1#include "models.h"
2
3
4
5llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
6 llm_graph_context(params) {
7 const int64_t n_embd_head = hparams.n_embd_head_v;
8
9 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
10 GGML_ASSERT(n_embd_head == hparams.n_rot);
11
12 ggml_tensor * cur;
13 ggml_tensor * inpL;
14
15 inpL = build_inp_embd(tok_embd: model.tok_embd);
16
17 // inp_pos - contains the positions
18 ggml_tensor * inp_pos = build_inp_pos();
19
20 auto * inp_attn = build_attn_inp_kv();
21
22 const float kq_scale =
23 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 // norm
31 cur = build_norm(cur: inpL, mw: model.layers[il].attn_norm, NULL, type: LLM_NORM_RMS, il);
32 cb(cur, name: "attn_norm", il);
33
34 // self-attention
35 {
36 // rope freq factors for llama3; may return nullptr for llama2 and other models
37 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
38
39 // compute Q and K and RoPE them
40 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
41 cb(cur: Qcur, name: "Qcur", il);
42 if (model.layers[il].bq) {
43 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
44 cb(cur: Qcur, name: "Qcur", il);
45 }
46 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
47 cb(cur: Kcur, name: "Kcur", il);
48 if (model.layers[il].bk) {
49 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
50 cb(cur: Kcur, name: "Kcur", il);
51 }
52 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
53 cb(cur: Vcur, name: "Vcur", il);
54 if (model.layers[il].bv) {
55 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
56 cb(cur: Vcur, name: "Vcur", il);
57 }
58 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
59 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
60 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
61
62 Qcur = ggml_rope_ext(ctx: ctx0, a: Qcur, b: inp_pos, c: rope_factors, 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 Kcur = ggml_rope_ext(ctx: ctx0, a: Kcur, b: inp_pos, c: rope_factors, 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 cb(cur: Qcur, name: "Qcur", il);
69 cb(cur: Kcur, name: "Kcur", il);
70 cb(cur: Vcur, name: "Vcur", il);
71
72 cur = build_attn(inp: inp_attn,
73 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
74 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale, il);
75 }
76 if (il == n_layer - 1 && inp_out_ids) {
77 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
78 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
79 }
80 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
81 cb(cur: ffn_inp, name: "ffn_inp", il);
82
83 cur = build_norm(cur: ffn_inp, mw: model.layers[il].ffn_norm, NULL, type: LLM_NORM_RMS, il);
84 cb(cur, name: "ffn_norm", il);
85
86 if ((uint32_t) il < hparams.n_layer_dense_lead) {
87 cur = build_ffn(cur,
88 up: model.layers[il].ffn_up, NULL, NULL,
89 gate: model.layers[il].ffn_gate, NULL, NULL,
90 down: model.layers[il].ffn_down, NULL, NULL,
91 NULL, type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
92 cb(cur, name: "ffn_out", il);
93 } else {
94 // MoE branch
95 ggml_tensor * moe_out = build_moe_ffn(cur,
96 gate_inp: model.layers[il].ffn_gate_inp,
97 up_exps: model.layers[il].ffn_up_exps,
98 gate_exps: model.layers[il].ffn_gate_exps,
99 down_exps: model.layers[il].ffn_down_exps,
100 exp_probs_b: nullptr,
101 n_expert, n_expert_used,
102 type_op: LLM_FFN_SILU, norm_w: false,
103 scale_w: false, w_scale: hparams.expert_weights_scale,
104 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
105 il);
106 cb(cur: moe_out, name: "ffn_moe_out", il);
107
108 // FFN shared expert
109 {
110 ggml_tensor * ffn_shexp =
111 build_ffn(cur,
112 up: model.layers[il].ffn_up_shexp, NULL, NULL,
113 gate: model.layers[il].ffn_gate_shexp, NULL, NULL,
114 down: model.layers[il].ffn_down_shexp, NULL, NULL,
115 NULL, type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
116 cb(cur: ffn_shexp, name: "ffn_shexp", il);
117
118 cur = ggml_add(ctx: ctx0, a: moe_out, b: ffn_shexp);
119 cb(cur, name: "ffn_out", il);
120 }
121 }
122 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
123
124 cur = build_cvec(cur, il);
125 cb(cur, name: "l_out", il);
126
127 // input for next layer
128 inpL = cur;
129 }
130 cur = inpL;
131
132 cur = build_norm(cur, mw: model.output_norm, NULL, type: LLM_NORM_RMS, il: -1);
133
134 cb(cur, name: "result_norm", il: -1);
135 res->t_embd = cur;
136
137 // lm_head
138 cur = build_lora_mm(w: model.output, cur);
139
140 cb(cur, name: "result_output", il: -1);
141 res->t_logits = cur;
142
143 ggml_build_forward_expand(cgraph: gf, tensor: cur);
144}
145