ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-02 10:00:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Hadamard transforms for K-cache - CPU only (#1033)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-12-04 06:51:11 +01:00
committed by GitHub
parent 0581f90c0f
commit 18fdd80eaf
13 changed files with 155 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/llama.cpp
View File

@@ -4048,6 +4048,7 @@ struct llama_context_params llama_context_default_params() {
/*.min_experts =*/ -1,
/*.thtesh_experts =*/ 0.0f,
/*.only_active_experts =*/ false,
/*.k_cache_hadamard =*/ false,
/*.abort_callback =*/ nullptr,
/*.abort_callback_data =*/ nullptr,
/*.offload_policy =*/ nullptr,
@@ -4297,6 +4298,11 @@ struct llama_context * llama_new_context_with_model(
return nullptr;
}
if (params.k_cache_hadamard && !ggml_is_quantized(params.type_k)) {
LLAMA_LOG_WARN("%s: there is no point in Hadamard transforms with not quantized K-cache. Turning Hadamard off\n", __func__);
params.k_cache_hadamard = false;
}
llama_context * ctx = new llama_context(*model);
// add devices to ctx->cparams from model
@@ -4330,6 +4336,7 @@ struct llama_context * llama_new_context_with_model(
cparams.fused_mmad = params.fused_mmad;
cparams.rope_cache = params.rope_cache;
cparams.graph_reuse = params.graph_reuse;
cparams.k_cache_hadamard = params.k_cache_hadamard;
cparams.min_experts = params.min_experts;
cparams.thresh_experts = params.thresh_experts;
cparams.cuda_params = params.cuda_params;
@@ -4417,6 +4424,7 @@ struct llama_context * llama_new_context_with_model(
LLAMA_LOG_INFO("%s: fused_mmad = %d\n", __func__, cparams.fused_mmad);
LLAMA_LOG_INFO("%s: rope_cache = %d\n", __func__, cparams.rope_cache);
LLAMA_LOG_INFO("%s: graph_reuse = %d\n", __func__, cparams.graph_reuse);
LLAMA_LOG_INFO("%s: k_cache_hadam = %d\n", __func__, cparams.k_cache_hadamard);
LLAMA_LOG_INFO("%s: ser = %d, %g\n", __func__, cparams.min_experts, cparams.thresh_experts);
LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base);
LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale);