Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_plamo3::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);6    if (found_swa && hparams.n_swa > 0) {7        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;8        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);9        load_swa_pattern(ml, 8);10    } else {11        hparams.swa_type = LLAMA_SWA_TYPE_NONE;12    }13 14    switch (hparams.n_layer()) {15        case 24: type = LLM_TYPE_2B; break;16        default: type = LLM_TYPE_UNKNOWN;17    }18}19 20void llama_model_plamo3::load_arch_tensors(llama_model_loader &) {21    LLAMA_LOAD_LOCALS;22 23    const int64_t head_dim_q = hparams.n_embd_head_k();24    const int64_t head_dim_v = hparams.n_embd_head_v();25 26    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);27 28    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);29    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);30    if (output == NULL) {31        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);32    }33 34    for (int i = 0; i < n_layer; ++i) {35        auto & layer = layers[i];36 37        const int64_t num_attention_heads = hparams.n_head(i);38        const int64_t num_key_value_heads = hparams.n_head_kv(i);39        const int64_t q_proj_dim = num_attention_heads * head_dim_q;40        const int64_t k_proj_dim = num_key_value_heads * head_dim_q;41        const int64_t v_proj_dim = num_key_value_heads * head_dim_v;42        const int64_t n_ff_cur   = hparams.n_ff(i);43 44        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);45        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i),46                {n_embd,q_proj_dim + k_proj_dim + v_proj_dim}, 0);47        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {head_dim_q}, 0);48        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {head_dim_q}, 0);49        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {num_attention_heads * head_dim_v, n_embd}, 0);50        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, i), {n_embd}, 0);51 52        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);53        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, i), {n_embd}, 0);54 55        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff_cur * 2}, 0);56        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff_cur, n_embd}, 0);57    }58}59 60std::unique_ptr<llm_graph_context> llama_model_plamo3::build_arch_graph(const llm_graph_params & params) const {61    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {62        return std::make_unique<graph<true>> (*this, params);63    } else {64        return std::make_unique<graph<false>>(*this, params);65    }66}67 68template <bool iswa>69llama_model_plamo3::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) :70    llm_graph_context(params) {71    const int64_t head_dim_q = hparams.n_embd_head_k();72    const int64_t head_dim_v = hparams.n_embd_head_v();73 74    ggml_tensor * cur;75    ggml_tensor * inpL = build_inp_embd(model.tok_embd);76    ggml_tensor * inp_pos = build_inp_pos();77 78    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;79    inp_attn_type * inp_attn = nullptr;80 81    if constexpr (iswa) {82        inp_attn = build_attn_inp_kv_iswa();83    } else {84        inp_attn = build_attn_inp_kv();85    }86 87    ggml_tensor * inp_out_ids = build_inp_out_ids();88 89    for (int il = 0; il < n_layer; ++il) {90        ggml_tensor * residual = inpL;91 92        float freq_base_l  = 0.0f;93        float freq_scale_l = 0.0f;94        if constexpr (iswa) {95            freq_base_l  = model.get_rope_freq_base (cparams, il);96            freq_scale_l = model.get_rope_freq_scale(cparams, il);97        } else {98            freq_base_l  = freq_base;99            freq_scale_l = freq_scale;100        }101 102        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);103        cb(cur, "attn_norm", il);104 105        ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);106        cb(cur, "wqkv", il);107 108        const int32_t n_head    = hparams.n_head(il);109        const int32_t n_head_kv = hparams.n_head_kv(il);110 111        const int64_t q_offset = 0;112        const int64_t k_offset = head_dim_q * n_head;113        const int64_t v_offset = k_offset + head_dim_q * n_head_kv;114 115        ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head, n_tokens,116                head_dim_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv));117        ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head_kv, n_tokens,118                head_dim_q * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv));119        ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, head_dim_v, n_head_kv, n_tokens,120                head_dim_v * sizeof(float), qkv->nb[1], v_offset * ggml_element_size(qkv));121 122        cb(Qcur, "Qcur", il);123        cb(Kcur, "Kcur", il);124        cb(Vcur, "Vcur", il);125 126        Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);127        cb(Qcur, "attn_q_norm", il);128        Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);129        cb(Kcur, "attn_k_norm", il);130 131        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,132                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,133                ext_factor, attn_factor, beta_fast, beta_slow);134        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,135                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,136                ext_factor, attn_factor, beta_fast, beta_slow);137 138        const float attn_scale = 1.0f / sqrtf(float(head_dim_q));139 140        cur = build_attn(inp_attn,141                model.layers[il].wo, NULL, model.layers[il].wo_s,142                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, attn_scale, il);143        cb(cur, "attn_out", il);144 145        if (il == n_layer - 1 && inp_out_ids) {146            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);147            residual = ggml_get_rows(ctx0, residual, inp_out_ids);148        }149 150        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);151        cb(cur, "attn_post_norm", il);152 153        cur = ggml_add(ctx0, cur, residual);154        cb(cur, "attn_residual", il);155 156        residual = cur;157 158        cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);159        cb(cur, "ffn_norm", il);160 161        cur = build_ffn(cur,162                model.layers[il].ffn_up,   NULL, NULL,163                NULL,                      NULL, NULL,164                model.layers[il].ffn_down, NULL, NULL,165                NULL,166                LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);167        cb(cur, "ffn_out", il);168 169        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);170        cb(cur, "ffn_post_norm", il);171 172        cur = ggml_add(ctx0, cur, residual);173        cb(cur, "ffn_residual", il);174 175        cur = build_cvec(cur, il);176        cb(cur, "l_out", il);177 178        // input for next layer179        inpL = cur;180    }181 182    cur = inpL;183 184    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);185    res->t_embd = cur;186 187    cur = build_lora_mm(model.output, cur, model.output_s);188    res->t_logits = cur;189 190    ggml_build_forward_expand(gf, cur);191}192 193// Explicit template instantiations194template struct llama_model_plamo3::graph<false>;195template struct llama_model_plamo3::graph<true>;196