From e57047827581af6cf09f1bca6e40c53937ebda62 Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Fri, 28 Nov 2025 13:05:11 +0800
Subject: [PATCH] limit cuda arches + scale tweaks

---
 CMakeLists.txt              | 2 ++
 Makefile                    | 6 ++++--
 ggml/src/ggml-cuda/fattn.cu | 4 ++++
 koboldcpp.py                | 6 ++++--
 4 files changed, 14 insertions(+), 4 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 4ae9e1640..1a9847f79 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -134,9 +134,11 @@ if (LLAMA_CUBLAS)
         message("CUDA Toolkit Version: ${CUDAToolkit_VERSION}")
         if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 13)
                 add_compile_definitions(GGML_CUDA_USE_GRAPHS) #try enable cuda graphs on cu12 build
+                add_compile_definitions(KCPP_LIMIT_CUDA_MAX_ARCH=860)
                 set(CMAKE_CUDA_ARCHITECTURES "75-virtual;80-virtual;86-virtual") # lowest CUDA 13 standard
         elseif(CUDAToolkit_VERSION VERSION_GREATER 12)
                 add_compile_definitions(GGML_CUDA_USE_GRAPHS) #try enable cuda graphs on cu12 build
+                add_compile_definitions(KCPP_LIMIT_CUDA_MAX_ARCH=800)
                 set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual") # lowest CUDA 12 standard + lowest for integer intrinsics
         else()
                 add_compile_definitions(KCPP_LIMIT_CUDA_MAX_ARCH=750) #will cause issues with ggml_cuda_highest_compiled_arch if removed
diff --git a/Makefile b/Makefile
index 5fad3d148..f083bea29 100644
--- a/Makefile
+++ b/Makefile
@@ -236,13 +236,15 @@ NVCCFLAGS += -Wno-deprecated-gpu-targets \
              -gencode arch=compute_61,code=compute_61 \
              -gencode arch=compute_70,code=compute_70 \
              -gencode arch=compute_75,code=compute_75 \
-             -gencode arch=compute_80,code=compute_80
+             -gencode arch=compute_80,code=compute_80 \
+			 -DKCPP_LIMIT_CUDA_MAX_ARCH=800
 
 else ifdef LLAMA_ARCHES_CU13
 NVCCFLAGS += -Wno-deprecated-gpu-targets \
              -gencode arch=compute_75,code=compute_75 \
              -gencode arch=compute_80,code=compute_80 \
-             -gencode arch=compute_86,code=compute_86
+             -gencode arch=compute_86,code=compute_86 \
+			 -DKCPP_LIMIT_CUDA_MAX_ARCH=860
 
 else
 NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=all
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 7e61ffa17..e0e4c91df 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -310,6 +310,10 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
             {
                 return BEST_FATTN_KERNEL_WMMA_F16;
             }
+            else
+            {
+                return BEST_FATTN_KERNEL_NONE;
+            }
         }
 
         return BEST_FATTN_KERNEL_MMA_F16;
diff --git a/koboldcpp.py b/koboldcpp.py
index 4b7c830cd..98c526195 100755
--- a/koboldcpp.py
+++ b/koboldcpp.py
@@ -5001,13 +5001,15 @@ def show_gui():
                     lastpos = ("+"+str(lparr[1])) if (len(lparr)==2) else ""
                     previous_event_width = new_width
                     previous_event_height = new_height
-                    windowwidth = math.floor(original_windowwidth*max(smallratio,min(incr_w,smallratio*1.2))) #allow slight extension past legal width
+                    windowwidth = math.floor(original_windowwidth*smallratio)
                     windowwidth = max(256, min(1024, windowwidth))
                     windowheight = math.floor(original_windowheight*smallratio)
                     windowheight = max(256, min(1024, windowheight))
                     root.geometry(str(windowwidth) + "x" + str(windowheight) + str(lastpos))
+                    if corrupt_scaler:
+                        smallratio = min(smallratio, 1)*0.98 #don't allow scaling beyond normal
                     ctk.set_widget_scaling(smallratio)
-                    root.after(20, clearesizing)
+                    root.after(5, clearesizing)
                     changerunmode(1,1,1)
                     togglerope(1,1,1)
                     toggleflashattn(1,1,1)