ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-02 01:50:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Offload only activated experts to the GPU (#698)

Browse Source 
* Offload only activated experts

* This seems to do the trick for -fmoe

* Do not recalculate activated expers for fused up/gate

* Log out of bounds access details

* Add a command line argument

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-09-04 12:22:30 +02:00
committed by GitHub
parent 144d456717
commit 13c3b6412e
8 changed files with 155 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
common/common.cpp
View File

@@ -1378,6 +1378,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
}
return true;
}
if (arg == "--offload-only-active-experts" || arg == "-ooae") {
params.only_active_exps = true;
return true;
}
if (arg == "--host") {
CHECK_ARG
params.hostname = argv[i];
@@ -2746,6 +2750,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.fused_up_gate = params.fused_up_gate;
cparams.min_experts = params.min_experts;
cparams.thresh_experts = params.thresh_experts;
cparams.only_active_experts = params.only_active_exps;
cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
cparams.type_v = kv_cache_type_from_str(params.cache_type_v);