1#include "models.h"
2
3llm_build_orion::llm_build_orion(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
6 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
7 GGML_ASSERT(n_embd_head == hparams.n_rot);
8
9 ggml_tensor * cur;
10 ggml_tensor * inpL;
11
12 inpL = build_inp_embd(tok_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 * inp_out_ids = build_inp_out_ids();
20
21 for (int il = 0; il < n_layer; ++il) {
22 ggml_tensor * inpSA = inpL;
23
24 // norm
25 cur = build_norm(cur: inpL,
26 mw: model.layers[il].attn_norm, mb: model.layers[il].attn_norm_b,
27 type: LLM_NORM, il);
28 cb(cur, name: "attn_norm", il);
29
30 // self-attention
31 {
32 // compute Q and K and RoPE them
33 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
34 cb(cur: Qcur, name: "Qcur", il);
35 // if (model.layers[il].bq) {
36 // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
37 // cb(Qcur, "Qcur", il);
38 // }
39
40 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
41 cb(cur: Kcur, name: "Kcur", il);
42 // if (model.layers[il].bk) {
43 // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
44 // cb(Kcur, "Kcur", il);
45 // }
46
47 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
48 cb(cur: Vcur, name: "Vcur", il);
49 // if (model.layers[il].bv) {
50 // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
51 // cb(Vcur, "Vcur", il);
52 // }
53
54 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
55 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
56 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
57
58 Qcur = ggml_rope_ext(
59 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
60 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
61 ext_factor, attn_factor, beta_fast, beta_slow
62 );
63
64 Kcur = ggml_rope_ext(
65 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
66 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
67 ext_factor, attn_factor, beta_fast, beta_slow
68 );
69
70 cb(cur: Qcur, name: "Qcur", il);
71 cb(cur: Kcur, name: "Kcur", il);
72 cb(cur: Vcur, name: "Vcur", il);
73
74 cur = build_attn(inp: inp_attn,
75 wo: model.layers[il].wo, NULL,
76 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);
77 }
78 if (il == n_layer - 1 && inp_out_ids) {
79 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
80 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
81 }
82 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
83 cb(cur: ffn_inp, name: "ffn_inp", il);
84
85 // feed-forward network
86 cur = build_norm(cur: ffn_inp,
87 mw: model.layers[il].ffn_norm, mb: model.layers[il].ffn_norm_b,
88 type: LLM_NORM, il);
89 cb(cur, name: "ffn_norm", il);
90
91 cur = build_ffn(cur,
92 up: model.layers[il].ffn_up, NULL, NULL,
93 gate: model.layers[il].ffn_gate, NULL, NULL,
94 down: model.layers[il].ffn_down, NULL, NULL,
95 NULL,
96 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
97 cb(cur, name: "ffn_out", il);
98
99 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
100
101 cur = build_cvec(cur, il);
102 cb(cur, name: "l_out", il);
103
104 // input for next layer
105 inpL = cur;
106 }
107 cur = inpL;
108
109 cur = build_norm(cur,
110 mw: model.output_norm, mb: model.output_norm_b,
111 type: LLM_NORM, il: -1);
112
113 cb(cur, name: "result_norm", il: -1);
114 res->t_embd = cur;
115
116 // lm_head
117 cur = build_lora_mm(w: model.output, cur);
118
119 cb(cur, name: "result_output", il: -1);
120 res->t_logits = cur;
121
122 ggml_build_forward_expand(cgraph: gf, tensor: cur);
123}
124