echodict/llama.cpp

1#include "models.h"2 3llm_build_openelm::llm_build_openelm(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    inpL = build_inp_embd(model.tok_embd);11 12    // inp_pos - contains the positions13    ggml_tensor * inp_pos = build_inp_pos();14 15    auto * inp_attn = build_attn_inp_kv();16 17    ggml_tensor * inp_out_ids = build_inp_out_ids();18 19    for (int il = 0; il < n_layer; ++il) {20        const int64_t n_head    = hparams.n_head(il);21        const int64_t n_head_kv = hparams.n_head_kv(il);22        const int64_t n_head_qkv = 2*n_head_kv + n_head;23 24        cur = inpL;25        ggml_tensor * residual = cur;26 27        // norm28        cur = build_norm(inpL,29                model.layers[il].attn_norm, NULL,30                LLM_NORM_RMS, il);31        cb(cur, "attn_norm", il);32 33        // self-attention34        {35            cur = build_lora_mm(model.layers[il].wqkv, cur);36            cb(cur, "wqkv", il);37 38            cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens);39 40            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, cur->nb[1], cur->nb[2], 0);41            cb(Qcur, "Qcur", il);42 43            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head);44            cb(Kcur, "Kcur", il);45 46            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv));47            cb(Vcur, "Vcur", il);48 49            Qcur = build_norm(Qcur,50                    model.layers[il].attn_q_norm, NULL,51                    LLM_NORM_RMS, il);52            cb(Qcur, "Qcur", il);53 54            Kcur = build_norm(Kcur,55                    model.layers[il].attn_k_norm, NULL,56                    LLM_NORM_RMS, il);57            cb(Kcur, "Kcur", il);58 59            Qcur = ggml_rope_ext(60                    ctx0, Qcur, inp_pos, NULL,61                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,62                    ext_factor, attn_factor, beta_fast, beta_slow63                    );64 65            Kcur = ggml_rope_ext(66                    ctx0, Kcur, inp_pos, NULL,67                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,68                    ext_factor, attn_factor, beta_fast, beta_slow69                    );70 71            cb(Qcur, "Qcur", il);72            cb(Kcur, "Kcur", il);73            cb(Qcur, "Vcur", il);74 75            cur = build_attn(inp_attn,76                    model.layers[il].wo, NULL, model.layers[il].wo_s,77                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);78        }79        if (il == n_layer - 1 && inp_out_ids) {80            residual = ggml_get_rows(ctx0, residual, inp_out_ids);81            cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);82        }83        ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur);84        cb(ffn_inp, "ffn_inp", il);85 86        // feed-forward network87        {88            cur = build_norm(ffn_inp,89                    model.layers[il].ffn_norm, NULL,90                    LLM_NORM_RMS, il);91            cb(cur, "ffn_norm", il);92 93            cur = build_ffn(cur,94                    model.layers[il].ffn_up,   NULL, NULL,95                    model.layers[il].ffn_gate, NULL, NULL,96                    model.layers[il].ffn_down, NULL, NULL,97                    NULL,98                    LLM_FFN_SILU, LLM_FFN_PAR, il);99            cb(cur, "ffn_out", il);100        }101        cur = ggml_add(ctx0, cur, ffn_inp);102 103        cur = build_cvec(cur, il);104        cb(cur, "l_out", il);105 106        inpL = cur;107    }108    cur = inpL;109 110    // norm111    cur = build_norm(cur,112            model.output_norm, NULL,113            LLM_NORM_RMS, -1);114 115    cb(cur, "result_norm", -1);116    res->t_embd = cur;117 118    cur = build_lora_mm(model.output, cur);119 120    cb(cur, "result_output", -1);121    res->t_logits = cur;122 123    ggml_build_forward_expand(gf, cur);124}125