1#include "models.h"
2
3llm_build_step35_iswa::llm_build_step35_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4 ggml_tensor * cur;
5 ggml_tensor * inpL;
6
7 inpL = build_inp_embd(model.tok_embd);
8 ggml_tensor * inp_pos = build_inp_pos();
9 auto * inp_attn = build_attn_inp_kv_iswa();
10 ggml_tensor * inp_out_ids = build_inp_out_ids();
11
12 for (int il = 0; il < n_layer; ++il) {
13 ggml_tensor * inpSA = inpL;
14
15 const uint32_t n_head_l = hparams.n_head(il);
16 const uint32_t n_head_kv_l = hparams.n_head_kv(il);
17
18 const float freq_base_l = model.get_rope_freq_base(cparams, il);
19 const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
20
21 cur = inpL;
22
23 // dump pre-attn RMSNorm input to pinpoint layer boundary issues
24 cb(cur, "attn_norm_in", il);
25
26 // self-attention
27 {
28 cur = build_norm(cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
29 cb(cur, "attn_norm", il);
30 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
31 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
32 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
33
34 cb(Qcur, "Qcur", il);
35 cb(Kcur, "Kcur", il);
36 cb(Vcur, "Vcur", il);
37
38 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);
39 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
40 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
41
42 // Q/K per-head RMSNorm (Step35 q_norm / k_norm)
43 if (model.layers[il].attn_q_norm) {
44 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
45 cb(Qcur, "Qcur_normed", il);
46 }
47 if (model.layers[il].attn_k_norm) {
48 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
49 cb(Kcur, "Kcur_normed", il);
50 }
51
52 // RoPE (partial rotary factors per layer)
53 const bool is_swa = hparams.is_swa(il);
54 ggml_tensor * rope_factors = is_swa ? nullptr : model.get_rope_factors(cparams, il);
55 const int64_t n_rot_l = is_swa ? hparams.n_rot : (hparams.n_rot / 2);
56 Qcur = ggml_rope_ext(
57 ctx0, Qcur, inp_pos, rope_factors,
58 n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
59 ext_factor, attn_factor, beta_fast, beta_slow
60 );
61 Kcur = ggml_rope_ext(
62 ctx0, Kcur, inp_pos, rope_factors,
63 n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
64 ext_factor, attn_factor, beta_fast, beta_slow
65 );
66 cb(Qcur, "Qcur_pos", il);
67 cb(Kcur, "Kcur_pos", il);
68
69 const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k));
70 ggml_tensor * attn_out = build_attn(inp_attn,
71 nullptr, nullptr,
72 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
73 cb(attn_out, "attn_out", il);
74 // head-wise attention gate: sigmoid(g_proj(x)) in torch
75 if (model.layers[il].wqkv_gate) {
76 ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, cur); // [n_head_l, n_tokens]
77 cb(gate, "attn_gate", il);
78
79 gate = ggml_sigmoid(ctx0, gate);
80 cb(gate, "attn_gate_sigmoid", il);
81
82 // reshape + broadcast to [n_embd_head_v, n_head_l, n_tokens]
83 ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, attn_out, n_embd_head_v, n_head_l, n_tokens);
84 ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate, 1, n_head_l, n_tokens);
85 cb(gate_3d, "attn_gate_3d", il);
86
87 attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);
88 cb(attn_3d, "attn_gated_3d", il);
89
90 attn_out = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v * n_head_l, n_tokens);
91 cb(attn_out, "attn_gated", il);
92 }
93
94 // output projection
95 cur = build_lora_mm(model.layers[il].wo, attn_out);
96 cb(cur, "attn_proj", il);
97 }
98
99 if (il == n_layer - 1 && inp_out_ids) {
100 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
101 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
102 }
103
104 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
105 cb(ffn_inp, "ffn_inp", il);
106
107 cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
108 cb(cur, "ffn_norm", il);
109
110 // feed-forward
111 if (model.layers[il].ffn_gate_inp == nullptr) {
112 // dense MLP
113 cur = build_ffn(cur,
114 model.layers[il].ffn_up, model.layers[il].ffn_up_b, nullptr,
115 model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, nullptr,
116 model.layers[il].ffn_down, model.layers[il].ffn_down_b, nullptr,
117 nullptr,
118 LLM_FFN_SILU, LLM_FFN_PAR, il);
119 cb(cur, "ffn_out", il);
120 } else {
121 // MoE routed experts
122 const bool norm_w = hparams.expert_weights_norm;
123 const float w_scale = hparams.expert_weights_scale;
124 const bool scale_w = w_scale != 0.0f;
125 ggml_tensor * moe_out = build_moe_ffn(cur,
126 model.layers[il].ffn_gate_inp,
127 model.layers[il].ffn_up_exps,
128 model.layers[il].ffn_gate_exps,
129 model.layers[il].ffn_down_exps,
130 model.layers[il].ffn_exp_probs_b,
131 n_expert, n_expert_used,
132 LLM_FFN_SILU,
133 norm_w, scale_w, w_scale,
134 (llama_expert_gating_func_type) hparams.expert_gating_func,
135 il);
136 cb(moe_out, "ffn_moe_out", il);
137
138 // shared expert MLP (always added on MoE layers in Step35)
139 ggml_tensor * sh_out = build_ffn(cur,
140 model.layers[il].ffn_up_shexp, nullptr, nullptr,
141 model.layers[il].ffn_gate_shexp, nullptr, nullptr,
142 model.layers[il].ffn_down_shexp, nullptr, nullptr,
143 nullptr,
144 LLM_FFN_SILU, LLM_FFN_PAR, il);
145 cb(sh_out, "ffn_shared_out", il);
146
147 cur = ggml_add(ctx0, moe_out, sh_out);
148 cb(cur, "ffn_out", il);
149 }
150 cur = ggml_add(ctx0, cur, ffn_inp);
151 cur = build_cvec(cur, il);
152 cb(cur, "l_out", il);
153
154 inpL = cur;
155 }
156
157 cur = inpL;
158
159 cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
160 cb(cur, "result_norm", -1);
161 res->t_embd = cur;
162
163 cur = build_lora_mm(model.output, cur);
164 cb(cur, "result_output", -1);
165 res->t_logits = cur;
166
167 ggml_build_forward_expand(gf, cur);
168}