From 287f3cdb53d4c663d3108c6ad6e4e45624c45104 Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Wed, 22 Jul 2026 20:49:41 +0800
Subject: [PATCH] fix device enumeration issues

---
 koboldcpp.py | 32 +++++++++++++++++++-------------
 1 file changed, 19 insertions(+), 13 deletions(-)

diff --git a/koboldcpp.py b/koboldcpp.py
index 285d8ff7b..362b280fb 100644
--- a/koboldcpp.py
+++ b/koboldcpp.py
@@ -1020,12 +1020,10 @@ def init_library():
     handle.set_environment_variable.restype = ctypes.c_int
     handle.set_environment_variable.argtypes = [ctypes.c_char_p, ctypes.c_char_p]
 
-def set_backend_props(inputs):
-    # we must force an explicit tensor split
-    # otherwise the default will divide equally and multigpu crap will slow it down badly
-    inputs.kcpp_main_gpu = -1
+def prepare_backend_visibility():
+    kcpp_main_gpu = -1
     if(args.maingpu is not None and args.maingpu>=0):
-        inputs.kcpp_main_gpu = args.maingpu
+        kcpp_main_gpu = args.maingpu
 
     if args.usecuda:
         os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
@@ -1033,35 +1031,42 @@ def set_backend_props(inputs):
         if (args.usecuda and "0" in args.usecuda):
             os.environ["CUDA_VISIBLE_DEVICES"] = "0"
             os.environ["HIP_VISIBLE_DEVICES"] = "0"
-            inputs.kcpp_main_gpu = 0
+            kcpp_main_gpu = 0
         elif (args.usecuda and "1" in args.usecuda):
             os.environ["CUDA_VISIBLE_DEVICES"] = "1"
             os.environ["HIP_VISIBLE_DEVICES"] = "1"
-            inputs.kcpp_main_gpu = 0
+            kcpp_main_gpu = 0
         elif (args.usecuda and "2" in args.usecuda):
             os.environ["CUDA_VISIBLE_DEVICES"] = "2"
             os.environ["HIP_VISIBLE_DEVICES"] = "2"
-            inputs.kcpp_main_gpu = 0
+            kcpp_main_gpu = 0
         elif (args.usecuda and "3" in args.usecuda):
             os.environ["CUDA_VISIBLE_DEVICES"] = "3"
             os.environ["HIP_VISIBLE_DEVICES"] = "3"
-            inputs.kcpp_main_gpu = 0
+            kcpp_main_gpu = 0
     else:
         if(args.maingpu is None or args.maingpu<0):
             if (args.usecuda and "0" in args.usecuda):
-                inputs.kcpp_main_gpu = 0
+                kcpp_main_gpu = 0
             elif (args.usecuda and "1" in args.usecuda):
-                inputs.kcpp_main_gpu = 1
+                kcpp_main_gpu = 1
             elif (args.usecuda and "2" in args.usecuda):
-                inputs.kcpp_main_gpu = 2
+                kcpp_main_gpu = 2
             elif (args.usecuda and "3" in args.usecuda):
-                inputs.kcpp_main_gpu = 3
+                kcpp_main_gpu = 3
 
     if "GGML_VK_VISIBLE_DEVICES" not in os.environ:
         if args.usevulkan: # is an empty array if using vulkan without defined gpu
             vulkangpus = ','.join([str(g) for g in args.usevulkan])
             handle.set_environment_variable("GGML_VK_VISIBLE_DEVICES".encode("UTF-8"),vulkangpus.encode("UTF-8"))
 
+    return kcpp_main_gpu
+
+def set_backend_props(inputs):
+    # we must force an explicit tensor split
+    # otherwise the default will divide equally and multigpu crap will slow it down badly
+    inputs.kcpp_main_gpu = prepare_backend_visibility()
+
     # set universal flags
     inputs.devices_override = (args.device if args.device else "").encode("UTF-8")
     inputs.quiet = args.quiet
@@ -11965,6 +11970,7 @@ def kcpp_main_process(launch_args, g_memory=None, gui_launcher=False):
             friendlysdmodelname = os.path.basename(imgmodel)
             friendlysdmodelname = os.path.splitext(friendlysdmodelname)[0]
             friendlysdmodelname = sanitize_string(friendlysdmodelname)
+            prepare_backend_visibility()
             cached_sd_info = sd_get_info()
             loadok = sd_load_model(imgmodel,imgvae,imgt5xxl,imgclip1,imgclip2,imgphotomaker,imgupscaler,imgaudiovae)
             print("Load Image Model OK: " + str(loadok))