Felipe97/llama-cpp-compiled
01.1k
1#include "models.h"2 3void llama_model_step35::load_arch_hparams(llama_model_loader & ml) {4 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5 6 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;7 8 // full_attention layer only use half of the RoPE dimensions9 hparams.n_rot_full = hparams.n_rot_full / 2;10 11 // MoE + SWA parameters12 ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);13 ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);14 ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false);15 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);16 ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);17 18 // Step35 uses sigmoid gating by default (if not set in GGUF)19 if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {20 hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;21 }22 23 ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);24 ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);25 26 ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);27 28 ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_EXP, hparams.swiglu_clamp_exp, hparams.n_layer_all, false);29 ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_SHEXP, hparams.swiglu_clamp_shexp, hparams.n_layer_all, false);30 31 switch (hparams.n_layer()) {32 case 45: type = LLM_TYPE_196B_A11B; break;33 default: type = LLM_TYPE_UNKNOWN;34 }35}36 37void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {38 LLAMA_LOAD_LOCALS;39 40 const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);41 // Trunk-only: the GGUF declares MTP layers in metadata but the actual MTP42 // tensors live in a separate file (e.g. user split target/draft). Mark43 // MTP tensors NOT_REQUIRED so the trunk loads cleanly.44 const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";45 const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);46 const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;47 int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;48 49 if (!ml.load_mtp) {50 mtp_flags |= TENSOR_SKIP;51 }52 53 tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);54 55 // output56 output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);57 output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, trunk_flags);58 59 // STEP35 supports per-layer partial RoPE dims; rope factors are stored as a single shared tensor60 // ("rope_freqs.weight") and ggml uses only the first (n_rot_l/2) entries per layer.61 uint32_t n_rot_max = 0;62 for (int i = 0; i < n_layer; ++i) {63 n_rot_max = std::max(n_rot_max, hparams.n_rot(i));64 }65 if (n_rot_max == 0) {66 n_rot_max = n_rot;67 }68 69 auto load_block_trunk = [&](int i, int flags) {70 auto & layer = layers[i];71 72 const uint32_t n_head_l = hparams.n_head(i);73 const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);74 const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);75 76 layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);77 layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED);78 layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED);79 80 // optional rope factors (llama3) / longrope tensors81 if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {82 layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));83 layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));84 } else {85 layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));86 }87 88 create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_l, n_embd_k_gqa, n_embd_v_gqa, flags);89 layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_v * n_head_l, n_embd}, flags);90 91 // head-wise attention gate (Step35 self_attn.g_proj)92 layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, TENSOR_NOT_REQUIRED);93 94 layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);95 96 // dense MLP (leading dense blocks)97 layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);98 layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, TENSOR_NOT_REQUIRED);99 layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);100 101 // MoE routed experts + selection bias (router_bias)102 const int64_t n_ff_exp = hparams.n_ff_exp();103 layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);104 layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);105 layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED);106 layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);107 layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);108 109 // shared expert MLP110 layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED);111 layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED);112 layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED);113 };114 115 auto load_block_mtp = [&](int i) {116 auto & layer = layers[i];117 118 const uint32_t n_head_l = hparams.n_head(i);119 const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);120 const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);121 122 // The MTP block is a full Step3p5 decoder layer (mtp_block) plus the123 // NextN-specific wiring (enorm/hnorm/eh_proj + optional shared head).124 // Multi-block MTP: every declared MTP block is required (the draft chain125 // runs all n_layer_nextn heads), so each block uses the captured126 // `mtp_flags` directly — already NOT_REQUIRED for a trunk-only GGUF,127 // which keeps that path correct.128 129 layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, mtp_flags);130 layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED);131 layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, TENSOR_NOT_REQUIRED);132 133 if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {134 layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | TENSOR_DUPLICATED);135 layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | TENSOR_DUPLICATED);136 } else {137 layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot_max/2}, TENSOR_NOT_REQUIRED | TENSOR_DUPLICATED);138 }139 140 create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_l, n_embd_k_gqa, n_embd_v_gqa, mtp_flags);141 layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_v * n_head_l, n_embd}, mtp_flags);142 143 layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_l}, TENSOR_NOT_REQUIRED);144 145 layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, mtp_flags);146 147 // dense MLP (leading dense blocks) — present if the MTP block isn't MoE148 layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);149 layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, TENSOR_NOT_REQUIRED);150 layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);151 152 // MoE routed experts + selection bias (router_bias)153 const int64_t n_ff_exp = hparams.n_ff_exp();154 layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);155 layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);156 layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED);157 layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);158 layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);159 160 layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED);161 layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, TENSOR_NOT_REQUIRED);162 layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED);163 164 // NextN-specific tensors that define the MTP block.165 layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, mtp_flags);166 layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, mtp_flags);167 layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, mtp_flags);168 layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);169 layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);170 layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED);171 };172 173 for (int i = 0; i < n_layer; ++i) {174 load_block_trunk(i, trunk_flags);175 }176 // All n_layer_nextn MTP blocks are required — the multi-block draft chain177 // runs every head (head k at offset k). The GGUF declares the count via178 // step35.nextn_predict_layers.179 for (int i = n_layer; i < n_layer_all; ++i) {180 load_block_mtp(i);181 }182}183 184std::unique_ptr<llm_graph_context> llama_model_step35::build_arch_graph(const llm_graph_params & params) const {185 if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {186 return std::make_unique<graph_mtp>(*this, params);187 }188 return std::make_unique<graph>(*this, params);189}190 191llama_model_step35::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {192 ggml_tensor * cur;193 ggml_tensor * inpL;194 195 inpL = build_inp_embd(model.tok_embd);196 ggml_tensor * inp_pos = build_inp_pos();197 auto * inp_attn = build_attn_inp_kv_iswa();198 ggml_tensor * inp_out_ids = build_inp_out_ids();199 200 // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.201 for (int il = 0; il < n_layer; ++il) {202 ggml_tensor * inpSA = inpL;203 204 const uint32_t n_head_l = hparams.n_head(il);205 const uint32_t n_head_kv_l = hparams.n_head_kv(il);206 207 const float freq_base_l = model.get_rope_freq_base(cparams, il);208 const float freq_scale_l = model.get_rope_freq_scale(cparams, il);209 210 cur = inpL;211 212 // dump pre-attn RMSNorm input to pinpoint layer boundary issues213 cb(cur, "attn_norm_in", il);214 215 // self-attention216 {217 cur = build_norm(cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);218 cb(cur, "attn_norm", il);219 auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,220 n_embd_head_k, n_head_l,221 n_embd_head_k, n_head_kv_l,222 n_embd_head_v, n_head_kv_l,223 il, false);224 225 cb(Qcur, "Qcur", il);226 cb(Kcur, "Kcur", il);227 cb(Vcur, "Vcur", il);228 229 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);230 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);231 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);232 233 // Q/K per-head RMSNorm (Step35 q_norm / k_norm)234 if (model.layers[il].attn_q_norm) {235 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);236 cb(Qcur, "Qcur_normed", il);237 }238 if (model.layers[il].attn_k_norm) {239 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);240 cb(Kcur, "Kcur_normed", il);241 }242 243 // RoPE (partial rotary factors per layer)244 const bool is_swa = hparams.is_swa(il);245 ggml_tensor * rope_factors = is_swa ? nullptr : model.get_rope_factors(cparams, il);246 const int64_t n_rot_l = hparams.n_rot(il);247 Qcur = ggml_rope_ext(248 ctx0, Qcur, inp_pos, rope_factors,249 n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,250 ext_factor, attn_factor, beta_fast, beta_slow251 );252 Kcur = ggml_rope_ext(253 ctx0, Kcur, inp_pos, rope_factors,254 n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,255 ext_factor, attn_factor, beta_fast, beta_slow256 );257 cb(Qcur, "Qcur_pos", il);258 cb(Kcur, "Kcur_pos", il);259 260 const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k));261 ggml_tensor * attn_out = build_attn(inp_attn,262 nullptr, nullptr, nullptr,263 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);264 cb(attn_out, "attn_out", il);265 // head-wise attention gate: sigmoid(g_proj(x)) in torch266 if (model.layers[il].wqkv_gate) {267 ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, cur); // [n_head_l, n_tokens]268 cb(gate, "attn_gate", il);269 270 gate = ggml_sigmoid(ctx0, gate);271 cb(gate, "attn_gate_sigmoid", il);272 273 // reshape + broadcast to [n_embd_head_v, n_head_l, n_tokens]274 ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, attn_out, n_embd_head_v, n_head_l, n_tokens);275 ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate, 1, n_head_l, n_tokens);276 cb(gate_3d, "attn_gate_3d", il);277 278 attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);279 cb(attn_3d, "attn_gated_3d", il);280 281 attn_out = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v * n_head_l, n_tokens);282 cb(attn_out, "attn_gated", il);283 }284 285 // output projection286 cur = build_lora_mm(model.layers[il].wo, attn_out, model.layers[il].wo_s);287 cb(cur, "attn_proj", il);288 }289 290 if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {291 cur = ggml_get_rows(ctx0, cur, inp_out_ids);292 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);293 }294 295 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);296 cb(ffn_inp, "ffn_inp", il);297 298 cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);299 cb(cur, "ffn_norm", il);300 301 // feed-forward302 if (model.layers[il].ffn_gate_inp == nullptr) {303 // dense MLP304 cur = build_ffn(cur,305 model.layers[il].ffn_up, model.layers[il].ffn_up_b, nullptr,306 model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, nullptr,307 model.layers[il].ffn_down, model.layers[il].ffn_down_b, nullptr,308 nullptr,309 LLM_FFN_SILU, LLM_FFN_PAR, il);310 cb(cur, "ffn_out", il);311 } else {312 // MoE routed experts313 ggml_tensor * moe_out = build_moe_ffn(cur,314 model.layers[il].ffn_gate_inp,315 model.layers[il].ffn_up_exps,316 model.layers[il].ffn_gate_exps,317 model.layers[il].ffn_down_exps,318 model.layers[il].ffn_exp_probs_b,319 n_expert, n_expert_used,320 LLM_FFN_SILU, hparams.expert_weights_norm,321 hparams.expert_weights_scale,322 (llama_expert_gating_func_type) hparams.expert_gating_func,323 il);324 cb(moe_out, "ffn_moe_out", il);325 326 // shared expert MLP (always added on MoE layers in Step35)327 ggml_tensor * sh_out = build_ffn(cur,328 model.layers[il].ffn_up_shexp, nullptr, nullptr,329 model.layers[il].ffn_gate_shexp, nullptr, nullptr,330 model.layers[il].ffn_down_shexp, nullptr, nullptr,331 nullptr,332 LLM_FFN_SILU, LLM_FFN_PAR, il);333 cb(sh_out, "ffn_shared_out", il);334 335 cur = ggml_add(ctx0, moe_out, sh_out);336 cb(cur, "ffn_out", il);337 }338 cur = ggml_add(ctx0, cur, ffn_inp);339 340 cur = build_cvec(cur, il);341 cb(cur, "l_out", il);342 343 // input for next layer344 inpL = cur;345 }346 347 cur = inpL;348 349 cb(cur, "h_nextn", -1);350 res->t_h_nextn = cur;351 352 if (!cparams.embeddings_nextn_masked && inp_out_ids) {353 cur = ggml_get_rows(ctx0, cur, inp_out_ids);354 }355 356 cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);357 cb(cur, "result_norm", -1);358 res->t_embd = cur;359 360 cur = build_lora_mm(model.output, cur, model.output_s);361 cb(cur, "result_output", -1);362 res->t_logits = cur;363 364 ggml_build_forward_expand(gf, cur);365}366 367// LLM_GRAPH_TYPE_DECODER_MTP draft head for Step3p5 (MoE)368llama_model_step35::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)369 : llm_graph_context(params) {370 GGML_ASSERT(hparams.n_layer_nextn > 0 && "STEP35 MTP requires n_layer_nextn > 0");371 372 // Multi-block MTP: the DECODER_MTP graph runs the MTP head selected by373 // cparams.nextn_layer_offset (0 = first trained head). The speculative driver374 // bumps the offset per draft step to chain heads 45->46->47. offset 0 keeps375 // single-block behavior identical to before.376 const int il = hparams.n_layer() + cparams.nextn_layer_offset;377 GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&378 cparams.nextn_layer_offset < (int) hparams.n_layer_nextn &&379 "nextn_layer_offset out of range [0, n_layer_nextn)");380 const auto & layer = model.layers[il];381 382 GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");383 GGML_ASSERT(layer.nextn.enorm && "MTP block missing nextn.enorm");384 GGML_ASSERT(layer.nextn.hnorm && "MTP block missing nextn.hnorm");385 386 const uint32_t n_head_l = hparams.n_head(il);387 const uint32_t n_head_kv_l = hparams.n_head_kv(il);388 389 const float freq_base_l = model.get_rope_freq_base(cparams, il);390 const float freq_scale_l = model.get_rope_freq_scale(cparams, il);391 392 auto inp = std::make_unique<llm_graph_input_embd>(hparams.n_embd);393 394 inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);395 ggml_set_input(inp->tokens);396 397 inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);398 ggml_set_input(inp->embd);399 ggml_set_name(inp->embd, "mtp_h_input");400 401 ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;402 403 ggml_tensor * h_input = inp->embd;404 ggml_tensor * tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);405 cb(tok_embd, "mtp_tok_embd", il);406 407 res->add_input(std::move(inp));408 409 ggml_tensor * inp_pos = build_inp_pos();410 auto * inp_attn = build_attn_inp_kv_iswa();411 412 ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);413 cb(h_norm, "mtp_hnorm", il);414 415 ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);416 cb(e_norm, "mtp_enorm", il);417 418 ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);419 cb(concat, "mtp_concat", il);420 421 ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat);422 cb(cur, "mtp_eh_proj", il);423 424 ggml_tensor * inpSA = cur;425 426 // mtp_block: full Step3p5 decoder layer (attention with optional head-wise gate, then MoE/dense FFN)427 cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);428 cb(cur, "mtp_attn_norm", il);429 430 auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur,431 n_embd_head_k, n_head_l,432 n_embd_head_k, n_head_kv_l,433 n_embd_head_v, n_head_kv_l,434 il, false);435 cb(Qcur, "mtp_Qcur", il);436 cb(Kcur, "mtp_Kcur", il);437 cb(Vcur, "mtp_Vcur", il);438 439 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);440 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);441 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);442 443 if (layer.attn_q_norm) {444 Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);445 cb(Qcur, "mtp_Qcur_normed", il);446 }447 if (layer.attn_k_norm) {448 Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);449 cb(Kcur, "mtp_Kcur_normed", il);450 }451 452 const bool is_swa = hparams.is_swa(il);453 ggml_tensor * rope_factors = is_swa ? nullptr : model.get_rope_factors(cparams, il);454 const int64_t n_rot_l = hparams.n_rot(il);455 456 Qcur = ggml_rope_ext(457 ctx0, Qcur, inp_pos, rope_factors,458 n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,459 ext_factor, attn_factor, beta_fast, beta_slow);460 Kcur = ggml_rope_ext(461 ctx0, Kcur, inp_pos, rope_factors,462 n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,463 ext_factor, attn_factor, beta_fast, beta_slow);464 cb(Qcur, "mtp_Qcur_pos", il);465 cb(Kcur, "mtp_Kcur_pos", il);466 467 const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k));468 ggml_tensor * attn_out = build_attn(inp_attn,469 nullptr, nullptr, nullptr,470 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);471 cb(attn_out, "mtp_attn_out", il);472 473 // head-wise attention gate: sigmoid(g_proj(x))474 if (layer.wqkv_gate) {475 ggml_tensor * gate = build_lora_mm(layer.wqkv_gate, cur); // [n_head_l, n_tokens]476 cb(gate, "mtp_attn_gate", il);477 478 gate = ggml_sigmoid(ctx0, gate);479 cb(gate, "mtp_attn_gate_sigmoid", il);480 481 ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, attn_out, n_embd_head_v, n_head_l, n_tokens);482 ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate, 1, n_head_l, n_tokens);483 cb(gate_3d, "mtp_attn_gate_3d", il);484 485 attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);486 cb(attn_3d, "mtp_attn_gated_3d", il);487 488 attn_out = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v * n_head_l, n_tokens);489 cb(attn_out, "mtp_attn_gated", il);490 }491 492 cur = build_lora_mm(layer.wo, attn_out, layer.wo_s);493 cb(cur, "mtp_attn_proj", il);494 495 cur = ggml_add(ctx0, cur, inpSA);496 cb(cur, "mtp_attn_residual", il);497 498 ggml_tensor * ffn_inp = cur;499 cur = build_norm(cur, layer.ffn_norm, nullptr, LLM_NORM_RMS, il);500 cb(cur, "mtp_ffn_norm", il);501 502 // FFN: dense MLP or MoE (mirrors trunk path)503 if (layer.ffn_gate_inp == nullptr) {504 cur = build_ffn(cur,505 layer.ffn_up, layer.ffn_up_b, nullptr,506 layer.ffn_gate, layer.ffn_gate_b, nullptr,507 layer.ffn_down, layer.ffn_down_b, nullptr,508 nullptr,509 LLM_FFN_SILU, LLM_FFN_PAR, il);510 cb(cur, "mtp_ffn_out", il);511 } else {512 ggml_tensor * moe_out = build_moe_ffn(cur,513 layer.ffn_gate_inp,514 layer.ffn_up_exps,515 layer.ffn_gate_exps,516 layer.ffn_down_exps,517 layer.ffn_exp_probs_b,518 n_expert, n_expert_used,519 LLM_FFN_SILU, hparams.expert_weights_norm,520 hparams.expert_weights_scale,521 (llama_expert_gating_func_type) hparams.expert_gating_func,522 il);523 cb(moe_out, "mtp_ffn_moe_out", il);524 525 ggml_tensor * sh_out = build_ffn(cur,526 layer.ffn_up_shexp, nullptr, nullptr,527 layer.ffn_gate_shexp, nullptr, nullptr,528 layer.ffn_down_shexp, nullptr, nullptr,529 nullptr,530 LLM_FFN_SILU, LLM_FFN_PAR, il);531 cb(sh_out, "mtp_ffn_shared_out", il);532 533 cur = ggml_add(ctx0, moe_out, sh_out);534 cb(cur, "mtp_ffn_out", il);535 }536 cur = ggml_add(ctx0, cur, ffn_inp);537 cb(cur, "mtp_post_ffn", il);538 539 ggml_tensor * inp_out_ids = build_inp_out_ids();540 cur = ggml_get_rows(ctx0, cur, inp_out_ids);541 542 // Pre-norm hidden state: used by the AR draft loop to seed the next MTP step.543 cb(cur, "h_nextn", -1);544 res->t_h_nextn = cur;545 546 ggml_tensor * head_norm_w = layer.nextn.shared_head_norm547 ? layer.nextn.shared_head_norm548 : model.output_norm;549 GGML_ASSERT(head_norm_w && "STEP35 MTP: missing both nextn.shared_head_norm and output_norm");550 cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);551 cb(cur, "mtp_shared_head_norm", -1);552 553 ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;554 GGML_ASSERT(head_w && "STEP35 MTP: missing LM head (nextn.shared_head_head or model.output)");555 cur = build_lora_mm(head_w, cur);556 cb(cur, "result_output", -1);557 558 res->t_logits = cur;559 ggml_build_forward_expand(gf, cur);560}561