echodict/llama.cpp

1#include "models.h"2 3#include <float.h>4 5llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {6    const int64_t n_embd_head = hparams.n_embd_head_v();7 8    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());9    GGML_ASSERT(n_embd_head == n_rot);10 11    ggml_tensor * cur;12    ggml_tensor * inpL;13 14    inpL = build_inp_embd(model.tok_embd);15 16    // inp_pos - contains the positions17    ggml_tensor * inp_pos = build_inp_pos();18 19    auto * inp_attn = build_attn_inp_kv();20 21    ggml_tensor * inp_out_ids = build_inp_out_ids();22 23    for (int il = 0; il < n_layer; ++il) {24        ggml_tensor * inpSA = inpL;25 26        // norm27        if (hparams.swin_norm) {28            cur = inpL;29        } else {30            cur = build_norm(inpL,31                    model.layers[il].attn_norm, NULL,32                    LLM_NORM_RMS, il);33            cb(cur, "attn_norm", il);34        }35 36        // self-attention37        {38            // compute Q and K and RoPE them39            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,40                    n_embd_head, n_head, n_head_kv, il);41 42            if (model.layers[il].attn_q_norm) {43                Qcur = build_norm(Qcur,44                        model.layers[il].attn_q_norm,45                        model.layers[il].attn_q_norm_b,46                        LLM_NORM, il);47                cb(Qcur, "Qcur", il);48            }49 50            if (model.layers[il].attn_k_norm) {51                Kcur = build_norm(Kcur,52                        model.layers[il].attn_k_norm,53                        model.layers[il].attn_k_norm_b,54                        LLM_NORM, il);55                cb(Kcur, "Kcur", il);56            }57 58            Qcur = ggml_rope_ext(59                    ctx0, Qcur, inp_pos, nullptr,60                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,61                    ext_factor, attn_factor, beta_fast, beta_slow62                    );63 64            Kcur = ggml_rope_ext(65                    ctx0, Kcur, inp_pos, nullptr,66                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,67                    ext_factor, attn_factor, beta_fast, beta_slow68                    );69 70            cb(Qcur, "Qcur", il);71            cb(Kcur, "Kcur", il);72            cb(Vcur, "Vcur", il);73 74            cur = build_attn(inp_attn,75                    model.layers[il].wo, nullptr, model.layers[il].wo_s,76                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);77        }78 79        if (il == n_layer - 1 && inp_out_ids) {80            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);81            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);82        }83 84        if (hparams.swin_norm) {85            cur = build_norm(cur,86                    model.layers[il].attn_norm, NULL,87                    LLM_NORM_RMS, il);88        }89 90        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);91        cb(ffn_inp, "ffn_inp", il);92 93        // feed-forward network94        if (!hparams.swin_norm) {95            cur = build_norm(ffn_inp,96                    model.layers[il].ffn_norm, NULL,97                    LLM_NORM_RMS, il);98            cb(cur, "ffn_norm", il);99        }100 101        cur = build_ffn(cur,102                model.layers[il].ffn_up,   NULL, NULL,103                model.layers[il].ffn_gate, NULL, NULL,104                model.layers[il].ffn_down, NULL, NULL,105                NULL,106                LLM_FFN_SILU, LLM_FFN_PAR, il);107        cb(cur, "ffn_out", il);108 109        if (hparams.swin_norm) {110            cur = build_norm(cur,111                    model.layers[il].ffn_norm, NULL,112                    LLM_NORM_RMS, il);113            cb(cur, "ffn_norm", il);114        }115 116        cur = ggml_add(ctx0, cur, ffn_inp);117        cb(cur, "ffn_out", il);118 119        cur = build_cvec(cur, il);120        cb(cur, "l_out", il);121 122        // input for next layer123        inpL = cur;124    }125 126    cur = inpL;127 128    cur = build_norm(cur,129            model.output_norm, NULL,130            LLM_NORM_RMS, -1);131 132    cb(cur, "result_norm", -1);133    res->t_embd = cur;134 135    // lm_head136    cur = build_lora_mm(model.output, cur);137    cb(cur, "result_output_with_img_logits", -1);138 139    // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.140    // Needs to be removed once image outputs are supported.141    int img_token_end_idx = 8196;142    int img_token_start_idx = 4;143    int num_img_tokens = img_token_end_idx - img_token_start_idx;144    // creates 1d tensor of size num_img_tokens and values -FLT_MAX,145    // which ensures that text token values are always at least larger than image token values146    ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);147    img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);148    cb(img_logits, "img_logits", -1);149 150    cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);151 152    cb(cur, "result_output", -1);153    res->t_logits = cur;154 155    ggml_build_forward_expand(gf, cur);156}157