1#include "ggml.h"
2#include "models.h"
3
4#define CHUNK_SIZE 64
5
6llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_graph_params & params) :
7 llm_graph_context_mamba(params), model(model) {
8 ggml_tensor * cur;
9 ggml_tensor * inpL;
10
11 inpL = build_inp_embd(model.tok_embd);
12 cb(inpL, "model.embed_tokens", -1);
13
14 auto * inp = build_inp_mem_hybrid();
15
16 ggml_tensor * inp_pos = build_inp_pos();
17 ggml_tensor * inp_out_ids = build_inp_out_ids();
18
19 ggml_tensor * causal_mask =
20 ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
21 GGML_TRI_TYPE_LOWER);
22
23 ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
24 ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
25
26 ggml_build_forward_expand(gf, causal_mask);
27 ggml_build_forward_expand(gf, identity);
28 ggml_build_forward_expand(gf, diag_mask);
29
30 for (int il = 0; il < n_layer; ++il) {
31 ggml_tensor * inpSA = inpL;
32
33 cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
34 cb(cur, "attn_norm", il);
35
36 // Determine layer type and build appropriate attention mechanism
37 if (hparams.is_recurrent(il)) {
38 // Linear attention layer (gated delta net)
39 cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
40 } else {
41 // Full attention layer
42 cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
43 }
44
45 if (il == n_layer - 1 && inp_out_ids) {
46 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
47 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
48 }
49
50 // Residual connection
51 cur = ggml_add(ctx0, cur, inpSA);
52 cb(cur, "attn_residual", il);
53
54 // Save the tensor before post-attention norm for residual connection
55 ggml_tensor * ffn_residual = cur;
56
57 // Post-attention norm
58 ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);
59 cb(attn_post_norm, "attn_post_norm", il);
60
61 // FFN layer (MoE or dense) - without residual connection
62 cur = build_layer_ffn(attn_post_norm, il);
63 cb(cur, "ffn_out", il);
64
65 // Residual connection for FFN - add to the tensor from before post_attention_layernorm
66 cur = ggml_add(ctx0, cur, ffn_residual);
67 cb(cur, "post_moe", il);
68
69 // Input for next layer
70 inpL = cur;
71 }
72 cur = inpL;
73
74 // Final norm
75 cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
76
77 cb(cur, "result_norm", -1);
78 res->t_embd = cur;
79
80 // LM head
81 cur = build_lora_mm(model.output, cur);
82
83 cb(cur, "result_output", -1);
84 res->t_logits = cur;
85
86 ggml_build_forward_expand(gf, cur);
87}
88
89// utility to get one slice from the third dimension
90// input dim: [x, y, c, b]
91// output dim: [x, y, 1, b]
92static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
93 return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
94 t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
95}
96
97std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chunking(
98 ggml_tensor * q,
99 ggml_tensor * k,
100 ggml_tensor * v,
101 ggml_tensor * g,
102 ggml_tensor * beta,
103 ggml_tensor * state,
104 ggml_tensor * causal_mask,
105 ggml_tensor * identity,
106 ggml_tensor * diag_mask,
107 int il) {
108 const int64_t S_k = q->ne[0];
109 const int64_t H_k = q->ne[1];
110 const int64_t n_tokens = q->ne[2];
111 const int64_t n_seqs = q->ne[3];
112
113 const int64_t S_v = v->ne[0];
114 const int64_t H_v = v->ne[1];
115
116 GGML_ASSERT(v->ne[2] == n_tokens);
117 GGML_ASSERT(k->ne[2] == n_tokens);
118 GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
119 GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
120 GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
121
122 GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
123 GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
124
125 GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
126
127 const float eps_norm = hparams.f_norm_rms_eps;
128
129 q = ggml_l2_norm(ctx0, q, eps_norm);
130 k = ggml_l2_norm(ctx0, k, eps_norm);
131
132 const float scale = 1.0f / sqrtf(S_v);
133
134 q = ggml_scale(ctx0, q, scale);
135
136 beta = ggml_sigmoid(ctx0, beta);
137
138 cb(q, "q_in", il);
139 cb(k, "k_in", il);
140 cb(v, "v_in", il);
141 cb(beta, "beta_in", il);
142 cb(g, "g_in", il);
143
144 q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
145 k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
146 v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
147 g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
148
149 beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
150 state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
151
152 cb(q, "q_perm", il);
153 cb(k, "k_perm", il);
154 cb(v, "v_perm", il);
155 cb(beta, "beta_perm", il);
156 cb(g, "g_perm", il);
157 cb(state, "state_in", il);
158
159 GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
160 GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
161 GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
162 GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
163
164 // Do padding
165 const int64_t chunk_size = CHUNK_SIZE;
166
167 const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
168 const int64_t n_chunks = (n_tokens + pad) / chunk_size;
169
170 q = ggml_pad(ctx0, q, 0, pad, 0, 0);
171 k = ggml_pad(ctx0, k, 0, pad, 0, 0);
172 v = ggml_pad(ctx0, v, 0, pad, 0, 0);
173 g = ggml_pad(ctx0, g, pad, 0, 0, 0);
174 beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
175
176 cb(q, "q_pad", il);
177 cb(k, "k_pad", il);
178 cb(v, "v_pad", il);
179 cb(beta, "beta_pad", il);
180 cb(g, "g_pad", il);
181
182 ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
183 ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
184
185 cb(v_beta, "v_beta", il);
186 cb(k_beta, "k_beta", il);
187
188 q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs);
189 k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs);
190 k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
191 v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs);
192 v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
193
194 g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
195 beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
196
197 ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
198 cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
199
200 ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
201 ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
202
203 ggml_tensor * gcs_j_broadcast =
204 ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
205
206 ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
207 cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
208
209 decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
210 decay_mask = ggml_exp(ctx0, decay_mask);
211 decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
212
213 ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
214
215 ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
216 ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
217 cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
218
219 ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
220 ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
221
222 ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
223 attn = ggml_mul(ctx0, lin_solve, causal_mask);
224 attn = ggml_add(ctx0, attn, identity);
225 cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
226
227 v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
228
229 ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
230 ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t);
231
232 ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
233 cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
234
235 ggml_tensor * k_cumdecay =
236 ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
237 cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
238
239 ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
240 attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
241 attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
242 cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
243
244
245 // vectorized calculation of key_gdiff
246 // improved from the chunked version:
247 // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
248 // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
249 // key_gdiff = key * g_diff.unsqueeze(-1)
250 // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
251 // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
252
253 // get last element in g_cumsum along chunk_size dimension (ne0)
254 // example: [[x, y, z, ..., last], ...] -> [[last], ...]
255 ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
256 g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
257 (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
258 g_last = ggml_cont(ctx0, g_last);
259 cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
260
261 ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
262 cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
263
264 ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
265 cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
266
267 ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
268 ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
269 1, chunk_size, n_chunks, g_diff_exp->ne[3]);
270
271 ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
272 cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
273
274 ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
275 cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
276
277
278 // state to be updated per chunk
279 ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
280 cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
281
282 // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
283 ggml_tensor * core_attn_out = nullptr;
284
285 for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
286 // shape: (S_k, chunk_size, 1, H_k * n_seqs)
287 ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
288
289 // shape: (S_v, chunk_size, 1, H_v * n_seqs)
290 ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
291
292 // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
293 ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
294
295 // shape: (chunk_size, 1, H_v * n_seqs)
296 ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
297
298 // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
299 // replaced by precomputed attn_kq
300 ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
301 cb(attn_chunk, "attn_chunk", il);
302
303 ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
304
305 // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
306 ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
307 cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
308
309 // v_new = v_i - v_prime
310 ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
311 ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
312 cb(v_new, "v_new_chunk", il);
313
314 // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
315 ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk);
316 ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
317 cb(attn_inter, "attn_inter_chunk", il);
318
319 // core_attn_out[:, :, i] = attn_inter + attn @ v_new
320 ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
321 cb(v_attn, "v_attn_chunk", il);
322
323 ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
324 cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
325
326 core_attn_out = core_attn_out == nullptr
327 ? core_attn_out_chunk
328 : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
329
330 // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
331 ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
332 //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
333 ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
334
335 // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
336 ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
337 new_state = ggml_add(ctx0,
338 ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
339 ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
340 }
341
342 // truncate padded tokens
343 ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
344 S_v, n_tokens, H_v, n_seqs,
345 ggml_row_size(core_attn_out->type, S_v),
346 ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
347 ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
348 output_tokens = ggml_cont(ctx0, output_tokens);
349 cb(output_tokens, "output_tokens", il);
350
351 // permute back to (S_v, H_v, n_tokens, n_seqs)
352 output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
353 output_tokens = ggml_cont(ctx0, output_tokens);
354
355 return {output_tokens, new_state};
356}
357
358std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_autoregressive(
359 ggml_tensor * q,
360 ggml_tensor * k,
361 ggml_tensor * v,
362 ggml_tensor * g,
363 ggml_tensor * beta,
364 ggml_tensor * state,
365 int il) {
366 const int64_t S_k = q->ne[0];
367 const int64_t H_k = q->ne[1];
368 const int64_t n_tokens = q->ne[2];
369 const int64_t n_seqs = q->ne[3];
370
371 const int64_t S_v = v->ne[0];
372 const int64_t H_v = v->ne[1];
373
374 GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing
375 GGML_ASSERT(v->ne[2] == n_tokens);
376 GGML_ASSERT(k->ne[2] == n_tokens);
377 GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
378 GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
379 GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
380
381 GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
382 GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
383
384 GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
385
386 const float eps_norm = hparams.f_norm_rms_eps;
387
388 q = ggml_l2_norm(ctx0, q, eps_norm);
389 k = ggml_l2_norm(ctx0, k, eps_norm);
390
391 const float scale = 1.0f / sqrtf(S_v);
392
393 q = ggml_scale(ctx0, q, scale);
394 beta = ggml_sigmoid(ctx0, beta);
395
396 cb(q, "q_in", il);
397 cb(k, "k_in", il);
398 cb(v, "v_in", il);
399 cb(beta, "beta_in", il);
400 cb(g, "g_in", il);
401
402 state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
403
404 ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
405 ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
406
407 // Apply exponential to g_t
408 g_t = ggml_exp(ctx0, g_t);
409
410 // Apply the gated delta rule for the single timestep
411 // last_recurrent_state = last_recurrent_state * g_t
412 state = ggml_mul(ctx0, state, g_t);
413
414 // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
415 ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
416 ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed);
417 // we need to sum over dim=-2, so we transpose, sum, then transpose again
418 kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
419
420 // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
421 ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
422 // delta = (v_t - kv_mem) * beta_t
423 ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs]
424 ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t);
425
426 // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
427 ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
428 state = ggml_add(ctx0, state, k_t_delta);
429
430 // Compute the attention output
431 // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
432 ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t
433 ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed);
434 // again, since it's over dim = -2, transpose, sum, transpose back
435 ggml_tensor * core_attn_out =
436 ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
437
438 // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
439 cb(core_attn_out, "output_tokens", il);
440 cb(state, "new_state", il);
441
442 return {core_attn_out, state};
443}
444
445ggml_tensor * llm_build_qwen3next::build_norm_gated(
446 ggml_tensor * input,
447 ggml_tensor * weights,
448 ggml_tensor * gate,
449 int layer) {
450 ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer);
451 ggml_tensor * gated_silu = ggml_silu(ctx0, gate);
452
453 return ggml_mul(ctx0, normalized, gated_silu);
454}
455
456ggml_tensor * llm_build_qwen3next::build_layer_attn(
457 llm_graph_input_attn_kv * inp,
458 ggml_tensor * cur,
459 ggml_tensor * inp_pos,
460 int il) {
461 const int64_t n_embd_head = hparams.n_embd_head_v;
462 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
463
464 // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
465
466 // Qwen3Next uses a single Q projection that outputs query + gate
467 ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur);
468 cb(Qcur_full, "Qcur_full", il);
469
470 Qcur_full = ggml_reshape_4d(ctx0, Qcur_full, n_embd_head * 2, n_head, n_tokens, 1);
471
472 // Split Q projection into query and gate
473 // The split should be along dimension 0 (the feature dimension)
474 ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
475 Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);
476 ggml_tensor * gate =
477 ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,
478 Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));
479 cb(Qcur, "Qcur", il);
480 cb(gate, "gate", il);
481
482 // Now reshape Qcur to [n_embd_head, n_head, n_tokens] for multi-head attention
483 Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
484 cb(Qcur, "Qcur_reshaped", il);
485
486 // Apply Q normalization
487 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
488 cb(Qcur, "Qcur_normed", il);
489
490 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
491 cb(Kcur, "Kcur", il);
492
493 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
494 cb(Vcur, "Vcur", il);
495
496 // Apply K normalization
497 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
498 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
499 cb(Kcur, "Kcur_normed", il);
500
501 // Reshape gate to [n_embd, n_tokens] for the sigmoid gating (flatten the heads)
502 gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
503 cb(gate, "gate_reshaped", il);
504
505 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
506
507 // Apply RoPE
508 Qcur = ggml_rope_ext(
509 ctx0, Qcur, inp_pos, nullptr,
510 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
511 ext_factor, attn_factor, beta_fast, beta_slow);
512
513 Kcur = ggml_rope_ext(
514 ctx0, Kcur, inp_pos, nullptr,
515 n_rot, rope_type, n_ctx_orig, freq_base,
516 freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
517
518 cb(Qcur, "Qcur", il);
519 cb(Kcur, "Kcur", il);
520 cb(Vcur, "Vcur", il);
521
522 // Attention computation
523 const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
524
525 cur = build_attn(inp,
526 nullptr, nullptr,
527 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
528 cb(cur, "attn_pregate", il);
529
530 ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate);
531 cb(gate_sigmoid, "gate_sigmoid", il);
532
533 cur = ggml_mul(ctx0, cur, gate_sigmoid);
534 cb(cur, "attn_gated", il);
535
536 cur = build_lora_mm(model.layers[il].wo, cur);
537 cb(cur, "attn_output", il);
538
539 return cur;
540}
541
542std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(
543 ggml_tensor * input,
544 int il) {
545 const int64_t d_inner = hparams.ssm_d_inner;
546 const int64_t n_seqs = ubatch.n_seqs;
547 const int64_t head_k_dim = hparams.ssm_d_state;
548 const int64_t num_k_heads = hparams.ssm_n_group;
549 const int64_t num_v_heads = hparams.ssm_dt_rank;
550 const int64_t head_v_dim = d_inner / num_v_heads;
551 const int64_t n_seq_tokens = ubatch.n_seq_tokens;
552
553 if (model.layers[il].wqkv) {
554 // optimized path
555 ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);
556 qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);
557 cb(qkv_mixed, "linear_attn_qkv_mixed", il);
558
559 ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);
560 cb(z, "z", il);
561
562 return { qkv_mixed, z };
563
564 } else {
565 // legacy (slower) path
566 ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input);
567 cb(mixed_qkvz, "linear_attn_mixed_qkvz", il);
568
569 int64_t qkvz_new_dim = 2 * head_k_dim + 2 * head_v_dim * (num_v_heads / num_k_heads);
570 ggml_tensor * mixed_qkvz_reshaped = ggml_reshape_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);
571
572 // Split mixed_qkvz into query, key, value, z
573 int64_t split_sizes_qkvz[4] = {
574 head_k_dim, // query size
575 head_k_dim, // key size
576 head_v_dim * num_v_heads / num_k_heads, // value size
577 head_v_dim * num_v_heads / num_k_heads // z size
578 };
579
580 ggml_tensor * query =
581 ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[0], num_k_heads, n_seq_tokens, n_seqs,
582 mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], 0);
583 cb(query, "q", il);
584
585 ggml_tensor * key = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[1], num_k_heads, n_seq_tokens, n_seqs,
586 mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],
587 split_sizes_qkvz[0] * ggml_element_size(mixed_qkvz_reshaped));
588 cb(key, "k", il);
589
590 ggml_tensor * value =
591 ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[2], num_k_heads, n_seq_tokens, n_seqs,
592 mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],
593 (split_sizes_qkvz[0] + split_sizes_qkvz[1]) * ggml_element_size(mixed_qkvz_reshaped));
594 cb(value, "v", il);
595
596 ggml_tensor * z = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[3], num_k_heads, n_seq_tokens, n_seqs,
597 mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],
598 (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * ggml_element_size(mixed_qkvz_reshaped));
599 z = ggml_cont(ctx0, z);
600 cb(z, "z", il);
601
602 // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions
603 // query: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]
604 ggml_tensor * query_flat = ggml_cont_3d(ctx0, query, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);
605 cb(query_flat, "query_flat", il);
606
607 // key: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]
608 ggml_tensor * key_flat = ggml_cont_3d(ctx0, key, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);
609 cb(key_flat, "key_flat", il);
610
611 // value_reshaped: [head_v_dim, num_v_heads, n_tokens, n_seqs] -> [head_v_dim * num_v_heads, n_tokens, n_seqs]
612 ggml_tensor * value_flat = ggml_cont_3d(ctx0, value, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
613 cb(value_flat, "value_flat", il);
614
615 // Now concatenate along the feature dimension (dim 0) to get [conv_dim, n_tokens, n_seqs]
616 ggml_tensor * qkv_mixed = ggml_concat(ctx0, query_flat, key_flat, 0);
617 qkv_mixed = ggml_concat(ctx0, qkv_mixed, value_flat, 0);
618 cb(qkv_mixed, "qkv_mixed", il);
619
620 return { qkv_mixed, z };
621 }
622}
623
624ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
625 llm_graph_input_rs * inp,
626 ggml_tensor * cur,
627 ggml_tensor * causal_mask,
628 ggml_tensor * identity,
629 ggml_tensor * diag_mask,
630 int il) {
631 const auto * mctx_cur = inp->mctx;
632
633 const int64_t d_inner = hparams.ssm_d_inner;
634 const int64_t n_seqs = ubatch.n_seqs;
635 const int64_t head_k_dim = hparams.ssm_d_state;
636 const int64_t num_k_heads = hparams.ssm_n_group;
637 const int64_t num_v_heads = hparams.ssm_dt_rank;
638 const int64_t head_v_dim = d_inner / num_v_heads;
639 const int64_t n_seq_tokens = ubatch.n_seq_tokens;
640
641 const auto kv_head = mctx_cur->get_head();
642
643 GGML_ASSERT(n_seqs != 0);
644 GGML_ASSERT(ubatch.equal_seqs());
645 GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
646
647 // Input projections
648 auto qkvz = build_qkvz(cur, il);
649 ggml_tensor * qkv_mixed = qkvz.first;
650 ggml_tensor * z = qkvz.second;
651
652 ggml_tensor * mixed_ba = build_lora_mm(model.layers[il].ssm_beta_alpha, cur);
653 cb(mixed_ba, "linear_attn_mixed_ba", il);
654
655 // Reshape mixed_ba: [batch, seq_len, hidden_size] -> [batch, seq_len, num_k_heads, 2*num_v_heads/num_k_heads]
656 int64_t ba_new_dim = 2 * num_v_heads / num_k_heads;
657 ggml_tensor * mixed_ba_reshaped = ggml_reshape_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);
658
659 // Split mixed_ba into b and a (beta and alpha parameters)
660 int64_t split_sizes_ba[2] = {
661 num_v_heads / num_k_heads, // beta size
662 num_v_heads / num_k_heads // alpha size
663 };
664
665 ggml_tensor * b = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[0], num_k_heads, n_seq_tokens, n_seqs,
666 mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3], 0);
667 cb(b, "b", il);
668
669 ggml_tensor * a = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[1], num_k_heads, n_seq_tokens, n_seqs,
670 mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3],
671 split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped));
672 cb(a, "a", il);
673
674 ggml_tensor * beta = ggml_cont_4d(ctx0, b, num_v_heads, 1, n_seq_tokens, n_seqs);
675
676 // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]
677 ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
678
679 ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
680 ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
681 cb(alpha_softplus, "a_softplus", il);
682 ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
683 cb(gate, "gate", il);
684
685 // Get convolution states from cache
686 ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
687 ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
688
689 // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
690
691 // Build the convolution states tensor
692 ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
693 cb(conv_states, "conv_states", il);
694
695 // Calculate convolution kernel size
696 ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d;
697 const int64_t conv_kernel_size = conv_kernel->ne[0];
698 const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
699 conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
700 cb(conv_states, "conv_states_reshaped", il);
701
702 qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
703 cb(qkv_mixed, "qkv_mixed_permuted", il);
704
705 ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
706 cb(conv_input, "conv_input", il);
707
708 // Update convolution state cache
709 // Extract the last (conv_kernel_size - 1) states from conv_input
710 ggml_tensor * last_conv_states =
711 ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1],
712 conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input));
713 cb(last_conv_states, "last_conv_states", il);
714
715 ggml_tensor * state_update_target =
716 ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs,
717 kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all));
718 cb(state_update_target, "state_update_target", il);
719
720 ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
721 cb(conv_states_all, "conv_states_updated", il);
722
723 // Apply SSM convolution
724 ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
725 cb(conv_output_proper, "conv_output_raw", il);
726
727 ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper);
728 cb(conv_output_silu, "conv_output_silu", il);
729
730 ggml_tensor * conv_qkv_mix = conv_output_silu;
731
732 // Calculate the total conv dimension
733 int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads;
734 int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
735
736 // Extract the convolved Q, K, V from conv_output
737 ggml_tensor * q_conv =
738 ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
739 cb(q_conv, "q_conv", il);
740 ggml_tensor * k_conv =
741 ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
742 head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
743 cb(k_conv, "k_conv", il);
744 ggml_tensor * v_conv =
745 ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
746 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
747 cb(v_conv, "v_conv", il);
748
749 // Unsqueeze them
750 q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
751 k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
752 v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
753
754 ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
755 state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
756 cb(state, "state_predelta", il);
757
758 // if head keys and value keys are different, repeat to force tensors into matching shapes
759 if (num_k_heads != num_v_heads) {
760 GGML_ASSERT(num_v_heads % num_k_heads == 0);
761 int64_t repeat_factor = num_v_heads / num_k_heads;
762
763 // repeat interleave: reshape to (repeat part, 1, remaining part), do repeat, then reshape back
764 ggml_tensor * q_reshaped = ggml_reshape_3d(ctx0, q_conv, head_k_dim, 1, num_k_heads * n_seq_tokens * n_seqs);
765 ggml_tensor * k_reshaped = ggml_reshape_3d(ctx0, k_conv, head_k_dim, 1, num_k_heads * n_seq_tokens * n_seqs);
766
767 // Repeat along the third dimension (the new dimension with size 1)
768 ggml_tensor * q_repeated =
769 ggml_repeat_4d(ctx0, q_reshaped, head_k_dim, repeat_factor, num_k_heads * n_seq_tokens * n_seqs, 1);
770 ggml_tensor * k_repeated =
771 ggml_repeat_4d(ctx0, k_reshaped, head_k_dim, repeat_factor, num_k_heads * n_seq_tokens * n_seqs, 1);
772
773 // Reshape back to merge the head and repeat dimensions
774 // From [head_dim, num_k_heads, repeat_factor, n_seq_tokens * n_seqs]
775 // Back to [head_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs]
776 q_conv = ggml_reshape_4d(ctx0, q_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);
777 k_conv = ggml_reshape_4d(ctx0, k_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);
778 }
779
780 cb(q_conv, "q_conv_predelta", il);
781 cb(k_conv, "k_conv_predelta", il);
782 cb(v_conv, "v_conv_predelta", il);
783
784 // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
785 std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
786 if (n_seq_tokens == 1) {
787 attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
788 } else {
789 attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
790 }
791 ggml_tensor * output = attn_out.first;
792 ggml_tensor * new_state = attn_out.second;
793 cb(output, "attn_output", il);
794 cb(new_state, "new_state", il);
795
796 // Update the recurrent states
797 ggml_build_forward_expand(gf,
798 ggml_cpy(ctx0, new_state,
799 ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
800 kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
801
802 // Reshape both attn_out_final and z to 2D tensors for normalization
803 // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
804 ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
805
806 // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
807 ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
808
809 // Apply gated normalization: self.norm(core_attn_out, z)
810 ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
811
812 // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
813 ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
814 cb(final_output, "final_output", il);
815
816 // Output projection
817 cur = build_lora_mm(model.layers[il].ssm_out, final_output);
818 cb(cur, "linear_attn_out", il);
819
820 // Reshape back to original dimensions
821 cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
822 return cur;
823}
824
825ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int il) {
826 // Check if this is an MoE layer
827 if (model.layers[il].ffn_gate_inp != nullptr) {
828 // MoE branch
829 ggml_tensor * moe_out =
830 build_moe_ffn(cur,
831 model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
832 model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
833 nullptr,
834 n_expert, n_expert_used, LLM_FFN_SILU,
835 true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
836 cb(moe_out, "ffn_moe_out", il);
837
838 // Add shared experts if present - following Qwen3Next reference implementation
839 if (model.layers[il].ffn_up_shexp != nullptr) {
840 ggml_tensor * ffn_shexp =
841 build_ffn(cur,
842 model.layers[il].ffn_up_shexp, NULL, NULL,
843 model.layers[il].ffn_gate_shexp, NULL, NULL,
844 model.layers[il].ffn_down_shexp, NULL, NULL,
845 NULL,
846 LLM_FFN_SILU, LLM_FFN_PAR, il);
847 cb(ffn_shexp, "ffn_shexp", il);
848
849 // Apply shared expert gating as in the reference implementation
850 // The shared expert has its own gate that is sigmoided
851 // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token)
852 ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
853 cb(shared_gate, "shared_expert_gate", il);
854
855 // Apply sigmoid to the gate
856 shared_gate = ggml_sigmoid(ctx0, shared_gate);
857 cb(shared_gate, "shared_expert_gate_sigmoid", il);
858
859 // Apply the gate to the shared expert output
860 ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);
861 cb(ffn_shexp, "ffn_shexp_gated", il);
862
863 cur = ggml_add(ctx0, moe_out, ffn_shexp);
864 cb(cur, "ffn_out", il);
865 } else {
866 cur = moe_out;
867 }
868 } else {
869 // Dense FFN branch (not currently used I believe)
870 cur = build_ffn(cur,
871 model.layers[il].ffn_up, NULL, NULL,
872 model.layers[il].ffn_gate, NULL, NULL,
873 model.layers[il].ffn_down, NULL, NULL,
874 NULL,
875 LLM_FFN_SILU, LLM_FFN_PAR, il);
876 cb(cur, "ffn_out", il);
877 }
878 return cur;
879}