char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Georgi Gerganovhparams : refactor `hparams.n_layer` (#24060)7acb4e8cd

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
3.6 KiB66 linesraw
1#include "models.h"
2
3void llama_model_jina_bert_v2::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
5    hparams.f_max_alibi_bias = 8.0f;
6
7    switch (hparams.n_layer()) {
8        case 4:  type = LLM_TYPE_33M;  break; // jina-embeddings-small
9        case 12: type = LLM_TYPE_137M; break; // jina-embeddings-base
10        default: type = LLM_TYPE_UNKNOWN;
11    }
12}
13
14void llama_model_jina_bert_v2::load_arch_tensors(llama_model_loader & ml) {
15    LLAMA_LOAD_LOCALS;
16
17    tok_embd  = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0); // word_embeddings
18    type_embd = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0); // token_type_embeddings
19
20    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0); // LayerNorm
21    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0); // LayerNorm bias
22
23    cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, 1}, TENSOR_NOT_REQUIRED);
24    cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {1},         TENSOR_NOT_REQUIRED);
25    for (int i = 0; i < n_layer; ++i) {
26        auto & layer = layers[i]; // JinaBertLayer
27
28        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
29
30        layer.attn_q_norm   = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
31        layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
32
33        layer.attn_k_norm   = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
34        layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
35
36        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); //output_dens
37        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0); //output_dens
38
39        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0); //output_norm
40        layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias",   i), {n_embd}, 0);
41
42        layer.attn_norm_2   = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
43        layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
44
45        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
46
47        const auto tn_ffn_up_weight = tn(LLM_TENSOR_FFN_UP, "weight", i);
48        ggml_tensor * t_ffn_up = ml.get_tensor_meta(tn_ffn_up_weight.str().c_str());
49        const int64_t n_ffn_up = t_ffn_up ? t_ffn_up->ne[1] : n_ff;
50
51        GGML_ASSERT(n_ffn_up == n_ff || n_ffn_up == n_ff * 2);
52        layer.ffn_up   = create_tensor(tn_ffn_up_weight, {n_embd, n_ffn_up}, 0);
53        layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ffn_up}, TENSOR_NOT_REQUIRED);
54
55        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
56        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, 0);
57
58        layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
59        layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias",   i), {n_embd}, 0);
60    }
61}
62
63std::unique_ptr<llm_graph_context> llama_model_jina_bert_v2::build_arch_graph(const llm_graph_params & params) const {
64    return std::make_unique<graph>(*this, params);
65}
66