Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit 'e38b7c6e9e4453e3b3e96d76e38bc2ccb6bce458' into concedo_experimental

Browse Source
This commit is contained in:
Concedo
2025-10-16 12:08:03 +08:00
parent f3b0ed157b e38b7c6e9e
commit 2cee3b2055
4 changed files with 87 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+5 -6
View File
@@ -11420,8 +11420,8 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
}
};
struct llm_build_gemma_embedding_iswa : public llm_graph_context {
llm_build_gemma_embedding_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
struct llm_build_gemma_embedding : public llm_graph_context {
llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_k;
ggml_tensor * cur;
@@ -11438,8 +11438,7 @@ struct llm_build_gemma_embedding_iswa : public llm_graph_context {
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
// TODO: support cacheless iSWA embeddings [TAG_NO_CACHE_ISWA]
auto * inp_attn = build_attn_inp_kv_iswa();
auto * inp_attn = build_attn_inp_no_cache();
ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -19440,7 +19439,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
case LLM_ARCH_NOMIC_BERT_MOE:
case LLM_ARCH_NEO_BERT:
case LLM_ARCH_WAVTOKENIZER_DEC:
//case LLM_ARCH_GEMMA_EMBEDDING: // TODO: disabled until the cacheless SWA logic is fixed [TAG_NO_CACHE_ISWA]
case LLM_ARCH_GEMMA_EMBEDDING:
case LLM_ARCH_DREAM:
case LLM_ARCH_LLADA:
case LLM_ARCH_LLADA_MOE:
@@ -19733,7 +19732,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
} break;
case LLM_ARCH_GEMMA_EMBEDDING:
{
llm = std::make_unique<llm_build_gemma_embedding_iswa>(*this, params);
llm = std::make_unique<llm_build_gemma_embedding>(*this, params);
} break;
case LLM_ARCH_STARCODER2:
{