echodict/llama.cpp

1#include "models.h"2 3llm_build_qwen2vl::llm_build_qwen2vl(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    int sections[4];20    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);21 22    ggml_tensor * inp_out_ids = build_inp_out_ids();23 24    for (int il = 0; il < n_layer; ++il) {25        ggml_tensor * inpSA = inpL;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            // compute Q and K and RoPE them36            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,37                    n_embd_head, n_head, n_head_kv, il);38 39            Qcur = ggml_rope_multi(40                    ctx0, Qcur, inp_pos, nullptr,41                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,42                    ext_factor, attn_factor, beta_fast, beta_slow43                    );44 45            Kcur = ggml_rope_multi(46                    ctx0, Kcur, inp_pos, nullptr,47                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,48                    ext_factor, attn_factor, beta_fast, beta_slow49                    );50 51            cb(Qcur, "Qcur", il);52            cb(Kcur, "Kcur", il);53            cb(Vcur, "Vcur", il);54 55            cur = build_attn(inp_attn,56                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,57                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);58        }59        if (il == n_layer - 1 && inp_out_ids) {60            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);61            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);62        }63        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);64        cb(ffn_inp, "ffn_inp", il);65 66        // feed-forward network67        cur = build_norm(ffn_inp,68                model.layers[il].ffn_norm, NULL,69                LLM_NORM_RMS, il);70        cb(cur, "ffn_norm", il);71 72        cur = build_ffn(cur,73                model.layers[il].ffn_up,   NULL, NULL,74                model.layers[il].ffn_gate, NULL, NULL,75                model.layers[il].ffn_down, NULL, NULL,76                NULL,77                LLM_FFN_SILU, LLM_FFN_PAR, il);78        cb(cur, "ffn_out", il);79 80        cur = ggml_add(ctx0, cur, ffn_inp);81 82        cur = build_cvec(cur, il);83        cb(cur, "l_out", il);84 85        // input for next layer86        inpL = cur;87    }88    cur = inpL;89 90    cur = build_norm(cur,91            model.output_norm, NULL,92            LLM_NORM_RMS, -1);93 94    cb(cur, "result_norm", -1);95    res->t_embd = cur;96 97    // lm_head98    cur = build_lora_mm(model.output, cur);99 100    cb(cur, "result_output", -1);101    res->t_logits = cur;102 103    ggml_build_forward_expand(gf, cur);104}105