echodict/llama.cpp

1#include "models.h"2 3llm_build_neo_bert::llm_build_neo_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {4    const int64_t n_embd_head = hparams.n_embd_head_v();5 6    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());7 8    ggml_tensor * cur;9    ggml_tensor * inpL;10    ggml_tensor * inp_pos = build_inp_pos();11 12    // construct input embeddings (token, type, position)13    inpL = build_inp_embd(model.tok_embd);14    cb(inpL, "inp_embd", -1);15 16    auto * inp_attn = build_attn_inp_no_cache();17 18    ggml_tensor * inp_out_ids = build_inp_out_ids();19 20    for (int il = 0; il < n_layer; ++il) {21        ggml_tensor * cur = inpL;22 23        // pre-norm24        cur = build_norm(inpL,25                model.layers[il].attn_norm, NULL,26                LLM_NORM_RMS, il);27 28        {29            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,30                    n_embd_head, n_head, n_head_kv, il);31 32            // RoPE33            Qcur = ggml_rope_ext(34                    ctx0, Qcur, inp_pos, nullptr,35                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,36                    ext_factor, attn_factor, beta_fast, beta_slow37                    );38 39            Kcur = ggml_rope_ext(40                    ctx0, Kcur, inp_pos, nullptr,41                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,42                    ext_factor, attn_factor, beta_fast, beta_slow43                    );44 45            cb(Qcur, "Qcur", il);46            cb(Kcur, "Kcur", il);47            cb(Vcur, "Vcur", il);48 49            cur = build_attn(inp_attn,50                    model.layers[il].wo, nullptr, model.layers[il].wo_s,51                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);52            cb(cur, "kqv_out", il);53        }54        if (il == n_layer - 1 && inp_out_ids) {55            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);56            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);57        }58        // re-add the layer input59        cur = ggml_add(ctx0, cur, inpL);60 61        ggml_tensor * ffn_inp = cur;62        cb(ffn_inp, "ffn_inp", il);63 64        // pre-norm65        cur = build_norm(ffn_inp,66                model.layers[il].ffn_norm, NULL,67                LLM_NORM_RMS, il);68        cb(cur, "ffn_norm", il);69 70        // feed-forward network71        cur = build_ffn(cur,72                model.layers[il].ffn_up,73                NULL, NULL, NULL, NULL, NULL,74                model.layers[il].ffn_down,75                NULL, NULL, NULL,76                LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);77 78        // attentions bypass the intermediate layer79        cur = ggml_add(ctx0, cur, ffn_inp);80 81        // input for next layer82        inpL = cur;83    }84    cur = inpL;85 86    cur = build_norm(cur,87            model.output_norm_enc, NULL,88            LLM_NORM_RMS, -1);89 90    cb(cur, "result_embd", -1);91    res->t_embd = cur;92 93    ggml_build_forward_expand(gf, cur);94}95