1#include "models.h"
  2
  3
  4llm_build_qwen::llm_build_qwen(const llama_model & model, const llm_graph_params & params) : 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
  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        cur = build_norm(inpL,
 25                model.layers[il].attn_norm, NULL,
 26                LLM_NORM_RMS, il);
 27        cb(cur, "attn_norm", il);
 28
 29        // self-attention
 30        {
 31            cur = build_lora_mm(model.layers[il].wqkv, cur);
 32            cb(cur, "wqkv", il);
 33
 34            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
 35            cb(cur, "bqkv", il);
 36
 37            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
 38            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
 39            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 2*sizeof(float)*(n_embd));
 40
 41            // using mode = 2 for neox mode
 42            Qcur = ggml_rope_ext(
 43                    ctx0, Qcur, inp_pos, nullptr,
 44                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 45                    ext_factor, attn_factor, beta_fast, beta_slow
 46                    );
 47
 48            Kcur = ggml_rope_ext(
 49                    ctx0, Kcur, inp_pos, nullptr,
 50                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 51                    ext_factor, attn_factor, beta_fast, beta_slow
 52                    );
 53
 54            cb(Qcur, "Qcur", il);
 55            cb(Kcur, "Kcur", il);
 56            cb(Vcur, "Vcur", il);
 57
 58            cur = build_attn(inp_attn,
 59                    model.layers[il].wo, NULL,
 60                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 61        }
 62        if (il == n_layer - 1 && inp_out_ids) {
 63            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 forward
 70        {
 71            cur = build_norm(ffn_inp,
 72                    model.layers[il].ffn_norm, NULL,
 73                    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,
 81                    LLM_FFN_SILU, LLM_FFN_PAR, il);
 82            cb(cur, "ffn_out", il);
 83        }
 84        cur = ggml_add(ctx0, cur, ffn_inp);
 85
 86        cur = build_cvec(cur, il);
 87        cb(cur, "l_out", il);
 88
 89        // input for next layer
 90        inpL = cur;
 91    }
 92    cur = inpL;
 93
 94    cur = build_norm(cur,
 95            model.output_norm, NULL,
 96            LLM_NORM_RMS, -1);
 97
 98    cb(cur, "result_norm", -1);
 99    res->t_embd = cur;
100
101    // lm_head
102    cur = build_lora_mm(model.output, cur);
103
104    cb(cur, "result_output", -1);
105    res->t_logits = cur;
106
107    ggml_build_forward_expand(gf, cur);
108}