1#include "models.h"
  2
  3
  4llm_build_gemma::llm_build_gemma(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_tensor * cur;
  8    ggml_tensor * inpL;
  9
 10    inpL = build_inp_embd(model.tok_embd);
 11
 12    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
 13    cb(inpL, "inp_scaled", -1);
 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        // norm
 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            // compute Q and K and RoPE them
 32            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
 33            cb(Qcur, "Qcur", il);
 34
 35            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 36            cb(Kcur, "Kcur", il);
 37
 38            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 39            cb(Vcur, "Vcur", il);
 40
 41            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
 42            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 43            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 44
 45            Qcur = ggml_rope_ext(
 46                    ctx0, Qcur, inp_pos, nullptr,
 47                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 48                    ext_factor, attn_factor, beta_fast, beta_slow);
 49
 50            Kcur = ggml_rope_ext(
 51                    ctx0, Kcur, 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            cb(Qcur, "Qcur", il);
 56            cb(Kcur, "Kcur", il);
 57            cb(Vcur, "Vcur", il);
 58
 59            Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
 60            cb(Qcur, "Qcur_scaled", il);
 61
 62            cur = build_attn(inp_attn,
 63                    model.layers[il].wo, NULL,
 64                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
 65        }
 66        if (il == n_layer - 1 && inp_out_ids) {
 67            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
 68            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
 69        }
 70        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
 71        cb(sa_out, "sa_out", il);
 72
 73        cur = build_norm(sa_out,
 74                model.layers[il].ffn_norm, NULL,
 75                LLM_NORM_RMS, il);
 76        cb(cur, "ffn_norm", il);
 77
 78        // feed-forward network
 79        {
 80            cur = build_ffn(cur,
 81                    model.layers[il].ffn_up,   NULL, NULL,
 82                    model.layers[il].ffn_gate, NULL, NULL,
 83                    model.layers[il].ffn_down, NULL, NULL,
 84                    NULL,
 85                    LLM_FFN_GELU, LLM_FFN_PAR, il);
 86            cb(cur, "ffn_out", il);
 87        }
 88        cur = ggml_add(ctx0, cur, sa_out);
 89
 90        cur = build_cvec(cur, il);
 91        cb(cur, "l_out", il);
 92
 93        // input for next layer
 94        inpL = cur;
 95    }
 96    cur = inpL;
 97
 98    cur = build_norm(cur,
 99            model.output_norm, NULL,
100            LLM_NORM_RMS, -1);
101
102    cb(cur, "result_norm", -1);
103    res->t_embd = cur;
104
105    // lm_head
106    cur = build_lora_mm(model.output, cur);
107
108    cb(cur, "result_output", -1);
109    res->t_logits = cur;
110
111    ggml_build_forward_expand(gf, cur);
112}