Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_jais2::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);5 6    switch (hparams.n_layer()) {7        case 32: type = LLM_TYPE_8B; break;8        case 68: type = LLM_TYPE_70B; break;9        default: type = LLM_TYPE_UNKNOWN;10    }11}12 13void llama_model_jais2::load_arch_tensors(llama_model_loader &) {14    LLAMA_LOAD_LOCALS;15 16    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);17 18    // output19    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);20    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);21    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);22    if (!output) {23        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);24    }25 26    for (int i = 0; i < n_layer; ++i) {27        auto & layer = layers[i];28 29        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);30        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);31 32        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);33        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);34 35        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);36 37        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);38        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);39 40        // Jais-2 uses simple MLP (no gate) with biases41        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);42        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);43        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);44        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);45    }46}47 48std::unique_ptr<llm_graph_context> llama_model_jais2::build_arch_graph(const llm_graph_params & params) const {49    return std::make_unique<graph>(*this, params);50}51 52// JAIS-2 model graph builder53// Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings54llama_model_jais2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {55    const int64_t n_embd_head = hparams.n_embd_head_v();56 57    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());58    GGML_ASSERT(n_embd_head == n_rot);59 60    ggml_tensor * cur;61    ggml_tensor * inpL;62 63    inpL = build_inp_embd(model.tok_embd);64 65    // inp_pos - contains the positions66    ggml_tensor * inp_pos = build_inp_pos();67 68    // KV input for attention69    auto * inp_attn = build_attn_inp_kv();70 71    ggml_tensor * inp_out_ids = build_inp_out_ids();72 73    for (int il = 0; il < n_layer; ++il) {74        // Pre-attention LayerNorm75        cur = build_norm(inpL,76                model.layers[il].attn_norm,77                model.layers[il].attn_norm_b,78                LLM_NORM, il);79        cb(cur, "attn_norm", il);80 81        // Self-attention with separate Q, K, V projections82        {83            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,84                    n_embd_head, n_head, n_head_kv, il);85 86            // Apply RoPE87            Qcur = ggml_rope_ext(88                ctx0, Qcur, inp_pos, nullptr,89                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,90                ext_factor, attn_factor, beta_fast, beta_slow91            );92 93            Kcur = ggml_rope_ext(94                ctx0, Kcur, inp_pos, nullptr,95                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,96                ext_factor, attn_factor, beta_fast, beta_slow97            );98 99            cb(Qcur, "Qcur_rope", il);100            cb(Kcur, "Kcur_rope", il);101 102            cur = build_attn(inp_attn,103                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,104                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);105        }106 107        if (il == n_layer - 1 && inp_out_ids) {108            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);109            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);110        }111 112        // Residual connection113        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);114        cb(ffn_inp, "ffn_inp", il);115 116        // Pre-FFN LayerNorm117        cur = build_norm(ffn_inp,118                model.layers[il].ffn_norm,119                model.layers[il].ffn_norm_b,120                LLM_NORM, il);121        cb(cur, "ffn_norm", il);122 123        // FFN with relu2 activation (ReLU squared) - no gate projection124        // up -> relu2 -> down125        cur = build_ffn(cur,126                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,127                NULL, NULL, NULL,  // no gate128                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,129                NULL,130                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);131        cb(cur, "ffn_out", il);132 133        // Residual connection134        inpL = ggml_add(ctx0, cur, ffn_inp);135        inpL = build_cvec(inpL, il);136        cb(inpL, "l_out", il);137    }138 139    // Final LayerNorm140    cur = build_norm(inpL,141            model.output_norm,142            model.output_norm_b,143            LLM_NORM, -1);144    cb(cur, "result_norm", -1);145 146    res->t_embd = cur;147 148    // Output projection149    cur = build_lora_mm(model.output, cur, model.output_s);150    cb(cur, "result_output", -1);151 152    res->t_logits = cur;153 154    ggml_build_forward_expand(gf, cur);155}156