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(tok_embd: model.tok_embd);
10
11 inpL = ggml_scale(ctx: ctx0, a: inpL, s: sqrtf(x: n_embd));
12 cb(cur: inpL, name: "inp_scaled", il: -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 // norm
23 cur = build_norm(cur: inpL,
24 mw: model.layers[il].attn_norm, NULL,
25 type: LLM_NORM_RMS, il);
26 cb(cur, name: "attn_norm", il);
27
28 // self-attention
29 {
30 // compute Q and K and RoPE them
31 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
32 cb(cur: Qcur, name: "Qcur", il);
33
34 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
35 cb(cur: Kcur, name: "Kcur", il);
36
37 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
38 cb(cur: Vcur, name: "Vcur", il);
39
40 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
41 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
42 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
43
44 Qcur = ggml_rope_ext(
45 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
46 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
47 ext_factor, attn_factor, beta_fast, beta_slow);
48
49 Kcur = ggml_rope_ext(
50 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
51 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
52 ext_factor, attn_factor, beta_fast, beta_slow);
53
54 cb(cur: Qcur, name: "Qcur", il);
55 cb(cur: Kcur, name: "Kcur", il);
56 cb(cur: Vcur, name: "Vcur", il);
57
58 Qcur = ggml_scale(ctx: ctx0, a: Qcur, s: hparams.f_attention_scale);
59
60 cur = build_attn(inp: inp_attn,
61 wo: model.layers[il].wo, NULL,
62 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale: 1.0f, il);
63 }
64 if (il == n_layer - 1 && inp_out_ids) {
65 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
66 inpL = ggml_get_rows(ctx: ctx0, a: inpL, b: inp_out_ids);
67 }
68 cur = build_norm(cur,
69 mw: model.layers[il].attn_post_norm, NULL,
70 type: LLM_NORM_RMS, il);
71 cb(cur, name: "attn_post_norm", il);
72
73 ggml_tensor * sa_out = ggml_add(ctx: ctx0, a: cur, b: inpL);
74 cb(cur: sa_out, name: "sa_out", il);
75
76 cur = build_norm(cur: sa_out,
77 mw: model.layers[il].ffn_norm, NULL,
78 type: LLM_NORM_RMS, il);
79 cb(cur, name: "ffn_norm", il);
80
81 // feed-forward network
82 {
83 cur = build_ffn(cur,
84 up: model.layers[il].ffn_up, NULL, NULL,
85 gate: model.layers[il].ffn_gate, NULL, NULL,
86 down: model.layers[il].ffn_down, NULL, NULL,
87 NULL,
88 type_op: LLM_FFN_GELU, type_gate: LLM_FFN_PAR, il);
89 cb(cur, name: "ffn_out", il);
90 }
91 cur = build_norm(cur,
92 mw: model.layers[il].ffn_post_norm, NULL,
93 type: LLM_NORM_RMS, il: -1);
94 cb(cur, name: "ffn_post_norm", il: -1);
95
96 cur = ggml_add(ctx: ctx0, a: cur, b: sa_out);
97
98 cur = build_cvec(cur, il);
99 cb(cur, name: "l_out", il);
100
101 // input for next layer
102 inpL = cur;
103 }
104 cur = inpL;
105
106 cur = build_norm(cur,
107 mw: model.output_norm, NULL,
108 type: LLM_NORM_RMS, il: -1);
109
110 cb(cur, name: "result_norm", il: -1);
111 res->t_embd = cur;
112
113 // lm_head
114 cur = build_lora_mm(w: model.output, cur);
115
116 // final logit soft-capping
117 cur = ggml_scale(ctx: ctx0, a: cur, s: 1.0f / hparams.f_final_logit_softcapping);
118 cur = ggml_tanh(ctx: ctx0, a: cur);
119 cur = ggml_scale(ctx: ctx0, a: cur, s: hparams.f_final_logit_softcapping);
120
121 cb(cur, name: "result_output", il: -1);
122 res->t_logits = cur;
123
124 ggml_build_forward_expand(cgraph: gf, tensor: cur);
125}
126