1#include "models.h"
  2
  3
  4llm_build_pangu_embedded::llm_build_pangu_embedded(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    GGML_ASSERT(n_embd_head == hparams.n_rot);
  9
 10    ggml_tensor * cur;
 11    ggml_tensor * inpL;
 12
 13    inpL = build_inp_embd(model.tok_embd);
 14
 15    // inp_pos - contains the positions
 16    ggml_tensor * inp_pos = build_inp_pos();
 17
 18    auto * inp_attn = build_attn_inp_kv();
 19
 20    ggml_tensor * inp_out_ids = build_inp_out_ids();
 21
 22    for (int il = 0; il < n_layer; ++il) {
 23        ggml_tensor * inpSA = inpL;
 24
 25        // norm
 26        cur = build_norm(inpL,
 27                model.layers[il].attn_norm, NULL,
 28                LLM_NORM_RMS, il);
 29        cb(cur, "attn_norm", il);
 30
 31        // self attention
 32        {
 33            // compute Q and K and RoPE them
 34            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
 35            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
 36            cb(Qcur, "Qcur", il);
 37
 38            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 39            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
 40            cb(Kcur, "Kcur", il);
 41
 42            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 43            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
 44            cb(Vcur, "Vcur", il);
 45
 46            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
 47            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 48            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 49
 50            Qcur = ggml_rope_ext(
 51                    ctx0, Qcur, inp_pos, nullptr,
 52                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 53                    ext_factor, attn_factor, beta_fast, beta_slow
 54                    );
 55
 56            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,
 57                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 58                    ext_factor, attn_factor, beta_fast, beta_slow
 59                    );
 60
 61            cb(Qcur, "Qcur", il);
 62            cb(Kcur, "Kcur", il);
 63            cb(Vcur, "Vcur", il);
 64
 65            cur = build_attn(inp_attn,
 66                    model.layers[il].wo, model.layers[il].bo,
 67                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 68        }
 69
 70        if (il == n_layer - 1 && inp_out_ids) {
 71            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
 72            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
 73        }
 74
 75        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
 76        cb(ffn_inp, "ffn_inp", il);
 77
 78        // feed-forward network
 79        cur = build_norm(ffn_inp,
 80                model.layers[il].ffn_norm, NULL,
 81                LLM_NORM_RMS, il);
 82        cb(cur, "ffn_norm", il);
 83
 84        cur = build_ffn(cur,
 85                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
 86                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
 87                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
 88                NULL,
 89                LLM_FFN_SILU, LLM_FFN_PAR, il);
 90
 91        cur = ggml_add(ctx0, cur, ffn_inp);
 92        cb(cur, "ffn_out", il);
 93
 94        cur = build_cvec(cur, il);
 95        cb(cur, "l_out", il);
 96
 97        // input for next layer
 98        inpL = cur;
 99    }
100
101    cur = inpL;
102
103    cur = build_norm(cur,
104            model.output_norm, NULL,
105            LLM_NORM_RMS, -1);
106
107    cb(cur, "result_norm", -1);
108    res->t_embd = cur;
109
110    // lm_head
111    cur = build_lora_mm(model.output, cur);
112
113    if (model.output_b != nullptr) {
114        cur = ggml_add(ctx0, cur, model.output_b);
115    }
116
117    cb(cur, "result_output", -1);
118    res->t_logits = cur;
119
120    ggml_build_forward_expand(gf, cur);
121}