echodict/llama.cpp

1#include "models.h"2 3llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :4    llm_graph_context(params) {5    const int64_t n_embd_head = hparams.n_embd_head_v();6 7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());8    GGML_ASSERT(n_embd_head == n_rot);9 10    ggml_tensor * cur;11    ggml_tensor * inpL;12 13    inpL = build_inp_embd(model.tok_embd);14 15    // inp_pos - contains the positions16    ggml_tensor * inp_pos = build_inp_pos();17 18    auto * inp_attn = build_attn_inp_kv();19 20    ggml_tensor * inp_out_ids = build_inp_out_ids();21 22    for (int il = 0; il < n_layer; ++il) {23        ggml_tensor * inpSA = inpL;24 25        // norm26        {27            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);28            cb(cur, "attn_norm", il);29        }30        // self-attention31        {32            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,33                    n_embd_head, n_head, n_head_kv, il);34 35            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,36                                 ext_factor, attn_factor, beta_fast, beta_slow);37 38            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,39                                 ext_factor, attn_factor, beta_fast, beta_slow);40 41            cb(Qcur, "Qcur", il);42            cb(Kcur, "Kcur", il);43            cb(Vcur, "Vcur", il);44 45            cur = build_attn(inp_attn,46                    model.layers[il].wo, NULL, model.layers[il].wo_s,47                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);48        }49        if (il == n_layer - 1) {50            // skip computing output for unused tokens51            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);52            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);53        }54        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);55        cb(ffn_inp, "ffn_inp", il);56 57        // feed-forward network58        {59            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);60            cb(cur, "ffn_norm", il);61 62            cur = build_ffn(cur,63                    model.layers[il].ffn_up, NULL, NULL,64                    model.layers[il].ffn_gate, NULL, NULL,65                    model.layers[il].ffn_down, NULL, NULL,66                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);67            cb(cur, "ffn_out", il);68        }69        cur = ggml_add(ctx0, cur, ffn_inp);70 71        cur = build_cvec(cur, il);72        cb(cur, "l_out", il);73 74        // input for next layer75        inpL = cur;76    }77    cur = inpL;78 79    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);80 81    cb(cur, "result_norm", -1);82    res->t_embd = cur;83 84    // lm_head85    cur = build_lora_mm(model.output, cur);86 87    cb(cur, "result_output", -1);88    res->t_logits = cur;89 90    ggml_build_forward_expand(gf, cur);91}92