echodict/llama.cpp

1#include "models.h"2 3llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :4    llm_build_mamba_base(params) {5    const int64_t n_embd_head = hparams.n_embd_head_v();6 7    ggml_tensor * cur;8    ggml_tensor * inpL;9 10    inpL = build_inp_embd(model.tok_embd);11 12    // inp_pos - contains the positions13    ggml_tensor * inp_pos = build_inp_pos();14 15    // Build the inputs in the recurrent & kv cache16    auto * inp = build_inp_mem_hybrid();17 18    const float kq_scale =19        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;20 21    ggml_tensor * inp_out_ids = build_inp_out_ids();22 23    for (int il = 0; il < n_layer; ++il) {24        ggml_tensor * inpSA = inpL;25 26        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);27        cb(cur, "attn_norm", il);28 29        // self-attention30        auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,31                n_embd_head, n_head, n_head_kv, il);32 33        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,34                             ext_factor, attn_factor, beta_fast, beta_slow);35 36        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale,37                             ext_factor, attn_factor, beta_fast, beta_slow);38 39        cb(Qcur, "Qcur-post-rope", il);40        cb(Kcur, "Kcur-post-rope", il);41        cb(Vcur, "Vcur-post-rope", il);42 43        ggml_tensor * attn_out = build_attn(inp->get_attn(),44                                    model.layers[il].wo, NULL, model.layers[il].wo_s,45                                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);46        cb(attn_out, "attn_out", il);47 48        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);49        // Mamba2 layer50        cb(cur, "ssm_in", il);51 52        ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);53        cb(ssm_out, "ssm_out", il);54 55        // // Aggregation56        cur   = ggml_add(ctx0, attn_out, ssm_out);57        inpSA = ggml_add(ctx0, cur, inpSA);58        cb(cur, "layer_out", il);59 60        if (il == n_layer - 1 && inp_out_ids) {61            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);62            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);63        }64        ggml_tensor * ffn_inp = inpSA;65        cb(ffn_inp, "ffn_inp", il);66 67        // feed-forward network68        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);69        cb(cur, "ffn_norm", il);70 71        cur = build_ffn(cur,72                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,73                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,74                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,75                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);76        cb(cur, "ffn_out", il);77 78        cur = ggml_add(ctx0, cur, inpSA);79 80        cur = build_cvec(cur, il);81        cb(cur, "l_out", il);82 83        // input for next layer84        inpL = cur;85    }86    cur = inpL;87 88    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);89 90    cb(cur, "result_norm", -1);91    res->t_embd = cur;92 93    // lm_head94    cur = build_lora_mm(model.output, cur);95 96    cb(cur, "result_output", -1);97    res->t_logits = cur;98 99    ggml_build_forward_expand(gf, cur);100}101