1#include "models.h"
2
3llm_build_nemotron::llm_build_nemotron(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,
27 mb: model.layers[il].attn_norm_b,
28 type: LLM_NORM, 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 cb(cur: Qcur, name: "Qcur", il);
36 if (model.layers[il].bq) {
37 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
38 cb(cur: Qcur, name: "Qcur", il);
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(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
44 cb(cur: Kcur, name: "Kcur", il);
45 }
46 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
47 cb(cur: Vcur, name: "Vcur", il);
48 if (model.layers[il].bv) {
49 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
50 cb(cur: Vcur, name: "Vcur", il);
51 }
52 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
53 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
54 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
55
56 Qcur = ggml_rope_ext(
57 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
58 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
59 ext_factor, attn_factor, beta_fast, beta_slow
60 );
61
62 Kcur = ggml_rope_ext(
63 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
64 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
65 ext_factor, attn_factor, beta_fast, beta_slow
66 );
67
68 cb(cur: Qcur, name: "Qcur", il);
69 cb(cur: Kcur, name: "Kcur", il);
70 cb(cur: Vcur, name: "Vcur", il);
71
72 cur = build_attn(inp: inp_attn,
73 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
74 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);
75 }
76 if (il == n_layer - 1 && inp_out_ids) {
77 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
78 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
79 }
80 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
81 cb(cur: ffn_inp, name: "ffn_inp", il);
82
83 // feed-forward network
84 cur = build_norm(cur: ffn_inp,
85 mw: model.layers[il].ffn_norm,
86 mb: model.layers[il].ffn_norm_b,
87 type: LLM_NORM, il);
88 cb(cur, name: "ffn_norm", il);
89
90 cur = build_ffn(cur,
91 up: model.layers[il].ffn_up, up_b: model.layers[il].ffn_up_b, NULL,
92 NULL, NULL, NULL,
93 down: model.layers[il].ffn_down, down_b: model.layers[il].ffn_down_b, NULL,
94 NULL,
95 type_op: LLM_FFN_RELU_SQR, type_gate: LLM_FFN_SEQ, il);
96
97 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
98 cb(cur, name: "ffn_out", il);
99
100 cur = build_cvec(cur, il);
101 cb(cur, name: "l_out", il);
102
103 // input for next layer
104 inpL = cur;
105 }
106 cur = inpL;
107
108 cur = build_norm(cur,
109 mw: model.output_norm, mb: model.output_norm_b,
110 type: LLM_NORM, il: -1);
111
112 cb(cur, name: "result_norm", il: -1);
113 res->t_embd = cur;
114
115 // lm_head
116 cur = build_lora_mm(w: model.output, cur);
117
118 cb(cur, name: "result_output", il: -1);
119 res->t_logits = cur;
120
121 ggml_build_forward_expand(cgraph: gf, tensor: cur);
122}
123