echodict/llama.cpp

1#include "models.h"2 3llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :4    llm_graph_context(params) {5    //copied from qwen26    const int64_t n_embd_head = hparams.n_embd_head_v();7 8    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());9    GGML_ASSERT(n_embd_head == n_rot);10 11    ggml_tensor * cur;12    ggml_tensor * inpL;13 14    inpL = build_inp_embd(model.tok_embd);15 16    // inp_pos - contains the positions17    ggml_tensor * inp_pos = build_inp_pos();18 19    auto * inp_attn = build_attn_inp_no_cache();20 21    ggml_tensor * inp_out_ids = build_inp_out_ids();22 23    for (int il = 0; il < n_layer; ++il) {24        ggml_tensor * inpSA = inpL;25 26        // norm27        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, model.layers[il].wo_b, 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 && inp_out_ids) {50            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);51            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);52        }53        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);54        cb(ffn_inp, "ffn_inp", il);55 56        // feed-forward network57        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);58        cb(cur, "ffn_norm", il);59 60        cur = build_ffn(cur,61            model.layers[il].ffn_up, NULL, NULL,62            model.layers[il].ffn_gate, NULL, NULL,63            model.layers[il].ffn_down, NULL, NULL,64            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);65        cb(cur, "ffn_out", il);66 67        cur = ggml_add(ctx0, cur, ffn_inp);68 69        cur = build_cvec(cur, il);70        cb(cur, "l_out", il);71 72        // input for next layer73        inpL = cur;74    }75    cur = inpL;76 77    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);78 79    cb(cur, "result_norm", -1);80    res->t_embd = cur;81 82    // lm_head83    cur = build_lora_mm(model.output, cur);84 85    cb(cur, "result_output", -1);86    res->t_logits = cur;87 88    ggml_build_forward_expand(gf, cur);89}90