1#include "models.h"
2
3
4
5llm_build_grovemoe::llm_build_grovemoe(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 const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;
9
10 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
11 GGML_ASSERT(n_embd_head == hparams.n_rot);
12
13 ggml_tensor * cur;
14 ggml_tensor * inpL;
15
16 inpL = build_inp_embd(tok_embd: model.tok_embd);
17
18 // inp_pos - contains the positions
19 ggml_tensor * inp_pos = build_inp_pos();
20
21 auto * inp_attn = build_attn_inp_kv();
22
23 ggml_tensor * inp_out_ids = build_inp_out_ids();
24
25 for (int il = 0; il < n_layer; ++il) {
26 ggml_tensor * inpSA = inpL;
27
28 // norm
29 cur = build_norm(cur: inpL, mw: model.layers[il].attn_norm, NULL, 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
38 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
39 cb(cur: Kcur, name: "Kcur", il);
40
41 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
42 cb(cur: Vcur, name: "Vcur", il);
43
44 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
45 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
46 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
47
48 Qcur = build_norm(cur: Qcur, mw: model.layers[il].attn_q_norm, NULL, type: LLM_NORM_RMS, il);
49 cb(cur: Qcur, name: "Qcur_normed", il);
50
51 Qcur = ggml_rope_ext(ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr, n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
52 ext_factor, attn_factor, beta_fast, beta_slow);
53
54 Kcur = build_norm(cur: Kcur, mw: model.layers[il].attn_k_norm, NULL, type: LLM_NORM_RMS, il);
55 cb(cur: Kcur, name: "Kcur_normed", il);
56
57 Kcur = ggml_rope_ext(ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr, n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
58 ext_factor, attn_factor, beta_fast, beta_slow);
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: nullptr, v_mla: nullptr, kq_scale: 1.0f / sqrtf(x: float(n_embd_head)), il);
67 }
68
69 if (il == n_layer - 1 && inp_out_ids) {
70 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
71 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
72 }
73
74 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
75 cb(cur: ffn_inp, name: "ffn_inp", il);
76
77 // MoE branch
78 cur = build_norm(cur: ffn_inp, mw: model.layers[il].ffn_norm, NULL, type: LLM_NORM_RMS, il);
79 cb(cur, name: "ffn_norm", il);
80
81 ggml_tensor * probs = build_lora_mm(w: model.layers[il].ffn_gate_inp, cur); // [n_expert, n_tokens]
82 cb(cur: probs, name: "ffn_moe_logits", il);
83
84 ggml_tensor * moe_out =
85 build_moe_ffn(cur,
86 gate_inp: nullptr,
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: true,
93 scale_w: false, w_scale: 0.0,
94 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
95 il,
96 probs_in: probs);
97 cb(cur: moe_out, name: "ffn_moe_out", il);
98 cur = moe_out;
99
100 // TODO: Only do the expert selection and weights once
101 moe_out = build_moe_ffn(cur,
102 gate_inp: nullptr,
103 up_exps: model.layers[il].ffn_up_chexps,
104 gate_exps: model.layers[il].ffn_gate_chexps,
105 down_exps: model.layers[il].ffn_down_chexps,
106 exp_probs_b: nullptr,
107 n_expert: n_chunk_expert, n_expert_used: n_expert_used > n_chunk_expert ? n_chunk_expert : n_expert_used,
108 type_op: LLM_FFN_SILU, norm_w: true,
109 scale_w: false, w_scale: 0.0,
110 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
111 il,
112 probs_in: probs);
113 cb(cur: moe_out, name: "ffn_adj_moe_out", il);
114
115 cur = ggml_add(ctx: ctx0, a: cur, b: ggml_scale(ctx: ctx0, a: moe_out, s: hparams.expert_group_scale));
116 cb(cur, name: "ffn_final_moe_out", il);
117
118 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
119
120 cur = build_cvec(cur, il);
121 cb(cur, name: "l_out", il);
122
123 // input for next layer
124 inpL = cur;
125 }
126
127 cur = inpL;
128
129 cur = build_norm(cur, mw: model.output_norm, NULL, type: LLM_NORM_RMS, il: -1);
130
131 cb(cur, name: "result_norm", il: -1);
132 res->t_embd = cur;
133
134 // lm_head
135 cur = build_lora_mm(w: model.output, cur);
136
137 cb(cur, name: "result_output", il: -1);
138 res->t_logits = cur;
139
140 ggml_build_forward_expand(cgraph: gf, tensor: cur);
141}
142