llama.cpp
https://github.com/ggerganov/llama.cpp

1. struct ggml_cgraph * build_llama()

/home/yongqiang/llm_work/llama_cpp_25_01_05/llama.cpp/src/llama.cpp

    switch (model.arch) {
        case LLM_ARCH_LLAMA:
        case LLM_ARCH_MINICPM:
        case LLM_ARCH_GRANITE:
        case LLM_ARCH_GRANITE_MOE:
            {
                result = llm.build_llama();
            } break;
        case LLM_ARCH_DECI:
            {
                result = llm.build_deci();
            } break;
        case LLM_ARCH_BAICHUAN:
            {
                result = llm.build_baichuan();
            } break;
        case LLM_ARCH_FALCON:
            {
                result = llm.build_falcon();
            } break;
        case LLM_ARCH_GROK:
            {
                result = llm.build_grok();
            } break;
        case LLM_ARCH_STARCODER:
            {
                result = llm.build_starcoder();
            } break;
        case LLM_ARCH_REFACT:
            {
                result = llm.build_refact();
            } break;
        case LLM_ARCH_BERT:
        case LLM_ARCH_JINA_BERT_V2:
        case LLM_ARCH_NOMIC_BERT:
            {
                result = llm.build_bert();
            } break;
        case LLM_ARCH_BLOOM:
            {
                result = llm.build_bloom();
            } break;
        case LLM_ARCH_MPT:
            {
                result = llm.build_mpt();
            } break;
         case LLM_ARCH_STABLELM:
            {
                result = llm.build_stablelm();
            } break;
        case LLM_ARCH_QWEN:
            {
                result = llm.build_qwen();
            } break;
        case LLM_ARCH_QWEN2:
            {
                result = llm.build_qwen2();
            } break;
        case LLM_ARCH_QWEN2VL:
            {
                lctx.n_pos_per_token = 4;
                result = llm.build_qwen2vl();
            } break;
        case LLM_ARCH_QWEN2MOE:
            {
                result = llm.build_qwen2moe();
            } break;
        case LLM_ARCH_PHI2:
            {
                result = llm.build_phi2();
            } break;
        case LLM_ARCH_PHI3:
        case LLM_ARCH_PHIMOE:
            {
                result = llm.build_phi3();
            } break;
        case LLM_ARCH_PLAMO:
            {
                result = llm.build_plamo();
            } break;
        case LLM_ARCH_GPT2:
            {
                result = llm.build_gpt2();
            } break;
        case LLM_ARCH_CODESHELL:
            {
                result = llm.build_codeshell();
            } break;
        case LLM_ARCH_ORION:
            {
                result = llm.build_orion();
            } break;
        case LLM_ARCH_INTERNLM2:
            {
                result = llm.build_internlm2();
            } break;
        case LLM_ARCH_MINICPM3:
            {
                result = llm.build_minicpm3();
            } break;
        case LLM_ARCH_GEMMA:
            {
                result = llm.build_gemma();
            } break;
        case LLM_ARCH_GEMMA2:
            {
                result = llm.build_gemma2();
            } break;
        case LLM_ARCH_STARCODER2:
            {
                result = llm.build_starcoder2();
            } break;
        case LLM_ARCH_MAMBA:
            {
                result = llm.build_mamba();
            } break;
        case LLM_ARCH_XVERSE:
            {
                result = llm.build_xverse();
            } break;
        case LLM_ARCH_COMMAND_R:
            {
                result = llm.build_command_r();
            } break;
        case LLM_ARCH_COHERE2:
            {
                result = llm.build_cohere2();
            } break;
        case LLM_ARCH_DBRX:
            {
                result = llm.build_dbrx();
            } break;
        case LLM_ARCH_OLMO:
            {
                result = llm.build_olmo();
            } break;
        case LLM_ARCH_OLMO2:
            {
                result = llm.build_olmo2();
            } break;
        case LLM_ARCH_OLMOE:
            {
                result = llm.build_olmoe();
            } break;
        case LLM_ARCH_OPENELM:
            {
                result = llm.build_openelm();
            } break;
        case LLM_ARCH_GPTNEOX:
            {
                result = llm.build_gptneox();
            } break;
        case LLM_ARCH_ARCTIC:
            {
                result = llm.build_arctic();
            } break;
        case LLM_ARCH_DEEPSEEK:
            {
                result = llm.build_deepseek();
            } break;
        case LLM_ARCH_DEEPSEEK2:
            {
                result = llm.build_deepseek2();
            } break;
        case LLM_ARCH_CHATGLM:
            {
                result = llm.build_chatglm();
            } break;
        case LLM_ARCH_BITNET:
            {
                result = llm.build_bitnet();
            } break;
        case LLM_ARCH_T5:
            {
                if (lctx.is_encoding) {
                    result = llm.build_t5_enc();
                } else {
                    result = llm.build_t5_dec();
                }
            } break;
        case LLM_ARCH_T5ENCODER:
            {
                result = llm.build_t5_enc();
            } break;
        case LLM_ARCH_JAIS:
            {
                result = llm.build_jais();
            } break;
        case LLM_ARCH_NEMOTRON:
            {
                result = llm.build_nemotron();
            } break;
        case LLM_ARCH_EXAONE:
            {
                result = llm.build_exaone();
            } break;
        case LLM_ARCH_RWKV6:
            {
                result = llm.build_rwkv6();
            } break;
        case LLM_ARCH_RWKV6QWEN2:
            {
                result = llm.build_rwkv6qwen2();
            } break;
        case LLM_ARCH_CHAMELEON:
            {
                result = llm.build_chameleon();
            } break;
        case LLM_ARCH_WAVTOKENIZER_DEC:
            {
                result = llm.build_wavtokenizer_dec();
            } break;
        default:
            GGML_ABORT("fatal error");
    }

    struct ggml_cgraph * build_llama() {
        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);

        // mutable variable, needed during the last layer of the computation to skip unused tokens
        int32_t n_tokens = this->n_tokens;

        const int64_t n_embd_head = hparams.n_embd_head_v;
        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
        GGML_ASSERT(n_embd_head == hparams.n_rot);

        struct ggml_tensor * cur;
        struct ggml_tensor * inpL;

        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);

        // inp_pos - contains the positions
        struct ggml_tensor * inp_pos = build_inp_pos();

        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();

        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
        for (int il = 0; il < n_layer; ++il) {
            struct ggml_tensor * inpSA = inpL;

            // norm
            cur = llm_build_norm(ctx0, inpL, hparams,
                    model.layers[il].attn_norm, NULL,
                    LLM_NORM_RMS, cb, il);
            cb(cur, "attn_norm", il);

            // self-attention
            {
                // rope freq factors for llama3; may return nullptr for llama2 and other models
                struct ggml_tensor * rope_factors = build_rope_factors(il);

                // compute Q and K and RoPE them
                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
                cb(Qcur, "Qcur", il);
                if (model.layers[il].bq) {
                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
                    cb(Qcur, "Qcur", il);
                }

                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
                cb(Kcur, "Kcur", il);
                if (model.layers[il].bk) {
                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
                    cb(Kcur, "Kcur", il);
                }

                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
                cb(Vcur, "Vcur", il);
                if (model.layers[il].bv) {
                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                    cb(Vcur, "Vcur", il);
                }

                Qcur = ggml_rope_ext(
                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow
                );
                cb(Qcur, "Qcur", il);

                Kcur = ggml_rope_ext(
                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow
                );
                cb(Kcur, "Kcur", il);

                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
                        model.layers[il].wo, model.layers[il].bo,
                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
            }

            if (il == n_layer - 1) {
                // skip computing output for unused tokens
                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
                n_tokens = n_outputs;
                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
            }

            // For Granite architecture
            if (hparams.f_residual_scale) {
                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
            }

            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
            cb(ffn_inp, "ffn_inp", il);

            // feed-forward network
            if (model.layers[il].ffn_gate_inp == nullptr) {

                cur = llm_build_norm(ctx0, ffn_inp, hparams,
                        model.layers[il].ffn_norm, NULL,
                        LLM_NORM_RMS, cb, il);
                cb(cur, "ffn_norm", il);

                cur = llm_build_ffn(ctx0, lctx, cur,
                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                        NULL,
                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
                cb(cur, "ffn_out", il);
            } else {
                // MoE branch
                cur = llm_build_norm(ctx0, ffn_inp, hparams,
                        model.layers[il].ffn_norm, NULL,
                        LLM_NORM_RMS, cb, il);
                cb(cur, "ffn_norm", il);

                cur = llm_build_moe_ffn(ctx0, lctx, cur,
                        model.layers[il].ffn_gate_inp,
                        model.layers[il].ffn_up_exps,
                        model.layers[il].ffn_gate_exps,
                        model.layers[il].ffn_down_exps,
                        nullptr,
                        n_expert, n_expert_used,
                        LLM_FFN_SILU, true,
                        false, 0.0,
                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
                        cb, il);
                cb(cur, "ffn_moe_out", il);
            }

            // For Granite architecture
            if (hparams.f_residual_scale) {
                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
            }

            cur = ggml_add(ctx0, cur, ffn_inp);
            cb(cur, "ffn_out", il);

            cur = lctx.cvec.apply_to(ctx0, cur, il);
            cb(cur, "l_out", il);

            // input for next layer
            inpL = cur;
        }

        cur = inpL;

        cur = llm_build_norm(ctx0, cur, hparams,
                model.output_norm, NULL,
                LLM_NORM_RMS, cb, -1);
        cb(cur, "result_norm", -1);

        // lm_head
        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);

        // For Granite architecture
        if (hparams.f_logit_scale) {
            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
        }

        cb(cur, "result_output", -1);

        ggml_build_forward_expand(gf, cur);

        return gf;
    }

References

[1] Yongqiang Cheng, https://yongqiang.blog.csdn.net/
[2] huggingface/gguf, https://github.com/huggingface/huggingface.js/tree/main/packages/gguf