echodict/llama.cpp

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 == 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 tensors31        ggml_tensor *wqkv, *wo, *wo_s;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            wo_s     = model.layers[il].wo_s;38            ffn_gate = model.layers[il].ffn_gate;39            ffn_down = model.layers[il].ffn_down;40            ffn_up   = model.layers[il].ffn_up;41        } else {42            wqkv     = model.layers[il].visexp_attn_wqkv;43            wo       = model.layers[il].visexp_attn_wo;44            wo_s     = nullptr;45            ffn_gate = model.layers[il].visexp_ffn_gate;46            ffn_down = model.layers[il].visexp_ffn_down;47            ffn_up   = model.layers[il].visexp_ffn_up;48        }49 50        ggml_tensor * inpSA = inpL;51        cur = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);52 53        // build self attention54        {55            ggml_tensor * qkv = build_lora_mm(wqkv, cur);56 57            // split qkv into Q, K, V along the first dimension58            ggml_tensor * Qcur =59                ggml_view_3d(ctx0, qkv, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), qkv->nb[1], 0);60            ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),61                                              qkv->nb[1], n_embd * ggml_element_size(qkv));62            ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),63                                              qkv->nb[1], 2 * n_embd * ggml_element_size(qkv));64 65            Qcur = ggml_rope(ctx0, Qcur, inp_pos, n_embd_head, rope_type);66            Kcur = ggml_rope(ctx0, Kcur, inp_pos, n_embd_head, rope_type);67 68            cur = build_attn(inp_attn,69                wo, nullptr, wo_s,70                Qcur, Kcur, Vcur,71                nullptr, nullptr, nullptr,72                kq_scale, il);73            cb(cur, "attn_out", il);74        }75 76        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);77        cb(ffn_inp, "ffn_inp", il);78 79        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);80        cb(cur, "ffn_norm", il);81 82        cur = build_ffn(cur,83                ffn_up, NULL, NULL,84                ffn_gate, NULL, NULL,85                ffn_down, NULL, NULL,86                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);87 88        cur = ggml_add(ctx0, cur, ffn_inp);89        cb(cur, "ffn_out", il);90 91        cur = build_cvec(cur, il);92        cb(cur, "l_out", il);93 94        // input for next layer95        inpL = cur;96    }97 98    cur = inpL;99 100    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);101    cb(cur, "result_norm", -1);102    res->t_embd = cur;103 104    cur = build_lora_mm(model.output, cur);105    cb(cur, "result_output", -1);106    res->t_logits = cur;107    ggml_build_forward_expand(gf, cur);108}109