echodict/llama.cpp

1#include "models.h"2 3llm_build_glm4_moe::llm_build_glm4_moe(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            // Apply Q/K norm if available (GLM-4.5 355B variant)45            if (model.layers[il].attn_q_norm) {46                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);47                cb(Qcur, "Qcur_normed", il);48            }49            if (model.layers[il].attn_k_norm) {50                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);51                cb(Kcur, "Kcur_normed", il);52            }53 54            if (use_mrope) {55                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,56                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,57                            ext_factor, attn_factor, beta_fast, beta_slow);58 59                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,60                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,61                            ext_factor, attn_factor, beta_fast, beta_slow);62            } else {63                // Normal RoPE64                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,65                                    rope_type, n_ctx_orig, freq_base, freq_scale,66                                    ext_factor, attn_factor, beta_fast, beta_slow);67 68                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,69                                    rope_type, n_ctx_orig, freq_base, freq_scale,70                                    ext_factor, attn_factor, beta_fast, beta_slow);71            }72 73            cb(Qcur, "Qcur", il);74            cb(Kcur, "Kcur", il);75            cb(Vcur, "Vcur", il);76 77            cur = build_attn(inp_attn,78                    model.layers[il].wo, NULL, model.layers[il].wo_s,79                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);80        }81        if (il == n_transformer_layers - 1 && inp_out_ids) {82            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);83            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);84        }85        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);86        cb(ffn_inp, "ffn_inp", il);87 88        // Post-attention norm89        cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);90        cb(cur, "post_attn_norm", il);91 92        // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)93        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {94            // Dense FFN layer95            cur = build_ffn(cur,96                    model.layers[il].ffn_up,   NULL, NULL,97                    model.layers[il].ffn_gate, NULL, NULL,98                    model.layers[il].ffn_down, NULL, NULL,99                    NULL,100                    LLM_FFN_SILU, LLM_FFN_PAR, il);101            cb(cur, "ffn_out", il);102        } else {103            // Process routed experts using existing MoE infrastructure104            ggml_tensor * routed_out = build_moe_ffn(cur,105                    model.layers[il].ffn_gate_inp,106                    model.layers[il].ffn_up_exps,107                    model.layers[il].ffn_gate_exps,108                    model.layers[il].ffn_down_exps,109                    model.layers[il].ffn_exp_probs_b,110                    n_expert, n_expert_used,111                    LLM_FFN_SILU, hparams.expert_weights_norm,112                    hparams.expert_weights_scale,113                    (llama_expert_gating_func_type) hparams.expert_gating_func,114                    il);115            cb(routed_out, "ffn_moe_out", il);116 117            // Process shared expert on original input118            ggml_tensor * shared_out = build_ffn(cur,119                    model.layers[il].ffn_up_shexp,   NULL, NULL,120                    model.layers[il].ffn_gate_shexp, NULL, NULL,121                    model.layers[il].ffn_down_shexp, NULL, NULL,122                    NULL,123                    LLM_FFN_SILU, LLM_FFN_PAR, il);124            cb(shared_out, "ffn_shexp_out", il);125 126            // Final output: routed_output + shared_output127            cur = ggml_add(ctx0, routed_out, shared_out);128            cb(cur, "ffn_out", il);129        }130        cur = ggml_add(ctx0, cur, ffn_inp);131 132        cur = build_cvec(cur, il);133        cb(cur, "l_out", il);134 135        // input for next layer136        inpL = cur;137    }138    cur = inpL;139    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);140 141    cb(cur, "result_norm", -1);142    res->t_embd = cur;143 144    // lm_head145    cur = build_lora_mm(model.output, cur);146 147    cb(cur, "result_output", -1);148    res->t_logits = cur;149 150    ggml_build_forward_expand(gf, cur);151}152