1#include "models.h"
2
3
4llm_build_gemma::llm_build_gemma(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_tensor * cur;
8 ggml_tensor * inpL;
9
10 inpL = build_inp_embd(tok_embd: model.tok_embd);
11
12 inpL = ggml_scale(ctx: ctx0, a: inpL, s: sqrtf(x: n_embd));
13 cb(cur: inpL, name: "inp_scaled", il: -1);
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 ggml_tensor * inp_out_ids = build_inp_out_ids();
21
22 for (int il = 0; il < n_layer; ++il) {
23 // norm
24 cur = build_norm(cur: inpL,
25 mw: model.layers[il].attn_norm, NULL,
26 type: LLM_NORM_RMS, il);
27 cb(cur, name: "attn_norm", il);
28
29 // self-attention
30 {
31 // compute Q and K and RoPE them
32 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
33 cb(cur: Qcur, name: "Qcur", il);
34
35 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
36 cb(cur: Kcur, name: "Kcur", il);
37
38 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
39 cb(cur: Vcur, name: "Vcur", il);
40
41 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
42 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
43 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
44
45 Qcur = ggml_rope_ext(
46 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
47 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
48 ext_factor, attn_factor, beta_fast, beta_slow);
49
50 Kcur = ggml_rope_ext(
51 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
52 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
53 ext_factor, attn_factor, beta_fast, beta_slow);
54
55 cb(cur: Qcur, name: "Qcur", il);
56 cb(cur: Kcur, name: "Kcur", il);
57 cb(cur: Vcur, name: "Vcur", il);
58
59 Qcur = ggml_scale(ctx: ctx0, a: Qcur, s: 1.0f / sqrtf(x: float(n_embd_head)));
60 cb(cur: Qcur, name: "Qcur_scaled", il);
61
62 cur = build_attn(inp: inp_attn,
63 wo: model.layers[il].wo, NULL,
64 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale: 1.0f, il);
65 }
66 if (il == n_layer - 1 && inp_out_ids) {
67 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
68 inpL = ggml_get_rows(ctx: ctx0, a: inpL, b: inp_out_ids);
69 }
70 ggml_tensor * sa_out = ggml_add(ctx: ctx0, a: cur, b: inpL);
71 cb(cur: sa_out, name: "sa_out", il);
72
73 cur = build_norm(cur: sa_out,
74 mw: model.layers[il].ffn_norm, NULL,
75 type: LLM_NORM_RMS, il);
76 cb(cur, name: "ffn_norm", il);
77
78 // feed-forward network
79 {
80 cur = build_ffn(cur,
81 up: model.layers[il].ffn_up, NULL, NULL,
82 gate: model.layers[il].ffn_gate, NULL, NULL,
83 down: model.layers[il].ffn_down, NULL, NULL,
84 NULL,
85 type_op: LLM_FFN_GELU, type_gate: LLM_FFN_PAR, il);
86 cb(cur, name: "ffn_out", il);
87 }
88 cur = ggml_add(ctx: ctx0, a: cur, b: sa_out);
89
90 cur = build_cvec(cur, il);
91 cb(cur, name: "l_out", il);
92
93 // input for next layer
94 inpL = cur;
95 }
96 cur = inpL;
97
98 cur = build_norm(cur,
99 mw: model.output_norm, NULL,
100 type: LLM_NORM_RMS, il: -1);
101
102 cb(cur, name: "result_norm", il: -1);
103 res->t_embd = cur;
104
105 // lm_head
106 cur = build_lora_mm(w: model.output, cur);
107
108 cb(cur, name: "result_output", il: -1);
109 res->t_logits = cur;
110
111 ggml_build_forward_expand(cgraph: gf, tensor: cur);
112}
113