1#include "models.h"
2
3
4
5llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
6 const int64_t n_embd_head = hparams.n_embd_head_v;
7 const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
8
9 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
10
11 int sections[4];
12 std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
13
14 ggml_tensor * cur;
15 ggml_tensor * inpL;
16
17 inpL = build_inp_embd(model.tok_embd);
18
19 bool use_mrope = hparams.use_mrope();
20 if (ubatch.embd && !use_mrope) {
21 // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
22 GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
23 }
24
25 // inp_pos - contains the positions
26 ggml_tensor * inp_pos = build_inp_pos();
27
28 auto * inp_attn = build_attn_inp_kv();
29
30 ggml_tensor * inp_out_ids = build_inp_out_ids();
31
32 for (int il = 0; il < n_layer; ++il) {
33 ggml_tensor * inpSA = inpL;
34
35 // Pre-attention norm
36 cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
37 cb(cur, "attn_norm", il);
38
39 // self-attention
40 {
41 ggml_tensor * Qcur = nullptr;
42 ggml_tensor * Kcur = nullptr;
43 ggml_tensor * Vcur = nullptr;
44
45 if (model.layers[il].wqkv == nullptr) {
46 Qcur = build_lora_mm(model.layers[il].wq, cur);
47 if (model.layers[il].bq) {
48 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
49 }
50 Kcur = build_lora_mm(model.layers[il].wk, cur);
51 if (model.layers[il].bk) {
52 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
53 }
54 Vcur = build_lora_mm(model.layers[il].wv, cur);
55 if (model.layers[il].bv) {
56 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
57 }
58 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
59 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
60 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
61 } else {
62 cur = build_lora_mm(model.layers[il].wqkv, cur);
63 cb(cur, "wqkv", il);
64 if (model.layers[il].bqkv) {
65 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
66 cb(cur, "bqkv", il);
67 }
68 Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
69 0 * sizeof(float) * (n_embd));
70 Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
71 cur->nb[1], 1 * sizeof(float) * (n_embd));
72 Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
73 cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
74 }
75
76 if (use_mrope) {
77 Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
78 n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
79 ext_factor, attn_factor, beta_fast, beta_slow);
80
81 Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
82 n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
83 ext_factor, attn_factor, beta_fast, beta_slow);
84 } else {
85 // Normal RoPE
86 Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
87 rope_type, n_ctx_orig, freq_base, freq_scale,
88 ext_factor, attn_factor, beta_fast, beta_slow);
89
90 Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
91 rope_type, n_ctx_orig, freq_base, freq_scale,
92 ext_factor, attn_factor, beta_fast, beta_slow);
93 }
94
95 cb(Qcur, "Qcur", il);
96 cb(Kcur, "Kcur", il);
97 cb(Vcur, "Vcur", il);
98
99 cur = build_attn(inp_attn,
100 model.layers[il].wo, NULL,
101 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
102 }
103 if (il == n_layer - 1 && inp_out_ids) {
104 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
105 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
106 }
107 // Post-attention norm (new!)
108 cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
109 cb(cur, "post_attn_norm", il);
110
111 // Add the input (residual connection after post-attention norm)
112 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
113 cb(ffn_inp, "ffn_inp", il);
114
115 // FF
116 {
117 // Pre-MLP norm
118 cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
119 cb(cur, "ffn_norm", il);
120
121 // MLP
122 cur = build_ffn(cur,
123 model.layers[il].ffn_up, NULL, NULL,
124 NULL, NULL, NULL,
125 model.layers[il].ffn_down, NULL, NULL,
126 NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
127 cb(cur, "ffn_out", il);
128
129 // Post-MLP norm
130 cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
131 cb(cur, "post_mlp_norm", il);
132 }
133 // Add residual connection after post-MLP norm
134 inpL = ggml_add(ctx0, cur, ffn_inp);
135 cb(inpL, "l_out", il);
136 }
137 // Final norm
138 cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
139
140 cb(cur, "result_norm", -1);
141 res->t_embd = cur;
142
143 // Output projection
144 cur = build_lora_mm(model.output, cur);
145
146 cb(cur, "result_output", -1);
147 res->t_logits = cur;
148
149 ggml_build_forward_expand(gf, cur);
150}