Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_hy_v3::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);5    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);6    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);7    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func, false);8    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);9    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);10 11    // HY V3 uses a sigmoid router with expert selection bias by default12    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {13        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;14    }15 16    switch (hparams.n_layer()) {17        case 48: type = LLM_TYPE_30B_A3B; break;18        default: type = LLM_TYPE_UNKNOWN;19    }20}21 22void llama_model_hy_v3::load_arch_tensors(llama_model_loader & ml) {23    LLAMA_LOAD_LOCALS;24 25    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);26    // Trunk-only: the GGUF declares MTP layers in metadata but the actual MTP27    // tensors live in a separate file (e.g. user split target/draft). Mark28    // MTP tensors NOT_REQUIRED so the trunk loads cleanly.29    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";30    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);31    const int trunk_flags = mtp_only   ? TENSOR_NOT_REQUIRED : 0;32    int mtp_flags         = trunk_only ? TENSOR_NOT_REQUIRED : 0;33 34    if (!ml.load_mtp) {35        mtp_flags |= TENSOR_SKIP;36    }37 38    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);39 40    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);41    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);42    if (output == NULL) {43        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);44    }45 46    auto load_block = [&](int i, int flags) {47        auto & layer = layers[i];48        const int64_t n_ff_exp   = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / (n_expert_used > 0 ? n_expert_used : 1);49        const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp;50 51        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);52 53        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);54        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags);55 56        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);57        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);58 59        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);60 61        // dense FFN (leading dense blocks, first_k_dense_replace)62        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);63        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, TENSOR_NOT_REQUIRED);64        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);65 66        // MoE routed experts (sigmoid router + expert selection bias)67        layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);68        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B,           i), {n_expert}, TENSOR_NOT_REQUIRED);69        layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,   "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED);70        create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, TENSOR_NOT_REQUIRED);71 72        // shared expert (always active, no gate)73        layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED);74        layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED);75        layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED);76    };77 78    for (int i = 0; i < n_layer; ++i) {79        load_block(i, trunk_flags);80    }81 82    // NextN/MTP block(s): a full hy_v3 decoder block plus the NextN projections.83    for (int i = n_layer; i < n_layer_all; ++i) {84        auto & layer = layers[i];85 86        load_block(i, mtp_flags);87 88        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), { 2 * n_embd, n_embd }, mtp_flags);89        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), { n_embd },             mtp_flags);90        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), { n_embd },             mtp_flags);91        layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), { n_embd, n_vocab },    TENSOR_NOT_REQUIRED);92        layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab },    TENSOR_NOT_REQUIRED);93        // hy_v3 stores the MTP block's trailing final_layernorm here (applied94        // after the decoder block, before the shared LM head).95        layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd },             TENSOR_NOT_REQUIRED);96    }97}98 99std::unique_ptr<llm_graph_context> llama_model_hy_v3::build_arch_graph(const llm_graph_params & params) const {100    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {101        return std::make_unique<graph_mtp>(*this, params);102    }103    return std::make_unique<graph>(*this, params);104}105 106llama_model_hy_v3::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {107    const int64_t n_embd_head = hparams.n_embd_head_v();108 109    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());110    GGML_ASSERT(n_embd_head == n_rot);111 112    ggml_tensor * cur;113    ggml_tensor * inpL;114 115    inpL = build_inp_embd(model.tok_embd);116    ggml_tensor * inp_pos = build_inp_pos();117    auto * inp_attn = build_attn_inp_kv();118    ggml_tensor * inp_out_ids = build_inp_out_ids();119 120    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));121 122    // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.123    for (int il = 0; il < n_layer; ++il) {124        ggml_tensor * inpSA = inpL;125 126        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);127        cb(cur, "attn_norm", il);128 129        // self-attention130        {131            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);132 133            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head, n_head_kv, il);134 135            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);136            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);137 138            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors,139                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,140                    ext_factor, attn_factor, beta_fast, beta_slow);141            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors,142                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,143                    ext_factor, attn_factor, beta_fast, beta_slow);144 145            cur = build_attn(inp_attn,146                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,147                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);148            cb(cur, "attn_out", il);149        }150 151        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {152            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);153            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);154        }155 156        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);157        cb(ffn_inp, "ffn_inp", il);158 159        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);160        cb(cur, "ffn_norm", il);161 162        if (model.layers[il].ffn_gate_inp == nullptr) {163            // dense FFN (leading dense blocks)164            cur = build_ffn(cur,165                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,166                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,167                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,168                    nullptr,169                    LLM_FFN_SILU, LLM_FFN_PAR, il);170            cb(cur, "ffn_dense_out", il);171        } else {172            // MoE routed experts (sigmoid gating + expert selection bias)173            ggml_tensor * moe_out = build_moe_ffn(cur,174                    model.layers[il].ffn_gate_inp,175                    model.layers[il].ffn_up_exps,176                    model.layers[il].ffn_gate_exps,177                    model.layers[il].ffn_down_exps,178                    model.layers[il].ffn_exp_probs_b,179                    n_expert, n_expert_used,180                    LLM_FFN_SILU,181                    hparams.expert_weights_norm,182                    hparams.expert_weights_scale,183                    (llama_expert_gating_func_type) hparams.expert_gating_func,184                    il,185                    nullptr, model.layers[il].ffn_gate_up_exps,186                    model.layers[il].ffn_up_exps_s,187                    model.layers[il].ffn_gate_exps_s,188                    model.layers[il].ffn_down_exps_s);189            cb(moe_out, "ffn_moe_out", il);190 191            // shared expert (always active, no gate)192            ggml_tensor * sh_out = build_ffn(cur,193                    model.layers[il].ffn_up_shexp,   nullptr, model.layers[il].ffn_up_shexp_s,194                    model.layers[il].ffn_gate_shexp, nullptr, model.layers[il].ffn_gate_shexp_s,195                    model.layers[il].ffn_down_shexp, nullptr, model.layers[il].ffn_down_shexp_s,196                    nullptr,197                    LLM_FFN_SILU, LLM_FFN_PAR, il);198            cb(sh_out, "ffn_shared_out", il);199 200            cur = ggml_add(ctx0, moe_out, sh_out);201            cb(cur, "ffn_out", il);202        }203 204        cur = ggml_add(ctx0, cur, ffn_inp);205        cur = build_cvec(cur, il);206        cb(cur, "l_out", il);207 208        inpL = cur;209    }210 211    cur = build_norm(inpL, model.output_norm, nullptr, LLM_NORM_RMS, -1);212 213    // Post-final-norm hidden state: what the MTP draft head's hnorm consumes.214    // vLLM feeds the target model's normed output states, and the MTP layer215    // itself returns final_layernorm(h), so the chained state is post-norm.216    cb(cur, "h_nextn", -1);217    res->t_h_nextn = cur;218 219    if (!cparams.embeddings_nextn_masked && inp_out_ids) {220        cur = ggml_get_rows(ctx0, cur, inp_out_ids);221    }222 223    cb(cur, "result_norm", -1);224    res->t_embd = cur;225 226    cur = build_lora_mm(model.output, cur, model.output_s);227    cb(cur, "result_output", -1);228    res->t_logits = cur;229 230    ggml_build_forward_expand(gf, cur);231}232 233// LLM_GRAPH_TYPE_DECODER_MTP draft head for HY V3 (MoE).234// Semantics mirror vLLM's HYV3MultiTokenPredictorLayer (hy_v3_mtp.py):235//   enorm(embed) + hnorm(prev_hidden) -> concat(e, h) -> eh_proj ->236//   hy_v3 decoder block -> final_layernorm (stored as nextn.shared_head_norm) ->237//   shared LM head (the main model's lm_head; the checkpoint has no separate238//   MTP head or MTP embeddings).239llama_model_hy_v3::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)240    : llm_graph_context(params) {241    GGML_ASSERT(hparams.n_layer_nextn > 0 && "HY_V3 MTP requires n_layer_nextn > 0");242 243    const int64_t n_embd_head = hparams.n_embd_head_v();244    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());245    GGML_ASSERT(n_embd_head == n_rot);246 247    const int il = hparams.n_layer() + cparams.nextn_layer_offset;248    GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&249                cparams.nextn_layer_offset < (int) hparams.n_layer_nextn &&250                "nextn_layer_offset out of range [0, n_layer_nextn)");251    const auto & layer = model.layers[il];252 253    GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");254    GGML_ASSERT(layer.nextn.enorm   && "MTP block missing nextn.enorm");255    GGML_ASSERT(layer.nextn.hnorm   && "MTP block missing nextn.hnorm");256 257    auto inp = std::make_unique<llm_graph_input_embd>(hparams.n_embd);258 259    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);260    ggml_set_input(inp->tokens);261 262    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);263    ggml_set_input(inp->embd);264    ggml_set_name(inp->embd, "mtp_h_input");265 266    ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;267 268    ggml_tensor * h_input  = inp->embd;269    ggml_tensor * tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);270    cb(tok_embd, "mtp_tok_embd", il);271 272    res->add_input(std::move(inp));273 274    ggml_tensor * inp_pos     = build_inp_pos();275    ggml_tensor * inp_out_ids = build_inp_out_ids();276    auto * inp_attn           = build_attn_inp_kv();277 278    ggml_tensor * h_norm = build_norm(h_input, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);279    cb(h_norm, "mtp_hnorm", il);280 281    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);282    cb(e_norm, "mtp_enorm", il);283 284    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);285    cb(concat, "mtp_concat", il);286 287    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat);288    cb(cur, "mtp_eh_proj", il);289 290    ggml_tensor * inpSA = cur;291 292    // mtp_block: a full hy_v3 decoder layer (mirrors the trunk graph)293    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);294    cb(cur, "mtp_attn_norm", il);295 296    {297        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);298 299        auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, n_head, n_head_kv, il);300 301        Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);302        Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);303 304        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors,305                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,306                ext_factor, attn_factor, beta_fast, beta_slow);307        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors,308                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,309                ext_factor, attn_factor, beta_fast, beta_slow);310 311        const float kq_scale = 1.0f / sqrtf(float(n_embd_head));312 313        cur = build_attn(inp_attn,314                layer.wo, layer.wo_b, layer.wo_s,315                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);316        cb(cur, "mtp_attn_out", il);317    }318 319    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);320    cb(ffn_inp, "mtp_ffn_inp", il);321 322    cur = build_norm(ffn_inp, layer.ffn_norm, nullptr, LLM_NORM_RMS, il);323    cb(cur, "mtp_ffn_norm", il);324 325    if (layer.ffn_gate_inp == nullptr) {326        cur = build_ffn(cur,327                layer.ffn_up,   layer.ffn_up_b,   layer.ffn_up_s,328                layer.ffn_gate, layer.ffn_gate_b, layer.ffn_gate_s,329                layer.ffn_down, layer.ffn_down_b, layer.ffn_down_s,330                nullptr,331                LLM_FFN_SILU, LLM_FFN_PAR, il);332        cb(cur, "mtp_ffn_dense_out", il);333    } else {334        ggml_tensor * moe_out = build_moe_ffn(cur,335                layer.ffn_gate_inp,336                layer.ffn_up_exps,337                layer.ffn_gate_exps,338                layer.ffn_down_exps,339                layer.ffn_exp_probs_b,340                n_expert, n_expert_used,341                LLM_FFN_SILU,342                hparams.expert_weights_norm,343                hparams.expert_weights_scale,344                (llama_expert_gating_func_type) hparams.expert_gating_func,345                il,346                nullptr, layer.ffn_gate_up_exps,347                layer.ffn_up_exps_s,348                layer.ffn_gate_exps_s,349                layer.ffn_down_exps_s);350        cb(moe_out, "mtp_ffn_moe_out", il);351 352        ggml_tensor * sh_out = build_ffn(cur,353                layer.ffn_up_shexp,   nullptr, layer.ffn_up_shexp_s,354                layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s,355                layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s,356                nullptr,357                LLM_FFN_SILU, LLM_FFN_PAR, il);358        cb(sh_out, "mtp_ffn_shared_out", il);359 360        cur = ggml_add(ctx0, moe_out, sh_out);361        cb(cur, "mtp_ffn_out", il);362    }363 364    cur = ggml_add(ctx0, cur, ffn_inp);365    cb(cur, "mtp_post_ffn", il);366 367    // final_layernorm applied after the decoder block, before the shared head.368    // The post-norm hidden state seeds the next MTP step (matches vLLM, where369    // HYV3MultiTokenPredictorLayer returns final_layernorm(h)).370    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm371            ? layer.nextn.shared_head_norm372            : model.output_norm;373    GGML_ASSERT(head_norm_w && "HY_V3 MTP: missing both nextn.shared_head_norm and output_norm");374    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);375 376    cb(cur, "h_nextn", -1);377    res->t_h_nextn = cur;378 379    cur = ggml_get_rows(ctx0, cur, inp_out_ids);380    cb(cur, "mtp_shared_head_norm", -1);381 382    ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;383    ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s;384    GGML_ASSERT(head_w && "HY_V3 MTP: missing LM head (nextn.shared_head_head or model.output)");385    cur = build_lora_mm(head_w, cur, head_s);386    cb(cur, "result_output", -1);387 388    res->t_logits = cur;389    ggml_build_forward_expand(gf, cur);390}391