echodict/llama.cpp

1#include "models.h"2 3llm_build_modern_bert::llm_build_modern_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    // embed layer norm17    inpL = build_norm(inpL, model.tok_norm, nullptr, LLM_NORM, 0);18    cb(inpL, "inp_norm", 0);19 20    ggml_tensor * inp_out_ids = build_inp_out_ids();21 22    auto * inp_attn = build_attn_inp_no_cache();23 24    for (int il = 0; il < n_layer; ++il) {25        const float freq_base_l  = model.get_rope_freq_base(cparams, il);26        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);27 28        cur = inpL;29 30        // attention layer norm31        if (model.layers[il].attn_norm) {32            cur = build_norm(inpL,33                    model.layers[il].attn_norm, NULL,34                    LLM_NORM, il);35            cb(cur, "attn_norm", il);36        }37 38        // self attention39        auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,40                n_embd_head, n_head, n_head_kv, il);41 42        // RoPE43        Qcur = ggml_rope_ext(44                ctx0, Qcur, inp_pos, nullptr,45                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,46                ext_factor, attn_factor, beta_fast, beta_slow47                );48 49        Kcur = ggml_rope_ext(50                ctx0, Kcur, inp_pos, nullptr,51                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,52                ext_factor, attn_factor, beta_fast, beta_slow53                );54 55        cb(Qcur, "Qcur", il);56        cb(Kcur, "Kcur", il);57        cb(Vcur, "Vcur", il);58 59        cur = build_attn(inp_attn,60                    model.layers[il].wo, nullptr, model.layers[il].wo_s,61                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);62        cb(cur, "kqv_out", il);63 64        if (il == n_layer - 1 && inp_out_ids) {65            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);66            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);67        }68 69        // re-add the layer input70        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);71        cb(ffn_inp, "ffn_inp", il);72 73        // attention layer norm74        cur = build_norm(ffn_inp,75                model.layers[il].ffn_norm, NULL,76                LLM_NORM, il);77        cb(cur, "ffn_norm", il);78 79        cur = build_ffn(cur,80                model.layers[il].ffn_up,   NULL, NULL,81                NULL,                      NULL, NULL,82                model.layers[il].ffn_down, NULL, NULL,83                NULL,84                LLM_FFN_GEGLU, LLM_FFN_SEQ, il);85 86        // attentions bypass the intermediate layer87        cur = ggml_add(ctx0, cur, ffn_inp);88 89        // input for next layer90        inpL = cur;91    }92 93    cur = inpL;94 95    cur = build_norm(cur,96            model.output_norm, NULL,97            LLM_NORM, -1);98    cb(cur, "final_norm_out", -1);99 100    res->t_embd = cur;101    ggml_build_forward_expand(gf, cur);102}103