1#include "models.h"
2
3template <bool iswa>
4llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
5 const int64_t n_embd_head = hparams.n_embd_head_k;
6
7 ggml_tensor * cur;
8 ggml_tensor * inpL;
9
10 inpL = build_inp_embd(model.tok_embd);
11
12 // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
13 inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
14 cb(inpL, "inp_scaled", -1);
15
16 // inp_pos - contains the positions
17 ggml_tensor * inp_pos = build_inp_pos();
18
19 // TODO: is causal == true correct? might need some changes
20 using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
21 inp_attn_type * inp_attn = nullptr;
22
23 if constexpr (iswa) {
24 inp_attn = build_attn_inp_kv_iswa();
25 } else {
26 inp_attn = build_attn_inp_kv();
27 }
28
29 ggml_tensor * inp_out_ids = build_inp_out_ids();
30
31 for (int il = 0; il < n_layer; ++il) {
32 float freq_base_l = 0.0f;
33 float freq_scale_l = 0.0f;
34
35 if constexpr (iswa) {
36 freq_base_l = model.get_rope_freq_base (cparams, il);
37 freq_scale_l = model.get_rope_freq_scale(cparams, il);
38 } else {
39 freq_base_l = freq_base;
40 freq_scale_l = freq_scale;
41 }
42
43 // norm
44 cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
45 cb(cur, "attn_norm", il);
46
47 // self-attention
48 {
49 // compute Q and K and RoPE them
50 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
51 cb(Qcur, "Qcur", il);
52
53 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
54 cb(Kcur, "Kcur", il);
55
56 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
57 cb(Vcur, "Vcur", il);
58
59 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
60 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
61 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
62
63 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
64 cb(Qcur, "Qcur_normed", il);
65
66 Qcur = ggml_rope_ext(
67 ctx0, Qcur, inp_pos, nullptr,
68 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
69 ext_factor, attn_factor, beta_fast, beta_slow);
70
71 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
72 cb(Kcur, "Kcur_normed", il);
73
74 Kcur = ggml_rope_ext(
75 ctx0, Kcur, inp_pos, nullptr,
76 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
77 ext_factor, attn_factor, beta_fast, beta_slow);
78
79 cb(Qcur, "Qcur", il);
80 cb(Kcur, "Kcur", il);
81 cb(Vcur, "Vcur", il);
82
83 // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
84 Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
85
86 cur = build_attn(inp_attn,
87 model.layers[il].wo, NULL,
88 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
89 }
90 if (il == n_layer - 1 && inp_out_ids) {
91 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
92 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
93 }
94 cur = build_norm(cur,
95 model.layers[il].attn_post_norm, NULL,
96 LLM_NORM_RMS, il);
97 cb(cur, "attn_post_norm", il);
98
99 ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
100 cb(sa_out, "sa_out", il);
101
102 cur = build_norm(sa_out,
103 model.layers[il].ffn_norm, NULL,
104 LLM_NORM_RMS, il);
105 cb(cur, "ffn_norm", il);
106
107 // feed-forward network
108 {
109 cur = build_ffn(cur,
110 model.layers[il].ffn_up, NULL, NULL,
111 model.layers[il].ffn_gate, NULL, NULL,
112 model.layers[il].ffn_down, NULL, NULL,
113 NULL,
114 LLM_FFN_GELU, LLM_FFN_PAR, il);
115 cb(cur, "ffn_out", il);
116 }
117 cur = build_norm(cur,
118 model.layers[il].ffn_post_norm, NULL,
119 LLM_NORM_RMS, -1);
120 cb(cur, "ffn_post_norm", il);
121
122 cur = ggml_add(ctx0, cur, sa_out);
123
124 cur = build_cvec(cur, il);
125 cb(cur, "l_out", il);
126
127 // input for next layer
128 inpL = cur;
129 }
130 cur = inpL;
131
132 cur = build_norm(cur,
133 model.output_norm, NULL,
134 LLM_NORM_RMS, -1);
135
136 cb(cur, "result_norm", -1);
137 res->t_embd = cur;
138
139 // lm_head
140 cur = build_lora_mm(model.output, cur);
141
142 if (hparams.f_final_logit_softcapping) {
143 cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
144 cur = ggml_tanh(ctx0, cur);
145 cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
146 }
147
148 cb(cur, "result_output", -1);
149 res->t_logits = cur;
150
151 ggml_build_forward_expand(gf, cur);
152}
153
154template struct llm_build_gemma3<false>;
155template struct llm_build_gemma3<true>;