1#include "models.h"
2
3
4
5llm_build_mpt::llm_build_mpt(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
6 const int64_t n_embd_head = hparams.n_embd_head_v;
7 const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
8
9 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
10
11 ggml_tensor * cur;
12 ggml_tensor * pos;
13 ggml_tensor * inpL;
14
15 inpL = build_inp_embd(model.tok_embd);
16
17 auto * inp_attn = build_attn_inp_kv();
18
19 if (model.pos_embd) {
20 // inp_pos - contains the positions
21 ggml_tensor * inp_pos = build_inp_pos();
22 pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
23 cb(pos, "pos_embd", -1);
24
25 inpL = ggml_add(ctx0, inpL, pos);
26 cb(inpL, "inpL", -1);
27 }
28
29 ggml_tensor * inp_out_ids = build_inp_out_ids();
30
31 for (int il = 0; il < n_layer; ++il) {
32 ggml_tensor * attn_norm;
33
34 attn_norm = build_norm(inpL, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, il);
35 cb(attn_norm, "attn_norm", il);
36
37 // self-attention
38 {
39 cur = attn_norm;
40
41 cur = build_lora_mm(model.layers[il].wqkv, cur);
42 cb(cur, "wqkv", il);
43
44 if (model.layers[il].bqkv) {
45 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
46 cb(cur, "bqkv", il);
47 }
48
49 if (hparams.f_clamp_kqv > 0.0f) {
50 cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
51 cb(cur, "wqkv_clamped", il);
52 }
53
54 ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float),
55 cur->nb[1], 0 * sizeof(float) * (n_embd));
56 ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
57 cur->nb[1], 1 * sizeof(float) * (n_embd));
58 ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
59 cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
60
61 // Q/K Layernorm
62 if (model.layers[il].attn_q_norm) {
63 Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);
64 Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);
65
66 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);
67
68 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);
69
70 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
71 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
72 }
73
74 cb(Qcur, "Qcur", il);
75 cb(Kcur, "Kcur", il);
76 cb(Vcur, "Vcur", il);
77
78 cur = build_attn(inp_attn,
79 model.layers[il].wo, model.layers[il].bo,
80 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
81 }
82
83 if (il == n_layer - 1 && inp_out_ids) {
84 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
85 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
86 }
87
88 // Add the input
89 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
90 cb(ffn_inp, "ffn_inp", il);
91
92 // feed forward
93 {
94 cur = build_norm(ffn_inp, model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, LLM_NORM, il);
95 cb(cur, "ffn_norm", il);
96 cur = build_ffn(cur,
97 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
98 NULL, NULL, NULL,
99 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
100 model.layers[il].ffn_act, LLM_FFN_GELU, LLM_FFN_SEQ, il);
101 cb(cur, "ffn_out", il);
102 }
103
104 cur = ggml_add(ctx0, cur, ffn_inp);
105
106 cur = build_cvec(cur, il);
107 cb(cur, "l_out", il);
108
109 // input for next layer
110 inpL = cur;
111 }
112
113 cur = inpL;
114
115 cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
116
117 cb(cur, "result_norm", -1);
118 res->t_embd = cur;
119
120 cur = build_lora_mm(model.output, cur);
121
122 cb(cur, "result_output", -1);
123 res->t_logits = cur;
124
125 ggml_build_forward_expand(gf, cur);
126}