Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2#include "llama-memory-recurrent.h"3 4void llama_model_qwen3next::load_arch_hparams(llama_model_loader & ml) {5    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp, false);6    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);7    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);8 9    // Load linear attention (gated delta net) parameters10    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);11    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);12    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);13    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);14    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);15 16    // NextN/MTP: extra decoder block appended beyond the main stack17    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);18    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all");19 20    // Mark recurrent layers (linear attention layers).21    if (!ml.get_key_or_arr(LLM_KV_ATTENTION_RECURRENT_LAYERS, hparams.is_recr_impl, hparams.n_layer_all, false)) {22        uint32_t full_attn_interval = 4;23        ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);24        for (uint32_t i = 0; i < hparams.n_layer_all; ++i) {25            hparams.is_recr_impl[i] = (i < hparams.n_layer()) && ((i + 1) % full_attn_interval != 0);26        }27    }28 29    switch (hparams.n_layer()) {30        case 48: type = LLM_TYPE_80B_A3B; break;31        default: type = LLM_TYPE_UNKNOWN;32    }33}34 35void llama_model_qwen3next::load_arch_tensors(llama_model_loader & ml) {36    LLAMA_LOAD_LOCALS;37 38    if (n_expert == 0) {39        throw std::runtime_error(arch_name() + " model cannot have zero experts");40    }41 42    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);43    const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;44    int mtp_flags = !ml.load_mtp ? TENSOR_SKIP : 0;45 46    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);47 48    // output49    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);50    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);51 52    // if output is NULL, init from the input tok embed53    if (output == NULL) {54        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);55    }56 57    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;58 59    // Calculate dimensions from hyperparameters60    const int64_t head_k_dim = hparams.ssm_d_state;61    const int64_t head_v_dim = hparams.ssm_d_state;62    const int64_t n_k_heads  = hparams.ssm_n_group;63    const int64_t n_v_heads  = hparams.ssm_dt_rank;64    const int64_t key_dim    = head_k_dim * n_k_heads;65    const int64_t value_dim  = head_v_dim * n_v_heads;66    const int64_t conv_dim   = key_dim * 2 + value_dim;67 68    // Calculate projection sizes69    const int64_t qkvz_dim = key_dim * 2 + value_dim * 2;70    const int64_t ba_dim   = n_v_heads * 2;71 72    auto load_block_trunk = [&](int il, int flags) {73        auto & layer = layers[il];74        const uint32_t n_ff_shexp = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : hparams.n_ff(il);75 76        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", il), { n_embd }, flags);77        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, flags);78 79        if (!hparams.is_recr(il)) {80            // Attention layers81            create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, flags);82            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, flags);83            // Q/K normalization for attention layers84            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, flags);85            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, flags);86        } else {87            // Linear attention (gated delta net) specific tensors88            // Create tensors with calculated dimensions89            // note: ssm_in is used by legacy GGUF90            layer.ssm_in         = create_tensor(tn(LLM_TENSOR_SSM_IN,         "weight", il), { n_embd, qkvz_dim }, TENSOR_NOT_REQUIRED | flags);91            layer.wqkv           = create_tensor(tn(LLM_TENSOR_ATTN_QKV,       "weight", il), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED | flags);92            layer.wqkv_gate      = create_tensor(tn(LLM_TENSOR_ATTN_GATE,      "weight", il), { n_embd, value_dim }, TENSOR_NOT_REQUIRED | flags);93            layer.ssm_conv1d     = create_tensor(tn(LLM_TENSOR_SSM_CONV1D,     "weight", il), { hparams.ssm_d_conv, conv_dim }, flags);94            layer.ssm_dt         = create_tensor(tn(LLM_TENSOR_SSM_DT,         "bias",   il), { hparams.ssm_dt_rank }, flags);95            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN,             il), { hparams.ssm_dt_rank }, flags);96            layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", il), { n_embd, ba_dim }, flags);97            layer.ssm_norm       = create_tensor(tn(LLM_TENSOR_SSM_NORM,       "weight", il), { head_v_dim }, flags);98            layer.ssm_out        = create_tensor(tn(LLM_TENSOR_SSM_OUT,        "weight", il), { value_dim, n_embd }, flags);99        }100 101        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", il), { n_embd, n_expert }, flags);102        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { n_ff_exp, n_embd, n_expert }, flags);103        create_tensor_gate_up_exps(layer, il, n_embd, n_ff_exp, n_expert, flags);104 105        // Shared experts106        layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", il), { n_embd }, flags);107        layer.ffn_gate_shexp     = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP,     "weight", il), { n_embd, n_ff_shexp }, flags);108        layer.ffn_up_shexp       = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,       "weight", il), { n_embd, n_ff_shexp }, flags);109        layer.ffn_down_shexp     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP,     "weight", il), { n_ff_shexp, n_embd }, flags);110    };111 112    auto load_block_mtp = [&](int il) {113        // MTP head is identical to the trunk block (full attention + FFN)114        load_block_trunk(il, mtp_flags);115 116        auto & layer = layers[il];117 118        // NextN-specific tensors that define the MTP block.119        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", il), { 2 * n_embd, n_embd }, mtp_flags);120        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", il), { n_embd },             mtp_flags);121        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", il), { n_embd },             mtp_flags);122        layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", il), { n_embd, n_vocab },    mtp_flags | TENSOR_NOT_REQUIRED);123        layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", il), { n_embd, n_vocab },    mtp_flags | TENSOR_NOT_REQUIRED);124        layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd },             mtp_flags | TENSOR_NOT_REQUIRED);125    };126 127    for (int i = 0; i < n_layer; i++) {128        load_block_trunk(i, trunk_flags);129    }130    for (int i = n_layer; i < n_layer_all; i++) {131        load_block_mtp(i);132    }133}134 135std::unique_ptr<llm_graph_context> llama_model_qwen3next::build_arch_graph(const llm_graph_params & params) const {136    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {137        return std::make_unique<graph_mtp>(*this, params);138    }139    return std::make_unique<graph>(*this, params);140}141 142llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_params & params) :143    llm_build_delta_net_base(params), model(model) {144    ggml_tensor * cur;145    ggml_tensor * inpL;146 147    inpL = build_inp_embd(model.tok_embd);148    cb(inpL, "model.embed_tokens", -1);149 150    auto * inp = build_inp_mem_hybrid();151 152    ggml_tensor * inp_pos     = build_inp_pos();153    ggml_tensor * inp_out_ids = build_inp_out_ids();154 155    // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.156    for (int il = 0; il < n_layer; ++il) {157        res->t_layer_inp[il] = inpL;158 159        ggml_tensor * inpSA = inpL;160 161        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);162        cb(cur, "attn_norm", il);163 164        ggml_build_forward_expand(gf, cur);165 166        // Determine layer type and build appropriate attention mechanism167        if (hparams.is_recr(il)) {168            // Linear attention layer (gated delta net)169            cur = build_layer_attn_linear(inp->get_recr(), cur, il);170        } else {171            // Full attention layer172            cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);173        }174 175        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {176            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);177            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);178        }179 180        // Residual connection181        cur = ggml_add(ctx0, cur, inpSA);182        cb(cur, "attn_residual", il);183 184        // Save the tensor before post-attention norm for residual connection185        ggml_tensor * ffn_residual = cur;186 187        // Post-attention norm188        ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);189        cb(attn_post_norm, "attn_post_norm", il);190 191        // FFN layer (MoE or dense) - without residual connection192        cur = build_layer_ffn(attn_post_norm, il);193        cb(cur, "ffn_out", il);194 195        // Residual connection for FFN - add to the tensor from before post_attention_layernorm196        cur = ggml_add(ctx0, cur, ffn_residual);197        cb(cur, "post_moe", il);198 199        cur = build_cvec(cur, il);200        cb(cur, "l_out", il);201 202        // Input for next layer203        inpL = cur;204    }205    cur = inpL;206 207    // post-norm hidden state is input to both the LM head and the MTP head208    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);209 210    cb(cur, "h_nextn", -1);211    res->t_h_nextn = cur;212 213    if (!cparams.embeddings_nextn_masked && inp_out_ids) {214        cur = ggml_get_rows(ctx0, cur, inp_out_ids);215    }216 217    cb(cur, "result_norm", -1);218    res->t_embd = cur;219 220    // LM head221    cur = build_lora_mm(model.output, cur, model.output_s);222 223    cb(cur, "result_output", -1);224    res->t_logits = cur;225 226    ggml_build_forward_expand(gf, cur);227}228 229ggml_tensor * llama_model_qwen3next::graph::build_norm_gated(230        ggml_tensor * input,231        ggml_tensor * weights,232        ggml_tensor * gate,233        int           layer) {234    ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer);235    ggml_tensor * gated_silu = ggml_silu(ctx0, gate);236 237    return ggml_mul(ctx0, normalized, gated_silu);238}239 240ggml_tensor * llama_model_qwen3next::graph::build_layer_attn(241        llm_graph_input_attn_kv * inp,242        ggml_tensor *             cur,243        ggml_tensor *             inp_pos,244        int                       il) {245    const int64_t n_embd_head = hparams.n_embd_head_v();246    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());247 248    // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention249 250    // Qwen3Next uses a single Q projection that outputs query + gate251    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);252    cb(Qcur_full, "Qcur_full", il);253 254    Qcur_full = ggml_reshape_4d(ctx0, Qcur_full, n_embd_head * 2, n_head, n_tokens, 1);255 256    // Split Q projection into query and gate257    // The split should be along dimension 0 (the feature dimension)258    ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,259                                            Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);260    cb(Qcur, "Qcur_view", il);261 262    ggml_tensor * gate =263        ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,264                     Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));265    cb(gate, "gate", il);266 267    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);268    cb(Kcur, "Kcur", il);269 270    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);271    cb(Vcur, "Vcur", il);272 273    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);274    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);275 276    Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);277    cb(Qcur, "Qcur_normed", il);278 279    Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);280    cb(Kcur, "Kcur_normed", il);281 282    Qcur = ggml_rope_ext(283            ctx0, Qcur, inp_pos, nullptr,284            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,285            ext_factor, attn_factor, beta_fast, beta_slow);286 287    Kcur = ggml_rope_ext(288            ctx0, Kcur, inp_pos, nullptr,289            n_rot, rope_type, n_ctx_orig, freq_base,290            freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);291 292    cb(Qcur, "Qcur", il);293    cb(Kcur, "Kcur", il);294    cb(Vcur, "Vcur", il);295 296    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;297 298    cur = build_attn(inp,299                nullptr, nullptr, nullptr,300                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);301    cb(cur, "attn_pregate", il);302 303    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont304    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);305 306    gate = ggml_sigmoid(ctx0, gate);307    cb(gate, "gate_sigmoid", il);308 309    cur = ggml_mul(ctx0, cur, gate);310    cb(cur, "attn_gated", il);311 312    cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);313    cb(cur, "attn_output", il);314 315    return cur;316}317 318std::pair<ggml_tensor *, ggml_tensor *> llama_model_qwen3next::graph::build_qkvz(319                ggml_tensor * input,320                        int   il) {321    const int64_t d_inner      = hparams.ssm_d_inner;322    const int64_t n_seqs       = ubatch.n_seqs;323    const int64_t head_k_dim   = hparams.ssm_d_state;324    const int64_t num_k_heads  = hparams.ssm_n_group;325    const int64_t num_v_heads  = hparams.ssm_dt_rank;326    const int64_t head_v_dim   = d_inner / num_v_heads;327    const int64_t n_seq_tokens = ubatch.n_seq_tokens;328 329    if (model.layers[il].wqkv) {330        // optimized path331        ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);332        qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);333        cb(qkv_mixed, "linear_attn_qkv_mixed", il);334 335        ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);336        cb(z, "z", il);337 338        return { qkv_mixed, z };339    } else {340        // legacy (slower) path341        ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input);342        cb(mixed_qkvz, "linear_attn_mixed_qkvz", il);343 344        int64_t       qkvz_new_dim        = 2 * head_k_dim + 2 * head_v_dim * (num_v_heads / num_k_heads);345        ggml_tensor * mixed_qkvz_reshaped = ggml_reshape_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);346 347        // Split mixed_qkvz into query, key, value, z348        int64_t split_sizes_qkvz[4] = {349            head_k_dim,                              // query size350            head_k_dim,                              // key size351            head_v_dim * num_v_heads / num_k_heads,  // value size352            head_v_dim * num_v_heads / num_k_heads   // z size353        };354 355        ggml_tensor * query =356            ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[0], num_k_heads, n_seq_tokens, n_seqs,357                        mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], 0);358        cb(query, "q", il);359 360        ggml_tensor * key = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[1], num_k_heads, n_seq_tokens, n_seqs,361                                        mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],362                                        split_sizes_qkvz[0] * ggml_element_size(mixed_qkvz_reshaped));363        cb(key, "k", il);364 365        ggml_tensor * value =366            ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[2], num_k_heads, n_seq_tokens, n_seqs,367                        mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],368                        (split_sizes_qkvz[0] + split_sizes_qkvz[1]) * ggml_element_size(mixed_qkvz_reshaped));369        cb(value, "v", il);370 371        ggml_tensor * z = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[3], num_k_heads, n_seq_tokens, n_seqs,372                                    mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],373                                    (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * ggml_element_size(mixed_qkvz_reshaped));374        z = ggml_cont(ctx0, z);375        cb(z, "z", il);376 377        // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions378        // query: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]379        ggml_tensor * query_flat = ggml_cont_3d(ctx0, query, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);380        cb(query_flat, "query_flat", il);381 382        // key: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]383        ggml_tensor * key_flat = ggml_cont_3d(ctx0, key, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);384        cb(key_flat, "key_flat", il);385 386        // value_reshaped: [head_v_dim, num_v_heads, n_tokens, n_seqs] -> [head_v_dim * num_v_heads, n_tokens, n_seqs]387        ggml_tensor * value_flat = ggml_cont_3d(ctx0, value, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);388        cb(value_flat, "value_flat", il);389 390        // Now concatenate along the feature dimension (dim 0) to get [conv_dim, n_tokens, n_seqs]391        ggml_tensor * qkv_mixed = ggml_concat(ctx0, query_flat, key_flat, 0);392        qkv_mixed               = ggml_concat(ctx0, qkv_mixed, value_flat, 0);393        cb(qkv_mixed, "qkv_mixed", il);394 395        return { qkv_mixed, z };396    }397}398 399ggml_tensor * llama_model_qwen3next::graph::build_layer_attn_linear(400        llm_graph_input_rs * inp,401        ggml_tensor *        cur,402        int                  il) {403    const auto * mctx_cur = inp->mctx;404 405    const int64_t d_inner      = hparams.ssm_d_inner;406    const int64_t n_seqs       = ubatch.n_seqs;407    const int64_t head_k_dim   = hparams.ssm_d_state;408    const int64_t num_k_heads  = hparams.ssm_n_group;409    const int64_t num_v_heads  = hparams.ssm_dt_rank;410    const int64_t head_v_dim   = d_inner / num_v_heads;411    const int64_t n_seq_tokens = ubatch.n_seq_tokens;412 413    GGML_ASSERT(n_seqs != 0);414    GGML_ASSERT(ubatch.equal_seqs());415    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);416 417    // Input projections418    auto qkvz = build_qkvz(cur, il);419    ggml_tensor * qkv_mixed = qkvz.first;420    ggml_tensor * z         = qkvz.second;421 422    ggml_tensor * mixed_ba = build_lora_mm(model.layers[il].ssm_beta_alpha, cur);423    cb(mixed_ba, "linear_attn_mixed_ba", il);424 425    // Reshape mixed_ba: [batch, seq_len, hidden_size] -> [batch, seq_len, num_k_heads, 2*num_v_heads/num_k_heads]426    int64_t       ba_new_dim        = 2 * num_v_heads / num_k_heads;427    ggml_tensor * mixed_ba_reshaped = ggml_reshape_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);428 429    // Split mixed_ba into b and a (beta and alpha parameters)430    int64_t split_sizes_ba[2] = {431        num_v_heads / num_k_heads,  // beta size432        num_v_heads / num_k_heads   // alpha size433    };434 435    ggml_tensor * b = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[0], num_k_heads, n_seq_tokens, n_seqs,436                                   mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3], 0);437    cb(b, "b", il);438 439    ggml_tensor * a = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[1], num_k_heads, n_seq_tokens, n_seqs,440                                   mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3],441                                   split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped));442    cb(a, "a", il);443 444    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont445    b = ggml_cont(ctx0, b);446 447    ggml_tensor * beta = ggml_sigmoid(ctx0, b);448 449    // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]450    ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);451 452    ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);453    ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);454    cb(alpha_softplus, "a_softplus", il);455 456    ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a);  // -A_log.exp() * softplus457    cb(gate, "gate", il);458 459    beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);460    gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);461 462    ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);463    ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);464 465    ggml_tensor * conv_kernel      = model.layers[il].ssm_conv1d;466    const int64_t conv_kernel_size = conv_kernel->ne[0];467    const int64_t conv_channels    = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;468 469    ggml_tensor * conv_input = build_conv_state(inp, conv_states_all, qkv_mixed, conv_kernel_size, conv_channels, il);470 471    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);472    state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);473    cb(state, "state_predelta", il);474 475    ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);476    cb(conv_output_proper, "conv_output_raw", il);477 478    ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper);479    cb(conv_output_silu, "conv_output_silu", il);480 481    ggml_tensor * conv_qkv_mix = conv_output_silu;482 483    // Calculate the total conv dimension484    int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads;485    int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);486 487    // Extract the convolved Q, K, V from conv_output488    ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,489            ggml_row_size(conv_qkv_mix->type, head_k_dim),490            nb1_qkv,491            nb1_qkv * n_seq_tokens,492            0);493 494    ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,495            ggml_row_size(conv_qkv_mix->type, head_k_dim),496            nb1_qkv,497            nb1_qkv * n_seq_tokens,498            head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));499 500    ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,501            ggml_row_size(conv_qkv_mix->type, head_v_dim),502            nb1_qkv,503            nb1_qkv * n_seq_tokens,504            ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));505 506    cb(q_conv, "q_conv", il);507    cb(k_conv, "k_conv", il);508    cb(v_conv, "v_conv", il);509 510    const float eps_norm = hparams.f_norm_rms_eps;511 512    q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);513    k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);514 515    //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);516    //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);517    //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);518 519    // if head keys and value keys are different, repeat to force tensors into matching shapes520    // TODO: avoid repeats for fused GDN, needs broadcast configuration for GDN op [TAG_GGML_GDN_BCAST]521    if (num_k_heads != num_v_heads) {522        GGML_ASSERT(num_v_heads % num_k_heads == 0);523        int64_t repeat_factor = num_v_heads / num_k_heads;524 525        // repeat interleave: reshape to (repeat part, 1, remaining part...), do repeat, then reshape back526        ggml_tensor * q_reshaped = ggml_reshape_4d(ctx0, q_conv, head_k_dim, 1, num_k_heads, n_seq_tokens * n_seqs);527        ggml_tensor * k_reshaped = ggml_reshape_4d(ctx0, k_conv, head_k_dim, 1, num_k_heads, n_seq_tokens * n_seqs);528 529        // Repeat along the third dimension (the new dimension with size 1)530        ggml_tensor * q_repeated =531            ggml_repeat_4d(ctx0, q_reshaped, head_k_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs);532        ggml_tensor * k_repeated =533            ggml_repeat_4d(ctx0, k_reshaped, head_k_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs);534 535        // Reshape back to merge the head and repeat dimensions536        // From [head_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs]537        // Back to [head_dim, repeat_factor * num_k_heads, n_seq_tokens, n_seqs]538        q_conv = ggml_reshape_4d(ctx0, q_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);539        k_conv = ggml_reshape_4d(ctx0, k_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);540    }541 542    cb(q_conv, "q_conv_predelta", il);543    cb(k_conv, "k_conv_predelta", il);544    cb(v_conv, "v_conv_predelta", il);545 546    ggml_tensor * output = build_recurrent_attn(inp, ssm_states_all, q_conv, k_conv, v_conv, gate, beta, state, il);547 548    // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]549    ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);550 551    // Apply gated normalization: self.norm(core_attn_out, z)552    ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);553 554    // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]555    ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);556    cb(final_output, "final_output", il);557 558    // Output projection559    cur = build_lora_mm(model.layers[il].ssm_out, final_output);560    cb(cur, "linear_attn_out", il);561 562    // Reshape back to original dimensions563    cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);564 565    return cur;566}567 568ggml_tensor * llama_model_qwen3next::graph::build_layer_ffn(ggml_tensor * cur, const int il) {569    // Check if this is an MoE layer570    if (model.layers[il].ffn_gate_inp != nullptr) {571        // MoE branch572        ggml_tensor * moe_out =573            build_moe_ffn(cur,574                model.layers[il].ffn_gate_inp,575                model.layers[il].ffn_up_exps,576                model.layers[il].ffn_gate_exps,577                model.layers[il].ffn_down_exps,578                nullptr,579                n_expert, n_expert_used,580                LLM_FFN_SILU, true,581                hparams.expert_weights_scale,582                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,583                nullptr, model.layers[il].ffn_gate_up_exps,584                model.layers[il].ffn_up_exps_s,585                model.layers[il].ffn_gate_exps_s,586                model.layers[il].ffn_down_exps_s);587        cb(moe_out, "ffn_moe_out", il);588 589        // Add shared experts if present - following Qwen3Next reference implementation590        if (model.layers[il].ffn_up_shexp != nullptr) {591            ggml_tensor * ffn_shexp =592                build_ffn(cur,593                    model.layers[il].ffn_up_shexp,   NULL, model.layers[il].ffn_up_shexp_s,594                    model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,595                    model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,596                    NULL,597                    LLM_FFN_SILU, LLM_FFN_PAR, il);598            cb(ffn_shexp, "ffn_shexp", il);599 600            // Apply shared expert gating as in the reference implementation601            // The shared expert has its own gate that is sigmoided602            // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token)603            ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);604            cb(shared_gate, "shared_expert_gate", il);605 606            shared_gate = ggml_sigmoid(ctx0, shared_gate);607            cb(shared_gate, "shared_expert_gate_sigmoid", il);608 609            ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);610            cb(ffn_shexp, "ffn_shexp_gated", il);611 612            cur = ggml_add(ctx0, moe_out, ffn_shexp);613            cb(cur, "ffn_out", il);614        } else {615            cur = moe_out;616        }617    } else {618        // Dense FFN branch (not currently used I believe)619        cur = build_ffn(cur,620            model.layers[il].ffn_up, NULL, NULL,621            model.layers[il].ffn_gate, NULL, NULL,622            model.layers[il].ffn_down, NULL, NULL,623            NULL,624            LLM_FFN_SILU, LLM_FFN_PAR, il);625        cb(cur, "ffn_out", il);626    }627    return cur;628}629 630// LLM_GRAPH_TYPE_DECODER_MTP draft head for Qwen3-Next631llama_model_qwen3next::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)632    : llm_graph_context(params) {633    GGML_ASSERT(hparams.n_layer_nextn > 0 && "QWEN3NEXT MTP requires n_layer_nextn > 0");634    GGML_ASSERT(hparams.n_layer_nextn == 1 && "QWEN3NEXT MTP currently only supports a single MTP block");635 636    const int64_t n_embd_head = hparams.n_embd_head_v();637    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());638 639    const int il = hparams.n_layer();640    const auto & layer = model.layers[il];641 642    GGML_ASSERT(layer.nextn.eh_proj    && "MTP block missing nextn.eh_proj");643    GGML_ASSERT(layer.nextn.enorm      && "MTP block missing nextn.enorm");644    GGML_ASSERT(layer.nextn.hnorm      && "MTP block missing nextn.hnorm");645    GGML_ASSERT(layer.ffn_gate_inp     && "MTP block missing ffn_gate_inp");646 647    // TODO: extract in a common llm_graph_context::build_inp_embd_h()648    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);649 650    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);651    ggml_set_input(inp->tokens);652 653    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);654    ggml_set_input(inp->embd);655 656    // TODO: make static using `ggml_build_forward_select()`657    //       see llm_graph_context::build_inp_embd() for reference658    ggml_tensor * tok_embd;659    if (ubatch.token) {660        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;661 662        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);663    } else {664        tok_embd = inp->embd;665    }666    cb(tok_embd, "mtp_tok_embd", il);667 668    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);669    ggml_set_input(inp->h);670    ggml_set_name(inp->h, "mtp_h_input");671 672    ggml_tensor * h_embd = inp->h;673 674    res->add_input(std::move(inp));675 676    ggml_tensor * inp_pos     = build_inp_pos();677    ggml_tensor * inp_out_ids = build_inp_out_ids();678 679    auto * inp_attn = build_attn_inp_kv();680 681    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);682    cb(h_norm, "mtp_hnorm", il);683 684    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);685    cb(e_norm, "mtp_enorm", il);686 687    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);688    cb(concat, "mtp_concat", il);689 690    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);691    cb(cur, "mtp_eh_proj", il);692 693    ggml_tensor * inpSA = cur;694 695    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);696    cb(cur, "mtp_attn_norm", il);697 698    ggml_tensor * Qcur_full = build_lora_mm(layer.wq, cur, layer.wq_s);699    cb(Qcur_full, "mtp_Qcur_full", il);700 701    ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full,702            n_embd_head, n_head, n_tokens,703            ggml_element_size(Qcur_full) * n_embd_head * 2,704            ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,705            0);706    Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);707    cb(Qcur, "mtp_Qcur_normed", il);708 709    ggml_tensor * Kcur = build_lora_mm(layer.wk, cur, layer.wk_s);710    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);711    Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);712    cb(Kcur, "mtp_Kcur_normed", il);713 714    ggml_tensor * Vcur = build_lora_mm(layer.wv, cur, layer.wv_s);715    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);716 717    Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,718            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,719            ext_factor, attn_factor, beta_fast, beta_slow);720    Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,721            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,722            ext_factor, attn_factor, beta_fast, beta_slow);723 724    cb(Qcur, "mtp_Qcur", il);725    cb(Kcur, "mtp_Kcur", il);726    cb(Vcur, "mtp_Vcur", il);727 728    const float kq_scale = hparams.f_attention_scale == 0.0f729            ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;730 731    cur = build_attn(inp_attn,732            nullptr, nullptr, nullptr,733            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);734    cb(cur, "mtp_attn_pregate", il);735 736    ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full,737            n_embd_head, n_head, n_tokens,738            ggml_element_size(Qcur_full) * n_embd_head * 2,739            ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,740            ggml_element_size(Qcur_full) * n_embd_head);741 742    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont743    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);744    cb(gate, "mtp_gate", il);745 746    cur = ggml_mul(ctx0, cur, ggml_sigmoid(ctx0, gate));747    cur = build_lora_mm(layer.wo, cur, layer.wo_s);748    cb(cur, "mtp_attn_out", il);749 750    if (inp_out_ids) {751        cur   = ggml_get_rows(ctx0, cur,   inp_out_ids);752        inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);753    }754 755    cur = ggml_add(ctx0, cur, inpSA);756    cb(cur, "mtp_attn_residual", il);757 758    ggml_tensor * ffn_residual = cur;759    cur = build_norm(cur, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il);760    cb(cur, "mtp_attn_post_norm", il);761 762    // MoE FFN — routed experts plus gated shared expert (mirrors the trunk).763    ggml_tensor * moe_out =764        build_moe_ffn(cur,765            layer.ffn_gate_inp,766            layer.ffn_up_exps,767            layer.ffn_gate_exps,768            layer.ffn_down_exps,769            nullptr,770            n_expert, n_expert_used,771            LLM_FFN_SILU, true,772            hparams.expert_weights_scale,773            LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,774            nullptr, layer.ffn_gate_up_exps,775            layer.ffn_up_exps_s,776            layer.ffn_gate_exps_s,777            layer.ffn_down_exps_s);778    cb(moe_out, "mtp_ffn_moe_out", il);779 780    if (layer.ffn_up_shexp != nullptr) {781        ggml_tensor * ffn_shexp =782            build_ffn(cur,783                layer.ffn_up_shexp,   nullptr, layer.ffn_up_shexp_s,784                layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s,785                layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s,786                nullptr,787                LLM_FFN_SILU, LLM_FFN_PAR, il);788        cb(ffn_shexp, "mtp_ffn_shexp", il);789 790        ggml_tensor * shared_gate = build_lora_mm(layer.ffn_gate_inp_shexp, cur);791        shared_gate = ggml_sigmoid(ctx0, shared_gate);792        cb(shared_gate, "mtp_shared_expert_gate_sigmoid", il);793 794        ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);795        cb(ffn_shexp, "mtp_ffn_shexp_gated", il);796 797        cur = ggml_add(ctx0, moe_out, ffn_shexp);798    } else {799        cur = moe_out;800    }801    cb(cur, "mtp_ffn_out", il);802 803    cur = ggml_add(ctx0, cur, ffn_residual);804    cb(cur, "mtp_post_ffn", il);805 806    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm807            ? layer.nextn.shared_head_norm808            : model.output_norm;809    GGML_ASSERT(head_norm_w && "QWEN3NEXT MTP: missing both nextn.shared_head_norm and output_norm");810    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);811 812    cb(cur, "h_nextn", -1);813    res->t_h_nextn = cur;814 815    ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;816    ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s;817    GGML_ASSERT(head_w && "QWEN3NEXT MTP: missing LM head (nextn.shared_head_head or model.output)");818    cur = build_lora_mm(head_w, cur, head_s);819    cb(cur, "result_output", -1);820 821    res->t_logits = cur;822    ggml_build_forward_expand(gf, cur);823}824