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(tok_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(cur: inpL,
20 mw: model.layers[il].attn_norm,
21 mb: model.layers[il].attn_norm_b,
22 type: LLM_NORM, il);
23 cb(cur, name: "attn_norm", il);
24
25 // self-attention
26 {
27 cur = build_lora_mm(w: model.layers[il].wqkv, cur);
28 cb(cur, name: "wqkv", il);
29
30 cur = ggml_add(ctx: ctx0, a: cur, b: model.layers[il].bqkv);
31 cb(cur, name: "bqkv", il);
32
33 ggml_tensor * Qcur = ggml_view_3d(ctx: ctx0, a: cur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens, nb1: n_embd_head*sizeof(float), nb2: cur->nb[1], offset: 0*cur->nb[0]*(n_embd));
34 ggml_tensor * Kcur = ggml_view_3d(ctx: ctx0, a: cur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens, nb1: n_embd_head*sizeof(float), nb2: cur->nb[1], offset: 1*cur->nb[0]*(n_embd));
35 ggml_tensor * Vcur = ggml_view_3d(ctx: ctx0, a: cur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens, nb1: n_embd_head*sizeof(float), nb2: cur->nb[1], offset: 1*cur->nb[0]*(n_embd + n_embd_gqa));
36
37 cb(cur: Qcur, name: "Qcur", il);
38 cb(cur: Kcur, name: "Kcur", il);
39 cb(cur: Vcur, name: "Vcur", il);
40
41 cur = build_attn(inp: inp_attn,
42 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
43 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale: 1.0f/float(n_embd_head), il);
44 }
45 if (il == n_layer - 1 && inp_out_ids) {
46 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
47 inpL = ggml_get_rows(ctx: ctx0, a: inpL, b: inp_out_ids);
48 }
49 // add the input
50 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpL);
51 cb(cur: ffn_inp, name: "ffn_inp", il);
52
53 // FF
54 {
55 cur = build_norm(cur: ffn_inp,
56 mw: model.layers[il].ffn_norm,
57 mb: model.layers[il].ffn_norm_b,
58 type: LLM_NORM, il);
59 cb(cur, name: "ffn_norm", il);
60
61 cur = build_ffn(cur,
62 up: model.layers[il].ffn_up, up_b: model.layers[il].ffn_up_b, NULL,
63 gate: model.layers[il].ffn_gate, gate_b: model.layers[il].ffn_gate_b, NULL,
64 down: model.layers[il].ffn_down, down_b: model.layers[il].ffn_down_b, NULL,
65 NULL,
66 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
67 cb(cur, name: "ffn_out", il);
68 }
69 inpL = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
70 cb(cur: inpL, name: "l_out", il);
71 }
72 cur = build_norm(cur: inpL,
73 mw: model.output_norm,
74 mb: model.output_norm_b,
75 type: LLM_NORM, il: -1);
76
77 cb(cur, name: "result_norm", il: -1);
78 res->t_embd = cur;
79
80 cur = build_lora_mm(w: model.output, cur);
81
82 cb(cur, name: "result_output", il: -1);
83 res->t_logits = cur;
84
85 ggml_build_forward_expand(cgraph: gf, tensor: cur);
86}
87