ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llama : add RobertaForSequenceClassification reranker support (#13875)

Browse source
This commit is contained in:
Sigbjørn Skjæret 2025-05-29 08:15:01 +02:00 • committed by GitHub
parent 1b8fb8152d
commit 6385b843a8
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 24 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

2
src/llama-arch.cpp
View file

@ -174,6 +174,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_CONVNEXT_EMBEDDING_LENGTH, "%s.convnext.embedding_length" },
{ LLM_KV_CONVNEXT_BLOCK_COUNT, "%s.convnext.block_count" },
{ LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
{ LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" },
{ LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" },
{ LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" },

2
src/llama-arch.h
View file

@ -213,6 +213,8 @@ enum llm_kv {
LLM_KV_CONVNEXT_EMBEDDING_LENGTH,
LLM_KV_CONVNEXT_BLOCK_COUNT,
LLM_KV_CLASSIFIER_OUTPUT_LABELS,
// deprecated:
LLM_KV_TOKENIZER_PREFIX_ID,
LLM_KV_TOKENIZER_SUFFIX_ID,

3
src/llama-hparams.h
View file

@ -131,6 +131,9 @@ struct llama_hparams {
bool attn_soft_cap = false;
bool use_kq_norm = true;
// for Classifiers
uint32_t n_cls_out = 1;
// llama4
uint32_t n_moe_layer_step = 0;
uint32_t n_no_rope_layer_step = 4;

5
src/llama-model.cpp
View file

@ -683,6 +683,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn);
ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false);
ml.get_arr_n(LLM_KV_CLASSIFIER_OUTPUT_LABELS, hparams.n_cls_out, false);
switch (hparams.n_layer) {
case 3:
@ -2121,8 +2122,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
cls = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"), {n_embd}, TENSOR_NOT_REQUIRED);
cls_out = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, 1}, TENSOR_NOT_REQUIRED);
cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"), {1}, TENSOR_NOT_REQUIRED);
cls_out = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"), {hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
}
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);