allow mmproj to be run on cpu

expose.h

@@ -45,6 +45,7 @@ struct load_model_inputs
     const int draft_gpulayers = 999;
     const float draft_gpusplit[tensor_split_max] = {};
     const char * mmproj_filename = nullptr;
+    const bool mmproj_cpu = false;
     const int visionmaxres = 2048;
     const bool use_mmap = false;
     const bool use_mlock = false;

gpttype_adapter.cpp

@@ -2320,6 +2320,11 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
                 printf("Clip will use CPU for this model!\n");
             }
             #endif
+            if(inputs.mmproj_cpu)
+            {
+                set_clip_uses_gpu(false);
+                printf("Clip forced to use CPU!\n");
+            }
             clp_ctx = clip_model_load(mmproj_filename.c_str(), /*verbosity=*/ 1);
             if(clp_ctx == nullptr) {
                 fprintf(stderr, "%s: error: failed to load mmproj model!\n", __func__);

koboldcpp.py

@@ -165,6 +165,7 @@ class load_model_inputs(ctypes.Structure):
                 ("draft_gpulayers", ctypes.c_int),
                 ("draft_gpusplit", ctypes.c_float * tensor_split_max),
                 ("mmproj_filename", ctypes.c_char_p),
+                ("mmproj_cpu", ctypes.c_bool),
                 ("visionmaxres", ctypes.c_int),
                 ("use_mmap", ctypes.c_bool),
                 ("use_mlock", ctypes.c_bool),
@@ -1184,6 +1185,7 @@ def load_model(model_filename):
         else:
             inputs.draft_gpusplit[n] = 0
     inputs.mmproj_filename = args.mmproj.encode("UTF-8") if args.mmproj else "".encode("UTF-8")
+    inputs.mmproj_cpu = (True if args.mmprojcpu else False)
     inputs.visionmaxres = (512 if args.visionmaxres < 512 else (2048 if args.visionmaxres > 2048 else args.visionmaxres))
     inputs.use_smartcontext = args.smartcontext
     inputs.use_contextshift = (0 if args.noshift else 1)
@@ -3879,6 +3881,7 @@ def show_gui():
     preloadstory_var = ctk.StringVar()
     savedatafile_var = ctk.StringVar()
     mmproj_var = ctk.StringVar()
+    mmprojcpu_var = ctk.IntVar(value=0)
     visionmaxres_var = ctk.StringVar(value=str(default_visionmaxres))
     draftmodel_var = ctk.StringVar()
     draftamount_var = ctk.StringVar(value=str(default_draft_amount))
@@ -4408,7 +4411,8 @@ def show_gui():
     makefileentry(model_tab, "Text Lora:", "Select Lora File",lora_var, 3,width=280,singlerow=True,tooltiptxt="Select an optional GGML Text LoRA adapter to use.\nLeave blank to skip.")
     makefileentry(model_tab, "Lora Base:", "Select Lora Base File", lora_base_var, 5,width=280,singlerow=True,tooltiptxt="Select an optional F16 GGML Text LoRA base file to use.\nLeave blank to skip.")
     makefileentry(model_tab, "Vision mmproj:", "Select Vision mmproj File", mmproj_var, 7,width=280,singlerow=True,tooltiptxt="Select a mmproj file to use for vision models like LLaVA.\nLeave blank to skip.")
-    makelabelentry(model_tab, "Vision MaxRes:", visionmaxres_var, 9, padx=100, singleline=True, tooltip=f"Clamp MMProj vision maximum allowed resolution. Allowed values are between 512 to 2048 px (default {default_visionmaxres}).")
+    makecheckbox(model_tab, "Vision Force CPU", mmprojcpu_var, 9, tooltiptxt="Force CLIP for Vision mmproj always on CPU.")
+    makelabelentry(model_tab, "Vision MaxRes:", visionmaxres_var, 9, padx=320, singleline=True, tooltip=f"Clamp MMProj vision maximum allowed resolution. Allowed values are between 512 to 2048 px (default {default_visionmaxres}).", labelpadx=220)
     makefileentry(model_tab, "Draft Model:", "Select Speculative Text Model File", draftmodel_var, 11,width=280,singlerow=True,tooltiptxt="Select a draft text model file to use for speculative decoding.\nLeave blank to skip.")
     makelabelentry(model_tab, "Draft Amount: ", draftamount_var, 13, 50,padx=100,singleline=True,tooltip="How many tokens to draft per chunk before verifying results")
     makelabelentry(model_tab, "Splits: ", draftgpusplit_str_vars, 13, 50,padx=210,singleline=True,tooltip="Distribution of draft model layers. Leave blank to follow main model's gpu split. Only works if multi-gpu (All) selected in main model.", labelpadx=160)
@@ -4694,6 +4698,7 @@ def show_gui():
         except Exception:
             pass
         args.mmproj = None if mmproj_var.get() == "" else mmproj_var.get()
+        args.mmprojcpu = (mmprojcpu_var.get()==1)
         args.visionmaxres = int(visionmaxres_var.get()) if visionmaxres_var.get()!="" else default_visionmaxres
         args.draftmodel = None if draftmodel_var.get() == "" else draftmodel_var.get()
         args.draftamount = int(draftamount_var.get()) if draftamount_var.get()!="" else default_draft_amount
@@ -4886,6 +4891,7 @@ def show_gui():
                 lora_var.set(dict["lora"][0])
 
         mmproj_var.set(dict["mmproj"] if ("mmproj" in dict and dict["mmproj"]) else "")
+        mmprojcpu_var.set(1 if ("mmprojcpu" in dict and dict["mmprojcpu"]) else 0)
         if "visionmaxres" in dict and dict["visionmaxres"]:
             visionmaxres_var.set(dict["visionmaxres"])
         draftmodel_var.set(dict["draftmodel"] if ("draftmodel" in dict and dict["draftmodel"]) else "")
@@ -6575,6 +6581,7 @@ if __name__ == '__main__':
     advparser.add_argument("--ssl", help="Allows all content to be served over SSL instead. A valid UNENCRYPTED SSL cert and key .pem files must be provided", metavar=('[cert_pem]', '[key_pem]'), nargs='+')
     advparser.add_argument("--nocertify", help="Allows insecure SSL connections. Use this if you have cert errors and need to bypass certificate restrictions.", action='store_true')
     advparser.add_argument("--mmproj", metavar=('[filename]'), help="Select a multimodal projector file for vision models like LLaVA.", default="")
+    advparser.add_argument("--mmprojcpu", help="Force CLIP for Vision mmproj always on CPU.", action='store_true')
     advparser.add_argument("--visionmaxres", metavar=('[max px]'), help="Clamp MMProj vision maximum allowed resolution. Allowed values are between 512 to 2048 px (default 1024).", type=int, default=default_visionmaxres)
     advparser.add_argument("--draftmodel", metavar=('[filename]'), help="Load a small draft model for speculative decoding. It will be fully offloaded. Vocab must match the main model.", default="")
     advparser.add_argument("--draftamount", metavar=('[tokens]'), help="How many tokens to draft per chunk before verifying results", type=int, default=default_draft_amount)