1#include "models.h"
  2
  3llm_build_plamo::llm_build_plamo(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        // norm
 23        cur = build_norm(inpL,
 24                model.layers[il].attn_norm, NULL,
 25                LLM_NORM_RMS, il);
 26        cb(cur, "attn_norm", il);
 27
 28        ggml_tensor * sa_inp = cur;
 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
 36            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 37            cb(Kcur, "Kcur", il);
 38
 39            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 40            cb(Vcur, "Vcur", il);
 41
 42            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
 43            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 44            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 45
 46            Qcur = ggml_rope_ext(
 47                    ctx0, Qcur, inp_pos, nullptr,
 48                    n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
 49                    ext_factor, attn_factor, beta_fast, beta_slow
 50                    );
 51
 52            Kcur = ggml_rope_ext(
 53                    ctx0, Kcur, inp_pos, nullptr,
 54                    n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
 55                    ext_factor, attn_factor, beta_fast, beta_slow
 56                    );
 57
 58            cb(Qcur, "Qcur", il);
 59            cb(Kcur, "Kcur", il);
 60            cb(Vcur, "Vcur", il);
 61
 62            cur = build_attn(inp_attn,
 63                    model.layers[il].wo, NULL,
 64                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 65        }
 66        if (il == n_layer - 1 && inp_out_ids) {
 67            cur    = ggml_get_rows(ctx0,    cur, inp_out_ids);
 68            sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids);
 69            inpL   = ggml_get_rows(ctx0,   inpL, inp_out_ids);
 70        }
 71        ggml_tensor * sa_out = cur;
 72
 73        cur = sa_inp;
 74
 75        // feed-forward network
 76        {
 77            cur = build_ffn(cur,
 78                    model.layers[il].ffn_up,   NULL, NULL,
 79                    model.layers[il].ffn_gate, NULL, NULL,
 80                    model.layers[il].ffn_down, NULL, NULL,
 81                    NULL,
 82                    LLM_FFN_SILU, LLM_FFN_PAR, il);
 83            cb(cur, "ffn_out", il);
 84        }
 85        cur = ggml_add(ctx0, cur, sa_out);
 86        cur = ggml_add(ctx0, cur, inpL);
 87
 88        cur = build_cvec(cur, il);
 89        cb(cur, "l_out", il);
 90
 91        // input for next layer
 92        inpL = cur;
 93    }
 94    cur = inpL;
 95
 96    cur = build_norm(cur,
 97            model.output_norm, NULL,
 98            LLM_NORM_RMS, -1);
 99
100    cb(cur, "result_norm", -1);
101    res->t_embd = cur;
102
103    // lm_head
104    cur = build_lora_mm(model.output, cur);
105
106    cb(cur, "result_output", -1);
107    res->t_logits = cur;
108
109    ggml_build_forward_expand(gf, cur);
110}