tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-29 04:35:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : add support for BertForSequenceClassification reranker (#13858)

Browse Source 
* convert: add support for BertForSequenceClassification

* add support for reranking using BertForSequenceClassification

* merge checks of eos and sep

* fix lint

---------

Co-authored-by: dinhhuy <huy.dinh@brains-tech.co.jp>
This commit is contained in:
Đinh Trọng Huy
2025-05-29 02:01:58 +09:00
committed by GitHub
parent aa6dff05be
commit e0e3aa231d
4 changed files with 42 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
src/llama-graph.cpp
View File

@ -1562,20 +1562,25 @@ void llm_graph_context::build_pooling(
ggml_tensor * inp_cls = build_inp_cls();
inp = ggml_get_rows(ctx0, inp, inp_cls);
// classification head
// https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
GGML_ASSERT(cls != nullptr);
GGML_ASSERT(cls_b != nullptr);
if (cls != nullptr && cls_b != nullptr) {
// classification head
// https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
cur = ggml_tanh(ctx0, cur);
cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
cur = ggml_tanh(ctx0, cur);
// some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
// https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
if (cls_out) {
// some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
// https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
if (cls_out) {
GGML_ASSERT(cls_out_b != nullptr);
cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
}
} else if (cls_out) {
// Single layer classification head (direct projection)
// https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
GGML_ASSERT(cls_out_b != nullptr);
cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, inp), cls_out_b);
} else {
GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
}
} break;
default: