echodict/llama.cpp

1#include "models.h"2 3template <bool iswa>4llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) :5    llm_graph_context(params) {6    const int64_t n_embd_head = hparams.n_embd_head_k();7 8    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());9    GGML_ASSERT(n_embd_head == n_rot);10 11    ggml_tensor * cur;12    ggml_tensor * inpL;13 14    inpL = build_inp_embd(model.tok_embd);15 16    // inp_pos - contains the positions17    ggml_tensor * inp_pos = build_inp_pos();18 19    using inp_attn_type      = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;20    inp_attn_type * inp_attn = nullptr;21 22    if constexpr (iswa) {23        inp_attn = build_attn_inp_kv_iswa();24    } else {25        inp_attn = build_attn_inp_kv();26    }27    ggml_tensor * inp_out_ids = build_inp_out_ids();28 29    for (int il = 0; il < n_layer; ++il) {30        ggml_tensor * inpSA = inpL;31 32        // use RoPE for SWA layers or non-SWA models33        const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;34 35        cur = inpL;36 37        // self-attention38        {39            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);40 41            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,42                    n_embd_head, n_head, n_head_kv, il);43 44            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);45            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);46            cb(Qcur, "Qcur_normed", il);47            cb(Kcur, "Kcur_normed", il);48 49            if (use_rope) {50                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,51                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);52 53                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,54                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);55            }56            cb(Qcur, "Qcur", il);57            cb(Kcur, "Kcur", il);58            cb(Vcur, "Vcur", il);59 60            cur = build_attn(inp_attn,61                    model.layers[il].wo, NULL, model.layers[il].wo_s,62                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);63            cb(cur, "attn_out", il);64        }65        if (il == n_layer - 1 && inp_out_ids) {66            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);67            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);68        }69        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);70        cb(cur, "attn_post_norm", il);71 72        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);73        cb(ffn_inp, "ffn_inp", il);74 75        // feed-forward network76        cur = build_ffn(ffn_inp,77                model.layers[il].ffn_up, NULL, NULL,78                model.layers[il].ffn_gate, NULL, NULL,79                model.layers[il].ffn_down, NULL, NULL, NULL,80                LLM_FFN_SILU, LLM_FFN_PAR, il);81        cb(cur, "ffn_out", il);82 83        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);84        cb(cur, "ffn_post_norm", -1);85 86        cur = ggml_add(ctx0, cur, ffn_inp);87 88        cur = build_cvec(cur, il);89        cb(cur, "l_out", il);90 91        // input for next layer92        inpL = cur;93    }94    cur = inpL;95 96    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);97 98    cb(cur, "result_norm", -1);99    res->t_embd = cur;100 101    // lm_head102    cur = build_lora_mm(model.output, cur);103 104    cb(cur, "result_output", -1);105    res->t_logits = cur;106 107    ggml_build_forward_expand(gf, cur);108}109 110// Explicit template instantiations111template struct llm_build_exaone4<false>;112template struct llm_build_exaone4<true>;113