echodict/llama.cpp

1#include "models.h"2 3llm_build_jais::llm_build_jais(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 11    inpL = build_inp_embd(model.tok_embd);12 13    auto * inp_attn = build_attn_inp_kv();14 15    ggml_tensor * inp_out_ids = build_inp_out_ids();16 17    for (int il = 0; il < n_layer; ++il) {18        cur = build_norm(inpL,19                model.layers[il].attn_norm,20                model.layers[il].attn_norm_b,21                LLM_NORM, il);22        cb(cur, "attn_norm", il);23 24        // self-attention25        {26            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,27                    n_embd_head, n_head, n_head_kv, il);28 29            cur = build_attn(inp_attn,30                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,31                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il);32        }33        if (il == n_layer - 1 && inp_out_ids) {34            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);35            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);36        }37        // add the input38        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);39        cb(ffn_inp, "ffn_inp", il);40 41        // FF42        {43            cur = build_norm(ffn_inp,44                    model.layers[il].ffn_norm,45                    model.layers[il].ffn_norm_b,46                    LLM_NORM, il);47            cb(cur, "ffn_norm", il);48 49            cur = build_ffn(cur,50                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,51                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,52                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,53                    NULL,54                    LLM_FFN_SILU, LLM_FFN_PAR, il);55            cb(cur, "ffn_out", il);56        }57 58        cur = ggml_add(ctx0, cur, ffn_inp);59 60        cur = build_cvec(cur, il);61        cb(cur, "l_out", il);62 63        // input for next layer64        inpL = cur;65    }66    cur = build_norm(inpL,67            model.output_norm,68            model.output_norm_b,69            LLM_NORM, -1);70 71    cb(cur, "result_norm", -1);72    res->t_embd = cur;73 74    cur = build_lora_mm(model.output, cur);75 76    cb(cur, "result_output", -1);77    res->t_logits = cur;78 79    ggml_build_forward_expand(gf, cur);80}81