1#include "models.h"
2
3llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4 const int64_t n_embd_head = hparams.n_embd_head_k;
5
6 ggml_tensor * cur;
7 ggml_tensor * inpL;
8
9 inpL = build_inp_embd(model.tok_embd);
10
11 inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
12 cb(inpL, "inp_scaled", -1);
13
14 // inp_pos - contains the positions
15 ggml_tensor * inp_pos = build_inp_pos();
16
17 auto * inp_attn = build_attn_inp_kv_iswa();
18
19 ggml_tensor * inp_out_ids = build_inp_out_ids();
20
21 for (int il = 0; il < n_layer; ++il) {
22 const float freq_base_l = model.get_rope_freq_base (cparams, il);
23 const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
24
25 // norm
26 cur = build_norm(inpL,
27 model.layers[il].attn_norm, NULL,
28 LLM_NORM_RMS, il);
29 cb(cur, "attn_norm", il);
30
31 // self-attention
32 {
33 // compute Q and K and RoPE them
34 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
35 cb(Qcur, "Qcur", il);
36
37 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
38 cb(Kcur, "Kcur", il);
39
40 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
41 cb(Vcur, "Vcur", il);
42
43 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
44 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
45 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
46
47 Qcur = ggml_rope_ext(
48 ctx0, Qcur, inp_pos, nullptr,
49 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
50 ext_factor, attn_factor, beta_fast, beta_slow);
51
52 Kcur = ggml_rope_ext(
53 ctx0, Kcur, inp_pos, nullptr,
54 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
55 ext_factor, attn_factor, beta_fast, beta_slow);
56
57 cb(Qcur, "Qcur", il);
58 cb(Kcur, "Kcur", il);
59 cb(Vcur, "Vcur", il);
60
61 Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
62
63 cur = build_attn(inp_attn,
64 model.layers[il].wo, NULL,
65 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
66 }
67 if (il == n_layer - 1 && inp_out_ids) {
68 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
69 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
70 }
71 cur = build_norm(cur,
72 model.layers[il].attn_post_norm, NULL,
73 LLM_NORM_RMS, il);
74 cb(cur, "attn_post_norm", il);
75
76 ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
77 cb(sa_out, "sa_out", il);
78
79 cur = build_norm(sa_out,
80 model.layers[il].ffn_norm, NULL,
81 LLM_NORM_RMS, il);
82 cb(cur, "ffn_norm", il);
83
84 // feed-forward network
85 {
86 cur = build_ffn(cur,
87 model.layers[il].ffn_up, NULL, NULL,
88 model.layers[il].ffn_gate, NULL, NULL,
89 model.layers[il].ffn_down, NULL, NULL,
90 NULL,
91 LLM_FFN_GELU, LLM_FFN_PAR, il);
92 cb(cur, "ffn_out", il);
93 }
94 cur = build_norm(cur,
95 model.layers[il].ffn_post_norm, NULL,
96 LLM_NORM_RMS, -1);
97 cb(cur, "ffn_post_norm", -1);
98
99 cur = ggml_add(ctx0, cur, sa_out);
100
101 cur = build_cvec(cur, il);
102 cb(cur, "l_out", il);
103
104 // input for next layer
105 inpL = cur;
106 }
107 cur = inpL;
108
109 cur = build_norm(cur,
110 model.output_norm, NULL,
111 LLM_NORM_RMS, -1);
112
113 cb(cur, "result_norm", -1);
114 res->t_embd = cur;
115
116 // lm_head
117 cur = build_lora_mm(model.output, cur);
118
119 // final logit soft-capping
120 cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
121 cur = ggml_tanh(ctx0, cur);
122 cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
123
124 cb(cur, "result_output", -1);
125 res->t_logits = cur;
126
127 ggml_build_forward_expand(gf, cur);
128}