Felipe97/llama-cpp-compiled
01.1k
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_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, 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 // Mark recurrent layers (linear attention layers).17 if (!ml.get_key_or_arr(LLM_KV_ATTENTION_RECURRENT_LAYERS, hparams.is_recr_impl, hparams.n_layer_all, false)) {18 uint32_t full_attn_interval = 4;19 ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);20 for (uint32_t i = 0; i < hparams.n_layer_all; ++i) {21 hparams.is_recr_impl[i] = (i < hparams.n_layer()) && ((i + 1) % full_attn_interval != 0);22 }23 }24 25 switch (hparams.n_layer()) {26 case 48: type = LLM_TYPE_80B_A3B; break;27 default: type = LLM_TYPE_UNKNOWN;28 }29}30 31void llama_model_qwen3next::load_arch_tensors(llama_model_loader & ml) {32 LLAMA_LOAD_LOCALS;33 34 if (n_expert == 0) {35 throw std::runtime_error(arch_name() + " model cannot have zero experts");36 }37 38 const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);39 const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;40 int mtp_flags = !ml.load_mtp ? TENSOR_SKIP : 0;41 42 tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);43 44 // output45 output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);46 output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);47 48 // if output is NULL, init from the input tok embed49 if (output == NULL) {50 output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);51 }52 53 const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;54 55 // Calculate dimensions from hyperparameters56 const int64_t head_k_dim = hparams.ssm_d_state;57 const int64_t head_v_dim = hparams.ssm_d_state;58 const int64_t n_k_heads = hparams.ssm_n_group;59 const int64_t n_v_heads = hparams.ssm_dt_rank;60 const int64_t key_dim = head_k_dim * n_k_heads;61 const int64_t value_dim = head_v_dim * n_v_heads;62 const int64_t conv_dim = key_dim * 2 + value_dim;63 64 // Calculate projection sizes65 const int64_t qkvz_dim = key_dim * 2 + value_dim * 2;66 const int64_t ba_dim = n_v_heads * 2;67 68 auto load_block_trunk = [&](int il, int flags) {69 auto & layer = layers[il];70 const uint32_t n_ff_shexp = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : hparams.n_ff(il);71 72 layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, flags);73 layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, flags);74 75 if (!hparams.is_recr(il)) {76 // Attention layers77 create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, flags);78 layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, flags);79 // Q/K normalization for attention layers80 layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, flags);81 layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, flags);82 } else {83 // Linear attention (gated delta net) specific tensors84 // Create tensors with calculated dimensions85 // note: ssm_in is used by legacy GGUF86 layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", il), { n_embd, qkvz_dim }, TENSOR_NOT_REQUIRED | flags);87 layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", il), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED | flags);88 layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", il), { n_embd, value_dim }, TENSOR_NOT_REQUIRED | flags);89 layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", il), { hparams.ssm_d_conv, conv_dim }, flags);90 layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", il), { hparams.ssm_dt_rank }, flags);91 layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, il), { hparams.ssm_dt_rank }, flags);92 layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", il), { n_embd, ba_dim }, flags);93 layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", il), { head_v_dim }, flags);94 layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", il), { value_dim, n_embd }, flags);95 }96 97 layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", il), { n_embd, n_expert }, flags);98 layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { n_ff_exp, n_embd, n_expert }, flags);99 create_tensor_gate_up_exps(layer, il, n_embd, n_ff_exp, n_expert, flags);100 101 // Shared experts102 layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", il), { n_embd }, flags);103 layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", il), { n_embd, n_ff_shexp }, flags);104 layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", il), { n_embd, n_ff_shexp }, flags);105 layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", il), { n_ff_shexp, n_embd }, flags);106 };107 108 auto load_block_mtp = [&](int il) {109 // MTP head is identical to the trunk block (full attention + FFN)110 load_block_trunk(il, mtp_flags);111 112 auto & layer = layers[il];113 114 // NextN-specific tensors that define the MTP block.115 layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", il), { 2 * n_embd, n_embd }, mtp_flags);116 layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", il), { n_embd }, mtp_flags);117 layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", il), { n_embd }, mtp_flags);118 layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", il), { n_embd, n_vocab }, mtp_flags | TENSOR_NOT_REQUIRED);119 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);120 layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd }, mtp_flags | TENSOR_NOT_REQUIRED);121 };122 123 for (int i = 0; i < n_layer; i++) {124 load_block_trunk(i, trunk_flags);125 }126 for (int i = n_layer; i < n_layer_all; i++) {127 load_block_mtp(i);128 }129}130 131std::unique_ptr<llm_graph_context> llama_model_qwen3next::build_arch_graph(const llm_graph_params & params) const {132 if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {133 return std::make_unique<graph_mtp>(*this, params);134 }135 return std::make_unique<graph>(*this, params);136}137 138llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_params & params) :139 llm_build_delta_net_base(params), model(model) {140 ggml_tensor * cur;141 ggml_tensor * inpL;142 143 inpL = build_inp_embd(model.tok_embd);144 cb(inpL, "model.embed_tokens", -1);145 146 auto * inp = build_inp_mem_hybrid();147 148 ggml_tensor * inp_pos = build_inp_pos();149 ggml_tensor * inp_out_ids = build_inp_out_ids();150 151 // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.152 for (int il = 0; il < n_layer; ++il) {153 res->t_layer_inp[il] = inpL;154 155 ggml_tensor * inpSA = inpL;156 157 cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);158 cb(cur, "attn_norm", il);159 160 ggml_build_forward_expand(gf, cur);161 162 // Determine layer type and build appropriate attention mechanism163 if (hparams.is_recr(il)) {164 // Linear attention layer (gated delta net)165 cur = build_layer_attn_linear(inp->get_recr(), cur, il);166 } else {167 // Full attention layer168 cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);169 }170 171 if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {172 cur = ggml_get_rows(ctx0, cur, inp_out_ids);173 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);174 }175 176 // Residual connection177 cur = ggml_add(ctx0, cur, inpSA);178 cb(cur, "attn_residual", il);179 180 // Save the tensor before post-attention norm for residual connection181 ggml_tensor * ffn_residual = cur;182 183 // Post-attention norm184 ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);185 cb(attn_post_norm, "attn_post_norm", il);186 187 // FFN layer (MoE or dense) - without residual connection188 cur = build_layer_ffn(attn_post_norm, il);189 cb(cur, "ffn_out", il);190 191 // Residual connection for FFN - add to the tensor from before post_attention_layernorm192 cur = ggml_add(ctx0, cur, ffn_residual);193 cb(cur, "post_moe", il);194 195 cur = build_cvec(cur, il);196 cb(cur, "l_out", il);197 198 // Input for next layer199 inpL = cur;200 }201 cur = inpL;202 203 // post-norm hidden state is input to both the LM head and the MTP head204 cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);205 206 cb(cur, "h_nextn", -1);207 res->t_h_nextn = cur;208 209 if (!cparams.embeddings_nextn_masked && inp_out_ids) {210 cur = ggml_get_rows(ctx0, cur, inp_out_ids);211 }212 213 cb(cur, "result_norm", -1);214 res->t_embd = cur;215 216 // LM head217 cur = build_lora_mm(model.output, cur, model.output_s);218 219 cb(cur, "result_output", -1);220 res->t_logits = cur;221 222 ggml_build_forward_expand(gf, cur);223}224 225ggml_tensor * llama_model_qwen3next::graph::build_norm_gated(226 ggml_tensor * input,227 ggml_tensor * weights,228 ggml_tensor * gate,229 int layer) {230 ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer);231 ggml_tensor * gated_silu = ggml_silu(ctx0, gate);232 233 return ggml_mul(ctx0, normalized, gated_silu);234}235 236ggml_tensor * llama_model_qwen3next::graph::build_layer_attn(237 llm_graph_input_attn_kv * inp,238 ggml_tensor * cur,239 ggml_tensor * inp_pos,240 int il) {241 const int64_t n_embd_head = hparams.n_embd_head_v();242 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());243 244 // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention245 246 // Qwen3Next uses a single Q projection that outputs query + gate247 auto [Qcur_full, Kcur, Vcur] = build_qkv(model.layers[il], cur,248 n_embd_head * 2, n_head,249 n_embd_head, n_head_kv,250 n_embd_head, n_head_kv,251 il, false);252 cb(Qcur_full, "Qcur_full", il);253 cb(Kcur, "Kcur", il);254 cb(Vcur, "Vcur", il);255 256 Qcur_full = ggml_reshape_4d(ctx0, Qcur_full, n_embd_head * 2, n_head, n_tokens, 1);257 258 // Split Q projection into query and gate259 // The split should be along dimension 0 (the feature dimension)260 ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,261 Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0);262 cb(Qcur, "Qcur_view", il);263 264 ggml_tensor * gate =265 ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1,266 Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));267 cb(gate, "gate", il);268 269 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);270 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);271 272 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);273 cb(Qcur, "Qcur_normed", il);274 275 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);276 cb(Kcur, "Kcur_normed", il);277 278 Qcur = ggml_rope_ext(279 ctx0, Qcur, inp_pos, nullptr,280 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,281 ext_factor, attn_factor, beta_fast, beta_slow);282 283 Kcur = ggml_rope_ext(284 ctx0, Kcur, inp_pos, nullptr,285 n_rot, rope_type, n_ctx_orig, freq_base,286 freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);287 288 cb(Qcur, "Qcur", il);289 cb(Kcur, "Kcur", il);290 cb(Vcur, "Vcur", il);291 292 const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;293 294 cur = build_attn(inp,295 nullptr, nullptr, nullptr,296 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);297 cb(cur, "attn_pregate", il);298 299 // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont300 gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);301 302 gate = ggml_sigmoid(ctx0, gate);303 cb(gate, "gate_sigmoid", il);304 305 cur = ggml_mul(ctx0, cur, gate);306 cb(cur, "attn_gated", il);307 308 cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);309 cb(cur, "attn_output", il);310 311 return cur;312}313 314std::pair<ggml_tensor *, ggml_tensor *> llama_model_qwen3next::graph::build_qkvz(315 ggml_tensor * input,316 int il) {317 const int64_t d_inner = hparams.ssm_d_inner;318 const int64_t n_seqs = ubatch.n_seqs;319 const int64_t head_k_dim = hparams.ssm_d_state;320 const int64_t num_k_heads = hparams.ssm_n_group;321 const int64_t num_v_heads = hparams.ssm_dt_rank;322 const int64_t head_v_dim = d_inner / num_v_heads;323 const int64_t n_seq_tokens = ubatch.n_seq_tokens;324 325 if (model.layers[il].wqkv) {326 // optimized path327 ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input);328 qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs);329 cb(qkv_mixed, "linear_attn_qkv_mixed", il);330 331 ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input);332 cb(z, "z", il);333 334 return { qkv_mixed, z };335 } else {336 // legacy (slower) path337 ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input);338 cb(mixed_qkvz, "linear_attn_mixed_qkvz", il);339 340 int64_t qkvz_new_dim = 2 * head_k_dim + 2 * head_v_dim * (num_v_heads / num_k_heads);341 ggml_tensor * mixed_qkvz_reshaped = ggml_reshape_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);342 343 // Split mixed_qkvz into query, key, value, z344 int64_t split_sizes_qkvz[4] = {345 head_k_dim, // query size346 head_k_dim, // key size347 head_v_dim * num_v_heads / num_k_heads, // value size348 head_v_dim * num_v_heads / num_k_heads // z size349 };350 351 ggml_tensor * query =352 ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[0], num_k_heads, n_seq_tokens, n_seqs,353 mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3], 0);354 cb(query, "q", il);355 356 ggml_tensor * key = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[1], num_k_heads, n_seq_tokens, n_seqs,357 mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],358 split_sizes_qkvz[0] * ggml_element_size(mixed_qkvz_reshaped));359 cb(key, "k", il);360 361 ggml_tensor * value =362 ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[2], num_k_heads, n_seq_tokens, n_seqs,363 mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],364 (split_sizes_qkvz[0] + split_sizes_qkvz[1]) * ggml_element_size(mixed_qkvz_reshaped));365 cb(value, "v", il);366 367 ggml_tensor * z = ggml_view_4d(ctx0, mixed_qkvz_reshaped, split_sizes_qkvz[3], num_k_heads, n_seq_tokens, n_seqs,368 mixed_qkvz_reshaped->nb[1], mixed_qkvz_reshaped->nb[2], mixed_qkvz_reshaped->nb[3],369 (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * ggml_element_size(mixed_qkvz_reshaped));370 z = ggml_cont(ctx0, z);371 cb(z, "z", il);372 373 // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions374 // query: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]375 ggml_tensor * query_flat = ggml_cont_3d(ctx0, query, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);376 cb(query_flat, "query_flat", il);377 378 // key: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]379 ggml_tensor * key_flat = ggml_cont_3d(ctx0, key, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);380 cb(key_flat, "key_flat", il);381 382 // value_reshaped: [head_v_dim, num_v_heads, n_tokens, n_seqs] -> [head_v_dim * num_v_heads, n_tokens, n_seqs]383 ggml_tensor * value_flat = ggml_cont_3d(ctx0, value, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);384 cb(value_flat, "value_flat", il);385 386 // Now concatenate along the feature dimension (dim 0) to get [conv_dim, n_tokens, n_seqs]387 ggml_tensor * qkv_mixed = ggml_concat(ctx0, query_flat, key_flat, 0);388 qkv_mixed = ggml_concat(ctx0, qkv_mixed, value_flat, 0);389 cb(qkv_mixed, "qkv_mixed", il);390 391 return { qkv_mixed, z };392 }393}394 395ggml_tensor * llama_model_qwen3next::graph::build_layer_attn_linear(396 llm_graph_input_rs * inp,397 ggml_tensor * cur,398 int il) {399 const auto * mctx_cur = inp->mctx;400 401 const int64_t d_inner = hparams.ssm_d_inner;402 const int64_t n_seqs = ubatch.n_seqs;403 const int64_t head_k_dim = hparams.ssm_d_state;404 const int64_t num_k_heads = hparams.ssm_n_group;405 const int64_t num_v_heads = hparams.ssm_dt_rank;406 const int64_t head_v_dim = d_inner / num_v_heads;407 const int64_t n_seq_tokens = ubatch.n_seq_tokens;408 409 GGML_ASSERT(n_seqs != 0);410 GGML_ASSERT(ubatch.equal_seqs());411 GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);412 413 // Input projections414 auto qkvz = build_qkvz(cur, il);415 ggml_tensor * qkv_mixed = qkvz.first;416 ggml_tensor * z = qkvz.second;417 418 ggml_tensor * mixed_ba = build_lora_mm(model.layers[il].ssm_beta_alpha, cur);419 cb(mixed_ba, "linear_attn_mixed_ba", il);420 421 // Reshape mixed_ba: [batch, seq_len, hidden_size] -> [batch, seq_len, num_k_heads, 2*num_v_heads/num_k_heads]422 int64_t ba_new_dim = 2 * num_v_heads / num_k_heads;423 ggml_tensor * mixed_ba_reshaped = ggml_reshape_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);424 425 // Split mixed_ba into b and a (beta and alpha parameters)426 int64_t split_sizes_ba[2] = {427 num_v_heads / num_k_heads, // beta size428 num_v_heads / num_k_heads // alpha size429 };430 431 ggml_tensor * b = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[0], num_k_heads, n_seq_tokens, n_seqs,432 mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3], 0);433 cb(b, "b", il);434 435 ggml_tensor * a = ggml_view_4d(ctx0, mixed_ba_reshaped, split_sizes_ba[1], num_k_heads, n_seq_tokens, n_seqs,436 mixed_ba_reshaped->nb[1], mixed_ba_reshaped->nb[2], mixed_ba_reshaped->nb[3],437 split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped));438 cb(a, "a", il);439 440 // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont441 b = ggml_cont(ctx0, b);442 443 ggml_tensor * beta = ggml_sigmoid(ctx0, b);444 445 // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]446 ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);447 448 ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);449 ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);450 cb(alpha_softplus, "a_softplus", il);451 452 ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus453 cb(gate, "gate", il);454 455 beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);456 gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);457 458 ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);459 ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);460 461 ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d;462 const int64_t conv_kernel_size = conv_kernel->ne[0];463 const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;464 465 ggml_tensor * conv_input = build_conv_state(inp, conv_states_all, qkv_mixed, conv_kernel_size, conv_channels, il);466 467 ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);468 state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);469 cb(state, "state_predelta", il);470 471 ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);472 cb(conv_output_proper, "conv_output_raw", il);473 474 ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper);475 cb(conv_output_silu, "conv_output_silu", il);476 477 ggml_tensor * conv_qkv_mix = conv_output_silu;478 479 // Calculate the total conv dimension480 int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads;481 int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);482 483 // Extract the convolved Q, K, V from conv_output484 ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,485 ggml_row_size(conv_qkv_mix->type, head_k_dim),486 nb1_qkv,487 nb1_qkv * n_seq_tokens,488 0);489 490 ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,491 ggml_row_size(conv_qkv_mix->type, head_k_dim),492 nb1_qkv,493 nb1_qkv * n_seq_tokens,494 head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));495 496 ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,497 ggml_row_size(conv_qkv_mix->type, head_v_dim),498 nb1_qkv,499 nb1_qkv * n_seq_tokens,500 ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));501 502 cb(q_conv, "q_conv", il);503 cb(k_conv, "k_conv", il);504 cb(v_conv, "v_conv", il);505 506 507 const float eps_norm = hparams.f_norm_rms_eps;508 509 q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm);510 k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm);511 512 //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);513 //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);514 //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);515 516 // if head keys and value keys are different, repeat to force tensors into matching shapes517 // TODO: avoid repeats for fused GDN, needs broadcast configuration for GDN op [TAG_GGML_GDN_BCAST]518 if (num_k_heads != num_v_heads) {519 GGML_ASSERT(num_v_heads % num_k_heads == 0);520 int64_t repeat_factor = num_v_heads / num_k_heads;521 522 // repeat interleave: reshape to (repeat part, 1, remaining part...), do repeat, then reshape back523 ggml_tensor * q_reshaped = ggml_reshape_4d(ctx0, q_conv, head_k_dim, 1, num_k_heads, n_seq_tokens * n_seqs);524 ggml_tensor * k_reshaped = ggml_reshape_4d(ctx0, k_conv, head_k_dim, 1, num_k_heads, n_seq_tokens * n_seqs);525 526 // Repeat along the third dimension (the new dimension with size 1)527 ggml_tensor * q_repeated =528 ggml_repeat_4d(ctx0, q_reshaped, head_k_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs);529 ggml_tensor * k_repeated =530 ggml_repeat_4d(ctx0, k_reshaped, head_k_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs);531 532 // Reshape back to merge the head and repeat dimensions533 // From [head_dim, repeat_factor, num_k_heads, n_seq_tokens * n_seqs]534 // Back to [head_dim, repeat_factor * num_k_heads, n_seq_tokens, n_seqs]535 q_conv = ggml_reshape_4d(ctx0, q_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);536 k_conv = ggml_reshape_4d(ctx0, k_repeated, head_k_dim, num_k_heads * repeat_factor, n_seq_tokens, n_seqs);537 }538 539 cb(q_conv, "q_conv_predelta", il);540 cb(k_conv, "k_conv_predelta", il);541 cb(v_conv, "v_conv_predelta", il);542 543 ggml_tensor * output = build_recurrent_attn(inp, ssm_states_all, q_conv, k_conv, v_conv, gate, beta, state, il);544 545 // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]546 ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);547 548 // Apply gated normalization: self.norm(core_attn_out, z)549 ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);550 551 // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]552 ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);553 cb(final_output, "final_output", il);554 555 // Output projection556 cur = build_lora_mm(model.layers[il].ssm_out, final_output);557 cb(cur, "linear_attn_out", il);558 559 // Reshape back to original dimensions560 cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);561 562 return cur;563}564 565ggml_tensor * llama_model_qwen3next::graph::build_layer_ffn(ggml_tensor * cur, const int il) {566 // Check if this is an MoE layer567 if (model.layers[il].ffn_gate_inp != nullptr) {568 // MoE branch569 ggml_tensor * moe_out =570 build_moe_ffn(cur,571 model.layers[il].ffn_gate_inp,572 model.layers[il].ffn_up_exps,573 model.layers[il].ffn_gate_exps,574 model.layers[il].ffn_down_exps,575 nullptr,576 n_expert, n_expert_used,577 LLM_FFN_SILU, true,578 hparams.expert_weights_scale,579 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,580 nullptr, model.layers[il].ffn_gate_up_exps,581 model.layers[il].ffn_up_exps_s,582 model.layers[il].ffn_gate_exps_s,583 model.layers[il].ffn_down_exps_s);584 cb(moe_out, "ffn_moe_out", il);585 586 // Add shared experts if present - following Qwen3Next reference implementation587 if (model.layers[il].ffn_up_shexp != nullptr) {588 ggml_tensor * ffn_shexp =589 build_ffn(cur,590 model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s,591 model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,592 model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,593 NULL,594 LLM_FFN_SILU, LLM_FFN_PAR, il);595 cb(ffn_shexp, "ffn_shexp", il);596 597 // Apply shared expert gating as in the reference implementation598 // The shared expert has its own gate that is sigmoided599 // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token)600 ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur);601 cb(shared_gate, "shared_expert_gate", il);602 603 shared_gate = ggml_sigmoid(ctx0, shared_gate);604 cb(shared_gate, "shared_expert_gate_sigmoid", il);605 606 ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);607 cb(ffn_shexp, "ffn_shexp_gated", il);608 609 cur = ggml_add(ctx0, moe_out, ffn_shexp);610 cb(cur, "ffn_out", il);611 } else {612 cur = moe_out;613 }614 } else {615 // Dense FFN branch (not currently used I believe)616 cur = build_ffn(cur,617 model.layers[il].ffn_up, NULL, NULL,618 model.layers[il].ffn_gate, NULL, NULL,619 model.layers[il].ffn_down, NULL, NULL,620 NULL,621 LLM_FFN_SILU, LLM_FFN_PAR, il);622 cb(cur, "ffn_out", il);623 }624 return cur;625}626 627// LLM_GRAPH_TYPE_DECODER_MTP draft head for Qwen3-Next628llama_model_qwen3next::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)629 : llm_graph_context(params) {630 GGML_ASSERT(hparams.n_layer_nextn > 0 && "QWEN3NEXT MTP requires n_layer_nextn > 0");631 GGML_ASSERT(hparams.n_layer_nextn == 1 && "QWEN3NEXT MTP currently only supports a single MTP block");632 633 const int64_t n_embd_head = hparams.n_embd_head_v();634 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());635 636 const int il = hparams.n_layer();637 const auto & layer = model.layers[il];638 639 GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");640 GGML_ASSERT(layer.nextn.enorm && "MTP block missing nextn.enorm");641 GGML_ASSERT(layer.nextn.hnorm && "MTP block missing nextn.hnorm");642 GGML_ASSERT(layer.ffn_gate_inp && "MTP block missing ffn_gate_inp");643 644 // TODO: extract in a common llm_graph_context::build_inp_embd_h()645 auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);646 647 inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);648 ggml_set_input(inp->tokens);649 650 inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);651 ggml_set_input(inp->embd);652 653 // TODO: make static using `ggml_build_forward_select()`654 // see llm_graph_context::build_inp_embd() for reference655 ggml_tensor * tok_embd;656 if (ubatch.token) {657 ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;658 659 tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);660 } else {661 tok_embd = inp->embd;662 }663 cb(tok_embd, "mtp_tok_embd", il);664 665 inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);666 ggml_set_input(inp->h);667 ggml_set_name(inp->h, "mtp_h_input");668 669 ggml_tensor * h_embd = inp->h;670 671 res->add_input(std::move(inp));672 673 ggml_tensor * inp_pos = build_inp_pos();674 ggml_tensor * inp_out_ids = build_inp_out_ids();675 676 auto * inp_attn = build_attn_inp_kv();677 678 ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);679 cb(h_norm, "mtp_hnorm", il);680 681 ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);682 cb(e_norm, "mtp_enorm", il);683 684 ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);685 cb(concat, "mtp_concat", il);686 687 ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);688 cb(cur, "mtp_eh_proj", il);689 690 ggml_tensor * inpSA = cur;691 692 cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);693 cb(cur, "mtp_attn_norm", il);694 695 auto [Qcur_full, Kcur, Vcur] = build_qkv(layer, cur,696 n_embd_head * 2, n_head,697 n_embd_head, n_head_kv,698 n_embd_head, n_head_kv,699 il, false);700 cb(Qcur_full, "mtp_Qcur_full", il);701 702 ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full,703 n_embd_head, n_head, n_tokens,704 ggml_element_size(Qcur_full) * n_embd_head * 2,705 ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,706 0);707 Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);708 cb(Qcur, "mtp_Qcur_normed", il);709 710 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 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);715 716 Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,717 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,718 ext_factor, attn_factor, beta_fast, beta_slow);719 Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,720 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,721 ext_factor, attn_factor, beta_fast, beta_slow);722 723 cb(Qcur, "mtp_Qcur", il);724 cb(Kcur, "mtp_Kcur", il);725 cb(Vcur, "mtp_Vcur", il);726 727 const float kq_scale = hparams.f_attention_scale == 0.0f728 ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;729 730 cur = build_attn(inp_attn,731 nullptr, nullptr, nullptr,732 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);733 cb(cur, "mtp_attn_pregate", il);734 735 ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full,736 n_embd_head, n_head, n_tokens,737 ggml_element_size(Qcur_full) * n_embd_head * 2,738 ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,739 ggml_element_size(Qcur_full) * n_embd_head);740 741 // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont742 gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);743 cb(gate, "mtp_gate", il);744 745 cur = ggml_mul(ctx0, cur, ggml_sigmoid(ctx0, gate));746 cur = build_lora_mm(layer.wo, cur, layer.wo_s);747 cb(cur, "mtp_attn_out", il);748 749 if (inp_out_ids) {750 cur = ggml_get_rows(ctx0, cur, inp_out_ids);751 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);752 }753 754 cur = ggml_add(ctx0, cur, inpSA);755 cb(cur, "mtp_attn_residual", il);756 757 ggml_tensor * ffn_residual = cur;758 cur = build_norm(cur, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il);759 cb(cur, "mtp_attn_post_norm", il);760 761 // MoE FFN — routed experts plus gated shared expert (mirrors the trunk).762 ggml_tensor * moe_out =763 build_moe_ffn(cur,764 layer.ffn_gate_inp,765 layer.ffn_up_exps,766 layer.ffn_gate_exps,767 layer.ffn_down_exps,768 nullptr,769 n_expert, n_expert_used,770 LLM_FFN_SILU, true,771 hparams.expert_weights_scale,772 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,773 nullptr, layer.ffn_gate_up_exps,774 layer.ffn_up_exps_s,775 layer.ffn_gate_exps_s,776 layer.ffn_down_exps_s);777 cb(moe_out, "mtp_ffn_moe_out", il);778 779 if (layer.ffn_up_shexp != nullptr) {780 ggml_tensor * ffn_shexp =781 build_ffn(cur,782 layer.ffn_up_shexp, nullptr, layer.ffn_up_shexp_s,783 layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s,784 layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s,785 nullptr,786 LLM_FFN_SILU, LLM_FFN_PAR, il);787 cb(ffn_shexp, "mtp_ffn_shexp", il);788 789 ggml_tensor * shared_gate = build_lora_mm(layer.ffn_gate_inp_shexp, cur);790 shared_gate = ggml_sigmoid(ctx0, shared_gate);791 cb(shared_gate, "mtp_shared_expert_gate_sigmoid", il);792 793 ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);794 cb(ffn_shexp, "mtp_ffn_shexp_gated", il);795 796 cur = ggml_add(ctx0, moe_out, ffn_shexp);797 } else {798 cur = moe_out;799 }800 cb(cur, "mtp_ffn_out", il);801 802 cur = ggml_add(ctx0, cur, ffn_residual);803 cb(cur, "mtp_post_ffn", il);804 805 ggml_tensor * head_norm_w = layer.nextn.shared_head_norm806 ? layer.nextn.shared_head_norm807 : model.output_norm;808 GGML_ASSERT(head_norm_w && "QWEN3NEXT MTP: missing both nextn.shared_head_norm and output_norm");809 cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);810 811 cb(cur, "h_nextn", -1);812 res->t_h_nextn = cur;813 814 ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;815 ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s;816 GGML_ASSERT(head_w && "QWEN3NEXT MTP: missing LM head (nextn.shared_head_head or model.output)");817 cur = build_lora_mm(head_w, cur, head_s);818 cb(cur, "result_output", -1);819 820 res->t_logits = cur;821 ggml_build_forward_expand(gf, cur);822}823