echodict/llama.cpp

1#include "models.h"2 3llm_build_glm4::llm_build_glm4(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    int sections[4];9    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);10 11    ggml_tensor * cur;12    ggml_tensor * inpL;13 14    inpL = build_inp_embd(model.tok_embd);15 16    bool use_mrope = hparams.use_mrope();17    if (ubatch.embd && !use_mrope) {18        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results19        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");20    }21 22    // inp_pos - contains the positions23    ggml_tensor * inp_pos = build_inp_pos();24 25    auto * inp_attn = build_attn_inp_kv();26 27    ggml_tensor * inp_out_ids = build_inp_out_ids();28 29    // Only process up to last layer (skip final NextN layer)30    // Final layer tensors are loaded but not processed in forward pass31    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;32    for (int il = 0; il < n_transformer_layers; ++il) {33        ggml_tensor * inpSA = inpL;34 35        // Pre-attention norm36        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);37        cb(cur, "attn_norm", il);38 39        // self-attention40        {41            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,42                    n_embd_head, n_head, n_head_kv, il);43 44            if (use_mrope) {45                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,46                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,47                            ext_factor, attn_factor, beta_fast, beta_slow);48 49                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,50                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,51                            ext_factor, attn_factor, beta_fast, beta_slow);52            } else {53                // Normal RoPE54                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,55                                    rope_type, n_ctx_orig, freq_base, freq_scale,56                                    ext_factor, attn_factor, beta_fast, beta_slow);57 58                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,59                                    rope_type, n_ctx_orig, freq_base, freq_scale,60                                    ext_factor, attn_factor, beta_fast, beta_slow);61            }62 63            cb(Qcur, "Qcur", il);64            cb(Kcur, "Kcur", il);65            cb(Vcur, "Vcur", il);66 67            cur = build_attn(inp_attn,68                    model.layers[il].wo, NULL, model.layers[il].wo_s,69                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);70        }71        if (il == n_transformer_layers - 1 && inp_out_ids) {72            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);73            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);74        }75        // Post-attention norm (new!)76        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);77        cb(cur, "post_attn_norm", il);78 79        // Add the input (residual connection after post-attention norm)80        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);81        cb(ffn_inp, "ffn_inp", il);82 83        // FF84        {85            // Pre-MLP norm86            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);87            cb(cur, "ffn_norm", il);88 89            // MLP90            cur = build_ffn(cur,91                    model.layers[il].ffn_up, NULL, NULL,92                    NULL, NULL, NULL,93                    model.layers[il].ffn_down, NULL, NULL,94                    NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);95            cb(cur, "ffn_out", il);96 97            // Post-MLP norm98            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);99            cb(cur, "post_mlp_norm", 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        // input for next layer107        inpL = cur;108    }109    // Final norm110    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);111 112    cb(cur, "result_norm", -1);113    res->t_embd = cur;114 115    // Output projection116    cur = build_lora_mm(model.output, cur);117 118    cb(cur, "result_output", -1);119    res->t_logits = cur;120 121    ggml_build_forward_expand(gf, cur);122}123