echodict/llama.cpp

1#include "models.h"2 3llm_build_bloom::llm_build_bloom(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 * inpL;10 11    inpL = build_inp_embd(model.tok_embd);12 13    auto * inp_attn = build_attn_inp_kv();14 15    inpL = build_norm(inpL,16            model.tok_norm,17            model.tok_norm_b,18            LLM_NORM, 0);19    cb(inpL, "inp_norm", 0);20 21    ggml_tensor * inp_out_ids = build_inp_out_ids();22 23    for (int il = 0; il < n_layer; ++il) {24        cur = build_norm(inpL,25                model.layers[il].attn_norm,26                model.layers[il].attn_norm_b,27                LLM_NORM, il);28        cb(cur, "attn_norm", il);29 30        // self-attention31        {32            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,33                    n_embd_head, n_head, n_head_kv, il);34 35            cur = build_attn(inp_attn,36                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,37                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);38        }39 40        if (il == n_layer - 1 && inp_out_ids) {41            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);42            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);43        }44 45        // Add the input46        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);47        cb(ffn_inp, "ffn_inp", il);48 49        // FF50        {51            cur = build_norm(ffn_inp,52                    model.layers[il].ffn_norm,53                    model.layers[il].ffn_norm_b,54                    LLM_NORM, il);55            cb(cur, "ffn_norm", il);56 57            cur = build_ffn(cur,58                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,59                    NULL,                      NULL,                        NULL,60                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,61                    NULL,62                    LLM_FFN_GELU, LLM_FFN_SEQ, il);63            cb(cur, "ffn_out", il);64        }65 66        cur = ggml_add(ctx0, cur, ffn_inp);67 68        cur = build_cvec(cur, il);69        cb(cur, "l_out", il);70 71        // input for next layer72        inpL = cur;73    }74 75    cur = build_norm(inpL,76            model.output_norm,77            model.output_norm_b,78            LLM_NORM, -1);79 80    cb(cur, "result_norm", -1);81    res->t_embd = cur;82 83    cur = build_lora_mm(model.output, cur);84 85    cb(cur, "result_output", -1);86    res->t_logits = cur;87 88    ggml_build_forward_expand(gf, cur);89}90