1
2#include "models.h"
3
4llm_build_mimo2_iswa::llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
5 ggml_tensor * cur;
6 ggml_tensor * inpL;
7
8 inpL = build_inp_embd(model.tok_embd);
9
10 ggml_tensor * inp_pos = build_inp_pos();
11 auto * inp_attn = build_attn_inp_kv_iswa();
12 ggml_tensor * inp_out_ids = build_inp_out_ids();
13
14 for (int il = 0; il < n_layer; ++il) {
15 ggml_tensor * inpSA = inpL;
16
17 uint32_t n_head_l = hparams.n_head(il);
18 uint32_t n_head_kv_l = hparams.n_head_kv(il);
19 const float freq_base_l = model.get_rope_freq_base(cparams, il);
20 const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
21
22 cur = inpL;
23
24 // self_attention
25 {
26 cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
27 cb(cur, "attn_norm", il);
28
29 // compute Q and K and RoPE them
30 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
31 cb(Qcur, "Qcur", il);
32
33 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
34 cb(Kcur, "Kcur", il);
35
36 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
37 cb(Vcur, "Vcur", il);
38
39 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);
40 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
41 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
42
43 Qcur = ggml_rope_ext(
44 ctx0, Qcur, inp_pos, nullptr,
45 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
46 ext_factor, attn_factor, beta_fast, beta_slow
47 );
48
49 Kcur = ggml_rope_ext(
50 ctx0, Kcur, inp_pos, nullptr,
51 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
52 ext_factor, attn_factor, beta_fast, beta_slow
53 );
54
55 cb(Qcur, "Qcur", il);
56 cb(Kcur, "Kcur", il);
57 cb(Vcur, "Vcur", il);
58
59 ggml_tensor * sinks = model.layers[il].attn_sinks;
60
61 cur = build_attn(inp_attn,
62 model.layers[il].wo, NULL,
63 Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), il);
64 }
65
66 if (il == n_layer - 1 && inp_out_ids) {
67 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
68 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
69 }
70
71 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
72 cb(ffn_inp, "ffn_inp", il);
73
74 cur = build_norm(ffn_inp,
75 model.layers[il].ffn_norm, NULL,
76 LLM_NORM_RMS, il);
77 cb(cur, "ffn_norm", il);
78
79 // feed-forward network
80 if (model.layers[il].ffn_gate_inp == nullptr) {
81 // dense branch
82 cur = build_ffn(cur,
83 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
84 model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
85 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
86 NULL,
87 LLM_FFN_SILU, LLM_FFN_PAR, il);
88 cb(cur, "ffn_out", il);
89 } else {
90 // MoE branch
91 cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
92 model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
93 model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, true, false,
94 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, il);
95 cb(cur, "ffn_moe_out", il);
96 }
97
98 cur = ggml_add(ctx0, cur, ffn_inp);
99
100 cur = build_cvec(cur, il);
101 cb(cur, "l_out", il);
102
103 // input for next layer
104 inpL = cur;
105 }
106
107 cur = inpL;
108
109 cur = build_norm(cur,
110 model.output_norm, NULL,
111 LLM_NORM_RMS, -1);
112
113 cb(cur, "result_norm", -1);
114 res->t_embd = cur;
115
116 // lm_head
117 cur = build_lora_mm(model.output, cur);
118
119 cb(cur, "result_output", -1);
120 res->t_logits = cur;
121
122 ggml_build_forward_expand(gf, cur);
123}