echodict/llama.cpp

1#include "models.h"2 3llm_build_paddleocr::llm_build_paddleocr(const llama_model & model, const llm_graph_params & params) :4    llm_graph_context(params) {5 6    // NOTE: same with qwen2vl.cpp, but bias tensors are optional7 8    const int64_t n_embd_head = hparams.n_embd_head_v();9 10    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());11    GGML_ASSERT(n_embd_head == n_rot);12 13    ggml_tensor * cur;14    ggml_tensor * inpL;15 16    inpL = build_inp_embd(model.tok_embd);17 18    int sections[4];19    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);20 21    // inp_pos - contains the positions22    ggml_tensor * inp_pos = build_inp_pos();23 24    auto * inp_attn = build_attn_inp_kv();25 26    ggml_tensor * inp_out_ids = build_inp_out_ids();27 28    for (int il = 0; il < n_layer; ++il) {29        ggml_tensor * inpSA = inpL;30 31        // norm32        {33            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);34            cb(cur, "attn_norm", il);35        }36        // self-attention37        {38            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,39                    n_embd_head, n_head, n_head_kv, il);40 41            Qcur = ggml_rope_multi(42                    ctx0, Qcur, inp_pos, nullptr,43                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,44                    ext_factor, attn_factor, beta_fast, beta_slow45                    );46 47            Kcur = ggml_rope_multi(48                    ctx0, Kcur, inp_pos, nullptr,49                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,50                    ext_factor, attn_factor, beta_fast, beta_slow51                    );52 53            cb(Qcur, "Qcur", il);54            cb(Kcur, "Kcur", il);55            cb(Vcur, "Vcur", il);56 57            cur = build_attn(inp_attn,58                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,59                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);60        }61        if (il == n_layer - 1) {62            // skip computing output for unused tokens63            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);64            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);65        }66        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);67        cb(ffn_inp, "ffn_inp", il);68 69        // feed-forward network70        {71            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);72            cb(cur, "ffn_norm", il);73 74            cur = build_ffn(cur,75                    model.layers[il].ffn_up, NULL, NULL,76                    model.layers[il].ffn_gate, NULL, NULL,77                    model.layers[il].ffn_down, NULL, NULL,78                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);79            cb(cur, "ffn_out", il);80        }81        cur = ggml_add(ctx0, cur, ffn_inp);82 83        cur = build_cvec(cur, il);84        cb(cur, "l_out", il);85 86        // input for next layer87        inpL = cur;88    }89    cur = inpL;90 91    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);92 93    cb(cur, "result_norm", -1);94    res->t_embd = cur;95 96    // lm_head97    cur = build_lora_mm(model.output, cur);98 99    cb(cur, "result_output", -1);100    res->t_logits = cur;101 102    ggml_build_forward_expand(gf, cur);103}104