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    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
  6
  7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
  8
  9    ggml_tensor * cur;
 10    ggml_tensor * inpL;
 11    ggml_tensor * inp_pos = build_inp_pos();
 12
 13    // construct input embeddings (token, type, position)
 14    inpL = build_inp_embd(model.tok_embd);
 15    cb(inpL, "inp_embd", -1);
 16
 17    // embed layer norm
 18    inpL = build_norm(inpL, model.tok_norm, nullptr, LLM_NORM, -1);
 19    cb(inpL, "inp_norm", -1);
 20
 21    ggml_tensor * inp_out_ids = build_inp_out_ids();
 22
 23    auto * inp_attn = build_attn_inp_no_cache();
 24
 25    for (int il = 0; il < n_layer; ++il) {
 26        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
 27        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
 28
 29        cur = inpL;
 30
 31        // attention layer norm
 32        if (model.layers[il].attn_norm) {
 33            cur = build_norm(inpL,
 34                    model.layers[il].attn_norm, NULL,
 35                    LLM_NORM, il);
 36            cb(cur, "attn_norm", il);
 37        }
 38
 39        // self attention
 40        cur = build_lora_mm(model.layers[il].wqkv, cur);
 41        cb(cur, "wqkv", il);
 42
 43        const size_t type_size = ggml_type_size(cur->type);
 44
 45        ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*type_size, cur->nb[1], 0*type_size*(n_embd));
 46        ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*type_size, cur->nb[1], 1*type_size*(n_embd));
 47        ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*type_size, cur->nb[1], 1*type_size*(n_embd + n_embd_gqa));
 48
 49        // RoPE
 50        Qcur = ggml_rope_ext(
 51                ctx0, Qcur, inp_pos, nullptr,
 52                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
 53                ext_factor, attn_factor, beta_fast, beta_slow
 54                );
 55
 56        Kcur = ggml_rope_ext(
 57                ctx0, Kcur, inp_pos, nullptr,
 58                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
 59                ext_factor, attn_factor, beta_fast, beta_slow
 60                );
 61
 62        cb(Qcur, "Qcur", il);
 63        cb(Kcur, "Kcur", il);
 64        cb(Vcur, "Vcur", il);
 65
 66        cur = build_attn(inp_attn,
 67                    model.layers[il].wo, nullptr,
 68                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 69        cb(cur, "kqv_out", il);
 70
 71        if (il == n_layer - 1 && inp_out_ids) {
 72            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
 73            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
 74        }
 75
 76        // re-add the layer input
 77        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
 78        cb(ffn_inp, "ffn_inp", il);
 79
 80        // attention layer norm
 81        cur = build_norm(ffn_inp,
 82                model.layers[il].ffn_norm, NULL,
 83                LLM_NORM, il);
 84        cb(cur, "ffn_norm", il);
 85
 86        cur = build_ffn(cur,
 87                model.layers[il].ffn_up,   NULL, NULL,
 88                NULL,                      NULL, NULL,
 89                model.layers[il].ffn_down, NULL, NULL,
 90                NULL,
 91                LLM_FFN_GEGLU, LLM_FFN_SEQ, il);
 92
 93        // attentions bypass the intermediate layer
 94        cur = ggml_add(ctx0, cur, ffn_inp);
 95
 96        // input for next layer
 97        inpL = cur;
 98    }
 99
100    cur = inpL;
101
102    cur = build_norm(cur,
103            model.output_norm, NULL,
104            LLM_NORM, -1);
105    cb(cur, "final_norm_out", -1);
106
107    if (hparams.pooling_type == LLAMA_POOLING_TYPE_CLS) {
108        // extracting cls token
109        cur = ggml_view_1d(ctx0, cur, hparams.n_embd, 0);
110        cb(cur, "cls_pooled_embd", -1);
111    }
112
113    cb(cur, "res_embd", -1);
114    res->t_embd = cur;
115    ggml_build_forward_expand(gf, cur);
116}