1#include "models.h"
  2
  3llm_build_qwen2::llm_build_qwen2(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    GGML_ASSERT(n_embd_head == hparams.n_rot);
  8
  9    ggml_tensor * cur;
 10    ggml_tensor * inpL;
 11
 12    inpL = build_inp_embd(model.tok_embd);
 13
 14    // inp_pos - contains the positions
 15    ggml_tensor * inp_pos = build_inp_pos();
 16
 17    auto * inp_attn = build_attn_inp_kv();
 18
 19    ggml_tensor * inp_out_ids = build_inp_out_ids();
 20
 21    for (int il = 0; il < n_layer; ++il) {
 22        ggml_tensor * inpSA = inpL;
 23
 24        // norm
 25        cur = build_norm(inpL,
 26                model.layers[il].attn_norm, NULL,
 27                LLM_NORM_RMS, il);
 28        cb(cur, "attn_norm", il);
 29
 30        // self-attention
 31        {
 32            // compute Q and K and RoPE them
 33            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
 34            cb(Qcur, "Qcur", il);
 35            if (model.layers[il].bq) {
 36                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
 37                cb(Qcur, "Qcur", il);
 38            }
 39
 40            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 41            cb(Kcur, "Kcur", il);
 42            if (model.layers[il].bk) {
 43                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
 44                cb(Kcur, "Kcur", il);
 45            }
 46
 47            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 48            cb(Vcur, "Vcur", il);
 49            if (model.layers[il].bv) {
 50                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
 51                cb(Vcur, "Vcur", il);
 52            }
 53
 54            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
 55            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 56            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 57
 58            Qcur = ggml_rope_ext(
 59                    ctx0, Qcur, inp_pos, nullptr,
 60                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 61                    ext_factor, attn_factor, beta_fast, beta_slow
 62                    );
 63
 64            Kcur = ggml_rope_ext(
 65                    ctx0, Kcur, inp_pos, nullptr,
 66                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 67                    ext_factor, attn_factor, beta_fast, beta_slow
 68                    );
 69
 70            cb(Qcur, "Qcur", il);
 71            cb(Kcur, "Kcur", il);
 72            cb(Vcur, "Vcur", il);
 73
 74            cur = build_attn(inp_attn,
 75                    model.layers[il].wo, model.layers[il].bo,
 76                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 77        }
 78        if (il == n_layer - 1 && inp_out_ids) {
 79            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
 80            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
 81        }
 82        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
 83        cb(ffn_inp, "ffn_inp", il);
 84
 85        // feed-forward network
 86        cur = build_norm(ffn_inp,
 87                model.layers[il].ffn_norm, NULL,
 88                LLM_NORM_RMS, il);
 89        cb(cur, "ffn_norm", il);
 90
 91        cur = build_ffn(cur,
 92                model.layers[il].ffn_up,   NULL, NULL,
 93                model.layers[il].ffn_gate, NULL, NULL,
 94                model.layers[il].ffn_down, NULL, NULL,
 95                NULL,
 96                LLM_FFN_SILU, LLM_FFN_PAR, il);
 97        cb(cur, "ffn_out", il);
 98
 99        cur = ggml_add(ctx0, cur, ffn_inp);
100
101        cur = build_cvec(cur, il);
102        cb(cur, "l_out", il);
103
104        // input for next layer
105        inpL = cur;
106    }
107    cur = inpL;
108
109    cur = build_norm(cur,
110            model.output_norm, NULL,
111            LLM_NORM_RMS, -1);
112
113    cb(cur, "result_norm", -1);
114    res->t_embd = cur;
115
116    // lm_head
117    cur = build_lora_mm(model.output, cur);
118
119    if (model.output_b != nullptr) {
120        cur = ggml_add(ctx0, cur, model.output_b);
121    }
122    cb(cur, "result_output", -1);
123    res->t_logits = cur;
124
125    ggml_build_forward_expand(gf, cur);
126}