char-fork/llama.cpp

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

Iggy Jacksonmodels: clean up some dead switch branches in old models (#28669)d344123fe

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
2.3 KiB51 linesraw
1#include "models.h"
2
3void llama_model_nomic_bert::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 12:
8            type = LLM_TYPE_137M; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_nomic_bert::load_arch_tensors(llama_model_loader &) {
14    LLAMA_LOAD_LOCALS;
15
16    if (n_token_types == 0) {
17        throw std::runtime_error(arch_name() + " model needs to define token type count");
18    }
19    tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
20    type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);
21
22    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
23    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
24
25    for (int i = 0; i < n_layer; ++i) {
26        auto & layer = layers[i];
27
28        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
29
30        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
31        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
32
33        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
34        layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);
35
36        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
37        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
38        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
39        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
40
41        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
42
43        layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
44        layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);
45    }
46}
47
48std::unique_ptr<llm_graph_context> llama_model_nomic_bert::build_arch_graph(const llm_graph_params & params) const {
49    return std::make_unique<graph>(*this, params);
50}
51