1#include "models.h"
2
3
4llm_build_pangu_embedded::llm_build_pangu_embedded(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
7 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
8 GGML_ASSERT(n_embd_head == hparams.n_rot);
9
10 ggml_tensor * cur;
11 ggml_tensor * inpL;
12
13 inpL = build_inp_embd(tok_embd: model.tok_embd);
14
15 // inp_pos - contains the positions
16 ggml_tensor * inp_pos = build_inp_pos();
17
18 auto * inp_attn = build_attn_inp_kv();
19
20 ggml_tensor * inp_out_ids = build_inp_out_ids();
21
22 for (int il = 0; il < n_layer; ++il) {
23 ggml_tensor * inpSA = inpL;
24
25 // norm
26 cur = build_norm(cur: inpL,
27 mw: model.layers[il].attn_norm, NULL,
28 type: LLM_NORM_RMS, il);
29 cb(cur, name: "attn_norm", il);
30
31 // self attention
32 {
33 // compute Q and K and RoPE them
34 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
35 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
36 cb(cur: Qcur, name: "Qcur", il);
37
38 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
39 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
40 cb(cur: Kcur, name: "Kcur", il);
41
42 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
43 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
44 cb(cur: Vcur, name: "Vcur", il);
45
46 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
47 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
48 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
49
50 Qcur = ggml_rope_ext(
51 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
52 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
53 ext_factor, attn_factor, beta_fast, beta_slow
54 );
55
56 Kcur = ggml_rope_ext(ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
57 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
58 ext_factor, attn_factor, beta_fast, beta_slow
59 );
60
61 cb(cur: Qcur, name: "Qcur", il);
62 cb(cur: Kcur, name: "Kcur", il);
63 cb(cur: Vcur, name: "Vcur", il);
64
65 cur = build_attn(inp: inp_attn,
66 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
67 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);
68 }
69
70 if (il == n_layer - 1 && inp_out_ids) {
71 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
72 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
73 }
74
75 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
76 cb(cur: ffn_inp, name: "ffn_inp", il);
77
78 // feed-forward network
79 cur = build_norm(cur: ffn_inp,
80 mw: model.layers[il].ffn_norm, NULL,
81 type: LLM_NORM_RMS, il);
82 cb(cur, name: "ffn_norm", il);
83
84 cur = build_ffn(cur,
85 up: model.layers[il].ffn_up, up_b: model.layers[il].ffn_up_b, NULL,
86 gate: model.layers[il].ffn_gate, gate_b: model.layers[il].ffn_gate_b, NULL,
87 down: model.layers[il].ffn_down, down_b: model.layers[il].ffn_down_b, NULL,
88 NULL,
89 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
90
91 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
92 cb(cur, name: "ffn_out", il);
93
94 cur = build_cvec(cur, il);
95 cb(cur, name: "l_out", il);
96
97 // input for next layer
98 inpL = cur;
99 }
100
101 cur = inpL;
102
103 cur = build_norm(cur,
104 mw: model.output_norm, NULL,
105 type: LLM_NORM_RMS, il: -1);
106
107 cb(cur, name: "result_norm", il: -1);
108 res->t_embd = cur;
109
110 // lm_head
111 cur = build_lora_mm(w: model.output, cur);
112
113 if (model.output_b != nullptr) {
114 cur = ggml_add(ctx: ctx0, a: cur, b: model.output_b);
115 }
116
117 cb(cur, name: "result_output", il: -1);
118 res->t_logits = cur;
119
120 ggml_build_forward_expand(cgraph: gf, tensor: cur);
121}
122