echodict/llama.cpp

1#include "models.h"2 3llm_build_qwen3moe::llm_build_qwen3moe(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    ggml_tensor * inp_out_ids = build_inp_out_ids();20 21    for (int il = 0; il < n_layer; ++il) {22        ggml_tensor * inpSA = inpL;23 24        // norm25        cur = build_norm(inpL,26                model.layers[il].attn_norm, NULL,27                LLM_NORM_RMS, il);28        cb(cur, "attn_norm", il);29 30        // self_attention31        {32            // compute Q and K and RoPE them33            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,34                    n_embd_head, n_head, n_head_kv, il);35 36            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);37            cb(Qcur, "Qcur_normed", il);38 39            Qcur = ggml_rope_ext(40                    ctx0, Qcur, inp_pos, nullptr,41                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,42                    ext_factor, attn_factor, beta_fast, beta_slow43                    );44 45            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);46            cb(Kcur, "Kcur_normed", il);47 48            Kcur = ggml_rope_ext(49                    ctx0, Kcur, inp_pos, nullptr,50                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,51                    ext_factor, attn_factor, beta_fast, beta_slow52                    );53 54            cb(Qcur, "Qcur", il);55            cb(Kcur, "Kcur", il);56            cb(Vcur, "Vcur", il);57 58            cur = build_attn(inp_attn,59                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,60                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);61            if (model.layers[il].wo_s) {62                cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);63            }64        }65        if (il == n_layer - 1 && inp_out_ids) {66            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);67            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);68        }69        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);70        cb(ffn_inp, "ffn_inp", il);71 72        // MoE branch73        cur = build_norm(ffn_inp,74                model.layers[il].ffn_norm, NULL,75                LLM_NORM_RMS, il);76        cb(cur, "ffn_norm", il);77 78        ggml_tensor * moe_out =79            build_moe_ffn(cur,80                    model.layers[il].ffn_gate_inp,81                    model.layers[il].ffn_up_exps,82                    model.layers[il].ffn_gate_exps,83                    model.layers[il].ffn_down_exps,84                    nullptr,85                    n_expert, n_expert_used,86                    LLM_FFN_SILU, true,87                    hparams.expert_weights_scale,88                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,89                    il,90                    nullptr, nullptr,91                    model.layers[il].ffn_up_exps_s,92                    model.layers[il].ffn_gate_exps_s,93                    model.layers[il].ffn_down_exps_s);94        cb(moe_out, "ffn_moe_out", il);95        cur = moe_out;96 97        cur = ggml_add(ctx0, cur, ffn_inp);98 99        cur = build_cvec(cur, il);100        cb(cur, "l_out", il);101 102        // input for next layer103        inpL = cur;104    }105    cur = inpL;106 107    cur = build_norm(cur,108            model.output_norm, NULL,109            LLM_NORM_RMS, -1);110 111    cb(cur, "result_norm", -1);112    res->t_embd = cur;113 114    // lm_head115    cur = build_lora_mm(model.output, cur);116 117    cb(cur, "result_output", -1);118    res->t_logits = cur;119 120    ggml_build_forward_expand(gf, cur);121}122