1#include "models.h"
2
3llm_build_llama_iswa::llm_build_llama_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_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 // temperature tuning
18 ggml_tensor * inp_attn_scale = nullptr;
19 inp_attn_scale = build_inp_attn_scale();
20
21 auto * inp_attn = build_attn_inp_kv_iswa();
22
23 const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(x: float(n_embd_head)) : hparams.f_attention_scale;
24
25 ggml_tensor * inp_out_ids = build_inp_out_ids();
26
27 for (int il = 0; il < n_layer; ++il) {
28 ggml_tensor * inpSA = inpL;
29
30 const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
31 (il + 1) % hparams.n_no_rope_layer_step != 0;
32
33 // norm
34 cur = build_norm(cur: inpL,
35 mw: model.layers[il].attn_norm, NULL,
36 type: LLM_NORM_RMS, il);
37 cb(cur, name: "attn_norm", il);
38
39 // self-attention
40 {
41 // rope freq factors for llama3; may return nullptr for llama2 and other models
42 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
43
44 // compute Q and K and RoPE them
45 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
46 cb(cur: Qcur, name: "Qcur", il);
47 if (model.layers[il].bq) {
48 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
49 cb(cur: Qcur, name: "Qcur", il);
50 }
51 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
52 cb(cur: Kcur, name: "Kcur", il);
53 if (model.layers[il].bk) {
54 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
55 cb(cur: Kcur, name: "Kcur", il);
56 }
57 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
58 cb(cur: Vcur, name: "Vcur", il);
59 if (model.layers[il].bv) {
60 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
61 cb(cur: Vcur, name: "Vcur", il);
62 }
63 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
64 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
65 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
66
67 if (use_rope) {
68 Qcur = ggml_rope_ext(
69 ctx: ctx0, a: Qcur, b: inp_pos, c: rope_factors,
70 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
71 ext_factor, attn_factor, beta_fast, beta_slow
72 );
73
74 Kcur = ggml_rope_ext(
75 ctx: ctx0, a: Kcur, b: inp_pos, c: rope_factors,
76 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
77 ext_factor, attn_factor, beta_fast, beta_slow
78 );
79 } else if (inp_attn_scale) {
80 Qcur = ggml_mul(ctx: ctx0, a: Qcur, b: inp_attn_scale);
81 }
82 cb(cur: Qcur, name: "Qcur", il);
83 cb(cur: Kcur, name: "Kcur", il);
84 cb(cur: Vcur, name: "Vcur", il);
85
86 if (use_rope && hparams.use_kq_norm) {
87 // Llama4TextL2Norm
88 Qcur = ggml_rms_norm(ctx: ctx0, a: Qcur, eps: hparams.f_norm_rms_eps);
89 Kcur = ggml_rms_norm(ctx: ctx0, a: Kcur, eps: hparams.f_norm_rms_eps);
90 cb(cur: Qcur, name: "Qcur_normed", il);
91 cb(cur: Kcur, name: "Kcur_normed", il);
92 }
93 cur = build_attn(inp: inp_attn,
94 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
95 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale, il);
96 cb(cur, name: "attn_out", il);
97 }
98 if (il == n_layer - 1 && inp_out_ids) {
99 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
100 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
101 }
102 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
103 cb(cur: ffn_inp, name: "ffn_inp", il);
104
105 // feed-forward network (non-MoE)
106 if (model.layers[il].ffn_gate_inp == nullptr) {
107 cur = build_norm(cur: ffn_inp,
108 mw: model.layers[il].ffn_norm, NULL,
109 type: LLM_NORM_RMS, il);
110 cb(cur, name: "ffn_norm", il);
111
112 cur = build_ffn(cur,
113 up: model.layers[il].ffn_up, up_b: model.layers[il].ffn_up_b, NULL,
114 gate: model.layers[il].ffn_gate, gate_b: model.layers[il].ffn_gate_b, NULL,
115 down: model.layers[il].ffn_down, down_b: model.layers[il].ffn_down_b, NULL,
116 NULL,
117 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
118 cb(cur, name: "ffn_out", il);
119 } else {
120 ggml_tensor * ffn_inp_normed = build_norm(cur: ffn_inp,
121 mw: model.layers[il].ffn_norm, NULL,
122 type: LLM_NORM_RMS, il);
123 cb(cur, name: "ffn_norm", il);
124
125 ggml_tensor * moe_out = build_moe_ffn(cur: ffn_inp_normed,
126 gate_inp: model.layers[il].ffn_gate_inp,
127 up_exps: model.layers[il].ffn_up_exps,
128 gate_exps: model.layers[il].ffn_gate_exps,
129 down_exps: model.layers[il].ffn_down_exps,
130 exp_probs_b: nullptr,
131 n_expert, n_expert_used,
132 type_op: LLM_FFN_SILU, norm_w: false,
133 scale_w: false, w_scale: 0.0,
134 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
135 il);
136
137 // Shared experts
138 ggml_tensor * shexp_out = build_ffn(cur: ffn_inp_normed,
139 up: model.layers[il].ffn_up_shexp, NULL, NULL,
140 gate: model.layers[il].ffn_gate_shexp, NULL, NULL,
141 down: model.layers[il].ffn_down_shexp, NULL, NULL,
142 NULL,
143 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
144 cb(cur: shexp_out, name: "ffn_moe_shexp", il);
145
146 cur = ggml_add(ctx: ctx0, a: moe_out, b: shexp_out);
147 cb(cur, name: "ffn_moe_out_merged", il);
148 }
149 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
150 cb(cur, name: "ffn_out", il);
151
152 cur = build_cvec(cur, il);
153 cb(cur, name: "l_out", il);
154
155 // input for next layer
156 inpL = cur;
157 }
158 cur = inpL;
159
160 cur = build_norm(cur,
161 mw: model.output_norm, NULL,
162 type: LLM_NORM_RMS, il: -1);
163
164 cb(cur, name: "result_norm", il: -1);
165 res->t_embd = cur;
166
167 // lm_head
168 cur = build_lora_mm(w: model.output, cur);
169
170 cb(cur, name: "result_output", il: -1);
171 res->t_logits = cur;
172
173 ggml_build_forward_expand(cgraph: gf, tensor: cur);
174}
175