1#include "models.h"
2
3
4
5llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params) :
6 llm_graph_context_mamba(params) {
7 const int64_t n_embd_head = hparams.n_embd_head_v;
8 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
9
10 ggml_tensor * cur;
11 ggml_tensor * inpL;
12
13 inpL = build_inp_embd(model.tok_embd);
14 ggml_build_forward_expand(gf, inpL);
15
16 auto * inp = build_inp_mem_hybrid();
17
18 ggml_tensor * inp_out_ids = build_inp_out_ids();
19
20 for (int il = 0; il < n_layer; ++il) {
21 struct ggml_tensor * inpSA = inpL;
22
23 // norm
24 cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
25 cb(cur, "attn_norm", il);
26
27 if (hparams.is_recurrent(il)) {
28 // ssm layer //
29 cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);
30 } else if (hparams.n_ff(il) == 0) {
31 // attention layer //
32 cur = build_attention_layer(cur, inp->get_attn(), model, n_embd_head, il);
33 } else {
34 cur = build_ffn_layer(cur, model, il);
35 }
36
37 if (il == n_layer - 1 && inp_out_ids) {
38 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
39 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
40 }
41
42 // add residual
43 cur = ggml_add(ctx0, cur, inpSA);
44 cb(cur, "nemotron_h_block_out", il);
45
46 // input for next layer
47 inpL = cur;
48 }
49
50 cur = inpL;
51
52 cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
53
54 cb(cur, "result_norm", -1);
55 res->t_embd = cur;
56
57 // lm_head
58 cur = build_lora_mm(model.output, cur);
59 cb(cur, "result_output", -1);
60 res->t_logits = cur;
61
62 ggml_build_forward_expand(gf, cur);
63}
64
65ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor * cur,
66 llm_graph_input_attn_kv * inp_attn,
67 const llama_model & model,
68 const int64_t n_embd_head,
69 const int il) {
70 // compute Q and K
71 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
72 cb(Qcur, "Qcur", il);
73 if (model.layers[il].bq) {
74 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
75 cb(Qcur, "Qcur", il);
76 }
77
78 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
79 cb(Kcur, "Kcur", il);
80 if (model.layers[il].bk) {
81 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
82 cb(Kcur, "Kcur", il);
83 }
84
85 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
86 cb(Vcur, "Vcur", il);
87 if (model.layers[il].bv) {
88 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
89 cb(Vcur, "Vcur", il);
90 }
91
92 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens);
93 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
94 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
95
96 cb(Qcur, "Qcur", il);
97 cb(Kcur, "Kcur", il);
98 cb(Vcur, "Vcur", il);
99
100 const float kq_scale =
101 hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
102 cur = build_attn(inp_attn,
103 model.layers[il].wo, model.layers[il].bo,
104 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
105 cb(cur, "attn_out", il);
106 return cur;
107}
108
109ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) {
110 if (model.layers[il].ffn_gate_inp == nullptr) {
111 cur = build_ffn(cur,
112 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
113 NULL, NULL, NULL,
114 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
115 NULL,
116 LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
117 cb(cur, "ffn_out", il);
118 } else {
119 ggml_tensor * ffn_inp = cur;
120 ggml_tensor * moe_out =
121 build_moe_ffn(ffn_inp,
122 model.layers[il].ffn_gate_inp,
123 model.layers[il].ffn_up_exps,
124 nullptr, // no gate
125 model.layers[il].ffn_down_exps,
126 model.layers[il].ffn_exp_probs_b,
127 n_expert, n_expert_used,
128 LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
129 true, hparams.expert_weights_scale,
130 LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
131 il);
132 cb(moe_out, "ffn_moe_out", il);
133
134 ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
135 model.layers[il].ffn_up_shexp, NULL, NULL,
136 NULL /* no gate */ , NULL, NULL,
137 model.layers[il].ffn_down_shexp, NULL, NULL,
138 NULL,
139 LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
140 cb(ffn_shexp, "ffn_shexp", il);
141
142 cur = ggml_add(ctx0, moe_out, ffn_shexp);
143 cb(cur, "ffn_out", il);
144 }
145
146 cur = build_cvec(cur, il);
147 cb(cur, "l_out", il);
148
149 return cur;
150}