echodict/llama.cpp

1#include "models.h"2 3llm_build_ernie4_5_moe::llm_build_ernie4_5_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_v();6 7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());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 = build_attn_inp_kv();19 20    ggml_tensor * inp_out_ids = build_inp_out_ids();21 22    GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");23    for (int il = 0; il < n_layer; ++il) {24        ggml_tensor * inpSA = inpL;25        // norm26        {27            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);28            cb(cur, "attn_norm", il);29        }30        // self-attention31        {32            // compute Q and K and RoPE them33            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,34                    n_embd_head, n_head, n_head_kv, il);35 36            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,37                                 ext_factor, attn_factor, beta_fast, beta_slow);38 39            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,40                                 ext_factor, attn_factor, beta_fast, beta_slow);41 42            cb(Qcur, "Qcur", il);43            cb(Kcur, "Kcur", il);44            cb(Vcur, "Vcur", il);45 46            cur = build_attn(inp_attn,47                    model.layers[il].wo, NULL, model.layers[il].wo_s,48                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);49            cb(cur, "attn_out", il);50        }51        if (il == n_layer - 1 && inp_out_ids) {52            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);53            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);54        }55        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);56        cb(ffn_inp, "ffn_inp", il);57 58        // feed-forward network59        bool is_moe_layer =60            static_cast<uint32_t>(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0;61 62        if (!is_moe_layer) {63            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);64            cb(cur, "ffn_norm", il);65 66            cur = build_ffn(cur,67                    model.layers[il].ffn_up, NULL, NULL,68                    model.layers[il].ffn_gate, NULL, NULL,69                    model.layers[il].ffn_down, NULL, NULL,70                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);71            cb(cur, "ffn_out", il);72        } else {73            // MoE branch74            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);75            cb(cur, "ffn_norm", il);76 77            ggml_tensor * moe_out = build_moe_ffn(cur,78                                        model.layers[il].ffn_gate_inp,79                                        model.layers[il].ffn_up_exps,80                                        model.layers[il].ffn_gate_exps,81                                        model.layers[il].ffn_down_exps,82                                        model.layers[il].ffn_exp_probs_b,83                                        n_expert, n_expert_used,84                                        LLM_FFN_SILU, true,85                                        hparams.expert_weights_scale,86                                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,87                                        il);88            cb(moe_out, "ffn_moe_out", il);89 90            // Shared expert (if present)91            if (hparams.n_ff_shexp > 0) {92                ggml_tensor * ffn_shexp =93                    build_ffn(cur,94                        model.layers[il].ffn_up_shexp, NULL, NULL,95                        model.layers[il].ffn_gate_shexp, NULL, NULL,96                        model.layers[il].ffn_down_shexp, NULL, NULL,97                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);98                cb(ffn_shexp, "ffn_shexp", il);99 100                cur = ggml_add(ctx0, moe_out, ffn_shexp);101            } else {102                cur = moe_out;103            }104            cb(cur, "ffn_out", il);105        }106        cur = ggml_add(ctx0, cur, ffn_inp);107        cb(cur, "ffn_out", il);108 109        cur = build_cvec(cur, il);110        cb(cur, "l_out", il);111 112        // input for next layer113        inpL = cur;114    }115    cur = inpL;116 117    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);118 119    cb(cur, "result_norm", -1);120    res->t_embd = cur;121 122    // lm_head123    cur = build_lora_mm(model.output, cur);124 125    cb(cur, "result_output", -1);126    res->t_logits = cur;127 128    ggml_build_forward_expand(gf, cur);129}130