no host by default

expose.h

@@ -51,6 +51,7 @@ struct load_model_inputs
     const int visionmaxtokens = -1;
     const bool use_mmap = false;
     const bool use_mlock = false;
+    const bool no_host = false;
     const bool use_mtp = false;
     const bool use_smartcontext = false;
     const bool use_contextshift = false;

gpttype_adapter.cpp

@@ -3231,6 +3231,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
         }
         model_params.load_mode = inputs.use_mlock ? LLAMA_LOAD_MODE_MLOCK : (inputs.use_mmap ? LLAMA_LOAD_MODE_MMAP : LLAMA_LOAD_MODE_NONE);
         model_params.n_gpu_layers = inputs.gpulayers;
+        model_params.no_host = inputs.no_host;
         kcpp_permit_any_repack = (inputs.use_mmap?false:true);
 
         //set device overrides if needed

koboldcpp.py

@@ -273,6 +273,7 @@ class load_model_inputs(ctypes.Structure):
                 ("visionmaxtokens", ctypes.c_int),
                 ("use_mmap", ctypes.c_bool),
                 ("use_mlock", ctypes.c_bool),
+                ("no_host", ctypes.c_bool),
                 ("use_mtp", ctypes.c_bool),
                 ("use_smartcontext", ctypes.c_bool),
                 ("use_contextshift", ctypes.c_bool),
@@ -1969,6 +1970,7 @@ def load_model(model_filename):
     inputs.blasthreads = args.blasthreads
     inputs.use_mmap = args.usemmap
     inputs.use_mlock = args.usemlock
+    inputs.no_host = True
     inputs.use_mtp = args.usemtp
     inputs.lora_filename = "".encode("UTF-8")
     inputs.lora_multiplier = args.loramult