1#include "models.h"
2
3
4llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
5 const int64_t n_embd_head = hparams.n_embd_head_v;
6 const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
7
8 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
9 GGML_ASSERT(n_embd_head == hparams.n_rot);
10
11 ggml_tensor * cur;
12 ggml_tensor * inpL;
13
14 inpL = build_inp_embd(tok_embd: model.tok_embd);
15
16 // inp_pos - contains the positions
17 ggml_tensor * inp_pos = build_inp_pos();
18
19 auto * inp_attn = build_attn_inp_kv();
20
21 ggml_tensor * inp_out_ids = build_inp_out_ids();
22
23 for (int il = 0; il < n_layer; ++il) {
24 ggml_tensor * attn_norm;
25
26 attn_norm = build_norm(cur: inpL,
27 mw: model.layers[il].attn_norm,
28 mb: model.layers[il].attn_norm_b,
29 type: LLM_NORM, il);
30 cb(cur: attn_norm, name: "attn_norm", il);
31
32 // self-attention
33 {
34 if (model.layers[il].attn_norm_2) {
35 // Falcon-40B
36 cur = build_norm(cur: inpL,
37 mw: model.layers[il].attn_norm_2,
38 mb: model.layers[il].attn_norm_2_b,
39 type: LLM_NORM, il);
40 cb(cur, name: "attn_norm_2", il);
41 } else {
42 cur = attn_norm;
43 }
44
45 cur = build_lora_mm(w: model.layers[il].wqkv, cur);
46 cb(cur, name: "wqkv", il);
47
48 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*sizeof(float)*(n_embd));
49 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*sizeof(float)*(n_embd));
50 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*sizeof(float)*(n_embd + n_embd_gqa));
51
52 // using mode = 2 for neox mode
53 Qcur = ggml_rope_ext(
54 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
55 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
56 ext_factor, attn_factor, beta_fast, beta_slow
57 );
58
59 Kcur = ggml_rope_ext(
60 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
61 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
62 ext_factor, attn_factor, beta_fast, beta_slow
63 );
64
65 cb(cur: Qcur, name: "Qcur", il);
66 cb(cur: Kcur, name: "Kcur", il);
67 cb(cur: Vcur, name: "Vcur", il);
68
69 cur = build_attn(inp: inp_attn,
70 wo: model.layers[il].wo, NULL,
71 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale: 1.0f/sqrtf(x: float(n_embd_head)), il);
72 }
73
74 if (il == n_layer - 1 && inp_out_ids) {
75 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
76 inpL = ggml_get_rows(ctx: ctx0, a: inpL, b: inp_out_ids);
77 attn_norm = ggml_get_rows(ctx: ctx0, a: attn_norm, b: inp_out_ids);
78 }
79
80 ggml_tensor * ffn_inp = cur;
81
82 // feed forward
83 {
84 cur = build_ffn(cur: attn_norm, // !! use the attn norm, not the result
85 up: model.layers[il].ffn_up, NULL, NULL,
86 NULL, NULL, NULL,
87 down: model.layers[il].ffn_down, NULL, NULL,
88 NULL,
89 type_op: LLM_FFN_GELU, type_gate: LLM_FFN_SEQ, il);
90 cb(cur, name: "ffn_out", il);
91 }
92
93 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
94 cur = ggml_add(ctx: ctx0, a: cur, b: inpL);
95
96 cur = build_cvec(cur, il);
97 cb(cur, name: "l_out", il);
98
99 // input for next layer
100 inpL = cur;
101 }
102
103 cur = inpL;
104
105 // norm
106 cur = build_norm(cur,
107 mw: model.output_norm,
108 mb: model.output_norm_b,
109 type: LLM_NORM, il: -1);
110
111 cb(cur, name: "result_norm", il: -1);
112 res->t_embd = cur;
113
114 cur = build_lora_mm(w: model.output, cur);
115
116 cb(cur, name: "result_output", il: -1);
117 res->t_logits = cur;
118
119 ggml_build_forward_expand(cgraph: gf, tensor: cur);
120}
121