1#include "models.h"
2
3llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4 const int64_t n_embd_head = hparams.n_embd_head_v;
5
6 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
7 GGML_ASSERT(n_embd_head == hparams.n_rot);
8
9 ggml_tensor * cur;
10 ggml_tensor * inpL;
11
12 inpL = build_inp_embd(tok_embd: model.tok_embd);
13
14 // inp_pos - contains the positions
15 ggml_tensor * inp_pos = build_inp_pos();
16
17 auto * inp_attn = build_attn_inp_kv();
18
19 const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(x: float(n_embd_head)) : hparams.f_attention_scale;
20
21 ggml_tensor * inp_out_ids = build_inp_out_ids();
22
23 for (int il = 0; il < n_layer; ++il) {
24 ggml_tensor * inpSA = inpL;
25
26 // norm
27 cur = build_norm(cur: inpL,
28 mw: model.layers[il].attn_norm, NULL,
29 type: LLM_NORM_RMS, il);
30 cb(cur, name: "attn_norm", il);
31
32 // self-attention
33 {
34 // rope freq factors for llama3; may return nullptr for llama2 and other models
35 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
36
37 // compute Q and K and RoPE them
38 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
39 cb(cur: Qcur, name: "Qcur", il);
40 if (model.layers[il].bq) {
41 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
42 cb(cur: Qcur, name: "Qcur", il);
43 }
44 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
45 cb(cur: Kcur, name: "Kcur", il);
46 if (model.layers[il].bk) {
47 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
48 cb(cur: Kcur, name: "Kcur", il);
49 }
50 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
51 cb(cur: Vcur, name: "Vcur", il);
52 if (model.layers[il].bv) {
53 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
54 cb(cur: Vcur, name: "Vcur", il);
55 }
56 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
57 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
58 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
59
60 Qcur = ggml_rope_ext(
61 ctx: ctx0, a: Qcur, b: inp_pos, c: rope_factors,
62 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
63 ext_factor, attn_factor, beta_fast, beta_slow
64 );
65
66 Kcur = ggml_rope_ext(
67 ctx: ctx0, a: Kcur, b: inp_pos, c: rope_factors,
68 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
69 ext_factor, attn_factor, beta_fast, beta_slow
70 );
71
72 cb(cur: Qcur, name: "Qcur", il);
73 cb(cur: Kcur, name: "Kcur", il);
74 cb(cur: Vcur, name: "Vcur", il);
75
76 if (hparams.use_kq_norm) {
77 // Llama4TextL2Norm
78 Qcur = ggml_rms_norm(ctx: ctx0, a: Qcur, eps: hparams.f_norm_rms_eps);
79 Kcur = ggml_rms_norm(ctx: ctx0, a: Kcur, eps: hparams.f_norm_rms_eps);
80 cb(cur: Qcur, name: "Qcur_normed", il);
81 cb(cur: Kcur, name: "Kcur_normed", il);
82 }
83 cur = build_attn(inp: inp_attn,
84 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
85 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale, il);
86 cb(cur, name: "attn_out", il);
87 }
88 if (il == n_layer - 1 && inp_out_ids) {
89 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
90 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
91 }
92 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
93 cb(cur: ffn_inp, name: "ffn_inp", il);
94
95 // feed-forward network (non-MoE)
96 if (model.layers[il].ffn_gate_inp == nullptr) {
97
98 cur = build_norm(cur: ffn_inp,
99 mw: model.layers[il].ffn_norm, NULL,
100 type: LLM_NORM_RMS, il);
101 cb(cur, name: "ffn_norm", il);
102
103 cur = build_ffn(cur,
104 up: model.layers[il].ffn_up, up_b: model.layers[il].ffn_up_b, NULL,
105 gate: model.layers[il].ffn_gate, gate_b: model.layers[il].ffn_gate_b, NULL,
106 down: model.layers[il].ffn_down, down_b: model.layers[il].ffn_down_b, NULL,
107 NULL,
108 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
109 cb(cur, name: "ffn_out", il);
110 } else {
111 // MoE branch
112 cur = build_norm(cur: ffn_inp,
113 mw: model.layers[il].ffn_norm, NULL,
114 type: LLM_NORM_RMS, il);
115 cb(cur, name: "ffn_norm", il);
116
117 cur = build_moe_ffn(cur,
118 gate_inp: model.layers[il].ffn_gate_inp,
119 up_exps: model.layers[il].ffn_up_exps,
120 gate_exps: model.layers[il].ffn_gate_exps,
121 down_exps: model.layers[il].ffn_down_exps,
122 exp_probs_b: nullptr,
123 n_expert, n_expert_used,
124 type_op: LLM_FFN_SILU, norm_w: true,
125 scale_w: false, w_scale: 0.0,
126 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
127 il);
128 cb(cur, name: "ffn_moe_out", il);
129 }
130 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
131 cb(cur, name: "ffn_out", il);
132
133 cur = build_cvec(cur, il);
134 cb(cur, name: "l_out", il);
135
136 // input for next layer
137 inpL = cur;
138 }
139 cur = inpL;
140
141 cur = build_norm(cur,
142 mw: model.output_norm, NULL,
143 type: LLM_NORM_RMS, il: -1);
144
145 cb(cur, name: "result_norm", il: -1);
146 res->t_embd = cur;
147
148 // lm_head
149 cur = build_lora_mm(w: model.output, cur);
150
151 cb(cur, name: "result_output", il: -1);
152 res->t_logits = cur;
153
154 ggml_build_forward_expand(cgraph: gf, tensor: cur);
155}
156