EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

kv-cache : fix SWA checks + disable cacheless iSWA (#15811)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-09-05 10:39:22 +03:00
committed by GitHub
parent 5d6688de08
commit c610b6c11b
9 changed files with 29 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/llama-model.cpp
View File

@@ -11084,7 +11084,8 @@ 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();
@@ -18632,7 +18633,7 @@ 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:
{
@@ -18681,6 +18682,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
/* attn_kv_size */ cparams.n_ctx,
/* attn_n_pad */ padding,
/* attn_n_swa */ hparams.n_swa,
/* attn_swa_type */ hparams.swa_type,
/* recurrent_type_k */ GGML_TYPE_F32,
/* recurrent_type_v */ GGML_TYPE_F32,
/* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
@@ -18750,6 +18752,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
cparams.n_seq_max,
padding,
hparams.n_swa,
hparams.swa_type,
nullptr,
nullptr);
}