vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-13 07:09:03 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

graph : support cacheless embeddings with FA and iSWA

Browse source
This commit is contained in:
Georgi Gerganov 2025-10-12 10:34:44 +03:00
parent 41aac5c69b
commit d4d465bce4
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
4 changed files with 108 additions and 51 deletions
Download patch file Download diff file Expand all files Collapse all files

5
src/llama-model.cpp
View file

@ -11376,8 +11376,7 @@ 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();
@ -19378,7 +19377,7 @@ 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: