echodict/llama.cpp

1#include "models.h"2 3llm_build_mpt::llm_build_mpt(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 8    ggml_tensor * cur;9    ggml_tensor * pos;10    ggml_tensor * inpL;11 12    inpL = build_inp_embd(model.tok_embd);13 14    auto * inp_attn = build_attn_inp_kv();15 16    if (model.pos_embd) {17        // inp_pos - contains the positions18        ggml_tensor * inp_pos = build_inp_pos();19        pos                   = ggml_get_rows(ctx0, model.pos_embd, inp_pos);20        cb(pos, "pos_embd", -1);21 22        inpL = ggml_add(ctx0, inpL, pos);23        cb(inpL, "inpL", -1);24    }25 26    ggml_tensor * inp_out_ids = build_inp_out_ids();27 28    for (int il = 0; il < n_layer; ++il) {29        ggml_tensor * attn_norm;30 31        attn_norm = build_norm(inpL, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, il);32        cb(attn_norm, "attn_norm", il);33 34        // self-attention35        {36            cur = attn_norm;37 38            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,39                    n_embd_head, n_head, n_head_kv, il);40 41            // Q/K Layernorm42            if (model.layers[il].attn_q_norm) {43                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);44                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);45 46                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);47 48                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);49 50                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);51                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);52            }53 54            cb(Qcur, "Qcur", il);55            cb(Kcur, "Kcur", il);56            cb(Vcur, "Vcur", il);57 58            cur = build_attn(inp_attn,59                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,60                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);61        }62 63        if (il == n_layer - 1 && inp_out_ids) {64            cur  = ggml_get_rows(ctx0, cur, inp_out_ids);65            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);66        }67 68        // Add the input69        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);70        cb(ffn_inp, "ffn_inp", il);71 72        // feed forward73        {74            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, LLM_NORM, il);75            cb(cur, "ffn_norm", il);76            cur = build_ffn(cur,77                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,78                NULL, NULL, NULL,79                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,80                model.layers[il].ffn_act, LLM_FFN_GELU, LLM_FFN_SEQ, il);81            cb(cur, "ffn_out", il);82        }83 84        cur = ggml_add(ctx0, cur, ffn_inp);85 86        cur = build_cvec(cur, il);87        cb(cur, "l_out", il);88 89        // input for next layer90        inpL = cur;91    }92 93    cur = inpL;94 95    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);96 97    cb(cur, "result_norm", -1);98    res->t_embd = cur;99 100    cur = build_lora_mm(model.output, cur);101 102    cb(cur, "result_output", -1);103    res->t_logits = cur;104 105    ggml_build_forward_expand(gf, cur);106}107