1#include "models.h"
  2
  3llm_build_bloom::llm_build_bloom(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    const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
  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    auto * inp_attn = build_attn_inp_kv();
 15
 16    inpL = build_norm(inpL,
 17            model.tok_norm,
 18            model.tok_norm_b,
 19            LLM_NORM, -1);
 20    cb(inpL, "inp_norm", -1);
 21
 22    ggml_tensor * inp_out_ids = build_inp_out_ids();
 23
 24    for (int il = 0; il < n_layer; ++il) {
 25        cur = build_norm(inpL,
 26                model.layers[il].attn_norm,
 27                model.layers[il].attn_norm_b,
 28                LLM_NORM, il);
 29        cb(cur, "attn_norm", il);
 30
 31        // self-attention
 32        {
 33            cur = build_lora_mm(model.layers[il].wqkv, cur);
 34            cb(cur, "wqkv", il);
 35
 36            cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
 37            cb(cur, "bqkv", il);
 38
 39            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));
 40            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));
 41            ggml_tensor * Vcur = 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 + n_embd_gqa));
 42
 43            cb(Qcur, "Qcur", il);
 44            cb(Kcur, "Kcur", il);
 45            cb(Vcur, "Vcur", il);
 46
 47            cur = build_attn(inp_attn,
 48                    model.layers[il].wo, model.layers[il].bo,
 49                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 50        }
 51
 52        if (il == n_layer - 1 && inp_out_ids) {
 53            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
 54            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
 55        }
 56
 57        // Add the input
 58        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
 59        cb(ffn_inp, "ffn_inp", il);
 60
 61        // FF
 62        {
 63            cur = build_norm(ffn_inp,
 64                    model.layers[il].ffn_norm,
 65                    model.layers[il].ffn_norm_b,
 66                    LLM_NORM, il);
 67            cb(cur, "ffn_norm", il);
 68
 69            cur = build_ffn(cur,
 70                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
 71                    NULL,                      NULL,                        NULL,
 72                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
 73                    NULL,
 74                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
 75            cb(cur, "ffn_out", il);
 76        }
 77
 78        cur = ggml_add(ctx0, cur, ffn_inp);
 79
 80        cur = build_cvec(cur, il);
 81        cb(cur, "l_out", il);
 82
 83        // input for next layer
 84        inpL = cur;
 85    }
 86
 87    cur = build_norm(inpL,
 88            model.output_norm,
 89            model.output_norm_b,
 90            LLM_NORM, -1);
 91
 92    cb(cur, "result_norm", -1);
 93    res->t_embd = cur;
 94
 95    cur = build_lora_mm(model.output, cur);
 96
 97    cb(cur, "result_output", -1);
 98    res->t_logits = cur;
 99
100    ggml_build_forward_expand(gf, cur);
101}