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(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(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 const float freq_base_l = model.get_rope_freq_base (cparams, il);
29 const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
30
31 ggml_tensor * inpSA = inpL;
32
33 // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
34 const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
35 (il + 1) % hparams.n_no_rope_layer_step != 0;
36
37 // norm
38 cur = build_norm(inpL,
39 model.layers[il].attn_norm, NULL,
40 LLM_NORM_RMS, il);
41 cb(cur, "attn_norm", il);
42
43 // self-attention
44 {
45 // rope freq factors for llama3; may return nullptr for llama2 and other models
46 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
47
48 // compute Q and K and RoPE them
49 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
50 cb(Qcur, "Qcur", il);
51 if (model.layers[il].bq) {
52 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
53 cb(Qcur, "Qcur", il);
54 }
55 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
56 cb(Kcur, "Kcur", il);
57 if (model.layers[il].bk) {
58 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
59 cb(Kcur, "Kcur", il);
60 }
61 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
62 cb(Vcur, "Vcur", il);
63 if (model.layers[il].bv) {
64 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
65 cb(Vcur, "Vcur", il);
66 }
67 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
68 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
69 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
70
71 if (use_rope) {
72 Qcur = ggml_rope_ext(
73 ctx0, Qcur, inp_pos, rope_factors,
74 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
75 ext_factor, attn_factor, beta_fast, beta_slow
76 );
77
78 Kcur = ggml_rope_ext(
79 ctx0, Kcur, inp_pos, rope_factors,
80 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
81 ext_factor, attn_factor, beta_fast, beta_slow
82 );
83 } else if (inp_attn_scale) {
84 Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
85 }
86 cb(Qcur, "Qcur", il);
87 cb(Kcur, "Kcur", il);
88 cb(Vcur, "Vcur", il);
89
90 if (use_rope && hparams.use_kq_norm) {
91 // Llama4TextL2Norm
92 Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
93 Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
94 cb(Qcur, "Qcur_normed", il);
95 cb(Kcur, "Kcur_normed", il);
96 }
97 cur = build_attn(inp_attn,
98 model.layers[il].wo, model.layers[il].bo,
99 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
100 cb(cur, "attn_out", il);
101 }
102 if (il == n_layer - 1 && inp_out_ids) {
103 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
104 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
105 }
106 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
107 cb(ffn_inp, "ffn_inp", il);
108
109 // feed-forward network (non-MoE)
110 if (model.layers[il].ffn_gate_inp == nullptr) {
111 cur = build_norm(ffn_inp,
112 model.layers[il].ffn_norm, NULL,
113 LLM_NORM_RMS, il);
114 cb(cur, "ffn_norm", il);
115
116 cur = build_ffn(cur,
117 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
118 model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
119 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
120 NULL,
121 LLM_FFN_SILU, LLM_FFN_PAR, il);
122 cb(cur, "ffn_out", il);
123 } else {
124 ggml_tensor * ffn_inp_normed = build_norm(ffn_inp,
125 model.layers[il].ffn_norm, NULL,
126 LLM_NORM_RMS, il);
127 cb(cur, "ffn_norm", il);
128
129 ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed,
130 model.layers[il].ffn_gate_inp,
131 model.layers[il].ffn_up_exps,
132 model.layers[il].ffn_gate_exps,
133 model.layers[il].ffn_down_exps,
134 nullptr,
135 n_expert, n_expert_used,
136 LLM_FFN_SILU, false,
137 false, 0.0,
138 LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
139 il);
140
141 // Shared experts
142 ggml_tensor * shexp_out = build_ffn(ffn_inp_normed,
143 model.layers[il].ffn_up_shexp, NULL, NULL,
144 model.layers[il].ffn_gate_shexp, NULL, NULL,
145 model.layers[il].ffn_down_shexp, NULL, NULL,
146 NULL,
147 LLM_FFN_SILU, LLM_FFN_PAR, il);
148 cb(shexp_out, "ffn_moe_shexp", il);
149
150 cur = ggml_add(ctx0, moe_out, shexp_out);
151 cb(cur, "ffn_moe_out_merged", il);
152 }
153 cur = ggml_add(ctx0, cur, ffn_inp);
154 cb(cur, "ffn_out", il);
155
156 cur = build_cvec(cur, il);
157 cb(cur, "l_out", il);
158
159 // input for next layer
160 inpL = cur;
161 }
162 cur = inpL;
163
164 cur = build_norm(cur,
165 model.output_norm, NULL,
166 LLM_NORM_RMS, -1);
167
168 cb(cur, "result_norm", -1);
169 res->t_embd = cur;
170
171 // lm_head
172 cur = build_lora_mm(model.output, cur);
173
174 cb(cur, "result_output", -1);
175 res->t_logits = cur;
176
177 ggml_build_forward_expand(gf, cur);
178}