1#include "models.h"
2
3llm_build_seed_oss::llm_build_seed_oss(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 // compute Q and K and RoPE them
35 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
36 cb(cur: Qcur, name: "Qcur", il);
37 if (model.layers[il].bq) {
38 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
39 cb(cur: Qcur, name: "Qcur", il);
40 }
41 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
42 cb(cur: Kcur, name: "Kcur", il);
43 if (model.layers[il].bk) {
44 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
45 cb(cur: Kcur, name: "Kcur", il);
46 }
47 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
48 cb(cur: Vcur, name: "Vcur", il);
49 if (model.layers[il].bv) {
50 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
51 cb(cur: Vcur, name: "Vcur", il);
52 }
53 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
54 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
55 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
56
57 Qcur = ggml_rope_ext(
58 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
59 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
60 ext_factor, attn_factor, beta_fast, beta_slow
61 );
62
63 Kcur = ggml_rope_ext(
64 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
65 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
66 ext_factor, attn_factor, beta_fast, beta_slow
67 );
68
69 cb(cur: Qcur, name: "Qcur", il);
70 cb(cur: Kcur, name: "Kcur", il);
71 cb(cur: Vcur, name: "Vcur", il);
72
73 cur = build_attn(inp: inp_attn,
74 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
75 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale, il);
76 cb(cur, name: "attn_out", il);
77 }
78 if (il == n_layer - 1 && inp_out_ids) {
79 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
80 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
81 }
82 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
83 cb(cur: ffn_inp, name: "ffn_inp", il);
84
85 // feed-forward network
86 cur = build_norm(cur: ffn_inp,
87 mw: model.layers[il].attn_post_norm, NULL,
88 type: LLM_NORM_RMS, il);
89 cb(cur, name: "attn_post_norm", il);
90
91 cur = build_ffn(cur,
92 up: model.layers[il].ffn_up, NULL, NULL,
93 gate: model.layers[il].ffn_gate, NULL, NULL,
94 down: model.layers[il].ffn_down, NULL, NULL,
95 NULL,
96 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
97 cb(cur, name: "ffn_out", il);
98
99 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
100 cb(cur, name: "ffn_out", il);
101
102 cur = build_cvec(cur, il);
103 cb(cur, name: "l_out", il);
104
105 // input for next layer
106 inpL = cur;
107 }
108 cur = inpL;
109
110 cur = build_norm(cur,
111 mw: model.output_norm, NULL,
112 type: LLM_NORM_RMS, il: -1);
113
114 cb(cur, name: "result_norm", il: -1);
115 res->t_embd = cur;
116
117 // lm_head
118 cur = build_lora_mm(w: model.output, cur);
119
120 cb(cur, name: "result_output", il: -1);
121 res->t_logits = cur;
122
123 ggml_build_forward_expand(cgraph: gf, tensor: cur);
124}
125