Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_llama::load_arch_hparams(llama_model_loader & ml) {4    uint32_t n_vocab = 0;5    ml.get_key(LLM_KV_VOCAB_SIZE, n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, n_vocab, false);6 7    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);8 9    if (hparams.n_expert == 8) {10        switch (hparams.n_layer()) {11            case 32: type = LLM_TYPE_8x7B; break;12            case 56: type = LLM_TYPE_8x22B; break;13            default: type = LLM_TYPE_UNKNOWN;14        }15    } else {16        switch (hparams.n_layer()) {17            case 16: type = LLM_TYPE_1B; break; // Llama 3.2 1B18            case 22: type = LLM_TYPE_1B; break;19            case 26: type = LLM_TYPE_3B; break;20            case 28: type = LLM_TYPE_3B; break; // Llama 3.2 3B21            case 30: type = LLM_TYPE_256M; break; // smoldocling 256M22            // granite uses a vocab with len 4915223            case 32: type = n_vocab == 49152 ? LLM_TYPE_3B : (n_vocab < 40000 ? LLM_TYPE_7B : LLM_TYPE_8B); break;24            case 36: type = LLM_TYPE_8B; break; // granite25            case 40: type = LLM_TYPE_13B; break;26            case 48: type = LLM_TYPE_34B; break;27            case 60: type = LLM_TYPE_30B; break;28            case 80: type = hparams.n_head() == hparams.n_head_kv() ? LLM_TYPE_65B : LLM_TYPE_70B; break;29            default: type = LLM_TYPE_UNKNOWN;30        }31    }32}33 34void llama_model_llama::load_arch_tensors(llama_model_loader &) {35    LLAMA_LOAD_LOCALS;36 37    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);38 39    // output40    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);41    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);42 43    // if output is NULL, init from the input tok embed44    if (output == NULL) {45        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);46    }47 48    for (int i = 0; i < n_layer; ++i) {49        auto & layer = layers[i];50 51        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);52 53        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);54        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);55 56        // optional bias tensors57        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);58 59        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);60 61        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {62            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));63            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));64        }65        else {66            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));67        }68 69        if (n_expert == 0) {70            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);71            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);72            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);73 74            // optional MLP bias75            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);76            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);77            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);78        } else {79            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);80            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);81            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);82            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);83 84            // For Granite MoE Shared85            if (hparams.n_ff_shexp > 0) {86                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);87                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);88                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);89            }90        }91    }92}93 94std::unique_ptr<llm_graph_context> llama_model_llama::build_arch_graph(const llm_graph_params & params) const {95    return std::make_unique<graph<false>>(*this, params);96}97 98template <bool embed>99llama_model_llama::graph<embed>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {100    const int64_t n_embd_head = hparams.n_embd_head_v();101 102    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());103    GGML_ASSERT(n_embd_head == n_rot);104 105    ggml_tensor * cur;106    ggml_tensor * inpL;107 108    inpL = build_inp_embd(model.tok_embd);109 110    // inp_pos - contains the positions111    ggml_tensor * inp_pos = build_inp_pos();112 113    using inp_attn_type = std::conditional_t<embed, llm_graph_input_attn_no_cache, llm_graph_input_attn_kv>;114 115    inp_attn_type * inp_attn = nullptr;116    if constexpr (embed) {117        inp_attn = build_attn_inp_no_cache();118    } else {119        inp_attn = build_attn_inp_kv();120    }121 122    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;123 124    ggml_tensor * inp_out_ids = build_inp_out_ids();125 126    for (int il = 0; il < n_layer; ++il) {127        res->t_layer_inp[il] = inpL;128 129        ggml_tensor * inpSA = inpL;130 131        // norm132        cur = build_norm(inpL,133                model.layers[il].attn_norm, NULL,134                LLM_NORM_RMS, il);135        cb(cur, "attn_norm", il);136 137        // self-attention138        {139            // rope freq factors for llama3; may return nullptr for llama2 and other models140            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);141 142            // compute Q and K and RoPE them143            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,144                    n_embd_head, n_head, n_head_kv, il);145 146            Qcur = ggml_rope_ext(147                    ctx0, Qcur, inp_pos, rope_factors,148                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,149                    ext_factor, attn_factor, beta_fast, beta_slow150                    );151 152            Kcur = ggml_rope_ext(153                    ctx0, Kcur, inp_pos, rope_factors,154                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,155                    ext_factor, attn_factor, beta_fast, beta_slow156                    );157 158            cb(Qcur, "Qcur", il);159            cb(Kcur, "Kcur", il);160            cb(Vcur, "Vcur", il);161 162            if (hparams.use_kq_norm) {163                // Llama4TextL2Norm164                Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);165                Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);166                cb(Qcur, "Qcur_normed", il);167                cb(Kcur, "Kcur_normed", il);168            }169            cur = build_attn(inp_attn,170                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,171                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);172            cb(cur, "attn_out", il);173        }174        if (il == n_layer - 1 && inp_out_ids) {175            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);176            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);177        }178        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);179        cb(ffn_inp, "ffn_inp", il);180 181        // feed-forward network (non-MoE)182        if (model.layers[il].ffn_gate_inp == nullptr) {183 184            cur = build_norm(ffn_inp,185                    model.layers[il].ffn_norm, NULL,186                    LLM_NORM_RMS, il);187            cb(cur, "ffn_norm", il);188 189            cur = build_ffn(cur,190                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,191                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,192                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,193                    NULL,194                    LLM_FFN_SILU, LLM_FFN_PAR, il);195            cb(cur, "ffn_out", il);196        } else {197            // MoE branch198            cur = build_norm(ffn_inp,199                    model.layers[il].ffn_norm, NULL,200                    LLM_NORM_RMS, il);201            cb(cur, "ffn_norm", il);202 203            cur = build_moe_ffn(cur,204                    model.layers[il].ffn_gate_inp,205                    model.layers[il].ffn_up_exps,206                    model.layers[il].ffn_gate_exps,207                    model.layers[il].ffn_down_exps,208                    nullptr,209                    n_expert, n_expert_used,210                    LLM_FFN_SILU, true,211                    hparams.expert_weights_scale,212                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,213                    il,214                    nullptr, nullptr,215                    model.layers[il].ffn_up_exps_s,216                    model.layers[il].ffn_gate_exps_s,217                    model.layers[il].ffn_down_exps_s);218            cb(cur, "ffn_moe_out", il);219        }220        cur = ggml_add(ctx0, cur, ffn_inp);221        cb(cur, "ffn_out", il);222 223        cur = build_cvec(cur, il);224        cb(cur, "l_out", il);225 226        // input for next layer227        inpL = cur;228    }229    cur = inpL;230 231    cur = build_norm(cur,232            model.output_norm, NULL,233            LLM_NORM_RMS, -1);234 235    cb(cur, "result_norm", -1);236    res->t_embd = cur;237 238    if constexpr (!embed) {239        // lm_head240        cur = build_lora_mm(model.output, cur, model.output_s);241 242        cb(cur, "result_output", -1);243        res->t_logits = cur;244    }245 246    ggml_build_forward_expand(gf, cur);247}248 249template struct llama_model_llama::graph<false>;250template struct llama_model_llama::graph<true>;251