1#include "models.h"
2
3
4
5llm_build_deci::llm_build_deci(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
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 const float kq_scale =
22 hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(x: float(n_embd_head)) : hparams.f_attention_scale;
23
24 ggml_tensor * inp_out_ids = build_inp_out_ids();
25
26 for (int il = 0; il < n_layer; ++il) {
27 ggml_tensor * inpSA = inpL;
28 const int64_t n_head_kv = hparams.n_head_kv(il);
29 const int64_t n_head = hparams.n_head(il);
30 const int64_t n_ff = hparams.n_ff(il);
31
32 if (n_head == 0) {
33 // attention-free layer of Llama-3_1-Nemotron-51B
34 cur = inpL;
35 } else {
36 // norm
37 cur = build_norm(cur: inpL, mw: model.layers[il].attn_norm, NULL, type: LLM_NORM_RMS, il);
38 cb(cur, name: "attn_norm", il);
39 }
40 if (n_head > 0 && n_head_kv == 0) {
41 // "linear attention" of Llama-3_1-Nemotron-51B
42 cur = build_lora_mm(w: model.layers[il].wo, cur);
43 cb(cur, name: "wo", il);
44 } else if (n_head > 0) {
45 // self-attention
46 // rope freq factors for llama3; may return nullptr for llama2 and other models
47 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
48
49 // compute Q and K and RoPE them
50 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
51 cb(cur: Qcur, name: "Qcur", il);
52 if (model.layers[il].bq) {
53 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
54 cb(cur: Qcur, name: "Qcur", il);
55 }
56 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
57 cb(cur: Kcur, name: "Kcur", il);
58 if (model.layers[il].bk) {
59 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
60 cb(cur: Kcur, name: "Kcur", il);
61 }
62 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
63 cb(cur: Vcur, name: "Vcur", il);
64 if (model.layers[il].bv) {
65 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
66 cb(cur: Vcur, name: "Vcur", il);
67 }
68 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
69 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
70 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
71
72 Qcur = ggml_rope_ext(ctx: ctx0, a: Qcur, b: inp_pos, c: rope_factors, n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
73 ext_factor, attn_factor, beta_fast, beta_slow);
74
75 Kcur = ggml_rope_ext(ctx: ctx0, a: Kcur, b: inp_pos, c: rope_factors, n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
76 ext_factor, attn_factor, beta_fast, beta_slow);
77
78 cb(cur: Qcur, name: "Qcur", il);
79 cb(cur: Kcur, name: "Kcur", il);
80 cb(cur: Vcur, name: "Vcur", il);
81
82 cur = build_attn(inp: inp_attn,
83 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
84 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale, il);
85 }
86 if (il == n_layer - 1 && inp_out_ids) {
87 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
88 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
89 }
90 // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B
91 if (n_ff == 0) {
92 continue;
93 }
94 // modified to support attention-free layer of Llama-3_1-Nemotron-51B
95 ggml_tensor * ffn_inp = cur;
96 if (n_head > 0) {
97 ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
98 cb(cur: ffn_inp, name: "ffn_inp", il);
99 }
100 // feed-forward network
101 if (model.layers[il].ffn_gate_inp == nullptr) {
102 cur = build_norm(cur: ffn_inp, mw: model.layers[il].ffn_norm, NULL, type: LLM_NORM_RMS, il);
103 cb(cur, name: "ffn_norm", il);
104
105 cur = build_ffn(cur,
106 up: model.layers[il].ffn_up, up_b: model.layers[il].ffn_up_b, NULL,
107 gate: model.layers[il].ffn_gate, gate_b: model.layers[il].ffn_gate_b, NULL,
108 down: model.layers[il].ffn_down, down_b: model.layers[il].ffn_down_b, NULL,
109 NULL, type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
110 cb(cur, name: "ffn_out", il);
111 }
112 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
113 cb(cur, name: "ffn_out", il);
114
115 cur = build_cvec(cur, il);
116 cb(cur, name: "l_out", il);
117
118 // input for next layer
119 inpL = cur;
120 }
121 cur = inpL;
122
123 cur = build_norm(cur, mw: model.output_norm, NULL, type: LLM_NORM_RMS, il: -1);
124
125 cb(cur, name: "result_norm", il: -1);
126 res->t_embd = cur;
127
128 // lm_head
129 cur = build_lora_mm(w: model.output, cur);
130
131 cb(cur, name: "result_output", il: -1);
132 res->t_logits = cur;
133
134 ggml_build_forward_expand(cgraph: gf, tensor: cur);
135}
136