echodict/llama.cpp

1#include "models.h"2 3llm_build_gpt2::llm_build_gpt2(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 * pos;10    ggml_tensor * inpL;11 12    inpL = build_inp_embd(model.tok_embd);13 14    // inp_pos - contains the positions15    ggml_tensor * inp_pos = build_inp_pos();16 17    auto * inp_attn = build_attn_inp_kv();18 19    pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);20    cb(pos, "pos_embd", -1);21 22    inpL = ggml_add(ctx0, inpL, pos);23    cb(inpL, "inpL", -1);24 25    ggml_tensor * inp_out_ids = build_inp_out_ids();26 27    for (int il = 0; il < n_layer; ++il) {28        cur = build_norm(inpL,29                model.layers[il].attn_norm,30                model.layers[il].attn_norm_b,31                LLM_NORM, il);32        cb(cur, "attn_norm", il);33 34        // self-attention35        {36            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,37                    n_embd_head, n_head, n_head_kv, il);38 39            cur = build_attn(inp_attn,40                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,41                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);42        }43 44        if (il == n_layer - 1 && inp_out_ids) {45            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);46            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);47        }48 49        // add the input50        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);51        cb(ffn_inp, "ffn_inp", il);52 53        // FF54        {55            cur = build_norm(ffn_inp,56                    model.layers[il].ffn_norm,57                    model.layers[il].ffn_norm_b,58                    LLM_NORM, il);59            cb(cur, "ffn_norm", il);60 61            cur = build_ffn(cur,62                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,63                    NULL,                      NULL,                        NULL,64                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,65                    NULL,66                    LLM_FFN_GELU, LLM_FFN_SEQ, il);67            cb(cur, "ffn_out", il);68        }69 70        cur = ggml_add(ctx0, cur, ffn_inp);71 72        cur = build_cvec(cur, il);73        cb(cur, "l_out", il);74 75        // input for next layer76        inpL = cur;77    }78 79    cur = build_norm(inpL,80            model.output_norm,81            model.output_norm_b,82            LLM_NORM, -1);83 84    cb(cur, "result_norm", -1);85    res->t_embd = cur;86 87    cur = build_lora_mm(model.output, cur);88 89    cb(cur, "result_output", -1);90    res->t_logits = cur;91 92    ggml_build_forward_expand(gf, cur);93}94