1#include "models.h"
2
3llm_build_llada_moe::llm_build_llada_moe(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_no_cache();
18
19 ggml_tensor * inp_out_ids = build_inp_out_ids();
20
21 for (int il = 0; il < n_layer; ++il) {
22 ggml_tensor * inpSA = inpL;
23
24 // norm
25 cur = build_norm(cur: inpL,
26 mw: model.layers[il].attn_norm, NULL,
27 type: LLM_NORM_RMS, il);
28 cb(cur, name: "attn_norm", il);
29
30 // self_attention
31 {
32 // compute Q and K and RoPE them
33 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
34 cb(cur: Qcur, name: "Qcur", il);
35
36 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
37 cb(cur: Kcur, name: "Kcur", il);
38
39 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
40 cb(cur: Vcur, name: "Vcur", il);
41
42 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
43 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
44 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
45
46 Qcur = build_norm(cur: Qcur, mw: model.layers[il].attn_q_norm, NULL, type: LLM_NORM_RMS, il);
47 cb(cur: Qcur, name: "Qcur_normed", il);
48
49 Kcur = build_norm(cur: Kcur, mw: model.layers[il].attn_k_norm, NULL, type: LLM_NORM_RMS, il);
50 cb(cur: Kcur, name: "Kcur_normed", il);
51
52 Qcur = ggml_rope_ext(
53 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
54 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
55 ext_factor, attn_factor, beta_fast, beta_slow
56 );
57
58 Kcur = ggml_rope_ext(
59 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
60 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
61 ext_factor, attn_factor, beta_fast, beta_slow
62 );
63
64 cb(cur: Qcur, name: "Qcur", il);
65 cb(cur: Kcur, name: "Kcur", il);
66 cb(cur: Vcur, name: "Vcur", il);
67
68 cur = build_attn(inp: inp_attn,
69 wo: model.layers[il].wo, NULL,
70 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale: 1.0f/sqrtf(x: float(n_embd_head)), il);
71 }
72 if (il == n_layer - 1 && inp_out_ids) {
73 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
74 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
75 }
76 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
77 cb(cur: ffn_inp, name: "ffn_inp", il);
78
79 // MoE branch
80 cur = build_norm(cur: ffn_inp,
81 mw: model.layers[il].ffn_norm, NULL,
82 type: LLM_NORM_RMS, il);
83 cb(cur, name: "ffn_norm", il);
84
85 cur = build_moe_ffn(cur,
86 gate_inp: model.layers[il].ffn_gate_inp,
87 up_exps: model.layers[il].ffn_up_exps,
88 gate_exps: model.layers[il].ffn_gate_exps,
89 down_exps: model.layers[il].ffn_down_exps,
90 exp_probs_b: nullptr,
91 n_expert, n_expert_used,
92 type_op: LLM_FFN_SILU, norm_w: false,
93 scale_w: false, w_scale: 0.0,
94 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
95 il);
96 cb(cur, name: "ffn_moe_out", il);
97
98 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
99
100 cur = build_cvec(cur, il);
101 cb(cur, name: "l_out", il);
102
103 // input for next layer
104 inpL = cur;
105 }
106 cur = inpL;
107
108 cur = build_norm(cur,
109 mw: model.output_norm, NULL,
110 type: LLM_NORM_RMS, il: -1);
111
112 cb(cur, name: "result_norm", il: -1);
113 res->t_embd = cur;
114
115 // lm_head
116 cur = build_lora_mm(w: model.output, cur);
117
118 cb(cur, name: "result_output", il: -1);
119 res->t_logits = cur;
120
121 ggml_build_forward_expand(cgraph: gf, tensor: cur);
122}
123