Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_pangu_embed::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5 6    switch (hparams.n_layer()) {7        case 26: type = LLM_TYPE_1B; break; // openPangu-Embedded-1B-V1.18        case 34: type = LLM_TYPE_7B; break; // openPangu-Embedded-7B-V1.19        default: type = LLM_TYPE_UNKNOWN;10    }11}12 13void llama_model_pangu_embed::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      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);21 22    // if output is NULL, init from the input tok embed23    if (output == NULL) {24        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);25    }26 27    for (int i = 0; i < n_layer; ++i) {28        auto & layer = layers[i];29 30        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);31 32        // weight tensors33        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);34        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);35 36        // bias tensors37        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);38 39        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);40 41        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {42            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));43            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));44        } else {45            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));46        }47 48        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);49        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);50        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);51    }52}53 54std::unique_ptr<llm_graph_context> llama_model_pangu_embed::build_arch_graph(const llm_graph_params & params) const {55    return std::make_unique<graph>(*this, params);56}57 58llama_model_pangu_embed::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {59    const int64_t n_embd_head = hparams.n_embd_head_v();60 61    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());62    GGML_ASSERT(n_embd_head == n_rot);63 64    ggml_tensor * cur;65    ggml_tensor * inpL;66 67    inpL = build_inp_embd(model.tok_embd);68 69    // inp_pos - contains the positions70    ggml_tensor * inp_pos = build_inp_pos();71 72    auto * inp_attn = build_attn_inp_kv();73 74    ggml_tensor * inp_out_ids = build_inp_out_ids();75 76    for (int il = 0; il < n_layer; ++il) {77        ggml_tensor * inpSA = inpL;78 79        // norm80        cur = build_norm(inpL,81                model.layers[il].attn_norm, NULL,82                LLM_NORM_RMS, il);83        cb(cur, "attn_norm", il);84 85        // self attention86        {87            // compute Q and K and RoPE them88            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,89                    n_embd_head, n_head, n_head_kv, il);90 91            Qcur = ggml_rope_ext(92                    ctx0, Qcur, inp_pos, nullptr,93                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,94                    ext_factor, attn_factor, beta_fast, beta_slow95                    );96 97            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,98                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,99                    ext_factor, attn_factor, beta_fast, beta_slow100                    );101 102            cb(Qcur, "Qcur", il);103            cb(Kcur, "Kcur", il);104            cb(Vcur, "Vcur", il);105 106            cur = build_attn(inp_attn,107                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,108                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);109        }110 111        if (il == n_layer - 1 && inp_out_ids) {112            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);113            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);114        }115 116        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);117        cb(ffn_inp, "ffn_inp", il);118 119        // feed-forward network120        cur = build_norm(ffn_inp,121                model.layers[il].ffn_norm, NULL,122                LLM_NORM_RMS, il);123        cb(cur, "ffn_norm", il);124 125        cur = build_ffn(cur,126                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,127                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,128                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,129                NULL,130                LLM_FFN_SILU, LLM_FFN_PAR, il);131 132        cur = ggml_add(ctx0, cur, ffn_inp);133        cb(cur, "ffn_out", il);134 135        cur = build_cvec(cur, il);136        cb(cur, "l_out", il);137 138        // input for next layer139        inpL = cur;140    }141 142    cur = inpL;143 144    cur = build_norm(cur,145            model.output_norm, NULL,146            LLM_NORM_RMS, -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    if (model.output_b != nullptr) {155        cur = ggml_add(ctx0, cur, model.output_b);156    }157 158    cb(cur, "result_output", -1);159    res->t_logits = cur;160 161    ggml_build_forward_expand(gf, cur);162}163