1#include "models.h"
2
3
4llm_build_arcee::llm_build_arcee(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(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 const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
21
22 ggml_tensor * inp_out_ids = build_inp_out_ids();
23
24 for (int il = 0; il < n_layer; ++il) {
25 ggml_tensor * inpSA = inpL;
26
27 // norm
28 cur = build_norm(inpL,
29 model.layers[il].attn_norm, NULL,
30 LLM_NORM_RMS, il);
31 cb(cur, "attn_norm", il);
32
33 // self-attention
34 {
35 // rope freq factors for llama3; may return nullptr for llama2 and other models
36 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
37
38 // compute Q and K and RoPE them
39 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
40 cb(Qcur, "Qcur", il);
41 if (model.layers[il].bq) {
42 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
43 cb(Qcur, "Qcur", il);
44 }
45
46 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
47 cb(Kcur, "Kcur", il);
48 if (model.layers[il].bk) {
49 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
50 cb(Kcur, "Kcur", il);
51 }
52
53 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
54 cb(Vcur, "Vcur", il);
55 if (model.layers[il].bv) {
56 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
57 cb(Vcur, "Vcur", il);
58 }
59
60 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
61 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
62 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
63
64 Qcur = ggml_rope_ext(
65 ctx0, Qcur, inp_pos, rope_factors,
66 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
67 ext_factor, attn_factor, beta_fast, beta_slow
68 );
69
70 Kcur = ggml_rope_ext(
71 ctx0, Kcur, inp_pos, rope_factors,
72 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
73 ext_factor, attn_factor, beta_fast, beta_slow
74 );
75
76 cb(Qcur, "Qcur", il);
77 cb(Kcur, "Kcur", il);
78 cb(Vcur, "Vcur", il);
79
80 cur = build_attn(inp_attn,
81 model.layers[il].wo, model.layers[il].bo,
82 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
83 cb(cur, "attn_out", il);
84 }
85
86 if (il == n_layer - 1 && inp_out_ids) {
87 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
88 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
89 }
90
91 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
92 cb(ffn_inp, "ffn_inp", il);
93
94 // feed-forward network
95 // ARCEE uses relu^2 instead of silu
96 cur = build_norm(ffn_inp,
97 model.layers[il].ffn_norm, NULL,
98 LLM_NORM_RMS, il);
99 cb(cur, "ffn_norm", il);
100
101 cur = build_ffn(cur,
102 model.layers[il].ffn_up, NULL, NULL,
103 NULL, NULL, NULL,
104 model.layers[il].ffn_down, NULL, NULL,
105 NULL,
106 LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
107 cb(cur, "ffn_out", il);
108
109 cur = ggml_add(ctx0, cur, ffn_inp);
110 cb(cur, "ffn_out", il);
111
112 cur = build_cvec(cur, il);
113 cb(cur, "l_out", il);
114
115 // input for next layer
116 inpL = cur;
117 }
118
119 cur = inpL;
120
121 cur = build_norm(cur,
122 model.output_norm, NULL,
123 LLM_NORM_RMS, -1);
124
125 cb(cur, "result_norm", -1);
126 res->t_embd = cur;
127
128 // lm_head
129 cur = build_lora_mm(model.output, cur);
130
131 cb(cur, "result_output", -1);
132 res->t_logits = cur;
133
134 ggml_build_forward_expand(gf, cur);
135}