1#include "models.h"
  2
  3
  4
  5llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
  6    llm_graph_context(params) {
  7    //copied from qwen2
  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 == hparams.n_rot);
 12
 13    ggml_tensor * cur;
 14    ggml_tensor * inpL;
 15
 16    inpL = build_inp_embd(model.tok_embd);
 17
 18    // inp_pos - contains the positions
 19    ggml_tensor * inp_pos = build_inp_pos();
 20
 21    auto * inp_attn = build_attn_inp_no_cache();
 22
 23    ggml_tensor * inp_out_ids = build_inp_out_ids();
 24
 25    for (int il = 0; il < n_layer; ++il) {
 26        ggml_tensor * inpSA = inpL;
 27
 28        // norm
 29        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 30        cb(cur, "attn_norm", il);
 31
 32        // self-attention
 33        {
 34            // compute Q and K and RoPE them
 35            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
 36            Qcur               = ggml_add(ctx0, Qcur, model.layers[il].bq);
 37            cb(Qcur, "Qcur", il);
 38
 39            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 40            Kcur               = ggml_add(ctx0, Kcur, model.layers[il].bk);
 41            cb(Kcur, "Kcur", il);
 42
 43            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 44            Vcur               = ggml_add(ctx0, Vcur, model.layers[il].bv);
 45            cb(Vcur, "Vcur", il);
 46
 47            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 48            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 49            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 50
 51            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 52                                 ext_factor, attn_factor, beta_fast, beta_slow);
 53
 54            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 55                                 ext_factor, attn_factor, beta_fast, beta_slow);
 56
 57            cb(Qcur, "Qcur", il);
 58            cb(Kcur, "Kcur", il);
 59            cb(Vcur, "Vcur", il);
 60
 61            cur = build_attn(inp_attn,
 62                    model.layers[il].wo, model.layers[il].bo,
 63                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
 64        }
 65        if (il == n_layer - 1 && inp_out_ids) {
 66            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
 67            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
 68        }
 69        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
 70        cb(ffn_inp, "ffn_inp", il);
 71
 72        // feed-forward network
 73        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
 74        cb(cur, "ffn_norm", il);
 75
 76        cur = build_ffn(cur,
 77            model.layers[il].ffn_up, NULL, NULL,
 78            model.layers[il].ffn_gate, NULL, NULL,
 79            model.layers[il].ffn_down, NULL, NULL,
 80            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
 81        cb(cur, "ffn_out", il);
 82
 83        cur = ggml_add(ctx0, cur, ffn_inp);
 84
 85        cur = build_cvec(cur, il);
 86        cb(cur, "l_out", il);
 87
 88        // input for next layer
 89        inpL = cur;
 90    }
 91    cur = inpL;
 92
 93    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 94
 95    cb(cur, "result_norm", -1);
 96    res->t_embd = cur;
 97
 98    // lm_head
 99    cur = build_lora_mm(model.output, cur);
100
101    cb(cur, "result_output", -1);
102    res->t_logits = cur;
103
104    ggml_build_forward_expand(gf, cur);
105}