1#include "models.h"
2
3// RND1 is a Qwen3Moe AR model converted to diffusion model.
4llm_build_rnd1::llm_build_rnd1(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
5 const int64_t n_embd_head = hparams.n_embd_head_v;
6
7 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
8 GGML_ASSERT(n_embd_head == hparams.n_rot);
9
10 ggml_tensor * cur;
11 ggml_tensor * inpL;
12
13 inpL = build_inp_embd(model.tok_embd);
14
15 // inp_pos - contains the positions
16 ggml_tensor * inp_pos = build_inp_pos();
17
18 // Non-causal attention for diffusion
19 auto * inp_attn = build_attn_inp_no_cache();
20
21 ggml_tensor * inp_out_ids = build_inp_out_ids();
22
23 for (int il = 0; il < n_layer; ++il) {
24 ggml_tensor * inpSA = inpL;
25
26 // norm
27 cur = build_norm(inpL,
28 model.layers[il].attn_norm, NULL,
29 LLM_NORM_RMS, il);
30 cb(cur, "attn_norm", il);
31
32 // self_attention
33 {
34 // compute Q and K and RoPE them
35 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
36 cb(Qcur, "Qcur", il);
37
38 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
39 cb(Kcur, "Kcur", il);
40
41 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
42 cb(Vcur, "Vcur", il);
43
44 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
45 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
46 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
47
48 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
49 cb(Qcur, "Qcur_normed", il);
50
51 Qcur = ggml_rope_ext(
52 ctx0, Qcur, inp_pos, nullptr,
53 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
54 ext_factor, attn_factor, beta_fast, beta_slow
55 );
56
57 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
58 cb(Kcur, "Kcur_normed", il);
59
60 Kcur = ggml_rope_ext(
61 ctx0, Kcur, inp_pos, nullptr,
62 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
63 ext_factor, attn_factor, beta_fast, beta_slow
64 );
65
66 cb(Qcur, "Qcur", il);
67 cb(Kcur, "Kcur", il);
68 cb(Vcur, "Vcur", il);
69
70 cur = build_attn(inp_attn,
71 model.layers[il].wo, model.layers[il].bo,
72 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
73 }
74 if (il == n_layer - 1 && inp_out_ids) {
75 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
76 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
77 }
78 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
79 cb(ffn_inp, "ffn_inp", il);
80
81 // MoE branch
82 cur = build_norm(ffn_inp,
83 model.layers[il].ffn_norm, NULL,
84 LLM_NORM_RMS, il);
85 cb(cur, "ffn_norm", il);
86
87 ggml_tensor * moe_out =
88 build_moe_ffn(cur,
89 model.layers[il].ffn_gate_inp,
90 model.layers[il].ffn_up_exps,
91 model.layers[il].ffn_gate_exps,
92 model.layers[il].ffn_down_exps,
93 nullptr,
94 n_expert, n_expert_used,
95 LLM_FFN_SILU, true,
96 false, 0.0,
97 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
98 il);
99 cb(moe_out, "ffn_moe_out", il);
100 cur = moe_out;
101
102 cur = ggml_add(ctx0, cur, ffn_inp);
103
104 cur = build_cvec(cur, il);
105 cb(cur, "l_out", il);
106
107 // input for next layer
108 inpL = cur;
109 }
110 cur = inpL;
111
112 cur = build_norm(cur,
113 model.output_norm, NULL,
114 LLM_NORM_RMS, -1);
115
116 cb(cur, "result_norm", -1);
117 res->t_embd = cur;
118
119 // lm_head
120 cur = build_lora_mm(model.output, cur);
121
122 cb(cur, "result_output", -1);
123 res->t_logits = cur;
124
125 ggml_build_forward_expand(gf, cur);
126}