Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_deepseek2::load_arch_hparams(llama_model_loader & ml) {4    uint32_t n_vocab = 0;5    ml.get_key(LLM_KV_VOCAB_SIZE, n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, n_vocab, false);6 7    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B, Kanana-2-30B-A3B8    const bool is_lite = (hparams.n_layer() == 27 || hparams.n_layer() == 26 || (hparams.n_layer() == 48 && n_vocab == 128256));9 10    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);11    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);12    if (!is_lite) {13        ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);14    }15    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);16    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);17    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);18    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);19    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);20    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);21    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);22    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,         hparams.expert_gating_func, false);23    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {24        // for compatibility with existing DeepSeek V2 and V2.5 GGUFs25        // that have no expert_gating_func model parameter set26        if ((hparams.n_layer() == 47 || hparams.n_layer() == 48) && n_vocab == 154880) {27            // GLM 4.7 Lite28            hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;29        } else {30            hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;31        }32    }33 34    if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false)) {35        // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]36        // cancel the factor from the convert script37        hparams.rope_yarn_log_mul /= 0.1f;38    }39 40    // NextN/MTP41    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);42    GGML_ASSERT(hparams.n_layer_nextn == 0 ||43        hparams.n_layer() + hparams.n_layer_nextn == hparams.n_layer_all);44 45    // (optional) temperature tuning - used by mistral-large46    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);47    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false); // FIXME why not use temperature_length?48 49    hparams.f_attn_temp_offset = 0.0f;50 51    switch (hparams.n_layer()) {52        case 27: type = LLM_TYPE_16B; break;53        case 47: type = LLM_TYPE_30B_A3B; break;54        case 60: type = LLM_TYPE_236B; break;55        case 61: type = LLM_TYPE_671B; break;56        default: type = LLM_TYPE_UNKNOWN;57    }58}59 60void llama_model_deepseek2::load_arch_tensors(llama_model_loader & ml) {61    LLAMA_LOAD_LOCALS;62    const int64_t n_expert_shared = hparams.n_expert_shared;63 64    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);65    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";66    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);67    const int trunk_flags = mtp_only  ? TENSOR_NOT_REQUIRED : 0;68    int       mtp_flags   = trunk_only ? TENSOR_NOT_REQUIRED : 0;69 70    if (!ml.load_mtp) {71        mtp_flags |= TENSOR_SKIP;72    }73 74    const bool is_mla = hparams.is_mla();75 76    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA77    const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();78    const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();79 80    const int64_t n_embd_head_qk_rope = hparams.n_rot();81    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;82    GGML_ASSERT(n_embd_head_qk_nope >= 1);83 84    const int64_t q_lora_rank  = hparams.n_lora_q;85    const int64_t kv_lora_rank = hparams.n_lora_kv;86 87    const int64_t n_ff_exp        = hparams.n_ff_exp;88 89    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);90 91    // output92    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);93    // try to load output.weight, if not found, use token_embd (tied embeddings)94    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);95    if (!output) {96        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);97    }98 99    for (int i = 0; i < n_layer_all; ++i) {100        auto & layer = layers[i];101        const int flags = i < n_layer ? trunk_flags : mtp_flags;102 103        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);104        if (q_lora_rank > 0) {105            layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, flags);106        }107 108        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);109 110        if (q_lora_rank > 0) {111            layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);112            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, flags);113        } else {114            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, flags);115        }116 117        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);118 119        // note: only old legacy GGUF files will have the unsplit wkv_b tensor in120        if (is_mla) {121            layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, flags);122            layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, flags);123        } else {124            layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, flags);125        }126 127        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, flags);128 129        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);130 131        if (i < (int) hparams.n_layer_dense_lead) {132            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);133            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);134            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);135        } else {136            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);137            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);138 139            if (n_expert == 0) {140                throw std::runtime_error("n_expert must be > 0");141            }142            if (n_expert_used == 0) {143                throw std::runtime_error("n_expert_used must be > 0");144            }145 146            // MoE branch147            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);148            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, flags);149 150            // Shared expert branch151            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);152            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, flags);153            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);154        }155 156        // NextN/MTP tensors157        if (i >= n_layer) {158            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, mtp_flags);159            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, mtp_flags);160            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, mtp_flags);161            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);162            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);163            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);164        }165    }166}167 168std::unique_ptr<llm_graph_context> llama_model_deepseek2::build_arch_graph(const llm_graph_params & params) const {169    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {170        return std::make_unique<graph_mtp>(*this, params);171    }172    return std::make_unique<graph>(*this, params);173}174 175llama_model_deepseek2::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) :176    llm_graph_context(params) {177    GGML_ASSERT(hparams.n_layer_nextn > 0 && "GLM4 MTP requires n_layer_nextn > 0");178    GGML_ASSERT(hparams.n_layer_nextn == 1 && "GLM4 MTP currently only supports a single MTP block");179    GGML_ASSERT(hparams.is_mla() && "GLM4 MTP requires MLA");180    GGML_ASSERT(hparams.f_attn_temp_scale == 0.0f && "GLM4 MTP does not support attention temperature scaling");181 182    // The appended MTP block is stored immediately after the main decoder layers.183    const int il = hparams.n_layer();184    const auto & layer = model.layers[il];185 186    GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");187    GGML_ASSERT(layer.nextn.enorm   && "MTP block missing nextn.enorm");188    GGML_ASSERT(layer.nextn.hnorm   && "MTP block missing nextn.hnorm");189 190    GGML_ASSERT((uint32_t) il >= hparams.n_layer_dense_lead && "GLM4 MTP block expected to use MoE FFN");191 192    const int64_t n_embd_head_k_mla   = hparams.n_embd_head_k_mla();193    const int64_t n_embd_head_qk_rope = hparams.n_rot();194    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;195    const int64_t kv_lora_rank        = hparams.n_lora_kv;196 197    GGML_ASSERT(n_embd_head_qk_nope >= 1);198    GGML_ASSERT(hparams.n_lora_q > 0);199    GGML_ASSERT(layer.wq_a);200    GGML_ASSERT(layer.attn_q_a_norm);201    GGML_ASSERT(layer.wq_b);202    GGML_ASSERT(layer.wkv_a_mqa);203    GGML_ASSERT(layer.attn_kv_a_norm);204    GGML_ASSERT(layer.wk_b);205 206    const bool has_split_exps =207            layer.ffn_up_exps   != nullptr &&208            layer.ffn_gate_exps != nullptr;209 210    const bool has_fused_exps = layer.ffn_gate_up_exps != nullptr;211 212    GGML_ASSERT(has_split_exps || has_fused_exps);213    GGML_ASSERT(layer.ffn_norm);214    GGML_ASSERT(layer.ffn_gate_inp);215    GGML_ASSERT(layer.ffn_down_exps);216    GGML_ASSERT(layer.ffn_gate_shexp);217    GGML_ASSERT(layer.ffn_down_shexp);218    GGML_ASSERT(layer.ffn_up_shexp);219 220    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);221 222    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);223    ggml_set_input(inp->tokens);224 225    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);226    ggml_set_input(inp->embd);227 228    ggml_tensor * tok_embd;229    if (ubatch.token) {230        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens231                ? layer.nextn.embed_tokens232                : model.tok_embd;233 234        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);235    } else {236        tok_embd = inp->embd;237    }238    cb(tok_embd, "mtp_tok_embd", il);239 240    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);241    ggml_set_input(inp->h);242    ggml_set_name(inp->h, "mtp_h_input");243 244    ggml_tensor * h_embd = inp->h;245 246    res->add_input(std::move(inp));247 248    ggml_tensor * inp_pos     = build_inp_pos();249    ggml_tensor * inp_out_ids = build_inp_out_ids();250 251    auto * inp_attn_k = build_attn_inp_k();252 253    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);254    cb(h_norm, "mtp_hnorm", il);255 256    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);257    cb(e_norm, "mtp_enorm", il);258 259    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, 0);260    cb(concat, "mtp_concat", il);261 262    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);263    cb(cur, "mtp_eh_proj", il);264 265    ggml_tensor * inpSA = cur;266 267    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);268    cb(cur, "mtp_attn_norm", il);269 270    ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_a, cur);271    cb(q, "mtp_q_a", il);272 273    q = build_norm(q, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);274    cb(q, "mtp_q_a_norm", il);275 276    q = ggml_mul_mat(ctx0, layer.wq_b, q);277    cb(q, "mtp_q_b", il);278 279    ggml_tensor * q_nope =280        ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,281                ggml_row_size(q->type, n_embd_head_k_mla),282                ggml_row_size(q->type, n_embd_head_k_mla) * n_head, 0);283    cb(q_nope, "mtp_q_nope", il);284 285    ggml_tensor * q_pe =286        ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,287                ggml_row_size(q->type, n_embd_head_k_mla),288                ggml_row_size(q->type, n_embd_head_k_mla) * n_head,289                ggml_row_size(q->type, n_embd_head_qk_nope));290    cb(q_pe, "mtp_q_pe", il);291 292    ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);293    cb(kv_cmpr_pe, "mtp_kv_cmpr_pe", il);294 295    ggml_tensor * kv_cmpr =296        ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,297                ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);298    cb(kv_cmpr, "mtp_kv_cmpr", il);299 300    ggml_tensor * k_pe =301        ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,302                ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),303                ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),304                ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));305    cb(k_pe, "mtp_k_pe", il);306 307    kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);308    cb(kv_cmpr, "mtp_kv_cmpr_norm", il);309 310    GGML_ASSERT(ext_factor >= 0.0f);311 312    const float attn_factor_org =313            attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));314 315    const float mscale =316            attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));317 318    const float kq_scale =319            1.0f * mscale * mscale / sqrtf(float(n_embd_head_k_mla));320 321    q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr,322            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,323            ext_factor, attn_factor, beta_fast, beta_slow);324    cb(q_pe, "mtp_q_pe_rope", il);325 326    k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr,327            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,328            ext_factor, attn_factor, beta_fast, beta_slow);329    cb(k_pe, "mtp_k_pe_rope", il);330 331    q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);332    cb(q_nope, "mtp_q_nope_perm", il);333 334    ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);335    cb(q_nope_absorbed, "mtp_q_nope_absorbed", il);336 337    q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);338    cb(q_nope_absorbed, "mtp_q_nope_absorbed_perm", il);339 340    ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);341    cb(Qcur, "mtp_Qcur", il);342 343    kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, hparams.n_lora_kv, 1, n_tokens);344    cb(kv_cmpr, "mtp_kv_cmpr_reshape", il);345 346    ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);347    cb(Kcur, "mtp_Kcur", il);348 349    ggml_tensor * Vcur = kv_cmpr;350    cb(Vcur, "mtp_Vcur", il);351 352    cur = build_attn(inp_attn_k,353            layer.wo, nullptr, layer.wo_s,354            Qcur, Kcur, Vcur, nullptr, nullptr, layer.wv_b, kq_scale, il);355    cb(cur, "mtp_attn_out", il);356 357    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);358    cb(ffn_inp, "mtp_ffn_inp", il);359 360    cur = build_norm(ffn_inp, layer.ffn_norm, nullptr, LLM_NORM_RMS, il);361    cb(cur, "mtp_ffn_norm", il);362 363    ggml_tensor * moe_out = build_moe_ffn(cur,364            layer.ffn_gate_inp,365            layer.ffn_up_exps,366            layer.ffn_gate_exps,367            layer.ffn_down_exps,368            layer.ffn_exp_probs_b,369            n_expert, n_expert_used,370            LLM_FFN_SILU, hparams.expert_weights_norm,371            hparams.expert_weights_scale,372            (llama_expert_gating_func_type) hparams.expert_gating_func,373            il,374            nullptr,375            layer.ffn_gate_up_exps);376    cb(moe_out, "mtp_ffn_moe_out", il);377 378    ggml_tensor * ffn_shexp = build_ffn(cur,379            layer.ffn_up_shexp, nullptr, nullptr,380            layer.ffn_gate_shexp, nullptr, nullptr,381            layer.ffn_down_shexp, nullptr, nullptr,382            nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);383    cb(ffn_shexp, "mtp_ffn_shexp", il);384 385    cur = ggml_add(ctx0, moe_out, ffn_shexp);386    cb(cur, "mtp_ffn_out", il);387 388    cur = ggml_add(ctx0, cur, ffn_inp);389    cb(cur, "mtp_post_ffn", il);390 391    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm392            ? layer.nextn.shared_head_norm393            : model.output_norm;394    GGML_ASSERT(head_norm_w && "GLM4 MTP: missing both nextn.shared_head_norm and output_norm");395 396    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);397    cb(cur, "h_nextn", -1);398    res->t_h_nextn = cur;399 400    if (inp_out_ids) {401        cur = ggml_get_rows(ctx0, cur, inp_out_ids);402    }403    cb(cur, "mtp_shared_head_norm", -1);404 405    ggml_tensor * head_w = layer.nextn.shared_head_head406            ? layer.nextn.shared_head_head407            : model.output;408 409    ggml_tensor * head_s = layer.nextn.shared_head_head410            ? layer.nextn.shared_head_head_s411            : model.output_s;412 413    GGML_ASSERT(head_w && "GLM4 MTP: missing LM head (nextn.shared_head_head or model.output)");414 415    cur = build_lora_mm(head_w, cur, head_s);416    cb(cur, "result_output", -1);417 418    res->t_logits = cur;419    ggml_build_forward_expand(gf, cur);420}421 422llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_params & params) :423    llm_graph_context(params) {424    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B425    bool is_ocr = model.arch == LLM_ARCH_DEEPSEEK2OCR;426 427    const bool is_mla = hparams.is_mla();428 429    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA430    const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();431    const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();432 433    const int64_t n_embd_head_qk_rope = hparams.n_rot();434    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;435 436    const uint32_t kv_lora_rank = hparams.n_lora_kv;437 438    // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.439    // See https://github.com/ggml-org/llama.cpp/discussions/7416 for detailed explanation.440    // And also: https://github.com/ggml-org/llama.cpp/pull/17945 [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]441 442    // first cancel the adjustment from llama_hparams::yarn_attn_factor_adjust to get the original attn_factor443    GGML_ASSERT(ext_factor >= 0.0f);444    const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));445 446    // use the original attn_factor to pre-scale the kq_scale447    const float mscale   = attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));448    const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));449 450    ggml_tensor * cur;451    ggml_tensor * inpL;452 453    // {n_embd, n_tokens}454    inpL = build_inp_embd(model.tok_embd);455 456    // (optional) temperature tuning - used by mistral-large457    ggml_tensor * inp_attn_scale = nullptr;458    if (hparams.f_attn_temp_scale != 0.0f) {459        inp_attn_scale = build_inp_attn_scale();460    }461 462    // inp_pos - contains the positions463    ggml_tensor * inp_pos = build_inp_pos();464 465    auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr;466    auto * inp_attn_k  =  is_mla ? build_attn_inp_k()  : nullptr;467 468    ggml_tensor * inp_out_ids = build_inp_out_ids();469 470    for (int il = 0; il < n_layer; ++il) {471        ggml_tensor * inpSA = inpL;472 473        // norm474        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);475        cb(cur, "attn_norm", il);476 477        // self_attention478        if (is_ocr) {479            const int n_embed_head = hparams.n_embd / hparams.n_head();480            const int ocr_rope_type = GGML_ROPE_TYPE_NEOX;481            GGML_ASSERT(n_embed_head == n_embd_head_k && n_embed_head == n_embd_head_v);482 483            ggml_tensor * Qcur = NULL;484            ggml_tensor * Kcur = NULL;485            ggml_tensor * Vcur = NULL;486 487            Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);488            Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);489            Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);490            cb(Qcur, "q", il);491            cb(Kcur, "k", il);492            cb(Vcur, "v", il);493 494            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embed_head, n_head, n_tokens);495            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embed_head, n_head, n_tokens);496            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embed_head, n_head, n_tokens);497 498            GGML_ASSERT(fabs(freq_base - 10000.0) < 1e-4);499            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_embed_head, ocr_rope_type, 0, freq_base, 1, 0, 1, 0, 0);500            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_embed_head, ocr_rope_type, 0, freq_base, 1, 0, 1, 0, 0);501            cb(Qcur, "q_pe", il);502            cb(Kcur, "k_pe", il);503 504            cur = build_attn(inp_attn_kv,505                        model.layers[il].wo, NULL, model.layers[il].wo_s,506                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);507            cb(cur, "attn_out", il);508        }509        else {510            ggml_tensor * q = NULL;511 512            const bool is_lite = model.layers[il].wq;513 514            if (!is_lite) {515                q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);516                cb(q, "q", il);517 518                q = build_norm(q, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);519                cb(q, "q", il);520 521                q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);522                cb(q, "q", il);523            } else {524                q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);525                cb(q, "q", il);526            }527            // split into {n_embd_head_qk_nope, n_head, n_tokens}528            ggml_tensor * q_nope =529                ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),530                             ggml_row_size(q->type, n_embd_head_k) * n_head, 0);531            cb(q_nope, "q_nope", il);532 533            // and {n_embd_head_qk_rope, n_head, n_tokens}534            ggml_tensor * q_pe = ggml_view_3d(535                ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),536                ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));537            cb(q_pe, "q_pe", il);538 539            ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);540            cb(kv_cmpr_pe, "kv_cmpr_pe", il);541 542            // split into {kv_lora_rank, n_tokens}543            ggml_tensor * kv_cmpr =544                ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,545                             ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);546            cb(kv_cmpr, "kv_cmpr", il);547 548            // and {n_embd_head_qk_rope, 1, n_tokens}549            ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,550                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),551                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),552                                              ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));553            cb(k_pe, "k_pe", il);554 555            q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,556                                 ext_factor, attn_factor, beta_fast, beta_slow);557            cb(q_pe, "q_pe", il);558 559            k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,560                                 ext_factor, attn_factor, beta_fast, beta_slow);561            cb(k_pe, "k_pe", il);562 563            kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);564            cb(kv_cmpr, "kv_cmpr", il);565 566            if (is_mla) {567                // {n_embd_head_qk_nope, n_tokens, n_head}568                q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);569                cb(q_nope, "q_nope_perm", il);570 571                // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}572                ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);573                cb(q_nope_absorbed, "q_nope_absorbed", il);574 575                // {kv_lora_rank, n_head, n_tokens}576                q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);577                cb(q_nope_absorbed, "q_nope_absorbed_perm", il);578 579                // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}580                // note: rope must go first for in-place context shifting in build_rope_shift()581                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);582                cb(Qcur, "Qcur", il);583 584                kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);585                cb(kv_cmpr, "kv_cmpr_reshape", il);586 587                // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}588                ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);589                cb(Kcur, "Kcur", il);590 591                // {kv_lora_rank, 1, n_tokens}592                ggml_tensor * Vcur = kv_cmpr;593                cb(Vcur, "Vcur", il);594 595                if (inp_attn_scale) {596                    // apply llama 4 temperature scaling597                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);598                    cb(Qcur, "Qcur_attn_temp_scaled", il);599                }600 601                // note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group)602                cur = build_attn(inp_attn_k,603                        model.layers[il].wo, NULL, model.layers[il].wo_s,604                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);605            } else {606                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);607                cb(kv, "kv", il);608 609                // split into {n_embd_head_qk_nope, n_head, n_tokens}610                ggml_tensor * k_nope =611                    ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,612                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),613                                 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, 0);614                cb(k_nope, "k_nope_view", il);615 616                // and {n_embd_head_v, n_head, n_tokens}617                ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, n_embd_head_v, n_head, n_tokens,618                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),619                                                  ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,620                                                  ggml_row_size(kv->type, n_embd_head_qk_nope));621                cb(Vcur, "Vcur_view", il);622 623                Vcur = ggml_cont(ctx0, Vcur);624                cb(Vcur, "Vcur_cont", il);625 626                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope, q_pe, 0);627                cb(Qcur, "Qcur", il);628 629                ggml_tensor * Kcur = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);630                cb(Kcur, "Kcur", il);631 632                if (inp_attn_scale) {633                    // apply llama 4 temperature scaling634                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);635                    cb(Qcur, "Qcur_attn_temp_scaled", il);636                }637 638                // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)639                cur = build_attn(inp_attn_kv,640                            model.layers[il].wo, NULL, model.layers[il].wo_s,641                            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);642            }643        }644        if (il == n_layer - 1 && inp_out_ids && (!cparams.embeddings_nextn || cparams.embeddings_nextn_masked)) {645            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);646            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);647        }648        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);649        cb(ffn_inp, "ffn_inp", il);650 651        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);652        cb(cur, "ffn_norm", il);653 654        if ((uint32_t) il < hparams.n_layer_dense_lead) {655            cur = build_ffn(cur,656                model.layers[il].ffn_up, NULL, NULL,657                model.layers[il].ffn_gate, NULL, NULL,658                model.layers[il].ffn_down, NULL, NULL,659                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);660            cb(cur, "ffn_out", il);661        } else {662            // MoE branch663            ggml_tensor * moe_out = build_moe_ffn(cur,664                model.layers[il].ffn_gate_inp,665                model.layers[il].ffn_up_exps,666                model.layers[il].ffn_gate_exps,667                model.layers[il].ffn_down_exps,668                model.layers[il].ffn_exp_probs_b,669                n_expert, n_expert_used,670                LLM_FFN_SILU, hparams.expert_weights_norm,671                hparams.expert_weights_scale,672                (llama_expert_gating_func_type) hparams.expert_gating_func,673                il,674                nullptr,675                model.layers[il].ffn_gate_up_exps);676            cb(moe_out, "ffn_moe_out", il);677 678            // FFN shared expert679            {680                ggml_tensor * ffn_shexp =681                    build_ffn(cur,682                        model.layers[il].ffn_up_shexp, NULL, NULL,683                        model.layers[il].ffn_gate_shexp, NULL, NULL,684                        model.layers[il].ffn_down_shexp, NULL, NULL,685                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);686                cb(ffn_shexp, "ffn_shexp", il);687 688                cur = ggml_add(ctx0, moe_out, ffn_shexp);689                cb(cur, "ffn_out", il);690            }691        }692        cur = ggml_add(ctx0, cur, ffn_inp);693 694        cur = build_cvec(cur, il);695        cb(cur, "l_out", il);696 697        // input for next layer698        inpL = cur;699    }700    cur = inpL;701 702    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);703 704    cb(cur, "h_nextn", -1);705    res->t_h_nextn = cur;706 707    if (cparams.embeddings_nextn && !cparams.embeddings_nextn_masked && inp_out_ids) {708        cur = ggml_get_rows(ctx0, cur, inp_out_ids);709    }710 711    cb(cur, "result_norm", -1);712    res->t_embd = cur;713 714    // lm_head715    cur = ggml_mul_mat(ctx0, model.output, cur);716 717    cb(cur, "result_output", -1);718    res->t_logits = cur;719 720    ggml_build_forward_expand(gf, cur);721}722