Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_smallthinker::load_arch_hparams(llama_model_loader & ml) {4    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);5 6    if (found_swa && hparams.n_swa > 0) {7        hparams.swa_type    = LLAMA_SWA_TYPE_STANDARD;8        hparams.n_swa       = 4096;9        load_swa_pattern(ml, 4, true);10 11        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;12        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;13        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);14    } else {15        hparams.swa_type             = LLAMA_SWA_TYPE_NONE;16        hparams.n_no_rope_layer_step = hparams.n_layer();17    }18 19    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);20    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);21    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);22 23    switch (hparams.n_layer()) {24        case 32: type = LLM_TYPE_4B;  break;25        case 52: type = LLM_TYPE_20B; break;26        default: type = LLM_TYPE_UNKNOWN;27    }28}29 30void llama_model_smallthinker::load_arch_tensors(llama_model_loader &) {31    LLAMA_LOAD_LOCALS;32 33    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);34 35    // output36    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);37    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);38 39    // if output is NULL, init from the input tok embed40    if (output == NULL) {41        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);42    }43 44    for (int i = 0; i < n_layer; ++i) {45        auto & layer = layers[i];46 47        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);48 49        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);50        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);51 52        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);53 54        GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for SMALLTHINKER");55        GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for SMALLTHINKER");56 57        // MoE branch58        const int64_t n_ff_exp = hparams.n_ff_exp();59        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, 0);60        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);61        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);62        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0);63    }64}65 66std::unique_ptr<llm_graph_context> llama_model_smallthinker::build_arch_graph(const llm_graph_params & params) const {67    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {68        return std::make_unique<graph<true>> (*this, params);69    } else {70        return std::make_unique<graph<false>>(*this, params);71    }72}73 74template <bool iswa>75llama_model_smallthinker::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){76    const int64_t n_embd_head = hparams.n_embd_head_v();77 78    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());79    GGML_ASSERT(n_embd_head == n_rot);80 81    ggml_tensor * cur;82    ggml_tensor * inpL;83 84    inpL = build_inp_embd(model.tok_embd);85 86    // inp_pos - contains the positions87    ggml_tensor * inp_pos = build_inp_pos();88 89    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;90    inp_attn_type * inp_attn = nullptr;91 92    if constexpr (iswa) {93        inp_attn = build_attn_inp_kv_iswa();94    } else {95        inp_attn = build_attn_inp_kv();96    }97    ggml_tensor * inp_out_ids = build_inp_out_ids();98 99    for (int il = 0; il < n_layer; ++il) {100        const float freq_base_l  = model.get_rope_freq_base (cparams, il);101        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);102 103        ggml_tensor * inpSA  = inpL;104 105        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous106        const bool use_rope = hparams.n_no_rope_layer_step == n_layer ||107                              il % hparams.n_no_rope_layer_step != 0;108 109        ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]110        cb(probs, "ffn_moe_logits", il);111 112        // norm113        cur = build_norm(inpL,model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);114        cb(cur, "attn_norm", il);115 116        // self_attention117        {118            // compute Q and K and RoPE them119            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,120                    n_embd_head, n_head, n_head_kv, il);121 122            if (use_rope) {123                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,124                                    ext_factor, attn_factor, beta_fast, beta_slow);125 126                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,127                                    ext_factor, attn_factor, beta_fast, beta_slow);128            }129            cb(Qcur, "Qcur", il);130            cb(Kcur, "Kcur", il);131 132            cur = build_attn(inp_attn,133                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,134                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);135        }136        if (il == n_layer - 1 && inp_out_ids) {137            cur = ggml_get_rows(ctx0, cur, inp_out_ids);138            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);139            probs = ggml_get_rows(ctx0, probs, inp_out_ids);140        }141        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);142        cb(ffn_inp, "ffn_inp", il);143 144        // MoE branch145        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);146        cb(cur, "ffn_norm", il);147 148        ggml_tensor * ffn_out =149            build_moe_ffn(cur,150                    nullptr,151                    model.layers[il].ffn_up_exps,152                    model.layers[il].ffn_gate_exps,153                    model.layers[il].ffn_down_exps,154                    nullptr,155                    n_expert, n_expert_used,156                    LLM_FFN_RELU, true,157                    hparams.expert_weights_scale,158                    static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),159                    il, probs);160 161        cb(ffn_out, "ffn_out", il);162        cur = ffn_out;163 164        cur = ggml_add(ctx0, cur, ffn_inp);165 166        cur = build_cvec(cur, il);167        cb(cur, "l_out", il);168 169        // input for next layer170        inpL = cur;171    }172    cur = inpL;173 174    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);175    cb(cur, "result_norm", -1);176    res->t_embd = cur;177 178    // lm_head179    cur = build_lora_mm(model.output, cur, model.output_s);180    cb(cur, "result_output", -1);181    res->t_logits = cur;182 183    ggml_build_forward_expand(gf, cur);184}185 186// Explicit template instantiations187template struct llama_model_smallthinker::graph<false>;188template struct llama_model_smallthinker::graph<true>;189