Merge branch 'concedo_experimental' into esocrok

Author: Nexesenex

Makefile

@@ -471,8 +471,13 @@ endif
 
 ifdef NO_VULKAN_EXTENSIONS
 VKGEN_NOEXT_ADD = -DNO_VULKAN_EXTENSIONS
+VKGEN_SUFFIX = -noext
+else
+VKGEN_SUFFIX =
 endif
 VKGEN_NOEXT_FORCE = -DNO_VULKAN_EXTENSIONS
+VKGEN_HPP = ggml/src/ggml-vulkan-shaders$(VKGEN_SUFFIX).hpp
+VKGEN_CPP = ggml/src/ggml-vulkan-shaders$(VKGEN_SUFFIX).cpp
 
 #
 # Print build information
@@ -685,9 +690,9 @@ ggml_v3-opencl.o: otherarch/ggml_v3-opencl.cpp otherarch/ggml_v3-opencl.h
 	$(CXX) $(CXXFLAGS) $(CLBLAST_FLAGS) -c $< -o $@
 
 #vulkan
-ggml-vulkan.o: ggml/src/ggml-vulkan/ggml-vulkan.cpp ggml/include/ggml-vulkan.h ggml/src/ggml-vulkan-shaders.cpp
+ggml-vulkan.o: ggml/src/ggml-vulkan/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(VKGEN_CPP)
 	$(CXX) $(CXXFLAGS) $(VKGEN_NOEXT_ADD) $(VULKAN_FLAGS) -c $< -o $@
-ggml-vulkan-shaders.o: ggml/src/ggml-vulkan-shaders.cpp ggml/include/ggml-vulkan.h
+ggml-vulkan-shaders.o: $(VKGEN_CPP) ggml/include/ggml-vulkan.h
 	$(CXX) $(CXXFLAGS) $(VKGEN_NOEXT_ADD) $(VULKAN_FLAGS) -c $< -o $@
 ggml-vulkan-noext.o: ggml/src/ggml-vulkan/ggml-vulkan.cpp ggml/include/ggml-vulkan.h ggml/src/ggml-vulkan-shaders-noext.cpp
 	$(CXX) $(CXXFLAGS) $(VKGEN_NOEXT_FORCE) $(VULKAN_FLAGS) -c $< -o $@
@@ -785,7 +790,7 @@ vulkan-shaders-gen: ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
 	$(CXX) $(CXXFLAGS) $(VKGEN_NOEXT_ADD) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 ifeq ($(OS),Windows_NT)
 	@echo 'Now rebuilding vulkan shaders for Windows...'
-	$(shell) vulkan-shaders-gen --glslc glslc --input-dir ggml/src/ggml-vulkan/vulkan-shaders --target-hpp ggml/src/ggml-vulkan-shaders.hpp --target-cpp ggml/src/ggml-vulkan-shaders.cpp --output-dir vulkan-spv-tmp
+	$(shell) vulkan-shaders-gen --glslc glslc --input-dir ggml/src/ggml-vulkan/vulkan-shaders --target-hpp $(VKGEN_HPP) --target-cpp $(VKGEN_CPP) --output-dir vulkan-spv-tmp
 	@echo 'Vulkan Shaders Rebuilt for Windows...'
 else
 	@echo 'Now rebuilding vulkan shaders for Linux...'
@@ -819,7 +824,7 @@ else
 		echo "Error: No usable glslc found. Vulkan shaders cannot be compiled!"; \
 	else \
 		echo "Using GLSLC: $$GLSLC_BIN"; \
-		./vulkan-shaders-gen --glslc "$$GLSLC_BIN" --input-dir ggml/src/ggml-vulkan/vulkan-shaders --target-hpp ggml/src/ggml-vulkan-shaders.hpp --target-cpp ggml/src/ggml-vulkan-shaders.cpp --output-dir vulkan-spv-tmp; \
+		./vulkan-shaders-gen --glslc "$$GLSLC_BIN" --input-dir ggml/src/ggml-vulkan/vulkan-shaders --target-hpp $(VKGEN_HPP) --target-cpp $(VKGEN_CPP) --output-dir vulkan-spv-tmp; \
 	fi
 	@echo 'Vulkan Shaders Rebuilt for Linux...'
 endif

koboldcpp.py

@@ -69,7 +69,7 @@
 extra_images_max = 4
 
 # global vars
-KcppVersion = "1.99.4"
+KcppVersion = "1.100"
 showdebug = True
 kcpp_instance = None #global running instance
 global_memory = {"tunnel_url": "", "restart_target":"", "input_to_exit":False, "load_complete":False, "restart_model": "", "currentConfig": None, "modelOverride": None, "currentModel": None}
@@ -1731,7 +1731,7 @@ def sd_load_model(model_filename,vae_filename,lora_filename,t5xxl_filename,clipl
     inputs.flash_attention = args.sdflashattention
     inputs.offload_cpu = args.sdoffloadcpu
     inputs.vae_cpu = args.sdvaecpu
-    inputs.clip_cpu = args.sdclipcpu
+    inputs.clip_cpu = False if args.sdclipgpu else True
     sdconvdirect = sd_convdirect_option(args.sdconvdirect)
     inputs.diffusion_conv_direct = sdconvdirect == 'full'
     inputs.vae_conv_direct = sdconvdirect in ['vaeonly', 'full']
@@ -5718,7 +5718,7 @@ def hide_tooltip(event):
     sd_flash_attention_var = ctk.IntVar(value=0)
     sd_offload_cpu_var = ctk.IntVar(value=0)
     sd_vae_cpu_var = ctk.IntVar(value=0)
-    sd_clip_cpu_var = ctk.IntVar(value=0)
+    sd_clip_gpu_var = ctk.IntVar(value=0)
     sd_vaeauto_var = ctk.IntVar(value=0)
     sd_tiled_vae_var = ctk.StringVar(value=str(default_vae_tile_threshold))
     sd_convdirect_var = ctk.StringVar(value=str(sd_convdirect_choices[0]))
@@ -6503,7 +6503,7 @@ def toggletaesd(a,b,c):
     makecheckbox(images_tab, "SD Flash Attention", sd_flash_attention_var, 44,padx=230, tooltiptxt="Enable Flash Attention for image diffusion. May save memory or improve performance.")
     makecheckbox(images_tab, "Model CPU Offload", sd_offload_cpu_var, 50,padx=8, tooltiptxt="Offload image weights in RAM to save VRAM, swap into VRAM when needed.")
     makecheckbox(images_tab, "VAE on CPU", sd_vae_cpu_var, 50,padx=160, tooltiptxt="Force VAE to CPU only for image generation.")
-    makecheckbox(images_tab, "CLIP on CPU", sd_clip_cpu_var, 50,padx=280, tooltiptxt="Force CLIP to CPU only for image generation.")
+    makecheckbox(images_tab, "CLIP on GPU", sd_clip_gpu_var, 50,padx=280, tooltiptxt="Put CLIP and T5 to GPU for image generation. Otherwise, CLIP will use CPU.")
 
     # audio tab
     audio_tab = tabcontent["Audio"]
@@ -6751,8 +6751,8 @@ def export_vars():
             args.sdoffloadcpu = True
         if sd_vae_cpu_var.get()==1:
             args.sdvaecpu = True
-        if sd_clip_cpu_var.get()==1:
-            args.sdclipcpu = True
+        if sd_clip_gpu_var.get()==1:
+            args.sdclipgpu = True
         args.sdthreads = (0 if sd_threads_var.get()=="" else int(sd_threads_var.get()))
         args.sdclamped = (0 if int(sd_clamped_var.get())<=0 else int(sd_clamped_var.get()))
         args.sdclampedsoft = (0 if int(sd_clamped_soft_var.get())<=0 else int(sd_clamped_soft_var.get()))
@@ -6997,7 +6997,7 @@ def import_vars(dict):
         sd_flash_attention_var.set(1 if ("sdflashattention" in dict and dict["sdflashattention"]) else 0)
         sd_offload_cpu_var.set(1 if ("sdoffloadcpu" in dict and dict["sdoffloadcpu"]) else 0)
         sd_vae_cpu_var.set(1 if ("sdvaecpu" in dict and dict["sdvaecpu"]) else 0)
-        sd_clip_cpu_var.set(1 if ("sdclipcpu" in dict and dict["sdclipcpu"]) else 0)
+        sd_clip_gpu_var.set(1 if ("sdclipgpu" in dict and dict["sdclipgpu"]) else 0)
         sd_convdirect_var.set(sd_convdirect_option(dict.get("sdconvdirect")))
         sd_vae_var.set(dict["sdvae"] if ("sdvae" in dict and dict["sdvae"]) else "")
         sd_t5xxl_var.set(dict["sdt5xxl"] if ("sdt5xxl" in dict and dict["sdt5xxl"]) else "")
@@ -8868,7 +8868,7 @@ def range_checker(arg: str):
     sdparsergroup.add_argument("--sdflashattention", help="Enables Flash Attention for image generation.", action='store_true')
     sdparsergroup.add_argument("--sdoffloadcpu", help="Offload image weights in RAM to save VRAM, swap into VRAM when needed.", action='store_true')
     sdparsergroup.add_argument("--sdvaecpu", help="Force VAE to CPU only for image generation.", action='store_true')
-    sdparsergroup.add_argument("--sdclipcpu", help="Force CLIP to CPU only for image generation.", action='store_true')
+    sdparsergroup.add_argument("--sdclipgpu", help="Put CLIP and T5 to GPU for image generation. Otherwise, CLIP will use CPU.", action='store_true')
     sdparsergroup.add_argument("--sdconvdirect", help="Enables Conv2D Direct. May improve performance or reduce memory usage. Might crash if not supported by the backend. Can be 'off' (default) to disable, 'full' to turn it on for all operations, or 'vaeonly' to enable only for the VAE.", type=sd_convdirect_option, choices=sd_convdirect_choices, default=sd_convdirect_choices[0])
     sdparsergroupvae = sdparsergroup.add_mutually_exclusive_group()
     sdparsergroupvae.add_argument("--sdvae", metavar=('[filename]'), help="Specify an image generation safetensors VAE which replaces the one in the model.", default="")