 2
 3llm_build_jais::llm_build_jais(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    auto * inp_attn = build_attn_inp_kv();
15
16    ggml_tensor * inp_out_ids = build_inp_out_ids();
17
18    for (int il = 0; il < n_layer; ++il) {
19        cur = build_norm(inpL,
20                model.layers[il].attn_norm,
21                model.layers[il].attn_norm_b,
22                LLM_NORM, il);
23        cb(cur, "attn_norm", il);
24
25        // self-attention
26        {
27            cur = build_lora_mm(model.layers[il].wqkv, cur);
28            cb(cur, "wqkv", il);
29
30            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
31            cb(cur, "bqkv", il);
32
33            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*cur->nb[0]*(n_embd));
34            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*cur->nb[0]*(n_embd));
35            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*cur->nb[0]*(n_embd + n_embd_gqa));
36
37            cb(Qcur, "Qcur", il);
38            cb(Kcur, "Kcur", il);
39            cb(Vcur, "Vcur", il);
40
41            cur = build_attn(inp_attn,
42                    model.layers[il].wo, model.layers[il].bo,
43                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il);
44        }
45        if (il == n_layer - 1 && inp_out_ids) {
46            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
47            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
48        }
49        // add the input
50        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
51        cb(ffn_inp, "ffn_inp", il);
52
53        // FF
54        {
55            cur = build_norm(ffn_inp,
56                    model.layers[il].ffn_norm,
57                    model.layers[il].ffn_norm_b,
58                    LLM_NORM, il);
59            cb(cur, "ffn_norm", il);
60
61            cur = build_ffn(cur,
62                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
63                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
64                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
65                    NULL,
66                    LLM_FFN_SILU, LLM_FFN_PAR, il);
67            cb(cur, "ffn_out", il);
68        }
69        inpL = ggml_add(ctx0, cur, ffn_inp);
70        cb(inpL, "l_out", il);
71    }
72    cur = build_norm(inpL,
73            model.output_norm,
74            model.output_norm_b,
75            LLM_NORM, -1);
76
77    cb(cur, "result_norm", -1);
78    res->t_embd = cur;
79
80    cur = build_lora_mm(model.output, cur);
81
82    cb(cur, "result_output", -1);
83    res->t_logits = cur;
84
85    ggml_build_forward_expand(gf, cur);
86}