1#include "models.h"
  2
  3llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
  4    llm_graph_context(params) {
  5    const int64_t n_embd_head = hparams.n_embd_head_v;
  6    const float   kq_scale    = 1.0f / sqrtf(float(n_embd_head));
  7
  8    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
  9    GGML_ASSERT(n_embd_head == hparams.n_rot);
 10
 11    ggml_tensor * inpL;
 12    ggml_tensor * cur;
 13
 14    inpL = build_inp_embd(model.tok_embd);
 15
 16    ggml_tensor * inp_pos = build_inp_pos();
 17
 18    auto * inp_attn = build_attn_inp_kv();
 19
 20    // check ubatch to see if we have input tokens (text)
 21    // or an input embedding vector (image)
 22    bool is_text;
 23    if (ubatch.token) {
 24        is_text = true;
 25    } else {
 26        is_text = false;
 27    }
 28
 29    for (int il = 0; il < n_layer; ++il) {
 30        // get either the text or image weight tensors
 31        ggml_tensor *wqkv, *wo;
 32        ggml_tensor *ffn_gate, *ffn_down, *ffn_up;
 33
 34        if (is_text) {
 35            wqkv     = model.layers[il].wqkv;
 36            wo       = model.layers[il].wo;
 37            ffn_gate = model.layers[il].ffn_gate;
 38            ffn_down = model.layers[il].ffn_down;
 39            ffn_up   = model.layers[il].ffn_up;
 40        } else {
 41            wqkv     = model.layers[il].visexp_attn_wqkv;
 42            wo       = model.layers[il].visexp_attn_wo;
 43            ffn_gate = model.layers[il].visexp_ffn_gate;
 44            ffn_down = model.layers[il].visexp_ffn_down;
 45            ffn_up   = model.layers[il].visexp_ffn_up;
 46        }
 47
 48        ggml_tensor * inpSA = inpL;
 49        cur = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 50
 51        // build self attention
 52        {
 53            ggml_tensor * qkv = build_lora_mm(wqkv, cur);
 54
 55            // split qkv into Q, K, V along the first dimension
 56            ggml_tensor * Qcur =
 57                ggml_view_3d(ctx0, qkv, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), qkv->nb[1], 0);
 58            ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
 59                                              qkv->nb[1], n_embd * ggml_element_size(qkv));
 60            ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
 61                                              qkv->nb[1], 2 * n_embd * ggml_element_size(qkv));
 62
 63            Qcur = ggml_rope(ctx0, Qcur, inp_pos, n_embd_head, rope_type);
 64            Kcur = ggml_rope(ctx0, Kcur, inp_pos, n_embd_head, rope_type);
 65
 66            cur = build_attn(inp_attn,
 67                wo, nullptr,
 68                Qcur, Kcur, Vcur,
 69                nullptr, nullptr, nullptr,
 70                kq_scale, il);
 71            cb(cur, "attn_out", il);
 72        }
 73
 74        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
 75        cb(ffn_inp, "ffn_inp", il);
 76
 77        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
 78        cb(cur, "ffn_norm", il);
 79
 80        cur = build_ffn(cur,
 81                ffn_up, NULL, NULL,
 82                ffn_gate, NULL, NULL,
 83                ffn_down, NULL, NULL,
 84                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
 85
 86        cur = ggml_add(ctx0, cur, ffn_inp);
 87        cb(cur, "ffn_out", il);
 88
 89        inpL = cur;
 90    }
 91
 92    cur = inpL;
 93
 94    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 95    cb(cur, "result_norm", -1);
 96    res->t_embd = cur;
 97
 98    cur = build_lora_mm(model.output, cur);
 99    cb(cur, "result_output", -1);
100    res->t_logits = cur;
101    ggml_build_forward_expand(gf, cur);
102}