char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Yaniss Amazouzmodel: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (#25444)c61b98b87

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
24.0 KiB476 linesraw
1#include "models.h"
2
3#include "llama-kv-cache.h"
4#include "llama-kv-cache-dsa.h"
5
6// note: code adapted from deepseek32.cpp (DSA indexer + absorbed MLA) and step35.cpp (head-wise output gate)
7
8void llama_model_dots3note::load_arch_hparams(llama_model_loader & ml) {
9    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
10    hparams.f_norm_eps = 1e-6;  // eps for the indexer k_norm layer norm
11
12    // MoE parameters
13    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
14    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
15    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,  hparams.n_layer_dense_lead);
16    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
17    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);
18    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,         hparams.expert_gating_func);
19
20    // MLA parameters of the full-attention layers
21    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,      hparams.n_lora_q);
22    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
23    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl);
24    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl);
25
26    // MLA parameters of the sliding-window layers
27    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK_SWA,     hparams.n_lora_kv_swa);
28    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA_SWA,   hparams.n_embd_head_k_mla_swa);
29    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA_SWA, hparams.n_embd_head_v_mla_swa);
30
31    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
32    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
33    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,       hparams.rope_freq_base_train_swa);
34    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);
35
36    // DSA parameters
37    ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head);
38    ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size);
39    ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K,      hparams.indexer_top_k);
40    ml.get_arr(LLM_KV_ATTENTION_INDEXER_TYPES, hparams.is_indexer_full_impl);
41
42    switch (hparams.n_layer()) {
43        case 46: type = LLM_TYPE_288B_A19B; break;
44        default: type = LLM_TYPE_UNKNOWN;
45    }
46}
47
48void llama_model_dots3note::load_arch_tensors(llama_model_loader & ml) {
49    LLAMA_LOAD_LOCALS;
50    GGML_UNUSED(ml);
51
52    if (!hparams.is_mla()) {
53        throw std::runtime_error("DOTS3NOTE architecture requires MLA");
54    }
55
56    const int64_t n_embd_head_qk_rope = hparams.n_rot();
57
58    const int64_t q_lora_rank     = hparams.n_lora_q;
59    const int64_t n_ff_exp        = hparams.n_ff_exp();
60    const int64_t n_expert_shared = hparams.n_expert_shared;
61
62    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
63
64    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
65    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
66    if (!output) {
67        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
68    }
69
70    for (int i = 0; i < n_layer_all; ++i) {
71        auto & layer = layers[i];
72
73        const bool is_mtp = i >= n_layer;
74        // the NextN/MTP block uses the sliding-attention geometry
75        const bool is_swa = is_mtp || hparams.is_swa(i);
76
77        // MTP tensors are preserved in the GGUF but there is no MTP graph yet
78        const int flags = is_mtp ? TENSOR_SKIP | TENSOR_NOT_REQUIRED : 0;
79
80        const int64_t n_head_l = hparams.n_head(i);
81
82        const int64_t kv_lora_rank       = is_swa ? hparams.n_lora_kv_swa         : hparams.n_lora_kv;
83        const int64_t n_embd_head_k_mla  = is_swa ? hparams.n_embd_head_k_mla_swa : hparams.n_embd_head_k_mla();
84        const int64_t n_embd_head_v_mla  = is_swa ? hparams.n_embd_head_v_mla_swa : hparams.n_embd_head_v_mla();
85        const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
86
87        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, flags);
88        layer.attn_q_a_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM,  "weight", i), {q_lora_rank}, flags);
89        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);
90        // norm applied on the shared rope key before rope
91        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_qk_rope}, flags);
92
93        layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);
94        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head_l * n_embd_head_k_mla}, flags);
95
96        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);
97
98        layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head_l}, flags);
99        layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head_l}, flags);
100
101        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head_l * n_embd_head_v_mla, n_embd}, flags);
102
103        // head-wise sigmoid output gate
104        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, flags);
105
106        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
107
108        // DSA indexer
109        if (!is_mtp && hparams.is_indexer_full(i)) {
110            layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags);
111            layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags);
112            layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags);
113            layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags);
114            layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
115        }
116
117        if (is_mtp || i < (int) hparams.n_layer_dense_lead) {
118            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
119            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
120            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
121        } else {
122            if (n_expert == 0 || n_expert_used == 0) {
123                throw std::runtime_error("n_expert and n_expert_used must be > 0");
124            }
125
126            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, flags);
127            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, flags);
128
129            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
130            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
131            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
132
133            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
134            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, flags);
135            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
136        }
137
138        if (is_mtp) {
139            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), { 2 * n_embd, n_embd }, flags);
140            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), { n_embd }, flags);
141            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), { n_embd }, flags);
142            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), { n_embd, n_vocab }, flags);
143            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags);
144        }
145    }
146}
147
148std::unique_ptr<llm_graph_context> llama_model_dots3note::build_arch_graph(const llm_graph_params & params) const {
149    return std::make_unique<graph>(*this, params);
150}
151
152llama_model_dots3note::graph::graph(const llama_model & model, const llm_graph_params & params) :
153    llm_graph_context(params) {
154    GGML_ASSERT(hparams.is_mla());
155
156    const int64_t n_embd_head_qk_rope = hparams.n_rot();
157
158    const int64_t n_indexer_head      = hparams.indexer_n_head;
159    const int64_t n_embd_indexer_head = hparams.indexer_head_size;
160    const uint32_t n_indexer_top_k = hparams.indexer_top_k;
161
162    // the indexer head layout is [rope | nope]
163    GGML_ASSERT(hparams.n_rot() <= n_embd_indexer_head);
164
165    ggml_tensor * cur;
166    ggml_tensor * inpL;
167
168    inpL = build_inp_embd(model.tok_embd);
169
170    ggml_tensor * inp_pos = build_inp_pos();
171
172    llm_graph_input_attn_k_dsa_iswa * inp_attn = build_attn_inp_k_dsa_iswa();
173
174    ggml_tensor * inp_out_ids = build_inp_out_ids();
175
176    for (int il = 0; il < n_layer; ++il) {
177        ggml_tensor * inpSA = inpL;
178
179        const bool is_swa = hparams.is_swa(il);
180
181        const int64_t n_head_l = hparams.n_head(il);
182
183        const int64_t kv_lora_rank        = is_swa ? hparams.n_lora_kv_swa         : hparams.n_lora_kv;
184        const int64_t n_embd_head_k_mla   = is_swa ? hparams.n_embd_head_k_mla_swa : hparams.n_embd_head_k_mla();
185        const int64_t n_embd_head_v_mla   = is_swa ? hparams.n_embd_head_v_mla_swa : hparams.n_embd_head_v_mla();
186        const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
187
188        const float kq_scale    = 1.0f/sqrtf(float(n_embd_head_k_mla));
189        const float freq_base_l = model.get_rope_freq_base(cparams, il);
190
191        // norm
192        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
193        cb(cur, "attn_norm", il);
194
195        // self_attention
196        {
197            ggml_tensor * attn_inp = cur;
198
199            ggml_tensor * qr = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
200            cb(qr, "qr", il);
201
202            qr = build_norm(qr, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
203            cb(qr, "qr", il);
204
205            ggml_tensor * top_k = nullptr;
206
207            // lightning indexer (full-attention layers only)
208            if (!is_swa) {
209                ggml_tensor * indexer_q = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_q_b, qr);
210                cb(indexer_q, "indexer_q", il);
211
212                // {n_embd_indexer_head, n_indexer_head, n_tokens}
213                indexer_q = ggml_reshape_3d(ctx0, indexer_q, n_embd_indexer_head, n_indexer_head, n_tokens);
214                indexer_q = ggml_rope_ext(ctx0, indexer_q, inp_pos, nullptr, n_rot,
215                                     LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
216                                     ext_factor, attn_factor, beta_fast, beta_slow);
217                cb(indexer_q, "indexer_q", il);
218
219                ggml_tensor * indexer_k = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_k, cur);
220                cb(indexer_k, "indexer_k", il);
221
222                indexer_k = build_norm(indexer_k, model.layers[il].indexer_k_norm, model.layers[il].indexer_k_norm_b, LLM_NORM, il);
223                cb(indexer_k, "indexer_k", il);
224
225                // {n_embd_indexer_head, 1, n_tokens}
226                indexer_k = ggml_reshape_3d(ctx0, indexer_k, n_embd_indexer_head, 1, n_tokens);
227                indexer_k = ggml_rope_ext(ctx0, indexer_k, inp_pos, nullptr, n_rot,
228                                     LLAMA_ROPE_TYPE_NEOX, n_ctx_orig, freq_base, freq_scale,
229                                     ext_factor, attn_factor, beta_fast, beta_slow);
230                cb(indexer_k, "indexer_k", il);
231
232                // perform Hadamard transform on indexer q and k
233                indexer_q = ggml_mul_mat(ctx0, inp_attn->get_dsa()->self_k_rot_lid, indexer_q);
234                cb(indexer_q, "indexer_q", il);
235                indexer_k = ggml_mul_mat(ctx0, inp_attn->get_dsa()->self_k_rot_lid, indexer_k);
236                cb(indexer_k, "indexer_k", il);
237
238                // store indexer keys to KV cache
239                const auto * mctx_lid = inp_attn->get_dsa()->mctx->get_lid();
240                const auto & k_idxs_lid = inp_attn->get_dsa()->get_k_idxs_lid();
241                ggml_build_forward_expand(gf, mctx_lid->cpy_k(ctx0, indexer_k, k_idxs_lid, il));
242
243                ggml_tensor * indexer_weights = ggml_mul_mat(ctx0, model.layers[il].indexer_proj, cur);
244                cb(indexer_weights, "indexer_weights", il);
245
246                indexer_k = mctx_lid->get_k(ctx0, il);
247
248                // split the batch into streams if needed
249                const auto n_stream = indexer_k->ne[3];
250                indexer_q = ggml_view_4d(ctx0, indexer_q, indexer_q->ne[0], indexer_q->ne[1], indexer_q->ne[2]/n_stream, n_stream, indexer_q->nb[1], indexer_q->nb[2], indexer_q->nb[3]/n_stream, 0);
251                indexer_weights = ggml_view_4d(ctx0, indexer_weights, indexer_weights->ne[0], indexer_weights->ne[1]/n_stream, indexer_weights->ne[2], n_stream, indexer_weights->nb[1], indexer_weights->nb[2]/n_stream, indexer_weights->nb[3]/n_stream, 0);
252
253                // pre-scale weights to avoid scaling operations on huge indexer_score tensor
254                indexer_weights = ggml_scale(ctx0, indexer_weights, 1.0f / sqrtf(float(n_embd_indexer_head * n_indexer_head)));
255                cb(indexer_weights, "indexer_weights", il);
256
257                ggml_tensor * indexer_score = nullptr;
258                if (cparams.fused_lid) {
259                    indexer_score = ggml_lightning_indexer(ctx0, indexer_q, indexer_k, indexer_weights, inp_attn->get_dsa()->get_kq_mask_lid());
260                    cb(indexer_score, "indexer_score", il);
261                    res->add_fused_node({LLM_FUSED_OP_LIGHTNING_INDEXER, indexer_score, il});
262                } else {
263                    indexer_q = ggml_permute(ctx0, indexer_q, 0, 2, 1, 3);
264                    cb(indexer_q, "indexer_q", il);
265                    indexer_k = ggml_permute(ctx0, indexer_k, 0, 2, 1, 3);
266                    cb(indexer_k, "indexer_k", il);
267
268                    ggml_tensor * indexer_kq = ggml_mul_mat(ctx0, indexer_k, indexer_q);
269                    cb(indexer_kq, "indexer_kq", il);
270
271                    // ReLU requires contiguous tensors
272                    indexer_kq = ggml_cont(ctx0, ggml_permute(ctx0, indexer_kq, 2, 1, 0, 3));
273                    cb(indexer_kq, "indexer_kq", il);
274
275                    indexer_score = ggml_relu(ctx0, indexer_kq);
276                    cb(indexer_score, "indexer_score", il);
277
278                    indexer_score = ggml_mul(ctx0, indexer_score, indexer_weights);
279                    cb(indexer_score, "indexer_score", il);
280
281                    // sum by q n_indexer_head dimension
282                    indexer_score = ggml_sum_rows(ctx0, indexer_score);
283                    cb(indexer_score, "indexer_score", il);
284
285                    // permute result to match KQ mask
286                    indexer_score = ggml_cont(ctx0, ggml_permute(ctx0, indexer_score, 2, 1, 0, 3));
287                    cb(indexer_score, "indexer_score", il);
288
289                    ggml_tensor * indexer_kq_mask = inp_attn->get_dsa()->get_kq_mask_lid();
290                    indexer_score = ggml_add(ctx0, indexer_score, indexer_kq_mask);
291                    cb(indexer_score, "indexer_score", il);
292                }
293
294                // get indices of top k indexer scores
295                uint32_t n_top_k = indexer_score->ne[0] < n_indexer_top_k ? indexer_score->ne[0] : n_indexer_top_k;
296                top_k = ggml_cont(ctx0, ggml_top_k(ctx0, indexer_score, n_top_k));
297                cb(top_k, "top_k", il);
298            }
299
300            ggml_tensor * q = ggml_mul_mat(ctx0, model.layers[il].wq_b, qr);
301            cb(q, "q", il);
302
303            // split into {n_embd_head_qk_nope, n_head_l, n_tokens}
304            ggml_tensor * q_nope =
305                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head_l, n_tokens, ggml_row_size(q->type, n_embd_head_k_mla),
306                             ggml_row_size(q->type, n_embd_head_k_mla) * n_head_l, 0);
307            cb(q_nope, "q_nope", il);
308
309            // and {n_embd_head_qk_rope, n_head_l, n_tokens}
310            ggml_tensor * q_pe = ggml_view_3d(
311                ctx0, q, n_embd_head_qk_rope, n_head_l, n_tokens, ggml_row_size(q->type, n_embd_head_k_mla),
312                ggml_row_size(q->type, n_embd_head_k_mla) * n_head_l, ggml_row_size(q->type, n_embd_head_qk_nope));
313            cb(q_pe, "q_pe", il);
314
315            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
316            cb(kv_cmpr_pe, "kv_cmpr_pe", il);
317
318            // split into {kv_lora_rank, n_tokens}
319            ggml_tensor * kv_cmpr =
320                ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
321                             ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
322            cb(kv_cmpr, "kv_cmpr", il);
323
324            // and {n_embd_head_qk_rope, 1, n_tokens}
325            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
326                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
327                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
328                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
329            cb(k_pe, "k_pe", il);
330
331            // norm on the shared rope key, applied before rope
332            k_pe = build_norm(k_pe, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
333            cb(k_pe, "k_pe", il);
334
335            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale,
336                                 ext_factor, attn_factor, beta_fast, beta_slow);
337            cb(q_pe, "q_pe", il);
338
339            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale,
340                                 ext_factor, attn_factor, beta_fast, beta_slow);
341            cb(k_pe, "k_pe", il);
342
343            kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
344            cb(kv_cmpr, "kv_cmpr", il);
345
346            // MLA attention with the absorption optimization
347            {
348                // {n_embd_head_qk_nope, n_tokens, n_head_l}
349                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
350                cb(q_nope, "q_nope_perm", il);
351
352                // {n_embd_head_qk_nope, kv_lora_rank, n_head_l} x {n_embd_head_qk_nope, n_tokens, n_head_l}
353                ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
354                cb(q_nope_absorbed, "q_nope_absorbed", il);
355
356                // {kv_lora_rank, n_head_l, n_tokens}
357                q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
358                cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
359
360                // {n_embd_head_qk_rope + kv_lora_rank, n_head_l, n_tokens}
361                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
362                cb(Qcur, "Qcur", il);
363
364                kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
365                cb(kv_cmpr, "kv_cmpr_reshape", il);
366
367                // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
368                ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
369                cb(Kcur, "Kcur", il);
370
371                // {kv_lora_rank, 1, n_tokens}
372                ggml_tensor * Vcur = kv_cmpr;
373                cb(Vcur, "Vcur", il);
374
375                // apply the head-wise output gate before o_proj, so wo stays out of build_attn
376                if (is_swa) {
377                    cur = build_attn(inp_attn->get_swa(),
378                            nullptr, nullptr, nullptr,
379                            Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
380                } else {
381                    cur = build_attn(inp_attn->get_dsa(),
382                            nullptr, nullptr, nullptr,
383                            Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, top_k, kq_scale, il);
384                }
385                cb(cur, "attn_out", il);
386
387                ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
388                cb(gate, "attn_gate", il);
389
390                gate = ggml_sigmoid(ctx0, gate);
391                cb(gate, "attn_gate_sigmoid", il);
392
393                // broadcast the per-head gate over the head dimension
394                ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, cur, n_embd_head_v_mla, n_head_l, n_tokens);
395                ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate,                1, n_head_l, n_tokens);
396                attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);
397                cb(attn_3d, "attn_gated", il);
398
399                cur = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v_mla * n_head_l, n_tokens);
400
401                cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
402                cb(cur, "attn_output", il);
403            }
404        }
405
406        if (il == n_layer - 1 && inp_out_ids) {
407            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
408            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
409        }
410
411        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
412        cb(ffn_inp, "ffn_inp", il);
413
414        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
415        cb(cur, "ffn_norm", il);
416
417        if ((uint32_t) il < hparams.n_layer_dense_lead) {
418            cur = build_ffn(cur,
419                model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
420                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
421                model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
422                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
423            cb(cur, "ffn_out", il);
424        } else {
425            ggml_tensor * moe_out = build_moe_ffn(cur,
426                model.layers[il].ffn_gate_inp,
427                model.layers[il].ffn_up_exps,
428                model.layers[il].ffn_gate_exps,
429                model.layers[il].ffn_down_exps,
430                model.layers[il].ffn_exp_probs_b,
431                n_expert, n_expert_used,
432                LLM_FFN_SILU, hparams.expert_weights_norm,
433                hparams.expert_weights_scale,
434                (llama_expert_gating_func_type) hparams.expert_gating_func,
435                il,
436                nullptr,
437                model.layers[il].ffn_gate_up_exps,
438                model.layers[il].ffn_up_exps_s,
439                model.layers[il].ffn_gate_exps_s,
440                model.layers[il].ffn_down_exps_s);
441            cb(moe_out, "ffn_moe_out", il);
442
443            ggml_tensor * ffn_shexp =
444                build_ffn(cur,
445                    model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s,
446                    model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,
447                    model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,
448                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
449            cb(ffn_shexp, "ffn_shexp", il);
450
451            cur = ggml_add(ctx0, moe_out, ffn_shexp);
452            cb(cur, "ffn_out", il);
453        }
454
455        cur = ggml_add(ctx0, cur, ffn_inp);
456
457        cur = build_cvec(cur, il);
458        cb(cur, "l_out", il);
459
460        inpL = cur;
461    }
462
463    cur = inpL;
464
465    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
466
467    cb(cur, "result_norm", -1);
468    res->t_embd = cur;
469
470    cur = ggml_mul_mat(ctx0, model.output, cur);
471
472    cb(cur, "result_output", -1);
473    res->t_logits = cur;
474
475    ggml_build_forward_expand(gf, cur);
476}