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

Revert "graph : support cacheless embeddings with FA and iSWA"

Browse Source 
This reverts commit d4d465bce4.
This commit is contained in:
Concedo
2025-10-16 12:07:48 +08:00
parent 1ff97f8a00
commit f3b0ed157b
4 changed files with 51 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+3 -2
View File
@@ -11438,7 +11438,8 @@ struct llm_build_gemma_embedding_iswa : public llm_graph_context {
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_no_cache();
// TODO: support cacheless iSWA embeddings [TAG_NO_CACHE_ISWA]
auto * inp_attn = build_attn_inp_kv_iswa();
ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -19439,7 +19440,7 @@ 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:
//case LLM_ARCH_GEMMA_EMBEDDING: // TODO: disabled until the cacheless SWA logic is fixed [TAG_NO_CACHE_ISWA]
case LLM_ARCH_DREAM:
case LLM_ARCH_LLADA:
case LLM_ARCH_LLADA_MOE: