echodict/llama.cpp

1#include "models.h"2 3llm_build_eurobert::llm_build_eurobert(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    inpL = build_inp_embd(model.tok_embd);13    cb(inpL, "inp_embd", -1);14 15    auto * inp_attn = build_attn_inp_no_cache();16 17    ggml_tensor * inp_out_ids = build_inp_out_ids();18 19    for (int il = 0; il < n_layer; ++il) {20        ggml_tensor * cur = inpL;21 22        cur = build_norm(inpL,23                model.layers[il].attn_norm, NULL,24                LLM_NORM_RMS, il);25 26        {27            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,28                    n_embd_head, n_head, n_head_kv, il);29 30            Qcur = ggml_rope_ext(31                    ctx0, Qcur, inp_pos, nullptr,32                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,33                    ext_factor, attn_factor, beta_fast, beta_slow34                    );35 36            Kcur = ggml_rope_ext(37                    ctx0, Kcur, inp_pos, nullptr,38                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,39                    ext_factor, attn_factor, beta_fast, beta_slow40                    );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, nullptr, model.layers[il].wo_s,48                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);49            cb(cur, "kqv_out", il);50        }51 52        if (il == n_layer - 1 && inp_out_ids) {53            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);54            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);55        }56 57        cur = ggml_add(ctx0, cur, inpL);58 59        ggml_tensor * ffn_inp = cur;60        cb(ffn_inp, "ffn_inp", il);61 62        cur = build_norm(ffn_inp,63                model.layers[il].ffn_norm, NULL,64                LLM_NORM_RMS, il);65        cb(cur, "ffn_norm", il);66 67        cur = build_ffn(cur,68                model.layers[il].ffn_up, NULL, NULL,69                model.layers[il].ffn_gate, NULL, NULL,70                model.layers[il].ffn_down, NULL, NULL,71                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);72        cb(cur, "ffn_out", il);73 74        cur = ggml_add(ctx0, cur, ffn_inp);75 76        // input for next layer77        inpL = cur;78    }79    cur = inpL;80 81    cur = build_norm(cur,82            model.output_norm, NULL,83            LLM_NORM_RMS, -1);84 85    cb(cur, "result_embd", -1);86    res->t_embd = cur;87 88    ggml_build_forward_expand(gf, cur);89}90