Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_exaone_moe::load_arch_hparams(llama_model_loader & ml) {4    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;5    hparams.n_swa = 128;6    uint32_t swa_period = 4;7    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);8    hparams.set_swa_pattern(swa_period);9    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;10    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;11 12    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,                hparams.rope_freq_base_train_swa, false);13    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);14    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);15    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);16    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);17    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);18    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);19    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);20    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);21    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);22 23    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);24    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_impl");25 26    switch (hparams.n_layer()) {27        case 32: type = LLM_TYPE_30B_A3B; break;28        case 48: type = LLM_TYPE_235B_A22B; break;29        default: type = LLM_TYPE_UNKNOWN;30    }31}32 33void llama_model_exaone_moe::load_arch_tensors(llama_model_loader &) {34    LLAMA_LOAD_LOCALS;35 36    const int64_t n_ff_exp       = hparams.n_ff_exp;37    const int64_t n_ff_shexp     = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : n_ff_exp;38    const int64_t head_dim       = hparams.n_embd_head_k();39    const int64_t n_qo_dim       = n_head * head_dim;40    const int64_t n_kv_dim       = n_head_kv * head_dim;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}, 0);47 48    if (output == NULL) {49        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);50    }51 52    for (int i = 0; i < n_layer_all; ++i) {53        int flags = 0;54        if (i >= n_layer) {55            // skip all tensors in the NextN layers56            flags |= TENSOR_SKIP;57        }58 59        auto & layer = layers[i];60        create_tensor_qkv(layer, i, n_embd, n_qo_dim, n_kv_dim, n_kv_dim, flags);61        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_qo_dim, n_embd}, flags);62 63        layer.rope_freqs   = create_tensor(tn(LLM_TENSOR_ROPE_FREQS,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0) | flags);64 65        layer.attn_norm    = create_tensor(tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, flags);66        layer.attn_q_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);67        layer.attn_k_norm  = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);68 69        layer.ffn_norm     = create_tensor(tn(LLM_TENSOR_FFN_NORM,    "weight", i), {n_embd}, flags);70 71        // dense layers for first n_layer_dense_lead layers or nextn_predict_layers layers at the end72        if (i < (int) hparams.n_layer_dense_lead || (i >= n_layer)) {73            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);74            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, flags);75            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, flags);76        } else {77            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, flags);78            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);79 80            if (n_expert == 0) {81                throw std::runtime_error("n_expert must be > 0");82            }83            if (n_expert_used == 0) {84                throw std::runtime_error("n_expert_used must be > 0");85            }86 87            layer.ffn_gate_exps  = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS,  "weight", i), {n_embd, n_ff_exp, n_expert}, flags);88            layer.ffn_down_exps  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,  "weight", i), {n_ff_exp, n_embd, n_expert}, flags);89            layer.ffn_up_exps    = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,    "weight", i), {n_embd, n_ff_exp, n_expert}, flags);90 91            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);92            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, flags);93            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, flags);94        }95 96        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers97        if (i >= n_layer) {98            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);99            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,   "weight", i), {n_embd}, flags);100            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,   "weight", i), {n_embd}, flags);101 102            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);103            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), {n_embd, n_vocab}, flags | TENSOR_NOT_REQUIRED);104            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), {n_embd, n_vocab}, flags | TENSOR_NOT_REQUIRED);105        }106    }107}108 109std::unique_ptr<llm_graph_context> llama_model_exaone_moe::build_arch_graph(const llm_graph_params & params) const {110    return std::make_unique<graph>(*this, params);111}112 113llama_model_exaone_moe::graph::graph(const llama_model & model, const llm_graph_params & params) :114    llm_graph_context(params) {115    const int64_t n_embd_head = hparams.n_embd_head_k();116 117    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());118    GGML_ASSERT(n_embd_head == n_rot);119 120    ggml_tensor * cur;121    ggml_tensor * inpL;122 123    inpL = build_inp_embd(model.tok_embd);124 125    // inp_pos - contains the positions126    ggml_tensor * inp_pos = build_inp_pos();127 128    auto * inp_attn_iswa = build_attn_inp_kv_iswa();129 130    ggml_tensor * inp_out_ids = build_inp_out_ids();131 132    for (int il = 0; il < n_layer; ++il) {133        ggml_tensor * inpSA = inpL;134 135        // use RoPE for SWA layers136        const bool is_local_layer = hparams.is_swa(il);137 138        // norm139        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);140        cb(cur, "attn_norm", il);141 142        // self-attention143        {144            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);145 146            // compute Q and K and RoPE them147            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,148                    n_embd_head, n_head, n_head_kv, il);149 150            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);151            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);152            cb(Qcur, "Qcur_normed", il);153            cb(Kcur, "Kcur_normed", il);154 155            if (is_local_layer) {156                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,157                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);158 159                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,160                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);161            }162            cb(Qcur, "Qcur", il);163            cb(Kcur, "Kcur", il);164            cb(Vcur, "Vcur", il);165 166            cur = build_attn(inp_attn_iswa,167                model.layers[il].wo, NULL, model.layers[il].wo_s,168                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);169            cb(cur, "attn_out", il);170        }171        if (il == n_layer - 1 && inp_out_ids) {172            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);173            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);174        }175        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);176        cb(ffn_inp, "ffn_inp", il);177 178        // norm179        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);180        cb(cur, "ffn_norm", il);181 182        // feed-forward network183        if (model.layers[il].ffn_gate_inp == nullptr) {184            // dense branch185            cur = build_ffn(cur,186                    model.layers[il].ffn_up, NULL, NULL,187                    model.layers[il].ffn_gate, NULL, NULL,188                    model.layers[il].ffn_down, NULL, NULL, NULL,189                    LLM_FFN_SILU, LLM_FFN_PAR, il);190            cb(cur, "ffn_out", il);191        } else {192            // MoE branch193            ggml_tensor * moe_out = build_moe_ffn(cur,194                model.layers[il].ffn_gate_inp,195                model.layers[il].ffn_up_exps,196                model.layers[il].ffn_gate_exps,197                model.layers[il].ffn_down_exps,198                model.layers[il].ffn_exp_probs_b,199                n_expert, n_expert_used,200                LLM_FFN_SILU, hparams.expert_weights_norm,201                hparams.expert_weights_scale,202                (llama_expert_gating_func_type) hparams.expert_gating_func,203                il);204            cb(moe_out, "ffn_moe_out", il);205 206            // FFN shared expert207            {208                ggml_tensor * ffn_shexp =209                    build_ffn(cur,210                        model.layers[il].ffn_up_shexp, NULL, NULL,211                        model.layers[il].ffn_gate_shexp, NULL, NULL,212                        model.layers[il].ffn_down_shexp, NULL, NULL,213                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);214                cb(ffn_shexp, "ffn_shexp", il);215 216                cur = ggml_add(ctx0, moe_out, ffn_shexp);217                cb(cur, "ffn_out", il);218            }219        }220 221        cur = ggml_add(ctx0, cur, ffn_inp);222 223        cur = build_cvec(cur, il);224        cb(cur, "l_out", il);225 226        // input for next layer227        inpL = cur;228    }229    cur = inpL;230 231    // final norm232    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);233 234    cb(cur, "result_norm", -1);235    res->t_embd = cur;236 237    // lm_head238    cur = build_lora_mm(model.output, cur, model.output_s);239 240    cb(cur, "result_output", -1);241    res->t_logits = cur;242 243    ggml_build_forward_expand(gf, cur);244}245