1#include "models.h"
2
3
4
5llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
6 llm_graph_context(params) {
7 //copied from qwen2
8 const int64_t n_embd_head = hparams.n_embd_head_v;
9
10 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
11 GGML_ASSERT(n_embd_head == hparams.n_rot);
12
13 ggml_tensor * cur;
14 ggml_tensor * inpL;
15
16 inpL = build_inp_embd(tok_embd: model.tok_embd);
17
18 // inp_pos - contains the positions
19 ggml_tensor * inp_pos = build_inp_pos();
20
21 auto * inp_attn = build_attn_inp_no_cache();
22
23 ggml_tensor * inp_out_ids = build_inp_out_ids();
24
25 for (int il = 0; il < n_layer; ++il) {
26 ggml_tensor * inpSA = inpL;
27
28 // norm
29 cur = build_norm(cur: inpL, mw: model.layers[il].attn_norm, NULL, type: LLM_NORM_RMS, il);
30 cb(cur, name: "attn_norm", il);
31
32 // self-attention
33 {
34 // compute Q and K and RoPE them
35 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
36 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
37 cb(cur: Qcur, name: "Qcur", il);
38
39 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
40 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
41 cb(cur: Kcur, name: "Kcur", il);
42
43 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
44 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
45 cb(cur: Vcur, name: "Vcur", il);
46
47 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
48 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
49 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
50
51 Qcur = ggml_rope_ext(ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr, n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
52 ext_factor, attn_factor, beta_fast, beta_slow);
53
54 Kcur = ggml_rope_ext(ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr, n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
55 ext_factor, attn_factor, beta_fast, beta_slow);
56
57 cb(cur: Qcur, name: "Qcur", il);
58 cb(cur: Kcur, name: "Kcur", il);
59 cb(cur: Vcur, name: "Vcur", il);
60
61 cur = build_attn(inp: inp_attn,
62 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
63 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);
64 }
65 if (il == n_layer - 1 && inp_out_ids) {
66 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
67 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
68 }
69 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
70 cb(cur: ffn_inp, name: "ffn_inp", il);
71
72 // feed-forward network
73 cur = build_norm(cur: ffn_inp, mw: model.layers[il].ffn_norm, NULL, type: LLM_NORM_RMS, il);
74 cb(cur, name: "ffn_norm", il);
75
76 cur = build_ffn(cur,
77 up: model.layers[il].ffn_up, NULL, NULL,
78 gate: model.layers[il].ffn_gate, NULL, NULL,
79 down: model.layers[il].ffn_down, NULL, NULL,
80 NULL, type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
81 cb(cur, name: "ffn_out", il);
82
83 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
84
85 cur = build_cvec(cur, il);
86 cb(cur, name: "l_out", il);
87
88 // input for next layer
89 inpL = cur;
90 }
91 cur = inpL;
92
93 cur = build_norm(cur, mw: model.output_norm, NULL, type: LLM_NORM_RMS, il: -1);
94
95 cb(cur, name: "result_norm", il: -1);
96 res->t_embd = cur;
97
98 // lm_head
99 cur = build_lora_mm(w: model.output, cur);
100
101 cb(cur, name: "result_output", il: -1);
102 res->t_logits = cur;
103
104 ggml_build_forward_expand(cgraph: gf, tensor: cur);
105}
106