Brunobkr/llama.cpp_AlgMor24_github

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        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);33        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);34        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);35        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);36 37        // attention biases - all have shape n_embd (output dimension of projections)38        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0);39        layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd}, 0);40        layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd}, 0);41        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);42 43        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);44        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);45 46        // Jais-2 uses simple MLP (no gate) with biases47        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);48        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);49        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);50        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);51    }52}53 54std::unique_ptr<llm_graph_context> llama_model_jais2::build_arch_graph(const llm_graph_params & params) const {55    return std::make_unique<graph>(*this, params);56}57 58// JAIS-2 model graph builder59// Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings60llama_model_jais2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {61    const int64_t n_embd_head = hparams.n_embd_head_v();62 63    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());64    GGML_ASSERT(n_embd_head == n_rot);65 66    ggml_tensor * cur;67    ggml_tensor * inpL;68 69    inpL = build_inp_embd(model.tok_embd);70 71    // inp_pos - contains the positions72    ggml_tensor * inp_pos = build_inp_pos();73 74    // KV input for attention75    auto * inp_attn = build_attn_inp_kv();76 77    ggml_tensor * inp_out_ids = build_inp_out_ids();78 79    for (int il = 0; il < n_layer; ++il) {80        // Pre-attention LayerNorm81        cur = build_norm(inpL,82                model.layers[il].attn_norm,83                model.layers[il].attn_norm_b,84                LLM_NORM, il);85        cb(cur, "attn_norm", il);86 87        // Self-attention with separate Q, K, V projections88        {89            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,90                    n_embd_head, n_head, n_head_kv, il);91 92            // Apply RoPE93            Qcur = ggml_rope_ext(94                ctx0, Qcur, 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            Kcur = ggml_rope_ext(100                ctx0, Kcur, inp_pos, nullptr,101                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,102                ext_factor, attn_factor, beta_fast, beta_slow103            );104 105            cb(Qcur, "Qcur_rope", il);106            cb(Kcur, "Kcur_rope", il);107 108            cur = build_attn(inp_attn,109                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,110                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);111        }112 113        if (il == n_layer - 1 && inp_out_ids) {114            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);115            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);116        }117 118        // Residual connection119        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);120        cb(ffn_inp, "ffn_inp", il);121 122        // Pre-FFN LayerNorm123        cur = build_norm(ffn_inp,124                model.layers[il].ffn_norm,125                model.layers[il].ffn_norm_b,126                LLM_NORM, il);127        cb(cur, "ffn_norm", il);128 129        // FFN with relu2 activation (ReLU squared) - no gate projection130        // up -> relu2 -> down131        cur = build_ffn(cur,132                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,133                NULL, NULL, NULL,  // no gate134                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,135                NULL,136                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);137        cb(cur, "ffn_out", il);138 139        // Residual connection140        inpL = ggml_add(ctx0, cur, ffn_inp);141        inpL = build_cvec(inpL, il);142        cb(inpL, "l_out", il);143    }144 145    // Final LayerNorm146    cur = build_norm(inpL,147            model.output_norm,148            model.output_norm_b,149            LLM_NORM, -1);150    cb(cur, "result_norm", -1);151 152    res->t_embd = cur;153 154    // Output projection155    cur = build_lora_mm(model.output, cur, model.output_s);156    cb(cur, "result_output", -1);157 158    res->t_logits = cur;159 160    ggml_build_forward_expand(gf, cur);161}162