Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_starcoder2::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 30: type = LLM_TYPE_3B; break;8        case 32: type = LLM_TYPE_7B; break;9        case 40: type = LLM_TYPE_15B; break;10        case 52: type = LLM_TYPE_20B; break; // granite11        case 88: type = LLM_TYPE_34B; break; // granite12        default: type = LLM_TYPE_UNKNOWN;13    }14}15 16void llama_model_starcoder2::load_arch_tensors(llama_model_loader &) {17    LLAMA_LOAD_LOCALS;18 19    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);20 21    // output22    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);23    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);24 25    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);26    // if output is NULL, init from the input tok embed27    if (output == NULL) {28        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);29    }30 31    for (int i = 0; i < n_layer; ++i) {32        auto & layer = layers[i];33 34        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);35        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);36 37        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);38        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);39 40        // optional bias tensors41        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        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);47        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);48 49        // optional bias tensors50        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);51        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP ,  "bias", i), {  n_ff}, 0);52    }53}54 55std::unique_ptr<llm_graph_context> llama_model_starcoder2::build_arch_graph(const llm_graph_params & params) const {56    return std::make_unique<graph>(*this, params);57}58 59llama_model_starcoder2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {60    const int64_t n_embd_head = hparams.n_embd_head_v();61 62    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());63    GGML_ASSERT(n_embd_head == n_rot);64 65    ggml_tensor * cur;66    ggml_tensor * inpL;67 68    inpL = build_inp_embd(model.tok_embd);69 70    // inp_pos - contains the positions71    ggml_tensor * inp_pos = build_inp_pos();72 73    auto * inp_attn = build_attn_inp_kv();74 75    ggml_tensor * inp_out_ids = build_inp_out_ids();76 77    for (int il = 0; il < n_layer; ++il) {78        ggml_tensor * inpSA = inpL;79 80        // norm81        cur = build_norm(inpL,82                model.layers[il].attn_norm, model.layers[il].attn_norm_b,83                LLM_NORM, il);84        cb(cur, "attn_norm", il);85 86        // self-attention87        {88            // compute Q and K and RoPE them89            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,90                    n_embd_head, n_head, n_head_kv, il);91 92            Qcur = ggml_rope_ext(93                    ctx0, Qcur, inp_pos, nullptr,94                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,95                    ext_factor, attn_factor, beta_fast, beta_slow96                    );97 98            Kcur = ggml_rope_ext(99                    ctx0, Kcur, inp_pos, nullptr,100                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,101                    ext_factor, attn_factor, beta_fast, beta_slow102                    );103 104            cb(Qcur, "Qcur", il);105            cb(Kcur, "Kcur", il);106            cb(Vcur, "Vcur", 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        if (il == n_layer - 1 && inp_out_ids) {113            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);114            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);115        }116        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);117        cb(ffn_inp, "ffn_inp", il);118 119        // feed-forward network120 121        cur = build_norm(ffn_inp,122                model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,123                LLM_NORM, il);124        cb(cur, "ffn_norm", il);125 126        cur = build_ffn(cur,127                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,128                NULL,                      NULL,                        NULL,129                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,130                NULL,131                LLM_FFN_GELU, LLM_FFN_SEQ, il);132        cb(cur, "ffn_out", il);133 134        cur = ggml_add(ctx0, cur, ffn_inp);135 136        cur = build_cvec(cur, il);137        cb(cur, "l_out", il);138 139        // input for next layer140        inpL = cur;141    }142    cur = inpL;143 144    cur = build_norm(cur,145            model.output_norm, model.output_norm_b,146            LLM_NORM, -1);147 148    cb(cur, "result_norm", -1);149    res->t_embd = cur;150 151    // lm_head152    cur = build_lora_mm(model.output, cur, model.output_s);153 154    cb(cur, "result_output", -1);155    res->t_logits = cur;156 157    ggml_build_forward_expand(gf, cur);158}159