echodict/llama.cpp

1#include "models.h"2 3llm_build_falcon::llm_build_falcon(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 == 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 positions15    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        ggml_tensor * attn_norm;23 24        attn_norm = build_norm(inpL,25                model.layers[il].attn_norm,26                model.layers[il].attn_norm_b,27                LLM_NORM, il);28        cb(attn_norm, "attn_norm", il);29 30        // self-attention31        {32            if (model.layers[il].attn_norm_2) {33                // Falcon-40B34                cur = build_norm(inpL,35                        model.layers[il].attn_norm_2,36                        model.layers[il].attn_norm_2_b,37                        LLM_NORM, il);38                cb(cur, "attn_norm_2", il);39            } else {40                cur = attn_norm;41            }42 43            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,44                    n_embd_head, n_head, n_head_kv, il);45 46            // using mode = 2 for neox mode47            Qcur = ggml_rope_ext(48                    ctx0, Qcur, inp_pos, nullptr,49                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,50                    ext_factor, attn_factor, beta_fast, beta_slow51                    );52 53            Kcur = ggml_rope_ext(54                    ctx0, Kcur, inp_pos, nullptr,55                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,56                    ext_factor, attn_factor, beta_fast, beta_slow57                    );58 59            cb(Qcur, "Qcur", il);60            cb(Kcur, "Kcur", il);61            cb(Vcur, "Vcur", il);62 63            cur = build_attn(inp_attn,64                    model.layers[il].wo, NULL, model.layers[il].wo_s,65                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);66        }67 68        if (il == n_layer - 1 && inp_out_ids) {69            cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);70            inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);71            attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);72        }73 74        ggml_tensor * ffn_inp = cur;75 76        // feed forward77        {78            cur = build_ffn(attn_norm, // !! use the attn norm, not the result79                    model.layers[il].ffn_up,   NULL, NULL,80                    NULL,                      NULL, NULL,81                    model.layers[il].ffn_down, NULL, NULL,82                    NULL,83                    LLM_FFN_GELU, LLM_FFN_SEQ, il);84            cb(cur, "ffn_out", il);85        }86 87        cur = ggml_add(ctx0, cur, ffn_inp);88        cur = ggml_add(ctx0, cur, inpL);89 90        cur = build_cvec(cur, il);91        cb(cur, "l_out", il);92 93        // input for next layer94        inpL = cur;95    }96 97    cur = inpL;98 99    // norm100    cur = build_norm(cur,101            model.output_norm,102            model.output_norm_b,103            LLM_NORM, -1);104 105    cb(cur, "result_norm", -1);106    res->t_embd = cur;107 108    cur = build_lora_mm(model.output, cur);109 110    cb(cur, "result_output", -1);111    res->t_logits = cur;112 113    ggml_build_forward_expand(gf, cur);114}115