1#include "models.h"
2
3llm_build_openai_moe_iswa::llm_build_openai_moe_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4 ggml_tensor * cur;
5 ggml_tensor * inpL;
6
7 inpL = build_inp_embd(tok_embd: model.tok_embd);
8
9 // inp_pos - contains the positions
10 ggml_tensor * inp_pos = build_inp_pos();
11
12 auto * inp_attn = build_attn_inp_kv_iswa();
13
14 for (int il = 0; il < n_layer; ++il) {
15 ggml_tensor * inpSA = inpL;
16
17 // norm
18 cur = build_norm(cur: inpL,
19 mw: model.layers[il].attn_norm, mb: nullptr,
20 type: LLM_NORM_RMS, il);
21 cb(cur, name: "attn_norm", il);
22
23 // self-attention
24 {
25 // compute Q and K and RoPE them
26 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
27 cb(cur: Qcur, name: "Qcur", il);
28 if (model.layers[il].bq) {
29 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
30 cb(cur: Qcur, name: "Qcur", il);
31 }
32 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
33 cb(cur: Kcur, name: "Kcur", il);
34 if (model.layers[il].bk) {
35 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
36 cb(cur: Kcur, name: "Kcur", il);
37 }
38 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
39 cb(cur: Vcur, name: "Vcur", il);
40 if (model.layers[il].bv) {
41 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
42 cb(cur: Vcur, name: "Vcur", il);
43 }
44 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_rot, ne1: n_head, ne2: n_tokens);
45 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_rot, ne1: n_head_kv, ne2: n_tokens);
46 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_rot, ne1: n_head_kv, ne2: n_tokens);
47
48 Qcur = ggml_rope_ext(
49 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
50 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
51 ext_factor, attn_factor, beta_fast, beta_slow
52 );
53
54 Kcur = ggml_rope_ext(
55 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
56 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
57 ext_factor, attn_factor, beta_fast, beta_slow
58 );
59
60 cb(cur: Qcur, name: "Qcur", il);
61 cb(cur: Kcur, name: "Kcur", il);
62 cb(cur: Vcur, name: "Vcur", il);
63
64 cur = build_attn(inp: inp_attn,
65 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
66 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: model.layers[il].attn_sinks, v_mla: nullptr, kq_scale: 1.0f/sqrtf(x: float(n_rot)), il);
67
68 cb(cur, name: "attn_out", il);
69 }
70 if (il == n_layer - 1) {
71 // skip computing output for unused tokens
72 ggml_tensor * inp_out_ids = build_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 cur = ffn_inp;
80 cur = build_norm(cur,
81 mw: model.layers[il].attn_post_norm, mb: nullptr,
82 type: LLM_NORM_RMS, il);
83 cb(cur, name: "attn_post_norm", il);
84
85 // MoE branch
86 cur = build_moe_ffn(cur,
87 gate_inp: model.layers[il].ffn_gate_inp, gate_inp_b: model.layers[il].ffn_gate_inp_b,
88 up_exps: model.layers[il].ffn_up_exps, up_exps_b: model.layers[il].ffn_up_exps_b,
89 gate_exps: model.layers[il].ffn_gate_exps, gate_exps_b: model.layers[il].ffn_gate_exps_b,
90 down_exps: model.layers[il].ffn_down_exps, down_exps_b: model.layers[il].ffn_down_exps_b,
91 exp_probs_b: nullptr,
92 n_expert, n_expert_used,
93 type_op: LLM_FFN_SWIGLU_OAI_MOE, norm_w: false,
94 scale_w: false, w_scale: 0.0,
95 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT,
96 il);
97 cb(cur, name: "ffn_moe_out", il);
98
99 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
100
101 cur = build_cvec(cur, il);
102 cb(cur, name: "l_out", il);
103
104 // input for next layer
105 inpL = cur;
106 }
107 cur = inpL;
108
109 cur = build_norm(cur,
110 mw: model.output_norm, NULL,
111 type: LLM_NORM_RMS, il: -1);
112
113 cb(cur, name: "result_norm", il: -1);
114 res->t_embd = cur;
115
116 // lm_head
117 cur = build_lora_mm(w: model.output, cur);
118
119 cb(cur, name: "result_output", il: -1);
120 res->t_logits = cur;
121
122 ggml_build_forward_expand(cgraph: gf, tensor: cur);
123}
124