echodict/llama.cpp

1#include "models.h"2 3llm_build_exaone_moe::llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params) :4    llm_graph_context(params) {5    const int64_t n_embd_head = hparams.n_embd_head_k();6 7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());8    GGML_ASSERT(n_embd_head == n_rot);9 10    ggml_tensor * cur;11    ggml_tensor * inpL;12 13    inpL = build_inp_embd(model.tok_embd);14 15    // inp_pos - contains the positions16    ggml_tensor * inp_pos = build_inp_pos();17 18    auto * inp_attn_iswa = build_attn_inp_kv_iswa();19 20    ggml_tensor * inp_out_ids = build_inp_out_ids();21 22    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;23    for (int il = 0; il < n_transformer_layers; ++il) {24        ggml_tensor * inpSA = inpL;25 26        // use RoPE for SWA layers27        const bool is_local_layer = hparams.is_swa(il);28 29        // norm30        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);31        cb(cur, "attn_norm", il);32 33        // self-attention34        {35            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);36 37            // compute Q and K and RoPE them38            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,39                    n_embd_head, n_head, n_head_kv, il);40 41            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);42            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);43            cb(Qcur, "Qcur_normed", il);44            cb(Kcur, "Kcur_normed", il);45 46            if (is_local_layer) {47                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,48                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);49 50                Kcur = ggml_rope_ext(ctx0, Kcur, 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            cb(Qcur, "Qcur", il);54            cb(Kcur, "Kcur", il);55            cb(Vcur, "Vcur", il);56 57            cur = build_attn(inp_attn_iswa,58                model.layers[il].wo, NULL, model.layers[il].wo_s,59                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);60            cb(cur, "attn_out", il);61        }62        if (il == n_transformer_layers - 1 && inp_out_ids) {63            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);64            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);65        }66        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);67        cb(ffn_inp, "ffn_inp", il);68 69        // norm70        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);71        cb(cur, "ffn_norm", il);72 73        // feed-forward network74        if (model.layers[il].ffn_gate_inp == nullptr) {75            // dense branch76            cur = build_ffn(cur,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        } else {83            // MoE branch84            ggml_tensor * moe_out = build_moe_ffn(cur,85                model.layers[il].ffn_gate_inp,86                model.layers[il].ffn_up_exps,87                model.layers[il].ffn_gate_exps,88                model.layers[il].ffn_down_exps,89                model.layers[il].ffn_exp_probs_b,90                n_expert, n_expert_used,91                LLM_FFN_SILU, hparams.expert_weights_norm,92                hparams.expert_weights_scale,93                (llama_expert_gating_func_type) hparams.expert_gating_func,94                il);95            cb(moe_out, "ffn_moe_out", il);96 97            // FFN shared expert98            {99                ggml_tensor * ffn_shexp =100                    build_ffn(cur,101                        model.layers[il].ffn_up_shexp, NULL, NULL,102                        model.layers[il].ffn_gate_shexp, NULL, NULL,103                        model.layers[il].ffn_down_shexp, NULL, NULL,104                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);105                cb(ffn_shexp, "ffn_shexp", il);106 107                cur = ggml_add(ctx0, moe_out, ffn_shexp);108                cb(cur, "ffn_out", il);109            }110        }111 112        cur = ggml_add(ctx0, cur, ffn_inp);113 114        cur = build_cvec(cur, il);115        cb(cur, "l_out", il);116 117        // input for next layer118        inpL = cur;119    }120    cur = inpL;121 122    // final norm123    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);124 125    cb(cur, "result_norm", -1);126    res->t_embd = cur;127 128    // lm_head129    cur = build_lora_mm(model.output, cur);130 131    cb(cur, "result_output", -1);132    res->t_logits = cur;133 134    ggml_build_forward_expand(gf, cur);135}136