Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_spark2_5::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);6 7    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;8    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);9 10    hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;11    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;12    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);13 14    switch (hparams.n_layer()) {15        case 28: type = LLM_TYPE_1_7B; break;16        default: type = LLM_TYPE_UNKNOWN;17    }18}19 20void llama_model_spark2_5::load_arch_tensors(llama_model_loader &) {21    LLAMA_LOAD_LOCALS;22 23    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);24 25    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);26    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);27    if (output == nullptr) {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        const int64_t n_head_i = hparams.n_head(i);35        const int64_t n_head_kv_i = hparams.n_head_kv(i);36        const int64_t n_embd_q = hparams.n_embd_head_k(i) * n_head_i;37        const int64_t n_embd_k = hparams.n_embd_head_k(i) * n_head_kv_i;38        const int64_t n_embd_v = hparams.n_embd_head_v(i) * n_head_kv_i;39 40        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);41        create_tensor_qkv(layer, i, n_embd, n_embd_q, n_embd_k, n_embd_v, 0);42        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_head_i}, 0);43        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_q, n_embd}, 0);44 45        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);46        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);47        layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);48        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);49    }50}51 52std::unique_ptr<llm_graph_context> llama_model_spark2_5::build_arch_graph(const llm_graph_params & params) const {53    return std::make_unique<graph>(*this, params);54}55 56llama_model_spark2_5::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {57    const int64_t n_embd_head = hparams.n_embd_head_v();58 59    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());60    GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_STANDARD);61 62    ggml_tensor * inpL = build_inp_embd(model.tok_embd);63    ggml_tensor * inp_pos = build_inp_pos();64    auto * inp_attn = build_attn_inp_kv_iswa();65    ggml_tensor * inp_out_ids = build_inp_out_ids();66 67    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));68 69    for (int il = 0; il < n_layer; ++il) {70        ggml_tensor * inpSA = inpL;71        ggml_tensor * cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);72        cb(cur, "attn_norm", il);73 74        const int64_t n_head_i = hparams.n_head(il);75        const int64_t n_head_kv_i = hparams.n_head_kv(il);76        const int64_t n_rot_i = hparams.n_rot(il);77        const float freq_base_i = model.get_rope_freq_base(cparams, il);78        const float freq_scale_i = model.get_rope_freq_scale(cparams, il);79 80        ggml_tensor * attn_inp = cur;81        auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head_i, n_head_kv_i, il);82 83        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,84                n_rot_i, rope_type, n_ctx_orig, freq_base_i, freq_scale_i,85                ext_factor, attn_factor, beta_fast, beta_slow);86        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,87                n_rot_i, rope_type, n_ctx_orig, freq_base_i, freq_scale_i,88                ext_factor, attn_factor, beta_fast, beta_slow);89        cb(Qcur, "Qcur_rope", il);90        cb(Kcur, "Kcur_rope", il);91 92        cur = build_attn(inp_attn,93                nullptr, nullptr, nullptr,94                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);95        cb(cur, "attn_out", il);96 97        ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);98        gate = ggml_sigmoid(ctx0, gate);99        cb(gate, "attn_gate", il);100 101        const int64_t n_tokens_i = cur->ne[1];102        cur = ggml_reshape_3d(ctx0, cur, n_embd_head, n_head_i, n_tokens_i);103        gate = ggml_reshape_3d(ctx0, gate, 1, n_head_i, n_tokens_i);104        cur = ggml_mul(ctx0, cur, gate);105        cur = ggml_reshape_2d(ctx0, cur, n_embd_head * n_head_i, n_tokens_i);106        cb(cur, "attn_gated", il);107 108        cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);109        cb(cur, "attn_out_proj", il);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        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);120        cb(cur, "ffn_norm", il);121 122        cur = build_ffn(cur,123                model.layers[il].ffn_up, nullptr, nullptr,124                model.layers[il].ffn_gate, nullptr, nullptr,125                model.layers[il].ffn_down, nullptr, nullptr,126                nullptr,127                LLM_FFN_GELU, LLM_FFN_PAR, il);128        cb(cur, "ffn_out", il);129 130        cur = ggml_add(ctx0, cur, ffn_inp);131        cur = build_cvec(cur, il);132        cb(cur, "l_out", il);133 134        inpL = cur;135    }136 137    ggml_tensor * cur = build_norm(inpL, model.output_norm, nullptr, LLM_NORM_RMS, -1);138    cb(cur, "result_norm", -1);139    res->t_embd = cur;140 141    cur = build_lora_mm(model.output, cur);142    cb(cur, "result_output", -1);143    res->t_logits = cur;144 145    ggml_build_forward_expand(gf, cur);146}147