1#include "ggml.h"
2#include "models.h"
3
4#define CHUNK_SIZE 64
5
6llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) :
7 llm_graph_context_mamba(params), model(model) {
8 const int64_t n_embd_head = hparams.n_embd_head_v;
9
10 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
11
12 int sections[4];
13 std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
14
15 ggml_tensor * cur;
16 ggml_tensor * inpL;
17
18 inpL = build_inp_embd(model.tok_embd);
19
20 cb(inpL, "model.input_embed", -1);
21
22 auto * inp = build_inp_mem_hybrid();
23
24 ggml_tensor * inp_pos = build_inp_pos();
25 ggml_tensor * inp_out_ids = build_inp_out_ids();
26
27 ggml_tensor * causal_mask =
28 ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
29 GGML_TRI_TYPE_LOWER);
30
31 ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
32 ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
33
34 ggml_build_forward_expand(gf, causal_mask);
35 ggml_build_forward_expand(gf, identity);
36 ggml_build_forward_expand(gf, diag_mask);
37
38 for (int il = 0; il < n_layer; ++il) {
39 ggml_tensor * inpSA = inpL;
40
41 cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
42 cb(cur, "attn_norm", il);
43
44 // Determine layer type and build appropriate attention mechanism
45 if (hparams.is_recurrent(il)) {
46 // Linear attention layer (gated delta net)
47 cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
48 } else {
49 // Full attention layer
50 cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
51 }
52
53 if (il == n_layer - 1 && inp_out_ids) {
54 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
55 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
56 }
57
58 // Residual connection
59 cur = ggml_add(ctx0, cur, inpSA);
60 cb(cur, "attn_residual", il);
61
62 // Save the tensor before post-attention norm for residual connection
63 ggml_tensor * ffn_residual = cur;
64
65 // Post-attention norm
66 ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);
67 cb(attn_post_norm, "attn_post_norm", il);
68
69 // MOE FFN layer
70 cur = build_layer_ffn(attn_post_norm, il);
71 cb(cur, "ffn_out", il);
72
73 // Residual connection for FFN - add to the tensor from before post_attention_layernorm
74 cur = ggml_add(ctx0, cur, ffn_residual);
75 cb(cur, "post_moe", il);
76
77 // Input for next layer
78 inpL = cur;
79 }
80 cur = inpL;
81
82 // Final norm
83 cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
84
85 cb(cur, "result_norm", -1);
86 res->t_embd = cur;
87
88 // LM head
89 cur = build_lora_mm(model.output, cur);
90
91 cb(cur, "result_output", -1);
92 res->t_logits = cur;
93
94 ggml_build_forward_expand(gf, cur);
95}
96
97// utility to get one slice from the third dimension
98// input dim: [x, y, c, b]
99// output dim: [x, y, 1, b]
100static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
101 return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
102 t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
103}
104
105std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_delta_net_chunking(
106 ggml_tensor * q,
107 ggml_tensor * k,
108 ggml_tensor * v,
109 ggml_tensor * g,
110 ggml_tensor * beta,
111 ggml_tensor * state,
112 ggml_tensor * causal_mask,
113 ggml_tensor * identity,
114 ggml_tensor * diag_mask,
115 int il) {
116 const int64_t S_k = q->ne[0];
117 const int64_t H_k = q->ne[1];
118 const int64_t n_tokens = q->ne[2];
119 const int64_t n_seqs = q->ne[3];
120
121 const int64_t S_v = v->ne[0];
122 const int64_t H_v = v->ne[1];
123
124 GGML_ASSERT(v->ne[2] == n_tokens);
125 GGML_ASSERT(k->ne[2] == n_tokens);
126 GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
127 GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
128 GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
129
130 GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
131 GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
132
133 GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
134
135 const float eps_norm = hparams.f_norm_rms_eps;
136
137 q = ggml_l2_norm(ctx0, q, eps_norm);
138 k = ggml_l2_norm(ctx0, k, eps_norm);
139
140 const float scale = 1.0f / sqrtf(S_v);
141
142 q = ggml_scale(ctx0, q, scale);
143
144 beta = ggml_sigmoid(ctx0, beta);
145
146 cb(q, "q_in", il);
147 cb(k, "k_in", il);
148 cb(v, "v_in", il);
149 cb(beta, "beta_in", il);
150 cb(g, "g_in", il);
151
152 q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
153 k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
154 v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
155 g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
156
157 beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
158 state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
159
160 cb(q, "q_perm", il);
161 cb(k, "k_perm", il);
162 cb(v, "v_perm", il);
163 cb(beta, "beta_perm", il);
164 cb(g, "g_perm", il);
165 cb(state, "state_in", il);
166
167 GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
168 GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
169 GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
170 GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
171
172 // Do padding
173 const int64_t chunk_size = CHUNK_SIZE;
174
175 const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
176 const int64_t n_chunks = (n_tokens + pad) / chunk_size;
177
178 q = ggml_pad(ctx0, q, 0, pad, 0, 0);
179 k = ggml_pad(ctx0, k, 0, pad, 0, 0);
180 v = ggml_pad(ctx0, v, 0, pad, 0, 0);
181 g = ggml_pad(ctx0, g, pad, 0, 0, 0);
182 beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
183
184 cb(q, "q_pad", il);
185 cb(k, "k_pad", il);
186 cb(v, "v_pad", il);
187 cb(beta, "beta_pad", il);
188 cb(g, "g_pad", il);
189
190 ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
191 ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
192
193 cb(v_beta, "v_beta", il);
194 cb(k_beta, "k_beta", il);
195
196 q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs);
197 k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs);
198 k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
199 v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs);
200 v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
201
202 g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
203 beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
204
205 ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
206 cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
207
208 ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
209 ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
210
211 ggml_tensor * gcs_j_broadcast =
212 ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
213
214 ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
215 cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
216
217 decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
218 decay_mask = ggml_exp(ctx0, decay_mask);
219 decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
220
221 ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
222
223 ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
224 ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
225 cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
226
227 ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
228 ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
229
230 ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
231 attn = ggml_mul(ctx0, lin_solve, causal_mask);
232 attn = ggml_add(ctx0, attn, identity);
233 cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
234
235 v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
236
237 ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
238 ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t);
239
240 ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
241 cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
242
243 ggml_tensor * k_cumdecay =
244 ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
245 cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
246
247 ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
248 attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
249 attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
250 cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
251
252
253 // vectorized calculation of key_gdiff
254 // improved from the chunked version:
255 // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
256 // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
257 // key_gdiff = key * g_diff.unsqueeze(-1)
258 // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
259 // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
260
261 // get last element in g_cumsum along chunk_size dimension (ne0)
262 // example: [[x, y, z, ..., last], ...] -> [[last], ...]
263 ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
264 g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
265 (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
266 g_last = ggml_cont(ctx0, g_last);
267 cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
268
269 ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
270 cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
271
272 ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
273 cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
274
275 ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
276 ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
277 1, chunk_size, n_chunks, g_diff_exp->ne[3]);
278
279 ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
280 cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
281
282 ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
283 cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
284
285
286 // state to be updated per chunk
287 ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
288 cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
289
290 // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
291 ggml_tensor * core_attn_out = nullptr;
292
293 for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
294 // shape: (S_k, chunk_size, 1, H_k * n_seqs)
295 ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
296
297 // shape: (S_v, chunk_size, 1, H_v * n_seqs)
298 ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
299
300 // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
301 ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
302
303 // shape: (chunk_size, 1, H_v * n_seqs)
304 ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
305
306 // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
307 // replaced by precomputed attn_kq
308 ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
309 cb(attn_chunk, "attn_chunk", il);
310
311 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);
312
313 // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
314 ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
315 cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
316
317 // v_new = v_i - v_prime
318 ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
319 ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
320 cb(v_new, "v_new_chunk", il);
321
322 // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
323 ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk);
324 ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
325 cb(attn_inter, "attn_inter_chunk", il);
326
327 // core_attn_out[:, :, i] = attn_inter + attn @ v_new
328 ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
329 cb(v_attn, "v_attn_chunk", il);
330
331 ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
332 cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
333
334 core_attn_out = core_attn_out == nullptr
335 ? core_attn_out_chunk
336 : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
337
338 // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
339 ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
340 //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
341 ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
342
343 // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
344 ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
345 new_state = ggml_add(ctx0,
346 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)),
347 ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
348 }
349
350 // truncate padded tokens
351 ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
352 S_v, n_tokens, H_v, n_seqs,
353 ggml_row_size(core_attn_out->type, S_v),
354 ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
355 ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
356 output_tokens = ggml_cont(ctx0, output_tokens);
357 cb(output_tokens, "output_tokens", il);
358
359 // permute back to (S_v, H_v, n_tokens, n_seqs)
360 output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
361 output_tokens = ggml_cont(ctx0, output_tokens);
362
363 return {output_tokens, new_state};
364}
365
366std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_delta_net_autoregressive(
367 ggml_tensor * q,
368 ggml_tensor * k,
369 ggml_tensor * v,
370 ggml_tensor * g,
371 ggml_tensor * beta,
372 ggml_tensor * state,
373 int il) {
374 const int64_t S_k = q->ne[0];
375 const int64_t H_k = q->ne[1];
376 const int64_t n_tokens = q->ne[2];
377 const int64_t n_seqs = q->ne[3];
378
379 const int64_t S_v = v->ne[0];
380 const int64_t H_v = v->ne[1];
381
382 GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing
383 GGML_ASSERT(v->ne[2] == n_tokens);
384 GGML_ASSERT(k->ne[2] == n_tokens);
385 GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
386 GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
387 GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
388
389 GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
390 GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
391
392 GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case
393
394 const float eps_norm = hparams.f_norm_rms_eps;
395
396 q = ggml_l2_norm(ctx0, q, eps_norm);
397 k = ggml_l2_norm(ctx0, k, eps_norm);
398
399 const float scale = 1.0f / sqrtf(S_v);
400
401 q = ggml_scale(ctx0, q, scale);
402 beta = ggml_sigmoid(ctx0, beta);
403
404 cb(q, "q_in", il);
405 cb(k, "k_in", il);
406 cb(v, "v_in", il);
407 cb(beta, "beta_in", il);
408 cb(g, "g_in", il);
409
410 state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
411
412 ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
413 ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
414
415 // Apply exponential to g_t
416 g_t = ggml_exp(ctx0, g_t);
417
418 // Apply the gated delta rule for the single timestep
419 // last_recurrent_state = last_recurrent_state * g_t
420 state = ggml_mul(ctx0, state, g_t);
421
422 // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
423 ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
424 ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed);
425 // we need to sum over dim=-2, so we transpose, sum, then transpose again
426 kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
427
428 // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
429 ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
430 // delta = (v_t - kv_mem) * beta_t
431 ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs]
432 ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t);
433
434 // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
435 ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
436 state = ggml_add(ctx0, state, k_t_delta);
437
438 // Compute the attention output
439 // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
440 ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t
441 ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed);
442 // again, since it's over dim = -2, transpose, sum, transpose back
443 ggml_tensor * core_attn_out =
444 ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
445
446 // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
447 cb(core_attn_out, "output_tokens", il);
448 cb(state, "new_state", il);
449
450 return {core_attn_out, state};
451}
452
453std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_qkvz(
454 ggml_tensor * input,
455 int il) {
456 const int64_t n_seqs = ubatch.n_seqs;
457 const int64_t n_seq_tokens = ubatch.n_seq_tokens;
458
459 ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);
460 qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);
461 cb(qkv_mixed, "linear_attn_qkv_mixed", il);
462
463 ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);
464 cb(z, "z", il);
465
466 return { qkv_mixed, z };
467}
468
469ggml_tensor * llm_build_qwen35moe::build_norm_gated(
470 ggml_tensor * input,
471 ggml_tensor * weights,
472 ggml_tensor * gate,
473 int layer) {
474 ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer);
475 ggml_tensor * gated_silu = ggml_silu(ctx0, gate);
476
477 return ggml_mul(ctx0, normalized, gated_silu);
478}
479
480ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
481 llm_graph_input_attn_kv * inp,
482 ggml_tensor * cur,
483 ggml_tensor * inp_pos,
484 int * sections,
485 int il) {
486 const int64_t n_embd_head = hparams.n_embd_head_v;
487 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
488
489 // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
490
491 // Qwen3Next uses a single Q projection that outputs query + gate
492 ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ]
493 cb(Qcur_full, "Qcur_full", il);
494
495 ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens,
496 ggml_element_size(Qcur_full) * n_embd_head * 2,
497 ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, 0);
498 cb(Qcur, "Qcur_reshaped", il);
499
500 // Apply Q normalization
501 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
502 cb(Qcur, "Qcur_normed", il);
503
504 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
505 cb(Kcur, "Kcur", il);
506
507 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
508 cb(Vcur, "Vcur", il);
509
510 // Apply K normalization
511 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
512 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
513 cb(Kcur, "Kcur_normed", il);
514
515 ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens,
516 ggml_element_size(Qcur_full) * n_embd_head * 2,
517 ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,
518 ggml_element_size(Qcur_full) * n_embd_head);
519 gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
520 cb(gate, "gate_reshaped", il);
521
522 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
523
524 // Apply IMRoPE
525 Qcur = ggml_rope_multi(
526 ctx0, Qcur, inp_pos, nullptr,
527 n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
528 ext_factor, attn_factor, beta_fast, beta_slow
529 );
530
531 Kcur = ggml_rope_multi(
532 ctx0, Kcur, inp_pos, nullptr,
533 n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
534 ext_factor, attn_factor, beta_fast, beta_slow
535 );
536
537 cb(Qcur, "Qcur", il);
538 cb(Kcur, "Kcur", il);
539 cb(Vcur, "Vcur", il);
540
541 // Attention computation
542 const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
543
544 cur = build_attn(inp,
545 nullptr, nullptr,
546 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
547 cb(cur, "attn_pregate", il);
548
549 ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate);
550 cb(gate_sigmoid, "gate_sigmoid", il);
551
552 cur = ggml_mul(ctx0, cur, gate_sigmoid);
553 cb(cur, "attn_gated", il);
554
555 cur = build_lora_mm(model.layers[il].wo, cur);
556 cb(cur, "attn_output", il);
557
558 return cur;
559}
560
561ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
562 llm_graph_input_rs * inp,
563 ggml_tensor * cur,
564 ggml_tensor * causal_mask,
565 ggml_tensor * identity,
566 ggml_tensor * diag_mask,
567 int il) {
568 const auto * mctx_cur = inp->mctx;
569
570 const int64_t d_inner = hparams.ssm_d_inner;
571 const int64_t n_seqs = ubatch.n_seqs;
572 const int64_t head_k_dim = hparams.ssm_d_state;
573 const int64_t num_k_heads = hparams.ssm_n_group;
574 const int64_t num_v_heads = hparams.ssm_dt_rank;
575 const int64_t head_v_dim = d_inner / num_v_heads;
576 const int64_t n_seq_tokens = ubatch.n_seq_tokens;
577
578 const auto kv_head = mctx_cur->get_head();
579
580 GGML_ASSERT(n_seqs != 0);
581 GGML_ASSERT(ubatch.equal_seqs());
582 GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
583
584 // Input projections
585 auto qkvz = build_qkvz(cur, il);
586 ggml_tensor * qkv_mixed = qkvz.first;
587 ggml_tensor * z = qkvz.second;
588
589 ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
590 beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
591 cb(beta, "beta", il);
592 ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
593 alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
594 cb(alpha, "alpha", il);
595
596 ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
597 ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
598 cb(alpha_softplus, "a_softplus", il);
599 ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus
600 cb(gate, "gate", il);
601
602 // Get convolution states from cache
603 ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
604 ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
605
606 // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
607
608 // Build the convolution states tensor
609 ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
610 cb(conv_states, "conv_states", il);
611
612 // Calculate convolution kernel size
613 ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d;
614 const int64_t conv_kernel_size = conv_kernel->ne[0];
615 const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
616 conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
617 cb(conv_states, "conv_states_reshaped", il);
618
619 qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
620 cb(qkv_mixed, "qkv_mixed_permuted", il);
621
622 ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
623 cb(conv_input, "conv_input", il);
624
625 // Update convolution state cache
626 // Extract the last (conv_kernel_size - 1) states from conv_input
627 ggml_tensor * last_conv_states =
628 ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1],
629 conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input));
630 cb(last_conv_states, "last_conv_states", il);
631
632 ggml_tensor * state_update_target =
633 ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs,
634 kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all));
635 cb(state_update_target, "state_update_target", il);
636
637 ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
638 cb(conv_states_all, "conv_states_updated", il);
639
640 // Apply SSM convolution
641 ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
642 cb(conv_output_proper, "conv_output_raw", il);
643
644 ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper);
645 cb(conv_output_silu, "conv_output_silu", il);
646
647 ggml_tensor * conv_qkv_mix = conv_output_silu;
648
649 // Calculate the total conv dimension
650 int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads;
651 int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
652
653 // Extract the convolved Q, K, V from conv_output
654 ggml_tensor * q_conv =
655 ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
656 cb(q_conv, "q_conv", il);
657 ggml_tensor * k_conv =
658 ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
659 head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
660 cb(k_conv, "k_conv", il);
661 ggml_tensor * v_conv =
662 ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
663 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
664 cb(v_conv, "v_conv", il);
665
666 // Unsqueeze them
667 q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
668 k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
669 v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
670
671 ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
672 state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
673 cb(state, "state_predelta", il);
674
675 // if head keys and value keys are different, repeat Q/K to match V's head count
676 // V heads are in tiled order (from conversion), so simple tiled repeat works
677 if (num_k_heads != num_v_heads) {
678 GGML_ASSERT(num_v_heads % num_k_heads == 0);
679 q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
680 k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
681 }
682
683 cb(q_conv, "q_conv_predelta", il);
684 cb(k_conv, "k_conv_predelta", il);
685 cb(v_conv, "v_conv_predelta", il);
686
687 // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
688 std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
689 if (n_seq_tokens == 1) {
690 attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
691 } else {
692 attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
693 }
694 ggml_tensor * output = attn_out.first;
695 ggml_tensor * new_state = attn_out.second;
696 cb(output, "attn_output", il);
697 cb(new_state, "new_state", il);
698
699 // Update the recurrent states
700 ggml_build_forward_expand(gf,
701 ggml_cpy(ctx0, new_state,
702 ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
703 kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
704
705 // Reshape both attn_out_final and z to 2D tensors for normalization
706 // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
707 ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
708
709 // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
710 ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
711
712 // Apply gated normalization: self.norm(core_attn_out, z)
713 ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
714
715 // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
716 ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
717 cb(final_output, "final_output", il);
718
719 // Output projection
720 cur = build_lora_mm(model.layers[il].ssm_out, final_output);
721 cb(cur, "linear_attn_out", il);
722
723 // Reshape back to original dimensions
724 cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
725 return cur;
726}
727
728ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int il) {
729 // Check if this is an MoE layer
730 GGML_ASSERT(model.layers[il].ffn_gate_inp != nullptr);
731
732 ggml_tensor * moe_out =
733 build_moe_ffn(cur,
734 model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
735 model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
736 nullptr,
737 n_expert, n_expert_used, LLM_FFN_SILU,
738 true, false, 0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
739 cb(moe_out, "ffn_moe_out", il);
740
741 // Add shared experts if present - following Qwen3Next reference implementation
742 if (model.layers[il].ffn_up_shexp != nullptr) {
743 ggml_tensor * ffn_shexp =
744 build_ffn(cur,
745 model.layers[il].ffn_up_shexp, NULL, NULL,
746 model.layers[il].ffn_gate_shexp, NULL, NULL,
747 model.layers[il].ffn_down_shexp, NULL, NULL,
748 NULL,
749 LLM_FFN_SILU, LLM_FFN_PAR, il);
750 cb(ffn_shexp, "ffn_shexp", il);
751
752 // Apply shared expert gating as in the reference implementation
753 // The shared expert has its own gate that is sigmoided
754 // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token)
755 ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);
756 cb(shared_gate, "shared_expert_gate", il);
757
758 // Apply sigmoid to the gate
759 shared_gate = ggml_sigmoid(ctx0, shared_gate);
760 cb(shared_gate, "shared_expert_gate_sigmoid", il);
761
762
763 // Apply the gate to the shared expert output
764 ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);
765 cb(ffn_shexp, "ffn_shexp_gated", il);
766
767 cur = ggml_add(ctx0, moe_out, ffn_shexp);
768 cb(cur, "ffn_out", il);
769 } else {
770 cur = moe_out;
771 }
772
773 return cur;
774}