echodict/llama.cpp

1#include "models.h"2 3#include "llama-memory-recurrent.h"4 5llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_graph_params & params) :6    llm_build_delta_net_base(params), model(model) {7    ggml_tensor * cur;8    ggml_tensor * inpL;9 10    inpL = build_inp_embd(model.tok_embd);11    cb(inpL, "model.embed_tokens", -1);12 13    auto * inp = build_inp_mem_hybrid();14 15    ggml_tensor * inp_pos     = build_inp_pos();16    ggml_tensor * inp_out_ids = build_inp_out_ids();17 18    for (int il = 0; il < n_layer; ++il) {19        ggml_tensor * inpSA = inpL;20 21        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);22        cb(cur, "attn_norm", il);23 24        ggml_build_forward_expand(gf, cur);25 26        // Determine layer type and build appropriate attention mechanism27        if (hparams.is_recurrent(il)) {28            // Linear attention layer (gated delta net)29            cur = build_layer_attn_linear(inp->get_recr(), cur, il);30        } else {31            // Full attention layer32            cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);33        }34 35        if (il == n_layer - 1 && inp_out_ids) {36            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);37            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);38        }39 40        // Residual connection41        cur = ggml_add(ctx0, cur, inpSA);42        cb(cur, "attn_residual", il);43 44        // Save the tensor before post-attention norm for residual connection45        ggml_tensor * ffn_residual = cur;46 47        // Post-attention norm48        ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);49        cb(attn_post_norm, "attn_post_norm", il);50 51        // FFN layer (MoE or dense) - without residual connection52        cur = build_layer_ffn(attn_post_norm, il);53        cb(cur, "ffn_out", il);54 55        // Residual connection for FFN - add to the tensor from before post_attention_layernorm56        cur = ggml_add(ctx0, cur, ffn_residual);57        cb(cur, "post_moe", il);58 59        cur = build_cvec(cur, il);60        cb(cur, "l_out", il);61 62        // Input for next layer63        inpL = cur;64    }65    cur = inpL;66 67    // Final norm68    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);69 70    cb(cur, "result_norm", -1);71    res->t_embd = cur;72 73    // LM head74    cur = build_lora_mm(model.output, cur);75 76    cb(cur, "result_output", -1);77    res->t_logits = cur;78 79    ggml_build_forward_expand(gf, cur);80}81 82// utility to get one slice from the third dimension83// input dim:  [x, y, c, b]84// output dim: [x, y, 1, b]85static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {86    return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],87        t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);88}89 90ggml_tensor * llm_build_qwen3next::build_norm_gated(91        ggml_tensor * input,92        ggml_tensor * weights,93        ggml_tensor * gate,94        int           layer) {95    ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer);96    ggml_tensor * gated_silu = ggml_silu(ctx0, gate);97 98    return ggml_mul(ctx0, normalized, gated_silu);99}100 101ggml_tensor * llm_build_qwen3next::build_layer_attn(102        llm_graph_input_attn_kv * inp,103        ggml_tensor *             cur,104        ggml_tensor *             inp_pos,105        int                       il) {106    const int64_t n_embd_head = hparams.n_embd_head_v();107    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());108 109    // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention110 111    // Qwen3Next uses a single Q projection that outputs query + gate112    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur);113    cb(Qcur_full, "Qcur_full", il);114 115    Qcur_full = ggml_reshape_4d(ctx0, Qcur_full, n_embd_head * 2, n_head, n_tokens, 1);116 117    // Split Q projection into query and gate118    // The split should be along dimension 0 (the feature dimension)119    ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,120                                            Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);121    cb(Qcur, "Qcur_view", il);122 123    ggml_tensor * gate =124        ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,125                     Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));126    cb(gate, "gate", il);127 128    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);129    cb(Kcur, "Kcur", il);130 131    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);132    cb(Vcur, "Vcur", il);133 134    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);135    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);136 137    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);138    cb(Qcur, "Qcur_normed", il);139 140    Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);141    cb(Kcur, "Kcur_normed", il);142 143    Qcur = ggml_rope_ext(144            ctx0, Qcur, inp_pos, nullptr,145            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,146            ext_factor, attn_factor, beta_fast, beta_slow);147 148    Kcur = ggml_rope_ext(149            ctx0, Kcur, inp_pos, nullptr,150            n_rot, rope_type, n_ctx_orig, freq_base,151            freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);152 153    cb(Qcur, "Qcur", il);154    cb(Kcur, "Kcur", il);155    cb(Vcur, "Vcur", il);156 157    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;158 159    cur = build_attn(inp,160                nullptr, nullptr, nullptr,161                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);162    cb(cur, "attn_pregate", il);163 164    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont165    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);166 167    gate = ggml_sigmoid(ctx0, gate);168    cb(gate, "gate_sigmoid", il);169 170    gate = ggml_reshape_2d(ctx0, gate, n_embd_head * n_head, n_tokens);171 172    cur = ggml_mul(ctx0, cur, gate);173    cb(cur, "attn_gated", il);174 175    cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);176    cb(cur, "attn_output", il);177 178    return cur;179}180 181std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_qkvz(182                ggml_tensor * input,183                        int   il) {184    const int64_t d_inner      = hparams.ssm_d_inner;185    const int64_t n_seqs       = ubatch.n_seqs;186    const int64_t head_k_dim   = hparams.ssm_d_state;187    const int64_t num_k_heads  = hparams.ssm_n_group;188    const int64_t num_v_heads  = hparams.ssm_dt_rank;189    const int64_t head_v_dim   = d_inner / num_v_heads;190    const int64_t n_seq_tokens = ubatch.n_seq_tokens;191 192    if (model.layers[il].wqkv) {193        // optimized path194        ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);195        qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);196        cb(qkv_mixed, "linear_attn_qkv_mixed", il);197 198        ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);199        cb(z, "z", il);200 201        return { qkv_mixed, z };202    } else {203        // legacy (slower) path204        ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input);205        cb(mixed_qkvz, "linear_attn_mixed_qkvz", il);206 207        int64_t       qkvz_new_dim        = 2 * head_k_dim + 2 * head_v_dim * (num_v_heads / num_k_heads);208        ggml_tensor * mixed_qkvz_reshaped = ggml_reshape_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);209 210        // Split mixed_qkvz into query, key, value, z211        int64_t split_sizes_qkvz[4] = {212            head_k_dim,                              // query size213            head_k_dim,                              // key size214            head_v_dim * num_v_heads / num_k_heads,  // value size215            head_v_dim * num_v_heads / num_k_heads   // z size216        };217 218        ggml_tensor * query =219            ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[0], num_k_heads, n_seq_tokens, n_seqs,220                        mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], 0);221        cb(query, "q", il);222 223        ggml_tensor * key = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[1], num_k_heads, n_seq_tokens, n_seqs,224                                        mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],225                                        split_sizes_qkvz[0] * ggml_element_size(mixed_qkvz_reshaped));226        cb(key, "k", il);227 228        ggml_tensor * value =229            ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[2], num_k_heads, n_seq_tokens, n_seqs,230                        mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],231                        (split_sizes_qkvz[0] + split_sizes_qkvz[1]) * ggml_element_size(mixed_qkvz_reshaped));232        cb(value, "v", il);233 234        ggml_tensor * z = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[3], num_k_heads, n_seq_tokens, n_seqs,235                                    mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],236                                    (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * ggml_element_size(mixed_qkvz_reshaped));237        z = ggml_cont(ctx0, z);238        cb(z, "z", il);239 240        // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions241        // query: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]242        ggml_tensor * query_flat = ggml_cont_3d(ctx0, query, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);243        cb(query_flat, "query_flat", il);244 245        // key: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]246        ggml_tensor * key_flat = ggml_cont_3d(ctx0, key, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);247        cb(key_flat, "key_flat", il);248 249        // value_reshaped: [head_v_dim, num_v_heads, n_tokens, n_seqs] -> [head_v_dim * num_v_heads, n_tokens, n_seqs]250        ggml_tensor * value_flat = ggml_cont_3d(ctx0, value, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);251        cb(value_flat, "value_flat", il);252 253        // Now concatenate along the feature dimension (dim 0) to get [conv_dim, n_tokens, n_seqs]254        ggml_tensor * qkv_mixed = ggml_concat(ctx0, query_flat, key_flat, 0);255        qkv_mixed               = ggml_concat(ctx0, qkv_mixed, value_flat, 0);256        cb(qkv_mixed, "qkv_mixed", il);257 258        return { qkv_mixed, z };259    }260}261 262ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(263        llm_graph_input_rs * inp,264        ggml_tensor *        cur,265        int                  il) {266    const auto * mctx_cur = inp->mctx;267 268    const int64_t d_inner      = hparams.ssm_d_inner;269    const int64_t n_seqs       = ubatch.n_seqs;270    const int64_t head_k_dim   = hparams.ssm_d_state;271    const int64_t num_k_heads  = hparams.ssm_n_group;272    const int64_t num_v_heads  = hparams.ssm_dt_rank;273    const int64_t head_v_dim   = d_inner / num_v_heads;274    const int64_t n_seq_tokens = ubatch.n_seq_tokens;275 276    const auto kv_head = mctx_cur->get_head();277 278    GGML_ASSERT(n_seqs != 0);279    GGML_ASSERT(ubatch.equal_seqs());280    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);281 282    // Input projections283    auto qkvz = build_qkvz(cur, il);284    ggml_tensor * qkv_mixed = qkvz.first;285    ggml_tensor * z         = qkvz.second;286 287    ggml_tensor * mixed_ba = build_lora_mm(model.layers[il].ssm_beta_alpha, cur);288    cb(mixed_ba, "linear_attn_mixed_ba", il);289 290    // Reshape mixed_ba: [batch, seq_len, hidden_size] -> [batch, seq_len, num_k_heads, 2*num_v_heads/num_k_heads]291    int64_t       ba_new_dim        = 2 * num_v_heads / num_k_heads;292    ggml_tensor * mixed_ba_reshaped = ggml_reshape_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);293 294    // Split mixed_ba into b and a (beta and alpha parameters)295    int64_t split_sizes_ba[2] = {296        num_v_heads / num_k_heads,  // beta size297        num_v_heads / num_k_heads   // alpha size298    };299 300    ggml_tensor * b = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[0], num_k_heads, n_seq_tokens, n_seqs,301                                   mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3], 0);302    cb(b, "b", il);303 304    ggml_tensor * a = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[1], num_k_heads, n_seq_tokens, n_seqs,305                                   mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3],306                                   split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped));307    cb(a, "a", il);308 309    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont310    b = ggml_cont(ctx0, b);311 312    ggml_tensor * beta = ggml_sigmoid(ctx0, b);313 314    // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]315    ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);316 317    ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);318    ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);319    cb(alpha_softplus, "a_softplus", il);320 321    ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a);  // -A_log.exp() * softplus322    cb(gate, "gate", il);323 324    beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);325    gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);326 327    // Get convolution states from cache328    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);329    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);330 331    // Build the convolution states tensor332    ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);333    cb(conv_states, "conv_states", il);334 335    // Calculate convolution kernel size336    ggml_tensor * conv_kernel      = model.layers[il].ssm_conv1d;337    const int64_t conv_kernel_size = conv_kernel->ne[0];338    const int64_t conv_channels    = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;339 340    conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);341    cb(conv_states, "conv_states_reshaped", il);342 343    qkv_mixed = ggml_transpose(ctx0, qkv_mixed);344    cb(qkv_mixed, "qkv_mixed_transposed", il);345 346    ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);347    cb(conv_input, "conv_input", il);348 349    // Update convolution state cache350    // Extract the last (conv_kernel_size - 1) states from conv_input351    ggml_tensor * last_conv_states =352        ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1],353                     conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input));354    cb(last_conv_states, "last_conv_states", il);355 356    ggml_tensor * state_update_target =357        ggml_view_2d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels, n_seqs, conv_states_all->nb[1],358                     kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all));359    cb(state_update_target, "state_update_target", il);360 361    ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));362 363    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);364    state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);365    cb(state, "state_predelta", il);366 367    ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);368    cb(conv_output_proper, "conv_output_raw", il);369 370    ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper);371    cb(conv_output_silu, "conv_output_silu", il);372 373    ggml_tensor * conv_qkv_mix = conv_output_silu;374 375    // Calculate the total conv dimension376    int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads;377    int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);378 379    // Extract the convolved Q, K, V from conv_output380    ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,381            ggml_row_size(conv_qkv_mix->type, head_k_dim),382            nb1_qkv,383            nb1_qkv * n_seq_tokens,384            0);385 386    ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,387            ggml_row_size(conv_qkv_mix->type, head_k_dim),388            nb1_qkv,389            nb1_qkv * n_seq_tokens,390            head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));391 392    ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,393            ggml_row_size(conv_qkv_mix->type, head_v_dim),394            nb1_qkv,395            nb1_qkv * n_seq_tokens,396            ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));397 398    cb(q_conv, "q_conv", il);399    cb(k_conv, "k_conv", il);400    cb(v_conv, "v_conv", il);401 402    const float eps_norm = hparams.f_norm_rms_eps;403 404    q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);405    k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);406 407    //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);408    //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);409    //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);410 411    // if head keys and value keys are different, repeat to force tensors into matching shapes412    // TODO: avoid repeats for fused GDN, needs broadcast configuration for GDN op [TAG_GGML_GDN_BCAST]413    if (num_k_heads != num_v_heads) {414        GGML_ASSERT(num_v_heads % num_k_heads == 0);415        int64_t repeat_factor = num_v_heads / num_k_heads;416 417        // repeat interleave: reshape to (repeat part, 1, remaining part...), do repeat, then reshape back418        ggml_tensor * q_reshaped = ggml_reshape_4d(ctx0, q_conv, head_k_dim, 1, num_k_heads, n_seq_tokens * n_seqs);419        ggml_tensor * k_reshaped = ggml_reshape_4d(ctx0, k_conv, head_k_dim, 1, num_k_heads, n_seq_tokens * n_seqs);420 421        // Repeat along the third dimension (the new dimension with size 1)422        ggml_tensor * q_repeated =423            ggml_repeat_4d(ctx0, q_reshaped, head_k_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs);424        ggml_tensor * k_repeated =425            ggml_repeat_4d(ctx0, k_reshaped, head_k_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs);426 427        // Reshape back to merge the head and repeat dimensions428        // From [head_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs]429        // Back to [head_dim, repeat_factor * num_k_heads, n_seq_tokens, n_seqs]430        q_conv = ggml_reshape_4d(ctx0, q_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);431        k_conv = ggml_reshape_4d(ctx0, k_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);432    }433 434    cb(q_conv, "q_conv_predelta", il);435    cb(k_conv, "k_conv_predelta", il);436    cb(v_conv, "v_conv_predelta", il);437 438    auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);439 440    ggml_tensor * output    = attn_out.first;441    ggml_tensor * new_state = attn_out.second;442    cb(output, "attn_output", il);443    cb(new_state, "new_state", il);444 445    // Update the recurrent states446    ggml_build_forward_expand(gf,447            ggml_cpy(ctx0, new_state,448                ggml_view_2d(ctx0, ssm_states_all, hparams.n_embd_s(), n_seqs, ssm_states_all->nb[1],449                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));450 451    // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]452    ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);453 454    // Apply gated normalization: self.norm(core_attn_out, z)455    ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);456 457    // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]458    ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);459    cb(final_output, "final_output", il);460 461    // Output projection462    cur = build_lora_mm(model.layers[il].ssm_out, final_output);463    cb(cur, "linear_attn_out", il);464 465    // Reshape back to original dimensions466    cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);467 468    return cur;469}470 471ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int il) {472    // Check if this is an MoE layer473    if (model.layers[il].ffn_gate_inp != nullptr) {474        // MoE branch475        ggml_tensor * moe_out =476            build_moe_ffn(cur,477                model.layers[il].ffn_gate_inp,478                model.layers[il].ffn_up_exps,479                model.layers[il].ffn_gate_exps,480                model.layers[il].ffn_down_exps,481                nullptr,482                n_expert, n_expert_used,483                LLM_FFN_SILU, true,484                hparams.expert_weights_scale,485                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,486                nullptr, model.layers[il].ffn_gate_up_exps);487        cb(moe_out, "ffn_moe_out", il);488 489        // Add shared experts if present - following Qwen3Next reference implementation490        if (model.layers[il].ffn_up_shexp != nullptr) {491            ggml_tensor * ffn_shexp =492                build_ffn(cur,493                    model.layers[il].ffn_up_shexp,   NULL, NULL,494                    model.layers[il].ffn_gate_shexp, NULL, NULL,495                    model.layers[il].ffn_down_shexp, NULL, NULL,496                    NULL,497                    LLM_FFN_SILU, LLM_FFN_PAR, il);498            cb(ffn_shexp, "ffn_shexp", il);499 500            // Apply shared expert gating as in the reference implementation501            // The shared expert has its own gate that is sigmoided502            // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token)503            ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);504            cb(shared_gate, "shared_expert_gate", il);505 506            shared_gate = ggml_sigmoid(ctx0, shared_gate);507            cb(shared_gate, "shared_expert_gate_sigmoid", il);508 509            ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);510            cb(ffn_shexp, "ffn_shexp_gated", il);511 512            cur = ggml_add(ctx0, moe_out, ffn_shexp);513            cb(cur, "ffn_out", il);514        } else {515            cur = moe_out;516        }517    } else {518        // Dense FFN branch (not currently used I believe)519        cur = build_ffn(cur,520            model.layers[il].ffn_up, NULL, NULL,521            model.layers[il].ffn_gate, NULL, NULL,522            model.layers[il].ffn_down, NULL, NULL,523            NULL,524            LLM_FFN_SILU, LLM_FFN_PAR, il);525        cb(cur, "ffn_out", il);526    }527    return cur;528}529