diff --git a/CMakeLists.txt b/CMakeLists.txt
index c133be87b..3d21f1b56 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -106,7 +106,12 @@ if (LLAMA_CUBLAS)
if (LLAMA_CUDA_F16 OR LLAMA_CUDA_DMMV_F16)
set(CMAKE_CUDA_ARCHITECTURES "60;61;70") # needed for f16 CUDA intrinsics
else()
- set(CMAKE_CUDA_ARCHITECTURES "37;52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+ message("CUDA version: ${CUDA_VERSION_MAJOR}")
+ if(CUDA_VERSION_MAJOR GREATER_EQUAL 12)
+ set(CMAKE_CUDA_ARCHITECTURES "52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+ else()
+ set(CMAKE_CUDA_ARCHITECTURES "37;52;61;70") # lowest CUDA 12 standard + lowest for integer intrinsics
+ endif()
endif()
endif()
message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
diff --git a/klite.embd b/klite.embd
index 2ce9c166b..153b4af7d 100644
--- a/klite.embd
+++ b/klite.embd
@@ -25,7 +25,7 @@ Kobold Lite is under the AGPL v3.0 License for the purposes of koboldcpp. Please
-
diff --git a/koboldcpp.py b/koboldcpp.py
index ed50f3708..34a619cd7 100755
--- a/koboldcpp.py
+++ b/koboldcpp.py
@@ -1689,6 +1689,7 @@ def main(args):
if args.hordeconfig and len(args.hordeconfig)>4:
horde_thread = threading.Thread(target=run_horde_worker,args=(args,args.hordeconfig[3],args.hordeconfig[4]))
+ horde_thread.setDaemon(True)
horde_thread.start()
print(f"Please connect to custom endpoint at {epurl}")
diff --git a/llama.cpp b/llama.cpp
index e35af038f..61c1f7998 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -1081,7 +1081,7 @@ static void llama_model_load_internal(
// LLaMAv2
// TODO: temporary until GGUF
//patch for llama2 gqa
- if (model.type == e_model::MODEL_65B && hparams.n_mult == 4096) {
+ if (model.type == e_model::MODEL_65B && (hparams.n_mult >= 4096 && hparams.n_mult != 5504)) {
fprintf(stderr, "%s: Applying KCPP Patch for 70B model, setting GQA to 8\n", __func__);
n_gqa = 8;
}