Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_exaone4::load_arch_hparams(llama_model_loader & ml) {4    if (hparams.n_layer() == 64) {    // 32B5        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;6        hparams.n_swa = 4096;7        load_swa_pattern(ml, 4);8 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        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);12    }13 14    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);15    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);16 17    switch (hparams.n_layer()) {18        case 30: type = LLM_TYPE_1_2B; break;19        case 64: type = LLM_TYPE_32B; break;20        default: type = LLM_TYPE_UNKNOWN;21    }22}23 24void llama_model_exaone4::load_arch_tensors(llama_model_loader &) {25    LLAMA_LOAD_LOCALS;26 27    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);28 29    // output30    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);31    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);32 33    // if output is NULL, init from the input tok embed34    if (output == NULL) {35        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);36    }37 38    for (int i = 0; i < n_layer_all; ++i) {39        const bool is_nextn = i >= n_layer;40        int flags = 0;41        if (is_nextn) {42            // NextN/MTP layers are preserved in GGUF but are not executed yet.43            flags |= TENSOR_SKIP;44        }45 46        auto & layer = layers[i];47 48        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);49        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, flags);50 51        if (!is_nextn) {52            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));53        }54 55        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, flags);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_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);60        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);61        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);62        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, flags);63 64        if (is_nextn) {65            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);66            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,   "weight", i), {n_embd}, flags);67            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,   "weight", i), {n_embd}, flags);68            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);69        }70    }71}72 73std::unique_ptr<llm_graph_context> llama_model_exaone4::build_arch_graph(const llm_graph_params & params) const {74    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {75        return std::make_unique<graph<true>>(*this, params);76    } else {77        return std::make_unique<graph<false>>(*this, params);78    }79}80 81template <bool iswa>82llama_model_exaone4::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) :83    llm_graph_context(params) {84    const int64_t n_embd_head = hparams.n_embd_head_k();85 86    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());87    GGML_ASSERT(n_embd_head == n_rot);88 89    ggml_tensor * cur;90    ggml_tensor * inpL;91 92    inpL = build_inp_embd(model.tok_embd);93 94    // inp_pos - contains the positions95    ggml_tensor * inp_pos = build_inp_pos();96 97    using inp_attn_type      = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;98    inp_attn_type * inp_attn = nullptr;99 100    if constexpr (iswa) {101        inp_attn = build_attn_inp_kv_iswa();102    } else {103        inp_attn = build_attn_inp_kv();104    }105    ggml_tensor * inp_out_ids = build_inp_out_ids();106 107    for (int il = 0; il < n_layer; ++il) {108        ggml_tensor * inpSA = inpL;109 110        // use RoPE for SWA layers or non-SWA models111        const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;112 113        cur = inpL;114 115        // self-attention116        {117            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);118 119            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,120                    n_embd_head, n_head, n_head_kv, il);121 122            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);123            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);124            cb(Qcur, "Qcur_normed", il);125            cb(Kcur, "Kcur_normed", il);126 127            if (use_rope) {128                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,129                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);130 131                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,132                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);133            }134            cb(Qcur, "Qcur", il);135            cb(Kcur, "Kcur", il);136            cb(Vcur, "Vcur", il);137 138            cur = build_attn(inp_attn,139                    model.layers[il].wo, NULL, model.layers[il].wo_s,140                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);141            cb(cur, "attn_out", il);142        }143        if (il == n_layer - 1 && inp_out_ids) {144            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);145            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);146        }147        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);148        cb(cur, "attn_post_norm", il);149 150        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);151        cb(ffn_inp, "ffn_inp", il);152 153        // feed-forward network154        cur = build_ffn(ffn_inp,155                model.layers[il].ffn_up, NULL, NULL,156                model.layers[il].ffn_gate, NULL, NULL,157                model.layers[il].ffn_down, NULL, NULL, NULL,158                LLM_FFN_SILU, LLM_FFN_PAR, il);159        cb(cur, "ffn_out", il);160 161        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);162        cb(cur, "ffn_post_norm", -1);163 164        cur = ggml_add(ctx0, cur, ffn_inp);165 166        cur = build_cvec(cur, il);167        cb(cur, "l_out", il);168 169        // input for next layer170        inpL = cur;171    }172    cur = inpL;173 174    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);175 176    cb(cur, "result_norm", -1);177    res->t_embd = cur;178 179    // lm_head180    cur = build_lora_mm(model.output, cur, model.output_s);181 182    cb(cur, "result_output", -1);183    res->t_logits = cur;184 185    ggml_build_forward_expand(gf, cur);186}187 188// Explicit template instantiations189template struct llama_model_exaone4::graph<false>;190template struct llama_model_exaone4::graph<true>;191